fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31
  32 #define MLOG_MASK_PREFIX ML_FILE_IO
  33 #include <cluster/masklog.h>
  34
  35 #include "ocfs2.h"
  36
  37 #include "alloc.h"
  38 #include "aops.h"
  39 #include "dlmglue.h"
  40 #include "extent_map.h"
  41 #include "file.h"
  42 #include "inode.h"
  43 #include "journal.h"
  44 #include "suballoc.h"
  45 #include "super.h"
  46 #include "symlink.h"
  47 #include "refcounttree.h"
  48
  49 #include "buffer_head_io.h"
  50
  51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  52                                    struct buffer_head *bh_result, int create)
  53 {
  54         int err = -EIO;
  55         int status;
  56         struct ocfs2_dinode *fe = NULL;
  57         struct buffer_head *bh = NULL;
  58         struct buffer_head *buffer_cache_bh = NULL;
  59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  60         void *kaddr;
  61
  62         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  63                    (unsigned long long)iblock, bh_result, create);
  64
  65         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  66
  67         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  68                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  69                      (unsigned long long)iblock);
  70                 goto bail;
  71         }
  72
  73         status = ocfs2_read_inode_block(inode, &bh);
  74         if (status < 0) {
  75                 mlog_errno(status);
  76                 goto bail;
  77         }
  78         fe = (struct ocfs2_dinode *) bh->b_data;
  79
  80         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  81                                                     le32_to_cpu(fe->i_clusters))) {
  82                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  83                      "%llu\n", (unsigned long long)iblock);
  84                 goto bail;
  85         }
  86
  87         /* We don't use the page cache to create symlink data, so if
  88          * need be, copy it over from the buffer cache. */
  89         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  90                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  91                             iblock;
  92                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  93                 if (!buffer_cache_bh) {
  94                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  95                         goto bail;
  96                 }
  97
  98                 /* we haven't locked out transactions, so a commit
  99                  * could've happened. Since we've got a reference on
 100                  * the bh, even if it commits while we're doing the
 101                  * copy, the data is still good. */
 102                 if (buffer_jbd(buffer_cache_bh)
 103                     && ocfs2_inode_is_new(inode)) {
 104                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 105                         if (!kaddr) {
 106                                 mlog(ML_ERROR, "couldn't kmap!\n");
 107                                 goto bail;
 108                         }
 109                         memcpy(kaddr + (bh_result->b_size * iblock),
 110                                buffer_cache_bh->b_data,
 111                                bh_result->b_size);
 112                         kunmap_atomic(kaddr, KM_USER0);
 113                         set_buffer_uptodate(bh_result);
 114                 }
 115                 brelse(buffer_cache_bh);
 116         }
 117
 118         map_bh(bh_result, inode->i_sb,
 119                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 120
 121         err = 0;
 122
 123 bail:
 124         brelse(bh);
 125
 126         mlog_exit(err);
 127         return err;
 128 }
 129
 130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 131                     struct buffer_head *bh_result, int create)
 132 {
 133         int err = 0;
 134         unsigned int ext_flags;
 135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 136         u64 p_blkno, count, past_eof;
 137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 138
 139         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 140                    (unsigned long long)iblock, bh_result, create);
 141
 142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 144                      inode, inode->i_ino);
 145
 146         if (S_ISLNK(inode->i_mode)) {
 147                 /* this always does I/O for some reason. */
 148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 149                 goto bail;
 150         }
 151
 152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 153                                           &ext_flags);
 154         if (err) {
 155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 157                      (unsigned long long)p_blkno);
 158                 goto bail;
 159         }
 160
 161         if (max_blocks < count)
 162                 count = max_blocks;
 163
 164         /*
 165          * ocfs2 never allocates in this function - the only time we
 166          * need to use BH_New is when we're extending i_size on a file
 167          * system which doesn't support holes, in which case BH_New
 168          * allows block_prepare_write() to zero.
 169          *
 170          * If we see this on a sparse file system, then a truncate has
 171          * raced us and removed the cluster. In this case, we clear
 172          * the buffers dirty and uptodate bits and let the buffer code
 173          * ignore it as a hole.
 174          */
 175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 176                 clear_buffer_dirty(bh_result);
 177                 clear_buffer_uptodate(bh_result);
 178                 goto bail;
 179         }
 180
 181         /* Treat the unwritten extent as a hole for zeroing purposes. */
 182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 183                 map_bh(bh_result, inode->i_sb, p_blkno);
 184
 185         bh_result->b_size = count << inode->i_blkbits;
 186
 187         if (!ocfs2_sparse_alloc(osb)) {
 188                 if (p_blkno == 0) {
 189                         err = -EIO;
 190                         mlog(ML_ERROR,
 191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 192                              (unsigned long long)iblock,
 193                              (unsigned long long)p_blkno,
 194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 196                         dump_stack();
 197                         goto bail;
 198                 }
 199
 200                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 201                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 202                      (unsigned long long)past_eof);
 203
 204                 if (create && (iblock >= past_eof))
 205                         set_buffer_new(bh_result);
 206         }
 207
 208 bail:
 209         if (err < 0)
 210                 err = -EIO;
 211
 212         mlog_exit(err);
 213         return err;
 214 }
 215
 216 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 217                            struct buffer_head *di_bh)
 218 {
 219         void *kaddr;
 220         loff_t size;
 221         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 222
 223         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 224                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 225                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 226                 return -EROFS;
 227         }
 228
 229         size = i_size_read(inode);
 230
 231         if (size > PAGE_CACHE_SIZE ||
 232             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 233                 ocfs2_error(inode->i_sb,
 234                             "Inode %llu has with inline data has bad size: %Lu",
 235                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 236                             (unsigned long long)size);
 237                 return -EROFS;
 238         }
 239
 240         kaddr = kmap_atomic(page, KM_USER0);
 241         if (size)
 242                 memcpy(kaddr, di->id2.i_data.id_data, size);
 243         /* Clear the remaining part of the page */
 244         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 245         flush_dcache_page(page);
 246         kunmap_atomic(kaddr, KM_USER0);
 247
 248         SetPageUptodate(page);
 249
 250         return 0;
 251 }
 252
 253 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 254 {
 255         int ret;
 256         struct buffer_head *di_bh = NULL;
 257
 258         BUG_ON(!PageLocked(page));
 259         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 260
 261         ret = ocfs2_read_inode_block(inode, &di_bh);
 262         if (ret) {
 263                 mlog_errno(ret);
 264                 goto out;
 265         }
 266
 267         ret = ocfs2_read_inline_data(inode, page, di_bh);
 268 out:
 269         unlock_page(page);
 270
 271         brelse(di_bh);
 272         return ret;
 273 }
 274
 275 static int ocfs2_readpage(struct file *file, struct page *page)
 276 {
 277         struct inode *inode = page->mapping->host;
 278         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 279         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 280         int ret, unlock = 1;
 281
 282         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 283
 284         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 285         if (ret != 0) {
 286                 if (ret == AOP_TRUNCATED_PAGE)
 287                         unlock = 0;
 288                 mlog_errno(ret);
 289                 goto out;
 290         }
 291
 292         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 293                 ret = AOP_TRUNCATED_PAGE;
 294                 goto out_inode_unlock;
 295         }
 296
 297         /*
 298          * i_size might have just been updated as we grabed the meta lock.  We
 299          * might now be discovering a truncate that hit on another node.
