Merge branch 'for-linus' of git://neil.brown.name/md
[linux-2.6] / fs / jbd2 / commit.c
1 /*
2  * linux/fs/jbd2/commit.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Journal commit routines for the generic filesystem journaling code;
13  * part of the ext2fs journaling system.
14  */
15
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/crc32.h>
25 #include <linux/writeback.h>
26 #include <linux/backing-dev.h>
27
28 /*
29  * Default IO end handler for temporary BJ_IO buffer_heads.
30  */
31 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
32 {
33         BUFFER_TRACE(bh, "");
34         if (uptodate)
35                 set_buffer_uptodate(bh);
36         else
37                 clear_buffer_uptodate(bh);
38         unlock_buffer(bh);
39 }
40
41 /*
42  * When an ext4 file is truncated, it is possible that some pages are not
43  * successfully freed, because they are attached to a committing transaction.
44  * After the transaction commits, these pages are left on the LRU, with no
45  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
46  * by the VM, but their apparent absence upsets the VM accounting, and it makes
47  * the numbers in /proc/meminfo look odd.
48  *
49  * So here, we have a buffer which has just come off the forget list.  Look to
50  * see if we can strip all buffers from the backing page.
51  *
52  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
53  * caller provided us with a ref against the buffer, and we drop that here.
54  */
55 static void release_buffer_page(struct buffer_head *bh)
56 {
57         struct page *page;
58
59         if (buffer_dirty(bh))
60                 goto nope;
61         if (atomic_read(&bh->b_count) != 1)
62                 goto nope;
63         page = bh->b_page;
64         if (!page)
65                 goto nope;
66         if (page->mapping)
67                 goto nope;
68
69         /* OK, it's a truncated page */
70         if (TestSetPageLocked(page))
71                 goto nope;
72
73         page_cache_get(page);
74         __brelse(bh);
75         try_to_free_buffers(page);
76         unlock_page(page);
77         page_cache_release(page);
78         return;
79
80 nope:
81         __brelse(bh);
82 }
83
84 /*
85  * Done it all: now submit the commit record.  We should have
86  * cleaned up our previous buffers by now, so if we are in abort
87  * mode we can now just skip the rest of the journal write
88  * entirely.
89  *
90  * Returns 1 if the journal needs to be aborted or 0 on success
91  */
92 static int journal_submit_commit_record(journal_t *journal,
93                                         transaction_t *commit_transaction,
94                                         struct buffer_head **cbh,
95                                         __u32 crc32_sum)
96 {
97         struct journal_head *descriptor;
98         struct commit_header *tmp;
99         struct buffer_head *bh;
100         int ret;
101         int barrier_done = 0;
102         struct timespec now = current_kernel_time();
103
104         if (is_journal_aborted(journal))
105                 return 0;
106
107         descriptor = jbd2_journal_get_descriptor_buffer(journal);
108         if (!descriptor)
109                 return 1;
110
111         bh = jh2bh(descriptor);
112
113         tmp = (struct commit_header *)bh->b_data;
114         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
115         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
116         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
117         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
118         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
119
120         if (JBD2_HAS_COMPAT_FEATURE(journal,
121                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
122                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
123                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
124                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
125         }
126
127         JBUFFER_TRACE(descriptor, "submit commit block");
128         lock_buffer(bh);
129         get_bh(bh);
130         set_buffer_dirty(bh);
131         set_buffer_uptodate(bh);
132         bh->b_end_io = journal_end_buffer_io_sync;
133
134         if (journal->j_flags & JBD2_BARRIER &&
135                 !JBD2_HAS_INCOMPAT_FEATURE(journal,
136                                          JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
137                 set_buffer_ordered(bh);
138                 barrier_done = 1;
139         }
140         ret = submit_bh(WRITE, bh);
141         if (barrier_done)
142                 clear_buffer_ordered(bh);
143
144         /* is it possible for another commit to fail at roughly
145          * the same time as this one?  If so, we don't want to
146          * trust the barrier flag in the super, but instead want
147          * to remember if we sent a barrier request
148          */
149         if (ret == -EOPNOTSUPP && barrier_done) {
150                 char b[BDEVNAME_SIZE];
151
152                 printk(KERN_WARNING
153                         "JBD: barrier-based sync failed on %s - "
154                         "disabling barriers\n",
155                         bdevname(journal->j_dev, b));
156                 spin_lock(&journal->j_state_lock);
157                 journal->j_flags &= ~JBD2_BARRIER;
158                 spin_unlock(&journal->j_state_lock);
159
160                 /* And try again, without the barrier */
161                 lock_buffer(bh);
162                 set_buffer_uptodate(bh);
163                 set_buffer_dirty(bh);
164                 ret = submit_bh(WRITE, bh);
165         }
166         *cbh = bh;
167         return ret;
168 }
169
170 /*
171  * This function along with journal_submit_commit_record
172  * allows to write the commit record asynchronously.
