ext3: add an option to control error handling on file data
[linux-3.10.git] / 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 successfully 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 journal->j_list_lock.  The caller provided us with a ref
49  * 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 (!trylock_page(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  * Decrement reference counter for data buffer. If it has been marked
82  * 'BH_Freed', release it and the page to which it belongs if possible.
83  */
84 static void release_data_buffer(struct buffer_head *bh)
85 {
86         if (buffer_freed(bh)) {
87                 clear_buffer_freed(bh);
88                 release_buffer_page(bh);
89         } else
90                 put_bh(bh);
91 }
92
93 /*
94  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
95  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
96  * return 0.  j_list_lock is dropped in this case.
97  */
98 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
99 {
100         if (!jbd_trylock_bh_state(bh)) {
101                 spin_unlock(&journal->j_list_lock);
102                 schedule();
103                 return 0;
104         }
105         return 1;
106 }
107
108 /* Done it all: now write the commit record.  We should have
109  * cleaned up our previous buffers by now, so if we are in abort
110  * mode we can now just skip the rest of the journal write
111  * entirely.
112  *
113  * Returns 1 if the journal needs to be aborted or 0 on success
114  */
115 static int journal_write_commit_record(journal_t *journal,
116                                         transaction_t *commit_transaction)
117 {
118         struct journal_head *descriptor;
119         struct buffer_head *bh;
120         journal_header_t *header;
121         int ret;
122         int barrier_done = 0;
123
124         if (is_journal_aborted(journal))
125                 return 0;
126
127         descriptor = journal_get_descriptor_buffer(journal);
128         if (!descriptor)
129                 return 1;
130
131         bh = jh2bh(descriptor);
132
133         header = (journal_header_t *)(bh->b_data);
134         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
135         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
136         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
137
138         JBUFFER_TRACE(descriptor, "write commit block");
139         set_buffer_dirty(bh);
140         if (journal->j_flags & JFS_BARRIER) {
141                 set_buffer_ordered(bh);
142                 barrier_done = 1;
143         }
144         ret = sync_dirty_buffer(bh);
145         if (barrier_done)
146                 clear_buffer_ordered(bh);
147         /* is it possible for another commit to fail at roughly
148          * the same time as this one?  If so, we don't want to
149          * trust the barrier flag in the super, but instead want
150          * to remember if we sent a barrier request
151          */
152         if (ret == -EOPNOTSUPP && barrier_done) {
153                 char b[BDEVNAME_SIZE];
154
155                 printk(KERN_WARNING
156                         "JBD: barrier-based sync failed on %s - "
157                         "disabling barriers\n",
158                         bdevname(journal->j_dev, b));
159                 spin_lock(&journal->j_state_lock);
160                 journal->j_flags &= ~JFS_BARRIER;
161                 spin_unlock(&journal->j_state_lock);
162
163                 /* And try again, without the barrier */
164                 set_buffer_uptodate(bh);
165                 set_buffer_dirty(bh);
166                 ret = sync_dirty_buffer(bh);
167         }
168         put_bh(bh);             /* One for getblk() */
169         journal_put_journal_head(descriptor);
170
171         return (ret == -EIO);
172 }
173
174 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
175 {
176         int i;
177
178         for (i = 0; i < bufs; i++) {
179                 wbuf[i]->b_end_io = end_buffer_write_sync;
180                 /* We use-up our safety reference in submit_bh() */
181                 submit_bh(WRITE, wbuf[i]);
182         }
183 }
184
185 /*
186  *  Submit all the data buffers to disk
187  */
188 static int journal_submit_data_buffers(journal_t *journal,
189                                 transaction_t *commit_transaction)
190 {
191         struct journal_head *jh;
192         struct buffer_head *bh;
193         int locked;
194         int bufs = 0;
195         struct buffer_head **wbuf = journal->j_wbuf;
196         int err = 0;
197
198         /*
199          * Whenever we unlock the journal and sleep, things can get added
200          * onto ->t_sync_datalist, so we have to keep looping back to
201          * write_out_data until we *know* that the list is empty.
