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
 201         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 202         mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 203              (unsigned long long)past_eof);
 204         if (create && (iblock >= past_eof))
 205                 set_buffer_new(bh_result);
 206
 207 bail:
 208         if (err < 0)
 209                 err = -EIO;
 210
 211         mlog_exit(err);
 212         return err;
 213 }
 214
 215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 216                            struct buffer_head *di_bh)
 217 {
 218         void *kaddr;
 219         loff_t size;
 220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 221
 222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 225                 return -EROFS;
 226         }
 227
 228         size = i_size_read(inode);
 229
 230         if (size > PAGE_CACHE_SIZE ||
 231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 232                 ocfs2_error(inode->i_sb,
 233                             "Inode %llu has with inline data has bad size: %Lu",
 234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 235                             (unsigned long long)size);
 236                 return -EROFS;
 237         }
 238
 239         kaddr = kmap_atomic(page, KM_USER0);
 240         if (size)
 241                 memcpy(kaddr, di->id2.i_data.id_data, size);
 242         /* Clear the remaining part of the page */
 243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 244         flush_dcache_page(page);
 245         kunmap_atomic(kaddr, KM_USER0);
 246
 247         SetPageUptodate(page);
 248
 249         return 0;
 250 }
 251
 252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 253 {
 254         int ret;
 255         struct buffer_head *di_bh = NULL;
 256
 257         BUG_ON(!PageLocked(page));
 258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 259
 260         ret = ocfs2_read_inode_block(inode, &di_bh);
 261         if (ret) {
 262                 mlog_errno(ret);
 263                 goto out;
 264         }
 265
 266         ret = ocfs2_read_inline_data(inode, page, di_bh);
 267 out:
 268         unlock_page(page);
 269
 270         brelse(di_bh);
 271         return ret;
 272 }
 273
 274 static int ocfs2_readpage(struct file *file, struct page *page)
 275 {
 276         struct inode *inode = page->mapping->host;
 277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 279         int ret, unlock = 1;
 280
 281         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 282
 283         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 284         if (ret != 0) {
 285                 if (ret == AOP_TRUNCATED_PAGE)
 286                         unlock = 0;
 287                 mlog_errno(ret);
 288                 goto out;
 289         }
 290
 291         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 292                 ret = AOP_TRUNCATED_PAGE;
 293                 goto out_inode_unlock;
 294         }
 295
 296         /*
 297          * i_size might have just been updated as we grabed the meta lock.  We
 298          * might now be discovering a truncate that hit on another node.
 299          * block_read_full_page->get_block freaks out if it is asked to read
 300          * beyond the end of a file, so we check here.  Callers
 301          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 302          * and notice that the page they just read isn't needed.
 303          *
 304          * XXX sys_readahead() seems to get that wrong?
 305          */
 306         if (start >= i_size_read(inode)) {
 307                 zero_user(page, 0, PAGE_SIZE);
 308                 SetPageUptodate(page);
 309                 ret = 0;
 310                 goto out_alloc;
 311         }
 312
 313         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 314                 ret = ocfs2_readpage_inline(inode, page);
 315         else
 316                 ret = block_read_full_page(page, ocfs2_get_block);
 317         unlock = 0;
 318
 319 out_alloc:
 320         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 321 out_inode_unlock:
 322         ocfs2_inode_unlock(inode, 0);
 323 out:
 324         if (unlock)
 325                 unlock_page(page);
 326         mlog_exit(ret);
 327         return ret;
 328 }
 329
 330 /*
 331  * This is used only for read-ahead. Failures or difficult to handle
 332  * situations are safe to ignore.
 333  *
 334  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 335  * are quite large (243 extents on 4k blocks), so most inodes don't
 336  * grow out to a tree. If need be, detecting boundary extents could
 337  * trivially be added in a future version of ocfs2_get_block().
 338  */
 339 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 340                            struct list_head *pages, unsigned nr_pages)
 341 {
 342         int ret, err = -EIO;
 343         struct inode *inode = mapping->host;
 344         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 345         loff_t start;
 346         struct page *last;
 347
 348         /*
 349          * Use the nonblocking flag for the dlm code to avoid page
 350          * lock inversion, but don't bother with retrying.
 351          */
 352         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 353         if (ret)
 354                 return err;
 355
 356         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 357                 ocfs2_inode_unlock(inode, 0);
 358                 return err;
 359         }
 360
 361         /*
 362          * Don't bother with inline-data. There isn't anything
 363          * to read-ahead in that case anyway...
 364          */
 365         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 366                 goto out_unlock;
 367
 368         /*
 369          * Check whether a remote node truncated this file - we just
 370          * drop out in that case as it's not worth handling here.
 371          */
 372         last = list_entry(pages->prev, struct page, lru);
 373         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 374         if (start >= i_size_read(inode))
 375                 goto out_unlock;
 376
 377         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 378
 379 out_unlock:
 380         up_read(&oi->ip_alloc_sem);
 381         ocfs2_inode_unlock(inode, 0);
 382
 383         return err;
 384 }
 385
 386 /* Note: Because we don't support holes, our allocation has
 387  * already happened (allocation writes zeros to the file data)
 388  * so we don't have to worry about ordered writes in
 389  * ocfs2_writepage.
 390  *
 391  * ->writepage is called during the process of invalidating the page cache
 392  * during blocked lock processing.  It can't block on any cluster locks
 393  * to during block mapping.  It's relying on the fact that the block
 394  * mapping can't have disappeared under the dirty pages that it is
 395  * being asked to write back.
 396  */
 397 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 398 {
 399         int ret;
 400
 401         mlog_entry("(0x%p)\n", page);
 402
 403         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 404
 405         mlog_exit(ret);
 406
 407         return ret;
 408 }
 409
 410 /*
 411  * This is called from ocfs2_write_zero_page() which has handled it's
 412  * own cluster locking and has ensured allocation exists for those
 413  * blocks to be written.
