GNU Linux-libre 6.1.86-gnu
[releases.git] / fs / jbd2 / commit.c
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * linux/fs/jbd2/commit.c
4  *
5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6  *
7  * Copyright 1998 Red Hat corp --- All Rights Reserved
8  *
9  * Journal commit routines for the generic filesystem journaling code;
10  * part of the ext2fs journaling system.
11  */
12
13 #include <linux/time.h>
14 #include <linux/fs.h>
15 #include <linux/jbd2.h>
16 #include <linux/errno.h>
17 #include <linux/slab.h>
18 #include <linux/mm.h>
19 #include <linux/pagemap.h>
20 #include <linux/jiffies.h>
21 #include <linux/crc32.h>
22 #include <linux/writeback.h>
23 #include <linux/backing-dev.h>
24 #include <linux/bio.h>
25 #include <linux/blkdev.h>
26 #include <linux/bitops.h>
27 #include <trace/events/jbd2.h>
28
29 /*
30  * IO end handler for temporary buffer_heads handling writes to the journal.
31  */
32 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
33 {
34         struct buffer_head *orig_bh = bh->b_private;
35
36         BUFFER_TRACE(bh, "");
37         if (uptodate)
38                 set_buffer_uptodate(bh);
39         else
40                 clear_buffer_uptodate(bh);
41         if (orig_bh) {
42                 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
43                 smp_mb__after_atomic();
44                 wake_up_bit(&orig_bh->b_state, BH_Shadow);
45         }
46         unlock_buffer(bh);
47 }
48
49 /*
50  * When an ext4 file is truncated, it is possible that some pages are not
51  * successfully freed, because they are attached to a committing transaction.
52  * After the transaction commits, these pages are left on the LRU, with no
53  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
54  * by the VM, but their apparent absence upsets the VM accounting, and it makes
55  * the numbers in /proc/meminfo look odd.
56  *
57  * So here, we have a buffer which has just come off the forget list.  Look to
58  * see if we can strip all buffers from the backing page.
59  *
60  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
61  * caller provided us with a ref against the buffer, and we drop that here.
62  */
63 static void release_buffer_page(struct buffer_head *bh)
64 {
65         struct folio *folio;
66         struct page *page;
67
68         if (buffer_dirty(bh))
69                 goto nope;
70         if (atomic_read(&bh->b_count) != 1)
71                 goto nope;
72         page = bh->b_page;
73         if (!page)
74                 goto nope;
75         folio = page_folio(page);
76         if (folio->mapping)
77                 goto nope;
78
79         /* OK, it's a truncated page */
80         if (!folio_trylock(folio))
81                 goto nope;
82
83         folio_get(folio);
84         __brelse(bh);
85         try_to_free_buffers(folio);
86         folio_unlock(folio);
87         folio_put(folio);
88         return;
89
90 nope:
91         __brelse(bh);
92 }
93
94 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
95 {
96         struct commit_header *h;
97         __u32 csum;
98
99         if (!jbd2_journal_has_csum_v2or3(j))
100                 return;
101
102         h = (struct commit_header *)(bh->b_data);
103         h->h_chksum_type = 0;
104         h->h_chksum_size = 0;
105         h->h_chksum[0] = 0;
106         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
107         h->h_chksum[0] = cpu_to_be32(csum);
108 }
109
110 /*
111  * Done it all: now submit the commit record.  We should have
112  * cleaned up our previous buffers by now, so if we are in abort
113  * mode we can now just skip the rest of the journal write
114  * entirely.
115  *
116  * Returns 1 if the journal needs to be aborted or 0 on success
117  */
118 static int journal_submit_commit_record(journal_t *journal,
119                                         transaction_t *commit_transaction,
120                                         struct buffer_head **cbh,
121                                         __u32 crc32_sum)
122 {
123         struct commit_header *tmp;
124         struct buffer_head *bh;
125         struct timespec64 now;
126         blk_opf_t write_flags = REQ_OP_WRITE | JBD2_JOURNAL_REQ_FLAGS;
127
128         *cbh = NULL;
129
130         if (is_journal_aborted(journal))
131                 return 0;
132
133         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
134                                                 JBD2_COMMIT_BLOCK);
135         if (!bh)
136                 return 1;
137
138         tmp = (struct commit_header *)bh->b_data;
139         ktime_get_coarse_real_ts64(&now);
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                 write_flags |= REQ_PREFLUSH | REQ_FUA;
159
160         submit_bh(write_flags, bh);
161         *cbh = bh;
162         return 0;
163 }
164
165 /*
166  * This function along with journal_submit_commit_record
167  * allows to write the commit record asynchronously.
168  */
169 static int journal_wait_on_commit_record(journal_t *journal,
170                                          struct buffer_head *bh)
171 {
172         int ret = 0;
173
174         clear_buffer_dirty(bh);
175         wait_on_buffer(bh);
176
177         if (unlikely(!buffer_uptodate(bh)))
178                 ret = -EIO;
179         put_bh(bh);            /* One for getblk() */
180
181         return ret;
182 }
183
184 /*
185  * write the filemap data using writepage() address_space_operations.
186  * We don't do block allocation here even for delalloc. We don't
187  * use writepages() because with delayed allocation we may be doing
188  * block allocation in writepages().