 300          * block_read_full_page->get_block freaks out if it is asked to read
 301          * beyond the end of a file, so we check here.  Callers
 302          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 303          * and notice that the page they just read isn't needed.
 304          *
 305          * XXX sys_readahead() seems to get that wrong?
 306          */
 307         if (start >= i_size_read(inode)) {
 308                 zero_user(page, 0, PAGE_SIZE);
 309                 SetPageUptodate(page);
 310                 ret = 0;
 311                 goto out_alloc;
 312         }
 313
 314         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 315                 ret = ocfs2_readpage_inline(inode, page);
 316         else
 317                 ret = block_read_full_page(page, ocfs2_get_block);
 318         unlock = 0;
 319
 320 out_alloc:
 321         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 322 out_inode_unlock:
 323         ocfs2_inode_unlock(inode, 0);
 324 out:
 325         if (unlock)
 326                 unlock_page(page);
 327         mlog_exit(ret);
 328         return ret;
 329 }
 330
 331 /*
 332  * This is used only for read-ahead. Failures or difficult to handle
 333  * situations are safe to ignore.
 334  *
 335  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 336  * are quite large (243 extents on 4k blocks), so most inodes don't
 337  * grow out to a tree. If need be, detecting boundary extents could
 338  * trivially be added in a future version of ocfs2_get_block().
 339  */
 340 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 341                            struct list_head *pages, unsigned nr_pages)
 342 {
 343         int ret, err = -EIO;
 344         struct inode *inode = mapping->host;
 345         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 346         loff_t start;
 347         struct page *last;
 348
 349         /*
 350          * Use the nonblocking flag for the dlm code to avoid page
 351          * lock inversion, but don't bother with retrying.
 352          */
 353         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 354         if (ret)
 355                 return err;
 356
 357         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 358                 ocfs2_inode_unlock(inode, 0);
 359                 return err;
 360         }
 361
 362         /*
 363          * Don't bother with inline-data. There isn't anything
 364          * to read-ahead in that case anyway...
 365          */
 366         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 367                 goto out_unlock;
 368
 369         /*
 370          * Check whether a remote node truncated this file - we just
 371          * drop out in that case as it's not worth handling here.
 372          */
 373         last = list_entry(pages->prev, struct page, lru);
 374         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 375         if (start >= i_size_read(inode))
 376                 goto out_unlock;
 377
 378         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 379
 380 out_unlock:
 381         up_read(&oi->ip_alloc_sem);
 382         ocfs2_inode_unlock(inode, 0);
 383
 384         return err;
 385 }
 386
 387 /* Note: Because we don't support holes, our allocation has
 388  * already happened (allocation writes zeros to the file data)
 389  * so we don't have to worry about ordered writes in
 390  * ocfs2_writepage.
 391  *
 392  * ->writepage is called during the process of invalidating the page cache
 393  * during blocked lock processing.  It can't block on any cluster locks
 394  * to during block mapping.  It's relying on the fact that the block
 395  * mapping can't have disappeared under the dirty pages that it is
 396  * being asked to write back.
 397  */
 398 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 399 {
 400         int ret;
 401
 402         mlog_entry("(0x%p)\n", page);
 403
 404         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 405
 406         mlog_exit(ret);
 407
 408         return ret;
 409 }
 410
 411 /*
 412  * This is called from ocfs2_write_zero_page() which has handled it's
 413  * own cluster locking and has ensured allocation exists for those
 414  * blocks to be written.
 415  */
 416 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 417                                unsigned from, unsigned to)
 418 {
 419         int ret;
 420
 421         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 422
 423         return ret;
 424 }
 425
 426 /* Taken from ext3. We don't necessarily need the full blown
 427  * functionality yet, but IMHO it's better to cut and paste the whole
 428  * thing so we can avoid introducing our own bugs (and easily pick up
 429  * their fixes when they happen) --Mark */
 430 int walk_page_buffers(  handle_t *handle,
 431                         struct buffer_head *head,
 432                         unsigned from,
 433                         unsigned to,
 434                         int *partial,
 435                         int (*fn)(      handle_t *handle,
 436                                         struct buffer_head *bh))
 437 {
 438         struct buffer_head *bh;
 439         unsigned block_start, block_end;
 440         unsigned blocksize = head->b_size;
 441         int err, ret = 0;
 442         struct buffer_head *next;
 443
 444         for (   bh = head, block_start = 0;
 445                 ret == 0 && (bh != head || !block_start);
 446                 block_start = block_end, bh = next)
 447         {
 448                 next = bh->b_this_page;
 449                 block_end = block_start + blocksize;
 450                 if (block_end <= from || block_start >= to) {
 451                         if (partial && !buffer_uptodate(bh))
 452                                 *partial = 1;
 453                         continue;
 454                 }
 455                 err = (*fn)(handle, bh);
 456                 if (!ret)
 457                         ret = err;
 458         }
 459         return ret;
 460 }
 461
 462 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
 463                                                          struct page *page,
 464                                                          unsigned from,
 465                                                          unsigned to)
 466 {
 467         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 468         handle_t *handle;
 469         int ret = 0;
 470
 471         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 472         if (IS_ERR(handle)) {
 473                 ret = -ENOMEM;
 474                 mlog_errno(ret);
 475                 goto out;
 476         }
 477
 478         if (ocfs2_should_order_data(inode)) {
 479                 ret = ocfs2_jbd2_file_inode(handle, inode);
 480                 if (ret < 0)
 481                         mlog_errno(ret);
 482         }
 483 out:
 484         if (ret) {
 485                 if (!IS_ERR(handle))
 486                         ocfs2_commit_trans(osb, handle);
 487                 handle = ERR_PTR(ret);
 488         }
 489         return handle;
 490 }
 491
 492 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 493 {
 494         sector_t status;
 495         u64 p_blkno = 0;
 496         int err = 0;
 497         struct inode *inode = mapping->host;
 498
 499         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 500
 501         /* We don't need to lock journal system files, since they aren't
 502          * accessed concurrently from multiple nodes.
 503          */
 504         if (!INODE_JOURNAL(inode)) {
 505                 err = ocfs2_inode_lock(inode, NULL, 0);
 506                 if (err) {
 507                         if (err != -ENOENT)
 508                                 mlog_errno(err);
 509                         goto bail;
 510                 }
 511                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 512         }
 513
 514         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 515                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 516                                                   NULL);
 517
 518         if (!INODE_JOURNAL(inode)) {
 519                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 520                 ocfs2_inode_unlock(inode, 0);
 521         }
 522
 523         if (err) {
 524                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 525                      (unsigned long long)block);
 526                 mlog_errno(err);
 527                 goto bail;
 528         }
 529
 530 bail:
 531         status = err ? 0 : p_blkno;
 532
 533         mlog_exit((int)status);
 534
 535         return status;
 536 }
 537
 538 /*
 539  * TODO: Make this into a generic get_blocks function.
 540  *
 541  * From do_direct_io in direct-io.c:
 542  *  "So what we do is to permit the ->get_blocks function to populate
 543  *   bh.b_size with the size of IO which is permitted at this offset and
 544  *   this i_blkbits."