 414  */
 415 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 416                                unsigned from, unsigned to)
 417 {
 418         int ret;
 419
 420         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 421
 422         return ret;
 423 }
 424
 425 /* Taken from ext3. We don't necessarily need the full blown
 426  * functionality yet, but IMHO it's better to cut and paste the whole
 427  * thing so we can avoid introducing our own bugs (and easily pick up
 428  * their fixes when they happen) --Mark */
 429 int walk_page_buffers(  handle_t *handle,
 430                         struct buffer_head *head,
 431                         unsigned from,
 432                         unsigned to,
 433                         int *partial,
 434                         int (*fn)(      handle_t *handle,
 435                                         struct buffer_head *bh))
 436 {
 437         struct buffer_head *bh;
 438         unsigned block_start, block_end;
 439         unsigned blocksize = head->b_size;
 440         int err, ret = 0;
 441         struct buffer_head *next;
 442
 443         for (   bh = head, block_start = 0;
 444                 ret == 0 && (bh != head || !block_start);
 445                 block_start = block_end, bh = next)
 446         {
 447                 next = bh->b_this_page;
 448                 block_end = block_start + blocksize;
 449                 if (block_end <= from || block_start >= to) {
 450                         if (partial && !buffer_uptodate(bh))
 451                                 *partial = 1;
 452                         continue;
 453                 }
 454                 err = (*fn)(handle, bh);
 455                 if (!ret)
 456                         ret = err;
 457         }
 458         return ret;
 459 }
 460
 461 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 462 {
 463         sector_t status;
 464         u64 p_blkno = 0;
 465         int err = 0;
 466         struct inode *inode = mapping->host;
 467
 468         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 469
 470         /* We don't need to lock journal system files, since they aren't
 471          * accessed concurrently from multiple nodes.
 472          */
 473         if (!INODE_JOURNAL(inode)) {
 474                 err = ocfs2_inode_lock(inode, NULL, 0);
 475                 if (err) {
 476                         if (err != -ENOENT)
 477                                 mlog_errno(err);
 478                         goto bail;
 479                 }
 480                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 481         }
 482
 483         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 484                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 485                                                   NULL);
 486
 487         if (!INODE_JOURNAL(inode)) {
 488                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 489                 ocfs2_inode_unlock(inode, 0);
 490         }
 491
 492         if (err) {
 493                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 494                      (unsigned long long)block);
 495                 mlog_errno(err);
 496                 goto bail;
 497         }
 498
 499 bail:
 500         status = err ? 0 : p_blkno;
 501
 502         mlog_exit((int)status);
 503
 504         return status;
 505 }
 506
 507 /*
 508  * TODO: Make this into a generic get_blocks function.
 509  *
 510  * From do_direct_io in direct-io.c:
 511  *  "So what we do is to permit the ->get_blocks function to populate
 512  *   bh.b_size with the size of IO which is permitted at this offset and
 513  *   this i_blkbits."
 514  *
 515  * This function is called directly from get_more_blocks in direct-io.c.
 516  *
 517  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 518  *                                      fs_count, map_bh, dio->rw == WRITE);
 519  *
 520  * Note that we never bother to allocate blocks here, and thus ignore the
 521  * create argument.
 522  */
 523 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 524                                      struct buffer_head *bh_result, int create)
 525 {
 526         int ret;
 527         u64 p_blkno, inode_blocks, contig_blocks;
 528         unsigned int ext_flags;
 529         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 530         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 531
 532         /* This function won't even be called if the request isn't all
 533          * nicely aligned and of the right size, so there's no need
 534          * for us to check any of that. */
 535
 536         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 537
 538         /* This figures out the size of the next contiguous block, and
 539          * our logical offset */
 540         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 541                                           &contig_blocks, &ext_flags);
 542         if (ret) {
 543                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 544                      (unsigned long long)iblock);
 545                 ret = -EIO;
 546                 goto bail;
 547         }
 548
 549         /* We should already CoW the refcounted extent in case of create. */
 550         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 551
 552         /*
 553          * get_more_blocks() expects us to describe a hole by clearing
 554          * the mapped bit on bh_result().
 555          *
 556          * Consider an unwritten extent as a hole.
 557          */
 558         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 559                 map_bh(bh_result, inode->i_sb, p_blkno);
 560         else
 561                 clear_buffer_mapped(bh_result);
 562
 563         /* make sure we don't map more than max_blocks blocks here as
 564            that's all the kernel will handle at this point. */
 565         if (max_blocks < contig_blocks)
 566                 contig_blocks = max_blocks;
 567         bh_result->b_size = contig_blocks << blocksize_bits;
 568 bail:
 569         return ret;
 570 }
 571
 572 /*
 573  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 574  * particularly interested in the aio/dio case.  Like the core uses
 575  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 576  * truncation on another.
 577  */
 578 static void ocfs2_dio_end_io(struct kiocb *iocb,
 579                              loff_t offset,
 580                              ssize_t bytes,
 581                              void *private,
 582                              int ret,
 583                              bool is_async)
 584 {
 585         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 586         int level;
 587
 588         /* this io's submitter should not have unlocked this before we could */
 589         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 590
 591         ocfs2_iocb_clear_rw_locked(iocb);
 592
 593         level = ocfs2_iocb_rw_locked_level(iocb);
 594         if (!level)
 595                 up_read(&inode->i_alloc_sem);
 596         ocfs2_rw_unlock(inode, level);
 597
 598         if (is_async)
 599                 aio_complete(iocb, ret, 0);
 600 }
 601
 602 /*
 603  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 604  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 605  * do journalled data.
 606  */
 607 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 608 {
 609         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 610
 611         jbd2_journal_invalidatepage(journal, page, offset);
 612 }
 613
 614 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 615 {
 616         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 617
 618         if (!page_has_buffers(page))
 619                 return 0;
 620         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 621 }
 622
 623 static ssize_t ocfs2_direct_IO(int rw,
 624                                struct kiocb *iocb,
 625                                const struct iovec *iov,
 626                                loff_t offset,
 627                                unsigned long nr_segs)
 628 {
 629         struct file *file = iocb->ki_filp;
 630         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 631         int ret;
 632
 633         mlog_entry_void();
 634
 635         /*
 636          * Fallback to buffered I/O if we see an inode without
 637          * extents.
 638          */
 639         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 640                 return 0;
 641
 642         /* Fallback to buffered I/O if we are appending. */
 643         if (i_size_read(inode) <= offset)
 644                 return 0;
 645
 646         ret = __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 647                                    iov, offset, nr_segs,
 648                                    ocfs2_direct_IO_get_blocks,
 649                                    ocfs2_dio_end_io, NULL, 0);
 650
 651         mlog_exit(ret);
 652         return ret;
 653 }
 654
 655 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 656                                             u32 cpos,
 657                                             unsigned int *start,
 658                                             unsigned int *end)
 659 {
 660         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 661
 662         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 663                 unsigned int cpp;
 664
 665                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 666
 667                 cluster_start = cpos % cpp;
 668                 cluster_start = cluster_start << osb->s_clustersize_bits;
 669
 670                 cluster_end = cluster_start + osb->s_clustersize;
 671         }
 672
 673         BUG_ON(cluster_start > PAGE_SIZE);
 674         BUG_ON(cluster_end > PAGE_SIZE);
 675
 676         if (start)
 677                 *start = cluster_start;
 678         if (end)
 679                 *end = cluster_end;
 680 }
 681
 682 /*
 683  * 'from' and 'to' are the region in the page to avoid zeroing.
