xfs: rearrange some code in xfs_bmap for better locality
authorDave Chinner <dchinner@redhat.com>
Mon, 25 Feb 2013 01:31:26 +0000 (12:31 +1100)
committerBen Myers <bpm@sgi.com>
Thu, 7 Mar 2013 18:35:22 +0000 (12:35 -0600)
xfs_bmap.c is a big file, and some of the related code is spread all
throughout the file requiring function prototypes for static
function and jumping all through the file to follow a single call
path. Rearrange the code so that:

a) related functionality is grouped together; and
b) functions are grouped in call dependency order

While the diffstat is large, there are no code changes in the patch;
it is just moving the functionality around and removing the function
prototypes at the top of the file. The resulting layout of the code
is as follows (top of file to bottom):

- miscellaneous helper functions
- extent tree block counting routines
- debug/sanity checking code
- bmap free list manipulation functions
- inode fork format manipulation functions
- internal/external extent tree seach functions
- extent tree manipulation functions used during allocation
- functions used during extent read/allocate/removal
  operations (i.e. xfs_bmapi_write, xfs_bmapi_read,
  xfs_bunmapi and xfs_getbmap)

This means that following logic paths through the bmapi code is much
simpler - most of the code relevant to a specific operation is now
clustered together rather than spread all over the file....

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
fs/xfs/xfs_bmap.c

index b44af9211bd99027930215a316f3044dc79f1bf7..d490fe8fd19f39e6da4ac1e32e24aed393d3ca85 100644 (file)
 kmem_zone_t            *xfs_bmap_free_item_zone;
 
 /*
- * Prototypes for internal bmap routines.
+ * Miscellaneous helper functions
  */
 
-#ifdef DEBUG
-STATIC void
-xfs_bmap_check_leaf_extents(
-       struct xfs_btree_cur    *cur,
-       struct xfs_inode        *ip,
-       int                     whichfork);
-#else
-#define xfs_bmap_check_leaf_extents(cur, ip, whichfork)                do { } while (0)
-#endif
-
-
-/*
- * Called from xfs_bmap_add_attrfork to handle extents format files.
- */
-STATIC int                                     /* error */
-xfs_bmap_add_attrfork_extents(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first block allocated */
-       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
-       int                     *flags);        /* inode logging flags */
-
-/*
- * Called from xfs_bmap_add_attrfork to handle local format files.
- */
-STATIC int                                     /* error */
-xfs_bmap_add_attrfork_local(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first block allocated */
-       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
-       int                     *flags);        /* inode logging flags */
-
-/*
- * xfs_bmap_alloc is called by xfs_bmapi to allocate an extent for a file.
- * It figures out where to ask the underlying allocator to put the new extent.
- */
-STATIC int                             /* error */
-xfs_bmap_alloc(
-       xfs_bmalloca_t          *ap);   /* bmap alloc argument struct */
-
-/*
- * Transform a btree format file with only one leaf node, where the
- * extents list will fit in the inode, into an extents format file.
- * Since the file extents are already in-core, all we have to do is
- * give up the space for the btree root and pitch the leaf block.
- */
-STATIC int                             /* error */
-xfs_bmap_btree_to_extents(
-       xfs_trans_t             *tp,    /* transaction pointer */
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_btree_cur_t         *cur,   /* btree cursor */
-       int                     *logflagsp, /* inode logging flags */
-       int                     whichfork); /* data or attr fork */
-
-/*
- * Remove the entry "free" from the free item list.  Prev points to the
- * previous entry, unless "free" is the head of the list.
- */
-STATIC void
-xfs_bmap_del_free(
-       xfs_bmap_free_t         *flist, /* free item list header */
-       xfs_bmap_free_item_t    *prev,  /* previous item on list, if any */
-       xfs_bmap_free_item_t    *free); /* list item to be freed */
-
-/*
- * Convert an extents-format file into a btree-format file.
- * The new file will have a root block (in the inode) and a single child block.
- */
-STATIC int                                     /* error */
-xfs_bmap_extents_to_btree(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first-block-allocated */
-       xfs_bmap_free_t         *flist,         /* blocks freed in xaction */
-       xfs_btree_cur_t         **curp,         /* cursor returned to caller */
-       int                     wasdel,         /* converting a delayed alloc */
-       int                     *logflagsp,     /* inode logging flags */
-       int                     whichfork);     /* data or attr fork */
-
-/*
- * Convert a local file to an extents file.
- * This code is sort of bogus, since the file data needs to get
- * logged so it won't be lost.  The bmap-level manipulations are ok, though.
- */
-STATIC int                             /* error */
-xfs_bmap_local_to_extents(
-       xfs_trans_t     *tp,            /* transaction pointer */
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_fsblock_t   *firstblock,    /* first block allocated in xaction */
-       xfs_extlen_t    total,          /* total blocks needed by transaction */
-       int             *logflagsp,     /* inode logging flags */
-       int             whichfork,      /* data or attr fork */
-       void            (*init_fn)(struct xfs_buf *bp,
-                                  struct xfs_inode *ip,
-                                  struct xfs_ifork *ifp));
-
-/*
- * Search the extents list for the inode, for the extent containing bno.
- * If bno lies in a hole, point to the next entry.  If bno lies past eof,
- * *eofp will be set, and *prevp will contain the last entry (null if none).
- * Else, *lastxp will be set to the index of the found
- * entry; *gotp will contain the entry.
- */
-STATIC xfs_bmbt_rec_host_t *           /* pointer to found extent entry */
-xfs_bmap_search_extents(
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_fileoff_t   bno,            /* block number searched for */
-       int             whichfork,      /* data or attr fork */
-       int             *eofp,          /* out: end of file found */
-       xfs_extnum_t    *lastxp,        /* out: last extent index */
-       xfs_bmbt_irec_t *gotp,          /* out: extent entry found */
-       xfs_bmbt_irec_t *prevp);        /* out: previous extent entry found */
-
-/*
- * Compute the worst-case number of indirect blocks that will be used
- * for ip's delayed extent of length "len".
- */
-STATIC xfs_filblks_t
-xfs_bmap_worst_indlen(
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_filblks_t           len);   /* delayed extent length */
-
-#ifdef DEBUG
 /*
- * Perform various validation checks on the values being returned
- * from xfs_bmapi().
+ * Compute and fill in the value of the maximum depth of a bmap btree
+ * in this filesystem.  Done once, during mount.
  */
-STATIC void
-xfs_bmap_validate_ret(
-       xfs_fileoff_t           bno,
-       xfs_filblks_t           len,
-       int                     flags,
-       xfs_bmbt_irec_t         *mval,
-       int                     nmap,
-       int                     ret_nmap);
-#else
-#define        xfs_bmap_validate_ret(bno,len,flags,mval,onmap,nmap)
-#endif /* DEBUG */
-
-STATIC int
-xfs_bmap_count_tree(
-       xfs_mount_t     *mp,
-       xfs_trans_t     *tp,
-       xfs_ifork_t     *ifp,
-       xfs_fsblock_t   blockno,
-       int             levelin,
-       int             *count);
-
-STATIC void
-xfs_bmap_count_leaves(
-       xfs_ifork_t             *ifp,
-       xfs_extnum_t            idx,
-       int                     numrecs,
-       int                     *count);
+void
+xfs_bmap_compute_maxlevels(
+       xfs_mount_t     *mp,            /* file system mount structure */
+       int             whichfork)      /* data or attr fork */
+{
+       int             level;          /* btree level */
+       uint            maxblocks;      /* max blocks at this level */
+       uint            maxleafents;    /* max leaf entries possible */
+       int             maxrootrecs;    /* max records in root block */
+       int             minleafrecs;    /* min records in leaf block */
+       int             minnoderecs;    /* min records in node block */
+       int             sz;             /* root block size */
 
-STATIC void
-xfs_bmap_disk_count_leaves(
-       struct xfs_mount        *mp,
-       struct xfs_btree_block  *block,
-       int                     numrecs,
-       int                     *count);
+       /*
+        * The maximum number of extents in a file, hence the maximum
+        * number of leaf entries, is controlled by the type of di_nextents
+        * (a signed 32-bit number, xfs_extnum_t), or by di_anextents
+        * (a signed 16-bit number, xfs_aextnum_t).
+        *
+        * Note that we can no longer assume that if we are in ATTR1 that
+        * the fork offset of all the inodes will be
+        * (xfs_default_attroffset(ip) >> 3) because we could have mounted
+        * with ATTR2 and then mounted back with ATTR1, keeping the
+        * di_forkoff's fixed but probably at various positions. Therefore,
+        * for both ATTR1 and ATTR2 we have to assume the worst case scenario
+        * of a minimum size available.
+        */
+       if (whichfork == XFS_DATA_FORK) {
+               maxleafents = MAXEXTNUM;
+               sz = XFS_BMDR_SPACE_CALC(MINDBTPTRS);
+       } else {
+               maxleafents = MAXAEXTNUM;
+               sz = XFS_BMDR_SPACE_CALC(MINABTPTRS);
+       }
+       maxrootrecs = xfs_bmdr_maxrecs(mp, sz, 0);
+       minleafrecs = mp->m_bmap_dmnr[0];
+       minnoderecs = mp->m_bmap_dmnr[1];
+       maxblocks = (maxleafents + minleafrecs - 1) / minleafrecs;
+       for (level = 1; maxblocks > 1; level++) {
+               if (maxblocks <= maxrootrecs)
+                       maxblocks = 1;
+               else
+                       maxblocks = (maxblocks + minnoderecs - 1) / minnoderecs;
+       }
+       mp->m_bm_maxlevels[whichfork] = level;
+}
 
 /*
- * Bmap internal routines.
+ * Convert the given file system block to a disk block.  We have to treat it
+ * differently based on whether the file is a real time file or not, because the
+ * bmap code does.
  */
+xfs_daddr_t
+xfs_fsb_to_db(struct xfs_inode *ip, xfs_fsblock_t fsb)
+{
+       return (XFS_IS_REALTIME_INODE(ip) ? \
+                (xfs_daddr_t)XFS_FSB_TO_BB((ip)->i_mount, (fsb)) : \
+                XFS_FSB_TO_DADDR((ip)->i_mount, (fsb)));
+}
 
 STATIC int                             /* error */
 xfs_bmbt_lookup_eq(
@@ -290,5935 +187,5914 @@ xfs_bmbt_update(
 }
 
 /*
- * Called from xfs_bmap_add_attrfork to handle btree format files.
+ * Compute the worst-case number of indirect blocks that will be used
+ * for ip's delayed extent of length "len".
  */
-STATIC int                                     /* error */
-xfs_bmap_add_attrfork_btree(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first block allocated */
-       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
-       int                     *flags)         /* inode logging flags */
+STATIC xfs_filblks_t
+xfs_bmap_worst_indlen(
+       xfs_inode_t     *ip,            /* incore inode pointer */
+       xfs_filblks_t   len)            /* delayed extent length */
 {
-       xfs_btree_cur_t         *cur;           /* btree cursor */
-       int                     error;          /* error return value */
-       xfs_mount_t             *mp;            /* file system mount struct */
-       int                     stat;           /* newroot status */
+       int             level;          /* btree level number */
+       int             maxrecs;        /* maximum record count at this level */
+       xfs_mount_t     *mp;            /* mount structure */
+       xfs_filblks_t   rval;           /* return value */
 
        mp = ip->i_mount;
-       if (ip->i_df.if_broot_bytes <= XFS_IFORK_DSIZE(ip))
-               *flags |= XFS_ILOG_DBROOT;
-       else {
-               cur = xfs_bmbt_init_cursor(mp, tp, ip, XFS_DATA_FORK);
-               cur->bc_private.b.flist = flist;
-               cur->bc_private.b.firstblock = *firstblock;
-               if ((error = xfs_bmbt_lookup_ge(cur, 0, 0, 0, &stat)))
-                       goto error0;
-               /* must be at least one entry */
-               XFS_WANT_CORRUPTED_GOTO(stat == 1, error0);
-               if ((error = xfs_btree_new_iroot(cur, flags, &stat)))
-                       goto error0;
-               if (stat == 0) {
-                       xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
-                       return XFS_ERROR(ENOSPC);
-               }
-               *firstblock = cur->bc_private.b.firstblock;
-               cur->bc_private.b.allocated = 0;
-               xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
+       maxrecs = mp->m_bmap_dmxr[0];
+       for (level = 0, rval = 0;
+            level < XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK);
+            level++) {
+               len += maxrecs - 1;
+               do_div(len, maxrecs);
+               rval += len;
+               if (len == 1)
+                       return rval + XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK) -
+                               level - 1;
+               if (level == 0)
+                       maxrecs = mp->m_bmap_dmxr[1];
        }
-       return 0;
-error0:
-       xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
-       return error;
+       return rval;
 }
 
 /*
- * Called from xfs_bmap_add_attrfork to handle extents format files.
+ * Calculate the default attribute fork offset for newly created inodes.
  */
-STATIC int                                     /* error */
-xfs_bmap_add_attrfork_extents(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first block allocated */
-       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
-       int                     *flags)         /* inode logging flags */
+uint
+xfs_default_attroffset(
+       struct xfs_inode        *ip)
 {
-       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
-       int                     error;          /* error return value */
+       struct xfs_mount        *mp = ip->i_mount;
+       uint                    offset;
 
-       if (ip->i_d.di_nextents * sizeof(xfs_bmbt_rec_t) <= XFS_IFORK_DSIZE(ip))
-               return 0;
-       cur = NULL;
-       error = xfs_bmap_extents_to_btree(tp, ip, firstblock, flist, &cur, 0,
-               flags, XFS_DATA_FORK);
-       if (cur) {
-               cur->bc_private.b.allocated = 0;
-               xfs_btree_del_cursor(cur,
-                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       if (mp->m_sb.sb_inodesize == 256) {
+               offset = XFS_LITINO(mp) -
+                               XFS_BMDR_SPACE_CALC(MINABTPTRS);
+       } else {
+               offset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
        }
-       return error;
+
+       ASSERT(offset < XFS_LITINO(mp));
+       return offset;
 }
 
 /*
- * Block initialisation functions for local to extent format conversion.
- * As these get more complex, they will be moved to the relevant files,
- * but for now they are too simple to worry about.
+ * Helper routine to reset inode di_forkoff field when switching
+ * attribute fork from local to extent format - we reset it where
+ * possible to make space available for inline data fork extents.
  */
 STATIC void
-xfs_bmap_local_to_extents_init_fn(
-       struct xfs_buf          *bp,
-       struct xfs_inode        *ip,
-       struct xfs_ifork        *ifp)
+xfs_bmap_forkoff_reset(
+       xfs_mount_t     *mp,
+       xfs_inode_t     *ip,
+       int             whichfork)
 {
-       bp->b_ops = &xfs_bmbt_buf_ops;
-       memcpy(bp->b_addr, ifp->if_u1.if_data, ifp->if_bytes);
+       if (whichfork == XFS_ATTR_FORK &&
+           ip->i_d.di_format != XFS_DINODE_FMT_DEV &&
+           ip->i_d.di_format != XFS_DINODE_FMT_UUID &&
+           ip->i_d.di_format != XFS_DINODE_FMT_BTREE) {
+               uint    dfl_forkoff = xfs_default_attroffset(ip) >> 3;
+
+               if (dfl_forkoff > ip->i_d.di_forkoff)
+                       ip->i_d.di_forkoff = dfl_forkoff;
+       }
 }
 
+/*
+ * Extent tree block counting routines.
+ */
+
+/*
+ * Count leaf blocks given a range of extent records.
+ */
 STATIC void
-xfs_symlink_local_to_remote(
-       struct xfs_buf          *bp,
-       struct xfs_inode        *ip,
-       struct xfs_ifork        *ifp)
+xfs_bmap_count_leaves(
+       xfs_ifork_t             *ifp,
+       xfs_extnum_t            idx,
+       int                     numrecs,
+       int                     *count)
 {
-       /* remote symlink blocks are not verifiable until CRCs come along */
-       bp->b_ops = NULL;
-       memcpy(bp->b_addr, ifp->if_u1.if_data, ifp->if_bytes);
+       int             b;
+
+       for (b = 0; b < numrecs; b++) {
+               xfs_bmbt_rec_host_t *frp = xfs_iext_get_ext(ifp, idx + b);
+               *count += xfs_bmbt_get_blockcount(frp);
+       }
 }
 
 /*
- * Called from xfs_bmap_add_attrfork to handle local format files. Each
- * different data fork content type needs a different callout to do the
- * conversion. Some are basic and only require special block initialisation
- * callouts for the data formating, others (directories) are so specialised they
- * handle everything themselves.
- *
- * XXX (dgc): investigate whether directory conversion can use the generic
- * formatting callout. It should be possible - it's just a very complex
- * formatter. it would also require passing the transaction through to the init
- * function.
+ * Count leaf blocks given a range of extent records originally
+ * in btree format.
  */
-STATIC int                                     /* error */
-xfs_bmap_add_attrfork_local(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first block allocated */
-       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
-       int                     *flags)         /* inode logging flags */
+STATIC void
+xfs_bmap_disk_count_leaves(
+       struct xfs_mount        *mp,
+       struct xfs_btree_block  *block,
+       int                     numrecs,
+       int                     *count)
 {
-       xfs_da_args_t           dargs;          /* args for dir/attr code */
-
-       if (ip->i_df.if_bytes <= XFS_IFORK_DSIZE(ip))
-               return 0;
+       int             b;
+       xfs_bmbt_rec_t  *frp;
 
-       if (S_ISDIR(ip->i_d.di_mode)) {
-               memset(&dargs, 0, sizeof(dargs));
-               dargs.dp = ip;
-               dargs.firstblock = firstblock;
-               dargs.flist = flist;
-               dargs.total = ip->i_mount->m_dirblkfsbs;
-               dargs.whichfork = XFS_DATA_FORK;
-               dargs.trans = tp;
-               return xfs_dir2_sf_to_block(&dargs);
+       for (b = 1; b <= numrecs; b++) {
+               frp = XFS_BMBT_REC_ADDR(mp, block, b);
+               *count += xfs_bmbt_disk_get_blockcount(frp);
        }
-
-       if (S_ISLNK(ip->i_d.di_mode))
-               return xfs_bmap_local_to_extents(tp, ip, firstblock, 1,
-                                                flags, XFS_DATA_FORK,
-                                                xfs_symlink_local_to_remote);
-
-       return xfs_bmap_local_to_extents(tp, ip, firstblock, 1, flags,
-                                        XFS_DATA_FORK,
-                                        xfs_bmap_local_to_extents_init_fn);
 }
 
 /*
- * Convert a delayed allocation to a real allocation.
+ * Recursively walks each level of a btree
+ * to count total fsblocks is use.
  */
-STATIC int                             /* error */
-xfs_bmap_add_extent_delay_real(
-       struct xfs_bmalloca     *bma)
+STATIC int                                     /* error */
+xfs_bmap_count_tree(
+       xfs_mount_t     *mp,            /* file system mount point */
+       xfs_trans_t     *tp,            /* transaction pointer */
+       xfs_ifork_t     *ifp,           /* inode fork pointer */
+       xfs_fsblock_t   blockno,        /* file system block number */
+       int             levelin,        /* level in btree */
+       int             *count)         /* Count of blocks */
 {
-       struct xfs_bmbt_irec    *new = &bma->got;
-       int                     diff;   /* temp value */
-       xfs_bmbt_rec_host_t     *ep;    /* extent entry for idx */
-       int                     error;  /* error return value */
-       int                     i;      /* temp state */
-       xfs_ifork_t             *ifp;   /* inode fork pointer */
-       xfs_fileoff_t           new_endoff;     /* end offset of new entry */
-       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */
-                                       /* left is 0, right is 1, prev is 2 */
-       int                     rval=0; /* return value (logging flags) */
-       int                     state = 0;/* state bits, accessed thru macros */
-       xfs_filblks_t           da_new; /* new count del alloc blocks used */
-       xfs_filblks_t           da_old; /* old count del alloc blocks used */
-       xfs_filblks_t           temp=0; /* value for da_new calculations */
-       xfs_filblks_t           temp2=0;/* value for da_new calculations */
-       int                     tmp_rval;       /* partial logging flags */
+       int                     error;
+       xfs_buf_t               *bp, *nbp;
+       int                     level = levelin;
+       __be64                  *pp;
+       xfs_fsblock_t           bno = blockno;
+       xfs_fsblock_t           nextbno;
+       struct xfs_btree_block  *block, *nextblock;
+       int                     numrecs;
 
-       ifp = XFS_IFORK_PTR(bma->ip, XFS_DATA_FORK);
+       error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp, XFS_BMAP_BTREE_REF,
+                                               &xfs_bmbt_buf_ops);
+       if (error)
+               return error;
+       *count += 1;
+       block = XFS_BUF_TO_BLOCK(bp);
 
-       ASSERT(bma->idx >= 0);
-       ASSERT(bma->idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
-       ASSERT(!isnullstartblock(new->br_startblock));
-       ASSERT(!bma->cur ||
-              (bma->cur->bc_private.b.flags & XFS_BTCUR_BPRV_WASDEL));
+       if (--level) {
+               /* Not at node above leaves, count this level of nodes */
+               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
+               while (nextbno != NULLFSBLOCK) {
+                       error = xfs_btree_read_bufl(mp, tp, nextbno, 0, &nbp,
+                                               XFS_BMAP_BTREE_REF,
+                                               &xfs_bmbt_buf_ops);
+                       if (error)
+                               return error;
+                       *count += 1;
+                       nextblock = XFS_BUF_TO_BLOCK(nbp);
+                       nextbno = be64_to_cpu(nextblock->bb_u.l.bb_rightsib);
+                       xfs_trans_brelse(tp, nbp);
+               }
 
-       XFS_STATS_INC(xs_add_exlist);
+               /* Dive to the next level */
+               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
+               bno = be64_to_cpu(*pp);
+               if (unlikely((error =
+                    xfs_bmap_count_tree(mp, tp, ifp, bno, level, count)) < 0)) {
+                       xfs_trans_brelse(tp, bp);
+                       XFS_ERROR_REPORT("xfs_bmap_count_tree(1)",
+                                        XFS_ERRLEVEL_LOW, mp);
+                       return XFS_ERROR(EFSCORRUPTED);
+               }
+               xfs_trans_brelse(tp, bp);
+       } else {
+               /* count all level 1 nodes and their leaves */
+               for (;;) {
+                       nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
+                       numrecs = be16_to_cpu(block->bb_numrecs);
+                       xfs_bmap_disk_count_leaves(mp, block, numrecs, count);
+                       xfs_trans_brelse(tp, bp);
+                       if (nextbno == NULLFSBLOCK)
+                               break;
+                       bno = nextbno;
+                       error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
+                                               XFS_BMAP_BTREE_REF,
+                                               &xfs_bmbt_buf_ops);
+                       if (error)
+                               return error;
+                       *count += 1;
+                       block = XFS_BUF_TO_BLOCK(bp);
+               }
+       }
+       return 0;
+}
 
-#define        LEFT            r[0]
-#define        RIGHT           r[1]
-#define        PREV            r[2]
+/*
+ * Count fsblocks of the given fork.
+ */
+int                                            /* error */
+xfs_bmap_count_blocks(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       xfs_inode_t             *ip,            /* incore inode */
+       int                     whichfork,      /* data or attr fork */
+       int                     *count)         /* out: count of blocks */
+{
+       struct xfs_btree_block  *block; /* current btree block */
+       xfs_fsblock_t           bno;    /* block # of "block" */
+       xfs_ifork_t             *ifp;   /* fork structure */
+       int                     level;  /* btree level, for checking */
+       xfs_mount_t             *mp;    /* file system mount structure */
+       __be64                  *pp;    /* pointer to block address */
+
+       bno = NULLFSBLOCK;
+       mp = ip->i_mount;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       if ( XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS ) {
+               xfs_bmap_count_leaves(ifp, 0,
+                       ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t),
+                       count);
+               return 0;
+       }
 
        /*
-        * Set up a bunch of variables to make the tests simpler.
+        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
         */
-       ep = xfs_iext_get_ext(ifp, bma->idx);
-       xfs_bmbt_get_all(ep, &PREV);
-       new_endoff = new->br_startoff + new->br_blockcount;
-       ASSERT(PREV.br_startoff <= new->br_startoff);
-       ASSERT(PREV.br_startoff + PREV.br_blockcount >= new_endoff);
+       block = ifp->if_broot;
+       level = be16_to_cpu(block->bb_level);
+       ASSERT(level > 0);
+       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
+       bno = be64_to_cpu(*pp);
+       ASSERT(bno != NULLDFSBNO);
+       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
+       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
 
-       da_old = startblockval(PREV.br_startblock);
-       da_new = 0;
+       if (unlikely(xfs_bmap_count_tree(mp, tp, ifp, bno, level, count) < 0)) {
+               XFS_ERROR_REPORT("xfs_bmap_count_blocks(2)", XFS_ERRLEVEL_LOW,
+                                mp);
+               return XFS_ERROR(EFSCORRUPTED);
+       }
 
-       /*
-        * Set flags determining what part of the previous delayed allocation
-        * extent is being replaced by a real allocation.
-        */
-       if (PREV.br_startoff == new->br_startoff)
-               state |= BMAP_LEFT_FILLING;
-       if (PREV.br_startoff + PREV.br_blockcount == new_endoff)
-               state |= BMAP_RIGHT_FILLING;
+       return 0;
+}
 
-       /*
-        * Check and set flags if this segment has a left neighbor.
-        * Don't set contiguous if the combined extent would be too large.
-        */
-       if (bma->idx > 0) {
-               state |= BMAP_LEFT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1), &LEFT);
+/*
+ * Debug/sanity checking code
+ */
 
-               if (isnullstartblock(LEFT.br_startblock))
-                       state |= BMAP_LEFT_DELAY;
-       }
+STATIC int
+xfs_bmap_sanity_check(
+       struct xfs_mount        *mp,
+       struct xfs_buf          *bp,
+       int                     level)
+{
+       struct xfs_btree_block  *block = XFS_BUF_TO_BLOCK(bp);
 
-       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
-           LEFT.br_startoff + LEFT.br_blockcount == new->br_startoff &&
-           LEFT.br_startblock + LEFT.br_blockcount == new->br_startblock &&
-           LEFT.br_state == new->br_state &&
-           LEFT.br_blockcount + new->br_blockcount <= MAXEXTLEN)
-               state |= BMAP_LEFT_CONTIG;
+       if (block->bb_magic != cpu_to_be32(XFS_BMAP_MAGIC) ||
+           be16_to_cpu(block->bb_level) != level ||
+           be16_to_cpu(block->bb_numrecs) == 0 ||
+           be16_to_cpu(block->bb_numrecs) > mp->m_bmap_dmxr[level != 0])
+               return 0;
+       return 1;
+}
 
-       /*
-        * Check and set flags if this segment has a right neighbor.
-        * Don't set contiguous if the combined extent would be too large.
-        * Also check for all-three-contiguous being too large.
-        */
-       if (bma->idx < bma->ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t) - 1) {
-               state |= BMAP_RIGHT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx + 1), &RIGHT);
+#ifdef DEBUG
+STATIC struct xfs_buf *
+xfs_bmap_get_bp(
+       struct xfs_btree_cur    *cur,
+       xfs_fsblock_t           bno)
+{
+       struct xfs_log_item_desc *lidp;
+       int                     i;
 
-               if (isnullstartblock(RIGHT.br_startblock))
-                       state |= BMAP_RIGHT_DELAY;
+       if (!cur)
+               return NULL;
+
+       for (i = 0; i < XFS_BTREE_MAXLEVELS; i++) {
+               if (!cur->bc_bufs[i])
+                       break;
+               if (XFS_BUF_ADDR(cur->bc_bufs[i]) == bno)
+                       return cur->bc_bufs[i];
        }
 
-       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
-           new_endoff == RIGHT.br_startoff &&
-           new->br_startblock + new->br_blockcount == RIGHT.br_startblock &&
-           new->br_state == RIGHT.br_state &&
-           new->br_blockcount + RIGHT.br_blockcount <= MAXEXTLEN &&
-           ((state & (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
-                      BMAP_RIGHT_FILLING)) !=
-                     (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
-                      BMAP_RIGHT_FILLING) ||
-            LEFT.br_blockcount + new->br_blockcount + RIGHT.br_blockcount
-                       <= MAXEXTLEN))
-               state |= BMAP_RIGHT_CONTIG;
+       /* Chase down all the log items to see if the bp is there */
+       list_for_each_entry(lidp, &cur->bc_tp->t_items, lid_trans) {
+               struct xfs_buf_log_item *bip;
+               bip = (struct xfs_buf_log_item *)lidp->lid_item;
+               if (bip->bli_item.li_type == XFS_LI_BUF &&
+                   XFS_BUF_ADDR(bip->bli_buf) == bno)
+                       return bip->bli_buf;
+       }
 
-       error = 0;
-       /*
-        * Switch out based on the FILLING and CONTIG state bits.
-        */
-       switch (state & (BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
-                        BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG)) {
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
-            BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
-               /*
-                * Filling in all of a previously delayed allocation extent.
-                * The left and right neighbors are both contiguous with new.
-                */
-               bma->idx--;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
-                       LEFT.br_blockcount + PREV.br_blockcount +
-                       RIGHT.br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+       return NULL;
+}
 
-               xfs_iext_remove(bma->ip, bma->idx + 1, 2, state);
-               bma->ip->i_d.di_nextents--;
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
-                                       RIGHT.br_startblock,
-                                       RIGHT.br_blockcount, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_btree_delete(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_btree_decrement(bma->cur, 0, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
-                                       LEFT.br_startblock,
-                                       LEFT.br_blockcount +
-                                       PREV.br_blockcount +
-                                       RIGHT.br_blockcount, LEFT.br_state);
-                       if (error)
-                               goto done;
-               }
-               break;
+STATIC void
+xfs_check_block(
+       struct xfs_btree_block  *block,
+       xfs_mount_t             *mp,
+       int                     root,
+       short                   sz)
+{
+       int                     i, j, dmxr;
+       __be64                  *pp, *thispa;   /* pointer to block address */
+       xfs_bmbt_key_t          *prevp, *keyp;
 
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
-               /*
-                * Filling in all of a previously delayed allocation extent.
-                * The left neighbor is contiguous, the right is not.
-                */
-               bma->idx--;
+       ASSERT(be16_to_cpu(block->bb_level) > 0);
 
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
-                       LEFT.br_blockcount + PREV.br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+       prevp = NULL;
+       for( i = 1; i <= xfs_btree_get_numrecs(block); i++) {
+               dmxr = mp->m_bmap_dmxr[0];
+               keyp = XFS_BMBT_KEY_ADDR(mp, block, i);
 
-               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur, LEFT.br_startoff,
-                                       LEFT.br_startblock, LEFT.br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
-                                       LEFT.br_startblock,
-                                       LEFT.br_blockcount +
-                                       PREV.br_blockcount, LEFT.br_state);
-                       if (error)
-                               goto done;
+               if (prevp) {
+                       ASSERT(be64_to_cpu(prevp->br_startoff) <
+                              be64_to_cpu(keyp->br_startoff));
                }
-               break;
+               prevp = keyp;
 
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * Filling in all of a previously delayed allocation extent.
-                * The right neighbor is contiguous, the left is not.
+                * Compare the block numbers to see if there are dups.
                 */
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_startblock(ep, new->br_startblock);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount + RIGHT.br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+               if (root)
+                       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, i, sz);
+               else
+                       pp = XFS_BMBT_PTR_ADDR(mp, block, i, dmxr);
 
-               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
-                                       RIGHT.br_startblock,
-                                       RIGHT.br_blockcount, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, PREV.br_startoff,
-                                       new->br_startblock,
-                                       PREV.br_blockcount +
-                                       RIGHT.br_blockcount, PREV.br_state);
-                       if (error)
-                               goto done;
+               for (j = i+1; j <= be16_to_cpu(block->bb_numrecs); j++) {
+                       if (root)
+                               thispa = XFS_BMAP_BROOT_PTR_ADDR(mp, block, j, sz);
+                       else
+                               thispa = XFS_BMBT_PTR_ADDR(mp, block, j, dmxr);
+                       if (*thispa == *pp) {
+                               xfs_warn(mp, "%s: thispa(%d) == pp(%d) %Ld",
+                                       __func__, j, i,
+                                       (unsigned long long)be64_to_cpu(*thispa));
+                               panic("%s: ptrs are equal in node\n",
+                                       __func__);
+                       }
                }
-               break;
+       }
+}
 
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING:
-               /*
-                * Filling in all of a previously delayed allocation extent.
-                * Neither the left nor right neighbors are contiguous with
-                * the new one.
-                */
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_startblock(ep, new->br_startblock);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+/*
+ * Check that the extents for the inode ip are in the right order in all
+ * btree leaves.
+ */
 
-               bma->ip->i_d.di_nextents++;
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
-                       error = xfs_btree_insert(bma->cur, &i);
+STATIC void
+xfs_bmap_check_leaf_extents(
+       xfs_btree_cur_t         *cur,   /* btree cursor or null */
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       int                     whichfork)      /* data or attr fork */
+{
+       struct xfs_btree_block  *block; /* current btree block */
+       xfs_fsblock_t           bno;    /* block # of "block" */
+       xfs_buf_t               *bp;    /* buffer for "block" */
+       int                     error;  /* error return value */
+       xfs_extnum_t            i=0, j; /* index into the extents list */
+       xfs_ifork_t             *ifp;   /* fork structure */
+       int                     level;  /* btree level, for checking */
+       xfs_mount_t             *mp;    /* file system mount structure */
+       __be64                  *pp;    /* pointer to block address */
+       xfs_bmbt_rec_t          *ep;    /* pointer to current extent */
+       xfs_bmbt_rec_t          last = {0, 0}; /* last extent in prev block */
+       xfs_bmbt_rec_t          *nextp; /* pointer to next extent */
+       int                     bp_release = 0;
+
+       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE) {
+               return;
+       }
+
+       bno = NULLFSBLOCK;
+       mp = ip->i_mount;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       block = ifp->if_broot;
+       /*
+        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
+        */
+       level = be16_to_cpu(block->bb_level);
+       ASSERT(level > 0);
+       xfs_check_block(block, mp, 1, ifp->if_broot_bytes);
+       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
+       bno = be64_to_cpu(*pp);
+
+       ASSERT(bno != NULLDFSBNO);
+       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
+       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
+
+       /*
+        * Go down the tree until leaf level is reached, following the first
+        * pointer (leftmost) at each level.
+        */
+       while (level-- > 0) {
+               /* See if buf is in cur first */
+               bp_release = 0;
+               bp = xfs_bmap_get_bp(cur, XFS_FSB_TO_DADDR(mp, bno));
+               if (!bp) {
+                       bp_release = 1;
+                       error = xfs_btree_read_bufl(mp, NULL, bno, 0, &bp,
+                                               XFS_BMAP_BTREE_REF,
+                                               &xfs_bmbt_buf_ops);
                        if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                               goto error_norelse;
                }
-               break;
+               block = XFS_BUF_TO_BLOCK(bp);
+               XFS_WANT_CORRUPTED_GOTO(
+                       xfs_bmap_sanity_check(mp, bp, level),
+                       error0);
+               if (level == 0)
+                       break;
 
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG:
                /*
-                * Filling in the first part of a previous delayed allocation.
-                * The left neighbor is contiguous.
+                * Check this block for basic sanity (increasing keys and
+                * no duplicate blocks).
                 */
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx - 1, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx - 1),
-                       LEFT.br_blockcount + new->br_blockcount);
-               xfs_bmbt_set_startoff(ep,
-                       PREV.br_startoff + new->br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx - 1, state, _THIS_IP_);
 
-               temp = PREV.br_blockcount - new->br_blockcount;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep, temp);
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur, LEFT.br_startoff,
-                                       LEFT.br_startblock, LEFT.br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
-                                       LEFT.br_startblock,
-                                       LEFT.br_blockcount +
-                                       new->br_blockcount,
-                                       LEFT.br_state);
-                       if (error)
-                               goto done;
+               xfs_check_block(block, mp, 0, 0);
+               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
+               bno = be64_to_cpu(*pp);
+               XFS_WANT_CORRUPTED_GOTO(XFS_FSB_SANITY_CHECK(mp, bno), error0);
+               if (bp_release) {
+                       bp_release = 0;
+                       xfs_trans_brelse(NULL, bp);
                }
-               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
-                       startblockval(PREV.br_startblock));
-               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+       }
 
-               bma->idx--;
-               break;
+       /*
+        * Here with bp and block set to the leftmost leaf node in the tree.
+        */
+       i = 0;
+
+       /*
+        * Loop over all leaf nodes checking that all extents are in the right order.
+        */
+       for (;;) {
+               xfs_fsblock_t   nextbno;
+               xfs_extnum_t    num_recs;
+
+
+               num_recs = xfs_btree_get_numrecs(block);
 
-       case BMAP_LEFT_FILLING:
                /*
-                * Filling in the first part of a previous delayed allocation.
-                * The left neighbor is not contiguous.
+                * Read-ahead the next leaf block, if any.
                 */
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_startoff(ep, new_endoff);
-               temp = PREV.br_blockcount - new->br_blockcount;
-               xfs_bmbt_set_blockcount(ep, temp);
-               xfs_iext_insert(bma->ip, bma->idx, 1, new, state);
-               bma->ip->i_d.di_nextents++;
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
-                       error = xfs_btree_insert(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-               }
 
-               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
-                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
-                                       bma->firstblock, bma->flist,
-                                       &bma->cur, 1, &tmp_rval, XFS_DATA_FORK);
-                       rval |= tmp_rval;
-                       if (error)
-                               goto done;
-               }
-               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
-                       startblockval(PREV.br_startblock) -
-                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
-               ep = xfs_iext_get_ext(ifp, bma->idx + 1);
-               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
-               break;
+               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
 
-       case BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * Filling in the last part of a previous delayed allocation.
-                * The right neighbor is contiguous with the new allocation.
+                * Check all the extents to make sure they are OK.
+                * If we had a previous block, the last entry should
+                * conform with the first entry in this one.
                 */
-               temp = PREV.br_blockcount - new->br_blockcount;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep, temp);
-               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, bma->idx + 1),
-                       new->br_startoff, new->br_startblock,
-                       new->br_blockcount + RIGHT.br_blockcount,
-                       RIGHT.br_state);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
-                                       RIGHT.br_startblock,
-                                       RIGHT.br_blockcount, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, new->br_startoff,
-                                       new->br_startblock,
-                                       new->br_blockcount +
-                                       RIGHT.br_blockcount,
-                                       RIGHT.br_state);
-                       if (error)
-                               goto done;
-               }
-
-               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
-                       startblockval(PREV.br_startblock));
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-               bma->idx++;
-               break;
+               ep = XFS_BMBT_REC_ADDR(mp, block, 1);
+               if (i) {
+                       ASSERT(xfs_bmbt_disk_get_startoff(&last) +
+                              xfs_bmbt_disk_get_blockcount(&last) <=
+                              xfs_bmbt_disk_get_startoff(ep));
+               }
+               for (j = 1; j < num_recs; j++) {
+                       nextp = XFS_BMBT_REC_ADDR(mp, block, j + 1);
+                       ASSERT(xfs_bmbt_disk_get_startoff(ep) +
+                              xfs_bmbt_disk_get_blockcount(ep) <=
+                              xfs_bmbt_disk_get_startoff(nextp));
+                       ep = nextp;
+               }
 
-       case BMAP_RIGHT_FILLING:
+               last = *ep;
+               i += num_recs;
+               if (bp_release) {
+                       bp_release = 0;
+                       xfs_trans_brelse(NULL, bp);
+               }
+               bno = nextbno;
                /*
-                * Filling in the last part of a previous delayed allocation.
-                * The right neighbor is not contiguous.
+                * If we've reached the end, stop.
                 */
-               temp = PREV.br_blockcount - new->br_blockcount;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep, temp);
-               xfs_iext_insert(bma->ip, bma->idx + 1, 1, new, state);
-               bma->ip->i_d.di_nextents++;
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
-                       error = xfs_btree_insert(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-               }
-
-               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
-                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
-                               bma->firstblock, bma->flist, &bma->cur, 1,
-                               &tmp_rval, XFS_DATA_FORK);
-                       rval |= tmp_rval;
-                       if (error)
-                               goto done;
-               }
-               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
-                       startblockval(PREV.br_startblock) -
-                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
-               ep = xfs_iext_get_ext(ifp, bma->idx);
-               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
-
-               bma->idx++;
-               break;
-
-       case 0:
-               /*
-                * Filling in the middle part of a previous delayed allocation.
-                * Contiguity is impossible here.
-                * This case is avoided almost all the time.
-                *
-                * We start with a delayed allocation:
-                *
-                * +ddddddddddddddddddddddddddddddddddddddddddddddddddddddd+
-                *  PREV @ idx
-                *
-                * and we are allocating:
-                *                     +rrrrrrrrrrrrrrrrr+
-                *                            new
-                *
-                * and we set it up for insertion as:
-                * +ddddddddddddddddddd+rrrrrrrrrrrrrrrrr+ddddddddddddddddd+
-                *                            new
-                *  PREV @ idx          LEFT              RIGHT
-                *                      inserted at idx + 1
-                */
-               temp = new->br_startoff - PREV.br_startoff;
-               temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, 0, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep, temp);      /* truncate PREV */
-               LEFT = *new;
-               RIGHT.br_state = PREV.br_state;
-               RIGHT.br_startblock = nullstartblock(
-                               (int)xfs_bmap_worst_indlen(bma->ip, temp2));
-               RIGHT.br_startoff = new_endoff;
-               RIGHT.br_blockcount = temp2;
-               /* insert LEFT (r[0]) and RIGHT (r[1]) at the same time */
-               xfs_iext_insert(bma->ip, bma->idx + 1, 2, &LEFT, state);
-               bma->ip->i_d.di_nextents++;
-               if (bma->cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
-                       error = xfs_btree_insert(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-               }
+               if (bno == NULLFSBLOCK)
+                       break;
 
-               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
-                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
-                                       bma->firstblock, bma->flist, &bma->cur,
-                                       1, &tmp_rval, XFS_DATA_FORK);
-                       rval |= tmp_rval;
-                       if (error)
-                               goto done;
-               }
-               temp = xfs_bmap_worst_indlen(bma->ip, temp);
-               temp2 = xfs_bmap_worst_indlen(bma->ip, temp2);
-               diff = (int)(temp + temp2 - startblockval(PREV.br_startblock) -
-                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
-               if (diff > 0) {
-                       error = xfs_icsb_modify_counters(bma->ip->i_mount,
-                                       XFS_SBS_FDBLOCKS,
-                                       -((int64_t)diff), 0);
-                       ASSERT(!error);
+               bp_release = 0;
+               bp = xfs_bmap_get_bp(cur, XFS_FSB_TO_DADDR(mp, bno));
+               if (!bp) {
+                       bp_release = 1;
+                       error = xfs_btree_read_bufl(mp, NULL, bno, 0, &bp,
+                                               XFS_BMAP_BTREE_REF,
+                                               &xfs_bmbt_buf_ops);
                        if (error)
-                               goto done;
+                               goto error_norelse;
                }
-
-               ep = xfs_iext_get_ext(ifp, bma->idx);
-               xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx + 2, state, _THIS_IP_);
-               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, bma->idx + 2),
-                       nullstartblock((int)temp2));
-               trace_xfs_bmap_post_update(bma->ip, bma->idx + 2, state, _THIS_IP_);
-
-               bma->idx++;
-               da_new = temp + temp2;
-               break;
-
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_CONTIG:
-       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
-       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_LEFT_CONTIG:
-       case BMAP_RIGHT_CONTIG:
-               /*
-                * These cases are all impossible.
-                */
-               ASSERT(0);
-       }
-
-       /* convert to a btree if necessary */
-       if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
-               int     tmp_logflags;   /* partial log flag return val */
-
-               ASSERT(bma->cur == NULL);
-               error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
-                               bma->firstblock, bma->flist, &bma->cur,
-                               da_old > 0, &tmp_logflags, XFS_DATA_FORK);
-               bma->logflags |= tmp_logflags;
-               if (error)
-                       goto done;
+               block = XFS_BUF_TO_BLOCK(bp);
        }
-
-       /* adjust for changes in reserved delayed indirect blocks */
-       if (da_old || da_new) {
-               temp = da_new;
-               if (bma->cur)
-                       temp += bma->cur->bc_private.b.allocated;
-               ASSERT(temp <= da_old);
-               if (temp < da_old)
-                       xfs_icsb_modify_counters(bma->ip->i_mount,
-                                       XFS_SBS_FDBLOCKS,
-                                       (int64_t)(da_old - temp), 0);
+       if (bp_release) {
+               bp_release = 0;
+               xfs_trans_brelse(NULL, bp);
        }
+       return;
 
-       /* clear out the allocated field, done with it now in any case. */
-       if (bma->cur)
-               bma->cur->bc_private.b.allocated = 0;
-
-       xfs_bmap_check_leaf_extents(bma->cur, bma->ip, XFS_DATA_FORK);
-done:
-       bma->logflags |= rval;
-       return error;
-#undef LEFT
-#undef RIGHT
-#undef PREV
+error0:
+       xfs_warn(mp, "%s: at error0", __func__);
+       if (bp_release)
+               xfs_trans_brelse(NULL, bp);
+error_norelse:
+       xfs_warn(mp, "%s: BAD after btree leaves for %d extents",
+               __func__, i);
+       panic("%s: CORRUPTED BTREE OR SOMETHING", __func__);
+       return;
 }
 
 /*
- * Convert an unwritten allocation to a real allocation or vice versa.
+ * Add bmap trace insert entries for all the contents of the extent records.
  */
-STATIC int                             /* error */
-xfs_bmap_add_extent_unwritten_real(
-       struct xfs_trans        *tp,
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_extnum_t            *idx,   /* extent number to update/insert */
-       xfs_btree_cur_t         **curp, /* if *curp is null, not a btree */
-       xfs_bmbt_irec_t         *new,   /* new data to add to file extents */
-       xfs_fsblock_t           *first, /* pointer to firstblock variable */
-       xfs_bmap_free_t         *flist, /* list of extents to be freed */
-       int                     *logflagsp) /* inode logging flags */
+void
+xfs_bmap_trace_exlist(
+       xfs_inode_t     *ip,            /* incore inode pointer */
+       xfs_extnum_t    cnt,            /* count of entries in the list */
+       int             whichfork,      /* data or attr fork */
+       unsigned long   caller_ip)
 {
-       xfs_btree_cur_t         *cur;   /* btree cursor */
-       xfs_bmbt_rec_host_t     *ep;    /* extent entry for idx */
-       int                     error;  /* error return value */
-       int                     i;      /* temp state */
-       xfs_ifork_t             *ifp;   /* inode fork pointer */
-       xfs_fileoff_t           new_endoff;     /* end offset of new entry */
-       xfs_exntst_t            newext; /* new extent state */
-       xfs_exntst_t            oldext; /* old extent state */
-       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */
-                                       /* left is 0, right is 1, prev is 2 */
-       int                     rval=0; /* return value (logging flags) */
-       int                     state = 0;/* state bits, accessed thru macros */
-
-       *logflagsp = 0;
+       xfs_extnum_t    idx;            /* extent record index */
+       xfs_ifork_t     *ifp;           /* inode fork pointer */
+       int             state = 0;
 
-       cur = *curp;
-       ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
+       if (whichfork == XFS_ATTR_FORK)
+               state |= BMAP_ATTRFORK;
 
-       ASSERT(*idx >= 0);
-       ASSERT(*idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
-       ASSERT(!isnullstartblock(new->br_startblock));
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT(cnt == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)));
+       for (idx = 0; idx < cnt; idx++)
+               trace_xfs_extlist(ip, idx, whichfork, caller_ip);
+}
 
-       XFS_STATS_INC(xs_add_exlist);
+/*
+ * Validate that the bmbt_irecs being returned from bmapi are valid
+ * given the callers original parameters.  Specifically check the
+ * ranges of the returned irecs to ensure that they only extent beyond
+ * the given parameters if the XFS_BMAPI_ENTIRE flag was set.
+ */
+STATIC void
+xfs_bmap_validate_ret(
+       xfs_fileoff_t           bno,
+       xfs_filblks_t           len,
+       int                     flags,
+       xfs_bmbt_irec_t         *mval,
+       int                     nmap,
+       int                     ret_nmap)
+{
+       int                     i;              /* index to map values */
 
-#define        LEFT            r[0]
-#define        RIGHT           r[1]
-#define        PREV            r[2]
+       ASSERT(ret_nmap <= nmap);
 
-       /*
-        * Set up a bunch of variables to make the tests simpler.
-        */
-       error = 0;
-       ep = xfs_iext_get_ext(ifp, *idx);
-       xfs_bmbt_get_all(ep, &PREV);
-       newext = new->br_state;
-       oldext = (newext == XFS_EXT_UNWRITTEN) ?
-               XFS_EXT_NORM : XFS_EXT_UNWRITTEN;
-       ASSERT(PREV.br_state == oldext);
-       new_endoff = new->br_startoff + new->br_blockcount;
-       ASSERT(PREV.br_startoff <= new->br_startoff);
-       ASSERT(PREV.br_startoff + PREV.br_blockcount >= new_endoff);
+       for (i = 0; i < ret_nmap; i++) {
+               ASSERT(mval[i].br_blockcount > 0);
+               if (!(flags & XFS_BMAPI_ENTIRE)) {
+                       ASSERT(mval[i].br_startoff >= bno);
+                       ASSERT(mval[i].br_blockcount <= len);
+                       ASSERT(mval[i].br_startoff + mval[i].br_blockcount <=
+                              bno + len);
+               } else {
+                       ASSERT(mval[i].br_startoff < bno + len);
+                       ASSERT(mval[i].br_startoff + mval[i].br_blockcount >
+                              bno);
+               }
+               ASSERT(i == 0 ||
+                      mval[i - 1].br_startoff + mval[i - 1].br_blockcount ==
+                      mval[i].br_startoff);
+               ASSERT(mval[i].br_startblock != DELAYSTARTBLOCK &&
+                      mval[i].br_startblock != HOLESTARTBLOCK);
+               ASSERT(mval[i].br_state == XFS_EXT_NORM ||
+                      mval[i].br_state == XFS_EXT_UNWRITTEN);
+       }
+}
 
-       /*
-        * Set flags determining what part of the previous oldext allocation
-        * extent is being replaced by a newext allocation.
-        */
-       if (PREV.br_startoff == new->br_startoff)
-               state |= BMAP_LEFT_FILLING;
-       if (PREV.br_startoff + PREV.br_blockcount == new_endoff)
-               state |= BMAP_RIGHT_FILLING;
+#else
+#define xfs_bmap_check_leaf_extents(cur, ip, whichfork)                do { } while (0)
+#define        xfs_bmap_validate_ret(bno,len,flags,mval,onmap,nmap)
+#endif /* DEBUG */
 
-       /*
-        * Check and set flags if this segment has a left neighbor.
-        * Don't set contiguous if the combined extent would be too large.
-        */
-       if (*idx > 0) {
-               state |= BMAP_LEFT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx - 1), &LEFT);
+/*
+ * bmap free list manipulation functions
+ */
 
-               if (isnullstartblock(LEFT.br_startblock))
-                       state |= BMAP_LEFT_DELAY;
+/*
+ * Add the extent to the list of extents to be free at transaction end.
+ * The list is maintained sorted (by block number).
+ */
+void
+xfs_bmap_add_free(
+       xfs_fsblock_t           bno,            /* fs block number of extent */
+       xfs_filblks_t           len,            /* length of extent */
+       xfs_bmap_free_t         *flist,         /* list of extents */
+       xfs_mount_t             *mp)            /* mount point structure */
+{
+       xfs_bmap_free_item_t    *cur;           /* current (next) element */
+       xfs_bmap_free_item_t    *new;           /* new element */
+       xfs_bmap_free_item_t    *prev;          /* previous element */
+#ifdef DEBUG
+       xfs_agnumber_t          agno;
+       xfs_agblock_t           agbno;
+
+       ASSERT(bno != NULLFSBLOCK);
+       ASSERT(len > 0);
+       ASSERT(len <= MAXEXTLEN);
+       ASSERT(!isnullstartblock(bno));
+       agno = XFS_FSB_TO_AGNO(mp, bno);
+       agbno = XFS_FSB_TO_AGBNO(mp, bno);
+       ASSERT(agno < mp->m_sb.sb_agcount);
+       ASSERT(agbno < mp->m_sb.sb_agblocks);
+       ASSERT(len < mp->m_sb.sb_agblocks);
+       ASSERT(agbno + len <= mp->m_sb.sb_agblocks);
+#endif
+       ASSERT(xfs_bmap_free_item_zone != NULL);
+       new = kmem_zone_alloc(xfs_bmap_free_item_zone, KM_SLEEP);
+       new->xbfi_startblock = bno;
+       new->xbfi_blockcount = (xfs_extlen_t)len;
+       for (prev = NULL, cur = flist->xbf_first;
+            cur != NULL;
+            prev = cur, cur = cur->xbfi_next) {
+               if (cur->xbfi_startblock >= bno)
+                       break;
        }
+       if (prev)
+               prev->xbfi_next = new;
+       else
+               flist->xbf_first = new;
+       new->xbfi_next = cur;
+       flist->xbf_count++;
+}
 