 545  *
 546  * This function is called directly from get_more_blocks in direct-io.c.
 547  *
 548  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 549  *                                      fs_count, map_bh, dio->rw == WRITE);
 550  */
 551 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 552                                      struct buffer_head *bh_result, int create)
 553 {
 554         int ret;
 555         u64 p_blkno, inode_blocks, contig_blocks;
 556         unsigned int ext_flags;
 557         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 558         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 559
 560         /* This function won't even be called if the request isn't all
 561          * nicely aligned and of the right size, so there's no need
 562          * for us to check any of that. */
 563
 564         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 565
 566         /*
 567          * Any write past EOF is not allowed because we'd be extending.
 568          */
 569         if (create && (iblock + max_blocks) > inode_blocks) {
 570                 ret = -EIO;
 571                 goto bail;
 572         }
 573
 574         /* This figures out the size of the next contiguous block, and
 575          * our logical offset */
 576         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 577                                           &contig_blocks, &ext_flags);
 578         if (ret) {
 579                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 580                      (unsigned long long)iblock);
 581                 ret = -EIO;
 582                 goto bail;
 583         }
 584
 585         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno && create) {
 586                 ocfs2_error(inode->i_sb,
 587                             "Inode %llu has a hole at block %llu\n",
 588                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 589                             (unsigned long long)iblock);
 590                 ret = -EROFS;
 591                 goto bail;
 592         }
 593
 594         /* We should already CoW the refcounted extent in case of create. */
 595         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 596
 597         /*
 598          * get_more_blocks() expects us to describe a hole by clearing
 599          * the mapped bit on bh_result().
 600          *
 601          * Consider an unwritten extent as a hole.
 602          */
 603         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 604                 map_bh(bh_result, inode->i_sb, p_blkno);
 605         else {
 606                 /*
 607                  * ocfs2_prepare_inode_for_write() should have caught
 608                  * the case where we'd be filling a hole and triggered
 609                  * a buffered write instead.
 610                  */
 611                 if (create) {
 612                         ret = -EIO;
 613                         mlog_errno(ret);
 614                         goto bail;
 615                 }
 616
 617                 clear_buffer_mapped(bh_result);
 618         }
 619
 620         /* make sure we don't map more than max_blocks blocks here as
 621            that's all the kernel will handle at this point. */
 622         if (max_blocks < contig_blocks)
 623                 contig_blocks = max_blocks;
 624         bh_result->b_size = contig_blocks << blocksize_bits;
 625 bail:
 626         return ret;
 627 }
 628
 629 /*
 630  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 631  * particularly interested in the aio/dio case.  Like the core uses
 632  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 633  * truncation on another.
 634  */
 635 static void ocfs2_dio_end_io(struct kiocb *iocb,
 636                              loff_t offset,
 637                              ssize_t bytes,
 638                              void *private)
 639 {
 640         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 641         int level;
 642
 643         /* this io's submitter should not have unlocked this before we could */
 644         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 645
 646         ocfs2_iocb_clear_rw_locked(iocb);
 647
 648         level = ocfs2_iocb_rw_locked_level(iocb);
 649         if (!level)
 650                 up_read(&inode->i_alloc_sem);
 651         ocfs2_rw_unlock(inode, level);
 652 }
 653
 654 /*
 655  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 656  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 657  * do journalled data.
 658  */
 659 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 660 {
 661         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 662
 663         jbd2_journal_invalidatepage(journal, page, offset);
 664 }
 665
 666 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 667 {
 668         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 669
 670         if (!page_has_buffers(page))
 671                 return 0;
 672         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 673 }
 674
 675 static ssize_t ocfs2_direct_IO(int rw,
 676                                struct kiocb *iocb,
 677                                const struct iovec *iov,
 678                                loff_t offset,
 679                                unsigned long nr_segs)
 680 {
 681         struct file *file = iocb->ki_filp;
 682         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 683         int ret;
 684
 685         mlog_entry_void();
 686
 687         /*
 688          * Fallback to buffered I/O if we see an inode without
 689          * extents.
 690          */
 691         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 692                 return 0;
 693
 694         /* Fallback to buffered I/O if we are appending. */
 695         if (i_size_read(inode) <= offset)
 696                 return 0;
 697
 698         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 699                                             inode->i_sb->s_bdev, iov, offset,
 700                                             nr_segs,
 701                                             ocfs2_direct_IO_get_blocks,
 702                                             ocfs2_dio_end_io);
 703
 704         mlog_exit(ret);
 705         return ret;
 706 }
 707
 708 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 709                                             u32 cpos,
 710                                             unsigned int *start,
 711                                             unsigned int *end)
 712 {
 713         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 714
 715         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 716                 unsigned int cpp;
 717
 718                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 719
 720                 cluster_start = cpos % cpp;
 721                 cluster_start = cluster_start << osb->s_clustersize_bits;
 722
 723                 cluster_end = cluster_start + osb->s_clustersize;
 724         }
 725
 726         BUG_ON(cluster_start > PAGE_SIZE);
 727         BUG_ON(cluster_end > PAGE_SIZE);
 728
 729         if (start)
 730                 *start = cluster_start;
 731         if (end)
 732                 *end = cluster_end;
 733 }
 734
 735 /*
 736  * 'from' and 'to' are the region in the page to avoid zeroing.
 737  *
 738  * If pagesize > clustersize, this function will avoid zeroing outside
 739  * of the cluster boundary.
 740  *
 741  * from == to == 0 is code for "zero the entire cluster region"
 742  */
 743 static void ocfs2_clear_page_regions(struct page *page,
 744                                      struct ocfs2_super *osb, u32 cpos,
 745                                      unsigned from, unsigned to)
 746 {
 747         void *kaddr;
 748         unsigned int cluster_start, cluster_end;
 749
 750         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 751
 752         kaddr = kmap_atomic(page, KM_USER0);
 753
 754         if (from || to) {
 755                 if (from > cluster_start)
 756                         memset(kaddr + cluster_start, 0, from - cluster_start);
 757                 if (to < cluster_end)
 758                         memset(kaddr + to, 0, cluster_end - to);
 759         } else {
 760                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 761         }
 762
 763         kunmap_atomic(kaddr, KM_USER0);
 764 }
 765
 766 /*
 767  * Nonsparse file systems fully allocate before we get to the write
 768  * code. This prevents ocfs2_write() from tagging the write as an
 769  * allocating one, which means ocfs2_map_page_blocks() might try to
 770  * read-in the blocks at the tail of our file. Avoid reading them by
 771  * testing i_size against each block offset.
 772  */
 773 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 774                                  unsigned int block_start)
 775 {
 776         u64 offset = page_offset(page) + block_start;
 777
 778         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 779                 return 1;
 780
 781         if (i_size_read(inode) > offset)
 782                 return 1;
 783
 784         return 0;
 785 }
 786
 787 /*
 788  * Some of this taken from block_prepare_write(). We already have our
 789  * mapping by now though, and the entire write will be allocating or
 790  * it won't, so not much need to use BH_New.
 791  *
 792  * This will also skip zeroing, which is handled externally.
