fs/ocfs2/aops.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public
  17  * License along with this program; if not, write to the
  18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  19  * Boston, MA 021110-1307, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/slab.h>
  24 #include <linux/highmem.h>
  25 #include <linux/pagemap.h>
  26 #include <asm/byteorder.h>
  27 #include <linux/swap.h>
  28 #include <linux/pipe_fs_i.h>
  29 #include <linux/mpage.h>
  30 #include <linux/quotaops.h>
  31
  32 #define MLOG_MASK_PREFIX ML_FILE_IO
  33 #include <cluster/masklog.h>
  34
  35 #include "ocfs2.h"
  36
  37 #include "alloc.h"
  38 #include "aops.h"
  39 #include "dlmglue.h"
  40 #include "extent_map.h"
  41 #include "file.h"
  42 #include "inode.h"
  43 #include "journal.h"
  44 #include "suballoc.h"
  45 #include "super.h"
  46 #include "symlink.h"
  47 #include "refcounttree.h"
  48
  49 #include "buffer_head_io.h"
  50
  51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
  52                                    struct buffer_head *bh_result, int create)
  53 {
  54         int err = -EIO;
  55         int status;
  56         struct ocfs2_dinode *fe = NULL;
  57         struct buffer_head *bh = NULL;
  58         struct buffer_head *buffer_cache_bh = NULL;
  59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  60         void *kaddr;
  61
  62         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
  63                    (unsigned long long)iblock, bh_result, create);
  64
  65         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
  66
  67         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
  68                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
  69                      (unsigned long long)iblock);
  70                 goto bail;
  71         }
  72
  73         status = ocfs2_read_inode_block(inode, &bh);
  74         if (status < 0) {
  75                 mlog_errno(status);
  76                 goto bail;
  77         }
  78         fe = (struct ocfs2_dinode *) bh->b_data;
  79
  80         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
  81                                                     le32_to_cpu(fe->i_clusters))) {
  82                 mlog(ML_ERROR, "block offset is outside the allocated size: "
  83                      "%llu\n", (unsigned long long)iblock);
  84                 goto bail;
  85         }
  86
  87         /* We don't use the page cache to create symlink data, so if
  88          * need be, copy it over from the buffer cache. */
  89         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
  90                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
  91                             iblock;
  92                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
  93                 if (!buffer_cache_bh) {
  94                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
  95                         goto bail;
  96                 }
  97
  98                 /* we haven't locked out transactions, so a commit
  99                  * could've happened. Since we've got a reference on
 100                  * the bh, even if it commits while we're doing the
 101                  * copy, the data is still good. */
 102                 if (buffer_jbd(buffer_cache_bh)
 103                     && ocfs2_inode_is_new(inode)) {
 104                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
 105                         if (!kaddr) {
 106                                 mlog(ML_ERROR, "couldn't kmap!\n");
 107                                 goto bail;
 108                         }
 109                         memcpy(kaddr + (bh_result->b_size * iblock),
 110                                buffer_cache_bh->b_data,
 111                                bh_result->b_size);
 112                         kunmap_atomic(kaddr, KM_USER0);
 113                         set_buffer_uptodate(bh_result);
 114                 }
 115                 brelse(buffer_cache_bh);
 116         }
 117
 118         map_bh(bh_result, inode->i_sb,
 119                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
 120
 121         err = 0;
 122
 123 bail:
 124         brelse(bh);
 125
 126         mlog_exit(err);
 127         return err;
 128 }
 129
 130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
 131                     struct buffer_head *bh_result, int create)
 132 {
 133         int err = 0;
 134         unsigned int ext_flags;
 135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
 136         u64 p_blkno, count, past_eof;
 137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 138
 139         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
 140                    (unsigned long long)iblock, bh_result, create);
 141
 142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
 143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
 144                      inode, inode->i_ino);
 145
 146         if (S_ISLNK(inode->i_mode)) {
 147                 /* this always does I/O for some reason. */
 148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
 149                 goto bail;
 150         }
 151
 152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
 153                                           &ext_flags);
 154         if (err) {
 155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
 156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
 157                      (unsigned long long)p_blkno);
 158                 goto bail;
 159         }
 160
 161         if (max_blocks < count)
 162                 count = max_blocks;
 163
 164         /*
 165          * ocfs2 never allocates in this function - the only time we
 166          * need to use BH_New is when we're extending i_size on a file
 167          * system which doesn't support holes, in which case BH_New
 168          * allows block_prepare_write() to zero.
 169          *
 170          * If we see this on a sparse file system, then a truncate has
 171          * raced us and removed the cluster. In this case, we clear
 172          * the buffers dirty and uptodate bits and let the buffer code
 173          * ignore it as a hole.
 174          */
 175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
 176                 clear_buffer_dirty(bh_result);
 177                 clear_buffer_uptodate(bh_result);
 178                 goto bail;
 179         }
 180
 181         /* Treat the unwritten extent as a hole for zeroing purposes. */
 182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 183                 map_bh(bh_result, inode->i_sb, p_blkno);
 184
 185         bh_result->b_size = count << inode->i_blkbits;
 186
 187         if (!ocfs2_sparse_alloc(osb)) {
 188                 if (p_blkno == 0) {
 189                         err = -EIO;
 190                         mlog(ML_ERROR,
 191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
 192                              (unsigned long long)iblock,
 193                              (unsigned long long)p_blkno,
 194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
 195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
 196                         dump_stack();
 197                         goto bail;
 198                 }
 199         }
 200
 201         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 202         mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
 203              (unsigned long long)past_eof);
 204         if (create && (iblock >= past_eof))
 205                 set_buffer_new(bh_result);
 206
 207 bail:
 208         if (err < 0)
 209                 err = -EIO;
 210
 211         mlog_exit(err);
 212         return err;
 213 }
 214
 215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
 216                            struct buffer_head *di_bh)
 217 {
 218         void *kaddr;
 219         loff_t size;
 220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 221
 222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
 223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
 224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
 225                 return -EROFS;
 226         }
 227
 228         size = i_size_read(inode);
 229
 230         if (size > PAGE_CACHE_SIZE ||
 231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
 232                 ocfs2_error(inode->i_sb,
 233                             "Inode %llu has with inline data has bad size: %Lu",
 234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
 235                             (unsigned long long)size);
 236                 return -EROFS;
 237         }
 238
 239         kaddr = kmap_atomic(page, KM_USER0);
 240         if (size)
 241                 memcpy(kaddr, di->id2.i_data.id_data, size);
 242         /* Clear the remaining part of the page */
 243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
 244         flush_dcache_page(page);
 245         kunmap_atomic(kaddr, KM_USER0);
 246
 247         SetPageUptodate(page);
 248
 249         return 0;
 250 }
 251
 252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
 253 {
 254         int ret;
 255         struct buffer_head *di_bh = NULL;
 256
 257         BUG_ON(!PageLocked(page));
 258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
 259
 260         ret = ocfs2_read_inode_block(inode, &di_bh);
 261         if (ret) {
 262                 mlog_errno(ret);
 263                 goto out;
 264         }
 265
 266         ret = ocfs2_read_inline_data(inode, page, di_bh);
 267 out:
 268         unlock_page(page);
 269
 270         brelse(di_bh);
 271         return ret;
 272 }
 273
 274 static int ocfs2_readpage(struct file *file, struct page *page)
 275 {
 276         struct inode *inode = page->mapping->host;
 277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
 279         int ret, unlock = 1;
 280
 281         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
 282
 283         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
 284         if (ret != 0) {
 285                 if (ret == AOP_TRUNCATED_PAGE)
 286                         unlock = 0;
 287                 mlog_errno(ret);
 288                 goto out;
 289         }
 290
 291         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 292                 ret = AOP_TRUNCATED_PAGE;
 293                 goto out_inode_unlock;
 294         }
 295
 296         /*
 297          * i_size might have just been updated as we grabed the meta lock.  We
 298          * might now be discovering a truncate that hit on another node.