189  */
190 int jbd2_journal_submit_inode_data_buffers(struct jbd2_inode *jinode)
191 {
192         struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
193         struct writeback_control wbc = {
194                 .sync_mode =  WB_SYNC_ALL,
195                 .nr_to_write = mapping->nrpages * 2,
196                 .range_start = jinode->i_dirty_start,
197                 .range_end = jinode->i_dirty_end,
198         };
199
200         /*
201          * submit the inode data buffers. We use writepage
202          * instead of writepages. Because writepages can do
203          * block allocation with delalloc. We need to write
204          * only allocated blocks here.
205          */
206         return generic_writepages(mapping, &wbc);
207 }
208
209 /* Send all the data buffers related to an inode */
210 int jbd2_submit_inode_data(struct jbd2_inode *jinode)
211 {
212
213         if (!jinode || !(jinode->i_flags & JI_WRITE_DATA))
214                 return 0;
215
216         trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
217         return jbd2_journal_submit_inode_data_buffers(jinode);
218
219 }
220 EXPORT_SYMBOL(jbd2_submit_inode_data);
221
222 int jbd2_wait_inode_data(journal_t *journal, struct jbd2_inode *jinode)
223 {
224         if (!jinode || !(jinode->i_flags & JI_WAIT_DATA) ||
225                 !jinode->i_vfs_inode || !jinode->i_vfs_inode->i_mapping)
226                 return 0;
227         return filemap_fdatawait_range_keep_errors(
228                 jinode->i_vfs_inode->i_mapping, jinode->i_dirty_start,
229                 jinode->i_dirty_end);
230 }
231 EXPORT_SYMBOL(jbd2_wait_inode_data);
232
233 /*
234  * Submit all the data buffers of inode associated with the transaction to
235  * disk.
236  *
237  * We are in a committing transaction. Therefore no new inode can be added to
238  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
239  * operate on from being released while we write out pages.
240  */
241 static int journal_submit_data_buffers(journal_t *journal,
242                 transaction_t *commit_transaction)
243 {
244         struct jbd2_inode *jinode;
245         int err, ret = 0;
246
247         spin_lock(&journal->j_list_lock);
248         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
249                 if (!(jinode->i_flags & JI_WRITE_DATA))
250                         continue;
251                 jinode->i_flags |= JI_COMMIT_RUNNING;
252                 spin_unlock(&journal->j_list_lock);
253                 /* submit the inode data buffers. */
254                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
255                 if (journal->j_submit_inode_data_buffers) {
256                         err = journal->j_submit_inode_data_buffers(jinode);
257                         if (!ret)
258                                 ret = err;
259                 }
260                 spin_lock(&journal->j_list_lock);
261                 J_ASSERT(jinode->i_transaction == commit_transaction);
262                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
263                 smp_mb();
264                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
265         }
266         spin_unlock(&journal->j_list_lock);
267         return ret;
268 }
269
270 int jbd2_journal_finish_inode_data_buffers(struct jbd2_inode *jinode)
271 {
272         struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
273
274         return filemap_fdatawait_range_keep_errors(mapping,
275                                                    jinode->i_dirty_start,
276                                                    jinode->i_dirty_end);
277 }
278
279 /*
280  * Wait for data submitted for writeout, refile inodes to proper
281  * transaction if needed.
282  *
283  */
284 static int journal_finish_inode_data_buffers(journal_t *journal,
285                 transaction_t *commit_transaction)
286 {
287         struct jbd2_inode *jinode, *next_i;
288         int err, ret = 0;
289
290         /* For locking, see the comment in journal_submit_data_buffers() */
291         spin_lock(&journal->j_list_lock);
292         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
293                 if (!(jinode->i_flags & JI_WAIT_DATA))
294                         continue;
295                 jinode->i_flags |= JI_COMMIT_RUNNING;
296                 spin_unlock(&journal->j_list_lock);
297                 /* wait for the inode data buffers writeout. */
298                 if (journal->j_finish_inode_data_buffers) {
299                         err = journal->j_finish_inode_data_buffers(jinode);
300                         if (!ret)
301                                 ret = err;
302                 }
303                 cond_resched();
304                 spin_lock(&journal->j_list_lock);
305                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
306                 smp_mb();
307                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
308         }
309
310         /* Now refile inode to proper lists */
311         list_for_each_entry_safe(jinode, next_i,
312                                  &commit_transaction->t_inode_list, i_list) {
313                 list_del(&jinode->i_list);
314                 if (jinode->i_next_transaction) {
315                         jinode->i_transaction = jinode->i_next_transaction;
316                         jinode->i_next_transaction = NULL;
317                         list_add(&jinode->i_list,
318                                 &jinode->i_transaction->t_inode_list);
319                 } else {
320                         jinode->i_transaction = NULL;
321                         jinode->i_dirty_start = 0;
322                         jinode->i_dirty_end = 0;
323                 }
324         }
325         spin_unlock(&journal->j_list_lock);
326
327         return ret;
328 }
329
330 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
331 {
332         struct page *page = bh->b_page;
333         char *addr;
334         __u32 checksum;
335
336         addr = kmap_atomic(page);
337         checksum = crc32_be(crc32_sum,
338                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
339         kunmap_atomic(addr);
340
341         return checksum;
342 }
343
344 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
345                                    unsigned long long block)
346 {
347         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
348         if (jbd2_has_feature_64bit(j))
349                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
350 }
351
352 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
353                                     struct buffer_head *bh, __u32 sequence)
354 {
355         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
356         struct page *page = bh->b_page;
357         __u8 *addr;
358         __u32 csum32;
359         __be32 seq;
360
361         if (!jbd2_journal_has_csum_v2or3(j))
362                 return;
363
364         seq = cpu_to_be32(sequence);
365         addr = kmap_atomic(page);
366         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
367         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
368                              bh->b_size);
369         kunmap_atomic(addr);
370
371         if (jbd2_has_feature_csum3(j))
372                 tag3->t_checksum = cpu_to_be32(csum32);
373         else
374                 tag->t_checksum = cpu_to_be16(csum32);
375 }
376 /*
377  * jbd2_journal_commit_transaction
378  *
379  * The primary function for committing a transaction to the log.  This
380  * function is called by the journal thread to begin a complete commit.
