GNU Linux-libre 4.9.308-gnu1
[releases.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 #include <linux/writeback.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bio.h>
28 #include <linux/blkdev.h>
29 #include <linux/bitops.h>
30 #include <trace/events/jbd2.h>
31
32 /*
33  * IO end handler for temporary buffer_heads handling writes to the journal.
34  */
35 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
36 {
37         struct buffer_head *orig_bh = bh->b_private;
38
39         BUFFER_TRACE(bh, "");
40         if (uptodate)
41                 set_buffer_uptodate(bh);
42         else
43                 clear_buffer_uptodate(bh);
44         if (orig_bh) {
45                 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
46                 smp_mb__after_atomic();
47                 wake_up_bit(&orig_bh->b_state, BH_Shadow);
48         }
49         unlock_buffer(bh);
50 }
51
52 /*
53  * When an ext4 file is truncated, it is possible that some pages are not
54  * successfully freed, because they are attached to a committing transaction.
55  * After the transaction commits, these pages are left on the LRU, with no
56  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
57  * by the VM, but their apparent absence upsets the VM accounting, and it makes
58  * the numbers in /proc/meminfo look odd.
59  *
60  * So here, we have a buffer which has just come off the forget list.  Look to
61  * see if we can strip all buffers from the backing page.
62  *
63  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
64  * caller provided us with a ref against the buffer, and we drop that here.
65  */
66 static void release_buffer_page(struct buffer_head *bh)
67 {
68         struct page *page;
69
70         if (buffer_dirty(bh))
71                 goto nope;
72         if (atomic_read(&bh->b_count) != 1)
73                 goto nope;
74         page = bh->b_page;
75         if (!page)
76                 goto nope;
77         if (page->mapping)
78                 goto nope;
79
80         /* OK, it's a truncated page */
81         if (!trylock_page(page))
82                 goto nope;
83
84         get_page(page);
85         __brelse(bh);
86         try_to_free_buffers(page);
87         unlock_page(page);
88         put_page(page);
89         return;
90
91 nope:
92         __brelse(bh);
93 }
94
95 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
96 {
97         struct commit_header *h;
98         __u32 csum;
99
100         if (!jbd2_journal_has_csum_v2or3(j))
101                 return;
102
103         h = (struct commit_header *)(bh->b_data);
104         h->h_chksum_type = 0;
105         h->h_chksum_size = 0;
106         h->h_chksum[0] = 0;
107         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
108         h->h_chksum[0] = cpu_to_be32(csum);
109 }
110
111 /*
112  * Done it all: now submit the commit record.  We should have
113  * cleaned up our previous buffers by now, so if we are in abort
114  * mode we can now just skip the rest of the journal write
115  * entirely.
116  *
117  * Returns 1 if the journal needs to be aborted or 0 on success
118  */
119 static int journal_submit_commit_record(journal_t *journal,
120                                         transaction_t *commit_transaction,
121                                         struct buffer_head **cbh,
122                                         __u32 crc32_sum)
123 {
124         struct commit_header *tmp;
125         struct buffer_head *bh;
126         int ret;
127         struct timespec64 now = current_kernel_time64();
128
129         *cbh = NULL;
130
131         if (is_journal_aborted(journal))
132                 return 0;
133
134         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
135                                                 JBD2_COMMIT_BLOCK);
136         if (!bh)
137                 return 1;
138
139         tmp = (struct commit_header *)bh->b_data;
140         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
141         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
142
143         if (jbd2_has_feature_checksum(journal)) {
144                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
145                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
146                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
147         }
148         jbd2_commit_block_csum_set(journal, bh);
149
150         BUFFER_TRACE(bh, "submit commit block");
151         lock_buffer(bh);
152         clear_buffer_dirty(bh);
153         set_buffer_uptodate(bh);
154         bh->b_end_io = journal_end_buffer_io_sync;
155
156         if (journal->j_flags & JBD2_BARRIER &&
157             !jbd2_has_feature_async_commit(journal))
158                 ret = submit_bh(REQ_OP_WRITE, WRITE_SYNC | WRITE_FLUSH_FUA, bh);
159         else
160                 ret = submit_bh(REQ_OP_WRITE, WRITE_SYNC, bh);
161
162         *cbh = bh;
163         return ret;
164 }
165
166 /*
167  * This function along with journal_submit_commit_record
168  * allows to write the commit record asynchronously.
169  */
170 static int journal_wait_on_commit_record(journal_t *journal,
171                                          struct buffer_head *bh)
172 {
173         int ret = 0;
174
175         clear_buffer_dirty(bh);
176         wait_on_buffer(bh);
177
178         if (unlikely(!buffer_uptodate(bh)))
179                 ret = -EIO;
180         put_bh(bh);            /* One for getblk() */
181
182         return ret;
183 }
184
185 /*
186  * write the filemap data using writepage() address_space_operations.