 299          * block_read_full_page->get_block freaks out if it is asked to read
 300          * beyond the end of a file, so we check here.  Callers
 301          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
 302          * and notice that the page they just read isn't needed.
 303          *
 304          * XXX sys_readahead() seems to get that wrong?
 305          */
 306         if (start >= i_size_read(inode)) {
 307                 zero_user(page, 0, PAGE_SIZE);
 308                 SetPageUptodate(page);
 309                 ret = 0;
 310                 goto out_alloc;
 311         }
 312
 313         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 314                 ret = ocfs2_readpage_inline(inode, page);
 315         else
 316                 ret = block_read_full_page(page, ocfs2_get_block);
 317         unlock = 0;
 318
 319 out_alloc:
 320         up_read(&OCFS2_I(inode)->ip_alloc_sem);
 321 out_inode_unlock:
 322         ocfs2_inode_unlock(inode, 0);
 323 out:
 324         if (unlock)
 325                 unlock_page(page);
 326         mlog_exit(ret);
 327         return ret;
 328 }
 329
 330 /*
 331  * This is used only for read-ahead. Failures or difficult to handle
 332  * situations are safe to ignore.
 333  *
 334  * Right now, we don't bother with BH_Boundary - in-inode extent lists
 335  * are quite large (243 extents on 4k blocks), so most inodes don't
 336  * grow out to a tree. If need be, detecting boundary extents could
 337  * trivially be added in a future version of ocfs2_get_block().
 338  */
 339 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
 340                            struct list_head *pages, unsigned nr_pages)
 341 {
 342         int ret, err = -EIO;
 343         struct inode *inode = mapping->host;
 344         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 345         loff_t start;
 346         struct page *last;
 347
 348         /*
 349          * Use the nonblocking flag for the dlm code to avoid page
 350          * lock inversion, but don't bother with retrying.
 351          */
 352         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
 353         if (ret)
 354                 return err;
 355
 356         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
 357                 ocfs2_inode_unlock(inode, 0);
 358                 return err;
 359         }
 360
 361         /*
 362          * Don't bother with inline-data. There isn't anything
 363          * to read-ahead in that case anyway...
 364          */
 365         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 366                 goto out_unlock;
 367
 368         /*
 369          * Check whether a remote node truncated this file - we just
 370          * drop out in that case as it's not worth handling here.
 371          */
 372         last = list_entry(pages->prev, struct page, lru);
 373         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
 374         if (start >= i_size_read(inode))
 375                 goto out_unlock;
 376
 377         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
 378
 379 out_unlock:
 380         up_read(&oi->ip_alloc_sem);
 381         ocfs2_inode_unlock(inode, 0);
 382
 383         return err;
 384 }
 385
 386 /* Note: Because we don't support holes, our allocation has
 387  * already happened (allocation writes zeros to the file data)
 388  * so we don't have to worry about ordered writes in
 389  * ocfs2_writepage.
 390  *
 391  * ->writepage is called during the process of invalidating the page cache
 392  * during blocked lock processing.  It can't block on any cluster locks
 393  * to during block mapping.  It's relying on the fact that the block
 394  * mapping can't have disappeared under the dirty pages that it is
 395  * being asked to write back.
 396  */
 397 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
 398 {
 399         int ret;
 400
 401         mlog_entry("(0x%p)\n", page);
 402
 403         ret = block_write_full_page(page, ocfs2_get_block, wbc);
 404
 405         mlog_exit(ret);
 406
 407         return ret;
 408 }
 409
 410 /*
 411  * This is called from ocfs2_write_zero_page() which has handled it's
 412  * own cluster locking and has ensured allocation exists for those
 413  * blocks to be written.
 414  */
 415 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
 416                                unsigned from, unsigned to)
 417 {
 418         int ret;
 419
 420         ret = block_prepare_write(page, from, to, ocfs2_get_block);
 421
 422         return ret;
 423 }
 424
 425 /* Taken from ext3. We don't necessarily need the full blown
 426  * functionality yet, but IMHO it's better to cut and paste the whole
 427  * thing so we can avoid introducing our own bugs (and easily pick up
 428  * their fixes when they happen) --Mark */
 429 int walk_page_buffers(  handle_t *handle,
 430                         struct buffer_head *head,
 431                         unsigned from,
 432                         unsigned to,
 433                         int *partial,
 434                         int (*fn)(      handle_t *handle,
 435                                         struct buffer_head *bh))
 436 {
 437         struct buffer_head *bh;
 438         unsigned block_start, block_end;
 439         unsigned blocksize = head->b_size;
 440         int err, ret = 0;
 441         struct buffer_head *next;
 442
 443         for (   bh = head, block_start = 0;
 444                 ret == 0 && (bh != head || !block_start);
 445                 block_start = block_end, bh = next)
 446         {
 447                 next = bh->b_this_page;
 448                 block_end = block_start + blocksize;
 449                 if (block_end <= from || block_start >= to) {
 450                         if (partial && !buffer_uptodate(bh))
 451                                 *partial = 1;
 452                         continue;
 453                 }
 454                 err = (*fn)(handle, bh);
 455                 if (!ret)
 456                         ret = err;
 457         }
 458         return ret;
 459 }
 460
 461 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
 462 {
 463         sector_t status;
 464         u64 p_blkno = 0;
 465         int err = 0;
 466         struct inode *inode = mapping->host;
 467
 468         mlog_entry("(block = %llu)\n", (unsigned long long)block);
 469
 470         /* We don't need to lock journal system files, since they aren't
 471          * accessed concurrently from multiple nodes.
 472          */
 473         if (!INODE_JOURNAL(inode)) {
 474                 err = ocfs2_inode_lock(inode, NULL, 0);
 475                 if (err) {
 476                         if (err != -ENOENT)
 477                                 mlog_errno(err);
 478                         goto bail;
 479                 }
 480                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
 481         }
 482
 483         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
 484                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
 485                                                   NULL);
 486
 487         if (!INODE_JOURNAL(inode)) {
 488                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
 489                 ocfs2_inode_unlock(inode, 0);
 490         }
 491
 492         if (err) {
 493                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
 494                      (unsigned long long)block);
 495                 mlog_errno(err);
 496                 goto bail;
 497         }
 498
 499 bail:
 500         status = err ? 0 : p_blkno;
 501
 502         mlog_exit((int)status);
 503
 504         return status;
 505 }
 506
 507 /*
 508  * TODO: Make this into a generic get_blocks function.
 509  *
 510  * From do_direct_io in direct-io.c:
 511  *  "So what we do is to permit the ->get_blocks function to populate
 512  *   bh.b_size with the size of IO which is permitted at this offset and
 513  *   this i_blkbits."