 793  */
 794 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 795                           struct inode *inode, unsigned int from,
 796                           unsigned int to, int new)
 797 {
 798         int ret = 0;
 799         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 800         unsigned int block_end, block_start;
 801         unsigned int bsize = 1 << inode->i_blkbits;
 802
 803         if (!page_has_buffers(page))
 804                 create_empty_buffers(page, bsize, 0);
 805
 806         head = page_buffers(page);
 807         for (bh = head, block_start = 0; bh != head || !block_start;
 808              bh = bh->b_this_page, block_start += bsize) {
 809                 block_end = block_start + bsize;
 810
 811                 clear_buffer_new(bh);
 812
 813                 /*
 814                  * Ignore blocks outside of our i/o range -
 815                  * they may belong to unallocated clusters.
 816                  */
 817                 if (block_start >= to || block_end <= from) {
 818                         if (PageUptodate(page))
 819                                 set_buffer_uptodate(bh);
 820                         continue;
 821                 }
 822
 823                 /*
 824                  * For an allocating write with cluster size >= page
 825                  * size, we always write the entire page.
 826                  */
 827                 if (new)
 828                         set_buffer_new(bh);
 829
 830                 if (!buffer_mapped(bh)) {
 831                         map_bh(bh, inode->i_sb, *p_blkno);
 832                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 833                 }
 834
 835                 if (PageUptodate(page)) {
 836                         if (!buffer_uptodate(bh))
 837                                 set_buffer_uptodate(bh);
 838                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 839                            !buffer_new(bh) &&
 840                            ocfs2_should_read_blk(inode, page, block_start) &&
 841                            (block_start < from || block_end > to)) {
 842                         ll_rw_block(READ, 1, &bh);
 843                         *wait_bh++=bh;
 844                 }
 845
 846                 *p_blkno = *p_blkno + 1;
 847         }
 848
 849         /*
 850          * If we issued read requests - let them complete.
 851          */
 852         while(wait_bh > wait) {
 853                 wait_on_buffer(*--wait_bh);
 854                 if (!buffer_uptodate(*wait_bh))
 855                         ret = -EIO;
 856         }
 857
 858         if (ret == 0 || !new)
 859                 return ret;
 860
 861         /*
 862          * If we get -EIO above, zero out any newly allocated blocks
 863          * to avoid exposing stale data.
 864          */
 865         bh = head;
 866         block_start = 0;
 867         do {
 868                 block_end = block_start + bsize;
 869                 if (block_end <= from)
 870                         goto next_bh;
 871                 if (block_start >= to)
 872                         break;
 873
 874                 zero_user(page, block_start, bh->b_size);
 875                 set_buffer_uptodate(bh);
 876                 mark_buffer_dirty(bh);
 877
 878 next_bh:
 879                 block_start = block_end;
 880                 bh = bh->b_this_page;
 881         } while (bh != head);
 882
 883         return ret;
 884 }
 885
 886 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 887 #define OCFS2_MAX_CTXT_PAGES    1
 888 #else
 889 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 890 #endif
 891
 892 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 893
 894 /*
 895  * Describe the state of a single cluster to be written to.
 896  */
 897 struct ocfs2_write_cluster_desc {
 898         u32             c_cpos;
 899         u32             c_phys;
 900         /*
 901          * Give this a unique field because c_phys eventually gets
 902          * filled.
 903          */
 904         unsigned        c_new;
 905         unsigned        c_unwritten;
 906         unsigned        c_needs_zero;
 907 };
 908
 909 struct ocfs2_write_ctxt {
 910         /* Logical cluster position / len of write */
 911         u32                             w_cpos;
 912         u32                             w_clen;
 913
 914         /* First cluster allocated in a nonsparse extend */
 915         u32                             w_first_new_cpos;
 916
 917         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 918
 919         /*
 920          * This is true if page_size > cluster_size.
 921          *
 922          * It triggers a set of special cases during write which might
 923          * have to deal with allocating writes to partial pages.
 924          */
 925         unsigned int                    w_large_pages;
 926
 927         /*
 928          * Pages involved in this write.
 929          *
 930          * w_target_page is the page being written to by the user.
 931          *
 932          * w_pages is an array of pages which always contains
 933          * w_target_page, and in the case of an allocating write with
 934          * page_size < cluster size, it will contain zero'd and mapped
 935          * pages adjacent to w_target_page which need to be written
 936          * out in so that future reads from that region will get
 937          * zero's.
 938          */
 939         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 940         unsigned int                    w_num_pages;
 941         struct page                     *w_target_page;
 942
 943         /*
 944          * ocfs2_write_end() uses this to know what the real range to
 945          * write in the target should be.
 946          */
 947         unsigned int                    w_target_from;
 948         unsigned int                    w_target_to;
 949
 950         /*
 951          * We could use journal_current_handle() but this is cleaner,
 952          * IMHO -Mark
 953          */
 954         handle_t                        *w_handle;
 955
 956         struct buffer_head              *w_di_bh;
 957
 958         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 959 };
 960
 961 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 962 {
 963         int i;
 964
 965         for(i = 0; i < num_pages; i++) {
 966                 if (pages[i]) {
 967                         unlock_page(pages[i]);
 968                         mark_page_accessed(pages[i]);
 969                         page_cache_release(pages[i]);
 970                 }
 971         }
 972 }
 973
 974 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 975 {
 976         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 977
 978         brelse(wc->w_di_bh);
 979         kfree(wc);
 980 }
 981
 982 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 983                                   struct ocfs2_super *osb, loff_t pos,
 984                                   unsigned len, struct buffer_head *di_bh)
 985 {
 986         u32 cend;
 987         struct ocfs2_write_ctxt *wc;
 988
 989         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 990         if (!wc)
 991                 return -ENOMEM;
 992
 993         wc->w_cpos = pos >> osb->s_clustersize_bits;
 994         wc->w_first_new_cpos = UINT_MAX;
 995         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 996         wc->w_clen = cend - wc->w_cpos + 1;
 997         get_bh(di_bh);
 998         wc->w_di_bh = di_bh;
 999
1000         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1001                 wc->w_large_pages = 1;
1002         else
1003                 wc->w_large_pages = 0;
1004
1005         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1006
1007         *wcp = wc;
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * If a page has any new buffers, zero them out here, and mark them uptodate
1014  * and dirty so they'll be written out (in order to prevent uninitialised
1015  * block data from leaking). And clear the new bit.
1016  */
1017 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1018 {
1019         unsigned int block_start, block_end;
1020         struct buffer_head *head, *bh;
1021
1022         BUG_ON(!PageLocked(page));
1023         if (!page_has_buffers(page))
1024                 return;
1025
1026         bh = head = page_buffers(page);
1027         block_start = 0;
1028         do {
1029                 block_end = block_start + bh->b_size;
1030
1031                 if (buffer_new(bh)) {
1032                         if (block_end > from && block_start < to) {
1033                                 if (!PageUptodate(page)) {
1034                                         unsigned start, end;
1035
1036                                         start = max(from, block_start);
1037                                         end = min(to, block_end);
1038
1039                                         zero_user_segment(page, start, end);
1040                                         set_buffer_uptodate(bh);
1041                                 }
1042
1043                                 clear_buffer_new(bh);
1044                                 mark_buffer_dirty(bh);
1045                         }
1046                 }
1047
1048                 block_start = block_end;
1049                 bh = bh->b_this_page;
1050         } while (bh != head);
1051 }
1052
1053 /*
1054  * Only called when we have a failure during allocating write to write
1055  * zero's to the newly allocated region.