187  * We don't do block allocation here even for delalloc. We don't
188  * use writepages() because with dealyed allocation we may be doing
189  * block allocation in writepages().
190  */
191 static int journal_submit_inode_data_buffers(struct address_space *mapping)
192 {
193         int ret;
194         struct writeback_control wbc = {
195                 .sync_mode =  WB_SYNC_ALL,
196                 .nr_to_write = mapping->nrpages * 2,
197                 .range_start = 0,
198                 .range_end = i_size_read(mapping->host),
199         };
200
201         ret = generic_writepages(mapping, &wbc);
202         return ret;
203 }
204
205 /*
206  * Submit all the data buffers of inode associated with the transaction to
207  * disk.
208  *
209  * We are in a committing transaction. Therefore no new inode can be added to
210  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
211  * operate on from being released while we write out pages.
212  */
213 static int journal_submit_data_buffers(journal_t *journal,
214                 transaction_t *commit_transaction)
215 {
216         struct jbd2_inode *jinode;
217         int err, ret = 0;
218         struct address_space *mapping;
219
220         spin_lock(&journal->j_list_lock);
221         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
222                 if (!(jinode->i_flags & JI_WRITE_DATA))
223                         continue;
224                 mapping = jinode->i_vfs_inode->i_mapping;
225                 jinode->i_flags |= JI_COMMIT_RUNNING;
226                 spin_unlock(&journal->j_list_lock);
227                 /*
228                  * submit the inode data buffers. We use writepage
229                  * instead of writepages. Because writepages can do
230                  * block allocation  with delalloc. We need to write
231                  * only allocated blocks here.
232                  */
233                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
234                 err = journal_submit_inode_data_buffers(mapping);
235                 if (!ret)
236                         ret = err;
237                 spin_lock(&journal->j_list_lock);
238                 J_ASSERT(jinode->i_transaction == commit_transaction);
239                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
240                 smp_mb();
241                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
242         }
243         spin_unlock(&journal->j_list_lock);
244         return ret;
245 }
246
247 /*
248  * Wait for data submitted for writeout, refile inodes to proper
249  * transaction if needed.
250  *
251  */
252 static int journal_finish_inode_data_buffers(journal_t *journal,
253                 transaction_t *commit_transaction)
254 {
255         struct jbd2_inode *jinode, *next_i;
256         int err, ret = 0;
257
258         /* For locking, see the comment in journal_submit_data_buffers() */
259         spin_lock(&journal->j_list_lock);
260         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
261                 if (!(jinode->i_flags & JI_WAIT_DATA))
262                         continue;
263                 jinode->i_flags |= JI_COMMIT_RUNNING;
264                 spin_unlock(&journal->j_list_lock);
265                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
266                 if (err) {
267                         /*
268                          * Because AS_EIO is cleared by
269                          * filemap_fdatawait_range(), set it again so
270                          * that user process can get -EIO from fsync().
271                          */
272                         mapping_set_error(jinode->i_vfs_inode->i_mapping, -EIO);
273
274                         if (!ret)
275                                 ret = err;
276                 }
277                 spin_lock(&journal->j_list_lock);
278                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
279                 smp_mb();
280                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
281         }
282
283         /* Now refile inode to proper lists */
284         list_for_each_entry_safe(jinode, next_i,
285                                  &commit_transaction->t_inode_list, i_list) {
286                 list_del(&jinode->i_list);
287                 if (jinode->i_next_transaction) {
288                         jinode->i_transaction = jinode->i_next_transaction;
289                         jinode->i_next_transaction = NULL;
290                         list_add(&jinode->i_list,
291                                 &jinode->i_transaction->t_inode_list);
292                 } else {
293                         jinode->i_transaction = NULL;
294                 }
295         }
296         spin_unlock(&journal->j_list_lock);
297
298         return ret;
299 }
300
301 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
302 {
303         struct page *page = bh->b_page;
304         char *addr;
305         __u32 checksum;
306
307         addr = kmap_atomic(page);
308         checksum = crc32_be(crc32_sum,
309                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
310         kunmap_atomic(addr);
311
312         return checksum;
313 }
314
315 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
316                                    unsigned long long block)
317 {
318         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
319         if (jbd2_has_feature_64bit(j))
320                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
321 }
322
323 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
324                                     struct buffer_head *bh, __u32 sequence)
325 {
326         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
327         struct page *page = bh->b_page;
328         __u8 *addr;
329         __u32 csum32;
330         __be32 seq;
331
332         if (!jbd2_journal_has_csum_v2or3(j))
333                 return;
334
335         seq = cpu_to_be32(sequence);
336         addr = kmap_atomic(page);
337         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
338         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
339                              bh->b_size);
340         kunmap_atomic(addr);
341
342         if (jbd2_has_feature_csum3(j))
343                 tag3->t_checksum = cpu_to_be32(csum32);
344         else
345                 tag->t_checksum = cpu_to_be16(csum32);
346 }
347 /*
348  * jbd2_journal_commit_transaction
349  *
350  * The primary function for committing a transaction to the log.  This
351  * function is called by the journal thread to begin a complete commit.
