jbd2: Call journal commit callback without holding j_list_lock
[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/marker.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/crc32.h>
26 #include <linux/writeback.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bio.h>
29
30 /*
31  * Default IO end handler for temporary BJ_IO buffer_heads.
32  */
33 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
34 {
35         BUFFER_TRACE(bh, "");
36         if (uptodate)
37                 set_buffer_uptodate(bh);
38         else
39                 clear_buffer_uptodate(bh);
40         unlock_buffer(bh);
41 }
42
43 /*
44  * When an ext4 file is truncated, it is possible that some pages are not
45  * successfully freed, because they are attached to a committing transaction.
46  * After the transaction commits, these pages are left on the LRU, with no
47  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
48  * by the VM, but their apparent absence upsets the VM accounting, and it makes
49  * the numbers in /proc/meminfo look odd.
50  *
51  * So here, we have a buffer which has just come off the forget list.  Look to
52  * see if we can strip all buffers from the backing page.
53  *
54  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
55  * caller provided us with a ref against the buffer, and we drop that here.
56  */
57 static void release_buffer_page(struct buffer_head *bh)
58 {
59         struct page *page;
60
61         if (buffer_dirty(bh))
62                 goto nope;
63         if (atomic_read(&bh->b_count) != 1)
64                 goto nope;
65         page = bh->b_page;
66         if (!page)
67                 goto nope;
68         if (page->mapping)
69                 goto nope;
70
71         /* OK, it's a truncated page */
72         if (!trylock_page(page))
73                 goto nope;
74
75         page_cache_get(page);
76         __brelse(bh);
77         try_to_free_buffers(page);
78         unlock_page(page);
79         page_cache_release(page);
80         return;
81
82 nope:
83         __brelse(bh);
84 }
85
86 /*
87  * Done it all: now submit the commit record.  We should have
88  * cleaned up our previous buffers by now, so if we are in abort
89  * mode we can now just skip the rest of the journal write
90  * entirely.
91  *
92  * Returns 1 if the journal needs to be aborted or 0 on success
93  */
94 static int journal_submit_commit_record(journal_t *journal,
95                                         transaction_t *commit_transaction,
96                                         struct buffer_head **cbh,
97                                         __u32 crc32_sum)
98 {
99         struct journal_head *descriptor;
100         struct commit_header *tmp;
101         struct buffer_head *bh;
102         int ret;
103         int barrier_done = 0;
104         struct timespec now = current_kernel_time();
105
106         if (is_journal_aborted(journal))
107                 return 0;
108
109         descriptor = jbd2_journal_get_descriptor_buffer(journal);
110         if (!descriptor)
111                 return 1;
112
113         bh = jh2bh(descriptor);
114
115         tmp = (struct commit_header *)bh->b_data;
116         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
117         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
118         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
119         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
120         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
121
122         if (JBD2_HAS_COMPAT_FEATURE(journal,
123                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
124                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
125                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
126                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
127         }
128
129         JBUFFER_TRACE(descriptor, "submit commit block");
130         lock_buffer(bh);
131         clear_buffer_dirty(bh);
132         set_buffer_uptodate(bh);
133         bh->b_end_io = journal_end_buffer_io_sync;
134
135         if (journal->j_flags & JBD2_BARRIER &&
136                 !JBD2_HAS_INCOMPAT_FEATURE(journal,
137                                          JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
138                 set_buffer_ordered(bh);
139                 barrier_done = 1;
140         }
141         ret = submit_bh(WRITE, bh);
142         if (barrier_done)
143                 clear_buffer_ordered(bh);
144
145         /* is it possible for another commit to fail at roughly
146          * the same time as this one?  If so, we don't want to
147          * trust the barrier flag in the super, but instead want
148          * to remember if we sent a barrier request
149          */
150         if (ret == -EOPNOTSUPP && barrier_done) {
151                 printk(KERN_WARNING
152                        "JBD: barrier-based sync failed on %s - "
153                        "disabling barriers\n", journal->j_devname);
154                 spin_lock(&journal->j_state_lock);
155                 journal->j_flags &= ~JBD2_BARRIER;
156                 spin_unlock(&journal->j_state_lock);
157
158                 /* And try again, without the barrier */
159                 lock_buffer(bh);
160                 set_buffer_uptodate(bh);
161                 clear_buffer_dirty(bh);
162                 ret = submit_bh(WRITE, bh);
163         }
164         *cbh = bh;
165         return ret;
166 }
167
168 /*
169  * This function along with journal_submit_commit_record
170  * allows to write the commit record asynchronously.
