jbd: Fix assertion failure in commit code due to lacking transaction credits
[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/mm.h>
21 #include <linux/pagemap.h>
22 #include <linux/bio.h>
23 #include <linux/blkdev.h>
24 #include <trace/events/jbd.h>
25
26 /*
27  * Default IO end handler for temporary BJ_IO buffer_heads.
28  */
29 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
30 {
31         BUFFER_TRACE(bh, "");
32         if (uptodate)
33                 set_buffer_uptodate(bh);
34         else
35                 clear_buffer_uptodate(bh);
36         unlock_buffer(bh);
37 }
38
39 /*
40  * When an ext3-ordered file is truncated, it is possible that many pages are
41  * not successfully freed, because they are attached to a committing transaction.
42  * After the transaction commits, these pages are left on the LRU, with no
43  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
44  * by the VM, but their apparent absence upsets the VM accounting, and it makes
45  * the numbers in /proc/meminfo look odd.
46  *
47  * So here, we have a buffer which has just come off the forget list.  Look to
48  * see if we can strip all buffers from the backing page.
49  *
50  * Called under journal->j_list_lock.  The caller provided us with a ref
51  * against the buffer, and we drop that here.
52  */
53 static void release_buffer_page(struct buffer_head *bh)
54 {
55         struct page *page;
56
57         if (buffer_dirty(bh))
58                 goto nope;
59         if (atomic_read(&bh->b_count) != 1)
60                 goto nope;
61         page = bh->b_page;
62         if (!page)
63                 goto nope;
64         if (page->mapping)
65                 goto nope;
66
67         /* OK, it's a truncated page */
68         if (!trylock_page(page))
69                 goto nope;
70
71         page_cache_get(page);
72         __brelse(bh);
73         try_to_free_buffers(page);
74         unlock_page(page);
75         page_cache_release(page);
76         return;
77
78 nope:
79         __brelse(bh);
80 }
81
82 /*
83  * Decrement reference counter for data buffer. If it has been marked
84  * 'BH_Freed', release it and the page to which it belongs if possible.
85  */
86 static void release_data_buffer(struct buffer_head *bh)
87 {
88         if (buffer_freed(bh)) {
89                 WARN_ON_ONCE(buffer_dirty(bh));
90                 clear_buffer_freed(bh);
91                 clear_buffer_mapped(bh);
92                 clear_buffer_new(bh);
93                 clear_buffer_req(bh);
94                 bh->b_bdev = NULL;
95                 release_buffer_page(bh);
96         } else
97                 put_bh(bh);
98 }
99
100 /*
101  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
102  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
103  * return 0.  j_list_lock is dropped in this case.
104  */
105 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
106 {
107         if (!jbd_trylock_bh_state(bh)) {
108                 spin_unlock(&journal->j_list_lock);
109                 schedule();
110                 return 0;
111         }
112         return 1;
113 }
114
115 /* Done it all: now write the commit record.  We should have
116  * cleaned up our previous buffers by now, so if we are in abort
117  * mode we can now just skip the rest of the journal write
118  * entirely.
119  *
120  * Returns 1 if the journal needs to be aborted or 0 on success
121  */
122 static int journal_write_commit_record(journal_t *journal,
123                                         transaction_t *commit_transaction)
124 {
125         struct journal_head *descriptor;
126         struct buffer_head *bh;
127         journal_header_t *header;
128         int ret;
129
130         if (is_journal_aborted(journal))
131                 return 0;
132
133         descriptor = journal_get_descriptor_buffer(journal);
134         if (!descriptor)
135                 return 1;
136
137         bh = jh2bh(descriptor);
138
139         header = (journal_header_t *)(bh->b_data);
140         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
141         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
142         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
143
144         JBUFFER_TRACE(descriptor, "write commit block");
145         set_buffer_dirty(bh);
146
147         if (journal->j_flags & JFS_BARRIER)
148                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
149         else
150                 ret = sync_dirty_buffer(bh);
151
152         put_bh(bh);             /* One for getblk() */
153         journal_put_journal_head(descriptor);
154
155         return (ret == -EIO);
156 }
157
158 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
159                                    int write_op)
160 {
161         int i;
162
163         for (i = 0; i < bufs; i++) {
164                 wbuf[i]->b_end_io = end_buffer_write_sync;
165                 /* We use-up our safety reference in submit_bh() */
166                 submit_bh(write_op, wbuf[i]);
167         }
168 }
169
170 /*
171  *  Submit all the data buffers to disk
172  */
173 static int journal_submit_data_buffers(journal_t *journal,
174                                        transaction_t *commit_transaction,
175                                        int write_op)
176 {
177         struct journal_head *jh;
178         struct buffer_head *bh;
179         int locked;
180         int bufs = 0;
181         struct buffer_head **wbuf = journal->j_wbuf;
182         int err = 0;
183
184         /*
185          * Whenever we unlock the journal and sleep, things can get added
186          * onto ->t_sync_datalist, so we have to keep looping back to
187          * write_out_data until we *know* that the list is empty.
