fs/ext4/extents.c

   1 /*
   2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3  * Written by Alex Tomas <alex@clusterfs.com>
   4  *
   5  * Architecture independence:
   6  *   Copyright (c) 2005, Bull S.A.
   7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public Licens
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  21  */
  22
  23 /*
  24  * Extents support for EXT4
  25  *
  26  * TODO:
  27  *   - ext4*_error() should be used in some situations
  28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
  29  *   - smart tree reduction
  30  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/time.h>
  34 #include <linux/jbd2.h>
  35 #include <linux/highuid.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/falloc.h>
  41 #include <asm/uaccess.h>
  42 #include <linux/fiemap.h>
  43 #include "ext4_jbd2.h"
  44 #include "ext4_extents.h"
  45 #include "xattr.h"
  46
  47 #include <trace/events/ext4.h>
  48
  49 /*
  50  * used by extent splitting.
  51  */
  52 #define EXT4_EXT_MAY_ZEROOUT    0x1  /* safe to zeroout if split fails \
  53                                         due to ENOSPC */
  54 #define EXT4_EXT_MARK_UNINIT1   0x2  /* mark first half uninitialized */
  55 #define EXT4_EXT_MARK_UNINIT2   0x4  /* mark second half uninitialized */
  56
  57 #define EXT4_EXT_DATA_VALID1    0x8  /* first half contains valid data */
  58 #define EXT4_EXT_DATA_VALID2    0x10 /* second half contains valid data */
  59
  60 static __le32 ext4_extent_block_csum(struct inode *inode,
  61                                      struct ext4_extent_header *eh)
  62 {
  63         struct ext4_inode_info *ei = EXT4_I(inode);
  64         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
  65         __u32 csum;
  66
  67         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
  68                            EXT4_EXTENT_TAIL_OFFSET(eh));
  69         return cpu_to_le32(csum);
  70 }
  71
  72 static int ext4_extent_block_csum_verify(struct inode *inode,
  73                                          struct ext4_extent_header *eh)
  74 {
  75         struct ext4_extent_tail *et;
  76
  77         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  78                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  79                 return 1;
  80
  81         et = find_ext4_extent_tail(eh);
  82         if (et->et_checksum != ext4_extent_block_csum(inode, eh))
  83                 return 0;
  84         return 1;
  85 }
  86
  87 static void ext4_extent_block_csum_set(struct inode *inode,
  88                                        struct ext4_extent_header *eh)
  89 {
  90         struct ext4_extent_tail *et;
  91
  92         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
  93                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
  94                 return;
  95
  96         et = find_ext4_extent_tail(eh);
  97         et->et_checksum = ext4_extent_block_csum(inode, eh);
  98 }
  99
 100 static int ext4_split_extent(handle_t *handle,
 101                                 struct inode *inode,
 102                                 struct ext4_ext_path *path,
 103                                 struct ext4_map_blocks *map,
 104                                 int split_flag,
 105                                 int flags);
 106
 107 static int ext4_split_extent_at(handle_t *handle,
 108                              struct inode *inode,
 109                              struct ext4_ext_path *path,
 110                              ext4_lblk_t split,
 111                              int split_flag,
 112                              int flags);
 113
 114 static int ext4_find_delayed_extent(struct inode *inode,
 115                                     struct extent_status *newes);
 116
 117 static int ext4_ext_truncate_extend_restart(handle_t *handle,
 118                                             struct inode *inode,
 119                                             int needed)
 120 {
 121         int err;
 122
 123         if (!ext4_handle_valid(handle))
 124                 return 0;
 125         if (handle->h_buffer_credits > needed)
 126                 return 0;
 127         err = ext4_journal_extend(handle, needed);
 128         if (err <= 0)
 129                 return err;
 130         err = ext4_truncate_restart_trans(handle, inode, needed);
 131         if (err == 0)
 132                 err = -EAGAIN;
 133
 134         return err;
 135 }
 136
 137 /*
 138  * could return:
 139  *  - EROFS
 140  *  - ENOMEM
 141  */
 142 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
 143                                 struct ext4_ext_path *path)
 144 {
 145         if (path->p_bh) {
 146                 /* path points to block */
 147                 return ext4_journal_get_write_access(handle, path->p_bh);
 148         }
 149         /* path points to leaf/index in inode body */
 150         /* we use in-core data, no need to protect them */
 151         return 0;
 152 }
 153
 154 /*
 155  * could return:
 156  *  - EROFS
 157  *  - ENOMEM
 158  *  - EIO
 159  */
 160 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
 161                      struct inode *inode, struct ext4_ext_path *path)
 162 {
 163         int err;
 164         if (path->p_bh) {
 165                 ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
 166                 /* path points to block */
 167                 err = __ext4_handle_dirty_metadata(where, line, handle,
 168                                                    inode, path->p_bh);
 169         } else {
 170                 /* path points to leaf/index in inode body */
 171                 err = ext4_mark_inode_dirty(handle, inode);
 172         }
 173         return err;
 174 }
 175
 176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
 177                               struct ext4_ext_path *path,
 178                               ext4_lblk_t block)
 179 {
 180         if (path) {
 181                 int depth = path->p_depth;
 182                 struct ext4_extent *ex;
 183
 184                 /*
 185                  * Try to predict block placement assuming that we are
 186                  * filling in a file which will eventually be
 187                  * non-sparse --- i.e., in the case of libbfd writing
 188                  * an ELF object sections out-of-order but in a way
 189                  * the eventually results in a contiguous object or
 190                  * executable file, or some database extending a table
 191                  * space file.  However, this is actually somewhat
 192                  * non-ideal if we are writing a sparse file such as
 193                  * qemu or KVM writing a raw image file that is going
 194                  * to stay fairly sparse, since it will end up
 195                  * fragmenting the file system's free space.  Maybe we
 196                  * should have some hueristics or some way to allow
 197                  * userspace to pass a hint to file system,
 198                  * especially if the latter case turns out to be
 199                  * common.
 200                  */
 201                 ex = path[depth].p_ext;
 202                 if (ex) {
 203                         ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
 204                         ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
 205
 206                         if (block > ext_block)
 207                                 return ext_pblk + (block - ext_block);
 208                         else
 209                                 return ext_pblk - (ext_block - block);
 210                 }
 211
 212                 /* it looks like index is empty;
 213                  * try to find starting block from index itself */
 214                 if (path[depth].p_bh)
 215                         return path[depth].p_bh->b_blocknr;
 216         }
 217
 218         /* OK. use inode's group */
 219         return ext4_inode_to_goal_block(inode);
 220 }
 221
 222 /*
 223  * Allocation for a meta data block
 224  */
 225 static ext4_fsblk_t
 226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
 227                         struct ext4_ext_path *path,
 228                         struct ext4_extent *ex, int *err, unsigned int flags)
 229 {
 230         ext4_fsblk_t goal, newblock;
 231
 232         goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
 233         newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
 234                                         NULL, err);
 235         return newblock;
 236 }
 237
 238 static inline int ext4_ext_space_block(struct inode *inode, int check)
 239 {
 240         int size;
 241
 242         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 243                         / sizeof(struct ext4_extent);
 244 #ifdef AGGRESSIVE_TEST
 245         if (!check && size > 6)
 246                 size = 6;
 247 #endif
 248         return size;
 249 }
 250
 251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
 252 {
 253         int size;
 254
 255         size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 256                         / sizeof(struct ext4_extent_idx);
 257 #ifdef AGGRESSIVE_TEST
 258         if (!check && size > 5)
 259                 size = 5;
 260 #endif
 261         return size;
 262 }
 263
 264 static inline int ext4_ext_space_root(struct inode *inode, int check)
 265 {
 266         int size;
 267
 268         size = sizeof(EXT4_I(inode)->i_data);
 269         size -= sizeof(struct ext4_extent_header);
 270         size /= sizeof(struct ext4_extent);
 271 #ifdef AGGRESSIVE_TEST
 272         if (!check && size > 3)
 273                 size = 3;
 274 #endif
 275         return size;
 276 }
 277
 278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
 279 {
 280         int size;
 281
 282         size = sizeof(EXT4_I(inode)->i_data);
 283         size -= sizeof(struct ext4_extent_header);
 284         size /= sizeof(struct ext4_extent_idx);
 285 #ifdef AGGRESSIVE_TEST
 286         if (!check && size > 4)
 287                 size = 4;
 288 #endif
 289         return size;
 290 }
 291
 292 /*
 293  * Calculate the number of metadata blocks needed
 294  * to allocate @blocks
 295  * Worse case is one block per extent
 296  */
 297 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
 298 {
 299         struct ext4_inode_info *ei = EXT4_I(inode);
 300         int idxs;
 301
 302         idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
 303                 / sizeof(struct ext4_extent_idx));
 304
 305         /*
 306          * If the new delayed allocation block is contiguous with the
 307          * previous da block, it can share index blocks with the
 308          * previous block, so we only need to allocate a new index
 309          * block every idxs leaf blocks.  At ldxs**2 blocks, we need
 310          * an additional index block, and at ldxs**3 blocks, yet
 311          * another index blocks.
 312          */
 313         if (ei->i_da_metadata_calc_len &&
 314             ei->i_da_metadata_calc_last_lblock+1 == lblock) {
 315                 int num = 0;
 316
 317                 if ((ei->i_da_metadata_calc_len % idxs) == 0)
 318                         num++;
 319                 if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
 320                         num++;
 321                 if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
 322                         num++;
 323                         ei->i_da_metadata_calc_len = 0;
 324                 } else
 325                         ei->i_da_metadata_calc_len++;
 326                 ei->i_da_metadata_calc_last_lblock++;
 327                 return num;
 328         }
 329
 330         /*
 331          * In the worst case we need a new set of index blocks at
 332          * every level of the inode's extent tree.
 333          */
 334         ei->i_da_metadata_calc_len = 1;
 335         ei->i_da_metadata_calc_last_lblock = lblock;
 336         return ext_depth(inode) + 1;
 337 }
 338
 339 static int
 340 ext4_ext_max_entries(struct inode *inode, int depth)
 341 {
 342         int max;
 343
 344         if (depth == ext_depth(inode)) {
 345                 if (depth == 0)
 346                         max = ext4_ext_space_root(inode, 1);
 347                 else
 348                         max = ext4_ext_space_root_idx(inode, 1);
 349         } else {
 350                 if (depth == 0)
 351                         max = ext4_ext_space_block(inode, 1);
 352                 else
 353                         max = ext4_ext_space_block_idx(inode, 1);
 354         }
 355
 356         return max;
 357 }
 358
 359 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
 360 {
 361         ext4_fsblk_t block = ext4_ext_pblock(ext);
 362         int len = ext4_ext_get_actual_len(ext);
 363         ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
 364         ext4_lblk_t last = lblock + len - 1;
 365
 366         if (lblock > last)
 367                 return 0;
 368         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
 369 }
 370
 371 static int ext4_valid_extent_idx(struct inode *inode,
 372                                 struct ext4_extent_idx *ext_idx)
 373 {
 374         ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
 375
 376         return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
 377 }
 378
 379 static int ext4_valid_extent_entries(struct inode *inode,
 380                                 struct ext4_extent_header *eh,
 381                                 int depth)
 382 {
 383         unsigned short entries;
 384         if (eh->eh_entries == 0)
 385                 return 1;
 386
 387         entries = le16_to_cpu(eh->eh_entries);
 388
 389         if (depth == 0) {
 390                 /* leaf entries */
 391                 struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
 392                 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
 393                 ext4_fsblk_t pblock = 0;
 394                 ext4_lblk_t lblock = 0;
 395                 ext4_lblk_t prev = 0;
 396                 int len = 0;
 397                 while (entries) {
 398                         if (!ext4_valid_extent(inode, ext))
 399                                 return 0;
 400
 401                         /* Check for overlapping extents */
 402                         lblock = le32_to_cpu(ext->ee_block);
 403                         len = ext4_ext_get_actual_len(ext);
 404                         if ((lblock <= prev) && prev) {
 405                                 pblock = ext4_ext_pblock(ext);
 406                                 es->s_last_error_block = cpu_to_le64(pblock);
 407                                 return 0;
 408                         }
 409                         ext++;
 410                         entries--;
 411                         prev = lblock + len - 1;
 412                 }
 413         } else {
 414                 struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
 415                 while (entries) {
 416                         if (!ext4_valid_extent_idx(inode, ext_idx))
 417                                 return 0;
 418                         ext_idx++;
 419                         entries--;
 420                 }
 421         }
 422         return 1;
 423 }
 424
 425 static int __ext4_ext_check(const char *function, unsigned int line,
 426                             struct inode *inode, struct ext4_extent_header *eh,
 427                             int depth)
 428 {
 429         const char *error_msg;
 430         int max = 0;
 431
 432         if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
 433                 error_msg = "invalid magic";
 434                 goto corrupted;
 435         }
 436         if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
 437                 error_msg = "unexpected eh_depth";
 438                 goto corrupted;
 439         }
 440         if (unlikely(eh->eh_max == 0)) {
 441                 error_msg = "invalid eh_max";
 442                 goto corrupted;
 443         }
 444         max = ext4_ext_max_entries(inode, depth);
 445         if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
 446                 error_msg = "too large eh_max";
 447                 goto corrupted;
 448         }
 449         if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
 450                 error_msg = "invalid eh_entries";
 451                 goto corrupted;
 452         }
 453         if (!ext4_valid_extent_entries(inode, eh, depth)) {
 454                 error_msg = "invalid extent entries";
 455                 goto corrupted;
 456         }
 457         /* Verify checksum on non-root extent tree nodes */
 458         if (ext_depth(inode) != depth &&
 459             !ext4_extent_block_csum_verify(inode, eh)) {
 460                 error_msg = "extent tree corrupted";
 461                 goto corrupted;
 462         }
 463         return 0;
 464
 465 corrupted:
 466         ext4_error_inode(inode, function, line, 0,
 467                         "bad header/extent: %s - magic %x, "
 468                         "entries %u, max %u(%u), depth %u(%u)",
 469                         error_msg, le16_to_cpu(eh->eh_magic),
 470                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
 471                         max, le16_to_cpu(eh->eh_depth), depth);
 472
 473         return -EIO;
 474 }
 475
 476 #define ext4_ext_check(inode, eh, depth)        \
 477         __ext4_ext_check(__func__, __LINE__, inode, eh, depth)
 478
 479 int ext4_ext_check_inode(struct inode *inode)
 480 {
 481         return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode));
 482 }
 483
 484 static int __ext4_ext_check_block(const char *function, unsigned int line,
 485                                   struct inode *inode,
 486                                   struct ext4_extent_header *eh,
 487                                   int depth,
 488                                   struct buffer_head *bh)
 489 {
 490         int ret;
 491
 492         if (buffer_verified(bh))
 493                 return 0;
 494         ret = ext4_ext_check(inode, eh, depth);
 495         if (ret)
 496                 return ret;
 497         set_buffer_verified(bh);
 498         return ret;
 499 }
 500
 501 #define ext4_ext_check_block(inode, eh, depth, bh)      \
 502         __ext4_ext_check_block(__func__, __LINE__, inode, eh, depth, bh)
 503
 504 #ifdef EXT_DEBUG
 505 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
 506 {
 507         int k, l = path->p_depth;
 508
 509         ext_debug("path:");
 510         for (k = 0; k <= l; k++, path++) {
 511                 if (path->p_idx) {
 512                   ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
 513                             ext4_idx_pblock(path->p_idx));
 514                 } else if (path->p_ext) {
 515                         ext_debug("  %d:[%d]%d:%llu ",
 516                                   le32_to_cpu(path->p_ext->ee_block),
 517                                   ext4_ext_is_uninitialized(path->p_ext),
 518                                   ext4_ext_get_actual_len(path->p_ext),
 519                                   ext4_ext_pblock(path->p_ext));
 520                 } else
 521                         ext_debug("  []");
 522         }
 523         ext_debug("\n");
 524 }
 525
 526 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
 527 {
 528         int depth = ext_depth(inode);
 529         struct ext4_extent_header *eh;
 530         struct ext4_extent *ex;
 531         int i;
 532
 533         if (!path)
 534                 return;
 535
 536         eh = path[depth].p_hdr;
 537         ex = EXT_FIRST_EXTENT(eh);
 538
 539         ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
 540
 541         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
 542                 ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
 543                           ext4_ext_is_uninitialized(ex),
 544                           ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
 545         }
 546         ext_debug("\n");
 547 }
 548
 549 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
 550                         ext4_fsblk_t newblock, int level)
 551 {
 552         int depth = ext_depth(inode);
 553         struct ext4_extent *ex;
 554
 555         if (depth != level) {
 556                 struct ext4_extent_idx *idx;
 557                 idx = path[level].p_idx;
 558                 while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
 559                         ext_debug("%d: move %d:%llu in new index %llu\n", level,
 560                                         le32_to_cpu(idx->ei_block),
 561                                         ext4_idx_pblock(idx),
 562                                         newblock);
 563                         idx++;
 564                 }
 565
 566                 return;
 567         }
 568
 569         ex = path[depth].p_ext;
 570         while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
 571                 ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
 572                                 le32_to_cpu(ex->ee_block),
 573                                 ext4_ext_pblock(ex),
 574                                 ext4_ext_is_uninitialized(ex),
 575                                 ext4_ext_get_actual_len(ex),
 576                                 newblock);
 577                 ex++;
 578         }
 579 }
 580
 581 #else
 582 #define ext4_ext_show_path(inode, path)
 583 #define ext4_ext_show_leaf(inode, path)
 584 #define ext4_ext_show_move(inode, path, newblock, level)
 585 #endif
 586
 587 void ext4_ext_drop_refs(struct ext4_ext_path *path)
 588 {
 589         int depth = path->p_depth;
 590         int i;
 591
 592         for (i = 0; i <= depth; i++, path++)
 593                 if (path->p_bh) {
 594                         brelse(path->p_bh);
 595                         path->p_bh = NULL;
 596                 }
 597 }
 598
 599 /*
 600  * ext4_ext_binsearch_idx:
 601  * binary search for the closest index of the given block
 602  * the header must be checked before calling this
 603  */
 604 static void
 605 ext4_ext_binsearch_idx(struct inode *inode,
 606                         struct ext4_ext_path *path, ext4_lblk_t block)
 607 {
 608         struct ext4_extent_header *eh = path->p_hdr;
 609         struct ext4_extent_idx *r, *l, *m;
 610
 611
 612         ext_debug("binsearch for %u(idx):  ", block);
 613
 614         l = EXT_FIRST_INDEX(eh) + 1;
 615         r = EXT_LAST_INDEX(eh);
 616         while (l <= r) {
 617                 m = l + (r - l) / 2;
 618                 if (block < le32_to_cpu(m->ei_block))
 619                         r = m - 1;
 620                 else
 621                         l = m + 1;
 622                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
 623                                 m, le32_to_cpu(m->ei_block),
 624                                 r, le32_to_cpu(r->ei_block));
 625         }
 626
 627         path->p_idx = l - 1;
 628         ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
 629                   ext4_idx_pblock(path->p_idx));
 630
 631 #ifdef CHECK_BINSEARCH
 632         {
