[PATCH 1/2] ocfs2: add flock lock type
[linux-2.6] / fs / jbd2 / commit.c
1 /*
2  * linux/fs/jbd2/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/jbd2.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 = jbd2_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(JBD2_MAGIC_NUMBER);
123                 tmp->h_blocktype = cpu_to_be32(JBD2_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 & JBD2_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 &= ~JBD2_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         jbd2_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                         __jbd2_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                         __jbd2_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                         __jbd2_journal_unfile_buffer(jh);
258                         jbd_unlock_bh_state(bh);
259                         if (locked)
260                                 unlock_buffer(bh);
261                         jbd2_journal_remove_journal_head(bh);
262                         /* Once for our safety reference, once for
263                          * jbd2_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 static inline void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
278                                    unsigned long long block)
279 {
280         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
281         if (tag_bytes > JBD2_TAG_SIZE32)
282                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
283 }
284
285 /*
286  * jbd2_journal_commit_transaction
287  *
288  * The primary function for committing a transaction to the log.  This
289  * function is called by the journal thread to begin a complete commit.
290  */
291 void jbd2_journal_commit_transaction(journal_t *journal)
292 {
293         transaction_t *commit_transaction;
294         struct journal_head *jh, *new_jh, *descriptor;
295         struct buffer_head **wbuf = journal->j_wbuf;
296         int bufs;
297         int flags;
298         int err;
299         unsigned long long blocknr;
300         char *tagp = NULL;
301         journal_header_t *header;
302         journal_block_tag_t *tag = NULL;
303         int space_left = 0;
304         int first_tag = 0;
305         int tag_flag;
306         int i;
307         int tag_bytes = journal_tag_bytes(journal);
308
309         /*
310          * First job: lock down the current transaction and wait for
311          * all outstanding updates to complete.
312          */
313
314 #ifdef COMMIT_STATS
315         spin_lock(&journal->j_list_lock);
316         summarise_journal_usage(journal);
317         spin_unlock(&journal->j_list_lock);
318 #endif
319
320         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
321         if (journal->j_flags & JBD2_FLUSHED) {
322                 jbd_debug(3, "super block updated\n");
323                 jbd2_journal_update_superblock(journal, 1);
324         } else {
325                 jbd_debug(3, "superblock not updated\n");
326         }
327
328         J_ASSERT(journal->j_running_transaction != NULL);
329         J_ASSERT(journal->j_committing_transaction == NULL);
330
331         commit_transaction = journal->j_running_transaction;
332         J_ASSERT(commit_transaction->t_state == T_RUNNING);
333
334         jbd_debug(1, "JBD: starting commit of transaction %d\n",
335                         commit_transaction->t_tid);
336
337         spin_lock(&journal->j_state_lock);
338         commit_transaction->t_state = T_LOCKED;
339
340         spin_lock(&commit_transaction->t_handle_lock);
341         while (commit_transaction->t_updates) {
342                 DEFINE_WAIT(wait);
343
344                 prepare_to_wait(&journal->j_wait_updates, &wait,
345                                         TASK_UNINTERRUPTIBLE);
346                 if (commit_transaction->t_updates) {
347                         spin_unlock(&commit_transaction->t_handle_lock);
348                         spin_unlock(&journal->j_state_lock);
349                         schedule();
350                         spin_lock(&journal->j_state_lock);
351                         spin_lock(&commit_transaction->t_handle_lock);
352                 }
353                 finish_wait(&journal->j_wait_updates, &wait);
354         }
355         spin_unlock(&commit_transaction->t_handle_lock);
356
357         J_ASSERT (commit_transaction->t_outstanding_credits <=
358                         journal->j_max_transaction_buffers);
359
360         /*
361          * First thing we are allowed to do is to discard any remaining
362          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
363          * that there are no such buffers: if a large filesystem
364          * operation like a truncate needs to split itself over multiple
365          * transactions, then it may try to do a jbd2_journal_restart() while
366          * there are still BJ_Reserved buffers outstanding.  These must
367          * be released cleanly from the current transaction.
