Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
aops.c
Go to the documentation of this file.
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);
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);
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 
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);
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);
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  /*
294  * Unlock the page and cycle ip_alloc_sem so that we don't
295  * busyloop waiting for ip_alloc_sem to unlock
296  */
297  ret = AOP_TRUNCATED_PAGE;
298  unlock_page(page);
299  unlock = 0;
300  down_read(&oi->ip_alloc_sem);
301  up_read(&oi->ip_alloc_sem);
302  goto out_inode_unlock;
303  }
304 
305  /*
306  * i_size might have just been updated as we grabed the meta lock. We
307  * might now be discovering a truncate that hit on another node.
308  * block_read_full_page->get_block freaks out if it is asked to read
309  * beyond the end of a file, so we check here. Callers
310  * (generic_file_read, vm_ops->fault) are clever enough to check i_size
311  * and notice that the page they just read isn't needed.
312  *
313  * XXX sys_readahead() seems to get that wrong?
314  */
315  if (start >= i_size_read(inode)) {
316  zero_user(page, 0, PAGE_SIZE);
317  SetPageUptodate(page);
318  ret = 0;
319  goto out_alloc;
320  }
321 
323  ret = ocfs2_readpage_inline(inode, page);
324  else
326  unlock = 0;
327 
328 out_alloc:
329  up_read(&OCFS2_I(inode)->ip_alloc_sem);
330 out_inode_unlock:
331  ocfs2_inode_unlock(inode, 0);
332 out:
333  if (unlock)
334  unlock_page(page);
335  return ret;
336 }
337 
338 /*
339  * This is used only for read-ahead. Failures or difficult to handle
340  * situations are safe to ignore.
341  *
342  * Right now, we don't bother with BH_Boundary - in-inode extent lists
343  * are quite large (243 extents on 4k blocks), so most inodes don't
344  * grow out to a tree. If need be, detecting boundary extents could
345  * trivially be added in a future version of ocfs2_get_block().
346  */
347 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
348  struct list_head *pages, unsigned nr_pages)
349 {
350  int ret, err = -EIO;
351  struct inode *inode = mapping->host;
352  struct ocfs2_inode_info *oi = OCFS2_I(inode);
353  loff_t start;
354  struct page *last;
355 
356  /*
357  * Use the nonblocking flag for the dlm code to avoid page
358  * lock inversion, but don't bother with retrying.
359  */
361  if (ret)
362  return err;
363 
364  if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
365  ocfs2_inode_unlock(inode, 0);
366  return err;
367  }
368 
369  /*
370  * Don't bother with inline-data. There isn't anything
371  * to read-ahead in that case anyway...
372  */
374  goto out_unlock;
375 
376  /*
377  * Check whether a remote node truncated this file - we just
378  * drop out in that case as it's not worth handling here.
379  */
380  last = list_entry(pages->prev, struct page, lru);
381  start = (loff_t)last->index << PAGE_CACHE_SHIFT;
382  if (start >= i_size_read(inode))
383  goto out_unlock;
384 
385  err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
386 
387 out_unlock:
388  up_read(&oi->ip_alloc_sem);
389  ocfs2_inode_unlock(inode, 0);
390 
391  return err;
392 }
393 
394 /* Note: Because we don't support holes, our allocation has
395  * already happened (allocation writes zeros to the file data)
396  * so we don't have to worry about ordered writes in
397  * ocfs2_writepage.
398  *
399  * ->writepage is called during the process of invalidating the page cache
400  * during blocked lock processing. It can't block on any cluster locks
401  * to during block mapping. It's relying on the fact that the block
402  * mapping can't have disappeared under the dirty pages that it is
403  * being asked to write back.
404  */
405 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
406 {
407  trace_ocfs2_writepage(
408  (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
409  page->index);
410 
411  return block_write_full_page(page, ocfs2_get_block, wbc);
412 }
413 
414 /* Taken from ext3. We don't necessarily need the full blown
415  * functionality yet, but IMHO it's better to cut and paste the whole
416  * thing so we can avoid introducing our own bugs (and easily pick up
417  * their fixes when they happen) --Mark */
418 int walk_page_buffers( handle_t *handle,
419  struct buffer_head *head,
420  unsigned from,
421  unsigned to,
422  int *partial,
423  int (*fn)( handle_t *handle,
424  struct buffer_head *bh))
425 {
426  struct buffer_head *bh;
427  unsigned block_start, block_end;
428  unsigned blocksize = head->b_size;
429  int err, ret = 0;
430  struct buffer_head *next;
431 
432  for ( bh = head, block_start = 0;
433  ret == 0 && (bh != head || !block_start);
434  block_start = block_end, bh = next)
435  {
436  next = bh->b_this_page;
437  block_end = block_start + blocksize;
438  if (block_end <= from || block_start >= to) {
439  if (partial && !buffer_uptodate(bh))
440  *partial = 1;
441  continue;
442  }
443  err = (*fn)(handle, bh);
444  if (!ret)
445  ret = err;
446  }
447  return ret;
448 }
449 
450 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
451 {
453  u64 p_blkno = 0;
454  int err = 0;
455  struct inode *inode = mapping->host;
456 
457  trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
458  (unsigned long long)block);
459 
460  /* We don't need to lock journal system files, since they aren't
461  * accessed concurrently from multiple nodes.
462  */
463  if (!INODE_JOURNAL(inode)) {
464  err = ocfs2_inode_lock(inode, NULL, 0);
465  if (err) {
466  if (err != -ENOENT)
467  mlog_errno(err);
468  goto bail;
469  }
470  down_read(&OCFS2_I(inode)->ip_alloc_sem);
471  }
472 
473  if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
474  err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
475  NULL);
476 
477  if (!INODE_JOURNAL(inode)) {
478  up_read(&OCFS2_I(inode)->ip_alloc_sem);
479  ocfs2_inode_unlock(inode, 0);
480  }
481 
482  if (err) {
483  mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
484  (unsigned long long)block);
485  mlog_errno(err);
486  goto bail;
487  }
488 
489 bail:
490  status = err ? 0 : p_blkno;
491 
492  return status;
493 }
494 
495 /*
496  * TODO: Make this into a generic get_blocks function.
497  *
498  * From do_direct_io in direct-io.c:
499  * "So what we do is to permit the ->get_blocks function to populate
500  * bh.b_size with the size of IO which is permitted at this offset and
501  * this i_blkbits."
502  *
503  * This function is called directly from get_more_blocks in direct-io.c.