 633                 struct ext4_extent_idx *chix, *ix;
 634                 int k;
 635
 636                 chix = ix = EXT_FIRST_INDEX(eh);
 637                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
 638                   if (k != 0 &&
 639                       le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
 640                                 printk(KERN_DEBUG "k=%d, ix=0x%p, "
 641                                        "first=0x%p\n", k,
 642                                        ix, EXT_FIRST_INDEX(eh));
 643                                 printk(KERN_DEBUG "%u <= %u\n",
 644                                        le32_to_cpu(ix->ei_block),
 645                                        le32_to_cpu(ix[-1].ei_block));
 646                         }
 647                         BUG_ON(k && le32_to_cpu(ix->ei_block)
 648                                            <= le32_to_cpu(ix[-1].ei_block));
 649                         if (block < le32_to_cpu(ix->ei_block))
 650                                 break;
 651                         chix = ix;
 652                 }
 653                 BUG_ON(chix != path->p_idx);
 654         }
 655 #endif
 656
 657 }
 658
 659 /*
 660  * ext4_ext_binsearch:
 661  * binary search for closest extent of the given block
 662  * the header must be checked before calling this
 663  */
 664 static void
 665 ext4_ext_binsearch(struct inode *inode,
 666                 struct ext4_ext_path *path, ext4_lblk_t block)
 667 {
 668         struct ext4_extent_header *eh = path->p_hdr;
 669         struct ext4_extent *r, *l, *m;
 670
 671         if (eh->eh_entries == 0) {
 672                 /*
 673                  * this leaf is empty:
 674                  * we get such a leaf in split/add case
 675                  */
 676                 return;
 677         }
 678
 679         ext_debug("binsearch for %u:  ", block);
 680
 681         l = EXT_FIRST_EXTENT(eh) + 1;
 682         r = EXT_LAST_EXTENT(eh);
 683
 684         while (l <= r) {
 685                 m = l + (r - l) / 2;
 686                 if (block < le32_to_cpu(m->ee_block))
 687                         r = m - 1;
 688                 else
 689                         l = m + 1;
 690                 ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
 691                                 m, le32_to_cpu(m->ee_block),
 692                                 r, le32_to_cpu(r->ee_block));
 693         }
 694
 695         path->p_ext = l - 1;
 696         ext_debug("  -> %d:%llu:[%d]%d ",
 697                         le32_to_cpu(path->p_ext->ee_block),
 698                         ext4_ext_pblock(path->p_ext),
 699                         ext4_ext_is_uninitialized(path->p_ext),
 700                         ext4_ext_get_actual_len(path->p_ext));
 701
 702 #ifdef CHECK_BINSEARCH
 703         {
 704                 struct ext4_extent *chex, *ex;
 705                 int k;
 706
 707                 chex = ex = EXT_FIRST_EXTENT(eh);
 708                 for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
 709                         BUG_ON(k && le32_to_cpu(ex->ee_block)
 710                                           <= le32_to_cpu(ex[-1].ee_block));
 711                         if (block < le32_to_cpu(ex->ee_block))
 712                                 break;
 713                         chex = ex;
 714                 }
 715                 BUG_ON(chex != path->p_ext);
 716         }
 717 #endif
 718
 719 }
 720
 721 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
 722 {
 723         struct ext4_extent_header *eh;
 724
 725         eh = ext_inode_hdr(inode);
 726         eh->eh_depth = 0;
 727         eh->eh_entries = 0;
 728         eh->eh_magic = EXT4_EXT_MAGIC;
 729         eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
 730         ext4_mark_inode_dirty(handle, inode);
 731         return 0;
 732 }
 733
 734 struct ext4_ext_path *
 735 ext4_ext_find_extent(struct inode *inode, ext4_lblk_t block,
 736                                         struct ext4_ext_path *path)
 737 {
 738         struct ext4_extent_header *eh;
 739         struct buffer_head *bh;
 740         short int depth, i, ppos = 0, alloc = 0;
 741         int ret;
 742
 743         eh = ext_inode_hdr(inode);
 744         depth = ext_depth(inode);
 745
 746         /* account possible depth increase */
 747         if (!path) {
 748                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
 749                                 GFP_NOFS);
 750                 if (!path)
 751                         return ERR_PTR(-ENOMEM);
 752                 alloc = 1;
 753         }
 754         path[0].p_hdr = eh;
 755         path[0].p_bh = NULL;
 756
 757         i = depth;
 758         /* walk through the tree */
 759         while (i) {
 760                 ext_debug("depth %d: num %d, max %d\n",
 761                           ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
 762
 763                 ext4_ext_binsearch_idx(inode, path + ppos, block);
 764                 path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
 765                 path[ppos].p_depth = i;
 766                 path[ppos].p_ext = NULL;
 767
 768                 bh = sb_getblk(inode->i_sb, path[ppos].p_block);
 769                 if (unlikely(!bh)) {
 770                         ret = -ENOMEM;
 771                         goto err;
 772                 }
 773                 if (!bh_uptodate_or_lock(bh)) {
 774                         trace_ext4_ext_load_extent(inode, block,
 775                                                 path[ppos].p_block);
 776                         ret = bh_submit_read(bh);
 777                         if (ret < 0) {
 778                                 put_bh(bh);
 779                                 goto err;
 780                         }
 781                 }
 782                 eh = ext_block_hdr(bh);
 783                 ppos++;
 784                 if (unlikely(ppos > depth)) {
 785                         put_bh(bh);
 786                         EXT4_ERROR_INODE(inode,
 787                                          "ppos %d > depth %d", ppos, depth);
 788                         ret = -EIO;
 789                         goto err;
 790                 }
 791                 path[ppos].p_bh = bh;
 792                 path[ppos].p_hdr = eh;
 793                 i--;
 794
 795                 ret = ext4_ext_check_block(inode, eh, i, bh);
 796                 if (ret < 0)
 797                         goto err;
 798         }
 799
 800         path[ppos].p_depth = i;
 801         path[ppos].p_ext = NULL;
 802         path[ppos].p_idx = NULL;
 803
 804         /* find extent */
 805         ext4_ext_binsearch(inode, path + ppos, block);
 806         /* if not an empty leaf */
 807         if (path[ppos].p_ext)
 808                 path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
 809
 810         ext4_ext_show_path(inode, path);
 811
 812         return path;
 813
 814 err:
 815         ext4_ext_drop_refs(path);
 816         if (alloc)
 817                 kfree(path);
 818         return ERR_PTR(ret);
 819 }
 820
 821 /*
 822  * ext4_ext_insert_index:
 823  * insert new index [@logical;@ptr] into the block at @curp;
 824  * check where to insert: before @curp or after @curp
 825  */
 826 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
 827                                  struct ext4_ext_path *curp,
 828                                  int logical, ext4_fsblk_t ptr)
 829 {
 830         struct ext4_extent_idx *ix;
 831         int len, err;
 832
 833         err = ext4_ext_get_access(handle, inode, curp);
 834         if (err)
 835                 return err;
 836
 837         if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
 838                 EXT4_ERROR_INODE(inode,
 839                                  "logical %d == ei_block %d!",
 840                                  logical, le32_to_cpu(curp->p_idx->ei_block));
 841                 return -EIO;
 842         }
 843
 844         if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
 845                              >= le16_to_cpu(curp->p_hdr->eh_max))) {
 846                 EXT4_ERROR_INODE(inode,
 847                                  "eh_entries %d >= eh_max %d!",
 848                                  le16_to_cpu(curp->p_hdr->eh_entries),
 849                                  le16_to_cpu(curp->p_hdr->eh_max));
 850                 return -EIO;
 851         }
 852
 853         if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
 854                 /* insert after */
 855                 ext_debug("insert new index %d after: %llu\n", logical, ptr);
 856                 ix = curp->p_idx + 1;
 857         } else {
 858                 /* insert before */
 859                 ext_debug("insert new index %d before: %llu\n", logical, ptr);
 860                 ix = curp->p_idx;
 861         }
 862
 863         len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
 864         BUG_ON(len < 0);
 865         if (len > 0) {
 866                 ext_debug("insert new index %d: "
 867                                 "move %d indices from 0x%p to 0x%p\n",
 868                                 logical, len, ix, ix + 1);
 869                 memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
 870         }
 871
 872         if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
 873                 EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
 874                 return -EIO;
 875         }
 876
 877         ix->ei_block = cpu_to_le32(logical);
 878         ext4_idx_store_pblock(ix, ptr);
 879         le16_add_cpu(&curp->p_hdr->eh_entries, 1);
 880
 881         if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
 882                 EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
 883                 return -EIO;
 884         }
 885
 886         err = ext4_ext_dirty(handle, inode, curp);
 887         ext4_std_error(inode->i_sb, err);
 888
 889         return err;
 890 }
 891
 892 /*
 893  * ext4_ext_split:
 894  * inserts new subtree into the path, using free index entry
 895  * at depth @at:
 896  * - allocates all needed blocks (new leaf and all intermediate index blocks)
 897  * - makes decision where to split
 898  * - moves remaining extents and index entries (right to the split point)
 899  *   into the newly allocated blocks
 900  * - initializes subtree
 901  */
 902 static int ext4_ext_split(handle_t *handle, struct inode *inode,
 903                           unsigned int flags,
 904                           struct ext4_ext_path *path,
 905                           struct ext4_extent *newext, int at)
 906 {
 907         struct buffer_head *bh = NULL;
 908         int depth = ext_depth(inode);
 909         struct ext4_extent_header *neh;
 910         struct ext4_extent_idx *fidx;
 911         int i = at, k, m, a;
 912         ext4_fsblk_t newblock, oldblock;
 913         __le32 border;
 914         ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
 915         int err = 0;
 916
 917         /* make decision: where to split? */
 918         /* FIXME: now decision is simplest: at current extent */
 919
 920         /* if current leaf will be split, then we should use
 921          * border from split point */
 922         if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
 923                 EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
 924                 return -EIO;
 925         }
 926         if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
 927                 border = path[depth].p_ext[1].ee_block;
 928                 ext_debug("leaf will be split."
 929                                 " next leaf starts at %d\n",
 930                                   le32_to_cpu(border));
 931         } else {
 932                 border = newext->ee_block;
 933                 ext_debug("leaf will be added."
 934                                 " next leaf starts at %d\n",
 935                                 le32_to_cpu(border));
 936         }
 937
 938         /*
 939          * If error occurs, then we break processing
 940          * and mark filesystem read-only. index won't
 941          * be inserted and tree will be in consistent
 942          * state. Next mount will repair buffers too.
 943          */
 944
 945         /*
 946          * Get array to track all allocated blocks.
 947          * We need this to handle errors and free blocks
 948          * upon them.
 949          */
 950         ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
 951         if (!ablocks)
 952                 return -ENOMEM;
 953
 954         /* allocate all needed blocks */
 955         ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
 956         for (a = 0; a < depth - at; a++) {
 957                 newblock = ext4_ext_new_meta_block(handle, inode, path,
 958                                                    newext, &err, flags);
 959                 if (newblock == 0)
 960                         goto cleanup;
 961                 ablocks[a] = newblock;
 962         }
 963
 964         /* initialize new leaf */
 965         newblock = ablocks[--a];
 966         if (unlikely(newblock == 0)) {
 967                 EXT4_ERROR_INODE(inode, "newblock == 0!");
 968                 err = -EIO;
 969                 goto cleanup;
 970         }
 971         bh = sb_getblk(inode->i_sb, newblock);
 972         if (unlikely(!bh)) {
 973                 err = -ENOMEM;
 974                 goto cleanup;
 975         }
 976         lock_buffer(bh);
 977
 978         err = ext4_journal_get_create_access(handle, bh);
 979         if (err)
 980                 goto cleanup;
 981
 982         neh = ext_block_hdr(bh);
 983         neh->eh_entries = 0;
 984         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
 985         neh->eh_magic = EXT4_EXT_MAGIC;
 986         neh->eh_depth = 0;
 987
 988         /* move remainder of path[depth] to the new leaf */
 989         if (unlikely(path[depth].p_hdr->eh_entries !=
 990                      path[depth].p_hdr->eh_max)) {
 991                 EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
 992                                  path[depth].p_hdr->eh_entries,
 993                                  path[depth].p_hdr->eh_max);
 994                 err = -EIO;
 995                 goto cleanup;
 996         }
 997         /* start copy from next extent */
 998         m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
 999         ext4_ext_show_move(inode, path, newblock, depth);
1000         if (m) {
1001                 struct ext4_extent *ex;
1002                 ex = EXT_FIRST_EXTENT(neh);
1003                 memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1004                 le16_add_cpu(&neh->eh_entries, m);
1005         }
1006
1007         ext4_extent_block_csum_set(inode, neh);
1008         set_buffer_uptodate(bh);
1009         unlock_buffer(bh);
1010
1011         err = ext4_handle_dirty_metadata(handle, inode, bh);
1012         if (err)
1013                 goto cleanup;
1014         brelse(bh);
1015         bh = NULL;
1016
1017         /* correct old leaf */
1018         if (m) {
1019                 err = ext4_ext_get_access(handle, inode, path + depth);
1020                 if (err)
1021                         goto cleanup;
1022                 le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1023                 err = ext4_ext_dirty(handle, inode, path + depth);
1024                 if (err)
1025                         goto cleanup;
1026
1027         }
1028
1029         /* create intermediate indexes */
1030         k = depth - at - 1;
1031         if (unlikely(k < 0)) {
1032                 EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1033                 err = -EIO;
1034                 goto cleanup;
1035         }
1036         if (k)
1037                 ext_debug("create %d intermediate indices\n", k);
1038         /* insert new index into current index block */
1039         /* current depth stored in i var */
1040         i = depth - 1;
1041         while (k--) {
1042                 oldblock = newblock;
1043                 newblock = ablocks[--a];
1044                 bh = sb_getblk(inode->i_sb, newblock);
1045                 if (unlikely(!bh)) {
1046                         err = -ENOMEM;
1047                         goto cleanup;
1048                 }
1049                 lock_buffer(bh);
1050
1051                 err = ext4_journal_get_create_access(handle, bh);
1052                 if (err)
1053                         goto cleanup;
1054
1055                 neh = ext_block_hdr(bh);
1056                 neh->eh_entries = cpu_to_le16(1);
1057                 neh->eh_magic = EXT4_EXT_MAGIC;
1058                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1059                 neh->eh_depth = cpu_to_le16(depth - i);
1060                 fidx = EXT_FIRST_INDEX(neh);
1061                 fidx->ei_block = border;
1062                 ext4_idx_store_pblock(fidx, oldblock);
1063
1064                 ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1065                                 i, newblock, le32_to_cpu(border), oldblock);
1066
1067                 /* move remainder of path[i] to the new index block */
1068                 if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1069                                         EXT_LAST_INDEX(path[i].p_hdr))) {
1070                         EXT4_ERROR_INODE(inode,
1071                                          "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1072                                          le32_to_cpu(path[i].p_ext->ee_block));
1073                         err = -EIO;
1074                         goto cleanup;
1075                 }
1076                 /* start copy indexes */
1077                 m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1078                 ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1079                                 EXT_MAX_INDEX(path[i].p_hdr));
1080                 ext4_ext_show_move(inode, path, newblock, i);
1081                 if (m) {
1082                         memmove(++fidx, path[i].p_idx,
1083                                 sizeof(struct ext4_extent_idx) * m);
1084                         le16_add_cpu(&neh->eh_entries, m);
1085                 }
1086                 ext4_extent_block_csum_set(inode, neh);
1087                 set_buffer_uptodate(bh);
1088                 unlock_buffer(bh);
1089
1090                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1091                 if (err)
1092                         goto cleanup;
1093                 brelse(bh);
1094                 bh = NULL;
1095
1096                 /* correct old index */
1097                 if (m) {
1098                         err = ext4_ext_get_access(handle, inode, path + i);
1099                         if (err)
1100                                 goto cleanup;
1101                         le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1102                         err = ext4_ext_dirty(handle, inode, path + i);
1103                         if (err)
1104                                 goto cleanup;
1105                 }
1106
1107                 i--;
1108         }
1109
1110         /* insert new index */
1111         err = ext4_ext_insert_index(handle, inode, path + at,
1112                                     le32_to_cpu(border), newblock);
1113
1114 cleanup:
1115         if (bh) {
1116                 if (buffer_locked(bh))
1117                         unlock_buffer(bh);
1118                 brelse(bh);
1119         }
1120
1121         if (err) {
1122                 /* free all allocated blocks in error case */
1123                 for (i = 0; i < depth; i++) {
1124                         if (!ablocks[i])
1125                                 continue;
1126                         ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1127                                          EXT4_FREE_BLOCKS_METADATA);
1128                 }
1129         }
1130         kfree(ablocks);
1131
1132         return err;
1133 }
1134
1135 /*
1136  * ext4_ext_grow_indepth:
1137  * implements tree growing procedure:
1138  * - allocates new block
1139  * - moves top-level data (index block or leaf) into the new block
1140  * - initializes new top-level, creating index that points to the
1141  *   just created block
1142  */
1143 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1144                                  unsigned int flags,
1145                                  struct ext4_extent *newext)
1146 {
1147         struct ext4_extent_header *neh;
1148         struct buffer_head *bh;
1149         ext4_fsblk_t newblock;
1150         int err = 0;
1151
1152         newblock = ext4_ext_new_meta_block(handle, inode, NULL,
1153                 newext, &err, flags);
1154         if (newblock == 0)
1155                 return err;
1156
1157         bh = sb_getblk(inode->i_sb, newblock);
1158         if (unlikely(!bh))
1159                 return -ENOMEM;
1160         lock_buffer(bh);
1161
1162         err = ext4_journal_get_create_access(handle, bh);
1163         if (err) {
1164                 unlock_buffer(bh);
1165                 goto out;
1166         }
1167
1168         /* move top-level index/leaf into new block */
1169         memmove(bh->b_data, EXT4_I(inode)->i_data,
1170                 sizeof(EXT4_I(inode)->i_data));
1171
1172         /* set size of new block */
1173         neh = ext_block_hdr(bh);
1174         /* old root could have indexes or leaves
1175          * so calculate e_max right way */
1176         if (ext_depth(inode))
1177                 neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1178         else
1179                 neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1180         neh->eh_magic = EXT4_EXT_MAGIC;
1181         ext4_extent_block_csum_set(inode, neh);
1182         set_buffer_uptodate(bh);
1183         unlock_buffer(bh);
1184
1185         err = ext4_handle_dirty_metadata(handle, inode, bh);
1186         if (err)
1187                 goto out;
1188
1189         /* Update top-level index: num,max,pointer */
1190         neh = ext_inode_hdr(inode);
1191         neh->eh_entries = cpu_to_le16(1);
1192         ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1193         if (neh->eh_depth == 0) {
1194                 /* Root extent block becomes index block */
1195                 neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1196                 EXT_FIRST_INDEX(neh)->ei_block =
1197                         EXT_FIRST_EXTENT(neh)->ee_block;
1198         }
1199         ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1200                   le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1201                   le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1202                   ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1203
1204         le16_add_cpu(&neh->eh_depth, 1);
1205         ext4_mark_inode_dirty(handle, inode);
1206 out:
1207         brelse(bh);
1208
1209         return err;
1210 }
1211
1212 /*
1213  * ext4_ext_create_new_leaf:
1214  * finds empty index and adds new leaf.
1215  * if no free index is found, then it requests in-depth growing.
