Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/drzeus/mmc
[linux-2.6] / fs / jbd / commit.c
1 /*
2  * linux/fs/jbd/commit.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Journal commit routines for the generic filesystem journaling code;
13  * part of the ext2fs journaling system.
14  */
15
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/pagemap.h>
23
24 /*
25  * Default IO end handler for temporary BJ_IO buffer_heads.
26  */
27 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
28 {
29         BUFFER_TRACE(bh, "");
30         if (uptodate)
31                 set_buffer_uptodate(bh);
32         else
33                 clear_buffer_uptodate(bh);
34         unlock_buffer(bh);
35 }
36
37 /*
38  * When an ext3-ordered file is truncated, it is possible that many pages are
39  * not 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 (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  * 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 (test_set_buffer_locked(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 (TestSetPageLocked(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                 err = 0;
486         }
487
488         journal_write_revoke_records(journal, commit_transaction);
489
490         /*
491          * If we found any dirty or locked buffers, then we should have
492          * looped back up to the write_out_data label.  If there weren't
493          * any then journal_clean_data_list should have wiped the list
494          * clean by now, so check that it is in fact empty.
495          */
496         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
497
498         jbd_debug (3, "JBD: commit phase 3\n");
499
500         /*
501          * Way to go: we have now written out all of the data for a
502          * transaction!  Now comes the tricky part: we need to write out
503          * metadata.  Loop over the transaction's entire buffer list:
504          */
505         spin_lock(&journal->j_state_lock);
506         commit_transaction->t_state = T_COMMIT;
507         spin_unlock(&journal->j_state_lock);
508
509         J_ASSERT(commit_transaction->t_nr_buffers <=
510                  commit_transaction->t_outstanding_credits);
511
512         descriptor = NULL;
513         bufs = 0;
514         while (commit_transaction->t_buffers) {
515
516                 /* Find the next buffer to be journaled... */
517
518                 jh = commit_transaction->t_buffers;
519
520                 /* If we're in abort mode, we just un-journal the buffer and
521                    release it for background writing. */
522
523                 if (is_journal_aborted(journal)) {
524                         JBUFFER_TRACE(jh, "journal is aborting: refile");
525                         journal_refile_buffer(journal, jh);
526                         /* If that was the last one, we need to clean up
527                          * any descriptor buffers which may have been
528                          * already allocated, even if we are now
529                          * aborting. */
530                         if (!commit_transaction->t_buffers)
531                                 goto start_journal_io;
532                         continue;
533                 }
534
535                 /* Make sure we have a descriptor block in which to
536                    record the metadata buffer. */
537
538                 if (!descriptor) {
539                         struct buffer_head *bh;
540
541                         J_ASSERT (bufs == 0);
542
543                         jbd_debug(4, "JBD: get descriptor\n");
544
545                         descriptor = journal_get_descriptor_buffer(journal);
546                         if (!descriptor) {
547                                 journal_abort(journal, -EIO);
548                                 continue;
549                         }
550
551                         bh = jh2bh(descriptor);
552                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
553                                 (unsigned long long)bh->b_blocknr, bh->b_data);
554                         header = (journal_header_t *)&bh->b_data[0];
555                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
556                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
557                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
558
559                         tagp = &bh->b_data[sizeof(journal_header_t)];
560                         space_left = bh->b_size - sizeof(journal_header_t);
561                         first_tag = 1;
562                         set_buffer_jwrite(bh);
563                         set_buffer_dirty(bh);
564                         wbuf[bufs++] = bh;
565
566                         /* Record it so that we can wait for IO
567                            completion later */
568                         BUFFER_TRACE(bh, "ph3: file as descriptor");
569                         journal_file_buffer(descriptor, commit_transaction,
570                                         BJ_LogCtl);
571                 }
572
573                 /* Where is the buffer to be written? */
574
575                 err = journal_next_log_block(journal, &blocknr);
576                 /* If the block mapping failed, just abandon the buffer
577                    and repeat this loop: we'll fall into the
578                    refile-on-abort condition above. */
579                 if (err) {
580                         journal_abort(journal, err);
581                         continue;
582                 }
583
584                 /*
585                  * start_this_handle() uses t_outstanding_credits to determine
586                  * the free space in the log, but this counter is changed
587                  * by journal_next_log_block() also.
