Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / fs / jbd / commit.c
1 /*
2  * linux/fs/commit.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Journal commit routines for the generic filesystem journaling code;
13  * part of the ext2fs journaling system.
14  */
15
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/pagemap.h>
23 #include <linux/smp_lock.h>
24
25 /*
26  * Default IO end handler for temporary BJ_IO buffer_heads.
27  */
28 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
29 {
30         BUFFER_TRACE(bh, "");
31         if (uptodate)
32                 set_buffer_uptodate(bh);
33         else
34                 clear_buffer_uptodate(bh);
35         unlock_buffer(bh);
36 }
37
38 /*
39  * When an ext3-ordered file is truncated, it is possible that many pages are
40  * not sucessfully freed, because they are attached to a committing transaction.
41  * After the transaction commits, these pages are left on the LRU, with no
42  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
43  * by the VM, but their apparent absence upsets the VM accounting, and it makes
44  * the numbers in /proc/meminfo look odd.
45  *
46  * So here, we have a buffer which has just come off the forget list.  Look to
47  * see if we can strip all buffers from the backing page.
48  *
49  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
50  * caller provided us with a ref against the buffer, and we drop that here.
51  */
52 static void release_buffer_page(struct buffer_head *bh)
53 {
54         struct page *page;
55
56         if (buffer_dirty(bh))
57                 goto nope;
58         if (atomic_read(&bh->b_count) != 1)
59                 goto nope;
60         page = bh->b_page;
61         if (!page)
62                 goto nope;
63         if (page->mapping)
64                 goto nope;
65
66         /* OK, it's a truncated page */
67         if (TestSetPageLocked(page))
68                 goto nope;
69
70         page_cache_get(page);
71         __brelse(bh);
72         try_to_free_buffers(page);
73         unlock_page(page);
74         page_cache_release(page);
75         return;
76
77 nope:
78         __brelse(bh);
79 }
80
81 /*
82  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
83  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
84  * return 0.  j_list_lock is dropped in this case.
85  */
86 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
87 {
88         if (!jbd_trylock_bh_state(bh)) {
89                 spin_unlock(&journal->j_list_lock);
90                 schedule();
91                 return 0;
92         }
93         return 1;
94 }
95
96 /* Done it all: now write the commit record.  We should have
97  * cleaned up our previous buffers by now, so if we are in abort
98  * mode we can now just skip the rest of the journal write
99  * entirely.
100  *
101  * Returns 1 if the journal needs to be aborted or 0 on success
102  */
103 static int journal_write_commit_record(journal_t *journal,
104                                         transaction_t *commit_transaction)
105 {
106         struct journal_head *descriptor;
107         struct buffer_head *bh;
108         int i, ret;
109         int barrier_done = 0;
110
111         if (is_journal_aborted(journal))
112                 return 0;
113
114         descriptor = journal_get_descriptor_buffer(journal);
115         if (!descriptor)
116                 return 1;
117
118         bh = jh2bh(descriptor);
119
120         /* AKPM: buglet - add `i' to tmp! */
121         for (i = 0; i < bh->b_size; i += 512) {
122                 journal_header_t *tmp = (journal_header_t*)bh->b_data;
123                 tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
124                 tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
125                 tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
126         }
127
128         JBUFFER_TRACE(descriptor, "write commit block");
129         set_buffer_dirty(bh);
130         if (journal->j_flags & JFS_BARRIER) {
131                 set_buffer_ordered(bh);
132                 barrier_done = 1;
133         }
134         ret = sync_dirty_buffer(bh);
135         /* is it possible for another commit to fail at roughly
136          * the same time as this one?  If so, we don't want to
137          * trust the barrier flag in the super, but instead want
138          * to remember if we sent a barrier request
139          */
140         if (ret == -EOPNOTSUPP && barrier_done) {
141                 char b[BDEVNAME_SIZE];
142
143                 printk(KERN_WARNING
144                         "JBD: barrier-based sync failed on %s - "
145                         "disabling barriers\n",
146                         bdevname(journal->j_dev, b));
147                 spin_lock(&journal->j_state_lock);
148                 journal->j_flags &= ~JFS_BARRIER;
149                 spin_unlock(&journal->j_state_lock);
150
151                 /* And try again, without the barrier */
152                 clear_buffer_ordered(bh);
153                 set_buffer_uptodate(bh);
154                 set_buffer_dirty(bh);
155                 ret = sync_dirty_buffer(bh);
156         }
157         put_bh(bh);             /* One for getblk() */
158         journal_put_journal_head(descriptor);
159
160         return (ret == -EIO);
