Merge tag 'loongarch-6.2' of git://git.kernel.org/pub/scm/linux/kernel/git/chenhuacai...
[platform/kernel/linux-starfive.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 | REQ_SYNC;
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(journal_t *journal, struct jbd2_inode *jinode)
211 {
212         if (!jinode || !(jinode->i_flags & JI_WRITE_DATA))
213                 return 0;
214
215         trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
216         return journal->j_submit_inode_data_buffers(jinode);
217
218 }
219 EXPORT_SYMBOL(jbd2_submit_inode_data);
220
221 int jbd2_wait_inode_data(journal_t *journal, struct jbd2_inode *jinode)
222 {
223         if (!jinode || !(jinode->i_flags & JI_WAIT_DATA) ||
224                 !jinode->i_vfs_inode || !jinode->i_vfs_inode->i_mapping)
225                 return 0;
226         return filemap_fdatawait_range_keep_errors(
227                 jinode->i_vfs_inode->i_mapping, jinode->i_dirty_start,
228                 jinode->i_dirty_end);
229 }
230 EXPORT_SYMBOL(jbd2_wait_inode_data);
231
232 /*
233  * Submit all the data buffers of inode associated with the transaction to
234  * disk.
235  *
236  * We are in a committing transaction. Therefore no new inode can be added to
237  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
238  * operate on from being released while we write out pages.
239  */
240 static int journal_submit_data_buffers(journal_t *journal,
241                 transaction_t *commit_transaction)
242 {
243         struct jbd2_inode *jinode;
244         int err, ret = 0;
245
246         spin_lock(&journal->j_list_lock);
247         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
248                 if (!(jinode->i_flags & JI_WRITE_DATA))
249                         continue;
250                 jinode->i_flags |= JI_COMMIT_RUNNING;
251                 spin_unlock(&journal->j_list_lock);
252                 /* submit the inode data buffers. */
253                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
254                 if (journal->j_submit_inode_data_buffers) {
255                         err = journal->j_submit_inode_data_buffers(jinode);
256                         if (!ret)
257                                 ret = err;
258                 }
259                 spin_lock(&journal->j_list_lock);
260                 J_ASSERT(jinode->i_transaction == commit_transaction);
261                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
262                 smp_mb();
263                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
264         }
265         spin_unlock(&journal->j_list_lock);
266         return ret;
267 }
268
269 int jbd2_journal_finish_inode_data_buffers(struct jbd2_inode *jinode)
270 {
271         struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
272
273         return filemap_fdatawait_range_keep_errors(mapping,
274                                                    jinode->i_dirty_start,
275                                                    jinode->i_dirty_end);
276 }
277
278 /*
279  * Wait for data submitted for writeout, refile inodes to proper
280  * transaction if needed.
281  *
282  */
283 static int journal_finish_inode_data_buffers(journal_t *journal,
284                 transaction_t *commit_transaction)
285 {
286         struct jbd2_inode *jinode, *next_i;
287         int err, ret = 0;
288
289         /* For locking, see the comment in journal_submit_data_buffers() */
290         spin_lock(&journal->j_list_lock);
291         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
292                 if (!(jinode->i_flags & JI_WAIT_DATA))
293                         continue;
294                 jinode->i_flags |= JI_COMMIT_RUNNING;
295                 spin_unlock(&journal->j_list_lock);
296                 /* wait for the inode data buffers writeout. */
297                 if (journal->j_finish_inode_data_buffers) {
298                         err = journal->j_finish_inode_data_buffers(jinode);
299                         if (!ret)
300                                 ret = err;
301                 }
302                 spin_lock(&journal->j_list_lock);
303                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
304                 smp_mb();
305                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
306         }
307
308         /* Now refile inode to proper lists */
309         list_for_each_entry_safe(jinode, next_i,
310                                  &commit_transaction->t_inode_list, i_list) {
311                 list_del(&jinode->i_list);
312                 if (jinode->i_next_transaction) {
313                         jinode->i_transaction = jinode->i_next_transaction;
314                         jinode->i_next_transaction = NULL;
315                         list_add(&jinode->i_list,
316                                 &jinode->i_transaction->t_inode_list);
317                 } else {
318                         jinode->i_transaction = NULL;
319                         jinode->i_dirty_start = 0;
320                         jinode->i_dirty_end = 0;
321                 }
322         }
323         spin_unlock(&journal->j_list_lock);
324
325         return ret;
326 }
327
328 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
329 {
330         struct page *page = bh->b_page;
331         char *addr;
332         __u32 checksum;
333
334         addr = kmap_atomic(page);
335         checksum = crc32_be(crc32_sum,
336                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
337         kunmap_atomic(addr);
338
339         return checksum;
340 }
341
342 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
343                                    unsigned long long block)
344 {
345         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
346         if (jbd2_has_feature_64bit(j))
347                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
348 }
349
350 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
351                                     struct buffer_head *bh, __u32 sequence)
352 {
353         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
354         struct page *page = bh->b_page;
355         __u8 *addr;
356         __u32 csum32;
357         __be32 seq;
358
359         if (!jbd2_journal_has_csum_v2or3(j))
360                 return;
361
362         seq = cpu_to_be32(sequence);
363         addr = kmap_atomic(page);
364         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
365         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
366                              bh->b_size);
367         kunmap_atomic(addr);
368
369         if (jbd2_has_feature_csum3(j))
370                 tag3->t_checksum = cpu_to_be32(csum32);
371         else
372                 tag->t_checksum = cpu_to_be16(csum32);
373 }
374 /*
375  * jbd2_journal_commit_transaction
376  *
377  * The primary function for committing a transaction to the log.  This
378  * function is called by the journal thread to begin a complete commit.
