Merge branch 'upstream-linus' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik...
[linux-2.6] / fs / jbd / commit.c
1 /*
2  * linux/fs/jbd/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
24 /*
25  * Default IO end handler for temporary BJ_IO buffer_heads.
26  */
27 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
28 {
29         BUFFER_TRACE(bh, "");
30         if (uptodate)
31                 set_buffer_uptodate(bh);
32         else
33                 clear_buffer_uptodate(bh);
34         unlock_buffer(bh);
35 }
36
37 /*
38  * When an ext3-ordered file is truncated, it is possible that many pages are
39  * not sucessfully freed, because they are attached to a committing transaction.
40  * After the transaction commits, these pages are left on the LRU, with no
41  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
42  * by the VM, but their apparent absence upsets the VM accounting, and it makes
43  * the numbers in /proc/meminfo look odd.
44  *
45  * So here, we have a buffer which has just come off the forget list.  Look to
46  * see if we can strip all buffers from the backing page.
47  *
48  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
49  * caller provided us with a ref against the buffer, and we drop that here.
50  */
51 static void release_buffer_page(struct buffer_head *bh)
52 {
53         struct page *page;
54
55         if (buffer_dirty(bh))
56                 goto nope;
57         if (atomic_read(&bh->b_count) != 1)
58                 goto nope;
59         page = bh->b_page;
60         if (!page)
61                 goto nope;
62         if (page->mapping)
63                 goto nope;
64
65         /* OK, it's a truncated page */
66         if (TestSetPageLocked(page))
67                 goto nope;
68
69         page_cache_get(page);
70         __brelse(bh);
71         try_to_free_buffers(page);
72         unlock_page(page);
73         page_cache_release(page);
74         return;
75
76 nope:
77         __brelse(bh);
78 }
79
80 /*
81  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
82  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
83  * return 0.  j_list_lock is dropped in this case.
84  */
85 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
86 {
87         if (!jbd_trylock_bh_state(bh)) {
88                 spin_unlock(&journal->j_list_lock);
89                 schedule();
90                 return 0;
91         }
92         return 1;
93 }
94
95 /* Done it all: now write the commit record.  We should have
96  * cleaned up our previous buffers by now, so if we are in abort
97  * mode we can now just skip the rest of the journal write
98  * entirely.
99  *
100  * Returns 1 if the journal needs to be aborted or 0 on success
101  */
102 static int journal_write_commit_record(journal_t *journal,
103                                         transaction_t *commit_transaction)
104 {
105         struct journal_head *descriptor;
106         struct buffer_head *bh;
107         int i, ret;
108         int barrier_done = 0;
109
110         if (is_journal_aborted(journal))
111                 return 0;
112
113         descriptor = journal_get_descriptor_buffer(journal);
114         if (!descriptor)
115                 return 1;
116
117         bh = jh2bh(descriptor);
118
119         /* AKPM: buglet - add `i' to tmp! */
120         for (i = 0; i < bh->b_size; i += 512) {
121                 journal_header_t *tmp = (journal_header_t*)bh->b_data;
122                 tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
123                 tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
124                 tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
125         }
126
127         JBUFFER_TRACE(descriptor, "write commit block");
128         set_buffer_dirty(bh);
129         if (journal->j_flags & JFS_BARRIER) {
130                 set_buffer_ordered(bh);
131                 barrier_done = 1;
132         }
133         ret = sync_dirty_buffer(bh);
134         /* is it possible for another commit to fail at roughly
135          * the same time as this one?  If so, we don't want to
136          * trust the barrier flag in the super, but instead want
137          * to remember if we sent a barrier request
138          */
139         if (ret == -EOPNOTSUPP && barrier_done) {
140                 char b[BDEVNAME_SIZE];
141
142                 printk(KERN_WARNING
143                         "JBD: barrier-based sync failed on %s - "
144                         "disabling barriers\n",
145                         bdevname(journal->j_dev, b));
146                 spin_lock(&journal->j_state_lock);
147                 journal->j_flags &= ~JFS_BARRIER;
148                 spin_unlock(&journal->j_state_lock);
149
150                 /* And try again, without the barrier */
151                 clear_buffer_ordered(bh);
152                 set_buffer_uptodate(bh);
153                 set_buffer_dirty(bh);
154                 ret = sync_dirty_buffer(bh);
155         }
156         put_bh(bh);             /* One for getblk() */
157         journal_put_journal_head(descriptor);
158
159         return (ret == -EIO);
160 }
161
162 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
163 {
164         int i;
165
166         for (i = 0; i < bufs; i++) {
167                 wbuf[i]->b_end_io = end_buffer_write_sync;