188          *
189          * Cleanup any flushed data buffers from the data list.  Even in
190          * abort mode, we want to flush this out as soon as possible.
191          */
192 write_out_data:
193         cond_resched();
194         spin_lock(&journal->j_list_lock);
195
196         while (commit_transaction->t_sync_datalist) {
197                 jh = commit_transaction->t_sync_datalist;
198                 bh = jh2bh(jh);
199                 locked = 0;
200
201                 /* Get reference just to make sure buffer does not disappear
202                  * when we are forced to drop various locks */
203                 get_bh(bh);
204                 /* If the buffer is dirty, we need to submit IO and hence
205                  * we need the buffer lock. We try to lock the buffer without
206                  * blocking. If we fail, we need to drop j_list_lock and do
207                  * blocking lock_buffer().
208                  */
209                 if (buffer_dirty(bh)) {
210                         if (!trylock_buffer(bh)) {
211                                 BUFFER_TRACE(bh, "needs blocking lock");
212                                 spin_unlock(&journal->j_list_lock);
213                                 trace_jbd_do_submit_data(journal,
214                                                      commit_transaction);
215                                 /* Write out all data to prevent deadlocks */
216                                 journal_do_submit_data(wbuf, bufs, write_op);
217                                 bufs = 0;
218                                 lock_buffer(bh);
219                                 spin_lock(&journal->j_list_lock);
220                         }
221                         locked = 1;
222                 }
223                 /* We have to get bh_state lock. Again out of order, sigh. */
224                 if (!inverted_lock(journal, bh)) {
225                         jbd_lock_bh_state(bh);
226                         spin_lock(&journal->j_list_lock);
227                 }
228                 /* Someone already cleaned up the buffer? */
229                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
230                         || jh->b_transaction != commit_transaction
231                         || jh->b_jlist != BJ_SyncData) {
232                         jbd_unlock_bh_state(bh);
233                         if (locked)
234                                 unlock_buffer(bh);
235                         BUFFER_TRACE(bh, "already cleaned up");
236                         release_data_buffer(bh);
237                         continue;
238                 }
239                 if (locked && test_clear_buffer_dirty(bh)) {
240                         BUFFER_TRACE(bh, "needs writeout, adding to array");
241                         wbuf[bufs++] = bh;
242                         __journal_file_buffer(jh, commit_transaction,
243                                                 BJ_Locked);
244                         jbd_unlock_bh_state(bh);
245                         if (bufs == journal->j_wbufsize) {
246                                 spin_unlock(&journal->j_list_lock);
247                                 trace_jbd_do_submit_data(journal,
248                                                      commit_transaction);
249                                 journal_do_submit_data(wbuf, bufs, write_op);
250                                 bufs = 0;
251                                 goto write_out_data;
252                         }
253                 } else if (!locked && buffer_locked(bh)) {
254                         __journal_file_buffer(jh, commit_transaction,
255                                                 BJ_Locked);
256                         jbd_unlock_bh_state(bh);
257                         put_bh(bh);
258                 } else {
259                         BUFFER_TRACE(bh, "writeout complete: unfile");
260                         if (unlikely(!buffer_uptodate(bh)))
261                                 err = -EIO;
262                         __journal_unfile_buffer(jh);
263                         jbd_unlock_bh_state(bh);
264                         if (locked)
265                                 unlock_buffer(bh);
266                         release_data_buffer(bh);
267                 }
268
269                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
270                         spin_unlock(&journal->j_list_lock);
271                         goto write_out_data;
272                 }
273         }
274         spin_unlock(&journal->j_list_lock);
275         trace_jbd_do_submit_data(journal, commit_transaction);
276         journal_do_submit_data(wbuf, bufs, write_op);
277
278         return err;
279 }
280
281 /*
282  * journal_commit_transaction
283  *
284  * The primary function for committing a transaction to the log.  This
285  * function is called by the journal thread to begin a complete commit.