379  */
380 void jbd2_journal_commit_transaction(journal_t *journal)
381 {
382         struct transaction_stats_s stats;
383         transaction_t *commit_transaction;
384         struct journal_head *jh;
385         struct buffer_head *descriptor;
386         struct buffer_head **wbuf = journal->j_wbuf;
387         int bufs;
388         int flags;
389         int err;
390         unsigned long long blocknr;
391         ktime_t start_time;
392         u64 commit_time;
393         char *tagp = NULL;
394         journal_block_tag_t *tag = NULL;
395         int space_left = 0;
396         int first_tag = 0;
397         int tag_flag;
398         int i;
399         int tag_bytes = journal_tag_bytes(journal);
400         struct buffer_head *cbh = NULL; /* For transactional checksums */
401         __u32 crc32_sum = ~0;
402         struct blk_plug plug;
403         /* Tail of the journal */
404         unsigned long first_block;
405         tid_t first_tid;
406         int update_tail;
407         int csum_size = 0;
408         LIST_HEAD(io_bufs);
409         LIST_HEAD(log_bufs);
410
411         if (jbd2_journal_has_csum_v2or3(journal))
412                 csum_size = sizeof(struct jbd2_journal_block_tail);
413
414         /*
415          * First job: lock down the current transaction and wait for
416          * all outstanding updates to complete.
417          */
418
419         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
420         if (journal->j_flags & JBD2_FLUSHED) {
421                 jbd2_debug(3, "super block updated\n");
422                 mutex_lock_io(&journal->j_checkpoint_mutex);
423                 /*
424                  * We hold j_checkpoint_mutex so tail cannot change under us.
425                  * We don't need any special data guarantees for writing sb
426                  * since journal is empty and it is ok for write to be
427                  * flushed only with transaction commit.
428                  */
429                 jbd2_journal_update_sb_log_tail(journal,
430                                                 journal->j_tail_sequence,
431                                                 journal->j_tail,
432                                                 REQ_SYNC);
433                 mutex_unlock(&journal->j_checkpoint_mutex);
434         } else {
435                 jbd2_debug(3, "superblock not updated\n");
436         }
437
438         J_ASSERT(journal->j_running_transaction != NULL);
439         J_ASSERT(journal->j_committing_transaction == NULL);
440
441         write_lock(&journal->j_state_lock);
442         journal->j_flags |= JBD2_FULL_COMMIT_ONGOING;
443         while (journal->j_flags & JBD2_FAST_COMMIT_ONGOING) {
444                 DEFINE_WAIT(wait);
445
446                 prepare_to_wait(&journal->j_fc_wait, &wait,
447                                 TASK_UNINTERRUPTIBLE);
448                 write_unlock(&journal->j_state_lock);
449                 schedule();
450                 write_lock(&journal->j_state_lock);
451                 finish_wait(&journal->j_fc_wait, &wait);
452                 /*
453                  * TODO: by blocking fast commits here, we are increasing
454                  * fsync() latency slightly. Strictly speaking, we don't need
455                  * to block fast commits until the transaction enters T_FLUSH
456                  * state. So an optimization is possible where we block new fast
457                  * commits here and wait for existing ones to complete
458                  * just before we enter T_FLUSH. That way, the existing fast
459                  * commits and this full commit can proceed parallely.
