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 #include <cluster/masklog.h>
  33
  34 #include "ocfs2.h"
  35
  36 #include "alloc.h"
  37 #include "aops.h"
  38 #include "dlmglue.h"
  39 #include "extent_map.h"
  40 #include "file.h"
  41 #include "inode.h"
  42 #include "journal.h"
  43 #include "suballoc.h"
  44 #include "super.h"
  45 #include "symlink.h"
  46 #include "refcounttree.h"
  47 #include "ocfs2_trace.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         trace_ocfs2_symlink_get_block(
  63                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
  64                         (unsigned long long)iblock, bh_result, create);
  65
  66         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  67
  68         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  69                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  70                      (unsigned long long)iblock);
  71                 goto bail;
  72         }
  73
  74         status = ocfs2_read_inode_block(inode, &bh);
  75         if (status < 0) {
  76                 mlog_errno(status);
  77                 goto bail;
  78         }
  79         fe = (struct ocfs2_dinode *) bh->b_data;
  80
  81         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  82                                                     le32_to_cpu(fe->i_clusters))) {
  83                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  84                      "%llu\n", (unsigned long long)iblock);
  85                 goto bail;
  86         }
  87
  88         /* We don't use the page cache to create symlink data, so if
  89          * need be, copy it over from the buffer cache. */
  90         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  91                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  92                             iblock;
  93                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  94                 if (!buffer_cache_bh) {
  95                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  96                         goto bail;
  97                 }
  98
  99                 /* we haven't locked out transactions, so a commit
 100                  * could've happened. Since we've got a reference on
 101                  * the bh, even if it commits while we're doing the
 102                  * copy, the data is still good. */
 103                 if (buffer_jbd(buffer_cache_bh)
 104                     && ocfs2_inode_is_new(inode)) {
 105                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 106                         if (!kaddr) {
 107                                 mlog(ML_ERROR, "couldn't kmap!\n");
 108                                 goto bail;
 109                         }
 110                         memcpy(kaddr + (bh_result->b_size * iblock),
 111                                buffer_cache_bh->b_data,
 112                                bh_result->b_size);
 113                         kunmap_atomic(kaddr, KM_USER0);
 114                         set_buffer_uptodate(bh_result);
 115                 }
 116                 brelse(buffer_cache_bh);
 117         }
 118
 119         map_bh(bh_result, inode->i_sb,
 120                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 121
 122         err = 0;
 123
 124 bail:
 125         brelse(bh);
 126
 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         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
 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_write_begin() 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
 203         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
 204                                   (unsigned long long)past_eof);
 205         if (create && (iblock >= past_eof))
 206                 set_buffer_new(bh_result);
 207
 208 bail:
 209         if (err < 0)
 210                 err = -EIO;
 211
 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         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
 282                              (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         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         trace_ocfs2_writepage(
 400                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
 401                 page->index);
 402
 403         return block_write_full_page(page, ocfs2_get_block, wbc);
 404 }
 405
 406 /* Taken from ext3. We don't necessarily need the full blown
 407  * functionality yet, but IMHO it's better to cut and paste the whole
 408  * thing so we can avoid introducing our own bugs (and easily pick up
 409  * their fixes when they happen) --Mark */
 410 int walk_page_buffers(  handle_t *handle,
 411                         struct buffer_head *head,
 412                         unsigned from,
 413                         unsigned to,
 414                         int *partial,
 415                         int (*fn)(      handle_t *handle,
 416                                         struct buffer_head *bh))
 417 {
 418         struct buffer_head *bh;
 419         unsigned block_start, block_end;
 420         unsigned blocksize = head->b_size;
 421         int err, ret = 0;
 422         struct buffer_head *next;
 423
 424         for (   bh = head, block_start = 0;
 425                 ret == 0 && (bh != head || !block_start);
 426                 block_start = block_end, bh = next)
 427         {
 428                 next = bh->b_this_page;
 429                 block_end = block_start + blocksize;
 430                 if (block_end <= from || block_start >= to) {
 431                         if (partial && !buffer_uptodate(bh))
 432                                 *partial = 1;
 433                         continue;
 434                 }
 435                 err = (*fn)(handle, bh);
 436                 if (!ret)
 437                         ret = err;
 438         }
 439         return ret;
 440 }
 441
 442 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 443 {
 444         sector_t status;
 445         u64 p_blkno = 0;
 446         int err = 0;
 447         struct inode *inode = mapping->host;
 448
 449         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
 450                          (unsigned long long)block);
 451
 452         /* We don't need to lock journal system files, since they aren't
 453          * accessed concurrently from multiple nodes.
 454          */
 455         if (!INODE_JOURNAL(inode)) {
 456                 err = ocfs2_inode_lock(inode, NULL, 0);
 457                 if (err) {
 458                         if (err != -ENOENT)
 459                                 mlog_errno(err);
 460                         goto bail;
 461                 }
 462                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 463         }
 464
 465         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 466                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 467                                                   NULL);
 468
 469         if (!INODE_JOURNAL(inode)) {
 470                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 471                 ocfs2_inode_unlock(inode, 0);
 472         }
 473
 474         if (err) {
 475                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 476                      (unsigned long long)block);
 477                 mlog_errno(err);
 478                 goto bail;
 479         }
 480
 481 bail:
 482         status = err ? 0 : p_blkno;
 483
 484         return status;
 485 }
 486
 487 /*
 488  * TODO: Make this into a generic get_blocks function.
 489  *
 490  * From do_direct_io in direct-io.c:
 491  *  "So what we do is to permit the ->get_blocks function to populate
 492  *   bh.b_size with the size of IO which is permitted at this offset and
 493  *   this i_blkbits."
 494  *
 495  * This function is called directly from get_more_blocks in direct-io.c.
 496  *
 497  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 498  *                                      fs_count, map_bh, dio->rw == WRITE);
 499  *
 500  * Note that we never bother to allocate blocks here, and thus ignore the
 501  * create argument.
 502  */
 503 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 504                                      struct buffer_head *bh_result, int create)
 505 {
 506         int ret;
 507         u64 p_blkno, inode_blocks, contig_blocks;
 508         unsigned int ext_flags;
 509         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 510         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 511
 512         /* This function won't even be called if the request isn't all
 513          * nicely aligned and of the right size, so there's no need
 514          * for us to check any of that. */
 515
 516         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 517
 518         /* This figures out the size of the next contiguous block, and
 519          * our logical offset */
 520         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 521                                           &contig_blocks, &ext_flags);
 522         if (ret) {
 523                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 524                      (unsigned long long)iblock);
 525                 ret = -EIO;
 526                 goto bail;
 527         }
 528
 529         /* We should already CoW the refcounted extent in case of create. */
 530         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 531
 532         /*
 533          * get_more_blocks() expects us to describe a hole by clearing
 534          * the mapped bit on bh_result().
