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                 ocfs2_iocb_clear_sem_locked(iocb);
 572
 573         ocfs2_iocb_clear_rw_locked(iocb);
 574
 575         level = ocfs2_iocb_rw_locked_level(iocb);
 576         ocfs2_rw_unlock(inode, level);
 577
 578         if (is_async)
 579                 aio_complete(iocb, ret, 0);
 580         inode_dio_done(inode);
 581 }
 582
 583 /*
 584  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 585  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 586  * do journalled data.
 587  */
 588 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 589 {
 590         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 591
 592         jbd2_journal_invalidatepage(journal, page, offset);
 593 }
 594
 595 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 596 {
 597         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 598
 599         if (!page_has_buffers(page))
 600                 return 0;
 601         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 602 }
 603
 604 static ssize_t ocfs2_direct_IO(int rw,
 605                                struct kiocb *iocb,
 606                                const struct iovec *iov,
 607                                loff_t offset,
 608                                unsigned long nr_segs)
 609 {
 610         struct file *file = iocb->ki_filp;
 611         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 612
 613         /*
 614          * Fallback to buffered I/O if we see an inode without
 615          * extents.
 616          */
 617         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 618                 return 0;
 619
 620         /* Fallback to buffered I/O if we are appending. */
 621         if (i_size_read(inode) <= offset)
 622                 return 0;
 623
 624         return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
 625                                     iov, offset, nr_segs,
 626                                     ocfs2_direct_IO_get_blocks,
 627                                     ocfs2_dio_end_io, NULL, 0);
 628 }
 629
 630 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 631                                             u32 cpos,
 632                                             unsigned int *start,
 633                                             unsigned int *end)
 634 {
 635         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 636
 637         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 638                 unsigned int cpp;
 639
 640                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 641
 642                 cluster_start = cpos % cpp;
 643                 cluster_start = cluster_start << osb->s_clustersize_bits;
 644
 645                 cluster_end = cluster_start + osb->s_clustersize;
 646         }
 647
 648         BUG_ON(cluster_start > PAGE_SIZE);
 649         BUG_ON(cluster_end > PAGE_SIZE);
 650
 651         if (start)
 652                 *start = cluster_start;
 653         if (end)
 654                 *end = cluster_end;
 655 }
 656
 657 /*
 658  * 'from' and 'to' are the region in the page to avoid zeroing.
 659  *
 660  * If pagesize > clustersize, this function will avoid zeroing outside
 661  * of the cluster boundary.
 662  *
 663  * from == to == 0 is code for "zero the entire cluster region"
 664  */
 665 static void ocfs2_clear_page_regions(struct page *page,
 666                                      struct ocfs2_super *osb, u32 cpos,
 667                                      unsigned from, unsigned to)
 668 {
 669         void *kaddr;
 670         unsigned int cluster_start, cluster_end;
 671
 672         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 673
 674         kaddr = kmap_atomic(page, KM_USER0);
 675
 676         if (from || to) {
 677                 if (from > cluster_start)
 678                         memset(kaddr + cluster_start, 0, from - cluster_start);
 679                 if (to < cluster_end)
 680                         memset(kaddr + to, 0, cluster_end - to);
 681         } else {
 682                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 683         }
 684
 685         kunmap_atomic(kaddr, KM_USER0);
 686 }
 687
 688 /*
 689  * Nonsparse file systems fully allocate before we get to the write
 690  * code. This prevents ocfs2_write() from tagging the write as an
 691  * allocating one, which means ocfs2_map_page_blocks() might try to
 692  * read-in the blocks at the tail of our file. Avoid reading them by
 693  * testing i_size against each block offset.
 694  */
 695 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 696                                  unsigned int block_start)
 697 {
 698         u64 offset = page_offset(page) + block_start;
 699
 700         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 701                 return 1;
 702
 703         if (i_size_read(inode) > offset)
 704                 return 1;
 705
 706         return 0;
 707 }
 708
 709 /*
 710  * Some of this taken from __block_write_begin(). We already have our
 711  * mapping by now though, and the entire write will be allocating or
 712  * it won't, so not much need to use BH_New.
 713  *
 714  * This will also skip zeroing, which is handled externally.
 715  */
 716 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 717                           struct inode *inode, unsigned int from,
 718                           unsigned int to, int new)
 719 {
 720         int ret = 0;
 721         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 722         unsigned int block_end, block_start;
 723         unsigned int bsize = 1 << inode->i_blkbits;
 724
 725         if (!page_has_buffers(page))
 726                 create_empty_buffers(page, bsize, 0);
 727
 728         head = page_buffers(page);
 729         for (bh = head, block_start = 0; bh != head || !block_start;
 730              bh = bh->b_this_page, block_start += bsize) {
 731                 block_end = block_start + bsize;
 732
 733                 clear_buffer_new(bh);
 734
 735                 /*
 736                  * Ignore blocks outside of our i/o range -
 737                  * they may belong to unallocated clusters.
 738                  */
 739                 if (block_start >= to || block_end <= from) {
 740                         if (PageUptodate(page))
 741                                 set_buffer_uptodate(bh);
 742                         continue;
 743                 }
 744
 745                 /*
 746                  * For an allocating write with cluster size >= page
 747                  * size, we always write the entire page.