171  */
172 static int journal_wait_on_commit_record(journal_t *journal,
173                                          struct buffer_head *bh)
174 {
175         int ret = 0;
176
177 retry:
178         clear_buffer_dirty(bh);
179         wait_on_buffer(bh);
180         if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
181                 printk(KERN_WARNING
182                        "JBD2: wait_on_commit_record: sync failed on %s - "
183                        "disabling barriers\n", journal->j_devname);
184                 spin_lock(&journal->j_state_lock);
185                 journal->j_flags &= ~JBD2_BARRIER;
186                 spin_unlock(&journal->j_state_lock);
187
188                 lock_buffer(bh);
189                 clear_buffer_dirty(bh);
190                 set_buffer_uptodate(bh);
191                 bh->b_end_io = journal_end_buffer_io_sync;
192
193                 ret = submit_bh(WRITE_SYNC, bh);
194                 if (ret) {
195                         unlock_buffer(bh);
196                         return ret;
197                 }
198                 goto retry;
199         }
200
201         if (unlikely(!buffer_uptodate(bh)))
202                 ret = -EIO;
203         put_bh(bh);            /* One for getblk() */
204         jbd2_journal_put_journal_head(bh2jh(bh));
205
206         return ret;
207 }
208
209 /*
210  * write the filemap data using writepage() address_space_operations.
211  * We don't do block allocation here even for delalloc. We don't
212  * use writepages() because with dealyed allocation we may be doing
213  * block allocation in writepages().
214  */
215 static int journal_submit_inode_data_buffers(struct address_space *mapping)
216 {
217         int ret;
218         struct writeback_control wbc = {
219                 .sync_mode =  WB_SYNC_ALL,
220                 .nr_to_write = mapping->nrpages * 2,
221                 .range_start = 0,
222                 .range_end = i_size_read(mapping->host),
223                 .for_writepages = 1,
224         };
225
226         ret = generic_writepages(mapping, &wbc);
227         return ret;
228 }
229
230 /*
231  * Submit all the data buffers of inode associated with the transaction to
232  * disk.
233  *
234  * We are in a committing transaction. Therefore no new inode can be added to
235  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
236  * operate on from being released while we write out pages.
237  */
238 static int journal_submit_data_buffers(journal_t *journal,
239                 transaction_t *commit_transaction)
240 {
241         struct jbd2_inode *jinode;
242         int err, ret = 0;
243         struct address_space *mapping;
244
245         spin_lock(&journal->j_list_lock);
246         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
247                 mapping = jinode->i_vfs_inode->i_mapping;
248                 jinode->i_flags |= JI_COMMIT_RUNNING;
249                 spin_unlock(&journal->j_list_lock);
250                 /*
251                  * submit the inode data buffers. We use writepage
252                  * instead of writepages. Because writepages can do
253                  * block allocation  with delalloc. We need to write
254                  * only allocated blocks here.
255                  */
256                 err = journal_submit_inode_data_buffers(mapping);
257                 if (!ret)
258                         ret = err;
259                 spin_lock(&journal->j_list_lock);
260                 J_ASSERT(jinode->i_transaction == commit_transaction);
261                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
262                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
263         }
264         spin_unlock(&journal->j_list_lock);
265         return ret;
266 }
267
268 /*
269  * Wait for data submitted for writeout, refile inodes to proper
270  * transaction if needed.
271  *
272  */
273 static int journal_finish_inode_data_buffers(journal_t *journal,
274                 transaction_t *commit_transaction)
275 {
276         struct jbd2_inode *jinode, *next_i;
277         int err, ret = 0;
278
279         /* For locking, see the comment in journal_submit_data_buffers() */
280         spin_lock(&journal->j_list_lock);
281         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
282                 jinode->i_flags |= JI_COMMIT_RUNNING;
283                 spin_unlock(&journal->j_list_lock);
284                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
285                 if (err) {
286                         /*
287                          * Because AS_EIO is cleared by
288                          * wait_on_page_writeback_range(), set it again so
289                          * that user process can get -EIO from fsync().