1216  */
1217 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1218                                     unsigned int flags,
1219                                     struct ext4_ext_path *path,
1220                                     struct ext4_extent *newext)
1221 {
1222         struct ext4_ext_path *curp;
1223         int depth, i, err = 0;
1224
1225 repeat:
1226         i = depth = ext_depth(inode);
1227
1228         /* walk up to the tree and look for free index entry */
1229         curp = path + depth;
1230         while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1231                 i--;
1232                 curp--;
1233         }
1234
1235         /* we use already allocated block for index block,
1236          * so subsequent data blocks should be contiguous */
1237         if (EXT_HAS_FREE_INDEX(curp)) {
1238                 /* if we found index with free entry, then use that
1239                  * entry: create all needed subtree and add new leaf */
1240                 err = ext4_ext_split(handle, inode, flags, path, newext, i);
1241                 if (err)
1242                         goto out;
1243
1244                 /* refill path */
1245                 ext4_ext_drop_refs(path);
1246                 path = ext4_ext_find_extent(inode,
1247                                     (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1248                                     path);
1249                 if (IS_ERR(path))
1250                         err = PTR_ERR(path);
1251         } else {
1252                 /* tree is full, time to grow in depth */
1253                 err = ext4_ext_grow_indepth(handle, inode, flags, newext);
1254                 if (err)
1255                         goto out;
1256
1257                 /* refill path */
1258                 ext4_ext_drop_refs(path);
1259                 path = ext4_ext_find_extent(inode,
1260                                    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1261                                     path);
1262                 if (IS_ERR(path)) {
1263                         err = PTR_ERR(path);
1264                         goto out;
1265                 }
1266
1267                 /*
1268                  * only first (depth 0 -> 1) produces free space;
1269                  * in all other cases we have to split the grown tree
1270                  */
1271                 depth = ext_depth(inode);
1272                 if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1273                         /* now we need to split */
1274                         goto repeat;
1275                 }
1276         }
1277
1278 out:
1279         return err;
1280 }
1281
1282 /*
1283  * search the closest allocated block to the left for *logical
1284  * and returns it at @logical + it's physical address at @phys
1285  * if *logical is the smallest allocated block, the function
1286  * returns 0 at @phys
1287  * return value contains 0 (success) or error code
1288  */
1289 static int ext4_ext_search_left(struct inode *inode,
1290                                 struct ext4_ext_path *path,
1291                                 ext4_lblk_t *logical, ext4_fsblk_t *phys)
1292 {
1293         struct ext4_extent_idx *ix;
1294         struct ext4_extent *ex;
1295         int depth, ee_len;
1296
1297         if (unlikely(path == NULL)) {
1298                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1299                 return -EIO;
1300         }
1301         depth = path->p_depth;
1302         *phys = 0;
1303
1304         if (depth == 0 && path->p_ext == NULL)
1305                 return 0;
1306
1307         /* usually extent in the path covers blocks smaller
1308          * then *logical, but it can be that extent is the
1309          * first one in the file */
1310
1311         ex = path[depth].p_ext;
1312         ee_len = ext4_ext_get_actual_len(ex);
1313         if (*logical < le32_to_cpu(ex->ee_block)) {
1314                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1315                         EXT4_ERROR_INODE(inode,
1316                                          "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1317                                          *logical, le32_to_cpu(ex->ee_block));
1318                         return -EIO;
1319                 }
1320                 while (--depth >= 0) {
1321                         ix = path[depth].p_idx;
1322                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1323                                 EXT4_ERROR_INODE(inode,
1324                                   "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1325                                   ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1326                                   EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1327                 le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1328                                   depth);
1329                                 return -EIO;
1330                         }
1331                 }
1332                 return 0;
1333         }
1334
1335         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1336                 EXT4_ERROR_INODE(inode,
1337                                  "logical %d < ee_block %d + ee_len %d!",
1338                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1339                 return -EIO;
1340         }
1341
1342         *logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1343         *phys = ext4_ext_pblock(ex) + ee_len - 1;
1344         return 0;
1345 }
1346
1347 /*
1348  * search the closest allocated block to the right for *logical
1349  * and returns it at @logical + it's physical address at @phys
1350  * if *logical is the largest allocated block, the function
1351  * returns 0 at @phys
1352  * return value contains 0 (success) or error code
1353  */
1354 static int ext4_ext_search_right(struct inode *inode,
1355                                  struct ext4_ext_path *path,
1356                                  ext4_lblk_t *logical, ext4_fsblk_t *phys,
1357                                  struct ext4_extent **ret_ex)
1358 {
1359         struct buffer_head *bh = NULL;
1360         struct ext4_extent_header *eh;
1361         struct ext4_extent_idx *ix;
1362         struct ext4_extent *ex;
1363         ext4_fsblk_t block;
1364         int depth;      /* Note, NOT eh_depth; depth from top of tree */
1365         int ee_len;
1366
1367         if (unlikely(path == NULL)) {
1368                 EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1369                 return -EIO;
1370         }
1371         depth = path->p_depth;
1372         *phys = 0;
1373
1374         if (depth == 0 && path->p_ext == NULL)
1375                 return 0;
1376
1377         /* usually extent in the path covers blocks smaller
1378          * then *logical, but it can be that extent is the
1379          * first one in the file */
1380
1381         ex = path[depth].p_ext;
1382         ee_len = ext4_ext_get_actual_len(ex);
1383         if (*logical < le32_to_cpu(ex->ee_block)) {
1384                 if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1385                         EXT4_ERROR_INODE(inode,
1386                                          "first_extent(path[%d].p_hdr) != ex",
1387                                          depth);
1388                         return -EIO;
1389                 }
1390                 while (--depth >= 0) {
1391                         ix = path[depth].p_idx;
1392                         if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1393                                 EXT4_ERROR_INODE(inode,
1394                                                  "ix != EXT_FIRST_INDEX *logical %d!",
1395                                                  *logical);
1396                                 return -EIO;
1397                         }
1398                 }
1399                 goto found_extent;
1400         }
1401
1402         if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1403                 EXT4_ERROR_INODE(inode,
1404                                  "logical %d < ee_block %d + ee_len %d!",
1405                                  *logical, le32_to_cpu(ex->ee_block), ee_len);
1406                 return -EIO;
1407         }
1408
1409         if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1410                 /* next allocated block in this leaf */
1411                 ex++;
1412                 goto found_extent;
1413         }
1414
1415         /* go up and search for index to the right */
1416         while (--depth >= 0) {
1417                 ix = path[depth].p_idx;
1418                 if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1419                         goto got_index;
1420         }
1421
1422         /* we've gone up to the root and found no index to the right */
1423         return 0;
1424
1425 got_index:
1426         /* we've found index to the right, let's
1427          * follow it and find the closest allocated
1428          * block to the right */
1429         ix++;
1430         block = ext4_idx_pblock(ix);
1431         while (++depth < path->p_depth) {
1432                 bh = sb_bread(inode->i_sb, block);
1433                 if (bh == NULL)
1434                         return -EIO;
1435                 eh = ext_block_hdr(bh);
1436                 /* subtract from p_depth to get proper eh_depth */
1437                 if (ext4_ext_check_block(inode, eh,
1438                                          path->p_depth - depth, bh)) {
1439                         put_bh(bh);
1440                         return -EIO;
1441                 }
1442                 ix = EXT_FIRST_INDEX(eh);
1443                 block = ext4_idx_pblock(ix);
1444                 put_bh(bh);
1445         }
1446
1447         bh = sb_bread(inode->i_sb, block);
1448         if (bh == NULL)
1449                 return -EIO;
1450         eh = ext_block_hdr(bh);
1451         if (ext4_ext_check_block(inode, eh, path->p_depth - depth, bh)) {
1452                 put_bh(bh);
1453                 return -EIO;
1454         }
1455         ex = EXT_FIRST_EXTENT(eh);
1456 found_extent:
1457         *logical = le32_to_cpu(ex->ee_block);
1458         *phys = ext4_ext_pblock(ex);
1459         *ret_ex = ex;
1460         if (bh)
1461                 put_bh(bh);
1462         return 0;
1463 }
1464
1465 /*
1466  * ext4_ext_next_allocated_block:
1467  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1468  * NOTE: it considers block number from index entry as
1469  * allocated block. Thus, index entries have to be consistent
1470  * with leaves.
1471  */
1472 static ext4_lblk_t
1473 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1474 {
1475         int depth;
1476
1477         BUG_ON(path == NULL);
1478         depth = path->p_depth;
1479
1480         if (depth == 0 && path->p_ext == NULL)
1481                 return EXT_MAX_BLOCKS;
1482
1483         while (depth >= 0) {
1484                 if (depth == path->p_depth) {
1485                         /* leaf */
1486                         if (path[depth].p_ext &&
1487                                 path[depth].p_ext !=
1488                                         EXT_LAST_EXTENT(path[depth].p_hdr))
1489                           return le32_to_cpu(path[depth].p_ext[1].ee_block);
1490                 } else {
1491                         /* index */
1492                         if (path[depth].p_idx !=
1493                                         EXT_LAST_INDEX(path[depth].p_hdr))
1494                           return le32_to_cpu(path[depth].p_idx[1].ei_block);
1495                 }
1496                 depth--;
1497         }
1498
1499         return EXT_MAX_BLOCKS;
1500 }
1501
1502 /*
1503  * ext4_ext_next_leaf_block:
1504  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1505  */
1506 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1507 {
1508         int depth;
1509
1510         BUG_ON(path == NULL);
1511         depth = path->p_depth;
1512
1513         /* zero-tree has no leaf blocks at all */
1514         if (depth == 0)
1515                 return EXT_MAX_BLOCKS;
1516
1517         /* go to index block */
1518         depth--;
1519
1520         while (depth >= 0) {
1521                 if (path[depth].p_idx !=
1522                                 EXT_LAST_INDEX(path[depth].p_hdr))
1523                         return (ext4_lblk_t)
1524                                 le32_to_cpu(path[depth].p_idx[1].ei_block);
1525                 depth--;
1526         }
1527
1528         return EXT_MAX_BLOCKS;
1529 }
1530
1531 /*
1532  * ext4_ext_correct_indexes:
1533  * if leaf gets modified and modified extent is first in the leaf,
1534  * then we have to correct all indexes above.
1535  * TODO: do we need to correct tree in all cases?
1536  */
1537 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1538                                 struct ext4_ext_path *path)
1539 {
1540         struct ext4_extent_header *eh;
1541         int depth = ext_depth(inode);
1542         struct ext4_extent *ex;
1543         __le32 border;
1544         int k, err = 0;
1545
1546         eh = path[depth].p_hdr;
1547         ex = path[depth].p_ext;
1548
1549         if (unlikely(ex == NULL || eh == NULL)) {
1550                 EXT4_ERROR_INODE(inode,
1551                                  "ex %p == NULL or eh %p == NULL", ex, eh);
1552                 return -EIO;
1553         }
1554
1555         if (depth == 0) {
1556                 /* there is no tree at all */
1557                 return 0;
1558         }
1559
1560         if (ex != EXT_FIRST_EXTENT(eh)) {
1561                 /* we correct tree if first leaf got modified only */
1562                 return 0;
1563         }
1564
1565         /*
1566          * TODO: we need correction if border is smaller than current one
1567          */
1568         k = depth - 1;
1569         border = path[depth].p_ext->ee_block;
1570         err = ext4_ext_get_access(handle, inode, path + k);
1571         if (err)
1572                 return err;
1573         path[k].p_idx->ei_block = border;
1574         err = ext4_ext_dirty(handle, inode, path + k);
1575         if (err)
1576                 return err;
1577
1578         while (k--) {
1579                 /* change all left-side indexes */
1580                 if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1581                         break;
1582                 err = ext4_ext_get_access(handle, inode, path + k);
1583                 if (err)
1584                         break;
1585                 path[k].p_idx->ei_block = border;
1586                 err = ext4_ext_dirty(handle, inode, path + k);
1587                 if (err)
1588                         break;
1589         }
1590
1591         return err;
1592 }
1593
1594 int
1595 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1596                                 struct ext4_extent *ex2)
1597 {
1598         unsigned short ext1_ee_len, ext2_ee_len, max_len;
1599
1600         /*
1601          * Make sure that both extents are initialized. We don't merge
1602          * uninitialized extents so that we can be sure that end_io code has
1603          * the extent that was written properly split out and conversion to
1604          * initialized is trivial.
1605          */
1606         if (ext4_ext_is_uninitialized(ex1) || ext4_ext_is_uninitialized(ex2))
1607                 return 0;
1608
1609         if (ext4_ext_is_uninitialized(ex1))
1610                 max_len = EXT_UNINIT_MAX_LEN;
1611         else
1612                 max_len = EXT_INIT_MAX_LEN;
1613
1614         ext1_ee_len = ext4_ext_get_actual_len(ex1);
1615         ext2_ee_len = ext4_ext_get_actual_len(ex2);
1616
1617         if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1618                         le32_to_cpu(ex2->ee_block))
1619                 return 0;
1620
1621         /*
1622          * To allow future support for preallocated extents to be added
1623          * as an RO_COMPAT feature, refuse to merge to extents if
1624          * this can result in the top bit of ee_len being set.
1625          */
1626         if (ext1_ee_len + ext2_ee_len > max_len)
1627                 return 0;
1628 #ifdef AGGRESSIVE_TEST
1629         if (ext1_ee_len >= 4)
1630                 return 0;
1631 #endif
1632
1633         if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1634                 return 1;
1635         return 0;
1636 }
1637
1638 /*
1639  * This function tries to merge the "ex" extent to the next extent in the tree.
1640  * It always tries to merge towards right. If you want to merge towards
1641  * left, pass "ex - 1" as argument instead of "ex".
1642  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1643  * 1 if they got merged.
1644  */
1645 static int ext4_ext_try_to_merge_right(struct inode *inode,
1646                                  struct ext4_ext_path *path,
1647                                  struct ext4_extent *ex)
1648 {
1649         struct ext4_extent_header *eh;
1650         unsigned int depth, len;
1651         int merge_done = 0;
1652         int uninitialized = 0;
1653
1654         depth = ext_depth(inode);
1655         BUG_ON(path[depth].p_hdr == NULL);
1656         eh = path[depth].p_hdr;
1657
1658         while (ex < EXT_LAST_EXTENT(eh)) {
1659                 if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1660                         break;
1661                 /* merge with next extent! */
1662                 if (ext4_ext_is_uninitialized(ex))
1663                         uninitialized = 1;
1664                 ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1665                                 + ext4_ext_get_actual_len(ex + 1));
1666                 if (uninitialized)
1667                         ext4_ext_mark_uninitialized(ex);
1668
1669                 if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1670                         len = (EXT_LAST_EXTENT(eh) - ex - 1)
1671                                 * sizeof(struct ext4_extent);
1672                         memmove(ex + 1, ex + 2, len);
1673                 }
1674                 le16_add_cpu(&eh->eh_entries, -1);
1675                 merge_done = 1;
1676                 WARN_ON(eh->eh_entries == 0);
1677                 if (!eh->eh_entries)
1678                         EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1679         }
1680
1681         return merge_done;
1682 }
1683
1684 /*
1685  * This function does a very simple check to see if we can collapse
1686  * an extent tree with a single extent tree leaf block into the inode.
1687  */
1688 static void ext4_ext_try_to_merge_up(handle_t *handle,
1689                                      struct inode *inode,
1690                                      struct ext4_ext_path *path)
1691 {
1692         size_t s;
1693         unsigned max_root = ext4_ext_space_root(inode, 0);
1694         ext4_fsblk_t blk;
1695
1696         if ((path[0].p_depth != 1) ||
1697             (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1698             (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1699                 return;
1700
1701         /*
1702          * We need to modify the block allocation bitmap and the block
1703          * group descriptor to release the extent tree block.  If we
1704          * can't get the journal credits, give up.
1705          */
1706         if (ext4_journal_extend(handle, 2))
1707                 return;
1708
1709         /*
1710          * Copy the extent data up to the inode
1711          */
1712         blk = ext4_idx_pblock(path[0].p_idx);
1713         s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1714                 sizeof(struct ext4_extent_idx);
1715         s += sizeof(struct ext4_extent_header);
1716
1717         memcpy(path[0].p_hdr, path[1].p_hdr, s);
1718         path[0].p_depth = 0;
1719         path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1720                 (path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1721         path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1722
1723         brelse(path[1].p_bh);
1724         ext4_free_blocks(handle, inode, NULL, blk, 1,
1725                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1726 }
1727
1728 /*
1729  * This function tries to merge the @ex extent to neighbours in the tree.
1730  * return 1 if merge left else 0.
1731  */
1732 static void ext4_ext_try_to_merge(handle_t *handle,
1733                                   struct inode *inode,
1734                                   struct ext4_ext_path *path,
1735                                   struct ext4_extent *ex) {
1736         struct ext4_extent_header *eh;
1737         unsigned int depth;
1738         int merge_done = 0;
1739
1740         depth = ext_depth(inode);
1741         BUG_ON(path[depth].p_hdr == NULL);
1742         eh = path[depth].p_hdr;
1743
1744         if (ex > EXT_FIRST_EXTENT(eh))
1745                 merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1746
1747         if (!merge_done)
1748                 (void) ext4_ext_try_to_merge_right(inode, path, ex);
1749
1750         ext4_ext_try_to_merge_up(handle, inode, path);
1751 }
1752
1753 /*
1754  * check if a portion of the "newext" extent overlaps with an
1755  * existing extent.
1756  *
1757  * If there is an overlap discovered, it updates the length of the newext
1758  * such that there will be no overlap, and then returns 1.
1759  * If there is no overlap found, it returns 0.
1760  */
1761 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1762                                            struct inode *inode,
1763                                            struct ext4_extent *newext,
1764                                            struct ext4_ext_path *path)
1765 {
1766         ext4_lblk_t b1, b2;
1767         unsigned int depth, len1;
1768         unsigned int ret = 0;
1769
1770         b1 = le32_to_cpu(newext->ee_block);
1771         len1 = ext4_ext_get_actual_len(newext);
1772         depth = ext_depth(inode);
1773         if (!path[depth].p_ext)
1774                 goto out;
1775         b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1776
1777         /*
1778          * get the next allocated block if the extent in the path
1779          * is before the requested block(s)
1780          */
1781         if (b2 < b1) {
1782                 b2 = ext4_ext_next_allocated_block(path);
1783                 if (b2 == EXT_MAX_BLOCKS)
1784                         goto out;
1785                 b2 = EXT4_LBLK_CMASK(sbi, b2);
1786         }
1787
1788         /* check for wrap through zero on extent logical start block*/
1789         if (b1 + len1 < b1) {
1790                 len1 = EXT_MAX_BLOCKS - b1;
1791                 newext->ee_len = cpu_to_le16(len1);
1792                 ret = 1;
1793         }
1794
1795         /* check for overlap */
1796         if (b1 + len1 > b2) {
1797                 newext->ee_len = cpu_to_le16(b2 - b1);
1798                 ret = 1;
1799         }
1800 out:
1801         return ret;
1802 }
1803
1804 /*
1805  * ext4_ext_insert_extent:
1806  * tries to merge requsted extent into the existing extent or
1807  * inserts requested extent as new one into the tree,
1808  * creating new leaf in the no-space case.
1809  */
1810 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1811                                 struct ext4_ext_path *path,
1812                                 struct ext4_extent *newext, int flag)
1813 {
1814         struct ext4_extent_header *eh;
1815         struct ext4_extent *ex, *fex;
1816         struct ext4_extent *nearex; /* nearest extent */
1817         struct ext4_ext_path *npath = NULL;
1818         int depth, len, err;
1819         ext4_lblk_t next;
1820         unsigned uninitialized = 0;
1821         int flags = 0;
1822
1823         if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1824                 EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1825                 return -EIO;
1826         }
1827         depth = ext_depth(inode);
1828         ex = path[depth].p_ext;
1829         eh = path[depth].p_hdr;
1830         if (unlikely(path[depth].p_hdr == NULL)) {
1831                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1832                 return -EIO;
1833         }
1834
1835         /* try to insert block into found extent and return */
1836         if (ex && !(flag & EXT4_GET_BLOCKS_PRE_IO)) {
1837
1838                 /*
1839                  * Try to see whether we should rather test the extent on
1840                  * right from ex, or from the left of ex. This is because
1841                  * ext4_ext_find_extent() can return either extent on the
1842                  * left, or on the right from the searched position. This
1843                  * will make merging more effective.
1844                  */
1845                 if (ex < EXT_LAST_EXTENT(eh) &&
1846                     (le32_to_cpu(ex->ee_block) +
1847                     ext4_ext_get_actual_len(ex) <
1848                     le32_to_cpu(newext->ee_block))) {
1849                         ex += 1;
1850                         goto prepend;
1851                 } else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1852                            (le32_to_cpu(newext->ee_block) +
1853                            ext4_ext_get_actual_len(newext) <
1854                            le32_to_cpu(ex->ee_block)))
1855                         ex -= 1;
1856
1857                 /* Try to append newex to the ex */
1858                 if (ext4_can_extents_be_merged(inode, ex, newext)) {
1859                         ext_debug("append [%d]%d block to %u:[%d]%d"
1860                                   "(from %llu)\n",
1861                                   ext4_ext_is_uninitialized(newext),
1862                                   ext4_ext_get_actual_len(newext),
1863                                   le32_to_cpu(ex->ee_block),
1864                                   ext4_ext_is_uninitialized(ex),
1865                                   ext4_ext_get_actual_len(ex),
1866                                   ext4_ext_pblock(ex));
1867                         err = ext4_ext_get_access(handle, inode,
1868                                                   path + depth);
1869                         if (err)
1870                                 return err;
1871
1872                         /*
1873                          * ext4_can_extents_be_merged should have checked
1874                          * that either both extents are uninitialized, or
1875                          * both aren't. Thus we need to check only one of
1876                          * them here.
1877                          */
1878                         if (ext4_ext_is_uninitialized(ex))
1879                                 uninitialized = 1;
1880                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1881                                         + ext4_ext_get_actual_len(newext));
1882                         if (uninitialized)
1883                                 ext4_ext_mark_uninitialized(ex);
1884                         eh = path[depth].p_hdr;
1885                         nearex = ex;
1886                         goto merge;
1887                 }
1888
1889 prepend:
1890                 /* Try to prepend newex to the ex */
1891                 if (ext4_can_extents_be_merged(inode, newext, ex)) {
1892                         ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
1893                                   "(from %llu)\n",
1894                                   le32_to_cpu(newext->ee_block),
1895                                   ext4_ext_is_uninitialized(newext),
1896                                   ext4_ext_get_actual_len(newext),
1897                                   le32_to_cpu(ex->ee_block),
1898                                   ext4_ext_is_uninitialized(ex),
1899                                   ext4_ext_get_actual_len(ex),
1900                                   ext4_ext_pblock(ex));
1901                         err = ext4_ext_get_access(handle, inode,
1902                                                   path + depth);
1903                         if (err)
1904                                 return err;
1905
1906                         /*
1907                          * ext4_can_extents_be_merged should have checked
1908                          * that either both extents are uninitialized, or
1909                          * both aren't. Thus we need to check only one of
1910                          * them here.