 514  *
 515  * This function is called directly from get_more_blocks in direct-io.c.
 516  *
 517  * called like this: dio->get_blocks(dio->inode, fs_startblk,
 518  *                                      fs_count, map_bh, dio->rw == WRITE);
 519  *
 520  * Note that we never bother to allocate blocks here, and thus ignore the
 521  * create argument.
 522  */
 523 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 524                                      struct buffer_head *bh_result, int create)
 525 {
 526         int ret;
 527         u64 p_blkno, inode_blocks, contig_blocks;
 528         unsigned int ext_flags;
 529         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
 530         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
 531
 532         /* This function won't even be called if the request isn't all
 533          * nicely aligned and of the right size, so there's no need
 534          * for us to check any of that. */
 535
 536         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
 537
 538         /* This figures out the size of the next contiguous block, and
 539          * our logical offset */
 540         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
 541                                           &contig_blocks, &ext_flags);
 542         if (ret) {
 543                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
 544                      (unsigned long long)iblock);
 545                 ret = -EIO;
 546                 goto bail;
 547         }
 548
 549         /* We should already CoW the refcounted extent in case of create. */
 550         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
 551
 552         /*
 553          * get_more_blocks() expects us to describe a hole by clearing
 554          * the mapped bit on bh_result().
 555          *
 556          * Consider an unwritten extent as a hole.
 557          */
 558         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
 559                 map_bh(bh_result, inode->i_sb, p_blkno);
 560         else
 561                 clear_buffer_mapped(bh_result);
 562
 563         /* make sure we don't map more than max_blocks blocks here as
 564            that's all the kernel will handle at this point. */
 565         if (max_blocks < contig_blocks)
 566                 contig_blocks = max_blocks;
 567         bh_result->b_size = contig_blocks << blocksize_bits;
 568 bail:
 569         return ret;
 570 }
 571
 572 /*
 573  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
 574  * particularly interested in the aio/dio case.  Like the core uses
 575  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
 576  * truncation on another.
 577  */
 578 static void ocfs2_dio_end_io(struct kiocb *iocb,
 579                              loff_t offset,
 580                              ssize_t bytes,
 581                              void *private)
 582 {
 583         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
 584         int level;
 585
 586         /* this io's submitter should not have unlocked this before we could */
 587         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 588
 589         ocfs2_iocb_clear_rw_locked(iocb);
 590
 591         level = ocfs2_iocb_rw_locked_level(iocb);
 592         if (!level)
 593                 up_read(&inode->i_alloc_sem);
 594         ocfs2_rw_unlock(inode, level);
 595 }
 596
 597 /*
 598  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
 599  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
 600  * do journalled data.
 601  */
 602 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
 603 {
 604         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 605
 606         jbd2_journal_invalidatepage(journal, page, offset);
 607 }
 608
 609 static int ocfs2_releasepage(struct page *page, gfp_t wait)
 610 {
 611         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
 612
 613         if (!page_has_buffers(page))
 614                 return 0;
 615         return jbd2_journal_try_to_free_buffers(journal, page, wait);
 616 }
 617
 618 static ssize_t ocfs2_direct_IO(int rw,
 619                                struct kiocb *iocb,
 620                                const struct iovec *iov,
 621                                loff_t offset,
 622                                unsigned long nr_segs)
 623 {
 624         struct file *file = iocb->ki_filp;
 625         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
 626         int ret;
 627
 628         mlog_entry_void();
 629
 630         /*
 631          * Fallback to buffered I/O if we see an inode without
 632          * extents.
 633          */
 634         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 635                 return 0;
 636
 637         /* Fallback to buffered I/O if we are appending. */
 638         if (i_size_read(inode) <= offset)
 639                 return 0;
 640
 641         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
 642                                             inode->i_sb->s_bdev, iov, offset,
 643                                             nr_segs,
 644                                             ocfs2_direct_IO_get_blocks,
 645                                             ocfs2_dio_end_io);
 646
 647         mlog_exit(ret);
 648         return ret;
 649 }
 650
 651 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
 652                                             u32 cpos,
 653                                             unsigned int *start,
 654                                             unsigned int *end)
 655 {
 656         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
 657
 658         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
 659                 unsigned int cpp;
 660
 661                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
 662
 663                 cluster_start = cpos % cpp;
 664                 cluster_start = cluster_start << osb->s_clustersize_bits;
 665
 666                 cluster_end = cluster_start + osb->s_clustersize;
 667         }
 668
 669         BUG_ON(cluster_start > PAGE_SIZE);
 670         BUG_ON(cluster_end > PAGE_SIZE);
 671
 672         if (start)
 673                 *start = cluster_start;
 674         if (end)
 675                 *end = cluster_end;
 676 }
 677
 678 /*
 679  * 'from' and 'to' are the region in the page to avoid zeroing.
 680  *
 681  * If pagesize > clustersize, this function will avoid zeroing outside
 682  * of the cluster boundary.
 683  *
 684  * from == to == 0 is code for "zero the entire cluster region"
 685  */
 686 static void ocfs2_clear_page_regions(struct page *page,
 687                                      struct ocfs2_super *osb, u32 cpos,
 688                                      unsigned from, unsigned to)
 689 {
 690         void *kaddr;
 691         unsigned int cluster_start, cluster_end;
 692
 693         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
 694
 695         kaddr = kmap_atomic(page, KM_USER0);
 696
 697         if (from || to) {
 698                 if (from > cluster_start)
 699                         memset(kaddr + cluster_start, 0, from - cluster_start);
 700                 if (to < cluster_end)
 701                         memset(kaddr + to, 0, cluster_end - to);
 702         } else {
 703                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
 704         }
 705
 706         kunmap_atomic(kaddr, KM_USER0);
 707 }
 708
 709 /*
 710  * Nonsparse file systems fully allocate before we get to the write
 711  * code. This prevents ocfs2_write() from tagging the write as an
 712  * allocating one, which means ocfs2_map_page_blocks() might try to
 713  * read-in the blocks at the tail of our file. Avoid reading them by
 714  * testing i_size against each block offset.
 715  */
 716 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
 717                                  unsigned int block_start)
 718 {
 719         u64 offset = page_offset(page) + block_start;
 720
 721         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 722                 return 1;
 723
 724         if (i_size_read(inode) > offset)
 725                 return 1;
 726
 727         return 0;
 728 }
 729
 730 /*
 731  * Some of this taken from block_prepare_write(). We already have our
 732  * mapping by now though, and the entire write will be allocating or
 733  * it won't, so not much need to use BH_New.
 734  *
 735  * This will also skip zeroing, which is handled externally.
 736  */
 737 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
 738                           struct inode *inode, unsigned int from,
 739                           unsigned int to, int new)
 740 {
 741         int ret = 0;
 742         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
 743         unsigned int block_end, block_start;
 744         unsigned int bsize = 1 << inode->i_blkbits;
 745
 746         if (!page_has_buffers(page))
 747                 create_empty_buffers(page, bsize, 0);
 748
 749         head = page_buffers(page);
 750         for (bh = head, block_start = 0; bh != head || !block_start;
 751              bh = bh->b_this_page, block_start += bsize) {
 752                 block_end = block_start + bsize;
 753
 754                 clear_buffer_new(bh);
 755
 756                 /*
 757                  * Ignore blocks outside of our i/o range -
 758                  * they may belong to unallocated clusters.