290                          */
291                         set_bit(AS_EIO,
292                                 &jinode->i_vfs_inode->i_mapping->flags);
293
294                         if (!ret)
295                                 ret = err;
296                 }
297                 spin_lock(&journal->j_list_lock);
298                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
299                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
300         }
301
302         /* Now refile inode to proper lists */
303         list_for_each_entry_safe(jinode, next_i,
304                                  &commit_transaction->t_inode_list, i_list) {
305                 list_del(&jinode->i_list);
306                 if (jinode->i_next_transaction) {
307                         jinode->i_transaction = jinode->i_next_transaction;
308                         jinode->i_next_transaction = NULL;
309                         list_add(&jinode->i_list,
310                                 &jinode->i_transaction->t_inode_list);
311                 } else {
312                         jinode->i_transaction = NULL;
313                 }
314         }
315         spin_unlock(&journal->j_list_lock);
316
317         return ret;
318 }
319
320 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
321 {
322         struct page *page = bh->b_page;
323         char *addr;
324         __u32 checksum;
325
326         addr = kmap_atomic(page, KM_USER0);
327         checksum = crc32_be(crc32_sum,
328                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
329         kunmap_atomic(addr, KM_USER0);
330
331         return checksum;
332 }
333
334 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
335                                    unsigned long long block)
336 {
337         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
338         if (tag_bytes > JBD2_TAG_SIZE32)
339                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
340 }
341
342 /*
343  * jbd2_journal_commit_transaction
344  *
345  * The primary function for committing a transaction to the log.  This
346  * function is called by the journal thread to begin a complete commit.
347  */
348 void jbd2_journal_commit_transaction(journal_t *journal)
349 {
350         struct transaction_stats_s stats;
351         transaction_t *commit_transaction;
352         struct journal_head *jh, *new_jh, *descriptor;
353         struct buffer_head **wbuf = journal->j_wbuf;
354         int bufs;
355         int flags;
356         int err;
357         unsigned long long blocknr;
358         ktime_t start_time;
359         u64 commit_time;
360         char *tagp = NULL;
361         journal_header_t *header;
362         journal_block_tag_t *tag = NULL;
363         int space_left = 0;
364         int first_tag = 0;
365         int tag_flag;
366         int i, to_free = 0;
367         int tag_bytes = journal_tag_bytes(journal);
368         struct buffer_head *cbh = NULL; /* For transactional checksums */
369         __u32 crc32_sum = ~0;
370
371         /*
372          * First job: lock down the current transaction and wait for
373          * all outstanding updates to complete.
374          */
375
376 #ifdef COMMIT_STATS
377         spin_lock(&journal->j_list_lock);
378         summarise_journal_usage(journal);
379         spin_unlock(&journal->j_list_lock);
380 #endif
381
382         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
383         if (journal->j_flags & JBD2_FLUSHED) {
384                 jbd_debug(3, "super block updated\n");
385                 jbd2_journal_update_superblock(journal, 1);
386         } else {
387                 jbd_debug(3, "superblock not updated\n");
388         }
389
390         J_ASSERT(journal->j_running_transaction != NULL);
391         J_ASSERT(journal->j_committing_transaction == NULL);
392
393         commit_transaction = journal->j_running_transaction;
394         J_ASSERT(commit_transaction->t_state == T_RUNNING);
395
396         trace_mark(jbd2_start_commit, "dev %s transaction %d",
397                    journal->j_devname, commit_transaction->t_tid);
398         jbd_debug(1, "JBD: starting commit of transaction %d\n",
399                         commit_transaction->t_tid);
400
401         spin_lock(&journal->j_state_lock);
402         commit_transaction->t_state = T_LOCKED;
403
404         stats.u.run.rs_wait = commit_transaction->t_max_wait;
405         stats.u.run.rs_locked = jiffies;
406         stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
407                                                 stats.u.run.rs_locked);
408
409         spin_lock(&commit_transaction->t_handle_lock);
410         while (commit_transaction->t_updates) {
411                 DEFINE_WAIT(wait);
412
413                 prepare_to_wait(&journal->j_wait_updates, &wait,
414                                         TASK_UNINTERRUPTIBLE);
415                 if (commit_transaction->t_updates) {
416                         spin_unlock(&commit_transaction->t_handle_lock);
417                         spin_unlock(&journal->j_state_lock);
418                         schedule();
419                         spin_lock(&journal->j_state_lock);
420                         spin_lock(&commit_transaction->t_handle_lock);
421                 }
422                 finish_wait(&journal->j_wait_updates, &wait);
423         }
424         spin_unlock(&commit_transaction->t_handle_lock);
425
426         J_ASSERT (commit_transaction->t_outstanding_credits <=
427                         journal->j_max_transaction_buffers);
428
429         /*
430          * First thing we are allowed to do is to discard any remaining
431          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
432          * that there are no such buffers: if a large filesystem
433          * operation like a truncate needs to split itself over multiple
434          * transactions, then it may try to do a jbd2_journal_restart() while
435          * there are still BJ_Reserved buffers outstanding.  These must
436          * be released cleanly from the current transaction.