286  */
287 void journal_commit_transaction(journal_t *journal)
288 {
289         transaction_t *commit_transaction;
290         struct journal_head *jh, *new_jh, *descriptor;
291         struct buffer_head **wbuf = journal->j_wbuf;
292         int bufs;
293         int flags;
294         int err;
295         unsigned int blocknr;
296         ktime_t start_time;
297         u64 commit_time;
298         char *tagp = NULL;
299         journal_header_t *header;
300         journal_block_tag_t *tag = NULL;
301         int space_left = 0;
302         int first_tag = 0;
303         int tag_flag;
304         int i;
305         struct blk_plug plug;
306         int write_op = WRITE;
307
308         /*
309          * First job: lock down the current transaction and wait for
310          * all outstanding updates to complete.
311          */
312
313         /* Do we need to erase the effects of a prior journal_flush? */
314         if (journal->j_flags & JFS_FLUSHED) {
315                 jbd_debug(3, "super block updated\n");
316                 mutex_lock(&journal->j_checkpoint_mutex);
317                 /*
318                  * We hold j_checkpoint_mutex so tail cannot change under us.
319                  * We don't need any special data guarantees for writing sb
320                  * since journal is empty and it is ok for write to be
321                  * flushed only with transaction commit.
322                  */
323                 journal_update_sb_log_tail(journal, journal->j_tail_sequence,
324                                            journal->j_tail, WRITE_SYNC);
325                 mutex_unlock(&journal->j_checkpoint_mutex);
326         } else {
327                 jbd_debug(3, "superblock not updated\n");
328         }
329
330         J_ASSERT(journal->j_running_transaction != NULL);
331         J_ASSERT(journal->j_committing_transaction == NULL);
332
333         commit_transaction = journal->j_running_transaction;
334         J_ASSERT(commit_transaction->t_state == T_RUNNING);
335
336         trace_jbd_start_commit(journal, commit_transaction);
337         jbd_debug(1, "JBD: starting commit of transaction %d\n",
338                         commit_transaction->t_tid);
339
340         spin_lock(&journal->j_state_lock);
341         commit_transaction->t_state = T_LOCKED;
342
343         trace_jbd_commit_locking(journal, commit_transaction);
344         spin_lock(&commit_transaction->t_handle_lock);
345         while (commit_transaction->t_updates) {
346                 DEFINE_WAIT(wait);
347
348                 prepare_to_wait(&journal->j_wait_updates, &wait,
349                                         TASK_UNINTERRUPTIBLE);
350                 if (commit_transaction->t_updates) {
351                         spin_unlock(&commit_transaction->t_handle_lock);
352                         spin_unlock(&journal->j_state_lock);
353                         schedule();
354                         spin_lock(&journal->j_state_lock);
355                         spin_lock(&commit_transaction->t_handle_lock);
356                 }
357                 finish_wait(&journal->j_wait_updates, &wait);
358         }
359         spin_unlock(&commit_transaction->t_handle_lock);
360
361         J_ASSERT (commit_transaction->t_outstanding_credits <=
362                         journal->j_max_transaction_buffers);
363
364         /*
365          * First thing we are allowed to do is to discard any remaining
366          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
367          * that there are no such buffers: if a large filesystem
368          * operation like a truncate needs to split itself over multiple
369          * transactions, then it may try to do a journal_restart() while
370          * there are still BJ_Reserved buffers outstanding.  These must
371          * be released cleanly from the current transaction.
372          *
373          * In this case, the filesystem must still reserve write access
374          * again before modifying the buffer in the new transaction, but
375          * we do not require it to remember exactly which old buffers it
376          * has reserved.  This is consistent with the existing behaviour
377          * that multiple journal_get_write_access() calls to the same
378          * buffer are perfectly permissible.
379          */
380         while (commit_transaction->t_reserved_list) {
381                 jh = commit_transaction->t_reserved_list;
382                 JBUFFER_TRACE(jh, "reserved, unused: refile");
383                 /*
384                  * A journal_get_undo_access()+journal_release_buffer() may
385                  * leave undo-committed data.