 748                  */
 749                 if (new)
 750                         set_buffer_new(bh);
 751
 752                 if (!buffer_mapped(bh)) {
 753                         map_bh(bh, inode->i_sb, *p_blkno);
 754                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 755                 }
 756
 757                 if (PageUptodate(page)) {
 758                         if (!buffer_uptodate(bh))
 759                                 set_buffer_uptodate(bh);
 760                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 761                            !buffer_new(bh) &&
 762                            ocfs2_should_read_blk(inode, page, block_start) &&
 763                            (block_start < from || block_end > to)) {
 764                         ll_rw_block(READ, 1, &bh);
 765                         *wait_bh++=bh;
 766                 }
 767
 768                 *p_blkno = *p_blkno + 1;
 769         }
 770
 771         /*
 772          * If we issued read requests - let them complete.
 773          */
 774         while(wait_bh > wait) {
 775                 wait_on_buffer(*--wait_bh);
 776                 if (!buffer_uptodate(*wait_bh))
 777                         ret = -EIO;
 778         }
 779
 780         if (ret == 0 || !new)
 781                 return ret;
 782
 783         /*
 784          * If we get -EIO above, zero out any newly allocated blocks
 785          * to avoid exposing stale data.
 786          */
 787         bh = head;
 788         block_start = 0;
 789         do {
 790                 block_end = block_start + bsize;
 791                 if (block_end <= from)
 792                         goto next_bh;
 793                 if (block_start >= to)
 794                         break;
 795
 796                 zero_user(page, block_start, bh->b_size);
 797                 set_buffer_uptodate(bh);
 798                 mark_buffer_dirty(bh);
 799
 800 next_bh:
 801                 block_start = block_end;
 802                 bh = bh->b_this_page;
 803         } while (bh != head);
 804
 805         return ret;
 806 }
 807
 808 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 809 #define OCFS2_MAX_CTXT_PAGES    1
 810 #else
 811 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 812 #endif
 813
 814 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 815
 816 /*
 817  * Describe the state of a single cluster to be written to.
 818  */
 819 struct ocfs2_write_cluster_desc {
 820         u32             c_cpos;
 821         u32             c_phys;
 822         /*
 823          * Give this a unique field because c_phys eventually gets
 824          * filled.
 825          */
 826         unsigned        c_new;
 827         unsigned        c_unwritten;
 828         unsigned        c_needs_zero;
 829 };
 830
 831 struct ocfs2_write_ctxt {
 832         /* Logical cluster position / len of write */
 833         u32                             w_cpos;
 834         u32                             w_clen;
 835
 836         /* First cluster allocated in a nonsparse extend */
 837         u32                             w_first_new_cpos;
 838
 839         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 840
 841         /*
 842          * This is true if page_size > cluster_size.
 843          *
 844          * It triggers a set of special cases during write which might
 845          * have to deal with allocating writes to partial pages.
 846          */
 847         unsigned int                    w_large_pages;
 848
 849         /*
 850          * Pages involved in this write.
 851          *
 852          * w_target_page is the page being written to by the user.
 853          *
 854          * w_pages is an array of pages which always contains
 855          * w_target_page, and in the case of an allocating write with
 856          * page_size < cluster size, it will contain zero'd and mapped
 857          * pages adjacent to w_target_page which need to be written
 858          * out in so that future reads from that region will get
 859          * zero's.
 860          */
 861         unsigned int                    w_num_pages;
 862         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 863         struct page                     *w_target_page;
 864
 865         /*
 866          * w_target_locked is used for page_mkwrite path indicating no unlocking
 867          * against w_target_page in ocfs2_write_end_nolock.
 868          */
 869         unsigned int                    w_target_locked:1;
 870
 871         /*
 872          * ocfs2_write_end() uses this to know what the real range to
 873          * write in the target should be.
 874          */
 875         unsigned int                    w_target_from;
 876         unsigned int                    w_target_to;
 877
 878         /*
 879          * We could use journal_current_handle() but this is cleaner,
 880          * IMHO -Mark
 881          */
 882         handle_t                        *w_handle;
 883
 884         struct buffer_head              *w_di_bh;
 885
 886         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 887 };
 888
 889 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 890 {
 891         int i;
 892
 893         for(i = 0; i < num_pages; i++) {
 894                 if (pages[i]) {
 895                         unlock_page(pages[i]);
 896                         mark_page_accessed(pages[i]);
 897                         page_cache_release(pages[i]);
 898                 }
 899         }
 900 }
 901
 902 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 903 {
 904         int i;
 905
 906         /*
 907          * w_target_locked is only set to true in the page_mkwrite() case.
 908          * The intent is to allow us to lock the target page from write_begin()
 909          * to write_end(). The caller must hold a ref on w_target_page.
 910          */
 911         if (wc->w_target_locked) {
 912                 BUG_ON(!wc->w_target_page);
 913                 for (i = 0; i < wc->w_num_pages; i++) {
 914                         if (wc->w_target_page == wc->w_pages[i]) {
 915                                 wc->w_pages[i] = NULL;
 916                                 break;
 917                         }
 918                 }
 919                 mark_page_accessed(wc->w_target_page);
 920                 page_cache_release(wc->w_target_page);
 921         }
 922         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 923
 924         brelse(wc->w_di_bh);
 925         kfree(wc);
 926 }
 927
 928 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 929                                   struct ocfs2_super *osb, loff_t pos,
 930                                   unsigned len, struct buffer_head *di_bh)
 931 {
 932         u32 cend;
 933         struct ocfs2_write_ctxt *wc;
 934
 935         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 936         if (!wc)
 937                 return -ENOMEM;
 938
 939         wc->w_cpos = pos >> osb->s_clustersize_bits;
 940         wc->w_first_new_cpos = UINT_MAX;
 941         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 942         wc->w_clen = cend - wc->w_cpos + 1;
 943         get_bh(di_bh);
 944         wc->w_di_bh = di_bh;
 945
 946         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 947                 wc->w_large_pages = 1;
 948         else
 949                 wc->w_large_pages = 0;
 950
 951         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 952
 953         *wcp = wc;
 954
 955         return 0;
 956 }
 957
 958 /*
 959  * If a page has any new buffers, zero them out here, and mark them uptodate
 960  * and dirty so they'll be written out (in order to prevent uninitialised
 961  * block data from leaking). And clear the new bit.