381  */
382 void jbd2_journal_commit_transaction(journal_t *journal)
383 {
384         struct transaction_stats_s stats;
385         transaction_t *commit_transaction;
386         struct journal_head *jh;
387         struct buffer_head *descriptor;
388         struct buffer_head **wbuf = journal->j_wbuf;
389         int bufs;
390         int flags;
391         int err;
392         unsigned long long blocknr;
393         ktime_t start_time;
394         u64 commit_time;
395         char *tagp = NULL;
396         journal_block_tag_t *tag = NULL;
397         int space_left = 0;
398         int first_tag = 0;
399         int tag_flag;
400         int i;
401         int tag_bytes = journal_tag_bytes(journal);
402         struct buffer_head *cbh = NULL; /* For transactional checksums */
403         __u32 crc32_sum = ~0;
404         struct blk_plug plug;
405         /* Tail of the journal */
406         unsigned long first_block;
407         tid_t first_tid;
408         int update_tail;
409         int csum_size = 0;
410         LIST_HEAD(io_bufs);
411         LIST_HEAD(log_bufs);
412
413         if (jbd2_journal_has_csum_v2or3(journal))
414                 csum_size = sizeof(struct jbd2_journal_block_tail);
415
416         /*
417          * First job: lock down the current transaction and wait for
418          * all outstanding updates to complete.
419          */
420
421         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
422         if (journal->j_flags & JBD2_FLUSHED) {
423                 jbd2_debug(3, "super block updated\n");
424                 mutex_lock_io(&journal->j_checkpoint_mutex);
425                 /*
426                  * We hold j_checkpoint_mutex so tail cannot change under us.
427                  * We don't need any special data guarantees for writing sb
428                  * since journal is empty and it is ok for write to be
429                  * flushed only with transaction commit.
430                  */
431                 jbd2_journal_update_sb_log_tail(journal,
432                                                 journal->j_tail_sequence,
433                                                 journal->j_tail, 0);
434                 mutex_unlock(&journal->j_checkpoint_mutex);
435         } else {
436                 jbd2_debug(3, "superblock not updated\n");
437         }
438
439         J_ASSERT(journal->j_running_transaction != NULL);
440         J_ASSERT(journal->j_committing_transaction == NULL);
441
442         write_lock(&journal->j_state_lock);
443         journal->j_flags |= JBD2_FULL_COMMIT_ONGOING;
444         while (journal->j_flags & JBD2_FAST_COMMIT_ONGOING) {
445                 DEFINE_WAIT(wait);
446
447                 prepare_to_wait(&journal->j_fc_wait, &wait,
448                                 TASK_UNINTERRUPTIBLE);
449                 write_unlock(&journal->j_state_lock);
450                 schedule();
451                 write_lock(&journal->j_state_lock);
452                 finish_wait(&journal->j_fc_wait, &wait);
453                 /*
454                  * TODO: by blocking fast commits here, we are increasing
455                  * fsync() latency slightly. Strictly speaking, we don't need
456                  * to block fast commits until the transaction enters T_FLUSH
457                  * state. So an optimization is possible where we block new fast
458                  * commits here and wait for existing ones to complete
459                  * just before we enter T_FLUSH. That way, the existing fast
460                  * commits and this full commit can proceed parallely.
461                  */
462         }
463         write_unlock(&journal->j_state_lock);
464
465         commit_transaction = journal->j_running_transaction;
466
467         trace_jbd2_start_commit(journal, commit_transaction);
468         jbd2_debug(1, "JBD2: starting commit of transaction %d\n",
469                         commit_transaction->t_tid);
470
471         write_lock(&journal->j_state_lock);
472         journal->j_fc_off = 0;
473         J_ASSERT(commit_transaction->t_state == T_RUNNING);
474         commit_transaction->t_state = T_LOCKED;
475
476         trace_jbd2_commit_locking(journal, commit_transaction);
477         stats.run.rs_wait = commit_transaction->t_max_wait;
478         stats.run.rs_request_delay = 0;
479         stats.run.rs_locked = jiffies;
480         if (commit_transaction->t_requested)
481                 stats.run.rs_request_delay =
482                         jbd2_time_diff(commit_transaction->t_requested,
483                                        stats.run.rs_locked);
484         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
485                                               stats.run.rs_locked);
486
487         // waits for any t_updates to finish
488         jbd2_journal_wait_updates(journal);
489
490         commit_transaction->t_state = T_SWITCH;
491
492         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
493                         journal->j_max_transaction_buffers);
494
495         /*
496          * First thing we are allowed to do is to discard any remaining
497          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
498          * that there are no such buffers: if a large filesystem
499          * operation like a truncate needs to split itself over multiple
500          * transactions, then it may try to do a jbd2_journal_restart() while
501          * there are still BJ_Reserved buffers outstanding.  These must
502          * be released cleanly from the current transaction.