202          *
203          * Cleanup any flushed data buffers from the data list.  Even in
204          * abort mode, we want to flush this out as soon as possible.
205          */
206 write_out_data:
207         cond_resched();
208         spin_lock(&journal->j_list_lock);
209
210         while (commit_transaction->t_sync_datalist) {
211                 jh = commit_transaction->t_sync_datalist;
212                 bh = jh2bh(jh);
213                 locked = 0;
214
215                 /* Get reference just to make sure buffer does not disappear
216                  * when we are forced to drop various locks */
217                 get_bh(bh);
218                 /* If the buffer is dirty, we need to submit IO and hence
219                  * we need the buffer lock. We try to lock the buffer without
220                  * blocking. If we fail, we need to drop j_list_lock and do
221                  * blocking lock_buffer().
222                  */
223                 if (buffer_dirty(bh)) {
224                         if (!trylock_buffer(bh)) {
225                                 BUFFER_TRACE(bh, "needs blocking lock");
226                                 spin_unlock(&journal->j_list_lock);
227                                 /* Write out all data to prevent deadlocks */
228                                 journal_do_submit_data(wbuf, bufs);
229                                 bufs = 0;
230                                 lock_buffer(bh);
231                                 spin_lock(&journal->j_list_lock);
232                         }
233                         locked = 1;
234                 }
235                 /* We have to get bh_state lock. Again out of order, sigh. */
236                 if (!inverted_lock(journal, bh)) {
237                         jbd_lock_bh_state(bh);
238                         spin_lock(&journal->j_list_lock);
239                 }
240                 /* Someone already cleaned up the buffer? */
241                 if (!buffer_jbd(bh)
242                         || jh->b_transaction != commit_transaction
243                         || jh->b_jlist != BJ_SyncData) {
244                         jbd_unlock_bh_state(bh);
245                         if (locked)
246                                 unlock_buffer(bh);
247                         BUFFER_TRACE(bh, "already cleaned up");
248                         release_data_buffer(bh);
249                         continue;
250                 }
251                 if (locked && test_clear_buffer_dirty(bh)) {
252                         BUFFER_TRACE(bh, "needs writeout, adding to array");
253                         wbuf[bufs++] = bh;
254                         __journal_file_buffer(jh, commit_transaction,
255                                                 BJ_Locked);
256                         jbd_unlock_bh_state(bh);
257                         if (bufs == journal->j_wbufsize) {
258                                 spin_unlock(&journal->j_list_lock);
259                                 journal_do_submit_data(wbuf, bufs);
260                                 bufs = 0;
261                                 goto write_out_data;
262                         }
263                 } else if (!locked && buffer_locked(bh)) {
264                         __journal_file_buffer(jh, commit_transaction,
265                                                 BJ_Locked);
266                         jbd_unlock_bh_state(bh);
267                         put_bh(bh);
268                 } else {
269                         BUFFER_TRACE(bh, "writeout complete: unfile");
270                         if (unlikely(!buffer_uptodate(bh)))
271                                 err = -EIO;
272                         __journal_unfile_buffer(jh);
273                         jbd_unlock_bh_state(bh);
274                         if (locked)
275                                 unlock_buffer(bh);
276                         journal_remove_journal_head(bh);
277                         /* One for our safety reference, other for
278                          * journal_remove_journal_head() */
279                         put_bh(bh);
280                         release_data_buffer(bh);
281                 }
282
283                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
284                         spin_unlock(&journal->j_list_lock);
285                         goto write_out_data;
286                 }
287         }
288         spin_unlock(&journal->j_list_lock);
289         journal_do_submit_data(wbuf, bufs);
290
291         return err;
292 }
293
294 /*
295  * journal_commit_transaction
296  *
297  * The primary function for committing a transaction to the log.  This
298  * function is called by the journal thread to begin a complete commit.