168                 /* We use-up our safety reference in submit_bh() */
169                 submit_bh(WRITE, wbuf[i]);
170         }
171 }
172
173 /*
174  *  Submit all the data buffers to disk
175  */
176 static void journal_submit_data_buffers(journal_t *journal,
177                                 transaction_t *commit_transaction)
178 {
179         struct journal_head *jh;
180         struct buffer_head *bh;
181         int locked;
182         int bufs = 0;
183         struct buffer_head **wbuf = journal->j_wbuf;
184
185         /*
186          * Whenever we unlock the journal and sleep, things can get added
187          * onto ->t_sync_datalist, so we have to keep looping back to
188          * write_out_data until we *know* that the list is empty.
189          *
190          * Cleanup any flushed data buffers from the data list.  Even in
191          * abort mode, we want to flush this out as soon as possible.
192          */
193 write_out_data:
194         cond_resched();
195         spin_lock(&journal->j_list_lock);
196
197         while (commit_transaction->t_sync_datalist) {
198                 jh = commit_transaction->t_sync_datalist;
199                 bh = jh2bh(jh);
200                 locked = 0;
201
202                 /* Get reference just to make sure buffer does not disappear
203                  * when we are forced to drop various locks */
204                 get_bh(bh);
205                 /* If the buffer is dirty, we need to submit IO and hence
206                  * we need the buffer lock. We try to lock the buffer without
207                  * blocking. If we fail, we need to drop j_list_lock and do
208                  * blocking lock_buffer().
209                  */
210                 if (buffer_dirty(bh)) {
211                         if (test_set_buffer_locked(bh)) {
212                                 BUFFER_TRACE(bh, "needs blocking lock");
213                                 spin_unlock(&journal->j_list_lock);
214                                 /* Write out all data to prevent deadlocks */
215                                 journal_do_submit_data(wbuf, bufs);
216                                 bufs = 0;
217                                 lock_buffer(bh);
218                                 spin_lock(&journal->j_list_lock);
219                         }
220                         locked = 1;
221                 }
222                 /* We have to get bh_state lock. Again out of order, sigh. */
223                 if (!inverted_lock(journal, bh)) {
224                         jbd_lock_bh_state(bh);
225                         spin_lock(&journal->j_list_lock);
226                 }
227                 /* Someone already cleaned up the buffer? */
228                 if (!buffer_jbd(bh)
229                         || jh->b_transaction != commit_transaction
230                         || jh->b_jlist != BJ_SyncData) {
231                         jbd_unlock_bh_state(bh);
232                         if (locked)
233                                 unlock_buffer(bh);
234                         BUFFER_TRACE(bh, "already cleaned up");
235                         put_bh(bh);
236                         continue;
237                 }
238                 if (locked && test_clear_buffer_dirty(bh)) {
239                         BUFFER_TRACE(bh, "needs writeout, adding to array");
240                         wbuf[bufs++] = bh;
241                         __journal_file_buffer(jh, commit_transaction,
242                                                 BJ_Locked);
243                         jbd_unlock_bh_state(bh);
244                         if (bufs == journal->j_wbufsize) {
245                                 spin_unlock(&journal->j_list_lock);
246                                 journal_do_submit_data(wbuf, bufs);
247                                 bufs = 0;
248                                 goto write_out_data;
249                         }
250                 } else if (!locked && buffer_locked(bh)) {
251                         __journal_file_buffer(jh, commit_transaction,
252                                                 BJ_Locked);
253                         jbd_unlock_bh_state(bh);
254                         put_bh(bh);
255                 } else {
256                         BUFFER_TRACE(bh, "writeout complete: unfile");
257                         __journal_unfile_buffer(jh);
258                         jbd_unlock_bh_state(bh);
259                         if (locked)
260                                 unlock_buffer(bh);
261                         journal_remove_journal_head(bh);
262                         /* Once for our safety reference, once for
263                          * journal_remove_journal_head() */
264                         put_bh(bh);
265                         put_bh(bh);
266                 }
267
268                 if (lock_need_resched(&journal->j_list_lock)) {
269                         spin_unlock(&journal->j_list_lock);
270                         goto write_out_data;
271                 }
272         }
273         spin_unlock(&journal->j_list_lock);
274         journal_do_submit_data(wbuf, bufs);
275 }
276
277 /*
278  * journal_commit_transaction
279  *
280  * The primary function for committing a transaction to the log.  This
281  * function is called by the journal thread to begin a complete commit.
