Merge tag 'arc-5.15-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/vgupta/arc
[linux-2.6-microblaze.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 page *page;
66
67         if (buffer_dirty(bh))
68                 goto nope;
69         if (atomic_read(&bh->b_count) != 1)
70                 goto nope;
71         page = bh->b_page;
72         if (!page)
73                 goto nope;
74         if (page->mapping)
75                 goto nope;
76
77         /* OK, it's a truncated page */
78         if (!trylock_page(page))
79                 goto nope;
80
81         get_page(page);
82         __brelse(bh);
83         try_to_free_buffers(page);
84         unlock_page(page);
85         put_page(page);
86         return;
87
88 nope:
89         __brelse(bh);
90 }
91
92 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
93 {
94         struct commit_header *h;
95         __u32 csum;
96
97         if (!jbd2_journal_has_csum_v2or3(j))
98                 return;
99
100         h = (struct commit_header *)(bh->b_data);
101         h->h_chksum_type = 0;
102         h->h_chksum_size = 0;
103         h->h_chksum[0] = 0;
104         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
105         h->h_chksum[0] = cpu_to_be32(csum);
106 }
107
108 /*
109  * Done it all: now submit the commit record.  We should have
110  * cleaned up our previous buffers by now, so if we are in abort
111  * mode we can now just skip the rest of the journal write
112  * entirely.
113  *
114  * Returns 1 if the journal needs to be aborted or 0 on success
115  */
116 static int journal_submit_commit_record(journal_t *journal,
117                                         transaction_t *commit_transaction,
118                                         struct buffer_head **cbh,
119                                         __u32 crc32_sum)
120 {
121         struct commit_header *tmp;
122         struct buffer_head *bh;
123         int ret;
124         struct timespec64 now;
125
126         *cbh = NULL;
127
128         if (is_journal_aborted(journal))
129                 return 0;
130
131         bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
132                                                 JBD2_COMMIT_BLOCK);
133         if (!bh)
134                 return 1;
135
136         tmp = (struct commit_header *)bh->b_data;
137         ktime_get_coarse_real_ts64(&now);
138         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
139         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
140
141         if (jbd2_has_feature_checksum(journal)) {
142                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
143                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
144                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
145         }
146         jbd2_commit_block_csum_set(journal, bh);
147
148         BUFFER_TRACE(bh, "submit commit block");
149         lock_buffer(bh);
150         clear_buffer_dirty(bh);
151         set_buffer_uptodate(bh);
152         bh->b_end_io = journal_end_buffer_io_sync;
153
154         if (journal->j_flags & JBD2_BARRIER &&
155             !jbd2_has_feature_async_commit(journal))
156                 ret = submit_bh(REQ_OP_WRITE,
157                         REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh);
158         else
159                 ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
160
161         *cbh = bh;
162         return ret;
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                 spin_lock(&journal->j_list_lock);
304                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
305                 smp_mb();
306                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
307         }
308
309         /* Now refile inode to proper lists */
310         list_for_each_entry_safe(jinode, next_i,
311                                  &commit_transaction->t_inode_list, i_list) {
312                 list_del(&jinode->i_list);
313                 if (jinode->i_next_transaction) {
314                         jinode->i_transaction = jinode->i_next_transaction;
315                         jinode->i_next_transaction = NULL;
316                         list_add(&jinode->i_list,
317                                 &jinode->i_transaction->t_inode_list);
318                 } else {
319                         jinode->i_transaction = NULL;
320                         jinode->i_dirty_start = 0;
321                         jinode->i_dirty_end = 0;
322                 }
323         }
324         spin_unlock(&journal->j_list_lock);
325
326         return ret;
327 }
328
329 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
330 {
331         struct page *page = bh->b_page;
332         char *addr;
333         __u32 checksum;
334
335         addr = kmap_atomic(page);
336         checksum = crc32_be(crc32_sum,
337                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
338         kunmap_atomic(addr);
339
340         return checksum;
341 }
342
343 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
344                                    unsigned long long block)
345 {
346         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
347         if (jbd2_has_feature_64bit(j))
348                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
349 }
350
351 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
352                                     struct buffer_head *bh, __u32 sequence)
353 {
354         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
355         struct page *page = bh->b_page;
356         __u8 *addr;
357         __u32 csum32;
358         __be32 seq;
359
360         if (!jbd2_journal_has_csum_v2or3(j))
361                 return;
362
363         seq = cpu_to_be32(sequence);
364         addr = kmap_atomic(page);
365         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
366         csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
367                              bh->b_size);
368         kunmap_atomic(addr);
369
370         if (jbd2_has_feature_csum3(j))
371                 tag3->t_checksum = cpu_to_be32(csum32);
372         else
373                 tag->t_checksum = cpu_to_be16(csum32);
374 }
375 /*
376  * jbd2_journal_commit_transaction
377  *
378  * The primary function for committing a transaction to the log.  This
379  * function is called by the journal thread to begin a complete commit.