 759                  */
 760                 if (block_start >= to || block_end <= from) {
 761                         if (PageUptodate(page))
 762                                 set_buffer_uptodate(bh);
 763                         continue;
 764                 }
 765
 766                 /*
 767                  * For an allocating write with cluster size >= page
 768                  * size, we always write the entire page.
 769                  */
 770                 if (new)
 771                         set_buffer_new(bh);
 772
 773                 if (!buffer_mapped(bh)) {
 774                         map_bh(bh, inode->i_sb, *p_blkno);
 775                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 776                 }
 777
 778                 if (PageUptodate(page)) {
 779                         if (!buffer_uptodate(bh))
 780                                 set_buffer_uptodate(bh);
 781                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
 782                            !buffer_new(bh) &&
 783                            ocfs2_should_read_blk(inode, page, block_start) &&
 784                            (block_start < from || block_end > to)) {
 785                         ll_rw_block(READ, 1, &bh);
 786                         *wait_bh++=bh;
 787                 }
 788
 789                 *p_blkno = *p_blkno + 1;
 790         }
 791
 792         /*
 793          * If we issued read requests - let them complete.
 794          */
 795         while(wait_bh > wait) {
 796                 wait_on_buffer(*--wait_bh);
 797                 if (!buffer_uptodate(*wait_bh))
 798                         ret = -EIO;
 799         }
 800
 801         if (ret == 0 || !new)
 802                 return ret;
 803
 804         /*
 805          * If we get -EIO above, zero out any newly allocated blocks
 806          * to avoid exposing stale data.
 807          */
 808         bh = head;
 809         block_start = 0;
 810         do {
 811                 block_end = block_start + bsize;
 812                 if (block_end <= from)
 813                         goto next_bh;
 814                 if (block_start >= to)
 815                         break;
 816
 817                 zero_user(page, block_start, bh->b_size);
 818                 set_buffer_uptodate(bh);
 819                 mark_buffer_dirty(bh);
 820
 821 next_bh:
 822                 block_start = block_end;
 823                 bh = bh->b_this_page;
 824         } while (bh != head);
 825
 826         return ret;
 827 }
 828
 829 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
 830 #define OCFS2_MAX_CTXT_PAGES    1
 831 #else
 832 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
 833 #endif
 834
 835 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
 836
 837 /*
 838  * Describe the state of a single cluster to be written to.
 839  */
 840 struct ocfs2_write_cluster_desc {
 841         u32             c_cpos;
 842         u32             c_phys;
 843         /*
 844          * Give this a unique field because c_phys eventually gets
 845          * filled.
 846          */
 847         unsigned        c_new;
 848         unsigned        c_unwritten;
 849         unsigned        c_needs_zero;
 850 };
 851
 852 struct ocfs2_write_ctxt {
 853         /* Logical cluster position / len of write */
 854         u32                             w_cpos;
 855         u32                             w_clen;
 856
 857         /* First cluster allocated in a nonsparse extend */
 858         u32                             w_first_new_cpos;
 859
 860         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
 861
 862         /*
 863          * This is true if page_size > cluster_size.
 864          *
 865          * It triggers a set of special cases during write which might
 866          * have to deal with allocating writes to partial pages.
 867          */
 868         unsigned int                    w_large_pages;
 869
 870         /*
 871          * Pages involved in this write.
 872          *
 873          * w_target_page is the page being written to by the user.
 874          *
 875          * w_pages is an array of pages which always contains
 876          * w_target_page, and in the case of an allocating write with
 877          * page_size < cluster size, it will contain zero'd and mapped
 878          * pages adjacent to w_target_page which need to be written
 879          * out in so that future reads from that region will get
 880          * zero's.
 881          */
 882         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
 883         unsigned int                    w_num_pages;
 884         struct page                     *w_target_page;
 885
 886         /*
 887          * ocfs2_write_end() uses this to know what the real range to
 888          * write in the target should be.
 889          */
 890         unsigned int                    w_target_from;
 891         unsigned int                    w_target_to;
 892
 893         /*
 894          * We could use journal_current_handle() but this is cleaner,
 895          * IMHO -Mark
 896          */
 897         handle_t                        *w_handle;
 898
 899         struct buffer_head              *w_di_bh;
 900
 901         struct ocfs2_cached_dealloc_ctxt w_dealloc;
 902 };
 903
 904 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
 905 {
 906         int i;
 907
 908         for(i = 0; i < num_pages; i++) {
 909                 if (pages[i]) {
 910                         unlock_page(pages[i]);
 911                         mark_page_accessed(pages[i]);
 912                         page_cache_release(pages[i]);
 913                 }
 914         }
 915 }
 916
 917 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
 918 {
 919         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
 920
 921         brelse(wc->w_di_bh);
 922         kfree(wc);
 923 }
 924
 925 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
 926                                   struct ocfs2_super *osb, loff_t pos,
 927                                   unsigned len, struct buffer_head *di_bh)
 928 {
 929         u32 cend;
 930         struct ocfs2_write_ctxt *wc;
 931
 932         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
 933         if (!wc)
 934                 return -ENOMEM;
 935
 936         wc->w_cpos = pos >> osb->s_clustersize_bits;
 937         wc->w_first_new_cpos = UINT_MAX;
 938         cend = (pos + len - 1) >> osb->s_clustersize_bits;
 939         wc->w_clen = cend - wc->w_cpos + 1;
 940         get_bh(di_bh);
 941         wc->w_di_bh = di_bh;
 942
 943         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
 944                 wc->w_large_pages = 1;
 945         else
 946                 wc->w_large_pages = 0;
 947
 948         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
 949
 950         *wcp = wc;
 951
 952         return 0;
 953 }
 954
 955 /*
 956  * If a page has any new buffers, zero them out here, and mark them uptodate
 957  * and dirty so they'll be written out (in order to prevent uninitialised
 958  * block data from leaking). And clear the new bit.
 959  */
 960 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
 961 {
 962         unsigned int block_start, block_end;
 963         struct buffer_head *head, *bh;
 964
 965         BUG_ON(!PageLocked(page));
 966         if (!page_has_buffers(page))
 967                 return;
 968
 969         bh = head = page_buffers(page);
 970         block_start = 0;
 971         do {
 972                 block_end = block_start + bh->b_size;
 973
 974                 if (buffer_new(bh)) {
 975                         if (block_end > from && block_start < to) {
 976                                 if (!PageUptodate(page)) {
 977                                         unsigned start, end;
 978
 979                                         start = max(from, block_start);
 980                                         end = min(to, block_end);
 981
 982                                         zero_user_segment(page, start, end);
 983                                         set_buffer_uptodate(bh);
 984                                 }
 985
 986                                 clear_buffer_new(bh);
 987                                 mark_buffer_dirty(bh);
 988                         }
 989                 }
 990
 991                 block_start = block_end;
 992                 bh = bh->b_this_page;
 993         } while (bh != head);
 994 }
 995
 996 /*
 997  * Only called when we have a failure during allocating write to write
 998  * zero's to the newly allocated region.