-       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
-           LEFT.br_startoff + LEFT.br_blockcount == new->br_startoff &&
-           LEFT.br_startblock + LEFT.br_blockcount == new->br_startblock &&
-           LEFT.br_state == newext &&
-           LEFT.br_blockcount + new->br_blockcount <= MAXEXTLEN)
-               state |= BMAP_LEFT_CONTIG;
+/*
+ * Remove the entry "free" from the free item list.  Prev points to the
+ * previous entry, unless "free" is the head of the list.
+ */
+STATIC void
+xfs_bmap_del_free(
+       xfs_bmap_free_t         *flist, /* free item list header */
+       xfs_bmap_free_item_t    *prev,  /* previous item on list, if any */
+       xfs_bmap_free_item_t    *free)  /* list item to be freed */
+{
+       if (prev)
+               prev->xbfi_next = free->xbfi_next;
+       else
+               flist->xbf_first = free->xbfi_next;
+       flist->xbf_count--;
+       kmem_zone_free(xfs_bmap_free_item_zone, free);
+}
+
+
+/*
+ * Routine to be called at transaction's end by xfs_bmapi, xfs_bunmapi
+ * caller.  Frees all the extents that need freeing, which must be done
+ * last due to locking considerations.  We never free any extents in
+ * the first transaction.
+ *
+ * Return 1 if the given transaction was committed and a new one
+ * started, and 0 otherwise in the committed parameter.
+ */
+int                                            /* error */
+xfs_bmap_finish(
+       xfs_trans_t             **tp,           /* transaction pointer addr */
+       xfs_bmap_free_t         *flist,         /* i/o: list extents to free */
+       int                     *committed)     /* xact committed or not */
+{
+       xfs_efd_log_item_t      *efd;           /* extent free data */
+       xfs_efi_log_item_t      *efi;           /* extent free intention */
+       int                     error;          /* error return value */
+       xfs_bmap_free_item_t    *free;          /* free extent item */
+       unsigned int            logres;         /* new log reservation */
+       unsigned int            logcount;       /* new log count */
+       xfs_mount_t             *mp;            /* filesystem mount structure */
+       xfs_bmap_free_item_t    *next;          /* next item on free list */
+       xfs_trans_t             *ntp;           /* new transaction pointer */
 
+       ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
+       if (flist->xbf_count == 0) {
+               *committed = 0;
+               return 0;
+       }
+       ntp = *tp;
+       efi = xfs_trans_get_efi(ntp, flist->xbf_count);
+       for (free = flist->xbf_first; free; free = free->xbfi_next)
+               xfs_trans_log_efi_extent(ntp, efi, free->xbfi_startblock,
+                       free->xbfi_blockcount);
+       logres = ntp->t_log_res;
+       logcount = ntp->t_log_count;
+       ntp = xfs_trans_dup(*tp);
+       error = xfs_trans_commit(*tp, 0);
+       *tp = ntp;
+       *committed = 1;
        /*
-        * Check and set flags if this segment has a right neighbor.
-        * Don't set contiguous if the combined extent would be too large.
-        * Also check for all-three-contiguous being too large.
+        * We have a new transaction, so we should return committed=1,
+        * even though we're returning an error.
         */
-       if (*idx < ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t) - 1) {
-               state |= BMAP_RIGHT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx + 1), &RIGHT);
-               if (isnullstartblock(RIGHT.br_startblock))
-                       state |= BMAP_RIGHT_DELAY;
-       }
-
-       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
-           new_endoff == RIGHT.br_startoff &&
-           new->br_startblock + new->br_blockcount == RIGHT.br_startblock &&
-           newext == RIGHT.br_state &&
-           new->br_blockcount + RIGHT.br_blockcount <= MAXEXTLEN &&
-           ((state & (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
-                      BMAP_RIGHT_FILLING)) !=
-                     (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
-                      BMAP_RIGHT_FILLING) ||
-            LEFT.br_blockcount + new->br_blockcount + RIGHT.br_blockcount
-                       <= MAXEXTLEN))
-               state |= BMAP_RIGHT_CONTIG;
+       if (error)
+               return error;
 
        /*
-        * Switch out based on the FILLING and CONTIG state bits.
+        * transaction commit worked ok so we can drop the extra ticket
+        * reference that we gained in xfs_trans_dup()
         */
-       switch (state & (BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
-                        BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG)) {
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
-            BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
-               /*
-                * Setting all of a previous oldext extent to newext.
-                * The left and right neighbors are both contiguous with new.
-                */
-               --*idx;
+       xfs_log_ticket_put(ntp->t_ticket);
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx),
-                       LEFT.br_blockcount + PREV.br_blockcount +
-                       RIGHT.br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       if ((error = xfs_trans_reserve(ntp, 0, logres, 0, XFS_TRANS_PERM_LOG_RES,
+                       logcount)))
+               return error;
+       efd = xfs_trans_get_efd(ntp, efi, flist->xbf_count);
+       for (free = flist->xbf_first; free != NULL; free = next) {
+               next = free->xbfi_next;
+               if ((error = xfs_free_extent(ntp, free->xbfi_startblock,
+                               free->xbfi_blockcount))) {
+                       /*
+                        * The bmap free list will be cleaned up at a
+                        * higher level.  The EFI will be canceled when
+                        * this transaction is aborted.
+                        * Need to force shutdown here to make sure it
+                        * happens, since this transaction may not be
+                        * dirty yet.
+                        */
+                       mp = ntp->t_mountp;
+                       if (!XFS_FORCED_SHUTDOWN(mp))
+                               xfs_force_shutdown(mp,
+                                                  (error == EFSCORRUPTED) ?
+                                                  SHUTDOWN_CORRUPT_INCORE :
+                                                  SHUTDOWN_META_IO_ERROR);
+                       return error;
+               }
+               xfs_trans_log_efd_extent(ntp, efd, free->xbfi_startblock,
+                       free->xbfi_blockcount);
+               xfs_bmap_del_free(flist, NULL, free);
+       }
+       return 0;
+}
 
-               xfs_iext_remove(ip, *idx + 1, 2, state);
-               ip->i_d.di_nextents -= 2;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, RIGHT.br_startoff,
-                                       RIGHT.br_startblock,
-                                       RIGHT.br_blockcount, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_delete(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_decrement(cur, 0, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_delete(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_decrement(cur, 0, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, LEFT.br_startoff,
-                               LEFT.br_startblock,
-                               LEFT.br_blockcount + PREV.br_blockcount +
-                               RIGHT.br_blockcount, LEFT.br_state)))
-                               goto done;
-               }
-               break;
-
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
-               /*
-                * Setting all of a previous oldext extent to newext.
-                * The left neighbor is contiguous, the right is not.
-                */
-               --*idx;
+/*
+ * Free up any items left in the list.
+ */
+void
+xfs_bmap_cancel(
+       xfs_bmap_free_t         *flist) /* list of bmap_free_items */
+{
+       xfs_bmap_free_item_t    *free;  /* free list item */
+       xfs_bmap_free_item_t    *next;
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx),
-                       LEFT.br_blockcount + PREV.br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       if (flist->xbf_count == 0)
+               return;
+       ASSERT(flist->xbf_first != NULL);
+       for (free = flist->xbf_first; free; free = next) {
+               next = free->xbfi_next;
+               xfs_bmap_del_free(flist, NULL, free);
+       }
+       ASSERT(flist->xbf_count == 0);
+}
 
-               xfs_iext_remove(ip, *idx + 1, 1, state);
-               ip->i_d.di_nextents--;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock, PREV.br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_delete(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_decrement(cur, 0, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, LEFT.br_startoff,
-                               LEFT.br_startblock,
-                               LEFT.br_blockcount + PREV.br_blockcount,
-                               LEFT.br_state)))
-                               goto done;
-               }
-               break;
+/*
+ * Inode fork format manipulation functions
+ */
 
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
-               /*
-                * Setting all of a previous oldext extent to newext.
-                * The right neighbor is contiguous, the left is not.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount + RIGHT.br_blockcount);
-               xfs_bmbt_set_state(ep, newext);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-               xfs_iext_remove(ip, *idx + 1, 1, state);
-               ip->i_d.di_nextents--;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, RIGHT.br_startoff,
-                                       RIGHT.br_startblock,
-                                       RIGHT.br_blockcount, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_delete(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_btree_decrement(cur, 0, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
-                               new->br_startblock,
-                               new->br_blockcount + RIGHT.br_blockcount,
-                               newext)))
-                               goto done;
-               }
-               break;
+/*
+ * Transform a btree format file with only one leaf node, where the
+ * extents list will fit in the inode, into an extents format file.
+ * Since the file extents are already in-core, all we have to do is
+ * give up the space for the btree root and pitch the leaf block.
+ */
+STATIC int                             /* error */
+xfs_bmap_btree_to_extents(
+       xfs_trans_t             *tp,    /* transaction pointer */
+       xfs_inode_t             *ip,    /* incore inode pointer */
+       xfs_btree_cur_t         *cur,   /* btree cursor */
+       int                     *logflagsp, /* inode logging flags */
+       int                     whichfork)  /* data or attr fork */
+{
+       /* REFERENCED */
+       struct xfs_btree_block  *cblock;/* child btree block */
+       xfs_fsblock_t           cbno;   /* child block number */
+       xfs_buf_t               *cbp;   /* child block's buffer */
+       int                     error;  /* error return value */
+       xfs_ifork_t             *ifp;   /* inode fork data */
+       xfs_mount_t             *mp;    /* mount point structure */
+       __be64                  *pp;    /* ptr to block address */
+       struct xfs_btree_block  *rblock;/* root btree block */
 
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING:
-               /*
-                * Setting all of a previous oldext extent to newext.
-                * Neither the left nor right neighbors are contiguous with
-                * the new one.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_state(ep, newext);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       mp = ip->i_mount;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT(ifp->if_flags & XFS_IFEXTENTS);
+       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE);
+       rblock = ifp->if_broot;
+       ASSERT(be16_to_cpu(rblock->bb_level) == 1);
+       ASSERT(be16_to_cpu(rblock->bb_numrecs) == 1);
+       ASSERT(xfs_bmbt_maxrecs(mp, ifp->if_broot_bytes, 0) == 1);
+       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, rblock, 1, ifp->if_broot_bytes);
+       cbno = be64_to_cpu(*pp);
+       *logflagsp = 0;
+#ifdef DEBUG
+       if ((error = xfs_btree_check_lptr(cur, cbno, 1)))
+               return error;
+#endif
+       error = xfs_btree_read_bufl(mp, tp, cbno, 0, &cbp, XFS_BMAP_BTREE_REF,
+                               &xfs_bmbt_buf_ops);
+       if (error)
+               return error;
+       cblock = XFS_BUF_TO_BLOCK(cbp);
+       if ((error = xfs_btree_check_block(cur, cblock, 0, cbp)))
+               return error;
+       xfs_bmap_add_free(cbno, 1, cur->bc_private.b.flist, mp);
+       ip->i_d.di_nblocks--;
+       xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, -1L);
+       xfs_trans_binval(tp, cbp);
+       if (cur->bc_bufs[0] == cbp)
+               cur->bc_bufs[0] = NULL;
+       xfs_iroot_realloc(ip, -1, whichfork);
+       ASSERT(ifp->if_broot == NULL);
+       ASSERT((ifp->if_flags & XFS_IFBROOT) == 0);
+       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_EXTENTS);
+       *logflagsp = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
+       return 0;
+}
 
-               if (cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
-                               new->br_startblock, new->br_blockcount,
-                               newext)))
-                               goto done;
-               }
-               break;
+/*
+ * Convert an extents-format file into a btree-format file.
+ * The new file will have a root block (in the inode) and a single child block.
+ */
+STATIC int                                     /* error */
+xfs_bmap_extents_to_btree(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       xfs_fsblock_t           *firstblock,    /* first-block-allocated */
+       xfs_bmap_free_t         *flist,         /* blocks freed in xaction */
+       xfs_btree_cur_t         **curp,         /* cursor returned to caller */
+       int                     wasdel,         /* converting a delayed alloc */
+       int                     *logflagsp,     /* inode logging flags */
+       int                     whichfork)      /* data or attr fork */
+{
+       struct xfs_btree_block  *ablock;        /* allocated (child) bt block */
+       xfs_buf_t               *abp;           /* buffer for ablock */
+       xfs_alloc_arg_t         args;           /* allocation arguments */
+       xfs_bmbt_rec_t          *arp;           /* child record pointer */
+       struct xfs_btree_block  *block;         /* btree root block */
+       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
+       xfs_bmbt_rec_host_t     *ep;            /* extent record pointer */
+       int                     error;          /* error return value */
+       xfs_extnum_t            i, cnt;         /* extent record index */
+       xfs_ifork_t             *ifp;           /* inode fork pointer */
+       xfs_bmbt_key_t          *kp;            /* root block key pointer */
+       xfs_mount_t             *mp;            /* mount structure */
+       xfs_extnum_t            nextents;       /* number of file extents */
+       xfs_bmbt_ptr_t          *pp;            /* root block address pointer */
 
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG:
-               /*
-                * Setting the first part of a previous oldext extent to newext.
-                * The left neighbor is contiguous.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx - 1, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx - 1),
-                       LEFT.br_blockcount + new->br_blockcount);
-               xfs_bmbt_set_startoff(ep,
-                       PREV.br_startoff + new->br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx - 1, state, _THIS_IP_);
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS);
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_startblock(ep,
-                       new->br_startblock + new->br_blockcount);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount - new->br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       /*
+        * Make space in the inode incore.
+        */
+       xfs_iroot_realloc(ip, 1, whichfork);
+       ifp->if_flags |= XFS_IFBROOT;
 
-               --*idx;
+       /*
+        * Fill in the root.
+        */
+       block = ifp->if_broot;
+       block->bb_magic = cpu_to_be32(XFS_BMAP_MAGIC);
+       block->bb_level = cpu_to_be16(1);
+       block->bb_numrecs = cpu_to_be16(1);
+       block->bb_u.l.bb_leftsib = cpu_to_be64(NULLDFSBNO);
+       block->bb_u.l.bb_rightsib = cpu_to_be64(NULLDFSBNO);
 
-               if (cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock, PREV.br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur,
-                               PREV.br_startoff + new->br_blockcount,
-                               PREV.br_startblock + new->br_blockcount,
-                               PREV.br_blockcount - new->br_blockcount,
-                               oldext)))
-                               goto done;
-                       if ((error = xfs_btree_decrement(cur, 0, &i)))
-                               goto done;
-                       error = xfs_bmbt_update(cur, LEFT.br_startoff,
-                               LEFT.br_startblock,
-                               LEFT.br_blockcount + new->br_blockcount,
-                               LEFT.br_state);
-                       if (error)
-                               goto done;
-               }
-               break;
-
-       case BMAP_LEFT_FILLING:
-               /*
-                * Setting the first part of a previous oldext extent to newext.
-                * The left neighbor is not contiguous.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               ASSERT(ep && xfs_bmbt_get_state(ep) == oldext);
-               xfs_bmbt_set_startoff(ep, new_endoff);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount - new->br_blockcount);
-               xfs_bmbt_set_startblock(ep,
-                       new->br_startblock + new->br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-
-               xfs_iext_insert(ip, *idx, 1, new, state);
-               ip->i_d.di_nextents++;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock, PREV.br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur,
-                               PREV.br_startoff + new->br_blockcount,
-                               PREV.br_startblock + new->br_blockcount,
-                               PREV.br_blockcount - new->br_blockcount,
-                               oldext)))
-                               goto done;
-                       cur->bc_rec.b = *new;
-                       if ((error = xfs_btree_insert(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+       /*
+        * Need a cursor.  Can't allocate until bb_level is filled in.
+        */
+       mp = ip->i_mount;
+       cur = xfs_bmbt_init_cursor(mp, tp, ip, whichfork);
+       cur->bc_private.b.firstblock = *firstblock;
+       cur->bc_private.b.flist = flist;
+       cur->bc_private.b.flags = wasdel ? XFS_BTCUR_BPRV_WASDEL : 0;
+       /*
+        * Convert to a btree with two levels, one record in root.
+        */
+       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_BTREE);
+       memset(&args, 0, sizeof(args));
+       args.tp = tp;
+       args.mp = mp;
+       args.firstblock = *firstblock;
+       if (*firstblock == NULLFSBLOCK) {
+               args.type = XFS_ALLOCTYPE_START_BNO;
+               args.fsbno = XFS_INO_TO_FSB(mp, ip->i_ino);
+       } else if (flist->xbf_low) {
+               args.type = XFS_ALLOCTYPE_START_BNO;
+               args.fsbno = *firstblock;
+       } else {
+               args.type = XFS_ALLOCTYPE_NEAR_BNO;
+               args.fsbno = *firstblock;
+       }
+       args.minlen = args.maxlen = args.prod = 1;
+       args.wasdel = wasdel;
+       *logflagsp = 0;
+       if ((error = xfs_alloc_vextent(&args))) {
+               xfs_iroot_realloc(ip, -1, whichfork);
+               xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
+               return error;
+       }
+       /*
+        * Allocation can't fail, the space was reserved.
+        */
+       ASSERT(args.fsbno != NULLFSBLOCK);
+       ASSERT(*firstblock == NULLFSBLOCK ||
+              args.agno == XFS_FSB_TO_AGNO(mp, *firstblock) ||
+              (flist->xbf_low &&
+               args.agno > XFS_FSB_TO_AGNO(mp, *firstblock)));
+       *firstblock = cur->bc_private.b.firstblock = args.fsbno;
+       cur->bc_private.b.allocated++;
+       ip->i_d.di_nblocks++;
+       xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, 1L);
+       abp = xfs_btree_get_bufl(mp, tp, args.fsbno, 0);
+       /*
+        * Fill in the child block.
+        */
+       abp->b_ops = &xfs_bmbt_buf_ops;
+       ablock = XFS_BUF_TO_BLOCK(abp);
+       ablock->bb_magic = cpu_to_be32(XFS_BMAP_MAGIC);
+       ablock->bb_level = 0;
+       ablock->bb_u.l.bb_leftsib = cpu_to_be64(NULLDFSBNO);
+       ablock->bb_u.l.bb_rightsib = cpu_to_be64(NULLDFSBNO);
+       arp = XFS_BMBT_REC_ADDR(mp, ablock, 1);
+       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
+       for (cnt = i = 0; i < nextents; i++) {
+               ep = xfs_iext_get_ext(ifp, i);
+               if (!isnullstartblock(xfs_bmbt_get_startblock(ep))) {
+                       arp->l0 = cpu_to_be64(ep->l0);
+                       arp->l1 = cpu_to_be64(ep->l1);
+                       arp++; cnt++;
                }
-               break;
+       }
+       ASSERT(cnt == XFS_IFORK_NEXTENTS(ip, whichfork));
+       xfs_btree_set_numrecs(ablock, cnt);
 
-       case BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
-               /*
-                * Setting the last part of a previous oldext extent to newext.
-                * The right neighbor is contiguous with the new allocation.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount - new->br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       /*
+        * Fill in the root key and pointer.
+        */
+       kp = XFS_BMBT_KEY_ADDR(mp, block, 1);
+       arp = XFS_BMBT_REC_ADDR(mp, ablock, 1);
+       kp->br_startoff = cpu_to_be64(xfs_bmbt_disk_get_startoff(arp));
+       pp = XFS_BMBT_PTR_ADDR(mp, block, 1, xfs_bmbt_get_maxrecs(cur,
+                                               be16_to_cpu(block->bb_level)));
+       *pp = cpu_to_be64(args.fsbno);
 
-               ++*idx;
+       /*
+        * Do all this logging at the end so that
+        * the root is at the right level.
+        */
+       xfs_btree_log_block(cur, abp, XFS_BB_ALL_BITS);
+       xfs_btree_log_recs(cur, abp, 1, be16_to_cpu(ablock->bb_numrecs));
+       ASSERT(*curp == NULL);
+       *curp = cur;
+       *logflagsp = XFS_ILOG_CORE | xfs_ilog_fbroot(whichfork);
+       return 0;
+}
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, *idx),
-                       new->br_startoff, new->br_startblock,
-                       new->br_blockcount + RIGHT.br_blockcount, newext);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+/*
+ * Convert a local file to an extents file.
+ * This code is out of bounds for data forks of regular files,
+ * since the file data needs to get logged so things will stay consistent.
+ * (The bmap-level manipulations are ok, though).
+ */
+STATIC int                             /* error */
+xfs_bmap_local_to_extents(
+       xfs_trans_t     *tp,            /* transaction pointer */
+       xfs_inode_t     *ip,            /* incore inode pointer */
+       xfs_fsblock_t   *firstblock,    /* first block allocated in xaction */
+       xfs_extlen_t    total,          /* total blocks needed by transaction */
+       int             *logflagsp,     /* inode logging flags */
+       int             whichfork,
+       void            (*init_fn)(struct xfs_buf *bp,
+                                  struct xfs_inode *ip,
+                                  struct xfs_ifork *ifp))
+{
+       int             error;          /* error return value */
+       int             flags;          /* logging flags returned */
+       xfs_ifork_t     *ifp;           /* inode fork pointer */
 
-               if (cur == NULL)
-                       rval = XFS_ILOG_DEXT;
-               else {
-                       rval = 0;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock,
-                                       PREV.br_blockcount, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, PREV.br_startoff,
-                               PREV.br_startblock,
-                               PREV.br_blockcount - new->br_blockcount,
-                               oldext)))
-                               goto done;
-                       if ((error = xfs_btree_increment(cur, 0, &i)))
-                               goto done;
-                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
-                               new->br_startblock,
-                               new->br_blockcount + RIGHT.br_blockcount,
-                               newext)))
-                               goto done;
-               }
-               break;
+       /*
+        * We don't want to deal with the case of keeping inode data inline yet.
+        * So sending the data fork of a regular inode is invalid.
+        */
+       ASSERT(!(S_ISREG(ip->i_d.di_mode) && whichfork == XFS_DATA_FORK));
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL);
+       flags = 0;
+       error = 0;
+       if (ifp->if_bytes) {
+               xfs_alloc_arg_t args;   /* allocation arguments */
+               xfs_buf_t       *bp;    /* buffer for extent block */
+               xfs_bmbt_rec_host_t *ep;/* extent record pointer */
 
-       case BMAP_RIGHT_FILLING:
+               ASSERT((ifp->if_flags &
+                       (XFS_IFINLINE|XFS_IFEXTENTS|XFS_IFEXTIREC)) == XFS_IFINLINE);
+               memset(&args, 0, sizeof(args));
+               args.tp = tp;
+               args.mp = ip->i_mount;
+               args.firstblock = *firstblock;
                /*
-                * Setting the last part of a previous oldext extent to newext.
-                * The right neighbor is not contiguous.
+                * Allocate a block.  We know we need only one, since the
+                * file currently fits in an inode.
                 */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep,
-                       PREV.br_blockcount - new->br_blockcount);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               if (*firstblock == NULLFSBLOCK) {
+                       args.fsbno = XFS_INO_TO_FSB(args.mp, ip->i_ino);
+                       args.type = XFS_ALLOCTYPE_START_BNO;
+               } else {
+                       args.fsbno = *firstblock;
+                       args.type = XFS_ALLOCTYPE_NEAR_BNO;
+               }
+               args.total = total;
+               args.minlen = args.maxlen = args.prod = 1;
+               error = xfs_alloc_vextent(&args);
+               if (error)
+                       goto done;
 
-               ++*idx;
-               xfs_iext_insert(ip, *idx, 1, new, state);
+               /* Can't fail, the space was reserved. */
+               ASSERT(args.fsbno != NULLFSBLOCK);
+               ASSERT(args.len == 1);
+               *firstblock = args.fsbno;
+               bp = xfs_btree_get_bufl(args.mp, tp, args.fsbno, 0);
 
-               ip->i_d.di_nextents++;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock, PREV.br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       if ((error = xfs_bmbt_update(cur, PREV.br_startoff,
-                               PREV.br_startblock,
-                               PREV.br_blockcount - new->br_blockcount,
-                               oldext)))
-                               goto done;
-                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       cur->bc_rec.b.br_state = XFS_EXT_NORM;
-                       if ((error = xfs_btree_insert(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-               }
-               break;
-
-       case 0:
-               /*
-                * Setting the middle part of a previous oldext extent to
-                * newext.  Contiguity is impossible here.
-                * One extent becomes three extents.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep,
-                       new->br_startoff - PREV.br_startoff);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               /* initialise the block and copy the data */
+               init_fn(bp, ip, ifp);
 
-               r[0] = *new;
-               r[1].br_startoff = new_endoff;
-               r[1].br_blockcount =
-                       PREV.br_startoff + PREV.br_blockcount - new_endoff;
-               r[1].br_startblock = new->br_startblock + new->br_blockcount;
-               r[1].br_state = oldext;
+               /* account for the change in fork size and log everything */
+               xfs_trans_log_buf(tp, bp, 0, ifp->if_bytes - 1);
+               xfs_bmap_forkoff_reset(args.mp, ip, whichfork);
+               xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
+               xfs_iext_add(ifp, 0, 1);
+               ep = xfs_iext_get_ext(ifp, 0);
+               xfs_bmbt_set_allf(ep, 0, args.fsbno, 1, XFS_EXT_NORM);
+               trace_xfs_bmap_post_update(ip, 0,
+                               whichfork == XFS_ATTR_FORK ? BMAP_ATTRFORK : 0,
+                               _THIS_IP_);
+               XFS_IFORK_NEXT_SET(ip, whichfork, 1);
+               ip->i_d.di_nblocks = 1;
+               xfs_trans_mod_dquot_byino(tp, ip,
+                       XFS_TRANS_DQ_BCOUNT, 1L);
+               flags |= xfs_ilog_fext(whichfork);
+       } else {
+               ASSERT(XFS_IFORK_NEXTENTS(ip, whichfork) == 0);
+               xfs_bmap_forkoff_reset(ip->i_mount, ip, whichfork);
+       }
+       ifp->if_flags &= ~XFS_IFINLINE;
+       ifp->if_flags |= XFS_IFEXTENTS;
+       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_EXTENTS);
+       flags |= XFS_ILOG_CORE;
+done:
+       *logflagsp = flags;
+       return error;
+}
 
-               ++*idx;
-               xfs_iext_insert(ip, *idx, 2, &r[0], state);
+/*
+ * Called from xfs_bmap_add_attrfork to handle btree format files.
+ */
+STATIC int                                     /* error */
+xfs_bmap_add_attrfork_btree(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       xfs_fsblock_t           *firstblock,    /* first block allocated */
+       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
+       int                     *flags)         /* inode logging flags */
+{
+       xfs_btree_cur_t         *cur;           /* btree cursor */
+       int                     error;          /* error return value */
+       xfs_mount_t             *mp;            /* file system mount struct */
+       int                     stat;           /* newroot status */
 
-               ip->i_d.di_nextents += 2;
-               if (cur == NULL)
-                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
-               else {
-                       rval = XFS_ILOG_CORE;
-                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
-                                       PREV.br_startblock, PREV.br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       /* new right extent - oldext */
-                       if ((error = xfs_bmbt_update(cur, r[1].br_startoff,
-                               r[1].br_startblock, r[1].br_blockcount,
-                               r[1].br_state)))
-                               goto done;
-                       /* new left extent - oldext */
-                       cur->bc_rec.b = PREV;
-                       cur->bc_rec.b.br_blockcount =
-                               new->br_startoff - PREV.br_startoff;
-                       if ((error = xfs_btree_insert(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       /*
-                        * Reset the cursor to the position of the new extent
-                        * we are about to insert as we can't trust it after
-                        * the previous insert.
-                        */
-                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
-                                       new->br_startblock, new->br_blockcount,
-                                       &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       /* new middle extent - newext */
-                       cur->bc_rec.b.br_state = new->br_state;
-                       if ((error = xfs_btree_insert(cur, &i)))
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+       mp = ip->i_mount;
+       if (ip->i_df.if_broot_bytes <= XFS_IFORK_DSIZE(ip))
+               *flags |= XFS_ILOG_DBROOT;
+       else {
+               cur = xfs_bmbt_init_cursor(mp, tp, ip, XFS_DATA_FORK);
+               cur->bc_private.b.flist = flist;
+               cur->bc_private.b.firstblock = *firstblock;
+               if ((error = xfs_bmbt_lookup_ge(cur, 0, 0, 0, &stat)))
+                       goto error0;
+               /* must be at least one entry */
+               XFS_WANT_CORRUPTED_GOTO(stat == 1, error0);
+               if ((error = xfs_btree_new_iroot(cur, flags, &stat)))
+                       goto error0;
+               if (stat == 0) {
+                       xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
+                       return XFS_ERROR(ENOSPC);
                }
-               break;
-
-       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_LEFT_FILLING | BMAP_RIGHT_CONTIG:
-       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
-       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-       case BMAP_LEFT_CONTIG:
-       case BMAP_RIGHT_CONTIG:
-               /*
-                * These cases are all impossible.
-                */
-               ASSERT(0);
+               *firstblock = cur->bc_private.b.firstblock;
+               cur->bc_private.b.allocated = 0;
+               xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
        }
+       return 0;
+error0:
+       xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
+       return error;
+}
 
-       /* convert to a btree if necessary */
-       if (xfs_bmap_needs_btree(ip, XFS_DATA_FORK)) {
-               int     tmp_logflags;   /* partial log flag return val */
-
-               ASSERT(cur == NULL);
-               error = xfs_bmap_extents_to_btree(tp, ip, first, flist, &cur,
-                               0, &tmp_logflags, XFS_DATA_FORK);
-               *logflagsp |= tmp_logflags;
-               if (error)
-                       goto done;
-       }
+/*
+ * Called from xfs_bmap_add_attrfork to handle extents format files.
+ */
+STATIC int                                     /* error */
+xfs_bmap_add_attrfork_extents(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       xfs_fsblock_t           *firstblock,    /* first block allocated */
+       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
+       int                     *flags)         /* inode logging flags */
+{
+       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
+       int                     error;          /* error return value */
 
-       /* clear out the allocated field, done with it now in any case. */
+       if (ip->i_d.di_nextents * sizeof(xfs_bmbt_rec_t) <= XFS_IFORK_DSIZE(ip))
+               return 0;
+       cur = NULL;
+       error = xfs_bmap_extents_to_btree(tp, ip, firstblock, flist, &cur, 0,
+               flags, XFS_DATA_FORK);
        if (cur) {
                cur->bc_private.b.allocated = 0;
-               *curp = cur;
+               xfs_btree_del_cursor(cur,
+                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
        }
-
-       xfs_bmap_check_leaf_extents(*curp, ip, XFS_DATA_FORK);
-done:
-       *logflagsp |= rval;
        return error;
-#undef LEFT
-#undef RIGHT
-#undef PREV
 }
 
 /*
- * Convert a hole to a delayed allocation.
+ * Block initialisation functions for local to extent format conversion.
+ * As these get more complex, they will be moved to the relevant files,
+ * but for now they are too simple to worry about.
  */
 STATIC void
-xfs_bmap_add_extent_hole_delay(
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_extnum_t            *idx,   /* extent number to update/insert */
-       xfs_bmbt_irec_t         *new)   /* new data to add to file extents */
+xfs_bmap_local_to_extents_init_fn(
+       struct xfs_buf          *bp,
+       struct xfs_inode        *ip,
+       struct xfs_ifork        *ifp)
 {
-       xfs_ifork_t             *ifp;   /* inode fork pointer */
-       xfs_bmbt_irec_t         left;   /* left neighbor extent entry */
-       xfs_filblks_t           newlen=0;       /* new indirect size */
-       xfs_filblks_t           oldlen=0;       /* old indirect size */
-       xfs_bmbt_irec_t         right;  /* right neighbor extent entry */
-       int                     state;  /* state bits, accessed thru macros */
-       xfs_filblks_t           temp=0; /* temp for indirect calculations */
+       bp->b_ops = &xfs_bmbt_buf_ops;
+       memcpy(bp->b_addr, ifp->if_u1.if_data, ifp->if_bytes);
+}
 
-       ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
-       state = 0;
-       ASSERT(isnullstartblock(new->br_startblock));
+STATIC void
+xfs_symlink_local_to_remote(
+       struct xfs_buf          *bp,
+       struct xfs_inode        *ip,
+       struct xfs_ifork        *ifp)
+{
+       /* remote symlink blocks are not verifiable until CRCs come along */
+       bp->b_ops = NULL;
+       memcpy(bp->b_addr, ifp->if_u1.if_data, ifp->if_bytes);
+}
 
-       /*
-        * Check and set flags if this segment has a left neighbor
-        */
-       if (*idx > 0) {
-               state |= BMAP_LEFT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx - 1), &left);
+/*
+ * Called from xfs_bmap_add_attrfork to handle local format files. Each
+ * different data fork content type needs a different callout to do the
+ * conversion. Some are basic and only require special block initialisation
+ * callouts for the data formating, others (directories) are so specialised they
+ * handle everything themselves.
+ *
+ * XXX (dgc): investigate whether directory conversion can use the generic
+ * formatting callout. It should be possible - it's just a very complex
+ * formatter. it would also require passing the transaction through to the init
+ * function.
+ */
+STATIC int                                     /* error */
+xfs_bmap_add_attrfork_local(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       xfs_fsblock_t           *firstblock,    /* first block allocated */
+       xfs_bmap_free_t         *flist,         /* blocks to free at commit */
+       int                     *flags)         /* inode logging flags */
+{
+       xfs_da_args_t           dargs;          /* args for dir/attr code */
 
-               if (isnullstartblock(left.br_startblock))
-                       state |= BMAP_LEFT_DELAY;
-       }
-
-       /*
-        * Check and set flags if the current (right) segment exists.
-        * If it doesn't exist, we're converting the hole at end-of-file.
-        */
-       if (*idx < ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t)) {
-               state |= BMAP_RIGHT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx), &right);
+       if (ip->i_df.if_bytes <= XFS_IFORK_DSIZE(ip))
+               return 0;
 
-               if (isnullstartblock(right.br_startblock))
-                       state |= BMAP_RIGHT_DELAY;
+       if (S_ISDIR(ip->i_d.di_mode)) {
+               memset(&dargs, 0, sizeof(dargs));
+               dargs.dp = ip;
+               dargs.firstblock = firstblock;
+               dargs.flist = flist;
+               dargs.total = ip->i_mount->m_dirblkfsbs;
+               dargs.whichfork = XFS_DATA_FORK;
+               dargs.trans = tp;
+               return xfs_dir2_sf_to_block(&dargs);
        }
 
-       /*
-        * Set contiguity flags on the left and right neighbors.
-        * Don't let extents get too large, even if the pieces are contiguous.
-        */
-       if ((state & BMAP_LEFT_VALID) && (state & BMAP_LEFT_DELAY) &&
-           left.br_startoff + left.br_blockcount == new->br_startoff &&
-           left.br_blockcount + new->br_blockcount <= MAXEXTLEN)
-               state |= BMAP_LEFT_CONTIG;
-
-       if ((state & BMAP_RIGHT_VALID) && (state & BMAP_RIGHT_DELAY) &&
-           new->br_startoff + new->br_blockcount == right.br_startoff &&
-           new->br_blockcount + right.br_blockcount <= MAXEXTLEN &&
-           (!(state & BMAP_LEFT_CONTIG) ||
-            (left.br_blockcount + new->br_blockcount +
-             right.br_blockcount <= MAXEXTLEN)))
-               state |= BMAP_RIGHT_CONTIG;
+       if (S_ISLNK(ip->i_d.di_mode))
+               return xfs_bmap_local_to_extents(tp, ip, firstblock, 1,
+                                                flags, XFS_DATA_FORK,
+                                                xfs_symlink_local_to_remote);
 
-       /*
-        * Switch out based on the contiguity flags.
-        */
-       switch (state & (BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG)) {
-       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-               /*
-                * New allocation is contiguous with delayed allocations
-                * on the left and on the right.
-                * Merge all three into a single extent record.
-                */
-               --*idx;
-               temp = left.br_blockcount + new->br_blockcount +
-                       right.br_blockcount;
+       return xfs_bmap_local_to_extents(tp, ip, firstblock, 1, flags,
+                                        XFS_DATA_FORK,
+                                        xfs_bmap_local_to_extents_init_fn);
+}
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx), temp);
-               oldlen = startblockval(left.br_startblock) +
-                       startblockval(new->br_startblock) +
-                       startblockval(right.br_startblock);
-               newlen = xfs_bmap_worst_indlen(ip, temp);
-               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, *idx),
-                       nullstartblock((int)newlen));
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+/*
+ * Convert inode from non-attributed to attributed.
+ * Must not be in a transaction, ip must not be locked.
+ */
+int                                            /* error code */
+xfs_bmap_add_attrfork(
+       xfs_inode_t             *ip,            /* incore inode pointer */
+       int                     size,           /* space new attribute needs */
+       int                     rsvd)           /* xact may use reserved blks */
+{
+       xfs_fsblock_t           firstblock;     /* 1st block/ag allocated */
+       xfs_bmap_free_t         flist;          /* freed extent records */
+       xfs_mount_t             *mp;            /* mount structure */
+       xfs_trans_t             *tp;            /* transaction pointer */
+       int                     blks;           /* space reservation */
+       int                     version = 1;    /* superblock attr version */
+       int                     committed;      /* xaction was committed */
+       int                     logflags;       /* logging flags */
+       int                     error;          /* error return value */
 
-               xfs_iext_remove(ip, *idx + 1, 1, state);
-               break;
+       ASSERT(XFS_IFORK_Q(ip) == 0);
 
-       case BMAP_LEFT_CONTIG:
+       mp = ip->i_mount;
+       ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
+       tp = xfs_trans_alloc(mp, XFS_TRANS_ADDAFORK);
+       blks = XFS_ADDAFORK_SPACE_RES(mp);
+       if (rsvd)
+               tp->t_flags |= XFS_TRANS_RESERVE;
+       if ((error = xfs_trans_reserve(tp, blks, XFS_ADDAFORK_LOG_RES(mp), 0,
+                       XFS_TRANS_PERM_LOG_RES, XFS_ADDAFORK_LOG_COUNT)))
+               goto error0;
+       xfs_ilock(ip, XFS_ILOCK_EXCL);
+       error = xfs_trans_reserve_quota_nblks(tp, ip, blks, 0, rsvd ?
+                       XFS_QMOPT_RES_REGBLKS | XFS_QMOPT_FORCE_RES :
+                       XFS_QMOPT_RES_REGBLKS);
+       if (error) {
+               xfs_iunlock(ip, XFS_ILOCK_EXCL);
+               xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES);
+               return error;
+       }
+       if (XFS_IFORK_Q(ip))
+               goto error1;
+       if (ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS) {
                /*
-                * New allocation is contiguous with a delayed allocation
-                * on the left.
-                * Merge the new allocation with the left neighbor.
+                * For inodes coming from pre-6.2 filesystems.
                 */
-               --*idx;
-               temp = left.br_blockcount + new->br_blockcount;
+               ASSERT(ip->i_d.di_aformat == 0);
+               ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
+       }
+       ASSERT(ip->i_d.di_anextents == 0);
 
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx), temp);
-               oldlen = startblockval(left.br_startblock) +
-                       startblockval(new->br_startblock);
-               newlen = xfs_bmap_worst_indlen(ip, temp);
-               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, *idx),
-                       nullstartblock((int)newlen));
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-               break;
+       xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
+       xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 
-       case BMAP_RIGHT_CONTIG:
-               /*
-                * New allocation is contiguous with a delayed allocation
-                * on the right.
-                * Merge the new allocation with the right neighbor.
-                */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               temp = new->br_blockcount + right.br_blockcount;
-               oldlen = startblockval(new->br_startblock) +
-                       startblockval(right.br_startblock);
-               newlen = xfs_bmap_worst_indlen(ip, temp);
-               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, *idx),
-                       new->br_startoff,
-                       nullstartblock((int)newlen), temp, right.br_state);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+       switch (ip->i_d.di_format) {
+       case XFS_DINODE_FMT_DEV:
+               ip->i_d.di_forkoff = roundup(sizeof(xfs_dev_t), 8) >> 3;
+               break;
+       case XFS_DINODE_FMT_UUID:
+               ip->i_d.di_forkoff = roundup(sizeof(uuid_t), 8) >> 3;
+               break;
+       case XFS_DINODE_FMT_LOCAL:
+       case XFS_DINODE_FMT_EXTENTS:
+       case XFS_DINODE_FMT_BTREE:
+               ip->i_d.di_forkoff = xfs_attr_shortform_bytesfit(ip, size);
+               if (!ip->i_d.di_forkoff)
+                       ip->i_d.di_forkoff = xfs_default_attroffset(ip) >> 3;
+               else if (mp->m_flags & XFS_MOUNT_ATTR2)
+                       version = 2;
                break;
+       default:
+               ASSERT(0);
+               error = XFS_ERROR(EINVAL);
+               goto error1;
+       }
 
-       case 0:
-               /*
-                * New allocation is not contiguous with another
-                * delayed allocation.
-                * Insert a new entry.
-                */
-               oldlen = newlen = 0;
-               xfs_iext_insert(ip, *idx, 1, new, state);
+       ASSERT(ip->i_afp == NULL);
+       ip->i_afp = kmem_zone_zalloc(xfs_ifork_zone, KM_SLEEP);
+       ip->i_afp->if_flags = XFS_IFEXTENTS;
+       logflags = 0;
+       xfs_bmap_init(&flist, &firstblock);
+       switch (ip->i_d.di_format) {
+       case XFS_DINODE_FMT_LOCAL:
+               error = xfs_bmap_add_attrfork_local(tp, ip, &firstblock, &flist,
+                       &logflags);
+               break;
+       case XFS_DINODE_FMT_EXTENTS:
+               error = xfs_bmap_add_attrfork_extents(tp, ip, &firstblock,
+                       &flist, &logflags);
+               break;
+       case XFS_DINODE_FMT_BTREE:
+               error = xfs_bmap_add_attrfork_btree(tp, ip, &firstblock, &flist,
+                       &logflags);
+               break;
+       default:
+               error = 0;
                break;
        }
-       if (oldlen != newlen) {
-               ASSERT(oldlen > newlen);
-               xfs_icsb_modify_counters(ip->i_mount, XFS_SBS_FDBLOCKS,
-                       (int64_t)(oldlen - newlen), 0);
-               /*
-                * Nothing to do for disk quota accounting here.
-                */
+       if (logflags)
+               xfs_trans_log_inode(tp, ip, logflags);
+       if (error)
+               goto error2;
+       if (!xfs_sb_version_hasattr(&mp->m_sb) ||
+          (!xfs_sb_version_hasattr2(&mp->m_sb) && version == 2)) {
+               __int64_t sbfields = 0;
+
+               spin_lock(&mp->m_sb_lock);
+               if (!xfs_sb_version_hasattr(&mp->m_sb)) {
+                       xfs_sb_version_addattr(&mp->m_sb);
+                       sbfields |= XFS_SB_VERSIONNUM;
+               }
+               if (!xfs_sb_version_hasattr2(&mp->m_sb) && version == 2) {
+                       xfs_sb_version_addattr2(&mp->m_sb);
+                       sbfields |= (XFS_SB_VERSIONNUM | XFS_SB_FEATURES2);
+               }
+               if (sbfields) {
+                       spin_unlock(&mp->m_sb_lock);
+                       xfs_mod_sb(tp, sbfields);
+               } else
+                       spin_unlock(&mp->m_sb_lock);
        }
+
+       error = xfs_bmap_finish(&tp, &flist, &committed);
+       if (error)
+               goto error2;
+       return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
+error2:
+       xfs_bmap_cancel(&flist);
+error1:
+       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+error0:
+       xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
+       return error;
 }
 
 /*
- * Convert a hole to a real allocation.
+ * Internal and external extent tree search functions.
  */
-STATIC int                             /* error */
-xfs_bmap_add_extent_hole_real(
-       struct xfs_bmalloca     *bma,
-       int                     whichfork)
-{
-       struct xfs_bmbt_irec    *new = &bma->got;
-       int                     error;  /* error return value */
-       int                     i;      /* temp state */
-       xfs_ifork_t             *ifp;   /* inode fork pointer */
-       xfs_bmbt_irec_t         left;   /* left neighbor extent entry */
-       xfs_bmbt_irec_t         right;  /* right neighbor extent entry */
-       int                     rval=0; /* return value (logging flags) */
-       int                     state;  /* state bits, accessed thru macros */
-
-       ifp = XFS_IFORK_PTR(bma->ip, whichfork);
 
-       ASSERT(bma->idx >= 0);
-       ASSERT(bma->idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
-       ASSERT(!isnullstartblock(new->br_startblock));
-       ASSERT(!bma->cur ||
-              !(bma->cur->bc_private.b.flags & XFS_BTCUR_BPRV_WASDEL));
-
-       XFS_STATS_INC(xs_add_exlist);
-
-       state = 0;
-       if (whichfork == XFS_ATTR_FORK)
-               state |= BMAP_ATTRFORK;
+/*
+ * Read in the extents to if_extents.
+ * All inode fields are set up by caller, we just traverse the btree
+ * and copy the records in. If the file system cannot contain unwritten
+ * extents, the records are checked for no "state" flags.
+ */
+int                                    /* error */
+xfs_bmap_read_extents(
+       xfs_trans_t             *tp,    /* transaction pointer */
+       xfs_inode_t             *ip,    /* incore inode */
+       int                     whichfork) /* data or attr fork */
+{
+       struct xfs_btree_block  *block; /* current btree block */
+       xfs_fsblock_t           bno;    /* block # of "block" */
+       xfs_buf_t               *bp;    /* buffer for "block" */
+       int                     error;  /* error return value */
+       xfs_exntfmt_t           exntf;  /* XFS_EXTFMT_NOSTATE, if checking */
+       xfs_extnum_t            i, j;   /* index into the extents list */
+       xfs_ifork_t             *ifp;   /* fork structure */
+       int                     level;  /* btree level, for checking */
+       xfs_mount_t             *mp;    /* file system mount structure */
+       __be64                  *pp;    /* pointer to block address */
+       /* REFERENCED */
+       xfs_extnum_t            room;   /* number of entries there's room for */
 
+       bno = NULLFSBLOCK;
+       mp = ip->i_mount;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       exntf = (whichfork != XFS_DATA_FORK) ? XFS_EXTFMT_NOSTATE :
+                                       XFS_EXTFMT_INODE(ip);
+       block = ifp->if_broot;
        /*
-        * Check and set flags if this segment has a left neighbor.
+        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
         */
-       if (bma->idx > 0) {
-               state |= BMAP_LEFT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1), &left);
-               if (isnullstartblock(left.br_startblock))
-                       state |= BMAP_LEFT_DELAY;
-       }
-
+       level = be16_to_cpu(block->bb_level);
+       ASSERT(level > 0);
+       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
+       bno = be64_to_cpu(*pp);
+       ASSERT(bno != NULLDFSBNO);
+       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
+       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
        /*
-        * Check and set flags if this segment has a current value.
-        * Not true if we're inserting into the "hole" at eof.
+        * Go down the tree until leaf level is reached, following the first
+        * pointer (leftmost) at each level.
         */
-       if (bma->idx < ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)) {
-               state |= BMAP_RIGHT_VALID;
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &right);
-               if (isnullstartblock(right.br_startblock))
-                       state |= BMAP_RIGHT_DELAY;
+       while (level-- > 0) {
+               error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
+                               XFS_BMAP_BTREE_REF, &xfs_bmbt_buf_ops);
+               if (error)
+                       return error;
+               block = XFS_BUF_TO_BLOCK(bp);
+               XFS_WANT_CORRUPTED_GOTO(
+                       xfs_bmap_sanity_check(mp, bp, level),
+                       error0);
+               if (level == 0)
+                       break;
+               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
+               bno = be64_to_cpu(*pp);
+               XFS_WANT_CORRUPTED_GOTO(XFS_FSB_SANITY_CHECK(mp, bno), error0);
+               xfs_trans_brelse(tp, bp);
        }
-
        /*
-        * We're inserting a real allocation between "left" and "right".
-        * Set the contiguity flags.  Don't let extents get too large.
+        * Here with bp and block set to the leftmost leaf node in the tree.
         */
-       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
-           left.br_startoff + left.br_blockcount == new->br_startoff &&
-           left.br_startblock + left.br_blockcount == new->br_startblock &&
-           left.br_state == new->br_state &&
-           left.br_blockcount + new->br_blockcount <= MAXEXTLEN)
-               state |= BMAP_LEFT_CONTIG;
-
-       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
-           new->br_startoff + new->br_blockcount == right.br_startoff &&
-           new->br_startblock + new->br_blockcount == right.br_startblock &&
-           new->br_state == right.br_state &&
-           new->br_blockcount + right.br_blockcount <= MAXEXTLEN &&
-           (!(state & BMAP_LEFT_CONTIG) ||
-            left.br_blockcount + new->br_blockcount +
-            right.br_blockcount <= MAXEXTLEN))
-               state |= BMAP_RIGHT_CONTIG;
-
-       error = 0;
+       room = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
+       i = 0;
        /*
-        * Select which case we're in here, and implement it.
+        * Loop over all leaf nodes.  Copy information to the extent records.
         */
-       switch (state & (BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG)) {
-       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
-               /*
-                * New allocation is contiguous with real allocations on the
-                * left and on the right.
-                * Merge all three into a single extent record.
-                */
-               --bma->idx;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
-                       left.br_blockcount + new->br_blockcount +
-                       right.br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
-
-               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
+       for (;;) {
+               xfs_bmbt_rec_t  *frp;
+               xfs_fsblock_t   nextbno;
+               xfs_extnum_t    num_recs;
+               xfs_extnum_t    start;
 
-               XFS_IFORK_NEXT_SET(bma->ip, whichfork,
-                       XFS_IFORK_NEXTENTS(bma->ip, whichfork) - 1);
-               if (bma->cur == NULL) {
-                       rval = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
-               } else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur, right.br_startoff,
-                                       right.br_startblock, right.br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_btree_delete(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_btree_decrement(bma->cur, 0, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, left.br_startoff,
-                                       left.br_startblock,
-                                       left.br_blockcount +
-                                               new->br_blockcount +
-                                               right.br_blockcount,
-                                       left.br_state);
-                       if (error)
-                               goto done;
+               num_recs = xfs_btree_get_numrecs(block);
+               if (unlikely(i + num_recs > room)) {
+                       ASSERT(i + num_recs <= room);
+                       xfs_warn(ip->i_mount,
+                               "corrupt dinode %Lu, (btree extents).",
+                               (unsigned long long) ip->i_ino);
+                       XFS_CORRUPTION_ERROR("xfs_bmap_read_extents(1)",
+                               XFS_ERRLEVEL_LOW, ip->i_mount, block);
+                       goto error0;
                }
-               break;
-
-       case BMAP_LEFT_CONTIG:
+               XFS_WANT_CORRUPTED_GOTO(
+                       xfs_bmap_sanity_check(mp, bp, 0),
+                       error0);
                /*
-                * New allocation is contiguous with a real allocation
-                * on the left.
-                * Merge the new allocation with the left neighbor.
+                * Read-ahead the next leaf block, if any.
                 */
-               --bma->idx;
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
-                       left.br_blockcount + new->br_blockcount);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
-
-               if (bma->cur == NULL) {
-                       rval = xfs_ilog_fext(whichfork);
-               } else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur, left.br_startoff,
-                                       left.br_startblock, left.br_blockcount,
-                                       &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, left.br_startoff,
-                                       left.br_startblock,
-                                       left.br_blockcount +
-                                               new->br_blockcount,
-                                       left.br_state);
-                       if (error)
-                               goto done;
-               }
-               break;
-
-       case BMAP_RIGHT_CONTIG:
+               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
+               if (nextbno != NULLFSBLOCK)
+                       xfs_btree_reada_bufl(mp, nextbno, 1,
+                                            &xfs_bmbt_buf_ops);
                /*
-                * New allocation is contiguous with a real allocation
-                * on the right.
-                * Merge the new allocation with the right neighbor.
+                * Copy records into the extent records.
                 */
-               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
-               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, bma->idx),
-                       new->br_startoff, new->br_startblock,
-                       new->br_blockcount + right.br_blockcount,
-                       right.br_state);
-               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
-
-               if (bma->cur == NULL) {
-                       rval = xfs_ilog_fext(whichfork);
-               } else {
-                       rval = 0;
-                       error = xfs_bmbt_lookup_eq(bma->cur,
-                                       right.br_startoff,
-                                       right.br_startblock,
-                                       right.br_blockcount, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       error = xfs_bmbt_update(bma->cur, new->br_startoff,
-                                       new->br_startblock,
-                                       new->br_blockcount +
-                                               right.br_blockcount,
-                                       right.br_state);
-                       if (error)
-                               goto done;
+               frp = XFS_BMBT_REC_ADDR(mp, block, 1);
+               start = i;
+               for (j = 0; j < num_recs; j++, i++, frp++) {
+                       xfs_bmbt_rec_host_t *trp = xfs_iext_get_ext(ifp, i);
+                       trp->l0 = be64_to_cpu(frp->l0);
+                       trp->l1 = be64_to_cpu(frp->l1);
                }
-               break;
-
-       case 0:
+               if (exntf == XFS_EXTFMT_NOSTATE) {
+                       /*
+                        * Check all attribute bmap btree records and
+                        * any "older" data bmap btree records for a
+                        * set bit in the "extent flag" position.
+                        */
+                       if (unlikely(xfs_check_nostate_extents(ifp,
+                                       start, num_recs))) {
+                               XFS_ERROR_REPORT("xfs_bmap_read_extents(2)",
+                                                XFS_ERRLEVEL_LOW,
+                                                ip->i_mount);
+                               goto error0;
+                       }
+               }
+               xfs_trans_brelse(tp, bp);
+               bno = nextbno;
                /*
-                * New allocation is not contiguous with another
-                * real allocation.
-                * Insert a new entry.
+                * If we've reached the end, stop.
                 */
-               xfs_iext_insert(bma->ip, bma->idx, 1, new, state);
-               XFS_IFORK_NEXT_SET(bma->ip, whichfork,
-                       XFS_IFORK_NEXTENTS(bma->ip, whichfork) + 1);
-               if (bma->cur == NULL) {
-                       rval = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
-               } else {
-                       rval = XFS_ILOG_CORE;
-                       error = xfs_bmbt_lookup_eq(bma->cur,
-                                       new->br_startoff,
-                                       new->br_startblock,
-                                       new->br_blockcount, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
-                       bma->cur->bc_rec.b.br_state = new->br_state;
-                       error = xfs_btree_insert(bma->cur, &i);
-                       if (error)
-                               goto done;
-                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-               }
-               break;
+               if (bno == NULLFSBLOCK)
+                       break;
+               error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
+                               XFS_BMAP_BTREE_REF, &xfs_bmbt_buf_ops);
+               if (error)
+                       return error;
+               block = XFS_BUF_TO_BLOCK(bp);
        }
+       ASSERT(i == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)));
+       ASSERT(i == XFS_IFORK_NEXTENTS(ip, whichfork));
+       XFS_BMAP_TRACE_EXLIST(ip, i, whichfork);
+       return 0;
+error0:
+       xfs_trans_brelse(tp, bp);
+       return XFS_ERROR(EFSCORRUPTED);
+}
 