503          *
504          * In this case, the filesystem must still reserve write access
505          * again before modifying the buffer in the new transaction, but
506          * we do not require it to remember exactly which old buffers it
507          * has reserved.  This is consistent with the existing behaviour
508          * that multiple jbd2_journal_get_write_access() calls to the same
509          * buffer are perfectly permissible.
510          * We use journal->j_state_lock here to serialize processing of
511          * t_reserved_list with eviction of buffers from journal_unmap_buffer().
512          */
513         while (commit_transaction->t_reserved_list) {
514                 jh = commit_transaction->t_reserved_list;
515                 JBUFFER_TRACE(jh, "reserved, unused: refile");
516                 /*
517                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
518                  * leave undo-committed data.
519                  */
520                 if (jh->b_committed_data) {
521                         struct buffer_head *bh = jh2bh(jh);
522
523                         spin_lock(&jh->b_state_lock);
524                         jbd2_free(jh->b_committed_data, bh->b_size);
525                         jh->b_committed_data = NULL;
526                         spin_unlock(&jh->b_state_lock);
527                 }
528                 jbd2_journal_refile_buffer(journal, jh);
529         }
530
531         write_unlock(&journal->j_state_lock);
532         /*
533          * Now try to drop any written-back buffers from the journal's
534          * checkpoint lists.  We do this *before* commit because it potentially
535          * frees some memory
536          */
537         spin_lock(&journal->j_list_lock);
538         __jbd2_journal_clean_checkpoint_list(journal, false);
539         spin_unlock(&journal->j_list_lock);
540
541         jbd2_debug(3, "JBD2: commit phase 1\n");
542
543         /*
544          * Clear revoked flag to reflect there is no revoked buffers
545          * in the next transaction which is going to be started.
546          */
547         jbd2_clear_buffer_revoked_flags(journal);
548
549         /*
550          * Switch to a new revoke table.
551          */
552         jbd2_journal_switch_revoke_table(journal);
553
554         write_lock(&journal->j_state_lock);
555         /*
556          * Reserved credits cannot be claimed anymore, free them
557          */
558         atomic_sub(atomic_read(&journal->j_reserved_credits),
559                    &commit_transaction->t_outstanding_credits);
560
561         trace_jbd2_commit_flushing(journal, commit_transaction);
562         stats.run.rs_flushing = jiffies;
563         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
564                                              stats.run.rs_flushing);
565
566         commit_transaction->t_state = T_FLUSH;
567         journal->j_committing_transaction = commit_transaction;
568         journal->j_running_transaction = NULL;
569         start_time = ktime_get();
570         commit_transaction->t_log_start = journal->j_head;
571         wake_up_all(&journal->j_wait_transaction_locked);
572         write_unlock(&journal->j_state_lock);
573
574         jbd2_debug(3, "JBD2: commit phase 2a\n");
575
576         /*
577          * Now start flushing things to disk, in the order they appear
578          * on the transaction lists.  Data blocks go first.
579          */
580         err = journal_submit_data_buffers(journal, commit_transaction);
581         if (err)
582                 jbd2_journal_abort(journal, err);
583
584         blk_start_plug(&plug);
585         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
586
587         jbd2_debug(3, "JBD2: commit phase 2b\n");
588
589         /*
590          * Way to go: we have now written out all of the data for a
591          * transaction!  Now comes the tricky part: we need to write out
592          * metadata.  Loop over the transaction's entire buffer list:
593          */
594         write_lock(&journal->j_state_lock);
595         commit_transaction->t_state = T_COMMIT;
596         write_unlock(&journal->j_state_lock);
597
598         trace_jbd2_commit_logging(journal, commit_transaction);
599         stats.run.rs_logging = jiffies;
600         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
601                                                stats.run.rs_logging);
602         stats.run.rs_blocks = commit_transaction->t_nr_buffers;
603         stats.run.rs_blocks_logged = 0;
604
605         J_ASSERT(commit_transaction->t_nr_buffers <=
606                  atomic_read(&commit_transaction->t_outstanding_credits));
607
608         err = 0;
609         bufs = 0;
610         descriptor = NULL;
611         while (commit_transaction->t_buffers) {
612
613                 /* Find the next buffer to be journaled... */
614
615                 jh = commit_transaction->t_buffers;
616
617                 /* If we're in abort mode, we just un-journal the buffer and
618                    release it. */
619
620                 if (is_journal_aborted(journal)) {
621                         clear_buffer_jbddirty(jh2bh(jh));
622                         JBUFFER_TRACE(jh, "journal is aborting: refile");
623                         jbd2_buffer_abort_trigger(jh,
624                                                   jh->b_frozen_data ?
