jbd2: fix the way the b_modified flag is cleared
[linux-2.6.git] / fs / jbd2 / commit.c
1 /*
2  * linux/fs/jbd2/commit.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Journal commit routines for the generic filesystem journaling code;
13  * part of the ext2fs journaling system.
14  */
15
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/crc32.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 sucessfully 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 lock_journal(), and possibly under journal_datalist_lock.  The
51  * caller provided us with a ref 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 (TestSetPageLocked(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  * Try to acquire jbd_lock_bh_state() against the buffer, when j_list_lock is
84  * held.  For ranking reasons we must trylock.  If we lose, schedule away and
85  * return 0.  j_list_lock is dropped in this case.
86  */
87 static int inverted_lock(journal_t *journal, struct buffer_head *bh)
88 {
89         if (!jbd_trylock_bh_state(bh)) {
90                 spin_unlock(&journal->j_list_lock);
91                 schedule();
92                 return 0;
93         }
94         return 1;
95 }
96
97 /*
98  * Done it all: now submit the commit record.  We should have
99  * cleaned up our previous buffers by now, so if we are in abort
100  * mode we can now just skip the rest of the journal write
101  * entirely.
102  *
103  * Returns 1 if the journal needs to be aborted or 0 on success
104  */
105 static int journal_submit_commit_record(journal_t *journal,
106                                         transaction_t *commit_transaction,
107                                         struct buffer_head **cbh,
108                                         __u32 crc32_sum)
109 {
110         struct journal_head *descriptor;
111         struct commit_header *tmp;
112         struct buffer_head *bh;
113         int ret;
114         int barrier_done = 0;
115
116         if (is_journal_aborted(journal))
117                 return 0;
118
119         descriptor = jbd2_journal_get_descriptor_buffer(journal);
120         if (!descriptor)
121                 return 1;
122
123         bh = jh2bh(descriptor);
124
125         tmp = (struct commit_header *)bh->b_data;
126         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
127         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
128         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
129
130         if (JBD2_HAS_COMPAT_FEATURE(journal,
131                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
132                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
133                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
134                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
135         }
136
137         JBUFFER_TRACE(descriptor, "submit commit block");
138         lock_buffer(bh);
139         get_bh(bh);
140         set_buffer_dirty(bh);
141         set_buffer_uptodate(bh);
142         bh->b_end_io = journal_end_buffer_io_sync;
143
144         if (journal->j_flags & JBD2_BARRIER &&
145                 !JBD2_HAS_INCOMPAT_FEATURE(journal,
146                                          JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
147                 set_buffer_ordered(bh);
148                 barrier_done = 1;
149         }
150         ret = submit_bh(WRITE, bh);
151         if (barrier_done)
152                 clear_buffer_ordered(bh);
153
154         /* is it possible for another commit to fail at roughly
155          * the same time as this one?  If so, we don't want to
156          * trust the barrier flag in the super, but instead want
157          * to remember if we sent a barrier request
158          */
159         if (ret == -EOPNOTSUPP && barrier_done) {
160                 char b[BDEVNAME_SIZE];
161
162                 printk(KERN_WARNING
163                         "JBD: barrier-based sync failed on %s - "
164                         "disabling barriers\n",
165                         bdevname(journal->j_dev, b));
166                 spin_lock(&journal->j_state_lock);
167                 journal->j_flags &= ~JBD2_BARRIER;
168                 spin_unlock(&journal->j_state_lock);
169
170                 /* And try again, without the barrier */
171                 set_buffer_uptodate(bh);
172                 set_buffer_dirty(bh);
173                 ret = submit_bh(WRITE, bh);
174         }
175         *cbh = bh;
176         return ret;
177 }
178
179 /*
180  * This function along with journal_submit_commit_record
181  * allows to write the commit record asynchronously.
182  */
183 static int journal_wait_on_commit_record(struct buffer_head *bh)
184 {
185         int ret = 0;
186
187         clear_buffer_dirty(bh);
188         wait_on_buffer(bh);
189
190         if (unlikely(!buffer_uptodate(bh)))
191                 ret = -EIO;
192         put_bh(bh);            /* One for getblk() */
193         jbd2_journal_put_journal_head(bh2jh(bh));
194
195         return ret;
196 }
197
198 /*
199  * Wait for all submitted IO to complete.