 962  */
 963 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 964 {
 965         unsigned int block_start, block_end;
 966         struct buffer_head *head, *bh;
 967
 968         BUG_ON(!PageLocked(page));
 969         if (!page_has_buffers(page))
 970                 return;
 971
 972         bh = head = page_buffers(page);
 973         block_start = 0;
 974         do {
 975                 block_end = block_start + bh->b_size;
 976
 977                 if (buffer_new(bh)) {
 978                         if (block_end > from && block_start < to) {
 979                                 if (!PageUptodate(page)) {
 980                                         unsigned start, end;
 981
 982                                         start = max(from, block_start);
 983                                         end = min(to, block_end);
 984
 985                                         zero_user_segment(page, start, end);
 986                                         set_buffer_uptodate(bh);
 987                                 }
 988
 989                                 clear_buffer_new(bh);
 990                                 mark_buffer_dirty(bh);
 991                         }
 992                 }
 993
 994                 block_start = block_end;
 995                 bh = bh->b_this_page;
 996         } while (bh != head);
 997 }
 998
 999 /*
1000  * Only called when we have a failure during allocating write to write
1001  * zero's to the newly allocated region.
1002  */
1003 static void ocfs2_write_failure(struct inode *inode,
1004                                 struct ocfs2_write_ctxt *wc,
1005                                 loff_t user_pos, unsigned user_len)
1006 {
1007         int i;
1008         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1009                 to = user_pos + user_len;
1010         struct page *tmppage;
1011
1012         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1013
1014         for(i = 0; i < wc->w_num_pages; i++) {
1015                 tmppage = wc->w_pages[i];
1016
1017                 if (page_has_buffers(tmppage)) {
1018                         if (ocfs2_should_order_data(inode))
1019                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1020
1021                         block_commit_write(tmppage, from, to);
1022                 }
1023         }
1024 }
1025
1026 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1027                                         struct ocfs2_write_ctxt *wc,
1028                                         struct page *page, u32 cpos,
1029                                         loff_t user_pos, unsigned user_len,
1030                                         int new)
1031 {
1032         int ret;
1033         unsigned int map_from = 0, map_to = 0;
1034         unsigned int cluster_start, cluster_end;
1035         unsigned int user_data_from = 0, user_data_to = 0;
1036
1037         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1038                                         &cluster_start, &cluster_end);
1039
1040         /* treat the write as new if the a hole/lseek spanned across
1041          * the page boundary.
1042          */
1043         new = new | ((i_size_read(inode) <= page_offset(page)) &&
1044                         (page_offset(page) <= user_pos));
1045
1046         if (page == wc->w_target_page) {
1047                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1048                 map_to = map_from + user_len;
1049
1050                 if (new)
1051                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1052                                                     cluster_start, cluster_end,
1053                                                     new);
1054                 else
1055                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1056                                                     map_from, map_to, new);
1057                 if (ret) {
1058                         mlog_errno(ret);
1059                         goto out;
1060                 }
1061
1062                 user_data_from = map_from;
1063                 user_data_to = map_to;
1064                 if (new) {
1065                         map_from = cluster_start;
1066                         map_to = cluster_end;
1067                 }
1068         } else {
1069                 /*
1070                  * If we haven't allocated the new page yet, we
1071                  * shouldn't be writing it out without copying user
1072                  * data. This is likely a math error from the caller.
1073                  */
1074                 BUG_ON(!new);
1075
1076                 map_from = cluster_start;
1077                 map_to = cluster_end;
1078
1079                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1080                                             cluster_start, cluster_end, new);
1081                 if (ret) {
1082                         mlog_errno(ret);
1083                         goto out;
1084                 }
1085         }
1086
1087         /*
1088          * Parts of newly allocated pages need to be zero'd.
1089          *
1090          * Above, we have also rewritten 'to' and 'from' - as far as
1091          * the rest of the function is concerned, the entire cluster
1092          * range inside of a page needs to be written.
1093          *
1094          * We can skip this if the page is up to date - it's already
1095          * been zero'd from being read in as a hole.
1096          */
1097         if (new && !PageUptodate(page))
1098                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1099                                          cpos, user_data_from, user_data_to);
1100
1101         flush_dcache_page(page);
1102
1103 out:
1104         return ret;
1105 }
1106
1107 /*
1108  * This function will only grab one clusters worth of pages.
1109  */
1110 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1111                                       struct ocfs2_write_ctxt *wc,
1112                                       u32 cpos, loff_t user_pos,
1113                                       unsigned user_len, int new,
1114                                       struct page *mmap_page)
1115 {
1116         int ret = 0, i;
1117         unsigned long start, target_index, end_index, index;
1118         struct inode *inode = mapping->host;
1119         loff_t last_byte;
1120
1121         target_index = user_pos >> PAGE_CACHE_SHIFT;
1122
1123         /*
1124          * Figure out how many pages we'll be manipulating here. For
1125          * non allocating write, we just change the one
1126          * page. Otherwise, we'll need a whole clusters worth.  If we're
1127          * writing past i_size, we only need enough pages to cover the
1128          * last page of the write.
1129          */
1130         if (new) {
1131                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1132                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1133                 /*
1134                  * We need the index *past* the last page we could possibly
1135                  * touch.  This is the page past the end of the write or
1136                  * i_size, whichever is greater.