282  */
283 void journal_commit_transaction(journal_t *journal)
284 {
285         transaction_t *commit_transaction;
286         struct journal_head *jh, *new_jh, *descriptor;
287         struct buffer_head **wbuf = journal->j_wbuf;
288         int bufs;
289         int flags;
290         int err;
291         unsigned long blocknr;
292         char *tagp = NULL;
293         journal_header_t *header;
294         journal_block_tag_t *tag = NULL;
295         int space_left = 0;
296         int first_tag = 0;
297         int tag_flag;
298         int i;
299
300         /*
301          * First job: lock down the current transaction and wait for
302          * all outstanding updates to complete.
303          */
304
305 #ifdef COMMIT_STATS
306         spin_lock(&journal->j_list_lock);
307         summarise_journal_usage(journal);
308         spin_unlock(&journal->j_list_lock);
309 #endif
310
311         /* Do we need to erase the effects of a prior journal_flush? */
312         if (journal->j_flags & JFS_FLUSHED) {
313                 jbd_debug(3, "super block updated\n");
314                 journal_update_superblock(journal, 1);
315         } else {
316                 jbd_debug(3, "superblock not updated\n");
317         }
318
319         J_ASSERT(journal->j_running_transaction != NULL);
320         J_ASSERT(journal->j_committing_transaction == NULL);
321
322         commit_transaction = journal->j_running_transaction;
323         J_ASSERT(commit_transaction->t_state == T_RUNNING);
324
325         jbd_debug(1, "JBD: starting commit of transaction %d\n",
326                         commit_transaction->t_tid);
327
328         spin_lock(&journal->j_state_lock);
329         commit_transaction->t_state = T_LOCKED;
330
331         spin_lock(&commit_transaction->t_handle_lock);
332         while (commit_transaction->t_updates) {
333                 DEFINE_WAIT(wait);
334
335                 prepare_to_wait(&journal->j_wait_updates, &wait,
336                                         TASK_UNINTERRUPTIBLE);
337                 if (commit_transaction->t_updates) {
338                         spin_unlock(&commit_transaction->t_handle_lock);
339                         spin_unlock(&journal->j_state_lock);
340                         schedule();
341                         spin_lock(&journal->j_state_lock);
342                         spin_lock(&commit_transaction->t_handle_lock);
343                 }
344                 finish_wait(&journal->j_wait_updates, &wait);
345         }
346         spin_unlock(&commit_transaction->t_handle_lock);
347
348         J_ASSERT (commit_transaction->t_outstanding_credits <=
349                         journal->j_max_transaction_buffers);
350
351         /*
352          * First thing we are allowed to do is to discard any remaining
353          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
354          * that there are no such buffers: if a large filesystem
355          * operation like a truncate needs to split itself over multiple
356          * transactions, then it may try to do a journal_restart() while
357          * there are still BJ_Reserved buffers outstanding.  These must
358          * be released cleanly from the current transaction.