588                  */
589                 commit_transaction->t_outstanding_credits--;
590
591                 /* Bump b_count to prevent truncate from stumbling over
592                    the shadowed buffer!  @@@ This can go if we ever get
593                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
594                 atomic_inc(&jh2bh(jh)->b_count);
595
596                 /* Make a temporary IO buffer with which to write it out
597                    (this will requeue both the metadata buffer and the
598                    temporary IO buffer). new_bh goes on BJ_IO*/
599
600                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
601                 /*
602                  * akpm: journal_write_metadata_buffer() sets
603                  * new_bh->b_transaction to commit_transaction.
604                  * We need to clean this up before we release new_bh
605                  * (which is of type BJ_IO)
606                  */
607                 JBUFFER_TRACE(jh, "ph3: write metadata");
608                 flags = journal_write_metadata_buffer(commit_transaction,
609                                                       jh, &new_jh, blocknr);
610                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
611                 wbuf[bufs++] = jh2bh(new_jh);
612
613                 /* Record the new block's tag in the current descriptor
614                    buffer */
615
616                 tag_flag = 0;
617                 if (flags & 1)
618                         tag_flag |= JFS_FLAG_ESCAPE;
619                 if (!first_tag)
620                         tag_flag |= JFS_FLAG_SAME_UUID;
621
622                 tag = (journal_block_tag_t *) tagp;
623                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
624                 tag->t_flags = cpu_to_be32(tag_flag);
625                 tagp += sizeof(journal_block_tag_t);
626                 space_left -= sizeof(journal_block_tag_t);
627
628                 if (first_tag) {
629                         memcpy (tagp, journal->j_uuid, 16);
630                         tagp += 16;
631                         space_left -= 16;
632                         first_tag = 0;
633                 }
634
635                 /* If there's no more to do, or if the descriptor is full,
636                    let the IO rip! */
637
638                 if (bufs == journal->j_wbufsize ||
639                     commit_transaction->t_buffers == NULL ||
640                     space_left < sizeof(journal_block_tag_t) + 16) {
641
642                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
643
644                         /* Write an end-of-descriptor marker before
645                            submitting the IOs.  "tag" still points to
646                            the last tag we set up. */
647
648                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
649
650 start_journal_io:
651                         for (i = 0; i < bufs; i++) {
652                                 struct buffer_head *bh = wbuf[i];
653                                 lock_buffer(bh);
654                                 clear_buffer_dirty(bh);
655                                 set_buffer_uptodate(bh);
656                                 bh->b_end_io = journal_end_buffer_io_sync;
657                                 submit_bh(WRITE, bh);
658                         }
659                         cond_resched();
660
661                         /* Force a new descriptor to be generated next
662                            time round the loop. */
663                         descriptor = NULL;
664                         bufs = 0;
665                 }
666         }
667
668         /* Lo and behold: we have just managed to send a transaction to
669            the log.  Before we can commit it, wait for the IO so far to
670            complete.  Control buffers being written are on the
671            transaction's t_log_list queue, and metadata buffers are on
672            the t_iobuf_list queue.
673
674            Wait for the buffers in reverse order.  That way we are
675            less likely to be woken up until all IOs have completed, and
676            so we incur less scheduling load.
677         */
678
679         jbd_debug(3, "JBD: commit phase 4\n");
680
681         /*
682          * akpm: these are BJ_IO, and j_list_lock is not needed.
683          * See __journal_try_to_free_buffer.
684          */
685 wait_for_iobuf:
686         while (commit_transaction->t_iobuf_list != NULL) {
687                 struct buffer_head *bh;
688
689                 jh = commit_transaction->t_iobuf_list->b_tprev;
690                 bh = jh2bh(jh);
691                 if (buffer_locked(bh)) {
692                         wait_on_buffer(bh);
693                         goto wait_for_iobuf;
694                 }
695                 if (cond_resched())
696                         goto wait_for_iobuf;
697
698                 if (unlikely(!buffer_uptodate(bh)))
699                         err = -EIO;
700
701                 clear_buffer_jwrite(bh);
702
703                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
704                 journal_unfile_buffer(journal, jh);
705
706                 /*
707                  * ->t_iobuf_list should contain only dummy buffer_heads
708                  * which were created by journal_write_metadata_buffer().
