jbd: Fix oops in journal_remove_journal_head()
[linux-2.6.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                 clear_buffer_freed(bh);
90                 release_buffer_page(bh);
91         } else
92                 put_bh(bh);
93 }
94
95 /*
96  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
97  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
98  * return 0.  j_list_lock is dropped in this case.
99  */
100 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
101 {
102         if (!jbd_trylock_bh_state(bh)) {
103                 spin_unlock(&journal->j_list_lock);
104                 schedule();
105                 return 0;
106         }
107         return 1;
108 }
109
110 /* Done it all: now write the commit record.  We should have
111  * cleaned up our previous buffers by now, so if we are in abort
112  * mode we can now just skip the rest of the journal write
113  * entirely.
114  *
115  * Returns 1 if the journal needs to be aborted or 0 on success
116  */
117 static int journal_write_commit_record(journal_t *journal,
118                                         transaction_t *commit_transaction)
119 {
120         struct journal_head *descriptor;
121         struct buffer_head *bh;
122         journal_header_t *header;
123         int ret;
124
125         if (is_journal_aborted(journal))
126                 return 0;
127
128         descriptor = journal_get_descriptor_buffer(journal);
129         if (!descriptor)
130                 return 1;
131
132         bh = jh2bh(descriptor);
133
134         header = (journal_header_t *)(bh->b_data);
135         header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
136         header->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
137         header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
138
139         JBUFFER_TRACE(descriptor, "write commit block");
140         set_buffer_dirty(bh);
141
142         if (journal->j_flags & JFS_BARRIER)
143                 ret = __sync_dirty_buffer(bh, WRITE_SYNC | WRITE_FLUSH_FUA);
144         else
145                 ret = sync_dirty_buffer(bh);
146
147         put_bh(bh);             /* One for getblk() */
148         journal_put_journal_head(descriptor);
149
150         return (ret == -EIO);
151 }
152
153 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
154                                    int write_op)
155 {
156         int i;
157
158         for (i = 0; i < bufs; i++) {
159                 wbuf[i]->b_end_io = end_buffer_write_sync;
160                 /* We use-up our safety reference in submit_bh() */
161                 submit_bh(write_op, wbuf[i]);
162         }
163 }
164
165 /*
166  *  Submit all the data buffers to disk
167  */
168 static int journal_submit_data_buffers(journal_t *journal,
169                                        transaction_t *commit_transaction,
170                                        int write_op)
171 {
172         struct journal_head *jh;
173         struct buffer_head *bh;
174         int locked;
175         int bufs = 0;
176         struct buffer_head **wbuf = journal->j_wbuf;
177         int err = 0;
178
179         /*
180          * Whenever we unlock the journal and sleep, things can get added
181          * onto ->t_sync_datalist, so we have to keep looping back to
182          * write_out_data until we *know* that the list is empty.
183          *
184          * Cleanup any flushed data buffers from the data list.  Even in
185          * abort mode, we want to flush this out as soon as possible.
186          */
187 write_out_data:
188         cond_resched();
189         spin_lock(&journal->j_list_lock);
190
191         while (commit_transaction->t_sync_datalist) {
192                 jh = commit_transaction->t_sync_datalist;
193                 bh = jh2bh(jh);
194                 locked = 0;
195
196                 /* Get reference just to make sure buffer does not disappear
197                  * when we are forced to drop various locks */
198                 get_bh(bh);
199                 /* If the buffer is dirty, we need to submit IO and hence
200                  * we need the buffer lock. We try to lock the buffer without
201                  * blocking. If we fail, we need to drop j_list_lock and do
202                  * blocking lock_buffer().