1911                          */
1912                         if (ext4_ext_is_uninitialized(ex))
1913                                 uninitialized = 1;
1914                         ex->ee_block = newext->ee_block;
1915                         ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
1916                         ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1917                                         + ext4_ext_get_actual_len(newext));
1918                         if (uninitialized)
1919                                 ext4_ext_mark_uninitialized(ex);
1920                         eh = path[depth].p_hdr;
1921                         nearex = ex;
1922                         goto merge;
1923                 }
1924         }
1925
1926         depth = ext_depth(inode);
1927         eh = path[depth].p_hdr;
1928         if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
1929                 goto has_space;
1930
1931         /* probably next leaf has space for us? */
1932         fex = EXT_LAST_EXTENT(eh);
1933         next = EXT_MAX_BLOCKS;
1934         if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
1935                 next = ext4_ext_next_leaf_block(path);
1936         if (next != EXT_MAX_BLOCKS) {
1937                 ext_debug("next leaf block - %u\n", next);
1938                 BUG_ON(npath != NULL);
1939                 npath = ext4_ext_find_extent(inode, next, NULL);
1940                 if (IS_ERR(npath))
1941                         return PTR_ERR(npath);
1942                 BUG_ON(npath->p_depth != path->p_depth);
1943                 eh = npath[depth].p_hdr;
1944                 if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
1945                         ext_debug("next leaf isn't full(%d)\n",
1946                                   le16_to_cpu(eh->eh_entries));
1947                         path = npath;
1948                         goto has_space;
1949                 }
1950                 ext_debug("next leaf has no free space(%d,%d)\n",
1951                           le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
1952         }
1953
1954         /*
1955          * There is no free space in the found leaf.
1956          * We're gonna add a new leaf in the tree.
1957          */
1958         if (flag & EXT4_GET_BLOCKS_METADATA_NOFAIL)
1959                 flags = EXT4_MB_USE_RESERVED;
1960         err = ext4_ext_create_new_leaf(handle, inode, flags, path, newext);
1961         if (err)
1962                 goto cleanup;
1963         depth = ext_depth(inode);
1964         eh = path[depth].p_hdr;
1965
1966 has_space:
1967         nearex = path[depth].p_ext;
1968
1969         err = ext4_ext_get_access(handle, inode, path + depth);
1970         if (err)
1971                 goto cleanup;
1972
1973         if (!nearex) {
1974                 /* there is no extent in this leaf, create first one */
1975                 ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
1976                                 le32_to_cpu(newext->ee_block),
1977                                 ext4_ext_pblock(newext),
1978                                 ext4_ext_is_uninitialized(newext),
1979                                 ext4_ext_get_actual_len(newext));
1980                 nearex = EXT_FIRST_EXTENT(eh);
1981         } else {
1982                 if (le32_to_cpu(newext->ee_block)
1983                            > le32_to_cpu(nearex->ee_block)) {
1984                         /* Insert after */
1985                         ext_debug("insert %u:%llu:[%d]%d before: "
1986                                         "nearest %p\n",
1987                                         le32_to_cpu(newext->ee_block),
1988                                         ext4_ext_pblock(newext),
1989                                         ext4_ext_is_uninitialized(newext),
1990                                         ext4_ext_get_actual_len(newext),
1991                                         nearex);
1992                         nearex++;
1993                 } else {
1994                         /* Insert before */
1995                         BUG_ON(newext->ee_block == nearex->ee_block);
1996                         ext_debug("insert %u:%llu:[%d]%d after: "
1997                                         "nearest %p\n",
1998                                         le32_to_cpu(newext->ee_block),
1999                                         ext4_ext_pblock(newext),
2000                                         ext4_ext_is_uninitialized(newext),
2001                                         ext4_ext_get_actual_len(newext),
2002                                         nearex);
2003                 }
2004                 len = EXT_LAST_EXTENT(eh) - nearex + 1;
2005                 if (len > 0) {
2006                         ext_debug("insert %u:%llu:[%d]%d: "
2007                                         "move %d extents from 0x%p to 0x%p\n",
2008                                         le32_to_cpu(newext->ee_block),
2009                                         ext4_ext_pblock(newext),
2010                                         ext4_ext_is_uninitialized(newext),
2011                                         ext4_ext_get_actual_len(newext),
2012                                         len, nearex, nearex + 1);
2013                         memmove(nearex + 1, nearex,
2014                                 len * sizeof(struct ext4_extent));
2015                 }
2016         }
2017
2018         le16_add_cpu(&eh->eh_entries, 1);
2019         path[depth].p_ext = nearex;
2020         nearex->ee_block = newext->ee_block;
2021         ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2022         nearex->ee_len = newext->ee_len;
2023
2024 merge:
2025         /* try to merge extents */
2026         if (!(flag & EXT4_GET_BLOCKS_PRE_IO))
2027                 ext4_ext_try_to_merge(handle, inode, path, nearex);
2028
2029
2030         /* time to correct all indexes above */
2031         err = ext4_ext_correct_indexes(handle, inode, path);
2032         if (err)
2033                 goto cleanup;
2034
2035         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2036
2037 cleanup:
2038         if (npath) {
2039                 ext4_ext_drop_refs(npath);
2040                 kfree(npath);
2041         }
2042         return err;
2043 }
2044
2045 static int ext4_fill_fiemap_extents(struct inode *inode,
2046                                     ext4_lblk_t block, ext4_lblk_t num,
2047                                     struct fiemap_extent_info *fieinfo)
2048 {
2049         struct ext4_ext_path *path = NULL;
2050         struct ext4_extent *ex;
2051         struct extent_status es;
2052         ext4_lblk_t next, next_del, start = 0, end = 0;
2053         ext4_lblk_t last = block + num;
2054         int exists, depth = 0, err = 0;
2055         unsigned int flags = 0;
2056         unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2057
2058         while (block < last && block != EXT_MAX_BLOCKS) {
2059                 num = last - block;
2060                 /* find extent for this block */
2061                 down_read(&EXT4_I(inode)->i_data_sem);
2062
2063                 if (path && ext_depth(inode) != depth) {
2064                         /* depth was changed. we have to realloc path */
2065                         kfree(path);
2066                         path = NULL;
2067                 }
2068
2069                 path = ext4_ext_find_extent(inode, block, path);
2070                 if (IS_ERR(path)) {
2071                         up_read(&EXT4_I(inode)->i_data_sem);
2072                         err = PTR_ERR(path);
2073                         path = NULL;
2074                         break;
2075                 }
2076
2077                 depth = ext_depth(inode);
2078                 if (unlikely(path[depth].p_hdr == NULL)) {
2079                         up_read(&EXT4_I(inode)->i_data_sem);
2080                         EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2081                         err = -EIO;
2082                         break;
2083                 }
2084                 ex = path[depth].p_ext;
2085                 next = ext4_ext_next_allocated_block(path);
2086                 ext4_ext_drop_refs(path);
2087
2088                 flags = 0;
2089                 exists = 0;
2090                 if (!ex) {
2091                         /* there is no extent yet, so try to allocate
2092                          * all requested space */
2093                         start = block;
2094                         end = block + num;
2095                 } else if (le32_to_cpu(ex->ee_block) > block) {
2096                         /* need to allocate space before found extent */
2097                         start = block;
2098                         end = le32_to_cpu(ex->ee_block);
2099                         if (block + num < end)
2100                                 end = block + num;
2101                 } else if (block >= le32_to_cpu(ex->ee_block)
2102                                         + ext4_ext_get_actual_len(ex)) {
2103                         /* need to allocate space after found extent */
2104                         start = block;
2105                         end = block + num;
2106                         if (end >= next)
2107                                 end = next;
2108                 } else if (block >= le32_to_cpu(ex->ee_block)) {
2109                         /*
2110                          * some part of requested space is covered
2111                          * by found extent
2112                          */
2113                         start = block;
2114                         end = le32_to_cpu(ex->ee_block)
2115                                 + ext4_ext_get_actual_len(ex);
2116                         if (block + num < end)
2117                                 end = block + num;
2118                         exists = 1;
2119                 } else {
2120                         BUG();
2121                 }
2122                 BUG_ON(end <= start);
2123
2124                 if (!exists) {
2125                         es.es_lblk = start;
2126                         es.es_len = end - start;
2127                         es.es_pblk = 0;
2128                 } else {
2129                         es.es_lblk = le32_to_cpu(ex->ee_block);
2130                         es.es_len = ext4_ext_get_actual_len(ex);
2131                         es.es_pblk = ext4_ext_pblock(ex);
2132                         if (ext4_ext_is_uninitialized(ex))
2133                                 flags |= FIEMAP_EXTENT_UNWRITTEN;
2134                 }
2135
2136                 /*
2137                  * Find delayed extent and update es accordingly. We call
2138                  * it even in !exists case to find out whether es is the
2139                  * last existing extent or not.
2140                  */
2141                 next_del = ext4_find_delayed_extent(inode, &es);
2142                 if (!exists && next_del) {
2143                         exists = 1;
2144                         flags |= FIEMAP_EXTENT_DELALLOC;
2145                 }
2146                 up_read(&EXT4_I(inode)->i_data_sem);
2147
2148                 if (unlikely(es.es_len == 0)) {
2149                         EXT4_ERROR_INODE(inode, "es.es_len == 0");
2150                         err = -EIO;
2151                         break;
2152                 }
2153
2154                 /*
2155                  * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2156                  * we need to check next == EXT_MAX_BLOCKS because it is
2157                  * possible that an extent is with unwritten and delayed
2158                  * status due to when an extent is delayed allocated and
2159                  * is allocated by fallocate status tree will track both of
2160                  * them in a extent.
2161                  *
2162                  * So we could return a unwritten and delayed extent, and
2163                  * its block is equal to 'next'.
2164                  */
2165                 if (next == next_del && next == EXT_MAX_BLOCKS) {
2166                         flags |= FIEMAP_EXTENT_LAST;
2167                         if (unlikely(next_del != EXT_MAX_BLOCKS ||
2168                                      next != EXT_MAX_BLOCKS)) {
2169                                 EXT4_ERROR_INODE(inode,
2170                                                  "next extent == %u, next "
2171                                                  "delalloc extent = %u",
2172                                                  next, next_del);
2173                                 err = -EIO;
2174                                 break;
2175                         }
2176                 }
2177
2178                 if (exists) {
2179                         err = fiemap_fill_next_extent(fieinfo,
2180                                 (__u64)es.es_lblk << blksize_bits,
2181                                 (__u64)es.es_pblk << blksize_bits,
2182                                 (__u64)es.es_len << blksize_bits,
2183                                 flags);
2184                         if (err < 0)
2185                                 break;
2186                         if (err == 1) {
2187                                 err = 0;
2188                                 break;
2189                         }
2190                 }
2191
2192                 block = es.es_lblk + es.es_len;
2193         }
2194
2195         if (path) {
2196                 ext4_ext_drop_refs(path);
2197                 kfree(path);
2198         }
2199
2200         return err;
2201 }
2202
2203 /*
2204  * ext4_ext_put_gap_in_cache:
2205  * calculate boundaries of the gap that the requested block fits into
2206  * and cache this gap
2207  */
2208 static void
2209 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2210                                 ext4_lblk_t block)
2211 {
2212         int depth = ext_depth(inode);
2213         unsigned long len;
2214         ext4_lblk_t lblock;
2215         struct ext4_extent *ex;
2216
2217         ex = path[depth].p_ext;
2218         if (ex == NULL) {
2219                 /*
2220                  * there is no extent yet, so gap is [0;-] and we
2221                  * don't cache it
2222                  */
2223                 ext_debug("cache gap(whole file):");
2224         } else if (block < le32_to_cpu(ex->ee_block)) {
2225                 lblock = block;
2226                 len = le32_to_cpu(ex->ee_block) - block;
2227                 ext_debug("cache gap(before): %u [%u:%u]",
2228                                 block,
2229                                 le32_to_cpu(ex->ee_block),
2230                                  ext4_ext_get_actual_len(ex));
2231                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2232                         ext4_es_insert_extent(inode, lblock, len, ~0,
2233                                               EXTENT_STATUS_HOLE);
2234         } else if (block >= le32_to_cpu(ex->ee_block)
2235                         + ext4_ext_get_actual_len(ex)) {
2236                 ext4_lblk_t next;
2237                 lblock = le32_to_cpu(ex->ee_block)
2238                         + ext4_ext_get_actual_len(ex);
2239
2240                 next = ext4_ext_next_allocated_block(path);
2241                 ext_debug("cache gap(after): [%u:%u] %u",
2242                                 le32_to_cpu(ex->ee_block),
2243                                 ext4_ext_get_actual_len(ex),
2244                                 block);
2245                 BUG_ON(next == lblock);
2246                 len = next - lblock;
2247                 if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2248                         ext4_es_insert_extent(inode, lblock, len, ~0,
2249                                               EXTENT_STATUS_HOLE);
2250         } else {
2251                 lblock = len = 0;
2252                 BUG();
2253         }
2254
2255         ext_debug(" -> %u:%lu\n", lblock, len);
2256 }
2257
2258 /*
2259  * ext4_ext_rm_idx:
2260  * removes index from the index block.
2261  */
2262 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2263                         struct ext4_ext_path *path, int depth)
2264 {
2265         int err;
2266         ext4_fsblk_t leaf;
2267
2268         /* free index block */
2269         depth--;
2270         path = path + depth;
2271         leaf = ext4_idx_pblock(path->p_idx);
2272         if (unlikely(path->p_hdr->eh_entries == 0)) {
2273                 EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2274                 return -EIO;
2275         }
2276         err = ext4_ext_get_access(handle, inode, path);
2277         if (err)
2278                 return err;
2279
2280         if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2281                 int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2282                 len *= sizeof(struct ext4_extent_idx);
2283                 memmove(path->p_idx, path->p_idx + 1, len);
2284         }
2285
2286         le16_add_cpu(&path->p_hdr->eh_entries, -1);
2287         err = ext4_ext_dirty(handle, inode, path);
2288         if (err)
2289                 return err;
2290         ext_debug("index is empty, remove it, free block %llu\n", leaf);
2291         trace_ext4_ext_rm_idx(inode, leaf);
2292
2293         ext4_free_blocks(handle, inode, NULL, leaf, 1,
2294                          EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2295
2296         while (--depth >= 0) {
2297                 if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2298                         break;
2299                 path--;
2300                 err = ext4_ext_get_access(handle, inode, path);
2301                 if (err)
2302                         break;
2303                 path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2304                 err = ext4_ext_dirty(handle, inode, path);
2305                 if (err)
2306                         break;
2307         }
2308         return err;
2309 }
2310
2311 /*
2312  * ext4_ext_calc_credits_for_single_extent:
2313  * This routine returns max. credits that needed to insert an extent
2314  * to the extent tree.
2315  * When pass the actual path, the caller should calculate credits
2316  * under i_data_sem.
2317  */
2318 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2319                                                 struct ext4_ext_path *path)
2320 {
2321         if (path) {
2322                 int depth = ext_depth(inode);
2323                 int ret = 0;
2324
2325                 /* probably there is space in leaf? */
2326                 if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2327                                 < le16_to_cpu(path[depth].p_hdr->eh_max)) {
2328
2329                         /*
2330                          *  There are some space in the leaf tree, no
2331                          *  need to account for leaf block credit
2332                          *
2333                          *  bitmaps and block group descriptor blocks
2334                          *  and other metadata blocks still need to be
2335                          *  accounted.
2336                          */
2337                         /* 1 bitmap, 1 block group descriptor */
2338                         ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2339                         return ret;
2340                 }
2341         }
2342
2343         return ext4_chunk_trans_blocks(inode, nrblocks);
2344 }
2345
2346 /*
2347  * How many index/leaf blocks need to change/allocate to modify nrblocks?
2348  *
2349  * if nrblocks are fit in a single extent (chunk flag is 1), then
2350  * in the worse case, each tree level index/leaf need to be changed
2351  * if the tree split due to insert a new extent, then the old tree
2352  * index/leaf need to be updated too
2353  *
2354  * If the nrblocks are discontiguous, they could cause
2355  * the whole tree split more than once, but this is really rare.
2356  */
2357 int ext4_ext_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
2358 {
2359         int index;
2360         int depth;
2361
2362         /* If we are converting the inline data, only one is needed here. */
2363         if (ext4_has_inline_data(inode))
2364                 return 1;
2365
2366         depth = ext_depth(inode);
2367
2368         if (chunk)
2369                 index = depth * 2;
2370         else
2371                 index = depth * 3;
2372
2373         return index;
2374 }
2375
2376 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2377                               struct ext4_extent *ex,
2378                               ext4_fsblk_t *partial_cluster,
2379                               ext4_lblk_t from, ext4_lblk_t to)
2380 {
2381         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2382         unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2383         ext4_fsblk_t pblk;
2384         int flags = 0;
2385
2386         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2387                 flags |= EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2388         else if (ext4_should_journal_data(inode))
2389                 flags |= EXT4_FREE_BLOCKS_FORGET;
2390
2391         /*
2392          * For bigalloc file systems, we never free a partial cluster
2393          * at the beginning of the extent.  Instead, we make a note
2394          * that we tried freeing the cluster, and check to see if we
2395          * need to free it on a subsequent call to ext4_remove_blocks,
2396          * or at the end of the ext4_truncate() operation.
2397          */
2398         flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2399
2400         trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2401         /*
2402          * If we have a partial cluster, and it's different from the
2403          * cluster of the last block, we need to explicitly free the
2404          * partial cluster here.
2405          */
2406         pblk = ext4_ext_pblock(ex) + ee_len - 1;
2407         if (*partial_cluster && (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2408                 ext4_free_blocks(handle, inode, NULL,
2409                                  EXT4_C2B(sbi, *partial_cluster),
2410                                  sbi->s_cluster_ratio, flags);
2411                 *partial_cluster = 0;
2412         }
2413
2414 #ifdef EXTENTS_STATS
2415         {
2416                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2417                 spin_lock(&sbi->s_ext_stats_lock);
2418                 sbi->s_ext_blocks += ee_len;
2419                 sbi->s_ext_extents++;
2420                 if (ee_len < sbi->s_ext_min)
2421                         sbi->s_ext_min = ee_len;
2422                 if (ee_len > sbi->s_ext_max)
2423                         sbi->s_ext_max = ee_len;
2424                 if (ext_depth(inode) > sbi->s_depth_max)
2425                         sbi->s_depth_max = ext_depth(inode);
2426                 spin_unlock(&sbi->s_ext_stats_lock);
2427         }
2428 #endif
2429         if (from >= le32_to_cpu(ex->ee_block)
2430             && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2431                 /* tail removal */
2432                 ext4_lblk_t num;
2433
2434                 num = le32_to_cpu(ex->ee_block) + ee_len - from;
2435                 pblk = ext4_ext_pblock(ex) + ee_len - num;
2436                 ext_debug("free last %u blocks starting %llu\n", num, pblk);
2437                 ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2438                 /*
2439                  * If the block range to be freed didn't start at the
2440                  * beginning of a cluster, and we removed the entire
2441                  * extent, save the partial cluster here, since we
2442                  * might need to delete if we determine that the
2443                  * truncate operation has removed all of the blocks in
2444                  * the cluster.
2445                  */
2446                 if (EXT4_PBLK_COFF(sbi, pblk) &&
2447                     (ee_len == num))
2448                         *partial_cluster = EXT4_B2C(sbi, pblk);
2449                 else
2450                         *partial_cluster = 0;
2451         } else if (from == le32_to_cpu(ex->ee_block)
2452                    && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
2453                 /* head removal */
2454                 ext4_lblk_t num;
2455                 ext4_fsblk_t start;
2456
2457                 num = to - from;
2458                 start = ext4_ext_pblock(ex);
2459
2460                 ext_debug("free first %u blocks starting %llu\n", num, start);
2461                 ext4_free_blocks(handle, inode, NULL, start, num, flags);
2462
2463         } else {
2464                 printk(KERN_INFO "strange request: removal(2) "
2465                                 "%u-%u from %u:%u\n",
2466                                 from, to, le32_to_cpu(ex->ee_block), ee_len);
2467         }
2468         return 0;
2469 }
2470
2471
2472 /*
2473  * ext4_ext_rm_leaf() Removes the extents associated with the
2474  * blocks appearing between "start" and "end", and splits the extents
2475  * if "start" and "end" appear in the same extent
2476  *
2477  * @handle: The journal handle
2478  * @inode:  The files inode
2479  * @path:   The path to the leaf
2480  * @start:  The first block to remove
2481  * @end:   The last block to remove
2482  */
2483 static int
2484 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2485                  struct ext4_ext_path *path, ext4_fsblk_t *partial_cluster,
2486                  ext4_lblk_t start, ext4_lblk_t end)
2487 {
2488         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2489         int err = 0, correct_index = 0;
2490         int depth = ext_depth(inode), credits;
2491         struct ext4_extent_header *eh;
2492         ext4_lblk_t a, b;
2493         unsigned num;
2494         ext4_lblk_t ex_ee_block;
2495         unsigned short ex_ee_len;
2496         unsigned uninitialized = 0;
2497         struct ext4_extent *ex;
2498
2499         /* the header must be checked already in ext4_ext_remove_space() */
2500         ext_debug("truncate since %u in leaf to %u\n", start, end);
2501         if (!path[depth].p_hdr)
2502                 path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2503         eh = path[depth].p_hdr;
2504         if (unlikely(path[depth].p_hdr == NULL)) {
2505                 EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2506                 return -EIO;
2507         }
2508         /* find where to start removing */
2509         ex = EXT_LAST_EXTENT(eh);
2510
2511         ex_ee_block = le32_to_cpu(ex->ee_block);
2512         ex_ee_len = ext4_ext_get_actual_len(ex);
2513
2514         /*
2515          * If we're starting with an extent other than the last one in the
2516          * node, we need to see if it shares a cluster with the extent to
2517          * the right (towards the end of the file). If its leftmost cluster
2518          * is this extent's rightmost cluster and it is not cluster aligned,
2519          * we'll mark it as a partial that is not to be deallocated.