504  *
505  * called like this: dio->get_blocks(dio->inode, fs_startblk,
506  * fs_count, map_bh, dio->rw == WRITE);
507  *
508  * Note that we never bother to allocate blocks here, and thus ignore the
509  * create argument.
510  */
511 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
512  struct buffer_head *bh_result, int create)
513 {
514  int ret;
515  u64 p_blkno, inode_blocks, contig_blocks;
516  unsigned int ext_flags;
517  unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
518  unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
519 
520  /* This function won't even be called if the request isn't all
521  * nicely aligned and of the right size, so there's no need
522  * for us to check any of that. */
523 
524  inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
525 
526  /* This figures out the size of the next contiguous block, and
527  * our logical offset */
528  ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
529  &contig_blocks, &ext_flags);
530  if (ret) {
531  mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
532  (unsigned long long)iblock);
533  ret = -EIO;
534  goto bail;
535  }
536 
537  /* We should already CoW the refcounted extent in case of create. */
538  BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
539 
540  /*
541  * get_more_blocks() expects us to describe a hole by clearing
542  * the mapped bit on bh_result().
543  *
544  * Consider an unwritten extent as a hole.
545  */
546  if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
547  map_bh(bh_result, inode->i_sb, p_blkno);
548  else
549  clear_buffer_mapped(bh_result);
550 
551  /* make sure we don't map more than max_blocks blocks here as
552  that's all the kernel will handle at this point. */
553  if (max_blocks < contig_blocks)
554  contig_blocks = max_blocks;
555  bh_result->b_size = contig_blocks << blocksize_bits;
556 bail:
557  return ret;
558 }
559 
560 /*
561  * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
562  * particularly interested in the aio/dio case. We use the rw_lock DLM lock
563  * to protect io on one node from truncation on another.
564  */
565 static void ocfs2_dio_end_io(struct kiocb *iocb,
566  loff_t offset,
567  ssize_t bytes,
568  void *private,
569  int ret,
570  bool is_async)
571 {
572  struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
573  int level;
574  wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
575 
576  /* this io's submitter should not have unlocked this before we could */
578 
579  if (ocfs2_iocb_is_sem_locked(iocb))
581 
582  if (ocfs2_iocb_is_unaligned_aio(iocb)) {
584 
585  if (atomic_dec_and_test(&OCFS2_I(inode)->ip_unaligned_aio) &&
586  waitqueue_active(wq)) {
587  wake_up_all(wq);
588  }
589  }
590 
592 
593  level = ocfs2_iocb_rw_locked_level(iocb);
594  ocfs2_rw_unlock(inode, level);
595 
596  if (is_async)
597  aio_complete(iocb, ret, 0);
598  inode_dio_done(inode);
599 }
600 
601 /*
602  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
603  * from ext3. PageChecked() bits have been removed as OCFS2 does not
604  * do journalled data.
605  */
606 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
607 {
608  journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
609 
610  jbd2_journal_invalidatepage(journal, page, offset);
611 }
612 
613 static int ocfs2_releasepage(struct page *page, gfp_t wait)
614 {
615  journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
616 
617  if (!page_has_buffers(page))
618  return 0;
619  return jbd2_journal_try_to_free_buffers(journal, page, wait);
620 }
621 
622 static ssize_t ocfs2_direct_IO(int rw,
623  struct kiocb *iocb,
624  const struct iovec *iov,
625  loff_t offset,
626  unsigned long nr_segs)
627 {
628  struct file *file = iocb->ki_filp;
629  struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
630 
631  /*
632  * Fallback to buffered I/O if we see an inode without
633  * extents.
634  */
635  if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
636  return 0;
637 
638  /* Fallback to buffered I/O if we are appending. */
639  if (i_size_read(inode) <= offset)
640  return 0;
641 
642  return __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
643  iov, offset, nr_segs,
644  ocfs2_direct_IO_get_blocks,
645  ocfs2_dio_end_io, NULL, 0);
646 }
647 
648 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
649  u32 cpos,
650  unsigned int *start,
651  unsigned int *end)
652 {
653  unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
654 
656  unsigned int cpp;
657 
658  cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
659 
660  cluster_start = cpos % cpp;
661  cluster_start = cluster_start << osb->s_clustersize_bits;
662 
663  cluster_end = cluster_start + osb->s_clustersize;
664  }
665 
666  BUG_ON(cluster_start > PAGE_SIZE);
667  BUG_ON(cluster_end > PAGE_SIZE);
668 
669  if (start)
670  *start = cluster_start;
671  if (end)
672  *end = cluster_end;
673 }
674 
675 /*
676  * 'from' and 'to' are the region in the page to avoid zeroing.
677  *
678  * If pagesize > clustersize, this function will avoid zeroing outside
679  * of the cluster boundary.
680  *
681  * from == to == 0 is code for "zero the entire cluster region"
682  */
683 static void ocfs2_clear_page_regions(struct page *page,
684  struct ocfs2_super *osb, u32 cpos,
685  unsigned from, unsigned to)
686 {
687  void *kaddr;
688  unsigned int cluster_start, cluster_end;
689 
690  ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
691 
692  kaddr = kmap_atomic(page);
693 
694  if (from || to) {
695  if (from > cluster_start)
696  memset(kaddr + cluster_start, 0, from - cluster_start);
697  if (to < cluster_end)
698  memset(kaddr + to, 0, cluster_end - to);
699  } else {
700  memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
701  }
702 
703  kunmap_atomic(kaddr);
704 }
705 
706 /*
707  * Nonsparse file systems fully allocate before we get to the write
708  * code. This prevents ocfs2_write() from tagging the write as an
709  * allocating one, which means ocfs2_map_page_blocks() might try to
710  * read-in the blocks at the tail of our file. Avoid reading them by
711  * testing i_size against each block offset.
712  */
713 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
714  unsigned int block_start)
715 {
716  u64 offset = page_offset(page) + block_start;
717 
718  if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
719  return 1;
720 
721  if (i_size_read(inode) > offset)
722  return 1;
723 
724  return 0;
725 }
726 
727 /*
728  * Some of this taken from __block_write_begin(). We already have our
729  * mapping by now though, and the entire write will be allocating or
730  * it won't, so not much need to use BH_New.
731  *
732  * This will also skip zeroing, which is handled externally.