 535          *
 536          * Consider an unwritten extent as a hole.
 537          */
 538         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 539                 map_bh(bh_result, inode->i_sb, p_blkno);
 540         else
 541                 clear_buffer_mapped(bh_result);
 542
 543         /* make sure we don't map more than max_blocks blocks here as
 544            that's all the kernel will handle at this point. */
 545         if (max_blocks < contig_blocks)
 546                 contig_blocks = max_blocks;
 547         bh_result->b_size = contig_blocks << blocksize_bits;
 548 bail:
 549         return ret;
 550 }
 551
 552 /*
 553  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 554  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
 555  * to protect io on one node from truncation on another.
 556  */
 557 static void ocfs2_dio_end_io(struct kiocb *iocb,
 558                              loff_t offset,
 559                              ssize_t bytes,
 560                              void *private,
 561                              int ret,
 562                              bool is_async)
 563 {
 564         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 565         int level;
 566
 567         /* this io's submitter should not have unlocked this before we could */
 568         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 569
 570         if (ocfs2_iocb_is_sem_locked(iocb)) {
 571                 inode_dio_done(inode);
 572                 ocfs2_iocb_clear_sem_locked(iocb);
 573         }
 574
 575         ocfs2_iocb_clear_rw_locked(iocb);
 576
 577         level = ocfs2_iocb_rw_locked_level(iocb);
 578         ocfs2_rw_unlock(inode, level);
 579
 580         if (is_async)
 581                 aio_complete(iocb, ret, 0);
 582 }
 583
 584 /*
 585  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 586  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 587  * do journalled data.
 588  */
 589 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 590 {
 591         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 592
 593         jbd2_journal_invalidatepage(journal, page, offset);
 594 }
 595
 596 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 597 {
 598         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 599
 600         if (!page_has_buffers(page))
 601                 return 0;
 602         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 603 }
 604
 605 static ssize_t ocfs2_direct_IO(int rw,
 606                                struct kiocb *iocb,
 607                                const struct iovec *iov,
 608                                loff_t offset,
 609                                unsigned long nr_segs)
 610 {
 611         struct file *file = iocb->ki_filp;
 612         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 613
 614         /*
 615          * Fallback to buffered I/O if we see an inode without
 616          * extents.
 617          */
 618         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 619                 return 0;
 620
 621         /* Fallback to buffered I/O if we are appending. */
 622         if (i_size_read(inode) <= offset)
 623                 return 0;
 624
 625         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 626                                     iov, offset, nr_segs,
 627                                     ocfs2_direct_IO_get_blocks,
 628                                     ocfs2_dio_end_io, NULL, 0);
 629 }
 630
 631 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 632                                             u32 cpos,
 633                                             unsigned int *start,
 634                                             unsigned int *end)
 635 {
 636         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 637
 638         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 639                 unsigned int cpp;
 640
 641                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 642
 643                 cluster_start = cpos % cpp;
 644                 cluster_start = cluster_start << osb->s_clustersize_bits;
 645
 646                 cluster_end = cluster_start + osb->s_clustersize;
 647         }
 648
 649         BUG_ON(cluster_start > PAGE_SIZE);
 650         BUG_ON(cluster_end > PAGE_SIZE);
 651
 652         if (start)
 653                 *start = cluster_start;
 654         if (end)
 655                 *end = cluster_end;
 656 }
 657
 658 /*
 659  * 'from' and 'to' are the region in the page to avoid zeroing.
 660  *
 661  * If pagesize > clustersize, this function will avoid zeroing outside
 662  * of the cluster boundary.
 663  *
 664  * from == to == 0 is code for "zero the entire cluster region"
 665  */
 666 static void ocfs2_clear_page_regions(struct page *page,
 667                                      struct ocfs2_super *osb, u32 cpos,
 668                                      unsigned from, unsigned to)
 669 {
 670         void *kaddr;
 671         unsigned int cluster_start, cluster_end;
 672
 673         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 674
 675         kaddr = kmap_atomic(page, KM_USER0);
 676
 677         if (from || to) {
 678                 if (from > cluster_start)
 679                         memset(kaddr + cluster_start, 0, from - cluster_start);
 680                 if (to < cluster_end)
 681                         memset(kaddr + to, 0, cluster_end - to);
 682         } else {
 683                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 684         }
 685
 686         kunmap_atomic(kaddr, KM_USER0);
 687 }
 688
 689 /*
 690  * Nonsparse file systems fully allocate before we get to the write
 691  * code. This prevents ocfs2_write() from tagging the write as an
 692  * allocating one, which means ocfs2_map_page_blocks() might try to
 693  * read-in the blocks at the tail of our file. Avoid reading them by
 694  * testing i_size against each block offset.
 695  */
 696 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 697                                  unsigned int block_start)
 698 {
 699         u64 offset = page_offset(page) + block_start;
 700
 701         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 702                 return 1;
 703
 704         if (i_size_read(inode) > offset)
 705                 return 1;
 706
 707         return 0;
 708 }
 709
 710 /*
 711  * Some of this taken from __block_write_begin(). We already have our
 712  * mapping by now though, and the entire write will be allocating or
 713  * it won't, so not much need to use BH_New.
 714  *
 715  * This will also skip zeroing, which is handled externally.
 716  */
 717 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 718                           struct inode *inode, unsigned int from,
 719                           unsigned int to, int new)
 720 {
 721         int ret = 0;
 722         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 723         unsigned int block_end, block_start;
 724         unsigned int bsize = 1 << inode->i_blkbits;
 725
 726         if (!page_has_buffers(page))
 727                 create_empty_buffers(page, bsize, 0);
 728
 729         head = page_buffers(page);
 730         for (bh = head, block_start = 0; bh != head || !block_start;
 731              bh = bh->b_this_page, block_start += bsize) {
 732                 block_end = block_start + bsize;
 733
 734                 clear_buffer_new(bh);
 735
 736                 /*
 737                  * Ignore blocks outside of our i/o range -
 738                  * they may belong to unallocated clusters.