173  */
174 static int journal_wait_on_commit_record(struct buffer_head *bh)
175 {
176         int ret = 0;
177
178         clear_buffer_dirty(bh);
179         wait_on_buffer(bh);
180
181         if (unlikely(!buffer_uptodate(bh)))
182                 ret = -EIO;
183         put_bh(bh);            /* One for getblk() */
184         jbd2_journal_put_journal_head(bh2jh(bh));
185
186         return ret;
187 }
188
189 /*
190  * write the filemap data using writepage() address_space_operations.
191  * We don't do block allocation here even for delalloc. We don't
192  * use writepages() because with dealyed allocation we may be doing
193  * block allocation in writepages().
194  */
195 static int journal_submit_inode_data_buffers(struct address_space *mapping)
196 {
197         int ret;
198         struct writeback_control wbc = {
199                 .sync_mode =  WB_SYNC_ALL,
200                 .nr_to_write = mapping->nrpages * 2,
201                 .range_start = 0,
202                 .range_end = i_size_read(mapping->host),
203                 .for_writepages = 1,
204         };
205
206         ret = generic_writepages(mapping, &wbc);
207         return ret;
208 }
209
210 /*
211  * Submit all the data buffers of inode associated with the transaction to
212  * disk.
213  *
214  * We are in a committing transaction. Therefore no new inode can be added to
215  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
216  * operate on from being released while we write out pages.
217  */
218 static int journal_submit_data_buffers(journal_t *journal,
219                 transaction_t *commit_transaction)
220 {
221         struct jbd2_inode *jinode;
222         int err, ret = 0;
223         struct address_space *mapping;
224
225         spin_lock(&journal->j_list_lock);
226         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
227                 mapping = jinode->i_vfs_inode->i_mapping;
228                 jinode->i_flags |= JI_COMMIT_RUNNING;
229                 spin_unlock(&journal->j_list_lock);
230                 /*
231                  * submit the inode data buffers. We use writepage
232                  * instead of writepages. Because writepages can do
233                  * block allocation  with delalloc. We need to write
234                  * only allocated blocks here.
235                  */
236                 err = journal_submit_inode_data_buffers(mapping);
237                 if (!ret)
238                         ret = err;
239                 spin_lock(&journal->j_list_lock);
240                 J_ASSERT(jinode->i_transaction == commit_transaction);
241                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
242                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
243         }
244         spin_unlock(&journal->j_list_lock);
245         return ret;
246 }
247
248 /*
249  * Wait for data submitted for writeout, refile inodes to proper
250  * transaction if needed.