386                  */
387                 if (jh->b_committed_data) {
388                         struct buffer_head *bh = jh2bh(jh);
389
390                         jbd_lock_bh_state(bh);
391                         jbd_free(jh->b_committed_data, bh->b_size);
392                         jh->b_committed_data = NULL;
393                         jbd_unlock_bh_state(bh);
394                 }
395                 journal_refile_buffer(journal, jh);
396         }
397
398         /*
399          * Now try to drop any written-back buffers from the journal's
400          * checkpoint lists.  We do this *before* commit because it potentially
401          * frees some memory
402          */
403         spin_lock(&journal->j_list_lock);
404         __journal_clean_checkpoint_list(journal);
405         spin_unlock(&journal->j_list_lock);
406
407         jbd_debug (3, "JBD: commit phase 1\n");
408
409         /*
410          * Clear revoked flag to reflect there is no revoked buffers
411          * in the next transaction which is going to be started.
412          */
413         journal_clear_buffer_revoked_flags(journal);
414
415         /*
416          * Switch to a new revoke table.
417          */
418         journal_switch_revoke_table(journal);
419
420         trace_jbd_commit_flushing(journal, commit_transaction);
421         commit_transaction->t_state = T_FLUSH;
422         journal->j_committing_transaction = commit_transaction;
423         journal->j_running_transaction = NULL;
424         start_time = ktime_get();
425         commit_transaction->t_log_start = journal->j_head;
426         wake_up(&journal->j_wait_transaction_locked);
427         spin_unlock(&journal->j_state_lock);
428
429         jbd_debug (3, "JBD: commit phase 2\n");
430
431         if (tid_geq(journal->j_commit_waited, commit_transaction->t_tid))
432                 write_op = WRITE_SYNC;
433
434         /*
435          * Now start flushing things to disk, in the order they appear
436          * on the transaction lists.  Data blocks go first.
437          */
438         blk_start_plug(&plug);
439         err = journal_submit_data_buffers(journal, commit_transaction,
440                                           write_op);
441         blk_finish_plug(&plug);
442
443         /*
444          * Wait for all previously submitted IO to complete.
445          */
446         spin_lock(&journal->j_list_lock);
447         while (commit_transaction->t_locked_list) {
448                 struct buffer_head *bh;
449
450                 jh = commit_transaction->t_locked_list->b_tprev;
451                 bh = jh2bh(jh);
452                 get_bh(bh);
453                 if (buffer_locked(bh)) {
454                         spin_unlock(&journal->j_list_lock);
455                         wait_on_buffer(bh);
456                         spin_lock(&journal->j_list_lock);
457                 }
458                 if (unlikely(!buffer_uptodate(bh))) {
459                         if (!trylock_page(bh->b_page)) {
460                                 spin_unlock(&journal->j_list_lock);
461                                 lock_page(bh->b_page);
462                                 spin_lock(&journal->j_list_lock);
463                         }
464                         if (bh->b_page->mapping)
465                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
466
467                         unlock_page(bh->b_page);
468                         SetPageError(bh->b_page);
469                         err = -EIO;
470                 }
471                 if (!inverted_lock(journal, bh)) {
472                         put_bh(bh);
473                         spin_lock(&journal->j_list_lock);
474                         continue;
475                 }
476                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
477                     jh->b_transaction == commit_transaction &&
478                     jh->b_jlist == BJ_Locked)
479                         __journal_unfile_buffer(jh);
480                 jbd_unlock_bh_state(bh);
481                 release_data_buffer(bh);
482                 cond_resched_lock(&journal->j_list_lock);
483         }
484         spin_unlock(&journal->j_list_lock);
485
486         if (err) {
487                 char b[BDEVNAME_SIZE];
488
489                 printk(KERN_WARNING
490                         "JBD: Detected IO errors while flushing file data "
491                         "on %s\n", bdevname(journal->j_fs_dev, b));
492                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
493                         journal_abort(journal, err);
494                 err = 0;
495         }
496
497         blk_start_plug(&plug);
498
499         journal_write_revoke_records(journal, commit_transaction, write_op);
500
501         /*
502          * If we found any dirty or locked buffers, then we should have
503          * looped back up to the write_out_data label.  If there weren't
504          * any then journal_clean_data_list should have wiped the list
505          * clean by now, so check that it is in fact empty.