625                                                   jh->b_frozen_triggers :
626                                                   jh->b_triggers);
627                         jbd2_journal_refile_buffer(journal, jh);
628                         /* If that was the last one, we need to clean up
629                          * any descriptor buffers which may have been
630                          * already allocated, even if we are now
631                          * aborting. */
632                         if (!commit_transaction->t_buffers)
633                                 goto start_journal_io;
634                         continue;
635                 }
636
637                 /* Make sure we have a descriptor block in which to
638                    record the metadata buffer. */
639
640                 if (!descriptor) {
641                         J_ASSERT (bufs == 0);
642
643                         jbd2_debug(4, "JBD2: get descriptor\n");
644
645                         descriptor = jbd2_journal_get_descriptor_buffer(
646                                                         commit_transaction,
647                                                         JBD2_DESCRIPTOR_BLOCK);
648                         if (!descriptor) {
649                                 jbd2_journal_abort(journal, -EIO);
650                                 continue;
651                         }
652
653                         jbd2_debug(4, "JBD2: got buffer %llu (%p)\n",
654                                 (unsigned long long)descriptor->b_blocknr,
655                                 descriptor->b_data);
656                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
657                         space_left = descriptor->b_size -
658                                                 sizeof(journal_header_t);
659                         first_tag = 1;
660                         set_buffer_jwrite(descriptor);
661                         set_buffer_dirty(descriptor);
662                         wbuf[bufs++] = descriptor;
663
664                         /* Record it so that we can wait for IO
665                            completion later */
666                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
667                         jbd2_file_log_bh(&log_bufs, descriptor);
668                 }
669
670                 /* Where is the buffer to be written? */
671
672                 err = jbd2_journal_next_log_block(journal, &blocknr);
673                 /* If the block mapping failed, just abandon the buffer
674                    and repeat this loop: we'll fall into the
675                    refile-on-abort condition above. */
676                 if (err) {
677                         jbd2_journal_abort(journal, err);
678                         continue;
679                 }
680
681                 /*
682                  * start_this_handle() uses t_outstanding_credits to determine
683                  * the free space in the log.
684                  */
685                 atomic_dec(&commit_transaction->t_outstanding_credits);
686
687                 /* Bump b_count to prevent truncate from stumbling over
688                    the shadowed buffer!  @@@ This can go if we ever get
689                    rid of the shadow pairing of buffers. */
690                 atomic_inc(&jh2bh(jh)->b_count);
691
692                 /*
693                  * Make a temporary IO buffer with which to write it out
694                  * (this will requeue the metadata buffer to BJ_Shadow).
695                  */
696                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
697                 JBUFFER_TRACE(jh, "ph3: write metadata");
698                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
699                                                 jh, &wbuf[bufs], blocknr);
700                 if (flags < 0) {
701                         jbd2_journal_abort(journal, flags);
702                         continue;
703                 }
704                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
705
706                 /* Record the new block's tag in the current descriptor
707                    buffer */
708
709                 tag_flag = 0;
710                 if (flags & 1)
711                         tag_flag |= JBD2_FLAG_ESCAPE;
712                 if (!first_tag)
713                         tag_flag |= JBD2_FLAG_SAME_UUID;
714
715                 tag = (journal_block_tag_t *) tagp;
716                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
717                 tag->t_flags = cpu_to_be16(tag_flag);
718                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
719                                         commit_transaction->t_tid);
720                 tagp += tag_bytes;
721                 space_left -= tag_bytes;
722                 bufs++;
723
724                 if (first_tag) {
725                         memcpy (tagp, journal->j_uuid, 16);
726                         tagp += 16;
727                         space_left -= 16;
728                         first_tag = 0;
729                 }
730
731                 /* If there's no more to do, or if the descriptor is full,
732                    let the IO rip! */
733
734                 if (bufs == journal->j_wbufsize ||
735                     commit_transaction->t_buffers == NULL ||
736                     space_left < tag_bytes + 16 + csum_size) {
737
738                         jbd2_debug(4, "JBD2: Submit %d IOs\n", bufs);
739
740                         /* Write an end-of-descriptor marker before
741                            submitting the IOs.  "tag" still points to
742                            the last tag we set up. */
743
744                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
745 start_journal_io:
746                         if (descriptor)
747                                 jbd2_descriptor_block_csum_set(journal,
748                                                         descriptor);
749
750                         for (i = 0; i < bufs; i++) {
751                                 struct buffer_head *bh = wbuf[i];
752
753                                 /*
754                                  * Compute checksum.
755                                  */
756                                 if (jbd2_has_feature_checksum(journal)) {
757                                         crc32_sum =
758                                             jbd2_checksum_data(crc32_sum, bh);
759                                 }
760
761                                 lock_buffer(bh);
762                                 clear_buffer_dirty(bh);
763                                 set_buffer_uptodate(bh);
764                                 bh->b_end_io = journal_end_buffer_io_sync;
765                                 submit_bh(REQ_OP_WRITE | JBD2_JOURNAL_REQ_FLAGS,
766                                           bh);
767                         }
768                         cond_resched();
769
770                         /* Force a new descriptor to be generated next
771                            time round the loop. */
772                         descriptor = NULL;
773                         bufs = 0;
774                 }
775         }
776
777         err = journal_finish_inode_data_buffers(journal, commit_transaction);
778         if (err) {
779                 printk(KERN_WARNING
780                         "JBD2: Detected IO errors while flushing file data "
781                        "on %s\n", journal->j_devname);
782                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
783                         jbd2_journal_abort(journal, err);
784                 err = 0;
785         }
786
787         /*
788          * Get current oldest transaction in the log before we issue flush
789          * to the filesystem device. After the flush we can be sure that
790          * blocks of all older transactions are checkpointed to persistent
791          * storage and we will be safe to update journal start in the
792          * superblock with the numbers we get here.