437          *
438          * In this case, the filesystem must still reserve write access
439          * again before modifying the buffer in the new transaction, but
440          * we do not require it to remember exactly which old buffers it
441          * has reserved.  This is consistent with the existing behaviour
442          * that multiple jbd2_journal_get_write_access() calls to the same
443          * buffer are perfectly permissable.
444          */
445         while (commit_transaction->t_reserved_list) {
446                 jh = commit_transaction->t_reserved_list;
447                 JBUFFER_TRACE(jh, "reserved, unused: refile");
448                 /*
449                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
450                  * leave undo-committed data.
451                  */
452                 if (jh->b_committed_data) {
453                         struct buffer_head *bh = jh2bh(jh);
454
455                         jbd_lock_bh_state(bh);
456                         jbd2_free(jh->b_committed_data, bh->b_size);
457                         jh->b_committed_data = NULL;
458                         jbd_unlock_bh_state(bh);
459                 }
460                 jbd2_journal_refile_buffer(journal, jh);
461         }
462
463         /*
464          * Now try to drop any written-back buffers from the journal's
465          * checkpoint lists.  We do this *before* commit because it potentially
466          * frees some memory
467          */
468         spin_lock(&journal->j_list_lock);
469         __jbd2_journal_clean_checkpoint_list(journal);
470         spin_unlock(&journal->j_list_lock);
471
472         jbd_debug (3, "JBD: commit phase 1\n");
473
474         /*
475          * Switch to a new revoke table.
476          */
477         jbd2_journal_switch_revoke_table(journal);
478
479         stats.u.run.rs_flushing = jiffies;
480         stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
481                                                stats.u.run.rs_flushing);
482
483         commit_transaction->t_state = T_FLUSH;
484         journal->j_committing_transaction = commit_transaction;
485         journal->j_running_transaction = NULL;
486         start_time = ktime_get();
487         commit_transaction->t_log_start = journal->j_head;
488         wake_up(&journal->j_wait_transaction_locked);
489         spin_unlock(&journal->j_state_lock);
490
491         jbd_debug (3, "JBD: commit phase 2\n");
492
493         /*
494          * Now start flushing things to disk, in the order they appear
495          * on the transaction lists.  Data blocks go first.
496          */
497         err = journal_submit_data_buffers(journal, commit_transaction);
498         if (err)
499                 jbd2_journal_abort(journal, err);
500
501         jbd2_journal_write_revoke_records(journal, commit_transaction);
502
503         jbd_debug(3, "JBD: commit phase 2\n");
504
505         /*
506          * Way to go: we have now written out all of the data for a
507          * transaction!  Now comes the tricky part: we need to write out
508          * metadata.  Loop over the transaction's entire buffer list:
509          */
510         spin_lock(&journal->j_state_lock);
511         commit_transaction->t_state = T_COMMIT;
512         spin_unlock(&journal->j_state_lock);
513
514         stats.u.run.rs_logging = jiffies;
515         stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
516                                                  stats.u.run.rs_logging);
517         stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
518         stats.u.run.rs_blocks_logged = 0;
519
520         J_ASSERT(commit_transaction->t_nr_buffers <=
521                  commit_transaction->t_outstanding_credits);
522
523         err = 0;
524         descriptor = NULL;
525         bufs = 0;
526         while (commit_transaction->t_buffers) {
527
528                 /* Find the next buffer to be journaled... */
529
530                 jh = commit_transaction->t_buffers;
531
532                 /* If we're in abort mode, we just un-journal the buffer and
533                    release it. */
534
535                 if (is_journal_aborted(journal)) {
536                         clear_buffer_jbddirty(jh2bh(jh));
537                         JBUFFER_TRACE(jh, "journal is aborting: refile");
538                         jbd2_journal_refile_buffer(journal, jh);
539                         /* If that was the last one, we need to clean up
540                          * any descriptor buffers which may have been
541                          * already allocated, even if we are now
542                          * aborting. */
543                         if (!commit_transaction->t_buffers)
544                                 goto start_journal_io;
545                         continue;
546                 }
547
548                 /* Make sure we have a descriptor block in which to
549                    record the metadata buffer. */
550
551                 if (!descriptor) {
552                         struct buffer_head *bh;
553
554                         J_ASSERT (bufs == 0);
555
556                         jbd_debug(4, "JBD: get descriptor\n");
557
558                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