203                  */
204                 if (buffer_dirty(bh)) {
205                         if (!trylock_buffer(bh)) {
206                                 BUFFER_TRACE(bh, "needs blocking lock");
207                                 spin_unlock(&journal->j_list_lock);
208                                 trace_jbd_do_submit_data(journal,
209                                                      commit_transaction);
210                                 /* Write out all data to prevent deadlocks */
211                                 journal_do_submit_data(wbuf, bufs, write_op);
212                                 bufs = 0;
213                                 lock_buffer(bh);
214                                 spin_lock(&journal->j_list_lock);
215                         }
216                         locked = 1;
217                 }
218                 /* We have to get bh_state lock. Again out of order, sigh. */
219                 if (!inverted_lock(journal, bh)) {
220                         jbd_lock_bh_state(bh);
221                         spin_lock(&journal->j_list_lock);
222                 }
223                 /* Someone already cleaned up the buffer? */
224                 if (!buffer_jbd(bh) || bh2jh(bh) != jh
225                         || jh->b_transaction != commit_transaction
226                         || jh->b_jlist != BJ_SyncData) {
227                         jbd_unlock_bh_state(bh);
228                         if (locked)
229                                 unlock_buffer(bh);
230                         BUFFER_TRACE(bh, "already cleaned up");
231                         release_data_buffer(bh);
232                         continue;
233                 }
234                 if (locked && test_clear_buffer_dirty(bh)) {
235                         BUFFER_TRACE(bh, "needs writeout, adding to array");
236                         wbuf[bufs++] = bh;
237                         __journal_file_buffer(jh, commit_transaction,
238                                                 BJ_Locked);
239                         jbd_unlock_bh_state(bh);
240                         if (bufs == journal->j_wbufsize) {
241                                 spin_unlock(&journal->j_list_lock);
242                                 trace_jbd_do_submit_data(journal,
243                                                      commit_transaction);
244                                 journal_do_submit_data(wbuf, bufs, write_op);
245                                 bufs = 0;
246                                 goto write_out_data;
247                         }
248                 } else if (!locked && buffer_locked(bh)) {
249                         __journal_file_buffer(jh, commit_transaction,
250                                                 BJ_Locked);
251                         jbd_unlock_bh_state(bh);
252                         put_bh(bh);
253                 } else {
254                         BUFFER_TRACE(bh, "writeout complete: unfile");
255                         if (unlikely(!buffer_uptodate(bh)))
256                                 err = -EIO;
257                         __journal_unfile_buffer(jh);
258                         jbd_unlock_bh_state(bh);
259                         if (locked)
260                                 unlock_buffer(bh);
261                         release_data_buffer(bh);
262                 }
263
264                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
265                         spin_unlock(&journal->j_list_lock);
266                         goto write_out_data;
267                 }
268         }
269         spin_unlock(&journal->j_list_lock);
270         trace_jbd_do_submit_data(journal, commit_transaction);
271         journal_do_submit_data(wbuf, bufs, write_op);
272
273         return err;
274 }
275
276 /*
277  * journal_commit_transaction
278  *
279  * The primary function for committing a transaction to the log.  This
280  * function is called by the journal thread to begin a complete commit.
281  */
282 void journal_commit_transaction(journal_t *journal)
283 {
284         transaction_t *commit_transaction;
285         struct journal_head *jh, *new_jh, *descriptor;
286         struct buffer_head **wbuf = journal->j_wbuf;
287         int bufs;
288         int flags;
289         int err;
290         unsigned int blocknr;
291         ktime_t start_time;
292         u64 commit_time;
293         char *tagp = NULL;
294         journal_header_t *header;
295         journal_block_tag_t *tag = NULL;
296         int space_left = 0;
297         int first_tag = 0;
298         int tag_flag;
299         int i;
300         struct blk_plug plug;
301
302         /*
303          * First job: lock down the current transaction and wait for
304          * all outstanding updates to complete.