2520          */
2521
2522         if (ex != EXT_LAST_EXTENT(eh)) {
2523                 ext4_fsblk_t current_pblk, right_pblk;
2524                 long long current_cluster, right_cluster;
2525
2526                 current_pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2527                 current_cluster = (long long)EXT4_B2C(sbi, current_pblk);
2528                 right_pblk = ext4_ext_pblock(ex + 1);
2529                 right_cluster = (long long)EXT4_B2C(sbi, right_pblk);
2530                 if (current_cluster == right_cluster &&
2531                         EXT4_PBLK_COFF(sbi, right_pblk))
2532                         *partial_cluster = -right_cluster;
2533         }
2534
2535         trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2536
2537         while (ex >= EXT_FIRST_EXTENT(eh) &&
2538                         ex_ee_block + ex_ee_len > start) {
2539
2540                 if (ext4_ext_is_uninitialized(ex))
2541                         uninitialized = 1;
2542                 else
2543                         uninitialized = 0;
2544
2545                 ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2546                          uninitialized, ex_ee_len);
2547                 path[depth].p_ext = ex;
2548
2549                 a = ex_ee_block > start ? ex_ee_block : start;
2550                 b = ex_ee_block+ex_ee_len - 1 < end ?
2551                         ex_ee_block+ex_ee_len - 1 : end;
2552
2553                 ext_debug("  border %u:%u\n", a, b);
2554
2555                 /* If this extent is beyond the end of the hole, skip it */
2556                 if (end < ex_ee_block) {
2557                         ex--;
2558                         ex_ee_block = le32_to_cpu(ex->ee_block);
2559                         ex_ee_len = ext4_ext_get_actual_len(ex);
2560                         continue;
2561                 } else if (b != ex_ee_block + ex_ee_len - 1) {
2562                         EXT4_ERROR_INODE(inode,
2563                                          "can not handle truncate %u:%u "
2564                                          "on extent %u:%u",
2565                                          start, end, ex_ee_block,
2566                                          ex_ee_block + ex_ee_len - 1);
2567                         err = -EIO;
2568                         goto out;
2569                 } else if (a != ex_ee_block) {
2570                         /* remove tail of the extent */
2571                         num = a - ex_ee_block;
2572                 } else {
2573                         /* remove whole extent: excellent! */
2574                         num = 0;
2575                 }
2576                 /*
2577                  * 3 for leaf, sb, and inode plus 2 (bmap and group
2578                  * descriptor) for each block group; assume two block
2579                  * groups plus ex_ee_len/blocks_per_block_group for
2580                  * the worst case
2581                  */
2582                 credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2583                 if (ex == EXT_FIRST_EXTENT(eh)) {
2584                         correct_index = 1;
2585                         credits += (ext_depth(inode)) + 1;
2586                 }
2587                 credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2588
2589                 err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2590                 if (err)
2591                         goto out;
2592
2593                 err = ext4_ext_get_access(handle, inode, path + depth);
2594                 if (err)
2595                         goto out;
2596
2597                 err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2598                                          a, b);
2599                 if (err)
2600                         goto out;
2601
2602                 if (num == 0)
2603                         /* this extent is removed; mark slot entirely unused */
2604                         ext4_ext_store_pblock(ex, 0);
2605
2606                 ex->ee_len = cpu_to_le16(num);
2607                 /*
2608                  * Do not mark uninitialized if all the blocks in the
2609                  * extent have been removed.
2610                  */
2611                 if (uninitialized && num)
2612                         ext4_ext_mark_uninitialized(ex);
2613                 /*
2614                  * If the extent was completely released,
2615                  * we need to remove it from the leaf
2616                  */
2617                 if (num == 0) {
2618                         if (end != EXT_MAX_BLOCKS - 1) {
2619                                 /*
2620                                  * For hole punching, we need to scoot all the
2621                                  * extents up when an extent is removed so that
2622                                  * we dont have blank extents in the middle
2623                                  */
2624                                 memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2625                                         sizeof(struct ext4_extent));
2626
2627                                 /* Now get rid of the one at the end */
2628                                 memset(EXT_LAST_EXTENT(eh), 0,
2629                                         sizeof(struct ext4_extent));
2630                         }
2631                         le16_add_cpu(&eh->eh_entries, -1);
2632                 } else
2633                         *partial_cluster = 0;
2634
2635                 err = ext4_ext_dirty(handle, inode, path + depth);
2636                 if (err)
2637                         goto out;
2638
2639                 ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2640                                 ext4_ext_pblock(ex));
2641                 ex--;
2642                 ex_ee_block = le32_to_cpu(ex->ee_block);
2643                 ex_ee_len = ext4_ext_get_actual_len(ex);
2644         }
2645
2646         if (correct_index && eh->eh_entries)
2647                 err = ext4_ext_correct_indexes(handle, inode, path);
2648
2649         /*
2650          * If there is still a entry in the leaf node, check to see if
2651          * it references the partial cluster.  This is the only place
2652          * where it could; if it doesn't, we can free the cluster.
2653          */
2654         if (*partial_cluster && ex >= EXT_FIRST_EXTENT(eh) &&
2655             (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2656              *partial_cluster)) {
2657                 int flags = EXT4_FREE_BLOCKS_FORGET;
2658
2659                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2660                         flags |= EXT4_FREE_BLOCKS_METADATA;
2661
2662                 ext4_free_blocks(handle, inode, NULL,
2663                                  EXT4_C2B(sbi, *partial_cluster),
2664                                  sbi->s_cluster_ratio, flags);
2665                 *partial_cluster = 0;
2666         }
2667
2668         /* if this leaf is free, then we should
2669          * remove it from index block above */
2670         if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2671                 err = ext4_ext_rm_idx(handle, inode, path, depth);
2672
2673 out:
2674         return err;
2675 }
2676
2677 /*
2678  * ext4_ext_more_to_rm:
2679  * returns 1 if current index has to be freed (even partial)
2680  */
2681 static int
2682 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2683 {
2684         BUG_ON(path->p_idx == NULL);
2685
2686         if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2687                 return 0;
2688
2689         /*
2690          * if truncate on deeper level happened, it wasn't partial,
2691          * so we have to consider current index for truncation
2692          */
2693         if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2694                 return 0;
2695         return 1;
2696 }
2697
2698 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2699                           ext4_lblk_t end)
2700 {
2701         struct super_block *sb = inode->i_sb;
2702         int depth = ext_depth(inode);
2703         struct ext4_ext_path *path = NULL;
2704         ext4_fsblk_t partial_cluster = 0;
2705         handle_t *handle;
2706         int i = 0, err = 0;
2707
2708         ext_debug("truncate since %u to %u\n", start, end);
2709
2710         /* probably first extent we're gonna free will be last in block */
2711         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2712         if (IS_ERR(handle))
2713                 return PTR_ERR(handle);
2714
2715 again:
2716         trace_ext4_ext_remove_space(inode, start, depth);
2717
2718         /*
2719          * Check if we are removing extents inside the extent tree. If that
2720          * is the case, we are going to punch a hole inside the extent tree
2721          * so we have to check whether we need to split the extent covering
2722          * the last block to remove so we can easily remove the part of it
2723          * in ext4_ext_rm_leaf().
2724          */
2725         if (end < EXT_MAX_BLOCKS - 1) {
2726                 struct ext4_extent *ex;
2727                 ext4_lblk_t ee_block;
2728
2729                 /* find extent for this block */
2730                 path = ext4_ext_find_extent(inode, end, NULL);
2731                 if (IS_ERR(path)) {
2732                         ext4_journal_stop(handle);
2733                         return PTR_ERR(path);
2734                 }
2735                 depth = ext_depth(inode);
2736                 /* Leaf not may not exist only if inode has no blocks at all */
2737                 ex = path[depth].p_ext;
2738                 if (!ex) {
2739                         if (depth) {
2740                                 EXT4_ERROR_INODE(inode,
2741                                                  "path[%d].p_hdr == NULL",
2742                                                  depth);
2743                                 err = -EIO;
2744                         }
2745                         goto out;
2746                 }
2747
2748                 ee_block = le32_to_cpu(ex->ee_block);
2749
2750                 /*
2751                  * See if the last block is inside the extent, if so split
2752                  * the extent at 'end' block so we can easily remove the
2753                  * tail of the first part of the split extent in
2754                  * ext4_ext_rm_leaf().
2755                  */
2756                 if (end >= ee_block &&
2757                     end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2758                         int split_flag = 0;
2759
2760                         if (ext4_ext_is_uninitialized(ex))
2761                                 split_flag = EXT4_EXT_MARK_UNINIT1 |
2762                                              EXT4_EXT_MARK_UNINIT2;
2763
2764                         /*
2765                          * Split the extent in two so that 'end' is the last
2766                          * block in the first new extent. Also we should not
2767                          * fail removing space due to ENOSPC so try to use
2768                          * reserved block if that happens.
2769                          */
2770                         err = ext4_split_extent_at(handle, inode, path,
2771                                         end + 1, split_flag,
2772                                         EXT4_GET_BLOCKS_PRE_IO |
2773                                         EXT4_GET_BLOCKS_METADATA_NOFAIL);
2774
2775                         if (err < 0)
2776                                 goto out;
2777                 }
2778         }
2779         /*
2780          * We start scanning from right side, freeing all the blocks
2781          * after i_size and walking into the tree depth-wise.
2782          */
2783         depth = ext_depth(inode);
2784         if (path) {
2785                 int k = i = depth;
2786                 while (--k > 0)
2787                         path[k].p_block =
2788                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2789         } else {
2790                 path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2791                                GFP_NOFS);
2792                 if (path == NULL) {
2793                         ext4_journal_stop(handle);
2794                         return -ENOMEM;
2795                 }
2796                 path[0].p_depth = depth;
2797                 path[0].p_hdr = ext_inode_hdr(inode);
2798                 i = 0;
2799
2800                 if (ext4_ext_check(inode, path[0].p_hdr, depth)) {
2801                         err = -EIO;
2802                         goto out;
2803                 }
2804         }
2805         err = 0;
2806
2807         while (i >= 0 && err == 0) {
2808                 if (i == depth) {
2809                         /* this is leaf block */
2810                         err = ext4_ext_rm_leaf(handle, inode, path,
2811                                                &partial_cluster, start,
2812                                                end);
2813                         /* root level has p_bh == NULL, brelse() eats this */
2814                         brelse(path[i].p_bh);
2815                         path[i].p_bh = NULL;
2816                         i--;
2817                         continue;
2818                 }
2819
2820                 /* this is index block */
2821                 if (!path[i].p_hdr) {
2822                         ext_debug("initialize header\n");
2823                         path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2824                 }
2825
2826                 if (!path[i].p_idx) {
2827                         /* this level hasn't been touched yet */
2828                         path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2829                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2830                         ext_debug("init index ptr: hdr 0x%p, num %d\n",
2831                                   path[i].p_hdr,
2832                                   le16_to_cpu(path[i].p_hdr->eh_entries));
2833                 } else {
2834                         /* we were already here, see at next index */
2835                         path[i].p_idx--;
2836                 }
2837
2838                 ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2839                                 i, EXT_FIRST_INDEX(path[i].p_hdr),
2840                                 path[i].p_idx);
2841                 if (ext4_ext_more_to_rm(path + i)) {
2842                         struct buffer_head *bh;
2843                         /* go to the next level */
2844                         ext_debug("move to level %d (block %llu)\n",
2845                                   i + 1, ext4_idx_pblock(path[i].p_idx));
2846                         memset(path + i + 1, 0, sizeof(*path));
2847                         bh = sb_bread(sb, ext4_idx_pblock(path[i].p_idx));
2848                         if (!bh) {
2849                                 /* should we reset i_size? */
2850                                 err = -EIO;
2851                                 break;
2852                         }
2853                         if (WARN_ON(i + 1 > depth)) {
2854                                 err = -EIO;
2855                                 break;
2856                         }
2857                         if (ext4_ext_check_block(inode, ext_block_hdr(bh),
2858                                                         depth - i - 1, bh)) {
2859                                 err = -EIO;
2860                                 break;
2861                         }
2862                         path[i + 1].p_bh = bh;
2863
2864                         /* save actual number of indexes since this
2865                          * number is changed at the next iteration */
2866                         path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2867                         i++;
2868                 } else {
2869                         /* we finished processing this index, go up */
2870                         if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2871                                 /* index is empty, remove it;
2872                                  * handle must be already prepared by the
2873                                  * truncatei_leaf() */
2874                                 err = ext4_ext_rm_idx(handle, inode, path, i);
2875                         }
2876                         /* root level has p_bh == NULL, brelse() eats this */
2877                         brelse(path[i].p_bh);
2878                         path[i].p_bh = NULL;
2879                         i--;
2880                         ext_debug("return to level %d\n", i);
2881                 }
2882         }
2883
2884         trace_ext4_ext_remove_space_done(inode, start, depth, partial_cluster,
2885                         path->p_hdr->eh_entries);
2886
2887         /* If we still have something in the partial cluster and we have removed
2888          * even the first extent, then we should free the blocks in the partial
2889          * cluster as well. */
2890         if (partial_cluster && path->p_hdr->eh_entries == 0) {
2891                 int flags = EXT4_FREE_BLOCKS_FORGET;
2892
2893                 if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2894                         flags |= EXT4_FREE_BLOCKS_METADATA;
2895
2896                 ext4_free_blocks(handle, inode, NULL,
2897                                  EXT4_C2B(EXT4_SB(sb), partial_cluster),
2898                                  EXT4_SB(sb)->s_cluster_ratio, flags);
2899                 partial_cluster = 0;
2900         }
2901
2902         /* TODO: flexible tree reduction should be here */
2903         if (path->p_hdr->eh_entries == 0) {
2904                 /*
2905                  * truncate to zero freed all the tree,
2906                  * so we need to correct eh_depth
2907                  */
2908                 err = ext4_ext_get_access(handle, inode, path);
2909                 if (err == 0) {
2910                         ext_inode_hdr(inode)->eh_depth = 0;
2911                         ext_inode_hdr(inode)->eh_max =
2912                                 cpu_to_le16(ext4_ext_space_root(inode, 0));
2913                         err = ext4_ext_dirty(handle, inode, path);
2914                 }
2915         }
2916 out:
2917         ext4_ext_drop_refs(path);
2918         kfree(path);
2919         if (err == -EAGAIN) {
2920                 path = NULL;
2921                 goto again;
2922         }
2923         ext4_journal_stop(handle);
2924
2925         return err;
2926 }
2927
2928 /*
2929  * called at mount time
2930  */
2931 void ext4_ext_init(struct super_block *sb)
2932 {
2933         /*
2934          * possible initialization would be here
2935          */
2936
2937         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
2938 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
2939                 printk(KERN_INFO "EXT4-fs: file extents enabled"
2940 #ifdef AGGRESSIVE_TEST
2941                        ", aggressive tests"
2942 #endif
2943 #ifdef CHECK_BINSEARCH
2944                        ", check binsearch"
2945 #endif
2946 #ifdef EXTENTS_STATS
2947                        ", stats"
2948 #endif
2949                        "\n");
2950 #endif
2951 #ifdef EXTENTS_STATS
2952                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
2953                 EXT4_SB(sb)->s_ext_min = 1 << 30;
2954                 EXT4_SB(sb)->s_ext_max = 0;
2955 #endif
2956         }
2957 }
2958
2959 /*
2960  * called at umount time
2961  */
2962 void ext4_ext_release(struct super_block *sb)
2963 {
2964         if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
2965                 return;
2966
2967 #ifdef EXTENTS_STATS
2968         if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
2969                 struct ext4_sb_info *sbi = EXT4_SB(sb);
2970                 printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
2971                         sbi->s_ext_blocks, sbi->s_ext_extents,
2972                         sbi->s_ext_blocks / sbi->s_ext_extents);
2973                 printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
2974                         sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
2975         }
2976 #endif
2977 }
2978
2979 /* FIXME!! we need to try to merge to left or right after zero-out  */
2980 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
2981 {
2982         ext4_fsblk_t ee_pblock;
2983         unsigned int ee_len;
2984         int ret;
2985
2986         ee_len    = ext4_ext_get_actual_len(ex);
2987         ee_pblock = ext4_ext_pblock(ex);
2988
2989         ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
2990         if (ret > 0)
2991                 ret = 0;
2992
2993         return ret;
2994 }
2995
2996 /*
2997  * ext4_split_extent_at() splits an extent at given block.
2998  *
2999  * @handle: the journal handle
3000  * @inode: the file inode
3001  * @path: the path to the extent
3002  * @split: the logical block where the extent is splitted.
3003  * @split_flags: indicates if the extent could be zeroout if split fails, and
3004  *               the states(init or uninit) of new extents.
3005  * @flags: flags used to insert new extent to extent tree.
3006  *
3007  *
3008  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3009  * of which are deterimined by split_flag.
3010  *
3011  * There are two cases:
3012  *  a> the extent are splitted into two extent.
3013  *  b> split is not needed, and just mark the extent.
3014  *
3015  * return 0 on success.
3016  */
3017 static int ext4_split_extent_at(handle_t *handle,
3018                              struct inode *inode,
3019                              struct ext4_ext_path *path,
3020                              ext4_lblk_t split,
3021                              int split_flag,
3022                              int flags)
3023 {
3024         ext4_fsblk_t newblock;
3025         ext4_lblk_t ee_block;
3026         struct ext4_extent *ex, newex, orig_ex, zero_ex;
3027         struct ext4_extent *ex2 = NULL;
3028         unsigned int ee_len, depth;
3029         int err = 0;
3030
3031         BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3032                (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3033
3034         ext_debug("ext4_split_extents_at: inode %lu, logical"
3035                 "block %llu\n", inode->i_ino, (unsigned long long)split);
3036
3037         ext4_ext_show_leaf(inode, path);
3038
3039         depth = ext_depth(inode);
3040         ex = path[depth].p_ext;
3041         ee_block = le32_to_cpu(ex->ee_block);
3042         ee_len = ext4_ext_get_actual_len(ex);
3043         newblock = split - ee_block + ext4_ext_pblock(ex);
3044
3045         BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3046         BUG_ON(!ext4_ext_is_uninitialized(ex) &&
3047                split_flag & (EXT4_EXT_MAY_ZEROOUT |
3048                              EXT4_EXT_MARK_UNINIT1 |
3049                              EXT4_EXT_MARK_UNINIT2));
3050
3051         err = ext4_ext_get_access(handle, inode, path + depth);
3052         if (err)
3053                 goto out;
3054
3055         if (split == ee_block) {
3056                 /*
3057                  * case b: block @split is the block that the extent begins with
3058                  * then we just change the state of the extent, and splitting
3059                  * is not needed.