352  */
353 void jbd2_journal_commit_transaction(journal_t *journal)
354 {
355         struct transaction_stats_s stats;
356         transaction_t *commit_transaction;
357         struct journal_head *jh;
358         struct buffer_head *descriptor;
359         struct buffer_head **wbuf = journal->j_wbuf;
360         int bufs;
361         int flags;
362         int err;
363         unsigned long long blocknr;
364         ktime_t start_time;
365         u64 commit_time;
366         char *tagp = NULL;
367         journal_block_tag_t *tag = NULL;
368         int space_left = 0;
369         int first_tag = 0;
370         int tag_flag;
371         int i;
372         int tag_bytes = journal_tag_bytes(journal);
373         struct buffer_head *cbh = NULL; /* For transactional checksums */
374         __u32 crc32_sum = ~0;
375         struct blk_plug plug;
376         /* Tail of the journal */
377         unsigned long first_block;
378         tid_t first_tid;
379         int update_tail;
380         int csum_size = 0;
381         LIST_HEAD(io_bufs);
382         LIST_HEAD(log_bufs);
383
384         if (jbd2_journal_has_csum_v2or3(journal))
385                 csum_size = sizeof(struct jbd2_journal_block_tail);
386
387         /*
388          * First job: lock down the current transaction and wait for
389          * all outstanding updates to complete.
390          */
391
392         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
393         if (journal->j_flags & JBD2_FLUSHED) {
394                 jbd_debug(3, "super block updated\n");
395                 mutex_lock(&journal->j_checkpoint_mutex);
396                 /*
397                  * We hold j_checkpoint_mutex so tail cannot change under us.
398                  * We don't need any special data guarantees for writing sb
399                  * since journal is empty and it is ok for write to be
400                  * flushed only with transaction commit.
401                  */
402                 jbd2_journal_update_sb_log_tail(journal,
403                                                 journal->j_tail_sequence,
404                                                 journal->j_tail,
405                                                 WRITE_SYNC);
406                 mutex_unlock(&journal->j_checkpoint_mutex);
407         } else {
408                 jbd_debug(3, "superblock not updated\n");
409         }
410
411         J_ASSERT(journal->j_running_transaction != NULL);
412         J_ASSERT(journal->j_committing_transaction == NULL);
413
414         commit_transaction = journal->j_running_transaction;
415
416         trace_jbd2_start_commit(journal, commit_transaction);
417         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
418                         commit_transaction->t_tid);
419
420         write_lock(&journal->j_state_lock);
421         J_ASSERT(commit_transaction->t_state == T_RUNNING);
422         commit_transaction->t_state = T_LOCKED;
423
424         trace_jbd2_commit_locking(journal, commit_transaction);
425         stats.run.rs_wait = commit_transaction->t_max_wait;
426         stats.run.rs_request_delay = 0;
427         stats.run.rs_locked = jiffies;
428         if (commit_transaction->t_requested)
429                 stats.run.rs_request_delay =
430                         jbd2_time_diff(commit_transaction->t_requested,
431                                        stats.run.rs_locked);
432         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
433                                               stats.run.rs_locked);
434
435         spin_lock(&commit_transaction->t_handle_lock);
436         while (atomic_read(&commit_transaction->t_updates)) {
437                 DEFINE_WAIT(wait);
438
439                 prepare_to_wait(&journal->j_wait_updates, &wait,
440                                         TASK_UNINTERRUPTIBLE);
441                 if (atomic_read(&commit_transaction->t_updates)) {
442                         spin_unlock(&commit_transaction->t_handle_lock);
443                         write_unlock(&journal->j_state_lock);
444                         schedule();
445                         write_lock(&journal->j_state_lock);
446                         spin_lock(&commit_transaction->t_handle_lock);
447                 }
448                 finish_wait(&journal->j_wait_updates, &wait);
449         }
450         spin_unlock(&commit_transaction->t_handle_lock);
451
452         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
453                         journal->j_max_transaction_buffers);
454
455         /*
456          * First thing we are allowed to do is to discard any remaining
457          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
458          * that there are no such buffers: if a large filesystem
459          * operation like a truncate needs to split itself over multiple
460          * transactions, then it may try to do a jbd2_journal_restart() while
461          * there are still BJ_Reserved buffers outstanding.  These must
462          * be released cleanly from the current transaction.
463          *
464          * In this case, the filesystem must still reserve write access
465          * again before modifying the buffer in the new transaction, but
466          * we do not require it to remember exactly which old buffers it
467          * has reserved.  This is consistent with the existing behaviour
468          * that multiple jbd2_journal_get_write_access() calls to the same
469          * buffer are perfectly permissible.