506          */
507         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
508
509         jbd_debug (3, "JBD: commit phase 3\n");
510
511         /*
512          * Way to go: we have now written out all of the data for a
513          * transaction!  Now comes the tricky part: we need to write out
514          * metadata.  Loop over the transaction's entire buffer list:
515          */
516         spin_lock(&journal->j_state_lock);
517         commit_transaction->t_state = T_COMMIT;
518         spin_unlock(&journal->j_state_lock);
519
520         trace_jbd_commit_logging(journal, commit_transaction);
521         J_ASSERT(commit_transaction->t_nr_buffers <=
522                  commit_transaction->t_outstanding_credits);
523
524         descriptor = NULL;
525         bufs = 0;
526         while (commit_transaction->t_buffers) {
527
528                 /* Find the next buffer to be journaled... */
529
530                 jh = commit_transaction->t_buffers;
531
532                 /* If we're in abort mode, we just un-journal the buffer and
533                    release it. */
534
535                 if (is_journal_aborted(journal)) {
536                         clear_buffer_jbddirty(jh2bh(jh));
537                         JBUFFER_TRACE(jh, "journal is aborting: refile");
538                         journal_refile_buffer(journal, jh);
539                         /* If that was the last one, we need to clean up
540                          * any descriptor buffers which may have been
541                          * already allocated, even if we are now
542                          * aborting. */
543                         if (!commit_transaction->t_buffers)
544                                 goto start_journal_io;
545                         continue;
546                 }
547
548                 /* Make sure we have a descriptor block in which to
549                    record the metadata buffer. */
550
551                 if (!descriptor) {
552                         struct buffer_head *bh;
553
554                         J_ASSERT (bufs == 0);
555
556                         jbd_debug(4, "JBD: get descriptor\n");
557
558                         descriptor = journal_get_descriptor_buffer(journal);
559                         if (!descriptor) {
560                                 journal_abort(journal, -EIO);
561                                 continue;
562                         }
563
564                         bh = jh2bh(descriptor);
565                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
566                                 (unsigned long long)bh->b_blocknr, bh->b_data);
567                         header = (journal_header_t *)&bh->b_data[0];
568                         header->h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
569                         header->h_blocktype = cpu_to_be32(JFS_DESCRIPTOR_BLOCK);
570                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
571
572                         tagp = &bh->b_data[sizeof(journal_header_t)];
573                         space_left = bh->b_size - sizeof(journal_header_t);
574                         first_tag = 1;
575                         set_buffer_jwrite(bh);
576                         set_buffer_dirty(bh);
577                         wbuf[bufs++] = bh;
578
579                         /* Record it so that we can wait for IO
580                            completion later */
581                         BUFFER_TRACE(bh, "ph3: file as descriptor");
582                         journal_file_buffer(descriptor, commit_transaction,
583                                         BJ_LogCtl);
584                 }
585
586                 /* Where is the buffer to be written? */
587
588                 err = journal_next_log_block(journal, &blocknr);
589                 /* If the block mapping failed, just abandon the buffer
590                    and repeat this loop: we'll fall into the
591                    refile-on-abort condition above. */
592                 if (err) {
593                         journal_abort(journal, err);
594                         continue;
595                 }
596
597                 /*
598                  * start_this_handle() uses t_outstanding_credits to determine
599                  * the free space in the log, but this counter is changed
600                  * by journal_next_log_block() also.
601                  */
602                 commit_transaction->t_outstanding_credits--;
603
604                 /* Bump b_count to prevent truncate from stumbling over
605                    the shadowed buffer!  @@@ This can go if we ever get
606                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
607                 get_bh(jh2bh(jh));
608
609                 /* Make a temporary IO buffer with which to write it out
610                    (this will requeue both the metadata buffer and the
611                    temporary IO buffer). new_bh goes on BJ_IO*/
612
613                 set_buffer_jwrite(jh2bh(jh));
614                 /*
615                  * akpm: journal_write_metadata_buffer() sets
616                  * new_bh->b_transaction to commit_transaction.