299  */
300 void journal_commit_transaction(journal_t *journal)
301 {
302         transaction_t *commit_transaction;
303         struct journal_head *jh, *new_jh, *descriptor;
304         struct buffer_head **wbuf = journal->j_wbuf;
305         int bufs;
306         int flags;
307         int err;
308         unsigned long blocknr;
309         char *tagp = NULL;
310         journal_header_t *header;
311         journal_block_tag_t *tag = NULL;
312         int space_left = 0;
313         int first_tag = 0;
314         int tag_flag;
315         int i;
316
317         /*
318          * First job: lock down the current transaction and wait for
319          * all outstanding updates to complete.
320          */
321
322 #ifdef COMMIT_STATS
323         spin_lock(&journal->j_list_lock);
324         summarise_journal_usage(journal);
325         spin_unlock(&journal->j_list_lock);
326 #endif
327
328         /* Do we need to erase the effects of a prior journal_flush? */
329         if (journal->j_flags & JFS_FLUSHED) {
330                 jbd_debug(3, "super block updated\n");
331                 journal_update_superblock(journal, 1);
332         } else {
333                 jbd_debug(3, "superblock not updated\n");
334         }
335
336         J_ASSERT(journal->j_running_transaction != NULL);
337         J_ASSERT(journal->j_committing_transaction == NULL);
338
339         commit_transaction = journal->j_running_transaction;
340         J_ASSERT(commit_transaction->t_state == T_RUNNING);
341
342         jbd_debug(1, "JBD: starting commit of transaction %d\n",
343                         commit_transaction->t_tid);
344
345         spin_lock(&journal->j_state_lock);
346         commit_transaction->t_state = T_LOCKED;
347
348         spin_lock(&commit_transaction->t_handle_lock);
349         while (commit_transaction->t_updates) {
350                 DEFINE_WAIT(wait);
351
352                 prepare_to_wait(&journal->j_wait_updates, &wait,
353                                         TASK_UNINTERRUPTIBLE);
354                 if (commit_transaction->t_updates) {
355                         spin_unlock(&commit_transaction->t_handle_lock);
356                         spin_unlock(&journal->j_state_lock);
357                         schedule();
358                         spin_lock(&journal->j_state_lock);
359                         spin_lock(&commit_transaction->t_handle_lock);
360                 }
361                 finish_wait(&journal->j_wait_updates, &wait);
362         }
363         spin_unlock(&commit_transaction->t_handle_lock);
364
365         J_ASSERT (commit_transaction->t_outstanding_credits <=
366                         journal->j_max_transaction_buffers);
367
368         /*
369          * First thing we are allowed to do is to discard any remaining
370          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
371          * that there are no such buffers: if a large filesystem
372          * operation like a truncate needs to split itself over multiple
373          * transactions, then it may try to do a journal_restart() while
374          * there are still BJ_Reserved buffers outstanding.  These must
375          * be released cleanly from the current transaction.
376          *
377          * In this case, the filesystem must still reserve write access
378          * again before modifying the buffer in the new transaction, but
379          * we do not require it to remember exactly which old buffers it
380          * has reserved.  This is consistent with the existing behaviour
381          * that multiple journal_get_write_access() calls to the same
382          * buffer are perfectly permissable.
383          */
384         while (commit_transaction->t_reserved_list) {
385                 jh = commit_transaction->t_reserved_list;
386                 JBUFFER_TRACE(jh, "reserved, unused: refile");
387                 /*
388                  * A journal_get_undo_access()+journal_release_buffer() may
389                  * leave undo-committed data.