368          *
369          * In this case, the filesystem must still reserve write access
370          * again before modifying the buffer in the new transaction, but
371          * we do not require it to remember exactly which old buffers it
372          * has reserved.  This is consistent with the existing behaviour
373          * that multiple jbd2_journal_get_write_access() calls to the same
374          * buffer are perfectly permissable.
375          */
376         while (commit_transaction->t_reserved_list) {
377                 jh = commit_transaction->t_reserved_list;
378                 JBUFFER_TRACE(jh, "reserved, unused: refile");
379                 /*
380                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
381                  * leave undo-committed data.
382                  */
383                 if (jh->b_committed_data) {
384                         struct buffer_head *bh = jh2bh(jh);
385
386                         jbd_lock_bh_state(bh);
387                         jbd2_free(jh->b_committed_data, bh->b_size);
388                         jh->b_committed_data = NULL;
389                         jbd_unlock_bh_state(bh);
390                 }
391                 jbd2_journal_refile_buffer(journal, jh);
392         }
393
394         /*
395          * Now try to drop any written-back buffers from the journal's
396          * checkpoint lists.  We do this *before* commit because it potentially
397          * frees some memory
398          */
399         spin_lock(&journal->j_list_lock);
400         __jbd2_journal_clean_checkpoint_list(journal);
401         spin_unlock(&journal->j_list_lock);
402
403         jbd_debug (3, "JBD: commit phase 1\n");
404
405         /*
406          * Switch to a new revoke table.
407          */
408         jbd2_journal_switch_revoke_table(journal);
409
410         commit_transaction->t_state = T_FLUSH;
411         journal->j_committing_transaction = commit_transaction;
412         journal->j_running_transaction = NULL;
413         commit_transaction->t_log_start = journal->j_head;
414         wake_up(&journal->j_wait_transaction_locked);
415         spin_unlock(&journal->j_state_lock);
416
417         jbd_debug (3, "JBD: commit phase 2\n");
418
419         /*
420          * First, drop modified flag: all accesses to the buffers
421          * will be tracked for a new trasaction only -bzzz
422          */
423         spin_lock(&journal->j_list_lock);
424         if (commit_transaction->t_buffers) {
425                 new_jh = jh = commit_transaction->t_buffers->b_tnext;
426                 do {
427                         J_ASSERT_JH(new_jh, new_jh->b_modified == 1 ||
428                                         new_jh->b_modified == 0);
429                         new_jh->b_modified = 0;
430                         new_jh = new_jh->b_tnext;
431                 } while (new_jh != jh);
432         }
433         spin_unlock(&journal->j_list_lock);
434
435         /*
436          * Now start flushing things to disk, in the order they appear
437          * on the transaction lists.  Data blocks go first.
438          */
439         err = 0;
440         journal_submit_data_buffers(journal, commit_transaction);
441
442         /*
443          * Wait for all previously submitted IO to complete.
444          */
445         spin_lock(&journal->j_list_lock);
446         while (commit_transaction->t_locked_list) {
447                 struct buffer_head *bh;
448
449                 jh = commit_transaction->t_locked_list->b_tprev;
450                 bh = jh2bh(jh);
451                 get_bh(bh);
452                 if (buffer_locked(bh)) {
453                         spin_unlock(&journal->j_list_lock);
454                         wait_on_buffer(bh);
455                         if (unlikely(!buffer_uptodate(bh)))
456                                 err = -EIO;
457                         spin_lock(&journal->j_list_lock);
458                 }
459                 if (!inverted_lock(journal, bh)) {
460                         put_bh(bh);
461                         spin_lock(&journal->j_list_lock);
462                         continue;
463                 }
464                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
465                         __jbd2_journal_unfile_buffer(jh);
466                         jbd_unlock_bh_state(bh);
467                         jbd2_journal_remove_journal_head(bh);
468                         put_bh(bh);
469                 } else {
470                         jbd_unlock_bh_state(bh);
471                 }
472                 put_bh(bh);
473                 cond_resched_lock(&journal->j_list_lock);
474         }
475         spin_unlock(&journal->j_list_lock);
476
477         if (err)
478                 jbd2_journal_abort(journal, err);
479
480         jbd2_journal_write_revoke_records(journal, commit_transaction);
481
482         jbd_debug(3, "JBD: commit phase 2\n");
483
484         /*
485          * If we found any dirty or locked buffers, then we should have
486          * looped back up to the write_out_data label.  If there weren't
487          * any then journal_clean_data_list should have wiped the list
488          * clean by now, so check that it is in fact empty.
