Auto-update from upstream
[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 /*
164  * journal_commit_transaction
165  *
166  * The primary function for committing a transaction to the log.  This
167  * function is called by the journal thread to begin a complete commit.
168  */
169 void journal_commit_transaction(journal_t *journal)
170 {
171         transaction_t *commit_transaction;
172         struct journal_head *jh, *new_jh, *descriptor;
173         struct buffer_head **wbuf = journal->j_wbuf;
174         int bufs;
175         int flags;
176         int err;
177         unsigned long blocknr;
178         char *tagp = NULL;
179         journal_header_t *header;
180         journal_block_tag_t *tag = NULL;
181         int space_left = 0;
182         int first_tag = 0;
183         int tag_flag;
184         int i;
185
186         /*
187          * First job: lock down the current transaction and wait for
188          * all outstanding updates to complete.
189          */
190
191 #ifdef COMMIT_STATS
192         spin_lock(&journal->j_list_lock);
193         summarise_journal_usage(journal);
194         spin_unlock(&journal->j_list_lock);
195 #endif
196
197         /* Do we need to erase the effects of a prior journal_flush? */
198         if (journal->j_flags & JFS_FLUSHED) {
199                 jbd_debug(3, "super block updated\n");
200                 journal_update_superblock(journal, 1);
201         } else {
202                 jbd_debug(3, "superblock not updated\n");
203         }
204
205         J_ASSERT(journal->j_running_transaction != NULL);
206         J_ASSERT(journal->j_committing_transaction == NULL);
207
208         commit_transaction = journal->j_running_transaction;
209         J_ASSERT(commit_transaction->t_state == T_RUNNING);
210
211         jbd_debug(1, "JBD: starting commit of transaction %d\n",
212                         commit_transaction->t_tid);
213
214         spin_lock(&journal->j_state_lock);
215         commit_transaction->t_state = T_LOCKED;
216
217         spin_lock(&commit_transaction->t_handle_lock);
218         while (commit_transaction->t_updates) {
219                 DEFINE_WAIT(wait);
220
221                 prepare_to_wait(&journal->j_wait_updates, &wait,
222                                         TASK_UNINTERRUPTIBLE);
223                 if (commit_transaction->t_updates) {
224                         spin_unlock(&commit_transaction->t_handle_lock);
225                         spin_unlock(&journal->j_state_lock);
226                         schedule();
227                         spin_lock(&journal->j_state_lock);
228                         spin_lock(&commit_transaction->t_handle_lock);
229                 }
230                 finish_wait(&journal->j_wait_updates, &wait);
231         }
232         spin_unlock(&commit_transaction->t_handle_lock);
233
234         J_ASSERT (commit_transaction->t_outstanding_credits <=
235                         journal->j_max_transaction_buffers);
236
237         /*
238          * First thing we are allowed to do is to discard any remaining
239          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
240          * that there are no such buffers: if a large filesystem
241          * operation like a truncate needs to split itself over multiple
242          * transactions, then it may try to do a journal_restart() while
243          * there are still BJ_Reserved buffers outstanding.  These must
244          * be released cleanly from the current transaction.
245          *
246          * In this case, the filesystem must still reserve write access
247          * again before modifying the buffer in the new transaction, but
248          * we do not require it to remember exactly which old buffers it
249          * has reserved.  This is consistent with the existing behaviour
250          * that multiple journal_get_write_access() calls to the same
251          * buffer are perfectly permissable.
252          */
253         while (commit_transaction->t_reserved_list) {
254                 jh = commit_transaction->t_reserved_list;
255                 JBUFFER_TRACE(jh, "reserved, unused: refile");
256                 /*
257                  * A journal_get_undo_access()+journal_release_buffer() may
258                  * leave undo-committed data.
259                  */
260                 if (jh->b_committed_data) {
261                         struct buffer_head *bh = jh2bh(jh);
262
263                         jbd_lock_bh_state(bh);
264                         kfree(jh->b_committed_data);
265                         jh->b_committed_data = NULL;
266                         jbd_unlock_bh_state(bh);
267                 }
268                 journal_refile_buffer(journal, jh);
269         }
270
271         /*
272          * Now try to drop any written-back buffers from the journal's
273          * checkpoint lists.  We do this *before* commit because it potentially
274          * frees some memory
275          */
276         spin_lock(&journal->j_list_lock);
277         __journal_clean_checkpoint_list(journal);
278         spin_unlock(&journal->j_list_lock);
279
280         jbd_debug (3, "JBD: commit phase 1\n");
281
282         /*
283          * Switch to a new revoke table.