 999  */
1000 static void ocfs2_write_failure(struct inode *inode,
1001                                 struct ocfs2_write_ctxt *wc,
1002                                 loff_t user_pos, unsigned user_len)
1003 {
1004         int i;
1005         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1006                 to = user_pos + user_len;
1007         struct page *tmppage;
1008
1009         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1010
1011         for(i = 0; i < wc->w_num_pages; i++) {
1012                 tmppage = wc->w_pages[i];
1013
1014                 if (page_has_buffers(tmppage)) {
1015                         if (ocfs2_should_order_data(inode))
1016                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1017
1018                         block_commit_write(tmppage, from, to);
1019                 }
1020         }
1021 }
1022
1023 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1024                                         struct ocfs2_write_ctxt *wc,
1025                                         struct page *page, u32 cpos,
1026                                         loff_t user_pos, unsigned user_len,
1027                                         int new)
1028 {
1029         int ret;
1030         unsigned int map_from = 0, map_to = 0;
1031         unsigned int cluster_start, cluster_end;
1032         unsigned int user_data_from = 0, user_data_to = 0;
1033
1034         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1035                                         &cluster_start, &cluster_end);
1036
1037         if (page == wc->w_target_page) {
1038                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1039                 map_to = map_from + user_len;
1040
1041                 if (new)
1042                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1043                                                     cluster_start, cluster_end,
1044                                                     new);
1045                 else
1046                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1047                                                     map_from, map_to, new);
1048                 if (ret) {
1049                         mlog_errno(ret);
1050                         goto out;
1051                 }
1052
1053                 user_data_from = map_from;
1054                 user_data_to = map_to;
1055                 if (new) {
1056                         map_from = cluster_start;
1057                         map_to = cluster_end;
1058                 }
1059         } else {
1060                 /*
1061                  * If we haven't allocated the new page yet, we
1062                  * shouldn't be writing it out without copying user
1063                  * data. This is likely a math error from the caller.
1064                  */
1065                 BUG_ON(!new);
1066
1067                 map_from = cluster_start;
1068                 map_to = cluster_end;
1069
1070                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1071                                             cluster_start, cluster_end, new);
1072                 if (ret) {
1073                         mlog_errno(ret);
1074                         goto out;
1075                 }
1076         }
1077
1078         /*
1079          * Parts of newly allocated pages need to be zero'd.
1080          *
1081          * Above, we have also rewritten 'to' and 'from' - as far as
1082          * the rest of the function is concerned, the entire cluster
1083          * range inside of a page needs to be written.
1084          *
1085          * We can skip this if the page is up to date - it's already
1086          * been zero'd from being read in as a hole.
1087          */
1088         if (new && !PageUptodate(page))
1089                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1090                                          cpos, user_data_from, user_data_to);
1091
1092         flush_dcache_page(page);
1093
1094 out:
1095         return ret;
1096 }
1097
1098 /*
1099  * This function will only grab one clusters worth of pages.
1100  */
1101 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1102                                       struct ocfs2_write_ctxt *wc,
1103                                       u32 cpos, loff_t user_pos,
1104                                       unsigned user_len, int new,
1105                                       struct page *mmap_page)
1106 {
1107         int ret = 0, i;
1108         unsigned long start, target_index, end_index, index;
1109         struct inode *inode = mapping->host;
1110         loff_t last_byte;
1111
1112         target_index = user_pos >> PAGE_CACHE_SHIFT;
1113
1114         /*
1115          * Figure out how many pages we'll be manipulating here. For
1116          * non allocating write, we just change the one
1117          * page. Otherwise, we'll need a whole clusters worth.  If we're
1118          * writing past i_size, we only need enough pages to cover the
1119          * last page of the write.
1120          */
1121         if (new) {
1122                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1123                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1124                 /*
1125                  * We need the index *past* the last page we could possibly
1126                  * touch.  This is the page past the end of the write or
1127                  * i_size, whichever is greater.
1128                  */
1129                 last_byte = max(user_pos + user_len, i_size_read(inode));
1130                 BUG_ON(last_byte < 1);
1131                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1132                 if ((start + wc->w_num_pages) > end_index)
1133                         wc->w_num_pages = end_index - start;
1134         } else {
1135                 wc->w_num_pages = 1;
1136                 start = target_index;
1137         }
1138
1139         for(i = 0; i < wc->w_num_pages; i++) {
1140                 index = start + i;
1141
1142                 if (index == target_index && mmap_page) {
1143                         /*
1144                          * ocfs2_pagemkwrite() is a little different
1145                          * and wants us to directly use the page
1146                          * passed in.
1147                          */
1148                         lock_page(mmap_page);
1149
1150                         if (mmap_page->mapping != mapping) {
1151                                 unlock_page(mmap_page);
1152                                 /*
1153                                  * Sanity check - the locking in
1154                                  * ocfs2_pagemkwrite() should ensure
1155                                  * that this code doesn't trigger.
1156                                  */
1157                                 ret = -EINVAL;
1158                                 mlog_errno(ret);
1159                                 goto out;
1160                         }
1161
1162                         page_cache_get(mmap_page);
1163                         wc->w_pages[i] = mmap_page;
1164                 } else {
1165                         wc->w_pages[i] = find_or_create_page(mapping, index,
1166                                                              GFP_NOFS);
1167                         if (!wc->w_pages[i]) {
1168                                 ret = -ENOMEM;
1169                                 mlog_errno(ret);
1170                                 goto out;
1171                         }
1172                 }
1173
1174                 if (index == target_index)
1175                         wc->w_target_page = wc->w_pages[i];
1176         }
1177 out:
1178         return ret;
1179 }
1180
1181 /*
1182  * Prepare a single cluster for write one cluster into the file.
1183  */
1184 static int ocfs2_write_cluster(struct address_space *mapping,
1185                                u32 phys, unsigned int unwritten,
1186                                unsigned int should_zero,
1187                                struct ocfs2_alloc_context *data_ac,
1188                                struct ocfs2_alloc_context *meta_ac,
1189                                struct ocfs2_write_ctxt *wc, u32 cpos,
1190                                loff_t user_pos, unsigned user_len)
1191 {
1192         int ret, i, new;
1193         u64 v_blkno, p_blkno;
1194         struct inode *inode = mapping->host;
1195         struct ocfs2_extent_tree et;
1196
1197         new = phys == 0 ? 1 : 0;
1198         if (new) {
1199                 u32 tmp_pos;
1200
1201                 /*
1202                  * This is safe to call with the page locks - it won't take
1203                  * any additional semaphores or cluster locks.
1204                  */
1205                 tmp_pos = cpos;
1206                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1207                                            &tmp_pos, 1, 0, wc->w_di_bh,
1208                                            wc->w_handle, data_ac,
1209                                            meta_ac, NULL);
1210                 /*
1211                  * This shouldn't happen because we must have already
1212                  * calculated the correct meta data allocation required. The
1213                  * internal tree allocation code should know how to increase
1214                  * transaction credits itself.
1215                  *
1216                  * If need be, we could handle -EAGAIN for a
1217                  * RESTART_TRANS here.
1218                  */
1219                 mlog_bug_on_msg(ret == -EAGAIN,
1220                                 "Inode %llu: EAGAIN return during allocation.\n",
1221                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1222                 if (ret < 0) {
1223                         mlog_errno(ret);
1224                         goto out;
1225                 }
1226         } else if (unwritten) {
1227                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1228                                               wc->w_di_bh);
1229                 ret = ocfs2_mark_extent_written(inode, &et,
1230                                                 wc->w_handle, cpos, 1, phys,
1231                                                 meta_ac, &wc->w_dealloc);
1232                 if (ret < 0) {
1233                         mlog_errno(ret);
1234                         goto out;
1235                 }
1236         }
1237
1238         if (should_zero)
1239                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1240         else
1241                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1242
1243         /*
1244          * The only reason this should fail is due to an inability to
1245          * find the extent added.