-       /* convert to a btree if necessary */
-       if (xfs_bmap_needs_btree(bma->ip, whichfork)) {
-               int     tmp_logflags;   /* partial log flag return val */
 
-               ASSERT(bma->cur == NULL);
-               error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
-                               bma->firstblock, bma->flist, &bma->cur,
-                               0, &tmp_logflags, whichfork);
-               bma->logflags |= tmp_logflags;
-               if (error)
-                       goto done;
-       }
+/*
+ * Search the extent records for the entry containing block bno.
+ * If bno lies in a hole, point to the next entry.  If bno lies
+ * past eof, *eofp will be set, and *prevp will contain the last
+ * entry (null if none).  Else, *lastxp will be set to the index
+ * of the found entry; *gotp will contain the entry.
+ */
+STATIC xfs_bmbt_rec_host_t *           /* pointer to found extent entry */
+xfs_bmap_search_multi_extents(
+       xfs_ifork_t     *ifp,           /* inode fork pointer */
+       xfs_fileoff_t   bno,            /* block number searched for */
+       int             *eofp,          /* out: end of file found */
+       xfs_extnum_t    *lastxp,        /* out: last extent index */
+       xfs_bmbt_irec_t *gotp,          /* out: extent entry found */
+       xfs_bmbt_irec_t *prevp)         /* out: previous extent entry found */
+{
+       xfs_bmbt_rec_host_t *ep;                /* extent record pointer */
+       xfs_extnum_t    lastx;          /* last extent index */
 
-       /* clear out the allocated field, done with it now in any case. */
-       if (bma->cur)
-               bma->cur->bc_private.b.allocated = 0;
+       /*
+        * Initialize the extent entry structure to catch access to
+        * uninitialized br_startblock field.
+        */
+       gotp->br_startoff = 0xffa5a5a5a5a5a5a5LL;
+       gotp->br_blockcount = 0xa55a5a5a5a5a5a5aLL;
+       gotp->br_state = XFS_EXT_INVALID;
+#if XFS_BIG_BLKNOS
+       gotp->br_startblock = 0xffffa5a5a5a5a5a5LL;
+#else
+       gotp->br_startblock = 0xffffa5a5;
+#endif
+       prevp->br_startoff = NULLFILEOFF;
 
-       xfs_bmap_check_leaf_extents(bma->cur, bma->ip, whichfork);
-done:
-       bma->logflags |= rval;
-       return error;
+       ep = xfs_iext_bno_to_ext(ifp, bno, &lastx);
+       if (lastx > 0) {
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx - 1), prevp);
+       }
+       if (lastx < (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) {
+               xfs_bmbt_get_all(ep, gotp);
+               *eofp = 0;
+       } else {
+               if (lastx > 0) {
+                       *gotp = *prevp;
+               }
+               *eofp = 1;
+               ep = NULL;
+       }
+       *lastxp = lastx;
+       return ep;
 }
 
 /*
- * Adjust the size of the new extent based on di_extsize and rt extsize.
+ * Search the extents list for the inode, for the extent containing bno.
+ * If bno lies in a hole, point to the next entry.  If bno lies past eof,
+ * *eofp will be set, and *prevp will contain the last entry (null if none).
+ * Else, *lastxp will be set to the index of the found
+ * entry; *gotp will contain the entry.
  */
-STATIC int
-xfs_bmap_extsize_align(
-       xfs_mount_t     *mp,
-       xfs_bmbt_irec_t *gotp,          /* next extent pointer */
-       xfs_bmbt_irec_t *prevp,         /* previous extent pointer */
-       xfs_extlen_t    extsz,          /* align to this extent size */
-       int             rt,             /* is this a realtime inode? */
-       int             eof,            /* is extent at end-of-file? */
-       int             delay,          /* creating delalloc extent? */
-       int             convert,        /* overwriting unwritten extent? */
-       xfs_fileoff_t   *offp,          /* in/out: aligned offset */
-       xfs_extlen_t    *lenp)          /* in/out: aligned length */
+STATIC xfs_bmbt_rec_host_t *                 /* pointer to found extent entry */
+xfs_bmap_search_extents(
+       xfs_inode_t     *ip,            /* incore inode pointer */
+       xfs_fileoff_t   bno,            /* block number searched for */
+       int             fork,           /* data or attr fork */
+       int             *eofp,          /* out: end of file found */
+       xfs_extnum_t    *lastxp,        /* out: last extent index */
+       xfs_bmbt_irec_t *gotp,          /* out: extent entry found */
+       xfs_bmbt_irec_t *prevp)         /* out: previous extent entry found */
 {
-       xfs_fileoff_t   orig_off;       /* original offset */
-       xfs_extlen_t    orig_alen;      /* original length */
-       xfs_fileoff_t   orig_end;       /* original off+len */
-       xfs_fileoff_t   nexto;          /* next file offset */
-       xfs_fileoff_t   prevo;          /* previous file offset */
-       xfs_fileoff_t   align_off;      /* temp for offset */
-       xfs_extlen_t    align_alen;     /* temp for length */
-       xfs_extlen_t    temp;           /* temp for calculations */
+       xfs_ifork_t     *ifp;           /* inode fork pointer */
+       xfs_bmbt_rec_host_t  *ep;            /* extent record pointer */
 
-       if (convert)
-               return 0;
+       XFS_STATS_INC(xs_look_exlist);
+       ifp = XFS_IFORK_PTR(ip, fork);
 
-       orig_off = align_off = *offp;
-       orig_alen = align_alen = *lenp;
-       orig_end = orig_off + orig_alen;
+       ep = xfs_bmap_search_multi_extents(ifp, bno, eofp, lastxp, gotp, prevp);
 
-       /*
-        * If this request overlaps an existing extent, then don't
-        * attempt to perform any additional alignment.
-        */
-       if (!delay && !eof &&
-           (orig_off >= gotp->br_startoff) &&
-           (orig_end <= gotp->br_startoff + gotp->br_blockcount)) {
-               return 0;
+       if (unlikely(!(gotp->br_startblock) && (*lastxp != NULLEXTNUM) &&
+                    !(XFS_IS_REALTIME_INODE(ip) && fork == XFS_DATA_FORK))) {
+               xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
+                               "Access to block zero in inode %llu "
+                               "start_block: %llx start_off: %llx "
+                               "blkcnt: %llx extent-state: %x lastx: %x\n",
+                       (unsigned long long)ip->i_ino,
+                       (unsigned long long)gotp->br_startblock,
+                       (unsigned long long)gotp->br_startoff,
+                       (unsigned long long)gotp->br_blockcount,
+                       gotp->br_state, *lastxp);
+               *lastxp = NULLEXTNUM;
+               *eofp = 1;
+               return NULL;
        }
+       return ep;
+}
 
-       /*
-        * If the file offset is unaligned vs. the extent size
-        * we need to align it.  This will be possible unless
-        * the file was previously written with a kernel that didn't
-        * perform this alignment, or if a truncate shot us in the
-        * foot.
-        */
-       temp = do_mod(orig_off, extsz);
-       if (temp) {
-               align_alen += temp;
-               align_off -= temp;
+/*
+ * Returns the file-relative block number of the first unused block(s)
+ * in the file with at least "len" logically contiguous blocks free.
+ * This is the lowest-address hole if the file has holes, else the first block
+ * past the end of file.
+ * Return 0 if the file is currently local (in-inode).
+ */
+int                                            /* error */
+xfs_bmap_first_unused(
+       xfs_trans_t     *tp,                    /* transaction pointer */
+       xfs_inode_t     *ip,                    /* incore inode */
+       xfs_extlen_t    len,                    /* size of hole to find */
+       xfs_fileoff_t   *first_unused,          /* unused block */
+       int             whichfork)              /* data or attr fork */
+{
+       int             error;                  /* error return value */
+       int             idx;                    /* extent record index */
+       xfs_ifork_t     *ifp;                   /* inode fork pointer */
+       xfs_fileoff_t   lastaddr;               /* last block number seen */
+       xfs_fileoff_t   lowest;                 /* lowest useful block */
+       xfs_fileoff_t   max;                    /* starting useful block */
+       xfs_fileoff_t   off;                    /* offset for this block */
+       xfs_extnum_t    nextents;               /* number of extent entries */
+
+       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE ||
+              XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS ||
+              XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL);
+       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
+               *first_unused = 0;
+               return 0;
        }
-       /*
-        * Same adjustment for the end of the requested area.
-        */
-       if ((temp = (align_alen % extsz))) {
-               align_alen += extsz - temp;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
+           (error = xfs_iread_extents(tp, ip, whichfork)))
+               return error;
+       lowest = *first_unused;
+       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
+       for (idx = 0, lastaddr = 0, max = lowest; idx < nextents; idx++) {
+               xfs_bmbt_rec_host_t *ep = xfs_iext_get_ext(ifp, idx);
+               off = xfs_bmbt_get_startoff(ep);
+               /*
+                * See if the hole before this extent will work.
+                */
+               if (off >= lowest + len && off - max >= len) {
+                       *first_unused = max;
+                       return 0;
+               }
+               lastaddr = off + xfs_bmbt_get_blockcount(ep);
+               max = XFS_FILEOFF_MAX(lastaddr, lowest);
        }
-       /*
-        * If the previous block overlaps with this proposed allocation
-        * then move the start forward without adjusting the length.
-        */
-       if (prevp->br_startoff != NULLFILEOFF) {
-               if (prevp->br_startblock == HOLESTARTBLOCK)
-                       prevo = prevp->br_startoff;
-               else
-                       prevo = prevp->br_startoff + prevp->br_blockcount;
-       } else
-               prevo = 0;
-       if (align_off != orig_off && align_off < prevo)
-               align_off = prevo;
-       /*
-        * If the next block overlaps with this proposed allocation
-        * then move the start back without adjusting the length,
-        * but not before offset 0.
-        * This may of course make the start overlap previous block,
-        * and if we hit the offset 0 limit then the next block
-        * can still overlap too.
-        */
-       if (!eof && gotp->br_startoff != NULLFILEOFF) {
-               if ((delay && gotp->br_startblock == HOLESTARTBLOCK) ||
-                   (!delay && gotp->br_startblock == DELAYSTARTBLOCK))
-                       nexto = gotp->br_startoff + gotp->br_blockcount;
+       *first_unused = max;
+       return 0;
+}
+
+/*
+ * Returns the file-relative block number of the last block + 1 before
+ * last_block (input value) in the file.
+ * This is not based on i_size, it is based on the extent records.
+ * Returns 0 for local files, as they do not have extent records.
+ */
+int                                            /* error */
+xfs_bmap_last_before(
+       xfs_trans_t     *tp,                    /* transaction pointer */
+       xfs_inode_t     *ip,                    /* incore inode */
+       xfs_fileoff_t   *last_block,            /* last block */
+       int             whichfork)              /* data or attr fork */
+{
+       xfs_fileoff_t   bno;                    /* input file offset */
+       int             eof;                    /* hit end of file */
+       xfs_bmbt_rec_host_t *ep;                /* pointer to last extent */
+       int             error;                  /* error return value */
+       xfs_bmbt_irec_t got;                    /* current extent value */
+       xfs_ifork_t     *ifp;                   /* inode fork pointer */
+       xfs_extnum_t    lastx;                  /* last extent used */
+       xfs_bmbt_irec_t prev;                   /* previous extent value */
+
+       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
+              return XFS_ERROR(EIO);
+       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
+               *last_block = 0;
+               return 0;
+       }
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
+           (error = xfs_iread_extents(tp, ip, whichfork)))
+               return error;
+       bno = *last_block - 1;
+       ep = xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got,
+               &prev);
+       if (eof || xfs_bmbt_get_startoff(ep) > bno) {
+               if (prev.br_startoff == NULLFILEOFF)
+                       *last_block = 0;
                else
-                       nexto = gotp->br_startoff;
-       } else
-               nexto = NULLFILEOFF;
-       if (!eof &&
-           align_off + align_alen != orig_end &&
-           align_off + align_alen > nexto)
-               align_off = nexto > align_alen ? nexto - align_alen : 0;
-       /*
-        * If we're now overlapping the next or previous extent that
-        * means we can't fit an extsz piece in this hole.  Just move
-        * the start forward to the first valid spot and set
-        * the length so we hit the end.
-        */
-       if (align_off != orig_off && align_off < prevo)
-               align_off = prevo;
-       if (align_off + align_alen != orig_end &&
-           align_off + align_alen > nexto &&
-           nexto != NULLFILEOFF) {
-               ASSERT(nexto > prevo);
-               align_alen = nexto - align_off;
+                       *last_block = prev.br_startoff + prev.br_blockcount;
        }
-
        /*
-        * If realtime, and the result isn't a multiple of the realtime
-        * extent size we need to remove blocks until it is.
+        * Otherwise *last_block is already the right answer.
         */
-       if (rt && (temp = (align_alen % mp->m_sb.sb_rextsize))) {
-               /*
-                * We're not covering the original request, or
-                * we won't be able to once we fix the length.
-                */
-               if (orig_off < align_off ||
-                   orig_end > align_off + align_alen ||
-                   align_alen - temp < orig_alen)
-                       return XFS_ERROR(EINVAL);
-               /*
-                * Try to fix it by moving the start up.
-                */
-               if (align_off + temp <= orig_off) {
-                       align_alen -= temp;
-                       align_off += temp;
-               }
-               /*
-                * Try to fix it by moving the end in.
-                */
-               else if (align_off + align_alen - temp >= orig_end)
-                       align_alen -= temp;
-               /*
-                * Set the start to the minimum then trim the length.
-                */
-               else {
-                       align_alen -= orig_off - align_off;
-                       align_off = orig_off;
-                       align_alen -= align_alen % mp->m_sb.sb_rextsize;
-               }
-               /*
-                * Result doesn't cover the request, fail it.
-                */
-               if (orig_off < align_off || orig_end > align_off + align_alen)
-                       return XFS_ERROR(EINVAL);
-       } else {
-               ASSERT(orig_off >= align_off);
-               ASSERT(orig_end <= align_off + align_alen);
+       return 0;
+}
+
+STATIC int
+xfs_bmap_last_extent(
+       struct xfs_trans        *tp,
+       struct xfs_inode        *ip,
+       int                     whichfork,
+       struct xfs_bmbt_irec    *rec,
+       int                     *is_empty)
+{
+       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, whichfork);
+       int                     error;
+       int                     nextents;
+
+       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
+               error = xfs_iread_extents(tp, ip, whichfork);
+               if (error)
+                       return error;
        }
 
-#ifdef DEBUG
-       if (!eof && gotp->br_startoff != NULLFILEOFF)
-               ASSERT(align_off + align_alen <= gotp->br_startoff);
-       if (prevp->br_startoff != NULLFILEOFF)
-               ASSERT(align_off >= prevp->br_startoff + prevp->br_blockcount);
-#endif
+       nextents = ifp->if_bytes / sizeof(xfs_bmbt_rec_t);
+       if (nextents == 0) {
+               *is_empty = 1;
+               return 0;
+       }
 
-       *lenp = align_alen;
-       *offp = align_off;
+       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, nextents - 1), rec);
+       *is_empty = 0;
        return 0;
 }
 
-#define XFS_ALLOC_GAP_UNITS    4
-
-STATIC void
-xfs_bmap_adjacent(
-       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
+/*
+ * Check the last inode extent to determine whether this allocation will result
+ * in blocks being allocated at the end of the file. When we allocate new data
+ * blocks at the end of the file which do not start at the previous data block,
+ * we will try to align the new blocks at stripe unit boundaries.
+ *
+ * Returns 0 in bma->aeof if the file (fork) is empty as any new write will be
+ * at, or past the EOF.
+ */
+STATIC int
+xfs_bmap_isaeof(
+       struct xfs_bmalloca     *bma,
+       int                     whichfork)
 {
-       xfs_fsblock_t   adjust;         /* adjustment to block numbers */
-       xfs_agnumber_t  fb_agno;        /* ag number of ap->firstblock */
-       xfs_mount_t     *mp;            /* mount point structure */
-       int             nullfb;         /* true if ap->firstblock isn't set */
-       int             rt;             /* true if inode is realtime */
+       struct xfs_bmbt_irec    rec;
+       int                     is_empty;
+       int                     error;
 
-#define        ISVALID(x,y)    \
-       (rt ? \
-               (x) < mp->m_sb.sb_rblocks : \
-               XFS_FSB_TO_AGNO(mp, x) == XFS_FSB_TO_AGNO(mp, y) && \
-               XFS_FSB_TO_AGNO(mp, x) < mp->m_sb.sb_agcount && \
-               XFS_FSB_TO_AGBNO(mp, x) < mp->m_sb.sb_agblocks)
+       bma->aeof = 0;
+       error = xfs_bmap_last_extent(NULL, bma->ip, whichfork, &rec,
+                                    &is_empty);
+       if (error || is_empty)
+               return error;
 
-       mp = ap->ip->i_mount;
-       nullfb = *ap->firstblock == NULLFSBLOCK;
-       rt = XFS_IS_REALTIME_INODE(ap->ip) && ap->userdata;
-       fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp, *ap->firstblock);
-       /*
-        * If allocating at eof, and there's a previous real block,
-        * try to use its last block as our starting point.
-        */
-       if (ap->eof && ap->prev.br_startoff != NULLFILEOFF &&
-           !isnullstartblock(ap->prev.br_startblock) &&
-           ISVALID(ap->prev.br_startblock + ap->prev.br_blockcount,
-                   ap->prev.br_startblock)) {
-               ap->blkno = ap->prev.br_startblock + ap->prev.br_blockcount;
-               /*
-                * Adjust for the gap between prevp and us.
-                */
-               adjust = ap->offset -
-                       (ap->prev.br_startoff + ap->prev.br_blockcount);
-               if (adjust &&
-                   ISVALID(ap->blkno + adjust, ap->prev.br_startblock))
-                       ap->blkno += adjust;
-       }
        /*
-        * If not at eof, then compare the two neighbor blocks.
-        * Figure out whether either one gives us a good starting point,
-        * and pick the better one.
+        * Check if we are allocation or past the last extent, or at least into
+        * the last delayed allocated extent.
         */
-       else if (!ap->eof) {
-               xfs_fsblock_t   gotbno;         /* right side block number */
-               xfs_fsblock_t   gotdiff=0;      /* right side difference */
-               xfs_fsblock_t   prevbno;        /* left side block number */
-               xfs_fsblock_t   prevdiff=0;     /* left side difference */
+       bma->aeof = bma->offset >= rec.br_startoff + rec.br_blockcount ||
+               (bma->offset >= rec.br_startoff &&
+                isnullstartblock(rec.br_startblock));
+       return 0;
+}
 
-               /*
-                * If there's a previous (left) block, select a requested
-                * start block based on it.
-                */
-               if (ap->prev.br_startoff != NULLFILEOFF &&
-                   !isnullstartblock(ap->prev.br_startblock) &&
-                   (prevbno = ap->prev.br_startblock +
-                              ap->prev.br_blockcount) &&
-                   ISVALID(prevbno, ap->prev.br_startblock)) {
-                       /*
-                        * Calculate gap to end of previous block.
-                        */
-                       adjust = prevdiff = ap->offset -
-                               (ap->prev.br_startoff +
-                                ap->prev.br_blockcount);
-                       /*
-                        * Figure the startblock based on the previous block's
-                        * end and the gap size.
-                        * Heuristic!
-                        * If the gap is large relative to the piece we're
-                        * allocating, or using it gives us an invalid block
-                        * number, then just use the end of the previous block.
-                        */
-                       if (prevdiff <= XFS_ALLOC_GAP_UNITS * ap->length &&
-                           ISVALID(prevbno + prevdiff,
-                                   ap->prev.br_startblock))
-                               prevbno += adjust;
-                       else
-                               prevdiff += adjust;
-                       /*
-                        * If the firstblock forbids it, can't use it,
-                        * must use default.
-                        */
-                       if (!rt && !nullfb &&
-                           XFS_FSB_TO_AGNO(mp, prevbno) != fb_agno)
-                               prevbno = NULLFSBLOCK;
-               }
-               /*
-                * No previous block or can't follow it, just default.
-                */
-               else
-                       prevbno = NULLFSBLOCK;
-               /*
-                * If there's a following (right) block, select a requested
-                * start block based on it.
-                */
-               if (!isnullstartblock(ap->got.br_startblock)) {
-                       /*
-                        * Calculate gap to start of next block.
-                        */
-                       adjust = gotdiff = ap->got.br_startoff - ap->offset;
-                       /*
-                        * Figure the startblock based on the next block's
-                        * start and the gap size.
-                        */
-                       gotbno = ap->got.br_startblock;
-                       /*
-                        * Heuristic!
-                        * If the gap is large relative to the piece we're
-                        * allocating, or using it gives us an invalid block
-                        * number, then just use the start of the next block
-                        * offset by our length.
-                        */
-                       if (gotdiff <= XFS_ALLOC_GAP_UNITS * ap->length &&
-                           ISVALID(gotbno - gotdiff, gotbno))
-                               gotbno -= adjust;
-                       else if (ISVALID(gotbno - ap->length, gotbno)) {
-                               gotbno -= ap->length;
-                               gotdiff += adjust - ap->length;
-                       } else
-                               gotdiff += adjust;
-                       /*
-                        * If the firstblock forbids it, can't use it,
-                        * must use default.
-                        */
-                       if (!rt && !nullfb &&
-                           XFS_FSB_TO_AGNO(mp, gotbno) != fb_agno)
-                               gotbno = NULLFSBLOCK;
-               }
-               /*
-                * No next block, just default.
-                */
-               else
-                       gotbno = NULLFSBLOCK;
-               /*
-                * If both valid, pick the better one, else the only good
-                * one, else ap->blkno is already set (to 0 or the inode block).
-                */
-               if (prevbno != NULLFSBLOCK && gotbno != NULLFSBLOCK)
-                       ap->blkno = prevdiff <= gotdiff ? prevbno : gotbno;
-               else if (prevbno != NULLFSBLOCK)
-                       ap->blkno = prevbno;
-               else if (gotbno != NULLFSBLOCK)
-                       ap->blkno = gotbno;
-       }
-#undef ISVALID
-}
-
-STATIC int
-xfs_bmap_rtalloc(
-       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
+/*
+ * Check if the endoff is outside the last extent. If so the caller will grow
+ * the allocation to a stripe unit boundary.  All offsets are considered outside
+ * the end of file for an empty fork, so 1 is returned in *eof in that case.
+ */
+int
+xfs_bmap_eof(
+       struct xfs_inode        *ip,
+       xfs_fileoff_t           endoff,
+       int                     whichfork,
+       int                     *eof)
 {
-       xfs_alloctype_t atype = 0;      /* type for allocation routines */
-       int             error;          /* error return value */
-       xfs_mount_t     *mp;            /* mount point structure */
-       xfs_extlen_t    prod = 0;       /* product factor for allocators */
-       xfs_extlen_t    ralen = 0;      /* realtime allocation length */
-       xfs_extlen_t    align;          /* minimum allocation alignment */
-       xfs_rtblock_t   rtb;
+       struct xfs_bmbt_irec    rec;
+       int                     error;
 
-       mp = ap->ip->i_mount;
-       align = xfs_get_extsz_hint(ap->ip);
-       prod = align / mp->m_sb.sb_rextsize;
-       error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
-                                       align, 1, ap->eof, 0,
-                                       ap->conv, &ap->offset, &ap->length);
-       if (error)
+       error = xfs_bmap_last_extent(NULL, ip, whichfork, &rec, eof);
+       if (error || *eof)
                return error;
-       ASSERT(ap->length);
-       ASSERT(ap->length % mp->m_sb.sb_rextsize == 0);
 
-       /*
-        * If the offset & length are not perfectly aligned
-        * then kill prod, it will just get us in trouble.
-        */
-       if (do_mod(ap->offset, align) || ap->length % align)
-               prod = 1;
-       /*
-        * Set ralen to be the actual requested length in rtextents.
-        */
-       ralen = ap->length / mp->m_sb.sb_rextsize;
-       /*
-        * If the old value was close enough to MAXEXTLEN that
-        * we rounded up to it, cut it back so it's valid again.
-        * Note that if it's a really large request (bigger than
-        * MAXEXTLEN), we don't hear about that number, and can't
-        * adjust the starting point to match it.
-        */
-       if (ralen * mp->m_sb.sb_rextsize >= MAXEXTLEN)
-               ralen = MAXEXTLEN / mp->m_sb.sb_rextsize;
+       *eof = endoff >= rec.br_startoff + rec.br_blockcount;
+       return 0;
+}
 
-       /*
-        * Lock out other modifications to the RT bitmap inode.
-        */
-       xfs_ilock(mp->m_rbmip, XFS_ILOCK_EXCL);
-       xfs_trans_ijoin(ap->tp, mp->m_rbmip, XFS_ILOCK_EXCL);
+/*
+ * Returns the file-relative block number of the first block past eof in
+ * the file.  This is not based on i_size, it is based on the extent records.
+ * Returns 0 for local files, as they do not have extent records.
+ */
+int
+xfs_bmap_last_offset(
+       struct xfs_trans        *tp,
+       struct xfs_inode        *ip,
+       xfs_fileoff_t           *last_block,
+       int                     whichfork)
+{
+       struct xfs_bmbt_irec    rec;
+       int                     is_empty;
+       int                     error;
 
-       /*
-        * If it's an allocation to an empty file at offset 0,
-        * pick an extent that will space things out in the rt area.
-        */
-       if (ap->eof && ap->offset == 0) {
-               xfs_rtblock_t uninitialized_var(rtx); /* realtime extent no */
+       *last_block = 0;
 
-               error = xfs_rtpick_extent(mp, ap->tp, ralen, &rtx);
-               if (error)
-                       return error;
-               ap->blkno = rtx * mp->m_sb.sb_rextsize;
-       } else {
-               ap->blkno = 0;
-       }
+       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL)
+               return 0;
 
-       xfs_bmap_adjacent(ap);
+       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
+              return XFS_ERROR(EIO);
 
-       /*
-        * Realtime allocation, done through xfs_rtallocate_extent.
-        */
-       atype = ap->blkno == 0 ?  XFS_ALLOCTYPE_ANY_AG : XFS_ALLOCTYPE_NEAR_BNO;
-       do_div(ap->blkno, mp->m_sb.sb_rextsize);
-       rtb = ap->blkno;
-       ap->length = ralen;
-       if ((error = xfs_rtallocate_extent(ap->tp, ap->blkno, 1, ap->length,
-                               &ralen, atype, ap->wasdel, prod, &rtb)))
-               return error;
-       if (rtb == NULLFSBLOCK && prod > 1 &&
-           (error = xfs_rtallocate_extent(ap->tp, ap->blkno, 1,
-                                          ap->length, &ralen, atype,
-                                          ap->wasdel, 1, &rtb)))
+       error = xfs_bmap_last_extent(NULL, ip, whichfork, &rec, &is_empty);
+       if (error || is_empty)
                return error;
-       ap->blkno = rtb;
-       if (ap->blkno != NULLFSBLOCK) {
-               ap->blkno *= mp->m_sb.sb_rextsize;
-               ralen *= mp->m_sb.sb_rextsize;
-               ap->length = ralen;
-               ap->ip->i_d.di_nblocks += ralen;
-               xfs_trans_log_inode(ap->tp, ap->ip, XFS_ILOG_CORE);
-               if (ap->wasdel)
-                       ap->ip->i_delayed_blks -= ralen;
-               /*
-                * Adjust the disk quota also. This was reserved
-                * earlier.
-                */
-               xfs_trans_mod_dquot_byino(ap->tp, ap->ip,
-                       ap->wasdel ? XFS_TRANS_DQ_DELRTBCOUNT :
-                                       XFS_TRANS_DQ_RTBCOUNT, (long) ralen);
-       } else {
-               ap->length = 0;
-       }
+
+       *last_block = rec.br_startoff + rec.br_blockcount;
        return 0;
 }
 
-STATIC int
-xfs_bmap_btalloc_nullfb(
-       struct xfs_bmalloca     *ap,
-       struct xfs_alloc_arg    *args,
-       xfs_extlen_t            *blen)
+/*
+ * Returns whether the selected fork of the inode has exactly one
+ * block or not.  For the data fork we check this matches di_size,
+ * implying the file's range is 0..bsize-1.
+ */
+int                                    /* 1=>1 block, 0=>otherwise */
+xfs_bmap_one_block(
+       xfs_inode_t     *ip,            /* incore inode */
+       int             whichfork)      /* data or attr fork */
 {
-       struct xfs_mount        *mp = ap->ip->i_mount;
-       struct xfs_perag        *pag;
-       xfs_agnumber_t          ag, startag;
-       int                     notinit = 0;
-       int                     error;
+       xfs_bmbt_rec_host_t *ep;        /* ptr to fork's extent */
+       xfs_ifork_t     *ifp;           /* inode fork pointer */
+       int             rval;           /* return value */
+       xfs_bmbt_irec_t s;              /* internal version of extent */
 
-       if (ap->userdata && xfs_inode_is_filestream(ap->ip))
-               args->type = XFS_ALLOCTYPE_NEAR_BNO;
-       else
-               args->type = XFS_ALLOCTYPE_START_BNO;
-       args->total = ap->total;
+#ifndef DEBUG
+       if (whichfork == XFS_DATA_FORK)
+               return XFS_ISIZE(ip) == ip->i_mount->m_sb.sb_blocksize;
+#endif /* !DEBUG */
+       if (XFS_IFORK_NEXTENTS(ip, whichfork) != 1)
+               return 0;
+       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
+               return 0;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT(ifp->if_flags & XFS_IFEXTENTS);
+       ep = xfs_iext_get_ext(ifp, 0);
+       xfs_bmbt_get_all(ep, &s);
+       rval = s.br_startoff == 0 && s.br_blockcount == 1;
+       if (rval && whichfork == XFS_DATA_FORK)
+               ASSERT(XFS_ISIZE(ip) == ip->i_mount->m_sb.sb_blocksize);
+       return rval;
+}
 
-       /*
-        * Search for an allocation group with a single extent large enough
-        * for the request.  If one isn't found, then adjust the minimum
-        * allocation size to the largest space found.
-        */
-       startag = ag = XFS_FSB_TO_AGNO(mp, args->fsbno);
-       if (startag == NULLAGNUMBER)
-               startag = ag = 0;
+/*
+ * Extent tree manipulation functions used during allocation.
+ */
 
-       pag = xfs_perag_get(mp, ag);
-       while (*blen < args->maxlen) {
-               if (!pag->pagf_init) {
-                       error = xfs_alloc_pagf_init(mp, args->tp, ag,
-                                                   XFS_ALLOC_FLAG_TRYLOCK);
-                       if (error) {
-                               xfs_perag_put(pag);
-                               return error;
-                       }
-               }
-
-               /*
-                * See xfs_alloc_fix_freelist...
-                */
-               if (pag->pagf_init) {
-                       xfs_extlen_t    longest;
-                       longest = xfs_alloc_longest_free_extent(mp, pag);
-                       if (*blen < longest)
-                               *blen = longest;
-               } else
-                       notinit = 1;
+/*
+ * Convert a delayed allocation to a real allocation.
+ */
+STATIC int                             /* error */
+xfs_bmap_add_extent_delay_real(
+       struct xfs_bmalloca     *bma)
+{
+       struct xfs_bmbt_irec    *new = &bma->got;
+       int                     diff;   /* temp value */
+       xfs_bmbt_rec_host_t     *ep;    /* extent entry for idx */
+       int                     error;  /* error return value */
+       int                     i;      /* temp state */
+       xfs_ifork_t             *ifp;   /* inode fork pointer */
+       xfs_fileoff_t           new_endoff;     /* end offset of new entry */
+       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */
+                                       /* left is 0, right is 1, prev is 2 */
+       int                     rval=0; /* return value (logging flags) */
+       int                     state = 0;/* state bits, accessed thru macros */
+       xfs_filblks_t           da_new; /* new count del alloc blocks used */
+       xfs_filblks_t           da_old; /* old count del alloc blocks used */
+       xfs_filblks_t           temp=0; /* value for da_new calculations */
+       xfs_filblks_t           temp2=0;/* value for da_new calculations */
+       int                     tmp_rval;       /* partial logging flags */
 
-               if (xfs_inode_is_filestream(ap->ip)) {
-                       if (*blen >= args->maxlen)
-                               break;
+       ifp = XFS_IFORK_PTR(bma->ip, XFS_DATA_FORK);
 
-                       if (ap->userdata) {
-                               /*
-                                * If startag is an invalid AG, we've
-                                * come here once before and
-                                * xfs_filestream_new_ag picked the
-                                * best currently available.
-                                *
-                                * Don't continue looping, since we
-                                * could loop forever.
-                                */
-                               if (startag == NULLAGNUMBER)
-                                       break;
+       ASSERT(bma->idx >= 0);
+       ASSERT(bma->idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
+       ASSERT(!isnullstartblock(new->br_startblock));
+       ASSERT(!bma->cur ||
+              (bma->cur->bc_private.b.flags & XFS_BTCUR_BPRV_WASDEL));
 
-                               error = xfs_filestream_new_ag(ap, &ag);
-                               xfs_perag_put(pag);
-                               if (error)
-                                       return error;
+       XFS_STATS_INC(xs_add_exlist);
 
-                               /* loop again to set 'blen'*/
-                               startag = NULLAGNUMBER;
-                               pag = xfs_perag_get(mp, ag);
-                               continue;
-                       }
-               }
-               if (++ag == mp->m_sb.sb_agcount)
-                       ag = 0;
-               if (ag == startag)
-                       break;
-               xfs_perag_put(pag);
-               pag = xfs_perag_get(mp, ag);
-       }
-       xfs_perag_put(pag);
+#define        LEFT            r[0]
+#define        RIGHT           r[1]
+#define        PREV            r[2]
 
        /*
-        * Since the above loop did a BUF_TRYLOCK, it is
-        * possible that there is space for this request.
-        */
-       if (notinit || *blen < ap->minlen)
-               args->minlen = ap->minlen;
-       /*
-        * If the best seen length is less than the request
-        * length, use the best as the minimum.
+        * Set up a bunch of variables to make the tests simpler.
         */
-       else if (*blen < args->maxlen)
-               args->minlen = *blen;
+       ep = xfs_iext_get_ext(ifp, bma->idx);
+       xfs_bmbt_get_all(ep, &PREV);
+       new_endoff = new->br_startoff + new->br_blockcount;
+       ASSERT(PREV.br_startoff <= new->br_startoff);
+       ASSERT(PREV.br_startoff + PREV.br_blockcount >= new_endoff);
+
+       da_old = startblockval(PREV.br_startblock);
+       da_new = 0;
+
        /*
-        * Otherwise we've seen an extent as big as maxlen,
-        * use that as the minimum.
+        * Set flags determining what part of the previous delayed allocation
+        * extent is being replaced by a real allocation.
         */
-       else
-               args->minlen = args->maxlen;
+       if (PREV.br_startoff == new->br_startoff)
+               state |= BMAP_LEFT_FILLING;
+       if (PREV.br_startoff + PREV.br_blockcount == new_endoff)
+               state |= BMAP_RIGHT_FILLING;
 
        /*
-        * set the failure fallback case to look in the selected
-        * AG as the stream may have moved.
+        * Check and set flags if this segment has a left neighbor.
+        * Don't set contiguous if the combined extent would be too large.
         */
-       if (xfs_inode_is_filestream(ap->ip))
-               ap->blkno = args->fsbno = XFS_AGB_TO_FSB(mp, ag, 0);
-
-       return 0;
-}
-
-STATIC int
-xfs_bmap_btalloc(
-       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
-{
-       xfs_mount_t     *mp;            /* mount point structure */
-       xfs_alloctype_t atype = 0;      /* type for allocation routines */
-       xfs_extlen_t    align;          /* minimum allocation alignment */
-       xfs_agnumber_t  fb_agno;        /* ag number of ap->firstblock */
-       xfs_agnumber_t  ag;
-       xfs_alloc_arg_t args;
-       xfs_extlen_t    blen;
-       xfs_extlen_t    nextminlen = 0;
-       int             nullfb;         /* true if ap->firstblock isn't set */
-       int             isaligned;
-       int             tryagain;
-       int             error;
-
-       ASSERT(ap->length);
+       if (bma->idx > 0) {
+               state |= BMAP_LEFT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1), &LEFT);
 
-       mp = ap->ip->i_mount;
-       align = ap->userdata ? xfs_get_extsz_hint(ap->ip) : 0;
-       if (unlikely(align)) {
-               error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
-                                               align, 0, ap->eof, 0, ap->conv,
-                                               &ap->offset, &ap->length);
-               ASSERT(!error);
-               ASSERT(ap->length);
+               if (isnullstartblock(LEFT.br_startblock))
+                       state |= BMAP_LEFT_DELAY;
        }
-       nullfb = *ap->firstblock == NULLFSBLOCK;
-       fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp, *ap->firstblock);
-       if (nullfb) {
-               if (ap->userdata && xfs_inode_is_filestream(ap->ip)) {
-                       ag = xfs_filestream_lookup_ag(ap->ip);
-                       ag = (ag != NULLAGNUMBER) ? ag : 0;
-                       ap->blkno = XFS_AGB_TO_FSB(mp, ag, 0);
-               } else {
-                       ap->blkno = XFS_INO_TO_FSB(mp, ap->ip->i_ino);
-               }
-       } else
-               ap->blkno = *ap->firstblock;
 
-       xfs_bmap_adjacent(ap);
+       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
+           LEFT.br_startoff + LEFT.br_blockcount == new->br_startoff &&
+           LEFT.br_startblock + LEFT.br_blockcount == new->br_startblock &&
+           LEFT.br_state == new->br_state &&
+           LEFT.br_blockcount + new->br_blockcount <= MAXEXTLEN)
+               state |= BMAP_LEFT_CONTIG;
 
        /*
-        * If allowed, use ap->blkno; otherwise must use firstblock since
-        * it's in the right allocation group.
-        */
-       if (nullfb || XFS_FSB_TO_AGNO(mp, ap->blkno) == fb_agno)
-               ;
-       else
-               ap->blkno = *ap->firstblock;
-       /*
-        * Normal allocation, done through xfs_alloc_vextent.
+        * Check and set flags if this segment has a right neighbor.
+        * Don't set contiguous if the combined extent would be too large.
+        * Also check for all-three-contiguous being too large.
         */
-       tryagain = isaligned = 0;
-       memset(&args, 0, sizeof(args));
-       args.tp = ap->tp;
-       args.mp = mp;
-       args.fsbno = ap->blkno;
+       if (bma->idx < bma->ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t) - 1) {
+               state |= BMAP_RIGHT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx + 1), &RIGHT);
 
-       /* Trim the allocation back to the maximum an AG can fit. */
-       args.maxlen = MIN(ap->length, XFS_ALLOC_AG_MAX_USABLE(mp));
-       args.firstblock = *ap->firstblock;
-       blen = 0;
-       if (nullfb) {
-               error = xfs_bmap_btalloc_nullfb(ap, &args, &blen);
-               if (error)
-                       return error;
-       } else if (ap->flist->xbf_low) {
-               if (xfs_inode_is_filestream(ap->ip))
-                       args.type = XFS_ALLOCTYPE_FIRST_AG;
-               else
-                       args.type = XFS_ALLOCTYPE_START_BNO;
-               args.total = args.minlen = ap->minlen;
-       } else {
-               args.type = XFS_ALLOCTYPE_NEAR_BNO;
-               args.total = ap->total;
-               args.minlen = ap->minlen;
-       }
-       /* apply extent size hints if obtained earlier */
-       if (unlikely(align)) {
-               args.prod = align;
-               if ((args.mod = (xfs_extlen_t)do_mod(ap->offset, args.prod)))
-                       args.mod = (xfs_extlen_t)(args.prod - args.mod);
-       } else if (mp->m_sb.sb_blocksize >= PAGE_CACHE_SIZE) {
-               args.prod = 1;
-               args.mod = 0;
-       } else {
-               args.prod = PAGE_CACHE_SIZE >> mp->m_sb.sb_blocklog;
-               if ((args.mod = (xfs_extlen_t)(do_mod(ap->offset, args.prod))))
-                       args.mod = (xfs_extlen_t)(args.prod - args.mod);
+               if (isnullstartblock(RIGHT.br_startblock))
+                       state |= BMAP_RIGHT_DELAY;
        }
+
+       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
+           new_endoff == RIGHT.br_startoff &&
+           new->br_startblock + new->br_blockcount == RIGHT.br_startblock &&
+           new->br_state == RIGHT.br_state &&
+           new->br_blockcount + RIGHT.br_blockcount <= MAXEXTLEN &&
+           ((state & (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
+                      BMAP_RIGHT_FILLING)) !=
+                     (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
+                      BMAP_RIGHT_FILLING) ||
+            LEFT.br_blockcount + new->br_blockcount + RIGHT.br_blockcount
+                       <= MAXEXTLEN))
+               state |= BMAP_RIGHT_CONTIG;
+
+       error = 0;
        /*
-        * If we are not low on available data blocks, and the
-        * underlying logical volume manager is a stripe, and
-        * the file offset is zero then try to allocate data
-        * blocks on stripe unit boundary.
-        * NOTE: ap->aeof is only set if the allocation length
-        * is >= the stripe unit and the allocation offset is
-        * at the end of file.
+        * Switch out based on the FILLING and CONTIG state bits.
         */
-       if (!ap->flist->xbf_low && ap->aeof) {
-               if (!ap->offset) {
-                       args.alignment = mp->m_dalign;
-                       atype = args.type;
-                       isaligned = 1;
-                       /*
-                        * Adjust for alignment
-                        */
-                       if (blen > args.alignment && blen <= args.maxlen)
-                               args.minlen = blen - args.alignment;
-                       args.minalignslop = 0;
-               } else {
-                       /*
-                        * First try an exact bno allocation.
-                        * If it fails then do a near or start bno
-                        * allocation with alignment turned on.
-                        */
-                       atype = args.type;
-                       tryagain = 1;
-                       args.type = XFS_ALLOCTYPE_THIS_BNO;
-                       args.alignment = 1;
-                       /*
-                        * Compute the minlen+alignment for the
-                        * next case.  Set slop so that the value
-                        * of minlen+alignment+slop doesn't go up
-                        * between the calls.
-                        */
-                       if (blen > mp->m_dalign && blen <= args.maxlen)
-                               nextminlen = blen - mp->m_dalign;
-                       else
-                               nextminlen = args.minlen;
-                       if (nextminlen + mp->m_dalign > args.minlen + 1)
-                               args.minalignslop =
-                                       nextminlen + mp->m_dalign -
-                                       args.minlen - 1;
-                       else
-                               args.minalignslop = 0;
-               }
-       } else {
-               args.alignment = 1;
-               args.minalignslop = 0;
-       }
-       args.minleft = ap->minleft;
-       args.wasdel = ap->wasdel;
-       args.isfl = 0;
-       args.userdata = ap->userdata;
-       if ((error = xfs_alloc_vextent(&args)))
-               return error;
-       if (tryagain && args.fsbno == NULLFSBLOCK) {
+       switch (state & (BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
+                        BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG)) {
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
+            BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * Exact allocation failed. Now try with alignment
-                * turned on.
+                * Filling in all of a previously delayed allocation extent.
+                * The left and right neighbors are both contiguous with new.
                 */
-               args.type = atype;
-               args.fsbno = ap->blkno;
-               args.alignment = mp->m_dalign;
-               args.minlen = nextminlen;
-               args.minalignslop = 0;
-               isaligned = 1;
-               if ((error = xfs_alloc_vextent(&args)))
-                       return error;
-       }
-       if (isaligned && args.fsbno == NULLFSBLOCK) {
+               bma->idx--;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
+                       LEFT.br_blockcount + PREV.br_blockcount +
+                       RIGHT.br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+
+               xfs_iext_remove(bma->ip, bma->idx + 1, 2, state);
+               bma->ip->i_d.di_nextents--;
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
+                                       RIGHT.br_startblock,
+                                       RIGHT.br_blockcount, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_btree_delete(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_btree_decrement(bma->cur, 0, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
+                                       LEFT.br_startblock,
+                                       LEFT.br_blockcount +
+                                       PREV.br_blockcount +
+                                       RIGHT.br_blockcount, LEFT.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
+
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
                /*
-                * allocation failed, so turn off alignment and
-                * try again.
+                * Filling in all of a previously delayed allocation extent.
+                * The left neighbor is contiguous, the right is not.
                 */
-               args.type = atype;
-               args.fsbno = ap->blkno;
-               args.alignment = 0;
-               if ((error = xfs_alloc_vextent(&args)))
-                       return error;
-       }
-       if (args.fsbno == NULLFSBLOCK && nullfb &&
-           args.minlen > ap->minlen) {
-               args.minlen = ap->minlen;
-               args.type = XFS_ALLOCTYPE_START_BNO;
-               args.fsbno = ap->blkno;
-               if ((error = xfs_alloc_vextent(&args)))
-                       return error;
-       }
-       if (args.fsbno == NULLFSBLOCK && nullfb) {
-               args.fsbno = 0;
-               args.type = XFS_ALLOCTYPE_FIRST_AG;
-               args.total = ap->minlen;
-               args.minleft = 0;
-               if ((error = xfs_alloc_vextent(&args)))
-                       return error;
-               ap->flist->xbf_low = 1;
-       }
-       if (args.fsbno != NULLFSBLOCK) {
+               bma->idx--;
+
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
+                       LEFT.br_blockcount + PREV.br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+
+               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur, LEFT.br_startoff,
+                                       LEFT.br_startblock, LEFT.br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
+                                       LEFT.br_startblock,
+                                       LEFT.br_blockcount +
+                                       PREV.br_blockcount, LEFT.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
+
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * check the allocation happened at the same or higher AG than
-                * the first block that was allocated.
+                * Filling in all of a previously delayed allocation extent.
+                * The right neighbor is contiguous, the left is not.
                 */
-               ASSERT(*ap->firstblock == NULLFSBLOCK ||
-                      XFS_FSB_TO_AGNO(mp, *ap->firstblock) ==
-                      XFS_FSB_TO_AGNO(mp, args.fsbno) ||
-                      (ap->flist->xbf_low &&
-                       XFS_FSB_TO_AGNO(mp, *ap->firstblock) <
-                       XFS_FSB_TO_AGNO(mp, args.fsbno)));
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_startblock(ep, new->br_startblock);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount + RIGHT.br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-               ap->blkno = args.fsbno;
-               if (*ap->firstblock == NULLFSBLOCK)
-                       *ap->firstblock = args.fsbno;
-               ASSERT(nullfb || fb_agno == args.agno ||
-                      (ap->flist->xbf_low && fb_agno < args.agno));
-               ap->length = args.len;
-               ap->ip->i_d.di_nblocks += args.len;
-               xfs_trans_log_inode(ap->tp, ap->ip, XFS_ILOG_CORE);
-               if (ap->wasdel)
-                       ap->ip->i_delayed_blks -= args.len;
+               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
+                                       RIGHT.br_startblock,
+                                       RIGHT.br_blockcount, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, PREV.br_startoff,
+                                       new->br_startblock,
+                                       PREV.br_blockcount +
+                                       RIGHT.br_blockcount, PREV.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
+
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING:
                /*
-                * Adjust the disk quota also. This was reserved
-                * earlier.
+                * Filling in all of a previously delayed allocation extent.
+                * Neither the left nor right neighbors are contiguous with
+                * the new one.
                 */
-               xfs_trans_mod_dquot_byino(ap->tp, ap->ip,
-                       ap->wasdel ? XFS_TRANS_DQ_DELBCOUNT :
-                                       XFS_TRANS_DQ_BCOUNT,
-                       (long) args.len);
-       } else {
-               ap->blkno = NULLFSBLOCK;
-               ap->length = 0;
-       }
-       return 0;
-}
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_startblock(ep, new->br_startblock);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-/*
- * xfs_bmap_alloc is called by xfs_bmapi to allocate an extent for a file.
- * It figures out where to ask the underlying allocator to put the new extent.
- */
-STATIC int
-xfs_bmap_alloc(
-       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
-{
-       if (XFS_IS_REALTIME_INODE(ap->ip) && ap->userdata)
-               return xfs_bmap_rtalloc(ap);
-       return xfs_bmap_btalloc(ap);
-}
+               bma->ip->i_d.di_nextents++;
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
+                       error = xfs_btree_insert(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               }
+               break;
 