733  */
734 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
735  struct inode *inode, unsigned int from,
736  unsigned int to, int new)
737 {
738  int ret = 0;
739  struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
740  unsigned int block_end, block_start;
741  unsigned int bsize = 1 << inode->i_blkbits;
742 
743  if (!page_has_buffers(page))
744  create_empty_buffers(page, bsize, 0);
745 
746  head = page_buffers(page);
747  for (bh = head, block_start = 0; bh != head || !block_start;
748  bh = bh->b_this_page, block_start += bsize) {
749  block_end = block_start + bsize;
750 
751  clear_buffer_new(bh);
752 
753  /*
754  * Ignore blocks outside of our i/o range -
755  * they may belong to unallocated clusters.
756  */
757  if (block_start >= to || block_end <= from) {
758  if (PageUptodate(page))
759  set_buffer_uptodate(bh);
760  continue;
761  }
762 
763  /*
764  * For an allocating write with cluster size >= page
765  * size, we always write the entire page.
766  */
767  if (new)
768  set_buffer_new(bh);
769 
770  if (!buffer_mapped(bh)) {
771  map_bh(bh, inode->i_sb, *p_blkno);
772  unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
773  }
774 
775  if (PageUptodate(page)) {
776  if (!buffer_uptodate(bh))
777  set_buffer_uptodate(bh);
778  } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
779  !buffer_new(bh) &&
780  ocfs2_should_read_blk(inode, page, block_start) &&
781  (block_start < from || block_end > to)) {
782  ll_rw_block(READ, 1, &bh);
783  *wait_bh++=bh;
784  }
785 
786  *p_blkno = *p_blkno + 1;
787  }
788 
789  /*
790  * If we issued read requests - let them complete.
791  */
792  while(wait_bh > wait) {
793  wait_on_buffer(*--wait_bh);
794  if (!buffer_uptodate(*wait_bh))
795  ret = -EIO;
796  }
797 
798  if (ret == 0 || !new)
799  return ret;
800 
801  /*
802  * If we get -EIO above, zero out any newly allocated blocks
803  * to avoid exposing stale data.
804  */
805  bh = head;
806  block_start = 0;
807  do {
808  block_end = block_start + bsize;
809  if (block_end <= from)
810  goto next_bh;
811  if (block_start >= to)
812  break;
813 
814  zero_user(page, block_start, bh->b_size);
815  set_buffer_uptodate(bh);
816  mark_buffer_dirty(bh);
817 
818 next_bh:
819  block_start = block_end;
820  bh = bh->b_this_page;
821  } while (bh != head);
822 
823  return ret;
824 }
825 
826 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
827 #define OCFS2_MAX_CTXT_PAGES 1
828 #else
829 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
830 #endif
831 
832 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
833 
834 /*
835  * Describe the state of a single cluster to be written to.
836  */
840  /*
841  * Give this a unique field because c_phys eventually gets
842  * filled.
843  */
844  unsigned c_new;
845  unsigned c_unwritten;
846  unsigned c_needs_zero;
847 };
848 
850  /* Logical cluster position / len of write */
853 
854  /* First cluster allocated in a nonsparse extend */
856 
858 
859  /*
860  * This is true if page_size > cluster_size.
861  *
862  * It triggers a set of special cases during write which might
863  * have to deal with allocating writes to partial pages.
864  */
865  unsigned int w_large_pages;
866 
867  /*
868  * Pages involved in this write.
869  *
870  * w_target_page is the page being written to by the user.
871  *
872  * w_pages is an array of pages which always contains
873  * w_target_page, and in the case of an allocating write with
874  * page_size < cluster size, it will contain zero'd and mapped
875  * pages adjacent to w_target_page which need to be written
876  * out in so that future reads from that region will get
877  * zero's.
878  */
879  unsigned int w_num_pages;
881  struct page *w_target_page;
882 
883  /*
884  * w_target_locked is used for page_mkwrite path indicating no unlocking
885  * against w_target_page in ocfs2_write_end_nolock.
886  */
887  unsigned int w_target_locked:1;
888 
889  /*
890  * ocfs2_write_end() uses this to know what the real range to
891  * write in the target should be.
892  */
893  unsigned int w_target_from;
894  unsigned int w_target_to;
895 
896  /*
897  * We could use journal_current_handle() but this is cleaner,
898  * IMHO -Mark
899  */
900  handle_t *w_handle;
901 
902  struct buffer_head *w_di_bh;
903 
905 };
906 
907 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
908 {
909  int i;
910 
911  for(i = 0; i < num_pages; i++) {
912  if (pages[i]) {
913  unlock_page(pages[i]);
914  mark_page_accessed(pages[i]);
915  page_cache_release(pages[i]);
916  }
917  }
918 }
919 
920 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
921 {
922  int i;
923 
924  /*
925  * w_target_locked is only set to true in the page_mkwrite() case.
926  * The intent is to allow us to lock the target page from write_begin()
927  * to write_end(). The caller must hold a ref on w_target_page.
928  */
929  if (wc->w_target_locked) {
930  BUG_ON(!wc->w_target_page);
931  for (i = 0; i < wc->w_num_pages; i++) {
932  if (wc->w_target_page == wc->w_pages[i]) {
933  wc->w_pages[i] = NULL;
934  break;
935  }
936  }
939  }
941 
942  brelse(wc->w_di_bh);
943  kfree(wc);
944 }
945 
946 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
947  struct ocfs2_super *osb, loff_t pos,
948  unsigned len, struct buffer_head *di_bh)
949 {
950  u32 cend;
951  struct ocfs2_write_ctxt *wc;
952 
953  wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
954  if (!wc)
955  return -ENOMEM;
956 
957  wc->w_cpos = pos >> osb->s_clustersize_bits;
959  cend = (pos + len - 1) >> osb->s_clustersize_bits;
960  wc->w_clen = cend - wc->w_cpos + 1;
961  get_bh(di_bh);
962  wc->w_di_bh = di_bh;
963 
965  wc->w_large_pages = 1;
966  else
967  wc->w_large_pages = 0;
968 
969  ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
970 
971  *wcp = wc;
972 
973  return 0;
974 }
975 
976 /*
977  * If a page has any new buffers, zero them out here, and mark them uptodate
978  * and dirty so they'll be written out (in order to prevent uninitialised
979  * block data from leaking). And clear the new bit.
980  */
981 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
982 {
983  unsigned int block_start, block_end;
984  struct buffer_head *head, *bh;
985 
986  BUG_ON(!PageLocked(page));
987  if (!page_has_buffers(page))
988  return;
989 
990  bh = head = page_buffers(page);
991  block_start = 0;
992  do {
993  block_end = block_start + bh->b_size;
994 
995  if (buffer_new(bh)) {
996  if (block_end > from && block_start < to) {
997  if (!PageUptodate(page)) {
998  unsigned start, end;
999 
1000  start = max(from, block_start);
1001  end = min(to, block_end);
1002 
1003  zero_user_segment(page, start, end);
1004  set_buffer_uptodate(bh);
1005  }
1006 
1007  clear_buffer_new(bh);
1008  mark_buffer_dirty(bh);
1009  }
1010  }
1011 
1012  block_start = block_end;
1013  bh = bh->b_this_page;
1014  } while (bh != head);
1015 }
1016 
1017 /*
1018  * Only called when we have a failure during allocating write to write
1019  * zero's to the newly allocated region.