390                  */
391                 if (jh->b_committed_data) {
392                         struct buffer_head *bh = jh2bh(jh);
393
394                         jbd_lock_bh_state(bh);
395                         jbd_free(jh->b_committed_data, bh->b_size);
396                         jh->b_committed_data = NULL;
397                         jbd_unlock_bh_state(bh);
398                 }
399                 journal_refile_buffer(journal, jh);
400         }
401
402         /*
403          * Now try to drop any written-back buffers from the journal's
404          * checkpoint lists.  We do this *before* commit because it potentially
405          * frees some memory
406          */
407         spin_lock(&journal->j_list_lock);
408         __journal_clean_checkpoint_list(journal);
409         spin_unlock(&journal->j_list_lock);
410
411         jbd_debug (3, "JBD: commit phase 1\n");
412
413         /*
414          * Switch to a new revoke table.
415          */
416         journal_switch_revoke_table(journal);
417
418         commit_transaction->t_state = T_FLUSH;
419         journal->j_committing_transaction = commit_transaction;
420         journal->j_running_transaction = NULL;
421         commit_transaction->t_log_start = journal->j_head;
422         wake_up(&journal->j_wait_transaction_locked);
423         spin_unlock(&journal->j_state_lock);
424
425         jbd_debug (3, "JBD: commit phase 2\n");
426
427         /*
428          * Now start flushing things to disk, in the order they appear
429          * on the transaction lists.  Data blocks go first.
430          */
431         err = 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                         spin_lock(&journal->j_list_lock);
447                 }
448                 if (unlikely(!buffer_uptodate(bh))) {
449                         if (!trylock_page(bh->b_page)) {
450                                 spin_unlock(&journal->j_list_lock);
451                                 lock_page(bh->b_page);
452                                 spin_lock(&journal->j_list_lock);
453                         }
454                         if (bh->b_page->mapping)
455                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
456
457                         unlock_page(bh->b_page);
458                         SetPageError(bh->b_page);
459                         err = -EIO;
460                 }
461                 if (!inverted_lock(journal, bh)) {
462                         put_bh(bh);
463                         spin_lock(&journal->j_list_lock);
464                         continue;
465                 }
466                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
467                         __journal_unfile_buffer(jh);
468                         jbd_unlock_bh_state(bh);
469                         journal_remove_journal_head(bh);
470                         put_bh(bh);
471                 } else {
472                         jbd_unlock_bh_state(bh);
473                 }
474                 release_data_buffer(bh);
475                 cond_resched_lock(&journal->j_list_lock);
476         }
477         spin_unlock(&journal->j_list_lock);
478
479         if (err) {
480                 char b[BDEVNAME_SIZE];
481
482                 printk(KERN_WARNING
483                         "JBD: Detected IO errors while flushing file data "
484                         "on %s\n", bdevname(journal->j_fs_dev, b));
485                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
486                         journal_abort(journal, err);
487                 err = 0;
488         }
489
490         journal_write_revoke_records(journal, commit_transaction);
491
492         /*
493          * If we found any dirty or locked buffers, then we should have
494          * looped back up to the write_out_data label.  If there weren't
495          * any then journal_clean_data_list should have wiped the list
496          * clean by now, so check that it is in fact empty.