380  */
381 void jbd2_journal_commit_transaction(journal_t *journal)
382 {
383         struct transaction_stats_s stats;
384         transaction_t *commit_transaction;
385         struct journal_head *jh;
386         struct buffer_head *descriptor;
387         struct buffer_head **wbuf = journal->j_wbuf;
388         int bufs;
389         int flags;
390         int err;
391         unsigned long long blocknr;
392         ktime_t start_time;
393         u64 commit_time;
394         char *tagp = NULL;
395         journal_block_tag_t *tag = NULL;
396         int space_left = 0;
397         int first_tag = 0;
398         int tag_flag;
399         int i;
400         int tag_bytes = journal_tag_bytes(journal);
401         struct buffer_head *cbh = NULL; /* For transactional checksums */
402         __u32 crc32_sum = ~0;
403         struct blk_plug plug;
404         /* Tail of the journal */
405         unsigned long first_block;
406         tid_t first_tid;
407         int update_tail;
408         int csum_size = 0;
409         LIST_HEAD(io_bufs);
410         LIST_HEAD(log_bufs);
411
412         if (jbd2_journal_has_csum_v2or3(journal))
413                 csum_size = sizeof(struct jbd2_journal_block_tail);
414
415         /*
416          * First job: lock down the current transaction and wait for
417          * all outstanding updates to complete.
418          */
419
420         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
421         if (journal->j_flags & JBD2_FLUSHED) {
422                 jbd_debug(3, "super block updated\n");
423                 mutex_lock_io(&journal->j_checkpoint_mutex);
424                 /*
425                  * We hold j_checkpoint_mutex so tail cannot change under us.
426                  * We don't need any special data guarantees for writing sb
427                  * since journal is empty and it is ok for write to be
428                  * flushed only with transaction commit.
429                  */
430                 jbd2_journal_update_sb_log_tail(journal,
431                                                 journal->j_tail_sequence,
432                                                 journal->j_tail,
433                                                 REQ_SYNC);
434                 mutex_unlock(&journal->j_checkpoint_mutex);
435         } else {
436                 jbd_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         jbd_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         spin_lock(&commit_transaction->t_handle_lock);
488         while (atomic_read(&commit_transaction->t_updates)) {
489                 DEFINE_WAIT(wait);
490
491                 prepare_to_wait(&journal->j_wait_updates, &wait,
492                                         TASK_UNINTERRUPTIBLE);
493                 if (atomic_read(&commit_transaction->t_updates)) {
494                         spin_unlock(&commit_transaction->t_handle_lock);
495                         write_unlock(&journal->j_state_lock);
496                         schedule();
497                         write_lock(&journal->j_state_lock);
498                         spin_lock(&commit_transaction->t_handle_lock);
499                 }
500                 finish_wait(&journal->j_wait_updates, &wait);
501         }
502         spin_unlock(&commit_transaction->t_handle_lock);
503         commit_transaction->t_state = T_SWITCH;
504         write_unlock(&journal->j_state_lock);
505
506         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
507                         journal->j_max_transaction_buffers);
508
509         /*
510          * First thing we are allowed to do is to discard any remaining
511          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
512          * that there are no such buffers: if a large filesystem
513          * operation like a truncate needs to split itself over multiple
514          * transactions, then it may try to do a jbd2_journal_restart() while
515          * there are still BJ_Reserved buffers outstanding.  These must
516          * be released cleanly from the current transaction.
517          *
518          * In this case, the filesystem must still reserve write access
519          * again before modifying the buffer in the new transaction, but
520          * we do not require it to remember exactly which old buffers it
521          * has reserved.  This is consistent with the existing behaviour
522          * that multiple jbd2_journal_get_write_access() calls to the same
523          * buffer are perfectly permissible.