489          */
490         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
491
492         jbd_debug (3, "JBD: commit phase 3\n");
493
494         /*
495          * Way to go: we have now written out all of the data for a
496          * transaction!  Now comes the tricky part: we need to write out
497          * metadata.  Loop over the transaction's entire buffer list:
498          */
499         commit_transaction->t_state = T_COMMIT;
500
501         descriptor = NULL;
502         bufs = 0;
503         while (commit_transaction->t_buffers) {
504
505                 /* Find the next buffer to be journaled... */
506
507                 jh = commit_transaction->t_buffers;
508
509                 /* If we're in abort mode, we just un-journal the buffer and
510                    release it for background writing. */
511
512                 if (is_journal_aborted(journal)) {
513                         JBUFFER_TRACE(jh, "journal is aborting: refile");
514                         jbd2_journal_refile_buffer(journal, jh);
515                         /* If that was the last one, we need to clean up
516                          * any descriptor buffers which may have been
517                          * already allocated, even if we are now
518                          * aborting. */
519                         if (!commit_transaction->t_buffers)
520                                 goto start_journal_io;
521                         continue;
522                 }
523
524                 /* Make sure we have a descriptor block in which to
525                    record the metadata buffer. */
526
527                 if (!descriptor) {
528                         struct buffer_head *bh;
529
530                         J_ASSERT (bufs == 0);
531
532                         jbd_debug(4, "JBD: get descriptor\n");
533
534                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
535                         if (!descriptor) {
536                                 jbd2_journal_abort(journal, -EIO);
537                                 continue;
538                         }
539
540                         bh = jh2bh(descriptor);
541                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
542                                 (unsigned long long)bh->b_blocknr, bh->b_data);
543                         header = (journal_header_t *)&bh->b_data[0];
544                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
545                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
546                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
547
548                         tagp = &bh->b_data[sizeof(journal_header_t)];
549                         space_left = bh->b_size - sizeof(journal_header_t);
550                         first_tag = 1;
551                         set_buffer_jwrite(bh);
552                         set_buffer_dirty(bh);
553                         wbuf[bufs++] = bh;
554
555                         /* Record it so that we can wait for IO
556                            completion later */
557                         BUFFER_TRACE(bh, "ph3: file as descriptor");
558                         jbd2_journal_file_buffer(descriptor, commit_transaction,
559                                         BJ_LogCtl);
560                 }
561
562                 /* Where is the buffer to be written? */
563
564                 err = jbd2_journal_next_log_block(journal, &blocknr);
565                 /* If the block mapping failed, just abandon the buffer
566                    and repeat this loop: we'll fall into the
567                    refile-on-abort condition above. */
568                 if (err) {
569                         jbd2_journal_abort(journal, err);
570                         continue;
571                 }
572
573                 /*
574                  * start_this_handle() uses t_outstanding_credits to determine
575                  * the free space in the log, but this counter is changed
576                  * by jbd2_journal_next_log_block() also.
577                  */
578                 commit_transaction->t_outstanding_credits--;
579
580                 /* Bump b_count to prevent truncate from stumbling over
581                    the shadowed buffer!  @@@ This can go if we ever get
582                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
583                 atomic_inc(&jh2bh(jh)->b_count);
584
585                 /* Make a temporary IO buffer with which to write it out
586                    (this will requeue both the metadata buffer and the
587                    temporary IO buffer). new_bh goes on BJ_IO*/
588
589                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
590                 /*
591                  * akpm: jbd2_journal_write_metadata_buffer() sets
592                  * new_bh->b_transaction to commit_transaction.