1246          */
1247         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1248                                           NULL);
1249         if (ret < 0) {
1250                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1251                             "at logical block %llu",
1252                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1253                             (unsigned long long)v_blkno);
1254                 goto out;
1255         }
1256
1257         BUG_ON(p_blkno == 0);
1258
1259         for(i = 0; i < wc->w_num_pages; i++) {
1260                 int tmpret;
1261
1262                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1263                                                       wc->w_pages[i], cpos,
1264                                                       user_pos, user_len,
1265                                                       should_zero);
1266                 if (tmpret) {
1267                         mlog_errno(tmpret);
1268                         if (ret == 0)
1269                                 ret = tmpret;
1270                 }
1271         }
1272
1273         /*
1274          * We only have cleanup to do in case of allocating write.
1275          */
1276         if (ret && new)
1277                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1278
1279 out:
1280
1281         return ret;
1282 }
1283
1284 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1285                                        struct ocfs2_alloc_context *data_ac,
1286                                        struct ocfs2_alloc_context *meta_ac,
1287                                        struct ocfs2_write_ctxt *wc,
1288                                        loff_t pos, unsigned len)
1289 {
1290         int ret, i;
1291         loff_t cluster_off;
1292         unsigned int local_len = len;
1293         struct ocfs2_write_cluster_desc *desc;
1294         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1295
1296         for (i = 0; i < wc->w_clen; i++) {
1297                 desc = &wc->w_desc[i];
1298
1299                 /*
1300                  * We have to make sure that the total write passed in
1301                  * doesn't extend past a single cluster.
1302                  */
1303                 local_len = len;
1304                 cluster_off = pos & (osb->s_clustersize - 1);
1305                 if ((cluster_off + local_len) > osb->s_clustersize)
1306                         local_len = osb->s_clustersize - cluster_off;
1307
1308                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1309                                           desc->c_unwritten,
1310                                           desc->c_needs_zero,
1311                                           data_ac, meta_ac,
1312                                           wc, desc->c_cpos, pos, local_len);
1313                 if (ret) {
1314                         mlog_errno(ret);
1315                         goto out;
1316                 }
1317
1318                 len -= local_len;
1319                 pos += local_len;
1320         }
1321
1322         ret = 0;
1323 out:
1324         return ret;
1325 }
1326
1327 /*
1328  * ocfs2_write_end() wants to know which parts of the target page it
1329  * should complete the write on. It's easiest to compute them ahead of
1330  * time when a more complete view of the write is available.
1331  */
1332 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1333                                         struct ocfs2_write_ctxt *wc,
1334                                         loff_t pos, unsigned len, int alloc)
1335 {
1336         struct ocfs2_write_cluster_desc *desc;
1337
1338         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1339         wc->w_target_to = wc->w_target_from + len;
1340
1341         if (alloc == 0)
1342                 return;
1343
1344         /*
1345          * Allocating write - we may have different boundaries based
1346          * on page size and cluster size.
1347          *
1348          * NOTE: We can no longer compute one value from the other as
1349          * the actual write length and user provided length may be
1350          * different.
1351          */
1352
1353         if (wc->w_large_pages) {
1354                 /*
1355                  * We only care about the 1st and last cluster within
1356                  * our range and whether they should be zero'd or not. Either
1357                  * value may be extended out to the start/end of a
1358                  * newly allocated cluster.
1359                  */
1360                 desc = &wc->w_desc[0];
1361                 if (desc->c_needs_zero)
1362                         ocfs2_figure_cluster_boundaries(osb,
1363                                                         desc->c_cpos,
1364                                                         &wc->w_target_from,
1365                                                         NULL);
1366
1367                 desc = &wc->w_desc[wc->w_clen - 1];
1368                 if (desc->c_needs_zero)
1369                         ocfs2_figure_cluster_boundaries(osb,
1370                                                         desc->c_cpos,
1371                                                         NULL,
1372                                                         &wc->w_target_to);
1373         } else {
1374                 wc->w_target_from = 0;
1375                 wc->w_target_to = PAGE_CACHE_SIZE;
1376         }
1377 }
1378
1379 /*
1380  * Populate each single-cluster write descriptor in the write context
1381  * with information about the i/o to be done.
1382  *
1383  * Returns the number of clusters that will have to be allocated, as
1384  * well as a worst case estimate of the number of extent records that
1385  * would have to be created during a write to an unwritten region.
1386  */
1387 static int ocfs2_populate_write_desc(struct inode *inode,
1388                                      struct ocfs2_write_ctxt *wc,
1389                                      unsigned int *clusters_to_alloc,
1390                                      unsigned int *extents_to_split)
1391 {
1392         int ret;
1393         struct ocfs2_write_cluster_desc *desc;
1394         unsigned int num_clusters = 0;
1395         unsigned int ext_flags = 0;
1396         u32 phys = 0;
1397         int i;
1398
1399         *clusters_to_alloc = 0;
1400         *extents_to_split = 0;
1401
1402         for (i = 0; i < wc->w_clen; i++) {
1403                 desc = &wc->w_desc[i];
1404                 desc->c_cpos = wc->w_cpos + i;
1405
1406                 if (num_clusters == 0) {
1407                         /*
1408                          * Need to look up the next extent record.
1409                          */
1410                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1411                                                  &num_clusters, &ext_flags);
1412                         if (ret) {
1413                                 mlog_errno(ret);
1414                                 goto out;
1415                         }
1416
1417                         /* We should already CoW the refcountd extent. */
1418                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1419
1420                         /*
1421                          * Assume worst case - that we're writing in
1422                          * the middle of the extent.
1423                          *
1424                          * We can assume that the write proceeds from
1425                          * left to right, in which case the extent
1426                          * insert code is smart enough to coalesce the
1427                          * next splits into the previous records created.
1428                          */
1429                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1430                                 *extents_to_split = *extents_to_split + 2;
1431                 } else if (phys) {
1432                         /*
1433                          * Only increment phys if it doesn't describe
1434                          * a hole.
1435                          */
1436                         phys++;
1437                 }
1438
1439                 /*
1440                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1441                  * file that got extended.  w_first_new_cpos tells us
1442                  * where the newly allocated clusters are so we can
1443                  * zero them.
1444                  */
1445                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1446                         BUG_ON(phys == 0);
1447                         desc->c_needs_zero = 1;
1448                 }
1449
1450                 desc->c_phys = phys;
1451                 if (phys == 0) {
1452                         desc->c_new = 1;
1453                         desc->c_needs_zero = 1;
1454                         *clusters_to_alloc = *clusters_to_alloc + 1;
1455                 }
1456
1457                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1458                         desc->c_unwritten = 1;
1459                         desc->c_needs_zero = 1;
1460                 }
1461
1462                 num_clusters--;
1463         }
1464
1465         ret = 0;
1466 out:
1467         return ret;
1468 }
1469
1470 static int ocfs2_write_begin_inline(struct address_space *mapping,
1471                                     struct inode *inode,
1472                                     struct ocfs2_write_ctxt *wc)
1473 {
1474         int ret;
1475         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1476         struct page *page;
1477         handle_t *handle;
1478         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1479
1480         page = find_or_create_page(mapping, 0, GFP_NOFS);
1481         if (!page) {
1482                 ret = -ENOMEM;
1483                 mlog_errno(ret);
1484                 goto out;
1485         }
1486         /*
1487          * If we don't set w_num_pages then this page won't get unlocked
1488          * and freed on cleanup of the write context.