200  */
201 static int journal_wait_on_locked_list(journal_t *journal,
202                                        transaction_t *commit_transaction)
203 {
204         int ret = 0;
205         struct journal_head *jh;
206
207         while (commit_transaction->t_locked_list) {
208                 struct buffer_head *bh;
209
210                 jh = commit_transaction->t_locked_list->b_tprev;
211                 bh = jh2bh(jh);
212                 get_bh(bh);
213                 if (buffer_locked(bh)) {
214                         spin_unlock(&journal->j_list_lock);
215                         wait_on_buffer(bh);
216                         if (unlikely(!buffer_uptodate(bh)))
217                                 ret = -EIO;
218                         spin_lock(&journal->j_list_lock);
219                 }
220                 if (!inverted_lock(journal, bh)) {
221                         put_bh(bh);
222                         spin_lock(&journal->j_list_lock);
223                         continue;
224                 }
225                 if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
226                         __jbd2_journal_unfile_buffer(jh);
227                         jbd_unlock_bh_state(bh);
228                         jbd2_journal_remove_journal_head(bh);
229                         put_bh(bh);
230                 } else {
231                         jbd_unlock_bh_state(bh);
232                 }
233                 put_bh(bh);
234                 cond_resched_lock(&journal->j_list_lock);
235         }
236         return ret;
237   }
238
239 static void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
240 {
241         int i;
242
243         for (i = 0; i < bufs; i++) {
244                 wbuf[i]->b_end_io = end_buffer_write_sync;
245                 /* We use-up our safety reference in submit_bh() */
246                 submit_bh(WRITE, wbuf[i]);
247         }
248 }
249
250 /*
251  *  Submit all the data buffers to disk
252  */
253 static void journal_submit_data_buffers(journal_t *journal,
254                                 transaction_t *commit_transaction)
255 {
256         struct journal_head *jh;
257         struct buffer_head *bh;
258         int locked;
259         int bufs = 0;
260         struct buffer_head **wbuf = journal->j_wbuf;
261
262         /*
263          * Whenever we unlock the journal and sleep, things can get added
264          * onto ->t_sync_datalist, so we have to keep looping back to
265          * write_out_data until we *know* that the list is empty.
266          *
267          * Cleanup any flushed data buffers from the data list.  Even in
268          * abort mode, we want to flush this out as soon as possible.
269          */
270 write_out_data:
271         cond_resched();
272         spin_lock(&journal->j_list_lock);
273
274         while (commit_transaction->t_sync_datalist) {
275                 jh = commit_transaction->t_sync_datalist;
276                 bh = jh2bh(jh);
277                 locked = 0;
278
279                 /* Get reference just to make sure buffer does not disappear
280                  * when we are forced to drop various locks */
281                 get_bh(bh);
282                 /* If the buffer is dirty, we need to submit IO and hence
283                  * we need the buffer lock. We try to lock the buffer without
284                  * blocking. If we fail, we need to drop j_list_lock and do
285                  * blocking lock_buffer().
286                  */
287                 if (buffer_dirty(bh)) {
288                         if (test_set_buffer_locked(bh)) {
289                                 BUFFER_TRACE(bh, "needs blocking lock");
290                                 spin_unlock(&journal->j_list_lock);
291                                 /* Write out all data to prevent deadlocks */
292                                 journal_do_submit_data(wbuf, bufs);
293                                 bufs = 0;
294                                 lock_buffer(bh);
295                                 spin_lock(&journal->j_list_lock);
296                         }
297                         locked = 1;
298                 }
299                 /* We have to get bh_state lock. Again out of order, sigh. */
300                 if (!inverted_lock(journal, bh)) {
301                         jbd_lock_bh_state(bh);
302                         spin_lock(&journal->j_list_lock);
303                 }
304                 /* Someone already cleaned up the buffer? */
305                 if (!buffer_jbd(bh)
306                         || jh->b_transaction != commit_transaction
307                         || jh->b_jlist != BJ_SyncData) {
308                         jbd_unlock_bh_state(bh);
309                         if (locked)
310                                 unlock_buffer(bh);
311                         BUFFER_TRACE(bh, "already cleaned up");
312                         put_bh(bh);
313                         continue;
314                 }
315                 if (locked && test_clear_buffer_dirty(bh)) {
316                         BUFFER_TRACE(bh, "needs writeout, adding to array");
317                         wbuf[bufs++] = bh;
318                         __jbd2_journal_file_buffer(jh, commit_transaction,
319                                                 BJ_Locked);
320                         jbd_unlock_bh_state(bh);
321                         if (bufs == journal->j_wbufsize) {
322                                 spin_unlock(&journal->j_list_lock);
323                                 journal_do_submit_data(wbuf, bufs);
324                                 bufs = 0;
325                                 goto write_out_data;
326                         }
327                 } else if (!locked && buffer_locked(bh)) {