161 }
162
163 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
164 {
165         int i;
166
167         for (i = 0; i < bufs; i++) {
168                 wbuf[i]->b_end_io = end_buffer_write_sync;
169                 /* We use-up our safety reference in submit_bh() */
170                 submit_bh(WRITE, wbuf[i]);
171         }
172 }
173
174 /*
175  *  Submit all the data buffers to disk
176  */
177 static void journal_submit_data_buffers(journal_t *journal,
178                                 transaction_t *commit_transaction)
179 {
180         struct journal_head *jh;
181         struct buffer_head *bh;
182         int locked;
183         int bufs = 0;
184         struct buffer_head **wbuf = journal->j_wbuf;
185
186         /*
187          * Whenever we unlock the journal and sleep, things can get added
188          * onto ->t_sync_datalist, so we have to keep looping back to
189          * write_out_data until we *know* that the list is empty.
190          *
191          * Cleanup any flushed data buffers from the data list.  Even in
192          * abort mode, we want to flush this out as soon as possible.
193          */
194 write_out_data:
195         cond_resched();
196         spin_lock(&journal->j_list_lock);
197
198         while (commit_transaction->t_sync_datalist) {
199                 jh = commit_transaction->t_sync_datalist;
200                 bh = jh2bh(jh);
201                 locked = 0;
202
203                 /* Get reference just to make sure buffer does not disappear
204                  * when we are forced to drop various locks */
205                 get_bh(bh);
206                 /* If the buffer is dirty, we need to submit IO and hence
207                  * we need the buffer lock. We try to lock the buffer without
208                  * blocking. If we fail, we need to drop j_list_lock and do
209                  * blocking lock_buffer().
210                  */
211                 if (buffer_dirty(bh)) {
212                         if (test_set_buffer_locked(bh)) {
213                                 BUFFER_TRACE(bh, "needs blocking lock");
214                                 spin_unlock(&journal->j_list_lock);
215                                 /* Write out all data to prevent deadlocks */
216                                 journal_do_submit_data(wbuf, bufs);
217                                 bufs = 0;
218                                 lock_buffer(bh);
219                                 spin_lock(&journal->j_list_lock);
220                         }
221                         locked = 1;
222                 }
223                 /* We have to get bh_state lock. Again out of order, sigh. */
224                 if (!inverted_lock(journal, bh)) {
225                         jbd_lock_bh_state(bh);
226                         spin_lock(&journal->j_list_lock);
227                 }
228                 /* Someone already cleaned up the buffer? */
229                 if (!buffer_jbd(bh)
230                         || jh->b_transaction != commit_transaction
231                         || jh->b_jlist != BJ_SyncData) {
232                         jbd_unlock_bh_state(bh);
233                         if (locked)
234                                 unlock_buffer(bh);
235                         BUFFER_TRACE(bh, "already cleaned up");
236                         put_bh(bh);
237                         continue;
238                 }
239                 if (locked && test_clear_buffer_dirty(bh)) {
240                         BUFFER_TRACE(bh, "needs writeout, adding to array");
241                         wbuf[bufs++] = bh;
242                         __journal_file_buffer(jh, commit_transaction,
243                                                 BJ_Locked);
244                         jbd_unlock_bh_state(bh);
245                         if (bufs == journal->j_wbufsize) {
246                                 spin_unlock(&journal->j_list_lock);
247                                 journal_do_submit_data(wbuf, bufs);
248                                 bufs = 0;
249                                 goto write_out_data;
250                         }
251                 }
252                 else {
253                         BUFFER_TRACE(bh, "writeout complete: unfile");
254                         __journal_unfile_buffer(jh);
255                         jbd_unlock_bh_state(bh);
256                         if (locked)
257                                 unlock_buffer(bh);
258                         journal_remove_journal_head(bh);
259                         /* Once for our safety reference, once for
260                          * journal_remove_journal_head() */
261                         put_bh(bh);
262                         put_bh(bh);
263                 }
264
265                 if (lock_need_resched(&journal->j_list_lock)) {
266                         spin_unlock(&journal->j_list_lock);
267                         goto write_out_data;
268                 }
269         }
270         spin_unlock(&journal->j_list_lock);
271         journal_do_submit_data(wbuf, bufs);
272 }
273
274 /*
275  * journal_commit_transaction
276  *
277  * The primary function for committing a transaction to the log.  This
278  * function is called by the journal thread to begin a complete commit.