470          */
471         while (commit_transaction->t_reserved_list) {
472                 jh = commit_transaction->t_reserved_list;
473                 JBUFFER_TRACE(jh, "reserved, unused: refile");
474                 /*
475                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
476                  * leave undo-committed data.
477                  */
478                 if (jh->b_committed_data) {
479                         struct buffer_head *bh = jh2bh(jh);
480
481                         jbd_lock_bh_state(bh);
482                         jbd2_free(jh->b_committed_data, bh->b_size);
483                         jh->b_committed_data = NULL;
484                         jbd_unlock_bh_state(bh);
485                 }
486                 jbd2_journal_refile_buffer(journal, jh);
487         }
488
489         /*
490          * Now try to drop any written-back buffers from the journal's
491          * checkpoint lists.  We do this *before* commit because it potentially
492          * frees some memory
493          */
494         spin_lock(&journal->j_list_lock);
495         __jbd2_journal_clean_checkpoint_list(journal, false);
496         spin_unlock(&journal->j_list_lock);
497
498         jbd_debug(3, "JBD2: commit phase 1\n");
499
500         /*
501          * Clear revoked flag to reflect there is no revoked buffers
502          * in the next transaction which is going to be started.
503          */
504         jbd2_clear_buffer_revoked_flags(journal);
505
506         /*
507          * Switch to a new revoke table.
508          */
509         jbd2_journal_switch_revoke_table(journal);
510
511         /*
512          * Reserved credits cannot be claimed anymore, free them
513          */
514         atomic_sub(atomic_read(&journal->j_reserved_credits),
515                    &commit_transaction->t_outstanding_credits);
516
517         trace_jbd2_commit_flushing(journal, commit_transaction);
518         stats.run.rs_flushing = jiffies;
519         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
520                                              stats.run.rs_flushing);
521
522         commit_transaction->t_state = T_FLUSH;
523         journal->j_committing_transaction = commit_transaction;
524         journal->j_running_transaction = NULL;
525         start_time = ktime_get();
526         commit_transaction->t_log_start = journal->j_head;
527         wake_up(&journal->j_wait_transaction_locked);
528         write_unlock(&journal->j_state_lock);
529
530         jbd_debug(3, "JBD2: commit phase 2a\n");
531
532         /*
533          * Now start flushing things to disk, in the order they appear
534          * on the transaction lists.  Data blocks go first.
535          */
536         err = journal_submit_data_buffers(journal, commit_transaction);
537         if (err)
538                 jbd2_journal_abort(journal, err);
539
540         blk_start_plug(&plug);
541         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
542
543         jbd_debug(3, "JBD2: commit phase 2b\n");
544
545         /*
546          * Way to go: we have now written out all of the data for a
547          * transaction!  Now comes the tricky part: we need to write out
548          * metadata.  Loop over the transaction's entire buffer list:
549          */
550         write_lock(&journal->j_state_lock);
551         commit_transaction->t_state = T_COMMIT;
552         write_unlock(&journal->j_state_lock);
553
554         trace_jbd2_commit_logging(journal, commit_transaction);
555         stats.run.rs_logging = jiffies;
556         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
557                                                stats.run.rs_logging);
558         stats.run.rs_blocks =
559                 atomic_read(&commit_transaction->t_outstanding_credits);
560         stats.run.rs_blocks_logged = 0;
561
562         J_ASSERT(commit_transaction->t_nr_buffers <=
563                  atomic_read(&commit_transaction->t_outstanding_credits));
564
565         err = 0;
566         bufs = 0;
567         descriptor = NULL;
568         while (commit_transaction->t_buffers) {
569
570                 /* Find the next buffer to be journaled... */
571
572                 jh = commit_transaction->t_buffers;
573
574                 /* If we're in abort mode, we just un-journal the buffer and
575                    release it. */
576
577                 if (is_journal_aborted(journal)) {
578                         clear_buffer_jbddirty(jh2bh(jh));
579                         JBUFFER_TRACE(jh, "journal is aborting: refile");
580                         jbd2_buffer_abort_trigger(jh,
581                                                   jh->b_frozen_data ?