 684  *
 685  * If pagesize > clustersize, this function will avoid zeroing outside
 686  * of the cluster boundary.
 687  *
 688  * from == to == 0 is code for "zero the entire cluster region"
 689  */
 690 static void ocfs2_clear_page_regions(struct page *page,
 691                                      struct ocfs2_super *osb, u32 cpos,
 692                                      unsigned from, unsigned to)
 693 {
 694         void *kaddr;
 695         unsigned int cluster_start, cluster_end;
 696
 697         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 698
 699         kaddr = kmap_atomic(page, KM_USER0);
 700
 701         if (from || to) {
 702                 if (from > cluster_start)
 703                         memset(kaddr + cluster_start, 0, from - cluster_start);
 704                 if (to < cluster_end)
 705                         memset(kaddr + to, 0, cluster_end - to);
 706         } else {
 707                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 708         }
 709
 710         kunmap_atomic(kaddr, KM_USER0);
 711 }
 712
 713 /*
 714  * Nonsparse file systems fully allocate before we get to the write
 715  * code. This prevents ocfs2_write() from tagging the write as an
 716  * allocating one, which means ocfs2_map_page_blocks() might try to
 717  * read-in the blocks at the tail of our file. Avoid reading them by
 718  * testing i_size against each block offset.
 719  */
 720 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 721                                  unsigned int block_start)
 722 {
 723         u64 offset = page_offset(page) + block_start;
 724
 725         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 726                 return 1;
 727
 728         if (i_size_read(inode) > offset)
 729                 return 1;
 730
 731         return 0;
 732 }
 733
 734 /*
 735  * Some of this taken from block_prepare_write(). We already have our
 736  * mapping by now though, and the entire write will be allocating or
 737  * it won't, so not much need to use BH_New.
 738  *
 739  * This will also skip zeroing, which is handled externally.
 740  */
 741 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 742                           struct inode *inode, unsigned int from,
 743                           unsigned int to, int new)
 744 {
 745         int ret = 0;
 746         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 747         unsigned int block_end, block_start;
 748         unsigned int bsize = 1 << inode->i_blkbits;
 749
 750         if (!page_has_buffers(page))
 751                 create_empty_buffers(page, bsize, 0);
 752
 753         head = page_buffers(page);
 754         for (bh = head, block_start = 0; bh != head || !block_start;
 755              bh = bh->b_this_page, block_start += bsize) {
 756                 block_end = block_start + bsize;
 757
 758                 clear_buffer_new(bh);
 759
 760                 /*
 761                  * Ignore blocks outside of our i/o range -
 762                  * they may belong to unallocated clusters.
 763                  */
 764                 if (block_start >= to || block_end <= from) {
 765                         if (PageUptodate(page))
 766                                 set_buffer_uptodate(bh);
 767                         continue;
 768                 }
 769
 770                 /*
 771                  * For an allocating write with cluster size >= page
 772                  * size, we always write the entire page.
 773                  */
 774                 if (new)
 775                         set_buffer_new(bh);
 776
 777                 if (!buffer_mapped(bh)) {
 778                         map_bh(bh, inode->i_sb, *p_blkno);
 779                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 780                 }
 781
 782                 if (PageUptodate(page)) {
 783                         if (!buffer_uptodate(bh))
 784                                 set_buffer_uptodate(bh);
 785                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 786                            !buffer_new(bh) &&
 787                            ocfs2_should_read_blk(inode, page, block_start) &&
 788                            (block_start < from || block_end > to)) {
 789                         ll_rw_block(READ, 1, &bh);
 790                         *wait_bh++=bh;
 791                 }
 792
 793                 *p_blkno = *p_blkno + 1;
 794         }
 795
 796         /*
 797          * If we issued read requests - let them complete.
 798          */
 799         while(wait_bh > wait) {
 800                 wait_on_buffer(*--wait_bh);
 801                 if (!buffer_uptodate(*wait_bh))
 802                         ret = -EIO;
 803         }
 804
 805         if (ret == 0 || !new)
 806                 return ret;
 807
 808         /*
 809          * If we get -EIO above, zero out any newly allocated blocks
 810          * to avoid exposing stale data.
 811          */
 812         bh = head;
 813         block_start = 0;
 814         do {
 815                 block_end = block_start + bsize;
 816                 if (block_end <= from)
 817                         goto next_bh;
 818                 if (block_start >= to)
 819                         break;
 820
 821                 zero_user(page, block_start, bh->b_size);
 822                 set_buffer_uptodate(bh);
 823                 mark_buffer_dirty(bh);
 824
 825 next_bh:
 826                 block_start = block_end;
 827                 bh = bh->b_this_page;
 828         } while (bh != head);
 829
 830         return ret;
 831 }
 832
 833 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 834 #define OCFS2_MAX_CTXT_PAGES    1
 835 #else
 836 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 837 #endif
 838
 839 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 840
 841 /*
 842  * Describe the state of a single cluster to be written to.
 843  */
 844 struct ocfs2_write_cluster_desc {
 845         u32             c_cpos;
 846         u32             c_phys;
 847         /*
 848          * Give this a unique field because c_phys eventually gets
 849          * filled.
 850          */
 851         unsigned        c_new;
 852         unsigned        c_unwritten;
 853         unsigned        c_needs_zero;
 854 };
 855
 856 struct ocfs2_write_ctxt {
 857         /* Logical cluster position / len of write */
 858         u32                             w_cpos;
 859         u32                             w_clen;
 860
 861         /* First cluster allocated in a nonsparse extend */
 862         u32                             w_first_new_cpos;
 863
 864         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 865
 866         /*
 867          * This is true if page_size > cluster_size.
 868          *
 869          * It triggers a set of special cases during write which might
 870          * have to deal with allocating writes to partial pages.
 871          */
 872         unsigned int                    w_large_pages;
 873
 874         /*
 875          * Pages involved in this write.
 876          *
 877          * w_target_page is the page being written to by the user.
 878          *
 879          * w_pages is an array of pages which always contains
 880          * w_target_page, and in the case of an allocating write with
 881          * page_size < cluster size, it will contain zero'd and mapped
 882          * pages adjacent to w_target_page which need to be written
 883          * out in so that future reads from that region will get
 884          * zero's.