497          */
498         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
499
500         jbd_debug (3, "JBD: commit phase 3\n");
501
502         /*
503          * Way to go: we have now written out all of the data for a
504          * transaction!  Now comes the tricky part: we need to write out
505          * metadata.  Loop over the transaction's entire buffer list:
506          */
507         spin_lock(&journal->j_state_lock);
508         commit_transaction->t_state = T_COMMIT;
509         spin_unlock(&journal->j_state_lock);
510
511         J_ASSERT(commit_transaction->t_nr_buffers <=
512                  commit_transaction->t_outstanding_credits);
513
514         descriptor = NULL;
515         bufs = 0;
516         while (commit_transaction->t_buffers) {
517
518                 /* Find the next buffer to be journaled... */
519
520                 jh = commit_transaction->t_buffers;
521
522                 /* If we're in abort mode, we just un-journal the buffer and
523                    release it. */
524
525                 if (is_journal_aborted(journal)) {
526                         clear_buffer_jbddirty(jh2bh(jh));
527                         JBUFFER_TRACE(jh, "journal is aborting: refile");
528                         journal_refile_buffer(journal, jh);
529                         /* If that was the last one, we need to clean up
530                          * any descriptor buffers which may have been
531                          * already allocated, even if we are now
532                          * aborting. */
533                         if (!commit_transaction->t_buffers)
534                                 goto start_journal_io;
535                         continue;
536                 }
537
538                 /* Make sure we have a descriptor block in which to
539                    record the metadata buffer. */
540
541                 if (!descriptor) {
542                         struct buffer_head *bh;
543
544                         J_ASSERT (bufs == 0);
545
546                         jbd_debug(4, "JBD: get descriptor\n");
547
548                         descriptor = journal_get_descriptor_buffer(journal);
549                         if (!descriptor) {
550                                 journal_abort(journal, -EIO);
551                                 continue;
552                         }
553
554                         bh = jh2bh(descriptor);
555                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
556                                 (unsigned long long)bh->b_blocknr, bh->b_data);
557                         header = (journal_header_t *)&bh->b_data[0];
558                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
559                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
560                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
561
562                         tagp = &bh->b_data[sizeof(journal_header_t)];
563                         space_left = bh->b_size - sizeof(journal_header_t);
564                         first_tag = 1;
565                         set_buffer_jwrite(bh);
566                         set_buffer_dirty(bh);
567                         wbuf[bufs++] = bh;
568
569                         /* Record it so that we can wait for IO
570                            completion later */
571                         BUFFER_TRACE(bh, "ph3: file as descriptor");
572                         journal_file_buffer(descriptor, commit_transaction,
573                                         BJ_LogCtl);
574                 }
575
576                 /* Where is the buffer to be written? */
577
578                 err = journal_next_log_block(journal, &blocknr);
579                 /* If the block mapping failed, just abandon the buffer
580                    and repeat this loop: we'll fall into the
581                    refile-on-abort condition above. */
582                 if (err) {
583                         journal_abort(journal, err);
584                         continue;
585                 }
586
587                 /*
588                  * start_this_handle() uses t_outstanding_credits to determine
589                  * the free space in the log, but this counter is changed
590                  * by journal_next_log_block() also.
591                  */
592                 commit_transaction->t_outstanding_credits--;
593
594                 /* Bump b_count to prevent truncate from stumbling over
595                    the shadowed buffer!  @@@ This can go if we ever get
596                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
597                 atomic_inc(&jh2bh(jh)->b_count);
598
599                 /* Make a temporary IO buffer with which to write it out
600                    (this will requeue both the metadata buffer and the
601                    temporary IO buffer). new_bh goes on BJ_IO*/
602
603                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
604                 /*
605                  * akpm: journal_write_metadata_buffer() sets
606                  * new_bh->b_transaction to commit_transaction.