284          */
285         journal_switch_revoke_table(journal);
286
287         commit_transaction->t_state = T_FLUSH;
288         journal->j_committing_transaction = commit_transaction;
289         journal->j_running_transaction = NULL;
290         commit_transaction->t_log_start = journal->j_head;
291         wake_up(&journal->j_wait_transaction_locked);
292         spin_unlock(&journal->j_state_lock);
293
294         jbd_debug (3, "JBD: commit phase 2\n");
295
296         /*
297          * First, drop modified flag: all accesses to the buffers
298          * will be tracked for a new trasaction only -bzzz
299          */
300         spin_lock(&journal->j_list_lock);
301         if (commit_transaction->t_buffers) {
302                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
303                 do {
304                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
305                                         new_jh->b_modified == 0);
306                         new_jh->b_modified = 0;
307                         new_jh = new_jh->b_tnext;
308                 } while (new_jh != jh);
309         }
310         spin_unlock(&journal->j_list_lock);
311
312         /*
313          * Now start flushing things to disk, in the order they appear
314          * on the transaction lists.  Data blocks go first.
315          */
316
317         err = 0;
318         /*
319          * Whenever we unlock the journal and sleep, things can get added
320          * onto ->t_sync_datalist, so we have to keep looping back to
321          * write_out_data until we *know* that the list is empty.
322          */
323         bufs = 0;
324         /*
325          * Cleanup any flushed data buffers from the data list.  Even in
326          * abort mode, we want to flush this out as soon as possible.
327          */
328 write_out_data:
329         cond_resched();
330         spin_lock(&journal->j_list_lock);
331
332         while (commit_transaction->t_sync_datalist) {
333                 struct buffer_head *bh;
334
335                 jh = commit_transaction->t_sync_datalist;
336                 commit_transaction->t_sync_datalist = jh->b_tnext;
337                 bh = jh2bh(jh);
338                 if (buffer_locked(bh)) {
339                         BUFFER_TRACE(bh, "locked");
340                         if (!inverted_lock(journal, bh))
341                                 goto write_out_data;
342                         __journal_temp_unlink_buffer(jh);
343                         __journal_file_buffer(jh, commit_transaction,
344                                                 BJ_Locked);
345                         jbd_unlock_bh_state(bh);
346                         if (lock_need_resched(&journal->j_list_lock)) {
347                                 spin_unlock(&journal->j_list_lock);
348                                 goto write_out_data;
349                         }
350                 } else {
351                         if (buffer_dirty(bh)) {
352                                 BUFFER_TRACE(bh, "start journal writeout");
353                                 get_bh(bh);
354                                 wbuf[bufs++] = bh;
355                                 if (bufs == journal->j_wbufsize) {
356                                         jbd_debug(2, "submit %d writes\n",
357                                                         bufs);
358                                         spin_unlock(&journal->j_list_lock);
359                                         ll_rw_block(SWRITE, bufs, wbuf);
360                                         journal_brelse_array(wbuf, bufs);
361                                         bufs = 0;
362                                         goto write_out_data;
363                                 }
364                         } else {
365                                 BUFFER_TRACE(bh, "writeout complete: unfile");
366                                 if (!inverted_lock(journal, bh))
367                                         goto write_out_data;
368                                 __journal_unfile_buffer(jh);
369                                 jbd_unlock_bh_state(bh);
370                                 journal_remove_journal_head(bh);
371                                 put_bh(bh);
372                                 if (lock_need_resched(&journal->j_list_lock)) {
373                                         spin_unlock(&journal->j_list_lock);
374                                         goto write_out_data;
375                                 }
376                         }
377                 }
378         }
379
380         if (bufs) {
381                 spin_unlock(&journal->j_list_lock);
382                 ll_rw_block(SWRITE, bufs, wbuf);
383                 journal_brelse_array(wbuf, bufs);
384                 spin_lock(&journal->j_list_lock);
385         }
386
387         /*
388          * Wait for all previously submitted IO to complete.