524          */
525         while (commit_transaction->t_reserved_list) {
526                 jh = commit_transaction->t_reserved_list;
527                 JBUFFER_TRACE(jh, "reserved, unused: refile");
528                 /*
529                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
530                  * leave undo-committed data.
531                  */
532                 if (jh->b_committed_data) {
533                         struct buffer_head *bh = jh2bh(jh);
534
535                         spin_lock(&jh->b_state_lock);
536                         jbd2_free(jh->b_committed_data, bh->b_size);
537                         jh->b_committed_data = NULL;
538                         spin_unlock(&jh->b_state_lock);
539                 }
540                 jbd2_journal_refile_buffer(journal, jh);
541         }
542
543         /*
544          * Now try to drop any written-back buffers from the journal's
545          * checkpoint lists.  We do this *before* commit because it potentially
546          * frees some memory
547          */
548         spin_lock(&journal->j_list_lock);
549         __jbd2_journal_clean_checkpoint_list(journal, false);
550         spin_unlock(&journal->j_list_lock);
551
552         jbd_debug(3, "JBD2: commit phase 1\n");
553
554         /*
555          * Clear revoked flag to reflect there is no revoked buffers
556          * in the next transaction which is going to be started.
557          */
558         jbd2_clear_buffer_revoked_flags(journal);
559
560         /*
561          * Switch to a new revoke table.
562          */
563         jbd2_journal_switch_revoke_table(journal);
564
565         /*
566          * Reserved credits cannot be claimed anymore, free them
567          */
568         atomic_sub(atomic_read(&journal->j_reserved_credits),
569                    &commit_transaction->t_outstanding_credits);
570
571         write_lock(&journal->j_state_lock);
572         trace_jbd2_commit_flushing(journal, commit_transaction);
573         stats.run.rs_flushing = jiffies;
574         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
575                                              stats.run.rs_flushing);
576
577         commit_transaction->t_state = T_FLUSH;
578         journal->j_committing_transaction = commit_transaction;
579         journal->j_running_transaction = NULL;
580         start_time = ktime_get();
581         commit_transaction->t_log_start = journal->j_head;
582         wake_up(&journal->j_wait_transaction_locked);
583         write_unlock(&journal->j_state_lock);
584
585         jbd_debug(3, "JBD2: commit phase 2a\n");
586
587         /*
588          * Now start flushing things to disk, in the order they appear
589          * on the transaction lists.  Data blocks go first.
590          */
591         err = journal_submit_data_buffers(journal, commit_transaction);
592         if (err)
593                 jbd2_journal_abort(journal, err);
594
595         blk_start_plug(&plug);
596         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
597
598         jbd_debug(3, "JBD2: commit phase 2b\n");
599
600         /*
601          * Way to go: we have now written out all of the data for a
602          * transaction!  Now comes the tricky part: we need to write out
603          * metadata.  Loop over the transaction's entire buffer list:
604          */
605         write_lock(&journal->j_state_lock);
606         commit_transaction->t_state = T_COMMIT;
607         write_unlock(&journal->j_state_lock);
608
609         trace_jbd2_commit_logging(journal, commit_transaction);
610         stats.run.rs_logging = jiffies;
611         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
612                                                stats.run.rs_logging);
613         stats.run.rs_blocks = commit_transaction->t_nr_buffers;
614         stats.run.rs_blocks_logged = 0;
615
616         J_ASSERT(commit_transaction->t_nr_buffers <=
617                  atomic_read(&commit_transaction->t_outstanding_credits));
618
619         err = 0;
620         bufs = 0;
621         descriptor = NULL;
622         while (commit_transaction->t_buffers) {
623
624                 /* Find the next buffer to be journaled... */
625
626                 jh = commit_transaction->t_buffers;
627
628                 /* If we're in abort mode, we just un-journal the buffer and
629                    release it. */
630
631                 if (is_journal_aborted(journal)) {
632                         clear_buffer_jbddirty(jh2bh(jh));
633                         JBUFFER_TRACE(jh, "journal is aborting: refile");
634                         jbd2_buffer_abort_trigger(jh,
635                                                   jh->b_frozen_data ?