1137                  */
1138                 last_byte = max(user_pos + user_len, i_size_read(inode));
1139                 BUG_ON(last_byte < 1);
1140                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1141                 if ((start + wc->w_num_pages) > end_index)
1142                         wc->w_num_pages = end_index - start;
1143         } else {
1144                 wc->w_num_pages = 1;
1145                 start = target_index;
1146         }
1147
1148         for(i = 0; i < wc->w_num_pages; i++) {
1149                 index = start + i;
1150
1151                 if (index == target_index && mmap_page) {
1152                         /*
1153                          * ocfs2_pagemkwrite() is a little different
1154                          * and wants us to directly use the page
1155                          * passed in.
1156                          */
1157                         lock_page(mmap_page);
1158
1159                         /* Exit and let the caller retry */
1160                         if (mmap_page->mapping != mapping) {
1161                                 WARN_ON(mmap_page->mapping);
1162                                 unlock_page(mmap_page);
1163                                 ret = -EAGAIN;
1164                                 goto out;
1165                         }
1166
1167                         page_cache_get(mmap_page);
1168                         wc->w_pages[i] = mmap_page;
1169                         wc->w_target_locked = true;
1170                 } else {
1171                         wc->w_pages[i] = find_or_create_page(mapping, index,
1172                                                              GFP_NOFS);
1173                         if (!wc->w_pages[i]) {
1174                                 ret = -ENOMEM;
1175                                 mlog_errno(ret);
1176                                 goto out;
1177                         }
1178                 }
1179
1180                 if (index == target_index)
1181                         wc->w_target_page = wc->w_pages[i];
1182         }
1183 out:
1184         if (ret)
1185                 wc->w_target_locked = false;
1186         return ret;
1187 }
1188
1189 /*
1190  * Prepare a single cluster for write one cluster into the file.
1191  */
1192 static int ocfs2_write_cluster(struct address_space *mapping,
1193                                u32 phys, unsigned int unwritten,
1194                                unsigned int should_zero,
1195                                struct ocfs2_alloc_context *data_ac,
1196                                struct ocfs2_alloc_context *meta_ac,
1197                                struct ocfs2_write_ctxt *wc, u32 cpos,
1198                                loff_t user_pos, unsigned user_len)
1199 {
1200         int ret, i, new;
1201         u64 v_blkno, p_blkno;
1202         struct inode *inode = mapping->host;
1203         struct ocfs2_extent_tree et;
1204
1205         new = phys == 0 ? 1 : 0;
1206         if (new) {
1207                 u32 tmp_pos;
1208
1209                 /*
1210                  * This is safe to call with the page locks - it won't take
1211                  * any additional semaphores or cluster locks.
1212                  */
1213                 tmp_pos = cpos;
1214                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1215                                            &tmp_pos, 1, 0, wc->w_di_bh,
1216                                            wc->w_handle, data_ac,
1217                                            meta_ac, NULL);
1218                 /*
1219                  * This shouldn't happen because we must have already
1220                  * calculated the correct meta data allocation required. The
1221                  * internal tree allocation code should know how to increase
1222                  * transaction credits itself.
1223                  *
1224                  * If need be, we could handle -EAGAIN for a
1225                  * RESTART_TRANS here.
1226                  */
1227                 mlog_bug_on_msg(ret == -EAGAIN,
1228                                 "Inode %llu: EAGAIN return during allocation.\n",
1229                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1230                 if (ret < 0) {
1231                         mlog_errno(ret);
1232                         goto out;
1233                 }
1234         } else if (unwritten) {
1235                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1236                                               wc->w_di_bh);
1237                 ret = ocfs2_mark_extent_written(inode, &et,
1238                                                 wc->w_handle, cpos, 1, phys,
1239                                                 meta_ac, &wc->w_dealloc);
1240                 if (ret < 0) {
1241                         mlog_errno(ret);
1242                         goto out;
1243                 }
1244         }
1245
1246         if (should_zero)
1247                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1248         else
1249                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1250
1251         /*
1252          * The only reason this should fail is due to an inability to
1253          * find the extent added.
1254          */
1255         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1256                                           NULL);
1257         if (ret < 0) {
1258                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1259                             "at logical block %llu",
1260                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1261                             (unsigned long long)v_blkno);
1262                 goto out;
1263         }
1264
1265         BUG_ON(p_blkno == 0);
1266
1267         for(i = 0; i < wc->w_num_pages; i++) {
1268                 int tmpret;
1269
1270                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1271                                                       wc->w_pages[i], cpos,
1272                                                       user_pos, user_len,
1273                                                       should_zero);
1274                 if (tmpret) {
1275                         mlog_errno(tmpret);
1276                         if (ret == 0)
1277                                 ret = tmpret;
1278                 }
1279         }
1280
1281         /*
1282          * We only have cleanup to do in case of allocating write.
1283          */
1284         if (ret && new)
1285                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1286
1287 out:
1288
1289         return ret;
1290 }
1291
1292 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1293                                        struct ocfs2_alloc_context *data_ac,
1294                                        struct ocfs2_alloc_context *meta_ac,
1295                                        struct ocfs2_write_ctxt *wc,
1296                                        loff_t pos, unsigned len)
1297 {
1298         int ret, i;
1299         loff_t cluster_off;
1300         unsigned int local_len = len;
1301         struct ocfs2_write_cluster_desc *desc;
1302         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1303
1304         for (i = 0; i < wc->w_clen; i++) {
1305                 desc = &wc->w_desc[i];
1306
1307                 /*
1308                  * We have to make sure that the total write passed in
1309                  * doesn't extend past a single cluster.