793          */
794         update_tail =
795                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
796
797         write_lock(&journal->j_state_lock);
798         if (update_tail) {
799                 long freed = first_block - journal->j_tail;
800
801                 if (first_block < journal->j_tail)
802                         freed += journal->j_last - journal->j_first;
803                 /* Update tail only if we free significant amount of space */
804                 if (freed < jbd2_journal_get_max_txn_bufs(journal))
805                         update_tail = 0;
806         }
807         J_ASSERT(commit_transaction->t_state == T_COMMIT);
808         commit_transaction->t_state = T_COMMIT_DFLUSH;
809         write_unlock(&journal->j_state_lock);
810
811         /*
812          * If the journal is not located on the file system device,
813          * then we must flush the file system device before we issue
814          * the commit record
815          */
816         if (commit_transaction->t_need_data_flush &&
817             (journal->j_fs_dev != journal->j_dev) &&
818             (journal->j_flags & JBD2_BARRIER))
819                 blkdev_issue_flush(journal->j_fs_dev);
820
821         /* Done it all: now write the commit record asynchronously. */
822         if (jbd2_has_feature_async_commit(journal)) {
823                 err = journal_submit_commit_record(journal, commit_transaction,
824                                                  &cbh, crc32_sum);
825                 if (err)
826                         jbd2_journal_abort(journal, err);
827         }
828
829         blk_finish_plug(&plug);
830
831         /* Lo and behold: we have just managed to send a transaction to
832            the log.  Before we can commit it, wait for the IO so far to
833            complete.  Control buffers being written are on the
834            transaction's t_log_list queue, and metadata buffers are on
835            the io_bufs list.
836
837            Wait for the buffers in reverse order.  That way we are
838            less likely to be woken up until all IOs have completed, and
839            so we incur less scheduling load.
840         */
841
842         jbd2_debug(3, "JBD2: commit phase 3\n");
843
844         while (!list_empty(&io_bufs)) {
845                 struct buffer_head *bh = list_entry(io_bufs.prev,
846                                                     struct buffer_head,
847                                                     b_assoc_buffers);
848
849                 wait_on_buffer(bh);
850                 cond_resched();
851
852                 if (unlikely(!buffer_uptodate(bh)))
853                         err = -EIO;
854                 jbd2_unfile_log_bh(bh);
855                 stats.run.rs_blocks_logged++;
856
857                 /*
858                  * The list contains temporary buffer heads created by
859                  * jbd2_journal_write_metadata_buffer().
860                  */
861                 BUFFER_TRACE(bh, "dumping temporary bh");
862                 __brelse(bh);
863                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
864                 free_buffer_head(bh);
865
866                 /* We also have to refile the corresponding shadowed buffer */
867                 jh = commit_transaction->t_shadow_list->b_tprev;
868                 bh = jh2bh(jh);
869                 clear_buffer_jwrite(bh);
870                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
871                 J_ASSERT_BH(bh, !buffer_shadow(bh));
872
873                 /* The metadata is now released for reuse, but we need
874                    to remember it against this transaction so that when
875                    we finally commit, we can do any checkpointing
876                    required. */
877                 JBUFFER_TRACE(jh, "file as BJ_Forget");
878                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
879                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
880                 __brelse(bh);
881         }
882
883         J_ASSERT (commit_transaction->t_shadow_list == NULL);
884
885         jbd2_debug(3, "JBD2: commit phase 4\n");
886
887         /* Here we wait for the revoke record and descriptor record buffers */
888         while (!list_empty(&log_bufs)) {
889                 struct buffer_head *bh;
890
891                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
892                 wait_on_buffer(bh);
893                 cond_resched();
894
895                 if (unlikely(!buffer_uptodate(bh)))
896                         err = -EIO;
897
898                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
899                 clear_buffer_jwrite(bh);
900                 jbd2_unfile_log_bh(bh);
901                 stats.run.rs_blocks_logged++;
902                 __brelse(bh);           /* One for getblk */
903                 /* AKPM: bforget here */
904         }
905
906         if (err)
907                 jbd2_journal_abort(journal, err);
908
909         jbd2_debug(3, "JBD2: commit phase 5\n");
910         write_lock(&journal->j_state_lock);
911         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
912         commit_transaction->t_state = T_COMMIT_JFLUSH;
913         write_unlock(&journal->j_state_lock);
914
915         if (!jbd2_has_feature_async_commit(journal)) {
916                 err = journal_submit_commit_record(journal, commit_transaction,
917                                                 &cbh, crc32_sum);
918                 if (err)
919                         jbd2_journal_abort(journal, err);
920         }
921         if (cbh)
922                 err = journal_wait_on_commit_record(journal, cbh);
923         stats.run.rs_blocks_logged++;
924         if (jbd2_has_feature_async_commit(journal) &&
925             journal->j_flags & JBD2_BARRIER) {
926                 blkdev_issue_flush(journal->j_dev);
927         }
928
929         if (err)
930                 jbd2_journal_abort(journal, err);
931
932         WARN_ON_ONCE(
933                 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
934
935         /*
936          * Now disk caches for filesystem device are flushed so we are safe to
937          * erase checkpointed transactions from the log by updating journal
938          * superblock.