 885          */
 886         unsigned int                    w_num_pages;
 887         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 888         struct page                     *w_target_page;
 889
 890         /*
 891          * ocfs2_write_end() uses this to know what the real range to
 892          * write in the target should be.
 893          */
 894         unsigned int                    w_target_from;
 895         unsigned int                    w_target_to;
 896
 897         /*
 898          * We could use journal_current_handle() but this is cleaner,
 899          * IMHO -Mark
 900          */
 901         handle_t                        *w_handle;
 902
 903         struct buffer_head              *w_di_bh;
 904
 905         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 906 };
 907
 908 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 909 {
 910         int i;
 911
 912         for(i = 0; i < num_pages; i++) {
 913                 if (pages[i]) {
 914                         unlock_page(pages[i]);
 915                         mark_page_accessed(pages[i]);
 916                         page_cache_release(pages[i]);
 917                 }
 918         }
 919 }
 920
 921 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 922 {
 923         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 924
 925         brelse(wc->w_di_bh);
 926         kfree(wc);
 927 }
 928
 929 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 930                                   struct ocfs2_super *osb, loff_t pos,
 931                                   unsigned len, struct buffer_head *di_bh)
 932 {
 933         u32 cend;
 934         struct ocfs2_write_ctxt *wc;
 935
 936         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 937         if (!wc)
 938                 return -ENOMEM;
 939
 940         wc->w_cpos = pos >> osb->s_clustersize_bits;
 941         wc->w_first_new_cpos = UINT_MAX;
 942         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 943         wc->w_clen = cend - wc->w_cpos + 1;
 944         get_bh(di_bh);
 945         wc->w_di_bh = di_bh;
 946
 947         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 948                 wc->w_large_pages = 1;
 949         else
 950                 wc->w_large_pages = 0;
 951
 952         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 953
 954         *wcp = wc;
 955
 956         return 0;
 957 }
 958
 959 /*
 960  * If a page has any new buffers, zero them out here, and mark them uptodate
 961  * and dirty so they'll be written out (in order to prevent uninitialised
 962  * block data from leaking). And clear the new bit.
 963  */
 964 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 965 {
 966         unsigned int block_start, block_end;
 967         struct buffer_head *head, *bh;
 968
 969         BUG_ON(!PageLocked(page));
 970         if (!page_has_buffers(page))
 971                 return;
 972
 973         bh = head = page_buffers(page);
 974         block_start = 0;
 975         do {
 976                 block_end = block_start + bh->b_size;
 977
 978                 if (buffer_new(bh)) {
 979                         if (block_end > from && block_start < to) {
 980                                 if (!PageUptodate(page)) {
 981                                         unsigned start, end;
 982
 983                                         start = max(from, block_start);
 984                                         end = min(to, block_end);
 985
 986                                         zero_user_segment(page, start, end);
 987                                         set_buffer_uptodate(bh);
 988                                 }
 989
 990                                 clear_buffer_new(bh);
 991                                 mark_buffer_dirty(bh);
 992                         }
 993                 }
 994
 995                 block_start = block_end;
 996                 bh = bh->b_this_page;
 997         } while (bh != head);
 998 }
 999
1000 /*
1001  * Only called when we have a failure during allocating write to write
1002  * zero's to the newly allocated region.
1003  */
1004 static void ocfs2_write_failure(struct inode *inode,
1005                                 struct ocfs2_write_ctxt *wc,
1006                                 loff_t user_pos, unsigned user_len)
1007 {
1008         int i;
1009         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1010                 to = user_pos + user_len;
1011         struct page *tmppage;
1012
1013         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1014
1015         for(i = 0; i < wc->w_num_pages; i++) {
1016                 tmppage = wc->w_pages[i];
1017
1018                 if (page_has_buffers(tmppage)) {
1019                         if (ocfs2_should_order_data(inode))
1020                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1021
1022                         block_commit_write(tmppage, from, to);
1023                 }
1024         }
1025 }
1026
1027 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1028                                         struct ocfs2_write_ctxt *wc,
1029                                         struct page *page, u32 cpos,
1030                                         loff_t user_pos, unsigned user_len,
1031                                         int new)
1032 {
1033         int ret;
1034         unsigned int map_from = 0, map_to = 0;
1035         unsigned int cluster_start, cluster_end;
1036         unsigned int user_data_from = 0, user_data_to = 0;
1037
1038         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1039                                         &cluster_start, &cluster_end);
1040
1041         if (page == wc->w_target_page) {
1042                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1043                 map_to = map_from + user_len;
1044
1045                 if (new)
1046                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1047                                                     cluster_start, cluster_end,
1048                                                     new);
1049                 else
1050                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1051                                                     map_from, map_to, new);
1052                 if (ret) {
1053                         mlog_errno(ret);
1054                         goto out;
1055                 }
1056
1057                 user_data_from = map_from;
1058                 user_data_to = map_to;
1059                 if (new) {
1060                         map_from = cluster_start;
1061                         map_to = cluster_end;
1062                 }
1063         } else {
1064                 /*
1065                  * If we haven't allocated the new page yet, we
1066                  * shouldn't be writing it out without copying user
1067                  * data. This is likely a math error from the caller.
1068                  */
1069                 BUG_ON(!new);
1070
1071                 map_from = cluster_start;
1072                 map_to = cluster_end;
1073
1074                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1075                                             cluster_start, cluster_end, new);
1076                 if (ret) {
1077                         mlog_errno(ret);
1078                         goto out;
1079                 }
1080         }
1081
1082         /*
1083          * Parts of newly allocated pages need to be zero'd.
1084          *
1085          * Above, we have also rewritten 'to' and 'from' - as far as
1086          * the rest of the function is concerned, the entire cluster
1087          * range inside of a page needs to be written.
1088          *
1089          * We can skip this if the page is up to date - it's already
1090          * been zero'd from being read in as a hole.
1091          */
1092         if (new && !PageUptodate(page))
1093                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1094                                          cpos, user_data_from, user_data_to);
1095
1096         flush_dcache_page(page);
1097
1098 out:
1099         return ret;
1100 }
1101
1102 /*
1103  * This function will only grab one clusters worth of pages.
1104  */
1105 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1106                                       struct ocfs2_write_ctxt *wc,
1107                                       u32 cpos, loff_t user_pos,
1108                                       unsigned user_len, int new,
1109                                       struct page *mmap_page)
1110 {
1111         int ret = 0, i;
1112         unsigned long start, target_index, end_index, index;
1113         struct inode *inode = mapping->host;
1114         loff_t last_byte;
1115
1116         target_index = user_pos >> PAGE_CACHE_SHIFT;
1117
1118         /*
1119          * Figure out how many pages we'll be manipulating here. For
1120          * non allocating write, we just change the one
1121          * page. Otherwise, we'll need a whole clusters worth.  If we're
1122          * writing past i_size, we only need enough pages to cover the
1123          * last page of the write.