1020  */
1021 static void ocfs2_write_failure(struct inode *inode,
1022  struct ocfs2_write_ctxt *wc,
1023  loff_t user_pos, unsigned user_len)
1024 {
1025  int i;
1026  unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1027  to = user_pos + user_len;
1028  struct page *tmppage;
1029 
1030  ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1031 
1032  for(i = 0; i < wc->w_num_pages; i++) {
1033  tmppage = wc->w_pages[i];
1034 
1035  if (page_has_buffers(tmppage)) {
1036  if (ocfs2_should_order_data(inode))
1037  ocfs2_jbd2_file_inode(wc->w_handle, inode);
1038 
1039  block_commit_write(tmppage, from, to);
1040  }
1041  }
1042 }
1043 
1044 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1045  struct ocfs2_write_ctxt *wc,
1046  struct page *page, u32 cpos,
1047  loff_t user_pos, unsigned user_len,
1048  int new)
1049 {
1050  int ret;
1051  unsigned int map_from = 0, map_to = 0;
1052  unsigned int cluster_start, cluster_end;
1053  unsigned int user_data_from = 0, user_data_to = 0;
1054 
1055  ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1056  &cluster_start, &cluster_end);
1057 
1058  /* treat the write as new if the a hole/lseek spanned across
1059  * the page boundary.
1060  */
1061  new = new | ((i_size_read(inode) <= page_offset(page)) &&
1062  (page_offset(page) <= user_pos));
1063 
1064  if (page == wc->w_target_page) {
1065  map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1066  map_to = map_from + user_len;
1067 
1068  if (new)
1069  ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1070  cluster_start, cluster_end,
1071  new);
1072  else
1073  ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1074  map_from, map_to, new);
1075  if (ret) {
1076  mlog_errno(ret);
1077  goto out;
1078  }
1079 
1080  user_data_from = map_from;
1081  user_data_to = map_to;
1082  if (new) {
1083  map_from = cluster_start;
1084  map_to = cluster_end;
1085  }
1086  } else {
1087  /*
1088  * If we haven't allocated the new page yet, we
1089  * shouldn't be writing it out without copying user
1090  * data. This is likely a math error from the caller.
1091  */
1092  BUG_ON(!new);
1093 
1094  map_from = cluster_start;
1095  map_to = cluster_end;
1096 
1097  ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1098  cluster_start, cluster_end, new);
1099  if (ret) {
1100  mlog_errno(ret);
1101  goto out;
1102  }
1103  }
1104 
1105  /*
1106  * Parts of newly allocated pages need to be zero'd.
1107  *
1108  * Above, we have also rewritten 'to' and 'from' - as far as
1109  * the rest of the function is concerned, the entire cluster
1110  * range inside of a page needs to be written.
1111  *
1112  * We can skip this if the page is up to date - it's already
1113  * been zero'd from being read in as a hole.
1114  */
1115  if (new && !PageUptodate(page))
1116  ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1117  cpos, user_data_from, user_data_to);
1118 
1119  flush_dcache_page(page);
1120 
1121 out:
1122  return ret;
1123 }
1124 
1125 /*
1126  * This function will only grab one clusters worth of pages.
1127  */
1128 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1129  struct ocfs2_write_ctxt *wc,
1130  u32 cpos, loff_t user_pos,
1131  unsigned user_len, int new,
1132  struct page *mmap_page)
1133 {
1134  int ret = 0, i;
1135  unsigned long start, target_index, end_index, index;
1136  struct inode *inode = mapping->host;
1137  loff_t last_byte;
1138 
1139  target_index = user_pos >> PAGE_CACHE_SHIFT;
1140 
1141  /*
1142  * Figure out how many pages we'll be manipulating here. For
1143  * non allocating write, we just change the one
1144  * page. Otherwise, we'll need a whole clusters worth. If we're
1145  * writing past i_size, we only need enough pages to cover the
1146  * last page of the write.
1147  */
1148  if (new) {
1149  wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1150  start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1151  /*
1152  * We need the index *past* the last page we could possibly
1153  * touch. This is the page past the end of the write or
1154  * i_size, whichever is greater.
1155  */
1156  last_byte = max(user_pos + user_len, i_size_read(inode));
1157  BUG_ON(last_byte < 1);
1158  end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1159  if ((start + wc->w_num_pages) > end_index)
1160  wc->w_num_pages = end_index - start;
1161  } else {
1162  wc->w_num_pages = 1;
1163  start = target_index;
1164  }
1165 
1166  for(i = 0; i < wc->w_num_pages; i++) {
1167  index = start + i;
1168 
1169  if (index == target_index && mmap_page) {
1170  /*
1171  * ocfs2_pagemkwrite() is a little different
1172  * and wants us to directly use the page
1173  * passed in.
1174  */
1175  lock_page(mmap_page);
1176 
1177  /* Exit and let the caller retry */
1178  if (mmap_page->mapping != mapping) {
1179  WARN_ON(mmap_page->mapping);
1180  unlock_page(mmap_page);
1181  ret = -EAGAIN;
1182  goto out;
1183  }
1184 
1185  page_cache_get(mmap_page);
1186  wc->w_pages[i] = mmap_page;
1187  wc->w_target_locked = true;
1188  } else {
1189  wc->w_pages[i] = find_or_create_page(mapping, index,
1190  GFP_NOFS);
1191  if (!wc->w_pages[i]) {
1192  ret = -ENOMEM;
1193  mlog_errno(ret);
1194  goto out;
1195  }
1196  }
1197 
1198  if (index == target_index)
1199  wc->w_target_page = wc->w_pages[i];
1200  }
1201 out:
1202  if (ret)
1203  wc->w_target_locked = false;
1204  return ret;
1205 }
1206 
1207 /*
1208  * Prepare a single cluster for write one cluster into the file.