939          */
940         if (update_tail)
941                 jbd2_update_log_tail(journal, first_tid, first_block);
942
943         /* End of a transaction!  Finally, we can do checkpoint
944            processing: any buffers committed as a result of this
945            transaction can be removed from any checkpoint list it was on
946            before. */
947
948         jbd2_debug(3, "JBD2: commit phase 6\n");
949
950         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
951         J_ASSERT(commit_transaction->t_buffers == NULL);
952         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
953         J_ASSERT(commit_transaction->t_shadow_list == NULL);
954
955 restart_loop:
956         /*
957          * As there are other places (journal_unmap_buffer()) adding buffers
958          * to this list we have to be careful and hold the j_list_lock.
959          */
960         spin_lock(&journal->j_list_lock);
961         while (commit_transaction->t_forget) {
962                 transaction_t *cp_transaction;
963                 struct buffer_head *bh;
964                 int try_to_free = 0;
965                 bool drop_ref;
966
967                 jh = commit_transaction->t_forget;
968                 spin_unlock(&journal->j_list_lock);
969                 bh = jh2bh(jh);
970                 /*
971                  * Get a reference so that bh cannot be freed before we are
972                  * done with it.
973                  */
974                 get_bh(bh);
975                 spin_lock(&jh->b_state_lock);
976                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
977
978                 /*
979                  * If there is undo-protected committed data against
980                  * this buffer, then we can remove it now.  If it is a
981                  * buffer needing such protection, the old frozen_data
982                  * field now points to a committed version of the
983                  * buffer, so rotate that field to the new committed
984                  * data.
985                  *
986                  * Otherwise, we can just throw away the frozen data now.
987                  *
988                  * We also know that the frozen data has already fired
989                  * its triggers if they exist, so we can clear that too.
990                  */
991                 if (jh->b_committed_data) {
992                         jbd2_free(jh->b_committed_data, bh->b_size);
993                         jh->b_committed_data = NULL;
994                         if (jh->b_frozen_data) {
995                                 jh->b_committed_data = jh->b_frozen_data;
996                                 jh->b_frozen_data = NULL;
997                                 jh->b_frozen_triggers = NULL;
998                         }
999                 } else if (jh->b_frozen_data) {
1000                         jbd2_free(jh->b_frozen_data, bh->b_size);
1001                         jh->b_frozen_data = NULL;
1002                         jh->b_frozen_triggers = NULL;
1003                 }
1004
1005                 spin_lock(&journal->j_list_lock);
1006                 cp_transaction = jh->b_cp_transaction;
1007                 if (cp_transaction) {
1008                         JBUFFER_TRACE(jh, "remove from old cp transaction");
1009                         cp_transaction->t_chp_stats.cs_dropped++;
1010                         __jbd2_journal_remove_checkpoint(jh);
1011                 }
1012
1013                 /* Only re-checkpoint the buffer_head if it is marked
1014                  * dirty.  If the buffer was added to the BJ_Forget list
1015                  * by jbd2_journal_forget, it may no longer be dirty and
1016                  * there's no point in keeping a checkpoint record for
1017                  * it. */
1018
1019                 /*
1020                  * A buffer which has been freed while still being journaled
1021                  * by a previous transaction, refile the buffer to BJ_Forget of
1022                  * the running transaction. If the just committed transaction
1023                  * contains "add to orphan" operation, we can completely
1024                  * invalidate the buffer now. We are rather through in that
1025                  * since the buffer may be still accessible when blocksize <
1026                  * pagesize and it is attached to the last partial page.
1027                  */
1028                 if (buffer_freed(bh) && !jh->b_next_transaction) {
1029                         struct address_space *mapping;
1030
1031                         clear_buffer_freed(bh);
1032                         clear_buffer_jbddirty(bh);
1033
1034                         /*
1035                          * Block device buffers need to stay mapped all the
1036                          * time, so it is enough to clear buffer_jbddirty and
1037                          * buffer_freed bits. For the file mapping buffers (i.e.
1038                          * journalled data) we need to unmap buffer and clear
1039                          * more bits. We also need to be careful about the check
1040                          * because the data page mapping can get cleared under
1041                          * our hands. Note that if mapping == NULL, we don't
1042                          * need to make buffer unmapped because the page is
1043                          * already detached from the mapping and buffers cannot
1044                          * get reused.
1045                          */
1046                         mapping = READ_ONCE(bh->b_page->mapping);
1047                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1048                                 clear_buffer_mapped(bh);
1049                                 clear_buffer_new(bh);
1050                                 clear_buffer_req(bh);
1051                                 bh->b_bdev = NULL;
1052                         }
1053                 }
1054
1055                 if (buffer_jbddirty(bh)) {
1056                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1057                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1058                         if (is_journal_aborted(journal))
1059                                 clear_buffer_jbddirty(bh);
1060                 } else {
1061                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1062                         /*
1063                          * The buffer on BJ_Forget list and not jbddirty means
1064                          * it has been freed by this transaction and hence it
1065                          * could not have been reallocated until this
1066                          * transaction has committed. *BUT* it could be
1067                          * reallocated once we have written all the data to
1068                          * disk and before we process the buffer on BJ_Forget
1069                          * list.