-/*
- * Transform a btree format file with only one leaf node, where the
- * extents list will fit in the inode, into an extents format file.
- * Since the file extents are already in-core, all we have to do is
- * give up the space for the btree root and pitch the leaf block.
- */
-STATIC int                             /* error */
-xfs_bmap_btree_to_extents(
-       xfs_trans_t             *tp,    /* transaction pointer */
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_btree_cur_t         *cur,   /* btree cursor */
-       int                     *logflagsp, /* inode logging flags */
-       int                     whichfork)  /* data or attr fork */
-{
-       /* REFERENCED */
-       struct xfs_btree_block  *cblock;/* child btree block */
-       xfs_fsblock_t           cbno;   /* child block number */
-       xfs_buf_t               *cbp;   /* child block's buffer */
-       int                     error;  /* error return value */
-       xfs_ifork_t             *ifp;   /* inode fork data */
-       xfs_mount_t             *mp;    /* mount point structure */
-       __be64                  *pp;    /* ptr to block address */
-       struct xfs_btree_block  *rblock;/* root btree block */
-
-       mp = ip->i_mount;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT(ifp->if_flags & XFS_IFEXTENTS);
-       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE);
-       rblock = ifp->if_broot;
-       ASSERT(be16_to_cpu(rblock->bb_level) == 1);
-       ASSERT(be16_to_cpu(rblock->bb_numrecs) == 1);
-       ASSERT(xfs_bmbt_maxrecs(mp, ifp->if_broot_bytes, 0) == 1);
-       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, rblock, 1, ifp->if_broot_bytes);
-       cbno = be64_to_cpu(*pp);
-       *logflagsp = 0;
-#ifdef DEBUG
-       if ((error = xfs_btree_check_lptr(cur, cbno, 1)))
-               return error;
-#endif
-       error = xfs_btree_read_bufl(mp, tp, cbno, 0, &cbp, XFS_BMAP_BTREE_REF,
-                               &xfs_bmbt_buf_ops);
-       if (error)
-               return error;
-       cblock = XFS_BUF_TO_BLOCK(cbp);
-       if ((error = xfs_btree_check_block(cur, cblock, 0, cbp)))
-               return error;
-       xfs_bmap_add_free(cbno, 1, cur->bc_private.b.flist, mp);
-       ip->i_d.di_nblocks--;
-       xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, -1L);
-       xfs_trans_binval(tp, cbp);
-       if (cur->bc_bufs[0] == cbp)
-               cur->bc_bufs[0] = NULL;
-       xfs_iroot_realloc(ip, -1, whichfork);
-       ASSERT(ifp->if_broot == NULL);
-       ASSERT((ifp->if_flags & XFS_IFBROOT) == 0);
-       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_EXTENTS);
-       *logflagsp = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
-       return 0;
-}
-
-/*
- * Called by xfs_bmapi to update file extent records and the btree
- * after removing space (or undoing a delayed allocation).
- */
-STATIC int                             /* error */
-xfs_bmap_del_extent(
-       xfs_inode_t             *ip,    /* incore inode pointer */
-       xfs_trans_t             *tp,    /* current transaction pointer */
-       xfs_extnum_t            *idx,   /* extent number to update/delete */
-       xfs_bmap_free_t         *flist, /* list of extents to be freed */
-       xfs_btree_cur_t         *cur,   /* if null, not a btree */
-       xfs_bmbt_irec_t         *del,   /* data to remove from extents */
-       int                     *logflagsp, /* inode logging flags */
-       int                     whichfork) /* data or attr fork */
-{
-       xfs_filblks_t           da_new; /* new delay-alloc indirect blocks */
-       xfs_filblks_t           da_old; /* old delay-alloc indirect blocks */
-       xfs_fsblock_t           del_endblock=0; /* first block past del */
-       xfs_fileoff_t           del_endoff;     /* first offset past del */
-       int                     delay;  /* current block is delayed allocated */
-       int                     do_fx;  /* free extent at end of routine */
-       xfs_bmbt_rec_host_t     *ep;    /* current extent entry pointer */
-       int                     error;  /* error return value */
-       int                     flags;  /* inode logging flags */
-       xfs_bmbt_irec_t         got;    /* current extent entry */
-       xfs_fileoff_t           got_endoff;     /* first offset past got */
-       int                     i;      /* temp state */
-       xfs_ifork_t             *ifp;   /* inode fork pointer */
-       xfs_mount_t             *mp;    /* mount structure */
-       xfs_filblks_t           nblks;  /* quota/sb block count */
-       xfs_bmbt_irec_t         new;    /* new record to be inserted */
-       /* REFERENCED */
-       uint                    qfield; /* quota field to update */
-       xfs_filblks_t           temp;   /* for indirect length calculations */
-       xfs_filblks_t           temp2;  /* for indirect length calculations */
-       int                     state = 0;
-
-       XFS_STATS_INC(xs_del_exlist);
-
-       if (whichfork == XFS_ATTR_FORK)
-               state |= BMAP_ATTRFORK;
-
-       mp = ip->i_mount;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT((*idx >= 0) && (*idx < ifp->if_bytes /
-               (uint)sizeof(xfs_bmbt_rec_t)));
-       ASSERT(del->br_blockcount > 0);
-       ep = xfs_iext_get_ext(ifp, *idx);
-       xfs_bmbt_get_all(ep, &got);
-       ASSERT(got.br_startoff <= del->br_startoff);
-       del_endoff = del->br_startoff + del->br_blockcount;
-       got_endoff = got.br_startoff + got.br_blockcount;
-       ASSERT(got_endoff >= del_endoff);
-       delay = isnullstartblock(got.br_startblock);
-       ASSERT(isnullstartblock(del->br_startblock) == delay);
-       flags = 0;
-       qfield = 0;
-       error = 0;
-       /*
-        * If deleting a real allocation, must free up the disk space.
-        */
-       if (!delay) {
-               flags = XFS_ILOG_CORE;
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG:
                /*
-                * Realtime allocation.  Free it and record di_nblocks update.
+                * Filling in the first part of a previous delayed allocation.
+                * The left neighbor is contiguous.
                 */
-               if (whichfork == XFS_DATA_FORK && XFS_IS_REALTIME_INODE(ip)) {
-                       xfs_fsblock_t   bno;
-                       xfs_filblks_t   len;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx - 1, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx - 1),
+                       LEFT.br_blockcount + new->br_blockcount);
+               xfs_bmbt_set_startoff(ep,
+                       PREV.br_startoff + new->br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx - 1, state, _THIS_IP_);
 
-                       ASSERT(do_mod(del->br_blockcount,
-                                     mp->m_sb.sb_rextsize) == 0);
-                       ASSERT(do_mod(del->br_startblock,
-                                     mp->m_sb.sb_rextsize) == 0);
-                       bno = del->br_startblock;
-                       len = del->br_blockcount;
-                       do_div(bno, mp->m_sb.sb_rextsize);
-                       do_div(len, mp->m_sb.sb_rextsize);
-                       error = xfs_rtfree_extent(tp, bno, (xfs_extlen_t)len);
+               temp = PREV.br_blockcount - new->br_blockcount;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep, temp);
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur, LEFT.br_startoff,
+                                       LEFT.br_startblock, LEFT.br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, LEFT.br_startoff,
+                                       LEFT.br_startblock,
+                                       LEFT.br_blockcount +
+                                       new->br_blockcount,
+                                       LEFT.br_state);
                        if (error)
                                goto done;
-                       do_fx = 0;
-                       nblks = len * mp->m_sb.sb_rextsize;
-                       qfield = XFS_TRANS_DQ_RTBCOUNT;
                }
+               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
+                       startblockval(PREV.br_startblock));
+               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+
+               bma->idx--;
+               break;
+
+       case BMAP_LEFT_FILLING:
                /*
-                * Ordinary allocation.
+                * Filling in the first part of a previous delayed allocation.
+                * The left neighbor is not contiguous.
                 */
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_startoff(ep, new_endoff);
+               temp = PREV.br_blockcount - new->br_blockcount;
+               xfs_bmbt_set_blockcount(ep, temp);
+               xfs_iext_insert(bma->ip, bma->idx, 1, new, state);
+               bma->ip->i_d.di_nextents++;
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
                else {
-                       do_fx = 1;
-                       nblks = del->br_blockcount;
-                       qfield = XFS_TRANS_DQ_BCOUNT;
-               }
-               /*
-                * Set up del_endblock and cur for later.
-                */
-               del_endblock = del->br_startblock + del->br_blockcount;
-               if (cur) {
-                       if ((error = xfs_bmbt_lookup_eq(cur, got.br_startoff,
-                                       got.br_startblock, got.br_blockcount,
-                                       &i)))
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
+                       error = xfs_btree_insert(bma->cur, &i);
+                       if (error)
                                goto done;
                        XFS_WANT_CORRUPTED_GOTO(i == 1, done);
                }
-               da_old = da_new = 0;
-       } else {
-               da_old = startblockval(got.br_startblock);
-               da_new = 0;
-               nblks = 0;
-               do_fx = 0;
-       }
-       /*
-        * Set flag value to use in switch statement.
-        * Left-contig is 2, right-contig is 1.
-        */
-       switch (((got.br_startoff == del->br_startoff) << 1) |
-               (got_endoff == del_endoff)) {
-       case 3:
-               /*
-                * Matches the whole extent.  Delete the entry.
-                */
-               xfs_iext_remove(ip, *idx, 1,
-                               whichfork == XFS_ATTR_FORK ? BMAP_ATTRFORK : 0);
-               --*idx;
-               if (delay)
-                       break;
 
-               XFS_IFORK_NEXT_SET(ip, whichfork,
-                       XFS_IFORK_NEXTENTS(ip, whichfork) - 1);
-               flags |= XFS_ILOG_CORE;
-               if (!cur) {
-                       flags |= xfs_ilog_fext(whichfork);
-                       break;
+               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
+                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
+                                       bma->firstblock, bma->flist,
+                                       &bma->cur, 1, &tmp_rval, XFS_DATA_FORK);
+                       rval |= tmp_rval;
+                       if (error)
+                               goto done;
                }
-               if ((error = xfs_btree_delete(cur, &i)))
-                       goto done;
-               XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
+                       startblockval(PREV.br_startblock) -
+                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
+               ep = xfs_iext_get_ext(ifp, bma->idx + 1);
+               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
                break;
 
-       case 2:
+       case BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * Deleting the first part of the extent.
+                * Filling in the last part of a previous delayed allocation.
+                * The right neighbor is contiguous with the new allocation.
                 */
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_startoff(ep, del_endoff);
-               temp = got.br_blockcount - del->br_blockcount;
+               temp = PREV.br_blockcount - new->br_blockcount;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
                xfs_bmbt_set_blockcount(ep, temp);
-               if (delay) {
-                       temp = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(ip, temp),
-                               da_old);
-                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
-                       trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-                       da_new = temp;
-                       break;
-               }
-               xfs_bmbt_set_startblock(ep, del_endblock);
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-               if (!cur) {
-                       flags |= xfs_ilog_fext(whichfork);
-                       break;
+               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, bma->idx + 1),
+                       new->br_startoff, new->br_startblock,
+                       new->br_blockcount + RIGHT.br_blockcount,
+                       RIGHT.br_state);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx + 1, state, _THIS_IP_);
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur, RIGHT.br_startoff,
+                                       RIGHT.br_startblock,
+                                       RIGHT.br_blockcount, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, new->br_startoff,
+                                       new->br_startblock,
+                                       new->br_blockcount +
+                                       RIGHT.br_blockcount,
+                                       RIGHT.br_state);
+                       if (error)
+                               goto done;
                }
-               if ((error = xfs_bmbt_update(cur, del_endoff, del_endblock,
-                               got.br_blockcount - del->br_blockcount,
-                               got.br_state)))
-                       goto done;
+
+               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
+                       startblockval(PREV.br_startblock));
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+
+               bma->idx++;
                break;
 
-       case 1:
+       case BMAP_RIGHT_FILLING:
                /*
-                * Deleting the last part of the extent.
+                * Filling in the last part of a previous delayed allocation.
+                * The right neighbor is not contiguous.
                 */
-               temp = got.br_blockcount - del->br_blockcount;
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               temp = PREV.br_blockcount - new->br_blockcount;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
                xfs_bmbt_set_blockcount(ep, temp);
-               if (delay) {
-                       temp = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(ip, temp),
-                               da_old);
-                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
-                       trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-                       da_new = temp;
-                       break;
+               xfs_iext_insert(bma->ip, bma->idx + 1, 1, new, state);
+               bma->ip->i_d.di_nextents++;
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
+                       error = xfs_btree_insert(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
                }
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-               if (!cur) {
-                       flags |= xfs_ilog_fext(whichfork);
-                       break;
+
+               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
+                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
+                               bma->firstblock, bma->flist, &bma->cur, 1,
+                               &tmp_rval, XFS_DATA_FORK);
+                       rval |= tmp_rval;
+                       if (error)
+                               goto done;
                }
-               if ((error = xfs_bmbt_update(cur, got.br_startoff,
-                               got.br_startblock,
-                               got.br_blockcount - del->br_blockcount,
-                               got.br_state)))
-                       goto done;
+               da_new = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(bma->ip, temp),
+                       startblockval(PREV.br_startblock) -
+                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
+               ep = xfs_iext_get_ext(ifp, bma->idx);
+               xfs_bmbt_set_startblock(ep, nullstartblock(da_new));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+
+               bma->idx++;
                break;
 
        case 0:
                /*
-                * Deleting the middle of the extent.
+                * Filling in the middle part of a previous delayed allocation.
+                * Contiguity is impossible here.
+                * This case is avoided almost all the time.
+                *
+                * We start with a delayed allocation:
+                *
+                * +ddddddddddddddddddddddddddddddddddddddddddddddddddddddd+
+                *  PREV @ idx
+                *
+                * and we are allocating:
+                *                     +rrrrrrrrrrrrrrrrr+
+                *                            new
+                *
+                * and we set it up for insertion as:
+                * +ddddddddddddddddddd+rrrrrrrrrrrrrrrrr+ddddddddddddddddd+
+                *                            new
+                *  PREV @ idx          LEFT              RIGHT
+                *                      inserted at idx + 1
                 */
-               temp = del->br_startoff - got.br_startoff;
-               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
-               xfs_bmbt_set_blockcount(ep, temp);
-               new.br_startoff = del_endoff;
-               temp2 = got_endoff - del_endoff;
-               new.br_blockcount = temp2;
-               new.br_state = got.br_state;
-               if (!delay) {
-                       new.br_startblock = del_endblock;
-                       flags |= XFS_ILOG_CORE;
-                       if (cur) {
-                               if ((error = xfs_bmbt_update(cur,
-                                               got.br_startoff,
-                                               got.br_startblock, temp,
-                                               got.br_state)))
-                                       goto done;
-                               if ((error = xfs_btree_increment(cur, 0, &i)))
-                                       goto done;
-                               cur->bc_rec.b = new;
-                               error = xfs_btree_insert(cur, &i);
-                               if (error && error != ENOSPC)
-                                       goto done;
-                               /*
-                                * If get no-space back from btree insert,
-                                * it tried a split, and we have a zero
-                                * block reservation.
-                                * Fix up our state and return the error.
-                                */
-                               if (error == ENOSPC) {
-                                       /*
-                                        * Reset the cursor, don't trust
-                                        * it after any insert operation.
-                                        */
-                                       if ((error = xfs_bmbt_lookup_eq(cur,
-                                                       got.br_startoff,
-                                                       got.br_startblock,
-                                                       temp, &i)))
-                                               goto done;
-                                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                                       /*
-                                        * Update the btree record back
-                                        * to the original value.
-                                        */
-                                       if ((error = xfs_bmbt_update(cur,
-                                                       got.br_startoff,
-                                                       got.br_startblock,
-                                                       got.br_blockcount,
-                                                       got.br_state)))
-                                               goto done;
-                                       /*
-                                        * Reset the extent record back
-                                        * to the original value.
-                                        */
-                                       xfs_bmbt_set_blockcount(ep,
-                                               got.br_blockcount);
-                                       flags = 0;
-                                       error = XFS_ERROR(ENOSPC);
-                                       goto done;
-                               }
-                               XFS_WANT_CORRUPTED_GOTO(i == 1, done);
-                       } else
-                               flags |= xfs_ilog_fext(whichfork);
-                       XFS_IFORK_NEXT_SET(ip, whichfork,
-                               XFS_IFORK_NEXTENTS(ip, whichfork) + 1);
-               } else {
-                       ASSERT(whichfork == XFS_DATA_FORK);
-                       temp = xfs_bmap_worst_indlen(ip, temp);
-                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
-                       temp2 = xfs_bmap_worst_indlen(ip, temp2);
-                       new.br_startblock = nullstartblock((int)temp2);
-                       da_new = temp + temp2;
-                       while (da_new > da_old) {
-                               if (temp) {
-                                       temp--;
-                                       da_new--;
-                                       xfs_bmbt_set_startblock(ep,
-                                               nullstartblock((int)temp));
-                               }
-                               if (da_new == da_old)
-                                       break;
-                               if (temp2) {
-                                       temp2--;
-                                       da_new--;
-                                       new.br_startblock =
-                                               nullstartblock((int)temp2);
-                               }
-                       }
+               temp = new->br_startoff - PREV.br_startoff;
+               temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, 0, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep, temp);      /* truncate PREV */
+               LEFT = *new;
+               RIGHT.br_state = PREV.br_state;
+               RIGHT.br_startblock = nullstartblock(
+                               (int)xfs_bmap_worst_indlen(bma->ip, temp2));
+               RIGHT.br_startoff = new_endoff;
+               RIGHT.br_blockcount = temp2;
+               /* insert LEFT (r[0]) and RIGHT (r[1]) at the same time */
+               xfs_iext_insert(bma->ip, bma->idx + 1, 2, &LEFT, state);
+               bma->ip->i_d.di_nextents++;
+               if (bma->cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       bma->cur->bc_rec.b.br_state = XFS_EXT_NORM;
+                       error = xfs_btree_insert(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
                }
-               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
-               xfs_iext_insert(ip, *idx + 1, 1, &new, state);
-               ++*idx;
-               break;
-       }
-       /*
-        * If we need to, add to list of extents to delete.
-        */
-       if (do_fx)
-               xfs_bmap_add_free(del->br_startblock, del->br_blockcount, flist,
-                       mp);
-       /*
-        * Adjust inode # blocks in the file.
-        */
-       if (nblks)
-               ip->i_d.di_nblocks -= nblks;
-       /*
-        * Adjust quota data.
-        */
-       if (qfield)
-               xfs_trans_mod_dquot_byino(tp, ip, qfield, (long)-nblks);
 
-       /*
-        * Account for change in delayed indirect blocks.
-        * Nothing to do for disk quota accounting here.
-        */
-       ASSERT(da_old >= da_new);
-       if (da_old > da_new) {
-               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
-                       (int64_t)(da_old - da_new), 0);
+               if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
+                       error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
+                                       bma->firstblock, bma->flist, &bma->cur,
+                                       1, &tmp_rval, XFS_DATA_FORK);
+                       rval |= tmp_rval;
+                       if (error)
+                               goto done;
+               }
+               temp = xfs_bmap_worst_indlen(bma->ip, temp);
+               temp2 = xfs_bmap_worst_indlen(bma->ip, temp2);
+               diff = (int)(temp + temp2 - startblockval(PREV.br_startblock) -
+                       (bma->cur ? bma->cur->bc_private.b.allocated : 0));
+               if (diff > 0) {
+                       error = xfs_icsb_modify_counters(bma->ip->i_mount,
+                                       XFS_SBS_FDBLOCKS,
+                                       -((int64_t)diff), 0);
+                       ASSERT(!error);
+                       if (error)
+                               goto done;
+               }
+
+               ep = xfs_iext_get_ext(ifp, bma->idx);
+               xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx + 2, state, _THIS_IP_);
+               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, bma->idx + 2),
+                       nullstartblock((int)temp2));
+               trace_xfs_bmap_post_update(bma->ip, bma->idx + 2, state, _THIS_IP_);
+
+               bma->idx++;
+               da_new = temp + temp2;
+               break;
+
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_CONTIG:
+       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
+       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_LEFT_CONTIG:
+       case BMAP_RIGHT_CONTIG:
+               /*
+                * These cases are all impossible.
+                */
+               ASSERT(0);
+       }
+
+       /* convert to a btree if necessary */
+       if (xfs_bmap_needs_btree(bma->ip, XFS_DATA_FORK)) {
+               int     tmp_logflags;   /* partial log flag return val */
+
+               ASSERT(bma->cur == NULL);
+               error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
+                               bma->firstblock, bma->flist, &bma->cur,
+                               da_old > 0, &tmp_logflags, XFS_DATA_FORK);
+               bma->logflags |= tmp_logflags;
+               if (error)
+                       goto done;
+       }
+
+       /* adjust for changes in reserved delayed indirect blocks */
+       if (da_old || da_new) {
+               temp = da_new;
+               if (bma->cur)
+                       temp += bma->cur->bc_private.b.allocated;
+               ASSERT(temp <= da_old);
+               if (temp < da_old)
+                       xfs_icsb_modify_counters(bma->ip->i_mount,
+                                       XFS_SBS_FDBLOCKS,
+                                       (int64_t)(da_old - temp), 0);
        }
+
+       /* clear out the allocated field, done with it now in any case. */
+       if (bma->cur)
+               bma->cur->bc_private.b.allocated = 0;
+
+       xfs_bmap_check_leaf_extents(bma->cur, bma->ip, XFS_DATA_FORK);
 done:
-       *logflagsp = flags;
+       bma->logflags |= rval;
        return error;
+#undef LEFT
+#undef RIGHT
+#undef PREV
 }
 
 /*
- * Remove the entry "free" from the free item list.  Prev points to the
- * previous entry, unless "free" is the head of the list.
+ * Convert an unwritten allocation to a real allocation or vice versa.
  */
-STATIC void
-xfs_bmap_del_free(
-       xfs_bmap_free_t         *flist, /* free item list header */
-       xfs_bmap_free_item_t    *prev,  /* previous item on list, if any */
-       xfs_bmap_free_item_t    *free)  /* list item to be freed */
+STATIC int                             /* error */
+xfs_bmap_add_extent_unwritten_real(
+       struct xfs_trans        *tp,
+       xfs_inode_t             *ip,    /* incore inode pointer */
+       xfs_extnum_t            *idx,   /* extent number to update/insert */
+       xfs_btree_cur_t         **curp, /* if *curp is null, not a btree */
+       xfs_bmbt_irec_t         *new,   /* new data to add to file extents */
+       xfs_fsblock_t           *first, /* pointer to firstblock variable */
+       xfs_bmap_free_t         *flist, /* list of extents to be freed */
+       int                     *logflagsp) /* inode logging flags */
 {
-       if (prev)
-               prev->xbfi_next = free->xbfi_next;
-       else
-               flist->xbf_first = free->xbfi_next;
-       flist->xbf_count--;
-       kmem_zone_free(xfs_bmap_free_item_zone, free);
-}
+       xfs_btree_cur_t         *cur;   /* btree cursor */
+       xfs_bmbt_rec_host_t     *ep;    /* extent entry for idx */
+       int                     error;  /* error return value */
+       int                     i;      /* temp state */
+       xfs_ifork_t             *ifp;   /* inode fork pointer */
+       xfs_fileoff_t           new_endoff;     /* end offset of new entry */
+       xfs_exntst_t            newext; /* new extent state */
+       xfs_exntst_t            oldext; /* old extent state */
+       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */
+                                       /* left is 0, right is 1, prev is 2 */
+       int                     rval=0; /* return value (logging flags) */
+       int                     state = 0;/* state bits, accessed thru macros */
 
-/*
- * Convert an extents-format file into a btree-format file.
- * The new file will have a root block (in the inode) and a single child block.
- */
-STATIC int                                     /* error */
-xfs_bmap_extents_to_btree(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       xfs_fsblock_t           *firstblock,    /* first-block-allocated */
-       xfs_bmap_free_t         *flist,         /* blocks freed in xaction */
-       xfs_btree_cur_t         **curp,         /* cursor returned to caller */
-       int                     wasdel,         /* converting a delayed alloc */
-       int                     *logflagsp,     /* inode logging flags */
-       int                     whichfork)      /* data or attr fork */
-{
-       struct xfs_btree_block  *ablock;        /* allocated (child) bt block */
-       xfs_buf_t               *abp;           /* buffer for ablock */
-       xfs_alloc_arg_t         args;           /* allocation arguments */
-       xfs_bmbt_rec_t          *arp;           /* child record pointer */
-       struct xfs_btree_block  *block;         /* btree root block */
-       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
-       xfs_bmbt_rec_host_t     *ep;            /* extent record pointer */
-       int                     error;          /* error return value */
-       xfs_extnum_t            i, cnt;         /* extent record index */
-       xfs_ifork_t             *ifp;           /* inode fork pointer */
-       xfs_bmbt_key_t          *kp;            /* root block key pointer */
-       xfs_mount_t             *mp;            /* mount structure */
-       xfs_extnum_t            nextents;       /* number of file extents */
-       xfs_bmbt_ptr_t          *pp;            /* root block address pointer */
+       *logflagsp = 0;
 
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS);
+       cur = *curp;
+       ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
 
-       /*
-        * Make space in the inode incore.
-        */
-       xfs_iroot_realloc(ip, 1, whichfork);
-       ifp->if_flags |= XFS_IFBROOT;
+       ASSERT(*idx >= 0);
+       ASSERT(*idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
+       ASSERT(!isnullstartblock(new->br_startblock));
+
+       XFS_STATS_INC(xs_add_exlist);
+
+#define        LEFT            r[0]
+#define        RIGHT           r[1]
+#define        PREV            r[2]
 
        /*
-        * Fill in the root.
+        * Set up a bunch of variables to make the tests simpler.
         */
-       block = ifp->if_broot;
-       block->bb_magic = cpu_to_be32(XFS_BMAP_MAGIC);
-       block->bb_level = cpu_to_be16(1);
-       block->bb_numrecs = cpu_to_be16(1);
-       block->bb_u.l.bb_leftsib = cpu_to_be64(NULLDFSBNO);
-       block->bb_u.l.bb_rightsib = cpu_to_be64(NULLDFSBNO);
+       error = 0;
+       ep = xfs_iext_get_ext(ifp, *idx);
+       xfs_bmbt_get_all(ep, &PREV);
+       newext = new->br_state;
+       oldext = (newext == XFS_EXT_UNWRITTEN) ?
+               XFS_EXT_NORM : XFS_EXT_UNWRITTEN;
+       ASSERT(PREV.br_state == oldext);
+       new_endoff = new->br_startoff + new->br_blockcount;
+       ASSERT(PREV.br_startoff <= new->br_startoff);
+       ASSERT(PREV.br_startoff + PREV.br_blockcount >= new_endoff);
 
        /*
-        * Need a cursor.  Can't allocate until bb_level is filled in.
+        * Set flags determining what part of the previous oldext allocation
+        * extent is being replaced by a newext allocation.
         */
-       mp = ip->i_mount;
-       cur = xfs_bmbt_init_cursor(mp, tp, ip, whichfork);
-       cur->bc_private.b.firstblock = *firstblock;
-       cur->bc_private.b.flist = flist;
-       cur->bc_private.b.flags = wasdel ? XFS_BTCUR_BPRV_WASDEL : 0;
+       if (PREV.br_startoff == new->br_startoff)
+               state |= BMAP_LEFT_FILLING;
+       if (PREV.br_startoff + PREV.br_blockcount == new_endoff)
+               state |= BMAP_RIGHT_FILLING;
+
        /*
-        * Convert to a btree with two levels, one record in root.
+        * Check and set flags if this segment has a left neighbor.
+        * Don't set contiguous if the combined extent would be too large.
         */
-       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_BTREE);
-       memset(&args, 0, sizeof(args));
-       args.tp = tp;
-       args.mp = mp;
-       args.firstblock = *firstblock;
-       if (*firstblock == NULLFSBLOCK) {
-               args.type = XFS_ALLOCTYPE_START_BNO;
-               args.fsbno = XFS_INO_TO_FSB(mp, ip->i_ino);
-       } else if (flist->xbf_low) {
-               args.type = XFS_ALLOCTYPE_START_BNO;
-               args.fsbno = *firstblock;
-       } else {
-               args.type = XFS_ALLOCTYPE_NEAR_BNO;
-               args.fsbno = *firstblock;
-       }
-       args.minlen = args.maxlen = args.prod = 1;
-       args.wasdel = wasdel;
-       *logflagsp = 0;
-       if ((error = xfs_alloc_vextent(&args))) {
-               xfs_iroot_realloc(ip, -1, whichfork);
-               xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
-               return error;
+       if (*idx > 0) {
+               state |= BMAP_LEFT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx - 1), &LEFT);
+
+               if (isnullstartblock(LEFT.br_startblock))
+                       state |= BMAP_LEFT_DELAY;
        }
+
+       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
+           LEFT.br_startoff + LEFT.br_blockcount == new->br_startoff &&
+           LEFT.br_startblock + LEFT.br_blockcount == new->br_startblock &&
+           LEFT.br_state == newext &&
+           LEFT.br_blockcount + new->br_blockcount <= MAXEXTLEN)
+               state |= BMAP_LEFT_CONTIG;
+
        /*
-        * Allocation can't fail, the space was reserved.
+        * Check and set flags if this segment has a right neighbor.
+        * Don't set contiguous if the combined extent would be too large.
+        * Also check for all-three-contiguous being too large.
         */
-       ASSERT(args.fsbno != NULLFSBLOCK);
-       ASSERT(*firstblock == NULLFSBLOCK ||
-              args.agno == XFS_FSB_TO_AGNO(mp, *firstblock) ||
-              (flist->xbf_low &&
-               args.agno > XFS_FSB_TO_AGNO(mp, *firstblock)));
-       *firstblock = cur->bc_private.b.firstblock = args.fsbno;
-       cur->bc_private.b.allocated++;
-       ip->i_d.di_nblocks++;
-       xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, 1L);
-       abp = xfs_btree_get_bufl(mp, tp, args.fsbno, 0);
-       /*
-        * Fill in the child block.
-        */
-       abp->b_ops = &xfs_bmbt_buf_ops;
-       ablock = XFS_BUF_TO_BLOCK(abp);
-       ablock->bb_magic = cpu_to_be32(XFS_BMAP_MAGIC);
-       ablock->bb_level = 0;
-       ablock->bb_u.l.bb_leftsib = cpu_to_be64(NULLDFSBNO);
-       ablock->bb_u.l.bb_rightsib = cpu_to_be64(NULLDFSBNO);
-       arp = XFS_BMBT_REC_ADDR(mp, ablock, 1);
-       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
-       for (cnt = i = 0; i < nextents; i++) {
-               ep = xfs_iext_get_ext(ifp, i);
-               if (!isnullstartblock(xfs_bmbt_get_startblock(ep))) {
-                       arp->l0 = cpu_to_be64(ep->l0);
-                       arp->l1 = cpu_to_be64(ep->l1);
-                       arp++; cnt++;
-               }
+       if (*idx < ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t) - 1) {
+               state |= BMAP_RIGHT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx + 1), &RIGHT);
+               if (isnullstartblock(RIGHT.br_startblock))
+                       state |= BMAP_RIGHT_DELAY;
        }
-       ASSERT(cnt == XFS_IFORK_NEXTENTS(ip, whichfork));
-       xfs_btree_set_numrecs(ablock, cnt);
 
-       /*
-        * Fill in the root key and pointer.
-        */
-       kp = XFS_BMBT_KEY_ADDR(mp, block, 1);
-       arp = XFS_BMBT_REC_ADDR(mp, ablock, 1);
-       kp->br_startoff = cpu_to_be64(xfs_bmbt_disk_get_startoff(arp));
-       pp = XFS_BMBT_PTR_ADDR(mp, block, 1, xfs_bmbt_get_maxrecs(cur,
-                                               be16_to_cpu(block->bb_level)));
-       *pp = cpu_to_be64(args.fsbno);
+       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
+           new_endoff == RIGHT.br_startoff &&
+           new->br_startblock + new->br_blockcount == RIGHT.br_startblock &&
+           newext == RIGHT.br_state &&
+           new->br_blockcount + RIGHT.br_blockcount <= MAXEXTLEN &&
+           ((state & (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
+                      BMAP_RIGHT_FILLING)) !=
+                     (BMAP_LEFT_CONTIG | BMAP_LEFT_FILLING |
+                      BMAP_RIGHT_FILLING) ||
+            LEFT.br_blockcount + new->br_blockcount + RIGHT.br_blockcount
+                       <= MAXEXTLEN))
+               state |= BMAP_RIGHT_CONTIG;
 
        /*
-        * Do all this logging at the end so that
-        * the root is at the right level.
+        * Switch out based on the FILLING and CONTIG state bits.
         */
-       xfs_btree_log_block(cur, abp, XFS_BB_ALL_BITS);
-       xfs_btree_log_recs(cur, abp, 1, be16_to_cpu(ablock->bb_numrecs));
-       ASSERT(*curp == NULL);
-       *curp = cur;
-       *logflagsp = XFS_ILOG_CORE | xfs_ilog_fbroot(whichfork);
-       return 0;
-}
-
-/*
- * Calculate the default attribute fork offset for newly created inodes.
- */
-uint
-xfs_default_attroffset(
-       struct xfs_inode        *ip)
-{
-       struct xfs_mount        *mp = ip->i_mount;
-       uint                    offset;
-
-       if (mp->m_sb.sb_inodesize == 256) {
-               offset = XFS_LITINO(mp) -
-                               XFS_BMDR_SPACE_CALC(MINABTPTRS);
-       } else {
-               offset = XFS_BMDR_SPACE_CALC(6 * MINABTPTRS);
-       }
+       switch (state & (BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
+                        BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG)) {
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG |
+            BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
+               /*
+                * Setting all of a previous oldext extent to newext.
+                * The left and right neighbors are both contiguous with new.
+                */
+               --*idx;
 
-       ASSERT(offset < XFS_LITINO(mp));
-       return offset;
-}
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx),
+                       LEFT.br_blockcount + PREV.br_blockcount +
+                       RIGHT.br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-/*
- * Helper routine to reset inode di_forkoff field when switching
- * attribute fork from local to extent format - we reset it where
- * possible to make space available for inline data fork extents.
- */
-STATIC void
-xfs_bmap_forkoff_reset(
-       xfs_mount_t     *mp,
-       xfs_inode_t     *ip,
-       int             whichfork)
-{
-       if (whichfork == XFS_ATTR_FORK &&
-           ip->i_d.di_format != XFS_DINODE_FMT_DEV &&
-           ip->i_d.di_format != XFS_DINODE_FMT_UUID &&
-           ip->i_d.di_format != XFS_DINODE_FMT_BTREE) {
-               uint    dfl_forkoff = xfs_default_attroffset(ip) >> 3;
+               xfs_iext_remove(ip, *idx + 1, 2, state);
+               ip->i_d.di_nextents -= 2;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, RIGHT.br_startoff,
+                                       RIGHT.br_startblock,
+                                       RIGHT.br_blockcount, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_delete(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_decrement(cur, 0, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_delete(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_decrement(cur, 0, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, LEFT.br_startoff,
+                               LEFT.br_startblock,
+                               LEFT.br_blockcount + PREV.br_blockcount +
+                               RIGHT.br_blockcount, LEFT.br_state)))
+                               goto done;
+               }
+               break;
 
-               if (dfl_forkoff > ip->i_d.di_forkoff)
-                       ip->i_d.di_forkoff = dfl_forkoff;
-       }
-}
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
+               /*
+                * Setting all of a previous oldext extent to newext.
+                * The left neighbor is contiguous, the right is not.
+                */
+               --*idx;
 
-/*
- * Convert a local file to an extents file.
- * This code is out of bounds for data forks of regular files,
- * since the file data needs to get logged so things will stay consistent.
- * (The bmap-level manipulations are ok, though).
- */
-STATIC int                             /* error */
-xfs_bmap_local_to_extents(
-       xfs_trans_t     *tp,            /* transaction pointer */
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_fsblock_t   *firstblock,    /* first block allocated in xaction */
-       xfs_extlen_t    total,          /* total blocks needed by transaction */
-       int             *logflagsp,     /* inode logging flags */
-       int             whichfork,
-       void            (*init_fn)(struct xfs_buf *bp,
-                                  struct xfs_inode *ip,
-                                  struct xfs_ifork *ifp))
-{
-       int             error;          /* error return value */
-       int             flags;          /* logging flags returned */
-       xfs_ifork_t     *ifp;           /* inode fork pointer */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx),
+                       LEFT.br_blockcount + PREV.br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       /*
-        * We don't want to deal with the case of keeping inode data inline yet.
-        * So sending the data fork of a regular inode is invalid.
-        */
-       ASSERT(!(S_ISREG(ip->i_d.di_mode) && whichfork == XFS_DATA_FORK));
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL);
-       flags = 0;
-       error = 0;
-       if (ifp->if_bytes) {
-               xfs_alloc_arg_t args;   /* allocation arguments */
-               xfs_buf_t       *bp;    /* buffer for extent block */
-               xfs_bmbt_rec_host_t *ep;/* extent record pointer */
+               xfs_iext_remove(ip, *idx + 1, 1, state);
+               ip->i_d.di_nextents--;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock, PREV.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_delete(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_decrement(cur, 0, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, LEFT.br_startoff,
+                               LEFT.br_startblock,
+                               LEFT.br_blockcount + PREV.br_blockcount,
+                               LEFT.br_state)))
+                               goto done;
+               }
+               break;
 
-               ASSERT((ifp->if_flags &
-                       (XFS_IFINLINE|XFS_IFEXTENTS|XFS_IFEXTIREC)) == XFS_IFINLINE);
-               memset(&args, 0, sizeof(args));
-               args.tp = tp;
-               args.mp = ip->i_mount;
-               args.firstblock = *firstblock;
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
                /*
-                * Allocate a block.  We know we need only one, since the
-                * file currently fits in an inode.
+                * Setting all of a previous oldext extent to newext.
+                * The right neighbor is contiguous, the left is not.
                 */
-               if (*firstblock == NULLFSBLOCK) {
-                       args.fsbno = XFS_INO_TO_FSB(args.mp, ip->i_ino);
-                       args.type = XFS_ALLOCTYPE_START_BNO;
-               } else {
-                       args.fsbno = *firstblock;
-                       args.type = XFS_ALLOCTYPE_NEAR_BNO;
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount + RIGHT.br_blockcount);
+               xfs_bmbt_set_state(ep, newext);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               xfs_iext_remove(ip, *idx + 1, 1, state);
+               ip->i_d.di_nextents--;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, RIGHT.br_startoff,
+                                       RIGHT.br_startblock,
+                                       RIGHT.br_blockcount, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_delete(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_btree_decrement(cur, 0, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
+                               new->br_startblock,
+                               new->br_blockcount + RIGHT.br_blockcount,
+                               newext)))
+                               goto done;
                }
-               args.total = total;
-               args.minlen = args.maxlen = args.prod = 1;
-               error = xfs_alloc_vextent(&args);
-               if (error)
-                       goto done;
+               break;
 
-               /* Can't fail, the space was reserved. */
-               ASSERT(args.fsbno != NULLFSBLOCK);
-               ASSERT(args.len == 1);
-               *firstblock = args.fsbno;
-               bp = xfs_btree_get_bufl(args.mp, tp, args.fsbno, 0);
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING:
+               /*
+                * Setting all of a previous oldext extent to newext.
+                * Neither the left nor right neighbors are contiguous with
+                * the new one.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_state(ep, newext);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-               /* initialise the block and copy the data */
-               init_fn(bp, ip, ifp);
+               if (cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
+                               new->br_startblock, new->br_blockcount,
+                               newext)))
+                               goto done;
+               }
+               break;
 
-               /* account for the change in fork size and log everything */
-               xfs_trans_log_buf(tp, bp, 0, ifp->if_bytes - 1);
-               xfs_bmap_forkoff_reset(args.mp, ip, whichfork);
-               xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
-               xfs_iext_add(ifp, 0, 1);
-               ep = xfs_iext_get_ext(ifp, 0);
-               xfs_bmbt_set_allf(ep, 0, args.fsbno, 1, XFS_EXT_NORM);
-               trace_xfs_bmap_post_update(ip, 0,
-                               whichfork == XFS_ATTR_FORK ? BMAP_ATTRFORK : 0,
-                               _THIS_IP_);
-               XFS_IFORK_NEXT_SET(ip, whichfork, 1);
-               ip->i_d.di_nblocks = 1;
-               xfs_trans_mod_dquot_byino(tp, ip,
-                       XFS_TRANS_DQ_BCOUNT, 1L);
-               flags |= xfs_ilog_fext(whichfork);
-       } else {
-               ASSERT(XFS_IFORK_NEXTENTS(ip, whichfork) == 0);
-               xfs_bmap_forkoff_reset(ip->i_mount, ip, whichfork);
-       }
-       ifp->if_flags &= ~XFS_IFINLINE;
-       ifp->if_flags |= XFS_IFEXTENTS;
-       XFS_IFORK_FMT_SET(ip, whichfork, XFS_DINODE_FMT_EXTENTS);
-       flags |= XFS_ILOG_CORE;
-done:
-       *logflagsp = flags;
-       return error;
-}
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG:
+               /*
+                * Setting the first part of a previous oldext extent to newext.
+                * The left neighbor is contiguous.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx - 1, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx - 1),
+                       LEFT.br_blockcount + new->br_blockcount);
+               xfs_bmbt_set_startoff(ep,
+                       PREV.br_startoff + new->br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx - 1, state, _THIS_IP_);
 
-/*
- * Search the extent records for the entry containing block bno.
- * If bno lies in a hole, point to the next entry.  If bno lies
- * past eof, *eofp will be set, and *prevp will contain the last
- * entry (null if none).  Else, *lastxp will be set to the index
- * of the found entry; *gotp will contain the entry.
- */
-STATIC xfs_bmbt_rec_host_t *           /* pointer to found extent entry */
-xfs_bmap_search_multi_extents(
-       xfs_ifork_t     *ifp,           /* inode fork pointer */
-       xfs_fileoff_t   bno,            /* block number searched for */
-       int             *eofp,          /* out: end of file found */
-       xfs_extnum_t    *lastxp,        /* out: last extent index */
-       xfs_bmbt_irec_t *gotp,          /* out: extent entry found */
-       xfs_bmbt_irec_t *prevp)         /* out: previous extent entry found */
-{
-       xfs_bmbt_rec_host_t *ep;                /* extent record pointer */
-       xfs_extnum_t    lastx;          /* last extent index */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_startblock(ep,
+                       new->br_startblock + new->br_blockcount);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount - new->br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       /*
-        * Initialize the extent entry structure to catch access to
-        * uninitialized br_startblock field.
-        */
-       gotp->br_startoff = 0xffa5a5a5a5a5a5a5LL;
-       gotp->br_blockcount = 0xa55a5a5a5a5a5a5aLL;
-       gotp->br_state = XFS_EXT_INVALID;
-#if XFS_BIG_BLKNOS
-       gotp->br_startblock = 0xffffa5a5a5a5a5a5LL;
-#else
-       gotp->br_startblock = 0xffffa5a5;
-#endif
-       prevp->br_startoff = NULLFILEOFF;
+               --*idx;
 
-       ep = xfs_iext_bno_to_ext(ifp, bno, &lastx);
-       if (lastx > 0) {
-               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx - 1), prevp);
-       }
-       if (lastx < (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) {
-               xfs_bmbt_get_all(ep, gotp);
-               *eofp = 0;
-       } else {
-               if (lastx > 0) {
-                       *gotp = *prevp;
+               if (cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock, PREV.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur,
+                               PREV.br_startoff + new->br_blockcount,
+                               PREV.br_startblock + new->br_blockcount,
+                               PREV.br_blockcount - new->br_blockcount,
+                               oldext)))
+                               goto done;
+                       if ((error = xfs_btree_decrement(cur, 0, &i)))
+                               goto done;
+                       error = xfs_bmbt_update(cur, LEFT.br_startoff,
+                               LEFT.br_startblock,
+                               LEFT.br_blockcount + new->br_blockcount,
+                               LEFT.br_state);
+                       if (error)
+                               goto done;
                }
-               *eofp = 1;
-               ep = NULL;
-       }
-       *lastxp = lastx;
-       return ep;
-}
+               break;
 
-/*
- * Search the extents list for the inode, for the extent containing bno.
- * If bno lies in a hole, point to the next entry.  If bno lies past eof,
- * *eofp will be set, and *prevp will contain the last entry (null if none).
- * Else, *lastxp will be set to the index of the found
- * entry; *gotp will contain the entry.
- */
-STATIC xfs_bmbt_rec_host_t *                 /* pointer to found extent entry */
-xfs_bmap_search_extents(
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_fileoff_t   bno,            /* block number searched for */
-       int             fork,           /* data or attr fork */
-       int             *eofp,          /* out: end of file found */
-       xfs_extnum_t    *lastxp,        /* out: last extent index */
-       xfs_bmbt_irec_t *gotp,          /* out: extent entry found */
-       xfs_bmbt_irec_t *prevp)         /* out: previous extent entry found */
-{
-       xfs_ifork_t     *ifp;           /* inode fork pointer */
-       xfs_bmbt_rec_host_t  *ep;            /* extent record pointer */
+       case BMAP_LEFT_FILLING:
+               /*
+                * Setting the first part of a previous oldext extent to newext.
+                * The left neighbor is not contiguous.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               ASSERT(ep && xfs_bmbt_get_state(ep) == oldext);
+               xfs_bmbt_set_startoff(ep, new_endoff);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount - new->br_blockcount);
+               xfs_bmbt_set_startblock(ep,
+                       new->br_startblock + new->br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       XFS_STATS_INC(xs_look_exlist);
-       ifp = XFS_IFORK_PTR(ip, fork);
+               xfs_iext_insert(ip, *idx, 1, new, state);
+               ip->i_d.di_nextents++;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock, PREV.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur,
+                               PREV.br_startoff + new->br_blockcount,
+                               PREV.br_startblock + new->br_blockcount,
+                               PREV.br_blockcount - new->br_blockcount,
+                               oldext)))
+                               goto done;
+                       cur->bc_rec.b = *new;
+                       if ((error = xfs_btree_insert(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               }
+               break;
 
-       ep = xfs_bmap_search_multi_extents(ifp, bno, eofp, lastxp, gotp, prevp);
+       case BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
+               /*
+                * Setting the last part of a previous oldext extent to newext.
+                * The right neighbor is contiguous with the new allocation.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount - new->br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       if (unlikely(!(gotp->br_startblock) && (*lastxp != NULLEXTNUM) &&
-                    !(XFS_IS_REALTIME_INODE(ip) && fork == XFS_DATA_FORK))) {
-               xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
-                               "Access to block zero in inode %llu "
-                               "start_block: %llx start_off: %llx "
-                               "blkcnt: %llx extent-state: %x lastx: %x\n",
-                       (unsigned long long)ip->i_ino,
-                       (unsigned long long)gotp->br_startblock,
-                       (unsigned long long)gotp->br_startoff,
-                       (unsigned long long)gotp->br_blockcount,
-                       gotp->br_state, *lastxp);
-               *lastxp = NULLEXTNUM;
-               *eofp = 1;
-               return NULL;
-       }
-       return ep;
-}
+               ++*idx;
 
-/*
- * Compute the worst-case number of indirect blocks that will be used
- * for ip's delayed extent of length "len".
- */
-STATIC xfs_filblks_t
-xfs_bmap_worst_indlen(
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_filblks_t   len)            /* delayed extent length */
-{
-       int             level;          /* btree level number */
-       int             maxrecs;        /* maximum record count at this level */
-       xfs_mount_t     *mp;            /* mount structure */
-       xfs_filblks_t   rval;           /* return value */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, *idx),
+                       new->br_startoff, new->br_startblock,
+                       new->br_blockcount + RIGHT.br_blockcount, newext);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       mp = ip->i_mount;
-       maxrecs = mp->m_bmap_dmxr[0];
-       for (level = 0, rval = 0;
-            level < XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK);
-            level++) {
-               len += maxrecs - 1;
-               do_div(len, maxrecs);
-               rval += len;
-               if (len == 1)
-                       return rval + XFS_BM_MAXLEVELS(mp, XFS_DATA_FORK) -
-                               level - 1;
-               if (level == 0)
-                       maxrecs = mp->m_bmap_dmxr[1];
-       }
-       return rval;
-}
+               if (cur == NULL)
+                       rval = XFS_ILOG_DEXT;
+               else {
+                       rval = 0;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock,
+                                       PREV.br_blockcount, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, PREV.br_startoff,
+                               PREV.br_startblock,
+                               PREV.br_blockcount - new->br_blockcount,
+                               oldext)))
+                               goto done;
+                       if ((error = xfs_btree_increment(cur, 0, &i)))
+                               goto done;
+                       if ((error = xfs_bmbt_update(cur, new->br_startoff,
+                               new->br_startblock,
+                               new->br_blockcount + RIGHT.br_blockcount,
+                               newext)))
+                               goto done;
+               }
+               break;
 
-/*
- * Convert inode from non-attributed to attributed.
- * Must not be in a transaction, ip must not be locked.
- */
-int                                            /* error code */
-xfs_bmap_add_attrfork(
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       int                     size,           /* space new attribute needs */
-       int                     rsvd)           /* xact may use reserved blks */
-{
-       xfs_fsblock_t           firstblock;     /* 1st block/ag allocated */
-       xfs_bmap_free_t         flist;          /* freed extent records */
-       xfs_mount_t             *mp;            /* mount structure */
-       xfs_trans_t             *tp;            /* transaction pointer */
-       int                     blks;           /* space reservation */
-       int                     version = 1;    /* superblock attr version */
-       int                     committed;      /* xaction was committed */
-       int                     logflags;       /* logging flags */
-       int                     error;          /* error return value */
+       case BMAP_RIGHT_FILLING:
+               /*
+                * Setting the last part of a previous oldext extent to newext.
+                * The right neighbor is not contiguous.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep,
+                       PREV.br_blockcount - new->br_blockcount);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       ASSERT(XFS_IFORK_Q(ip) == 0);
+               ++*idx;
+               xfs_iext_insert(ip, *idx, 1, new, state);
 
-       mp = ip->i_mount;
-       ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
-       tp = xfs_trans_alloc(mp, XFS_TRANS_ADDAFORK);
-       blks = XFS_ADDAFORK_SPACE_RES(mp);
-       if (rsvd)
-               tp->t_flags |= XFS_TRANS_RESERVE;
-       if ((error = xfs_trans_reserve(tp, blks, XFS_ADDAFORK_LOG_RES(mp), 0,
-                       XFS_TRANS_PERM_LOG_RES, XFS_ADDAFORK_LOG_COUNT)))
-               goto error0;
-       xfs_ilock(ip, XFS_ILOCK_EXCL);
-       error = xfs_trans_reserve_quota_nblks(tp, ip, blks, 0, rsvd ?
-                       XFS_QMOPT_RES_REGBLKS | XFS_QMOPT_FORCE_RES :
-                       XFS_QMOPT_RES_REGBLKS);
-       if (error) {
-               xfs_iunlock(ip, XFS_ILOCK_EXCL);
-               xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES);
-               return error;
-       }
-       if (XFS_IFORK_Q(ip))
-               goto error1;
-       if (ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS) {
+               ip->i_d.di_nextents++;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock, PREV.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       if ((error = xfs_bmbt_update(cur, PREV.br_startoff,
+                               PREV.br_startblock,
+                               PREV.br_blockcount - new->br_blockcount,
+                               oldext)))
+                               goto done;
+                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       cur->bc_rec.b.br_state = XFS_EXT_NORM;
+                       if ((error = xfs_btree_insert(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               }
+               break;
+
+       case 0:
                /*
-                * For inodes coming from pre-6.2 filesystems.
+                * Setting the middle part of a previous oldext extent to
+                * newext.  Contiguity is impossible here.
+                * One extent becomes three extents.
                 */
-               ASSERT(ip->i_d.di_aformat == 0);
-               ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
-       }
-       ASSERT(ip->i_d.di_anextents == 0);
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep,
+                       new->br_startoff - PREV.br_startoff);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-       xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
-       xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
+               r[0] = *new;
+               r[1].br_startoff = new_endoff;
+               r[1].br_blockcount =
+                       PREV.br_startoff + PREV.br_blockcount - new_endoff;
+               r[1].br_startblock = new->br_startblock + new->br_blockcount;
+               r[1].br_state = oldext;
 
-       switch (ip->i_d.di_format) {
-       case XFS_DINODE_FMT_DEV:
-               ip->i_d.di_forkoff = roundup(sizeof(xfs_dev_t), 8) >> 3;
-               break;
-       case XFS_DINODE_FMT_UUID:
-               ip->i_d.di_forkoff = roundup(sizeof(uuid_t), 8) >> 3;
-               break;
-       case XFS_DINODE_FMT_LOCAL:
-       case XFS_DINODE_FMT_EXTENTS:
-       case XFS_DINODE_FMT_BTREE:
-               ip->i_d.di_forkoff = xfs_attr_shortform_bytesfit(ip, size);
-               if (!ip->i_d.di_forkoff)
-                       ip->i_d.di_forkoff = xfs_default_attroffset(ip) >> 3;
-               else if (mp->m_flags & XFS_MOUNT_ATTR2)
-                       version = 2;
+               ++*idx;
+               xfs_iext_insert(ip, *idx, 2, &r[0], state);
+
+               ip->i_d.di_nextents += 2;
+               if (cur == NULL)
+                       rval = XFS_ILOG_CORE | XFS_ILOG_DEXT;
+               else {
+                       rval = XFS_ILOG_CORE;
+                       if ((error = xfs_bmbt_lookup_eq(cur, PREV.br_startoff,
+                                       PREV.br_startblock, PREV.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       /* new right extent - oldext */
+                       if ((error = xfs_bmbt_update(cur, r[1].br_startoff,
+                               r[1].br_startblock, r[1].br_blockcount,
+                               r[1].br_state)))
+                               goto done;
+                       /* new left extent - oldext */
+                       cur->bc_rec.b = PREV;
+                       cur->bc_rec.b.br_blockcount =
+                               new->br_startoff - PREV.br_startoff;
+                       if ((error = xfs_btree_insert(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       /*
+                        * Reset the cursor to the position of the new extent
+                        * we are about to insert as we can't trust it after
+                        * the previous insert.
+                        */
+                       if ((error = xfs_bmbt_lookup_eq(cur, new->br_startoff,
+                                       new->br_startblock, new->br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       /* new middle extent - newext */
+                       cur->bc_rec.b.br_state = new->br_state;
+                       if ((error = xfs_btree_insert(cur, &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               }
                break;
-       default:
+
+       case BMAP_LEFT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_LEFT_FILLING | BMAP_RIGHT_CONTIG:
+       case BMAP_RIGHT_FILLING | BMAP_LEFT_CONTIG:
+       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+       case BMAP_LEFT_CONTIG:
+       case BMAP_RIGHT_CONTIG:
+               /*
+                * These cases are all impossible.
+                */
                ASSERT(0);
-               error = XFS_ERROR(EINVAL);
-               goto error1;
        }
 