389          */
390         while (commit_transaction->t_locked_list) {
391                 struct buffer_head *bh;
392
393                 jh = commit_transaction->t_locked_list->b_tprev;
394                 bh = jh2bh(jh);
395                 get_bh(bh);
396                 if (buffer_locked(bh)) {
397                         spin_unlock(&journal->j_list_lock);
398                         wait_on_buffer(bh);
399                         if (unlikely(!buffer_uptodate(bh)))
400                                 err = -EIO;
401                         spin_lock(&journal->j_list_lock);
402                 }
403                 if (!inverted_lock(journal, bh)) {
404                         put_bh(bh);
405                         spin_lock(&journal->j_list_lock);
406                         continue;
407                 }
408                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
409                         __journal_unfile_buffer(jh);
410                         jbd_unlock_bh_state(bh);
411                         journal_remove_journal_head(bh);
412                         put_bh(bh);
413                 } else {
414                         jbd_unlock_bh_state(bh);
415                 }
416                 put_bh(bh);
417                 cond_resched_lock(&journal->j_list_lock);
418         }
419         spin_unlock(&journal->j_list_lock);
420
421         if (err)
422                 __journal_abort_hard(journal);
423
424         journal_write_revoke_records(journal, commit_transaction);
425
426         jbd_debug(3, "JBD: commit phase 2\n");
427
428         /*
429          * If we found any dirty or locked buffers, then we should have
430          * looped back up to the write_out_data label.  If there weren't
431          * any then journal_clean_data_list should have wiped the list
432          * clean by now, so check that it is in fact empty.
433          */
434         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
435
436         jbd_debug (3, "JBD: commit phase 3\n");
437
438         /*
439          * Way to go: we have now written out all of the data for a
440          * transaction!  Now comes the tricky part: we need to write out
441          * metadata.  Loop over the transaction's entire buffer list:
442          */
443         commit_transaction->t_state = T_COMMIT;
444
445         descriptor = NULL;
446         bufs = 0;
447         while (commit_transaction->t_buffers) {
448
449                 /* Find the next buffer to be journaled... */
450
451                 jh = commit_transaction->t_buffers;
452
453                 /* If we're in abort mode, we just un-journal the buffer and
454                    release it for background writing. */
455
456                 if (is_journal_aborted(journal)) {
457                         JBUFFER_TRACE(jh, "journal is aborting: refile");
458                         journal_refile_buffer(journal, jh);
459                         /* If that was the last one, we need to clean up
460                          * any descriptor buffers which may have been
461                          * already allocated, even if we are now
462                          * aborting. */
463                         if (!commit_transaction->t_buffers)
464                                 goto start_journal_io;
465                         continue;
466                 }
467
468                 /* Make sure we have a descriptor block in which to
469                    record the metadata buffer. */
470
471                 if (!descriptor) {
472                         struct buffer_head *bh;
473
474                         J_ASSERT (bufs == 0);
475
476                         jbd_debug(4, "JBD: get descriptor\n");
477
478                         descriptor = journal_get_descriptor_buffer(journal);
479                         if (!descriptor) {
480                                 __journal_abort_hard(journal);
481                                 continue;
482                         }
483
484                         bh = jh2bh(descriptor);
485                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
486                                 (unsigned long long)bh->b_blocknr, bh->b_data);
487                         header = (journal_header_t *)&bh->b_data[0];
488                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
489                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
490                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
491
492                         tagp = &bh->b_data[sizeof(journal_header_t)];
493                         space_left = bh->b_size - sizeof(journal_header_t);
494                         first_tag = 1;
495                         set_buffer_jwrite(bh);
496                         set_buffer_dirty(bh);
497                         wbuf[bufs++] = bh;
498
499                         /* Record it so that we can wait for IO
500                            completion later */
501                         BUFFER_TRACE(bh, "ph3: file as descriptor");
502                         journal_file_buffer(descriptor, commit_transaction,
503                                         BJ_LogCtl);
504                 }
505
506                 /* Where is the buffer to be written? */
507
508                 err = journal_next_log_block(journal, &blocknr);
509                 /* If the block mapping failed, just abandon the buffer
510                    and repeat this loop: we'll fall into the
511                    refile-on-abort condition above. */
512                 if (err) {
513                         __journal_abort_hard(journal);
514                         continue;
515                 }
516
517                 /*
518                  * start_this_handle() uses t_outstanding_credits to determine
519                  * the free space in the log, but this counter is changed
520                  * by journal_next_log_block() also.