1310                  */
1311                 local_len = len;
1312                 cluster_off = pos & (osb->s_clustersize - 1);
1313                 if ((cluster_off + local_len) > osb->s_clustersize)
1314                         local_len = osb->s_clustersize - cluster_off;
1315
1316                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1317                                           desc->c_unwritten,
1318                                           desc->c_needs_zero,
1319                                           data_ac, meta_ac,
1320                                           wc, desc->c_cpos, pos, local_len);
1321                 if (ret) {
1322                         mlog_errno(ret);
1323                         goto out;
1324                 }
1325
1326                 len -= local_len;
1327                 pos += local_len;
1328         }
1329
1330         ret = 0;
1331 out:
1332         return ret;
1333 }
1334
1335 /*
1336  * ocfs2_write_end() wants to know which parts of the target page it
1337  * should complete the write on. It's easiest to compute them ahead of
1338  * time when a more complete view of the write is available.
1339  */
1340 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1341                                         struct ocfs2_write_ctxt *wc,
1342                                         loff_t pos, unsigned len, int alloc)
1343 {
1344         struct ocfs2_write_cluster_desc *desc;
1345
1346         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1347         wc->w_target_to = wc->w_target_from + len;
1348
1349         if (alloc == 0)
1350                 return;
1351
1352         /*
1353          * Allocating write - we may have different boundaries based
1354          * on page size and cluster size.
1355          *
1356          * NOTE: We can no longer compute one value from the other as
1357          * the actual write length and user provided length may be
1358          * different.
1359          */
1360
1361         if (wc->w_large_pages) {
1362                 /*
1363                  * We only care about the 1st and last cluster within
1364                  * our range and whether they should be zero'd or not. Either
1365                  * value may be extended out to the start/end of a
1366                  * newly allocated cluster.
1367                  */
1368                 desc = &wc->w_desc[0];
1369                 if (desc->c_needs_zero)
1370                         ocfs2_figure_cluster_boundaries(osb,
1371                                                         desc->c_cpos,
1372                                                         &wc->w_target_from,
1373                                                         NULL);
1374
1375                 desc = &wc->w_desc[wc->w_clen - 1];
1376                 if (desc->c_needs_zero)
1377                         ocfs2_figure_cluster_boundaries(osb,
1378                                                         desc->c_cpos,
1379                                                         NULL,
1380                                                         &wc->w_target_to);
1381         } else {
1382                 wc->w_target_from = 0;
1383                 wc->w_target_to = PAGE_CACHE_SIZE;
1384         }
1385 }
1386
1387 /*
1388  * Populate each single-cluster write descriptor in the write context
1389  * with information about the i/o to be done.
1390  *
1391  * Returns the number of clusters that will have to be allocated, as
1392  * well as a worst case estimate of the number of extent records that
1393  * would have to be created during a write to an unwritten region.
1394  */
1395 static int ocfs2_populate_write_desc(struct inode *inode,
1396                                      struct ocfs2_write_ctxt *wc,
1397                                      unsigned int *clusters_to_alloc,
1398                                      unsigned int *extents_to_split)
1399 {
1400         int ret;
1401         struct ocfs2_write_cluster_desc *desc;
1402         unsigned int num_clusters = 0;
1403         unsigned int ext_flags = 0;
1404         u32 phys = 0;
1405         int i;
1406
1407         *clusters_to_alloc = 0;
1408         *extents_to_split = 0;
1409
1410         for (i = 0; i < wc->w_clen; i++) {
1411                 desc = &wc->w_desc[i];
1412                 desc->c_cpos = wc->w_cpos + i;
1413
1414                 if (num_clusters == 0) {
1415                         /*
1416                          * Need to look up the next extent record.
1417                          */
1418                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1419                                                  &num_clusters, &ext_flags);
1420                         if (ret) {
1421                                 mlog_errno(ret);
1422                                 goto out;
1423                         }
1424
1425                         /* We should already CoW the refcountd extent. */
1426                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1427
1428                         /*
1429                          * Assume worst case - that we're writing in
1430                          * the middle of the extent.
1431                          *
1432                          * We can assume that the write proceeds from
1433                          * left to right, in which case the extent
1434                          * insert code is smart enough to coalesce the
1435                          * next splits into the previous records created.
1436                          */
1437                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1438                                 *extents_to_split = *extents_to_split + 2;
1439                 } else if (phys) {
1440                         /*
1441                          * Only increment phys if it doesn't describe
1442                          * a hole.
1443                          */
1444                         phys++;
1445                 }
1446
1447                 /*
1448                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1449                  * file that got extended.  w_first_new_cpos tells us
1450                  * where the newly allocated clusters are so we can
1451                  * zero them.
1452                  */
1453                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1454                         BUG_ON(phys == 0);
1455                         desc->c_needs_zero = 1;
1456                 }
1457
1458                 desc->c_phys = phys;
1459                 if (phys == 0) {
1460                         desc->c_new = 1;
1461                         desc->c_needs_zero = 1;
1462                         *clusters_to_alloc = *clusters_to_alloc + 1;
1463                 }
1464
1465                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1466                         desc->c_unwritten = 1;
1467                         desc->c_needs_zero = 1;
1468                 }
1469
1470                 num_clusters--;
1471         }
1472
1473         ret = 0;
1474 out:
1475         return ret;
1476 }
1477
1478 static int ocfs2_write_begin_inline(struct address_space *mapping,
1479                                     struct inode *inode,
1480                                     struct ocfs2_write_ctxt *wc)
1481 {
1482         int ret;
1483         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1484         struct page *page;
1485         handle_t *handle;
1486         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1487
1488         page = find_or_create_page(mapping, 0, GFP_NOFS);
1489         if (!page) {
1490                 ret = -ENOMEM;
1491                 mlog_errno(ret);
1492                 goto out;
1493         }
1494         /*
1495          * If we don't set w_num_pages then this page won't get unlocked
1496          * and freed on cleanup of the write context.