359          *
360          * In this case, the filesystem must still reserve write access
361          * again before modifying the buffer in the new transaction, but
362          * we do not require it to remember exactly which old buffers it
363          * has reserved.  This is consistent with the existing behaviour
364          * that multiple journal_get_write_access() calls to the same
365          * buffer are perfectly permissable.
366          */
367         while (commit_transaction->t_reserved_list) {
368                 jh = commit_transaction->t_reserved_list;
369                 JBUFFER_TRACE(jh, "reserved, unused: refile");
370                 /*
371                  * A journal_get_undo_access()+journal_release_buffer() may
372                  * leave undo-committed data.
373                  */
374                 if (jh->b_committed_data) {
375                         struct buffer_head *bh = jh2bh(jh);
376
377                         jbd_lock_bh_state(bh);
378                         jbd_slab_free(jh->b_committed_data, bh->b_size);
379                         jh->b_committed_data = NULL;
380                         jbd_unlock_bh_state(bh);
381                 }
382                 journal_refile_buffer(journal, jh);
383         }
384
385         /*
386          * Now try to drop any written-back buffers from the journal's
387          * checkpoint lists.  We do this *before* commit because it potentially
388          * frees some memory
389          */
390         spin_lock(&journal->j_list_lock);
391         __journal_clean_checkpoint_list(journal);
392         spin_unlock(&journal->j_list_lock);
393
394         jbd_debug (3, "JBD: commit phase 1\n");
395
396         /*
397          * Switch to a new revoke table.
398          */
399         journal_switch_revoke_table(journal);
400
401         commit_transaction->t_state = T_FLUSH;
402         journal->j_committing_transaction = commit_transaction;
403         journal->j_running_transaction = NULL;
404         commit_transaction->t_log_start = journal->j_head;
405         wake_up(&journal->j_wait_transaction_locked);
406         spin_unlock(&journal->j_state_lock);
407
408         jbd_debug (3, "JBD: commit phase 2\n");
409
410         /*
411          * First, drop modified flag: all accesses to the buffers
412          * will be tracked for a new trasaction only -bzzz
413          */
414         spin_lock(&journal->j_list_lock);
415         if (commit_transaction->t_buffers) {
416                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
417                 do {
418                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
419                                         new_jh->b_modified == 0);
420                         new_jh->b_modified = 0;
421                         new_jh = new_jh->b_tnext;
422                 } while (new_jh != jh);
423         }
424         spin_unlock(&journal->j_list_lock);
425
426         /*
427          * Now start flushing things to disk, in the order they appear
428          * on the transaction lists.  Data blocks go first.
429          */
430         err = 0;
431         journal_submit_data_buffers(journal, commit_transaction);
432
433         /*
434          * Wait for all previously submitted IO to complete.
435          */
436         spin_lock(&journal->j_list_lock);
437         while (commit_transaction->t_locked_list) {
438                 struct buffer_head *bh;
439
440                 jh = commit_transaction->t_locked_list->b_tprev;
441                 bh = jh2bh(jh);
442                 get_bh(bh);
443                 if (buffer_locked(bh)) {
444                         spin_unlock(&journal->j_list_lock);
445                         wait_on_buffer(bh);
446                         if (unlikely(!buffer_uptodate(bh)))
447                                 err = -EIO;
448                         spin_lock(&journal->j_list_lock);
449                 }
450                 if (!inverted_lock(journal, bh)) {
451                         put_bh(bh);
452                         spin_lock(&journal->j_list_lock);
453                         continue;
454                 }
455                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
456                         __journal_unfile_buffer(jh);
457                         jbd_unlock_bh_state(bh);
458                         journal_remove_journal_head(bh);
459                         put_bh(bh);
460                 } else {
461                         jbd_unlock_bh_state(bh);
462                 }
463                 put_bh(bh);
464                 cond_resched_lock(&journal->j_list_lock);
465         }
466         spin_unlock(&journal->j_list_lock);
467
468         if (err)
469                 __journal_abort_hard(journal);
470
471         journal_write_revoke_records(journal, commit_transaction);
472
473         jbd_debug(3, "JBD: commit phase 2\n");
474
475         /*
476          * If we found any dirty or locked buffers, then we should have
477          * looped back up to the write_out_data label.  If there weren't
478          * any then journal_clean_data_list should have wiped the list
479          * clean by now, so check that it is in fact empty.