1124          */
1125         if (new) {
1126                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1127                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1128                 /*
1129                  * We need the index *past* the last page we could possibly
1130                  * touch.  This is the page past the end of the write or
1131                  * i_size, whichever is greater.
1132                  */
1133                 last_byte = max(user_pos + user_len, i_size_read(inode));
1134                 BUG_ON(last_byte < 1);
1135                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1136                 if ((start + wc->w_num_pages) > end_index)
1137                         wc->w_num_pages = end_index - start;
1138         } else {
1139                 wc->w_num_pages = 1;
1140                 start = target_index;
1141         }
1142
1143         for(i = 0; i < wc->w_num_pages; i++) {
1144                 index = start + i;
1145
1146                 if (index == target_index && mmap_page) {
1147                         /*
1148                          * ocfs2_pagemkwrite() is a little different
1149                          * and wants us to directly use the page
1150                          * passed in.
1151                          */
1152                         lock_page(mmap_page);
1153
1154                         if (mmap_page->mapping != mapping) {
1155                                 unlock_page(mmap_page);
1156                                 /*
1157                                  * Sanity check - the locking in
1158                                  * ocfs2_pagemkwrite() should ensure
1159                                  * that this code doesn't trigger.
1160                                  */
1161                                 ret = -EINVAL;
1162                                 mlog_errno(ret);
1163                                 goto out;
1164                         }
1165
1166                         page_cache_get(mmap_page);
1167                         wc->w_pages[i] = mmap_page;
1168                 } else {
1169                         wc->w_pages[i] = find_or_create_page(mapping, index,
1170                                                              GFP_NOFS);
1171                         if (!wc->w_pages[i]) {
1172                                 ret = -ENOMEM;
1173                                 mlog_errno(ret);
1174                                 goto out;
1175                         }
1176                 }
1177
1178                 if (index == target_index)
1179                         wc->w_target_page = wc->w_pages[i];
1180         }
1181 out:
1182         return ret;
1183 }
1184
1185 /*
1186  * Prepare a single cluster for write one cluster into the file.
1187  */
1188 static int ocfs2_write_cluster(struct address_space *mapping,
1189                                u32 phys, unsigned int unwritten,
1190                                unsigned int should_zero,
1191                                struct ocfs2_alloc_context *data_ac,
1192                                struct ocfs2_alloc_context *meta_ac,
1193                                struct ocfs2_write_ctxt *wc, u32 cpos,
1194                                loff_t user_pos, unsigned user_len)
1195 {
1196         int ret, i, new;
1197         u64 v_blkno, p_blkno;
1198         struct inode *inode = mapping->host;
1199         struct ocfs2_extent_tree et;
1200
1201         new = phys == 0 ? 1 : 0;
1202         if (new) {
1203                 u32 tmp_pos;
1204
1205                 /*
1206                  * This is safe to call with the page locks - it won't take
1207                  * any additional semaphores or cluster locks.
1208                  */
1209                 tmp_pos = cpos;
1210                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1211                                            &tmp_pos, 1, 0, wc->w_di_bh,
1212                                            wc->w_handle, data_ac,
1213                                            meta_ac, NULL);
1214                 /*
1215                  * This shouldn't happen because we must have already
1216                  * calculated the correct meta data allocation required. The
1217                  * internal tree allocation code should know how to increase
1218                  * transaction credits itself.
1219                  *
1220                  * If need be, we could handle -EAGAIN for a
1221                  * RESTART_TRANS here.
1222                  */
1223                 mlog_bug_on_msg(ret == -EAGAIN,
1224                                 "Inode %llu: EAGAIN return during allocation.\n",
1225                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1226                 if (ret < 0) {
1227                         mlog_errno(ret);
1228                         goto out;
1229                 }
1230         } else if (unwritten) {
1231                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1232                                               wc->w_di_bh);
1233                 ret = ocfs2_mark_extent_written(inode, &et,
1234                                                 wc->w_handle, cpos, 1, phys,
1235                                                 meta_ac, &wc->w_dealloc);
1236                 if (ret < 0) {
1237                         mlog_errno(ret);
1238                         goto out;
1239                 }
1240         }
1241
1242         if (should_zero)
1243                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1244         else
1245                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1246
1247         /*
1248          * The only reason this should fail is due to an inability to
1249          * find the extent added.
1250          */
1251         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1252                                           NULL);
1253         if (ret < 0) {
1254                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1255                             "at logical block %llu",
1256                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1257                             (unsigned long long)v_blkno);
1258                 goto out;
1259         }
1260
1261         BUG_ON(p_blkno == 0);
1262
1263         for(i = 0; i < wc->w_num_pages; i++) {
1264                 int tmpret;
1265
1266                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1267                                                       wc->w_pages[i], cpos,
1268                                                       user_pos, user_len,
1269                                                       should_zero);
1270                 if (tmpret) {
1271                         mlog_errno(tmpret);
1272                         if (ret == 0)
1273                                 ret = tmpret;
1274                 }
1275         }
1276
1277         /*
1278          * We only have cleanup to do in case of allocating write.
1279          */
1280         if (ret && new)
1281                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1282
1283 out:
1284
1285         return ret;
1286 }
1287
1288 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1289                                        struct ocfs2_alloc_context *data_ac,
1290                                        struct ocfs2_alloc_context *meta_ac,
1291                                        struct ocfs2_write_ctxt *wc,
1292                                        loff_t pos, unsigned len)
1293 {
1294         int ret, i;
1295         loff_t cluster_off;
1296         unsigned int local_len = len;
1297         struct ocfs2_write_cluster_desc *desc;
1298         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1299
1300         for (i = 0; i < wc->w_clen; i++) {
1301                 desc = &wc->w_desc[i];
1302
1303                 /*
1304                  * We have to make sure that the total write passed in
1305                  * doesn't extend past a single cluster.
1306                  */
1307                 local_len = len;
1308                 cluster_off = pos & (osb->s_clustersize - 1);
1309                 if ((cluster_off + local_len) > osb->s_clustersize)
1310                         local_len = osb->s_clustersize - cluster_off;
1311
1312                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1313                                           desc->c_unwritten,
1314                                           desc->c_needs_zero,
1315                                           data_ac, meta_ac,
1316                                           wc, desc->c_cpos, pos, local_len);
1317                 if (ret) {
1318                         mlog_errno(ret);
1319                         goto out;
1320                 }
1321
1322                 len -= local_len;
1323                 pos += local_len;
1324         }
1325
1326         ret = 0;
1327 out:
1328         return ret;
1329 }
1330
1331 /*
1332  * ocfs2_write_end() wants to know which parts of the target page it
1333  * should complete the write on. It's easiest to compute them ahead of
1334  * time when a more complete view of the write is available.