251  *
252  */
253 static int journal_finish_inode_data_buffers(journal_t *journal,
254                 transaction_t *commit_transaction)
255 {
256         struct jbd2_inode *jinode, *next_i;
257         int err, ret = 0;
258
259         /* For locking, see the comment in journal_submit_data_buffers() */
260         spin_lock(&journal->j_list_lock);
261         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
262                 jinode->i_flags |= JI_COMMIT_RUNNING;
263                 spin_unlock(&journal->j_list_lock);
264                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
265                 if (!ret)
266                         ret = err;
267                 spin_lock(&journal->j_list_lock);
268                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
269                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
270         }
271
272         /* Now refile inode to proper lists */
273         list_for_each_entry_safe(jinode, next_i,
274                                  &commit_transaction->t_inode_list, i_list) {
275                 list_del(&jinode->i_list);
276                 if (jinode->i_next_transaction) {
277                         jinode->i_transaction = jinode->i_next_transaction;
278                         jinode->i_next_transaction = NULL;
279                         list_add(&jinode->i_list,
280                                 &jinode->i_transaction->t_inode_list);
281                 } else {
282                         jinode->i_transaction = NULL;
283                 }
284         }
285         spin_unlock(&journal->j_list_lock);
286
287         return ret;
288 }
289
290 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
291 {
292         struct page *page = bh->b_page;
293         char *addr;
294         __u32 checksum;
295
296         addr = kmap_atomic(page, KM_USER0);
297         checksum = crc32_be(crc32_sum,
298                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
299         kunmap_atomic(addr, KM_USER0);
300
301         return checksum;
302 }
303
304 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
305                                    unsigned long long block)
306 {
307         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
308         if (tag_bytes > JBD2_TAG_SIZE32)
309                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
310 }
311
312 /*
313  * jbd2_journal_commit_transaction
314  *
315  * The primary function for committing a transaction to the log.  This
316  * function is called by the journal thread to begin a complete commit.
317  */
318 void jbd2_journal_commit_transaction(journal_t *journal)
319 {
320         struct transaction_stats_s stats;
321         transaction_t *commit_transaction;
322         struct journal_head *jh, *new_jh, *descriptor;
323         struct buffer_head **wbuf = journal->j_wbuf;
324         int bufs;
325         int flags;
326         int err;
327         unsigned long long blocknr;
328         char *tagp = NULL;
329         journal_header_t *header;
330         journal_block_tag_t *tag = NULL;
331         int space_left = 0;
332         int first_tag = 0;
333         int tag_flag;
334         int i;
335         int tag_bytes = journal_tag_bytes(journal);
336         struct buffer_head *cbh = NULL; /* For transactional checksums */
337         __u32 crc32_sum = ~0;
338
339         /*
340          * First job: lock down the current transaction and wait for
341          * all outstanding updates to complete.
342          */
343
344 #ifdef COMMIT_STATS
345         spin_lock(&journal->j_list_lock);
346         summarise_journal_usage(journal);
347         spin_unlock(&journal->j_list_lock);
348 #endif
349
350         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
351         if (journal->j_flags & JBD2_FLUSHED) {
352                 jbd_debug(3, "super block updated\n");
353                 jbd2_journal_update_superblock(journal, 1);
354         } else {
355                 jbd_debug(3, "superblock not updated\n");
356         }
357
358         J_ASSERT(journal->j_running_transaction != NULL);
359         J_ASSERT(journal->j_committing_transaction == NULL);
360
361         commit_transaction = journal->j_running_transaction;
362         J_ASSERT(commit_transaction->t_state == T_RUNNING);
363
364         jbd_debug(1, "JBD: starting commit of transaction %d\n",
365                         commit_transaction->t_tid);
366
367         spin_lock(&journal->j_state_lock);
368         commit_transaction->t_state = T_LOCKED;
369
370         stats.u.run.rs_wait = commit_transaction->t_max_wait;
371         stats.u.run.rs_locked = jiffies;
372         stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
373                                                 stats.u.run.rs_locked);
374
375         spin_lock(&commit_transaction->t_handle_lock);
376         while (commit_transaction->t_updates) {
377                 DEFINE_WAIT(wait);
378
379                 prepare_to_wait(&journal->j_wait_updates, &wait,
380                                         TASK_UNINTERRUPTIBLE);
381                 if (commit_transaction->t_updates) {
382                         spin_unlock(&commit_transaction->t_handle_lock);
383                         spin_unlock(&journal->j_state_lock);
384                         schedule();
385                         spin_lock(&journal->j_state_lock);
386                         spin_lock(&commit_transaction->t_handle_lock);
387                 }
388                 finish_wait(&journal->j_wait_updates, &wait);
389         }
390         spin_unlock(&commit_transaction->t_handle_lock);
391
392         J_ASSERT (commit_transaction->t_outstanding_credits <=
393                         journal->j_max_transaction_buffers);
394
395         /*
396          * First thing we are allowed to do is to discard any remaining
397          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
398          * that there are no such buffers: if a large filesystem
399          * operation like a truncate needs to split itself over multiple
400          * transactions, then it may try to do a jbd2_journal_restart() while
401          * there are still BJ_Reserved buffers outstanding.  These must
402          * be released cleanly from the current transaction.