636                                                   jh->b_frozen_triggers :
637                                                   jh->b_triggers);
638                         jbd2_journal_refile_buffer(journal, jh);
639                         /* If that was the last one, we need to clean up
640                          * any descriptor buffers which may have been
641                          * already allocated, even if we are now
642                          * aborting. */
643                         if (!commit_transaction->t_buffers)
644                                 goto start_journal_io;
645                         continue;
646                 }
647
648                 /* Make sure we have a descriptor block in which to
649                    record the metadata buffer. */
650
651                 if (!descriptor) {
652                         J_ASSERT (bufs == 0);
653
654                         jbd_debug(4, "JBD2: get descriptor\n");
655
656                         descriptor = jbd2_journal_get_descriptor_buffer(
657                                                         commit_transaction,
658                                                         JBD2_DESCRIPTOR_BLOCK);
659                         if (!descriptor) {
660                                 jbd2_journal_abort(journal, -EIO);
661                                 continue;
662                         }
663
664                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
665                                 (unsigned long long)descriptor->b_blocknr,
666                                 descriptor->b_data);
667                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
668                         space_left = descriptor->b_size -
669                                                 sizeof(journal_header_t);
670                         first_tag = 1;
671                         set_buffer_jwrite(descriptor);
672                         set_buffer_dirty(descriptor);
673                         wbuf[bufs++] = descriptor;
674
675                         /* Record it so that we can wait for IO
676                            completion later */
677                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
678                         jbd2_file_log_bh(&log_bufs, descriptor);
679                 }
680
681                 /* Where is the buffer to be written? */
682
683                 err = jbd2_journal_next_log_block(journal, &blocknr);
684                 /* If the block mapping failed, just abandon the buffer
685                    and repeat this loop: we'll fall into the
686                    refile-on-abort condition above. */
687                 if (err) {
688                         jbd2_journal_abort(journal, err);
689                         continue;
690                 }
691
692                 /*
693                  * start_this_handle() uses t_outstanding_credits to determine
694                  * the free space in the log.
695                  */
696                 atomic_dec(&commit_transaction->t_outstanding_credits);
697
698                 /* Bump b_count to prevent truncate from stumbling over
699                    the shadowed buffer!  @@@ This can go if we ever get
700                    rid of the shadow pairing of buffers. */
701                 atomic_inc(&jh2bh(jh)->b_count);
702
703                 /*
704                  * Make a temporary IO buffer with which to write it out
705                  * (this will requeue the metadata buffer to BJ_Shadow).
706                  */
707                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
708                 JBUFFER_TRACE(jh, "ph3: write metadata");
709                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
710                                                 jh, &wbuf[bufs], blocknr);
711                 if (flags < 0) {
712                         jbd2_journal_abort(journal, flags);
713                         continue;
714                 }
715                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
716
717                 /* Record the new block's tag in the current descriptor
718                    buffer */
719
720                 tag_flag = 0;
721                 if (flags & 1)
722                         tag_flag |= JBD2_FLAG_ESCAPE;
723                 if (!first_tag)
724                         tag_flag |= JBD2_FLAG_SAME_UUID;
725
726                 tag = (journal_block_tag_t *) tagp;
727                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
728                 tag->t_flags = cpu_to_be16(tag_flag);
729                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
730                                         commit_transaction->t_tid);
731                 tagp += tag_bytes;
732                 space_left -= tag_bytes;
733                 bufs++;
734
735                 if (first_tag) {
736                         memcpy (tagp, journal->j_uuid, 16);
737                         tagp += 16;
738                         space_left -= 16;
739                         first_tag = 0;
740                 }
741
742                 /* If there's no more to do, or if the descriptor is full,
743                    let the IO rip! */
744
745                 if (bufs == journal->j_wbufsize ||
746                     commit_transaction->t_buffers == NULL ||
747                     space_left < tag_bytes + 16 + csum_size) {
748
749                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
750
751                         /* Write an end-of-descriptor marker before
752                            submitting the IOs.  "tag" still points to
753                            the last tag we set up. */
754
755                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
756 start_journal_io:
757                         if (descriptor)
758                                 jbd2_descriptor_block_csum_set(journal,
759                                                         descriptor);
760
761                         for (i = 0; i < bufs; i++) {
762                                 struct buffer_head *bh = wbuf[i];
763                                 /*
764                                  * Compute checksum.