521                  */
522                 commit_transaction->t_outstanding_credits--;
523
524                 /* Bump b_count to prevent truncate from stumbling over
525                    the shadowed buffer!  @@@ This can go if we ever get
526                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
527                 atomic_inc(&jh2bh(jh)->b_count);
528
529                 /* Make a temporary IO buffer with which to write it out
530                    (this will requeue both the metadata buffer and the
531                    temporary IO buffer). new_bh goes on BJ_IO*/
532
533                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
534                 /*
535                  * akpm: journal_write_metadata_buffer() sets
536                  * new_bh->b_transaction to commit_transaction.
537                  * We need to clean this up before we release new_bh
538                  * (which is of type BJ_IO)
539                  */
540                 JBUFFER_TRACE(jh, "ph3: write metadata");
541                 flags = journal_write_metadata_buffer(commit_transaction,
542                                                       jh, &new_jh, blocknr);
543                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
544                 wbuf[bufs++] = jh2bh(new_jh);
545
546                 /* Record the new block's tag in the current descriptor
547                    buffer */
548
549                 tag_flag = 0;
550                 if (flags & 1)
551                         tag_flag |= JFS_FLAG_ESCAPE;
552                 if (!first_tag)
553                         tag_flag |= JFS_FLAG_SAME_UUID;
554
555                 tag = (journal_block_tag_t *) tagp;
556                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
557                 tag->t_flags = cpu_to_be32(tag_flag);
558                 tagp += sizeof(journal_block_tag_t);
559                 space_left -= sizeof(journal_block_tag_t);
560
561                 if (first_tag) {
562                         memcpy (tagp, journal->j_uuid, 16);
563                         tagp += 16;
564                         space_left -= 16;
565                         first_tag = 0;
566                 }
567
568                 /* If there's no more to do, or if the descriptor is full,
569                    let the IO rip! */
570
571                 if (bufs == journal->j_wbufsize ||
572                     commit_transaction->t_buffers == NULL ||
573                     space_left < sizeof(journal_block_tag_t) + 16) {
574
575                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
576
577                         /* Write an end-of-descriptor marker before
578                            submitting the IOs.  "tag" still points to
579                            the last tag we set up. */
580
581                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
582
583 start_journal_io:
584                         for (i = 0; i < bufs; i++) {
585                                 struct buffer_head *bh = wbuf[i];
586                                 lock_buffer(bh);
587                                 clear_buffer_dirty(bh);
588                                 set_buffer_uptodate(bh);
589                                 bh->b_end_io = journal_end_buffer_io_sync;
590                                 submit_bh(WRITE, bh);
591                         }
592                         cond_resched();
593
594                         /* Force a new descriptor to be generated next
595                            time round the loop. */
596                         descriptor = NULL;
597                         bufs = 0;
598                 }
599         }
600
601         /* Lo and behold: we have just managed to send a transaction to
602            the log.  Before we can commit it, wait for the IO so far to
603            complete.  Control buffers being written are on the
604            transaction's t_log_list queue, and metadata buffers are on
605            the t_iobuf_list queue.
606
607            Wait for the buffers in reverse order.  That way we are
608            less likely to be woken up until all IOs have completed, and
609            so we incur less scheduling load.
610         */
611
612         jbd_debug(3, "JBD: commit phase 4\n");
613
614         /*
615          * akpm: these are BJ_IO, and j_list_lock is not needed.
616          * See __journal_try_to_free_buffer.
617          */
618 wait_for_iobuf:
619         while (commit_transaction->t_iobuf_list != NULL) {
620                 struct buffer_head *bh;
621
622                 jh = commit_transaction->t_iobuf_list->b_tprev;
623                 bh = jh2bh(jh);
624                 if (buffer_locked(bh)) {
625                         wait_on_buffer(bh);
626                         goto wait_for_iobuf;
627                 }
628                 if (cond_resched())
629                         goto wait_for_iobuf;
630
631                 if (unlikely(!buffer_uptodate(bh)))
632                         err = -EIO;
633
634                 clear_buffer_jwrite(bh);
635
636                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
637                 journal_unfile_buffer(journal, jh);
638
639                 /*
640                  * ->t_iobuf_list should contain only dummy buffer_heads
641                  * which were created by journal_write_metadata_buffer().