403          *
404          * In this case, the filesystem must still reserve write access
405          * again before modifying the buffer in the new transaction, but
406          * we do not require it to remember exactly which old buffers it
407          * has reserved.  This is consistent with the existing behaviour
408          * that multiple jbd2_journal_get_write_access() calls to the same
409          * buffer are perfectly permissable.
410          */
411         while (commit_transaction->t_reserved_list) {
412                 jh = commit_transaction->t_reserved_list;
413                 JBUFFER_TRACE(jh, "reserved, unused: refile");
414                 /*
415                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
416                  * leave undo-committed data.
417                  */
418                 if (jh->b_committed_data) {
419                         struct buffer_head *bh = jh2bh(jh);
420
421                         jbd_lock_bh_state(bh);
422                         jbd2_free(jh->b_committed_data, bh->b_size);
423                         jh->b_committed_data = NULL;
424                         jbd_unlock_bh_state(bh);
425                 }
426                 jbd2_journal_refile_buffer(journal, jh);
427         }
428
429         /*
430          * Now try to drop any written-back buffers from the journal's
431          * checkpoint lists.  We do this *before* commit because it potentially
432          * frees some memory
433          */
434         spin_lock(&journal->j_list_lock);
435         __jbd2_journal_clean_checkpoint_list(journal);
436         spin_unlock(&journal->j_list_lock);
437
438         jbd_debug (3, "JBD: commit phase 1\n");
439
440         /*
441          * Switch to a new revoke table.
442          */
443         jbd2_journal_switch_revoke_table(journal);
444
445         stats.u.run.rs_flushing = jiffies;
446         stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
447                                                stats.u.run.rs_flushing);
448
449         commit_transaction->t_state = T_FLUSH;
450         journal->j_committing_transaction = commit_transaction;
451         journal->j_running_transaction = NULL;
452         commit_transaction->t_log_start = journal->j_head;
453         wake_up(&journal->j_wait_transaction_locked);
454         spin_unlock(&journal->j_state_lock);
455
456         jbd_debug (3, "JBD: commit phase 2\n");
457
458         /*
459          * Now start flushing things to disk, in the order they appear
460          * on the transaction lists.  Data blocks go first.
461          */
462         err = journal_submit_data_buffers(journal, commit_transaction);
463         if (err)
464                 jbd2_journal_abort(journal, err);
465
466         jbd2_journal_write_revoke_records(journal, commit_transaction);
467
468         jbd_debug(3, "JBD: commit phase 2\n");
469
470         /*
471          * Way to go: we have now written out all of the data for a
472          * transaction!  Now comes the tricky part: we need to write out
473          * metadata.  Loop over the transaction's entire buffer list:
474          */
475         spin_lock(&journal->j_state_lock);
476         commit_transaction->t_state = T_COMMIT;
477         spin_unlock(&journal->j_state_lock);
478
479         stats.u.run.rs_logging = jiffies;
480         stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
481                                                  stats.u.run.rs_logging);
482         stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
483         stats.u.run.rs_blocks_logged = 0;
484
485         J_ASSERT(commit_transaction->t_nr_buffers <=
486                  commit_transaction->t_outstanding_credits);
487
488         err = 0;
489         descriptor = NULL;
490         bufs = 0;
491         while (commit_transaction->t_buffers) {
492
493                 /* Find the next buffer to be journaled... */
494
495                 jh = commit_transaction->t_buffers;
496
497                 /* If we're in abort mode, we just un-journal the buffer and
498                    release it for background writing. */
499
500                 if (is_journal_aborted(journal)) {
501                         JBUFFER_TRACE(jh, "journal is aborting: refile");
502                         jbd2_journal_refile_buffer(journal, jh);
503                         /* If that was the last one, we need to clean up
504                          * any descriptor buffers which may have been
505                          * already allocated, even if we are now
506                          * aborting. */
507                         if (!commit_transaction->t_buffers)
508                                 goto start_journal_io;
509                         continue;
510                 }
511
512                 /* Make sure we have a descriptor block in which to
513                    record the metadata buffer. */
514
515                 if (!descriptor) {
516                         struct buffer_head *bh;
517
518                         J_ASSERT (bufs == 0);
519
520                         jbd_debug(4, "JBD: get descriptor\n");
521
522                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