460                  */
461         }
462         write_unlock(&journal->j_state_lock);
463
464         commit_transaction = journal->j_running_transaction;
465
466         trace_jbd2_start_commit(journal, commit_transaction);
467         jbd2_debug(1, "JBD2: starting commit of transaction %d\n",
468                         commit_transaction->t_tid);
469
470         write_lock(&journal->j_state_lock);
471         journal->j_fc_off = 0;
472         J_ASSERT(commit_transaction->t_state == T_RUNNING);
473         commit_transaction->t_state = T_LOCKED;
474
475         trace_jbd2_commit_locking(journal, commit_transaction);
476         stats.run.rs_wait = commit_transaction->t_max_wait;
477         stats.run.rs_request_delay = 0;
478         stats.run.rs_locked = jiffies;
479         if (commit_transaction->t_requested)
480                 stats.run.rs_request_delay =
481                         jbd2_time_diff(commit_transaction->t_requested,
482                                        stats.run.rs_locked);
483         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
484                                               stats.run.rs_locked);
485
486         // waits for any t_updates to finish
487         jbd2_journal_wait_updates(journal);
488
489         commit_transaction->t_state = T_SWITCH;
490
491         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
492                         journal->j_max_transaction_buffers);
493
494         /*
495          * First thing we are allowed to do is to discard any remaining
496          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
497          * that there are no such buffers: if a large filesystem
498          * operation like a truncate needs to split itself over multiple
499          * transactions, then it may try to do a jbd2_journal_restart() while
500          * there are still BJ_Reserved buffers outstanding.  These must
501          * be released cleanly from the current transaction.
502          *
503          * In this case, the filesystem must still reserve write access
504          * again before modifying the buffer in the new transaction, but
505          * we do not require it to remember exactly which old buffers it
506          * has reserved.  This is consistent with the existing behaviour
507          * that multiple jbd2_journal_get_write_access() calls to the same
508          * buffer are perfectly permissible.
509          * We use journal->j_state_lock here to serialize processing of
510          * t_reserved_list with eviction of buffers from journal_unmap_buffer().
511          */
512         while (commit_transaction->t_reserved_list) {
513                 jh = commit_transaction->t_reserved_list;
514                 JBUFFER_TRACE(jh, "reserved, unused: refile");
515                 /*
516                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
517                  * leave undo-committed data.
518                  */
519                 if (jh->b_committed_data) {
520                         struct buffer_head *bh = jh2bh(jh);
521
522                         spin_lock(&jh->b_state_lock);
523                         jbd2_free(jh->b_committed_data, bh->b_size);
524                         jh->b_committed_data = NULL;
525                         spin_unlock(&jh->b_state_lock);
526                 }
527                 jbd2_journal_refile_buffer(journal, jh);
528         }
529
530         write_unlock(&journal->j_state_lock);
531         /*
532          * Now try to drop any written-back buffers from the journal's
533          * checkpoint lists.  We do this *before* commit because it potentially
534          * frees some memory
535          */
536         spin_lock(&journal->j_list_lock);
537         __jbd2_journal_clean_checkpoint_list(journal, false);
538         spin_unlock(&journal->j_list_lock);
539
540         jbd2_debug(3, "JBD2: commit phase 1\n");
541
542         /*
543          * Clear revoked flag to reflect there is no revoked buffers
544          * in the next transaction which is going to be started.
545          */
546         jbd2_clear_buffer_revoked_flags(journal);
547
548         /*
549          * Switch to a new revoke table.
550          */
551         jbd2_journal_switch_revoke_table(journal);
552
553         write_lock(&journal->j_state_lock);
554         /*
555          * Reserved credits cannot be claimed anymore, free them
556          */
557         atomic_sub(atomic_read(&journal->j_reserved_credits),
558                    &commit_transaction->t_outstanding_credits);
559
560         trace_jbd2_commit_flushing(journal, commit_transaction);
561         stats.run.rs_flushing = jiffies;
562         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
563                                              stats.run.rs_flushing);
564
565         commit_transaction->t_state = T_FLUSH;
566         journal->j_committing_transaction = commit_transaction;
567         journal->j_running_transaction = NULL;
568         start_time = ktime_get();
569         commit_transaction->t_log_start = journal->j_head;
570         wake_up_all(&journal->j_wait_transaction_locked);
571         write_unlock(&journal->j_state_lock);
572
573         jbd2_debug(3, "JBD2: commit phase 2a\n");
574
575         /*
576          * Now start flushing things to disk, in the order they appear
577          * on the transaction lists.  Data blocks go first.