642                  */
643                 BUFFER_TRACE(bh, "dumping temporary bh");
644                 journal_put_journal_head(jh);
645                 __brelse(bh);
646                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
647                 free_buffer_head(bh);
648
649                 /* We also have to unlock and free the corresponding
650                    shadowed buffer */
651                 jh = commit_transaction->t_shadow_list->b_tprev;
652                 bh = jh2bh(jh);
653                 clear_bit(BH_JWrite, &bh->b_state);
654                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
655
656                 /* The metadata is now released for reuse, but we need
657                    to remember it against this transaction so that when
658                    we finally commit, we can do any checkpointing
659                    required. */
660                 JBUFFER_TRACE(jh, "file as BJ_Forget");
661                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
662                 /* Wake up any transactions which were waiting for this
663                    IO to complete */
664                 wake_up_bit(&bh->b_state, BH_Unshadow);
665                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
666                 __brelse(bh);
667         }
668
669         J_ASSERT (commit_transaction->t_shadow_list == NULL);
670
671         jbd_debug(3, "JBD: commit phase 5\n");
672
673         /* Here we wait for the revoke record and descriptor record buffers */
674  wait_for_ctlbuf:
675         while (commit_transaction->t_log_list != NULL) {
676                 struct buffer_head *bh;
677
678                 jh = commit_transaction->t_log_list->b_tprev;
679                 bh = jh2bh(jh);
680                 if (buffer_locked(bh)) {
681                         wait_on_buffer(bh);
682                         goto wait_for_ctlbuf;
683                 }
684                 if (cond_resched())
685                         goto wait_for_ctlbuf;
686
687                 if (unlikely(!buffer_uptodate(bh)))
688                         err = -EIO;
689
690                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
691                 clear_buffer_jwrite(bh);
692                 journal_unfile_buffer(journal, jh);
693                 journal_put_journal_head(jh);
694                 __brelse(bh);           /* One for getblk */
695                 /* AKPM: bforget here */
696         }
697
698         jbd_debug(3, "JBD: commit phase 6\n");
699
700         if (journal_write_commit_record(journal, commit_transaction))
701                 err = -EIO;
702
703         if (err)
704                 __journal_abort_hard(journal);
705
706         /* End of a transaction!  Finally, we can do checkpoint
707            processing: any buffers committed as a result of this
708            transaction can be removed from any checkpoint list it was on
709            before. */
710
711         jbd_debug(3, "JBD: commit phase 7\n");
712
713         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
714         J_ASSERT(commit_transaction->t_buffers == NULL);
715         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
716         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
717         J_ASSERT(commit_transaction->t_shadow_list == NULL);
718         J_ASSERT(commit_transaction->t_log_list == NULL);
719
720 restart_loop:
721         /*
722          * As there are other places (journal_unmap_buffer()) adding buffers
723          * to this list we have to be careful and hold the j_list_lock.
724          */
725         spin_lock(&journal->j_list_lock);
726         while (commit_transaction->t_forget) {
727                 transaction_t *cp_transaction;
728                 struct buffer_head *bh;
729
730                 jh = commit_transaction->t_forget;
731                 spin_unlock(&journal->j_list_lock);
732                 bh = jh2bh(jh);
733                 jbd_lock_bh_state(bh);
734                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
735                         jh->b_transaction == journal->j_running_transaction);
736
737                 /*
738                  * If there is undo-protected committed data against
739                  * this buffer, then we can remove it now.  If it is a
740                  * buffer needing such protection, the old frozen_data
741                  * field now points to a committed version of the
742                  * buffer, so rotate that field to the new committed
743                  * data.
744                  *
745                  * Otherwise, we can just throw away the frozen data now.