 739                  */
 740                 if (block_start >= to || block_end <= from) {
 741                         if (PageUptodate(page))
 742                                 set_buffer_uptodate(bh);
 743                         continue;
 744                 }
 745
 746                 /*
 747                  * For an allocating write with cluster size >= page
 748                  * size, we always write the entire page.
 749                  */
 750                 if (new)
 751                         set_buffer_new(bh);
 752
 753                 if (!buffer_mapped(bh)) {
 754                         map_bh(bh, inode->i_sb, *p_blkno);
 755                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 756                 }
 757
 758                 if (PageUptodate(page)) {
 759                         if (!buffer_uptodate(bh))
 760                                 set_buffer_uptodate(bh);
 761                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 762                            !buffer_new(bh) &&
 763                            ocfs2_should_read_blk(inode, page, block_start) &&
 764                            (block_start < from || block_end > to)) {
 765                         ll_rw_block(READ, 1, &bh);
 766                         *wait_bh++=bh;
 767                 }
 768
 769                 *p_blkno = *p_blkno + 1;
 770         }
 771
 772         /*
 773          * If we issued read requests - let them complete.
 774          */
 775         while(wait_bh > wait) {
 776                 wait_on_buffer(*--wait_bh);
 777                 if (!buffer_uptodate(*wait_bh))
 778                         ret = -EIO;
 779         }
 780
 781         if (ret == 0 || !new)
 782                 return ret;
 783
 784         /*
 785          * If we get -EIO above, zero out any newly allocated blocks
 786          * to avoid exposing stale data.
 787          */
 788         bh = head;
 789         block_start = 0;
 790         do {
 791                 block_end = block_start + bsize;
 792                 if (block_end <= from)
 793                         goto next_bh;
 794                 if (block_start >= to)
 795                         break;
 796
 797                 zero_user(page, block_start, bh->b_size);
 798                 set_buffer_uptodate(bh);
 799                 mark_buffer_dirty(bh);
 800
 801 next_bh:
 802                 block_start = block_end;
 803                 bh = bh->b_this_page;
 804         } while (bh != head);
 805
 806         return ret;
 807 }
 808
 809 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 810 #define OCFS2_MAX_CTXT_PAGES    1
 811 #else
 812 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 813 #endif
 814
 815 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 816
 817 /*
 818  * Describe the state of a single cluster to be written to.
 819  */
 820 struct ocfs2_write_cluster_desc {
 821         u32             c_cpos;
 822         u32             c_phys;
 823         /*
 824          * Give this a unique field because c_phys eventually gets
 825          * filled.
 826          */
 827         unsigned        c_new;
 828         unsigned        c_unwritten;
 829         unsigned        c_needs_zero;
 830 };
 831
 832 struct ocfs2_write_ctxt {
 833         /* Logical cluster position / len of write */
 834         u32                             w_cpos;
 835         u32                             w_clen;
 836
 837         /* First cluster allocated in a nonsparse extend */
 838         u32                             w_first_new_cpos;
 839
 840         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 841
 842         /*
 843          * This is true if page_size > cluster_size.
 844          *
 845          * It triggers a set of special cases during write which might
 846          * have to deal with allocating writes to partial pages.
 847          */
 848         unsigned int                    w_large_pages;
 849
 850         /*
 851          * Pages involved in this write.
 852          *
 853          * w_target_page is the page being written to by the user.
 854          *
 855          * w_pages is an array of pages which always contains
 856          * w_target_page, and in the case of an allocating write with
 857          * page_size < cluster size, it will contain zero'd and mapped
 858          * pages adjacent to w_target_page which need to be written
 859          * out in so that future reads from that region will get
 860          * zero's.
 861          */
 862         unsigned int                    w_num_pages;
 863         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 864         struct page                     *w_target_page;
 865
 866         /*
 867          * ocfs2_write_end() uses this to know what the real range to
 868          * write in the target should be.
 869          */
 870         unsigned int                    w_target_from;
 871         unsigned int                    w_target_to;
 872
 873         /*
 874          * We could use journal_current_handle() but this is cleaner,
 875          * IMHO -Mark
 876          */
 877         handle_t                        *w_handle;
 878
 879         struct buffer_head              *w_di_bh;
 880
 881         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 882 };
 883
 884 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 885 {
 886         int i;
 887
 888         for(i = 0; i < num_pages; i++) {
 889                 if (pages[i]) {
 890                         unlock_page(pages[i]);
 891                         mark_page_accessed(pages[i]);
 892                         page_cache_release(pages[i]);
 893                 }
 894         }
 895 }
 896
 897 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 898 {
 899         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 900
 901         brelse(wc->w_di_bh);
 902         kfree(wc);
 903 }
 904
 905 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 906                                   struct ocfs2_super *osb, loff_t pos,
 907                                   unsigned len, struct buffer_head *di_bh)
 908 {
 909         u32 cend;
 910         struct ocfs2_write_ctxt *wc;
 911
 912         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 913         if (!wc)
 914                 return -ENOMEM;
 915
 916         wc->w_cpos = pos >> osb->s_clustersize_bits;
 917         wc->w_first_new_cpos = UINT_MAX;
 918         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 919         wc->w_clen = cend - wc->w_cpos + 1;
 920         get_bh(di_bh);
 921         wc->w_di_bh = di_bh;
 922
 923         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 924                 wc->w_large_pages = 1;
 925         else
 926                 wc->w_large_pages = 0;
 927
 928         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 929
 930         *wcp = wc;
 931
 932         return 0;
 933 }
 934
 935 /*
 936  * If a page has any new buffers, zero them out here, and mark them uptodate
 937  * and dirty so they'll be written out (in order to prevent uninitialised
 938  * block data from leaking). And clear the new bit.
 939  */
 940 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 941 {
 942         unsigned int block_start, block_end;
 943         struct buffer_head *head, *bh;
 944
 945         BUG_ON(!PageLocked(page));
 946         if (!page_has_buffers(page))
 947                 return;
 948
 949         bh = head = page_buffers(page);
 950         block_start = 0;
 951         do {
 952                 block_end = block_start + bh->b_size;
 953
 954                 if (buffer_new(bh)) {
 955                         if (block_end > from && block_start < to) {
 956                                 if (!PageUptodate(page)) {
 957                                         unsigned start, end;
 958
 959                                         start = max(from, block_start);
 960                                         end = min(to, block_end);
 961
 962                                         zero_user_segment(page, start, end);
 963                                         set_buffer_uptodate(bh);
 964                                 }
 965
 966                                 clear_buffer_new(bh);
 967                                 mark_buffer_dirty(bh);
 968                         }
 969                 }
 970
 971                 block_start = block_end;
 972                 bh = bh->b_this_page;
 973         } while (bh != head);
 974 }
 975
 976 /*
 977  * Only called when we have a failure during allocating write to write
 978  * zero's to the newly allocated region.