559                         if (!descriptor) {
560                                 jbd2_journal_abort(journal, -EIO);
561                                 continue;
562                         }
563
564                         bh = jh2bh(descriptor);
565                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
566                                 (unsigned long long)bh->b_blocknr, bh->b_data);
567                         header = (journal_header_t *)&bh->b_data[0];
568                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
569                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
570                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
571
572                         tagp = &bh->b_data[sizeof(journal_header_t)];
573                         space_left = bh->b_size - sizeof(journal_header_t);
574                         first_tag = 1;
575                         set_buffer_jwrite(bh);
576                         set_buffer_dirty(bh);
577                         wbuf[bufs++] = bh;
578
579                         /* Record it so that we can wait for IO
580                            completion later */
581                         BUFFER_TRACE(bh, "ph3: file as descriptor");
582                         jbd2_journal_file_buffer(descriptor, commit_transaction,
583                                         BJ_LogCtl);
584                 }
585
586                 /* Where is the buffer to be written? */
587
588                 err = jbd2_journal_next_log_block(journal, &blocknr);
589                 /* If the block mapping failed, just abandon the buffer
590                    and repeat this loop: we'll fall into the
591                    refile-on-abort condition above. */
592                 if (err) {
593                         jbd2_journal_abort(journal, err);
594                         continue;
595                 }
596
597                 /*
598                  * start_this_handle() uses t_outstanding_credits to determine
599                  * the free space in the log, but this counter is changed
600                  * by jbd2_journal_next_log_block() also.
601                  */
602                 commit_transaction->t_outstanding_credits--;
603
604                 /* Bump b_count to prevent truncate from stumbling over
605                    the shadowed buffer!  @@@ This can go if we ever get
606                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
607                 atomic_inc(&jh2bh(jh)->b_count);
608
609                 /* Make a temporary IO buffer with which to write it out
610                    (this will requeue both the metadata buffer and the
611                    temporary IO buffer). new_bh goes on BJ_IO*/
612
613                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
614                 /*
615                  * akpm: jbd2_journal_write_metadata_buffer() sets
616                  * new_bh->b_transaction to commit_transaction.
617                  * We need to clean this up before we release new_bh
618                  * (which is of type BJ_IO)
619                  */
620                 JBUFFER_TRACE(jh, "ph3: write metadata");
621                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
622                                                       jh, &new_jh, blocknr);
623                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
624                 wbuf[bufs++] = jh2bh(new_jh);
625
626                 /* Record the new block's tag in the current descriptor
627                    buffer */
628
629                 tag_flag = 0;
630                 if (flags & 1)
631                         tag_flag |= JBD2_FLAG_ESCAPE;
632                 if (!first_tag)
633                         tag_flag |= JBD2_FLAG_SAME_UUID;
634
635                 tag = (journal_block_tag_t *) tagp;
636                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
637                 tag->t_flags = cpu_to_be32(tag_flag);
638                 tagp += tag_bytes;
639                 space_left -= tag_bytes;
640
641                 if (first_tag) {
642                         memcpy (tagp, journal->j_uuid, 16);
643                         tagp += 16;
644                         space_left -= 16;
645                         first_tag = 0;
646                 }
647
648                 /* If there's no more to do, or if the descriptor is full,
649                    let the IO rip! */
650
651                 if (bufs == journal->j_wbufsize ||
652                     commit_transaction->t_buffers == NULL ||
653                     space_left < tag_bytes + 16) {
654
655                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
656
657                         /* Write an end-of-descriptor marker before
658                            submitting the IOs.  "tag" still points to
659                            the last tag we set up. */
660
661                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
662
663 start_journal_io:
664                         for (i = 0; i < bufs; i++) {
665                                 struct buffer_head *bh = wbuf[i];
666                                 /*
667                                  * Compute checksum.