-       ASSERT(ip->i_afp == NULL);
-       ip->i_afp = kmem_zone_zalloc(xfs_ifork_zone, KM_SLEEP);
-       ip->i_afp->if_flags = XFS_IFEXTENTS;
-       logflags = 0;
-       xfs_bmap_init(&flist, &firstblock);
-       switch (ip->i_d.di_format) {
-       case XFS_DINODE_FMT_LOCAL:
-               error = xfs_bmap_add_attrfork_local(tp, ip, &firstblock, &flist,
-                       &logflags);
-               break;
-       case XFS_DINODE_FMT_EXTENTS:
-               error = xfs_bmap_add_attrfork_extents(tp, ip, &firstblock,
-                       &flist, &logflags);
-               break;
-       case XFS_DINODE_FMT_BTREE:
-               error = xfs_bmap_add_attrfork_btree(tp, ip, &firstblock, &flist,
-                       &logflags);
-               break;
-       default:
-               error = 0;
-               break;
+       /* convert to a btree if necessary */
+       if (xfs_bmap_needs_btree(ip, XFS_DATA_FORK)) {
+               int     tmp_logflags;   /* partial log flag return val */
+
+               ASSERT(cur == NULL);
+               error = xfs_bmap_extents_to_btree(tp, ip, first, flist, &cur,
+                               0, &tmp_logflags, XFS_DATA_FORK);
+               *logflagsp |= tmp_logflags;
+               if (error)
+                       goto done;
        }
-       if (logflags)
-               xfs_trans_log_inode(tp, ip, logflags);
-       if (error)
-               goto error2;
-       if (!xfs_sb_version_hasattr(&mp->m_sb) ||
-          (!xfs_sb_version_hasattr2(&mp->m_sb) && version == 2)) {
-               __int64_t sbfields = 0;
 
-               spin_lock(&mp->m_sb_lock);
-               if (!xfs_sb_version_hasattr(&mp->m_sb)) {
-                       xfs_sb_version_addattr(&mp->m_sb);
-                       sbfields |= XFS_SB_VERSIONNUM;
-               }
-               if (!xfs_sb_version_hasattr2(&mp->m_sb) && version == 2) {
-                       xfs_sb_version_addattr2(&mp->m_sb);
-                       sbfields |= (XFS_SB_VERSIONNUM | XFS_SB_FEATURES2);
-               }
-               if (sbfields) {
-                       spin_unlock(&mp->m_sb_lock);
-                       xfs_mod_sb(tp, sbfields);
-               } else
-                       spin_unlock(&mp->m_sb_lock);
+       /* clear out the allocated field, done with it now in any case. */
+       if (cur) {
+               cur->bc_private.b.allocated = 0;
+               *curp = cur;
        }
 
-       error = xfs_bmap_finish(&tp, &flist, &committed);
-       if (error)
-               goto error2;
-       return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
-error2:
-       xfs_bmap_cancel(&flist);
-error1:
-       xfs_iunlock(ip, XFS_ILOCK_EXCL);
-error0:
-       xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
+       xfs_bmap_check_leaf_extents(*curp, ip, XFS_DATA_FORK);
+done:
+       *logflagsp |= rval;
        return error;
+#undef LEFT
+#undef RIGHT
+#undef PREV
 }
 
 /*
- * Add the extent to the list of extents to be free at transaction end.
- * The list is maintained sorted (by block number).
+ * Convert a hole to a delayed allocation.
  */
-/* ARGSUSED */
-void
-xfs_bmap_add_free(
-       xfs_fsblock_t           bno,            /* fs block number of extent */
-       xfs_filblks_t           len,            /* length of extent */
-       xfs_bmap_free_t         *flist,         /* list of extents */
-       xfs_mount_t             *mp)            /* mount point structure */
+STATIC void
+xfs_bmap_add_extent_hole_delay(
+       xfs_inode_t             *ip,    /* incore inode pointer */
+       xfs_extnum_t            *idx,   /* extent number to update/insert */
+       xfs_bmbt_irec_t         *new)   /* new data to add to file extents */
 {
-       xfs_bmap_free_item_t    *cur;           /* current (next) element */
-       xfs_bmap_free_item_t    *new;           /* new element */
-       xfs_bmap_free_item_t    *prev;          /* previous element */
-#ifdef DEBUG
-       xfs_agnumber_t          agno;
-       xfs_agblock_t           agbno;
-
-       ASSERT(bno != NULLFSBLOCK);
-       ASSERT(len > 0);
-       ASSERT(len <= MAXEXTLEN);
-       ASSERT(!isnullstartblock(bno));
-       agno = XFS_FSB_TO_AGNO(mp, bno);
-       agbno = XFS_FSB_TO_AGBNO(mp, bno);
-       ASSERT(agno < mp->m_sb.sb_agcount);
-       ASSERT(agbno < mp->m_sb.sb_agblocks);
-       ASSERT(len < mp->m_sb.sb_agblocks);
-       ASSERT(agbno + len <= mp->m_sb.sb_agblocks);
-#endif
-       ASSERT(xfs_bmap_free_item_zone != NULL);
-       new = kmem_zone_alloc(xfs_bmap_free_item_zone, KM_SLEEP);
-       new->xbfi_startblock = bno;
-       new->xbfi_blockcount = (xfs_extlen_t)len;
-       for (prev = NULL, cur = flist->xbf_first;
-            cur != NULL;
-            prev = cur, cur = cur->xbfi_next) {
-               if (cur->xbfi_startblock >= bno)
-                       break;
-       }
-       if (prev)
-               prev->xbfi_next = new;
-       else
-               flist->xbf_first = new;
-       new->xbfi_next = cur;
-       flist->xbf_count++;
-}
+       xfs_ifork_t             *ifp;   /* inode fork pointer */
+       xfs_bmbt_irec_t         left;   /* left neighbor extent entry */
+       xfs_filblks_t           newlen=0;       /* new indirect size */
+       xfs_filblks_t           oldlen=0;       /* old indirect size */
+       xfs_bmbt_irec_t         right;  /* right neighbor extent entry */
+       int                     state;  /* state bits, accessed thru macros */
+       xfs_filblks_t           temp=0; /* temp for indirect calculations */
 
-/*
- * Compute and fill in the value of the maximum depth of a bmap btree
- * in this filesystem.  Done once, during mount.
- */
-void
-xfs_bmap_compute_maxlevels(
-       xfs_mount_t     *mp,            /* file system mount structure */
-       int             whichfork)      /* data or attr fork */
-{
-       int             level;          /* btree level */
-       uint            maxblocks;      /* max blocks at this level */
-       uint            maxleafents;    /* max leaf entries possible */
-       int             maxrootrecs;    /* max records in root block */
-       int             minleafrecs;    /* min records in leaf block */
-       int             minnoderecs;    /* min records in node block */
-       int             sz;             /* root block size */
+       ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
+       state = 0;
+       ASSERT(isnullstartblock(new->br_startblock));
 
        /*
-        * The maximum number of extents in a file, hence the maximum
-        * number of leaf entries, is controlled by the type of di_nextents
-        * (a signed 32-bit number, xfs_extnum_t), or by di_anextents
-        * (a signed 16-bit number, xfs_aextnum_t).
-        *
-        * Note that we can no longer assume that if we are in ATTR1 that
-        * the fork offset of all the inodes will be
-        * (xfs_default_attroffset(ip) >> 3) because we could have mounted
-        * with ATTR2 and then mounted back with ATTR1, keeping the
-        * di_forkoff's fixed but probably at various positions. Therefore,
-        * for both ATTR1 and ATTR2 we have to assume the worst case scenario
-        * of a minimum size available.
+        * Check and set flags if this segment has a left neighbor
         */
-       if (whichfork == XFS_DATA_FORK) {
-               maxleafents = MAXEXTNUM;
-               sz = XFS_BMDR_SPACE_CALC(MINDBTPTRS);
-       } else {
-               maxleafents = MAXAEXTNUM;
-               sz = XFS_BMDR_SPACE_CALC(MINABTPTRS);
-       }
-       maxrootrecs = xfs_bmdr_maxrecs(mp, sz, 0);
-       minleafrecs = mp->m_bmap_dmnr[0];
-       minnoderecs = mp->m_bmap_dmnr[1];
-       maxblocks = (maxleafents + minleafrecs - 1) / minleafrecs;
-       for (level = 1; maxblocks > 1; level++) {
-               if (maxblocks <= maxrootrecs)
-                       maxblocks = 1;
-               else
-                       maxblocks = (maxblocks + minnoderecs - 1) / minnoderecs;
+       if (*idx > 0) {
+               state |= BMAP_LEFT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx - 1), &left);
+
+               if (isnullstartblock(left.br_startblock))
+                       state |= BMAP_LEFT_DELAY;
        }
-       mp->m_bm_maxlevels[whichfork] = level;
-}
 
-/*
- * Routine to be called at transaction's end by xfs_bmapi, xfs_bunmapi
- * caller.  Frees all the extents that need freeing, which must be done
- * last due to locking considerations.  We never free any extents in
- * the first transaction.
- *
- * Return 1 if the given transaction was committed and a new one
- * started, and 0 otherwise in the committed parameter.
- */
-int                                            /* error */
-xfs_bmap_finish(
-       xfs_trans_t             **tp,           /* transaction pointer addr */
-       xfs_bmap_free_t         *flist,         /* i/o: list extents to free */
-       int                     *committed)     /* xact committed or not */
-{
-       xfs_efd_log_item_t      *efd;           /* extent free data */
-       xfs_efi_log_item_t      *efi;           /* extent free intention */
-       int                     error;          /* error return value */
-       xfs_bmap_free_item_t    *free;          /* free extent item */
-       unsigned int            logres;         /* new log reservation */
-       unsigned int            logcount;       /* new log count */
-       xfs_mount_t             *mp;            /* filesystem mount structure */
-       xfs_bmap_free_item_t    *next;          /* next item on free list */
-       xfs_trans_t             *ntp;           /* new transaction pointer */
+       /*
+        * Check and set flags if the current (right) segment exists.
+        * If it doesn't exist, we're converting the hole at end-of-file.
+        */
+       if (*idx < ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t)) {
+               state |= BMAP_RIGHT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *idx), &right);
 
-       ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
-       if (flist->xbf_count == 0) {
-               *committed = 0;
-               return 0;
+               if (isnullstartblock(right.br_startblock))
+                       state |= BMAP_RIGHT_DELAY;
        }
-       ntp = *tp;
-       efi = xfs_trans_get_efi(ntp, flist->xbf_count);
-       for (free = flist->xbf_first; free; free = free->xbfi_next)
-               xfs_trans_log_efi_extent(ntp, efi, free->xbfi_startblock,
-                       free->xbfi_blockcount);
-       logres = ntp->t_log_res;
-       logcount = ntp->t_log_count;
-       ntp = xfs_trans_dup(*tp);
-       error = xfs_trans_commit(*tp, 0);
-       *tp = ntp;
-       *committed = 1;
+
        /*
-        * We have a new transaction, so we should return committed=1,
-        * even though we're returning an error.
+        * Set contiguity flags on the left and right neighbors.
+        * Don't let extents get too large, even if the pieces are contiguous.
         */
-       if (error)
-               return error;
+       if ((state & BMAP_LEFT_VALID) && (state & BMAP_LEFT_DELAY) &&
+           left.br_startoff + left.br_blockcount == new->br_startoff &&
+           left.br_blockcount + new->br_blockcount <= MAXEXTLEN)
+               state |= BMAP_LEFT_CONTIG;
+
+       if ((state & BMAP_RIGHT_VALID) && (state & BMAP_RIGHT_DELAY) &&
+           new->br_startoff + new->br_blockcount == right.br_startoff &&
+           new->br_blockcount + right.br_blockcount <= MAXEXTLEN &&
+           (!(state & BMAP_LEFT_CONTIG) ||
+            (left.br_blockcount + new->br_blockcount +
+             right.br_blockcount <= MAXEXTLEN)))
+               state |= BMAP_RIGHT_CONTIG;
 
        /*
-        * transaction commit worked ok so we can drop the extra ticket
-        * reference that we gained in xfs_trans_dup()
+        * Switch out based on the contiguity flags.
         */
-       xfs_log_ticket_put(ntp->t_ticket);
+       switch (state & (BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG)) {
+       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+               /*
+                * New allocation is contiguous with delayed allocations
+                * on the left and on the right.
+                * Merge all three into a single extent record.
+                */
+               --*idx;
+               temp = left.br_blockcount + new->br_blockcount +
+                       right.br_blockcount;
 
-       if ((error = xfs_trans_reserve(ntp, 0, logres, 0, XFS_TRANS_PERM_LOG_RES,
-                       logcount)))
-               return error;
-       efd = xfs_trans_get_efd(ntp, efi, flist->xbf_count);
-       for (free = flist->xbf_first; free != NULL; free = next) {
-               next = free->xbfi_next;
-               if ((error = xfs_free_extent(ntp, free->xbfi_startblock,
-                               free->xbfi_blockcount))) {
-                       /*
-                        * The bmap free list will be cleaned up at a
-                        * higher level.  The EFI will be canceled when
-                        * this transaction is aborted.
-                        * Need to force shutdown here to make sure it
-                        * happens, since this transaction may not be
-                        * dirty yet.
-                        */
-                       mp = ntp->t_mountp;
-                       if (!XFS_FORCED_SHUTDOWN(mp))
-                               xfs_force_shutdown(mp,
-                                                  (error == EFSCORRUPTED) ?
-                                                  SHUTDOWN_CORRUPT_INCORE :
-                                                  SHUTDOWN_META_IO_ERROR);
-                       return error;
-               }
-               xfs_trans_log_efd_extent(ntp, efd, free->xbfi_startblock,
-                       free->xbfi_blockcount);
-               xfs_bmap_del_free(flist, NULL, free);
-       }
-       return 0;
-}
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx), temp);
+               oldlen = startblockval(left.br_startblock) +
+                       startblockval(new->br_startblock) +
+                       startblockval(right.br_startblock);
+               newlen = xfs_bmap_worst_indlen(ip, temp);
+               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, *idx),
+                       nullstartblock((int)newlen));
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
 
-/*
- * Free up any items left in the list.
- */
-void
-xfs_bmap_cancel(
-       xfs_bmap_free_t         *flist) /* list of bmap_free_items */
-{
-       xfs_bmap_free_item_t    *free;  /* free list item */
-       xfs_bmap_free_item_t    *next;
+               xfs_iext_remove(ip, *idx + 1, 1, state);
+               break;
 
-       if (flist->xbf_count == 0)
-               return;
-       ASSERT(flist->xbf_first != NULL);
-       for (free = flist->xbf_first; free; free = next) {
-               next = free->xbfi_next;
-               xfs_bmap_del_free(flist, NULL, free);
-       }
-       ASSERT(flist->xbf_count == 0);
-}
+       case BMAP_LEFT_CONTIG:
+               /*
+                * New allocation is contiguous with a delayed allocation
+                * on the left.
+                * Merge the new allocation with the left neighbor.
+                */
+               --*idx;
+               temp = left.br_blockcount + new->br_blockcount;
 
-/*
- * Returns the file-relative block number of the first unused block(s)
- * in the file with at least "len" logically contiguous blocks free.
- * This is the lowest-address hole if the file has holes, else the first block
- * past the end of file.
- * Return 0 if the file is currently local (in-inode).
- */
-int                                            /* error */
-xfs_bmap_first_unused(
-       xfs_trans_t     *tp,                    /* transaction pointer */
-       xfs_inode_t     *ip,                    /* incore inode */
-       xfs_extlen_t    len,                    /* size of hole to find */
-       xfs_fileoff_t   *first_unused,          /* unused block */
-       int             whichfork)              /* data or attr fork */
-{
-       int             error;                  /* error return value */
-       int             idx;                    /* extent record index */
-       xfs_ifork_t     *ifp;                   /* inode fork pointer */
-       xfs_fileoff_t   lastaddr;               /* last block number seen */
-       xfs_fileoff_t   lowest;                 /* lowest useful block */
-       xfs_fileoff_t   max;                    /* starting useful block */
-       xfs_fileoff_t   off;                    /* offset for this block */
-       xfs_extnum_t    nextents;               /* number of extent entries */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, *idx), temp);
+               oldlen = startblockval(left.br_startblock) +
+                       startblockval(new->br_startblock);
+               newlen = xfs_bmap_worst_indlen(ip, temp);
+               xfs_bmbt_set_startblock(xfs_iext_get_ext(ifp, *idx),
+                       nullstartblock((int)newlen));
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               break;
 
-       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE ||
-              XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS ||
-              XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL);
-       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
-               *first_unused = 0;
-               return 0;
+       case BMAP_RIGHT_CONTIG:
+               /*
+                * New allocation is contiguous with a delayed allocation
+                * on the right.
+                * Merge the new allocation with the right neighbor.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               temp = new->br_blockcount + right.br_blockcount;
+               oldlen = startblockval(new->br_startblock) +
+                       startblockval(right.br_startblock);
+               newlen = xfs_bmap_worst_indlen(ip, temp);
+               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, *idx),
+                       new->br_startoff,
+                       nullstartblock((int)newlen), temp, right.br_state);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               break;
+
+       case 0:
+               /*
+                * New allocation is not contiguous with another
+                * delayed allocation.
+                * Insert a new entry.
+                */
+               oldlen = newlen = 0;
+               xfs_iext_insert(ip, *idx, 1, new, state);
+               break;
        }
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
-           (error = xfs_iread_extents(tp, ip, whichfork)))
-               return error;
-       lowest = *first_unused;
-       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
-       for (idx = 0, lastaddr = 0, max = lowest; idx < nextents; idx++) {
-               xfs_bmbt_rec_host_t *ep = xfs_iext_get_ext(ifp, idx);
-               off = xfs_bmbt_get_startoff(ep);
+       if (oldlen != newlen) {
+               ASSERT(oldlen > newlen);
+               xfs_icsb_modify_counters(ip->i_mount, XFS_SBS_FDBLOCKS,
+                       (int64_t)(oldlen - newlen), 0);
                /*
-                * See if the hole before this extent will work.
+                * Nothing to do for disk quota accounting here.
                 */
-               if (off >= lowest + len && off - max >= len) {
-                       *first_unused = max;
-                       return 0;
-               }
-               lastaddr = off + xfs_bmbt_get_blockcount(ep);
-               max = XFS_FILEOFF_MAX(lastaddr, lowest);
        }
-       *first_unused = max;
-       return 0;
 }
 
 /*
- * Returns the file-relative block number of the last block + 1 before
- * last_block (input value) in the file.
- * This is not based on i_size, it is based on the extent records.
- * Returns 0 for local files, as they do not have extent records.
+ * Convert a hole to a real allocation.
  */
-int                                            /* error */
-xfs_bmap_last_before(
-       xfs_trans_t     *tp,                    /* transaction pointer */
-       xfs_inode_t     *ip,                    /* incore inode */
-       xfs_fileoff_t   *last_block,            /* last block */
-       int             whichfork)              /* data or attr fork */
-{
-       xfs_fileoff_t   bno;                    /* input file offset */
-       int             eof;                    /* hit end of file */
-       xfs_bmbt_rec_host_t *ep;                /* pointer to last extent */
-       int             error;                  /* error return value */
-       xfs_bmbt_irec_t got;                    /* current extent value */
-       xfs_ifork_t     *ifp;                   /* inode fork pointer */
-       xfs_extnum_t    lastx;                  /* last extent used */
-       xfs_bmbt_irec_t prev;                   /* previous extent value */
-
-       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
-              return XFS_ERROR(EIO);
-       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
-               *last_block = 0;
-               return 0;
-       }
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
-           (error = xfs_iread_extents(tp, ip, whichfork)))
-               return error;
-       bno = *last_block - 1;
-       ep = xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got,
-               &prev);
-       if (eof || xfs_bmbt_get_startoff(ep) > bno) {
-               if (prev.br_startoff == NULLFILEOFF)
-                       *last_block = 0;
-               else
-                       *last_block = prev.br_startoff + prev.br_blockcount;
-       }
-       /*
-        * Otherwise *last_block is already the right answer.
-        */
-       return 0;
-}
-
-STATIC int
-xfs_bmap_last_extent(
-       struct xfs_trans        *tp,
-       struct xfs_inode        *ip,
-       int                     whichfork,
-       struct xfs_bmbt_irec    *rec,
-       int                     *is_empty)
+STATIC int                             /* error */
+xfs_bmap_add_extent_hole_real(
+       struct xfs_bmalloca     *bma,
+       int                     whichfork)
 {
-       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, whichfork);
-       int                     error;
-       int                     nextents;
+       struct xfs_bmbt_irec    *new = &bma->got;
+       int                     error;  /* error return value */
+       int                     i;      /* temp state */
+       xfs_ifork_t             *ifp;   /* inode fork pointer */
+       xfs_bmbt_irec_t         left;   /* left neighbor extent entry */
+       xfs_bmbt_irec_t         right;  /* right neighbor extent entry */
+       int                     rval=0; /* return value (logging flags) */
+       int                     state;  /* state bits, accessed thru macros */
 
-       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
-               error = xfs_iread_extents(tp, ip, whichfork);
-               if (error)
-                       return error;
-       }
+       ifp = XFS_IFORK_PTR(bma->ip, whichfork);
 
-       nextents = ifp->if_bytes / sizeof(xfs_bmbt_rec_t);
-       if (nextents == 0) {
-               *is_empty = 1;
-               return 0;
-       }
+       ASSERT(bma->idx >= 0);
+       ASSERT(bma->idx <= ifp->if_bytes / sizeof(struct xfs_bmbt_rec));
+       ASSERT(!isnullstartblock(new->br_startblock));
+       ASSERT(!bma->cur ||
+              !(bma->cur->bc_private.b.flags & XFS_BTCUR_BPRV_WASDEL));
 
-       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, nextents - 1), rec);
-       *is_empty = 0;
-       return 0;
-}
+       XFS_STATS_INC(xs_add_exlist);
 
-/*
- * Check the last inode extent to determine whether this allocation will result
- * in blocks being allocated at the end of the file. When we allocate new data
- * blocks at the end of the file which do not start at the previous data block,
- * we will try to align the new blocks at stripe unit boundaries.
- *
- * Returns 0 in bma->aeof if the file (fork) is empty as any new write will be
- * at, or past the EOF.
- */
-STATIC int
-xfs_bmap_isaeof(
-       struct xfs_bmalloca     *bma,
-       int                     whichfork)
-{
-       struct xfs_bmbt_irec    rec;
-       int                     is_empty;
-       int                     error;
+       state = 0;
+       if (whichfork == XFS_ATTR_FORK)
+               state |= BMAP_ATTRFORK;
 
-       bma->aeof = 0;
-       error = xfs_bmap_last_extent(NULL, bma->ip, whichfork, &rec,
-                                    &is_empty);
-       if (error || is_empty)
-               return error;
+       /*
+        * Check and set flags if this segment has a left neighbor.
+        */
+       if (bma->idx > 0) {
+               state |= BMAP_LEFT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1), &left);
+               if (isnullstartblock(left.br_startblock))
+                       state |= BMAP_LEFT_DELAY;
+       }
 
        /*
-        * Check if we are allocation or past the last extent, or at least into
-        * the last delayed allocated extent.
+        * Check and set flags if this segment has a current value.
+        * Not true if we're inserting into the "hole" at eof.
         */
-       bma->aeof = bma->offset >= rec.br_startoff + rec.br_blockcount ||
-               (bma->offset >= rec.br_startoff &&
-                isnullstartblock(rec.br_startblock));
-       return 0;
-}
+       if (bma->idx < ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)) {
+               state |= BMAP_RIGHT_VALID;
+               xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &right);
+               if (isnullstartblock(right.br_startblock))
+                       state |= BMAP_RIGHT_DELAY;
+       }
 
-/*
- * Check if the endoff is outside the last extent. If so the caller will grow
- * the allocation to a stripe unit boundary.  All offsets are considered outside
- * the end of file for an empty fork, so 1 is returned in *eof in that case.
- */
-int
-xfs_bmap_eof(
-       struct xfs_inode        *ip,
-       xfs_fileoff_t           endoff,
-       int                     whichfork,
-       int                     *eof)
-{
-       struct xfs_bmbt_irec    rec;
-       int                     error;
+       /*
+        * We're inserting a real allocation between "left" and "right".
+        * Set the contiguity flags.  Don't let extents get too large.
+        */
+       if ((state & BMAP_LEFT_VALID) && !(state & BMAP_LEFT_DELAY) &&
+           left.br_startoff + left.br_blockcount == new->br_startoff &&
+           left.br_startblock + left.br_blockcount == new->br_startblock &&
+           left.br_state == new->br_state &&
+           left.br_blockcount + new->br_blockcount <= MAXEXTLEN)
+               state |= BMAP_LEFT_CONTIG;
 
-       error = xfs_bmap_last_extent(NULL, ip, whichfork, &rec, eof);
-       if (error || *eof)
-               return error;
+       if ((state & BMAP_RIGHT_VALID) && !(state & BMAP_RIGHT_DELAY) &&
+           new->br_startoff + new->br_blockcount == right.br_startoff &&
+           new->br_startblock + new->br_blockcount == right.br_startblock &&
+           new->br_state == right.br_state &&
+           new->br_blockcount + right.br_blockcount <= MAXEXTLEN &&
+           (!(state & BMAP_LEFT_CONTIG) ||
+            left.br_blockcount + new->br_blockcount +
+            right.br_blockcount <= MAXEXTLEN))
+               state |= BMAP_RIGHT_CONTIG;
 
-       *eof = endoff >= rec.br_startoff + rec.br_blockcount;
-       return 0;
-}
+       error = 0;
+       /*
+        * Select which case we're in here, and implement it.
+        */
+       switch (state & (BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG)) {
+       case BMAP_LEFT_CONTIG | BMAP_RIGHT_CONTIG:
+               /*
+                * New allocation is contiguous with real allocations on the
+                * left and on the right.
+                * Merge all three into a single extent record.
+                */
+               --bma->idx;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
+                       left.br_blockcount + new->br_blockcount +
+                       right.br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-/*
- * Returns the file-relative block number of the first block past eof in
- * the file.  This is not based on i_size, it is based on the extent records.
- * Returns 0 for local files, as they do not have extent records.
- */
-int
-xfs_bmap_last_offset(
-       struct xfs_trans        *tp,
-       struct xfs_inode        *ip,
-       xfs_fileoff_t           *last_block,
-       int                     whichfork)
-{
-       struct xfs_bmbt_irec    rec;
-       int                     is_empty;
-       int                     error;
+               xfs_iext_remove(bma->ip, bma->idx + 1, 1, state);
 
-       *last_block = 0;
+               XFS_IFORK_NEXT_SET(bma->ip, whichfork,
+                       XFS_IFORK_NEXTENTS(bma->ip, whichfork) - 1);
+               if (bma->cur == NULL) {
+                       rval = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
+               } else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur, right.br_startoff,
+                                       right.br_startblock, right.br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_btree_delete(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_btree_decrement(bma->cur, 0, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, left.br_startoff,
+                                       left.br_startblock,
+                                       left.br_blockcount +
+                                               new->br_blockcount +
+                                               right.br_blockcount,
+                                       left.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
 
-       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL)
-               return 0;
+       case BMAP_LEFT_CONTIG:
+               /*
+                * New allocation is contiguous with a real allocation
+                * on the left.
+                * Merge the new allocation with the left neighbor.
+                */
+               --bma->idx;
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(xfs_iext_get_ext(ifp, bma->idx),
+                       left.br_blockcount + new->br_blockcount);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
-              return XFS_ERROR(EIO);
+               if (bma->cur == NULL) {
+                       rval = xfs_ilog_fext(whichfork);
+               } else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur, left.br_startoff,
+                                       left.br_startblock, left.br_blockcount,
+                                       &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, left.br_startoff,
+                                       left.br_startblock,
+                                       left.br_blockcount +
+                                               new->br_blockcount,
+                                       left.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
 
-       error = xfs_bmap_last_extent(NULL, ip, whichfork, &rec, &is_empty);
-       if (error || is_empty)
-               return error;
+       case BMAP_RIGHT_CONTIG:
+               /*
+                * New allocation is contiguous with a real allocation
+                * on the right.
+                * Merge the new allocation with the right neighbor.
+                */
+               trace_xfs_bmap_pre_update(bma->ip, bma->idx, state, _THIS_IP_);
+               xfs_bmbt_set_allf(xfs_iext_get_ext(ifp, bma->idx),
+                       new->br_startoff, new->br_startblock,
+                       new->br_blockcount + right.br_blockcount,
+                       right.br_state);
+               trace_xfs_bmap_post_update(bma->ip, bma->idx, state, _THIS_IP_);
 
-       *last_block = rec.br_startoff + rec.br_blockcount;
-       return 0;
+               if (bma->cur == NULL) {
+                       rval = xfs_ilog_fext(whichfork);
+               } else {
+                       rval = 0;
+                       error = xfs_bmbt_lookup_eq(bma->cur,
+                                       right.br_startoff,
+                                       right.br_startblock,
+                                       right.br_blockcount, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       error = xfs_bmbt_update(bma->cur, new->br_startoff,
+                                       new->br_startblock,
+                                       new->br_blockcount +
+                                               right.br_blockcount,
+                                       right.br_state);
+                       if (error)
+                               goto done;
+               }
+               break;
+
+       case 0:
+               /*
+                * New allocation is not contiguous with another
+                * real allocation.
+                * Insert a new entry.
+                */
+               xfs_iext_insert(bma->ip, bma->idx, 1, new, state);
+               XFS_IFORK_NEXT_SET(bma->ip, whichfork,
+                       XFS_IFORK_NEXTENTS(bma->ip, whichfork) + 1);
+               if (bma->cur == NULL) {
+                       rval = XFS_ILOG_CORE | xfs_ilog_fext(whichfork);
+               } else {
+                       rval = XFS_ILOG_CORE;
+                       error = xfs_bmbt_lookup_eq(bma->cur,
+                                       new->br_startoff,
+                                       new->br_startblock,
+                                       new->br_blockcount, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 0, done);
+                       bma->cur->bc_rec.b.br_state = new->br_state;
+                       error = xfs_btree_insert(bma->cur, &i);
+                       if (error)
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               }
+               break;
+       }
+
+       /* convert to a btree if necessary */
+       if (xfs_bmap_needs_btree(bma->ip, whichfork)) {
+               int     tmp_logflags;   /* partial log flag return val */
+
+               ASSERT(bma->cur == NULL);
+               error = xfs_bmap_extents_to_btree(bma->tp, bma->ip,
+                               bma->firstblock, bma->flist, &bma->cur,
+                               0, &tmp_logflags, whichfork);
+               bma->logflags |= tmp_logflags;
+               if (error)
+                       goto done;
+       }
+
+       /* clear out the allocated field, done with it now in any case. */
+       if (bma->cur)
+               bma->cur->bc_private.b.allocated = 0;
+
+       xfs_bmap_check_leaf_extents(bma->cur, bma->ip, whichfork);
+done:
+       bma->logflags |= rval;
+       return error;
 }
 
 /*
- * Returns whether the selected fork of the inode has exactly one
- * block or not.  For the data fork we check this matches di_size,
- * implying the file's range is 0..bsize-1.
+ * Functions used in the extent read, allocate and remove paths
  */
-int                                    /* 1=>1 block, 0=>otherwise */
-xfs_bmap_one_block(
-       xfs_inode_t     *ip,            /* incore inode */
-       int             whichfork)      /* data or attr fork */
-{
-       xfs_bmbt_rec_host_t *ep;        /* ptr to fork's extent */
-       xfs_ifork_t     *ifp;           /* inode fork pointer */
-       int             rval;           /* return value */
-       xfs_bmbt_irec_t s;              /* internal version of extent */
-
-#ifndef DEBUG
-       if (whichfork == XFS_DATA_FORK)
-               return XFS_ISIZE(ip) == ip->i_mount->m_sb.sb_blocksize;
-#endif /* !DEBUG */
-       if (XFS_IFORK_NEXTENTS(ip, whichfork) != 1)
-               return 0;
-       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
-               return 0;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT(ifp->if_flags & XFS_IFEXTENTS);
-       ep = xfs_iext_get_ext(ifp, 0);
-       xfs_bmbt_get_all(ep, &s);
-       rval = s.br_startoff == 0 && s.br_blockcount == 1;
-       if (rval && whichfork == XFS_DATA_FORK)
-               ASSERT(XFS_ISIZE(ip) == ip->i_mount->m_sb.sb_blocksize);
-       return rval;
-}
 
+/*
+ * Adjust the size of the new extent based on di_extsize and rt extsize.
+ */
 STATIC int
-xfs_bmap_sanity_check(
-       struct xfs_mount        *mp,
-       struct xfs_buf          *bp,
-       int                     level)
+xfs_bmap_extsize_align(
+       xfs_mount_t     *mp,
+       xfs_bmbt_irec_t *gotp,          /* next extent pointer */
+       xfs_bmbt_irec_t *prevp,         /* previous extent pointer */
+       xfs_extlen_t    extsz,          /* align to this extent size */
+       int             rt,             /* is this a realtime inode? */
+       int             eof,            /* is extent at end-of-file? */
+       int             delay,          /* creating delalloc extent? */
+       int             convert,        /* overwriting unwritten extent? */
+       xfs_fileoff_t   *offp,          /* in/out: aligned offset */
+       xfs_extlen_t    *lenp)          /* in/out: aligned length */
 {
-       struct xfs_btree_block  *block = XFS_BUF_TO_BLOCK(bp);
+       xfs_fileoff_t   orig_off;       /* original offset */
+       xfs_extlen_t    orig_alen;      /* original length */
+       xfs_fileoff_t   orig_end;       /* original off+len */
+       xfs_fileoff_t   nexto;          /* next file offset */
+       xfs_fileoff_t   prevo;          /* previous file offset */
+       xfs_fileoff_t   align_off;      /* temp for offset */
+       xfs_extlen_t    align_alen;     /* temp for length */
+       xfs_extlen_t    temp;           /* temp for calculations */
 
-       if (block->bb_magic != cpu_to_be32(XFS_BMAP_MAGIC) ||
-           be16_to_cpu(block->bb_level) != level ||
-           be16_to_cpu(block->bb_numrecs) == 0 ||
-           be16_to_cpu(block->bb_numrecs) > mp->m_bmap_dmxr[level != 0])
+       if (convert)
                return 0;
-       return 1;
-}
 
-/*
- * Read in the extents to if_extents.
- * All inode fields are set up by caller, we just traverse the btree
- * and copy the records in. If the file system cannot contain unwritten
- * extents, the records are checked for no "state" flags.
- */
-int                                    /* error */
-xfs_bmap_read_extents(
-       xfs_trans_t             *tp,    /* transaction pointer */
-       xfs_inode_t             *ip,    /* incore inode */
-       int                     whichfork) /* data or attr fork */
-{
-       struct xfs_btree_block  *block; /* current btree block */
-       xfs_fsblock_t           bno;    /* block # of "block" */
-       xfs_buf_t               *bp;    /* buffer for "block" */
-       int                     error;  /* error return value */
-       xfs_exntfmt_t           exntf;  /* XFS_EXTFMT_NOSTATE, if checking */
-       xfs_extnum_t            i, j;   /* index into the extents list */
-       xfs_ifork_t             *ifp;   /* fork structure */
-       int                     level;  /* btree level, for checking */
-       xfs_mount_t             *mp;    /* file system mount structure */
-       __be64                  *pp;    /* pointer to block address */
-       /* REFERENCED */
-       xfs_extnum_t            room;   /* number of entries there's room for */
+       orig_off = align_off = *offp;
+       orig_alen = align_alen = *lenp;
+       orig_end = orig_off + orig_alen;
 
-       bno = NULLFSBLOCK;
-       mp = ip->i_mount;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       exntf = (whichfork != XFS_DATA_FORK) ? XFS_EXTFMT_NOSTATE :
-                                       XFS_EXTFMT_INODE(ip);
-       block = ifp->if_broot;
        /*
-        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
+        * If this request overlaps an existing extent, then don't
+        * attempt to perform any additional alignment.
         */
-       level = be16_to_cpu(block->bb_level);
-       ASSERT(level > 0);
-       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
-       bno = be64_to_cpu(*pp);
-       ASSERT(bno != NULLDFSBNO);
-       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
-       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
+       if (!delay && !eof &&
+           (orig_off >= gotp->br_startoff) &&
+           (orig_end <= gotp->br_startoff + gotp->br_blockcount)) {
+               return 0;
+       }
+
        /*
-        * Go down the tree until leaf level is reached, following the first
-        * pointer (leftmost) at each level.
+        * If the file offset is unaligned vs. the extent size
+        * we need to align it.  This will be possible unless
+        * the file was previously written with a kernel that didn't
+        * perform this alignment, or if a truncate shot us in the
+        * foot.
         */
-       while (level-- > 0) {
-               error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
-                               XFS_BMAP_BTREE_REF, &xfs_bmbt_buf_ops);
-               if (error)
-                       return error;
-               block = XFS_BUF_TO_BLOCK(bp);
-               XFS_WANT_CORRUPTED_GOTO(
-                       xfs_bmap_sanity_check(mp, bp, level),
-                       error0);
-               if (level == 0)
-                       break;
-               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
-               bno = be64_to_cpu(*pp);
-               XFS_WANT_CORRUPTED_GOTO(XFS_FSB_SANITY_CHECK(mp, bno), error0);
-               xfs_trans_brelse(tp, bp);
+       temp = do_mod(orig_off, extsz);
+       if (temp) {
+               align_alen += temp;
+               align_off -= temp;
        }
        /*
-        * Here with bp and block set to the leftmost leaf node in the tree.
+        * Same adjustment for the end of the requested area.
         */
-       room = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
-       i = 0;
+       if ((temp = (align_alen % extsz))) {
+               align_alen += extsz - temp;
+       }
        /*
-        * Loop over all leaf nodes.  Copy information to the extent records.
+        * If the previous block overlaps with this proposed allocation
+        * then move the start forward without adjusting the length.
         */
-       for (;;) {
-               xfs_bmbt_rec_t  *frp;
-               xfs_fsblock_t   nextbno;
-               xfs_extnum_t    num_recs;
-               xfs_extnum_t    start;
+       if (prevp->br_startoff != NULLFILEOFF) {
+               if (prevp->br_startblock == HOLESTARTBLOCK)
+                       prevo = prevp->br_startoff;
+               else
+                       prevo = prevp->br_startoff + prevp->br_blockcount;
+       } else
+               prevo = 0;
+       if (align_off != orig_off && align_off < prevo)
+               align_off = prevo;
+       /*
+        * If the next block overlaps with this proposed allocation
+        * then move the start back without adjusting the length,
+        * but not before offset 0.
+        * This may of course make the start overlap previous block,
+        * and if we hit the offset 0 limit then the next block
+        * can still overlap too.
+        */
+       if (!eof && gotp->br_startoff != NULLFILEOFF) {
+               if ((delay && gotp->br_startblock == HOLESTARTBLOCK) ||
+                   (!delay && gotp->br_startblock == DELAYSTARTBLOCK))
+                       nexto = gotp->br_startoff + gotp->br_blockcount;
+               else
+                       nexto = gotp->br_startoff;
+       } else
+               nexto = NULLFILEOFF;
+       if (!eof &&
+           align_off + align_alen != orig_end &&
+           align_off + align_alen > nexto)
+               align_off = nexto > align_alen ? nexto - align_alen : 0;
+       /*
+        * If we're now overlapping the next or previous extent that
+        * means we can't fit an extsz piece in this hole.  Just move
+        * the start forward to the first valid spot and set
+        * the length so we hit the end.
+        */
+       if (align_off != orig_off && align_off < prevo)
+               align_off = prevo;
+       if (align_off + align_alen != orig_end &&
+           align_off + align_alen > nexto &&
+           nexto != NULLFILEOFF) {
+               ASSERT(nexto > prevo);
+               align_alen = nexto - align_off;
+       }
 
-               num_recs = xfs_btree_get_numrecs(block);
-               if (unlikely(i + num_recs > room)) {
-                       ASSERT(i + num_recs <= room);
-                       xfs_warn(ip->i_mount,
-                               "corrupt dinode %Lu, (btree extents).",
-                               (unsigned long long) ip->i_ino);
-                       XFS_CORRUPTION_ERROR("xfs_bmap_read_extents(1)",
-                               XFS_ERRLEVEL_LOW, ip->i_mount, block);
-                       goto error0;
-               }
-               XFS_WANT_CORRUPTED_GOTO(
-                       xfs_bmap_sanity_check(mp, bp, 0),
-                       error0);
+       /*
+        * If realtime, and the result isn't a multiple of the realtime
+        * extent size we need to remove blocks until it is.
+        */
+       if (rt && (temp = (align_alen % mp->m_sb.sb_rextsize))) {
                /*
-                * Read-ahead the next leaf block, if any.
+                * We're not covering the original request, or
+                * we won't be able to once we fix the length.
                 */
-               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
-               if (nextbno != NULLFSBLOCK)
-                       xfs_btree_reada_bufl(mp, nextbno, 1,
-                                            &xfs_bmbt_buf_ops);
+               if (orig_off < align_off ||
+                   orig_end > align_off + align_alen ||
+                   align_alen - temp < orig_alen)
+                       return XFS_ERROR(EINVAL);
                /*
-                * Copy records into the extent records.
+                * Try to fix it by moving the start up.
                 */
-               frp = XFS_BMBT_REC_ADDR(mp, block, 1);
-               start = i;
-               for (j = 0; j < num_recs; j++, i++, frp++) {
-                       xfs_bmbt_rec_host_t *trp = xfs_iext_get_ext(ifp, i);
-                       trp->l0 = be64_to_cpu(frp->l0);
-                       trp->l1 = be64_to_cpu(frp->l1);
+               if (align_off + temp <= orig_off) {
+                       align_alen -= temp;
+                       align_off += temp;
                }
-               if (exntf == XFS_EXTFMT_NOSTATE) {
-                       /*
-                        * Check all attribute bmap btree records and
-                        * any "older" data bmap btree records for a
-                        * set bit in the "extent flag" position.
-                        */
-                       if (unlikely(xfs_check_nostate_extents(ifp,
-                                       start, num_recs))) {
-                               XFS_ERROR_REPORT("xfs_bmap_read_extents(2)",
-                                                XFS_ERRLEVEL_LOW,
-                                                ip->i_mount);
-                               goto error0;
-                       }
+               /*
+                * Try to fix it by moving the end in.
+                */
+               else if (align_off + align_alen - temp >= orig_end)
+                       align_alen -= temp;
+               /*
+                * Set the start to the minimum then trim the length.
+                */
+               else {
+                       align_alen -= orig_off - align_off;
+                       align_off = orig_off;
+                       align_alen -= align_alen % mp->m_sb.sb_rextsize;
                }
-               xfs_trans_brelse(tp, bp);
-               bno = nextbno;
                /*
-                * If we've reached the end, stop.
+                * Result doesn't cover the request, fail it.
                 */
-               if (bno == NULLFSBLOCK)
-                       break;
-               error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
-                               XFS_BMAP_BTREE_REF, &xfs_bmbt_buf_ops);
-               if (error)
-                       return error;
-               block = XFS_BUF_TO_BLOCK(bp);
+               if (orig_off < align_off || orig_end > align_off + align_alen)
+                       return XFS_ERROR(EINVAL);
+       } else {
+               ASSERT(orig_off >= align_off);
+               ASSERT(orig_end <= align_off + align_alen);
        }
-       ASSERT(i == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)));
-       ASSERT(i == XFS_IFORK_NEXTENTS(ip, whichfork));
-       XFS_BMAP_TRACE_EXLIST(ip, i, whichfork);
-       return 0;
-error0:
-       xfs_trans_brelse(tp, bp);
-       return XFS_ERROR(EFSCORRUPTED);
-}
 
 #ifdef DEBUG
-/*
- * Add bmap trace insert entries for all the contents of the extent records.
- */
-void
-xfs_bmap_trace_exlist(
-       xfs_inode_t     *ip,            /* incore inode pointer */
-       xfs_extnum_t    cnt,            /* count of entries in the list */
-       int             whichfork,      /* data or attr fork */
-       unsigned long   caller_ip)
-{
-       xfs_extnum_t    idx;            /* extent record index */
-       xfs_ifork_t     *ifp;           /* inode fork pointer */
-       int             state = 0;
-
-       if (whichfork == XFS_ATTR_FORK)
-               state |= BMAP_ATTRFORK;
+       if (!eof && gotp->br_startoff != NULLFILEOFF)
+               ASSERT(align_off + align_alen <= gotp->br_startoff);
+       if (prevp->br_startoff != NULLFILEOFF)
+               ASSERT(align_off >= prevp->br_startoff + prevp->br_blockcount);
+#endif
 
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       ASSERT(cnt == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)));
-       for (idx = 0; idx < cnt; idx++)
-               trace_xfs_extlist(ip, idx, whichfork, caller_ip);
+       *lenp = align_alen;
+       *offp = align_off;
+       return 0;
 }
 
-/*
- * Validate that the bmbt_irecs being returned from bmapi are valid
- * given the callers original parameters.  Specifically check the
- * ranges of the returned irecs to ensure that they only extent beyond
- * the given parameters if the XFS_BMAPI_ENTIRE flag was set.
- */
+#define XFS_ALLOC_GAP_UNITS    4
+
 STATIC void
-xfs_bmap_validate_ret(
-       xfs_fileoff_t           bno,
-       xfs_filblks_t           len,
-       int                     flags,
-       xfs_bmbt_irec_t         *mval,
-       int                     nmap,
-       int                     ret_nmap)
+xfs_bmap_adjacent(
+       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
 {
-       int                     i;              /* index to map values */
+       xfs_fsblock_t   adjust;         /* adjustment to block numbers */
+       xfs_agnumber_t  fb_agno;        /* ag number of ap->firstblock */
+       xfs_mount_t     *mp;            /* mount point structure */
+       int             nullfb;         /* true if ap->firstblock isn't set */
+       int             rt;             /* true if inode is realtime */
 
-       ASSERT(ret_nmap <= nmap);
+#define        ISVALID(x,y)    \
+       (rt ? \
+               (x) < mp->m_sb.sb_rblocks : \
+               XFS_FSB_TO_AGNO(mp, x) == XFS_FSB_TO_AGNO(mp, y) && \
+               XFS_FSB_TO_AGNO(mp, x) < mp->m_sb.sb_agcount && \
+               XFS_FSB_TO_AGBNO(mp, x) < mp->m_sb.sb_agblocks)
 
-       for (i = 0; i < ret_nmap; i++) {
-               ASSERT(mval[i].br_blockcount > 0);
-               if (!(flags & XFS_BMAPI_ENTIRE)) {
-                       ASSERT(mval[i].br_startoff >= bno);
-                       ASSERT(mval[i].br_blockcount <= len);
-                       ASSERT(mval[i].br_startoff + mval[i].br_blockcount <=
-                              bno + len);
-               } else {
-                       ASSERT(mval[i].br_startoff < bno + len);
-                       ASSERT(mval[i].br_startoff + mval[i].br_blockcount >
-                              bno);
+       mp = ap->ip->i_mount;
+       nullfb = *ap->firstblock == NULLFSBLOCK;
+       rt = XFS_IS_REALTIME_INODE(ap->ip) && ap->userdata;
+       fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp, *ap->firstblock);
+       /*
+        * If allocating at eof, and there's a previous real block,
+        * try to use its last block as our starting point.
+        */
+       if (ap->eof && ap->prev.br_startoff != NULLFILEOFF &&
+           !isnullstartblock(ap->prev.br_startblock) &&
+           ISVALID(ap->prev.br_startblock + ap->prev.br_blockcount,
+                   ap->prev.br_startblock)) {
+               ap->blkno = ap->prev.br_startblock + ap->prev.br_blockcount;
+               /*
+                * Adjust for the gap between prevp and us.
+                */
+               adjust = ap->offset -
+                       (ap->prev.br_startoff + ap->prev.br_blockcount);
+               if (adjust &&
+                   ISVALID(ap->blkno + adjust, ap->prev.br_startblock))
+                       ap->blkno += adjust;
+       }
+       /*
+        * If not at eof, then compare the two neighbor blocks.
+        * Figure out whether either one gives us a good starting point,
+        * and pick the better one.
+        */
+       else if (!ap->eof) {
+               xfs_fsblock_t   gotbno;         /* right side block number */
+               xfs_fsblock_t   gotdiff=0;      /* right side difference */
+               xfs_fsblock_t   prevbno;        /* left side block number */
+               xfs_fsblock_t   prevdiff=0;     /* left side difference */
+
+               /*
+                * If there's a previous (left) block, select a requested
+                * start block based on it.
+                */
+               if (ap->prev.br_startoff != NULLFILEOFF &&
+                   !isnullstartblock(ap->prev.br_startblock) &&
+                   (prevbno = ap->prev.br_startblock +
+                              ap->prev.br_blockcount) &&
+                   ISVALID(prevbno, ap->prev.br_startblock)) {
+                       /*
+                        * Calculate gap to end of previous block.
+                        */
+                       adjust = prevdiff = ap->offset -
+                               (ap->prev.br_startoff +
+                                ap->prev.br_blockcount);
+                       /*
+                        * Figure the startblock based on the previous block's
+                        * end and the gap size.
+                        * Heuristic!
+                        * If the gap is large relative to the piece we're
+                        * allocating, or using it gives us an invalid block
+                        * number, then just use the end of the previous block.
+                        */
+                       if (prevdiff <= XFS_ALLOC_GAP_UNITS * ap->length &&
+                           ISVALID(prevbno + prevdiff,
+                                   ap->prev.br_startblock))
+                               prevbno += adjust;
+                       else
+                               prevdiff += adjust;
+                       /*
+                        * If the firstblock forbids it, can't use it,
+                        * must use default.
+                        */
+                       if (!rt && !nullfb &&
+                           XFS_FSB_TO_AGNO(mp, prevbno) != fb_agno)
+                               prevbno = NULLFSBLOCK;
                }
-               ASSERT(i == 0 ||
-                      mval[i - 1].br_startoff + mval[i - 1].br_blockcount ==
-                      mval[i].br_startoff);
-               ASSERT(mval[i].br_startblock != DELAYSTARTBLOCK &&
-                      mval[i].br_startblock != HOLESTARTBLOCK);
-               ASSERT(mval[i].br_state == XFS_EXT_NORM ||
-                      mval[i].br_state == XFS_EXT_UNWRITTEN);
+               /*
+                * No previous block or can't follow it, just default.
+                */
+               else
+                       prevbno = NULLFSBLOCK;
+               /*
+                * If there's a following (right) block, select a requested
+                * start block based on it.
+                */
+               if (!isnullstartblock(ap->got.br_startblock)) {
+                       /*
+                        * Calculate gap to start of next block.
+                        */
+                       adjust = gotdiff = ap->got.br_startoff - ap->offset;
+                       /*
+                        * Figure the startblock based on the next block's
+                        * start and the gap size.
+                        */
+                       gotbno = ap->got.br_startblock;
+                       /*
+                        * Heuristic!
+                        * If the gap is large relative to the piece we're
+                        * allocating, or using it gives us an invalid block
+                        * number, then just use the start of the next block
+                        * offset by our length.
+                        */
+                       if (gotdiff <= XFS_ALLOC_GAP_UNITS * ap->length &&
+                           ISVALID(gotbno - gotdiff, gotbno))
+                               gotbno -= adjust;
+                       else if (ISVALID(gotbno - ap->length, gotbno)) {
+                               gotbno -= ap->length;
+                               gotdiff += adjust - ap->length;
+                       } else
+                               gotdiff += adjust;
+                       /*
+                        * If the firstblock forbids it, can't use it,
+                        * must use default.
+                        */
+                       if (!rt && !nullfb &&
+                           XFS_FSB_TO_AGNO(mp, gotbno) != fb_agno)
+                               gotbno = NULLFSBLOCK;
+               }
+               /*
+                * No next block, just default.
+                */
+               else
+                       gotbno = NULLFSBLOCK;
+               /*
+                * If both valid, pick the better one, else the only good
+                * one, else ap->blkno is already set (to 0 or the inode block).
+                */
+               if (prevbno != NULLFSBLOCK && gotbno != NULLFSBLOCK)
+                       ap->blkno = prevdiff <= gotdiff ? prevbno : gotbno;
+               else if (prevbno != NULLFSBLOCK)
+                       ap->blkno = prevbno;
+               else if (gotbno != NULLFSBLOCK)
+                       ap->blkno = gotbno;
        }
+#undef ISVALID
 }
-#endif /* DEBUG */
-
 