 979  */
 980 static void ocfs2_write_failure(struct inode *inode,
 981                                 struct ocfs2_write_ctxt *wc,
 982                                 loff_t user_pos, unsigned user_len)
 983 {
 984         int i;
 985         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
 986                 to = user_pos + user_len;
 987         struct page *tmppage;
 988
 989         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
 990
 991         for(i = 0; i < wc->w_num_pages; i++) {
 992                 tmppage = wc->w_pages[i];
 993
 994                 if (page_has_buffers(tmppage)) {
 995                         if (ocfs2_should_order_data(inode))
 996                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
 997
 998                         block_commit_write(tmppage, from, to);
 999                 }
1000         }
1001 }
1002
1003 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1004                                         struct ocfs2_write_ctxt *wc,
1005                                         struct page *page, u32 cpos,
1006                                         loff_t user_pos, unsigned user_len,
1007                                         int new)
1008 {
1009         int ret;
1010         unsigned int map_from = 0, map_to = 0;
1011         unsigned int cluster_start, cluster_end;
1012         unsigned int user_data_from = 0, user_data_to = 0;
1013
1014         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1015                                         &cluster_start, &cluster_end);
1016
1017         /* treat the write as new if the a hole/lseek spanned across
1018          * the page boundary.
1019          */
1020         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1021                         (page_offset(page) <= user_pos));
1022
1023         if (page == wc->w_target_page) {
1024                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1025                 map_to = map_from + user_len;
1026
1027                 if (new)
1028                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1029                                                     cluster_start, cluster_end,
1030                                                     new);
1031                 else
1032                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1033                                                     map_from, map_to, new);
1034                 if (ret) {
1035                         mlog_errno(ret);
1036                         goto out;
1037                 }
1038
1039                 user_data_from = map_from;
1040                 user_data_to = map_to;
1041                 if (new) {
1042                         map_from = cluster_start;
1043                         map_to = cluster_end;
1044                 }
1045         } else {
1046                 /*
1047                  * If we haven't allocated the new page yet, we
1048                  * shouldn't be writing it out without copying user
1049                  * data. This is likely a math error from the caller.
1050                  */
1051                 BUG_ON(!new);
1052
1053                 map_from = cluster_start;
1054                 map_to = cluster_end;
1055
1056                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1057                                             cluster_start, cluster_end, new);
1058                 if (ret) {
1059                         mlog_errno(ret);
1060                         goto out;
1061                 }
1062         }
1063
1064         /*
1065          * Parts of newly allocated pages need to be zero'd.
1066          *
1067          * Above, we have also rewritten 'to' and 'from' - as far as
1068          * the rest of the function is concerned, the entire cluster
1069          * range inside of a page needs to be written.
1070          *
1071          * We can skip this if the page is up to date - it's already
1072          * been zero'd from being read in as a hole.
1073          */
1074         if (new && !PageUptodate(page))
1075                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1076                                          cpos, user_data_from, user_data_to);
1077
1078         flush_dcache_page(page);
1079
1080 out:
1081         return ret;
1082 }
1083
1084 /*
1085  * This function will only grab one clusters worth of pages.
1086  */
1087 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1088                                       struct ocfs2_write_ctxt *wc,
1089                                       u32 cpos, loff_t user_pos,
1090                                       unsigned user_len, int new,
1091                                       struct page *mmap_page)
1092 {
1093         int ret = 0, i;
1094         unsigned long start, target_index, end_index, index;
1095         struct inode *inode = mapping->host;
1096         loff_t last_byte;
1097
1098         target_index = user_pos >> PAGE_CACHE_SHIFT;
1099
1100         /*
1101          * Figure out how many pages we'll be manipulating here. For
1102          * non allocating write, we just change the one
1103          * page. Otherwise, we'll need a whole clusters worth.  If we're
1104          * writing past i_size, we only need enough pages to cover the
1105          * last page of the write.
1106          */
1107         if (new) {
1108                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1109                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1110                 /*
1111                  * We need the index *past* the last page we could possibly
1112                  * touch.  This is the page past the end of the write or
1113                  * i_size, whichever is greater.
1114                  */
1115                 last_byte = max(user_pos + user_len, i_size_read(inode));
1116                 BUG_ON(last_byte < 1);
1117                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1118                 if ((start + wc->w_num_pages) > end_index)
1119                         wc->w_num_pages = end_index - start;
1120         } else {
1121                 wc->w_num_pages = 1;
1122                 start = target_index;
1123         }
1124
1125         for(i = 0; i < wc->w_num_pages; i++) {
1126                 index = start + i;
1127
1128                 if (index == target_index && mmap_page) {
1129                         /*
1130                          * ocfs2_pagemkwrite() is a little different
1131                          * and wants us to directly use the page
1132                          * passed in.
1133                          */
1134                         lock_page(mmap_page);
1135
1136                         if (mmap_page->mapping != mapping) {
1137                                 unlock_page(mmap_page);
1138                                 /*
1139                                  * Sanity check - the locking in
1140                                  * ocfs2_pagemkwrite() should ensure
1141                                  * that this code doesn't trigger.
1142                                  */
1143                                 ret = -EINVAL;
1144                                 mlog_errno(ret);
1145                                 goto out;
1146                         }
1147
1148                         page_cache_get(mmap_page);
1149                         wc->w_pages[i] = mmap_page;
1150                 } else {
1151                         wc->w_pages[i] = find_or_create_page(mapping, index,
1152                                                              GFP_NOFS);
1153                         if (!wc->w_pages[i]) {
1154                                 ret = -ENOMEM;
1155                                 mlog_errno(ret);
1156                                 goto out;
1157                         }
1158                 }
1159
1160                 if (index == target_index)
1161                         wc->w_target_page = wc->w_pages[i];
1162         }
1163 out:
1164         return ret;
1165 }
1166
1167 /*
1168  * Prepare a single cluster for write one cluster into the file.