1209  */
1210 static int ocfs2_write_cluster(struct address_space *mapping,
1211  u32 phys, unsigned int unwritten,
1212  unsigned int should_zero,
1213  struct ocfs2_alloc_context *data_ac,
1214  struct ocfs2_alloc_context *meta_ac,
1215  struct ocfs2_write_ctxt *wc, u32 cpos,
1216  loff_t user_pos, unsigned user_len)
1217 {
1218  int ret, i, new;
1219  u64 v_blkno, p_blkno;
1220  struct inode *inode = mapping->host;
1221  struct ocfs2_extent_tree et;
1222 
1223  new = phys == 0 ? 1 : 0;
1224  if (new) {
1225  u32 tmp_pos;
1226 
1227  /*
1228  * This is safe to call with the page locks - it won't take
1229  * any additional semaphores or cluster locks.
1230  */
1231  tmp_pos = cpos;
1232  ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1233  &tmp_pos, 1, 0, wc->w_di_bh,
1234  wc->w_handle, data_ac,
1235  meta_ac, NULL);
1236  /*
1237  * This shouldn't happen because we must have already
1238  * calculated the correct meta data allocation required. The
1239  * internal tree allocation code should know how to increase
1240  * transaction credits itself.
1241  *
1242  * If need be, we could handle -EAGAIN for a
1243  * RESTART_TRANS here.
1244  */
1245  mlog_bug_on_msg(ret == -EAGAIN,
1246  "Inode %llu: EAGAIN return during allocation.\n",
1247  (unsigned long long)OCFS2_I(inode)->ip_blkno);
1248  if (ret < 0) {
1249  mlog_errno(ret);
1250  goto out;
1251  }
1252  } else if (unwritten) {
1253  ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1254  wc->w_di_bh);
1255  ret = ocfs2_mark_extent_written(inode, &et,
1256  wc->w_handle, cpos, 1, phys,
1257  meta_ac, &wc->w_dealloc);
1258  if (ret < 0) {
1259  mlog_errno(ret);
1260  goto out;
1261  }
1262  }
1263 
1264  if (should_zero)
1265  v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1266  else
1267  v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1268 
1269  /*
1270  * The only reason this should fail is due to an inability to
1271  * find the extent added.
1272  */
1273  ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1274  NULL);
1275  if (ret < 0) {
1276  ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1277  "at logical block %llu",
1278  (unsigned long long)OCFS2_I(inode)->ip_blkno,
1279  (unsigned long long)v_blkno);
1280  goto out;
1281  }
1282 
1283  BUG_ON(p_blkno == 0);
1284 
1285  for(i = 0; i < wc->w_num_pages; i++) {
1286  int tmpret;
1287 
1288  tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1289  wc->w_pages[i], cpos,
1290  user_pos, user_len,
1291  should_zero);
1292  if (tmpret) {
1293  mlog_errno(tmpret);
1294  if (ret == 0)
1295  ret = tmpret;
1296  }
1297  }
1298 
1299  /*
1300  * We only have cleanup to do in case of allocating write.
1301  */
1302  if (ret && new)
1303  ocfs2_write_failure(inode, wc, user_pos, user_len);
1304 
1305 out:
1306 
1307  return ret;
1308 }
1309 
1310 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1311  struct ocfs2_alloc_context *data_ac,
1312  struct ocfs2_alloc_context *meta_ac,
1313  struct ocfs2_write_ctxt *wc,
1314  loff_t pos, unsigned len)
1315 {
1316  int ret, i;
1317  loff_t cluster_off;
1318  unsigned int local_len = len;
1320  struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1321 
1322  for (i = 0; i < wc->w_clen; i++) {
1323  desc = &wc->w_desc[i];
1324 
1325  /*
1326  * We have to make sure that the total write passed in
1327  * doesn't extend past a single cluster.
1328  */
1329  local_len = len;
1330  cluster_off = pos & (osb->s_clustersize - 1);
1331  if ((cluster_off + local_len) > osb->s_clustersize)
1332  local_len = osb->s_clustersize - cluster_off;
1333 
1334  ret = ocfs2_write_cluster(mapping, desc->c_phys,
1335  desc->c_unwritten,
1336  desc->c_needs_zero,
1337  data_ac, meta_ac,
1338  wc, desc->c_cpos, pos, local_len);
1339  if (ret) {
1340  mlog_errno(ret);
1341  goto out;
1342  }
1343 
1344  len -= local_len;
1345  pos += local_len;
1346  }
1347 
1348  ret = 0;
1349 out:
1350  return ret;
1351 }
1352 
1353 /*
1354  * ocfs2_write_end() wants to know which parts of the target page it
1355  * should complete the write on. It's easiest to compute them ahead of
1356  * time when a more complete view of the write is available.
1357  */
1358 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1359  struct ocfs2_write_ctxt *wc,
1360  loff_t pos, unsigned len, int alloc)
1361 {
1362  struct ocfs2_write_cluster_desc *desc;
1363 
1364  wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1365  wc->w_target_to = wc->w_target_from + len;
1366 
1367  if (alloc == 0)
1368  return;
1369 
1370  /*
1371  * Allocating write - we may have different boundaries based
1372  * on page size and cluster size.
1373  *
1374  * NOTE: We can no longer compute one value from the other as
1375  * the actual write length and user provided length may be
1376  * different.
1377  */
1378 
1379  if (wc->w_large_pages) {
1380  /*
1381  * We only care about the 1st and last cluster within
1382  * our range and whether they should be zero'd or not. Either
1383  * value may be extended out to the start/end of a
1384  * newly allocated cluster.
1385  */
1386  desc = &wc->w_desc[0];
1387  if (desc->c_needs_zero)
1388  ocfs2_figure_cluster_boundaries(osb,
1389  desc->c_cpos,
1390  &wc->w_target_from,
1391  NULL);
1392 
1393  desc = &wc->w_desc[wc->w_clen - 1];
1394  if (desc->c_needs_zero)
1395  ocfs2_figure_cluster_boundaries(osb,
1396  desc->c_cpos,
1397  NULL,
1398  &wc->w_target_to);
1399  } else {
1400  wc->w_target_from = 0;
1402  }
1403 }
1404 
1405 /*
1406  * Populate each single-cluster write descriptor in the write context
1407  * with information about the i/o to be done.
1408  *
1409  * Returns the number of clusters that will have to be allocated, as
1410  * well as a worst case estimate of the number of extent records that
1411  * would have to be created during a write to an unwritten region.