582                                                   jh->b_frozen_triggers :
583                                                   jh->b_triggers);
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                         J_ASSERT (bufs == 0);
599
600                         jbd_debug(4, "JBD2: get descriptor\n");
601
602                         descriptor = jbd2_journal_get_descriptor_buffer(
603                                                         commit_transaction,
604                                                         JBD2_DESCRIPTOR_BLOCK);
605                         if (!descriptor) {
606                                 jbd2_journal_abort(journal, -EIO);
607                                 continue;
608                         }
609
610                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
611                                 (unsigned long long)descriptor->b_blocknr,
612                                 descriptor->b_data);
613                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
614                         space_left = descriptor->b_size -
615                                                 sizeof(journal_header_t);
616                         first_tag = 1;
617                         set_buffer_jwrite(descriptor);
618                         set_buffer_dirty(descriptor);
619                         wbuf[bufs++] = descriptor;
620
621                         /* Record it so that we can wait for IO
622                            completion later */
623                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
624                         jbd2_file_log_bh(&log_bufs, descriptor);
625                 }
626
627                 /* Where is the buffer to be written? */
628
629                 err = jbd2_journal_next_log_block(journal, &blocknr);
630                 /* If the block mapping failed, just abandon the buffer
631                    and repeat this loop: we'll fall into the
632                    refile-on-abort condition above. */
633                 if (err) {
634                         jbd2_journal_abort(journal, err);
635                         continue;
636                 }
637
638                 /*
639                  * start_this_handle() uses t_outstanding_credits to determine
640                  * the free space in the log, but this counter is changed
641                  * by jbd2_journal_next_log_block() also.
642                  */
643                 atomic_dec(&commit_transaction->t_outstanding_credits);
644
645                 /* Bump b_count to prevent truncate from stumbling over
646                    the shadowed buffer!  @@@ This can go if we ever get
647                    rid of the shadow pairing of buffers. */
648                 atomic_inc(&jh2bh(jh)->b_count);
649
650                 /*
651                  * Make a temporary IO buffer with which to write it out
652                  * (this will requeue the metadata buffer to BJ_Shadow).
653                  */
654                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
655                 JBUFFER_TRACE(jh, "ph3: write metadata");
656                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
657                                                 jh, &wbuf[bufs], blocknr);
658                 if (flags < 0) {
659                         jbd2_journal_abort(journal, flags);
660                         continue;
661                 }
662                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
663
664                 /* Record the new block's tag in the current descriptor
665                    buffer */
666
667                 tag_flag = 0;
668                 if (flags & 1)
669                         tag_flag |= JBD2_FLAG_ESCAPE;
670                 if (!first_tag)
671                         tag_flag |= JBD2_FLAG_SAME_UUID;
672
673                 tag = (journal_block_tag_t *) tagp;
674                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
675                 tag->t_flags = cpu_to_be16(tag_flag);
676                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
677                                         commit_transaction->t_tid);
678                 tagp += tag_bytes;
679                 space_left -= tag_bytes;
680                 bufs++;
681
682                 if (first_tag) {
683                         memcpy (tagp, journal->j_uuid, 16);
684                         tagp += 16;
685                         space_left -= 16;
686                         first_tag = 0;
687                 }
688
689                 /* If there's no more to do, or if the descriptor is full,
690                    let the IO rip! */
691
692                 if (bufs == journal->j_wbufsize ||
693                     commit_transaction->t_buffers == NULL ||
694                     space_left < tag_bytes + 16 + csum_size) {
695
696                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
697
698                         /* Write an end-of-descriptor marker before
699                            submitting the IOs.  "tag" still points to
700                            the last tag we set up. */
701
702                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
703 start_journal_io:
704                         if (descriptor)
705                                 jbd2_descriptor_block_csum_set(journal,
706                                                         descriptor);
707
708                         for (i = 0; i < bufs; i++) {
709                                 struct buffer_head *bh = wbuf[i];
710                                 /*
711                                  * Compute checksum.
712                                  */
713                                 if (jbd2_has_feature_checksum(journal)) {
714                                         crc32_sum =
715                                             jbd2_checksum_data(crc32_sum, bh);
716                                 }
717
718                                 lock_buffer(bh);
719                                 clear_buffer_dirty(bh);
720                                 set_buffer_uptodate(bh);
721                                 bh->b_end_io = journal_end_buffer_io_sync;
722                                 submit_bh(REQ_OP_WRITE, WRITE_SYNC, bh);
723                         }
724                         cond_resched();
725
726                         /* Force a new descriptor to be generated next
727                            time round the loop. */
728                         descriptor = NULL;
729                         bufs = 0;
730                 }
731         }
732
733         err = journal_finish_inode_data_buffers(journal, commit_transaction);
734         if (err) {
735                 printk(KERN_WARNING
736                         "JBD2: Detected IO errors while flushing file data "
737                        "on %s\n", journal->j_devname);
738                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
739                         jbd2_journal_abort(journal, err);
740                 err = 0;
741         }
742
743         /*
744          * Get current oldest transaction in the log before we issue flush
745          * to the filesystem device. After the flush we can be sure that
746          * blocks of all older transactions are checkpointed to persistent
747          * storage and we will be safe to update journal start in the
748          * superblock with the numbers we get here.