617                  * We need to clean this up before we release new_bh
618                  * (which is of type BJ_IO)
619                  */
620                 JBUFFER_TRACE(jh, "ph3: write metadata");
621                 flags = journal_write_metadata_buffer(commit_transaction,
622                                                       jh, &new_jh, blocknr);
623                 set_buffer_jwrite(jh2bh(new_jh));
624                 wbuf[bufs++] = jh2bh(new_jh);
625
626                 /* Record the new block's tag in the current descriptor
627                    buffer */
628
629                 tag_flag = 0;
630                 if (flags & 1)
631                         tag_flag |= JFS_FLAG_ESCAPE;
632                 if (!first_tag)
633                         tag_flag |= JFS_FLAG_SAME_UUID;
634
635                 tag = (journal_block_tag_t *) tagp;
636                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
637                 tag->t_flags = cpu_to_be32(tag_flag);
638                 tagp += sizeof(journal_block_tag_t);
639                 space_left -= sizeof(journal_block_tag_t);
640
641                 if (first_tag) {
642                         memcpy (tagp, journal->j_uuid, 16);
643                         tagp += 16;
644                         space_left -= 16;
645                         first_tag = 0;
646                 }
647
648                 /* If there's no more to do, or if the descriptor is full,
649                    let the IO rip! */
650
651                 if (bufs == journal->j_wbufsize ||
652                     commit_transaction->t_buffers == NULL ||
653                     space_left < sizeof(journal_block_tag_t) + 16) {
654
655                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
656
657                         /* Write an end-of-descriptor marker before
658                            submitting the IOs.  "tag" still points to
659                            the last tag we set up. */
660
661                         tag->t_flags |= cpu_to_be32(JFS_FLAG_LAST_TAG);
662
663 start_journal_io:
664                         for (i = 0; i < bufs; i++) {
665                                 struct buffer_head *bh = wbuf[i];
666                                 lock_buffer(bh);
667                                 clear_buffer_dirty(bh);
668                                 set_buffer_uptodate(bh);
669                                 bh->b_end_io = journal_end_buffer_io_sync;
670                                 submit_bh(write_op, bh);
671                         }
672                         cond_resched();
673
674                         /* Force a new descriptor to be generated next
675                            time round the loop. */
676                         descriptor = NULL;
677                         bufs = 0;
678                 }
679         }
680
681         blk_finish_plug(&plug);
682
683         /* Lo and behold: we have just managed to send a transaction to
684            the log.  Before we can commit it, wait for the IO so far to
685            complete.  Control buffers being written are on the
686            transaction's t_log_list queue, and metadata buffers are on
687            the t_iobuf_list queue.
688
689            Wait for the buffers in reverse order.  That way we are
690            less likely to be woken up until all IOs have completed, and
691            so we incur less scheduling load.
692         */
693
694         jbd_debug(3, "JBD: commit phase 4\n");
695
696         /*
697          * akpm: these are BJ_IO, and j_list_lock is not needed.
698          * See __journal_try_to_free_buffer.
699          */
700 wait_for_iobuf:
701         while (commit_transaction->t_iobuf_list != NULL) {
702                 struct buffer_head *bh;
703
704                 jh = commit_transaction->t_iobuf_list->b_tprev;
705                 bh = jh2bh(jh);
706                 if (buffer_locked(bh)) {
707                         wait_on_buffer(bh);
708                         goto wait_for_iobuf;
709                 }
710                 if (cond_resched())
711                         goto wait_for_iobuf;
712
713                 if (unlikely(!buffer_uptodate(bh)))
714                         err = -EIO;
715
716                 clear_buffer_jwrite(bh);
717
718                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
719                 journal_unfile_buffer(journal, jh);
720
721                 /*
722                  * ->t_iobuf_list should contain only dummy buffer_heads
723                  * which were created by journal_write_metadata_buffer().
724                  */
725                 BUFFER_TRACE(bh, "dumping temporary bh");
726                 journal_put_journal_head(jh);
727                 __brelse(bh);
728                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
729                 free_buffer_head(bh);
730
731                 /* We also have to unlock and free the corresponding
732                    shadowed buffer */
733                 jh = commit_transaction->t_shadow_list->b_tprev;
734                 bh = jh2bh(jh);
735                 clear_buffer_jwrite(bh);
736                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
737
738                 /* The metadata is now released for reuse, but we need
739                    to remember it against this transaction so that when
740                    we finally commit, we can do any checkpointing
741                    required. */
742                 JBUFFER_TRACE(jh, "file as BJ_Forget");
743                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
744                 /*
745                  * Wake up any transactions which were waiting for this
746                  * IO to complete. The barrier must be here so that changes
747                  * by journal_file_buffer() take effect before wake_up_bit()
748                  * does the waitqueue check.