523                         if (!descriptor) {
524                                 jbd2_journal_abort(journal, -EIO);
525                                 continue;
526                         }
527
528                         bh = jh2bh(descriptor);
529                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
530                                 (unsigned long long)bh->b_blocknr, bh->b_data);
531                         header = (journal_header_t *)&bh->b_data[0];
532                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
533                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
534                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
535
536                         tagp = &bh->b_data[sizeof(journal_header_t)];
537                         space_left = bh->b_size - sizeof(journal_header_t);
538                         first_tag = 1;
539                         set_buffer_jwrite(bh);
540                         set_buffer_dirty(bh);
541                         wbuf[bufs++] = bh;
542
543                         /* Record it so that we can wait for IO
544                            completion later */
545                         BUFFER_TRACE(bh, "ph3: file as descriptor");
546                         jbd2_journal_file_buffer(descriptor, commit_transaction,
547                                         BJ_LogCtl);
548                 }
549
550                 /* Where is the buffer to be written? */
551
552                 err = jbd2_journal_next_log_block(journal, &blocknr);
553                 /* If the block mapping failed, just abandon the buffer
554                    and repeat this loop: we'll fall into the
555                    refile-on-abort condition above. */
556                 if (err) {
557                         jbd2_journal_abort(journal, err);
558                         continue;
559                 }
560
561                 /*
562                  * start_this_handle() uses t_outstanding_credits to determine
563                  * the free space in the log, but this counter is changed
564                  * by jbd2_journal_next_log_block() also.
565                  */
566                 commit_transaction->t_outstanding_credits--;
567
568                 /* Bump b_count to prevent truncate from stumbling over
569                    the shadowed buffer!  @@@ This can go if we ever get
570                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
571                 atomic_inc(&jh2bh(jh)->b_count);
572
573                 /* Make a temporary IO buffer with which to write it out
574                    (this will requeue both the metadata buffer and the
575                    temporary IO buffer). new_bh goes on BJ_IO*/
576
577                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
578                 /*
579                  * akpm: jbd2_journal_write_metadata_buffer() sets
580                  * new_bh->b_transaction to commit_transaction.
581                  * We need to clean this up before we release new_bh
582                  * (which is of type BJ_IO)
583                  */
584                 JBUFFER_TRACE(jh, "ph3: write metadata");
585                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
586                                                       jh, &new_jh, blocknr);
587                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
588                 wbuf[bufs++] = jh2bh(new_jh);
589
590                 /* Record the new block's tag in the current descriptor
591                    buffer */
592
593                 tag_flag = 0;
594                 if (flags & 1)
595                         tag_flag |= JBD2_FLAG_ESCAPE;
596                 if (!first_tag)
597                         tag_flag |= JBD2_FLAG_SAME_UUID;
598
599                 tag = (journal_block_tag_t *) tagp;
600                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
601                 tag->t_flags = cpu_to_be32(tag_flag);
602                 tagp += tag_bytes;
603                 space_left -= tag_bytes;
604
605                 if (first_tag) {
606                         memcpy (tagp, journal->j_uuid, 16);
607                         tagp += 16;
608                         space_left -= 16;
609                         first_tag = 0;
610                 }
611
612                 /* If there's no more to do, or if the descriptor is full,
613                    let the IO rip! */
614
615                 if (bufs == journal->j_wbufsize ||
616                     commit_transaction->t_buffers == NULL ||
617                     space_left < tag_bytes + 16) {
618
619                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
620
621                         /* Write an end-of-descriptor marker before
622                            submitting the IOs.  "tag" still points to
623                            the last tag we set up. */
624
625                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
626
627 start_journal_io:
628                         for (i = 0; i < bufs; i++) {
629                                 struct buffer_head *bh = wbuf[i];
630                                 /*
631                                  * Compute checksum.