1412  */
1413 static int ocfs2_populate_write_desc(struct inode *inode,
1414  struct ocfs2_write_ctxt *wc,
1415  unsigned int *clusters_to_alloc,
1416  unsigned int *extents_to_split)
1417 {
1418  int ret;
1419  struct ocfs2_write_cluster_desc *desc;
1420  unsigned int num_clusters = 0;
1421  unsigned int ext_flags = 0;
1422  u32 phys = 0;
1423  int i;
1424 
1425  *clusters_to_alloc = 0;
1426  *extents_to_split = 0;
1427 
1428  for (i = 0; i < wc->w_clen; i++) {
1429  desc = &wc->w_desc[i];
1430  desc->c_cpos = wc->w_cpos + i;
1431 
1432  if (num_clusters == 0) {
1433  /*
1434  * Need to look up the next extent record.
1435  */
1436  ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1437  &num_clusters, &ext_flags);
1438  if (ret) {
1439  mlog_errno(ret);
1440  goto out;
1441  }
1442 
1443  /* We should already CoW the refcountd extent. */
1444  BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1445 
1446  /*
1447  * Assume worst case - that we're writing in
1448  * the middle of the extent.
1449  *
1450  * We can assume that the write proceeds from
1451  * left to right, in which case the extent
1452  * insert code is smart enough to coalesce the
1453  * next splits into the previous records created.
1454  */
1455  if (ext_flags & OCFS2_EXT_UNWRITTEN)
1456  *extents_to_split = *extents_to_split + 2;
1457  } else if (phys) {
1458  /*
1459  * Only increment phys if it doesn't describe
1460  * a hole.
1461  */
1462  phys++;
1463  }
1464 
1465  /*
1466  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1467  * file that got extended. w_first_new_cpos tells us
1468  * where the newly allocated clusters are so we can
1469  * zero them.
1470  */
1471  if (desc->c_cpos >= wc->w_first_new_cpos) {
1472  BUG_ON(phys == 0);
1473  desc->c_needs_zero = 1;
1474  }
1475 
1476  desc->c_phys = phys;
1477  if (phys == 0) {
1478  desc->c_new = 1;
1479  desc->c_needs_zero = 1;
1480  *clusters_to_alloc = *clusters_to_alloc + 1;
1481  }
1482 
1483  if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1484  desc->c_unwritten = 1;
1485  desc->c_needs_zero = 1;
1486  }
1487 
1488  num_clusters--;
1489  }
1490 
1491  ret = 0;
1492 out:
1493  return ret;
1494 }
1495 
1496 static int ocfs2_write_begin_inline(struct address_space *mapping,
1497  struct inode *inode,
1498  struct ocfs2_write_ctxt *wc)
1499 {
1500  int ret;
1501  struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1502  struct page *page;
1503  handle_t *handle;
1504  struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1505 
1506  page = find_or_create_page(mapping, 0, GFP_NOFS);
1507  if (!page) {
1508  ret = -ENOMEM;
1509  mlog_errno(ret);
1510  goto out;
1511  }
1512  /*
1513  * If we don't set w_num_pages then this page won't get unlocked
1514  * and freed on cleanup of the write context.
1515  */
1516  wc->w_pages[0] = wc->w_target_page = page;
1517  wc->w_num_pages = 1;
1518 
1520  if (IS_ERR(handle)) {
1521  ret = PTR_ERR(handle);
1522  mlog_errno(ret);
1523  goto out;
1524  }
1525 
1526  ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1528  if (ret) {
1529  ocfs2_commit_trans(osb, handle);
1530 
1531  mlog_errno(ret);
1532  goto out;
1533  }
1534 
1535  if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1536  ocfs2_set_inode_data_inline(inode, di);
1537 
1538  if (!PageUptodate(page)) {
1539  ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1540  if (ret) {
1541  ocfs2_commit_trans(osb, handle);
1542 
1543  goto out;
1544  }
1545  }
1546 
1547  wc->w_handle = handle;
1548 out:
1549  return ret;
1550 }
1551 
1552 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1553 {
1554  struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1555 
1556  if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1557  return 1;
1558  return 0;
1559 }
1560 
1561 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1562  struct inode *inode, loff_t pos,
1563  unsigned len, struct page *mmap_page,
1564  struct ocfs2_write_ctxt *wc)
1565 {
1566  int ret, written = 0;
1567  loff_t end = pos + len;
1568  struct ocfs2_inode_info *oi = OCFS2_I(inode);
1569  struct ocfs2_dinode *di = NULL;
1570 
1571  trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1572  len, (unsigned long long)pos,
1573  oi->ip_dyn_features);
1574 
1575  /*
1576  * Handle inodes which already have inline data 1st.
1577  */
1578  if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1579  if (mmap_page == NULL &&
1581  goto do_inline_write;
1582 
1583  /*
1584  * The write won't fit - we have to give this inode an
1585  * inline extent list now.
1586  */
1588  if (ret)
1589  mlog_errno(ret);
1590  goto out;
1591  }
1592 
1593  /*
1594  * Check whether the inode can accept inline data.
1595  */
1596  if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1597  return 0;
1598 
1599  /*
1600  * Check whether the write can fit.
1601  */
1602  di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1603  if (mmap_page ||
1604  end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1605  return 0;
1606 
1607 do_inline_write:
1608  ret = ocfs2_write_begin_inline(mapping, inode, wc);
1609  if (ret) {
1610  mlog_errno(ret);
1611  goto out;
1612  }
1613 
1614  /*
1615  * This signals to the caller that the data can be written
1616  * inline.
1617  */
1618  written = 1;
1619 out:
1620  return written ? written : ret;
1621 }
1622 
1623 /*
1624  * This function only does anything for file systems which can't
1625  * handle sparse files.
1626  *
1627  * What we want to do here is fill in any hole between the current end
1628  * of allocation and the end of our write. That way the rest of the
1629  * write path can treat it as an non-allocating write, which has no
1630  * special case code for sparse/nonsparse files.
1631  */
1632 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1633  struct buffer_head *di_bh,
1634  loff_t pos, unsigned len,
1635  struct ocfs2_write_ctxt *wc)
1636 {
1637  int ret;
1638  loff_t newsize = pos + len;
1639 
1640  BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1641 
1642  if (newsize <= i_size_read(inode))
1643  return 0;
1644 
1645  ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1646  if (ret)
1647  mlog_errno(ret);
1648 
1649  wc->w_first_new_cpos =
1650  ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1651 
1652  return ret;
1653 }
1654 
1655 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1656  loff_t pos)
1657 {
1658  int ret = 0;
1659 
1660  BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1661  if (pos > i_size_read(inode))
1662  ret = ocfs2_zero_extend(inode, di_bh, pos);
1663 
1664  return ret;
1665 }
1666 
1667 /*
1668  * Try to flush truncate logs if we can free enough clusters from it.