279  */
280 void journal_commit_transaction(journal_t *journal)
281 {
282         transaction_t *commit_transaction;
283         struct journal_head *jh, *new_jh, *descriptor;
284         struct buffer_head **wbuf = journal->j_wbuf;
285         int bufs;
286         int flags;
287         int err;
288         unsigned long blocknr;
289         char *tagp = NULL;
290         journal_header_t *header;
291         journal_block_tag_t *tag = NULL;
292         int space_left = 0;
293         int first_tag = 0;
294         int tag_flag;
295         int i;
296
297         /*
298          * First job: lock down the current transaction and wait for
299          * all outstanding updates to complete.
300          */
301
302 #ifdef COMMIT_STATS
303         spin_lock(&journal->j_list_lock);
304         summarise_journal_usage(journal);
305         spin_unlock(&journal->j_list_lock);
306 #endif
307
308         /* Do we need to erase the effects of a prior journal_flush? */
309         if (journal->j_flags & JFS_FLUSHED) {
310                 jbd_debug(3, "super block updated\n");
311                 journal_update_superblock(journal, 1);
312         } else {
313                 jbd_debug(3, "superblock not updated\n");
314         }
315
316         J_ASSERT(journal->j_running_transaction != NULL);
317         J_ASSERT(journal->j_committing_transaction == NULL);
318
319         commit_transaction = journal->j_running_transaction;
320         J_ASSERT(commit_transaction->t_state == T_RUNNING);
321
322         jbd_debug(1, "JBD: starting commit of transaction %d\n",
323                         commit_transaction->t_tid);
324
325         spin_lock(&journal->j_state_lock);
326         commit_transaction->t_state = T_LOCKED;
327
328         spin_lock(&commit_transaction->t_handle_lock);
329         while (commit_transaction->t_updates) {
330                 DEFINE_WAIT(wait);
331
332                 prepare_to_wait(&journal->j_wait_updates, &wait,
333                                         TASK_UNINTERRUPTIBLE);
334                 if (commit_transaction->t_updates) {
335                         spin_unlock(&commit_transaction->t_handle_lock);
336                         spin_unlock(&journal->j_state_lock);
337                         schedule();
338                         spin_lock(&journal->j_state_lock);
339                         spin_lock(&commit_transaction->t_handle_lock);
340                 }
341                 finish_wait(&journal->j_wait_updates, &wait);
342         }
343         spin_unlock(&commit_transaction->t_handle_lock);
344
345         J_ASSERT (commit_transaction->t_outstanding_credits <=
346                         journal->j_max_transaction_buffers);
347
348         /*
349          * First thing we are allowed to do is to discard any remaining
350          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
351          * that there are no such buffers: if a large filesystem
352          * operation like a truncate needs to split itself over multiple
353          * transactions, then it may try to do a journal_restart() while
354          * there are still BJ_Reserved buffers outstanding.  These must
355          * be released cleanly from the current transaction.
356          *
357          * In this case, the filesystem must still reserve write access
358          * again before modifying the buffer in the new transaction, but
359          * we do not require it to remember exactly which old buffers it
360          * has reserved.  This is consistent with the existing behaviour
361          * that multiple journal_get_write_access() calls to the same
362          * buffer are perfectly permissable.