1497          */
1498         wc->w_pages[0] = wc->w_target_page = page;
1499         wc->w_num_pages = 1;
1500
1501         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1502         if (IS_ERR(handle)) {
1503                 ret = PTR_ERR(handle);
1504                 mlog_errno(ret);
1505                 goto out;
1506         }
1507
1508         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1509                                       OCFS2_JOURNAL_ACCESS_WRITE);
1510         if (ret) {
1511                 ocfs2_commit_trans(osb, handle);
1512
1513                 mlog_errno(ret);
1514                 goto out;
1515         }
1516
1517         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1518                 ocfs2_set_inode_data_inline(inode, di);
1519
1520         if (!PageUptodate(page)) {
1521                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1522                 if (ret) {
1523                         ocfs2_commit_trans(osb, handle);
1524
1525                         goto out;
1526                 }
1527         }
1528
1529         wc->w_handle = handle;
1530 out:
1531         return ret;
1532 }
1533
1534 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1535 {
1536         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1537
1538         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1539                 return 1;
1540         return 0;
1541 }
1542
1543 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1544                                           struct inode *inode, loff_t pos,
1545                                           unsigned len, struct page *mmap_page,
1546                                           struct ocfs2_write_ctxt *wc)
1547 {
1548         int ret, written = 0;
1549         loff_t end = pos + len;
1550         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1551         struct ocfs2_dinode *di = NULL;
1552
1553         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1554                                              len, (unsigned long long)pos,
1555                                              oi->ip_dyn_features);
1556
1557         /*
1558          * Handle inodes which already have inline data 1st.
1559          */
1560         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1561                 if (mmap_page == NULL &&
1562                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1563                         goto do_inline_write;
1564
1565                 /*
1566                  * The write won't fit - we have to give this inode an
1567                  * inline extent list now.
1568                  */
1569                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1570                 if (ret)
1571                         mlog_errno(ret);
1572                 goto out;
1573         }
1574
1575         /*
1576          * Check whether the inode can accept inline data.
1577          */
1578         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1579                 return 0;
1580
1581         /*
1582          * Check whether the write can fit.
1583          */
1584         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1585         if (mmap_page ||
1586             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1587                 return 0;
1588
1589 do_inline_write:
1590         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1591         if (ret) {
1592                 mlog_errno(ret);
1593                 goto out;
1594         }
1595
1596         /*
1597          * This signals to the caller that the data can be written
1598          * inline.
1599          */
1600         written = 1;
1601 out:
1602         return written ? written : ret;
1603 }
1604
1605 /*
1606  * This function only does anything for file systems which can't
1607  * handle sparse files.
1608  *
1609  * What we want to do here is fill in any hole between the current end
1610  * of allocation and the end of our write. That way the rest of the
1611  * write path can treat it as an non-allocating write, which has no
1612  * special case code for sparse/nonsparse files.
1613  */
1614 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1615                                         struct buffer_head *di_bh,
1616                                         loff_t pos, unsigned len,
1617                                         struct ocfs2_write_ctxt *wc)
1618 {
1619         int ret;
1620         loff_t newsize = pos + len;
1621
1622         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1623
1624         if (newsize <= i_size_read(inode))
1625                 return 0;
1626
1627         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1628         if (ret)
1629                 mlog_errno(ret);
1630
1631         wc->w_first_new_cpos =
1632                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1633
1634         return ret;
1635 }
1636
1637 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1638                            loff_t pos)
1639 {
1640         int ret = 0;
1641
1642         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1643         if (pos > i_size_read(inode))
1644                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1645
1646         return ret;
1647 }
1648
1649 /*
1650  * Try to flush truncate logs if we can free enough clusters from it.
1651  * As for return value, "< 0" means error, "0" no space and "1" means
1652  * we have freed enough spaces and let the caller try to allocate again.
1653  */
1654 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1655                                           unsigned int needed)
1656 {
1657         tid_t target;
1658         int ret = 0;
1659         unsigned int truncated_clusters;
1660
1661         mutex_lock(&osb->osb_tl_inode->i_mutex);
1662         truncated_clusters = osb->truncated_clusters;
1663         mutex_unlock(&osb->osb_tl_inode->i_mutex);
1664
1665         /*
1666          * Check whether we can succeed in allocating if we free
1667          * the truncate log.