749          */
750         update_tail =
751                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
752
753         write_lock(&journal->j_state_lock);
754         if (update_tail) {
755                 long freed = first_block - journal->j_tail;
756
757                 if (first_block < journal->j_tail)
758                         freed += journal->j_last - journal->j_first;
759                 /* Update tail only if we free significant amount of space */
760                 if (freed < journal->j_maxlen / 4)
761                         update_tail = 0;
762         }
763         J_ASSERT(commit_transaction->t_state == T_COMMIT);
764         commit_transaction->t_state = T_COMMIT_DFLUSH;
765         write_unlock(&journal->j_state_lock);
766
767         /* 
768          * If the journal is not located on the file system device,
769          * then we must flush the file system device before we issue
770          * the commit record
771          */
772         if (commit_transaction->t_need_data_flush &&
773             (journal->j_fs_dev != journal->j_dev) &&
774             (journal->j_flags & JBD2_BARRIER))
775                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
776
777         /* Done it all: now write the commit record asynchronously. */
778         if (jbd2_has_feature_async_commit(journal)) {
779                 err = journal_submit_commit_record(journal, commit_transaction,
780                                                  &cbh, crc32_sum);
781                 if (err)
782                         jbd2_journal_abort(journal, err);
783         }
784
785         blk_finish_plug(&plug);
786
787         /* Lo and behold: we have just managed to send a transaction to
788            the log.  Before we can commit it, wait for the IO so far to
789            complete.  Control buffers being written are on the
790            transaction's t_log_list queue, and metadata buffers are on
791            the io_bufs list.
792
793            Wait for the buffers in reverse order.  That way we are
794            less likely to be woken up until all IOs have completed, and
795            so we incur less scheduling load.
796         */
797
798         jbd_debug(3, "JBD2: commit phase 3\n");
799
800         while (!list_empty(&io_bufs)) {
801                 struct buffer_head *bh = list_entry(io_bufs.prev,
802                                                     struct buffer_head,
803                                                     b_assoc_buffers);
804
805                 wait_on_buffer(bh);
806                 cond_resched();
807
808                 if (unlikely(!buffer_uptodate(bh)))
809                         err = -EIO;
810                 jbd2_unfile_log_bh(bh);
811                 stats.run.rs_blocks_logged++;
812
813                 /*
814                  * The list contains temporary buffer heads created by
815                  * jbd2_journal_write_metadata_buffer().
816                  */
817                 BUFFER_TRACE(bh, "dumping temporary bh");
818                 __brelse(bh);
819                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
820                 free_buffer_head(bh);
821
822                 /* We also have to refile the corresponding shadowed buffer */
823                 jh = commit_transaction->t_shadow_list->b_tprev;
824                 bh = jh2bh(jh);
825                 clear_buffer_jwrite(bh);
826                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
827                 J_ASSERT_BH(bh, !buffer_shadow(bh));
828
829                 /* The metadata is now released for reuse, but we need
830                    to remember it against this transaction so that when
831                    we finally commit, we can do any checkpointing
832                    required. */
833                 JBUFFER_TRACE(jh, "file as BJ_Forget");
834                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
835                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
836                 __brelse(bh);
837         }
838
839         J_ASSERT (commit_transaction->t_shadow_list == NULL);
840
841         jbd_debug(3, "JBD2: commit phase 4\n");
842
843         /* Here we wait for the revoke record and descriptor record buffers */
844         while (!list_empty(&log_bufs)) {
845                 struct buffer_head *bh;
846
847                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
848                 wait_on_buffer(bh);
849                 cond_resched();
850
851                 if (unlikely(!buffer_uptodate(bh)))
852                         err = -EIO;
853
854                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
855                 clear_buffer_jwrite(bh);
856                 jbd2_unfile_log_bh(bh);
857                 stats.run.rs_blocks_logged++;
858                 __brelse(bh);           /* One for getblk */
859                 /* AKPM: bforget here */
860         }
861
862         if (err)
863                 jbd2_journal_abort(journal, err);
864
865         jbd_debug(3, "JBD2: commit phase 5\n");
866         write_lock(&journal->j_state_lock);
867         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
868         commit_transaction->t_state = T_COMMIT_JFLUSH;
869         write_unlock(&journal->j_state_lock);
870
871         if (!jbd2_has_feature_async_commit(journal)) {
872                 err = journal_submit_commit_record(journal, commit_transaction,
873                                                 &cbh, crc32_sum);
874                 if (err)
875                         jbd2_journal_abort(journal, err);
876         }
877         if (cbh)
878                 err = journal_wait_on_commit_record(journal, cbh);
879         stats.run.rs_blocks_logged++;
880         if (jbd2_has_feature_async_commit(journal) &&
881             journal->j_flags & JBD2_BARRIER) {
882                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
883         }
884
885         if (err)
886                 jbd2_journal_abort(journal, err);
887
888         /*
889          * Now disk caches for filesystem device are flushed so we are safe to
890          * erase checkpointed transactions from the log by updating journal
891          * superblock.