765                                  */
766                                 if (jbd2_has_feature_checksum(journal)) {
767                                         crc32_sum =
768                                             jbd2_checksum_data(crc32_sum, bh);
769                                 }
770
771                                 lock_buffer(bh);
772                                 clear_buffer_dirty(bh);
773                                 set_buffer_uptodate(bh);
774                                 bh->b_end_io = journal_end_buffer_io_sync;
775                                 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
776                         }
777                         cond_resched();
778
779                         /* Force a new descriptor to be generated next
780                            time round the loop. */
781                         descriptor = NULL;
782                         bufs = 0;
783                 }
784         }
785
786         err = journal_finish_inode_data_buffers(journal, commit_transaction);
787         if (err) {
788                 printk(KERN_WARNING
789                         "JBD2: Detected IO errors while flushing file data "
790                        "on %s\n", journal->j_devname);
791                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
792                         jbd2_journal_abort(journal, err);
793                 err = 0;
794         }
795
796         /*
797          * Get current oldest transaction in the log before we issue flush
798          * to the filesystem device. After the flush we can be sure that
799          * blocks of all older transactions are checkpointed to persistent
800          * storage and we will be safe to update journal start in the
801          * superblock with the numbers we get here.
802          */
803         update_tail =
804                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
805
806         write_lock(&journal->j_state_lock);
807         if (update_tail) {
808                 long freed = first_block - journal->j_tail;
809
810                 if (first_block < journal->j_tail)
811                         freed += journal->j_last - journal->j_first;
812                 /* Update tail only if we free significant amount of space */
813                 if (freed < jbd2_journal_get_max_txn_bufs(journal))
814                         update_tail = 0;
815         }
816         J_ASSERT(commit_transaction->t_state == T_COMMIT);
817         commit_transaction->t_state = T_COMMIT_DFLUSH;
818         write_unlock(&journal->j_state_lock);
819
820         /* 
821          * If the journal is not located on the file system device,
822          * then we must flush the file system device before we issue
823          * the commit record
824          */
825         if (commit_transaction->t_need_data_flush &&
826             (journal->j_fs_dev != journal->j_dev) &&
827             (journal->j_flags & JBD2_BARRIER))
828                 blkdev_issue_flush(journal->j_fs_dev);
829
830         /* Done it all: now write the commit record asynchronously. */
831         if (jbd2_has_feature_async_commit(journal)) {
832                 err = journal_submit_commit_record(journal, commit_transaction,
833                                                  &cbh, crc32_sum);
834                 if (err)
835                         jbd2_journal_abort(journal, err);
836         }
837
838         blk_finish_plug(&plug);
839
840         /* Lo and behold: we have just managed to send a transaction to
841            the log.  Before we can commit it, wait for the IO so far to
842            complete.  Control buffers being written are on the
843            transaction's t_log_list queue, and metadata buffers are on
844            the io_bufs list.
845
846            Wait for the buffers in reverse order.  That way we are
847            less likely to be woken up until all IOs have completed, and
848            so we incur less scheduling load.
849         */
850
851         jbd_debug(3, "JBD2: commit phase 3\n");
852
853         while (!list_empty(&io_bufs)) {
854                 struct buffer_head *bh = list_entry(io_bufs.prev,
855                                                     struct buffer_head,
856                                                     b_assoc_buffers);
857
858                 wait_on_buffer(bh);
859                 cond_resched();
860
861                 if (unlikely(!buffer_uptodate(bh)))
862                         err = -EIO;
863                 jbd2_unfile_log_bh(bh);
864                 stats.run.rs_blocks_logged++;
865
866                 /*
867                  * The list contains temporary buffer heads created by
868                  * jbd2_journal_write_metadata_buffer().