1668          */
1669         if (truncated_clusters < needed)
1670                 goto out;
1671
1672         ret = ocfs2_flush_truncate_log(osb);
1673         if (ret) {
1674                 mlog_errno(ret);
1675                 goto out;
1676         }
1677
1678         if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1679                 jbd2_log_wait_commit(osb->journal->j_journal, target);
1680                 ret = 1;
1681         }
1682 out:
1683         return ret;
1684 }
1685
1686 int ocfs2_write_begin_nolock(struct file *filp,
1687                              struct address_space *mapping,
1688                              loff_t pos, unsigned len, unsigned flags,
1689                              struct page **pagep, void **fsdata,
1690                              struct buffer_head *di_bh, struct page *mmap_page)
1691 {
1692         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1693         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1694         struct ocfs2_write_ctxt *wc;
1695         struct inode *inode = mapping->host;
1696         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1697         struct ocfs2_dinode *di;
1698         struct ocfs2_alloc_context *data_ac = NULL;
1699         struct ocfs2_alloc_context *meta_ac = NULL;
1700         handle_t *handle;
1701         struct ocfs2_extent_tree et;
1702         int try_free = 1, ret1;
1703
1704 try_again:
1705         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1706         if (ret) {
1707                 mlog_errno(ret);
1708                 return ret;
1709         }
1710
1711         if (ocfs2_supports_inline_data(osb)) {
1712                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1713                                                      mmap_page, wc);
1714                 if (ret == 1) {
1715                         ret = 0;
1716                         goto success;
1717                 }
1718                 if (ret < 0) {
1719                         mlog_errno(ret);
1720                         goto out;
1721                 }
1722         }
1723
1724         if (ocfs2_sparse_alloc(osb))
1725                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1726         else
1727                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1728                                                    wc);
1729         if (ret) {
1730                 mlog_errno(ret);
1731                 goto out;
1732         }
1733
1734         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1735         if (ret < 0) {
1736                 mlog_errno(ret);
1737                 goto out;
1738         } else if (ret == 1) {
1739                 clusters_need = wc->w_clen;
1740                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1741                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1742                 if (ret) {
1743                         mlog_errno(ret);
1744                         goto out;
1745                 }
1746         }
1747
1748         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1749                                         &extents_to_split);
1750         if (ret) {
1751                 mlog_errno(ret);
1752                 goto out;
1753         }
1754         clusters_need += clusters_to_alloc;
1755
1756         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1757
1758         trace_ocfs2_write_begin_nolock(
1759                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1760                         (long long)i_size_read(inode),
1761                         le32_to_cpu(di->i_clusters),
1762                         pos, len, flags, mmap_page,
1763                         clusters_to_alloc, extents_to_split);
1764
1765         /*
1766          * We set w_target_from, w_target_to here so that
1767          * ocfs2_write_end() knows which range in the target page to
1768          * write out. An allocation requires that we write the entire
1769          * cluster range.
1770          */
1771         if (clusters_to_alloc || extents_to_split) {
1772                 /*
1773                  * XXX: We are stretching the limits of
1774                  * ocfs2_lock_allocators(). It greatly over-estimates
1775                  * the work to be done.
1776                  */
1777                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1778                                               wc->w_di_bh);
1779                 ret = ocfs2_lock_allocators(inode, &et,
1780                                             clusters_to_alloc, extents_to_split,
1781                                             &data_ac, &meta_ac);
1782                 if (ret) {
1783                         mlog_errno(ret);
1784                         goto out;
1785                 }
1786
1787                 if (data_ac)
1788                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1789
1790                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1791                                                     &di->id2.i_list,
1792                                                     clusters_to_alloc);
1793
1794         }
1795
1796         /*
1797          * We have to zero sparse allocated clusters, unwritten extent clusters,
1798          * and non-sparse clusters we just extended.  For non-sparse writes,
1799          * we know zeros will only be needed in the first and/or last cluster.
1800          */
1801         if (clusters_to_alloc || extents_to_split ||
1802             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1803                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1804                 cluster_of_pages = 1;
1805         else
1806                 cluster_of_pages = 0;
1807
1808         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1809
1810         handle = ocfs2_start_trans(osb, credits);
1811         if (IS_ERR(handle)) {
1812                 ret = PTR_ERR(handle);
1813                 mlog_errno(ret);
1814                 goto out;
1815         }
1816
1817         wc->w_handle = handle;
1818
1819         if (clusters_to_alloc) {
1820                 ret = dquot_alloc_space_nodirty(inode,
1821                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1822                 if (ret)
1823                         goto out_commit;
1824         }
1825         /*
1826          * We don't want this to fail in ocfs2_write_end(), so do it
1827          * here.
1828          */
1829         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1830                                       OCFS2_JOURNAL_ACCESS_WRITE);
1831         if (ret) {
1832                 mlog_errno(ret);
1833                 goto out_quota;
1834         }
1835
1836         /*
1837          * Fill our page array first. That way we've grabbed enough so
1838          * that we can zero and flush if we error after adding the
1839          * extent.
1840          */
1841         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1842                                          cluster_of_pages, mmap_page);
1843         if (ret && ret != -EAGAIN) {
1844                 mlog_errno(ret);
1845                 goto out_quota;
1846         }
1847
1848         /*
1849          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1850          * the target page. In this case, we exit with no error and no target
1851          * page. This will trigger the caller, page_mkwrite(), to re-try
1852          * the operation.
1853          */
1854         if (ret == -EAGAIN) {
1855                 BUG_ON(wc->w_target_page);
1856                 ret = 0;
1857                 goto out_quota;
1858         }
1859
1860         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1861                                           len);
1862         if (ret) {
1863                 mlog_errno(ret);
1864                 goto out_quota;
1865         }
1866
1867         if (data_ac)
1868                 ocfs2_free_alloc_context(data_ac);
1869         if (meta_ac)
1870                 ocfs2_free_alloc_context(meta_ac);
1871
1872 success:
1873         *pagep = wc->w_target_page;
1874         *fsdata = wc;
1875         return 0;
1876 out_quota:
1877         if (clusters_to_alloc)
1878                 dquot_free_space(inode,
1879                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1880 out_commit:
1881         ocfs2_commit_trans(osb, handle);
1882
1883 out:
1884         ocfs2_free_write_ctxt(wc);
1885
1886         if (data_ac)
1887                 ocfs2_free_alloc_context(data_ac);
1888         if (meta_ac)
1889                 ocfs2_free_alloc_context(meta_ac);
1890
1891         if (ret == -ENOSPC && try_free) {
1892                 /*
1893                  * Try to free some truncate log so that we can have enough
1894                  * clusters to allocate.