305          */
306
307         /* Do we need to erase the effects of a prior journal_flush? */
308         if (journal->j_flags & JFS_FLUSHED) {
309                 jbd_debug(3, "super block updated\n");
310                 journal_update_superblock(journal, 1);
311         } else {
312                 jbd_debug(3, "superblock not updated\n");
313         }
314
315         J_ASSERT(journal->j_running_transaction != NULL);
316         J_ASSERT(journal->j_committing_transaction == NULL);
317
318         commit_transaction = journal->j_running_transaction;
319         J_ASSERT(commit_transaction->t_state == T_RUNNING);
320
321         trace_jbd_start_commit(journal, commit_transaction);
322         jbd_debug(1, "JBD: starting commit of transaction %d\n",
323                         commit_transaction->t_tid);
324
325         spin_lock(&journal->j_state_lock);
326         commit_transaction->t_state = T_LOCKED;
327
328         trace_jbd_commit_locking(journal, commit_transaction);
329         spin_lock(&commit_transaction->t_handle_lock);
330         while (commit_transaction->t_updates) {
331                 DEFINE_WAIT(wait);
332
333                 prepare_to_wait(&journal->j_wait_updates, &wait,
334                                         TASK_UNINTERRUPTIBLE);
335                 if (commit_transaction->t_updates) {
336                         spin_unlock(&commit_transaction->t_handle_lock);
337                         spin_unlock(&journal->j_state_lock);
338                         schedule();
339                         spin_lock(&journal->j_state_lock);
340                         spin_lock(&commit_transaction->t_handle_lock);
341                 }
342                 finish_wait(&journal->j_wait_updates, &wait);
343         }
344         spin_unlock(&commit_transaction->t_handle_lock);
345
346         J_ASSERT (commit_transaction->t_outstanding_credits <=
347                         journal->j_max_transaction_buffers);
348
349         /*
350          * First thing we are allowed to do is to discard any remaining
351          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
352          * that there are no such buffers: if a large filesystem
353          * operation like a truncate needs to split itself over multiple
354          * transactions, then it may try to do a journal_restart() while
355          * there are still BJ_Reserved buffers outstanding.  These must
356          * be released cleanly from the current transaction.
357          *
358          * In this case, the filesystem must still reserve write access
359          * again before modifying the buffer in the new transaction, but
360          * we do not require it to remember exactly which old buffers it
361          * has reserved.  This is consistent with the existing behaviour
362          * that multiple journal_get_write_access() calls to the same
363          * buffer are perfectly permissible.
364          */
365         while (commit_transaction->t_reserved_list) {
366                 jh = commit_transaction->t_reserved_list;
367                 JBUFFER_TRACE(jh, "reserved, unused: refile");
368                 /*
369                  * A journal_get_undo_access()+journal_release_buffer() may
370                  * leave undo-committed data.
371                  */
372                 if (jh->b_committed_data) {
373                         struct buffer_head *bh = jh2bh(jh);
374
375                         jbd_lock_bh_state(bh);
376                         jbd_free(jh->b_committed_data, bh->b_size);
377                         jh->b_committed_data = NULL;
378                         jbd_unlock_bh_state(bh);
379                 }
380                 journal_refile_buffer(journal, jh);
381         }
382
383         /*
384          * Now try to drop any written-back buffers from the journal's
385          * checkpoint lists.  We do this *before* commit because it potentially
386          * frees some memory
387          */
388         spin_lock(&journal->j_list_lock);
389         __journal_clean_checkpoint_list(journal);
390         spin_unlock(&journal->j_list_lock);
391
392         jbd_debug (3, "JBD: commit phase 1\n");
393
394         /*
395          * Switch to a new revoke table.
396          */
397         journal_switch_revoke_table(journal);
398
399         trace_jbd_commit_flushing(journal, commit_transaction);
400         commit_transaction->t_state = T_FLUSH;
401         journal->j_committing_transaction = commit_transaction;
402         journal->j_running_transaction = NULL;
403         start_time = ktime_get();
404         commit_transaction->t_log_start = journal->j_head;
405         wake_up(&journal->j_wait_transaction_locked);
406         spin_unlock(&journal->j_state_lock);
407
408         jbd_debug (3, "JBD: commit phase 2\n");
409
410         /*
411          * Now start flushing things to disk, in the order they appear
412          * on the transaction lists.  Data blocks go first.
413          */
414         blk_start_plug(&plug);
415         err = journal_submit_data_buffers(journal, commit_transaction,
416                                           WRITE_SYNC);
417         blk_finish_plug(&plug);
418
419         /*
420          * Wait for all previously submitted IO to complete.