869                  */
870                 BUFFER_TRACE(bh, "dumping temporary bh");
871                 __brelse(bh);
872                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
873                 free_buffer_head(bh);
874
875                 /* We also have to refile the corresponding shadowed buffer */
876                 jh = commit_transaction->t_shadow_list->b_tprev;
877                 bh = jh2bh(jh);
878                 clear_buffer_jwrite(bh);
879                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
880                 J_ASSERT_BH(bh, !buffer_shadow(bh));
881
882                 /* The metadata is now released for reuse, but we need
883                    to remember it against this transaction so that when
884                    we finally commit, we can do any checkpointing
885                    required. */
886                 JBUFFER_TRACE(jh, "file as BJ_Forget");
887                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
888                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
889                 __brelse(bh);
890         }
891
892         J_ASSERT (commit_transaction->t_shadow_list == NULL);
893
894         jbd_debug(3, "JBD2: commit phase 4\n");
895
896         /* Here we wait for the revoke record and descriptor record buffers */
897         while (!list_empty(&log_bufs)) {
898                 struct buffer_head *bh;
899
900                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
901                 wait_on_buffer(bh);
902                 cond_resched();
903
904                 if (unlikely(!buffer_uptodate(bh)))
905                         err = -EIO;
906
907                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
908                 clear_buffer_jwrite(bh);
909                 jbd2_unfile_log_bh(bh);
910                 stats.run.rs_blocks_logged++;
911                 __brelse(bh);           /* One for getblk */
912                 /* AKPM: bforget here */
913         }
914
915         if (err)
916                 jbd2_journal_abort(journal, err);
917
918         jbd_debug(3, "JBD2: commit phase 5\n");
919         write_lock(&journal->j_state_lock);
920         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
921         commit_transaction->t_state = T_COMMIT_JFLUSH;
922         write_unlock(&journal->j_state_lock);
923
924         if (!jbd2_has_feature_async_commit(journal)) {
925                 err = journal_submit_commit_record(journal, commit_transaction,
926                                                 &cbh, crc32_sum);
927                 if (err)
928                         jbd2_journal_abort(journal, err);
929         }
930         if (cbh)
931                 err = journal_wait_on_commit_record(journal, cbh);
932         stats.run.rs_blocks_logged++;
933         if (jbd2_has_feature_async_commit(journal) &&
934             journal->j_flags & JBD2_BARRIER) {
935                 blkdev_issue_flush(journal->j_dev);
936         }
937
938         if (err)
939                 jbd2_journal_abort(journal, err);
940
941         WARN_ON_ONCE(
942                 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
943
944         /*
945          * Now disk caches for filesystem device are flushed so we are safe to
946          * erase checkpointed transactions from the log by updating journal
947          * superblock.
948          */
949         if (update_tail)
950                 jbd2_update_log_tail(journal, first_tid, first_block);
951
952         /* End of a transaction!  Finally, we can do checkpoint
953            processing: any buffers committed as a result of this
954            transaction can be removed from any checkpoint list it was on
955            before. */
956
957         jbd_debug(3, "JBD2: commit phase 6\n");
958
959         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
960         J_ASSERT(commit_transaction->t_buffers == NULL);
961         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
962         J_ASSERT(commit_transaction->t_shadow_list == NULL);
963
964 restart_loop:
965         /*
966          * As there are other places (journal_unmap_buffer()) adding buffers
967          * to this list we have to be careful and hold the j_list_lock.
968          */
969         spin_lock(&journal->j_list_lock);
970         while (commit_transaction->t_forget) {
971                 transaction_t *cp_transaction;
972                 struct buffer_head *bh;
973                 int try_to_free = 0;
974                 bool drop_ref;
975
976                 jh = commit_transaction->t_forget;
977                 spin_unlock(&journal->j_list_lock);
978                 bh = jh2bh(jh);
979                 /*
980                  * Get a reference so that bh cannot be freed before we are
981                  * done with it.
982                  */
983                 get_bh(bh);
984                 spin_lock(&jh->b_state_lock);
985                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
986
987                 /*
988                  * If there is undo-protected committed data against
989                  * this buffer, then we can remove it now.  If it is a
990                  * buffer needing such protection, the old frozen_data
991                  * field now points to a committed version of the
992                  * buffer, so rotate that field to the new committed
993                  * data.
994                  *
995                  * Otherwise, we can just throw away the frozen data now.
996                  *
997                  * We also know that the frozen data has already fired
998                  * its triggers if they exist, so we can clear that too.