480          */
481         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
482
483         jbd_debug (3, "JBD: commit phase 3\n");
484
485         /*
486          * Way to go: we have now written out all of the data for a
487          * transaction!  Now comes the tricky part: we need to write out
488          * metadata.  Loop over the transaction's entire buffer list:
489          */
490         commit_transaction->t_state = T_COMMIT;
491
492         descriptor = NULL;
493         bufs = 0;
494         while (commit_transaction->t_buffers) {
495
496                 /* Find the next buffer to be journaled... */
497
498                 jh = commit_transaction->t_buffers;
499
500                 /* If we're in abort mode, we just un-journal the buffer and
501                    release it for background writing. */
502
503                 if (is_journal_aborted(journal)) {
504                         JBUFFER_TRACE(jh, "journal is aborting: refile");
505                         journal_refile_buffer(journal, jh);
506                         /* If that was the last one, we need to clean up
507                          * any descriptor buffers which may have been
508                          * already allocated, even if we are now
509                          * aborting. */
510                         if (!commit_transaction->t_buffers)
511                                 goto start_journal_io;
512                         continue;
513                 }
514
515                 /* Make sure we have a descriptor block in which to
516                    record the metadata buffer. */
517
518                 if (!descriptor) {
519                         struct buffer_head *bh;
520
521                         J_ASSERT (bufs == 0);
522
523                         jbd_debug(4, "JBD: get descriptor\n");
524
525                         descriptor = journal_get_descriptor_buffer(journal);
526                         if (!descriptor) {
527                                 __journal_abort_hard(journal);
528                                 continue;
529                         }
530
531                         bh = jh2bh(descriptor);
532                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
533                                 (unsigned long long)bh->b_blocknr, bh->b_data);
534                         header = (journal_header_t *)&bh->b_data[0];
535                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
536                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
537                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
538
539                         tagp = &bh->b_data[sizeof(journal_header_t)];
540                         space_left = bh->b_size - sizeof(journal_header_t);
541                         first_tag = 1;
542                         set_buffer_jwrite(bh);
543                         set_buffer_dirty(bh);
544                         wbuf[bufs++] = bh;
545
546                         /* Record it so that we can wait for IO
547                            completion later */
548                         BUFFER_TRACE(bh, "ph3: file as descriptor");
549                         journal_file_buffer(descriptor, commit_transaction,
550                                         BJ_LogCtl);
551                 }
552
553                 /* Where is the buffer to be written? */
554
555                 err = journal_next_log_block(journal, &blocknr);
556                 /* If the block mapping failed, just abandon the buffer
557                    and repeat this loop: we'll fall into the
558                    refile-on-abort condition above. */
559                 if (err) {
560                         __journal_abort_hard(journal);
561                         continue;
562                 }
563
564                 /*
565                  * start_this_handle() uses t_outstanding_credits to determine
566                  * the free space in the log, but this counter is changed
567                  * by journal_next_log_block() also.
568                  */
569                 commit_transaction->t_outstanding_credits--;
570
571                 /* Bump b_count to prevent truncate from stumbling over
572                    the shadowed buffer!  @@@ This can go if we ever get
573                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
574                 atomic_inc(&jh2bh(jh)->b_count);
575
576                 /* Make a temporary IO buffer with which to write it out
577                    (this will requeue both the metadata buffer and the
578                    temporary IO buffer). new_bh goes on BJ_IO*/
579
580                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
581                 /*
582                  * akpm: journal_write_metadata_buffer() sets
583                  * new_bh->b_transaction to commit_transaction.