578          */
579         err = journal_submit_data_buffers(journal, commit_transaction);
580         if (err)
581                 jbd2_journal_abort(journal, err);
582
583         blk_start_plug(&plug);
584         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
585
586         jbd2_debug(3, "JBD2: commit phase 2b\n");
587
588         /*
589          * Way to go: we have now written out all of the data for a
590          * transaction!  Now comes the tricky part: we need to write out
591          * metadata.  Loop over the transaction's entire buffer list:
592          */
593         write_lock(&journal->j_state_lock);
594         commit_transaction->t_state = T_COMMIT;
595         write_unlock(&journal->j_state_lock);
596
597         trace_jbd2_commit_logging(journal, commit_transaction);
598         stats.run.rs_logging = jiffies;
599         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
600                                                stats.run.rs_logging);
601         stats.run.rs_blocks = commit_transaction->t_nr_buffers;
602         stats.run.rs_blocks_logged = 0;
603
604         J_ASSERT(commit_transaction->t_nr_buffers <=
605                  atomic_read(&commit_transaction->t_outstanding_credits));
606
607         err = 0;
608         bufs = 0;
609         descriptor = NULL;
610         while (commit_transaction->t_buffers) {
611
612                 /* Find the next buffer to be journaled... */
613
614                 jh = commit_transaction->t_buffers;
615
616                 /* If we're in abort mode, we just un-journal the buffer and
617                    release it. */
618
619                 if (is_journal_aborted(journal)) {
620                         clear_buffer_jbddirty(jh2bh(jh));
621                         JBUFFER_TRACE(jh, "journal is aborting: refile");
622                         jbd2_buffer_abort_trigger(jh,
623                                                   jh->b_frozen_data ?
624                                                   jh->b_frozen_triggers :
625                                                   jh->b_triggers);
626                         jbd2_journal_refile_buffer(journal, jh);
627                         /* If that was the last one, we need to clean up
628                          * any descriptor buffers which may have been
629                          * already allocated, even if we are now
630                          * aborting. */
631                         if (!commit_transaction->t_buffers)
632                                 goto start_journal_io;
633                         continue;
634                 }
635
636                 /* Make sure we have a descriptor block in which to
637                    record the metadata buffer. */
638
639                 if (!descriptor) {
640                         J_ASSERT (bufs == 0);
641
642                         jbd2_debug(4, "JBD2: get descriptor\n");
643
644                         descriptor = jbd2_journal_get_descriptor_buffer(
645                                                         commit_transaction,
646                                                         JBD2_DESCRIPTOR_BLOCK);
647                         if (!descriptor) {
648                                 jbd2_journal_abort(journal, -EIO);
649                                 continue;
650                         }
651
652                         jbd2_debug(4, "JBD2: got buffer %llu (%p)\n",
653                                 (unsigned long long)descriptor->b_blocknr,
654                                 descriptor->b_data);
655                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
656                         space_left = descriptor->b_size -
657                                                 sizeof(journal_header_t);
658                         first_tag = 1;
659                         set_buffer_jwrite(descriptor);
660                         set_buffer_dirty(descriptor);
661                         wbuf[bufs++] = descriptor;
662
663                         /* Record it so that we can wait for IO
664                            completion later */
665                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
666                         jbd2_file_log_bh(&log_bufs, descriptor);
667                 }
668
669                 /* Where is the buffer to be written? */
670
671                 err = jbd2_journal_next_log_block(journal, &blocknr);
672                 /* If the block mapping failed, just abandon the buffer
673                    and repeat this loop: we'll fall into the
674                    refile-on-abort condition above. */
675                 if (err) {
676                         jbd2_journal_abort(journal, err);
677                         continue;
678                 }
679
680                 /*
681                  * start_this_handle() uses t_outstanding_credits to determine
682                  * the free space in the log.
683                  */
684                 atomic_dec(&commit_transaction->t_outstanding_credits);
685
686                 /* Bump b_count to prevent truncate from stumbling over
687                    the shadowed buffer!  @@@ This can go if we ever get
688                    rid of the shadow pairing of buffers. */
689                 atomic_inc(&jh2bh(jh)->b_count);
690
691                 /*
692                  * Make a temporary IO buffer with which to write it out
693                  * (this will requeue the metadata buffer to BJ_Shadow).