607                  * We need to clean this up before we release new_bh
608                  * (which is of type BJ_IO)
609                  */
610                 JBUFFER_TRACE(jh, "ph3: write metadata");
611                 flags = journal_write_metadata_buffer(commit_transaction,
612                                                       jh, &new_jh, blocknr);
613                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
614                 wbuf[bufs++] = jh2bh(new_jh);
615
616                 /* Record the new block's tag in the current descriptor
617                    buffer */
618
619                 tag_flag = 0;
620                 if (flags & 1)
621                         tag_flag |= JFS_FLAG_ESCAPE;
622                 if (!first_tag)
623                         tag_flag |= JFS_FLAG_SAME_UUID;
624
625                 tag = (journal_block_tag_t *) tagp;
626                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
627                 tag->t_flags = cpu_to_be32(tag_flag);
628                 tagp += sizeof(journal_block_tag_t);
629                 space_left -= sizeof(journal_block_tag_t);
630
631                 if (first_tag) {
632                         memcpy (tagp, journal->j_uuid, 16);
633                         tagp += 16;
634                         space_left -= 16;
635                         first_tag = 0;
636                 }
637
638                 /* If there's no more to do, or if the descriptor is full,
639                    let the IO rip! */
640
641                 if (bufs == journal->j_wbufsize ||
642                     commit_transaction->t_buffers == NULL ||
643                     space_left < sizeof(journal_block_tag_t) + 16) {
644
645                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
646
647                         /* Write an end-of-descriptor marker before
648                            submitting the IOs.  "tag" still points to
649                            the last tag we set up. */
650
651                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
652
653 start_journal_io:
654                         for (i = 0; i < bufs; i++) {
655                                 struct buffer_head *bh = wbuf[i];
656                                 lock_buffer(bh);
657                                 clear_buffer_dirty(bh);
658                                 set_buffer_uptodate(bh);
659                                 bh->b_end_io = journal_end_buffer_io_sync;
660                                 submit_bh(WRITE, bh);
661                         }
662                         cond_resched();
663
664                         /* Force a new descriptor to be generated next
665                            time round the loop. */
666                         descriptor = NULL;
667                         bufs = 0;
668                 }
669         }
670
671         /* Lo and behold: we have just managed to send a transaction to
672            the log.  Before we can commit it, wait for the IO so far to
673            complete.  Control buffers being written are on the
674            transaction's t_log_list queue, and metadata buffers are on
675            the t_iobuf_list queue.
676
677            Wait for the buffers in reverse order.  That way we are
678            less likely to be woken up until all IOs have completed, and
679            so we incur less scheduling load.
680         */
681
682         jbd_debug(3, "JBD: commit phase 4\n");
683
684         /*
685          * akpm: these are BJ_IO, and j_list_lock is not needed.
686          * See __journal_try_to_free_buffer.
687          */
688 wait_for_iobuf:
689         while (commit_transaction->t_iobuf_list != NULL) {
690                 struct buffer_head *bh;
691
692                 jh = commit_transaction->t_iobuf_list->b_tprev;
693                 bh = jh2bh(jh);
694                 if (buffer_locked(bh)) {
695                         wait_on_buffer(bh);
696                         goto wait_for_iobuf;
697                 }
698                 if (cond_resched())
699                         goto wait_for_iobuf;
700
701                 if (unlikely(!buffer_uptodate(bh)))
702                         err = -EIO;
703
704                 clear_buffer_jwrite(bh);
705
706                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
707                 journal_unfile_buffer(journal, jh);
708
709                 /*
710                  * ->t_iobuf_list should contain only dummy buffer_heads
711                  * which were created by journal_write_metadata_buffer().
712                  */
713                 BUFFER_TRACE(bh, "dumping temporary bh");
714                 journal_put_journal_head(jh);
715                 __brelse(bh);
716                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
717                 free_buffer_head(bh);
718
719                 /* We also have to unlock and free the corresponding
720                    shadowed buffer */
721                 jh = commit_transaction->t_shadow_list->b_tprev;
722                 bh = jh2bh(jh);
723                 clear_bit(BH_JWrite, &bh->b_state);
724                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
725
726                 /* The metadata is now released for reuse, but we need