1489          */
1490         wc->w_pages[0] = wc->w_target_page = page;
1491         wc->w_num_pages = 1;
1492
1493         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1494         if (IS_ERR(handle)) {
1495                 ret = PTR_ERR(handle);
1496                 mlog_errno(ret);
1497                 goto out;
1498         }
1499
1500         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1501                                       OCFS2_JOURNAL_ACCESS_WRITE);
1502         if (ret) {
1503                 ocfs2_commit_trans(osb, handle);
1504
1505                 mlog_errno(ret);
1506                 goto out;
1507         }
1508
1509         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1510                 ocfs2_set_inode_data_inline(inode, di);
1511
1512         if (!PageUptodate(page)) {
1513                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1514                 if (ret) {
1515                         ocfs2_commit_trans(osb, handle);
1516
1517                         goto out;
1518                 }
1519         }
1520
1521         wc->w_handle = handle;
1522 out:
1523         return ret;
1524 }
1525
1526 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1527 {
1528         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1529
1530         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1531                 return 1;
1532         return 0;
1533 }
1534
1535 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1536                                           struct inode *inode, loff_t pos,
1537                                           unsigned len, struct page *mmap_page,
1538                                           struct ocfs2_write_ctxt *wc)
1539 {
1540         int ret, written = 0;
1541         loff_t end = pos + len;
1542         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1543         struct ocfs2_dinode *di = NULL;
1544
1545         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1546              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1547              oi->ip_dyn_features);
1548
1549         /*
1550          * Handle inodes which already have inline data 1st.
1551          */
1552         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1553                 if (mmap_page == NULL &&
1554                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1555                         goto do_inline_write;
1556
1557                 /*
1558                  * The write won't fit - we have to give this inode an
1559                  * inline extent list now.
1560                  */
1561                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1562                 if (ret)
1563                         mlog_errno(ret);
1564                 goto out;
1565         }
1566
1567         /*
1568          * Check whether the inode can accept inline data.
1569          */
1570         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1571                 return 0;
1572
1573         /*
1574          * Check whether the write can fit.
1575          */
1576         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1577         if (mmap_page ||
1578             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1579                 return 0;
1580
1581 do_inline_write:
1582         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1583         if (ret) {
1584                 mlog_errno(ret);
1585                 goto out;
1586         }
1587
1588         /*
1589          * This signals to the caller that the data can be written
1590          * inline.
1591          */
1592         written = 1;
1593 out:
1594         return written ? written : ret;
1595 }
1596
1597 /*
1598  * This function only does anything for file systems which can't
1599  * handle sparse files.
1600  *
1601  * What we want to do here is fill in any hole between the current end
1602  * of allocation and the end of our write. That way the rest of the
1603  * write path can treat it as an non-allocating write, which has no
1604  * special case code for sparse/nonsparse files.
1605  */
1606 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1607                                         struct buffer_head *di_bh,
1608                                         loff_t pos, unsigned len,
1609                                         struct ocfs2_write_ctxt *wc)
1610 {
1611         int ret;
1612         loff_t newsize = pos + len;
1613
1614         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1615
1616         if (newsize <= i_size_read(inode))
1617                 return 0;
1618
1619         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1620         if (ret)
1621                 mlog_errno(ret);
1622
1623         wc->w_first_new_cpos =
1624                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1625
1626         return ret;
1627 }
1628
1629 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1630                            loff_t pos)
1631 {
1632         int ret = 0;
1633
1634         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1635         if (pos > i_size_read(inode))
1636                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1637
1638         return ret;
1639 }
1640
1641 int ocfs2_write_begin_nolock(struct address_space *mapping,
1642                              loff_t pos, unsigned len, unsigned flags,
1643                              struct page **pagep, void **fsdata,
1644                              struct buffer_head *di_bh, struct page *mmap_page)
1645 {
1646         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1647         unsigned int clusters_to_alloc, extents_to_split;
1648         struct ocfs2_write_ctxt *wc;
1649         struct inode *inode = mapping->host;
1650         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1651         struct ocfs2_dinode *di;
1652         struct ocfs2_alloc_context *data_ac = NULL;
1653         struct ocfs2_alloc_context *meta_ac = NULL;
1654         handle_t *handle;
1655         struct ocfs2_extent_tree et;
1656
1657         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1658         if (ret) {
1659                 mlog_errno(ret);
1660                 return ret;
1661         }
1662
1663         if (ocfs2_supports_inline_data(osb)) {
1664                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1665                                                      mmap_page, wc);
1666                 if (ret == 1) {
1667                         ret = 0;
1668                         goto success;
1669                 }
1670                 if (ret < 0) {
1671                         mlog_errno(ret);
1672                         goto out;
1673                 }
1674         }
1675
1676         if (ocfs2_sparse_alloc(osb))
1677                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1678         else
1679                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1680                                                    wc);
1681         if (ret) {
1682                 mlog_errno(ret);
1683                 goto out;
1684         }
1685
1686         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1687         if (ret < 0) {
1688                 mlog_errno(ret);
1689                 goto out;
1690         } else if (ret == 1) {
1691                 ret = ocfs2_refcount_cow(inode, di_bh,
1692                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1693                 if (ret) {
1694                         mlog_errno(ret);
1695                         goto out;
1696                 }
1697         }
1698
1699         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1700                                         &extents_to_split);
1701         if (ret) {
1702                 mlog_errno(ret);
1703                 goto out;
1704         }
1705
1706         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1707
1708         /*
1709          * We set w_target_from, w_target_to here so that
1710          * ocfs2_write_end() knows which range in the target page to
1711          * write out. An allocation requires that we write the entire
1712          * cluster range.
1713          */
1714         if (clusters_to_alloc || extents_to_split) {
1715                 /*
1716                  * XXX: We are stretching the limits of
1717                  * ocfs2_lock_allocators(). It greatly over-estimates
1718                  * the work to be done.
1719                  */
1720                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1721                      " clusters_to_add = %u, extents_to_split = %u\n",
1722                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1723                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1724                      clusters_to_alloc, extents_to_split);
1725
1726                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1727                                               wc->w_di_bh);
1728                 ret = ocfs2_lock_allocators(inode, &et,
1729                                             clusters_to_alloc, extents_to_split,
1730                                             &data_ac, &meta_ac);
1731                 if (ret) {
1732                         mlog_errno(ret);
1733                         goto out;
1734                 }
1735
1736                 if (data_ac)
1737                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1738
1739                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1740                                                     &di->id2.i_list,
1741                                                     clusters_to_alloc);
1742
1743         }
1744
1745         /*
1746          * We have to zero sparse allocated clusters, unwritten extent clusters,
1747          * and non-sparse clusters we just extended.  For non-sparse writes,
1748          * we know zeros will only be needed in the first and/or last cluster.