1669  * As for return value, "< 0" means error, "0" no space and "1" means
1670  * we have freed enough spaces and let the caller try to allocate again.
1671  */
1672 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
1673  unsigned int needed)
1674 {
1675  tid_t target;
1676  int ret = 0;
1677  unsigned int truncated_clusters;
1678 
1679  mutex_lock(&osb->osb_tl_inode->i_mutex);
1680  truncated_clusters = osb->truncated_clusters;
1681  mutex_unlock(&osb->osb_tl_inode->i_mutex);
1682 
1683  /*
1684  * Check whether we can succeed in allocating if we free
1685  * the truncate log.
1686  */
1687  if (truncated_clusters < needed)
1688  goto out;
1689 
1690  ret = ocfs2_flush_truncate_log(osb);
1691  if (ret) {
1692  mlog_errno(ret);
1693  goto out;
1694  }
1695 
1696  if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
1697  jbd2_log_wait_commit(osb->journal->j_journal, target);
1698  ret = 1;
1699  }
1700 out:
1701  return ret;
1702 }
1703 
1704 int ocfs2_write_begin_nolock(struct file *filp,
1705  struct address_space *mapping,
1706  loff_t pos, unsigned len, unsigned flags,
1707  struct page **pagep, void **fsdata,
1708  struct buffer_head *di_bh, struct page *mmap_page)
1709 {
1710  int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1711  unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1712  struct ocfs2_write_ctxt *wc;
1713  struct inode *inode = mapping->host;
1714  struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1715  struct ocfs2_dinode *di;
1716  struct ocfs2_alloc_context *data_ac = NULL;
1717  struct ocfs2_alloc_context *meta_ac = NULL;
1718  handle_t *handle;
1719  struct ocfs2_extent_tree et;
1720  int try_free = 1, ret1;
1721 
1722 try_again:
1723  ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1724  if (ret) {
1725  mlog_errno(ret);
1726  return ret;
1727  }
1728 
1729  if (ocfs2_supports_inline_data(osb)) {
1730  ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1731  mmap_page, wc);
1732  if (ret == 1) {
1733  ret = 0;
1734  goto success;
1735  }
1736  if (ret < 0) {
1737  mlog_errno(ret);
1738  goto out;
1739  }
1740  }
1741 
1742  if (ocfs2_sparse_alloc(osb))
1743  ret = ocfs2_zero_tail(inode, di_bh, pos);
1744  else
1745  ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1746  wc);
1747  if (ret) {
1748  mlog_errno(ret);
1749  goto out;
1750  }
1751 
1752  ret = ocfs2_check_range_for_refcount(inode, pos, len);
1753  if (ret < 0) {
1754  mlog_errno(ret);
1755  goto out;
1756  } else if (ret == 1) {
1757  clusters_need = wc->w_clen;
1758  ret = ocfs2_refcount_cow(inode, filp, di_bh,
1759  wc->w_cpos, wc->w_clen, UINT_MAX);
1760  if (ret) {
1761  mlog_errno(ret);
1762  goto out;
1763  }
1764  }
1765 
1766  ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1767  &extents_to_split);
1768  if (ret) {
1769  mlog_errno(ret);
1770  goto out;
1771  }
1772  clusters_need += clusters_to_alloc;
1773 
1774  di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1775 
1776  trace_ocfs2_write_begin_nolock(
1777  (unsigned long long)OCFS2_I(inode)->ip_blkno,
1778  (long long)i_size_read(inode),
1779  le32_to_cpu(di->i_clusters),
1780  pos, len, flags, mmap_page,
1781  clusters_to_alloc, extents_to_split);
1782 
1783  /*
1784  * We set w_target_from, w_target_to here so that
1785  * ocfs2_write_end() knows which range in the target page to
1786  * write out. An allocation requires that we write the entire
1787  * cluster range.
1788  */
1789  if (clusters_to_alloc || extents_to_split) {
1790  /*
1791  * XXX: We are stretching the limits of
1792  * ocfs2_lock_allocators(). It greatly over-estimates
1793  * the work to be done.
1794  */
1795  ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1796  wc->w_di_bh);
1797  ret = ocfs2_lock_allocators(inode, &et,
1798  clusters_to_alloc, extents_to_split,
1799  &data_ac, &meta_ac);
1800  if (ret) {
1801  mlog_errno(ret);
1802  goto out;
1803  }
1804 
1805  if (data_ac)
1806  data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1807 
1808  credits = ocfs2_calc_extend_credits(inode->i_sb,
1809  &di->id2.i_list,
1810  clusters_to_alloc);
1811 
1812  }
1813 
1814  /*
1815  * We have to zero sparse allocated clusters, unwritten extent clusters,
1816  * and non-sparse clusters we just extended. For non-sparse writes,
1817  * we know zeros will only be needed in the first and/or last cluster.
1818  */
1819  if (clusters_to_alloc || extents_to_split ||
1820  (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1821  wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1822  cluster_of_pages = 1;
1823  else
1824  cluster_of_pages = 0;
1825 
1826  ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1827 
1828  handle = ocfs2_start_trans(osb, credits);
1829  if (IS_ERR(handle)) {
1830  ret = PTR_ERR(handle);
1831  mlog_errno(ret);
1832  goto out;
1833  }
1834 
1835  wc->w_handle = handle;
1836 
1837  if (clusters_to_alloc) {
1838  ret = dquot_alloc_space_nodirty(inode,
1839  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1840  if (ret)
1841  goto out_commit;
1842  }
1843  /*
1844  * We don't want this to fail in ocfs2_write_end(), so do it
1845  * here.
1846  */
1847  ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1849  if (ret) {
1850  mlog_errno(ret);
1851  goto out_quota;
1852  }
1853 
1854  /*
1855  * Fill our page array first. That way we've grabbed enough so
1856  * that we can zero and flush if we error after adding the
1857  * extent.
1858  */
1859  ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1860  cluster_of_pages, mmap_page);
1861  if (ret && ret != -EAGAIN) {
1862  mlog_errno(ret);
1863  goto out_quota;
1864  }
1865 
1866  /*
1867  * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1868  * the target page. In this case, we exit with no error and no target
1869  * page. This will trigger the caller, page_mkwrite(), to re-try
1870  * the operation.