363          */
364         while (commit_transaction->t_reserved_list) {
365                 jh = commit_transaction->t_reserved_list;
366                 JBUFFER_TRACE(jh, "reserved, unused: refile");
367                 /*
368                  * A journal_get_undo_access()+journal_release_buffer() may
369                  * leave undo-committed data.
370                  */
371                 if (jh->b_committed_data) {
372                         struct buffer_head *bh = jh2bh(jh);
373
374                         jbd_lock_bh_state(bh);
375                         jbd_slab_free(jh->b_committed_data, bh->b_size);
376                         jh->b_committed_data = NULL;
377                         jbd_unlock_bh_state(bh);
378                 }
379                 journal_refile_buffer(journal, jh);
380         }
381
382         /*
383          * Now try to drop any written-back buffers from the journal's
384          * checkpoint lists.  We do this *before* commit because it potentially
385          * frees some memory
386          */
387         spin_lock(&journal->j_list_lock);
388         __journal_clean_checkpoint_list(journal);
389         spin_unlock(&journal->j_list_lock);
390
391         jbd_debug (3, "JBD: commit phase 1\n");
392
393         /*
394          * Switch to a new revoke table.
395          */
396         journal_switch_revoke_table(journal);
397
398         commit_transaction->t_state = T_FLUSH;
399         journal->j_committing_transaction = commit_transaction;
400         journal->j_running_transaction = NULL;
401         commit_transaction->t_log_start = journal->j_head;
402         wake_up(&journal->j_wait_transaction_locked);
403         spin_unlock(&journal->j_state_lock);
404
405         jbd_debug (3, "JBD: commit phase 2\n");
406
407         /*
408          * First, drop modified flag: all accesses to the buffers
409          * will be tracked for a new trasaction only -bzzz
410          */
411         spin_lock(&journal->j_list_lock);
412         if (commit_transaction->t_buffers) {
413                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
414                 do {
415                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
416                                         new_jh->b_modified == 0);
417                         new_jh->b_modified = 0;
418                         new_jh = new_jh->b_tnext;
419                 } while (new_jh != jh);
420         }
421         spin_unlock(&journal->j_list_lock);
422
423         /*
424          * Now start flushing things to disk, in the order they appear
425          * on the transaction lists.  Data blocks go first.
426          */
427         err = 0;
428         journal_submit_data_buffers(journal, commit_transaction);
429
430         /*
431          * Wait for all previously submitted IO to complete.
432          */
433         spin_lock(&journal->j_list_lock);
434         while (commit_transaction->t_locked_list) {
435                 struct buffer_head *bh;
436
437                 jh = commit_transaction->t_locked_list->b_tprev;
438                 bh = jh2bh(jh);
439                 get_bh(bh);
440                 if (buffer_locked(bh)) {
441                         spin_unlock(&journal->j_list_lock);
442                         wait_on_buffer(bh);
443                         if (unlikely(!buffer_uptodate(bh)))
444                                 err = -EIO;
445                         spin_lock(&journal->j_list_lock);
446                 }
447                 if (!inverted_lock(journal, bh)) {
448                         put_bh(bh);
449                         spin_lock(&journal->j_list_lock);
450                         continue;
451                 }
452                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
453                         __journal_unfile_buffer(jh);
454                         jbd_unlock_bh_state(bh);
455                         journal_remove_journal_head(bh);
456                         put_bh(bh);
457                 } else {
458                         jbd_unlock_bh_state(bh);
459                 }
460                 put_bh(bh);
461                 cond_resched_lock(&journal->j_list_lock);
462         }
463         spin_unlock(&journal->j_list_lock);
464
465         if (err)
466                 __journal_abort_hard(journal);
467
468         journal_write_revoke_records(journal, commit_transaction);
469
470         jbd_debug(3, "JBD: commit phase 2\n");
471
472         /*
473          * If we found any dirty or locked buffers, then we should have
474          * looped back up to the write_out_data label.  If there weren't
475          * any then journal_clean_data_list should have wiped the list
476          * clean by now, so check that it is in fact empty.