421          */
422         spin_lock(&journal->j_list_lock);
423         while (commit_transaction->t_locked_list) {
424                 struct buffer_head *bh;
425
426                 jh = commit_transaction->t_locked_list->b_tprev;
427                 bh = jh2bh(jh);
428                 get_bh(bh);
429                 if (buffer_locked(bh)) {
430                         spin_unlock(&journal->j_list_lock);
431                         wait_on_buffer(bh);
432                         spin_lock(&journal->j_list_lock);
433                 }
434                 if (unlikely(!buffer_uptodate(bh))) {
435                         if (!trylock_page(bh->b_page)) {
436                                 spin_unlock(&journal->j_list_lock);
437                                 lock_page(bh->b_page);
438                                 spin_lock(&journal->j_list_lock);
439                         }
440                         if (bh->b_page->mapping)
441                                 set_bit(AS_EIO, &bh->b_page->mapping->flags);
442
443                         unlock_page(bh->b_page);
444                         SetPageError(bh->b_page);
445                         err = -EIO;
446                 }
447                 if (!inverted_lock(journal, bh)) {
448                         put_bh(bh);
449                         spin_lock(&journal->j_list_lock);
450                         continue;
451                 }
452                 if (buffer_jbd(bh) && bh2jh(bh) == jh &&
453                     jh->b_transaction == commit_transaction &&
454                     jh->b_jlist == BJ_Locked)
455                         __journal_unfile_buffer(jh);
456                 jbd_unlock_bh_state(bh);
457                 release_data_buffer(bh);
458                 cond_resched_lock(&journal->j_list_lock);
459         }
460         spin_unlock(&journal->j_list_lock);
461
462         if (err) {
463                 char b[BDEVNAME_SIZE];
464
465                 printk(KERN_WARNING
466                         "JBD: Detected IO errors while flushing file data "
467                         "on %s\n", bdevname(journal->j_fs_dev, b));
468                 if (journal->j_flags & JFS_ABORT_ON_SYNCDATA_ERR)
469                         journal_abort(journal, err);
470                 err = 0;
471         }
472
473         blk_start_plug(&plug);
474
475         journal_write_revoke_records(journal, commit_transaction, WRITE_SYNC);
476
477         /*
478          * If we found any dirty or locked buffers, then we should have
479          * looped back up to the write_out_data label.  If there weren't
480          * any then journal_clean_data_list should have wiped the list
481          * clean by now, so check that it is in fact empty.
482          */
483         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
484
485         jbd_debug (3, "JBD: commit phase 3\n");
486
487         /*
488          * Way to go: we have now written out all of the data for a
489          * transaction!  Now comes the tricky part: we need to write out
490          * metadata.  Loop over the transaction's entire buffer list:
491          */
492         spin_lock(&journal->j_state_lock);
493         commit_transaction->t_state = T_COMMIT;
494         spin_unlock(&journal->j_state_lock);
495
496         trace_jbd_commit_logging(journal, commit_transaction);
497         J_ASSERT(commit_transaction->t_nr_buffers <=
498                  commit_transaction->t_outstanding_credits);
499
500         descriptor = NULL;
501         bufs = 0;
502         while (commit_transaction->t_buffers) {
503
504                 /* Find the next buffer to be journaled... */
505
506                 jh = commit_transaction->t_buffers;
507
508                 /* If we're in abort mode, we just un-journal the buffer and
509                    release it. */
510
511                 if (is_journal_aborted(journal)) {
512                         clear_buffer_jbddirty(jh2bh(jh));
513                         JBUFFER_TRACE(jh, "journal is aborting: refile");
514                         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 = journal_get_descriptor_buffer(journal);
535                         if (!descriptor) {
536                                 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(JFS_MAGIC_NUMBER);
545                         header->h_blocktype = cpu_to_be32(JFS_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                         journal_file_buffer(descriptor, commit_transaction,
559                                         BJ_LogCtl);
560                 }
561
562                 /* Where is the buffer to be written? */
563
564                 err = 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                         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 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                 get_bh(jh2bh(jh));
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_buffer_jwrite(jh2bh(jh));
590                 /*
591                  * akpm: 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 = journal_write_metadata_buffer(commit_transaction,
598                                                       jh, &new_jh, blocknr);
599                 set_buffer_jwrite(jh2bh(new_jh));
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 |= JFS_FLAG_ESCAPE;
608                 if (!first_tag)
609                         tag_flag |= JFS_FLAG_SAME_UUID;
610
611                 tag = (journal_block_tag_t *) tagp;
612                 tag->t_blocknr = cpu_to_be32(jh2bh(jh)->b_blocknr);
613                 tag->t_flags = cpu_to_be32(tag_flag);
614                 tagp += sizeof(journal_block_tag_t);
615                 space_left -= sizeof(journal_block_tag_t);
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 < sizeof(journal_block_tag_t) + 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(JFS_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_SYNC, 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         blk_finish_plug(&plug);
658
659         /* Lo and behold: we have just managed to send a transaction to
660            the log.  Before we can commit it, wait for the IO so far to
661            complete.  Control buffers being written are on the
662            transaction's t_log_list queue, and metadata buffers are on
663            the t_iobuf_list queue.