668                                  */
669                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
670                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
671                                         crc32_sum =
672                                             jbd2_checksum_data(crc32_sum, bh);
673                                 }
674
675                                 lock_buffer(bh);
676                                 clear_buffer_dirty(bh);
677                                 set_buffer_uptodate(bh);
678                                 bh->b_end_io = journal_end_buffer_io_sync;
679                                 submit_bh(WRITE, bh);
680                         }
681                         cond_resched();
682                         stats.u.run.rs_blocks_logged += bufs;
683
684                         /* Force a new descriptor to be generated next
685                            time round the loop. */
686                         descriptor = NULL;
687                         bufs = 0;
688                 }
689         }
690
691         /* Done it all: now write the commit record asynchronously. */
692
693         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
694                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
695                 err = journal_submit_commit_record(journal, commit_transaction,
696                                                  &cbh, crc32_sum);
697                 if (err)
698                         __jbd2_journal_abort_hard(journal);
699         }
700
701         /*
702          * This is the right place to wait for data buffers both for ASYNC
703          * and !ASYNC commit. If commit is ASYNC, we need to wait only after
704          * the commit block went to disk (which happens above). If commit is
705          * SYNC, we need to wait for data buffers before we start writing
706          * commit block, which happens below in such setting.
707          */
708         err = journal_finish_inode_data_buffers(journal, commit_transaction);
709         if (err) {
710                 printk(KERN_WARNING
711                         "JBD2: Detected IO errors while flushing file data "
712                        "on %s\n", journal->j_devname);
713                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
714                         jbd2_journal_abort(journal, err);
715                 err = 0;
716         }
717
718         /* Lo and behold: we have just managed to send a transaction to
719            the log.  Before we can commit it, wait for the IO so far to
720            complete.  Control buffers being written are on the
721            transaction's t_log_list queue, and metadata buffers are on
722            the t_iobuf_list queue.
723
724            Wait for the buffers in reverse order.  That way we are
725            less likely to be woken up until all IOs have completed, and
726            so we incur less scheduling load.
727         */
728
729         jbd_debug(3, "JBD: commit phase 3\n");
730
731         /*
732          * akpm: these are BJ_IO, and j_list_lock is not needed.
733          * See __journal_try_to_free_buffer.
734          */
735 wait_for_iobuf:
736         while (commit_transaction->t_iobuf_list != NULL) {
737                 struct buffer_head *bh;
738
739                 jh = commit_transaction->t_iobuf_list->b_tprev;
740                 bh = jh2bh(jh);
741                 if (buffer_locked(bh)) {
742                         wait_on_buffer(bh);
743                         goto wait_for_iobuf;
744                 }
745                 if (cond_resched())
746                         goto wait_for_iobuf;
747
748                 if (unlikely(!buffer_uptodate(bh)))
749                         err = -EIO;
750
751                 clear_buffer_jwrite(bh);
752
753                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
754                 jbd2_journal_unfile_buffer(journal, jh);
755
756                 /*
757                  * ->t_iobuf_list should contain only dummy buffer_heads
758                  * which were created by jbd2_journal_write_metadata_buffer().
759                  */
760                 BUFFER_TRACE(bh, "dumping temporary bh");
761                 jbd2_journal_put_journal_head(jh);
762                 __brelse(bh);
763                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
764                 free_buffer_head(bh);
765
766                 /* We also have to unlock and free the corresponding
767                    shadowed buffer */
768                 jh = commit_transaction->t_shadow_list->b_tprev;
769                 bh = jh2bh(jh);
770                 clear_bit(BH_JWrite, &bh->b_state);
771                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
772
773                 /* The metadata is now released for reuse, but we need
774                    to remember it against this transaction so that when
775                    we finally commit, we can do any checkpointing
776                    required. */
777                 JBUFFER_TRACE(jh, "file as BJ_Forget");
778                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
779                 /* Wake up any transactions which were waiting for this
780                    IO to complete */
781                 wake_up_bit(&bh->b_state, BH_Unshadow);
782                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
783                 __brelse(bh);
784         }
785
786         J_ASSERT (commit_transaction->t_shadow_list == NULL);
787
788         jbd_debug(3, "JBD: commit phase 4\n");
789
790         /* Here we wait for the revoke record and descriptor record buffers */
791  wait_for_ctlbuf:
792         while (commit_transaction->t_log_list != NULL) {
793                 struct buffer_head *bh;
794
795                 jh = commit_transaction->t_log_list->b_tprev;
796                 bh = jh2bh(jh);
797                 if (buffer_locked(bh)) {
798                         wait_on_buffer(bh);
799                         goto wait_for_ctlbuf;
800                 }
801                 if (cond_resched())
802                         goto wait_for_ctlbuf;
803
804                 if (unlikely(!buffer_uptodate(bh)))
805                         err = -EIO;
806
807                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
808                 clear_buffer_jwrite(bh);
809                 jbd2_journal_unfile_buffer(journal, jh);
810                 jbd2_journal_put_journal_head(jh);
811                 __brelse(bh);           /* One for getblk */
812                 /* AKPM: bforget here */
813         }
814
815         if (err)
816                 jbd2_journal_abort(journal, err);
817
818         jbd_debug(3, "JBD: commit phase 5\n");
819
820         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
821                 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
822                 err = journal_submit_commit_record(journal, commit_transaction,
823                                                 &cbh, crc32_sum);
824                 if (err)
825                         __jbd2_journal_abort_hard(journal);
826         }
827         if (!err && !is_journal_aborted(journal))
828                 err = journal_wait_on_commit_record(journal, cbh);
829
830         if (err)
831                 jbd2_journal_abort(journal, err);
832
833         /* End of a transaction!  Finally, we can do checkpoint
834            processing: any buffers committed as a result of this
835            transaction can be removed from any checkpoint list it was on
836            before. */
837
838         jbd_debug(3, "JBD: commit phase 6\n");
839
840         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
841         J_ASSERT(commit_transaction->t_buffers == NULL);
842         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
843         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
844         J_ASSERT(commit_transaction->t_shadow_list == NULL);
845         J_ASSERT(commit_transaction->t_log_list == NULL);
846
847 restart_loop:
848         /*
849          * As there are other places (journal_unmap_buffer()) adding buffers
850          * to this list we have to be careful and hold the j_list_lock.