892          */
893         if (update_tail)
894                 jbd2_update_log_tail(journal, first_tid, first_block);
895
896         /* End of a transaction!  Finally, we can do checkpoint
897            processing: any buffers committed as a result of this
898            transaction can be removed from any checkpoint list it was on
899            before. */
900
901         jbd_debug(3, "JBD2: commit phase 6\n");
902
903         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
904         J_ASSERT(commit_transaction->t_buffers == NULL);
905         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
906         J_ASSERT(commit_transaction->t_shadow_list == NULL);
907
908 restart_loop:
909         /*
910          * As there are other places (journal_unmap_buffer()) adding buffers
911          * to this list we have to be careful and hold the j_list_lock.
912          */
913         spin_lock(&journal->j_list_lock);
914         while (commit_transaction->t_forget) {
915                 transaction_t *cp_transaction;
916                 struct buffer_head *bh;
917                 int try_to_free = 0;
918
919                 jh = commit_transaction->t_forget;
920                 spin_unlock(&journal->j_list_lock);
921                 bh = jh2bh(jh);
922                 /*
923                  * Get a reference so that bh cannot be freed before we are
924                  * done with it.
925                  */
926                 get_bh(bh);
927                 jbd_lock_bh_state(bh);
928                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
929
930                 /*
931                  * If there is undo-protected committed data against
932                  * this buffer, then we can remove it now.  If it is a
933                  * buffer needing such protection, the old frozen_data
934                  * field now points to a committed version of the
935                  * buffer, so rotate that field to the new committed
936                  * data.
937                  *
938                  * Otherwise, we can just throw away the frozen data now.
939                  *
940                  * We also know that the frozen data has already fired
941                  * its triggers if they exist, so we can clear that too.
942                  */
943                 if (jh->b_committed_data) {
944                         jbd2_free(jh->b_committed_data, bh->b_size);
945                         jh->b_committed_data = NULL;
946                         if (jh->b_frozen_data) {
947                                 jh->b_committed_data = jh->b_frozen_data;
948                                 jh->b_frozen_data = NULL;
949                                 jh->b_frozen_triggers = NULL;
950                         }
951                 } else if (jh->b_frozen_data) {
952                         jbd2_free(jh->b_frozen_data, bh->b_size);
953                         jh->b_frozen_data = NULL;
954                         jh->b_frozen_triggers = NULL;
955                 }
956
957                 spin_lock(&journal->j_list_lock);
958                 cp_transaction = jh->b_cp_transaction;
959                 if (cp_transaction) {
960                         JBUFFER_TRACE(jh, "remove from old cp transaction");
961                         cp_transaction->t_chp_stats.cs_dropped++;
962                         __jbd2_journal_remove_checkpoint(jh);
963                 }
964
965                 /* Only re-checkpoint the buffer_head if it is marked
966                  * dirty.  If the buffer was added to the BJ_Forget list
967                  * by jbd2_journal_forget, it may no longer be dirty and
968                  * there's no point in keeping a checkpoint record for
969                  * it. */
970
971                 /*
972                  * A buffer which has been freed while still being journaled
973                  * by a previous transaction, refile the buffer to BJ_Forget of
974                  * the running transaction. If the just committed transaction
975                  * contains "add to orphan" operation, we can completely
976                  * invalidate the buffer now. We are rather through in that
977                  * since the buffer may be still accessible when blocksize <
978                  * pagesize and it is attached to the last partial page.
979                  */
980                 if (buffer_freed(bh) && !jh->b_next_transaction) {
981                         struct address_space *mapping;
982
983                         clear_buffer_freed(bh);
984                         clear_buffer_jbddirty(bh);
985
986                         /*
987                          * Block device buffers need to stay mapped all the
988                          * time, so it is enough to clear buffer_jbddirty and
989                          * buffer_freed bits. For the file mapping buffers (i.e.
990                          * journalled data) we need to unmap buffer and clear
991                          * more bits. We also need to be careful about the check
992                          * because the data page mapping can get cleared under
993                          * our hands. Note that if mapping == NULL, we don't
994                          * need to make buffer unmapped because the page is
995                          * already detached from the mapping and buffers cannot
996                          * get reused.
997                          */
998                         mapping = READ_ONCE(bh->b_page->mapping);
999                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1000                                 clear_buffer_mapped(bh);
1001                                 clear_buffer_new(bh);
1002                                 clear_buffer_req(bh);
1003                                 bh->b_bdev = NULL;
1004                         }
1005                 }
1006
1007                 if (buffer_jbddirty(bh)) {
1008                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1009                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1010                         if (is_journal_aborted(journal))
1011                                 clear_buffer_jbddirty(bh);
1012                 } else {
1013                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1014                         /*
1015                          * The buffer on BJ_Forget list and not jbddirty means
1016                          * it has been freed by this transaction and hence it
1017                          * could not have been reallocated until this
1018                          * transaction has committed. *BUT* it could be
1019                          * reallocated once we have written all the data to
1020                          * disk and before we process the buffer on BJ_Forget
1021                          * list.