584                  * We need to clean this up before we release new_bh
585                  * (which is of type BJ_IO)
586                  */
587                 JBUFFER_TRACE(jh, "ph3: write metadata");
588                 flags = journal_write_metadata_buffer(commit_transaction,
589                                                       jh, &new_jh, blocknr);
590                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
591                 wbuf[bufs++] = jh2bh(new_jh);
592
593                 /* Record the new block's tag in the current descriptor
594                    buffer */
595
596                 tag_flag = 0;
597                 if (flags & 1)
598                         tag_flag |= JFS_FLAG_ESCAPE;
599                 if (!first_tag)
600                         tag_flag |= JFS_FLAG_SAME_UUID;
601
602                 tag = (journal_block_tag_t *) tagp;
603                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
604                 tag->t_flags = cpu_to_be32(tag_flag);
605                 tagp += sizeof(journal_block_tag_t);
606                 space_left -= sizeof(journal_block_tag_t);
607
608                 if (first_tag) {
609                         memcpy (tagp, journal->j_uuid, 16);
610                         tagp += 16;
611                         space_left -= 16;
612                         first_tag = 0;
613                 }
614
615                 /* If there's no more to do, or if the descriptor is full,
616                    let the IO rip! */
617
618                 if (bufs == journal->j_wbufsize ||
619                     commit_transaction->t_buffers == NULL ||
620                     space_left < sizeof(journal_block_tag_t) + 16) {
621
622                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
623
624                         /* Write an end-of-descriptor marker before
625                            submitting the IOs.  "tag" still points to
626                            the last tag we set up. */
627
628                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
629
630 start_journal_io:
631                         for (i = 0; i < bufs; i++) {
632                                 struct buffer_head *bh = wbuf[i];
633                                 lock_buffer(bh);
634                                 clear_buffer_dirty(bh);
635                                 set_buffer_uptodate(bh);
636                                 bh->b_end_io = journal_end_buffer_io_sync;
637                                 submit_bh(WRITE, bh);
638                         }
639                         cond_resched();
640
641                         /* Force a new descriptor to be generated next
642                            time round the loop. */
643                         descriptor = NULL;
644                         bufs = 0;
645                 }
646         }
647
648         /* Lo and behold: we have just managed to send a transaction to
649            the log.  Before we can commit it, wait for the IO so far to
650            complete.  Control buffers being written are on the
651            transaction's t_log_list queue, and metadata buffers are on
652            the t_iobuf_list queue.
653
654            Wait for the buffers in reverse order.  That way we are
655            less likely to be woken up until all IOs have completed, and
656            so we incur less scheduling load.
657         */
658
659         jbd_debug(3, "JBD: commit phase 4\n");
660
661         /*
662          * akpm: these are BJ_IO, and j_list_lock is not needed.
663          * See __journal_try_to_free_buffer.
664          */
665 wait_for_iobuf:
666         while (commit_transaction->t_iobuf_list != NULL) {
667                 struct buffer_head *bh;
668
669                 jh = commit_transaction->t_iobuf_list->b_tprev;
670                 bh = jh2bh(jh);
671                 if (buffer_locked(bh)) {
672                         wait_on_buffer(bh);
673                         goto wait_for_iobuf;
674                 }
675                 if (cond_resched())
676                         goto wait_for_iobuf;
677
678                 if (unlikely(!buffer_uptodate(bh)))
679                         err = -EIO;
680
681                 clear_buffer_jwrite(bh);
682
683                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
684                 journal_unfile_buffer(journal, jh);
685
686                 /*
687                  * ->t_iobuf_list should contain only dummy buffer_heads
688                  * which were created by journal_write_metadata_buffer().