1056  */
1057 static void ocfs2_write_failure(struct inode *inode,
1058                                 struct ocfs2_write_ctxt *wc,
1059                                 loff_t user_pos, unsigned user_len)
1060 {
1061         int i;
1062         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1063                 to = user_pos + user_len;
1064         struct page *tmppage;
1065
1066         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1067
1068         for(i = 0; i < wc->w_num_pages; i++) {
1069                 tmppage = wc->w_pages[i];
1070
1071                 if (page_has_buffers(tmppage)) {
1072                         if (ocfs2_should_order_data(inode))
1073                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1074
1075                         block_commit_write(tmppage, from, to);
1076                 }
1077         }
1078 }
1079
1080 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1081                                         struct ocfs2_write_ctxt *wc,
1082                                         struct page *page, u32 cpos,
1083                                         loff_t user_pos, unsigned user_len,
1084                                         int new)
1085 {
1086         int ret;
1087         unsigned int map_from = 0, map_to = 0;
1088         unsigned int cluster_start, cluster_end;
1089         unsigned int user_data_from = 0, user_data_to = 0;
1090
1091         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1092                                         &cluster_start, &cluster_end);
1093
1094         if (page == wc->w_target_page) {
1095                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1096                 map_to = map_from + user_len;
1097
1098                 if (new)
1099                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1100                                                     cluster_start, cluster_end,
1101                                                     new);
1102                 else
1103                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1104                                                     map_from, map_to, new);
1105                 if (ret) {
1106                         mlog_errno(ret);
1107                         goto out;
1108                 }
1109
1110                 user_data_from = map_from;
1111                 user_data_to = map_to;
1112                 if (new) {
1113                         map_from = cluster_start;
1114                         map_to = cluster_end;
1115                 }
1116         } else {
1117                 /*
1118                  * If we haven't allocated the new page yet, we
1119                  * shouldn't be writing it out without copying user
1120                  * data. This is likely a math error from the caller.
1121                  */
1122                 BUG_ON(!new);
1123
1124                 map_from = cluster_start;
1125                 map_to = cluster_end;
1126
1127                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1128                                             cluster_start, cluster_end, new);
1129                 if (ret) {
1130                         mlog_errno(ret);
1131                         goto out;
1132                 }
1133         }
1134
1135         /*
1136          * Parts of newly allocated pages need to be zero'd.
1137          *
1138          * Above, we have also rewritten 'to' and 'from' - as far as
1139          * the rest of the function is concerned, the entire cluster
1140          * range inside of a page needs to be written.
1141          *
1142          * We can skip this if the page is up to date - it's already
1143          * been zero'd from being read in as a hole.
1144          */
1145         if (new && !PageUptodate(page))
1146                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1147                                          cpos, user_data_from, user_data_to);
1148
1149         flush_dcache_page(page);
1150
1151 out:
1152         return ret;
1153 }
1154
1155 /*
1156  * This function will only grab one clusters worth of pages.
1157  */
1158 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1159                                       struct ocfs2_write_ctxt *wc,
1160                                       u32 cpos, loff_t user_pos, int new,
1161                                       struct page *mmap_page)
1162 {
1163         int ret = 0, i;
1164         unsigned long start, target_index, index;
1165         struct inode *inode = mapping->host;
1166
1167         target_index = user_pos >> PAGE_CACHE_SHIFT;
1168
1169         /*
1170          * Figure out how many pages we'll be manipulating here. For
1171          * non allocating write, we just change the one
1172          * page. Otherwise, we'll need a whole clusters worth.
1173          */
1174         if (new) {
1175                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1176                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1177         } else {
1178                 wc->w_num_pages = 1;
1179                 start = target_index;
1180         }
1181
1182         for(i = 0; i < wc->w_num_pages; i++) {
1183                 index = start + i;
1184
1185                 if (index == target_index && mmap_page) {
1186                         /*
1187                          * ocfs2_pagemkwrite() is a little different
1188                          * and wants us to directly use the page
1189                          * passed in.
1190                          */
1191                         lock_page(mmap_page);
1192
1193                         if (mmap_page->mapping != mapping) {
1194                                 unlock_page(mmap_page);
1195                                 /*
1196                                  * Sanity check - the locking in
1197                                  * ocfs2_pagemkwrite() should ensure
1198                                  * that this code doesn't trigger.
1199                                  */
1200                                 ret = -EINVAL;
1201                                 mlog_errno(ret);
1202                                 goto out;
1203                         }
1204
1205                         page_cache_get(mmap_page);
1206                         wc->w_pages[i] = mmap_page;
1207                 } else {
1208                         wc->w_pages[i] = find_or_create_page(mapping, index,
1209                                                              GFP_NOFS);
1210                         if (!wc->w_pages[i]) {
1211                                 ret = -ENOMEM;
1212                                 mlog_errno(ret);
1213                                 goto out;
1214                         }
1215                 }
1216
1217                 if (index == target_index)
1218                         wc->w_target_page = wc->w_pages[i];
1219         }
1220 out:
1221         return ret;
1222 }
1223
1224 /*
1225  * Prepare a single cluster for write one cluster into the file.
1226  */
1227 static int ocfs2_write_cluster(struct address_space *mapping,
1228                                u32 phys, unsigned int unwritten,
1229                                unsigned int should_zero,
1230                                struct ocfs2_alloc_context *data_ac,
1231                                struct ocfs2_alloc_context *meta_ac,
1232                                struct ocfs2_write_ctxt *wc, u32 cpos,
1233                                loff_t user_pos, unsigned user_len)
1234 {
1235         int ret, i, new;
1236         u64 v_blkno, p_blkno;
1237         struct inode *inode = mapping->host;
1238         struct ocfs2_extent_tree et;
1239
1240         new = phys == 0 ? 1 : 0;
1241         if (new) {
1242                 u32 tmp_pos;
1243
1244                 /*
1245                  * This is safe to call with the page locks - it won't take
1246                  * any additional semaphores or cluster locks.
1247                  */
1248                 tmp_pos = cpos;
1249                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1250                                            &tmp_pos, 1, 0, wc->w_di_bh,
1251                                            wc->w_handle, data_ac,
1252                                            meta_ac, NULL);
1253                 /*
1254                  * This shouldn't happen because we must have already
1255                  * calculated the correct meta data allocation required. The
1256                  * internal tree allocation code should know how to increase
1257                  * transaction credits itself.
1258                  *
1259                  * If need be, we could handle -EAGAIN for a
1260                  * RESTART_TRANS here.
1261                  */
1262                 mlog_bug_on_msg(ret == -EAGAIN,
1263                                 "Inode %llu: EAGAIN return during allocation.\n",
1264                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1265                 if (ret < 0) {
1266                         mlog_errno(ret);
1267                         goto out;
1268                 }
1269         } else if (unwritten) {
1270                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1271                                               wc->w_di_bh);
1272                 ret = ocfs2_mark_extent_written(inode, &et,
1273                                                 wc->w_handle, cpos, 1, phys,
1274                                                 meta_ac, &wc->w_dealloc);
1275                 if (ret < 0) {
1276                         mlog_errno(ret);
1277                         goto out;
1278                 }
1279         }
1280
1281         if (should_zero)
1282                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1283         else
1284                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1285
1286         /*
1287          * The only reason this should fail is due to an inability to
1288          * find the extent added.