664
665            Wait for the buffers in reverse order.  That way we are
666            less likely to be woken up until all IOs have completed, and
667            so we incur less scheduling load.
668         */
669
670         jbd_debug(3, "JBD: commit phase 4\n");
671
672         /*
673          * akpm: these are BJ_IO, and j_list_lock is not needed.
674          * See __journal_try_to_free_buffer.
675          */
676 wait_for_iobuf:
677         while (commit_transaction->t_iobuf_list != NULL) {
678                 struct buffer_head *bh;
679
680                 jh = commit_transaction->t_iobuf_list->b_tprev;
681                 bh = jh2bh(jh);
682                 if (buffer_locked(bh)) {
683                         wait_on_buffer(bh);
684                         goto wait_for_iobuf;
685                 }
686                 if (cond_resched())
687                         goto wait_for_iobuf;
688
689                 if (unlikely(!buffer_uptodate(bh)))
690                         err = -EIO;
691
692                 clear_buffer_jwrite(bh);
693
694                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
695                 journal_unfile_buffer(journal, jh);
696
697                 /*
698                  * ->t_iobuf_list should contain only dummy buffer_heads
699                  * which were created by journal_write_metadata_buffer().
700                  */
701                 BUFFER_TRACE(bh, "dumping temporary bh");
702                 journal_put_journal_head(jh);
703                 __brelse(bh);
704                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
705                 free_buffer_head(bh);
706
707                 /* We also have to unlock and free the corresponding
708                    shadowed buffer */
709                 jh = commit_transaction->t_shadow_list->b_tprev;
710                 bh = jh2bh(jh);
711                 clear_buffer_jwrite(bh);
712                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
713
714                 /* The metadata is now released for reuse, but we need
715                    to remember it against this transaction so that when
716                    we finally commit, we can do any checkpointing
717                    required. */
718                 JBUFFER_TRACE(jh, "file as BJ_Forget");
719                 journal_file_buffer(jh, commit_transaction, BJ_Forget);
720                 /*
721                  * Wake up any transactions which were waiting for this
722                  * IO to complete. The barrier must be here so that changes
723                  * by journal_file_buffer() take effect before wake_up_bit()
724                  * does the waitqueue check.