632                                  */
633                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
634                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
635                                         crc32_sum =
636                                             jbd2_checksum_data(crc32_sum, bh);
637                                 }
638
639                                 lock_buffer(bh);
640                                 clear_buffer_dirty(bh);
641                                 set_buffer_uptodate(bh);
642                                 bh->b_end_io = journal_end_buffer_io_sync;
643                                 submit_bh(WRITE, bh);
644                         }
645                         cond_resched();
646                         stats.u.run.rs_blocks_logged += bufs;
647
648                         /* Force a new descriptor to be generated next
649                            time round the loop. */
650                         descriptor = NULL;
651                         bufs = 0;
652                 }
653         }
654
655         /* Done it all: now write the commit record asynchronously. */
656
657         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
658                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
659                 err = journal_submit_commit_record(journal, commit_transaction,
660                                                  &cbh, crc32_sum);
661                 if (err)
662                         __jbd2_journal_abort_hard(journal);
663         }
664
665         /*
666          * This is the right place to wait for data buffers both for ASYNC
667          * and !ASYNC commit. If commit is ASYNC, we need to wait only after
668          * the commit block went to disk (which happens above). If commit is
669          * SYNC, we need to wait for data buffers before we start writing
670          * commit block, which happens below in such setting.
671          */
672         err = journal_finish_inode_data_buffers(journal, commit_transaction);
673         if (err)
674                 jbd2_journal_abort(journal, err);
675
676         /* Lo and behold: we have just managed to send a transaction to
677            the log.  Before we can commit it, wait for the IO so far to
678            complete.  Control buffers being written are on the
679            transaction's t_log_list queue, and metadata buffers are on
680            the t_iobuf_list queue.
681
682            Wait for the buffers in reverse order.  That way we are
683            less likely to be woken up until all IOs have completed, and
684            so we incur less scheduling load.
685         */
686
687         jbd_debug(3, "JBD: commit phase 3\n");
688
689         /*
690          * akpm: these are BJ_IO, and j_list_lock is not needed.
691          * See __journal_try_to_free_buffer.
692          */
693 wait_for_iobuf:
694         while (commit_transaction->t_iobuf_list != NULL) {
695                 struct buffer_head *bh;
696
697                 jh = commit_transaction->t_iobuf_list->b_tprev;
698                 bh = jh2bh(jh);
699                 if (buffer_locked(bh)) {
700                         wait_on_buffer(bh);
701                         goto wait_for_iobuf;
702                 }
703                 if (cond_resched())
704                         goto wait_for_iobuf;
705
706                 if (unlikely(!buffer_uptodate(bh)))
707                         err = -EIO;
708
709                 clear_buffer_jwrite(bh);
710
711                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
712                 jbd2_journal_unfile_buffer(journal, jh);
713
714                 /*
715                  * ->t_iobuf_list should contain only dummy buffer_heads
716                  * which were created by jbd2_journal_write_metadata_buffer().
717                  */
718                 BUFFER_TRACE(bh, "dumping temporary bh");
719                 jbd2_journal_put_journal_head(jh);
720                 __brelse(bh);
721                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
722                 free_buffer_head(bh);
723
724                 /* We also have to unlock and free the corresponding
725                    shadowed buffer */
726                 jh = commit_transaction->t_shadow_list->b_tprev;
727                 bh = jh2bh(jh);
728                 clear_bit(BH_JWrite, &bh->b_state);
729                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
730
731                 /* The metadata is now released for reuse, but we need
732                    to remember it against this transaction so that when
733                    we finally commit, we can do any checkpointing
734                    required. */
735                 JBUFFER_TRACE(jh, "file as BJ_Forget");
736                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
737                 /* Wake up any transactions which were waiting for this
738                    IO to complete */
739                 wake_up_bit(&bh->b_state, BH_Unshadow);
740                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
741                 __brelse(bh);
742         }
743
744         J_ASSERT (commit_transaction->t_shadow_list == NULL);
745
746         jbd_debug(3, "JBD: commit phase 4\n");
747
748         /* Here we wait for the revoke record and descriptor record buffers */
749  wait_for_ctlbuf:
750         while (commit_transaction->t_log_list != NULL) {
751                 struct buffer_head *bh;
752
753                 jh = commit_transaction->t_log_list->b_tprev;
754                 bh = jh2bh(jh);