1289          */
1290         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1291                                           NULL);
1292         if (ret < 0) {
1293                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1294                             "at logical block %llu",
1295                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1296                             (unsigned long long)v_blkno);
1297                 goto out;
1298         }
1299
1300         BUG_ON(p_blkno == 0);
1301
1302         for(i = 0; i < wc->w_num_pages; i++) {
1303                 int tmpret;
1304
1305                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1306                                                       wc->w_pages[i], cpos,
1307                                                       user_pos, user_len,
1308                                                       should_zero);
1309                 if (tmpret) {
1310                         mlog_errno(tmpret);
1311                         if (ret == 0)
1312                                 ret = tmpret;
1313                 }
1314         }
1315
1316         /*
1317          * We only have cleanup to do in case of allocating write.
1318          */
1319         if (ret && new)
1320                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1321
1322 out:
1323
1324         return ret;
1325 }
1326
1327 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1328                                        struct ocfs2_alloc_context *data_ac,
1329                                        struct ocfs2_alloc_context *meta_ac,
1330                                        struct ocfs2_write_ctxt *wc,
1331                                        loff_t pos, unsigned len)
1332 {
1333         int ret, i;
1334         loff_t cluster_off;
1335         unsigned int local_len = len;
1336         struct ocfs2_write_cluster_desc *desc;
1337         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1338
1339         for (i = 0; i < wc->w_clen; i++) {
1340                 desc = &wc->w_desc[i];
1341
1342                 /*
1343                  * We have to make sure that the total write passed in
1344                  * doesn't extend past a single cluster.
1345                  */
1346                 local_len = len;
1347                 cluster_off = pos & (osb->s_clustersize - 1);
1348                 if ((cluster_off + local_len) > osb->s_clustersize)
1349                         local_len = osb->s_clustersize - cluster_off;
1350
1351                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1352                                           desc->c_unwritten,
1353                                           desc->c_needs_zero,
1354                                           data_ac, meta_ac,
1355                                           wc, desc->c_cpos, pos, local_len);
1356                 if (ret) {
1357                         mlog_errno(ret);
1358                         goto out;
1359                 }
1360
1361                 len -= local_len;
1362                 pos += local_len;
1363         }
1364
1365         ret = 0;
1366 out:
1367         return ret;
1368 }
1369
1370 /*
1371  * ocfs2_write_end() wants to know which parts of the target page it
1372  * should complete the write on. It's easiest to compute them ahead of
1373  * time when a more complete view of the write is available.
1374  */
1375 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1376                                         struct ocfs2_write_ctxt *wc,
1377                                         loff_t pos, unsigned len, int alloc)
1378 {
1379         struct ocfs2_write_cluster_desc *desc;
1380
1381         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1382         wc->w_target_to = wc->w_target_from + len;
1383
1384         if (alloc == 0)
1385                 return;
1386
1387         /*
1388          * Allocating write - we may have different boundaries based
1389          * on page size and cluster size.
1390          *
1391          * NOTE: We can no longer compute one value from the other as
1392          * the actual write length and user provided length may be
1393          * different.
1394          */
1395
1396         if (wc->w_large_pages) {
1397                 /*
1398                  * We only care about the 1st and last cluster within
1399                  * our range and whether they should be zero'd or not. Either
1400                  * value may be extended out to the start/end of a
1401                  * newly allocated cluster.
1402                  */
1403                 desc = &wc->w_desc[0];
1404                 if (desc->c_needs_zero)
1405                         ocfs2_figure_cluster_boundaries(osb,
1406                                                         desc->c_cpos,
1407                                                         &wc->w_target_from,
1408                                                         NULL);
1409
1410                 desc = &wc->w_desc[wc->w_clen - 1];
1411                 if (desc->c_needs_zero)
1412                         ocfs2_figure_cluster_boundaries(osb,
1413                                                         desc->c_cpos,
1414                                                         NULL,
1415                                                         &wc->w_target_to);
1416         } else {
1417                 wc->w_target_from = 0;
1418                 wc->w_target_to = PAGE_CACHE_SIZE;
1419         }
1420 }
1421
1422 /*
1423  * Populate each single-cluster write descriptor in the write context
1424  * with information about the i/o to be done.
1425  *
1426  * Returns the number of clusters that will have to be allocated, as
1427  * well as a worst case estimate of the number of extent records that
1428  * would have to be created during a write to an unwritten region.
1429  */
1430 static int ocfs2_populate_write_desc(struct inode *inode,
1431                                      struct ocfs2_write_ctxt *wc,
1432                                      unsigned int *clusters_to_alloc,
1433                                      unsigned int *extents_to_split)
1434 {
1435         int ret;
1436         struct ocfs2_write_cluster_desc *desc;
1437         unsigned int num_clusters = 0;
1438         unsigned int ext_flags = 0;
1439         u32 phys = 0;
1440         int i;
1441
1442         *clusters_to_alloc = 0;
1443         *extents_to_split = 0;
1444
1445         for (i = 0; i < wc->w_clen; i++) {
1446                 desc = &wc->w_desc[i];
1447                 desc->c_cpos = wc->w_cpos + i;
1448
1449                 if (num_clusters == 0) {
1450                         /*
1451                          * Need to look up the next extent record.
1452                          */
1453                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1454                                                  &num_clusters, &ext_flags);
1455                         if (ret) {
1456                                 mlog_errno(ret);
1457                                 goto out;
1458                         }
1459
1460                         /* We should already CoW the refcountd extent. */
1461                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1462
1463                         /*
1464                          * Assume worst case - that we're writing in
1465                          * the middle of the extent.
1466                          *
1467                          * We can assume that the write proceeds from
1468                          * left to right, in which case the extent
1469                          * insert code is smart enough to coalesce the
1470                          * next splits into the previous records created.
1471                          */
1472                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1473                                 *extents_to_split = *extents_to_split + 2;
1474                 } else if (phys) {
1475                         /*
1476                          * Only increment phys if it doesn't describe
1477                          * a hole.
1478                          */
1479                         phys++;
1480                 }
1481
1482                 /*
1483                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1484                  * file that got extended.  w_first_new_cpos tells us
1485                  * where the newly allocated clusters are so we can
1486                  * zero them.
1487                  */
1488                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1489                         BUG_ON(phys == 0);
1490                         desc->c_needs_zero = 1;
1491                 }
1492
1493                 desc->c_phys = phys;
1494                 if (phys == 0) {
1495                         desc->c_new = 1;
1496                         desc->c_needs_zero = 1;
1497                         *clusters_to_alloc = *clusters_to_alloc + 1;
1498                 }
1499
1500                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1501                         desc->c_unwritten = 1;
1502                         desc->c_needs_zero = 1;
1503                 }
1504
1505                 num_clusters--;
1506         }
1507
1508         ret = 0;
1509 out:
1510         return ret;
1511 }
1512
1513 static int ocfs2_write_begin_inline(struct address_space *mapping,
1514                                     struct inode *inode,
1515                                     struct ocfs2_write_ctxt *wc)
1516 {
1517         int ret;
1518         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1519         struct page *page;
1520         handle_t *handle;
1521         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1522
1523         page = find_or_create_page(mapping, 0, GFP_NOFS);
1524         if (!page) {
1525                 ret = -ENOMEM;
1526                 mlog_errno(ret);
1527                 goto out;
1528         }
1529         /*
1530          * If we don't set w_num_pages then this page won't get unlocked
1531          * and freed on cleanup of the write context.