694                  */
695                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
696                 JBUFFER_TRACE(jh, "ph3: write metadata");
697                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
698                                                 jh, &wbuf[bufs], blocknr);
699                 if (flags < 0) {
700                         jbd2_journal_abort(journal, flags);
701                         continue;
702                 }
703                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
704
705                 /* Record the new block's tag in the current descriptor
706                    buffer */
707
708                 tag_flag = 0;
709                 if (flags & 1)
710                         tag_flag |= JBD2_FLAG_ESCAPE;
711                 if (!first_tag)
712                         tag_flag |= JBD2_FLAG_SAME_UUID;
713
714                 tag = (journal_block_tag_t *) tagp;
715                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
716                 tag->t_flags = cpu_to_be16(tag_flag);
717                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
718                                         commit_transaction->t_tid);
719                 tagp += tag_bytes;
720                 space_left -= tag_bytes;
721                 bufs++;
722
723                 if (first_tag) {
724                         memcpy (tagp, journal->j_uuid, 16);
725                         tagp += 16;
726                         space_left -= 16;
727                         first_tag = 0;
728                 }
729
730                 /* If there's no more to do, or if the descriptor is full,
731                    let the IO rip! */
732
733                 if (bufs == journal->j_wbufsize ||
734                     commit_transaction->t_buffers == NULL ||
735                     space_left < tag_bytes + 16 + csum_size) {
736
737                         jbd2_debug(4, "JBD2: Submit %d IOs\n", bufs);
738
739                         /* Write an end-of-descriptor marker before
740                            submitting the IOs.  "tag" still points to
741                            the last tag we set up. */
742
743                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
744 start_journal_io:
745                         if (descriptor)
746                                 jbd2_descriptor_block_csum_set(journal,
747                                                         descriptor);
748
749                         for (i = 0; i < bufs; i++) {
750                                 struct buffer_head *bh = wbuf[i];
751                                 /*
752                                  * Compute checksum.
753                                  */
754                                 if (jbd2_has_feature_checksum(journal)) {
755                                         crc32_sum =
756                                             jbd2_checksum_data(crc32_sum, bh);
757                                 }
758
759                                 lock_buffer(bh);
760                                 clear_buffer_dirty(bh);
761                                 set_buffer_uptodate(bh);
762                                 bh->b_end_io = journal_end_buffer_io_sync;
763                                 submit_bh(REQ_OP_WRITE | REQ_SYNC, bh);
764                         }
765                         cond_resched();
766
767                         /* Force a new descriptor to be generated next
768                            time round the loop. */
769                         descriptor = NULL;
770                         bufs = 0;
771                 }
772         }
773
774         err = journal_finish_inode_data_buffers(journal, commit_transaction);
775         if (err) {
776                 printk(KERN_WARNING
777                         "JBD2: Detected IO errors while flushing file data "
778                        "on %s\n", journal->j_devname);
779                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
780                         jbd2_journal_abort(journal, err);
781                 err = 0;
782         }
783
784         /*
785          * Get current oldest transaction in the log before we issue flush
786          * to the filesystem device. After the flush we can be sure that
787          * blocks of all older transactions are checkpointed to persistent
788          * storage and we will be safe to update journal start in the
789          * superblock with the numbers we get here.
790          */
791         update_tail =
792                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
793
794         write_lock(&journal->j_state_lock);
795         if (update_tail) {
796                 long freed = first_block - journal->j_tail;
797
798                 if (first_block < journal->j_tail)
799                         freed += journal->j_last - journal->j_first;
800                 /* Update tail only if we free significant amount of space */
801                 if (freed < jbd2_journal_get_max_txn_bufs(journal))
802                         update_tail = 0;
803         }
804         J_ASSERT(commit_transaction->t_state == T_COMMIT);
805         commit_transaction->t_state = T_COMMIT_DFLUSH;
806         write_unlock(&journal->j_state_lock);
807
808         /*
809          * If the journal is not located on the file system device,
810          * then we must flush the file system device before we issue
811          * the commit record
812          */
813         if (commit_transaction->t_need_data_flush &&
814             (journal->j_fs_dev != journal->j_dev) &&
815             (journal->j_flags & JBD2_BARRIER))
816                 blkdev_issue_flush(journal->j_fs_dev);
817
818         /* Done it all: now write the commit record asynchronously. */
819         if (jbd2_has_feature_async_commit(journal)) {
820                 err = journal_submit_commit_record(journal, commit_transaction,
821                                                  &cbh, crc32_sum);
822                 if (err)
823                         jbd2_journal_abort(journal, err);
824         }
825
826         blk_finish_plug(&plug);
827
828         /* Lo and behold: we have just managed to send a transaction to
829            the log.  Before we can commit it, wait for the IO so far to
830            complete.  Control buffers being written are on the
831            transaction's t_log_list queue, and metadata buffers are on
832            the io_bufs list.
833
834            Wait for the buffers in reverse order.  That way we are
835            less likely to be woken up until all IOs have completed, and
836            so we incur less scheduling load.