1169  */
1170 static int ocfs2_write_cluster(struct address_space *mapping,
1171                                u32 phys, unsigned int unwritten,
1172                                unsigned int should_zero,
1173                                struct ocfs2_alloc_context *data_ac,
1174                                struct ocfs2_alloc_context *meta_ac,
1175                                struct ocfs2_write_ctxt *wc, u32 cpos,
1176                                loff_t user_pos, unsigned user_len)
1177 {
1178         int ret, i, new;
1179         u64 v_blkno, p_blkno;
1180         struct inode *inode = mapping->host;
1181         struct ocfs2_extent_tree et;
1182
1183         new = phys == 0 ? 1 : 0;
1184         if (new) {
1185                 u32 tmp_pos;
1186
1187                 /*
1188                  * This is safe to call with the page locks - it won't take
1189                  * any additional semaphores or cluster locks.
1190                  */
1191                 tmp_pos = cpos;
1192                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1193                                            &tmp_pos, 1, 0, wc->w_di_bh,
1194                                            wc->w_handle, data_ac,
1195                                            meta_ac, NULL);
1196                 /*
1197                  * This shouldn't happen because we must have already
1198                  * calculated the correct meta data allocation required. The
1199                  * internal tree allocation code should know how to increase
1200                  * transaction credits itself.
1201                  *
1202                  * If need be, we could handle -EAGAIN for a
1203                  * RESTART_TRANS here.
1204                  */
1205                 mlog_bug_on_msg(ret == -EAGAIN,
1206                                 "Inode %llu: EAGAIN return during allocation.\n",
1207                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1208                 if (ret < 0) {
1209                         mlog_errno(ret);
1210                         goto out;
1211                 }
1212         } else if (unwritten) {
1213                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1214                                               wc->w_di_bh);
1215                 ret = ocfs2_mark_extent_written(inode, &et,
1216                                                 wc->w_handle, cpos, 1, phys,
1217                                                 meta_ac, &wc->w_dealloc);
1218                 if (ret < 0) {
1219                         mlog_errno(ret);
1220                         goto out;
1221                 }
1222         }
1223
1224         if (should_zero)
1225                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1226         else
1227                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1228
1229         /*
1230          * The only reason this should fail is due to an inability to
1231          * find the extent added.
1232          */
1233         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1234                                           NULL);
1235         if (ret < 0) {
1236                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1237                             "at logical block %llu",
1238                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1239                             (unsigned long long)v_blkno);
1240                 goto out;
1241         }
1242
1243         BUG_ON(p_blkno == 0);
1244
1245         for(i = 0; i < wc->w_num_pages; i++) {
1246                 int tmpret;
1247
1248                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1249                                                       wc->w_pages[i], cpos,
1250                                                       user_pos, user_len,
1251                                                       should_zero);
1252                 if (tmpret) {
1253                         mlog_errno(tmpret);
1254                         if (ret == 0)
1255                                 ret = tmpret;
1256                 }
1257         }
1258
1259         /*
1260          * We only have cleanup to do in case of allocating write.
1261          */
1262         if (ret && new)
1263                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1264
1265 out:
1266
1267         return ret;
1268 }
1269
1270 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1271                                        struct ocfs2_alloc_context *data_ac,
1272                                        struct ocfs2_alloc_context *meta_ac,
1273                                        struct ocfs2_write_ctxt *wc,
1274                                        loff_t pos, unsigned len)
1275 {
1276         int ret, i;
1277         loff_t cluster_off;
1278         unsigned int local_len = len;
1279         struct ocfs2_write_cluster_desc *desc;
1280         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1281
1282         for (i = 0; i < wc->w_clen; i++) {
1283                 desc = &wc->w_desc[i];
1284
1285                 /*
1286                  * We have to make sure that the total write passed in
1287                  * doesn't extend past a single cluster.
1288                  */
1289                 local_len = len;
1290                 cluster_off = pos & (osb->s_clustersize - 1);
1291                 if ((cluster_off + local_len) > osb->s_clustersize)
1292                         local_len = osb->s_clustersize - cluster_off;
1293
1294                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1295                                           desc->c_unwritten,
1296                                           desc->c_needs_zero,
1297                                           data_ac, meta_ac,
1298                                           wc, desc->c_cpos, pos, local_len);
1299                 if (ret) {
1300                         mlog_errno(ret);
1301                         goto out;
1302                 }
1303
1304                 len -= local_len;
1305                 pos += local_len;
1306         }
1307
1308         ret = 0;
1309 out:
1310         return ret;
1311 }
1312
1313 /*
1314  * ocfs2_write_end() wants to know which parts of the target page it
1315  * should complete the write on. It's easiest to compute them ahead of
1316  * time when a more complete view of the write is available.
1317  */
1318 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1319                                         struct ocfs2_write_ctxt *wc,
1320                                         loff_t pos, unsigned len, int alloc)
1321 {
1322         struct ocfs2_write_cluster_desc *desc;
1323
1324         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1325         wc->w_target_to = wc->w_target_from + len;
1326
1327         if (alloc == 0)
1328                 return;
1329
1330         /*
1331          * Allocating write - we may have different boundaries based
1332          * on page size and cluster size.
1333          *
1334          * NOTE: We can no longer compute one value from the other as
1335          * the actual write length and user provided length may be
1336          * different.
1337          */
1338
1339         if (wc->w_large_pages) {
1340                 /*
1341                  * We only care about the 1st and last cluster within
1342                  * our range and whether they should be zero'd or not. Either
1343                  * value may be extended out to the start/end of a
1344                  * newly allocated cluster.
1345                  */
1346                 desc = &wc->w_desc[0];
1347                 if (desc->c_needs_zero)
1348                         ocfs2_figure_cluster_boundaries(osb,
1349                                                         desc->c_cpos,
1350                                                         &wc->w_target_from,
1351                                                         NULL);
1352
1353                 desc = &wc->w_desc[wc->w_clen - 1];
1354                 if (desc->c_needs_zero)
1355                         ocfs2_figure_cluster_boundaries(osb,
1356                                                         desc->c_cpos,
1357                                                         NULL,
1358                                                         &wc->w_target_to);
1359         } else {
1360                 wc->w_target_from = 0;
1361                 wc->w_target_to = PAGE_CACHE_SIZE;
1362         }
1363 }
1364
1365 /*
1366  * Populate each single-cluster write descriptor in the write context
1367  * with information about the i/o to be done.