-/*
- * Trim the returned map to the required bounds
- */
-STATIC void
-xfs_bmapi_trim_map(
-       struct xfs_bmbt_irec    *mval,
-       struct xfs_bmbt_irec    *got,
-       xfs_fileoff_t           *bno,
-       xfs_filblks_t           len,
-       xfs_fileoff_t           obno,
-       xfs_fileoff_t           end,
-       int                     n,
-       int                     flags)
+STATIC int
+xfs_bmap_rtalloc(
+       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
 {
-       if ((flags & XFS_BMAPI_ENTIRE) ||
-           got->br_startoff + got->br_blockcount <= obno) {
-               *mval = *got;
-               if (isnullstartblock(got->br_startblock))
-                       mval->br_startblock = DELAYSTARTBLOCK;
-               return;
-       }
+       xfs_alloctype_t atype = 0;      /* type for allocation routines */
+       int             error;          /* error return value */
+       xfs_mount_t     *mp;            /* mount point structure */
+       xfs_extlen_t    prod = 0;       /* product factor for allocators */
+       xfs_extlen_t    ralen = 0;      /* realtime allocation length */
+       xfs_extlen_t    align;          /* minimum allocation alignment */
+       xfs_rtblock_t   rtb;
+
+       mp = ap->ip->i_mount;
+       align = xfs_get_extsz_hint(ap->ip);
+       prod = align / mp->m_sb.sb_rextsize;
+       error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
+                                       align, 1, ap->eof, 0,
+                                       ap->conv, &ap->offset, &ap->length);
+       if (error)
+               return error;
+       ASSERT(ap->length);
+       ASSERT(ap->length % mp->m_sb.sb_rextsize == 0);
 
-       if (obno > *bno)
-               *bno = obno;
-       ASSERT((*bno >= obno) || (n == 0));
-       ASSERT(*bno < end);
-       mval->br_startoff = *bno;
-       if (isnullstartblock(got->br_startblock))
-               mval->br_startblock = DELAYSTARTBLOCK;
-       else
-               mval->br_startblock = got->br_startblock +
-                                       (*bno - got->br_startoff);
        /*
-        * Return the minimum of what we got and what we asked for for
-        * the length.  We can use the len variable here because it is
-        * modified below and we could have been there before coming
-        * here if the first part of the allocation didn't overlap what
-        * was asked for.
+        * If the offset & length are not perfectly aligned
+        * then kill prod, it will just get us in trouble.
         */
-       mval->br_blockcount = XFS_FILBLKS_MIN(end - *bno,
-                       got->br_blockcount - (*bno - got->br_startoff));
-       mval->br_state = got->br_state;
-       ASSERT(mval->br_blockcount <= len);
-       return;
-}
-
-/*
- * Update and validate the extent map to return
- */
-STATIC void
-xfs_bmapi_update_map(
-       struct xfs_bmbt_irec    **map,
-       xfs_fileoff_t           *bno,
-       xfs_filblks_t           *len,
-       xfs_fileoff_t           obno,
-       xfs_fileoff_t           end,
-       int                     *n,
-       int                     flags)
-{
-       xfs_bmbt_irec_t *mval = *map;
-
-       ASSERT((flags & XFS_BMAPI_ENTIRE) ||
-              ((mval->br_startoff + mval->br_blockcount) <= end));
-       ASSERT((flags & XFS_BMAPI_ENTIRE) || (mval->br_blockcount <= *len) ||
-              (mval->br_startoff < obno));
-
-       *bno = mval->br_startoff + mval->br_blockcount;
-       *len = end - *bno;
-       if (*n > 0 && mval->br_startoff == mval[-1].br_startoff) {
-               /* update previous map with new information */
-               ASSERT(mval->br_startblock == mval[-1].br_startblock);
-               ASSERT(mval->br_blockcount > mval[-1].br_blockcount);
-               ASSERT(mval->br_state == mval[-1].br_state);
-               mval[-1].br_blockcount = mval->br_blockcount;
-               mval[-1].br_state = mval->br_state;
-       } else if (*n > 0 && mval->br_startblock != DELAYSTARTBLOCK &&
-                  mval[-1].br_startblock != DELAYSTARTBLOCK &&
-                  mval[-1].br_startblock != HOLESTARTBLOCK &&
-                  mval->br_startblock == mval[-1].br_startblock +
-                                         mval[-1].br_blockcount &&
-                  ((flags & XFS_BMAPI_IGSTATE) ||
-                       mval[-1].br_state == mval->br_state)) {
-               ASSERT(mval->br_startoff ==
-                      mval[-1].br_startoff + mval[-1].br_blockcount);
-               mval[-1].br_blockcount += mval->br_blockcount;
-       } else if (*n > 0 &&
-                  mval->br_startblock == DELAYSTARTBLOCK &&
-                  mval[-1].br_startblock == DELAYSTARTBLOCK &&
-                  mval->br_startoff ==
-                  mval[-1].br_startoff + mval[-1].br_blockcount) {
-               mval[-1].br_blockcount += mval->br_blockcount;
-               mval[-1].br_state = mval->br_state;
-       } else if (!((*n == 0) &&
-                    ((mval->br_startoff + mval->br_blockcount) <=
-                     obno))) {
-               mval++;
-               (*n)++;
-       }
-       *map = mval;
-}
-
-/*
- * Map file blocks to filesystem blocks without allocation.
- */
-int
-xfs_bmapi_read(
-       struct xfs_inode        *ip,
-       xfs_fileoff_t           bno,
-       xfs_filblks_t           len,
-       struct xfs_bmbt_irec    *mval,
-       int                     *nmap,
-       int                     flags)
-{
-       struct xfs_mount        *mp = ip->i_mount;
-       struct xfs_ifork        *ifp;
-       struct xfs_bmbt_irec    got;
-       struct xfs_bmbt_irec    prev;
-       xfs_fileoff_t           obno;
-       xfs_fileoff_t           end;
-       xfs_extnum_t            lastx;
-       int                     error;
-       int                     eof;
-       int                     n = 0;
-       int                     whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
-                                               XFS_ATTR_FORK : XFS_DATA_FORK;
-
-       ASSERT(*nmap >= 1);
-       ASSERT(!(flags & ~(XFS_BMAPI_ATTRFORK|XFS_BMAPI_ENTIRE|
-                          XFS_BMAPI_IGSTATE)));
-
-       if (unlikely(XFS_TEST_ERROR(
-           (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
-            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE),
-            mp, XFS_ERRTAG_BMAPIFORMAT, XFS_RANDOM_BMAPIFORMAT))) {
-               XFS_ERROR_REPORT("xfs_bmapi_read", XFS_ERRLEVEL_LOW, mp);
-               return XFS_ERROR(EFSCORRUPTED);
-       }
-
-       if (XFS_FORCED_SHUTDOWN(mp))
-               return XFS_ERROR(EIO);
+       if (do_mod(ap->offset, align) || ap->length % align)
+               prod = 1;
+       /*
+        * Set ralen to be the actual requested length in rtextents.
+        */
+       ralen = ap->length / mp->m_sb.sb_rextsize;
+       /*
+        * If the old value was close enough to MAXEXTLEN that
+        * we rounded up to it, cut it back so it's valid again.
+        * Note that if it's a really large request (bigger than
+        * MAXEXTLEN), we don't hear about that number, and can't
+        * adjust the starting point to match it.
+        */
+       if (ralen * mp->m_sb.sb_rextsize >= MAXEXTLEN)
+               ralen = MAXEXTLEN / mp->m_sb.sb_rextsize;
 
-       XFS_STATS_INC(xs_blk_mapr);
+       /*
+        * Lock out other modifications to the RT bitmap inode.
+        */
+       xfs_ilock(mp->m_rbmip, XFS_ILOCK_EXCL);
+       xfs_trans_ijoin(ap->tp, mp->m_rbmip, XFS_ILOCK_EXCL);
 
-       ifp = XFS_IFORK_PTR(ip, whichfork);
+       /*
+        * If it's an allocation to an empty file at offset 0,
+        * pick an extent that will space things out in the rt area.
+        */
+       if (ap->eof && ap->offset == 0) {
+               xfs_rtblock_t uninitialized_var(rtx); /* realtime extent no */
 
-       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
-               error = xfs_iread_extents(NULL, ip, whichfork);
+               error = xfs_rtpick_extent(mp, ap->tp, ralen, &rtx);
                if (error)
                        return error;
+               ap->blkno = rtx * mp->m_sb.sb_rextsize;
+       } else {
+               ap->blkno = 0;
        }
 
-       xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got, &prev);
-       end = bno + len;
-       obno = bno;
-
-       while (bno < end && n < *nmap) {
-               /* Reading past eof, act as though there's a hole up to end. */
-               if (eof)
-                       got.br_startoff = end;
-               if (got.br_startoff > bno) {
-                       /* Reading in a hole.  */
-                       mval->br_startoff = bno;
-                       mval->br_startblock = HOLESTARTBLOCK;
-                       mval->br_blockcount =
-                               XFS_FILBLKS_MIN(len, got.br_startoff - bno);
-                       mval->br_state = XFS_EXT_NORM;
-                       bno += mval->br_blockcount;
-                       len -= mval->br_blockcount;
-                       mval++;
-                       n++;
-                       continue;
-               }
-
-               /* set up the extent map to return. */
-               xfs_bmapi_trim_map(mval, &got, &bno, len, obno, end, n, flags);
-               xfs_bmapi_update_map(&mval, &bno, &len, obno, end, &n, flags);
-
-               /* If we're done, stop now. */
-               if (bno >= end || n >= *nmap)
-                       break;
+       xfs_bmap_adjacent(ap);
 
-               /* Else go on to the next record. */
-               if (++lastx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
-                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx), &got);
-               else
-                       eof = 1;
+       /*
+        * Realtime allocation, done through xfs_rtallocate_extent.
+        */
+       atype = ap->blkno == 0 ?  XFS_ALLOCTYPE_ANY_AG : XFS_ALLOCTYPE_NEAR_BNO;
+       do_div(ap->blkno, mp->m_sb.sb_rextsize);
+       rtb = ap->blkno;
+       ap->length = ralen;
+       if ((error = xfs_rtallocate_extent(ap->tp, ap->blkno, 1, ap->length,
+                               &ralen, atype, ap->wasdel, prod, &rtb)))
+               return error;
+       if (rtb == NULLFSBLOCK && prod > 1 &&
+           (error = xfs_rtallocate_extent(ap->tp, ap->blkno, 1,
+                                          ap->length, &ralen, atype,
+                                          ap->wasdel, 1, &rtb)))
+               return error;
+       ap->blkno = rtb;
+       if (ap->blkno != NULLFSBLOCK) {
+               ap->blkno *= mp->m_sb.sb_rextsize;
+               ralen *= mp->m_sb.sb_rextsize;
+               ap->length = ralen;
+               ap->ip->i_d.di_nblocks += ralen;
+               xfs_trans_log_inode(ap->tp, ap->ip, XFS_ILOG_CORE);
+               if (ap->wasdel)
+                       ap->ip->i_delayed_blks -= ralen;
+               /*
+                * Adjust the disk quota also. This was reserved
+                * earlier.
+                */
+               xfs_trans_mod_dquot_byino(ap->tp, ap->ip,
+                       ap->wasdel ? XFS_TRANS_DQ_DELRTBCOUNT :
+                                       XFS_TRANS_DQ_RTBCOUNT, (long) ralen);
+       } else {
+               ap->length = 0;
        }
-       *nmap = n;
        return 0;
 }
 
 STATIC int
-xfs_bmapi_reserve_delalloc(
-       struct xfs_inode        *ip,
-       xfs_fileoff_t           aoff,
-       xfs_filblks_t           len,
-       struct xfs_bmbt_irec    *got,
-       struct xfs_bmbt_irec    *prev,
-       xfs_extnum_t            *lastx,
-       int                     eof)
+xfs_bmap_btalloc_nullfb(
+       struct xfs_bmalloca     *ap,
+       struct xfs_alloc_arg    *args,
+       xfs_extlen_t            *blen)
 {
-       struct xfs_mount        *mp = ip->i_mount;
-       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
-       xfs_extlen_t            alen;
-       xfs_extlen_t            indlen;
-       char                    rt = XFS_IS_REALTIME_INODE(ip);
-       xfs_extlen_t            extsz;
+       struct xfs_mount        *mp = ap->ip->i_mount;
+       struct xfs_perag        *pag;
+       xfs_agnumber_t          ag, startag;
+       int                     notinit = 0;
        int                     error;
 
-       alen = XFS_FILBLKS_MIN(len, MAXEXTLEN);
-       if (!eof)
-               alen = XFS_FILBLKS_MIN(alen, got->br_startoff - aoff);
-
-       /* Figure out the extent size, adjust alen */
-       extsz = xfs_get_extsz_hint(ip);
-       if (extsz) {
-               /*
-                * Make sure we don't exceed a single extent length when we
-                * align the extent by reducing length we are going to
-                * allocate by the maximum amount extent size aligment may
-                * require.
-                */
-               alen = XFS_FILBLKS_MIN(len, MAXEXTLEN - (2 * extsz - 1));
-               error = xfs_bmap_extsize_align(mp, got, prev, extsz, rt, eof,
-                                              1, 0, &aoff, &alen);
-               ASSERT(!error);
-       }
-
-       if (rt)
-               extsz = alen / mp->m_sb.sb_rextsize;
+       if (ap->userdata && xfs_inode_is_filestream(ap->ip))
+               args->type = XFS_ALLOCTYPE_NEAR_BNO;
+       else
+               args->type = XFS_ALLOCTYPE_START_BNO;
+       args->total = ap->total;
 
        /*
-        * Make a transaction-less quota reservation for delayed allocation
-        * blocks.  This number gets adjusted later.  We return if we haven't
-        * allocated blocks already inside this loop.
+        * Search for an allocation group with a single extent large enough
+        * for the request.  If one isn't found, then adjust the minimum
+        * allocation size to the largest space found.
         */
-       error = xfs_trans_reserve_quota_nblks(NULL, ip, (long)alen, 0,
-                       rt ? XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS);
-       if (error)
-               return error;
+       startag = ag = XFS_FSB_TO_AGNO(mp, args->fsbno);
+       if (startag == NULLAGNUMBER)
+               startag = ag = 0;
 
-       /*
-        * Split changing sb for alen and indlen since they could be coming
-        * from different places.
-        */
-       indlen = (xfs_extlen_t)xfs_bmap_worst_indlen(ip, alen);
-       ASSERT(indlen > 0);
+       pag = xfs_perag_get(mp, ag);
+       while (*blen < args->maxlen) {
+               if (!pag->pagf_init) {
+                       error = xfs_alloc_pagf_init(mp, args->tp, ag,
+                                                   XFS_ALLOC_FLAG_TRYLOCK);
+                       if (error) {
+                               xfs_perag_put(pag);
+                               return error;
+                       }
+               }
 
-       if (rt) {
-               error = xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS,
-                                         -((int64_t)extsz), 0);
-       } else {
-               error = xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
-                                                -((int64_t)alen), 0);
-       }
+               /*
+                * See xfs_alloc_fix_freelist...
+                */
+               if (pag->pagf_init) {
+                       xfs_extlen_t    longest;
+                       longest = xfs_alloc_longest_free_extent(mp, pag);
+                       if (*blen < longest)
+                               *blen = longest;
+               } else
+                       notinit = 1;
 
-       if (error)
-               goto out_unreserve_quota;
+               if (xfs_inode_is_filestream(ap->ip)) {
+                       if (*blen >= args->maxlen)
+                               break;
 
-       error = xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
-                                        -((int64_t)indlen), 0);
-       if (error)
-               goto out_unreserve_blocks;
+                       if (ap->userdata) {
+                               /*
+                                * If startag is an invalid AG, we've
+                                * come here once before and
+                                * xfs_filestream_new_ag picked the
+                                * best currently available.
+                                *
+                                * Don't continue looping, since we
+                                * could loop forever.
+                                */
+                               if (startag == NULLAGNUMBER)
+                                       break;
 
+                               error = xfs_filestream_new_ag(ap, &ag);
+                               xfs_perag_put(pag);
+                               if (error)
+                                       return error;
 
-       ip->i_delayed_blks += alen;
+                               /* loop again to set 'blen'*/
+                               startag = NULLAGNUMBER;
+                               pag = xfs_perag_get(mp, ag);
+                               continue;
+                       }
+               }
+               if (++ag == mp->m_sb.sb_agcount)
+                       ag = 0;
+               if (ag == startag)
+                       break;
+               xfs_perag_put(pag);
+               pag = xfs_perag_get(mp, ag);
+       }
+       xfs_perag_put(pag);
 
-       got->br_startoff = aoff;
-       got->br_startblock = nullstartblock(indlen);
-       got->br_blockcount = alen;
-       got->br_state = XFS_EXT_NORM;
-       xfs_bmap_add_extent_hole_delay(ip, lastx, got);
+       /*
+        * Since the above loop did a BUF_TRYLOCK, it is
+        * possible that there is space for this request.
+        */
+       if (notinit || *blen < ap->minlen)
+               args->minlen = ap->minlen;
+       /*
+        * If the best seen length is less than the request
+        * length, use the best as the minimum.
+        */
+       else if (*blen < args->maxlen)
+               args->minlen = *blen;
+       /*
+        * Otherwise we've seen an extent as big as maxlen,
+        * use that as the minimum.
+        */
+       else
+               args->minlen = args->maxlen;
 
        /*
-        * Update our extent pointer, given that xfs_bmap_add_extent_hole_delay
-        * might have merged it into one of the neighbouring ones.
+        * set the failure fallback case to look in the selected
+        * AG as the stream may have moved.
         */
-       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *lastx), got);
+       if (xfs_inode_is_filestream(ap->ip))
+               ap->blkno = args->fsbno = XFS_AGB_TO_FSB(mp, ag, 0);
 
-       ASSERT(got->br_startoff <= aoff);
-       ASSERT(got->br_startoff + got->br_blockcount >= aoff + alen);
-       ASSERT(isnullstartblock(got->br_startblock));
-       ASSERT(got->br_state == XFS_EXT_NORM);
        return 0;
-
-out_unreserve_blocks:
-       if (rt)
-               xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS, extsz, 0);
-       else
-               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS, alen, 0);
-out_unreserve_quota:
-       if (XFS_IS_QUOTA_ON(mp))
-               xfs_trans_unreserve_quota_nblks(NULL, ip, (long)alen, 0, rt ?
-                               XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS);
-       return error;
 }
 
-/*
- * Map file blocks to filesystem blocks, adding delayed allocations as needed.
- */
-int
-xfs_bmapi_delay(
-       struct xfs_inode        *ip,    /* incore inode */
-       xfs_fileoff_t           bno,    /* starting file offs. mapped */
-       xfs_filblks_t           len,    /* length to map in file */
-       struct xfs_bmbt_irec    *mval,  /* output: map values */
-       int                     *nmap,  /* i/o: mval size/count */
-       int                     flags)  /* XFS_BMAPI_... */
+STATIC int
+xfs_bmap_btalloc(
+       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
 {
-       struct xfs_mount        *mp = ip->i_mount;
-       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
-       struct xfs_bmbt_irec    got;    /* current file extent record */
-       struct xfs_bmbt_irec    prev;   /* previous file extent record */
-       xfs_fileoff_t           obno;   /* old block number (offset) */
-       xfs_fileoff_t           end;    /* end of mapped file region */
-       xfs_extnum_t            lastx;  /* last useful extent number */
-       int                     eof;    /* we've hit the end of extents */
-       int                     n = 0;  /* current extent index */
-       int                     error = 0;
+       xfs_mount_t     *mp;            /* mount point structure */
+       xfs_alloctype_t atype = 0;      /* type for allocation routines */
+       xfs_extlen_t    align;          /* minimum allocation alignment */
+       xfs_agnumber_t  fb_agno;        /* ag number of ap->firstblock */
+       xfs_agnumber_t  ag;
+       xfs_alloc_arg_t args;
+       xfs_extlen_t    blen;
+       xfs_extlen_t    nextminlen = 0;
+       int             nullfb;         /* true if ap->firstblock isn't set */
+       int             isaligned;
+       int             tryagain;
+       int             error;
 
-       ASSERT(*nmap >= 1);
-       ASSERT(*nmap <= XFS_BMAP_MAX_NMAP);
-       ASSERT(!(flags & ~XFS_BMAPI_ENTIRE));
+       ASSERT(ap->length);
 
-       if (unlikely(XFS_TEST_ERROR(
-           (XFS_IFORK_FORMAT(ip, XFS_DATA_FORK) != XFS_DINODE_FMT_EXTENTS &&
-            XFS_IFORK_FORMAT(ip, XFS_DATA_FORK) != XFS_DINODE_FMT_BTREE),
-            mp, XFS_ERRTAG_BMAPIFORMAT, XFS_RANDOM_BMAPIFORMAT))) {
-               XFS_ERROR_REPORT("xfs_bmapi_delay", XFS_ERRLEVEL_LOW, mp);
-               return XFS_ERROR(EFSCORRUPTED);
+       mp = ap->ip->i_mount;
+       align = ap->userdata ? xfs_get_extsz_hint(ap->ip) : 0;
+       if (unlikely(align)) {
+               error = xfs_bmap_extsize_align(mp, &ap->got, &ap->prev,
+                                               align, 0, ap->eof, 0, ap->conv,
+                                               &ap->offset, &ap->length);
+               ASSERT(!error);
+               ASSERT(ap->length);
        }
+       nullfb = *ap->firstblock == NULLFSBLOCK;
+       fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp, *ap->firstblock);
+       if (nullfb) {
+               if (ap->userdata && xfs_inode_is_filestream(ap->ip)) {
+                       ag = xfs_filestream_lookup_ag(ap->ip);
+                       ag = (ag != NULLAGNUMBER) ? ag : 0;
+                       ap->blkno = XFS_AGB_TO_FSB(mp, ag, 0);
+               } else {
+                       ap->blkno = XFS_INO_TO_FSB(mp, ap->ip->i_ino);
+               }
+       } else
+               ap->blkno = *ap->firstblock;
 
-       if (XFS_FORCED_SHUTDOWN(mp))
-               return XFS_ERROR(EIO);
+       xfs_bmap_adjacent(ap);
 
-       XFS_STATS_INC(xs_blk_mapw);
+       /*
+        * If allowed, use ap->blkno; otherwise must use firstblock since
+        * it's in the right allocation group.
+        */
+       if (nullfb || XFS_FSB_TO_AGNO(mp, ap->blkno) == fb_agno)
+               ;
+       else
+               ap->blkno = *ap->firstblock;
+       /*
+        * Normal allocation, done through xfs_alloc_vextent.
+        */
+       tryagain = isaligned = 0;
+       memset(&args, 0, sizeof(args));
+       args.tp = ap->tp;
+       args.mp = mp;
+       args.fsbno = ap->blkno;
 
-       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
-               error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
+       /* Trim the allocation back to the maximum an AG can fit. */
+       args.maxlen = MIN(ap->length, XFS_ALLOC_AG_MAX_USABLE(mp));
+       args.firstblock = *ap->firstblock;
+       blen = 0;
+       if (nullfb) {
+               error = xfs_bmap_btalloc_nullfb(ap, &args, &blen);
                if (error)
                        return error;
+       } else if (ap->flist->xbf_low) {
+               if (xfs_inode_is_filestream(ap->ip))
+                       args.type = XFS_ALLOCTYPE_FIRST_AG;
+               else
+                       args.type = XFS_ALLOCTYPE_START_BNO;
+               args.total = args.minlen = ap->minlen;
+       } else {
+               args.type = XFS_ALLOCTYPE_NEAR_BNO;
+               args.total = ap->total;
+               args.minlen = ap->minlen;
        }
-
-       xfs_bmap_search_extents(ip, bno, XFS_DATA_FORK, &eof, &lastx, &got, &prev);
-       end = bno + len;
-       obno = bno;
-
-       while (bno < end && n < *nmap) {
-               if (eof || got.br_startoff > bno) {
-                       error = xfs_bmapi_reserve_delalloc(ip, bno, len, &got,
-                                                          &prev, &lastx, eof);
-                       if (error) {
-                               if (n == 0) {
-                                       *nmap = 0;
-                                       return error;
-                               }
-                               break;
-                       }
-               }
-
-               /* set up the extent map to return. */
-               xfs_bmapi_trim_map(mval, &got, &bno, len, obno, end, n, flags);
-               xfs_bmapi_update_map(&mval, &bno, &len, obno, end, &n, flags);
-
-               /* If we're done, stop now. */
-               if (bno >= end || n >= *nmap)
-                       break;
-
-               /* Else go on to the next record. */
-               prev = got;
-               if (++lastx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
-                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx), &got);
-               else
-                       eof = 1;
+       /* apply extent size hints if obtained earlier */
+       if (unlikely(align)) {
+               args.prod = align;
+               if ((args.mod = (xfs_extlen_t)do_mod(ap->offset, args.prod)))
+                       args.mod = (xfs_extlen_t)(args.prod - args.mod);
+       } else if (mp->m_sb.sb_blocksize >= PAGE_CACHE_SIZE) {
+               args.prod = 1;
+               args.mod = 0;
+       } else {
+               args.prod = PAGE_CACHE_SIZE >> mp->m_sb.sb_blocklog;
+               if ((args.mod = (xfs_extlen_t)(do_mod(ap->offset, args.prod))))
+                       args.mod = (xfs_extlen_t)(args.prod - args.mod);
        }
-
-       *nmap = n;
-       return 0;
-}
-
-
-STATIC int
-__xfs_bmapi_allocate(
-       struct xfs_bmalloca     *bma)
-{
-       struct xfs_mount        *mp = bma->ip->i_mount;
-       int                     whichfork = (bma->flags & XFS_BMAPI_ATTRFORK) ?
-                                               XFS_ATTR_FORK : XFS_DATA_FORK;
-       struct xfs_ifork        *ifp = XFS_IFORK_PTR(bma->ip, whichfork);
-       int                     tmp_logflags = 0;
-       int                     error;
-       int                     rt;
-
-       ASSERT(bma->length > 0);
-
-       rt = (whichfork == XFS_DATA_FORK) && XFS_IS_REALTIME_INODE(bma->ip);
-
        /*
-        * For the wasdelay case, we could also just allocate the stuff asked
-        * for in this bmap call but that wouldn't be as good.
+        * If we are not low on available data blocks, and the
+        * underlying logical volume manager is a stripe, and
+        * the file offset is zero then try to allocate data
+        * blocks on stripe unit boundary.
+        * NOTE: ap->aeof is only set if the allocation length
+        * is >= the stripe unit and the allocation offset is
+        * at the end of file.
         */
-       if (bma->wasdel) {
-               bma->length = (xfs_extlen_t)bma->got.br_blockcount;
-               bma->offset = bma->got.br_startoff;
-               if (bma->idx != NULLEXTNUM && bma->idx) {
-                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1),
-                                        &bma->prev);
+       if (!ap->flist->xbf_low && ap->aeof) {
+               if (!ap->offset) {
+                       args.alignment = mp->m_dalign;
+                       atype = args.type;
+                       isaligned = 1;
+                       /*
+                        * Adjust for alignment
+                        */
+                       if (blen > args.alignment && blen <= args.maxlen)
+                               args.minlen = blen - args.alignment;
+                       args.minalignslop = 0;
+               } else {
+                       /*
+                        * First try an exact bno allocation.
+                        * If it fails then do a near or start bno
+                        * allocation with alignment turned on.
+                        */
+                       atype = args.type;
+                       tryagain = 1;
+                       args.type = XFS_ALLOCTYPE_THIS_BNO;
+                       args.alignment = 1;
+                       /*
+                        * Compute the minlen+alignment for the
+                        * next case.  Set slop so that the value
+                        * of minlen+alignment+slop doesn't go up
+                        * between the calls.
+                        */
+                       if (blen > mp->m_dalign && blen <= args.maxlen)
+                               nextminlen = blen - mp->m_dalign;
+                       else
+                               nextminlen = args.minlen;
+                       if (nextminlen + mp->m_dalign > args.minlen + 1)
+                               args.minalignslop =
+                                       nextminlen + mp->m_dalign -
+                                       args.minlen - 1;
+                       else
+                               args.minalignslop = 0;
                }
        } else {
-               bma->length = XFS_FILBLKS_MIN(bma->length, MAXEXTLEN);
-               if (!bma->eof)
-                       bma->length = XFS_FILBLKS_MIN(bma->length,
-                                       bma->got.br_startoff - bma->offset);
+               args.alignment = 1;
+               args.minalignslop = 0;
        }
-
-       /*
-        * Indicate if this is the first user data in the file, or just any
-        * user data.
-        */
-       if (!(bma->flags & XFS_BMAPI_METADATA)) {
-               bma->userdata = (bma->offset == 0) ?
-                       XFS_ALLOC_INITIAL_USER_DATA : XFS_ALLOC_USERDATA;
+       args.minleft = ap->minleft;
+       args.wasdel = ap->wasdel;
+       args.isfl = 0;
+       args.userdata = ap->userdata;
+       if ((error = xfs_alloc_vextent(&args)))
+               return error;
+       if (tryagain && args.fsbno == NULLFSBLOCK) {
+               /*
+                * Exact allocation failed. Now try with alignment
+                * turned on.
+                */
+               args.type = atype;
+               args.fsbno = ap->blkno;
+               args.alignment = mp->m_dalign;
+               args.minlen = nextminlen;
+               args.minalignslop = 0;
+               isaligned = 1;
+               if ((error = xfs_alloc_vextent(&args)))
+                       return error;
        }
-
-       bma->minlen = (bma->flags & XFS_BMAPI_CONTIG) ? bma->length : 1;
-
-       /*
-        * Only want to do the alignment at the eof if it is userdata and
-        * allocation length is larger than a stripe unit.
-        */
-       if (mp->m_dalign && bma->length >= mp->m_dalign &&
-           !(bma->flags & XFS_BMAPI_METADATA) && whichfork == XFS_DATA_FORK) {
-               error = xfs_bmap_isaeof(bma, whichfork);
-               if (error)
+       if (isaligned && args.fsbno == NULLFSBLOCK) {
+               /*
+                * allocation failed, so turn off alignment and
+                * try again.
+                */
+               args.type = atype;
+               args.fsbno = ap->blkno;
+               args.alignment = 0;
+               if ((error = xfs_alloc_vextent(&args)))
                        return error;
        }
-
-       error = xfs_bmap_alloc(bma);
-       if (error)
-               return error;
-
-       if (bma->flist->xbf_low)
-               bma->minleft = 0;
-       if (bma->cur)
-               bma->cur->bc_private.b.firstblock = *bma->firstblock;
-       if (bma->blkno == NULLFSBLOCK)
-               return 0;
-       if ((ifp->if_flags & XFS_IFBROOT) && !bma->cur) {
-               bma->cur = xfs_bmbt_init_cursor(mp, bma->tp, bma->ip, whichfork);
-               bma->cur->bc_private.b.firstblock = *bma->firstblock;
-               bma->cur->bc_private.b.flist = bma->flist;
+       if (args.fsbno == NULLFSBLOCK && nullfb &&
+           args.minlen > ap->minlen) {
+               args.minlen = ap->minlen;
+               args.type = XFS_ALLOCTYPE_START_BNO;
+               args.fsbno = ap->blkno;
+               if ((error = xfs_alloc_vextent(&args)))
+                       return error;
        }
-       /*
-        * Bump the number of extents we've allocated
-        * in this call.
-        */
-       bma->nallocs++;
-
-       if (bma->cur)
-               bma->cur->bc_private.b.flags =
-                       bma->wasdel ? XFS_BTCUR_BPRV_WASDEL : 0;
-
-       bma->got.br_startoff = bma->offset;
-       bma->got.br_startblock = bma->blkno;
-       bma->got.br_blockcount = bma->length;
-       bma->got.br_state = XFS_EXT_NORM;
+       if (args.fsbno == NULLFSBLOCK && nullfb) {
+               args.fsbno = 0;
+               args.type = XFS_ALLOCTYPE_FIRST_AG;
+               args.total = ap->minlen;
+               args.minleft = 0;
+               if ((error = xfs_alloc_vextent(&args)))
+                       return error;
+               ap->flist->xbf_low = 1;
+       }
+       if (args.fsbno != NULLFSBLOCK) {
+               /*
+                * check the allocation happened at the same or higher AG than
+                * the first block that was allocated.
+                */
+               ASSERT(*ap->firstblock == NULLFSBLOCK ||
+                      XFS_FSB_TO_AGNO(mp, *ap->firstblock) ==
+                      XFS_FSB_TO_AGNO(mp, args.fsbno) ||
+                      (ap->flist->xbf_low &&
+                       XFS_FSB_TO_AGNO(mp, *ap->firstblock) <
+                       XFS_FSB_TO_AGNO(mp, args.fsbno)));
 
-       /*
-        * A wasdelay extent has been initialized, so shouldn't be flagged
-        * as unwritten.
-        */
-       if (!bma->wasdel && (bma->flags & XFS_BMAPI_PREALLOC) &&
-           xfs_sb_version_hasextflgbit(&mp->m_sb))
-               bma->got.br_state = XFS_EXT_UNWRITTEN;
+               ap->blkno = args.fsbno;
+               if (*ap->firstblock == NULLFSBLOCK)
+                       *ap->firstblock = args.fsbno;
+               ASSERT(nullfb || fb_agno == args.agno ||
+                      (ap->flist->xbf_low && fb_agno < args.agno));
+               ap->length = args.len;
+               ap->ip->i_d.di_nblocks += args.len;
+               xfs_trans_log_inode(ap->tp, ap->ip, XFS_ILOG_CORE);
+               if (ap->wasdel)
+                       ap->ip->i_delayed_blks -= args.len;
+               /*
+                * Adjust the disk quota also. This was reserved
+                * earlier.
+                */
+               xfs_trans_mod_dquot_byino(ap->tp, ap->ip,
+                       ap->wasdel ? XFS_TRANS_DQ_DELBCOUNT :
+                                       XFS_TRANS_DQ_BCOUNT,
+                       (long) args.len);
+       } else {
+               ap->blkno = NULLFSBLOCK;
+               ap->length = 0;
+       }
+       return 0;
+}
 
-       if (bma->wasdel)
-               error = xfs_bmap_add_extent_delay_real(bma);
-       else
-               error = xfs_bmap_add_extent_hole_real(bma, whichfork);
+/*
+ * xfs_bmap_alloc is called by xfs_bmapi to allocate an extent for a file.
+ * It figures out where to ask the underlying allocator to put the new extent.
+ */
+STATIC int
+xfs_bmap_alloc(
+       xfs_bmalloca_t  *ap)            /* bmap alloc argument struct */
+{
+       if (XFS_IS_REALTIME_INODE(ap->ip) && ap->userdata)
+               return xfs_bmap_rtalloc(ap);
+       return xfs_bmap_btalloc(ap);
+}
 
-       bma->logflags |= tmp_logflags;
-       if (error)
-               return error;
+/*
+ * Trim the returned map to the required bounds
+ */
+STATIC void
+xfs_bmapi_trim_map(
+       struct xfs_bmbt_irec    *mval,
+       struct xfs_bmbt_irec    *got,
+       xfs_fileoff_t           *bno,
+       xfs_filblks_t           len,
+       xfs_fileoff_t           obno,
+       xfs_fileoff_t           end,
+       int                     n,
+       int                     flags)
+{
+       if ((flags & XFS_BMAPI_ENTIRE) ||
+           got->br_startoff + got->br_blockcount <= obno) {
+               *mval = *got;
+               if (isnullstartblock(got->br_startblock))
+                       mval->br_startblock = DELAYSTARTBLOCK;
+               return;
+       }
 
+       if (obno > *bno)
+               *bno = obno;
+       ASSERT((*bno >= obno) || (n == 0));
+       ASSERT(*bno < end);
+       mval->br_startoff = *bno;
+       if (isnullstartblock(got->br_startblock))
+               mval->br_startblock = DELAYSTARTBLOCK;
+       else
+               mval->br_startblock = got->br_startblock +
+                                       (*bno - got->br_startoff);
        /*
-        * Update our extent pointer, given that xfs_bmap_add_extent_delay_real
-        * or xfs_bmap_add_extent_hole_real might have merged it into one of
-        * the neighbouring ones.
+        * Return the minimum of what we got and what we asked for for
+        * the length.  We can use the len variable here because it is
+        * modified below and we could have been there before coming
+        * here if the first part of the allocation didn't overlap what
+        * was asked for.
         */
-       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &bma->got);
-
-       ASSERT(bma->got.br_startoff <= bma->offset);
-       ASSERT(bma->got.br_startoff + bma->got.br_blockcount >=
-              bma->offset + bma->length);
-       ASSERT(bma->got.br_state == XFS_EXT_NORM ||
-              bma->got.br_state == XFS_EXT_UNWRITTEN);
-       return 0;
+       mval->br_blockcount = XFS_FILBLKS_MIN(end - *bno,
+                       got->br_blockcount - (*bno - got->br_startoff));
+       mval->br_state = got->br_state;
+       ASSERT(mval->br_blockcount <= len);
+       return;
 }
 
-static void
-xfs_bmapi_allocate_worker(
-       struct work_struct      *work)
+/*
+ * Update and validate the extent map to return
+ */
+STATIC void
+xfs_bmapi_update_map(
+       struct xfs_bmbt_irec    **map,
+       xfs_fileoff_t           *bno,
+       xfs_filblks_t           *len,
+       xfs_fileoff_t           obno,
+       xfs_fileoff_t           end,
+       int                     *n,
+       int                     flags)
 {
-       struct xfs_bmalloca     *args = container_of(work,
-                                               struct xfs_bmalloca, work);
-       unsigned long           pflags;
-
-       /* we are in a transaction context here */
-       current_set_flags_nested(&pflags, PF_FSTRANS);
+       xfs_bmbt_irec_t *mval = *map;
 
-       args->result = __xfs_bmapi_allocate(args);
-       complete(args->done);
+       ASSERT((flags & XFS_BMAPI_ENTIRE) ||
+              ((mval->br_startoff + mval->br_blockcount) <= end));
+       ASSERT((flags & XFS_BMAPI_ENTIRE) || (mval->br_blockcount <= *len) ||
+              (mval->br_startoff < obno));
 
-       current_restore_flags_nested(&pflags, PF_FSTRANS);
+       *bno = mval->br_startoff + mval->br_blockcount;
+       *len = end - *bno;
+       if (*n > 0 && mval->br_startoff == mval[-1].br_startoff) {
+               /* update previous map with new information */
+               ASSERT(mval->br_startblock == mval[-1].br_startblock);
+               ASSERT(mval->br_blockcount > mval[-1].br_blockcount);
+               ASSERT(mval->br_state == mval[-1].br_state);
+               mval[-1].br_blockcount = mval->br_blockcount;
+               mval[-1].br_state = mval->br_state;
+       } else if (*n > 0 && mval->br_startblock != DELAYSTARTBLOCK &&
+                  mval[-1].br_startblock != DELAYSTARTBLOCK &&
+                  mval[-1].br_startblock != HOLESTARTBLOCK &&
+                  mval->br_startblock == mval[-1].br_startblock +
+                                         mval[-1].br_blockcount &&
+                  ((flags & XFS_BMAPI_IGSTATE) ||
+                       mval[-1].br_state == mval->br_state)) {
+               ASSERT(mval->br_startoff ==
+                      mval[-1].br_startoff + mval[-1].br_blockcount);
+               mval[-1].br_blockcount += mval->br_blockcount;
+       } else if (*n > 0 &&
+                  mval->br_startblock == DELAYSTARTBLOCK &&
+                  mval[-1].br_startblock == DELAYSTARTBLOCK &&
+                  mval->br_startoff ==
+                  mval[-1].br_startoff + mval[-1].br_blockcount) {
+               mval[-1].br_blockcount += mval->br_blockcount;
+               mval[-1].br_state = mval->br_state;
+       } else if (!((*n == 0) &&
+                    ((mval->br_startoff + mval->br_blockcount) <=
+                     obno))) {
+               mval++;
+               (*n)++;
+       }
+       *map = mval;
 }
 
 /*
- * Some allocation requests often come in with little stack to work on. Push
- * them off to a worker thread so there is lots of stack to use. Otherwise just
- * call directly to avoid the context switch overhead here.
+ * Map file blocks to filesystem blocks without allocation.
  */
 int
-xfs_bmapi_allocate(
-       struct xfs_bmalloca     *args)
-{
-       DECLARE_COMPLETION_ONSTACK(done);
-
-       if (!args->stack_switch)
-               return __xfs_bmapi_allocate(args);
-
-
-       args->done = &done;
-       INIT_WORK_ONSTACK(&args->work, xfs_bmapi_allocate_worker);
-       queue_work(xfs_alloc_wq, &args->work);
-       wait_for_completion(&done);
-       return args->result;
-}
-
-STATIC int
-xfs_bmapi_convert_unwritten(
-       struct xfs_bmalloca     *bma,
-       struct xfs_bmbt_irec    *mval,
+xfs_bmapi_read(
+       struct xfs_inode        *ip,
+       xfs_fileoff_t           bno,
        xfs_filblks_t           len,
+       struct xfs_bmbt_irec    *mval,
+       int                     *nmap,
        int                     flags)
 {
+       struct xfs_mount        *mp = ip->i_mount;
+       struct xfs_ifork        *ifp;
+       struct xfs_bmbt_irec    got;
+       struct xfs_bmbt_irec    prev;
+       xfs_fileoff_t           obno;
+       xfs_fileoff_t           end;
+       xfs_extnum_t            lastx;
+       int                     error;
+       int                     eof;
+       int                     n = 0;
        int                     whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
                                                XFS_ATTR_FORK : XFS_DATA_FORK;
-       struct xfs_ifork        *ifp = XFS_IFORK_PTR(bma->ip, whichfork);
-       int                     tmp_logflags = 0;
-       int                     error;
 
-       /* check if we need to do unwritten->real conversion */
-       if (mval->br_state == XFS_EXT_UNWRITTEN &&
-           (flags & XFS_BMAPI_PREALLOC))
-               return 0;
+       ASSERT(*nmap >= 1);
+       ASSERT(!(flags & ~(XFS_BMAPI_ATTRFORK|XFS_BMAPI_ENTIRE|
+                          XFS_BMAPI_IGSTATE)));
 
-       /* check if we need to do real->unwritten conversion */
-       if (mval->br_state == XFS_EXT_NORM &&
-           (flags & (XFS_BMAPI_PREALLOC | XFS_BMAPI_CONVERT)) !=
-                       (XFS_BMAPI_PREALLOC | XFS_BMAPI_CONVERT))
-               return 0;
+       if (unlikely(XFS_TEST_ERROR(
+           (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
+            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE),
+            mp, XFS_ERRTAG_BMAPIFORMAT, XFS_RANDOM_BMAPIFORMAT))) {
+               XFS_ERROR_REPORT("xfs_bmapi_read", XFS_ERRLEVEL_LOW, mp);
+               return XFS_ERROR(EFSCORRUPTED);
+       }
 
-       /*
-        * Modify (by adding) the state flag, if writing.
-        */
-       ASSERT(mval->br_blockcount <= len);
-       if ((ifp->if_flags & XFS_IFBROOT) && !bma->cur) {
-               bma->cur = xfs_bmbt_init_cursor(bma->ip->i_mount, bma->tp,
-                                       bma->ip, whichfork);
-               bma->cur->bc_private.b.firstblock = *bma->firstblock;
-               bma->cur->bc_private.b.flist = bma->flist;
+       if (XFS_FORCED_SHUTDOWN(mp))
+               return XFS_ERROR(EIO);
+
+       XFS_STATS_INC(xs_blk_mapr);
+
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+
+       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
+               error = xfs_iread_extents(NULL, ip, whichfork);
+               if (error)
+                       return error;
        }
-       mval->br_state = (mval->br_state == XFS_EXT_UNWRITTEN)
-                               ? XFS_EXT_NORM : XFS_EXT_UNWRITTEN;
 
-       error = xfs_bmap_add_extent_unwritten_real(bma->tp, bma->ip, &bma->idx,
-                       &bma->cur, mval, bma->firstblock, bma->flist,
-                       &tmp_logflags);
-       bma->logflags |= tmp_logflags;
-       if (error)
-               return error;
+       xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got, &prev);
+       end = bno + len;
+       obno = bno;
 
-       /*
-        * Update our extent pointer, given that
-        * xfs_bmap_add_extent_unwritten_real might have merged it into one
-        * of the neighbouring ones.
-        */
-       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &bma->got);
+       while (bno < end && n < *nmap) {
+               /* Reading past eof, act as though there's a hole up to end. */
+               if (eof)
+                       got.br_startoff = end;
+               if (got.br_startoff > bno) {
+                       /* Reading in a hole.  */
+                       mval->br_startoff = bno;
+                       mval->br_startblock = HOLESTARTBLOCK;
+                       mval->br_blockcount =
+                               XFS_FILBLKS_MIN(len, got.br_startoff - bno);
+                       mval->br_state = XFS_EXT_NORM;
+                       bno += mval->br_blockcount;
+                       len -= mval->br_blockcount;
+                       mval++;
+                       n++;
+                       continue;
+               }
 
-       /*
-        * We may have combined previously unwritten space with written space,
-        * so generate another request.
-        */
-       if (mval->br_blockcount < len)
-               return EAGAIN;
+               /* set up the extent map to return. */
+               xfs_bmapi_trim_map(mval, &got, &bno, len, obno, end, n, flags);
+               xfs_bmapi_update_map(&mval, &bno, &len, obno, end, &n, flags);
+
+               /* If we're done, stop now. */
+               if (bno >= end || n >= *nmap)
+                       break;
+
+               /* Else go on to the next record. */
+               if (++lastx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
+                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx), &got);
+               else
+                       eof = 1;
+       }
+       *nmap = n;
        return 0;
 }
 
-/*
- * Map file blocks to filesystem blocks, and allocate blocks or convert the
- * extent state if necessary.  Details behaviour is controlled by the flags
- * parameter.  Only allocates blocks from a single allocation group, to avoid
- * locking problems.
- *
- * The returned value in "firstblock" from the first call in a transaction
- * must be remembered and presented to subsequent calls in "firstblock".
- * An upper bound for the number of blocks to be allocated is supplied to
- * the first call in "total"; if no allocation group has that many free
- * blocks then the call will fail (return NULLFSBLOCK in "firstblock").
- */
-int
-xfs_bmapi_write(
-       struct xfs_trans        *tp,            /* transaction pointer */
-       struct xfs_inode        *ip,            /* incore inode */
-       xfs_fileoff_t           bno,            /* starting file offs. mapped */
-       xfs_filblks_t           len,            /* length to map in file */
-       int                     flags,          /* XFS_BMAPI_... */
-       xfs_fsblock_t           *firstblock,    /* first allocated block
-                                                  controls a.g. for allocs */
-       xfs_extlen_t            total,          /* total blocks needed */
-       struct xfs_bmbt_irec    *mval,          /* output: map values */
-       int                     *nmap,          /* i/o: mval size/count */
-       struct xfs_bmap_free    *flist)         /* i/o: list extents to free */
+STATIC int
+xfs_bmapi_reserve_delalloc(
+       struct xfs_inode        *ip,
+       xfs_fileoff_t           aoff,
+       xfs_filblks_t           len,
+       struct xfs_bmbt_irec    *got,
+       struct xfs_bmbt_irec    *prev,
+       xfs_extnum_t            *lastx,
+       int                     eof)
 {
        struct xfs_mount        *mp = ip->i_mount;
-       struct xfs_ifork        *ifp;
-       struct xfs_bmalloca     bma = { 0 };    /* args for xfs_bmap_alloc */
-       xfs_fileoff_t           end;            /* end of mapped file region */
-       int                     eof;            /* after the end of extents */
-       int                     error;          /* error return */
-       int                     n;              /* current extent index */
-       xfs_fileoff_t           obno;           /* old block number (offset) */
-       int                     whichfork;      /* data or attr fork */
-       char                    inhole;         /* current location is hole in file */
-       char                    wasdelay;       /* old extent was delayed */
+       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
+       xfs_extlen_t            alen;
+       xfs_extlen_t            indlen;
+       char                    rt = XFS_IS_REALTIME_INODE(ip);
+       xfs_extlen_t            extsz;
+       int                     error;
 
-#ifdef DEBUG
-       xfs_fileoff_t           orig_bno;       /* original block number value */
-       int                     orig_flags;     /* original flags arg value */
-       xfs_filblks_t           orig_len;       /* original value of len arg */
-       struct xfs_bmbt_irec    *orig_mval;     /* original value of mval */
-       int                     orig_nmap;      /* original value of *nmap */
+       alen = XFS_FILBLKS_MIN(len, MAXEXTLEN);
+       if (!eof)
+               alen = XFS_FILBLKS_MIN(alen, got->br_startoff - aoff);
 
-       orig_bno = bno;
-       orig_len = len;
-       orig_flags = flags;
-       orig_mval = mval;
-       orig_nmap = *nmap;
-#endif
+       /* Figure out the extent size, adjust alen */
+       extsz = xfs_get_extsz_hint(ip);
+       if (extsz) {
+               /*
+                * Make sure we don't exceed a single extent length when we
+                * align the extent by reducing length we are going to
+                * allocate by the maximum amount extent size aligment may
+                * require.
+                */
+               alen = XFS_FILBLKS_MIN(len, MAXEXTLEN - (2 * extsz - 1));
+               error = xfs_bmap_extsize_align(mp, got, prev, extsz, rt, eof,
+                                              1, 0, &aoff, &alen);
+               ASSERT(!error);
+       }
+
+       if (rt)
+               extsz = alen / mp->m_sb.sb_rextsize;
+
+       /*
+        * Make a transaction-less quota reservation for delayed allocation
+        * blocks.  This number gets adjusted later.  We return if we haven't
+        * allocated blocks already inside this loop.
+        */
+       error = xfs_trans_reserve_quota_nblks(NULL, ip, (long)alen, 0,
+                       rt ? XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS);
+       if (error)
+               return error;
+
+       /*
+        * Split changing sb for alen and indlen since they could be coming
+        * from different places.
+        */
+       indlen = (xfs_extlen_t)xfs_bmap_worst_indlen(ip, alen);
+       ASSERT(indlen > 0);
+
+       if (rt) {
+               error = xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS,
+                                         -((int64_t)extsz), 0);
+       } else {
+               error = xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
+                                                -((int64_t)alen), 0);
+       }
+
+       if (error)
+               goto out_unreserve_quota;
+
+       error = xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
+                                        -((int64_t)indlen), 0);
+       if (error)
+               goto out_unreserve_blocks;
+
+
+       ip->i_delayed_blks += alen;
+
+       got->br_startoff = aoff;
+       got->br_startblock = nullstartblock(indlen);
+       got->br_blockcount = alen;
+       got->br_state = XFS_EXT_NORM;
+       xfs_bmap_add_extent_hole_delay(ip, lastx, got);
+
+       /*
+        * Update our extent pointer, given that xfs_bmap_add_extent_hole_delay
+        * might have merged it into one of the neighbouring ones.
+        */
+       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, *lastx), got);
+
+       ASSERT(got->br_startoff <= aoff);
+       ASSERT(got->br_startoff + got->br_blockcount >= aoff + alen);
+       ASSERT(isnullstartblock(got->br_startblock));
+       ASSERT(got->br_state == XFS_EXT_NORM);
+       return 0;
+
+out_unreserve_blocks:
+       if (rt)
+               xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS, extsz, 0);
+       else
+               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS, alen, 0);
+out_unreserve_quota:
+       if (XFS_IS_QUOTA_ON(mp))
+               xfs_trans_unreserve_quota_nblks(NULL, ip, (long)alen, 0, rt ?
+                               XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS);
+       return error;
+}
+
+/*
+ * Map file blocks to filesystem blocks, adding delayed allocations as needed.
+ */
+int
+xfs_bmapi_delay(
+       struct xfs_inode        *ip,    /* incore inode */
+       xfs_fileoff_t           bno,    /* starting file offs. mapped */
+       xfs_filblks_t           len,    /* length to map in file */
+       struct xfs_bmbt_irec    *mval,  /* output: map values */
+       int                     *nmap,  /* i/o: mval size/count */
+       int                     flags)  /* XFS_BMAPI_... */
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       struct xfs_ifork        *ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
+       struct xfs_bmbt_irec    got;    /* current file extent record */
+       struct xfs_bmbt_irec    prev;   /* previous file extent record */
+       xfs_fileoff_t           obno;   /* old block number (offset) */
+       xfs_fileoff_t           end;    /* end of mapped file region */
+       xfs_extnum_t            lastx;  /* last useful extent number */
+       int                     eof;    /* we've hit the end of extents */
+       int                     n = 0;  /* current extent index */
+       int                     error = 0;
 