328                         __jbd2_journal_file_buffer(jh, commit_transaction,
329                                                 BJ_Locked);
330                         jbd_unlock_bh_state(bh);
331                         put_bh(bh);
332                 } else {
333                         BUFFER_TRACE(bh, "writeout complete: unfile");
334                         __jbd2_journal_unfile_buffer(jh);
335                         jbd_unlock_bh_state(bh);
336                         if (locked)
337                                 unlock_buffer(bh);
338                         jbd2_journal_remove_journal_head(bh);
339                         /* Once for our safety reference, once for
340                          * jbd2_journal_remove_journal_head() */
341                         put_bh(bh);
342                         put_bh(bh);
343                 }
344
345                 if (need_resched() || spin_needbreak(&journal->j_list_lock)) {
346                         spin_unlock(&journal->j_list_lock);
347                         goto write_out_data;
348                 }
349         }
350         spin_unlock(&journal->j_list_lock);
351         journal_do_submit_data(wbuf, bufs);
352 }
353
354 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
355 {
356         struct page *page = bh->b_page;
357         char *addr;
358         __u32 checksum;
359
360         addr = kmap_atomic(page, KM_USER0);
361         checksum = crc32_be(crc32_sum,
362                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
363         kunmap_atomic(addr, KM_USER0);
364
365         return checksum;
366 }
367
368 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
369                                    unsigned long long block)
370 {
371         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
372         if (tag_bytes > JBD2_TAG_SIZE32)
373                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
374 }
375
376 /*
377  * jbd2_journal_commit_transaction
378  *
379  * The primary function for committing a transaction to the log.  This
380  * function is called by the journal thread to begin a complete commit.
381  */
382 void jbd2_journal_commit_transaction(journal_t *journal)
383 {
384         struct transaction_stats_s stats;
385         transaction_t *commit_transaction;
386         struct journal_head *jh, *new_jh, *descriptor;
387         struct buffer_head **wbuf = journal->j_wbuf;
388         int bufs;
389         int flags;
390         int err;
391         unsigned long long blocknr;
392         char *tagp = NULL;
393         journal_header_t *header;
394         journal_block_tag_t *tag = NULL;
395         int space_left = 0;
396         int first_tag = 0;
397         int tag_flag;
398         int i;
399         int tag_bytes = journal_tag_bytes(journal);
400         struct buffer_head *cbh = NULL; /* For transactional checksums */
401         __u32 crc32_sum = ~0;
402
403         /*
404          * First job: lock down the current transaction and wait for
405          * all outstanding updates to complete.
406          */
407
408 #ifdef COMMIT_STATS
409         spin_lock(&journal->j_list_lock);
410         summarise_journal_usage(journal);
411         spin_unlock(&journal->j_list_lock);
412 #endif
413
414         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
415         if (journal->j_flags & JBD2_FLUSHED) {
416                 jbd_debug(3, "super block updated\n");
417                 jbd2_journal_update_superblock(journal, 1);
418         } else {
419                 jbd_debug(3, "superblock not updated\n");
420         }
421
422         J_ASSERT(journal->j_running_transaction != NULL);
423         J_ASSERT(journal->j_committing_transaction == NULL);
424
425         commit_transaction = journal->j_running_transaction;
426         J_ASSERT(commit_transaction->t_state == T_RUNNING);
427
428         jbd_debug(1, "JBD: starting commit of transaction %d\n",
429                         commit_transaction->t_tid);
430
431         spin_lock(&journal->j_state_lock);
432         commit_transaction->t_state = T_LOCKED;
433
434         stats.u.run.rs_wait = commit_transaction->t_max_wait;
435         stats.u.run.rs_locked = jiffies;
436         stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
437                                                 stats.u.run.rs_locked);
438
439         spin_lock(&commit_transaction->t_handle_lock);
440         while (commit_transaction->t_updates) {
441                 DEFINE_WAIT(wait);
442
443                 prepare_to_wait(&journal->j_wait_updates, &wait,
444                                         TASK_UNINTERRUPTIBLE);
445                 if (commit_transaction->t_updates) {
446                         spin_unlock(&commit_transaction->t_handle_lock);
447                         spin_unlock(&journal->j_state_lock);
448                         schedule();
449                         spin_lock(&journal->j_state_lock);
450                         spin_lock(&commit_transaction->t_handle_lock);
451                 }
452                 finish_wait(&journal->j_wait_updates, &wait);
453         }
454         spin_unlock(&commit_transaction->t_handle_lock);
455
456         J_ASSERT (commit_transaction->t_outstanding_credits <=
457                         journal->j_max_transaction_buffers);
458
459         /*
460          * First thing we are allowed to do is to discard any remaining
461          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
462          * that there are no such buffers: if a large filesystem
463          * operation like a truncate needs to split itself over multiple
464          * transactions, then it may try to do a jbd2_journal_restart() while
465          * there are still BJ_Reserved buffers outstanding.  These must
466          * be released cleanly from the current transaction.