593                  * We need to clean this up before we release new_bh
594                  * (which is of type BJ_IO)
595                  */
596                 JBUFFER_TRACE(jh, "ph3: write metadata");
597                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
598                                                       jh, &new_jh, blocknr);
599                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
600                 wbuf[bufs++] = jh2bh(new_jh);
601
602                 /* Record the new block's tag in the current descriptor
603                    buffer */
604
605                 tag_flag = 0;
606                 if (flags & 1)
607                         tag_flag |= JBD2_FLAG_ESCAPE;
608                 if (!first_tag)
609                         tag_flag |= JBD2_FLAG_SAME_UUID;
610
611                 tag = (journal_block_tag_t *) tagp;
612                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
613                 tag->t_flags = cpu_to_be32(tag_flag);
614                 tagp += tag_bytes;
615                 space_left -= tag_bytes;
616
617                 if (first_tag) {
618                         memcpy (tagp, journal->j_uuid, 16);
619                         tagp += 16;
620                         space_left -= 16;
621                         first_tag = 0;
622                 }
623
624                 /* If there's no more to do, or if the descriptor is full,
625                    let the IO rip! */
626
627                 if (bufs == journal->j_wbufsize ||
628                     commit_transaction->t_buffers == NULL ||
629                     space_left < tag_bytes + 16) {
630
631                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
632
633                         /* Write an end-of-descriptor marker before
634                            submitting the IOs.  "tag" still points to
635                            the last tag we set up. */
636
637                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
638
639 start_journal_io:
640                         for (i = 0; i < bufs; i++) {
641                                 struct buffer_head *bh = wbuf[i];
642                                 lock_buffer(bh);
643                                 clear_buffer_dirty(bh);
644                                 set_buffer_uptodate(bh);
645                                 bh->b_end_io = journal_end_buffer_io_sync;
646                                 submit_bh(WRITE, bh);
647                         }
648                         cond_resched();
649
650                         /* Force a new descriptor to be generated next
651                            time round the loop. */
652                         descriptor = NULL;
653                         bufs = 0;
654                 }
655         }
656
657         /* Lo and behold: we have just managed to send a transaction to
658            the log.  Before we can commit it, wait for the IO so far to
659            complete.  Control buffers being written are on the
660            transaction's t_log_list queue, and metadata buffers are on
661            the t_iobuf_list queue.
662
663            Wait for the buffers in reverse order.  That way we are
664            less likely to be woken up until all IOs have completed, and
665            so we incur less scheduling load.
666         */
667
668         jbd_debug(3, "JBD: commit phase 4\n");
669
670         /*
671          * akpm: these are BJ_IO, and j_list_lock is not needed.
672          * See __journal_try_to_free_buffer.
673          */
674 wait_for_iobuf:
675         while (commit_transaction->t_iobuf_list != NULL) {
676                 struct buffer_head *bh;
677
678                 jh = commit_transaction->t_iobuf_list->b_tprev;
679                 bh = jh2bh(jh);
680                 if (buffer_locked(bh)) {
681                         wait_on_buffer(bh);
682                         goto wait_for_iobuf;
683                 }
684                 if (cond_resched())
685                         goto wait_for_iobuf;
686
687                 if (unlikely(!buffer_uptodate(bh)))
688                         err = -EIO;
689
690                 clear_buffer_jwrite(bh);
691
692                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
693                 jbd2_journal_unfile_buffer(journal, jh);
694
695                 /*
696                  * ->t_iobuf_list should contain only dummy buffer_heads
697                  * which were created by jbd2_journal_write_metadata_buffer().