1532          */
1533         wc->w_pages[0] = wc->w_target_page = page;
1534         wc->w_num_pages = 1;
1535
1536         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1537         if (IS_ERR(handle)) {
1538                 ret = PTR_ERR(handle);
1539                 mlog_errno(ret);
1540                 goto out;
1541         }
1542
1543         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1544                                       OCFS2_JOURNAL_ACCESS_WRITE);
1545         if (ret) {
1546                 ocfs2_commit_trans(osb, handle);
1547
1548                 mlog_errno(ret);
1549                 goto out;
1550         }
1551
1552         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1553                 ocfs2_set_inode_data_inline(inode, di);
1554
1555         if (!PageUptodate(page)) {
1556                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1557                 if (ret) {
1558                         ocfs2_commit_trans(osb, handle);
1559
1560                         goto out;
1561                 }
1562         }
1563
1564         wc->w_handle = handle;
1565 out:
1566         return ret;
1567 }
1568
1569 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1570 {
1571         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1572
1573         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1574                 return 1;
1575         return 0;
1576 }
1577
1578 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1579                                           struct inode *inode, loff_t pos,
1580                                           unsigned len, struct page *mmap_page,
1581                                           struct ocfs2_write_ctxt *wc)
1582 {
1583         int ret, written = 0;
1584         loff_t end = pos + len;
1585         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1586         struct ocfs2_dinode *di = NULL;
1587
1588         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1589              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1590              oi->ip_dyn_features);
1591
1592         /*
1593          * Handle inodes which already have inline data 1st.
1594          */
1595         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1596                 if (mmap_page == NULL &&
1597                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1598                         goto do_inline_write;
1599
1600                 /*
1601                  * The write won't fit - we have to give this inode an
1602                  * inline extent list now.
1603                  */
1604                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1605                 if (ret)
1606                         mlog_errno(ret);
1607                 goto out;
1608         }
1609
1610         /*
1611          * Check whether the inode can accept inline data.
1612          */
1613         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1614                 return 0;
1615
1616         /*
1617          * Check whether the write can fit.
1618          */
1619         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1620         if (mmap_page ||
1621             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1622                 return 0;
1623
1624 do_inline_write:
1625         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1626         if (ret) {
1627                 mlog_errno(ret);
1628                 goto out;
1629         }
1630
1631         /*
1632          * This signals to the caller that the data can be written
1633          * inline.
1634          */
1635         written = 1;
1636 out:
1637         return written ? written : ret;
1638 }
1639
1640 /*
1641  * This function only does anything for file systems which can't
1642  * handle sparse files.
1643  *
1644  * What we want to do here is fill in any hole between the current end
1645  * of allocation and the end of our write. That way the rest of the
1646  * write path can treat it as an non-allocating write, which has no
1647  * special case code for sparse/nonsparse files.
1648  */
1649 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1650                                         unsigned len,
1651                                         struct ocfs2_write_ctxt *wc)
1652 {
1653         int ret;
1654         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1655         loff_t newsize = pos + len;
1656
1657         if (ocfs2_sparse_alloc(osb))
1658                 return 0;
1659
1660         if (newsize <= i_size_read(inode))
1661                 return 0;
1662
1663         ret = ocfs2_extend_no_holes(inode, newsize, pos);
1664         if (ret)
1665                 mlog_errno(ret);
1666
1667         wc->w_first_new_cpos =
1668                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1669
1670         return ret;
1671 }
1672
1673 int ocfs2_write_begin_nolock(struct address_space *mapping,
1674                              loff_t pos, unsigned len, unsigned flags,
1675                              struct page **pagep, void **fsdata,
1676                              struct buffer_head *di_bh, struct page *mmap_page)
1677 {
1678         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1679         unsigned int clusters_to_alloc, extents_to_split;
1680         struct ocfs2_write_ctxt *wc;
1681         struct inode *inode = mapping->host;
1682         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1683         struct ocfs2_dinode *di;
1684         struct ocfs2_alloc_context *data_ac = NULL;
1685         struct ocfs2_alloc_context *meta_ac = NULL;
1686         handle_t *handle;
1687         struct ocfs2_extent_tree et;
1688
1689         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1690         if (ret) {
1691                 mlog_errno(ret);
1692                 return ret;
1693         }
1694
1695         if (ocfs2_supports_inline_data(osb)) {
1696                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1697                                                      mmap_page, wc);
1698                 if (ret == 1) {
1699                         ret = 0;
1700                         goto success;
1701                 }
1702                 if (ret < 0) {
1703                         mlog_errno(ret);
1704                         goto out;
1705                 }
1706         }
1707
1708         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1709         if (ret) {
1710                 mlog_errno(ret);
1711                 goto out;
1712         }
1713
1714         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1715         if (ret < 0) {
1716                 mlog_errno(ret);
1717                 goto out;
1718         } else if (ret == 1) {
1719                 ret = ocfs2_refcount_cow(inode, di_bh,
1720                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1721                 if (ret) {
1722                         mlog_errno(ret);
1723                         goto out;
1724                 }
1725         }
1726
1727         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1728                                         &extents_to_split);
1729         if (ret) {
1730                 mlog_errno(ret);
1731                 goto out;
1732         }
1733
1734         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1735
1736         /*
1737          * We set w_target_from, w_target_to here so that
1738          * ocfs2_write_end() knows which range in the target page to
1739          * write out. An allocation requires that we write the entire
1740          * cluster range.
1741          */
1742         if (clusters_to_alloc || extents_to_split) {
1743                 /*
1744                  * XXX: We are stretching the limits of
1745                  * ocfs2_lock_allocators(). It greatly over-estimates
1746                  * the work to be done.
1747                  */
1748                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1749                      " clusters_to_add = %u, extents_to_split = %u\n",
1750                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1751                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1752                      clusters_to_alloc, extents_to_split);
1753
1754                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1755                                               wc->w_di_bh);
1756                 ret = ocfs2_lock_allocators(inode, &et,
1757                                             clusters_to_alloc, extents_to_split,
1758                                             &data_ac, &meta_ac);
1759                 if (ret) {
1760                         mlog_errno(ret);
1761                         goto out;
1762                 }
1763
1764                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1765                                                     &di->id2.i_list,
1766                                                     clusters_to_alloc);
1767
1768         }
1769
1770         /*
1771          * We have to zero sparse allocated clusters, unwritten extent clusters,
1772          * and non-sparse clusters we just extended.  For non-sparse writes,
1773          * we know zeros will only be needed in the first and/or last cluster.