477          */
478         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
479
480         jbd_debug (3, "JBD: commit phase 3\n");
481
482         /*
483          * Way to go: we have now written out all of the data for a
484          * transaction!  Now comes the tricky part: we need to write out
485          * metadata.  Loop over the transaction's entire buffer list:
486          */
487         commit_transaction->t_state = T_COMMIT;
488
489         descriptor = NULL;
490         bufs = 0;
491         while (commit_transaction->t_buffers) {
492
493                 /* Find the next buffer to be journaled... */
494
495                 jh = commit_transaction->t_buffers;
496
497                 /* If we're in abort mode, we just un-journal the buffer and
498                    release it for background writing. */
499
500                 if (is_journal_aborted(journal)) {
501                         JBUFFER_TRACE(jh, "journal is aborting: refile");
502                         journal_refile_buffer(journal, jh);
503                         /* If that was the last one, we need to clean up
504                          * any descriptor buffers which may have been
505                          * already allocated, even if we are now
506                          * aborting. */
507                         if (!commit_transaction->t_buffers)
508                                 goto start_journal_io;
509                         continue;
510                 }
511
512                 /* Make sure we have a descriptor block in which to
513                    record the metadata buffer. */
514
515                 if (!descriptor) {
516                         struct buffer_head *bh;
517
518                         J_ASSERT (bufs == 0);
519
520                         jbd_debug(4, "JBD: get descriptor\n");
521
522                         descriptor = journal_get_descriptor_buffer(journal);
523                         if (!descriptor) {
524                                 __journal_abort_hard(journal);
525                                 continue;
526                         }
527
528                         bh = jh2bh(descriptor);
529                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
530                                 (unsigned long long)bh->b_blocknr, bh->b_data);
531                         header = (journal_header_t *)&bh->b_data[0];
532                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
533                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
534                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
535
536                         tagp = &bh->b_data[sizeof(journal_header_t)];
537                         space_left = bh->b_size - sizeof(journal_header_t);
538                         first_tag = 1;
539                         set_buffer_jwrite(bh);
540                         set_buffer_dirty(bh);
541                         wbuf[bufs++] = bh;
542
543                         /* Record it so that we can wait for IO
544                            completion later */
545                         BUFFER_TRACE(bh, "ph3: file as descriptor");
546                         journal_file_buffer(descriptor, commit_transaction,
547                                         BJ_LogCtl);
548                 }
549
550                 /* Where is the buffer to be written? */
551
552                 err = journal_next_log_block(journal, &blocknr);
553                 /* If the block mapping failed, just abandon the buffer
554                    and repeat this loop: we'll fall into the
555                    refile-on-abort condition above. */
556                 if (err) {
557                         __journal_abort_hard(journal);
558                         continue;
559                 }
560
561                 /*
562                  * start_this_handle() uses t_outstanding_credits to determine
563                  * the free space in the log, but this counter is changed
564                  * by journal_next_log_block() also.
565                  */
566                 commit_transaction->t_outstanding_credits--;
567
568                 /* Bump b_count to prevent truncate from stumbling over
569                    the shadowed buffer!  @@@ This can go if we ever get
570                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
571                 atomic_inc(&jh2bh(jh)->b_count);
572
573                 /* Make a temporary IO buffer with which to write it out
574                    (this will requeue both the metadata buffer and the
575                    temporary IO buffer). new_bh goes on BJ_IO*/
576
577                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
578                 /*
579                  * akpm: journal_write_metadata_buffer() sets
580                  * new_bh->b_transaction to commit_transaction.