467          *
468          * In this case, the filesystem must still reserve write access
469          * again before modifying the buffer in the new transaction, but
470          * we do not require it to remember exactly which old buffers it
471          * has reserved.  This is consistent with the existing behaviour
472          * that multiple jbd2_journal_get_write_access() calls to the same
473          * buffer are perfectly permissable.
474          */
475         while (commit_transaction->t_reserved_list) {
476                 jh = commit_transaction->t_reserved_list;
477                 JBUFFER_TRACE(jh, "reserved, unused: refile");
478                 /*
479                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
480                  * leave undo-committed data.
481                  */
482                 if (jh->b_committed_data) {
483                         struct buffer_head *bh = jh2bh(jh);
484
485                         jbd_lock_bh_state(bh);
486                         jbd2_free(jh->b_committed_data, bh->b_size);
487                         jh->b_committed_data = NULL;
488                         jbd_unlock_bh_state(bh);
489                 }
490                 jbd2_journal_refile_buffer(journal, jh);
491         }
492
493         /*
494          * Now try to drop any written-back buffers from the journal's
495          * checkpoint lists.  We do this *before* commit because it potentially
496          * frees some memory
497          */
498         spin_lock(&journal->j_list_lock);
499         __jbd2_journal_clean_checkpoint_list(journal);
500         spin_unlock(&journal->j_list_lock);
501
502         jbd_debug (3, "JBD: commit phase 1\n");
503
504         /*
505          * Switch to a new revoke table.
506          */
507         jbd2_journal_switch_revoke_table(journal);
508
509         stats.u.run.rs_flushing = jiffies;
510         stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
511                                                stats.u.run.rs_flushing);
512
513         commit_transaction->t_state = T_FLUSH;
514         journal->j_committing_transaction = commit_transaction;
515         journal->j_running_transaction = NULL;
516         commit_transaction->t_log_start = journal->j_head;
517         wake_up(&journal->j_wait_transaction_locked);
518         spin_unlock(&journal->j_state_lock);
519
520         jbd_debug (3, "JBD: commit phase 2\n");
521
522         /*
523          * Now start flushing things to disk, in the order they appear
524          * on the transaction lists.  Data blocks go first.
525          */
526         err = 0;
527         journal_submit_data_buffers(journal, commit_transaction);
528
529         /*
530          * Wait for all previously submitted IO to complete if commit
531          * record is to be written synchronously.
532          */
533         spin_lock(&journal->j_list_lock);
534         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
535                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
536                 err = journal_wait_on_locked_list(journal,
537                                                 commit_transaction);
538
539         spin_unlock(&journal->j_list_lock);
540
541         if (err)
542                 jbd2_journal_abort(journal, err);
543
544         jbd2_journal_write_revoke_records(journal, commit_transaction);
545
546         jbd_debug(3, "JBD: commit phase 2\n");
547
548         /*
549          * If we found any dirty or locked buffers, then we should have
550          * looped back up to the write_out_data label.  If there weren't
551          * any then journal_clean_data_list should have wiped the list
552          * clean by now, so check that it is in fact empty.