1774          */
1775         if (clusters_to_alloc || extents_to_split ||
1776             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1777                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1778                 cluster_of_pages = 1;
1779         else
1780                 cluster_of_pages = 0;
1781
1782         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1783
1784         handle = ocfs2_start_trans(osb, credits);
1785         if (IS_ERR(handle)) {
1786                 ret = PTR_ERR(handle);
1787                 mlog_errno(ret);
1788                 goto out;
1789         }
1790
1791         wc->w_handle = handle;
1792
1793         if (clusters_to_alloc && vfs_dq_alloc_space_nodirty(inode,
1794                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc))) {
1795                 ret = -EDQUOT;
1796                 goto out_commit;
1797         }
1798         /*
1799          * We don't want this to fail in ocfs2_write_end(), so do it
1800          * here.
1801          */
1802         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1803                                       OCFS2_JOURNAL_ACCESS_WRITE);
1804         if (ret) {
1805                 mlog_errno(ret);
1806                 goto out_quota;
1807         }
1808
1809         /*
1810          * Fill our page array first. That way we've grabbed enough so
1811          * that we can zero and flush if we error after adding the
1812          * extent.
1813          */
1814         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1815                                          cluster_of_pages, mmap_page);
1816         if (ret) {
1817                 mlog_errno(ret);
1818                 goto out_quota;
1819         }
1820
1821         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1822                                           len);
1823         if (ret) {
1824                 mlog_errno(ret);
1825                 goto out_quota;
1826         }
1827
1828         if (data_ac)
1829                 ocfs2_free_alloc_context(data_ac);
1830         if (meta_ac)
1831                 ocfs2_free_alloc_context(meta_ac);
1832
1833 success:
1834         *pagep = wc->w_target_page;
1835         *fsdata = wc;
1836         return 0;
1837 out_quota:
1838         if (clusters_to_alloc)
1839                 vfs_dq_free_space(inode,
1840                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1841 out_commit:
1842         ocfs2_commit_trans(osb, handle);
1843
1844 out:
1845         ocfs2_free_write_ctxt(wc);
1846
1847         if (data_ac)
1848                 ocfs2_free_alloc_context(data_ac);
1849         if (meta_ac)
1850                 ocfs2_free_alloc_context(meta_ac);
1851         return ret;
1852 }
1853
1854 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1855                              loff_t pos, unsigned len, unsigned flags,
1856                              struct page **pagep, void **fsdata)
1857 {
1858         int ret;
1859         struct buffer_head *di_bh = NULL;
1860         struct inode *inode = mapping->host;
1861
1862         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1863         if (ret) {
1864                 mlog_errno(ret);
1865                 return ret;
1866         }
1867
1868         /*
1869          * Take alloc sem here to prevent concurrent lookups. That way
1870          * the mapping, zeroing and tree manipulation within
1871          * ocfs2_write() will be safe against ->readpage(). This
1872          * should also serve to lock out allocation from a shared
1873          * writeable region.
1874          */
1875         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1876
1877         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1878                                        fsdata, di_bh, NULL);
1879         if (ret) {
1880                 mlog_errno(ret);
1881                 goto out_fail;
1882         }
1883
1884         brelse(di_bh);
1885
1886         return 0;
1887
1888 out_fail:
1889         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1890
1891         brelse(di_bh);
1892         ocfs2_inode_unlock(inode, 1);
1893
1894         return ret;
1895 }
1896
1897 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1898                                    unsigned len, unsigned *copied,
1899                                    struct ocfs2_dinode *di,
1900                                    struct ocfs2_write_ctxt *wc)
1901 {
1902         void *kaddr;
1903
1904         if (unlikely(*copied < len)) {
1905                 if (!PageUptodate(wc->w_target_page)) {
1906                         *copied = 0;
1907                         return;
1908                 }
1909         }
1910
1911         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1912         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1913         kunmap_atomic(kaddr, KM_USER0);
1914
1915         mlog(0, "Data written to inode at offset %llu. "
1916              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1917              (unsigned long long)pos, *copied,
1918              le16_to_cpu(di->id2.i_data.id_count),
1919              le16_to_cpu(di->i_dyn_features));
1920 }
1921
1922 int ocfs2_write_end_nolock(struct address_space *mapping,
1923                            loff_t pos, unsigned len, unsigned copied,
1924                            struct page *page, void *fsdata)
1925 {
1926         int i;
1927         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1928         struct inode *inode = mapping->host;
1929         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1930         struct ocfs2_write_ctxt *wc = fsdata;
1931         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1932         handle_t *handle = wc->w_handle;
1933         struct page *tmppage;
1934
1935         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1936                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1937                 goto out_write_size;
1938         }
1939
1940         if (unlikely(copied < len)) {
1941                 if (!PageUptodate(wc->w_target_page))
1942                         copied = 0;
1943
1944                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1945                                        start+len);
1946         }
1947         flush_dcache_page(wc->w_target_page);
1948
1949         for(i = 0; i < wc->w_num_pages; i++) {
1950                 tmppage = wc->w_pages[i];
1951
1952                 if (tmppage == wc->w_target_page) {
1953                         from = wc->w_target_from;
1954                         to = wc->w_target_to;
1955
1956                         BUG_ON(from > PAGE_CACHE_SIZE ||
1957                                to > PAGE_CACHE_SIZE ||
1958                                to < from);
1959                 } else {
1960                         /*
1961                          * Pages adjacent to the target (if any) imply
1962                          * a hole-filling write in which case we want
1963                          * to flush their entire range.
1964                          */
1965                         from = 0;
1966                         to = PAGE_CACHE_SIZE;
1967                 }
1968
1969                 if (page_has_buffers(tmppage)) {
1970                         if (ocfs2_should_order_data(inode))
1971                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1972                         block_commit_write(tmppage, from, to);
1973                 }
1974         }
1975
1976 out_write_size:
1977         pos += copied;
1978         if (pos > inode->i_size) {
1979                 i_size_write(inode, pos);
1980                 mark_inode_dirty(inode);
1981         }
1982         inode->i_blocks = ocfs2_inode_sector_count(inode);
1983         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1984         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1985         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1986         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1987         ocfs2_journal_dirty(handle, wc->w_di_bh);
1988
1989         ocfs2_commit_trans(osb, handle);
1990
1991         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1992
1993         ocfs2_free_write_ctxt(wc);
1994
1995         return copied;
1996 }
1997
1998 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1999                            loff_t pos, unsigned len, unsigned copied,
2000                            struct page *page, void *fsdata)
2001 {
2002         int ret;
2003         struct inode *inode = mapping->host;
2004
2005         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2006
2007         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2008         ocfs2_inode_unlock(inode, 1);
2009
2010         return ret;
2011 }
2012
2013 const struct address_space_operations ocfs2_aops = {
2014         .readpage               = ocfs2_readpage,
2015         .readpages              = ocfs2_readpages,
2016         .writepage              = ocfs2_writepage,
2017         .write_begin            = ocfs2_write_begin,
2018         .write_end              = ocfs2_write_end,
2019         .bmap                   = ocfs2_bmap,
2020         .sync_page              = block_sync_page,
2021         .direct_IO              = ocfs2_direct_IO,
2022         .invalidatepage         = ocfs2_invalidatepage,
2023         .releasepage            = ocfs2_releasepage,
2024         .migratepage            = buffer_migrate_page,
2025         .is_partially_uptodate  = block_is_partially_uptodate,
2026         .error_remove_page      = generic_error_remove_page,
2027 };