727                    to remember it against this transaction so that when
728                    we finally commit, we can do any checkpointing
729                    required. */
730                 JBUFFER_TRACE(jh, "file as BJ_Forget");
731                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
732                 /* Wake up any transactions which were waiting for this
733                    IO to complete */
734                 wake_up_bit(&bh->b_state, BH_Unshadow);
735                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
736                 __brelse(bh);
737         }
738
739         J_ASSERT (commit_transaction->t_shadow_list == NULL);
740
741         jbd_debug(3, "JBD: commit phase 5\n");
742
743         /* Here we wait for the revoke record and descriptor record buffers */
744  wait_for_ctlbuf:
745         while (commit_transaction->t_log_list != NULL) {
746                 struct buffer_head *bh;
747
748                 jh = commit_transaction->t_log_list->b_tprev;
749                 bh = jh2bh(jh);
750                 if (buffer_locked(bh)) {
751                         wait_on_buffer(bh);
752                         goto wait_for_ctlbuf;
753                 }
754                 if (cond_resched())
755                         goto wait_for_ctlbuf;
756
757                 if (unlikely(!buffer_uptodate(bh)))
758                         err = -EIO;
759
760                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
761                 clear_buffer_jwrite(bh);
762                 journal_unfile_buffer(journal, jh);
763                 journal_put_journal_head(jh);
764                 __brelse(bh);           /* One for getblk */
765                 /* AKPM: bforget here */
766         }
767
768         if (err)
769                 journal_abort(journal, err);
770
771         jbd_debug(3, "JBD: commit phase 6\n");
772
773         if (journal_write_commit_record(journal, commit_transaction))
774                 err = -EIO;
775
776         if (err)
777                 journal_abort(journal, err);
778
779         /* End of a transaction!  Finally, we can do checkpoint
780            processing: any buffers committed as a result of this
781            transaction can be removed from any checkpoint list it was on
782            before. */
783
784         jbd_debug(3, "JBD: commit phase 7\n");
785
786         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
787         J_ASSERT(commit_transaction->t_buffers == NULL);
788         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
789         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
790         J_ASSERT(commit_transaction->t_shadow_list == NULL);
791         J_ASSERT(commit_transaction->t_log_list == NULL);
792
793 restart_loop:
794         /*
795          * As there are other places (journal_unmap_buffer()) adding buffers
796          * to this list we have to be careful and hold the j_list_lock.
797          */
798         spin_lock(&journal->j_list_lock);
799         while (commit_transaction->t_forget) {
800                 transaction_t *cp_transaction;
801                 struct buffer_head *bh;
802
803                 jh = commit_transaction->t_forget;
804                 spin_unlock(&journal->j_list_lock);
805                 bh = jh2bh(jh);
806                 jbd_lock_bh_state(bh);
807                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
808                         jh->b_transaction == journal->j_running_transaction);
809
810                 /*
811                  * If there is undo-protected committed data against
812                  * this buffer, then we can remove it now.  If it is a
813                  * buffer needing such protection, the old frozen_data
814                  * field now points to a committed version of the
815                  * buffer, so rotate that field to the new committed
816                  * data.
817                  *
818                  * Otherwise, we can just throw away the frozen data now.
819                  */
820                 if (jh->b_committed_data) {
821                         jbd_free(jh->b_committed_data, bh->b_size);
822                         jh->b_committed_data = NULL;
823                         if (jh->b_frozen_data) {
824                                 jh->b_committed_data = jh->b_frozen_data;
825                                 jh->b_frozen_data = NULL;
826                         }
827                 } else if (jh->b_frozen_data) {
828                         jbd_free(jh->b_frozen_data, bh->b_size);
829                         jh->b_frozen_data = NULL;
830                 }
831
832                 spin_lock(&journal->j_list_lock);
833                 cp_transaction = jh->b_cp_transaction;
834                 if (cp_transaction) {
835                         JBUFFER_TRACE(jh, "remove from old cp transaction");
836                         __journal_remove_checkpoint(jh);
837                 }
838
839                 /* Only re-checkpoint the buffer_head if it is marked
840                  * dirty.  If the buffer was added to the BJ_Forget list
841                  * by journal_forget, it may no longer be dirty and