851          */
852         spin_lock(&journal->j_list_lock);
853         while (commit_transaction->t_forget) {
854                 transaction_t *cp_transaction;
855                 struct buffer_head *bh;
856
857                 jh = commit_transaction->t_forget;
858                 spin_unlock(&journal->j_list_lock);
859                 bh = jh2bh(jh);
860                 jbd_lock_bh_state(bh);
861                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction ||
862                         jh->b_transaction == journal->j_running_transaction);
863
864                 /*
865                  * If there is undo-protected committed data against
866                  * this buffer, then we can remove it now.  If it is a
867                  * buffer needing such protection, the old frozen_data
868                  * field now points to a committed version of the
869                  * buffer, so rotate that field to the new committed
870                  * data.
871                  *
872                  * Otherwise, we can just throw away the frozen data now.
873                  */
874                 if (jh->b_committed_data) {
875                         jbd2_free(jh->b_committed_data, bh->b_size);
876                         jh->b_committed_data = NULL;
877                         if (jh->b_frozen_data) {
878                                 jh->b_committed_data = jh->b_frozen_data;
879                                 jh->b_frozen_data = NULL;
880                         }
881                 } else if (jh->b_frozen_data) {
882                         jbd2_free(jh->b_frozen_data, bh->b_size);
883                         jh->b_frozen_data = NULL;
884                 }
885
886                 spin_lock(&journal->j_list_lock);
887                 cp_transaction = jh->b_cp_transaction;
888                 if (cp_transaction) {
889                         JBUFFER_TRACE(jh, "remove from old cp transaction");
890                         cp_transaction->t_chp_stats.cs_dropped++;
891                         __jbd2_journal_remove_checkpoint(jh);
892                 }
893
894                 /* Only re-checkpoint the buffer_head if it is marked
895                  * dirty.  If the buffer was added to the BJ_Forget list
896                  * by jbd2_journal_forget, it may no longer be dirty and
897                  * there's no point in keeping a checkpoint record for
898                  * it. */
899
900                 /* A buffer which has been freed while still being
901                  * journaled by a previous transaction may end up still
902                  * being dirty here, but we want to avoid writing back
903                  * that buffer in the future now that the last use has
904                  * been committed.  That's not only a performance gain,
905                  * it also stops aliasing problems if the buffer is left
906                  * behind for writeback and gets reallocated for another
907                  * use in a different page. */
908                 if (buffer_freed(bh)) {
909                         clear_buffer_freed(bh);
910                         clear_buffer_jbddirty(bh);
911                 }
912
913                 if (buffer_jbddirty(bh)) {
914                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
915                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
916                         if (is_journal_aborted(journal))
917                                 clear_buffer_jbddirty(bh);
918                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
919                         __jbd2_journal_refile_buffer(jh);
920                         jbd_unlock_bh_state(bh);
921                 } else {
922                         J_ASSERT_BH(bh, !buffer_dirty(bh));
923                         /* The buffer on BJ_Forget list and not jbddirty means
924                          * it has been freed by this transaction and hence it
925                          * could not have been reallocated until this
926                          * transaction has committed. *BUT* it could be
927                          * reallocated once we have written all the data to
928                          * disk and before we process the buffer on BJ_Forget
929                          * list. */
930                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
931                         __jbd2_journal_refile_buffer(jh);
932                         if (!jh->b_transaction) {
933                                 jbd_unlock_bh_state(bh);
934                                  /* needs a brelse */
935                                 jbd2_journal_remove_journal_head(bh);
936                                 release_buffer_page(bh);
937                         } else
938                                 jbd_unlock_bh_state(bh);
939                 }
940                 cond_resched_lock(&journal->j_list_lock);
941         }
942         spin_unlock(&journal->j_list_lock);
943         /*
944          * This is a bit sleazy.  We use j_list_lock to protect transition
945          * of a transaction into T_FINISHED state and calling
946          * __jbd2_journal_drop_transaction(). Otherwise we could race with
947          * other checkpointing code processing the transaction...