1895                  */
1896                 try_free = 0;
1897
1898                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1899                 if (ret1 == 1)
1900                         goto try_again;
1901
1902                 if (ret1 < 0)
1903                         mlog_errno(ret1);
1904         }
1905
1906         return ret;
1907 }
1908
1909 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1910                              loff_t pos, unsigned len, unsigned flags,
1911                              struct page **pagep, void **fsdata)
1912 {
1913         int ret;
1914         struct buffer_head *di_bh = NULL;
1915         struct inode *inode = mapping->host;
1916
1917         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1918         if (ret) {
1919                 mlog_errno(ret);
1920                 return ret;
1921         }
1922
1923         /*
1924          * Take alloc sem here to prevent concurrent lookups. That way
1925          * the mapping, zeroing and tree manipulation within
1926          * ocfs2_write() will be safe against ->readpage(). This
1927          * should also serve to lock out allocation from a shared
1928          * writeable region.
1929          */
1930         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1931
1932         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1933                                        fsdata, di_bh, NULL);
1934         if (ret) {
1935                 mlog_errno(ret);
1936                 goto out_fail;
1937         }
1938
1939         brelse(di_bh);
1940
1941         return 0;
1942
1943 out_fail:
1944         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1945
1946         brelse(di_bh);
1947         ocfs2_inode_unlock(inode, 1);
1948
1949         return ret;
1950 }
1951
1952 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1953                                    unsigned len, unsigned *copied,
1954                                    struct ocfs2_dinode *di,
1955                                    struct ocfs2_write_ctxt *wc)
1956 {
1957         void *kaddr;
1958
1959         if (unlikely(*copied < len)) {
1960                 if (!PageUptodate(wc->w_target_page)) {
1961                         *copied = 0;
1962                         return;
1963                 }
1964         }
1965
1966         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1967         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1968         kunmap_atomic(kaddr, KM_USER0);
1969
1970         trace_ocfs2_write_end_inline(
1971              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1972              (unsigned long long)pos, *copied,
1973              le16_to_cpu(di->id2.i_data.id_count),
1974              le16_to_cpu(di->i_dyn_features));
1975 }
1976
1977 int ocfs2_write_end_nolock(struct address_space *mapping,
1978                            loff_t pos, unsigned len, unsigned copied,
1979                            struct page *page, void *fsdata)
1980 {
1981         int i;
1982         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1983         struct inode *inode = mapping->host;
1984         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1985         struct ocfs2_write_ctxt *wc = fsdata;
1986         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1987         handle_t *handle = wc->w_handle;
1988         struct page *tmppage;
1989
1990         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1991                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1992                 goto out_write_size;
1993         }
1994
1995         if (unlikely(copied < len)) {
1996                 if (!PageUptodate(wc->w_target_page))
1997                         copied = 0;
1998
1999                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2000                                        start+len);
2001         }
2002         flush_dcache_page(wc->w_target_page);
2003
2004         for(i = 0; i < wc->w_num_pages; i++) {
2005                 tmppage = wc->w_pages[i];
2006
2007                 if (tmppage == wc->w_target_page) {
2008                         from = wc->w_target_from;
2009                         to = wc->w_target_to;
2010
2011                         BUG_ON(from > PAGE_CACHE_SIZE ||
2012                                to > PAGE_CACHE_SIZE ||
2013                                to < from);
2014                 } else {
2015                         /*
2016                          * Pages adjacent to the target (if any) imply
2017                          * a hole-filling write in which case we want
2018                          * to flush their entire range.
2019                          */
2020                         from = 0;
2021                         to = PAGE_CACHE_SIZE;
2022                 }
2023
2024                 if (page_has_buffers(tmppage)) {
2025                         if (ocfs2_should_order_data(inode))
2026                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2027                         block_commit_write(tmppage, from, to);
2028                 }
2029         }
2030
2031 out_write_size:
2032         pos += copied;
2033         if (pos > inode->i_size) {
2034                 i_size_write(inode, pos);
2035                 mark_inode_dirty(inode);
2036         }
2037         inode->i_blocks = ocfs2_inode_sector_count(inode);
2038         di->i_size = cpu_to_le64((u64)i_size_read(inode));
2039         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2040         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2041         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2042         ocfs2_journal_dirty(handle, wc->w_di_bh);
2043
2044         ocfs2_commit_trans(osb, handle);
2045
2046         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2047
2048         ocfs2_free_write_ctxt(wc);
2049
2050         return copied;
2051 }
2052
2053 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2054                            loff_t pos, unsigned len, unsigned copied,
2055                            struct page *page, void *fsdata)
2056 {
2057         int ret;
2058         struct inode *inode = mapping->host;
2059
2060         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2061
2062         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2063         ocfs2_inode_unlock(inode, 1);
2064
2065         return ret;
2066 }
2067
2068 const struct address_space_operations ocfs2_aops = {
2069         .readpage               = ocfs2_readpage,
2070         .readpages              = ocfs2_readpages,
2071         .writepage              = ocfs2_writepage,
2072         .write_begin            = ocfs2_write_begin,
2073         .write_end              = ocfs2_write_end,
2074         .bmap                   = ocfs2_bmap,
2075         .direct_IO              = ocfs2_direct_IO,
2076         .invalidatepage         = ocfs2_invalidatepage,
2077         .releasepage            = ocfs2_releasepage,
2078         .migratepage            = buffer_migrate_page,
2079         .is_partially_uptodate  = block_is_partially_uptodate,
2080         .error_remove_page      = generic_error_remove_page,
2081 };