709                  */
710                 BUFFER_TRACE(bh, "dumping temporary bh");
711                 journal_put_journal_head(jh);
712                 __brelse(bh);
713                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
714                 free_buffer_head(bh);
715
716                 /* We also have to unlock and free the corresponding
717                    shadowed buffer */
718                 jh = commit_transaction->t_shadow_list->b_tprev;
719                 bh = jh2bh(jh);
720                 clear_bit(BH_JWrite, &bh->b_state);
721                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
722
723                 /* The metadata is now released for reuse, but we need
724                    to remember it against this transaction so that when
725                    we finally commit, we can do any checkpointing
726                    required. */
727                 JBUFFER_TRACE(jh, "file as BJ_Forget");
728                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
729                 /* Wake up any transactions which were waiting for this
730                    IO to complete */
731                 wake_up_bit(&bh->b_state, BH_Unshadow);
732                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
733                 __brelse(bh);
734         }
735
736         J_ASSERT (commit_transaction->t_shadow_list == NULL);
737
738         jbd_debug(3, "JBD: commit phase 5\n");
739
740         /* Here we wait for the revoke record and descriptor record buffers */
741  wait_for_ctlbuf:
742         while (commit_transaction->t_log_list != NULL) {
743                 struct buffer_head *bh;
744
745                 jh = commit_transaction->t_log_list->b_tprev;
746                 bh = jh2bh(jh);
747                 if (buffer_locked(bh)) {
748                         wait_on_buffer(bh);
749                         goto wait_for_ctlbuf;
750                 }
751                 if (cond_resched())
752                         goto wait_for_ctlbuf;
753
754                 if (unlikely(!buffer_uptodate(bh)))
755                         err = -EIO;
756
757                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
758                 clear_buffer_jwrite(bh);
759                 journal_unfile_buffer(journal, jh);
760                 journal_put_journal_head(jh);
761                 __brelse(bh);           /* One for getblk */
762                 /* AKPM: bforget here */
763         }
764
765         jbd_debug(3, "JBD: commit phase 6\n");
766
767         if (journal_write_commit_record(journal, commit_transaction))
768                 err = -EIO;
769
770         if (err)
771                 journal_abort(journal, err);
772
773         /* End of a transaction!  Finally, we can do checkpoint
774            processing: any buffers committed as a result of this
775            transaction can be removed from any checkpoint list it was on
776            before. */
777
778         jbd_debug(3, "JBD: commit phase 7\n");
779
780         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
781         J_ASSERT(commit_transaction->t_buffers == NULL);
782         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
783         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
784         J_ASSERT(commit_transaction->t_shadow_list == NULL);
785         J_ASSERT(commit_transaction->t_log_list == NULL);
786
787 restart_loop:
788         /*
789          * As there are other places (journal_unmap_buffer()) adding buffers
790          * to this list we have to be careful and hold the j_list_lock.
791          */
792         spin_lock(&journal->j_list_lock);
793         while (commit_transaction->t_forget) {
794                 transaction_t *cp_transaction;
795                 struct buffer_head *bh;
796
797                 jh = commit_transaction->t_forget;
798                 spin_unlock(&journal->j_list_lock);
799                 bh = jh2bh(jh);
800                 jbd_lock_bh_state(bh);
801                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
802                         jh->b_transaction == journal->j_running_transaction);
803
804                 /*
805                  * If there is undo-protected committed data against
806                  * this buffer, then we can remove it now.  If it is a
807                  * buffer needing such protection, the old frozen_data
808                  * field now points to a committed version of the
809                  * buffer, so rotate that field to the new committed
810                  * data.
811                  *
812                  * Otherwise, we can just throw away the frozen data now.