581                  * We need to clean this up before we release new_bh
582                  * (which is of type BJ_IO)
583                  */
584                 JBUFFER_TRACE(jh, "ph3: write metadata");
585                 flags = journal_write_metadata_buffer(commit_transaction,
586                                                       jh, &new_jh, blocknr);
587                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
588                 wbuf[bufs++] = jh2bh(new_jh);
589
590                 /* Record the new block's tag in the current descriptor
591                    buffer */
592
593                 tag_flag = 0;
594                 if (flags & 1)
595                         tag_flag |= JFS_FLAG_ESCAPE;
596                 if (!first_tag)
597                         tag_flag |= JFS_FLAG_SAME_UUID;
598
599                 tag = (journal_block_tag_t *) tagp;
600                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
601                 tag->t_flags = cpu_to_be32(tag_flag);
602                 tagp += sizeof(journal_block_tag_t);
603                 space_left -= sizeof(journal_block_tag_t);
604
605                 if (first_tag) {
606                         memcpy (tagp, journal->j_uuid, 16);
607                         tagp += 16;
608                         space_left -= 16;
609                         first_tag = 0;
610                 }
611
612                 /* If there's no more to do, or if the descriptor is full,
613                    let the IO rip! */
614
615                 if (bufs == journal->j_wbufsize ||
616                     commit_transaction->t_buffers == NULL ||
617                     space_left < sizeof(journal_block_tag_t) + 16) {
618
619                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
620
621                         /* Write an end-of-descriptor marker before
622                            submitting the IOs.  "tag" still points to
623                            the last tag we set up. */
624
625                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
626
627 start_journal_io:
628                         for (i = 0; i < bufs; i++) {
629                                 struct buffer_head *bh = wbuf[i];
630                                 lock_buffer(bh);
631                                 clear_buffer_dirty(bh);
632                                 set_buffer_uptodate(bh);
633                                 bh->b_end_io = journal_end_buffer_io_sync;
634                                 submit_bh(WRITE, bh);
635                         }
636                         cond_resched();
637
638                         /* Force a new descriptor to be generated next
639                            time round the loop. */
640                         descriptor = NULL;
641                         bufs = 0;
642                 }
643         }
644
645         /* Lo and behold: we have just managed to send a transaction to
646            the log.  Before we can commit it, wait for the IO so far to
647            complete.  Control buffers being written are on the
648            transaction's t_log_list queue, and metadata buffers are on
649            the t_iobuf_list queue.
650
651            Wait for the buffers in reverse order.  That way we are
652            less likely to be woken up until all IOs have completed, and
653            so we incur less scheduling load.
654         */
655
656         jbd_debug(3, "JBD: commit phase 4\n");
657
658         /*
659          * akpm: these are BJ_IO, and j_list_lock is not needed.
660          * See __journal_try_to_free_buffer.
661          */
662 wait_for_iobuf:
663         while (commit_transaction->t_iobuf_list != NULL) {
664                 struct buffer_head *bh;
665
666                 jh = commit_transaction->t_iobuf_list->b_tprev;
667                 bh = jh2bh(jh);
668                 if (buffer_locked(bh)) {
669                         wait_on_buffer(bh);
670                         goto wait_for_iobuf;
671                 }
672                 if (cond_resched())
673                         goto wait_for_iobuf;
674
675                 if (unlikely(!buffer_uptodate(bh)))
676                         err = -EIO;
677
678                 clear_buffer_jwrite(bh);
679
680                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
681                 journal_unfile_buffer(journal, jh);
682
683                 /*
684                  * ->t_iobuf_list should contain only dummy buffer_heads
685                  * which were created by journal_write_metadata_buffer().