        ASSERT(*nmap >= 1);
        ASSERT(*nmap <= XFS_BMAP_MAX_NMAP);
-       ASSERT(!(flags & XFS_BMAPI_IGSTATE));
-       ASSERT(tp != NULL);
-       ASSERT(len > 0);
-
-       whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
-               XFS_ATTR_FORK : XFS_DATA_FORK;
+       ASSERT(!(flags & ~XFS_BMAPI_ENTIRE));
 
        if (unlikely(XFS_TEST_ERROR(
-           (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
-            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
-            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL),
+           (XFS_IFORK_FORMAT(ip, XFS_DATA_FORK) != XFS_DINODE_FMT_EXTENTS &&
+            XFS_IFORK_FORMAT(ip, XFS_DATA_FORK) != XFS_DINODE_FMT_BTREE),
             mp, XFS_ERRTAG_BMAPIFORMAT, XFS_RANDOM_BMAPIFORMAT))) {
-               XFS_ERROR_REPORT("xfs_bmapi_write", XFS_ERRLEVEL_LOW, mp);
+               XFS_ERROR_REPORT("xfs_bmapi_delay", XFS_ERRLEVEL_LOW, mp);
                return XFS_ERROR(EFSCORRUPTED);
        }
 
        if (XFS_FORCED_SHUTDOWN(mp))
                return XFS_ERROR(EIO);
 
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-
        XFS_STATS_INC(xs_blk_mapw);
 
-       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
-               /*
-                * XXX (dgc): This assumes we are only called for inodes that
-                * contain content neutral data in local format. Anything that
-                * contains caller-specific data in local format that needs
-                * transformation to move to a block format needs to do the
-                * conversion to extent format itself.
-                *
-                * Directory data forks and attribute forks handle this
-                * themselves, but with the addition of metadata verifiers every
-                * data fork in local format now contains caller specific data
-                * and as such conversion through this function is likely to be
-                * broken.
-                *
-                * The only likely user of this branch is for remote symlinks,
-                * but we cannot overwrite the data fork contents of the symlink
-                * (EEXIST occurs higher up the stack) and so it will never go
-                * from local format to extent format here. Hence I don't think
-                * this branch is ever executed intentionally and we should
-                * consider removing it and asserting that xfs_bmapi_write()
-                * cannot be called directly on local format forks. i.e. callers
-                * are completely responsible for local to extent format
-                * conversion, not xfs_bmapi_write().
-                */
-               error = xfs_bmap_local_to_extents(tp, ip, firstblock, total,
-                                       &bma.logflags, whichfork,
-                                       xfs_bmap_local_to_extents_init_fn);
-               if (error)
-                       goto error0;
-       }
-
-       if (*firstblock == NULLFSBLOCK) {
-               if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE)
-                       bma.minleft = be16_to_cpu(ifp->if_broot->bb_level) + 1;
-               else
-                       bma.minleft = 1;
-       } else {
-               bma.minleft = 0;
-       }
-
        if (!(ifp->if_flags & XFS_IFEXTENTS)) {
-               error = xfs_iread_extents(tp, ip, whichfork);
+               error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
                if (error)
-                       goto error0;
+                       return error;
        }
 
-       xfs_bmap_search_extents(ip, bno, whichfork, &eof, &bma.idx, &bma.got,
-                               &bma.prev);
-       n = 0;
+       xfs_bmap_search_extents(ip, bno, XFS_DATA_FORK, &eof, &lastx, &got, &prev);
        end = bno + len;
        obno = bno;
 
-       bma.tp = tp;
-       bma.ip = ip;
-       bma.total = total;
-       bma.userdata = 0;
-       bma.flist = flist;
-       bma.firstblock = firstblock;
-
-       if (flags & XFS_BMAPI_STACK_SWITCH)
-               bma.stack_switch = 1;
-
        while (bno < end && n < *nmap) {
-               inhole = eof || bma.got.br_startoff > bno;
-               wasdelay = !inhole && isnullstartblock(bma.got.br_startblock);
+               if (eof || got.br_startoff > bno) {
+                       error = xfs_bmapi_reserve_delalloc(ip, bno, len, &got,
+                                                          &prev, &lastx, eof);
+                       if (error) {
+                               if (n == 0) {
+                                       *nmap = 0;
+                                       return error;
+                               }
+                               break;
+                       }
+               }
 
-               /*
-                * First, deal with the hole before the allocated space
-                * that we found, if any.
-                */
-               if (inhole || wasdelay) {
-                       bma.eof = eof;
-                       bma.conv = !!(flags & XFS_BMAPI_CONVERT);
-                       bma.wasdel = wasdelay;
-                       bma.offset = bno;
-                       bma.flags = flags;
-
-                       /*
-                        * There's a 32/64 bit type mismatch between the
-                        * allocation length request (which can be 64 bits in
-                        * length) and the bma length request, which is
-                        * xfs_extlen_t and therefore 32 bits. Hence we have to
-                        * check for 32-bit overflows and handle them here.
-                        */
-                       if (len > (xfs_filblks_t)MAXEXTLEN)
-                               bma.length = MAXEXTLEN;
-                       else
-                               bma.length = len;
-
-                       ASSERT(len > 0);
-                       ASSERT(bma.length > 0);
-                       error = xfs_bmapi_allocate(&bma);
-                       if (error)
-                               goto error0;
-                       if (bma.blkno == NULLFSBLOCK)
-                               break;
-               }
-
-               /* Deal with the allocated space we found.  */
-               xfs_bmapi_trim_map(mval, &bma.got, &bno, len, obno,
-                                                       end, n, flags);
-
-               /* Execute unwritten extent conversion if necessary */
-               error = xfs_bmapi_convert_unwritten(&bma, mval, len, flags);
-               if (error == EAGAIN)
-                       continue;
-               if (error)
-                       goto error0;
-
-               /* update the extent map to return */
+               /* set up the extent map to return. */
+               xfs_bmapi_trim_map(mval, &got, &bno, len, obno, end, n, flags);
                xfs_bmapi_update_map(&mval, &bno, &len, obno, end, &n, flags);
 
-               /*
-                * If we're done, stop now.  Stop when we've allocated
-                * XFS_BMAP_MAX_NMAP extents no matter what.  Otherwise
-                * the transaction may get too big.
-                */
-               if (bno >= end || n >= *nmap || bma.nallocs >= *nmap)
+               /* If we're done, stop now. */
+               if (bno >= end || n >= *nmap)
                        break;
 
                /* Else go on to the next record. */
-               bma.prev = bma.got;
-               if (++bma.idx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t)) {
-                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma.idx),
-                                        &bma.got);
-               } else
+               prev = got;
+               if (++lastx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
+                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, lastx), &got);
+               else
                        eof = 1;
        }
+
        *nmap = n;
+       return 0;
+}
 
-       /*
-        * Transform from btree to extents, give it cur.
-        */
-       if (xfs_bmap_wants_extents(ip, whichfork)) {
-               int             tmp_logflags = 0;
 
-               ASSERT(bma.cur);
-               error = xfs_bmap_btree_to_extents(tp, ip, bma.cur,
-                       &tmp_logflags, whichfork);
-               bma.logflags |= tmp_logflags;
-               if (error)
-                       goto error0;
-       }
+STATIC int
+__xfs_bmapi_allocate(
+       struct xfs_bmalloca     *bma)
+{
+       struct xfs_mount        *mp = bma->ip->i_mount;
+       int                     whichfork = (bma->flags & XFS_BMAPI_ATTRFORK) ?
+                                               XFS_ATTR_FORK : XFS_DATA_FORK;
+       struct xfs_ifork        *ifp = XFS_IFORK_PTR(bma->ip, whichfork);
+       int                     tmp_logflags = 0;
+       int                     error;
+       int                     rt;
+
+       ASSERT(bma->length > 0);
+
+       rt = (whichfork == XFS_DATA_FORK) && XFS_IS_REALTIME_INODE(bma->ip);
 
-       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE ||
-              XFS_IFORK_NEXTENTS(ip, whichfork) >
-               XFS_IFORK_MAXEXT(ip, whichfork));
-       error = 0;
-error0:
-       /*
-        * Log everything.  Do this after conversion, there's no point in
-        * logging the extent records if we've converted to btree format.
-        */
-       if ((bma.logflags & xfs_ilog_fext(whichfork)) &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
-               bma.logflags &= ~xfs_ilog_fext(whichfork);
-       else if ((bma.logflags & xfs_ilog_fbroot(whichfork)) &&
-                XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)
-               bma.logflags &= ~xfs_ilog_fbroot(whichfork);
        /*
-        * Log whatever the flags say, even if error.  Otherwise we might miss
-        * detecting a case where the data is changed, there's an error,
-        * and it's not logged so we don't shutdown when we should.
+        * For the wasdelay case, we could also just allocate the stuff asked
+        * for in this bmap call but that wouldn't be as good.
         */
-       if (bma.logflags)
-               xfs_trans_log_inode(tp, ip, bma.logflags);
-
-       if (bma.cur) {
-               if (!error) {
-                       ASSERT(*firstblock == NULLFSBLOCK ||
-                              XFS_FSB_TO_AGNO(mp, *firstblock) ==
-                              XFS_FSB_TO_AGNO(mp,
-                                      bma.cur->bc_private.b.firstblock) ||
-                              (flist->xbf_low &&
-                               XFS_FSB_TO_AGNO(mp, *firstblock) <
-                               XFS_FSB_TO_AGNO(mp,
-                                       bma.cur->bc_private.b.firstblock)));
-                       *firstblock = bma.cur->bc_private.b.firstblock;
+       if (bma->wasdel) {
+               bma->length = (xfs_extlen_t)bma->got.br_blockcount;
+               bma->offset = bma->got.br_startoff;
+               if (bma->idx != NULLEXTNUM && bma->idx) {
+                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx - 1),
+                                        &bma->prev);
                }
-               xfs_btree_del_cursor(bma.cur,
-                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       } else {
+               bma->length = XFS_FILBLKS_MIN(bma->length, MAXEXTLEN);
+               if (!bma->eof)
+                       bma->length = XFS_FILBLKS_MIN(bma->length,
+                                       bma->got.br_startoff - bma->offset);
        }
-       if (!error)
-               xfs_bmap_validate_ret(orig_bno, orig_len, orig_flags, orig_mval,
-                       orig_nmap, *nmap);
-       return error;
-}
 
-/*
- * Unmap (remove) blocks from a file.
- * If nexts is nonzero then the number of extents to remove is limited to
- * that value.  If not all extents in the block range can be removed then
- * *done is set.
- */
-int                                            /* error */
-xfs_bunmapi(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       struct xfs_inode        *ip,            /* incore inode */
-       xfs_fileoff_t           bno,            /* starting offset to unmap */
-       xfs_filblks_t           len,            /* length to unmap in file */
-       int                     flags,          /* misc flags */
-       xfs_extnum_t            nexts,          /* number of extents max */
-       xfs_fsblock_t           *firstblock,    /* first allocated block
-                                                  controls a.g. for allocs */
-       xfs_bmap_free_t         *flist,         /* i/o: list extents to free */
-       int                     *done)          /* set if not done yet */
-{
-       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
-       xfs_bmbt_irec_t         del;            /* extent being deleted */
-       int                     eof;            /* is deleting at eof */
-       xfs_bmbt_rec_host_t     *ep;            /* extent record pointer */
-       int                     error;          /* error return value */
-       xfs_extnum_t            extno;          /* extent number in list */
-       xfs_bmbt_irec_t         got;            /* current extent record */
-       xfs_ifork_t             *ifp;           /* inode fork pointer */
-       int                     isrt;           /* freeing in rt area */
-       xfs_extnum_t            lastx;          /* last extent index used */
-       int                     logflags;       /* transaction logging flags */
-       xfs_extlen_t            mod;            /* rt extent offset */
-       xfs_mount_t             *mp;            /* mount structure */
-       xfs_extnum_t            nextents;       /* number of file extents */
-       xfs_bmbt_irec_t         prev;           /* previous extent record */
-       xfs_fileoff_t           start;          /* first file offset deleted */
-       int                     tmp_logflags;   /* partial logging flags */
-       int                     wasdel;         /* was a delayed alloc extent */
-       int                     whichfork;      /* data or attribute fork */
-       xfs_fsblock_t           sum;
+       /*
+        * Indicate if this is the first user data in the file, or just any
+        * user data.
+        */
+       if (!(bma->flags & XFS_BMAPI_METADATA)) {
+               bma->userdata = (bma->offset == 0) ?
+                       XFS_ALLOC_INITIAL_USER_DATA : XFS_ALLOC_USERDATA;
+       }
 
-       trace_xfs_bunmap(ip, bno, len, flags, _RET_IP_);
+       bma->minlen = (bma->flags & XFS_BMAPI_CONTIG) ? bma->length : 1;
 
-       whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
-               XFS_ATTR_FORK : XFS_DATA_FORK;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       if (unlikely(
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)) {
-               XFS_ERROR_REPORT("xfs_bunmapi", XFS_ERRLEVEL_LOW,
-                                ip->i_mount);
-               return XFS_ERROR(EFSCORRUPTED);
+       /*
+        * Only want to do the alignment at the eof if it is userdata and
+        * allocation length is larger than a stripe unit.
+        */
+       if (mp->m_dalign && bma->length >= mp->m_dalign &&
+           !(bma->flags & XFS_BMAPI_METADATA) && whichfork == XFS_DATA_FORK) {
+               error = xfs_bmap_isaeof(bma, whichfork);
+               if (error)
+                       return error;
        }
-       mp = ip->i_mount;
-       if (XFS_FORCED_SHUTDOWN(mp))
-               return XFS_ERROR(EIO);
-
-       ASSERT(len > 0);
-       ASSERT(nexts >= 0);
 
-       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
-           (error = xfs_iread_extents(tp, ip, whichfork)))
+       error = xfs_bmap_alloc(bma);
+       if (error)
                return error;
-       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
-       if (nextents == 0) {
-               *done = 1;
+
+       if (bma->flist->xbf_low)
+               bma->minleft = 0;
+       if (bma->cur)
+               bma->cur->bc_private.b.firstblock = *bma->firstblock;
+       if (bma->blkno == NULLFSBLOCK)
                return 0;
+       if ((ifp->if_flags & XFS_IFBROOT) && !bma->cur) {
+               bma->cur = xfs_bmbt_init_cursor(mp, bma->tp, bma->ip, whichfork);
+               bma->cur->bc_private.b.firstblock = *bma->firstblock;
+               bma->cur->bc_private.b.flist = bma->flist;
        }
-       XFS_STATS_INC(xs_blk_unmap);
-       isrt = (whichfork == XFS_DATA_FORK) && XFS_IS_REALTIME_INODE(ip);
-       start = bno;
-       bno = start + len - 1;
-       ep = xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got,
-               &prev);
+       /*
+        * Bump the number of extents we've allocated
+        * in this call.
+        */
+       bma->nallocs++;
+
+       if (bma->cur)
+               bma->cur->bc_private.b.flags =
+                       bma->wasdel ? XFS_BTCUR_BPRV_WASDEL : 0;
+
+       bma->got.br_startoff = bma->offset;
+       bma->got.br_startblock = bma->blkno;
+       bma->got.br_blockcount = bma->length;
+       bma->got.br_state = XFS_EXT_NORM;
 
        /*
-        * Check to see if the given block number is past the end of the
-        * file, back up to the last block if so...
+        * A wasdelay extent has been initialized, so shouldn't be flagged
+        * as unwritten.
         */
-       if (eof) {
-               ep = xfs_iext_get_ext(ifp, --lastx);
-               xfs_bmbt_get_all(ep, &got);
-               bno = got.br_startoff + got.br_blockcount - 1;
+       if (!bma->wasdel && (bma->flags & XFS_BMAPI_PREALLOC) &&
+           xfs_sb_version_hasextflgbit(&mp->m_sb))
+               bma->got.br_state = XFS_EXT_UNWRITTEN;
+
+       if (bma->wasdel)
+               error = xfs_bmap_add_extent_delay_real(bma);
+       else
+               error = xfs_bmap_add_extent_hole_real(bma, whichfork);
+
+       bma->logflags |= tmp_logflags;
+       if (error)
+               return error;
+
+       /*
+        * Update our extent pointer, given that xfs_bmap_add_extent_delay_real
+        * or xfs_bmap_add_extent_hole_real might have merged it into one of
+        * the neighbouring ones.
+        */
+       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &bma->got);
+
+       ASSERT(bma->got.br_startoff <= bma->offset);
+       ASSERT(bma->got.br_startoff + bma->got.br_blockcount >=
+              bma->offset + bma->length);
+       ASSERT(bma->got.br_state == XFS_EXT_NORM ||
+              bma->got.br_state == XFS_EXT_UNWRITTEN);
+       return 0;
+}
+
+static void
+xfs_bmapi_allocate_worker(
+       struct work_struct      *work)
+{
+       struct xfs_bmalloca     *args = container_of(work,
+                                               struct xfs_bmalloca, work);
+       unsigned long           pflags;
+
+       /* we are in a transaction context here */
+       current_set_flags_nested(&pflags, PF_FSTRANS);
+
+       args->result = __xfs_bmapi_allocate(args);
+       complete(args->done);
+
+       current_restore_flags_nested(&pflags, PF_FSTRANS);
+}
+
+/*
+ * Some allocation requests often come in with little stack to work on. Push
+ * them off to a worker thread so there is lots of stack to use. Otherwise just
+ * call directly to avoid the context switch overhead here.
+ */
+int
+xfs_bmapi_allocate(
+       struct xfs_bmalloca     *args)
+{
+       DECLARE_COMPLETION_ONSTACK(done);
+
+       if (!args->stack_switch)
+               return __xfs_bmapi_allocate(args);
+
+
+       args->done = &done;
+       INIT_WORK_ONSTACK(&args->work, xfs_bmapi_allocate_worker);
+       queue_work(xfs_alloc_wq, &args->work);
+       wait_for_completion(&done);
+       return args->result;
+}
+
+STATIC int
+xfs_bmapi_convert_unwritten(
+       struct xfs_bmalloca     *bma,
+       struct xfs_bmbt_irec    *mval,
+       xfs_filblks_t           len,
+       int                     flags)
+{
+       int                     whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
+                                               XFS_ATTR_FORK : XFS_DATA_FORK;
+       struct xfs_ifork        *ifp = XFS_IFORK_PTR(bma->ip, whichfork);
+       int                     tmp_logflags = 0;
+       int                     error;
+
+       /* check if we need to do unwritten->real conversion */
+       if (mval->br_state == XFS_EXT_UNWRITTEN &&
+           (flags & XFS_BMAPI_PREALLOC))
+               return 0;
+
+       /* check if we need to do real->unwritten conversion */
+       if (mval->br_state == XFS_EXT_NORM &&
+           (flags & (XFS_BMAPI_PREALLOC | XFS_BMAPI_CONVERT)) !=
+                       (XFS_BMAPI_PREALLOC | XFS_BMAPI_CONVERT))
+               return 0;
+
+       /*
+        * Modify (by adding) the state flag, if writing.
+        */
+       ASSERT(mval->br_blockcount <= len);
+       if ((ifp->if_flags & XFS_IFBROOT) && !bma->cur) {
+               bma->cur = xfs_bmbt_init_cursor(bma->ip->i_mount, bma->tp,
+                                       bma->ip, whichfork);
+               bma->cur->bc_private.b.firstblock = *bma->firstblock;
+               bma->cur->bc_private.b.flist = bma->flist;
        }
-       logflags = 0;
-       if (ifp->if_flags & XFS_IFBROOT) {
-               ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE);
-               cur = xfs_bmbt_init_cursor(mp, tp, ip, whichfork);
-               cur->bc_private.b.firstblock = *firstblock;
-               cur->bc_private.b.flist = flist;
-               cur->bc_private.b.flags = 0;
-       } else
-               cur = NULL;
+       mval->br_state = (mval->br_state == XFS_EXT_UNWRITTEN)
+                               ? XFS_EXT_NORM : XFS_EXT_UNWRITTEN;
 
-       if (isrt) {
-               /*
-                * Synchronize by locking the bitmap inode.
-                */
-               xfs_ilock(mp->m_rbmip, XFS_ILOCK_EXCL);
-               xfs_trans_ijoin(tp, mp->m_rbmip, XFS_ILOCK_EXCL);
+       error = xfs_bmap_add_extent_unwritten_real(bma->tp, bma->ip, &bma->idx,
+                       &bma->cur, mval, bma->firstblock, bma->flist,
+                       &tmp_logflags);
+       bma->logflags |= tmp_logflags;
+       if (error)
+               return error;
+
+       /*
+        * Update our extent pointer, given that
+        * xfs_bmap_add_extent_unwritten_real might have merged it into one
+        * of the neighbouring ones.
+        */
+       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma->idx), &bma->got);
+
+       /*
+        * We may have combined previously unwritten space with written space,
+        * so generate another request.
+        */
+       if (mval->br_blockcount < len)
+               return EAGAIN;
+       return 0;
+}
+
+/*
+ * Map file blocks to filesystem blocks, and allocate blocks or convert the
+ * extent state if necessary.  Details behaviour is controlled by the flags
+ * parameter.  Only allocates blocks from a single allocation group, to avoid
+ * locking problems.
+ *
+ * The returned value in "firstblock" from the first call in a transaction
+ * must be remembered and presented to subsequent calls in "firstblock".
+ * An upper bound for the number of blocks to be allocated is supplied to
+ * the first call in "total"; if no allocation group has that many free
+ * blocks then the call will fail (return NULLFSBLOCK in "firstblock").
+ */
+int
+xfs_bmapi_write(
+       struct xfs_trans        *tp,            /* transaction pointer */
+       struct xfs_inode        *ip,            /* incore inode */
+       xfs_fileoff_t           bno,            /* starting file offs. mapped */
+       xfs_filblks_t           len,            /* length to map in file */
+       int                     flags,          /* XFS_BMAPI_... */
+       xfs_fsblock_t           *firstblock,    /* first allocated block
+                                                  controls a.g. for allocs */
+       xfs_extlen_t            total,          /* total blocks needed */
+       struct xfs_bmbt_irec    *mval,          /* output: map values */
+       int                     *nmap,          /* i/o: mval size/count */
+       struct xfs_bmap_free    *flist)         /* i/o: list extents to free */
+{
+       struct xfs_mount        *mp = ip->i_mount;
+       struct xfs_ifork        *ifp;
+       struct xfs_bmalloca     bma = { 0 };    /* args for xfs_bmap_alloc */
+       xfs_fileoff_t           end;            /* end of mapped file region */
+       int                     eof;            /* after the end of extents */
+       int                     error;          /* error return */
+       int                     n;              /* current extent index */
+       xfs_fileoff_t           obno;           /* old block number (offset) */
+       int                     whichfork;      /* data or attr fork */
+       char                    inhole;         /* current location is hole in file */
+       char                    wasdelay;       /* old extent was delayed */
+
+#ifdef DEBUG
+       xfs_fileoff_t           orig_bno;       /* original block number value */
+       int                     orig_flags;     /* original flags arg value */
+       xfs_filblks_t           orig_len;       /* original value of len arg */
+       struct xfs_bmbt_irec    *orig_mval;     /* original value of mval */
+       int                     orig_nmap;      /* original value of *nmap */
+
+       orig_bno = bno;
+       orig_len = len;
+       orig_flags = flags;
+       orig_mval = mval;
+       orig_nmap = *nmap;
+#endif
+
+       ASSERT(*nmap >= 1);
+       ASSERT(*nmap <= XFS_BMAP_MAX_NMAP);
+       ASSERT(!(flags & XFS_BMAPI_IGSTATE));
+       ASSERT(tp != NULL);
+       ASSERT(len > 0);
+
+       whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
+               XFS_ATTR_FORK : XFS_DATA_FORK;
+
+       if (unlikely(XFS_TEST_ERROR(
+           (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
+            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE &&
+            XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL),
+            mp, XFS_ERRTAG_BMAPIFORMAT, XFS_RANDOM_BMAPIFORMAT))) {
+               XFS_ERROR_REPORT("xfs_bmapi_write", XFS_ERRLEVEL_LOW, mp);
+               return XFS_ERROR(EFSCORRUPTED);
        }
 
-       extno = 0;
-       while (bno != (xfs_fileoff_t)-1 && bno >= start && lastx >= 0 &&
-              (nexts == 0 || extno < nexts)) {
-               /*
-                * Is the found extent after a hole in which bno lives?
-                * Just back up to the previous extent, if so.
-                */
-               if (got.br_startoff > bno) {
-                       if (--lastx < 0)
-                               break;
-                       ep = xfs_iext_get_ext(ifp, lastx);
-                       xfs_bmbt_get_all(ep, &got);
-               }
+       if (XFS_FORCED_SHUTDOWN(mp))
+               return XFS_ERROR(EIO);
+
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+
+       XFS_STATS_INC(xs_blk_mapw);
+
+       if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_LOCAL) {
                /*
-                * Is the last block of this extent before the range
-                * we're supposed to delete?  If so, we're done.
+                * XXX (dgc): This assumes we are only called for inodes that
+                * contain content neutral data in local format. Anything that
+                * contains caller-specific data in local format that needs
+                * transformation to move to a block format needs to do the
+                * conversion to extent format itself.
+                *
+                * Directory data forks and attribute forks handle this
+                * themselves, but with the addition of metadata verifiers every
+                * data fork in local format now contains caller specific data
+                * and as such conversion through this function is likely to be
+                * broken.
+                *
+                * The only likely user of this branch is for remote symlinks,
+                * but we cannot overwrite the data fork contents of the symlink
+                * (EEXIST occurs higher up the stack) and so it will never go
+                * from local format to extent format here. Hence I don't think
+                * this branch is ever executed intentionally and we should
+                * consider removing it and asserting that xfs_bmapi_write()
+                * cannot be called directly on local format forks. i.e. callers
+                * are completely responsible for local to extent format
+                * conversion, not xfs_bmapi_write().
                 */
-               bno = XFS_FILEOFF_MIN(bno,
-                       got.br_startoff + got.br_blockcount - 1);
-               if (bno < start)
-                       break;
+               error = xfs_bmap_local_to_extents(tp, ip, firstblock, total,
+                                       &bma.logflags, whichfork,
+                                       xfs_bmap_local_to_extents_init_fn);
+               if (error)
+                       goto error0;
+       }
+
+       if (*firstblock == NULLFSBLOCK) {
+               if (XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE)
+                       bma.minleft = be16_to_cpu(ifp->if_broot->bb_level) + 1;
+               else
+                       bma.minleft = 1;
+       } else {
+               bma.minleft = 0;
+       }
+
+       if (!(ifp->if_flags & XFS_IFEXTENTS)) {
+               error = xfs_iread_extents(tp, ip, whichfork);
+               if (error)
+                       goto error0;
+       }
+
+       xfs_bmap_search_extents(ip, bno, whichfork, &eof, &bma.idx, &bma.got,
+                               &bma.prev);
+       n = 0;
+       end = bno + len;
+       obno = bno;
+
+       bma.tp = tp;
+       bma.ip = ip;
+       bma.total = total;
+       bma.userdata = 0;
+       bma.flist = flist;
+       bma.firstblock = firstblock;
+
+       if (flags & XFS_BMAPI_STACK_SWITCH)
+               bma.stack_switch = 1;
+
+       while (bno < end && n < *nmap) {
+               inhole = eof || bma.got.br_startoff > bno;
+               wasdelay = !inhole && isnullstartblock(bma.got.br_startblock);
+
                /*
-                * Then deal with the (possibly delayed) allocated space
-                * we found.
+                * First, deal with the hole before the allocated space
+                * that we found, if any.
                 */
-               ASSERT(ep != NULL);
-               del = got;
-               wasdel = isnullstartblock(del.br_startblock);
-               if (got.br_startoff < start) {
-                       del.br_startoff = start;
-                       del.br_blockcount -= start - got.br_startoff;
-                       if (!wasdel)
-                               del.br_startblock += start - got.br_startoff;
-               }
-               if (del.br_startoff + del.br_blockcount > bno + 1)
-                       del.br_blockcount = bno + 1 - del.br_startoff;
-               sum = del.br_startblock + del.br_blockcount;
-               if (isrt &&
-                   (mod = do_mod(sum, mp->m_sb.sb_rextsize))) {
-                       /*
-                        * Realtime extent not lined up at the end.
-                        * The extent could have been split into written
-                        * and unwritten pieces, or we could just be
-                        * unmapping part of it.  But we can't really
-                        * get rid of part of a realtime extent.
-                        */
-                       if (del.br_state == XFS_EXT_UNWRITTEN ||
-                           !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
-                               /*
-                                * This piece is unwritten, or we're not
-                                * using unwritten extents.  Skip over it.
-                                */
-                               ASSERT(bno >= mod);
-                               bno -= mod > del.br_blockcount ?
-                                       del.br_blockcount : mod;
-                               if (bno < got.br_startoff) {
-                                       if (--lastx >= 0)
-                                               xfs_bmbt_get_all(xfs_iext_get_ext(
-                                                       ifp, lastx), &got);
-                               }
-                               continue;
-                       }
-                       /*
-                        * It's written, turn it unwritten.
-                        * This is better than zeroing it.
-                        */
-                       ASSERT(del.br_state == XFS_EXT_NORM);
-                       ASSERT(xfs_trans_get_block_res(tp) > 0);
+               if (inhole || wasdelay) {
+                       bma.eof = eof;
+                       bma.conv = !!(flags & XFS_BMAPI_CONVERT);
+                       bma.wasdel = wasdelay;
+                       bma.offset = bno;
+                       bma.flags = flags;
+
                        /*
-                        * If this spans a realtime extent boundary,
-                        * chop it back to the start of the one we end at.
+                        * There's a 32/64 bit type mismatch between the
+                        * allocation length request (which can be 64 bits in
+                        * length) and the bma length request, which is
+                        * xfs_extlen_t and therefore 32 bits. Hence we have to
+                        * check for 32-bit overflows and handle them here.
                         */
-                       if (del.br_blockcount > mod) {
-                               del.br_startoff += del.br_blockcount - mod;
-                               del.br_startblock += del.br_blockcount - mod;
-                               del.br_blockcount = mod;
-                       }
-                       del.br_state = XFS_EXT_UNWRITTEN;
-                       error = xfs_bmap_add_extent_unwritten_real(tp, ip,
-                                       &lastx, &cur, &del, firstblock, flist,
-                                       &logflags);
+                       if (len > (xfs_filblks_t)MAXEXTLEN)
+                               bma.length = MAXEXTLEN;
+                       else
+                               bma.length = len;
+
+                       ASSERT(len > 0);
+                       ASSERT(bma.length > 0);
+                       error = xfs_bmapi_allocate(&bma);
                        if (error)
                                goto error0;
-                       goto nodelete;
+                       if (bma.blkno == NULLFSBLOCK)
+                               break;
                }
-               if (isrt && (mod = do_mod(del.br_startblock, mp->m_sb.sb_rextsize))) {
-                       /*
-                        * Realtime extent is lined up at the end but not
-                        * at the front.  We'll get rid of full extents if
-                        * we can.
-                        */
-                       mod = mp->m_sb.sb_rextsize - mod;
-                       if (del.br_blockcount > mod) {
-                               del.br_blockcount -= mod;
-                               del.br_startoff += mod;
-                               del.br_startblock += mod;
-                       } else if ((del.br_startoff == start &&
-                                   (del.br_state == XFS_EXT_UNWRITTEN ||
-                                    xfs_trans_get_block_res(tp) == 0)) ||
-                                  !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
-                               /*
-                                * Can't make it unwritten.  There isn't
-                                * a full extent here so just skip it.
-                                */
-                               ASSERT(bno >= del.br_blockcount);
-                               bno -= del.br_blockcount;
-                               if (got.br_startoff > bno) {
-                                       if (--lastx >= 0) {
-                                               ep = xfs_iext_get_ext(ifp,
-                                                                     lastx);
-                                               xfs_bmbt_get_all(ep, &got);
-                                       }
-                               }
-                               continue;
-                       } else if (del.br_state == XFS_EXT_UNWRITTEN) {
-                               /*
-                                * This one is already unwritten.
-                                * It must have a written left neighbor.
-                                * Unwrite the killed part of that one and
-                                * try again.
-                                */
-                               ASSERT(lastx > 0);
-                               xfs_bmbt_get_all(xfs_iext_get_ext(ifp,
-                                               lastx - 1), &prev);
-                               ASSERT(prev.br_state == XFS_EXT_NORM);
-                               ASSERT(!isnullstartblock(prev.br_startblock));
-                               ASSERT(del.br_startblock ==
-                                      prev.br_startblock + prev.br_blockcount);
-                               if (prev.br_startoff < start) {
-                                       mod = start - prev.br_startoff;
-                                       prev.br_blockcount -= mod;
-                                       prev.br_startblock += mod;
-                                       prev.br_startoff = start;
-                               }
-                               prev.br_state = XFS_EXT_UNWRITTEN;
-                               lastx--;
-                               error = xfs_bmap_add_extent_unwritten_real(tp,
-                                               ip, &lastx, &cur, &prev,
-                                               firstblock, flist, &logflags);
-                               if (error)
-                                       goto error0;
-                               goto nodelete;
-                       } else {
-                               ASSERT(del.br_state == XFS_EXT_NORM);
-                               del.br_state = XFS_EXT_UNWRITTEN;
-                               error = xfs_bmap_add_extent_unwritten_real(tp,
-                                               ip, &lastx, &cur, &del,
-                                               firstblock, flist, &logflags);
-                               if (error)
-                                       goto error0;
-                               goto nodelete;
-                       }
+
+               /* Deal with the allocated space we found.  */
+               xfs_bmapi_trim_map(mval, &bma.got, &bno, len, obno,
+                                                       end, n, flags);
+
+               /* Execute unwritten extent conversion if necessary */
+               error = xfs_bmapi_convert_unwritten(&bma, mval, len, flags);
+               if (error == EAGAIN)
+                       continue;
+               if (error)
+                       goto error0;
+
+               /* update the extent map to return */
+               xfs_bmapi_update_map(&mval, &bno, &len, obno, end, &n, flags);
+
+               /*
+                * If we're done, stop now.  Stop when we've allocated
+                * XFS_BMAP_MAX_NMAP extents no matter what.  Otherwise
+                * the transaction may get too big.
+                */
+               if (bno >= end || n >= *nmap || bma.nallocs >= *nmap)
+                       break;
+
+               /* Else go on to the next record. */
+               bma.prev = bma.got;
+               if (++bma.idx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t)) {
+                       xfs_bmbt_get_all(xfs_iext_get_ext(ifp, bma.idx),
+                                        &bma.got);
+               } else
+                       eof = 1;
+       }
+       *nmap = n;
+
+       /*
+        * Transform from btree to extents, give it cur.
+        */
+       if (xfs_bmap_wants_extents(ip, whichfork)) {
+               int             tmp_logflags = 0;
+
+               ASSERT(bma.cur);
+               error = xfs_bmap_btree_to_extents(tp, ip, bma.cur,
+                       &tmp_logflags, whichfork);
+               bma.logflags |= tmp_logflags;
+               if (error)
+                       goto error0;
+       }
+
+       ASSERT(XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE ||
+              XFS_IFORK_NEXTENTS(ip, whichfork) >
+               XFS_IFORK_MAXEXT(ip, whichfork));
+       error = 0;
+error0:
+       /*
+        * Log everything.  Do this after conversion, there's no point in
+        * logging the extent records if we've converted to btree format.
+        */
+       if ((bma.logflags & xfs_ilog_fext(whichfork)) &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
+               bma.logflags &= ~xfs_ilog_fext(whichfork);
+       else if ((bma.logflags & xfs_ilog_fbroot(whichfork)) &&
+                XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)
+               bma.logflags &= ~xfs_ilog_fbroot(whichfork);
+       /*
+        * Log whatever the flags say, even if error.  Otherwise we might miss
+        * detecting a case where the data is changed, there's an error,
+        * and it's not logged so we don't shutdown when we should.
+        */
+       if (bma.logflags)
+               xfs_trans_log_inode(tp, ip, bma.logflags);
+
+       if (bma.cur) {
+               if (!error) {
+                       ASSERT(*firstblock == NULLFSBLOCK ||
+                              XFS_FSB_TO_AGNO(mp, *firstblock) ==
+                              XFS_FSB_TO_AGNO(mp,
+                                      bma.cur->bc_private.b.firstblock) ||
+                              (flist->xbf_low &&
+                               XFS_FSB_TO_AGNO(mp, *firstblock) <
+                               XFS_FSB_TO_AGNO(mp,
+                                       bma.cur->bc_private.b.firstblock)));
+                       *firstblock = bma.cur->bc_private.b.firstblock;
                }
-               if (wasdel) {
-                       ASSERT(startblockval(del.br_startblock) > 0);
-                       /* Update realtime/data freespace, unreserve quota */
-                       if (isrt) {
-                               xfs_filblks_t rtexts;
+               xfs_btree_del_cursor(bma.cur,
+                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       }
+       if (!error)
+               xfs_bmap_validate_ret(orig_bno, orig_len, orig_flags, orig_mval,
+                       orig_nmap, *nmap);
+       return error;
+}
 
-                               rtexts = XFS_FSB_TO_B(mp, del.br_blockcount);
-                               do_div(rtexts, mp->m_sb.sb_rextsize);
-                               xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS,
-                                               (int64_t)rtexts, 0);
-                               (void)xfs_trans_reserve_quota_nblks(NULL,
-                                       ip, -((long)del.br_blockcount), 0,
-                                       XFS_QMOPT_RES_RTBLKS);
-                       } else {
-                               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
-                                               (int64_t)del.br_blockcount, 0);
-                               (void)xfs_trans_reserve_quota_nblks(NULL,
-                                       ip, -((long)del.br_blockcount), 0,
-                                       XFS_QMOPT_RES_REGBLKS);
-                       }
-                       ip->i_delayed_blks -= del.br_blockcount;
-                       if (cur)
-                               cur->bc_private.b.flags |=
-                                       XFS_BTCUR_BPRV_WASDEL;
-               } else if (cur)
-                       cur->bc_private.b.flags &= ~XFS_BTCUR_BPRV_WASDEL;
+/*
+ * Called by xfs_bmapi to update file extent records and the btree
+ * after removing space (or undoing a delayed allocation).
+ */
+STATIC int                             /* error */
+xfs_bmap_del_extent(
+       xfs_inode_t             *ip,    /* incore inode pointer */
+       xfs_trans_t             *tp,    /* current transaction pointer */
+       xfs_extnum_t            *idx,   /* extent number to update/delete */
+       xfs_bmap_free_t         *flist, /* list of extents to be freed */
+       xfs_btree_cur_t         *cur,   /* if null, not a btree */
+       xfs_bmbt_irec_t         *del,   /* data to remove from extents */
+       int                     *logflagsp, /* inode logging flags */
+       int                     whichfork) /* data or attr fork */
+{
+       xfs_filblks_t           da_new; /* new delay-alloc indirect blocks */
+       xfs_filblks_t           da_old; /* old delay-alloc indirect blocks */
+       xfs_fsblock_t           del_endblock=0; /* first block past del */
+       xfs_fileoff_t           del_endoff;     /* first offset past del */
+       int                     delay;  /* current block is delayed allocated */
+       int                     do_fx;  /* free extent at end of routine */
+       xfs_bmbt_rec_host_t     *ep;    /* current extent entry pointer */
+       int                     error;  /* error return value */
+       int                     flags;  /* inode logging flags */
+       xfs_bmbt_irec_t         got;    /* current extent entry */
+       xfs_fileoff_t           got_endoff;     /* first offset past got */
+       int                     i;      /* temp state */
+       xfs_ifork_t             *ifp;   /* inode fork pointer */
+       xfs_mount_t             *mp;    /* mount structure */
+       xfs_filblks_t           nblks;  /* quota/sb block count */
+       xfs_bmbt_irec_t         new;    /* new record to be inserted */
+       /* REFERENCED */
+       uint                    qfield; /* quota field to update */
+       xfs_filblks_t           temp;   /* for indirect length calculations */
+       xfs_filblks_t           temp2;  /* for indirect length calculations */
+       int                     state = 0;
+
+       XFS_STATS_INC(xs_del_exlist);
+
+       if (whichfork == XFS_ATTR_FORK)
+               state |= BMAP_ATTRFORK;
+
+       mp = ip->i_mount;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       ASSERT((*idx >= 0) && (*idx < ifp->if_bytes /
+               (uint)sizeof(xfs_bmbt_rec_t)));
+       ASSERT(del->br_blockcount > 0);
+       ep = xfs_iext_get_ext(ifp, *idx);
+       xfs_bmbt_get_all(ep, &got);
+       ASSERT(got.br_startoff <= del->br_startoff);
+       del_endoff = del->br_startoff + del->br_blockcount;
+       got_endoff = got.br_startoff + got.br_blockcount;
+       ASSERT(got_endoff >= del_endoff);
+       delay = isnullstartblock(got.br_startblock);
+       ASSERT(isnullstartblock(del->br_startblock) == delay);
+       flags = 0;
+       qfield = 0;
+       error = 0;
+       /*
+        * If deleting a real allocation, must free up the disk space.
+        */
+       if (!delay) {
+               flags = XFS_ILOG_CORE;
+               /*
+                * Realtime allocation.  Free it and record di_nblocks update.
+                */
+               if (whichfork == XFS_DATA_FORK && XFS_IS_REALTIME_INODE(ip)) {
+                       xfs_fsblock_t   bno;
+                       xfs_filblks_t   len;
+
+                       ASSERT(do_mod(del->br_blockcount,
+                                     mp->m_sb.sb_rextsize) == 0);
+                       ASSERT(do_mod(del->br_startblock,
+                                     mp->m_sb.sb_rextsize) == 0);
+                       bno = del->br_startblock;
+                       len = del->br_blockcount;
+                       do_div(bno, mp->m_sb.sb_rextsize);
+                       do_div(len, mp->m_sb.sb_rextsize);
+                       error = xfs_rtfree_extent(tp, bno, (xfs_extlen_t)len);
+                       if (error)
+                               goto done;
+                       do_fx = 0;
+                       nblks = len * mp->m_sb.sb_rextsize;
+                       qfield = XFS_TRANS_DQ_RTBCOUNT;
+               }
                /*
-                * If it's the case where the directory code is running
-                * with no block reservation, and the deleted block is in
-                * the middle of its extent, and the resulting insert
-                * of an extent would cause transformation to btree format,
-                * then reject it.  The calling code will then swap
-                * blocks around instead.
-                * We have to do this now, rather than waiting for the
-                * conversion to btree format, since the transaction
-                * will be dirty.
+                * Ordinary allocation.
                 */
-               if (!wasdel && xfs_trans_get_block_res(tp) == 0 &&
-                   XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS &&
-                   XFS_IFORK_NEXTENTS(ip, whichfork) >= /* Note the >= */
-                       XFS_IFORK_MAXEXT(ip, whichfork) &&
-                   del.br_startoff > got.br_startoff &&
-                   del.br_startoff + del.br_blockcount <
-                   got.br_startoff + got.br_blockcount) {
-                       error = XFS_ERROR(ENOSPC);
-                       goto error0;
+               else {
+                       do_fx = 1;
+                       nblks = del->br_blockcount;
+                       qfield = XFS_TRANS_DQ_BCOUNT;
                }
-               error = xfs_bmap_del_extent(ip, tp, &lastx, flist, cur, &del,
-                               &tmp_logflags, whichfork);
-               logflags |= tmp_logflags;
-               if (error)
-                       goto error0;
-               bno = del.br_startoff - 1;
-nodelete:
                /*
-                * If not done go on to the next (previous) record.
+                * Set up del_endblock and cur for later.
                 */
-               if (bno != (xfs_fileoff_t)-1 && bno >= start) {
-                       if (lastx >= 0) {
-                               ep = xfs_iext_get_ext(ifp, lastx);
-                               if (xfs_bmbt_get_startoff(ep) > bno) {
-                                       if (--lastx >= 0)
-                                               ep = xfs_iext_get_ext(ifp,
-                                                                     lastx);
-                               }
-                               xfs_bmbt_get_all(ep, &got);
-                       }
-                       extno++;
+               del_endblock = del->br_startblock + del->br_blockcount;
+               if (cur) {
+                       if ((error = xfs_bmbt_lookup_eq(cur, got.br_startoff,
+                                       got.br_startblock, got.br_blockcount,
+                                       &i)))
+                               goto done;
+                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
                }
-       }
-       *done = bno == (xfs_fileoff_t)-1 || bno < start || lastx < 0;
-
-       /*
-        * Convert to a btree if necessary.
-        */
-       if (xfs_bmap_needs_btree(ip, whichfork)) {
-               ASSERT(cur == NULL);
-               error = xfs_bmap_extents_to_btree(tp, ip, firstblock, flist,
-                       &cur, 0, &tmp_logflags, whichfork);
-               logflags |= tmp_logflags;
-               if (error)
-                       goto error0;
+               da_old = da_new = 0;
+       } else {
+               da_old = startblockval(got.br_startblock);
+               da_new = 0;
+               nblks = 0;
+               do_fx = 0;
        }
        /*
-        * transform from btree to extents, give it cur
+        * Set flag value to use in switch statement.
+        * Left-contig is 2, right-contig is 1.
         */
-       else if (xfs_bmap_wants_extents(ip, whichfork)) {
-               ASSERT(cur != NULL);
-               error = xfs_bmap_btree_to_extents(tp, ip, cur, &tmp_logflags,
-                       whichfork);
-               logflags |= tmp_logflags;
-               if (error)
-                       goto error0;
+       switch (((got.br_startoff == del->br_startoff) << 1) |
+               (got_endoff == del_endoff)) {
+       case 3:
+               /*
+                * Matches the whole extent.  Delete the entry.
+                */
+               xfs_iext_remove(ip, *idx, 1,
+                               whichfork == XFS_ATTR_FORK ? BMAP_ATTRFORK : 0);
+               --*idx;
+               if (delay)
+                       break;
+
+               XFS_IFORK_NEXT_SET(ip, whichfork,
+                       XFS_IFORK_NEXTENTS(ip, whichfork) - 1);
+               flags |= XFS_ILOG_CORE;
+               if (!cur) {
+                       flags |= xfs_ilog_fext(whichfork);
+                       break;
+               }
+               if ((error = xfs_btree_delete(cur, &i)))
+                       goto done;
+               XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+               break;
+
+       case 2:
+               /*
+                * Deleting the first part of the extent.
+                */
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_startoff(ep, del_endoff);
+               temp = got.br_blockcount - del->br_blockcount;
+               xfs_bmbt_set_blockcount(ep, temp);
+               if (delay) {
+                       temp = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(ip, temp),
+                               da_old);
+                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
+                       trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+                       da_new = temp;
+                       break;
+               }
+               xfs_bmbt_set_startblock(ep, del_endblock);
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               if (!cur) {
+                       flags |= xfs_ilog_fext(whichfork);
+                       break;
+               }
+               if ((error = xfs_bmbt_update(cur, del_endoff, del_endblock,
+                               got.br_blockcount - del->br_blockcount,
+                               got.br_state)))
+                       goto done;
+               break;
+
+       case 1:
+               /*
+                * Deleting the last part of the extent.
+                */
+               temp = got.br_blockcount - del->br_blockcount;
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep, temp);
+               if (delay) {
+                       temp = XFS_FILBLKS_MIN(xfs_bmap_worst_indlen(ip, temp),
+                               da_old);
+                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
+                       trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+                       da_new = temp;
+                       break;
+               }
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               if (!cur) {
+                       flags |= xfs_ilog_fext(whichfork);
+                       break;
+               }
+               if ((error = xfs_bmbt_update(cur, got.br_startoff,
+                               got.br_startblock,
+                               got.br_blockcount - del->br_blockcount,
+                               got.br_state)))
+                       goto done;
+               break;
+
+       case 0:
+               /*
+                * Deleting the middle of the extent.
+                */
+               temp = del->br_startoff - got.br_startoff;
+               trace_xfs_bmap_pre_update(ip, *idx, state, _THIS_IP_);
+               xfs_bmbt_set_blockcount(ep, temp);
+               new.br_startoff = del_endoff;
+               temp2 = got_endoff - del_endoff;
+               new.br_blockcount = temp2;
+               new.br_state = got.br_state;
+               if (!delay) {
+                       new.br_startblock = del_endblock;
+                       flags |= XFS_ILOG_CORE;
+                       if (cur) {
+                               if ((error = xfs_bmbt_update(cur,
+                                               got.br_startoff,
+                                               got.br_startblock, temp,
+                                               got.br_state)))
+                                       goto done;
+                               if ((error = xfs_btree_increment(cur, 0, &i)))
+                                       goto done;
+                               cur->bc_rec.b = new;
+                               error = xfs_btree_insert(cur, &i);
+                               if (error && error != ENOSPC)
+                                       goto done;
+                               /*
+                                * If get no-space back from btree insert,
+                                * it tried a split, and we have a zero
+                                * block reservation.
+                                * Fix up our state and return the error.
+                                */
+                               if (error == ENOSPC) {
+                                       /*
+                                        * Reset the cursor, don't trust
+                                        * it after any insert operation.
+                                        */
+                                       if ((error = xfs_bmbt_lookup_eq(cur,
+                                                       got.br_startoff,
+                                                       got.br_startblock,
+                                                       temp, &i)))
+                                               goto done;
+                                       XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                                       /*
+                                        * Update the btree record back
+                                        * to the original value.
+                                        */
+                                       if ((error = xfs_bmbt_update(cur,
+                                                       got.br_startoff,
+                                                       got.br_startblock,
+                                                       got.br_blockcount,
+                                                       got.br_state)))
+                                               goto done;
+                                       /*
+                                        * Reset the extent record back
+                                        * to the original value.
+                                        */
+                                       xfs_bmbt_set_blockcount(ep,
+                                               got.br_blockcount);
+                                       flags = 0;
+                                       error = XFS_ERROR(ENOSPC);
+                                       goto done;
+                               }
+                               XFS_WANT_CORRUPTED_GOTO(i == 1, done);
+                       } else
+                               flags |= xfs_ilog_fext(whichfork);
+                       XFS_IFORK_NEXT_SET(ip, whichfork,
+                               XFS_IFORK_NEXTENTS(ip, whichfork) + 1);
+               } else {
+                       ASSERT(whichfork == XFS_DATA_FORK);
+                       temp = xfs_bmap_worst_indlen(ip, temp);
+                       xfs_bmbt_set_startblock(ep, nullstartblock((int)temp));
+                       temp2 = xfs_bmap_worst_indlen(ip, temp2);
+                       new.br_startblock = nullstartblock((int)temp2);
+                       da_new = temp + temp2;
+                       while (da_new > da_old) {
+                               if (temp) {
+                                       temp--;
+                                       da_new--;
+                                       xfs_bmbt_set_startblock(ep,
+                                               nullstartblock((int)temp));
+                               }
+                               if (da_new == da_old)
+                                       break;
+                               if (temp2) {
+                                       temp2--;
+                                       da_new--;
+                                       new.br_startblock =
+                                               nullstartblock((int)temp2);
+                               }
+                       }
+               }
+               trace_xfs_bmap_post_update(ip, *idx, state, _THIS_IP_);
+               xfs_iext_insert(ip, *idx + 1, 1, &new, state);
+               ++*idx;
+               break;
        }
        /*
-        * transform from extents to local?
+        * If we need to, add to list of extents to delete.
         */
-       error = 0;
-error0:
+       if (do_fx)
+               xfs_bmap_add_free(del->br_startblock, del->br_blockcount, flist,
+                       mp);
        /*
-        * Log everything.  Do this after conversion, there's no point in
-        * logging the extent records if we've converted to btree format.
+        * Adjust inode # blocks in the file.
         */
-       if ((logflags & xfs_ilog_fext(whichfork)) &&
-           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
-               logflags &= ~xfs_ilog_fext(whichfork);
-       else if ((logflags & xfs_ilog_fbroot(whichfork)) &&
-                XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)
-               logflags &= ~xfs_ilog_fbroot(whichfork);
+       if (nblks)
+               ip->i_d.di_nblocks -= nblks;
        /*
-        * Log inode even in the error case, if the transaction
-        * is dirty we'll need to shut down the filesystem.
+        * Adjust quota data.
         */
-       if (logflags)
-               xfs_trans_log_inode(tp, ip, logflags);
-       if (cur) {
-               if (!error) {
-                       *firstblock = cur->bc_private.b.firstblock;
-                       cur->bc_private.b.allocated = 0;
-               }
-               xfs_btree_del_cursor(cur,
-                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       if (qfield)
+               xfs_trans_mod_dquot_byino(tp, ip, qfield, (long)-nblks);
+
+       /*
+        * Account for change in delayed indirect blocks.
+        * Nothing to do for disk quota accounting here.
+        */
+       ASSERT(da_old >= da_new);
+       if (da_old > da_new) {
+               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
+                       (int64_t)(da_old - da_new), 0);
        }
+done:
+       *logflagsp = flags;
        return error;
 }
 