725                  */
726                 smp_mb();
727                 wake_up_bit(&bh->b_state, BH_Unshadow);
728                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
729                 __brelse(bh);
730         }
731
732         J_ASSERT (commit_transaction->t_shadow_list == NULL);
733
734         jbd_debug(3, "JBD: commit phase 5\n");
735
736         /* Here we wait for the revoke record and descriptor record buffers */
737  wait_for_ctlbuf:
738         while (commit_transaction->t_log_list != NULL) {
739                 struct buffer_head *bh;
740
741                 jh = commit_transaction->t_log_list->b_tprev;
742                 bh = jh2bh(jh);
743                 if (buffer_locked(bh)) {
744                         wait_on_buffer(bh);
745                         goto wait_for_ctlbuf;
746                 }
747                 if (cond_resched())
748                         goto wait_for_ctlbuf;
749
750                 if (unlikely(!buffer_uptodate(bh)))
751                         err = -EIO;
752
753                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
754                 clear_buffer_jwrite(bh);
755                 journal_unfile_buffer(journal, jh);
756                 journal_put_journal_head(jh);
757                 __brelse(bh);           /* One for getblk */
758                 /* AKPM: bforget here */
759         }
760
761         if (err)
762                 journal_abort(journal, err);
763
764         jbd_debug(3, "JBD: commit phase 6\n");
765
766         /* All metadata is written, now write commit record and do cleanup */
767         spin_lock(&journal->j_state_lock);
768         J_ASSERT(commit_transaction->t_state == T_COMMIT);
769         commit_transaction->t_state = T_COMMIT_RECORD;
770         spin_unlock(&journal->j_state_lock);
771
772         if (journal_write_commit_record(journal, commit_transaction))
773                 err = -EIO;
774
775         if (err)
776                 journal_abort(journal, err);
777
778         /* End of a transaction!  Finally, we can do checkpoint
779            processing: any buffers committed as a result of this
780            transaction can be removed from any checkpoint list it was on
781            before. */
782
783         jbd_debug(3, "JBD: commit phase 7\n");
784
785         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
786         J_ASSERT(commit_transaction->t_buffers == NULL);
787         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
788         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
789         J_ASSERT(commit_transaction->t_shadow_list == NULL);
790         J_ASSERT(commit_transaction->t_log_list == NULL);
791
792 restart_loop:
793         /*
794          * As there are other places (journal_unmap_buffer()) adding buffers
795          * to this list we have to be careful and hold the j_list_lock.
796          */
797         spin_lock(&journal->j_list_lock);
798         while (commit_transaction->t_forget) {
799                 transaction_t *cp_transaction;
800                 struct buffer_head *bh;
801                 int try_to_free = 0;
802
803                 jh = commit_transaction->t_forget;
804                 spin_unlock(&journal->j_list_lock);
805                 bh = jh2bh(jh);
806                 /*
807                  * Get a reference so that bh cannot be freed before we are
808                  * done with it.
809                  */
810                 get_bh(bh);
811                 jbd_lock_bh_state(bh);
812                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
813                         jh->b_transaction == journal->j_running_transaction);
814
815                 /*
816                  * If there is undo-protected committed data against
817                  * this buffer, then we can remove it now.  If it is a
818                  * buffer needing such protection, the old frozen_data
819                  * field now points to a committed version of the
820                  * buffer, so rotate that field to the new committed
821                  * data.
822                  *
823                  * Otherwise, we can just throw away the frozen data now.
824                  */
825                 if (jh->b_committed_data) {
826                         jbd_free(jh->b_committed_data, bh->b_size);
827                         jh->b_committed_data = NULL;
828                         if (jh->b_frozen_data) {
829                                 jh->b_committed_data = jh->b_frozen_data;
830                                 jh->b_frozen_data = NULL;
831                         }
832                 } else if (jh->b_frozen_data) {
833                         jbd_free(jh->b_frozen_data, bh->b_size);
834                         jh->b_frozen_data = NULL;
835                 }
836
837                 spin_lock(&journal->j_list_lock);
838                 cp_transaction = jh->b_cp_transaction;
839                 if (cp_transaction) {
840                         JBUFFER_TRACE(jh, "remove from old cp transaction");
841                         __journal_remove_checkpoint(jh);
842                 }
843
844                 /* Only re-checkpoint the buffer_head if it is marked
845                  * dirty.  If the buffer was added to the BJ_Forget list
846                  * by journal_forget, it may no longer be dirty and
847                  * there's no point in keeping a checkpoint record for
848                  * it. */
849
850                 /* A buffer which has been freed while still being
851                  * journaled by a previous transaction may end up still
852                  * being dirty here, but we want to avoid writing back
853                  * that buffer in the future after the "add to orphan"
854                  * operation been committed,  That's not only a performance
855                  * gain, it also stops aliasing problems if the buffer is
856                  * left behind for writeback and gets reallocated for another
857                  * use in a different page. */
858                 if (buffer_freed(bh) && !jh->b_next_transaction) {
859                         clear_buffer_freed(bh);
860                         clear_buffer_jbddirty(bh);
861                 }
862
863                 if (buffer_jbddirty(bh)) {
864                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
865                         __journal_insert_checkpoint(jh, commit_transaction);
866                         if (is_journal_aborted(journal))
867                                 clear_buffer_jbddirty(bh);
868                 } else {
869                         J_ASSERT_BH(bh, !buffer_dirty(bh));
870                         /*
871                          * The buffer on BJ_Forget list and not jbddirty means
872                          * it has been freed by this transaction and hence it
873                          * could not have been reallocated until this
874                          * transaction has committed. *BUT* it could be
875                          * reallocated once we have written all the data to
876                          * disk and before we process the buffer on BJ_Forget
877                          * list.