689                  */
690                 BUFFER_TRACE(bh, "dumping temporary bh");
691                 journal_put_journal_head(jh);
692                 __brelse(bh);
693                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
694                 free_buffer_head(bh);
695
696                 /* We also have to unlock and free the corresponding
697                    shadowed buffer */
698                 jh = commit_transaction->t_shadow_list->b_tprev;
699                 bh = jh2bh(jh);
700                 clear_bit(BH_JWrite, &bh->b_state);
701                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
702
703                 /* The metadata is now released for reuse, but we need
704                    to remember it against this transaction so that when
705                    we finally commit, we can do any checkpointing
706                    required. */
707                 JBUFFER_TRACE(jh, "file as BJ_Forget");
708                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
709                 /* Wake up any transactions which were waiting for this
710                    IO to complete */
711                 wake_up_bit(&bh->b_state, BH_Unshadow);
712                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
713                 __brelse(bh);
714         }
715
716         J_ASSERT (commit_transaction->t_shadow_list == NULL);
717
718         jbd_debug(3, "JBD: commit phase 5\n");
719
720         /* Here we wait for the revoke record and descriptor record buffers */
721  wait_for_ctlbuf:
722         while (commit_transaction->t_log_list != NULL) {
723                 struct buffer_head *bh;
724
725                 jh = commit_transaction->t_log_list->b_tprev;
726                 bh = jh2bh(jh);
727                 if (buffer_locked(bh)) {
728                         wait_on_buffer(bh);
729                         goto wait_for_ctlbuf;
730                 }
731                 if (cond_resched())
732                         goto wait_for_ctlbuf;
733
734                 if (unlikely(!buffer_uptodate(bh)))
735                         err = -EIO;
736
737                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
738                 clear_buffer_jwrite(bh);
739                 journal_unfile_buffer(journal, jh);
740                 journal_put_journal_head(jh);
741                 __brelse(bh);           /* One for getblk */
742                 /* AKPM: bforget here */
743         }
744
745         jbd_debug(3, "JBD: commit phase 6\n");
746
747         if (journal_write_commit_record(journal, commit_transaction))
748                 err = -EIO;
749
750         if (err)
751                 __journal_abort_hard(journal);
752
753         /* End of a transaction!  Finally, we can do checkpoint
754            processing: any buffers committed as a result of this
755            transaction can be removed from any checkpoint list it was on
756            before. */
757
758         jbd_debug(3, "JBD: commit phase 7\n");
759
760         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
761         J_ASSERT(commit_transaction->t_buffers == NULL);
762         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
763         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
764         J_ASSERT(commit_transaction->t_shadow_list == NULL);
765         J_ASSERT(commit_transaction->t_log_list == NULL);
766
767 restart_loop:
768         /*
769          * As there are other places (journal_unmap_buffer()) adding buffers
770          * to this list we have to be careful and hold the j_list_lock.
771          */
772         spin_lock(&journal->j_list_lock);
773         while (commit_transaction->t_forget) {
774                 transaction_t *cp_transaction;
775                 struct buffer_head *bh;
776
777                 jh = commit_transaction->t_forget;
778                 spin_unlock(&journal->j_list_lock);
779                 bh = jh2bh(jh);
780                 jbd_lock_bh_state(bh);
781                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
782                         jh->b_transaction == journal->j_running_transaction);
783
784                 /*
785                  * If there is undo-protected committed data against
786                  * this buffer, then we can remove it now.  If it is a
787                  * buffer needing such protection, the old frozen_data
788                  * field now points to a committed version of the
789                  * buffer, so rotate that field to the new committed
790                  * data.
791                  *
792                  * Otherwise, we can just throw away the frozen data now.