1335  */
1336 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1337                                         struct ocfs2_write_ctxt *wc,
1338                                         loff_t pos, unsigned len, int alloc)
1339 {
1340         struct ocfs2_write_cluster_desc *desc;
1341
1342         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1343         wc->w_target_to = wc->w_target_from + len;
1344
1345         if (alloc == 0)
1346                 return;
1347
1348         /*
1349          * Allocating write - we may have different boundaries based
1350          * on page size and cluster size.
1351          *
1352          * NOTE: We can no longer compute one value from the other as
1353          * the actual write length and user provided length may be
1354          * different.
1355          */
1356
1357         if (wc->w_large_pages) {
1358                 /*
1359                  * We only care about the 1st and last cluster within
1360                  * our range and whether they should be zero'd or not. Either
1361                  * value may be extended out to the start/end of a
1362                  * newly allocated cluster.
1363                  */
1364                 desc = &wc->w_desc[0];
1365                 if (desc->c_needs_zero)
1366                         ocfs2_figure_cluster_boundaries(osb,
1367                                                         desc->c_cpos,
1368                                                         &wc->w_target_from,
1369                                                         NULL);
1370
1371                 desc = &wc->w_desc[wc->w_clen - 1];
1372                 if (desc->c_needs_zero)
1373                         ocfs2_figure_cluster_boundaries(osb,
1374                                                         desc->c_cpos,
1375                                                         NULL,
1376                                                         &wc->w_target_to);
1377         } else {
1378                 wc->w_target_from = 0;
1379                 wc->w_target_to = PAGE_CACHE_SIZE;
1380         }
1381 }
1382
1383 /*
1384  * Populate each single-cluster write descriptor in the write context
1385  * with information about the i/o to be done.
1386  *
1387  * Returns the number of clusters that will have to be allocated, as
1388  * well as a worst case estimate of the number of extent records that
1389  * would have to be created during a write to an unwritten region.
1390  */
1391 static int ocfs2_populate_write_desc(struct inode *inode,
1392                                      struct ocfs2_write_ctxt *wc,
1393                                      unsigned int *clusters_to_alloc,
1394                                      unsigned int *extents_to_split)
1395 {
1396         int ret;
1397         struct ocfs2_write_cluster_desc *desc;
1398         unsigned int num_clusters = 0;
1399         unsigned int ext_flags = 0;
1400         u32 phys = 0;
1401         int i;
1402
1403         *clusters_to_alloc = 0;
1404         *extents_to_split = 0;
1405
1406         for (i = 0; i < wc->w_clen; i++) {
1407                 desc = &wc->w_desc[i];
1408                 desc->c_cpos = wc->w_cpos + i;
1409
1410                 if (num_clusters == 0) {
1411                         /*
1412                          * Need to look up the next extent record.
1413                          */
1414                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1415                                                  &num_clusters, &ext_flags);
1416                         if (ret) {
1417                                 mlog_errno(ret);
1418                                 goto out;
1419                         }
1420
1421                         /* We should already CoW the refcountd extent. */
1422                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1423
1424                         /*
1425                          * Assume worst case - that we're writing in
1426                          * the middle of the extent.
1427                          *
1428                          * We can assume that the write proceeds from
1429                          * left to right, in which case the extent
1430                          * insert code is smart enough to coalesce the
1431                          * next splits into the previous records created.
1432                          */
1433                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1434                                 *extents_to_split = *extents_to_split + 2;
1435                 } else if (phys) {
1436                         /*
1437                          * Only increment phys if it doesn't describe
1438                          * a hole.
1439                          */
1440                         phys++;
1441                 }
1442
1443                 /*
1444                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1445                  * file that got extended.  w_first_new_cpos tells us
1446                  * where the newly allocated clusters are so we can
1447                  * zero them.
1448                  */
1449                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1450                         BUG_ON(phys == 0);
1451                         desc->c_needs_zero = 1;
1452                 }
1453
1454                 desc->c_phys = phys;
1455                 if (phys == 0) {
1456                         desc->c_new = 1;
1457                         desc->c_needs_zero = 1;
1458                         *clusters_to_alloc = *clusters_to_alloc + 1;
1459                 }
1460
1461                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1462                         desc->c_unwritten = 1;
1463                         desc->c_needs_zero = 1;
1464                 }
1465
1466                 num_clusters--;
1467         }
1468
1469         ret = 0;
1470 out:
1471         return ret;
1472 }
1473
1474 static int ocfs2_write_begin_inline(struct address_space *mapping,
1475                                     struct inode *inode,
1476                                     struct ocfs2_write_ctxt *wc)
1477 {
1478         int ret;
1479         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1480         struct page *page;
1481         handle_t *handle;
1482         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1483
1484         page = find_or_create_page(mapping, 0, GFP_NOFS);
1485         if (!page) {
1486                 ret = -ENOMEM;
1487                 mlog_errno(ret);
1488                 goto out;
1489         }
1490         /*
1491          * If we don't set w_num_pages then this page won't get unlocked
1492          * and freed on cleanup of the write context.
1493          */
1494         wc->w_pages[0] = wc->w_target_page = page;
1495         wc->w_num_pages = 1;
1496
1497         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1498         if (IS_ERR(handle)) {
1499                 ret = PTR_ERR(handle);
1500                 mlog_errno(ret);
1501                 goto out;
1502         }
1503
1504         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1505                                       OCFS2_JOURNAL_ACCESS_WRITE);
1506         if (ret) {
1507                 ocfs2_commit_trans(osb, handle);
1508
1509                 mlog_errno(ret);
1510                 goto out;
1511         }
1512
1513         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1514                 ocfs2_set_inode_data_inline(inode, di);
1515
1516         if (!PageUptodate(page)) {
1517                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1518                 if (ret) {
1519                         ocfs2_commit_trans(osb, handle);
1520
1521                         goto out;
1522                 }
1523         }
1524
1525         wc->w_handle = handle;
1526 out:
1527         return ret;
1528 }
1529
1530 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1531 {
1532         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1533
1534         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1535                 return 1;
1536         return 0;
1537 }
1538
1539 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1540                                           struct inode *inode, loff_t pos,
1541                                           unsigned len, struct page *mmap_page,
1542                                           struct ocfs2_write_ctxt *wc)
1543 {
1544         int ret, written = 0;
1545         loff_t end = pos + len;
1546         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1547         struct ocfs2_dinode *di = NULL;
1548
1549         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1550              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1551              oi->ip_dyn_features);
1552
1553         /*
1554          * Handle inodes which already have inline data 1st.