837         */
838
839         jbd2_debug(3, "JBD2: commit phase 3\n");
840
841         while (!list_empty(&io_bufs)) {
842                 struct buffer_head *bh = list_entry(io_bufs.prev,
843                                                     struct buffer_head,
844                                                     b_assoc_buffers);
845
846                 wait_on_buffer(bh);
847                 cond_resched();
848
849                 if (unlikely(!buffer_uptodate(bh)))
850                         err = -EIO;
851                 jbd2_unfile_log_bh(bh);
852                 stats.run.rs_blocks_logged++;
853
854                 /*
855                  * The list contains temporary buffer heads created by
856                  * jbd2_journal_write_metadata_buffer().
857                  */
858                 BUFFER_TRACE(bh, "dumping temporary bh");
859                 __brelse(bh);
860                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
861                 free_buffer_head(bh);
862
863                 /* We also have to refile the corresponding shadowed buffer */
864                 jh = commit_transaction->t_shadow_list->b_tprev;
865                 bh = jh2bh(jh);
866                 clear_buffer_jwrite(bh);
867                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
868                 J_ASSERT_BH(bh, !buffer_shadow(bh));
869
870                 /* The metadata is now released for reuse, but we need
871                    to remember it against this transaction so that when
872                    we finally commit, we can do any checkpointing
873                    required. */
874                 JBUFFER_TRACE(jh, "file as BJ_Forget");
875                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
876                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
877                 __brelse(bh);
878         }
879
880         J_ASSERT (commit_transaction->t_shadow_list == NULL);
881
882         jbd2_debug(3, "JBD2: commit phase 4\n");
883
884         /* Here we wait for the revoke record and descriptor record buffers */
885         while (!list_empty(&log_bufs)) {
886                 struct buffer_head *bh;
887
888                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
889                 wait_on_buffer(bh);
890                 cond_resched();
891
892                 if (unlikely(!buffer_uptodate(bh)))
893                         err = -EIO;
894
895                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
896                 clear_buffer_jwrite(bh);
897                 jbd2_unfile_log_bh(bh);
898                 stats.run.rs_blocks_logged++;
899                 __brelse(bh);           /* One for getblk */
900                 /* AKPM: bforget here */
901         }
902
903         if (err)
904                 jbd2_journal_abort(journal, err);
905
906         jbd2_debug(3, "JBD2: commit phase 5\n");
907         write_lock(&journal->j_state_lock);
908         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
909         commit_transaction->t_state = T_COMMIT_JFLUSH;
910         write_unlock(&journal->j_state_lock);
911
912         if (!jbd2_has_feature_async_commit(journal)) {
913                 err = journal_submit_commit_record(journal, commit_transaction,
914                                                 &cbh, crc32_sum);
915                 if (err)
916                         jbd2_journal_abort(journal, err);
917         }
918         if (cbh)
919                 err = journal_wait_on_commit_record(journal, cbh);
920         stats.run.rs_blocks_logged++;
921         if (jbd2_has_feature_async_commit(journal) &&
922             journal->j_flags & JBD2_BARRIER) {
923                 blkdev_issue_flush(journal->j_dev);
924         }
925
926         if (err)
927                 jbd2_journal_abort(journal, err);
928
929         WARN_ON_ONCE(
930                 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
931
932         /*
933          * Now disk caches for filesystem device are flushed so we are safe to
934          * erase checkpointed transactions from the log by updating journal
935          * superblock.
936          */
937         if (update_tail)
938                 jbd2_update_log_tail(journal, first_tid, first_block);
939
940         /* End of a transaction!  Finally, we can do checkpoint
941            processing: any buffers committed as a result of this
942            transaction can be removed from any checkpoint list it was on
943            before. */
944
945         jbd2_debug(3, "JBD2: commit phase 6\n");
946
947         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
948         J_ASSERT(commit_transaction->t_buffers == NULL);
949         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
950         J_ASSERT(commit_transaction->t_shadow_list == NULL);
951
952 restart_loop:
953         /*
954          * As there are other places (journal_unmap_buffer()) adding buffers
955          * to this list we have to be careful and hold the j_list_lock.
956          */
957         spin_lock(&journal->j_list_lock);
958         while (commit_transaction->t_forget) {
959                 transaction_t *cp_transaction;
960                 struct buffer_head *bh;
961                 int try_to_free = 0;
962                 bool drop_ref;
963
964                 jh = commit_transaction->t_forget;
965                 spin_unlock(&journal->j_list_lock);
966                 bh = jh2bh(jh);
967                 /*
968                  * Get a reference so that bh cannot be freed before we are
969                  * done with it.