878                          */
879                         if (!jh->b_next_transaction)
880                                 try_to_free = 1;
881                 }
882                 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
883                 __journal_refile_buffer(jh);
884                 jbd_unlock_bh_state(bh);
885                 if (try_to_free)
886                         release_buffer_page(bh);
887                 else
888                         __brelse(bh);
889                 cond_resched_lock(&journal->j_list_lock);
890         }
891         spin_unlock(&journal->j_list_lock);
892         /*
893          * This is a bit sleazy.  We use j_list_lock to protect transition
894          * of a transaction into T_FINISHED state and calling
895          * __journal_drop_transaction(). Otherwise we could race with
896          * other checkpointing code processing the transaction...
897          */
898         spin_lock(&journal->j_state_lock);
899         spin_lock(&journal->j_list_lock);
900         /*
901          * Now recheck if some buffers did not get attached to the transaction
902          * while the lock was dropped...
903          */
904         if (commit_transaction->t_forget) {
905                 spin_unlock(&journal->j_list_lock);
906                 spin_unlock(&journal->j_state_lock);
907                 goto restart_loop;
908         }
909
910         /* Done with this transaction! */
911
912         jbd_debug(3, "JBD: commit phase 8\n");
913
914         J_ASSERT(commit_transaction->t_state == T_COMMIT_RECORD);
915
916         commit_transaction->t_state = T_FINISHED;
917         J_ASSERT(commit_transaction == journal->j_committing_transaction);
918         journal->j_commit_sequence = commit_transaction->t_tid;
919         journal->j_committing_transaction = NULL;
920         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
921
922         /*
923          * weight the commit time higher than the average time so we don't
924          * react too strongly to vast changes in commit time
925          */
926         if (likely(journal->j_average_commit_time))
927                 journal->j_average_commit_time = (commit_time*3 +
928                                 journal->j_average_commit_time) / 4;
929         else
930                 journal->j_average_commit_time = commit_time;
931
932         spin_unlock(&journal->j_state_lock);
933
934         if (commit_transaction->t_checkpoint_list == NULL &&
935             commit_transaction->t_checkpoint_io_list == NULL) {
936                 __journal_drop_transaction(journal, commit_transaction);
937         } else {
938                 if (journal->j_checkpoint_transactions == NULL) {
939                         journal->j_checkpoint_transactions = commit_transaction;
940                         commit_transaction->t_cpnext = commit_transaction;
941                         commit_transaction->t_cpprev = commit_transaction;
942                 } else {
943                         commit_transaction->t_cpnext =
944                                 journal->j_checkpoint_transactions;
945                         commit_transaction->t_cpprev =
946                                 commit_transaction->t_cpnext->t_cpprev;
947                         commit_transaction->t_cpnext->t_cpprev =
948                                 commit_transaction;
949                         commit_transaction->t_cpprev->t_cpnext =
950                                 commit_transaction;
951                 }
952         }
953         spin_unlock(&journal->j_list_lock);
954
955         trace_jbd_end_commit(journal, commit_transaction);
956         jbd_debug(1, "JBD: commit %d complete, head %d\n",
957                   journal->j_commit_sequence, journal->j_tail_sequence);
958
959         wake_up(&journal->j_wait_done_commit);
960 }