813                  */
814                 if (jh->b_committed_data) {
815                         jbd_free(jh->b_committed_data, bh->b_size);
816                         jh->b_committed_data = NULL;
817                         if (jh->b_frozen_data) {
818                                 jh->b_committed_data = jh->b_frozen_data;
819                                 jh->b_frozen_data = NULL;
820                         }
821                 } else if (jh->b_frozen_data) {
822                         jbd_free(jh->b_frozen_data, bh->b_size);
823                         jh->b_frozen_data = NULL;
824                 }
825
826                 spin_lock(&journal->j_list_lock);
827                 cp_transaction = jh->b_cp_transaction;
828                 if (cp_transaction) {
829                         JBUFFER_TRACE(jh, "remove from old cp transaction");
830                         __journal_remove_checkpoint(jh);
831                 }
832
833                 /* Only re-checkpoint the buffer_head if it is marked
834                  * dirty.  If the buffer was added to the BJ_Forget list
835                  * by journal_forget, it may no longer be dirty and
836                  * there's no point in keeping a checkpoint record for
837                  * it. */
838
839                 /* A buffer which has been freed while still being
840                  * journaled by a previous transaction may end up still
841                  * being dirty here, but we want to avoid writing back
842                  * that buffer in the future now that the last use has
843                  * been committed.  That's not only a performance gain,
844                  * it also stops aliasing problems if the buffer is left
845                  * behind for writeback and gets reallocated for another
846                  * use in a different page. */
847                 if (buffer_freed(bh)) {
848                         clear_buffer_freed(bh);
849                         clear_buffer_jbddirty(bh);
850                 }
851
852                 if (buffer_jbddirty(bh)) {
853                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
854                         __journal_insert_checkpoint(jh, commit_transaction);
855                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
856                         __journal_refile_buffer(jh);
857                         jbd_unlock_bh_state(bh);
858                 } else {
859                         J_ASSERT_BH(bh, !buffer_dirty(bh));
860                         /* The buffer on BJ_Forget list and not jbddirty means
861                          * it has been freed by this transaction and hence it
862                          * could not have been reallocated until this
863                          * transaction has committed. *BUT* it could be
864                          * reallocated once we have written all the data to
865                          * disk and before we process the buffer on BJ_Forget
866                          * list. */
867                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
868                         __journal_refile_buffer(jh);
869                         if (!jh->b_transaction) {
870                                 jbd_unlock_bh_state(bh);
871                                  /* needs a brelse */
872                                 journal_remove_journal_head(bh);
873                                 release_buffer_page(bh);
874                         } else
875                                 jbd_unlock_bh_state(bh);
876                 }
877                 cond_resched_lock(&journal->j_list_lock);
878         }
879         spin_unlock(&journal->j_list_lock);
880         /*
881          * This is a bit sleazy.  We use j_list_lock to protect transition
882          * of a transaction into T_FINISHED state and calling
883          * __journal_drop_transaction(). Otherwise we could race with
884          * other checkpointing code processing the transaction...
885          */
886         spin_lock(&journal->j_state_lock);
887         spin_lock(&journal->j_list_lock);
888         /*
889          * Now recheck if some buffers did not get attached to the transaction
890          * while the lock was dropped...
891          */
892         if (commit_transaction->t_forget) {
893                 spin_unlock(&journal->j_list_lock);
894                 spin_unlock(&journal->j_state_lock);
895                 goto restart_loop;
896         }
897
898         /* Done with this transaction! */
899
900         jbd_debug(3, "JBD: commit phase 8\n");
901
902         J_ASSERT(commit_transaction->t_state == T_COMMIT);
903
904         commit_transaction->t_state = T_FINISHED;
905         J_ASSERT(commit_transaction == journal->j_committing_transaction);
906         journal->j_commit_sequence = commit_transaction->t_tid;
907         journal->j_committing_transaction = NULL;
908         spin_unlock(&journal->j_state_lock);
909
910         if (commit_transaction->t_checkpoint_list == NULL &&
911             commit_transaction->t_checkpoint_io_list == NULL) {
912                 __journal_drop_transaction(journal, commit_transaction);
913         } else {
914                 if (journal->j_checkpoint_transactions == NULL) {
915                         journal->j_checkpoint_transactions = commit_transaction;
916                         commit_transaction->t_cpnext = commit_transaction;
917                         commit_transaction->t_cpprev = commit_transaction;
918                 } else {
919                         commit_transaction->t_cpnext =
920                                 journal->j_checkpoint_transactions;
921                         commit_transaction->t_cpprev =
922                                 commit_transaction->t_cpnext->t_cpprev;
923                         commit_transaction->t_cpnext->t_cpprev =
924                                 commit_transaction;
925                         commit_transaction->t_cpprev->t_cpnext =
926                                 commit_transaction;
927                 }
928         }
929         spin_unlock(&journal->j_list_lock);
930
931         jbd_debug(1, "JBD: commit %d complete, head %d\n",
932                   journal->j_commit_sequence, journal->j_tail_sequence);
933
934         wake_up(&journal->j_wait_done_commit);
935 }