749                  */
750                 smp_mb();
751                 wake_up_bit(&bh->b_state, BH_Unshadow);
752                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
753                 __brelse(bh);
754         }
755
756         J_ASSERT (commit_transaction->t_shadow_list == NULL);
757
758         jbd_debug(3, "JBD: commit phase 5\n");
759
760         /* Here we wait for the revoke record and descriptor record buffers */
761  wait_for_ctlbuf:
762         while (commit_transaction->t_log_list != NULL) {
763                 struct buffer_head *bh;
764
765                 jh = commit_transaction->t_log_list->b_tprev;
766                 bh = jh2bh(jh);
767                 if (buffer_locked(bh)) {
768                         wait_on_buffer(bh);
769                         goto wait_for_ctlbuf;
770                 }
771                 if (cond_resched())
772                         goto wait_for_ctlbuf;
773
774                 if (unlikely(!buffer_uptodate(bh)))
775                         err = -EIO;
776
777                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
778                 clear_buffer_jwrite(bh);
779                 journal_unfile_buffer(journal, jh);
780                 journal_put_journal_head(jh);
781                 __brelse(bh);           /* One for getblk */
782                 /* AKPM: bforget here */
783         }
784
785         if (err)
786                 journal_abort(journal, err);
787
788         jbd_debug(3, "JBD: commit phase 6\n");
789
790         /* All metadata is written, now write commit record and do cleanup */
791         spin_lock(&journal->j_state_lock);
792         J_ASSERT(commit_transaction->t_state == T_COMMIT);
793         commit_transaction->t_state = T_COMMIT_RECORD;
794         spin_unlock(&journal->j_state_lock);
795
796         if (journal_write_commit_record(journal, commit_transaction))
797                 err = -EIO;
798
799         if (err)
800                 journal_abort(journal, err);
801
802         /* End of a transaction!  Finally, we can do checkpoint
803            processing: any buffers committed as a result of this
804            transaction can be removed from any checkpoint list it was on
805            before. */
806
807         jbd_debug(3, "JBD: commit phase 7\n");
808
809         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
810         J_ASSERT(commit_transaction->t_buffers == NULL);
811         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
812         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
813         J_ASSERT(commit_transaction->t_shadow_list == NULL);
814         J_ASSERT(commit_transaction->t_log_list == NULL);
815
816 restart_loop:
817         /*
818          * As there are other places (journal_unmap_buffer()) adding buffers
819          * to this list we have to be careful and hold the j_list_lock.
820          */
821         spin_lock(&journal->j_list_lock);
822         while (commit_transaction->t_forget) {
823                 transaction_t *cp_transaction;
824                 struct buffer_head *bh;
825                 int try_to_free = 0;
826
827                 jh = commit_transaction->t_forget;
828                 spin_unlock(&journal->j_list_lock);
829                 bh = jh2bh(jh);
830                 /*
831                  * Get a reference so that bh cannot be freed before we are
832                  * done with it.
833                  */
834                 get_bh(bh);
835                 jbd_lock_bh_state(bh);
836                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
837                         jh->b_transaction == journal->j_running_transaction);
838
839                 /*
840                  * If there is undo-protected committed data against
841                  * this buffer, then we can remove it now.  If it is a
842                  * buffer needing such protection, the old frozen_data
843                  * field now points to a committed version of the
844                  * buffer, so rotate that field to the new committed
845                  * data.
846                  *
847                  * Otherwise, we can just throw away the frozen data now.
848                  */
849                 if (jh->b_committed_data) {
850                         jbd_free(jh->b_committed_data, bh->b_size);
851                         jh->b_committed_data = NULL;
852                         if (jh->b_frozen_data) {
853                                 jh->b_committed_data = jh->b_frozen_data;
854                                 jh->b_frozen_data = NULL;
855                         }
856                 } else if (jh->b_frozen_data) {
857                         jbd_free(jh->b_frozen_data, bh->b_size);
858                         jh->b_frozen_data = NULL;
859                 }
860
861                 spin_lock(&journal->j_list_lock);
862                 cp_transaction = jh->b_cp_transaction;
863                 if (cp_transaction) {
864                         JBUFFER_TRACE(jh, "remove from old cp transaction");
865                         __journal_remove_checkpoint(jh);
866                 }
867
868                 /* Only re-checkpoint the buffer_head if it is marked
869                  * dirty.  If the buffer was added to the BJ_Forget list
870                  * by journal_forget, it may no longer be dirty and
871                  * there's no point in keeping a checkpoint record for
872                  * it. */
873
874                 /*
875                  * A buffer which has been freed while still being journaled by
876                  * a previous transaction.
877                  */
878                 if (buffer_freed(bh)) {
879                         /*
880                          * If the running transaction is the one containing
881                          * "add to orphan" operation (b_next_transaction !=
882                          * NULL), we have to wait for that transaction to
883                          * commit before we can really get rid of the buffer.