842                  * there's no point in keeping a checkpoint record for
843                  * it. */
844
845                 /* A buffer which has been freed while still being
846                  * journaled by a previous transaction may end up still
847                  * being dirty here, but we want to avoid writing back
848                  * that buffer in the future now that the last use has
849                  * been committed.  That's not only a performance gain,
850                  * it also stops aliasing problems if the buffer is left
851                  * behind for writeback and gets reallocated for another
852                  * use in a different page. */
853                 if (buffer_freed(bh)) {
854                         clear_buffer_freed(bh);
855                         clear_buffer_jbddirty(bh);
856                 }
857
858                 if (buffer_jbddirty(bh)) {
859                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
860                         __journal_insert_checkpoint(jh, commit_transaction);
861                         if (is_journal_aborted(journal))
862                                 clear_buffer_jbddirty(bh);
863                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
864                         __journal_refile_buffer(jh);
865                         jbd_unlock_bh_state(bh);
866                 } else {
867                         J_ASSERT_BH(bh, !buffer_dirty(bh));
868                         /* The buffer on BJ_Forget list and not jbddirty means
869                          * it has been freed by this transaction and hence it
870                          * could not have been reallocated until this
871                          * transaction has committed. *BUT* it could be
872                          * reallocated once we have written all the data to
873                          * disk and before we process the buffer on BJ_Forget
874                          * list. */
875                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
876                         __journal_refile_buffer(jh);
877                         if (!jh->b_transaction) {
878                                 jbd_unlock_bh_state(bh);
879                                  /* needs a brelse */
880                                 journal_remove_journal_head(bh);
881                                 release_buffer_page(bh);
882                         } else
883                                 jbd_unlock_bh_state(bh);
884                 }
885                 cond_resched_lock(&journal->j_list_lock);
886         }
887         spin_unlock(&journal->j_list_lock);
888         /*
889          * This is a bit sleazy.  We use j_list_lock to protect transition
890          * of a transaction into T_FINISHED state and calling
891          * __journal_drop_transaction(). Otherwise we could race with
892          * other checkpointing code processing the transaction...
893          */
894         spin_lock(&journal->j_state_lock);
895         spin_lock(&journal->j_list_lock);
896         /*
897          * Now recheck if some buffers did not get attached to the transaction
898          * while the lock was dropped...
899          */
900         if (commit_transaction->t_forget) {
901                 spin_unlock(&journal->j_list_lock);
902                 spin_unlock(&journal->j_state_lock);
903                 goto restart_loop;
904         }
905
906         /* Done with this transaction! */
907
908         jbd_debug(3, "JBD: commit phase 8\n");
909
910         J_ASSERT(commit_transaction->t_state == T_COMMIT);
911
912         commit_transaction->t_state = T_FINISHED;
913         J_ASSERT(commit_transaction == journal->j_committing_transaction);
914         journal->j_commit_sequence = commit_transaction->t_tid;
915         journal->j_committing_transaction = NULL;
916         spin_unlock(&journal->j_state_lock);
917
918         if (commit_transaction->t_checkpoint_list == NULL &&
919             commit_transaction->t_checkpoint_io_list == NULL) {
920                 __journal_drop_transaction(journal, commit_transaction);
921         } else {
922                 if (journal->j_checkpoint_transactions == NULL) {
923                         journal->j_checkpoint_transactions = commit_transaction;
924                         commit_transaction->t_cpnext = commit_transaction;
925                         commit_transaction->t_cpprev = commit_transaction;
926                 } else {
927                         commit_transaction->t_cpnext =
928                                 journal->j_checkpoint_transactions;
929                         commit_transaction->t_cpprev =
930                                 commit_transaction->t_cpnext->t_cpprev;
931                         commit_transaction->t_cpnext->t_cpprev =
932                                 commit_transaction;
933                         commit_transaction->t_cpprev->t_cpnext =
934                                 commit_transaction;
935                 }
936         }
937         spin_unlock(&journal->j_list_lock);
938
939         jbd_debug(1, "JBD: commit %d complete, head %d\n",
940                   journal->j_commit_sequence, journal->j_tail_sequence);
941
942         wake_up(&journal->j_wait_done_commit);
943 }