948          */
949         spin_lock(&journal->j_state_lock);
950         spin_lock(&journal->j_list_lock);
951         /*
952          * Now recheck if some buffers did not get attached to the transaction
953          * while the lock was dropped...
954          */
955         if (commit_transaction->t_forget) {
956                 spin_unlock(&journal->j_list_lock);
957                 spin_unlock(&journal->j_state_lock);
958                 goto restart_loop;
959         }
960
961         /* Done with this transaction! */
962
963         jbd_debug(3, "JBD: commit phase 7\n");
964
965         J_ASSERT(commit_transaction->t_state == T_COMMIT);
966
967         commit_transaction->t_start = jiffies;
968         stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
969                                                 commit_transaction->t_start);
970
971         /*
972          * File the transaction for history
973          */
974         stats.ts_type = JBD2_STATS_RUN;
975         stats.ts_tid = commit_transaction->t_tid;
976         stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
977         spin_lock(&journal->j_history_lock);
978         memcpy(journal->j_history + journal->j_history_cur, &stats,
979                         sizeof(stats));
980         if (++journal->j_history_cur == journal->j_history_max)
981                 journal->j_history_cur = 0;
982
983         /*
984          * Calculate overall stats
985          */
986         journal->j_stats.ts_tid++;
987         journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
988         journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
989         journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
990         journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
991         journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
992         journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
993         journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
994         journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
995         spin_unlock(&journal->j_history_lock);
996
997         commit_transaction->t_state = T_FINISHED;
998         J_ASSERT(commit_transaction == journal->j_committing_transaction);
999         journal->j_commit_sequence = commit_transaction->t_tid;
1000         journal->j_committing_transaction = NULL;
1001         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1002
1003         /*
1004          * weight the commit time higher than the average time so we don't
1005          * react too strongly to vast changes in the commit time
1006          */
1007         if (likely(journal->j_average_commit_time))
1008                 journal->j_average_commit_time = (commit_time +
1009                                 journal->j_average_commit_time*3) / 4;
1010         else
1011                 journal->j_average_commit_time = commit_time;
1012         spin_unlock(&journal->j_state_lock);
1013
1014         if (commit_transaction->t_checkpoint_list == NULL &&
1015             commit_transaction->t_checkpoint_io_list == NULL) {
1016                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1017                 to_free = 1;
1018         } else {
1019                 if (journal->j_checkpoint_transactions == NULL) {
1020                         journal->j_checkpoint_transactions = commit_transaction;
1021                         commit_transaction->t_cpnext = commit_transaction;
1022                         commit_transaction->t_cpprev = commit_transaction;
1023                 } else {
1024                         commit_transaction->t_cpnext =
1025                                 journal->j_checkpoint_transactions;
1026                         commit_transaction->t_cpprev =
1027                                 commit_transaction->t_cpnext->t_cpprev;
1028                         commit_transaction->t_cpnext->t_cpprev =
1029                                 commit_transaction;
1030                         commit_transaction->t_cpprev->t_cpnext =
1031                                 commit_transaction;
1032                 }
1033         }
1034         spin_unlock(&journal->j_list_lock);
1035
1036         if (journal->j_commit_callback)
1037                 journal->j_commit_callback(journal, commit_transaction);
1038
1039         trace_mark(jbd2_end_commit, "dev %s transaction %d head %d",
1040                    journal->j_devname, commit_transaction->t_tid,
1041                    journal->j_tail_sequence);
1042         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1043                   journal->j_commit_sequence, journal->j_tail_sequence);
1044         if (to_free)
1045                 kfree(commit_transaction);
1046
1047         wake_up(&journal->j_wait_done_commit);
1048 }