884                          * So just clear b_modified to not confuse transaction
885                          * credit accounting and refile the buffer to
886                          * BJ_Forget of the running transaction. If the just
887                          * committed transaction contains "add to orphan"
888                          * operation, we can completely invalidate the buffer
889                          * now. We are rather throughout in that since the
890                          * buffer may be still accessible when blocksize <
891                          * pagesize and it is attached to the last partial
892                          * page.
893                          */
894                         jh->b_modified = 0;
895                         if (!jh->b_next_transaction) {
896                                 clear_buffer_freed(bh);
897                                 clear_buffer_jbddirty(bh);
898                                 clear_buffer_mapped(bh);
899                                 clear_buffer_new(bh);
900                                 clear_buffer_req(bh);
901                                 bh->b_bdev = NULL;
902                         }
903                 }
904
905                 if (buffer_jbddirty(bh)) {
906                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
907                         __journal_insert_checkpoint(jh, commit_transaction);
908                         if (is_journal_aborted(journal))
909                                 clear_buffer_jbddirty(bh);
910                 } else {
911                         J_ASSERT_BH(bh, !buffer_dirty(bh));
912                         /*
913                          * The buffer on BJ_Forget list and not jbddirty means
914                          * it has been freed by this transaction and hence it
915                          * could not have been reallocated until this
916                          * transaction has committed. *BUT* it could be
917                          * reallocated once we have written all the data to
918                          * disk and before we process the buffer on BJ_Forget
919                          * list.
920                          */
921                         if (!jh->b_next_transaction)
922                                 try_to_free = 1;
923                 }
924                 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
925                 __journal_refile_buffer(jh);
926                 jbd_unlock_bh_state(bh);
927                 if (try_to_free)
928                         release_buffer_page(bh);
929                 else
930                         __brelse(bh);
931                 cond_resched_lock(&journal->j_list_lock);
932         }
933         spin_unlock(&journal->j_list_lock);
934         /*
935          * This is a bit sleazy.  We use j_list_lock to protect transition
936          * of a transaction into T_FINISHED state and calling
937          * __journal_drop_transaction(). Otherwise we could race with
938          * other checkpointing code processing the transaction...
939          */
940         spin_lock(&journal->j_state_lock);
941         spin_lock(&journal->j_list_lock);
942         /*
943          * Now recheck if some buffers did not get attached to the transaction
944          * while the lock was dropped...
945          */
946         if (commit_transaction->t_forget) {
947                 spin_unlock(&journal->j_list_lock);
948                 spin_unlock(&journal->j_state_lock);
949                 goto restart_loop;
950         }
951
952         /* Done with this transaction! */
953
954         jbd_debug(3, "JBD: commit phase 8\n");
955
956         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
957
958         commit_transaction->t_state = T_FINISHED;
959         J_ASSERT(commit_transaction == journal->j_committing_transaction);
960         journal->j_commit_sequence = commit_transaction->t_tid;
961         journal->j_committing_transaction = NULL;
962         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
963
964         /*
965          * weight the commit time higher than the average time so we don't
966          * react too strongly to vast changes in commit time
967          */
968         if (likely(journal->j_average_commit_time))
969                 journal->j_average_commit_time = (commit_time*3 +
970                                 journal->j_average_commit_time) / 4;
971         else
972                 journal->j_average_commit_time = commit_time;
973
974         spin_unlock(&journal->j_state_lock);
975
976         if (commit_transaction->t_checkpoint_list == NULL &&
977             commit_transaction->t_checkpoint_io_list == NULL) {
978                 __journal_drop_transaction(journal, commit_transaction);
979         } else {
980                 if (journal->j_checkpoint_transactions == NULL) {
981                         journal->j_checkpoint_transactions = commit_transaction;
982                         commit_transaction->t_cpnext = commit_transaction;
983                         commit_transaction->t_cpprev = commit_transaction;
984                 } else {
985                         commit_transaction->t_cpnext =
986                                 journal->j_checkpoint_transactions;
987                         commit_transaction->t_cpprev =
988                                 commit_transaction->t_cpnext->t_cpprev;
989                         commit_transaction->t_cpnext->t_cpprev =
990                                 commit_transaction;
991                         commit_transaction->t_cpprev->t_cpnext =
992                                 commit_transaction;
993                 }
994         }
995         spin_unlock(&journal->j_list_lock);
996
997         trace_jbd_end_commit(journal, commit_transaction);
998         jbd_debug(1, "JBD: commit %d complete, head %d\n",
999                   journal->j_commit_sequence, journal->j_tail_sequence);
1000
1001         wake_up(&journal->j_wait_done_commit);
1002 }