755                 if (buffer_locked(bh)) {
756                         wait_on_buffer(bh);
757                         goto wait_for_ctlbuf;
758                 }
759                 if (cond_resched())
760                         goto wait_for_ctlbuf;
761
762                 if (unlikely(!buffer_uptodate(bh)))
763                         err = -EIO;
764
765                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
766                 clear_buffer_jwrite(bh);
767                 jbd2_journal_unfile_buffer(journal, jh);
768                 jbd2_journal_put_journal_head(jh);
769                 __brelse(bh);           /* One for getblk */
770                 /* AKPM: bforget here */
771         }
772
773         jbd_debug(3, "JBD: commit phase 5\n");
774
775         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
776                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
777                 err = journal_submit_commit_record(journal, commit_transaction,
778                                                 &cbh, crc32_sum);
779                 if (err)
780                         __jbd2_journal_abort_hard(journal);
781         }
782         if (!err && !is_journal_aborted(journal))
783                 err = journal_wait_on_commit_record(cbh);
784
785         if (err)
786                 jbd2_journal_abort(journal, err);
787
788         /* End of a transaction!  Finally, we can do checkpoint
789            processing: any buffers committed as a result of this
790            transaction can be removed from any checkpoint list it was on
791            before. */
792
793         jbd_debug(3, "JBD: commit phase 6\n");
794
795         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
796         J_ASSERT(commit_transaction->t_buffers == NULL);
797         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
798         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
799         J_ASSERT(commit_transaction->t_shadow_list == NULL);
800         J_ASSERT(commit_transaction->t_log_list == NULL);
801
802 restart_loop:
803         /*
804          * As there are other places (journal_unmap_buffer()) adding buffers
805          * to this list we have to be careful and hold the j_list_lock.
806          */
807         spin_lock(&journal->j_list_lock);
808         while (commit_transaction->t_forget) {
809                 transaction_t *cp_transaction;
810                 struct buffer_head *bh;
811
812                 jh = commit_transaction->t_forget;
813                 spin_unlock(&journal->j_list_lock);
814                 bh = jh2bh(jh);
815                 jbd_lock_bh_state(bh);
816                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
817                         jh->b_transaction == journal->j_running_transaction);
818
819                 /*
820                  * If there is undo-protected committed data against
821                  * this buffer, then we can remove it now.  If it is a
822                  * buffer needing such protection, the old frozen_data
823                  * field now points to a committed version of the
824                  * buffer, so rotate that field to the new committed
825                  * data.
826                  *
827                  * Otherwise, we can just throw away the frozen data now.
828                  */
829                 if (jh->b_committed_data) {
830                         jbd2_free(jh->b_committed_data, bh->b_size);
831                         jh->b_committed_data = NULL;
832                         if (jh->b_frozen_data) {
833                                 jh->b_committed_data = jh->b_frozen_data;
834                                 jh->b_frozen_data = NULL;
835                         }
836                 } else if (jh->b_frozen_data) {
837                         jbd2_free(jh->b_frozen_data, bh->b_size);
838                         jh->b_frozen_data = NULL;
839                 }
840
841                 spin_lock(&journal->j_list_lock);
842                 cp_transaction = jh->b_cp_transaction;
843                 if (cp_transaction) {
844                         JBUFFER_TRACE(jh, "remove from old cp transaction");
845                         cp_transaction->t_chp_stats.cs_dropped++;
846                         __jbd2_journal_remove_checkpoint(jh);
847                 }
848
849                 /* Only re-checkpoint the buffer_head if it is marked
850                  * dirty.  If the buffer was added to the BJ_Forget list
851                  * by jbd2_journal_forget, it may no longer be dirty and
852                  * there's no point in keeping a checkpoint record for
853                  * it. */
854
855                 /* A buffer which has been freed while still being
856                  * journaled by a previous transaction may end up still
857                  * being dirty here, but we want to avoid writing back
858                  * that buffer in the future now that the last use has
859                  * been committed.  That's not only a performance gain,
860                  * it also stops aliasing problems if the buffer is left
861                  * behind for writeback and gets reallocated for another
862                  * use in a different page. */
863                 if (buffer_freed(bh)) {
864                         clear_buffer_freed(bh);
865                         clear_buffer_jbddirty(bh);
866                 }
867
868                 if (buffer_jbddirty(bh)) {
869                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
870                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