970                  */
971                 get_bh(bh);
972                 spin_lock(&jh->b_state_lock);
973                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
974
975                 /*
976                  * If there is undo-protected committed data against
977                  * this buffer, then we can remove it now.  If it is a
978                  * buffer needing such protection, the old frozen_data
979                  * field now points to a committed version of the
980                  * buffer, so rotate that field to the new committed
981                  * data.
982                  *
983                  * Otherwise, we can just throw away the frozen data now.
984                  *
985                  * We also know that the frozen data has already fired
986                  * its triggers if they exist, so we can clear that too.
987                  */
988                 if (jh->b_committed_data) {
989                         jbd2_free(jh->b_committed_data, bh->b_size);
990                         jh->b_committed_data = NULL;
991                         if (jh->b_frozen_data) {
992                                 jh->b_committed_data = jh->b_frozen_data;
993                                 jh->b_frozen_data = NULL;
994                                 jh->b_frozen_triggers = NULL;
995                         }
996                 } else if (jh->b_frozen_data) {
997                         jbd2_free(jh->b_frozen_data, bh->b_size);
998                         jh->b_frozen_data = NULL;
999                         jh->b_frozen_triggers = NULL;
1000                 }
1001
1002                 spin_lock(&journal->j_list_lock);
1003                 cp_transaction = jh->b_cp_transaction;
1004                 if (cp_transaction) {
1005                         JBUFFER_TRACE(jh, "remove from old cp transaction");
1006                         cp_transaction->t_chp_stats.cs_dropped++;
1007                         __jbd2_journal_remove_checkpoint(jh);
1008                 }
1009
1010                 /* Only re-checkpoint the buffer_head if it is marked
1011                  * dirty.  If the buffer was added to the BJ_Forget list
1012                  * by jbd2_journal_forget, it may no longer be dirty and
1013                  * there's no point in keeping a checkpoint record for
1014                  * it. */
1015
1016                 /*
1017                  * A buffer which has been freed while still being journaled
1018                  * by a previous transaction, refile the buffer to BJ_Forget of
1019                  * the running transaction. If the just committed transaction
1020                  * contains "add to orphan" operation, we can completely
1021                  * invalidate the buffer now. We are rather through in that
1022                  * since the buffer may be still accessible when blocksize <
1023                  * pagesize and it is attached to the last partial page.
1024                  */
1025                 if (buffer_freed(bh) && !jh->b_next_transaction) {
1026                         struct address_space *mapping;
1027
1028                         clear_buffer_freed(bh);
1029                         clear_buffer_jbddirty(bh);
1030
1031                         /*
1032                          * Block device buffers need to stay mapped all the
1033                          * time, so it is enough to clear buffer_jbddirty and
1034                          * buffer_freed bits. For the file mapping buffers (i.e.
1035                          * journalled data) we need to unmap buffer and clear
1036                          * more bits. We also need to be careful about the check
1037                          * because the data page mapping can get cleared under
1038                          * our hands. Note that if mapping == NULL, we don't
1039                          * need to make buffer unmapped because the page is
1040                          * already detached from the mapping and buffers cannot
1041                          * get reused.
1042                          */
1043                         mapping = READ_ONCE(bh->b_page->mapping);
1044                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1045                                 clear_buffer_mapped(bh);
1046                                 clear_buffer_new(bh);
1047                                 clear_buffer_req(bh);
1048                                 bh->b_bdev = NULL;
1049                         }
1050                 }
1051
1052                 if (buffer_jbddirty(bh)) {
1053                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1054                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1055                         if (is_journal_aborted(journal))
1056                                 clear_buffer_jbddirty(bh);
1057                 } else {
1058                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1059                         /*
1060                          * The buffer on BJ_Forget list and not jbddirty means
1061                          * it has been freed by this transaction and hence it
1062                          * could not have been reallocated until this
1063                          * transaction has committed. *BUT* it could be
1064                          * reallocated once we have written all the data to
1065                          * disk and before we process the buffer on BJ_Forget
1066                          * list.
1067                          */
1068                         if (!jh->b_next_transaction)
1069                                 try_to_free = 1;
1070                 }
1071                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1072                 drop_ref = __jbd2_journal_refile_buffer(jh);
1073                 spin_unlock(&jh->b_state_lock);
1074                 if (drop_ref)
1075                         jbd2_journal_put_journal_head(jh);
1076                 if (try_to_free)
1077                         release_buffer_page(bh);        /* Drops bh reference */
1078                 else
1079                         __brelse(bh);
1080                 cond_resched_lock(&journal->j_list_lock);
1081         }
1082         spin_unlock(&journal->j_list_lock);
1083         /*
1084          * This is a bit sleazy.  We use j_list_lock to protect transition
1085          * of a transaction into T_FINISHED state and calling
1086          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1087          * other checkpointing code processing the transaction...