553          */
554         J_ASSERT (commit_transaction->t_sync_datalist == NULL);
555
556         jbd_debug (3, "JBD: commit phase 3\n");
557
558         /*
559          * Way to go: we have now written out all of the data for a
560          * transaction!  Now comes the tricky part: we need to write out
561          * metadata.  Loop over the transaction's entire buffer list:
562          */
563         commit_transaction->t_state = T_COMMIT;
564
565         stats.u.run.rs_logging = jiffies;
566         stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
567                                                  stats.u.run.rs_logging);
568         stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
569         stats.u.run.rs_blocks_logged = 0;
570
571         descriptor = NULL;
572         bufs = 0;
573         while (commit_transaction->t_buffers) {
574
575                 /* Find the next buffer to be journaled... */
576
577                 jh = commit_transaction->t_buffers;
578
579                 /* If we're in abort mode, we just un-journal the buffer and
580                    release it for background writing. */
581
582                 if (is_journal_aborted(journal)) {
583                         JBUFFER_TRACE(jh, "journal is aborting: refile");
584                         jbd2_journal_refile_buffer(journal, jh);
585                         /* If that was the last one, we need to clean up
586                          * any descriptor buffers which may have been
587                          * already allocated, even if we are now
588                          * aborting. */
589                         if (!commit_transaction->t_buffers)
590                                 goto start_journal_io;
591                         continue;
592                 }
593
594                 /* Make sure we have a descriptor block in which to
595                    record the metadata buffer. */
596
597                 if (!descriptor) {
598                         struct buffer_head *bh;
599
600                         J_ASSERT (bufs == 0);
601
602                         jbd_debug(4, "JBD: get descriptor\n");
603
604                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
605                         if (!descriptor) {
606                                 jbd2_journal_abort(journal, -EIO);
607                                 continue;
608                         }
609
610                         bh = jh2bh(descriptor);
611                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
612                                 (unsigned long long)bh->b_blocknr, bh->b_data);
613                         header = (journal_header_t *)&bh->b_data[0];
614                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
615                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
616                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
617
618                         tagp = &bh->b_data[sizeof(journal_header_t)];
619                         space_left = bh->b_size - sizeof(journal_header_t);
620                         first_tag = 1;
621                         set_buffer_jwrite(bh);
622                         set_buffer_dirty(bh);
623                         wbuf[bufs++] = bh;
624
625                         /* Record it so that we can wait for IO
626                            completion later */
627                         BUFFER_TRACE(bh, "ph3: file as descriptor");
628                         jbd2_journal_file_buffer(descriptor, commit_transaction,
629                                         BJ_LogCtl);
630                 }
631
632                 /* Where is the buffer to be written? */
633
634                 err = jbd2_journal_next_log_block(journal, &blocknr);
635                 /* If the block mapping failed, just abandon the buffer
636                    and repeat this loop: we'll fall into the
637                    refile-on-abort condition above. */
638                 if (err) {
639                         jbd2_journal_abort(journal, err);
640                         continue;
641                 }
642
643                 /*
644                  * start_this_handle() uses t_outstanding_credits to determine
645                  * the free space in the log, but this counter is changed
646                  * by jbd2_journal_next_log_block() also.
647                  */
648                 commit_transaction->t_outstanding_credits--;
649
650                 /* Bump b_count to prevent truncate from stumbling over
651                    the shadowed buffer!  @@@ This can go if we ever get
652                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
653                 atomic_inc(&jh2bh(jh)->b_count);
654
655                 /* Make a temporary IO buffer with which to write it out
656                    (this will requeue both the metadata buffer and the
657                    temporary IO buffer). new_bh goes on BJ_IO*/
658
659                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
660                 /*
661                  * akpm: jbd2_journal_write_metadata_buffer() sets
662                  * new_bh->b_transaction to commit_transaction.
663                  * We need to clean this up before we release new_bh
664                  * (which is of type BJ_IO)
665                  */
666                 JBUFFER_TRACE(jh, "ph3: write metadata");
667                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
668                                                       jh, &new_jh, blocknr);
669                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
670                 wbuf[bufs++] = jh2bh(new_jh);
671
672                 /* Record the new block's tag in the current descriptor
673                    buffer */
674
675                 tag_flag = 0;
676                 if (flags & 1)
677                         tag_flag |= JBD2_FLAG_ESCAPE;
678                 if (!first_tag)
679                         tag_flag |= JBD2_FLAG_SAME_UUID;
680
681                 tag = (journal_block_tag_t *) tagp;
682                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
683                 tag->t_flags = cpu_to_be32(tag_flag);
684                 tagp += tag_bytes;
685                 space_left -= tag_bytes;
686
687                 if (first_tag) {
688                         memcpy (tagp, journal->j_uuid, 16);
689                         tagp += 16;
690                         space_left -= 16;
691                         first_tag = 0;
692                 }
693
694                 /* If there's no more to do, or if the descriptor is full,
695                    let the IO rip! */
696
697                 if (bufs == journal->j_wbufsize ||
698                     commit_transaction->t_buffers == NULL ||
699                     space_left < tag_bytes + 16) {
700
701                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
702
703                         /* Write an end-of-descriptor marker before
704                            submitting the IOs.  "tag" still points to
705                            the last tag we set up. */
706
707                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
708
709 start_journal_io:
710                         for (i = 0; i < bufs; i++) {
711                                 struct buffer_head *bh = wbuf[i];
712                                 /*
713                                  * Compute checksum.