698                  */
699                 BUFFER_TRACE(bh, "dumping temporary bh");
700                 jbd2_journal_put_journal_head(jh);
701                 __brelse(bh);
702                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
703                 free_buffer_head(bh);
704
705                 /* We also have to unlock and free the corresponding
706                    shadowed buffer */
707                 jh = commit_transaction->t_shadow_list->b_tprev;
708                 bh = jh2bh(jh);
709                 clear_bit(BH_JWrite, &bh->b_state);
710                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
711
712                 /* The metadata is now released for reuse, but we need
713                    to remember it against this transaction so that when
714                    we finally commit, we can do any checkpointing
715                    required. */
716                 JBUFFER_TRACE(jh, "file as BJ_Forget");
717                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
718                 /* Wake up any transactions which were waiting for this
719                    IO to complete */
720                 wake_up_bit(&bh->b_state, BH_Unshadow);
721                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
722                 __brelse(bh);
723         }
724
725         J_ASSERT (commit_transaction->t_shadow_list == NULL);
726
727         jbd_debug(3, "JBD: commit phase 5\n");
728
729         /* Here we wait for the revoke record and descriptor record buffers */
730  wait_for_ctlbuf:
731         while (commit_transaction->t_log_list != NULL) {
732                 struct buffer_head *bh;
733
734                 jh = commit_transaction->t_log_list->b_tprev;
735                 bh = jh2bh(jh);
736                 if (buffer_locked(bh)) {
737                         wait_on_buffer(bh);
738                         goto wait_for_ctlbuf;
739                 }
740                 if (cond_resched())
741                         goto wait_for_ctlbuf;
742
743                 if (unlikely(!buffer_uptodate(bh)))
744                         err = -EIO;
745
746                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
747                 clear_buffer_jwrite(bh);
748                 jbd2_journal_unfile_buffer(journal, jh);
749                 jbd2_journal_put_journal_head(jh);
750                 __brelse(bh);           /* One for getblk */
751                 /* AKPM: bforget here */
752         }
753
754         jbd_debug(3, "JBD: commit phase 6\n");
755
756         if (journal_write_commit_record(journal, commit_transaction))
757                 err = -EIO;
758
759         if (err)
760                 jbd2_journal_abort(journal, err);
761
762         /* End of a transaction!  Finally, we can do checkpoint
763            processing: any buffers committed as a result of this
764            transaction can be removed from any checkpoint list it was on
765            before. */
766
767         jbd_debug(3, "JBD: commit phase 7\n");
768
769         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
770         J_ASSERT(commit_transaction->t_buffers == NULL);
771         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
772         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
773         J_ASSERT(commit_transaction->t_shadow_list == NULL);
774         J_ASSERT(commit_transaction->t_log_list == NULL);
775
776 restart_loop:
777         /*
778          * As there are other places (journal_unmap_buffer()) adding buffers
779          * to this list we have to be careful and hold the j_list_lock.
780          */
781         spin_lock(&journal->j_list_lock);
782         while (commit_transaction->t_forget) {
783                 transaction_t *cp_transaction;
784                 struct buffer_head *bh;
785
786                 jh = commit_transaction->t_forget;
787                 spin_unlock(&journal->j_list_lock);
788                 bh = jh2bh(jh);
789                 jbd_lock_bh_state(bh);
790                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
791                         jh->b_transaction == journal->j_running_transaction);
792
793                 /*
794                  * If there is undo-protected committed data against
795                  * this buffer, then we can remove it now.  If it is a
796                  * buffer needing such protection, the old frozen_data
797                  * field now points to a committed version of the
798                  * buffer, so rotate that field to the new committed
799                  * data.
800                  *
801                  * Otherwise, we can just throw away the frozen data now.