1555          */
1556         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1557                 if (mmap_page == NULL &&
1558                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1559                         goto do_inline_write;
1560
1561                 /*
1562                  * The write won't fit - we have to give this inode an
1563                  * inline extent list now.
1564                  */
1565                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1566                 if (ret)
1567                         mlog_errno(ret);
1568                 goto out;
1569         }
1570
1571         /*
1572          * Check whether the inode can accept inline data.
1573          */
1574         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1575                 return 0;
1576
1577         /*
1578          * Check whether the write can fit.
1579          */
1580         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1581         if (mmap_page ||
1582             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1583                 return 0;
1584
1585 do_inline_write:
1586         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1587         if (ret) {
1588                 mlog_errno(ret);
1589                 goto out;
1590         }
1591
1592         /*
1593          * This signals to the caller that the data can be written
1594          * inline.
1595          */
1596         written = 1;
1597 out:
1598         return written ? written : ret;
1599 }
1600
1601 /*
1602  * This function only does anything for file systems which can't
1603  * handle sparse files.
1604  *
1605  * What we want to do here is fill in any hole between the current end
1606  * of allocation and the end of our write. That way the rest of the
1607  * write path can treat it as an non-allocating write, which has no
1608  * special case code for sparse/nonsparse files.
1609  */
1610 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1611                                         struct buffer_head *di_bh,
1612                                         loff_t pos, unsigned len,
1613                                         struct ocfs2_write_ctxt *wc)
1614 {
1615         int ret;
1616         loff_t newsize = pos + len;
1617
1618         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1619
1620         if (newsize <= i_size_read(inode))
1621                 return 0;
1622
1623         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1624         if (ret)
1625                 mlog_errno(ret);
1626
1627         wc->w_first_new_cpos =
1628                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1629
1630         return ret;
1631 }
1632
1633 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1634                            loff_t pos)
1635 {
1636         int ret = 0;
1637
1638         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1639         if (pos > i_size_read(inode))
1640                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1641
1642         return ret;
1643 }
1644
1645 int ocfs2_write_begin_nolock(struct file *filp,
1646                              struct address_space *mapping,
1647                              loff_t pos, unsigned len, unsigned flags,
1648                              struct page **pagep, void **fsdata,
1649                              struct buffer_head *di_bh, struct page *mmap_page)
1650 {
1651         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1652         unsigned int clusters_to_alloc, extents_to_split;
1653         struct ocfs2_write_ctxt *wc;
1654         struct inode *inode = mapping->host;
1655         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1656         struct ocfs2_dinode *di;
1657         struct ocfs2_alloc_context *data_ac = NULL;
1658         struct ocfs2_alloc_context *meta_ac = NULL;
1659         handle_t *handle;
1660         struct ocfs2_extent_tree et;
1661
1662         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1663         if (ret) {
1664                 mlog_errno(ret);
1665                 return ret;
1666         }
1667
1668         if (ocfs2_supports_inline_data(osb)) {
1669                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1670                                                      mmap_page, wc);
1671                 if (ret == 1) {
1672                         ret = 0;
1673                         goto success;
1674                 }
1675                 if (ret < 0) {
1676                         mlog_errno(ret);
1677                         goto out;
1678                 }
1679         }
1680
1681         if (ocfs2_sparse_alloc(osb))
1682                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1683         else
1684                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1685                                                    wc);
1686         if (ret) {
1687                 mlog_errno(ret);
1688                 goto out;
1689         }
1690
1691         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1692         if (ret < 0) {
1693                 mlog_errno(ret);
1694                 goto out;
1695         } else if (ret == 1) {
1696                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1697                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1698                 if (ret) {
1699                         mlog_errno(ret);
1700                         goto out;
1701                 }
1702         }
1703
1704         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1705                                         &extents_to_split);
1706         if (ret) {
1707                 mlog_errno(ret);
1708                 goto out;
1709         }
1710
1711         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1712
1713         /*
1714          * We set w_target_from, w_target_to here so that
1715          * ocfs2_write_end() knows which range in the target page to
1716          * write out. An allocation requires that we write the entire
1717          * cluster range.
1718          */
1719         if (clusters_to_alloc || extents_to_split) {
1720                 /*
1721                  * XXX: We are stretching the limits of
1722                  * ocfs2_lock_allocators(). It greatly over-estimates
1723                  * the work to be done.
1724                  */
1725                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1726                      " clusters_to_add = %u, extents_to_split = %u\n",
1727                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1728                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1729                      clusters_to_alloc, extents_to_split);
1730
1731                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1732                                               wc->w_di_bh);
1733                 ret = ocfs2_lock_allocators(inode, &et,
1734                                             clusters_to_alloc, extents_to_split,
1735                                             &data_ac, &meta_ac);
1736                 if (ret) {
1737                         mlog_errno(ret);
1738                         goto out;
1739                 }
1740
1741                 if (data_ac)
1742                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1743
1744                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1745                                                     &di->id2.i_list,
1746                                                     clusters_to_alloc);
1747
1748         }
1749
1750         /*
1751          * We have to zero sparse allocated clusters, unwritten extent clusters,
1752          * and non-sparse clusters we just extended.  For non-sparse writes,
1753          * we know zeros will only be needed in the first and/or last cluster.
1754          */
1755         if (clusters_to_alloc || extents_to_split ||
1756             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1757                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1758                 cluster_of_pages = 1;
1759         else
1760                 cluster_of_pages = 0;
1761
1762         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1763
1764         handle = ocfs2_start_trans(osb, credits);
1765         if (IS_ERR(handle)) {
1766                 ret = PTR_ERR(handle);
1767                 mlog_errno(ret);
1768                 goto out;
1769         }
1770
1771         wc->w_handle = handle;
1772
1773         if (clusters_to_alloc) {
1774                 ret = dquot_alloc_space_nodirty(inode,
1775                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1776                 if (ret)
1777                         goto out_commit;
1778         }
1779         /*
1780          * We don't want this to fail in ocfs2_write_end(), so do it
1781          * here.