3060                  */
3061                 if (split_flag & EXT4_EXT_MARK_UNINIT2)
3062                         ext4_ext_mark_uninitialized(ex);
3063                 else
3064                         ext4_ext_mark_initialized(ex);
3065
3066                 if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3067                         ext4_ext_try_to_merge(handle, inode, path, ex);
3068
3069                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3070                 goto out;
3071         }
3072
3073         /* case a */
3074         memcpy(&orig_ex, ex, sizeof(orig_ex));
3075         ex->ee_len = cpu_to_le16(split - ee_block);
3076         if (split_flag & EXT4_EXT_MARK_UNINIT1)
3077                 ext4_ext_mark_uninitialized(ex);
3078
3079         /*
3080          * path may lead to new leaf, not to original leaf any more
3081          * after ext4_ext_insert_extent() returns,
3082          */
3083         err = ext4_ext_dirty(handle, inode, path + depth);
3084         if (err)
3085                 goto fix_extent_len;
3086
3087         ex2 = &newex;
3088         ex2->ee_block = cpu_to_le32(split);
3089         ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3090         ext4_ext_store_pblock(ex2, newblock);
3091         if (split_flag & EXT4_EXT_MARK_UNINIT2)
3092                 ext4_ext_mark_uninitialized(ex2);
3093
3094         err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
3095         if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3096                 if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3097                         if (split_flag & EXT4_EXT_DATA_VALID1) {
3098                                 err = ext4_ext_zeroout(inode, ex2);
3099                                 zero_ex.ee_block = ex2->ee_block;
3100                                 zero_ex.ee_len = cpu_to_le16(
3101                                                 ext4_ext_get_actual_len(ex2));
3102                                 ext4_ext_store_pblock(&zero_ex,
3103                                                       ext4_ext_pblock(ex2));
3104                         } else {
3105                                 err = ext4_ext_zeroout(inode, ex);
3106                                 zero_ex.ee_block = ex->ee_block;
3107                                 zero_ex.ee_len = cpu_to_le16(
3108                                                 ext4_ext_get_actual_len(ex));
3109                                 ext4_ext_store_pblock(&zero_ex,
3110                                                       ext4_ext_pblock(ex));
3111                         }
3112                 } else {
3113                         err = ext4_ext_zeroout(inode, &orig_ex);
3114                         zero_ex.ee_block = orig_ex.ee_block;
3115                         zero_ex.ee_len = cpu_to_le16(
3116                                                 ext4_ext_get_actual_len(&orig_ex));
3117                         ext4_ext_store_pblock(&zero_ex,
3118                                               ext4_ext_pblock(&orig_ex));
3119                 }
3120
3121                 if (err)
3122                         goto fix_extent_len;
3123                 /* update the extent length and mark as initialized */
3124                 ex->ee_len = cpu_to_le16(ee_len);
3125                 ext4_ext_try_to_merge(handle, inode, path, ex);
3126                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3127                 if (err)
3128                         goto fix_extent_len;
3129
3130                 /* update extent status tree */
3131                 err = ext4_es_zeroout(inode, &zero_ex);
3132
3133                 goto out;
3134         } else if (err)
3135                 goto fix_extent_len;
3136
3137 out:
3138         ext4_ext_show_leaf(inode, path);
3139         return err;
3140
3141 fix_extent_len:
3142         ex->ee_len = orig_ex.ee_len;
3143         ext4_ext_dirty(handle, inode, path + depth);
3144         return err;
3145 }
3146
3147 /*
3148  * ext4_split_extents() splits an extent and mark extent which is covered
3149  * by @map as split_flags indicates
3150  *
3151  * It may result in splitting the extent into multiple extents (upto three)
3152  * There are three possibilities:
3153  *   a> There is no split required
3154  *   b> Splits in two extents: Split is happening at either end of the extent
3155  *   c> Splits in three extents: Somone is splitting in middle of the extent
3156  *
3157  */
3158 static int ext4_split_extent(handle_t *handle,
3159                               struct inode *inode,
3160                               struct ext4_ext_path *path,
3161                               struct ext4_map_blocks *map,
3162                               int split_flag,
3163                               int flags)
3164 {
3165         ext4_lblk_t ee_block;
3166         struct ext4_extent *ex;
3167         unsigned int ee_len, depth;
3168         int err = 0;
3169         int uninitialized;
3170         int split_flag1, flags1;
3171         int allocated = map->m_len;
3172
3173         depth = ext_depth(inode);
3174         ex = path[depth].p_ext;
3175         ee_block = le32_to_cpu(ex->ee_block);
3176         ee_len = ext4_ext_get_actual_len(ex);
3177         uninitialized = ext4_ext_is_uninitialized(ex);
3178
3179         if (map->m_lblk + map->m_len < ee_block + ee_len) {
3180                 split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3181                 flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3182                 if (uninitialized)
3183                         split_flag1 |= EXT4_EXT_MARK_UNINIT1 |
3184                                        EXT4_EXT_MARK_UNINIT2;
3185                 if (split_flag & EXT4_EXT_DATA_VALID2)
3186                         split_flag1 |= EXT4_EXT_DATA_VALID1;
3187                 err = ext4_split_extent_at(handle, inode, path,
3188                                 map->m_lblk + map->m_len, split_flag1, flags1);
3189                 if (err)
3190                         goto out;
3191         } else {
3192                 allocated = ee_len - (map->m_lblk - ee_block);
3193         }
3194         /*
3195          * Update path is required because previous ext4_split_extent_at() may
3196          * result in split of original leaf or extent zeroout.
3197          */
3198         ext4_ext_drop_refs(path);
3199         path = ext4_ext_find_extent(inode, map->m_lblk, path);
3200         if (IS_ERR(path))
3201                 return PTR_ERR(path);
3202         depth = ext_depth(inode);
3203         ex = path[depth].p_ext;
3204         uninitialized = ext4_ext_is_uninitialized(ex);
3205         split_flag1 = 0;
3206
3207         if (map->m_lblk >= ee_block) {
3208                 split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3209                 if (uninitialized) {
3210                         split_flag1 |= EXT4_EXT_MARK_UNINIT1;
3211                         split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3212                                                      EXT4_EXT_MARK_UNINIT2);
3213                 }
3214                 err = ext4_split_extent_at(handle, inode, path,
3215                                 map->m_lblk, split_flag1, flags);
3216                 if (err)
3217                         goto out;
3218         }
3219
3220         ext4_ext_show_leaf(inode, path);
3221 out:
3222         return err ? err : allocated;
3223 }
3224
3225 /*
3226  * This function is called by ext4_ext_map_blocks() if someone tries to write
3227  * to an uninitialized extent. It may result in splitting the uninitialized
3228  * extent into multiple extents (up to three - one initialized and two
3229  * uninitialized).
3230  * There are three possibilities:
3231  *   a> There is no split required: Entire extent should be initialized
3232  *   b> Splits in two extents: Write is happening at either end of the extent
3233  *   c> Splits in three extents: Somone is writing in middle of the extent
3234  *
3235  * Pre-conditions:
3236  *  - The extent pointed to by 'path' is uninitialized.
3237  *  - The extent pointed to by 'path' contains a superset
3238  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3239  *
3240  * Post-conditions on success:
3241  *  - the returned value is the number of blocks beyond map->l_lblk
3242  *    that are allocated and initialized.
3243  *    It is guaranteed to be >= map->m_len.
3244  */
3245 static int ext4_ext_convert_to_initialized(handle_t *handle,
3246                                            struct inode *inode,
3247                                            struct ext4_map_blocks *map,
3248                                            struct ext4_ext_path *path,
3249                                            int flags)
3250 {
3251         struct ext4_sb_info *sbi;
3252         struct ext4_extent_header *eh;
3253         struct ext4_map_blocks split_map;
3254         struct ext4_extent zero_ex;
3255         struct ext4_extent *ex, *abut_ex;
3256         ext4_lblk_t ee_block, eof_block;
3257         unsigned int ee_len, depth, map_len = map->m_len;
3258         int allocated = 0, max_zeroout = 0;
3259         int err = 0;
3260         int split_flag = 0;
3261
3262         ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3263                 "block %llu, max_blocks %u\n", inode->i_ino,
3264                 (unsigned long long)map->m_lblk, map_len);
3265
3266         sbi = EXT4_SB(inode->i_sb);
3267         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3268                 inode->i_sb->s_blocksize_bits;
3269         if (eof_block < map->m_lblk + map_len)
3270                 eof_block = map->m_lblk + map_len;
3271
3272         depth = ext_depth(inode);
3273         eh = path[depth].p_hdr;
3274         ex = path[depth].p_ext;
3275         ee_block = le32_to_cpu(ex->ee_block);
3276         ee_len = ext4_ext_get_actual_len(ex);
3277         zero_ex.ee_len = 0;
3278
3279         trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3280
3281         /* Pre-conditions */
3282         BUG_ON(!ext4_ext_is_uninitialized(ex));
3283         BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3284
3285         /*
3286          * Attempt to transfer newly initialized blocks from the currently
3287          * uninitialized extent to its neighbor. This is much cheaper
3288          * than an insertion followed by a merge as those involve costly
3289          * memmove() calls. Transferring to the left is the common case in
3290          * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3291          * followed by append writes.
3292          *
3293          * Limitations of the current logic:
3294          *  - L1: we do not deal with writes covering the whole extent.
3295          *    This would require removing the extent if the transfer
3296          *    is possible.
3297          *  - L2: we only attempt to merge with an extent stored in the
3298          *    same extent tree node.
3299          */
3300         if ((map->m_lblk == ee_block) &&
3301                 /* See if we can merge left */
3302                 (map_len < ee_len) &&           /*L1*/
3303                 (ex > EXT_FIRST_EXTENT(eh))) {  /*L2*/
3304                 ext4_lblk_t prev_lblk;
3305                 ext4_fsblk_t prev_pblk, ee_pblk;
3306                 unsigned int prev_len;
3307
3308                 abut_ex = ex - 1;
3309                 prev_lblk = le32_to_cpu(abut_ex->ee_block);
3310                 prev_len = ext4_ext_get_actual_len(abut_ex);
3311                 prev_pblk = ext4_ext_pblock(abut_ex);
3312                 ee_pblk = ext4_ext_pblock(ex);
3313
3314                 /*
3315                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3316                  * upon those conditions:
3317                  * - C1: abut_ex is initialized,
3318                  * - C2: abut_ex is logically abutting ex,
3319                  * - C3: abut_ex is physically abutting ex,
3320                  * - C4: abut_ex can receive the additional blocks without
3321                  *   overflowing the (initialized) length limit.
3322                  */
3323                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3324                         ((prev_lblk + prev_len) == ee_block) &&         /*C2*/
3325                         ((prev_pblk + prev_len) == ee_pblk) &&          /*C3*/
3326                         (prev_len < (EXT_INIT_MAX_LEN - map_len))) {    /*C4*/
3327                         err = ext4_ext_get_access(handle, inode, path + depth);
3328                         if (err)
3329                                 goto out;
3330
3331                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3332                                 map, ex, abut_ex);
3333
3334                         /* Shift the start of ex by 'map_len' blocks */
3335                         ex->ee_block = cpu_to_le32(ee_block + map_len);
3336                         ext4_ext_store_pblock(ex, ee_pblk + map_len);
3337                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3338                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3339
3340                         /* Extend abut_ex by 'map_len' blocks */
3341                         abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3342
3343                         /* Result: number of initialized blocks past m_lblk */
3344                         allocated = map_len;
3345                 }
3346         } else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3347                    (map_len < ee_len) &&        /*L1*/
3348                    ex < EXT_LAST_EXTENT(eh)) {  /*L2*/
3349                 /* See if we can merge right */
3350                 ext4_lblk_t next_lblk;
3351                 ext4_fsblk_t next_pblk, ee_pblk;
3352                 unsigned int next_len;
3353
3354                 abut_ex = ex + 1;
3355                 next_lblk = le32_to_cpu(abut_ex->ee_block);
3356                 next_len = ext4_ext_get_actual_len(abut_ex);
3357                 next_pblk = ext4_ext_pblock(abut_ex);
3358                 ee_pblk = ext4_ext_pblock(ex);
3359
3360                 /*
3361                  * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3362                  * upon those conditions:
3363                  * - C1: abut_ex is initialized,
3364                  * - C2: abut_ex is logically abutting ex,
3365                  * - C3: abut_ex is physically abutting ex,
3366                  * - C4: abut_ex can receive the additional blocks without
3367                  *   overflowing the (initialized) length limit.
3368                  */
3369                 if ((!ext4_ext_is_uninitialized(abut_ex)) &&            /*C1*/
3370                     ((map->m_lblk + map_len) == next_lblk) &&           /*C2*/
3371                     ((ee_pblk + ee_len) == next_pblk) &&                /*C3*/
3372                     (next_len < (EXT_INIT_MAX_LEN - map_len))) {        /*C4*/
3373                         err = ext4_ext_get_access(handle, inode, path + depth);
3374                         if (err)
3375                                 goto out;
3376
3377                         trace_ext4_ext_convert_to_initialized_fastpath(inode,
3378                                 map, ex, abut_ex);
3379
3380                         /* Shift the start of abut_ex by 'map_len' blocks */
3381                         abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3382                         ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3383                         ex->ee_len = cpu_to_le16(ee_len - map_len);
3384                         ext4_ext_mark_uninitialized(ex); /* Restore the flag */
3385
3386                         /* Extend abut_ex by 'map_len' blocks */
3387                         abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3388
3389                         /* Result: number of initialized blocks past m_lblk */
3390                         allocated = map_len;
3391                 }
3392         }
3393         if (allocated) {
3394                 /* Mark the block containing both extents as dirty */
3395                 ext4_ext_dirty(handle, inode, path + depth);
3396
3397                 /* Update path to point to the right extent */
3398                 path[depth].p_ext = abut_ex;
3399                 goto out;
3400         } else
3401                 allocated = ee_len - (map->m_lblk - ee_block);
3402
3403         WARN_ON(map->m_lblk < ee_block);
3404         /*
3405          * It is safe to convert extent to initialized via explicit
3406          * zeroout only if extent is fully insde i_size or new_size.
3407          */
3408         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3409
3410         if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3411                 max_zeroout = sbi->s_extent_max_zeroout_kb >>
3412                         (inode->i_sb->s_blocksize_bits - 10);
3413
3414         /* If extent is less than s_max_zeroout_kb, zeroout directly */
3415         if (max_zeroout && (ee_len <= max_zeroout)) {
3416                 err = ext4_ext_zeroout(inode, ex);
3417                 if (err)
3418                         goto out;
3419                 zero_ex.ee_block = ex->ee_block;
3420                 zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3421                 ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3422
3423                 err = ext4_ext_get_access(handle, inode, path + depth);
3424                 if (err)
3425                         goto out;
3426                 ext4_ext_mark_initialized(ex);
3427                 ext4_ext_try_to_merge(handle, inode, path, ex);
3428                 err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3429                 goto out;
3430         }
3431
3432         /*
3433          * four cases:
3434          * 1. split the extent into three extents.
3435          * 2. split the extent into two extents, zeroout the first half.
3436          * 3. split the extent into two extents, zeroout the second half.
3437          * 4. split the extent into two extents with out zeroout.
3438          */
3439         split_map.m_lblk = map->m_lblk;
3440         split_map.m_len = map->m_len;
3441
3442         if (max_zeroout && (allocated > map->m_len)) {
3443                 if (allocated <= max_zeroout) {
3444                         /* case 3 */
3445                         zero_ex.ee_block =
3446                                          cpu_to_le32(map->m_lblk);
3447                         zero_ex.ee_len = cpu_to_le16(allocated);
3448                         ext4_ext_store_pblock(&zero_ex,
3449                                 ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3450                         err = ext4_ext_zeroout(inode, &zero_ex);
3451                         if (err)
3452                                 goto out;
3453                         split_map.m_lblk = map->m_lblk;
3454                         split_map.m_len = allocated;
3455                 } else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3456                         /* case 2 */
3457                         if (map->m_lblk != ee_block) {
3458                                 zero_ex.ee_block = ex->ee_block;
3459                                 zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3460                                                         ee_block);
3461                                 ext4_ext_store_pblock(&zero_ex,
3462                                                       ext4_ext_pblock(ex));
3463                                 err = ext4_ext_zeroout(inode, &zero_ex);
3464                                 if (err)
3465                                         goto out;
3466                         }
3467
3468                         split_map.m_lblk = ee_block;
3469                         split_map.m_len = map->m_lblk - ee_block + map->m_len;
3470                         allocated = map->m_len;
3471                 }
3472         }
3473
3474         allocated = ext4_split_extent(handle, inode, path,
3475                                       &split_map, split_flag, flags);
3476         if (allocated < 0)
3477                 err = allocated;
3478
3479 out:
3480         /* If we have gotten a failure, don't zero out status tree */
3481         if (!err)
3482                 err = ext4_es_zeroout(inode, &zero_ex);
3483         return err ? err : allocated;
3484 }
3485
3486 /*
3487  * This function is called by ext4_ext_map_blocks() from
3488  * ext4_get_blocks_dio_write() when DIO to write
3489  * to an uninitialized extent.
3490  *
3491  * Writing to an uninitialized extent may result in splitting the uninitialized
3492  * extent into multiple initialized/uninitialized extents (up to three)
3493  * There are three possibilities:
3494  *   a> There is no split required: Entire extent should be uninitialized
3495  *   b> Splits in two extents: Write is happening at either end of the extent
3496  *   c> Splits in three extents: Somone is writing in middle of the extent
3497  *
3498  * One of more index blocks maybe needed if the extent tree grow after
3499  * the uninitialized extent split. To prevent ENOSPC occur at the IO
3500  * complete, we need to split the uninitialized extent before DIO submit
3501  * the IO. The uninitialized extent called at this time will be split
3502  * into three uninitialized extent(at most). After IO complete, the part
3503  * being filled will be convert to initialized by the end_io callback function
3504  * via ext4_convert_unwritten_extents().
3505  *
3506  * Returns the size of uninitialized extent to be written on success.
3507  */
3508 static int ext4_split_unwritten_extents(handle_t *handle,
3509                                         struct inode *inode,
3510                                         struct ext4_map_blocks *map,
3511                                         struct ext4_ext_path *path,
3512                                         int flags)
3513 {
3514         ext4_lblk_t eof_block;
3515         ext4_lblk_t ee_block;
3516         struct ext4_extent *ex;
3517         unsigned int ee_len;
3518         int split_flag = 0, depth;
3519
3520         ext_debug("ext4_split_unwritten_extents: inode %lu, logical"
3521                 "block %llu, max_blocks %u\n", inode->i_ino,
3522                 (unsigned long long)map->m_lblk, map->m_len);
3523
3524         eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3525                 inode->i_sb->s_blocksize_bits;
3526         if (eof_block < map->m_lblk + map->m_len)
3527                 eof_block = map->m_lblk + map->m_len;
3528         /*
3529          * It is safe to convert extent to initialized via explicit
3530          * zeroout only if extent is fully insde i_size or new_size.
3531          */
3532         depth = ext_depth(inode);
3533         ex = path[depth].p_ext;
3534         ee_block = le32_to_cpu(ex->ee_block);
3535         ee_len = ext4_ext_get_actual_len(ex);
3536
3537         split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3538         split_flag |= EXT4_EXT_MARK_UNINIT2;
3539         if (flags & EXT4_GET_BLOCKS_CONVERT)
3540                 split_flag |= EXT4_EXT_DATA_VALID2;
3541         flags |= EXT4_GET_BLOCKS_PRE_IO;
3542         return ext4_split_extent(handle, inode, path, map, split_flag, flags);
3543 }
3544
3545 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3546                                                 struct inode *inode,
3547                                                 struct ext4_map_blocks *map,
3548                                                 struct ext4_ext_path *path)
3549 {
3550         struct ext4_extent *ex;
3551         ext4_lblk_t ee_block;
3552         unsigned int ee_len;
3553         int depth;
3554         int err = 0;
3555
3556         depth = ext_depth(inode);
3557         ex = path[depth].p_ext;
3558         ee_block = le32_to_cpu(ex->ee_block);
3559         ee_len = ext4_ext_get_actual_len(ex);
3560
3561         ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3562                 "block %llu, max_blocks %u\n", inode->i_ino,
3563                   (unsigned long long)ee_block, ee_len);
3564
3565         /* If extent is larger than requested it is a clear sign that we still
3566          * have some extent state machine issues left. So extent_split is still
3567          * required.
3568          * TODO: Once all related issues will be fixed this situation should be
3569          * illegal.