1070                          */
1071                         if (!jh->b_next_transaction)
1072                                 try_to_free = 1;
1073                 }
1074                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1075                 drop_ref = __jbd2_journal_refile_buffer(jh);
1076                 spin_unlock(&jh->b_state_lock);
1077                 if (drop_ref)
1078                         jbd2_journal_put_journal_head(jh);
1079                 if (try_to_free)
1080                         release_buffer_page(bh);        /* Drops bh reference */
1081                 else
1082                         __brelse(bh);
1083                 cond_resched_lock(&journal->j_list_lock);
1084         }
1085         spin_unlock(&journal->j_list_lock);
1086         /*
1087          * This is a bit sleazy.  We use j_list_lock to protect transition
1088          * of a transaction into T_FINISHED state and calling
1089          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1090          * other checkpointing code processing the transaction...
1091          */
1092         write_lock(&journal->j_state_lock);
1093         spin_lock(&journal->j_list_lock);
1094         /*
1095          * Now recheck if some buffers did not get attached to the transaction
1096          * while the lock was dropped...
1097          */
1098         if (commit_transaction->t_forget) {
1099                 spin_unlock(&journal->j_list_lock);
1100                 write_unlock(&journal->j_state_lock);
1101                 goto restart_loop;
1102         }
1103
1104         /* Add the transaction to the checkpoint list
1105          * __journal_remove_checkpoint() can not destroy transaction
1106          * under us because it is not marked as T_FINISHED yet */
1107         if (journal->j_checkpoint_transactions == NULL) {
1108                 journal->j_checkpoint_transactions = commit_transaction;
1109                 commit_transaction->t_cpnext = commit_transaction;
1110                 commit_transaction->t_cpprev = commit_transaction;
1111         } else {
1112                 commit_transaction->t_cpnext =
1113                         journal->j_checkpoint_transactions;
1114                 commit_transaction->t_cpprev =
1115                         commit_transaction->t_cpnext->t_cpprev;
1116                 commit_transaction->t_cpnext->t_cpprev =
1117                         commit_transaction;
1118                 commit_transaction->t_cpprev->t_cpnext =
1119                                 commit_transaction;
1120         }
1121         spin_unlock(&journal->j_list_lock);
1122
1123         /* Done with this transaction! */
1124
1125         jbd2_debug(3, "JBD2: commit phase 7\n");
1126
1127         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1128
1129         commit_transaction->t_start = jiffies;
1130         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1131                                               commit_transaction->t_start);
1132
1133         /*
1134          * File the transaction statistics
1135          */
1136         stats.ts_tid = commit_transaction->t_tid;
1137         stats.run.rs_handle_count =
1138                 atomic_read(&commit_transaction->t_handle_count);
1139         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1140                              commit_transaction->t_tid, &stats.run);
1141         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1142
1143         commit_transaction->t_state = T_COMMIT_CALLBACK;
1144         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1145         journal->j_commit_sequence = commit_transaction->t_tid;
1146         journal->j_committing_transaction = NULL;
1147         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1148
1149         /*
1150          * weight the commit time higher than the average time so we don't
1151          * react too strongly to vast changes in the commit time
1152          */
1153         if (likely(journal->j_average_commit_time))
1154                 journal->j_average_commit_time = (commit_time +
1155                                 journal->j_average_commit_time*3) / 4;
1156         else
1157                 journal->j_average_commit_time = commit_time;
1158
1159         write_unlock(&journal->j_state_lock);
1160
1161         if (journal->j_commit_callback)
1162                 journal->j_commit_callback(journal, commit_transaction);
1163         if (journal->j_fc_cleanup_callback)
1164                 journal->j_fc_cleanup_callback(journal, 1, commit_transaction->t_tid);
1165
1166         trace_jbd2_end_commit(journal, commit_transaction);
1167         jbd2_debug(1, "JBD2: commit %d complete, head %d\n",
1168                   journal->j_commit_sequence, journal->j_tail_sequence);
1169
1170         write_lock(&journal->j_state_lock);
1171         journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1172         journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1173         spin_lock(&journal->j_list_lock);
1174         commit_transaction->t_state = T_FINISHED;
1175         /* Check if the transaction can be dropped now that we are finished */
1176         if (commit_transaction->t_checkpoint_list == NULL) {
1177                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1178                 jbd2_journal_free_transaction(commit_transaction);
1179         }
1180         spin_unlock(&journal->j_list_lock);
1181         write_unlock(&journal->j_state_lock);
1182         wake_up(&journal->j_wait_done_commit);
1183         wake_up(&journal->j_fc_wait);
1184
1185         /*
1186          * Calculate overall stats
1187          */
1188         spin_lock(&journal->j_history_lock);
1189         journal->j_stats.ts_tid++;
1190         journal->j_stats.ts_requested += stats.ts_requested;
1191         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1192         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1193         journal->j_stats.run.rs_running += stats.run.rs_running;
1194         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1195         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1196         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1197         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1198         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1199         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1200         spin_unlock(&journal->j_history_lock);
1201 }