793                  */
794                 if (jh->b_committed_data) {
795                         jbd_slab_free(jh->b_committed_data, bh->b_size);
796                         jh->b_committed_data = NULL;
797                         if (jh->b_frozen_data) {
798                                 jh->b_committed_data = jh->b_frozen_data;
799                                 jh->b_frozen_data = NULL;
800                         }
801                 } else if (jh->b_frozen_data) {
802                         jbd_slab_free(jh->b_frozen_data, bh->b_size);
803                         jh->b_frozen_data = NULL;
804                 }
805
806                 spin_lock(&journal->j_list_lock);
807                 cp_transaction = jh->b_cp_transaction;
808                 if (cp_transaction) {
809                         JBUFFER_TRACE(jh, "remove from old cp transaction");
810                         __journal_remove_checkpoint(jh);
811                 }
812
813                 /* Only re-checkpoint the buffer_head if it is marked
814                  * dirty.  If the buffer was added to the BJ_Forget list
815                  * by journal_forget, it may no longer be dirty and
816                  * there's no point in keeping a checkpoint record for
817                  * it. */
818
819                 /* A buffer which has been freed while still being
820                  * journaled by a previous transaction may end up still
821                  * being dirty here, but we want to avoid writing back
822                  * that buffer in the future now that the last use has
823                  * been committed.  That's not only a performance gain,
824                  * it also stops aliasing problems if the buffer is left
825                  * behind for writeback and gets reallocated for another
826                  * use in a different page. */
827                 if (buffer_freed(bh)) {
828                         clear_buffer_freed(bh);
829                         clear_buffer_jbddirty(bh);
830                 }
831
832                 if (buffer_jbddirty(bh)) {
833                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
834                         __journal_insert_checkpoint(jh, commit_transaction);
835                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
836                         __journal_refile_buffer(jh);
837                         jbd_unlock_bh_state(bh);
838                 } else {
839                         J_ASSERT_BH(bh, !buffer_dirty(bh));
840                         /* The buffer on BJ_Forget list and not jbddirty means
841                          * it has been freed by this transaction and hence it
842                          * could not have been reallocated until this
843                          * transaction has committed. *BUT* it could be
844                          * reallocated once we have written all the data to
845                          * disk and before we process the buffer on BJ_Forget
846                          * list. */
847                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
848                         __journal_refile_buffer(jh);
849                         if (!jh->b_transaction) {
850                                 jbd_unlock_bh_state(bh);
851                                  /* needs a brelse */
852                                 journal_remove_journal_head(bh);
853                                 release_buffer_page(bh);
854                         } else
855                                 jbd_unlock_bh_state(bh);
856                 }
857                 cond_resched_lock(&journal->j_list_lock);
858         }
859         spin_unlock(&journal->j_list_lock);
860         /*
861          * This is a bit sleazy.  We borrow j_list_lock to protect
862          * journal->j_committing_transaction in __journal_remove_checkpoint.
863          * Really, __journal_remove_checkpoint should be using j_state_lock but
864          * it's a bit hassle to hold that across __journal_remove_checkpoint
865          */
866         spin_lock(&journal->j_state_lock);
867         spin_lock(&journal->j_list_lock);
868         /*
869          * Now recheck if some buffers did not get attached to the transaction
870          * while the lock was dropped...
871          */
872         if (commit_transaction->t_forget) {
873                 spin_unlock(&journal->j_list_lock);
874                 spin_unlock(&journal->j_state_lock);
875                 goto restart_loop;
876         }
877
878         /* Done with this transaction! */
879
880         jbd_debug(3, "JBD: commit phase 8\n");
881
882         J_ASSERT(commit_transaction->t_state == T_COMMIT);
883
884         commit_transaction->t_state = T_FINISHED;
885         J_ASSERT(commit_transaction == journal->j_committing_transaction);
886         journal->j_commit_sequence = commit_transaction->t_tid;
887         journal->j_committing_transaction = NULL;
888         spin_unlock(&journal->j_state_lock);
889
890         if (commit_transaction->t_checkpoint_list == NULL &&
891             commit_transaction->t_checkpoint_io_list == NULL) {
892                 __journal_drop_transaction(journal, commit_transaction);
893         } else {
894                 if (journal->j_checkpoint_transactions == NULL) {
895                         journal->j_checkpoint_transactions = commit_transaction;
896                         commit_transaction->t_cpnext = commit_transaction;
897                         commit_transaction->t_cpprev = commit_transaction;
898                 } else {
899                         commit_transaction->t_cpnext =
900                                 journal->j_checkpoint_transactions;
901                         commit_transaction->t_cpprev =
902                                 commit_transaction->t_cpnext->t_cpprev;
903                         commit_transaction->t_cpnext->t_cpprev =
904                                 commit_transaction;
905                         commit_transaction->t_cpprev->t_cpnext =
906                                 commit_transaction;
907                 }
908         }
909         spin_unlock(&journal->j_list_lock);
910
911         jbd_debug(1, "JBD: commit %d complete, head %d\n",
912                   journal->j_commit_sequence, journal->j_tail_sequence);
913
914         wake_up(&journal->j_wait_done_commit);
915 }