3570          */
3571         if (ee_block != map->m_lblk || ee_len > map->m_len) {
3572 #ifdef EXT4_DEBUG
3573                 ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3574                              " len %u; IO logical block %llu, len %u\n",
3575                              inode->i_ino, (unsigned long long)ee_block, ee_len,
3576                              (unsigned long long)map->m_lblk, map->m_len);
3577 #endif
3578                 err = ext4_split_unwritten_extents(handle, inode, map, path,
3579                                                    EXT4_GET_BLOCKS_CONVERT);
3580                 if (err < 0)
3581                         goto out;
3582                 ext4_ext_drop_refs(path);
3583                 path = ext4_ext_find_extent(inode, map->m_lblk, path);
3584                 if (IS_ERR(path)) {
3585                         err = PTR_ERR(path);
3586                         goto out;
3587                 }
3588                 depth = ext_depth(inode);
3589                 ex = path[depth].p_ext;
3590         }
3591
3592         err = ext4_ext_get_access(handle, inode, path + depth);
3593         if (err)
3594                 goto out;
3595         /* first mark the extent as initialized */
3596         ext4_ext_mark_initialized(ex);
3597
3598         /* note: ext4_ext_correct_indexes() isn't needed here because
3599          * borders are not changed
3600          */
3601         ext4_ext_try_to_merge(handle, inode, path, ex);
3602
3603         /* Mark modified extent as dirty */
3604         err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3605 out:
3606         ext4_ext_show_leaf(inode, path);
3607         return err;
3608 }
3609
3610 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3611                         sector_t block, int count)
3612 {
3613         int i;
3614         for (i = 0; i < count; i++)
3615                 unmap_underlying_metadata(bdev, block + i);
3616 }
3617
3618 /*
3619  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3620  */
3621 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3622                               ext4_lblk_t lblk,
3623                               struct ext4_ext_path *path,
3624                               unsigned int len)
3625 {
3626         int i, depth;
3627         struct ext4_extent_header *eh;
3628         struct ext4_extent *last_ex;
3629
3630         if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3631                 return 0;
3632
3633         depth = ext_depth(inode);
3634         eh = path[depth].p_hdr;
3635
3636         /*
3637          * We're going to remove EOFBLOCKS_FL entirely in future so we
3638          * do not care for this case anymore. Simply remove the flag
3639          * if there are no extents.
3640          */
3641         if (unlikely(!eh->eh_entries))
3642                 goto out;
3643         last_ex = EXT_LAST_EXTENT(eh);
3644         /*
3645          * We should clear the EOFBLOCKS_FL flag if we are writing the
3646          * last block in the last extent in the file.  We test this by
3647          * first checking to see if the caller to
3648          * ext4_ext_get_blocks() was interested in the last block (or
3649          * a block beyond the last block) in the current extent.  If
3650          * this turns out to be false, we can bail out from this
3651          * function immediately.
3652          */
3653         if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3654             ext4_ext_get_actual_len(last_ex))
3655                 return 0;
3656         /*
3657          * If the caller does appear to be planning to write at or
3658          * beyond the end of the current extent, we then test to see
3659          * if the current extent is the last extent in the file, by
3660          * checking to make sure it was reached via the rightmost node
3661          * at each level of the tree.
3662          */
3663         for (i = depth-1; i >= 0; i--)
3664                 if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3665                         return 0;
3666 out:
3667         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3668         return ext4_mark_inode_dirty(handle, inode);
3669 }
3670
3671 /**
3672  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3673  *
3674  * Return 1 if there is a delalloc block in the range, otherwise 0.
3675  */
3676 int ext4_find_delalloc_range(struct inode *inode,
3677                              ext4_lblk_t lblk_start,
3678                              ext4_lblk_t lblk_end)
3679 {
3680         struct extent_status es;
3681
3682         ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3683         if (es.es_len == 0)
3684                 return 0; /* there is no delay extent in this tree */
3685         else if (es.es_lblk <= lblk_start &&
3686                  lblk_start < es.es_lblk + es.es_len)
3687                 return 1;
3688         else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3689                 return 1;
3690         else
3691                 return 0;
3692 }
3693
3694 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3695 {
3696         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3697         ext4_lblk_t lblk_start, lblk_end;
3698         lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3699         lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3700
3701         return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3702 }
3703
3704 /**
3705  * Determines how many complete clusters (out of those specified by the 'map')
3706  * are under delalloc and were reserved quota for.
3707  * This function is called when we are writing out the blocks that were
3708  * originally written with their allocation delayed, but then the space was
3709  * allocated using fallocate() before the delayed allocation could be resolved.
3710  * The cases to look for are:
3711  * ('=' indicated delayed allocated blocks
3712  *  '-' indicates non-delayed allocated blocks)
3713  * (a) partial clusters towards beginning and/or end outside of allocated range
3714  *     are not delalloc'ed.
3715  *      Ex:
3716  *      |----c---=|====c====|====c====|===-c----|
3717  *               |++++++ allocated ++++++|
3718  *      ==> 4 complete clusters in above example
3719  *
3720  * (b) partial cluster (outside of allocated range) towards either end is
3721  *     marked for delayed allocation. In this case, we will exclude that
3722  *     cluster.
3723  *      Ex:
3724  *      |----====c========|========c========|
3725  *           |++++++ allocated ++++++|
3726  *      ==> 1 complete clusters in above example
3727  *
3728  *      Ex:
3729  *      |================c================|
3730  *            |++++++ allocated ++++++|
3731  *      ==> 0 complete clusters in above example
3732  *
3733  * The ext4_da_update_reserve_space will be called only if we
3734  * determine here that there were some "entire" clusters that span
3735  * this 'allocated' range.
3736  * In the non-bigalloc case, this function will just end up returning num_blks
3737  * without ever calling ext4_find_delalloc_range.
3738  */
3739 static unsigned int
3740 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3741                            unsigned int num_blks)
3742 {
3743         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3744         ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3745         ext4_lblk_t lblk_from, lblk_to, c_offset;
3746         unsigned int allocated_clusters = 0;
3747
3748         alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3749         alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3750
3751         /* max possible clusters for this allocation */
3752         allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3753
3754         trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3755
3756         /* Check towards left side */
3757         c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3758         if (c_offset) {
3759                 lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3760                 lblk_to = lblk_from + c_offset - 1;
3761
3762                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3763                         allocated_clusters--;
3764         }
3765
3766         /* Now check towards right. */
3767         c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3768         if (allocated_clusters && c_offset) {
3769                 lblk_from = lblk_start + num_blks;
3770                 lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3771
3772                 if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3773                         allocated_clusters--;
3774         }
3775
3776         return allocated_clusters;
3777 }
3778
3779 static int
3780 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
3781                         struct ext4_map_blocks *map,
3782                         struct ext4_ext_path *path, int flags,
3783                         unsigned int allocated, ext4_fsblk_t newblock)
3784 {
3785         int ret = 0;
3786         int err = 0;
3787         ext4_io_end_t *io = ext4_inode_aio(inode);
3788
3789         ext_debug("ext4_ext_handle_uninitialized_extents: inode %lu, logical "
3790                   "block %llu, max_blocks %u, flags %x, allocated %u\n",
3791                   inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
3792                   flags, allocated);
3793         ext4_ext_show_leaf(inode, path);
3794
3795         /*
3796          * When writing into uninitialized space, we should not fail to
3797          * allocate metadata blocks for the new extent block if needed.
3798          */
3799         flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
3800
3801         trace_ext4_ext_handle_uninitialized_extents(inode, map, flags,
3802                                                     allocated, newblock);
3803
3804         /* get_block() before submit the IO, split the extent */
3805         if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
3806                 ret = ext4_split_unwritten_extents(handle, inode, map,
3807                                                    path, flags);
3808                 if (ret <= 0)
3809                         goto out;
3810                 /*
3811                  * Flag the inode(non aio case) or end_io struct (aio case)
3812                  * that this IO needs to conversion to written when IO is
3813                  * completed
3814                  */
3815                 if (io)
3816                         ext4_set_io_unwritten_flag(inode, io);
3817                 else
3818                         ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3819                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3820                 if (ext4_should_dioread_nolock(inode))
3821                         map->m_flags |= EXT4_MAP_UNINIT;
3822                 goto out;
3823         }
3824         /* IO end_io complete, convert the filled extent to written */
3825         if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
3826                 ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
3827                                                         path);
3828                 if (ret >= 0) {
3829                         ext4_update_inode_fsync_trans(handle, inode, 1);
3830                         err = check_eofblocks_fl(handle, inode, map->m_lblk,
3831                                                  path, map->m_len);
3832                 } else
3833                         err = ret;
3834                 map->m_flags |= EXT4_MAP_MAPPED;
3835                 if (allocated > map->m_len)
3836                         allocated = map->m_len;
3837                 map->m_len = allocated;
3838                 goto out2;
3839         }
3840         /* buffered IO case */
3841         /*
3842          * repeat fallocate creation request
3843          * we already have an unwritten extent
3844          */
3845         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT) {
3846                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3847                 goto map_out;
3848         }
3849
3850         /* buffered READ or buffered write_begin() lookup */
3851         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
3852                 /*
3853                  * We have blocks reserved already.  We
3854                  * return allocated blocks so that delalloc
3855                  * won't do block reservation for us.  But
3856                  * the buffer head will be unmapped so that
3857                  * a read from the block returns 0s.
3858                  */
3859                 map->m_flags |= EXT4_MAP_UNWRITTEN;
3860                 goto out1;
3861         }
3862
3863         /* buffered write, writepage time, convert*/
3864         ret = ext4_ext_convert_to_initialized(handle, inode, map, path, flags);
3865         if (ret >= 0)
3866                 ext4_update_inode_fsync_trans(handle, inode, 1);
3867 out:
3868         if (ret <= 0) {
3869                 err = ret;
3870                 goto out2;
3871         } else
3872                 allocated = ret;
3873         map->m_flags |= EXT4_MAP_NEW;
3874         /*
3875          * if we allocated more blocks than requested
3876          * we need to make sure we unmap the extra block
3877          * allocated. The actual needed block will get
3878          * unmapped later when we find the buffer_head marked
3879          * new.
3880          */
3881         if (allocated > map->m_len) {
3882                 unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
3883                                         newblock + map->m_len,
3884                                         allocated - map->m_len);
3885                 allocated = map->m_len;
3886         }
3887         map->m_len = allocated;
3888
3889         /*
3890          * If we have done fallocate with the offset that is already
3891          * delayed allocated, we would have block reservation
3892          * and quota reservation done in the delayed write path.
3893          * But fallocate would have already updated quota and block
3894          * count for this offset. So cancel these reservation
3895          */
3896         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
3897                 unsigned int reserved_clusters;
3898                 reserved_clusters = get_reserved_cluster_alloc(inode,
3899                                 map->m_lblk, map->m_len);
3900                 if (reserved_clusters)
3901                         ext4_da_update_reserve_space(inode,
3902                                                      reserved_clusters,
3903                                                      0);
3904         }
3905
3906 map_out:
3907         map->m_flags |= EXT4_MAP_MAPPED;
3908         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
3909                 err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
3910                                          map->m_len);
3911                 if (err < 0)
3912                         goto out2;
3913         }
3914 out1:
3915         if (allocated > map->m_len)
3916                 allocated = map->m_len;
3917         ext4_ext_show_leaf(inode, path);
3918         map->m_pblk = newblock;
3919         map->m_len = allocated;
3920 out2:
3921         if (path) {
3922                 ext4_ext_drop_refs(path);
3923                 kfree(path);
3924         }
3925         return err ? err : allocated;
3926 }
3927
3928 /*
3929  * get_implied_cluster_alloc - check to see if the requested
3930  * allocation (in the map structure) overlaps with a cluster already
3931  * allocated in an extent.
3932  *      @sb     The filesystem superblock structure
3933  *      @map    The requested lblk->pblk mapping
3934  *      @ex     The extent structure which might contain an implied
3935  *                      cluster allocation
3936  *
3937  * This function is called by ext4_ext_map_blocks() after we failed to
3938  * find blocks that were already in the inode's extent tree.  Hence,
3939  * we know that the beginning of the requested region cannot overlap
3940  * the extent from the inode's extent tree.  There are three cases we
3941  * want to catch.  The first is this case:
3942  *
3943  *               |--- cluster # N--|
3944  *    |--- extent ---|  |---- requested region ---|
3945  *                      |==========|
3946  *
3947  * The second case that we need to test for is this one:
3948  *
3949  *   |--------- cluster # N ----------------|
3950  *         |--- requested region --|   |------- extent ----|
3951  *         |=======================|
3952  *
3953  * The third case is when the requested region lies between two extents
3954  * within the same cluster:
3955  *          |------------- cluster # N-------------|
3956  * |----- ex -----|                  |---- ex_right ----|
3957  *                  |------ requested region ------|
3958  *                  |================|
3959  *
3960  * In each of the above cases, we need to set the map->m_pblk and
3961  * map->m_len so it corresponds to the return the extent labelled as
3962  * "|====|" from cluster #N, since it is already in use for data in
3963  * cluster EXT4_B2C(sbi, map->m_lblk).  We will then return 1 to
3964  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
3965  * as a new "allocated" block region.  Otherwise, we will return 0 and
3966  * ext4_ext_map_blocks() will then allocate one or more new clusters
3967  * by calling ext4_mb_new_blocks().
3968  */
3969 static int get_implied_cluster_alloc(struct super_block *sb,
3970                                      struct ext4_map_blocks *map,
3971                                      struct ext4_extent *ex,
3972                                      struct ext4_ext_path *path)
3973 {
3974         struct ext4_sb_info *sbi = EXT4_SB(sb);
3975         ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
3976         ext4_lblk_t ex_cluster_start, ex_cluster_end;
3977         ext4_lblk_t rr_cluster_start;
3978         ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
3979         ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
3980         unsigned short ee_len = ext4_ext_get_actual_len(ex);
3981
3982         /* The extent passed in that we are trying to match */
3983         ex_cluster_start = EXT4_B2C(sbi, ee_block);
3984         ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
3985
3986         /* The requested region passed into ext4_map_blocks() */
3987         rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
3988
3989         if ((rr_cluster_start == ex_cluster_end) ||
3990             (rr_cluster_start == ex_cluster_start)) {
3991                 if (rr_cluster_start == ex_cluster_end)
3992                         ee_start += ee_len - 1;
3993                 map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
3994                 map->m_len = min(map->m_len,
3995                                  (unsigned) sbi->s_cluster_ratio - c_offset);
3996                 /*
3997                  * Check for and handle this case:
3998                  *
3999                  *   |--------- cluster # N-------------|
4000                  *                     |------- extent ----|
4001                  *         |--- requested region ---|
4002                  *         |===========|
4003                  */
4004
4005                 if (map->m_lblk < ee_block)
4006                         map->m_len = min(map->m_len, ee_block - map->m_lblk);
4007
4008                 /*
4009                  * Check for the case where there is already another allocated
4010                  * block to the right of 'ex' but before the end of the cluster.
4011                  *
4012                  *          |------------- cluster # N-------------|
4013                  * |----- ex -----|                  |---- ex_right ----|
4014                  *                  |------ requested region ------|
4015                  *                  |================|
4016                  */
4017                 if (map->m_lblk > ee_block) {
4018                         ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4019                         map->m_len = min(map->m_len, next - map->m_lblk);
4020                 }
4021
4022                 trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4023                 return 1;
4024         }
4025
4026         trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4027         return 0;
4028 }
4029
4030
4031 /*
4032  * Block allocation/map/preallocation routine for extents based files
4033  *
4034  *
4035  * Need to be called with
4036  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4037  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4038  *
4039  * return > 0, number of of blocks already mapped/allocated
4040  *          if create == 0 and these are pre-allocated blocks
4041  *              buffer head is unmapped
4042  *          otherwise blocks are mapped
4043  *
4044  * return = 0, if plain look up failed (blocks have not been allocated)
4045  *          buffer head is unmapped
4046  *
4047  * return < 0, error case.
4048  */
4049 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4050                         struct ext4_map_blocks *map, int flags)
4051 {
4052         struct ext4_ext_path *path = NULL;
4053         struct ext4_extent newex, *ex, *ex2;
4054         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4055         ext4_fsblk_t newblock = 0;
4056         int free_on_err = 0, err = 0, depth, ret;
4057         unsigned int allocated = 0, offset = 0;
4058         unsigned int allocated_clusters = 0;
4059         struct ext4_allocation_request ar;
4060         ext4_io_end_t *io = ext4_inode_aio(inode);
4061         ext4_lblk_t cluster_offset;
4062         int set_unwritten = 0;
4063
4064         ext_debug("blocks %u/%u requested for inode %lu\n",
4065                   map->m_lblk, map->m_len, inode->i_ino);
4066         trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4067
4068         /* find extent for this block */
4069         path = ext4_ext_find_extent(inode, map->m_lblk, NULL);
4070         if (IS_ERR(path)) {
4071                 err = PTR_ERR(path);
4072                 path = NULL;
4073                 goto out2;
4074         }
4075
4076         depth = ext_depth(inode);
4077
4078         /*
4079          * consistent leaf must not be empty;
4080          * this situation is possible, though, _during_ tree modification;
4081          * this is why assert can't be put in ext4_ext_find_extent()
4082          */
4083         if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4084                 EXT4_ERROR_INODE(inode, "bad extent address "
4085                                  "lblock: %lu, depth: %d pblock %lld",
4086                                  (unsigned long) map->m_lblk, depth,
4087                                  path[depth].p_block);
4088                 err = -EIO;
4089                 goto out2;
4090         }
4091
4092         ex = path[depth].p_ext;
4093         if (ex) {
4094                 ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4095                 ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4096                 unsigned short ee_len;
4097
4098                 /*
4099                  * Uninitialized extents are treated as holes, except that
4100                  * we split out initialized portions during a write.
4101                  */
4102                 ee_len = ext4_ext_get_actual_len(ex);
4103
4104                 trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4105
4106                 /* if found extent covers block, simply return it */
4107                 if (in_range(map->m_lblk, ee_block, ee_len)) {
4108                         newblock = map->m_lblk - ee_block + ee_start;
4109                         /* number of remaining blocks in the extent */
4110                         allocated = ee_len - (map->m_lblk - ee_block);
4111                         ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4112                                   ee_block, ee_len, newblock);
4113
4114                         if (!ext4_ext_is_uninitialized(ex))
4115                                 goto out;
4116
4117                         ret = ext4_ext_handle_uninitialized_extents(
4118                                 handle, inode, map, path, flags,
4119                                 allocated, newblock);
4120                         if (ret < 0)
4121                                 err = ret;
4122                         else
4123                                 allocated = ret;
4124                         goto out3;
4125                 }
4126         }
4127
4128         if ((sbi->s_cluster_ratio > 1) &&
4129             ext4_find_delalloc_cluster(inode, map->m_lblk))
4130                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4131
4132         /*
4133          * requested block isn't allocated yet;
4134          * we couldn't try to create block if create flag is zero
4135          */
4136         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4137                 /*
4138                  * put just found gap into cache to speed up
4139                  * subsequent requests
4140                  */
4141                 if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4142                         ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4143                 goto out2;
4144         }
4145
4146         /*
4147          * Okay, we need to do block allocation.
4148          */
4149         map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4150         newex.ee_block = cpu_to_le32(map->m_lblk);
4151         cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4152
4153         /*
4154          * If we are doing bigalloc, check to see if the extent returned
4155          * by ext4_ext_find_extent() implies a cluster we can use.
4156          */
4157         if (cluster_offset && ex &&
4158             get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4159                 ar.len = allocated = map->m_len;
4160                 newblock = map->m_pblk;
4161                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4162                 goto got_allocated_blocks;
4163         }
4164
4165         /* find neighbour allocated blocks */
4166         ar.lleft = map->m_lblk;
4167         err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4168         if (err)
4169                 goto out2;
4170         ar.lright = map->m_lblk;
4171         ex2 = NULL;
4172         err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4173         if (err)
4174                 goto out2;
4175
4176         /* Check if the extent after searching to the right implies a
4177          * cluster we can use. */
4178         if ((sbi->s_cluster_ratio > 1) && ex2 &&
4179             get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4180                 ar.len = allocated = map->m_len;
4181                 newblock = map->m_pblk;
4182                 map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4183                 goto got_allocated_blocks;
4184         }
4185
4186         /*
4187          * See if request is beyond maximum number of blocks we can have in
4188          * a single extent. For an initialized extent this limit is
4189          * EXT_INIT_MAX_LEN and for an uninitialized extent this limit is
4190          * EXT_UNINIT_MAX_LEN.
4191          */
4192         if (map->m_len > EXT_INIT_MAX_LEN &&
4193             !(flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4194                 map->m_len = EXT_INIT_MAX_LEN;
4195         else if (map->m_len > EXT_UNINIT_MAX_LEN &&
4196                  (flags & EXT4_GET_BLOCKS_UNINIT_EXT))
4197                 map->m_len = EXT_UNINIT_MAX_LEN;
4198
4199         /* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4200         newex.ee_len = cpu_to_le16(map->m_len);
4201         err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4202         if (err)
4203                 allocated = ext4_ext_get_actual_len(&newex);
4204         else
4205                 allocated = map->m_len;
4206
4207         /* allocate new block */
4208         ar.inode = inode;
4209         ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4210         ar.logical = map->m_lblk;
4211         /*
4212          * We calculate the offset from the beginning of the cluster
4213          * for the logical block number, since when we allocate a
4214          * physical cluster, the physical block should start at the
4215          * same offset from the beginning of the cluster.  This is
4216          * needed so that future calls to get_implied_cluster_alloc()
4217          * work correctly.