714                                  */
715                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
716                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
717                                         crc32_sum =
718                                             jbd2_checksum_data(crc32_sum, bh);
719                                 }
720
721                                 lock_buffer(bh);
722                                 clear_buffer_dirty(bh);
723                                 set_buffer_uptodate(bh);
724                                 bh->b_end_io = journal_end_buffer_io_sync;
725                                 submit_bh(WRITE, bh);
726                         }
727                         cond_resched();
728                         stats.u.run.rs_blocks_logged += bufs;
729
730                         /* Force a new descriptor to be generated next
731                            time round the loop. */
732                         descriptor = NULL;
733                         bufs = 0;
734                 }
735         }
736
737         /* Done it all: now write the commit record asynchronously. */
738
739         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
740                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
741                 err = journal_submit_commit_record(journal, commit_transaction,
742                                                  &cbh, crc32_sum);
743                 if (err)
744                         __jbd2_journal_abort_hard(journal);
745
746                 spin_lock(&journal->j_list_lock);
747                 err = journal_wait_on_locked_list(journal,
748                                                 commit_transaction);
749                 spin_unlock(&journal->j_list_lock);
750                 if (err)
751                         __jbd2_journal_abort_hard(journal);
752         }
753
754         /* Lo and behold: we have just managed to send a transaction to
755            the log.  Before we can commit it, wait for the IO so far to
756            complete.  Control buffers being written are on the
757            transaction's t_log_list queue, and metadata buffers are on
758            the t_iobuf_list queue.
759
760            Wait for the buffers in reverse order.  That way we are
761            less likely to be woken up until all IOs have completed, and
762            so we incur less scheduling load.
763         */
764
765         jbd_debug(3, "JBD: commit phase 4\n");
766
767         /*
768          * akpm: these are BJ_IO, and j_list_lock is not needed.
769          * See __journal_try_to_free_buffer.
770          */
771 wait_for_iobuf:
772         while (commit_transaction->t_iobuf_list != NULL) {
773                 struct buffer_head *bh;
774
775                 jh = commit_transaction->t_iobuf_list->b_tprev;
776                 bh = jh2bh(jh);
777                 if (buffer_locked(bh)) {
778                         wait_on_buffer(bh);
779                         goto wait_for_iobuf;
780                 }
781                 if (cond_resched())
782                         goto wait_for_iobuf;
783
784                 if (unlikely(!buffer_uptodate(bh)))
785                         err = -EIO;
786
787                 clear_buffer_jwrite(bh);
788
789                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
790                 jbd2_journal_unfile_buffer(journal, jh);
791
792                 /*
793                  * ->t_iobuf_list should contain only dummy buffer_heads
794                  * which were created by jbd2_journal_write_metadata_buffer().