1782          */
1783         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1784                                       OCFS2_JOURNAL_ACCESS_WRITE);
1785         if (ret) {
1786                 mlog_errno(ret);
1787                 goto out_quota;
1788         }
1789
1790         /*
1791          * Fill our page array first. That way we've grabbed enough so
1792          * that we can zero and flush if we error after adding the
1793          * extent.
1794          */
1795         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1796                                          cluster_of_pages, mmap_page);
1797         if (ret) {
1798                 mlog_errno(ret);
1799                 goto out_quota;
1800         }
1801
1802         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1803                                           len);
1804         if (ret) {
1805                 mlog_errno(ret);
1806                 goto out_quota;
1807         }
1808
1809         if (data_ac)
1810                 ocfs2_free_alloc_context(data_ac);
1811         if (meta_ac)
1812                 ocfs2_free_alloc_context(meta_ac);
1813
1814 success:
1815         *pagep = wc->w_target_page;
1816         *fsdata = wc;
1817         return 0;
1818 out_quota:
1819         if (clusters_to_alloc)
1820                 dquot_free_space(inode,
1821                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1822 out_commit:
1823         ocfs2_commit_trans(osb, handle);
1824
1825 out:
1826         ocfs2_free_write_ctxt(wc);
1827
1828         if (data_ac)
1829                 ocfs2_free_alloc_context(data_ac);
1830         if (meta_ac)
1831                 ocfs2_free_alloc_context(meta_ac);
1832         return ret;
1833 }
1834
1835 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1836                              loff_t pos, unsigned len, unsigned flags,
1837                              struct page **pagep, void **fsdata)
1838 {
1839         int ret;
1840         struct buffer_head *di_bh = NULL;
1841         struct inode *inode = mapping->host;
1842
1843         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1844         if (ret) {
1845                 mlog_errno(ret);
1846                 return ret;
1847         }
1848
1849         /*
1850          * Take alloc sem here to prevent concurrent lookups. That way
1851          * the mapping, zeroing and tree manipulation within
1852          * ocfs2_write() will be safe against ->readpage(). This
1853          * should also serve to lock out allocation from a shared
1854          * writeable region.
1855          */
1856         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1857
1858         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1859                                        fsdata, di_bh, NULL);
1860         if (ret) {
1861                 mlog_errno(ret);
1862                 goto out_fail;
1863         }
1864
1865         brelse(di_bh);
1866
1867         return 0;
1868
1869 out_fail:
1870         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1871
1872         brelse(di_bh);
1873         ocfs2_inode_unlock(inode, 1);
1874
1875         return ret;
1876 }
1877
1878 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1879                                    unsigned len, unsigned *copied,
1880                                    struct ocfs2_dinode *di,
1881                                    struct ocfs2_write_ctxt *wc)
1882 {
1883         void *kaddr;
1884
1885         if (unlikely(*copied < len)) {
1886                 if (!PageUptodate(wc->w_target_page)) {
1887                         *copied = 0;
1888                         return;
1889                 }
1890         }
1891
1892         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1893         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1894         kunmap_atomic(kaddr, KM_USER0);
1895
1896         mlog(0, "Data written to inode at offset %llu. "
1897              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1898              (unsigned long long)pos, *copied,
1899              le16_to_cpu(di->id2.i_data.id_count),
1900              le16_to_cpu(di->i_dyn_features));
1901 }
1902
1903 int ocfs2_write_end_nolock(struct address_space *mapping,
1904                            loff_t pos, unsigned len, unsigned copied,
1905                            struct page *page, void *fsdata)
1906 {
1907         int i;
1908         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1909         struct inode *inode = mapping->host;
1910         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1911         struct ocfs2_write_ctxt *wc = fsdata;
1912         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1913         handle_t *handle = wc->w_handle;
1914         struct page *tmppage;
1915
1916         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1917                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1918                 goto out_write_size;
1919         }
1920
1921         if (unlikely(copied < len)) {
1922                 if (!PageUptodate(wc->w_target_page))
1923                         copied = 0;
1924
1925                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1926                                        start+len);
1927         }
1928         flush_dcache_page(wc->w_target_page);
1929
1930         for(i = 0; i < wc->w_num_pages; i++) {
1931                 tmppage = wc->w_pages[i];
1932
1933                 if (tmppage == wc->w_target_page) {
1934                         from = wc->w_target_from;
1935                         to = wc->w_target_to;
1936
1937                         BUG_ON(from > PAGE_CACHE_SIZE ||
1938                                to > PAGE_CACHE_SIZE ||
1939                                to < from);
1940                 } else {
1941                         /*
1942                          * Pages adjacent to the target (if any) imply
1943                          * a hole-filling write in which case we want
1944                          * to flush their entire range.
1945                          */
1946                         from = 0;
1947                         to = PAGE_CACHE_SIZE;
1948                 }
1949
1950                 if (page_has_buffers(tmppage)) {
1951                         if (ocfs2_should_order_data(inode))
1952                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1953                         block_commit_write(tmppage, from, to);
1954                 }
1955         }
1956
1957 out_write_size:
1958         pos += copied;
1959         if (pos > inode->i_size) {
1960                 i_size_write(inode, pos);
1961                 mark_inode_dirty(inode);
1962         }
1963         inode->i_blocks = ocfs2_inode_sector_count(inode);
1964         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1965         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1966         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1967         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1968         ocfs2_journal_dirty(handle, wc->w_di_bh);
1969
1970         ocfs2_commit_trans(osb, handle);
1971
1972         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1973
1974         ocfs2_free_write_ctxt(wc);
1975
1976         return copied;
1977 }
1978
1979 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1980                            loff_t pos, unsigned len, unsigned copied,
1981                            struct page *page, void *fsdata)
1982 {
1983         int ret;
1984         struct inode *inode = mapping->host;
1985
1986         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1987
1988         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1989         ocfs2_inode_unlock(inode, 1);
1990
1991         return ret;
1992 }
1993
1994 const struct address_space_operations ocfs2_aops = {
1995         .readpage               = ocfs2_readpage,
1996         .readpages              = ocfs2_readpages,
1997         .writepage              = ocfs2_writepage,
1998         .write_begin            = ocfs2_write_begin,
1999         .write_end              = ocfs2_write_end,
2000         .bmap                   = ocfs2_bmap,
2001         .sync_page              = block_sync_page,
2002         .direct_IO              = ocfs2_direct_IO,
2003         .invalidatepage         = ocfs2_invalidatepage,
2004         .releasepage            = ocfs2_releasepage,
2005         .migratepage            = buffer_migrate_page,
2006         .is_partially_uptodate  = block_is_partially_uptodate,
2007         .error_remove_page      = generic_error_remove_page,
2008 };