1749          */
1750         if (clusters_to_alloc || extents_to_split ||
1751             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1752                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1753                 cluster_of_pages = 1;
1754         else
1755                 cluster_of_pages = 0;
1756
1757         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1758
1759         handle = ocfs2_start_trans(osb, credits);
1760         if (IS_ERR(handle)) {
1761                 ret = PTR_ERR(handle);
1762                 mlog_errno(ret);
1763                 goto out;
1764         }
1765
1766         wc->w_handle = handle;
1767
1768         if (clusters_to_alloc) {
1769                 ret = dquot_alloc_space_nodirty(inode,
1770                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1771                 if (ret)
1772                         goto out_commit;
1773         }
1774         /*
1775          * We don't want this to fail in ocfs2_write_end(), so do it
1776          * here.
1777          */
1778         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1779                                       OCFS2_JOURNAL_ACCESS_WRITE);
1780         if (ret) {
1781                 mlog_errno(ret);
1782                 goto out_quota;
1783         }
1784
1785         /*
1786          * Fill our page array first. That way we've grabbed enough so
1787          * that we can zero and flush if we error after adding the
1788          * extent.
1789          */
1790         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1791                                          cluster_of_pages, mmap_page);
1792         if (ret) {
1793                 mlog_errno(ret);
1794                 goto out_quota;
1795         }
1796
1797         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1798                                           len);
1799         if (ret) {
1800                 mlog_errno(ret);
1801                 goto out_quota;
1802         }
1803
1804         if (data_ac)
1805                 ocfs2_free_alloc_context(data_ac);
1806         if (meta_ac)
1807                 ocfs2_free_alloc_context(meta_ac);
1808
1809 success:
1810         *pagep = wc->w_target_page;
1811         *fsdata = wc;
1812         return 0;
1813 out_quota:
1814         if (clusters_to_alloc)
1815                 dquot_free_space(inode,
1816                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1817 out_commit:
1818         ocfs2_commit_trans(osb, handle);
1819
1820 out:
1821         ocfs2_free_write_ctxt(wc);
1822
1823         if (data_ac)
1824                 ocfs2_free_alloc_context(data_ac);
1825         if (meta_ac)
1826                 ocfs2_free_alloc_context(meta_ac);
1827         return ret;
1828 }
1829
1830 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1831                              loff_t pos, unsigned len, unsigned flags,
1832                              struct page **pagep, void **fsdata)
1833 {
1834         int ret;
1835         struct buffer_head *di_bh = NULL;
1836         struct inode *inode = mapping->host;
1837
1838         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1839         if (ret) {
1840                 mlog_errno(ret);
1841                 return ret;
1842         }
1843
1844         /*
1845          * Take alloc sem here to prevent concurrent lookups. That way
1846          * the mapping, zeroing and tree manipulation within
1847          * ocfs2_write() will be safe against ->readpage(). This
1848          * should also serve to lock out allocation from a shared
1849          * writeable region.
1850          */
1851         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1852
1853         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1854                                        fsdata, di_bh, NULL);
1855         if (ret) {
1856                 mlog_errno(ret);
1857                 goto out_fail;
1858         }
1859
1860         brelse(di_bh);
1861
1862         return 0;
1863
1864 out_fail:
1865         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1866
1867         brelse(di_bh);
1868         ocfs2_inode_unlock(inode, 1);
1869
1870         return ret;
1871 }
1872
1873 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1874                                    unsigned len, unsigned *copied,
1875                                    struct ocfs2_dinode *di,
1876                                    struct ocfs2_write_ctxt *wc)
1877 {
1878         void *kaddr;
1879
1880         if (unlikely(*copied < len)) {
1881                 if (!PageUptodate(wc->w_target_page)) {
1882                         *copied = 0;
1883                         return;
1884                 }
1885         }
1886
1887         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1888         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1889         kunmap_atomic(kaddr, KM_USER0);
1890
1891         mlog(0, "Data written to inode at offset %llu. "
1892              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1893              (unsigned long long)pos, *copied,
1894              le16_to_cpu(di->id2.i_data.id_count),
1895              le16_to_cpu(di->i_dyn_features));
1896 }
1897
1898 int ocfs2_write_end_nolock(struct address_space *mapping,
1899                            loff_t pos, unsigned len, unsigned copied,
1900                            struct page *page, void *fsdata)
1901 {
1902         int i;
1903         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1904         struct inode *inode = mapping->host;
1905         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1906         struct ocfs2_write_ctxt *wc = fsdata;
1907         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1908         handle_t *handle = wc->w_handle;
1909         struct page *tmppage;
1910
1911         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1912                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1913                 goto out_write_size;
1914         }
1915
1916         if (unlikely(copied < len)) {
1917                 if (!PageUptodate(wc->w_target_page))
1918                         copied = 0;
1919
1920                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1921                                        start+len);
1922         }
1923         flush_dcache_page(wc->w_target_page);
1924
1925         for(i = 0; i < wc->w_num_pages; i++) {
1926                 tmppage = wc->w_pages[i];
1927
1928                 if (tmppage == wc->w_target_page) {
1929                         from = wc->w_target_from;
1930                         to = wc->w_target_to;
1931
1932                         BUG_ON(from > PAGE_CACHE_SIZE ||
1933                                to > PAGE_CACHE_SIZE ||
1934                                to < from);
1935                 } else {
1936                         /*
1937                          * Pages adjacent to the target (if any) imply
1938                          * a hole-filling write in which case we want
1939                          * to flush their entire range.
1940                          */
1941                         from = 0;
1942                         to = PAGE_CACHE_SIZE;
1943                 }
1944
1945                 if (page_has_buffers(tmppage)) {
1946                         if (ocfs2_should_order_data(inode))
1947                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1948                         block_commit_write(tmppage, from, to);
1949                 }
1950         }
1951
1952 out_write_size:
1953         pos += copied;
1954         if (pos > inode->i_size) {
1955                 i_size_write(inode, pos);
1956                 mark_inode_dirty(inode);
1957         }
1958         inode->i_blocks = ocfs2_inode_sector_count(inode);
1959         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1960         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1961         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1962         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1963         ocfs2_journal_dirty(handle, wc->w_di_bh);
1964
1965         ocfs2_commit_trans(osb, handle);
1966
1967         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1968
1969         ocfs2_free_write_ctxt(wc);
1970
1971         return copied;
1972 }
1973
1974 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1975                            loff_t pos, unsigned len, unsigned copied,
1976                            struct page *page, void *fsdata)
1977 {
1978         int ret;
1979         struct inode *inode = mapping->host;
1980
1981         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1982
1983         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1984         ocfs2_inode_unlock(inode, 1);
1985
1986         return ret;
1987 }
1988
1989 const struct address_space_operations ocfs2_aops = {
1990         .readpage               = ocfs2_readpage,
1991         .readpages              = ocfs2_readpages,
1992         .writepage              = ocfs2_writepage,
1993         .write_begin            = ocfs2_write_begin,
1994         .write_end              = ocfs2_write_end,
1995         .bmap                   = ocfs2_bmap,
1996         .sync_page              = block_sync_page,
1997         .direct_IO              = ocfs2_direct_IO,
1998         .invalidatepage         = ocfs2_invalidatepage,
1999         .releasepage            = ocfs2_releasepage,
2000         .migratepage            = buffer_migrate_page,
2001         .is_partially_uptodate  = block_is_partially_uptodate,
2002         .error_remove_page      = generic_error_remove_page,
2003 };