795                  */
796                 BUFFER_TRACE(bh, "dumping temporary bh");
797                 jbd2_journal_put_journal_head(jh);
798                 __brelse(bh);
799                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
800                 free_buffer_head(bh);
801
802                 /* We also have to unlock and free the corresponding
803                    shadowed buffer */
804                 jh = commit_transaction->t_shadow_list->b_tprev;
805                 bh = jh2bh(jh);
806                 clear_bit(BH_JWrite, &bh->b_state);
807                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
808
809                 /* The metadata is now released for reuse, but we need
810                    to remember it against this transaction so that when
811                    we finally commit, we can do any checkpointing
812                    required. */
813                 JBUFFER_TRACE(jh, "file as BJ_Forget");
814                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
815                 /* Wake up any transactions which were waiting for this
816                    IO to complete */
817                 wake_up_bit(&bh->b_state, BH_Unshadow);
818                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
819                 __brelse(bh);
820         }
821
822         J_ASSERT (commit_transaction->t_shadow_list == NULL);
823
824         jbd_debug(3, "JBD: commit phase 5\n");
825
826         /* Here we wait for the revoke record and descriptor record buffers */
827  wait_for_ctlbuf:
828         while (commit_transaction->t_log_list != NULL) {
829                 struct buffer_head *bh;
830
831                 jh = commit_transaction->t_log_list->b_tprev;
832                 bh = jh2bh(jh);
833                 if (buffer_locked(bh)) {
834                         wait_on_buffer(bh);
835                         goto wait_for_ctlbuf;
836                 }
837                 if (cond_resched())
838                         goto wait_for_ctlbuf;
839
840                 if (unlikely(!buffer_uptodate(bh)))
841                         err = -EIO;
842
843                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
844                 clear_buffer_jwrite(bh);
845                 jbd2_journal_unfile_buffer(journal, jh);
846                 jbd2_journal_put_journal_head(jh);
847                 __brelse(bh);           /* One for getblk */
848                 /* AKPM: bforget here */
849         }
850
851         jbd_debug(3, "JBD: commit phase 6\n");
852
853         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
854                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
855                 err = journal_submit_commit_record(journal, commit_transaction,
856                                                 &cbh, crc32_sum);
857                 if (err)
858                         __jbd2_journal_abort_hard(journal);
859         }
860         if (!err && !is_journal_aborted(journal))
861                 err = journal_wait_on_commit_record(cbh);
862
863         if (err)
864                 jbd2_journal_abort(journal, err);
865
866         /* End of a transaction!  Finally, we can do checkpoint
867            processing: any buffers committed as a result of this
868            transaction can be removed from any checkpoint list it was on
869            before. */
870
871         jbd_debug(3, "JBD: commit phase 7\n");
872
873         J_ASSERT(commit_transaction->t_sync_datalist == NULL);
874         J_ASSERT(commit_transaction->t_buffers == NULL);
875         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
876         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
877         J_ASSERT(commit_transaction->t_shadow_list == NULL);
878         J_ASSERT(commit_transaction->t_log_list == NULL);
879
880 restart_loop:
881         /*
882          * As there are other places (journal_unmap_buffer()) adding buffers
883          * to this list we have to be careful and hold the j_list_lock.
884          */
885         spin_lock(&journal->j_list_lock);
886         while (commit_transaction->t_forget) {
887                 transaction_t *cp_transaction;
888                 struct buffer_head *bh;
889
890                 jh = commit_transaction->t_forget;
891                 spin_unlock(&journal->j_list_lock);
892                 bh = jh2bh(jh);
893                 jbd_lock_bh_state(bh);
894                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
895                         jh->b_transaction == journal->j_running_transaction);
896
897                 /*
898                  * If there is undo-protected committed data against
899                  * this buffer, then we can remove it now.  If it is a
900                  * buffer needing such protection, the old frozen_data
901                  * field now points to a committed version of the
902                  * buffer, so rotate that field to the new committed
903                  * data.
904                  *
905                  * Otherwise, we can just throw away the frozen data now.
906                  */
907                 if (jh->b_committed_data) {
908                         jbd2_free(jh->b_committed_data, bh->b_size);
909                         jh->b_committed_data = NULL;
910                         if (jh->b_frozen_data) {
911                                 jh->b_committed_data = jh->b_frozen_data;
912                                 jh->b_frozen_data = NULL;
913                         }
914                 } else if (jh->b_frozen_data) {
915                         jbd2_free(jh->b_frozen_data, bh->b_size);
916                         jh->b_frozen_data = NULL;
917                 }
918
919                 spin_lock(&journal->j_list_lock);
920                 cp_transaction = jh->b_cp_transaction;
921                 if (cp_transaction) {
922                         JBUFFER_TRACE(jh, "remove from old cp transaction");
923                         cp_transaction->t_chp_stats.cs_dropped++;
924                         __jbd2_journal_remove_checkpoint(jh);
925                 }
926
927                 /* Only re-checkpoint the buffer_head if it is marked
928                  * dirty.  If the buffer was added to the BJ_Forget list
929                  * by jbd2_journal_forget, it may no longer be dirty and
930                  * there's no point in keeping a checkpoint record for
931                  * it. */
932
933                 /* A buffer which has been freed while still being