 /*
- * returns 1 for success, 0 if we failed to map the extent.
+ * Unmap (remove) blocks from a file.
+ * If nexts is nonzero then the number of extents to remove is limited to
+ * that value.  If not all extents in the block range can be removed then
+ * *done is set.
  */
-STATIC int
-xfs_getbmapx_fix_eof_hole(
-       xfs_inode_t             *ip,            /* xfs incore inode pointer */
-       struct getbmapx         *out,           /* output structure */
-       int                     prealloced,     /* this is a file with
-                                                * preallocated data space */
-       __int64_t               end,            /* last block requested */
-       xfs_fsblock_t           startblock)
+int                                            /* error */
+xfs_bunmapi(
+       xfs_trans_t             *tp,            /* transaction pointer */
+       struct xfs_inode        *ip,            /* incore inode */
+       xfs_fileoff_t           bno,            /* starting offset to unmap */
+       xfs_filblks_t           len,            /* length to unmap in file */
+       int                     flags,          /* misc flags */
+       xfs_extnum_t            nexts,          /* number of extents max */
+       xfs_fsblock_t           *firstblock,    /* first allocated block
+                                                  controls a.g. for allocs */
+       xfs_bmap_free_t         *flist,         /* i/o: list extents to free */
+       int                     *done)          /* set if not done yet */
 {
-       __int64_t               fixlen;
-       xfs_mount_t             *mp;            /* file system mount point */
+       xfs_btree_cur_t         *cur;           /* bmap btree cursor */
+       xfs_bmbt_irec_t         del;            /* extent being deleted */
+       int                     eof;            /* is deleting at eof */
+       xfs_bmbt_rec_host_t     *ep;            /* extent record pointer */
+       int                     error;          /* error return value */
+       xfs_extnum_t            extno;          /* extent number in list */
+       xfs_bmbt_irec_t         got;            /* current extent record */
        xfs_ifork_t             *ifp;           /* inode fork pointer */
-       xfs_extnum_t            lastx;          /* last extent pointer */
-       xfs_fileoff_t           fileblock;
-
-       if (startblock == HOLESTARTBLOCK) {
-               mp = ip->i_mount;
-               out->bmv_block = -1;
-               fixlen = XFS_FSB_TO_BB(mp, XFS_B_TO_FSB(mp, XFS_ISIZE(ip)));
-               fixlen -= out->bmv_offset;
-               if (prealloced && out->bmv_offset + out->bmv_length == end) {
-                       /* Came to hole at EOF. Trim it. */
-                       if (fixlen <= 0)
-                               return 0;
-                       out->bmv_length = fixlen;
-               }
-       } else {
-               if (startblock == DELAYSTARTBLOCK)
-                       out->bmv_block = -2;
-               else
-                       out->bmv_block = xfs_fsb_to_db(ip, startblock);
-               fileblock = XFS_BB_TO_FSB(ip->i_mount, out->bmv_offset);
-               ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
-               if (xfs_iext_bno_to_ext(ifp, fileblock, &lastx) &&
-                  (lastx == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))-1))
-                       out->bmv_oflags |= BMV_OF_LAST;
-       }
-
-       return 1;
-}
-
-/*
- * Get inode's extents as described in bmv, and format for output.
- * Calls formatter to fill the user's buffer until all extents
- * are mapped, until the passed-in bmv->bmv_count slots have
- * been filled, or until the formatter short-circuits the loop,
- * if it is tracking filled-in extents on its own.
- */
-int                                            /* error code */
-xfs_getbmap(
-       xfs_inode_t             *ip,
-       struct getbmapx         *bmv,           /* user bmap structure */
-       xfs_bmap_format_t       formatter,      /* format to user */
-       void                    *arg)           /* formatter arg */
-{
-       __int64_t               bmvend;         /* last block requested */
-       int                     error = 0;      /* return value */
-       __int64_t               fixlen;         /* length for -1 case */
-       int                     i;              /* extent number */
-       int                     lock;           /* lock state */
-       xfs_bmbt_irec_t         *map;           /* buffer for user's data */
-       xfs_mount_t             *mp;            /* file system mount point */
-       int                     nex;            /* # of user extents can do */
-       int                     nexleft;        /* # of user extents left */
-       int                     subnex;         /* # of bmapi's can do */
-       int                     nmap;           /* number of map entries */
-       struct getbmapx         *out;           /* output structure */
-       int                     whichfork;      /* data or attr fork */
-       int                     prealloced;     /* this is a file with
-                                                * preallocated data space */
-       int                     iflags;         /* interface flags */
-       int                     bmapi_flags;    /* flags for xfs_bmapi */
-       int                     cur_ext = 0;
-
-       mp = ip->i_mount;
-       iflags = bmv->bmv_iflags;
-       whichfork = iflags & BMV_IF_ATTRFORK ? XFS_ATTR_FORK : XFS_DATA_FORK;
-
-       if (whichfork == XFS_ATTR_FORK) {
-               if (XFS_IFORK_Q(ip)) {
-                       if (ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS &&
-                           ip->i_d.di_aformat != XFS_DINODE_FMT_BTREE &&
-                           ip->i_d.di_aformat != XFS_DINODE_FMT_LOCAL)
-                               return XFS_ERROR(EINVAL);
-               } else if (unlikely(
-                          ip->i_d.di_aformat != 0 &&
-                          ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS)) {
-                       XFS_ERROR_REPORT("xfs_getbmap", XFS_ERRLEVEL_LOW,
-                                        ip->i_mount);
-                       return XFS_ERROR(EFSCORRUPTED);
-               }
-
-               prealloced = 0;
-               fixlen = 1LL << 32;
-       } else {
-               if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
-                   ip->i_d.di_format != XFS_DINODE_FMT_BTREE &&
-                   ip->i_d.di_format != XFS_DINODE_FMT_LOCAL)
-                       return XFS_ERROR(EINVAL);
-
-               if (xfs_get_extsz_hint(ip) ||
-                   ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC|XFS_DIFLAG_APPEND)){
-                       prealloced = 1;
-                       fixlen = mp->m_super->s_maxbytes;
-               } else {
-                       prealloced = 0;
-                       fixlen = XFS_ISIZE(ip);
-               }
-       }
-
-       if (bmv->bmv_length == -1) {
-               fixlen = XFS_FSB_TO_BB(mp, XFS_B_TO_FSB(mp, fixlen));
-               bmv->bmv_length =
-                       max_t(__int64_t, fixlen - bmv->bmv_offset, 0);
-       } else if (bmv->bmv_length == 0) {
-               bmv->bmv_entries = 0;
-               return 0;
-       } else if (bmv->bmv_length < 0) {
-               return XFS_ERROR(EINVAL);
-       }
-
-       nex = bmv->bmv_count - 1;
-       if (nex <= 0)
-               return XFS_ERROR(EINVAL);
-       bmvend = bmv->bmv_offset + bmv->bmv_length;
-
-
-       if (bmv->bmv_count > ULONG_MAX / sizeof(struct getbmapx))
-               return XFS_ERROR(ENOMEM);
-       out = kmem_zalloc(bmv->bmv_count * sizeof(struct getbmapx), KM_MAYFAIL);
-       if (!out) {
-               out = kmem_zalloc_large(bmv->bmv_count *
-                                       sizeof(struct getbmapx));
-               if (!out)
-                       return XFS_ERROR(ENOMEM);
-       }
-
-       xfs_ilock(ip, XFS_IOLOCK_SHARED);
-       if (whichfork == XFS_DATA_FORK && !(iflags & BMV_IF_DELALLOC)) {
-               if (ip->i_delayed_blks || XFS_ISIZE(ip) > ip->i_d.di_size) {
-                       error = -filemap_write_and_wait(VFS_I(ip)->i_mapping);
-                       if (error)
-                               goto out_unlock_iolock;
-               }
-               /*
-                * even after flushing the inode, there can still be delalloc
-                * blocks on the inode beyond EOF due to speculative
-                * preallocation. These are not removed until the release
-                * function is called or the inode is inactivated. Hence we
-                * cannot assert here that ip->i_delayed_blks == 0.
-                */
-       }
-
-       lock = xfs_ilock_map_shared(ip);
-
-       /*
-        * Don't let nex be bigger than the number of extents
-        * we can have assuming alternating holes and real extents.
-        */
-       if (nex > XFS_IFORK_NEXTENTS(ip, whichfork) * 2 + 1)
-               nex = XFS_IFORK_NEXTENTS(ip, whichfork) * 2 + 1;
-
-       bmapi_flags = xfs_bmapi_aflag(whichfork);
-       if (!(iflags & BMV_IF_PREALLOC))
-               bmapi_flags |= XFS_BMAPI_IGSTATE;
-
-       /*
-        * Allocate enough space to handle "subnex" maps at a time.
-        */
-       error = ENOMEM;
-       subnex = 16;
-       map = kmem_alloc(subnex * sizeof(*map), KM_MAYFAIL | KM_NOFS);
-       if (!map)
-               goto out_unlock_ilock;
+       int                     isrt;           /* freeing in rt area */
+       xfs_extnum_t            lastx;          /* last extent index used */
+       int                     logflags;       /* transaction logging flags */
+       xfs_extlen_t            mod;            /* rt extent offset */
+       xfs_mount_t             *mp;            /* mount structure */
+       xfs_extnum_t            nextents;       /* number of file extents */
+       xfs_bmbt_irec_t         prev;           /* previous extent record */
+       xfs_fileoff_t           start;          /* first file offset deleted */
+       int                     tmp_logflags;   /* partial logging flags */
+       int                     wasdel;         /* was a delayed alloc extent */
+       int                     whichfork;      /* data or attribute fork */
+       xfs_fsblock_t           sum;
 
-       bmv->bmv_entries = 0;
+       trace_xfs_bunmap(ip, bno, len, flags, _RET_IP_);
 
-       if (XFS_IFORK_NEXTENTS(ip, whichfork) == 0 &&
-           (whichfork == XFS_ATTR_FORK || !(iflags & BMV_IF_DELALLOC))) {
-               error = 0;
-               goto out_free_map;
+       whichfork = (flags & XFS_BMAPI_ATTRFORK) ?
+               XFS_ATTR_FORK : XFS_DATA_FORK;
+       ifp = XFS_IFORK_PTR(ip, whichfork);
+       if (unlikely(
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)) {
+               XFS_ERROR_REPORT("xfs_bunmapi", XFS_ERRLEVEL_LOW,
+                                ip->i_mount);
+               return XFS_ERROR(EFSCORRUPTED);
        }
+       mp = ip->i_mount;
+       if (XFS_FORCED_SHUTDOWN(mp))
+               return XFS_ERROR(EIO);
 
-       nexleft = nex;
+       ASSERT(len > 0);
+       ASSERT(nexts >= 0);
 
-       do {
-               nmap = (nexleft > subnex) ? subnex : nexleft;
-               error = xfs_bmapi_read(ip, XFS_BB_TO_FSBT(mp, bmv->bmv_offset),
-                                      XFS_BB_TO_FSB(mp, bmv->bmv_length),
-                                      map, &nmap, bmapi_flags);
-               if (error)
-                       goto out_free_map;
-               ASSERT(nmap <= subnex);
+       if (!(ifp->if_flags & XFS_IFEXTENTS) &&
+           (error = xfs_iread_extents(tp, ip, whichfork)))
+               return error;
+       nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t);
+       if (nextents == 0) {
+               *done = 1;
+               return 0;
+       }
+       XFS_STATS_INC(xs_blk_unmap);
+       isrt = (whichfork == XFS_DATA_FORK) && XFS_IS_REALTIME_INODE(ip);
+       start = bno;
+       bno = start + len - 1;
+       ep = xfs_bmap_search_extents(ip, bno, whichfork, &eof, &lastx, &got,
+               &prev);
 
-               for (i = 0; i < nmap && nexleft && bmv->bmv_length; i++) {
-                       out[cur_ext].bmv_oflags = 0;
-                       if (map[i].br_state == XFS_EXT_UNWRITTEN)
-                               out[cur_ext].bmv_oflags |= BMV_OF_PREALLOC;
-                       else if (map[i].br_startblock == DELAYSTARTBLOCK)
-                               out[cur_ext].bmv_oflags |= BMV_OF_DELALLOC;
-                       out[cur_ext].bmv_offset =
-                               XFS_FSB_TO_BB(mp, map[i].br_startoff);
-                       out[cur_ext].bmv_length =
-                               XFS_FSB_TO_BB(mp, map[i].br_blockcount);
-                       out[cur_ext].bmv_unused1 = 0;
-                       out[cur_ext].bmv_unused2 = 0;
+       /*
+        * Check to see if the given block number is past the end of the
+        * file, back up to the last block if so...
+        */
+       if (eof) {
+               ep = xfs_iext_get_ext(ifp, --lastx);
+               xfs_bmbt_get_all(ep, &got);
+               bno = got.br_startoff + got.br_blockcount - 1;
+       }
+       logflags = 0;
+       if (ifp->if_flags & XFS_IFBROOT) {
+               ASSERT(XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_BTREE);
+               cur = xfs_bmbt_init_cursor(mp, tp, ip, whichfork);
+               cur->bc_private.b.firstblock = *firstblock;
+               cur->bc_private.b.flist = flist;
+               cur->bc_private.b.flags = 0;
+       } else
+               cur = NULL;
+
+       if (isrt) {
+               /*
+                * Synchronize by locking the bitmap inode.
+                */
+               xfs_ilock(mp->m_rbmip, XFS_ILOCK_EXCL);
+               xfs_trans_ijoin(tp, mp->m_rbmip, XFS_ILOCK_EXCL);
+       }
 
+       extno = 0;
+       while (bno != (xfs_fileoff_t)-1 && bno >= start && lastx >= 0 &&
+              (nexts == 0 || extno < nexts)) {
+               /*
+                * Is the found extent after a hole in which bno lives?
+                * Just back up to the previous extent, if so.
+                */
+               if (got.br_startoff > bno) {
+                       if (--lastx < 0)
+                               break;
+                       ep = xfs_iext_get_ext(ifp, lastx);
+                       xfs_bmbt_get_all(ep, &got);
+               }
+               /*
+                * Is the last block of this extent before the range
+                * we're supposed to delete?  If so, we're done.
+                */
+               bno = XFS_FILEOFF_MIN(bno,
+                       got.br_startoff + got.br_blockcount - 1);
+               if (bno < start)
+                       break;
+               /*
+                * Then deal with the (possibly delayed) allocated space
+                * we found.
+                */
+               ASSERT(ep != NULL);
+               del = got;
+               wasdel = isnullstartblock(del.br_startblock);
+               if (got.br_startoff < start) {
+                       del.br_startoff = start;
+                       del.br_blockcount -= start - got.br_startoff;
+                       if (!wasdel)
+                               del.br_startblock += start - got.br_startoff;
+               }
+               if (del.br_startoff + del.br_blockcount > bno + 1)
+                       del.br_blockcount = bno + 1 - del.br_startoff;
+               sum = del.br_startblock + del.br_blockcount;
+               if (isrt &&
+                   (mod = do_mod(sum, mp->m_sb.sb_rextsize))) {
                        /*
-                        * delayed allocation extents that start beyond EOF can
-                        * occur due to speculative EOF allocation when the
-                        * delalloc extent is larger than the largest freespace
-                        * extent at conversion time. These extents cannot be
-                        * converted by data writeback, so can exist here even
-                        * if we are not supposed to be finding delalloc
-                        * extents.
+                        * Realtime extent not lined up at the end.
+                        * The extent could have been split into written
+                        * and unwritten pieces, or we could just be
+                        * unmapping part of it.  But we can't really
+                        * get rid of part of a realtime extent.
                         */
-                       if (map[i].br_startblock == DELAYSTARTBLOCK &&
-                           map[i].br_startoff <= XFS_B_TO_FSB(mp, XFS_ISIZE(ip)))
-                               ASSERT((iflags & BMV_IF_DELALLOC) != 0);
-
-                        if (map[i].br_startblock == HOLESTARTBLOCK &&
-                           whichfork == XFS_ATTR_FORK) {
-                               /* came to the end of attribute fork */
-                               out[cur_ext].bmv_oflags |= BMV_OF_LAST;
-                               goto out_free_map;
+                       if (del.br_state == XFS_EXT_UNWRITTEN ||
+                           !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
+                               /*
+                                * This piece is unwritten, or we're not
+                                * using unwritten extents.  Skip over it.
+                                */
+                               ASSERT(bno >= mod);
+                               bno -= mod > del.br_blockcount ?
+                                       del.br_blockcount : mod;
+                               if (bno < got.br_startoff) {
+                                       if (--lastx >= 0)
+                                               xfs_bmbt_get_all(xfs_iext_get_ext(
+                                                       ifp, lastx), &got);
+                               }
+                               continue;
                        }
-
-                       if (!xfs_getbmapx_fix_eof_hole(ip, &out[cur_ext],
-                                       prealloced, bmvend,
-                                       map[i].br_startblock))
-                               goto out_free_map;
-
-                       bmv->bmv_offset =
-                               out[cur_ext].bmv_offset +
-                               out[cur_ext].bmv_length;
-                       bmv->bmv_length =
-                               max_t(__int64_t, 0, bmvend - bmv->bmv_offset);
-
                        /*
-                        * In case we don't want to return the hole,
-                        * don't increase cur_ext so that we can reuse
-                        * it in the next loop.
+                        * It's written, turn it unwritten.
+                        * This is better than zeroing it.
                         */
-                       if ((iflags & BMV_IF_NO_HOLES) &&
-                           map[i].br_startblock == HOLESTARTBLOCK) {
-                               memset(&out[cur_ext], 0, sizeof(out[cur_ext]));
+                       ASSERT(del.br_state == XFS_EXT_NORM);
+                       ASSERT(xfs_trans_get_block_res(tp) > 0);
+                       /*
+                        * If this spans a realtime extent boundary,
+                        * chop it back to the start of the one we end at.
+                        */
+                       if (del.br_blockcount > mod) {
+                               del.br_startoff += del.br_blockcount - mod;
+                               del.br_startblock += del.br_blockcount - mod;
+                               del.br_blockcount = mod;
+                       }
+                       del.br_state = XFS_EXT_UNWRITTEN;
+                       error = xfs_bmap_add_extent_unwritten_real(tp, ip,
+                                       &lastx, &cur, &del, firstblock, flist,
+                                       &logflags);
+                       if (error)
+                               goto error0;
+                       goto nodelete;
+               }
+               if (isrt && (mod = do_mod(del.br_startblock, mp->m_sb.sb_rextsize))) {
+                       /*
+                        * Realtime extent is lined up at the end but not
+                        * at the front.  We'll get rid of full extents if
+                        * we can.
+                        */
+                       mod = mp->m_sb.sb_rextsize - mod;
+                       if (del.br_blockcount > mod) {
+                               del.br_blockcount -= mod;
+                               del.br_startoff += mod;
+                               del.br_startblock += mod;
+                       } else if ((del.br_startoff == start &&
+                                   (del.br_state == XFS_EXT_UNWRITTEN ||
+                                    xfs_trans_get_block_res(tp) == 0)) ||
+                                  !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
+                               /*
+                                * Can't make it unwritten.  There isn't
+                                * a full extent here so just skip it.
+                                */
+                               ASSERT(bno >= del.br_blockcount);
+                               bno -= del.br_blockcount;
+                               if (got.br_startoff > bno) {
+                                       if (--lastx >= 0) {
+                                               ep = xfs_iext_get_ext(ifp,
+                                                                     lastx);
+                                               xfs_bmbt_get_all(ep, &got);
+                                       }
+                               }
                                continue;
+                       } else if (del.br_state == XFS_EXT_UNWRITTEN) {
+                               /*
+                                * This one is already unwritten.
+                                * It must have a written left neighbor.
+                                * Unwrite the killed part of that one and
+                                * try again.
+                                */
+                               ASSERT(lastx > 0);
+                               xfs_bmbt_get_all(xfs_iext_get_ext(ifp,
+                                               lastx - 1), &prev);
+                               ASSERT(prev.br_state == XFS_EXT_NORM);
+                               ASSERT(!isnullstartblock(prev.br_startblock));
+                               ASSERT(del.br_startblock ==
+                                      prev.br_startblock + prev.br_blockcount);
+                               if (prev.br_startoff < start) {
+                                       mod = start - prev.br_startoff;
+                                       prev.br_blockcount -= mod;
+                                       prev.br_startblock += mod;
+                                       prev.br_startoff = start;
+                               }
+                               prev.br_state = XFS_EXT_UNWRITTEN;
+                               lastx--;
+                               error = xfs_bmap_add_extent_unwritten_real(tp,
+                                               ip, &lastx, &cur, &prev,
+                                               firstblock, flist, &logflags);
+                               if (error)
+                                       goto error0;
+                               goto nodelete;
+                       } else {
+                               ASSERT(del.br_state == XFS_EXT_NORM);
+                               del.br_state = XFS_EXT_UNWRITTEN;
+                               error = xfs_bmap_add_extent_unwritten_real(tp,
+                                               ip, &lastx, &cur, &del,
+                                               firstblock, flist, &logflags);
+                               if (error)
+                                       goto error0;
+                               goto nodelete;
                        }
-
-                       nexleft--;
-                       bmv->bmv_entries++;
-                       cur_ext++;
                }
-       } while (nmap && nexleft && bmv->bmv_length);
-
- out_free_map:
-       kmem_free(map);
- out_unlock_ilock:
-       xfs_iunlock_map_shared(ip, lock);
- out_unlock_iolock:
-       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
-
-       for (i = 0; i < cur_ext; i++) {
-               int full = 0;   /* user array is full */
-
-               /* format results & advance arg */
-               error = formatter(&arg, &out[i], &full);
-               if (error || full)
-                       break;
-       }
-
-       if (is_vmalloc_addr(out))
-               kmem_free_large(out);
-       else
-               kmem_free(out);
-       return error;
-}
-
-#ifdef DEBUG
-STATIC struct xfs_buf *
-xfs_bmap_get_bp(
-       struct xfs_btree_cur    *cur,
-       xfs_fsblock_t           bno)
-{
-       struct xfs_log_item_desc *lidp;
-       int                     i;
-
-       if (!cur)
-               return NULL;
-
-       for (i = 0; i < XFS_BTREE_MAXLEVELS; i++) {
-               if (!cur->bc_bufs[i])
-                       break;
-               if (XFS_BUF_ADDR(cur->bc_bufs[i]) == bno)
-                       return cur->bc_bufs[i];
-       }
-
-       /* Chase down all the log items to see if the bp is there */
-       list_for_each_entry(lidp, &cur->bc_tp->t_items, lid_trans) {
-               struct xfs_buf_log_item *bip;
-               bip = (struct xfs_buf_log_item *)lidp->lid_item;
-               if (bip->bli_item.li_type == XFS_LI_BUF &&
-                   XFS_BUF_ADDR(bip->bli_buf) == bno)
-                       return bip->bli_buf;
-       }
-
-       return NULL;
-}
-
-STATIC void
-xfs_check_block(
-       struct xfs_btree_block  *block,
-       xfs_mount_t             *mp,
-       int                     root,
-       short                   sz)
-{
-       int                     i, j, dmxr;
-       __be64                  *pp, *thispa;   /* pointer to block address */
-       xfs_bmbt_key_t          *prevp, *keyp;
-
-       ASSERT(be16_to_cpu(block->bb_level) > 0);
-
-       prevp = NULL;
-       for( i = 1; i <= xfs_btree_get_numrecs(block); i++) {
-               dmxr = mp->m_bmap_dmxr[0];
-               keyp = XFS_BMBT_KEY_ADDR(mp, block, i);
+               if (wasdel) {
+                       ASSERT(startblockval(del.br_startblock) > 0);
+                       /* Update realtime/data freespace, unreserve quota */
+                       if (isrt) {
+                               xfs_filblks_t rtexts;
 
-               if (prevp) {
-                       ASSERT(be64_to_cpu(prevp->br_startoff) <
-                              be64_to_cpu(keyp->br_startoff));
+                               rtexts = XFS_FSB_TO_B(mp, del.br_blockcount);
+                               do_div(rtexts, mp->m_sb.sb_rextsize);
+                               xfs_mod_incore_sb(mp, XFS_SBS_FREXTENTS,
+                                               (int64_t)rtexts, 0);
+                               (void)xfs_trans_reserve_quota_nblks(NULL,
+                                       ip, -((long)del.br_blockcount), 0,
+                                       XFS_QMOPT_RES_RTBLKS);
+                       } else {
+                               xfs_icsb_modify_counters(mp, XFS_SBS_FDBLOCKS,
+                                               (int64_t)del.br_blockcount, 0);
+                               (void)xfs_trans_reserve_quota_nblks(NULL,
+                                       ip, -((long)del.br_blockcount), 0,
+                                       XFS_QMOPT_RES_REGBLKS);
+                       }
+                       ip->i_delayed_blks -= del.br_blockcount;
+                       if (cur)
+                               cur->bc_private.b.flags |=
+                                       XFS_BTCUR_BPRV_WASDEL;
+               } else if (cur)
+                       cur->bc_private.b.flags &= ~XFS_BTCUR_BPRV_WASDEL;
+               /*
+                * If it's the case where the directory code is running
+                * with no block reservation, and the deleted block is in
+                * the middle of its extent, and the resulting insert
+                * of an extent would cause transformation to btree format,
+                * then reject it.  The calling code will then swap
+                * blocks around instead.
+                * We have to do this now, rather than waiting for the
+                * conversion to btree format, since the transaction
+                * will be dirty.
+                */
+               if (!wasdel && xfs_trans_get_block_res(tp) == 0 &&
+                   XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS &&
+                   XFS_IFORK_NEXTENTS(ip, whichfork) >= /* Note the >= */
+                       XFS_IFORK_MAXEXT(ip, whichfork) &&
+                   del.br_startoff > got.br_startoff &&
+                   del.br_startoff + del.br_blockcount <
+                   got.br_startoff + got.br_blockcount) {
+                       error = XFS_ERROR(ENOSPC);
+                       goto error0;
                }
-               prevp = keyp;
-
+               error = xfs_bmap_del_extent(ip, tp, &lastx, flist, cur, &del,
+                               &tmp_logflags, whichfork);
+               logflags |= tmp_logflags;
+               if (error)
+                       goto error0;
+               bno = del.br_startoff - 1;
+nodelete:
                /*
-                * Compare the block numbers to see if there are dups.
+                * If not done go on to the next (previous) record.
                 */
-               if (root)
-                       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, i, sz);
-               else
-                       pp = XFS_BMBT_PTR_ADDR(mp, block, i, dmxr);
-
-               for (j = i+1; j <= be16_to_cpu(block->bb_numrecs); j++) {
-                       if (root)
-                               thispa = XFS_BMAP_BROOT_PTR_ADDR(mp, block, j, sz);
-                       else
-                               thispa = XFS_BMBT_PTR_ADDR(mp, block, j, dmxr);
-                       if (*thispa == *pp) {
-                               xfs_warn(mp, "%s: thispa(%d) == pp(%d) %Ld",
-                                       __func__, j, i,
-                                       (unsigned long long)be64_to_cpu(*thispa));
-                               panic("%s: ptrs are equal in node\n",
-                                       __func__);
+               if (bno != (xfs_fileoff_t)-1 && bno >= start) {
+                       if (lastx >= 0) {
+                               ep = xfs_iext_get_ext(ifp, lastx);
+                               if (xfs_bmbt_get_startoff(ep) > bno) {
+                                       if (--lastx >= 0)
+                                               ep = xfs_iext_get_ext(ifp,
+                                                                     lastx);
+                               }
+                               xfs_bmbt_get_all(ep, &got);
                        }
+                       extno++;
                }
        }
-}
-
-/*
- * Check that the extents for the inode ip are in the right order in all
- * btree leaves.
- */
-
-STATIC void
-xfs_bmap_check_leaf_extents(
-       xfs_btree_cur_t         *cur,   /* btree cursor or null */
-       xfs_inode_t             *ip,            /* incore inode pointer */
-       int                     whichfork)      /* data or attr fork */
-{
-       struct xfs_btree_block  *block; /* current btree block */
-       xfs_fsblock_t           bno;    /* block # of "block" */
-       xfs_buf_t               *bp;    /* buffer for "block" */
-       int                     error;  /* error return value */
-       xfs_extnum_t            i=0, j; /* index into the extents list */
-       xfs_ifork_t             *ifp;   /* fork structure */
-       int                     level;  /* btree level, for checking */
-       xfs_mount_t             *mp;    /* file system mount structure */
-       __be64                  *pp;    /* pointer to block address */
-       xfs_bmbt_rec_t          *ep;    /* pointer to current extent */
-       xfs_bmbt_rec_t          last = {0, 0}; /* last extent in prev block */
-       xfs_bmbt_rec_t          *nextp; /* pointer to next extent */
-       int                     bp_release = 0;
-
-       if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE) {
-               return;
-       }
-
-       bno = NULLFSBLOCK;
-       mp = ip->i_mount;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       block = ifp->if_broot;
-       /*
-        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
-        */
-       level = be16_to_cpu(block->bb_level);
-       ASSERT(level > 0);
-       xfs_check_block(block, mp, 1, ifp->if_broot_bytes);
-       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
-       bno = be64_to_cpu(*pp);
-
-       ASSERT(bno != NULLDFSBNO);
-       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
-       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
+       *done = bno == (xfs_fileoff_t)-1 || bno < start || lastx < 0;
 
        /*
-        * Go down the tree until leaf level is reached, following the first
-        * pointer (leftmost) at each level.
+        * Convert to a btree if necessary.
         */
-       while (level-- > 0) {
-               /* See if buf is in cur first */
-               bp_release = 0;
-               bp = xfs_bmap_get_bp(cur, XFS_FSB_TO_DADDR(mp, bno));
-               if (!bp) {
-                       bp_release = 1;
-                       error = xfs_btree_read_bufl(mp, NULL, bno, 0, &bp,
-                                               XFS_BMAP_BTREE_REF,
-                                               &xfs_bmbt_buf_ops);
-                       if (error)
-                               goto error_norelse;
-               }
-               block = XFS_BUF_TO_BLOCK(bp);
-               XFS_WANT_CORRUPTED_GOTO(
-                       xfs_bmap_sanity_check(mp, bp, level),
-                       error0);
-               if (level == 0)
-                       break;
-
-               /*
-                * Check this block for basic sanity (increasing keys and
-                * no duplicate blocks).
-                */
-
-               xfs_check_block(block, mp, 0, 0);
-               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
-               bno = be64_to_cpu(*pp);
-               XFS_WANT_CORRUPTED_GOTO(XFS_FSB_SANITY_CHECK(mp, bno), error0);
-               if (bp_release) {
-                       bp_release = 0;
-                       xfs_trans_brelse(NULL, bp);
-               }
+       if (xfs_bmap_needs_btree(ip, whichfork)) {
+               ASSERT(cur == NULL);
+               error = xfs_bmap_extents_to_btree(tp, ip, firstblock, flist,
+                       &cur, 0, &tmp_logflags, whichfork);
+               logflags |= tmp_logflags;
+               if (error)
+                       goto error0;
        }
-
        /*
-        * Here with bp and block set to the leftmost leaf node in the tree.
+        * transform from btree to extents, give it cur
         */
-       i = 0;
-
+       else if (xfs_bmap_wants_extents(ip, whichfork)) {
+               ASSERT(cur != NULL);
+               error = xfs_bmap_btree_to_extents(tp, ip, cur, &tmp_logflags,
+                       whichfork);
+               logflags |= tmp_logflags;
+               if (error)
+                       goto error0;
+       }
        /*
-        * Loop over all leaf nodes checking that all extents are in the right order.
+        * transform from extents to local?
         */
-       for (;;) {
-               xfs_fsblock_t   nextbno;
-               xfs_extnum_t    num_recs;
+       error = 0;
+error0:
+       /*
+        * Log everything.  Do this after conversion, there's no point in
+        * logging the extent records if we've converted to btree format.
+        */
+       if ((logflags & xfs_ilog_fext(whichfork)) &&
+           XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_EXTENTS)
+               logflags &= ~xfs_ilog_fext(whichfork);
+       else if ((logflags & xfs_ilog_fbroot(whichfork)) &&
+                XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_BTREE)
+               logflags &= ~xfs_ilog_fbroot(whichfork);
+       /*
+        * Log inode even in the error case, if the transaction
+        * is dirty we'll need to shut down the filesystem.
+        */
+       if (logflags)
+               xfs_trans_log_inode(tp, ip, logflags);
+       if (cur) {
+               if (!error) {
+                       *firstblock = cur->bc_private.b.firstblock;
+                       cur->bc_private.b.allocated = 0;
+               }
+               xfs_btree_del_cursor(cur,
+                       error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       }
+       return error;
+}
 
+/*
+ * returns 1 for success, 0 if we failed to map the extent.
+ */
+STATIC int
+xfs_getbmapx_fix_eof_hole(
+       xfs_inode_t             *ip,            /* xfs incore inode pointer */
+       struct getbmapx         *out,           /* output structure */
+       int                     prealloced,     /* this is a file with
+                                                * preallocated data space */
+       __int64_t               end,            /* last block requested */
+       xfs_fsblock_t           startblock)
+{
+       __int64_t               fixlen;
+       xfs_mount_t             *mp;            /* file system mount point */
+       xfs_ifork_t             *ifp;           /* inode fork pointer */
+       xfs_extnum_t            lastx;          /* last extent pointer */
+       xfs_fileoff_t           fileblock;
 
-               num_recs = xfs_btree_get_numrecs(block);
+       if (startblock == HOLESTARTBLOCK) {
+               mp = ip->i_mount;
+               out->bmv_block = -1;
+               fixlen = XFS_FSB_TO_BB(mp, XFS_B_TO_FSB(mp, XFS_ISIZE(ip)));
+               fixlen -= out->bmv_offset;
+               if (prealloced && out->bmv_offset + out->bmv_length == end) {
+                       /* Came to hole at EOF. Trim it. */
+                       if (fixlen <= 0)
+                               return 0;
+                       out->bmv_length = fixlen;
+               }
+       } else {
+               if (startblock == DELAYSTARTBLOCK)
+                       out->bmv_block = -2;
+               else
+                       out->bmv_block = xfs_fsb_to_db(ip, startblock);
+               fileblock = XFS_BB_TO_FSB(ip->i_mount, out->bmv_offset);
+               ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
+               if (xfs_iext_bno_to_ext(ifp, fileblock, &lastx) &&
+                  (lastx == (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))-1))
+                       out->bmv_oflags |= BMV_OF_LAST;
+       }
 
-               /*
-                * Read-ahead the next leaf block, if any.
-                */
+       return 1;
+}
 
-               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
+/*
+ * Get inode's extents as described in bmv, and format for output.
+ * Calls formatter to fill the user's buffer until all extents
+ * are mapped, until the passed-in bmv->bmv_count slots have
+ * been filled, or until the formatter short-circuits the loop,
+ * if it is tracking filled-in extents on its own.
+ */
+int                                            /* error code */
+xfs_getbmap(
+       xfs_inode_t             *ip,
+       struct getbmapx         *bmv,           /* user bmap structure */
+       xfs_bmap_format_t       formatter,      /* format to user */
+       void                    *arg)           /* formatter arg */
+{
+       __int64_t               bmvend;         /* last block requested */
+       int                     error = 0;      /* return value */
+       __int64_t               fixlen;         /* length for -1 case */
+       int                     i;              /* extent number */
+       int                     lock;           /* lock state */
+       xfs_bmbt_irec_t         *map;           /* buffer for user's data */
+       xfs_mount_t             *mp;            /* file system mount point */
+       int                     nex;            /* # of user extents can do */
+       int                     nexleft;        /* # of user extents left */
+       int                     subnex;         /* # of bmapi's can do */
+       int                     nmap;           /* number of map entries */
+       struct getbmapx         *out;           /* output structure */
+       int                     whichfork;      /* data or attr fork */
+       int                     prealloced;     /* this is a file with
+                                                * preallocated data space */
+       int                     iflags;         /* interface flags */
+       int                     bmapi_flags;    /* flags for xfs_bmapi */
+       int                     cur_ext = 0;
 
-               /*
-                * Check all the extents to make sure they are OK.
-                * If we had a previous block, the last entry should
-                * conform with the first entry in this one.
-                */
+       mp = ip->i_mount;
+       iflags = bmv->bmv_iflags;
+       whichfork = iflags & BMV_IF_ATTRFORK ? XFS_ATTR_FORK : XFS_DATA_FORK;
 
-               ep = XFS_BMBT_REC_ADDR(mp, block, 1);
-               if (i) {
-                       ASSERT(xfs_bmbt_disk_get_startoff(&last) +
-                              xfs_bmbt_disk_get_blockcount(&last) <=
-                              xfs_bmbt_disk_get_startoff(ep));
-               }
-               for (j = 1; j < num_recs; j++) {
-                       nextp = XFS_BMBT_REC_ADDR(mp, block, j + 1);
-                       ASSERT(xfs_bmbt_disk_get_startoff(ep) +
-                              xfs_bmbt_disk_get_blockcount(ep) <=
-                              xfs_bmbt_disk_get_startoff(nextp));
-                       ep = nextp;
+       if (whichfork == XFS_ATTR_FORK) {
+               if (XFS_IFORK_Q(ip)) {
+                       if (ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS &&
+                           ip->i_d.di_aformat != XFS_DINODE_FMT_BTREE &&
+                           ip->i_d.di_aformat != XFS_DINODE_FMT_LOCAL)
+                               return XFS_ERROR(EINVAL);
+               } else if (unlikely(
+                          ip->i_d.di_aformat != 0 &&
+                          ip->i_d.di_aformat != XFS_DINODE_FMT_EXTENTS)) {
+                       XFS_ERROR_REPORT("xfs_getbmap", XFS_ERRLEVEL_LOW,
+                                        ip->i_mount);
+                       return XFS_ERROR(EFSCORRUPTED);
                }
 
-               last = *ep;
-               i += num_recs;
-               if (bp_release) {
-                       bp_release = 0;
-                       xfs_trans_brelse(NULL, bp);
-               }
-               bno = nextbno;
-               /*
-                * If we've reached the end, stop.
-                */
-               if (bno == NULLFSBLOCK)
-                       break;
+               prealloced = 0;
+               fixlen = 1LL << 32;
+       } else {
+               if (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS &&
+                   ip->i_d.di_format != XFS_DINODE_FMT_BTREE &&
+                   ip->i_d.di_format != XFS_DINODE_FMT_LOCAL)
+                       return XFS_ERROR(EINVAL);
 
-               bp_release = 0;
-               bp = xfs_bmap_get_bp(cur, XFS_FSB_TO_DADDR(mp, bno));
-               if (!bp) {
-                       bp_release = 1;
-                       error = xfs_btree_read_bufl(mp, NULL, bno, 0, &bp,
-                                               XFS_BMAP_BTREE_REF,
-                                               &xfs_bmbt_buf_ops);
-                       if (error)
-                               goto error_norelse;
+               if (xfs_get_extsz_hint(ip) ||
+                   ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC|XFS_DIFLAG_APPEND)){
+                       prealloced = 1;
+                       fixlen = mp->m_super->s_maxbytes;
+               } else {
+                       prealloced = 0;
+                       fixlen = XFS_ISIZE(ip);
                }
-               block = XFS_BUF_TO_BLOCK(bp);
        }
-       if (bp_release) {
-               bp_release = 0;
-               xfs_trans_brelse(NULL, bp);
+
+       if (bmv->bmv_length == -1) {
+               fixlen = XFS_FSB_TO_BB(mp, XFS_B_TO_FSB(mp, fixlen));
+               bmv->bmv_length =
+                       max_t(__int64_t, fixlen - bmv->bmv_offset, 0);
+       } else if (bmv->bmv_length == 0) {
+               bmv->bmv_entries = 0;
+               return 0;
+       } else if (bmv->bmv_length < 0) {
+               return XFS_ERROR(EINVAL);
        }
-       return;
 
-error0:
-       xfs_warn(mp, "%s: at error0", __func__);
-       if (bp_release)
-               xfs_trans_brelse(NULL, bp);
-error_norelse:
-       xfs_warn(mp, "%s: BAD after btree leaves for %d extents",
-               __func__, i);
-       panic("%s: CORRUPTED BTREE OR SOMETHING", __func__);
-       return;
-}
-#endif
+       nex = bmv->bmv_count - 1;
+       if (nex <= 0)
+               return XFS_ERROR(EINVAL);
+       bmvend = bmv->bmv_offset + bmv->bmv_length;
 
-/*
- * Count fsblocks of the given fork.
- */
-int                                            /* error */
-xfs_bmap_count_blocks(
-       xfs_trans_t             *tp,            /* transaction pointer */
-       xfs_inode_t             *ip,            /* incore inode */
-       int                     whichfork,      /* data or attr fork */
-       int                     *count)         /* out: count of blocks */
-{
-       struct xfs_btree_block  *block; /* current btree block */
-       xfs_fsblock_t           bno;    /* block # of "block" */
-       xfs_ifork_t             *ifp;   /* fork structure */
-       int                     level;  /* btree level, for checking */
-       xfs_mount_t             *mp;    /* file system mount structure */
-       __be64                  *pp;    /* pointer to block address */
 
-       bno = NULLFSBLOCK;
-       mp = ip->i_mount;
-       ifp = XFS_IFORK_PTR(ip, whichfork);
-       if ( XFS_IFORK_FORMAT(ip, whichfork) == XFS_DINODE_FMT_EXTENTS ) {
-               xfs_bmap_count_leaves(ifp, 0,
-                       ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t),
-                       count);
-               return 0;
+       if (bmv->bmv_count > ULONG_MAX / sizeof(struct getbmapx))
+               return XFS_ERROR(ENOMEM);
+       out = kmem_zalloc(bmv->bmv_count * sizeof(struct getbmapx), KM_MAYFAIL);
+       if (!out) {
+               out = kmem_zalloc_large(bmv->bmv_count *
+                                       sizeof(struct getbmapx));
+               if (!out)
+                       return XFS_ERROR(ENOMEM);
+       }
+
+       xfs_ilock(ip, XFS_IOLOCK_SHARED);
+       if (whichfork == XFS_DATA_FORK && !(iflags & BMV_IF_DELALLOC)) {
+               if (ip->i_delayed_blks || XFS_ISIZE(ip) > ip->i_d.di_size) {
+                       error = -filemap_write_and_wait(VFS_I(ip)->i_mapping);
+                       if (error)
+                               goto out_unlock_iolock;
+               }
+               /*
+                * even after flushing the inode, there can still be delalloc
+                * blocks on the inode beyond EOF due to speculative
+                * preallocation. These are not removed until the release
+                * function is called or the inode is inactivated. Hence we
+                * cannot assert here that ip->i_delayed_blks == 0.
+                */
        }
 
+       lock = xfs_ilock_map_shared(ip);
+
        /*
-        * Root level must use BMAP_BROOT_PTR_ADDR macro to get ptr out.
+        * Don't let nex be bigger than the number of extents
+        * we can have assuming alternating holes and real extents.
         */
-       block = ifp->if_broot;
-       level = be16_to_cpu(block->bb_level);
-       ASSERT(level > 0);
-       pp = XFS_BMAP_BROOT_PTR_ADDR(mp, block, 1, ifp->if_broot_bytes);
-       bno = be64_to_cpu(*pp);
-       ASSERT(bno != NULLDFSBNO);
-       ASSERT(XFS_FSB_TO_AGNO(mp, bno) < mp->m_sb.sb_agcount);
-       ASSERT(XFS_FSB_TO_AGBNO(mp, bno) < mp->m_sb.sb_agblocks);
+       if (nex > XFS_IFORK_NEXTENTS(ip, whichfork) * 2 + 1)
+               nex = XFS_IFORK_NEXTENTS(ip, whichfork) * 2 + 1;
 
-       if (unlikely(xfs_bmap_count_tree(mp, tp, ifp, bno, level, count) < 0)) {
-               XFS_ERROR_REPORT("xfs_bmap_count_blocks(2)", XFS_ERRLEVEL_LOW,
-                                mp);
-               return XFS_ERROR(EFSCORRUPTED);
+       bmapi_flags = xfs_bmapi_aflag(whichfork);
+       if (!(iflags & BMV_IF_PREALLOC))
+               bmapi_flags |= XFS_BMAPI_IGSTATE;
+
+       /*
+        * Allocate enough space to handle "subnex" maps at a time.
+        */
+       error = ENOMEM;
+       subnex = 16;
+       map = kmem_alloc(subnex * sizeof(*map), KM_MAYFAIL | KM_NOFS);
+       if (!map)
+               goto out_unlock_ilock;
+
+       bmv->bmv_entries = 0;
+
+       if (XFS_IFORK_NEXTENTS(ip, whichfork) == 0 &&
+           (whichfork == XFS_ATTR_FORK || !(iflags & BMV_IF_DELALLOC))) {
+               error = 0;
+               goto out_free_map;
        }
 
-       return 0;
-}
+       nexleft = nex;
 
-/*
- * Recursively walks each level of a btree
- * to count total fsblocks is use.
- */
-STATIC int                                     /* error */
-xfs_bmap_count_tree(
-       xfs_mount_t     *mp,            /* file system mount point */
-       xfs_trans_t     *tp,            /* transaction pointer */
-       xfs_ifork_t     *ifp,           /* inode fork pointer */
-       xfs_fsblock_t   blockno,        /* file system block number */
-       int             levelin,        /* level in btree */
-       int             *count)         /* Count of blocks */
-{
-       int                     error;
-       xfs_buf_t               *bp, *nbp;
-       int                     level = levelin;
-       __be64                  *pp;
-       xfs_fsblock_t           bno = blockno;
-       xfs_fsblock_t           nextbno;
-       struct xfs_btree_block  *block, *nextblock;
-       int                     numrecs;
+       do {
+               nmap = (nexleft > subnex) ? subnex : nexleft;
+               error = xfs_bmapi_read(ip, XFS_BB_TO_FSBT(mp, bmv->bmv_offset),
+                                      XFS_BB_TO_FSB(mp, bmv->bmv_length),
+                                      map, &nmap, bmapi_flags);
+               if (error)
+                       goto out_free_map;
+               ASSERT(nmap <= subnex);
 
-       error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp, XFS_BMAP_BTREE_REF,
-                                               &xfs_bmbt_buf_ops);
-       if (error)
-               return error;
-       *count += 1;
-       block = XFS_BUF_TO_BLOCK(bp);
+               for (i = 0; i < nmap && nexleft && bmv->bmv_length; i++) {
+                       out[cur_ext].bmv_oflags = 0;
+                       if (map[i].br_state == XFS_EXT_UNWRITTEN)
+                               out[cur_ext].bmv_oflags |= BMV_OF_PREALLOC;
+                       else if (map[i].br_startblock == DELAYSTARTBLOCK)
+                               out[cur_ext].bmv_oflags |= BMV_OF_DELALLOC;
+                       out[cur_ext].bmv_offset =
+                               XFS_FSB_TO_BB(mp, map[i].br_startoff);
+                       out[cur_ext].bmv_length =
+                               XFS_FSB_TO_BB(mp, map[i].br_blockcount);
+                       out[cur_ext].bmv_unused1 = 0;
+                       out[cur_ext].bmv_unused2 = 0;
 
-       if (--level) {
-               /* Not at node above leaves, count this level of nodes */
-               nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
-               while (nextbno != NULLFSBLOCK) {
-                       error = xfs_btree_read_bufl(mp, tp, nextbno, 0, &nbp,
-                                               XFS_BMAP_BTREE_REF,
-                                               &xfs_bmbt_buf_ops);
-                       if (error)
-                               return error;
-                       *count += 1;
-                       nextblock = XFS_BUF_TO_BLOCK(nbp);
-                       nextbno = be64_to_cpu(nextblock->bb_u.l.bb_rightsib);
-                       xfs_trans_brelse(tp, nbp);
-               }
+                       /*
+                        * delayed allocation extents that start beyond EOF can
+                        * occur due to speculative EOF allocation when the
+                        * delalloc extent is larger than the largest freespace
+                        * extent at conversion time. These extents cannot be
+                        * converted by data writeback, so can exist here even
+                        * if we are not supposed to be finding delalloc
+                        * extents.
+                        */
+                       if (map[i].br_startblock == DELAYSTARTBLOCK &&
+                           map[i].br_startoff <= XFS_B_TO_FSB(mp, XFS_ISIZE(ip)))
+                               ASSERT((iflags & BMV_IF_DELALLOC) != 0);
 
-               /* Dive to the next level */
-               pp = XFS_BMBT_PTR_ADDR(mp, block, 1, mp->m_bmap_dmxr[1]);
-               bno = be64_to_cpu(*pp);
-               if (unlikely((error =
-                    xfs_bmap_count_tree(mp, tp, ifp, bno, level, count)) < 0)) {
-                       xfs_trans_brelse(tp, bp);
-                       XFS_ERROR_REPORT("xfs_bmap_count_tree(1)",
-                                        XFS_ERRLEVEL_LOW, mp);
-                       return XFS_ERROR(EFSCORRUPTED);
-               }
-               xfs_trans_brelse(tp, bp);
-       } else {
-               /* count all level 1 nodes and their leaves */
-               for (;;) {
-                       nextbno = be64_to_cpu(block->bb_u.l.bb_rightsib);
-                       numrecs = be16_to_cpu(block->bb_numrecs);
-                       xfs_bmap_disk_count_leaves(mp, block, numrecs, count);
-                       xfs_trans_brelse(tp, bp);
-                       if (nextbno == NULLFSBLOCK)
-                               break;
-                       bno = nextbno;
-                       error = xfs_btree_read_bufl(mp, tp, bno, 0, &bp,
-                                               XFS_BMAP_BTREE_REF,
-                                               &xfs_bmbt_buf_ops);
-                       if (error)
-                               return error;
-                       *count += 1;
-                       block = XFS_BUF_TO_BLOCK(bp);
-               }
-       }
-       return 0;
-}
+                        if (map[i].br_startblock == HOLESTARTBLOCK &&
+                           whichfork == XFS_ATTR_FORK) {
+                               /* came to the end of attribute fork */
+                               out[cur_ext].bmv_oflags |= BMV_OF_LAST;
+                               goto out_free_map;
+                       }
 
-/*
- * Count leaf blocks given a range of extent records.
- */
-STATIC void
-xfs_bmap_count_leaves(
-       xfs_ifork_t             *ifp,
-       xfs_extnum_t            idx,
-       int                     numrecs,
-       int                     *count)
-{
-       int             b;
+                       if (!xfs_getbmapx_fix_eof_hole(ip, &out[cur_ext],
+                                       prealloced, bmvend,
+                                       map[i].br_startblock))
+                               goto out_free_map;
 
-       for (b = 0; b < numrecs; b++) {
-               xfs_bmbt_rec_host_t *frp = xfs_iext_get_ext(ifp, idx + b);
-               *count += xfs_bmbt_get_blockcount(frp);
-       }
-}
+                       bmv->bmv_offset =
+                               out[cur_ext].bmv_offset +
+                               out[cur_ext].bmv_length;
+                       bmv->bmv_length =
+                               max_t(__int64_t, 0, bmvend - bmv->bmv_offset);
 
-/*
- * Count leaf blocks given a range of extent records originally
- * in btree format.
- */
-STATIC void
-xfs_bmap_disk_count_leaves(
-       struct xfs_mount        *mp,
-       struct xfs_btree_block  *block,
-       int                     numrecs,
-       int                     *count)
-{
-       int             b;
-       xfs_bmbt_rec_t  *frp;
+                       /*
+                        * In case we don't want to return the hole,
+                        * don't increase cur_ext so that we can reuse
+                        * it in the next loop.
+                        */
+                       if ((iflags & BMV_IF_NO_HOLES) &&
+                           map[i].br_startblock == HOLESTARTBLOCK) {
+                               memset(&out[cur_ext], 0, sizeof(out[cur_ext]));
+                               continue;
+                       }
 
-       for (b = 1; b <= numrecs; b++) {
-               frp = XFS_BMBT_REC_ADDR(mp, block, b);
-               *count += xfs_bmbt_disk_get_blockcount(frp);
+                       nexleft--;
+                       bmv->bmv_entries++;
+                       cur_ext++;
+               }
+       } while (nmap && nexleft && bmv->bmv_length);
+
+ out_free_map:
+       kmem_free(map);
+ out_unlock_ilock:
+       xfs_iunlock_map_shared(ip, lock);
+ out_unlock_iolock:
+       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
+
+       for (i = 0; i < cur_ext; i++) {
+               int full = 0;   /* user array is full */
+
+               /* format results & advance arg */
+               error = formatter(&arg, &out[i], &full);
+               if (error || full)
+                       break;
        }
+
+       if (is_vmalloc_addr(out))
+               kmem_free_large(out);
+       else
+               kmem_free(out);
+       return error;
 }
 
 /*
@@ -6295,16 +6171,3 @@ next_block:
 
        return error;
 }
-
-/*
- * Convert the given file system block to a disk block.  We have to treat it
- * differently based on whether the file is a real time file or not, because the
- * bmap code does.
- */
-xfs_daddr_t
-xfs_fsb_to_db(struct xfs_inode *ip, xfs_fsblock_t fsb)
-{
-       return (XFS_IS_REALTIME_INODE(ip) ? \
-                (xfs_daddr_t)XFS_FSB_TO_BB((ip)->i_mount, (fsb)) : \
-                XFS_FSB_TO_DADDR((ip)->i_mount, (fsb)));
-}