4218          */
4219         offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4220         ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4221         ar.goal -= offset;
4222         ar.logical -= offset;
4223         if (S_ISREG(inode->i_mode))
4224                 ar.flags = EXT4_MB_HINT_DATA;
4225         else
4226                 /* disable in-core preallocation for non-regular files */
4227                 ar.flags = 0;
4228         if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4229                 ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4230         newblock = ext4_mb_new_blocks(handle, &ar, &err);
4231         if (!newblock)
4232                 goto out2;
4233         ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4234                   ar.goal, newblock, allocated);
4235         free_on_err = 1;
4236         allocated_clusters = ar.len;
4237         ar.len = EXT4_C2B(sbi, ar.len) - offset;
4238         if (ar.len > allocated)
4239                 ar.len = allocated;
4240
4241 got_allocated_blocks:
4242         /* try to insert new extent into found leaf and return */
4243         ext4_ext_store_pblock(&newex, newblock + offset);
4244         newex.ee_len = cpu_to_le16(ar.len);
4245         /* Mark uninitialized */
4246         if (flags & EXT4_GET_BLOCKS_UNINIT_EXT){
4247                 ext4_ext_mark_uninitialized(&newex);
4248                 map->m_flags |= EXT4_MAP_UNWRITTEN;
4249                 /*
4250                  * io_end structure was created for every IO write to an
4251                  * uninitialized extent. To avoid unnecessary conversion,
4252                  * here we flag the IO that really needs the conversion.
4253                  * For non asycn direct IO case, flag the inode state
4254                  * that we need to perform conversion when IO is done.
4255                  */
4256                 if ((flags & EXT4_GET_BLOCKS_PRE_IO))
4257                         set_unwritten = 1;
4258                 if (ext4_should_dioread_nolock(inode))
4259                         map->m_flags |= EXT4_MAP_UNINIT;
4260         }
4261
4262         err = 0;
4263         if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4264                 err = check_eofblocks_fl(handle, inode, map->m_lblk,
4265                                          path, ar.len);
4266         if (!err)
4267                 err = ext4_ext_insert_extent(handle, inode, path,
4268                                              &newex, flags);
4269
4270         if (!err && set_unwritten) {
4271                 if (io)
4272                         ext4_set_io_unwritten_flag(inode, io);
4273                 else
4274                         ext4_set_inode_state(inode,
4275                                              EXT4_STATE_DIO_UNWRITTEN);
4276         }
4277
4278         if (err && free_on_err) {
4279                 int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4280                         EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4281                 /* free data blocks we just allocated */
4282                 /* not a good idea to call discard here directly,
4283                  * but otherwise we'd need to call it every free() */
4284                 ext4_discard_preallocations(inode);
4285                 ext4_free_blocks(handle, inode, NULL, ext4_ext_pblock(&newex),
4286                                  ext4_ext_get_actual_len(&newex), fb_flags);
4287                 goto out2;
4288         }
4289
4290         /* previous routine could use block we allocated */
4291         newblock = ext4_ext_pblock(&newex);
4292         allocated = ext4_ext_get_actual_len(&newex);
4293         if (allocated > map->m_len)
4294                 allocated = map->m_len;
4295         map->m_flags |= EXT4_MAP_NEW;
4296
4297         /*
4298          * Update reserved blocks/metadata blocks after successful
4299          * block allocation which had been deferred till now.
4300          */
4301         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4302                 unsigned int reserved_clusters;
4303                 /*
4304                  * Check how many clusters we had reserved this allocated range
4305                  */
4306                 reserved_clusters = get_reserved_cluster_alloc(inode,
4307                                                 map->m_lblk, allocated);
4308                 if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4309                         if (reserved_clusters) {
4310                                 /*
4311                                  * We have clusters reserved for this range.
4312                                  * But since we are not doing actual allocation
4313                                  * and are simply using blocks from previously
4314                                  * allocated cluster, we should release the
4315                                  * reservation and not claim quota.
4316                                  */
4317                                 ext4_da_update_reserve_space(inode,
4318                                                 reserved_clusters, 0);
4319                         }
4320                 } else {
4321                         BUG_ON(allocated_clusters < reserved_clusters);
4322                         if (reserved_clusters < allocated_clusters) {
4323                                 struct ext4_inode_info *ei = EXT4_I(inode);
4324                                 int reservation = allocated_clusters -
4325                                                   reserved_clusters;
4326                                 /*
4327                                  * It seems we claimed few clusters outside of
4328                                  * the range of this allocation. We should give
4329                                  * it back to the reservation pool. This can
4330                                  * happen in the following case:
4331                                  *
4332                                  * * Suppose s_cluster_ratio is 4 (i.e., each
4333                                  *   cluster has 4 blocks. Thus, the clusters
4334                                  *   are [0-3],[4-7],[8-11]...
4335                                  * * First comes delayed allocation write for
4336                                  *   logical blocks 10 & 11. Since there were no
4337                                  *   previous delayed allocated blocks in the
4338                                  *   range [8-11], we would reserve 1 cluster
4339                                  *   for this write.
4340                                  * * Next comes write for logical blocks 3 to 8.
4341                                  *   In this case, we will reserve 2 clusters
4342                                  *   (for [0-3] and [4-7]; and not for [8-11] as
4343                                  *   that range has a delayed allocated blocks.
4344                                  *   Thus total reserved clusters now becomes 3.
4345                                  * * Now, during the delayed allocation writeout
4346                                  *   time, we will first write blocks [3-8] and
4347                                  *   allocate 3 clusters for writing these
4348                                  *   blocks. Also, we would claim all these
4349                                  *   three clusters above.
4350                                  * * Now when we come here to writeout the
4351                                  *   blocks [10-11], we would expect to claim
4352                                  *   the reservation of 1 cluster we had made
4353                                  *   (and we would claim it since there are no
4354                                  *   more delayed allocated blocks in the range
4355                                  *   [8-11]. But our reserved cluster count had
4356                                  *   already gone to 0.
4357                                  *
4358                                  *   Thus, at the step 4 above when we determine
4359                                  *   that there are still some unwritten delayed
4360                                  *   allocated blocks outside of our current
4361                                  *   block range, we should increment the
4362                                  *   reserved clusters count so that when the
4363                                  *   remaining blocks finally gets written, we
4364                                  *   could claim them.
4365                                  */
4366                                 dquot_reserve_block(inode,
4367                                                 EXT4_C2B(sbi, reservation));
4368                                 spin_lock(&ei->i_block_reservation_lock);
4369                                 ei->i_reserved_data_blocks += reservation;
4370                                 spin_unlock(&ei->i_block_reservation_lock);
4371                         }
4372                         /*
4373                          * We will claim quota for all newly allocated blocks.
4374                          * We're updating the reserved space *after* the
4375                          * correction above so we do not accidentally free
4376                          * all the metadata reservation because we might
4377                          * actually need it later on.
4378                          */
4379                         ext4_da_update_reserve_space(inode, allocated_clusters,
4380                                                         1);
4381                 }
4382         }
4383
4384         /*
4385          * Cache the extent and update transaction to commit on fdatasync only
4386          * when it is _not_ an uninitialized extent.
4387          */
4388         if ((flags & EXT4_GET_BLOCKS_UNINIT_EXT) == 0)
4389                 ext4_update_inode_fsync_trans(handle, inode, 1);
4390         else
4391                 ext4_update_inode_fsync_trans(handle, inode, 0);
4392 out:
4393         if (allocated > map->m_len)
4394                 allocated = map->m_len;
4395         ext4_ext_show_leaf(inode, path);
4396         map->m_flags |= EXT4_MAP_MAPPED;
4397         map->m_pblk = newblock;
4398         map->m_len = allocated;
4399 out2:
4400         if (path) {
4401                 ext4_ext_drop_refs(path);
4402                 kfree(path);
4403         }
4404
4405 out3:
4406         trace_ext4_ext_map_blocks_exit(inode, map, err ? err : allocated);
4407
4408         return err ? err : allocated;
4409 }
4410
4411 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4412 {
4413         struct super_block *sb = inode->i_sb;
4414         ext4_lblk_t last_block;
4415         int err = 0;
4416
4417         /*
4418          * TODO: optimization is possible here.
4419          * Probably we need not scan at all,
4420          * because page truncation is enough.
4421          */
4422
4423         /* we have to know where to truncate from in crash case */
4424         EXT4_I(inode)->i_disksize = inode->i_size;
4425         ext4_mark_inode_dirty(handle, inode);
4426
4427         last_block = (inode->i_size + sb->s_blocksize - 1)
4428                         >> EXT4_BLOCK_SIZE_BITS(sb);
4429 retry:
4430         err = ext4_es_remove_extent(inode, last_block,
4431                                     EXT_MAX_BLOCKS - last_block);
4432         if (err == -ENOMEM) {
4433                 cond_resched();
4434                 congestion_wait(BLK_RW_ASYNC, HZ/50);
4435                 goto retry;
4436         }
4437         if (err) {
4438                 ext4_std_error(inode->i_sb, err);
4439                 return;
4440         }
4441         err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4442         ext4_std_error(inode->i_sb, err);
4443 }
4444
4445 static void ext4_falloc_update_inode(struct inode *inode,
4446                                 int mode, loff_t new_size, int update_ctime)
4447 {
4448         struct timespec now;
4449
4450         if (update_ctime) {
4451                 now = current_fs_time(inode->i_sb);
4452                 if (!timespec_equal(&inode->i_ctime, &now))
4453                         inode->i_ctime = now;
4454         }
4455         /*
4456          * Update only when preallocation was requested beyond
4457          * the file size.
4458          */
4459         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
4460                 if (new_size > i_size_read(inode))
4461                         i_size_write(inode, new_size);
4462                 if (new_size > EXT4_I(inode)->i_disksize)
4463                         ext4_update_i_disksize(inode, new_size);
4464         } else {
4465                 /*
4466                  * Mark that we allocate beyond EOF so the subsequent truncate
4467                  * can proceed even if the new size is the same as i_size.
4468                  */
4469                 if (new_size > i_size_read(inode))
4470                         ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4471         }
4472
4473 }
4474
4475 /*
4476  * preallocate space for a file. This implements ext4's fallocate file
4477  * operation, which gets called from sys_fallocate system call.
4478  * For block-mapped files, posix_fallocate should fall back to the method
4479  * of writing zeroes to the required new blocks (the same behavior which is
4480  * expected for file systems which do not support fallocate() system call).
4481  */
4482 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4483 {
4484         struct inode *inode = file_inode(file);
4485         handle_t *handle;
4486         loff_t new_size;
4487         unsigned int max_blocks;
4488         int ret = 0;
4489         int ret2 = 0;
4490         int retries = 0;
4491         int flags;
4492         struct ext4_map_blocks map;
4493         unsigned int credits, blkbits = inode->i_blkbits;
4494
4495         /* Return error if mode is not supported */
4496         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
4497                 return -EOPNOTSUPP;
4498
4499         if (mode & FALLOC_FL_PUNCH_HOLE)
4500                 return ext4_punch_hole(file, offset, len);
4501
4502         ret = ext4_convert_inline_data(inode);
4503         if (ret)
4504                 return ret;
4505
4506         /*
4507          * currently supporting (pre)allocate mode for extent-based
4508          * files _only_
4509          */
4510         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4511                 return -EOPNOTSUPP;
4512
4513         trace_ext4_fallocate_enter(inode, offset, len, mode);
4514         map.m_lblk = offset >> blkbits;
4515         /*
4516          * We can't just convert len to max_blocks because
4517          * If blocksize = 4096 offset = 3072 and len = 2048
4518          */
4519         max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4520                 - map.m_lblk;
4521         /*
4522          * credits to insert 1 extent into extent tree
4523          */
4524         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4525         mutex_lock(&inode->i_mutex);
4526         ret = inode_newsize_ok(inode, (len + offset));
4527         if (ret) {
4528                 mutex_unlock(&inode->i_mutex);
4529                 trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4530                 return ret;
4531         }
4532         flags = EXT4_GET_BLOCKS_CREATE_UNINIT_EXT;
4533         if (mode & FALLOC_FL_KEEP_SIZE)
4534                 flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4535         /*
4536          * Don't normalize the request if it can fit in one extent so
4537          * that it doesn't get unnecessarily split into multiple
4538          * extents.
4539          */
4540         if (len <= EXT_UNINIT_MAX_LEN << blkbits)
4541                 flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4542
4543 retry:
4544         while (ret >= 0 && ret < max_blocks) {
4545                 map.m_lblk = map.m_lblk + ret;
4546                 map.m_len = max_blocks = max_blocks - ret;
4547                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4548                                             credits);
4549                 if (IS_ERR(handle)) {
4550                         ret = PTR_ERR(handle);
4551                         break;
4552                 }
4553                 ret = ext4_map_blocks(handle, inode, &map, flags);
4554                 if (ret <= 0) {
4555 #ifdef EXT4FS_DEBUG
4556                         ext4_warning(inode->i_sb,
4557                                      "inode #%lu: block %u: len %u: "
4558                                      "ext4_ext_map_blocks returned %d",
4559                                      inode->i_ino, map.m_lblk,
4560                                      map.m_len, ret);
4561 #endif
4562                         ext4_mark_inode_dirty(handle, inode);
4563                         ret2 = ext4_journal_stop(handle);
4564                         break;
4565                 }
4566                 if ((map.m_lblk + ret) >= (EXT4_BLOCK_ALIGN(offset + len,
4567                                                 blkbits) >> blkbits))
4568                         new_size = offset + len;
4569                 else
4570                         new_size = ((loff_t) map.m_lblk + ret) << blkbits;
4571
4572                 ext4_falloc_update_inode(inode, mode, new_size,
4573                                          (map.m_flags & EXT4_MAP_NEW));
4574                 ext4_mark_inode_dirty(handle, inode);
4575                 if ((file->f_flags & O_SYNC) && ret >= max_blocks)
4576                         ext4_handle_sync(handle);
4577                 ret2 = ext4_journal_stop(handle);
4578                 if (ret2)
4579                         break;
4580         }
4581         if (ret == -ENOSPC &&
4582                         ext4_should_retry_alloc(inode->i_sb, &retries)) {
4583                 ret = 0;
4584                 goto retry;
4585         }
4586         mutex_unlock(&inode->i_mutex);
4587         trace_ext4_fallocate_exit(inode, offset, max_blocks,
4588                                 ret > 0 ? ret2 : ret);
4589         return ret > 0 ? ret2 : ret;
4590 }
4591
4592 /*
4593  * This function convert a range of blocks to written extents
4594  * The caller of this function will pass the start offset and the size.
4595  * all unwritten extents within this range will be converted to
4596  * written extents.
4597  *
4598  * This function is called from the direct IO end io call back
4599  * function, to convert the fallocated extents after IO is completed.
4600  * Returns 0 on success.
4601  */
4602 int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
4603                                     ssize_t len)
4604 {
4605         handle_t *handle;
4606         unsigned int max_blocks;
4607         int ret = 0;
4608         int ret2 = 0;
4609         struct ext4_map_blocks map;
4610         unsigned int credits, blkbits = inode->i_blkbits;
4611
4612         map.m_lblk = offset >> blkbits;
4613         /*
4614          * We can't just convert len to max_blocks because
4615          * If blocksize = 4096 offset = 3072 and len = 2048
4616          */
4617         max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
4618                       map.m_lblk);
4619         /*
4620          * credits to insert 1 extent into extent tree
4621          */
4622         credits = ext4_chunk_trans_blocks(inode, max_blocks);
4623         while (ret >= 0 && ret < max_blocks) {
4624                 map.m_lblk += ret;
4625                 map.m_len = (max_blocks -= ret);
4626                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
4627                 if (IS_ERR(handle)) {
4628                         ret = PTR_ERR(handle);
4629                         break;
4630                 }
4631                 ret = ext4_map_blocks(handle, inode, &map,
4632                                       EXT4_GET_BLOCKS_IO_CONVERT_EXT);
4633                 if (ret <= 0)
4634                         ext4_warning(inode->i_sb,
4635                                      "inode #%lu: block %u: len %u: "
4636                                      "ext4_ext_map_blocks returned %d",
4637                                      inode->i_ino, map.m_lblk,
4638                                      map.m_len, ret);
4639                 ext4_mark_inode_dirty(handle, inode);
4640                 ret2 = ext4_journal_stop(handle);
4641                 if (ret <= 0 || ret2 )
4642                         break;
4643         }
4644         return ret > 0 ? ret2 : ret;
4645 }
4646
4647 /*
4648  * If newes is not existing extent (newes->ec_pblk equals zero) find
4649  * delayed extent at start of newes and update newes accordingly and
4650  * return start of the next delayed extent.
4651  *
4652  * If newes is existing extent (newes->ec_pblk is not equal zero)
4653  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
4654  * extent found. Leave newes unmodified.
4655  */
4656 static int ext4_find_delayed_extent(struct inode *inode,
4657                                     struct extent_status *newes)
4658 {
4659         struct extent_status es;
4660         ext4_lblk_t block, next_del;
4661
4662         if (newes->es_pblk == 0) {
4663                 ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
4664                                 newes->es_lblk + newes->es_len - 1, &es);
4665
4666                 /*
4667                  * No extent in extent-tree contains block @newes->es_pblk,
4668                  * then the block may stay in 1)a hole or 2)delayed-extent.
4669                  */
4670                 if (es.es_len == 0)
4671                         /* A hole found. */
4672                         return 0;
4673
4674                 if (es.es_lblk > newes->es_lblk) {
4675                         /* A hole found. */
4676                         newes->es_len = min(es.es_lblk - newes->es_lblk,
4677                                             newes->es_len);
4678                         return 0;
4679                 }
4680
4681                 newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
4682         }
4683
4684         block = newes->es_lblk + newes->es_len;
4685         ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
4686         if (es.es_len == 0)
4687                 next_del = EXT_MAX_BLOCKS;
4688         else
4689                 next_del = es.es_lblk;
4690
4691         return next_del;
4692 }
4693 /* fiemap flags we can handle specified here */
4694 #define EXT4_FIEMAP_FLAGS       (FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
4695
4696 static int ext4_xattr_fiemap(struct inode *inode,
4697                                 struct fiemap_extent_info *fieinfo)
4698 {
4699         __u64 physical = 0;
4700         __u64 length;
4701         __u32 flags = FIEMAP_EXTENT_LAST;
4702         int blockbits = inode->i_sb->s_blocksize_bits;
4703         int error = 0;
4704
4705         /* in-inode? */
4706         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
4707                 struct ext4_iloc iloc;
4708                 int offset;     /* offset of xattr in inode */
4709
4710                 error = ext4_get_inode_loc(inode, &iloc);
4711                 if (error)
4712                         return error;
4713                 physical = (__u64)iloc.bh->b_blocknr << blockbits;
4714                 offset = EXT4_GOOD_OLD_INODE_SIZE +
4715                                 EXT4_I(inode)->i_extra_isize;
4716                 physical += offset;
4717                 length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
4718                 flags |= FIEMAP_EXTENT_DATA_INLINE;
4719                 brelse(iloc.bh);
4720         } else { /* external block */
4721                 physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
4722                 length = inode->i_sb->s_blocksize;
4723         }
4724
4725         if (physical)
4726                 error = fiemap_fill_next_extent(fieinfo, 0, physical,
4727                                                 length, flags);
4728         return (error < 0 ? error : 0);
4729 }
4730
4731 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4732                 __u64 start, __u64 len)
4733 {
4734         ext4_lblk_t start_blk;
4735         int error = 0;
4736
4737         if (ext4_has_inline_data(inode)) {
4738                 int has_inline = 1;
4739
4740                 error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
4741
4742                 if (has_inline)
4743                         return error;
4744         }
4745
4746         /* fallback to generic here if not in extents fmt */
4747         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4748                 return generic_block_fiemap(inode, fieinfo, start, len,
4749                         ext4_get_block);
4750
4751         if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
4752                 return -EBADR;
4753
4754         if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
4755                 error = ext4_xattr_fiemap(inode, fieinfo);
4756         } else {
4757                 ext4_lblk_t len_blks;
4758                 __u64 last_blk;
4759
4760                 start_blk = start >> inode->i_sb->s_blocksize_bits;
4761                 last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
4762                 if (last_blk >= EXT_MAX_BLOCKS)
4763                         last_blk = EXT_MAX_BLOCKS-1;
4764                 len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
4765
4766                 /*
4767                  * Walk the extent tree gathering extent information
4768                  * and pushing extents back to the user.
4769                  */
4770                 error = ext4_fill_fiemap_extents(inode, start_blk,
4771                                                  len_blks, fieinfo);
4772         }
4773
4774         return error;
4775 }