934                  * journaled by a previous transaction may end up still
935                  * being dirty here, but we want to avoid writing back
936                  * that buffer in the future now that the last use has
937                  * been committed.  That's not only a performance gain,
938                  * it also stops aliasing problems if the buffer is left
939                  * behind for writeback and gets reallocated for another
940                  * use in a different page. */
941                 if (buffer_freed(bh)) {
942                         clear_buffer_freed(bh);
943                         clear_buffer_jbddirty(bh);
944                 }
945
946                 if (buffer_jbddirty(bh)) {
947                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
948                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
949                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
950                         __jbd2_journal_refile_buffer(jh);
951                         jbd_unlock_bh_state(bh);
952                 } else {
953                         J_ASSERT_BH(bh, !buffer_dirty(bh));
954                         /* The buffer on BJ_Forget list and not jbddirty means
955                          * it has been freed by this transaction and hence it
956                          * could not have been reallocated until this
957                          * transaction has committed. *BUT* it could be
958                          * reallocated once we have written all the data to
959                          * disk and before we process the buffer on BJ_Forget
960                          * list. */
961                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
962                         __jbd2_journal_refile_buffer(jh);
963                         if (!jh->b_transaction) {
964                                 jbd_unlock_bh_state(bh);
965                                  /* needs a brelse */
966                                 jbd2_journal_remove_journal_head(bh);
967                                 release_buffer_page(bh);
968                         } else
969                                 jbd_unlock_bh_state(bh);
970                 }
971                 cond_resched_lock(&journal->j_list_lock);
972         }
973         spin_unlock(&journal->j_list_lock);
974         /*
975          * This is a bit sleazy.  We use j_list_lock to protect transition
976          * of a transaction into T_FINISHED state and calling
977          * __jbd2_journal_drop_transaction(). Otherwise we could race with
978          * other checkpointing code processing the transaction...
979          */
980         spin_lock(&journal->j_state_lock);
981         spin_lock(&journal->j_list_lock);
982         /*
983          * Now recheck if some buffers did not get attached to the transaction
984          * while the lock was dropped...
985          */
986         if (commit_transaction->t_forget) {
987                 spin_unlock(&journal->j_list_lock);
988                 spin_unlock(&journal->j_state_lock);
989                 goto restart_loop;
990         }
991
992         /* Done with this transaction! */
993
994         jbd_debug(3, "JBD: commit phase 8\n");
995
996         J_ASSERT(commit_transaction->t_state == T_COMMIT);
997
998         commit_transaction->t_start = jiffies;
999         stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
1000                                                 commit_transaction->t_start);
1001
1002         /*
1003          * File the transaction for history
1004          */
1005         stats.ts_type = JBD2_STATS_RUN;
1006         stats.ts_tid = commit_transaction->t_tid;
1007         stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
1008         spin_lock(&journal->j_history_lock);
1009         memcpy(journal->j_history + journal->j_history_cur, &stats,
1010                         sizeof(stats));
1011         if (++journal->j_history_cur == journal->j_history_max)
1012                 journal->j_history_cur = 0;
1013
1014         /*
1015          * Calculate overall stats
1016          */
1017         journal->j_stats.ts_tid++;
1018         journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
1019         journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
1020         journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
1021         journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
1022         journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
1023         journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
1024         journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
1025         journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
1026         spin_unlock(&journal->j_history_lock);
1027
1028         commit_transaction->t_state = T_FINISHED;
1029         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1030         journal->j_commit_sequence = commit_transaction->t_tid;
1031         journal->j_committing_transaction = NULL;
1032         spin_unlock(&journal->j_state_lock);
1033
1034         if (commit_transaction->t_checkpoint_list == NULL &&
1035             commit_transaction->t_checkpoint_io_list == NULL) {
1036                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1037         } else {
1038                 if (journal->j_checkpoint_transactions == NULL) {
1039                         journal->j_checkpoint_transactions = commit_transaction;
1040                         commit_transaction->t_cpnext = commit_transaction;
1041                         commit_transaction->t_cpprev = commit_transaction;
1042                 } else {
1043                         commit_transaction->t_cpnext =
1044                                 journal->j_checkpoint_transactions;
1045                         commit_transaction->t_cpprev =
1046                                 commit_transaction->t_cpnext->t_cpprev;
1047                         commit_transaction->t_cpnext->t_cpprev =
1048                                 commit_transaction;
1049                         commit_transaction->t_cpprev->t_cpnext =
1050                                 commit_transaction;
1051                 }
1052         }
1053         spin_unlock(&journal->j_list_lock);
1054
1055         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1056                   journal->j_commit_sequence, journal->j_tail_sequence);
1057
1058         wake_up(&journal->j_wait_done_commit);
1059 }