686                  */
687                 BUFFER_TRACE(bh, "dumping temporary bh");
688                 journal_put_journal_head(jh);
689                 __brelse(bh);
690                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
691                 free_buffer_head(bh);
692
693                 /* We also have to unlock and free the corresponding
694                    shadowed buffer */
695                 jh = commit_transaction->t_shadow_list->b_tprev;
696                 bh = jh2bh(jh);
697                 clear_bit(BH_JWrite, &bh->b_state);
698                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
699
700                 /* The metadata is now released for reuse, but we need
701                    to remember it against this transaction so that when
702                    we finally commit, we can do any checkpointing
703                    required. */
704                 JBUFFER_TRACE(jh, "file as BJ_Forget");
705                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
706                 /* Wake up any transactions which were waiting for this
707                    IO to complete */
708                 wake_up_bit(&bh->b_state, BH_Unshadow);
709                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
710                 __brelse(bh);
711         }
712
713         J_ASSERT (commit_transaction->t_shadow_list == NULL);
714
715         jbd_debug(3, "JBD: commit phase 5\n");
716
717         /* Here we wait for the revoke record and descriptor record buffers */
718  wait_for_ctlbuf:
719         while (commit_transaction->t_log_list != NULL) {
720                 struct buffer_head *bh;
721
722                 jh = commit_transaction->t_log_list->b_tprev;
723                 bh = jh2bh(jh);
724                 if (buffer_locked(bh)) {
725                         wait_on_buffer(bh);
726                         goto wait_for_ctlbuf;
727                 }
728                 if (cond_resched())
729                         goto wait_for_ctlbuf;
730
731                 if (unlikely(!buffer_uptodate(bh)))
732                         err = -EIO;
733
734                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
735                 clear_buffer_jwrite(bh);
736                 journal_unfile_buffer(journal, jh);
737                 journal_put_journal_head(jh);
738                 __brelse(bh);           /* One for getblk */
739                 /* AKPM: bforget here */
740         }
741
742         jbd_debug(3, "JBD: commit phase 6\n");
743
744         if (journal_write_commit_record(journal, commit_transaction))
745                 err = -EIO;
746
747         if (err)
748                 __journal_abort_hard(journal);
749
750         /* End of a transaction!  Finally, we can do checkpoint
751            processing: any buffers committed as a result of this
752            transaction can be removed from any checkpoint list it was on
753            before. */
754
755         jbd_debug(3, "JBD: commit phase 7\n");
756
757         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
758         J_ASSERT(commit_transaction->t_buffers == NULL);
759         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
760         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
761         J_ASSERT(commit_transaction->t_shadow_list == NULL);
762         J_ASSERT(commit_transaction->t_log_list == NULL);
763
764 restart_loop:
765         /*
766          * As there are other places (journal_unmap_buffer()) adding buffers
767          * to this list we have to be careful and hold the j_list_lock.
768          */
769         spin_lock(&journal->j_list_lock);
770         while (commit_transaction->t_forget) {
771                 transaction_t *cp_transaction;
772                 struct buffer_head *bh;
773
774                 jh = commit_transaction->t_forget;
775                 spin_unlock(&journal->j_list_lock);
776                 bh = jh2bh(jh);
777                 jbd_lock_bh_state(bh);
778                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
779                         jh->b_transaction == journal->j_running_transaction);
780
781                 /*
782                  * If there is undo-protected committed data against
783                  * this buffer, then we can remove it now.  If it is a
784                  * buffer needing such protection, the old frozen_data
785                  * field now points to a committed version of the
786                  * buffer, so rotate that field to the new committed
787                  * data.
788                  *
789                  * Otherwise, we can just throw away the frozen data now.