1368  *
1369  * Returns the number of clusters that will have to be allocated, as
1370  * well as a worst case estimate of the number of extent records that
1371  * would have to be created during a write to an unwritten region.
1372  */
1373 static int ocfs2_populate_write_desc(struct inode *inode,
1374                                      struct ocfs2_write_ctxt *wc,
1375                                      unsigned int *clusters_to_alloc,
1376                                      unsigned int *extents_to_split)
1377 {
1378         int ret;
1379         struct ocfs2_write_cluster_desc *desc;
1380         unsigned int num_clusters = 0;
1381         unsigned int ext_flags = 0;
1382         u32 phys = 0;
1383         int i;
1384
1385         *clusters_to_alloc = 0;
1386         *extents_to_split = 0;
1387
1388         for (i = 0; i < wc->w_clen; i++) {
1389                 desc = &wc->w_desc[i];
1390                 desc->c_cpos = wc->w_cpos + i;
1391
1392                 if (num_clusters == 0) {
1393                         /*
1394                          * Need to look up the next extent record.
1395                          */
1396                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1397                                                  &num_clusters, &ext_flags);
1398                         if (ret) {
1399                                 mlog_errno(ret);
1400                                 goto out;
1401                         }
1402
1403                         /* We should already CoW the refcountd extent. */
1404                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1405
1406                         /*
1407                          * Assume worst case - that we're writing in
1408                          * the middle of the extent.
1409                          *
1410                          * We can assume that the write proceeds from
1411                          * left to right, in which case the extent
1412                          * insert code is smart enough to coalesce the
1413                          * next splits into the previous records created.
1414                          */
1415                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1416                                 *extents_to_split = *extents_to_split + 2;
1417                 } else if (phys) {
1418                         /*
1419                          * Only increment phys if it doesn't describe
1420                          * a hole.
1421                          */
1422                         phys++;
1423                 }
1424
1425                 /*
1426                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1427                  * file that got extended.  w_first_new_cpos tells us
1428                  * where the newly allocated clusters are so we can
1429                  * zero them.
1430                  */
1431                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1432                         BUG_ON(phys == 0);
1433                         desc->c_needs_zero = 1;
1434                 }
1435
1436                 desc->c_phys = phys;
1437                 if (phys == 0) {
1438                         desc->c_new = 1;
1439                         desc->c_needs_zero = 1;
1440                         *clusters_to_alloc = *clusters_to_alloc + 1;
1441                 }
1442
1443                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1444                         desc->c_unwritten = 1;
1445                         desc->c_needs_zero = 1;
1446                 }
1447
1448                 num_clusters--;
1449         }
1450
1451         ret = 0;
1452 out:
1453         return ret;
1454 }
1455
1456 static int ocfs2_write_begin_inline(struct address_space *mapping,
1457                                     struct inode *inode,
1458                                     struct ocfs2_write_ctxt *wc)
1459 {
1460         int ret;
1461         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1462         struct page *page;
1463         handle_t *handle;
1464         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1465
1466         page = find_or_create_page(mapping, 0, GFP_NOFS);
1467         if (!page) {
1468                 ret = -ENOMEM;
1469                 mlog_errno(ret);
1470                 goto out;
1471         }
1472         /*
1473          * If we don't set w_num_pages then this page won't get unlocked
1474          * and freed on cleanup of the write context.
1475          */
1476         wc->w_pages[0] = wc->w_target_page = page;
1477         wc->w_num_pages = 1;
1478
1479         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1480         if (IS_ERR(handle)) {
1481                 ret = PTR_ERR(handle);
1482                 mlog_errno(ret);
1483                 goto out;
1484         }
1485
1486         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1487                                       OCFS2_JOURNAL_ACCESS_WRITE);
1488         if (ret) {
1489                 ocfs2_commit_trans(osb, handle);
1490
1491                 mlog_errno(ret);
1492                 goto out;
1493         }
1494
1495         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1496                 ocfs2_set_inode_data_inline(inode, di);
1497
1498         if (!PageUptodate(page)) {
1499                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1500                 if (ret) {
1501                         ocfs2_commit_trans(osb, handle);
1502
1503                         goto out;
1504                 }
1505         }
1506
1507         wc->w_handle = handle;
1508 out:
1509         return ret;
1510 }
1511
1512 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1513 {
1514         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1515
1516         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1517                 return 1;
1518         return 0;
1519 }
1520
1521 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1522                                           struct inode *inode, loff_t pos,
1523                                           unsigned len, struct page *mmap_page,
1524                                           struct ocfs2_write_ctxt *wc)
1525 {
1526         int ret, written = 0;
1527         loff_t end = pos + len;
1528         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1529         struct ocfs2_dinode *di = NULL;
1530
1531         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1532                                              len, (unsigned long long)pos,
1533                                              oi->ip_dyn_features);
1534
1535         /*
1536          * Handle inodes which already have inline data 1st.
1537          */
1538         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1539                 if (mmap_page == NULL &&
1540                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1541                         goto do_inline_write;
1542
1543                 /*
1544                  * The write won't fit - we have to give this inode an
1545                  * inline extent list now.
1546                  */
1547                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1548                 if (ret)
1549                         mlog_errno(ret);
1550                 goto out;
1551         }
1552
1553         /*
1554          * Check whether the inode can accept inline data.
1555          */
1556         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1557                 return 0;
1558
1559         /*
1560          * Check whether the write can fit.
1561          */
1562         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1563         if (mmap_page ||
1564             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1565                 return 0;
1566
1567 do_inline_write:
1568         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1569         if (ret) {
1570                 mlog_errno(ret);
1571                 goto out;
1572         }
1573
1574         /*
1575          * This signals to the caller that the data can be written
1576          * inline.
1577          */
1578         written = 1;
1579 out:
1580         return written ? written : ret;
1581 }
1582
1583 /*
1584  * This function only does anything for file systems which can't
1585  * handle sparse files.
1586  *
1587  * What we want to do here is fill in any hole between the current end
1588  * of allocation and the end of our write. That way the rest of the
1589  * write path can treat it as an non-allocating write, which has no
1590  * special case code for sparse/nonsparse files.