802                  */
803                 if (jh->b_committed_data) {
804                         jbd2_free(jh->b_committed_data, bh->b_size);
805                         jh->b_committed_data = NULL;
806                         if (jh->b_frozen_data) {
807                                 jh->b_committed_data = jh->b_frozen_data;
808                                 jh->b_frozen_data = NULL;
809                         }
810                 } else if (jh->b_frozen_data) {
811                         jbd2_free(jh->b_frozen_data, bh->b_size);
812                         jh->b_frozen_data = NULL;
813                 }
814
815                 spin_lock(&journal->j_list_lock);
816                 cp_transaction = jh->b_cp_transaction;
817                 if (cp_transaction) {
818                         JBUFFER_TRACE(jh, "remove from old cp transaction");
819                         __jbd2_journal_remove_checkpoint(jh);
820                 }
821
822                 /* Only re-checkpoint the buffer_head if it is marked
823                  * dirty.  If the buffer was added to the BJ_Forget list
824                  * by jbd2_journal_forget, it may no longer be dirty and
825                  * there's no point in keeping a checkpoint record for
826                  * it. */
827
828                 /* A buffer which has been freed while still being
829                  * journaled by a previous transaction may end up still
830                  * being dirty here, but we want to avoid writing back
831                  * that buffer in the future now that the last use has
832                  * been committed.  That's not only a performance gain,
833                  * it also stops aliasing problems if the buffer is left
834                  * behind for writeback and gets reallocated for another
835                  * use in a different page. */
836                 if (buffer_freed(bh)) {
837                         clear_buffer_freed(bh);
838                         clear_buffer_jbddirty(bh);
839                 }
840
841                 if (buffer_jbddirty(bh)) {
842                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
843                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
844                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
845                         __jbd2_journal_refile_buffer(jh);
846                         jbd_unlock_bh_state(bh);
847                 } else {
848                         J_ASSERT_BH(bh, !buffer_dirty(bh));
849                         /* The buffer on BJ_Forget list and not jbddirty means
850                          * it has been freed by this transaction and hence it
851                          * could not have been reallocated until this
852                          * transaction has committed. *BUT* it could be
853                          * reallocated once we have written all the data to
854                          * disk and before we process the buffer on BJ_Forget
855                          * list. */
856                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
857                         __jbd2_journal_refile_buffer(jh);
858                         if (!jh->b_transaction) {
859                                 jbd_unlock_bh_state(bh);
860                                  /* needs a brelse */
861                                 jbd2_journal_remove_journal_head(bh);
862                                 release_buffer_page(bh);
863                         } else
864                                 jbd_unlock_bh_state(bh);
865                 }
866                 cond_resched_lock(&journal->j_list_lock);
867         }
868         spin_unlock(&journal->j_list_lock);
869         /*
870          * This is a bit sleazy.  We borrow j_list_lock to protect
871          * journal->j_committing_transaction in __jbd2_journal_remove_checkpoint.
872          * Really, __jbd2_journal_remove_checkpoint should be using j_state_lock but
873          * it's a bit hassle to hold that across __jbd2_journal_remove_checkpoint
874          */
875         spin_lock(&journal->j_state_lock);
876         spin_lock(&journal->j_list_lock);
877         /*
878          * Now recheck if some buffers did not get attached to the transaction
879          * while the lock was dropped...
880          */
881         if (commit_transaction->t_forget) {
882                 spin_unlock(&journal->j_list_lock);
883                 spin_unlock(&journal->j_state_lock);
884                 goto restart_loop;
885         }
886
887         /* Done with this transaction! */
888
889         jbd_debug(3, "JBD: commit phase 8\n");
890
891         J_ASSERT(commit_transaction->t_state == T_COMMIT);
892
893         commit_transaction->t_state = T_FINISHED;
894         J_ASSERT(commit_transaction == journal->j_committing_transaction);
895         journal->j_commit_sequence = commit_transaction->t_tid;
896         journal->j_committing_transaction = NULL;
897         spin_unlock(&journal->j_state_lock);
898
899         if (commit_transaction->t_checkpoint_list == NULL &&
900             commit_transaction->t_checkpoint_io_list == NULL) {
901                 __jbd2_journal_drop_transaction(journal, commit_transaction);
902         } else {
903                 if (journal->j_checkpoint_transactions == NULL) {
904                         journal->j_checkpoint_transactions = commit_transaction;
905                         commit_transaction->t_cpnext = commit_transaction;
906                         commit_transaction->t_cpprev = commit_transaction;
907                 } else {
908                         commit_transaction->t_cpnext =
909                                 journal->j_checkpoint_transactions;
910                         commit_transaction->t_cpprev =
911                                 commit_transaction->t_cpnext->t_cpprev;
912                         commit_transaction->t_cpnext->t_cpprev =
913                                 commit_transaction;
914                         commit_transaction->t_cpprev->t_cpnext =
915                                 commit_transaction;
916                 }
917         }
918         spin_unlock(&journal->j_list_lock);
919
920         jbd_debug(1, "JBD: commit %d complete, head %d\n",
921                   journal->j_commit_sequence, journal->j_tail_sequence);
922
923         wake_up(&journal->j_wait_done_commit);
924 }