1022                          */
1023                         if (!jh->b_next_transaction)
1024                                 try_to_free = 1;
1025                 }
1026                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1027                 __jbd2_journal_refile_buffer(jh);
1028                 jbd_unlock_bh_state(bh);
1029                 if (try_to_free)
1030                         release_buffer_page(bh);        /* Drops bh reference */
1031                 else
1032                         __brelse(bh);
1033                 cond_resched_lock(&journal->j_list_lock);
1034         }
1035         spin_unlock(&journal->j_list_lock);
1036         /*
1037          * This is a bit sleazy.  We use j_list_lock to protect transition
1038          * of a transaction into T_FINISHED state and calling
1039          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1040          * other checkpointing code processing the transaction...
1041          */
1042         write_lock(&journal->j_state_lock);
1043         spin_lock(&journal->j_list_lock);
1044         /*
1045          * Now recheck if some buffers did not get attached to the transaction
1046          * while the lock was dropped...
1047          */
1048         if (commit_transaction->t_forget) {
1049                 spin_unlock(&journal->j_list_lock);
1050                 write_unlock(&journal->j_state_lock);
1051                 goto restart_loop;
1052         }
1053
1054         /* Add the transaction to the checkpoint list
1055          * __journal_remove_checkpoint() can not destroy transaction
1056          * under us because it is not marked as T_FINISHED yet */
1057         if (journal->j_checkpoint_transactions == NULL) {
1058                 journal->j_checkpoint_transactions = commit_transaction;
1059                 commit_transaction->t_cpnext = commit_transaction;
1060                 commit_transaction->t_cpprev = commit_transaction;
1061         } else {
1062                 commit_transaction->t_cpnext =
1063                         journal->j_checkpoint_transactions;
1064                 commit_transaction->t_cpprev =
1065                         commit_transaction->t_cpnext->t_cpprev;
1066                 commit_transaction->t_cpnext->t_cpprev =
1067                         commit_transaction;
1068                 commit_transaction->t_cpprev->t_cpnext =
1069                                 commit_transaction;
1070         }
1071         spin_unlock(&journal->j_list_lock);
1072
1073         /* Done with this transaction! */
1074
1075         jbd_debug(3, "JBD2: commit phase 7\n");
1076
1077         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1078
1079         commit_transaction->t_start = jiffies;
1080         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1081                                               commit_transaction->t_start);
1082
1083         /*
1084          * File the transaction statistics
1085          */
1086         stats.ts_tid = commit_transaction->t_tid;
1087         stats.run.rs_handle_count =
1088                 atomic_read(&commit_transaction->t_handle_count);
1089         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1090                              commit_transaction->t_tid, &stats.run);
1091         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1092
1093         commit_transaction->t_state = T_COMMIT_CALLBACK;
1094         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1095         journal->j_commit_sequence = commit_transaction->t_tid;
1096         journal->j_committing_transaction = NULL;
1097         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1098
1099         /*
1100          * weight the commit time higher than the average time so we don't
1101          * react too strongly to vast changes in the commit time
1102          */
1103         if (likely(journal->j_average_commit_time))
1104                 journal->j_average_commit_time = (commit_time +
1105                                 journal->j_average_commit_time*3) / 4;
1106         else
1107                 journal->j_average_commit_time = commit_time;
1108
1109         write_unlock(&journal->j_state_lock);
1110
1111         if (journal->j_commit_callback)
1112                 journal->j_commit_callback(journal, commit_transaction);
1113
1114         trace_jbd2_end_commit(journal, commit_transaction);
1115         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1116                   journal->j_commit_sequence, journal->j_tail_sequence);
1117
1118         write_lock(&journal->j_state_lock);
1119         spin_lock(&journal->j_list_lock);
1120         commit_transaction->t_state = T_FINISHED;
1121         /* Check if the transaction can be dropped now that we are finished */
1122         if (commit_transaction->t_checkpoint_list == NULL &&
1123             commit_transaction->t_checkpoint_io_list == NULL) {
1124                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1125                 jbd2_journal_free_transaction(commit_transaction);
1126         }
1127         spin_unlock(&journal->j_list_lock);
1128         write_unlock(&journal->j_state_lock);
1129         wake_up(&journal->j_wait_done_commit);
1130
1131         /*
1132          * Calculate overall stats
1133          */
1134         spin_lock(&journal->j_history_lock);
1135         journal->j_stats.ts_tid++;
1136         journal->j_stats.ts_requested += stats.ts_requested;
1137         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1138         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1139         journal->j_stats.run.rs_running += stats.run.rs_running;
1140         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1141         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1142         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1143         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1144         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1145         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1146         spin_unlock(&journal->j_history_lock);
1147 }