1591  */
1592 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1593                                         struct buffer_head *di_bh,
1594                                         loff_t pos, unsigned len,
1595                                         struct ocfs2_write_ctxt *wc)
1596 {
1597         int ret;
1598         loff_t newsize = pos + len;
1599
1600         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1601
1602         if (newsize <= i_size_read(inode))
1603                 return 0;
1604
1605         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1606         if (ret)
1607                 mlog_errno(ret);
1608
1609         wc->w_first_new_cpos =
1610                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1611
1612         return ret;
1613 }
1614
1615 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1616                            loff_t pos)
1617 {
1618         int ret = 0;
1619
1620         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1621         if (pos > i_size_read(inode))
1622                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1623
1624         return ret;
1625 }
1626
1627 /*
1628  * Try to flush truncate logs if we can free enough clusters from it.
1629  * As for return value, "< 0" means error, "0" no space and "1" means
1630  * we have freed enough spaces and let the caller try to allocate again.
1631  */
1632 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1633                                           unsigned int needed)
1634 {
1635         tid_t target;
1636         int ret = 0;
1637         unsigned int truncated_clusters;
1638
1639         mutex_lock(&osb->osb_tl_inode->i_mutex);
1640         truncated_clusters = osb->truncated_clusters;
1641         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1642
1643         /*
1644          * Check whether we can succeed in allocating if we free
1645          * the truncate log.
1646          */
1647         if (truncated_clusters < needed)
1648                 goto out;
1649
1650         ret = ocfs2_flush_truncate_log(osb);
1651         if (ret) {
1652                 mlog_errno(ret);
1653                 goto out;
1654         }
1655
1656         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1657                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1658                 ret = 1;
1659         }
1660 out:
1661         return ret;
1662 }
1663
1664 int ocfs2_write_begin_nolock(struct file *filp,
1665                              struct address_space *mapping,
1666                              loff_t pos, unsigned len, unsigned flags,
1667                              struct page **pagep, void **fsdata,
1668                              struct buffer_head *di_bh, struct page *mmap_page)
1669 {
1670         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1671         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1672         struct ocfs2_write_ctxt *wc;
1673         struct inode *inode = mapping->host;
1674         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1675         struct ocfs2_dinode *di;
1676         struct ocfs2_alloc_context *data_ac = NULL;
1677         struct ocfs2_alloc_context *meta_ac = NULL;
1678         handle_t *handle;
1679         struct ocfs2_extent_tree et;
1680         int try_free = 1, ret1;
1681
1682 try_again:
1683         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1684         if (ret) {
1685                 mlog_errno(ret);
1686                 return ret;
1687         }
1688
1689         if (ocfs2_supports_inline_data(osb)) {
1690                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1691                                                      mmap_page, wc);
1692                 if (ret == 1) {
1693                         ret = 0;
1694                         goto success;
1695                 }
1696                 if (ret < 0) {
1697                         mlog_errno(ret);
1698                         goto out;
1699                 }
1700         }
1701
1702         if (ocfs2_sparse_alloc(osb))
1703                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1704         else
1705                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1706                                                    wc);
1707         if (ret) {
1708                 mlog_errno(ret);
1709                 goto out;
1710         }
1711
1712         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1713         if (ret < 0) {
1714                 mlog_errno(ret);
1715                 goto out;
1716         } else if (ret == 1) {
1717                 clusters_need = wc->w_clen;
1718                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1719                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1720                 if (ret) {
1721                         mlog_errno(ret);
1722                         goto out;
1723                 }
1724         }
1725
1726         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1727                                         &extents_to_split);
1728         if (ret) {
1729                 mlog_errno(ret);
1730                 goto out;
1731         }
1732         clusters_need += clusters_to_alloc;
1733
1734         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1735
1736         trace_ocfs2_write_begin_nolock(
1737                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1738                         (long long)i_size_read(inode),
1739                         le32_to_cpu(di->i_clusters),
1740                         pos, len, flags, mmap_page,
1741                         clusters_to_alloc, extents_to_split);
1742
1743         /*
1744          * We set w_target_from, w_target_to here so that
1745          * ocfs2_write_end() knows which range in the target page to
1746          * write out. An allocation requires that we write the entire
1747          * cluster range.
1748          */
1749         if (clusters_to_alloc || extents_to_split) {
1750                 /*
1751                  * XXX: We are stretching the limits of
1752                  * ocfs2_lock_allocators(). It greatly over-estimates
1753                  * the work to be done.
1754                  */
1755                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1756                                               wc->w_di_bh);
1757                 ret = ocfs2_lock_allocators(inode, &et,
1758                                             clusters_to_alloc, extents_to_split,
1759                                             &data_ac, &meta_ac);
1760                 if (ret) {
1761                         mlog_errno(ret);
1762                         goto out;
1763                 }
1764
1765                 if (data_ac)
1766                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1767
1768                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1769                                                     &di->id2.i_list,
1770                                                     clusters_to_alloc);
1771
1772         }
1773
1774         /*
1775          * We have to zero sparse allocated clusters, unwritten extent clusters,
1776          * and non-sparse clusters we just extended.  For non-sparse writes,
1777          * we know zeros will only be needed in the first and/or last cluster.
1778          */
1779         if (clusters_to_alloc || extents_to_split ||
1780             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1781                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1782                 cluster_of_pages = 1;
1783         else
1784                 cluster_of_pages = 0;
1785
1786         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1787
1788         handle = ocfs2_start_trans(osb, credits);
1789         if (IS_ERR(handle)) {
1790                 ret = PTR_ERR(handle);
1791                 mlog_errno(ret);
1792                 goto out;
1793         }
1794
1795         wc->w_handle = handle;
1796
1797         if (clusters_to_alloc) {
1798                 ret = dquot_alloc_space_nodirty(inode,
1799                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1800                 if (ret)
1801                         goto out_commit;
1802         }
1803         /*
1804          * We don't want this to fail in ocfs2_write_end(), so do it
1805          * here.
1806          */
1807         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1808                                       OCFS2_JOURNAL_ACCESS_WRITE);
1809         if (ret) {
1810                 mlog_errno(ret);
1811                 goto out_quota;
1812         }
1813
1814         /*
1815          * Fill our page array first. That way we've grabbed enough so
1816          * that we can zero and flush if we error after adding the
1817          * extent.