871                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
872                         __jbd2_journal_refile_buffer(jh);
873                         jbd_unlock_bh_state(bh);
874                 } else {
875                         J_ASSERT_BH(bh, !buffer_dirty(bh));
876                         /* The buffer on BJ_Forget list and not jbddirty means
877                          * it has been freed by this transaction and hence it
878                          * could not have been reallocated until this
879                          * transaction has committed. *BUT* it could be
880                          * reallocated once we have written all the data to
881                          * disk and before we process the buffer on BJ_Forget
882                          * list. */
883                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
884                         __jbd2_journal_refile_buffer(jh);
885                         if (!jh->b_transaction) {
886                                 jbd_unlock_bh_state(bh);
887                                  /* needs a brelse */
888                                 jbd2_journal_remove_journal_head(bh);
889                                 release_buffer_page(bh);
890                         } else
891                                 jbd_unlock_bh_state(bh);
892                 }
893                 cond_resched_lock(&journal->j_list_lock);
894         }
895         spin_unlock(&journal->j_list_lock);
896         /*
897          * This is a bit sleazy.  We use j_list_lock to protect transition
898          * of a transaction into T_FINISHED state and calling
899          * __jbd2_journal_drop_transaction(). Otherwise we could race with
900          * other checkpointing code processing the transaction...
901          */
902         spin_lock(&journal->j_state_lock);
903         spin_lock(&journal->j_list_lock);
904         /*
905          * Now recheck if some buffers did not get attached to the transaction
906          * while the lock was dropped...
907          */
908         if (commit_transaction->t_forget) {
909                 spin_unlock(&journal->j_list_lock);
910                 spin_unlock(&journal->j_state_lock);
911                 goto restart_loop;
912         }
913
914         /* Done with this transaction! */
915
916         jbd_debug(3, "JBD: commit phase 7\n");
917
918         J_ASSERT(commit_transaction->t_state == T_COMMIT);
919
920         commit_transaction->t_start = jiffies;
921         stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
922                                                 commit_transaction->t_start);
923
924         /*
925          * File the transaction for history
926          */
927         stats.ts_type = JBD2_STATS_RUN;
928         stats.ts_tid = commit_transaction->t_tid;
929         stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
930         spin_lock(&journal->j_history_lock);
931         memcpy(journal->j_history + journal->j_history_cur, &stats,
932                         sizeof(stats));
933         if (++journal->j_history_cur == journal->j_history_max)
934                 journal->j_history_cur = 0;
935
936         /*
937          * Calculate overall stats
938          */
939         journal->j_stats.ts_tid++;
940         journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
941         journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
942         journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
943         journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
944         journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
945         journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
946         journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
947         journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
948         spin_unlock(&journal->j_history_lock);
949
950         commit_transaction->t_state = T_FINISHED;
951         J_ASSERT(commit_transaction == journal->j_committing_transaction);
952         journal->j_commit_sequence = commit_transaction->t_tid;
953         journal->j_committing_transaction = NULL;
954         spin_unlock(&journal->j_state_lock);
955
956         if (commit_transaction->t_checkpoint_list == NULL &&
957             commit_transaction->t_checkpoint_io_list == NULL) {
958                 __jbd2_journal_drop_transaction(journal, commit_transaction);
959         } else {
960                 if (journal->j_checkpoint_transactions == NULL) {
961                         journal->j_checkpoint_transactions = commit_transaction;
962                         commit_transaction->t_cpnext = commit_transaction;
963                         commit_transaction->t_cpprev = commit_transaction;
964                 } else {
965                         commit_transaction->t_cpnext =
966                                 journal->j_checkpoint_transactions;
967                         commit_transaction->t_cpprev =
968                                 commit_transaction->t_cpnext->t_cpprev;
969                         commit_transaction->t_cpnext->t_cpprev =
970                                 commit_transaction;
971                         commit_transaction->t_cpprev->t_cpnext =
972                                 commit_transaction;
973                 }
974         }
975         spin_unlock(&journal->j_list_lock);
976
977         jbd_debug(1, "JBD: commit %d complete, head %d\n",
978                   journal->j_commit_sequence, journal->j_tail_sequence);
979
980         wake_up(&journal->j_wait_done_commit);
981 }