1088          */
1089         write_lock(&journal->j_state_lock);
1090         spin_lock(&journal->j_list_lock);
1091         /*
1092          * Now recheck if some buffers did not get attached to the transaction
1093          * while the lock was dropped...
1094          */
1095         if (commit_transaction->t_forget) {
1096                 spin_unlock(&journal->j_list_lock);
1097                 write_unlock(&journal->j_state_lock);
1098                 goto restart_loop;
1099         }
1100
1101         /* Add the transaction to the checkpoint list
1102          * __journal_remove_checkpoint() can not destroy transaction
1103          * under us because it is not marked as T_FINISHED yet */
1104         if (journal->j_checkpoint_transactions == NULL) {
1105                 journal->j_checkpoint_transactions = commit_transaction;
1106                 commit_transaction->t_cpnext = commit_transaction;
1107                 commit_transaction->t_cpprev = commit_transaction;
1108         } else {
1109                 commit_transaction->t_cpnext =
1110                         journal->j_checkpoint_transactions;
1111                 commit_transaction->t_cpprev =
1112                         commit_transaction->t_cpnext->t_cpprev;
1113                 commit_transaction->t_cpnext->t_cpprev =
1114                         commit_transaction;
1115                 commit_transaction->t_cpprev->t_cpnext =
1116                                 commit_transaction;
1117         }
1118         spin_unlock(&journal->j_list_lock);
1119
1120         /* Done with this transaction! */
1121
1122         jbd2_debug(3, "JBD2: commit phase 7\n");
1123
1124         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1125
1126         commit_transaction->t_start = jiffies;
1127         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1128                                               commit_transaction->t_start);
1129
1130         /*
1131          * File the transaction statistics
1132          */
1133         stats.ts_tid = commit_transaction->t_tid;
1134         stats.run.rs_handle_count =
1135                 atomic_read(&commit_transaction->t_handle_count);
1136         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1137                              commit_transaction->t_tid, &stats.run);
1138         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1139
1140         commit_transaction->t_state = T_COMMIT_CALLBACK;
1141         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1142         journal->j_commit_sequence = commit_transaction->t_tid;
1143         journal->j_committing_transaction = NULL;
1144         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1145
1146         /*
1147          * weight the commit time higher than the average time so we don't
1148          * react too strongly to vast changes in the commit time
1149          */
1150         if (likely(journal->j_average_commit_time))
1151                 journal->j_average_commit_time = (commit_time +
1152                                 journal->j_average_commit_time*3) / 4;
1153         else
1154                 journal->j_average_commit_time = commit_time;
1155
1156         write_unlock(&journal->j_state_lock);
1157
1158         if (journal->j_commit_callback)
1159                 journal->j_commit_callback(journal, commit_transaction);
1160         if (journal->j_fc_cleanup_callback)
1161                 journal->j_fc_cleanup_callback(journal, 1, commit_transaction->t_tid);
1162
1163         trace_jbd2_end_commit(journal, commit_transaction);
1164         jbd2_debug(1, "JBD2: commit %d complete, head %d\n",
1165                   journal->j_commit_sequence, journal->j_tail_sequence);
1166
1167         write_lock(&journal->j_state_lock);
1168         journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1169         journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1170         spin_lock(&journal->j_list_lock);
1171         commit_transaction->t_state = T_FINISHED;
1172         /* Check if the transaction can be dropped now that we are finished */
1173         if (commit_transaction->t_checkpoint_list == NULL &&
1174             commit_transaction->t_checkpoint_io_list == NULL) {
1175                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1176                 jbd2_journal_free_transaction(commit_transaction);
1177         }
1178         spin_unlock(&journal->j_list_lock);
1179         write_unlock(&journal->j_state_lock);
1180         wake_up(&journal->j_wait_done_commit);
1181         wake_up(&journal->j_fc_wait);
1182
1183         /*
1184          * Calculate overall stats
1185          */
1186         spin_lock(&journal->j_history_lock);
1187         journal->j_stats.ts_tid++;
1188         journal->j_stats.ts_requested += stats.ts_requested;
1189         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1190         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1191         journal->j_stats.run.rs_running += stats.run.rs_running;
1192         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1193         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1194         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1195         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1196         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1197         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1198         spin_unlock(&journal->j_history_lock);
1199 }