746                  */
747                 if (jh->b_committed_data) {
748                         kfree(jh->b_committed_data);
749                         jh->b_committed_data = NULL;
750                         if (jh->b_frozen_data) {
751                                 jh->b_committed_data = jh->b_frozen_data;
752                                 jh->b_frozen_data = NULL;
753                         }
754                 } else if (jh->b_frozen_data) {
755                         kfree(jh->b_frozen_data);
756                         jh->b_frozen_data = NULL;
757                 }
758
759                 spin_lock(&journal->j_list_lock);
760                 cp_transaction = jh->b_cp_transaction;
761                 if (cp_transaction) {
762                         JBUFFER_TRACE(jh, "remove from old cp transaction");
763                         __journal_remove_checkpoint(jh);
764                 }
765
766                 /* Only re-checkpoint the buffer_head if it is marked
767                  * dirty.  If the buffer was added to the BJ_Forget list
768                  * by journal_forget, it may no longer be dirty and
769                  * there's no point in keeping a checkpoint record for
770                  * it. */
771
772                 /* A buffer which has been freed while still being
773                  * journaled by a previous transaction may end up still
774                  * being dirty here, but we want to avoid writing back
775                  * that buffer in the future now that the last use has
776                  * been committed.  That's not only a performance gain,
777                  * it also stops aliasing problems if the buffer is left
778                  * behind for writeback and gets reallocated for another
779                  * use in a different page. */
780                 if (buffer_freed(bh)) {
781                         clear_buffer_freed(bh);
782                         clear_buffer_jbddirty(bh);
783                 }
784
785                 if (buffer_jbddirty(bh)) {
786                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
787                         __journal_insert_checkpoint(jh, commit_transaction);
788                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
789                         __journal_refile_buffer(jh);
790                         jbd_unlock_bh_state(bh);
791                 } else {
792                         J_ASSERT_BH(bh, !buffer_dirty(bh));
793                         /* The buffer on BJ_Forget list and not jbddirty means
794                          * it has been freed by this transaction and hence it
795                          * could not have been reallocated until this
796                          * transaction has committed. *BUT* it could be
797                          * reallocated once we have written all the data to
798                          * disk and before we process the buffer on BJ_Forget
799                          * list. */
800                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
801                         __journal_refile_buffer(jh);
802                         if (!jh->b_transaction) {
803                                 jbd_unlock_bh_state(bh);
804                                  /* needs a brelse */
805                                 journal_remove_journal_head(bh);
806                                 release_buffer_page(bh);
807                         } else
808                                 jbd_unlock_bh_state(bh);
809                 }
810                 cond_resched_lock(&journal->j_list_lock);
811         }
812         spin_unlock(&journal->j_list_lock);
813         /*
814          * This is a bit sleazy.  We borrow j_list_lock to protect
815          * journal->j_committing_transaction in __journal_remove_checkpoint.
816          * Really, __journal_remove_checkpoint should be using j_state_lock but
817          * it's a bit hassle to hold that across __journal_remove_checkpoint
818          */
819         spin_lock(&journal->j_state_lock);
820         spin_lock(&journal->j_list_lock);
821         /*
822          * Now recheck if some buffers did not get attached to the transaction
823          * while the lock was dropped...
824          */
825         if (commit_transaction->t_forget) {
826                 spin_unlock(&journal->j_list_lock);
827                 spin_unlock(&journal->j_state_lock);
828                 goto restart_loop;
829         }
830
831         /* Done with this transaction! */
832
833         jbd_debug(3, "JBD: commit phase 8\n");
834
835         J_ASSERT(commit_transaction->t_state == T_COMMIT);
836
837         commit_transaction->t_state = T_FINISHED;
838         J_ASSERT(commit_transaction == journal->j_committing_transaction);
839         journal->j_commit_sequence = commit_transaction->t_tid;
840         journal->j_committing_transaction = NULL;
841         spin_unlock(&journal->j_state_lock);
842
843         if (commit_transaction->t_checkpoint_list == NULL) {
844                 __journal_drop_transaction(journal, commit_transaction);
845         } else {
846                 if (journal->j_checkpoint_transactions == NULL) {
847                         journal->j_checkpoint_transactions = commit_transaction;
848                         commit_transaction->t_cpnext = commit_transaction;
849                         commit_transaction->t_cpprev = commit_transaction;
850                 } else {
851                         commit_transaction->t_cpnext =
852                                 journal->j_checkpoint_transactions;
853                         commit_transaction->t_cpprev =
854                                 commit_transaction->t_cpnext->t_cpprev;
855                         commit_transaction->t_cpnext->t_cpprev =
856                                 commit_transaction;
857                         commit_transaction->t_cpprev->t_cpnext =
858                                 commit_transaction;
859                 }
860         }
861         spin_unlock(&journal->j_list_lock);
862
863         jbd_debug(1, "JBD: commit %d complete, head %d\n",
864                   journal->j_commit_sequence, journal->j_tail_sequence);
865
866         wake_up(&journal->j_wait_done_commit);
867 }