1871  */
1872  if (ret == -EAGAIN) {
1873  BUG_ON(wc->w_target_page);
1874  ret = 0;
1875  goto out_quota;
1876  }
1877 
1878  ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1879  len);
1880  if (ret) {
1881  mlog_errno(ret);
1882  goto out_quota;
1883  }
1884 
1885  if (data_ac)
1886  ocfs2_free_alloc_context(data_ac);
1887  if (meta_ac)
1888  ocfs2_free_alloc_context(meta_ac);
1889 
1890 success:
1891  *pagep = wc->w_target_page;
1892  *fsdata = wc;
1893  return 0;
1894 out_quota:
1895  if (clusters_to_alloc)
1896  dquot_free_space(inode,
1897  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1898 out_commit:
1899  ocfs2_commit_trans(osb, handle);
1900 
1901 out:
1902  ocfs2_free_write_ctxt(wc);
1903 
1904  if (data_ac)
1905  ocfs2_free_alloc_context(data_ac);
1906  if (meta_ac)
1907  ocfs2_free_alloc_context(meta_ac);
1908 
1909  if (ret == -ENOSPC && try_free) {
1910  /*
1911  * Try to free some truncate log so that we can have enough
1912  * clusters to allocate.
1913  */
1914  try_free = 0;
1915 
1916  ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1917  if (ret1 == 1)
1918  goto try_again;
1919 
1920  if (ret1 < 0)
1921  mlog_errno(ret1);
1922  }
1923 
1924  return ret;
1925 }
1926 
1927 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1928  loff_t pos, unsigned len, unsigned flags,
1929  struct page **pagep, void **fsdata)
1930 {
1931  int ret;
1932  struct buffer_head *di_bh = NULL;
1933  struct inode *inode = mapping->host;
1934 
1935  ret = ocfs2_inode_lock(inode, &di_bh, 1);
1936  if (ret) {
1937  mlog_errno(ret);
1938  return ret;
1939  }
1940 
1941  /*
1942  * Take alloc sem here to prevent concurrent lookups. That way
1943  * the mapping, zeroing and tree manipulation within
1944  * ocfs2_write() will be safe against ->readpage(). This
1945  * should also serve to lock out allocation from a shared
1946  * writeable region.
1947  */
1948  down_write(&OCFS2_I(inode)->ip_alloc_sem);
1949 
1950  ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1951  fsdata, di_bh, NULL);
1952  if (ret) {
1953  mlog_errno(ret);
1954  goto out_fail;
1955  }
1956 
1957  brelse(di_bh);
1958 
1959  return 0;
1960 
1961 out_fail:
1962  up_write(&OCFS2_I(inode)->ip_alloc_sem);
1963 
1964  brelse(di_bh);
1965  ocfs2_inode_unlock(inode, 1);
1966 
1967  return ret;
1968 }
1969 
1970 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1971  unsigned len, unsigned *copied,
1972  struct ocfs2_dinode *di,
1973  struct ocfs2_write_ctxt *wc)
1974 {
1975  void *kaddr;
1976 
1977  if (unlikely(*copied < len)) {
1978  if (!PageUptodate(wc->w_target_page)) {
1979  *copied = 0;
1980  return;
1981  }
1982  }
1983 
1984  kaddr = kmap_atomic(wc->w_target_page);
1985  memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1986  kunmap_atomic(kaddr);
1987 
1988  trace_ocfs2_write_end_inline(
1989  (unsigned long long)OCFS2_I(inode)->ip_blkno,
1990  (unsigned long long)pos, *copied,
1991  le16_to_cpu(di->id2.i_data.id_count),
1993 }
1994 
1996  loff_t pos, unsigned len, unsigned copied,
1997  struct page *page, void *fsdata)
1998 {
1999  int i;
2000  unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2001  struct inode *inode = mapping->host;
2002  struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2003  struct ocfs2_write_ctxt *wc = fsdata;
2004  struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2005  handle_t *handle = wc->w_handle;
2006  struct page *tmppage;
2007 
2008  if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2009  ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2010  goto out_write_size;
2011  }
2012 
2013  if (unlikely(copied < len)) {
2014  if (!PageUptodate(wc->w_target_page))
2015  copied = 0;
2016 
2017  ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2018  start+len);
2019  }
2021 
2022  for(i = 0; i < wc->w_num_pages; i++) {
2023  tmppage = wc->w_pages[i];
2024 
2025  if (tmppage == wc->w_target_page) {
2026  from = wc->w_target_from;
2027  to = wc->w_target_to;
2028 
2029  BUG_ON(from > PAGE_CACHE_SIZE ||
2030  to > PAGE_CACHE_SIZE ||
2031  to < from);
2032  } else {
2033  /*
2034  * Pages adjacent to the target (if any) imply
2035  * a hole-filling write in which case we want
2036  * to flush their entire range.
2037  */
2038  from = 0;
2039  to = PAGE_CACHE_SIZE;
2040  }
2041 
2042  if (page_has_buffers(tmppage)) {
2043  if (ocfs2_should_order_data(inode))
2044  ocfs2_jbd2_file_inode(wc->w_handle, inode);
2045  block_commit_write(tmppage, from, to);
2046  }
2047  }
2048 
2049 out_write_size:
2050  pos += copied;
2051  if (pos > inode->i_size) {
2052  i_size_write(inode, pos);
2053  mark_inode_dirty(inode);
2054  }
2055  inode->i_blocks = ocfs2_inode_sector_count(inode);
2056  di->i_size = cpu_to_le64((u64)i_size_read(inode));
2057  inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2058  di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2059  di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2060  ocfs2_journal_dirty(handle, wc->w_di_bh);
2061 
2062  ocfs2_commit_trans(osb, handle);
2063 
2064  ocfs2_run_deallocs(osb, &wc->w_dealloc);
2065 
2066  ocfs2_free_write_ctxt(wc);
2067 
2068  return copied;
2069 }
2070 
2071 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2072  loff_t pos, unsigned len, unsigned copied,
2073  struct page *page, void *fsdata)
2074 {
2075  int ret;
2076  struct inode *inode = mapping->host;
2077 
2078  ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2079 
2080  up_write(&OCFS2_I(inode)->ip_alloc_sem);
2081  ocfs2_inode_unlock(inode, 1);
2082 
2083  return ret;
2084 }
2085 
2087  .readpage = ocfs2_readpage,
2088  .readpages = ocfs2_readpages,
2089  .writepage = ocfs2_writepage,
2090  .write_begin = ocfs2_write_begin,
2091  .write_end = ocfs2_write_end,
2092  .bmap = ocfs2_bmap,
2093  .direct_IO = ocfs2_direct_IO,
2094  .invalidatepage = ocfs2_invalidatepage,
2095  .releasepage = ocfs2_releasepage,
2096  .migratepage = buffer_migrate_page,
2097  .is_partially_uptodate = block_is_partially_uptodate,
2098  .error_remove_page = generic_error_remove_page,
2099 };