790                  */
791                 if (jh->b_committed_data) {
792                         jbd_slab_free(jh->b_committed_data, bh->b_size);
793                         jh->b_committed_data = NULL;
794                         if (jh->b_frozen_data) {
795                                 jh->b_committed_data = jh->b_frozen_data;
796                                 jh->b_frozen_data = NULL;
797                         }
798                 } else if (jh->b_frozen_data) {
799                         jbd_slab_free(jh->b_frozen_data, bh->b_size);
800                         jh->b_frozen_data = NULL;
801                 }
802
803                 spin_lock(&journal->j_list_lock);
804                 cp_transaction = jh->b_cp_transaction;
805                 if (cp_transaction) {
806                         JBUFFER_TRACE(jh, "remove from old cp transaction");
807                         __journal_remove_checkpoint(jh);
808                 }
809
810                 /* Only re-checkpoint the buffer_head if it is marked
811                  * dirty.  If the buffer was added to the BJ_Forget list
812                  * by journal_forget, it may no longer be dirty and
813                  * there's no point in keeping a checkpoint record for
814                  * it. */
815
816                 /* A buffer which has been freed while still being
817                  * journaled by a previous transaction may end up still
818                  * being dirty here, but we want to avoid writing back
819                  * that buffer in the future now that the last use has
820                  * been committed.  That's not only a performance gain,
821                  * it also stops aliasing problems if the buffer is left
822                  * behind for writeback and gets reallocated for another
823                  * use in a different page. */
824                 if (buffer_freed(bh)) {
825                         clear_buffer_freed(bh);
826                         clear_buffer_jbddirty(bh);
827                 }
828
829                 if (buffer_jbddirty(bh)) {
830                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
831                         __journal_insert_checkpoint(jh, commit_transaction);
832                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
833                         __journal_refile_buffer(jh);
834                         jbd_unlock_bh_state(bh);
835                 } else {
836                         J_ASSERT_BH(bh, !buffer_dirty(bh));
837                         /* The buffer on BJ_Forget list and not jbddirty means
838                          * it has been freed by this transaction and hence it
839                          * could not have been reallocated until this
840                          * transaction has committed. *BUT* it could be
841                          * reallocated once we have written all the data to
842                          * disk and before we process the buffer on BJ_Forget
843                          * list. */
844                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
845                         __journal_refile_buffer(jh);
846                         if (!jh->b_transaction) {
847                                 jbd_unlock_bh_state(bh);
848                                  /* needs a brelse */
849                                 journal_remove_journal_head(bh);
850                                 release_buffer_page(bh);
851                         } else
852                                 jbd_unlock_bh_state(bh);
853                 }
854                 cond_resched_lock(&journal->j_list_lock);
855         }
856         spin_unlock(&journal->j_list_lock);
857         /*
858          * This is a bit sleazy.  We borrow j_list_lock to protect
859          * journal->j_committing_transaction in __journal_remove_checkpoint.
860          * Really, __journal_remove_checkpoint should be using j_state_lock but
861          * it's a bit hassle to hold that across __journal_remove_checkpoint
862          */
863         spin_lock(&journal->j_state_lock);
864         spin_lock(&journal->j_list_lock);
865         /*
866          * Now recheck if some buffers did not get attached to the transaction
867          * while the lock was dropped...
868          */
869         if (commit_transaction->t_forget) {
870                 spin_unlock(&journal->j_list_lock);
871                 spin_unlock(&journal->j_state_lock);
872                 goto restart_loop;
873         }
874
875         /* Done with this transaction! */
876
877         jbd_debug(3, "JBD: commit phase 8\n");
878
879         J_ASSERT(commit_transaction->t_state == T_COMMIT);
880
881         commit_transaction->t_state = T_FINISHED;
882         J_ASSERT(commit_transaction == journal->j_committing_transaction);
883         journal->j_commit_sequence = commit_transaction->t_tid;
884         journal->j_committing_transaction = NULL;
885         spin_unlock(&journal->j_state_lock);
886
887         if (commit_transaction->t_checkpoint_list == NULL) {
888                 __journal_drop_transaction(journal, commit_transaction);
889         } else {
890                 if (journal->j_checkpoint_transactions == NULL) {
891                         journal->j_checkpoint_transactions = commit_transaction;
892                         commit_transaction->t_cpnext = commit_transaction;
893                         commit_transaction->t_cpprev = commit_transaction;
894                 } else {
895                         commit_transaction->t_cpnext =
896                                 journal->j_checkpoint_transactions;
897                         commit_transaction->t_cpprev =
898                                 commit_transaction->t_cpnext->t_cpprev;
899                         commit_transaction->t_cpnext->t_cpprev =
900                                 commit_transaction;
901                         commit_transaction->t_cpprev->t_cpnext =
902                                 commit_transaction;
903                 }
904         }
905         spin_unlock(&journal->j_list_lock);
906
907         jbd_debug(1, "JBD: commit %d complete, head %d\n",
908                   journal->j_commit_sequence, journal->j_tail_sequence);
909
910         wake_up(&journal->j_wait_done_commit);
911 }