1818          */
1819         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1820                                          cluster_of_pages, mmap_page);
1821         if (ret) {
1822                 mlog_errno(ret);
1823                 goto out_quota;
1824         }
1825
1826         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1827                                           len);
1828         if (ret) {
1829                 mlog_errno(ret);
1830                 goto out_quota;
1831         }
1832
1833         if (data_ac)
1834                 ocfs2_free_alloc_context(data_ac);
1835         if (meta_ac)
1836                 ocfs2_free_alloc_context(meta_ac);
1837
1838 success:
1839         *pagep = wc->w_target_page;
1840         *fsdata = wc;
1841         return 0;
1842 out_quota:
1843         if (clusters_to_alloc)
1844                 dquot_free_space(inode,
1845                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1846 out_commit:
1847         ocfs2_commit_trans(osb, handle);
1848
1849 out:
1850         ocfs2_free_write_ctxt(wc);
1851
1852         if (data_ac)
1853                 ocfs2_free_alloc_context(data_ac);
1854         if (meta_ac)
1855                 ocfs2_free_alloc_context(meta_ac);
1856
1857         if (ret == -ENOSPC && try_free) {
1858                 /*
1859                  * Try to free some truncate log so that we can have enough
1860                  * clusters to allocate.
1861                  */
1862                 try_free = 0;
1863
1864                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1865                 if (ret1 == 1)
1866                         goto try_again;
1867
1868                 if (ret1 < 0)
1869                         mlog_errno(ret1);
1870         }
1871
1872         return ret;
1873 }
1874
1875 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1876                              loff_t pos, unsigned len, unsigned flags,
1877                              struct page **pagep, void **fsdata)
1878 {
1879         int ret;
1880         struct buffer_head *di_bh = NULL;
1881         struct inode *inode = mapping->host;
1882
1883         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1884         if (ret) {
1885                 mlog_errno(ret);
1886                 return ret;
1887         }
1888
1889         /*
1890          * Take alloc sem here to prevent concurrent lookups. That way
1891          * the mapping, zeroing and tree manipulation within
1892          * ocfs2_write() will be safe against ->readpage(). This
1893          * should also serve to lock out allocation from a shared
1894          * writeable region.
1895          */
1896         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1897
1898         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1899                                        fsdata, di_bh, NULL);
1900         if (ret) {
1901                 mlog_errno(ret);
1902                 goto out_fail;
1903         }
1904
1905         brelse(di_bh);
1906
1907         return 0;
1908
1909 out_fail:
1910         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1911
1912         brelse(di_bh);
1913         ocfs2_inode_unlock(inode, 1);
1914
1915         return ret;
1916 }
1917
1918 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1919                                    unsigned len, unsigned *copied,
1920                                    struct ocfs2_dinode *di,
1921                                    struct ocfs2_write_ctxt *wc)
1922 {
1923         void *kaddr;
1924
1925         if (unlikely(*copied < len)) {
1926                 if (!PageUptodate(wc->w_target_page)) {
1927                         *copied = 0;
1928                         return;
1929                 }
1930         }
1931
1932         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1933         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1934         kunmap_atomic(kaddr, KM_USER0);
1935
1936         trace_ocfs2_write_end_inline(
1937              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1938              (unsigned long long)pos, *copied,
1939              le16_to_cpu(di->id2.i_data.id_count),
1940              le16_to_cpu(di->i_dyn_features));
1941 }
1942
1943 int ocfs2_write_end_nolock(struct address_space *mapping,
1944                            loff_t pos, unsigned len, unsigned copied,
1945                            struct page *page, void *fsdata)
1946 {
1947         int i;
1948         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1949         struct inode *inode = mapping->host;
1950         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1951         struct ocfs2_write_ctxt *wc = fsdata;
1952         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1953         handle_t *handle = wc->w_handle;
1954         struct page *tmppage;
1955
1956         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1957                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1958                 goto out_write_size;
1959         }
1960
1961         if (unlikely(copied < len)) {
1962                 if (!PageUptodate(wc->w_target_page))
1963                         copied = 0;
1964
1965                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1966                                        start+len);
1967         }
1968         flush_dcache_page(wc->w_target_page);
1969
1970         for(i = 0; i < wc->w_num_pages; i++) {
1971                 tmppage = wc->w_pages[i];
1972
1973                 if (tmppage == wc->w_target_page) {
1974                         from = wc->w_target_from;
1975                         to = wc->w_target_to;
1976
1977                         BUG_ON(from > PAGE_CACHE_SIZE ||
1978                                to > PAGE_CACHE_SIZE ||
1979                                to < from);
1980                 } else {
1981                         /*
1982                          * Pages adjacent to the target (if any) imply
1983                          * a hole-filling write in which case we want
1984                          * to flush their entire range.
1985                          */
1986                         from = 0;
1987                         to = PAGE_CACHE_SIZE;
1988                 }
1989
1990                 if (page_has_buffers(tmppage)) {
1991                         if (ocfs2_should_order_data(inode))
1992                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1993                         block_commit_write(tmppage, from, to);
1994                 }
1995         }
1996
1997 out_write_size:
1998         pos += copied;
1999         if (pos > inode->i_size) {
2000                 i_size_write(inode, pos);
2001                 mark_inode_dirty(inode);
2002         }
2003         inode->i_blocks = ocfs2_inode_sector_count(inode);
2004         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2005         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2006         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2007         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2008         ocfs2_journal_dirty(handle, wc->w_di_bh);
2009
2010         ocfs2_commit_trans(osb, handle);
2011
2012         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2013
2014         ocfs2_free_write_ctxt(wc);
2015
2016         return copied;
2017 }
2018
2019 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2020                            loff_t pos, unsigned len, unsigned copied,
2021                            struct page *page, void *fsdata)
2022 {
2023         int ret;
2024         struct inode *inode = mapping->host;
2025
2026         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2027
2028         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2029         ocfs2_inode_unlock(inode, 1);
2030
2031         return ret;
2032 }
2033
2034 const struct address_space_operations ocfs2_aops = {
2035         .readpage               = ocfs2_readpage,
2036         .readpages              = ocfs2_readpages,
2037         .writepage              = ocfs2_writepage,
2038         .write_begin            = ocfs2_write_begin,
2039         .write_end              = ocfs2_write_end,
2040         .bmap                   = ocfs2_bmap,
2041         .direct_IO              = ocfs2_direct_IO,
2042         .invalidatepage         = ocfs2_invalidatepage,
2043         .releasepage            = ocfs2_releasepage,
2044         .migratepage            = buffer_migrate_page,
2045         .is_partially_uptodate  = block_is_partially_uptodate,
2046         .error_remove_page      = generic_error_remove_page,
2047 };