999                  */
1000                 if (jh->b_committed_data) {
1001                         jbd2_free(jh->b_committed_data, bh->b_size);
1002                         jh->b_committed_data = NULL;
1003                         if (jh->b_frozen_data) {
1004                                 jh->b_committed_data = jh->b_frozen_data;
1005                                 jh->b_frozen_data = NULL;
1006                                 jh->b_frozen_triggers = NULL;
1007                         }
1008                 } else if (jh->b_frozen_data) {
1009                         jbd2_free(jh->b_frozen_data, bh->b_size);
1010                         jh->b_frozen_data = NULL;
1011                         jh->b_frozen_triggers = NULL;
1012                 }
1013
1014                 spin_lock(&journal->j_list_lock);
1015                 cp_transaction = jh->b_cp_transaction;
1016                 if (cp_transaction) {
1017                         JBUFFER_TRACE(jh, "remove from old cp transaction");
1018                         cp_transaction->t_chp_stats.cs_dropped++;
1019                         __jbd2_journal_remove_checkpoint(jh);
1020                 }
1021
1022                 /* Only re-checkpoint the buffer_head if it is marked
1023                  * dirty.  If the buffer was added to the BJ_Forget list
1024                  * by jbd2_journal_forget, it may no longer be dirty and
1025                  * there's no point in keeping a checkpoint record for
1026                  * it. */
1027
1028                 /*
1029                  * A buffer which has been freed while still being journaled
1030                  * by a previous transaction, refile the buffer to BJ_Forget of
1031                  * the running transaction. If the just committed transaction
1032                  * contains "add to orphan" operation, we can completely
1033                  * invalidate the buffer now. We are rather through in that
1034                  * since the buffer may be still accessible when blocksize <
1035                  * pagesize and it is attached to the last partial page.
1036                  */
1037                 if (buffer_freed(bh) && !jh->b_next_transaction) {
1038                         struct address_space *mapping;
1039
1040                         clear_buffer_freed(bh);
1041                         clear_buffer_jbddirty(bh);
1042
1043                         /*
1044                          * Block device buffers need to stay mapped all the
1045                          * time, so it is enough to clear buffer_jbddirty and
1046                          * buffer_freed bits. For the file mapping buffers (i.e.
1047                          * journalled data) we need to unmap buffer and clear
1048                          * more bits. We also need to be careful about the check
1049                          * because the data page mapping can get cleared under
1050                          * our hands. Note that if mapping == NULL, we don't
1051                          * need to make buffer unmapped because the page is
1052                          * already detached from the mapping and buffers cannot
1053                          * get reused.
1054                          */
1055                         mapping = READ_ONCE(bh->b_page->mapping);
1056                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1057                                 clear_buffer_mapped(bh);
1058                                 clear_buffer_new(bh);
1059                                 clear_buffer_req(bh);
1060                                 bh->b_bdev = NULL;
1061                         }
1062                 }
1063
1064                 if (buffer_jbddirty(bh)) {
1065                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1066                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1067                         if (is_journal_aborted(journal))
1068                                 clear_buffer_jbddirty(bh);
1069                 } else {
1070                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1071                         /*
1072                          * The buffer on BJ_Forget list and not jbddirty means
1073                          * it has been freed by this transaction and hence it
1074                          * could not have been reallocated until this
1075                          * transaction has committed. *BUT* it could be
1076                          * reallocated once we have written all the data to
1077                          * disk and before we process the buffer on BJ_Forget
1078                          * list.
1079                          */
1080                         if (!jh->b_next_transaction)
1081                                 try_to_free = 1;
1082                 }
1083                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1084                 drop_ref = __jbd2_journal_refile_buffer(jh);
1085                 spin_unlock(&jh->b_state_lock);
1086                 if (drop_ref)
1087                         jbd2_journal_put_journal_head(jh);
1088                 if (try_to_free)
1089                         release_buffer_page(bh);        /* Drops bh reference */
1090                 else
1091                         __brelse(bh);
1092                 cond_resched_lock(&journal->j_list_lock);
1093         }
1094         spin_unlock(&journal->j_list_lock);
1095         /*
1096          * This is a bit sleazy.  We use j_list_lock to protect transition
1097          * of a transaction into T_FINISHED state and calling
1098          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1099          * other checkpointing code processing the transaction...
1100          */
1101         write_lock(&journal->j_state_lock);
1102         spin_lock(&journal->j_list_lock);
1103         /*
1104          * Now recheck if some buffers did not get attached to the transaction
1105          * while the lock was dropped...
1106          */
1107         if (commit_transaction->t_forget) {
1108                 spin_unlock(&journal->j_list_lock);
1109                 write_unlock(&journal->j_state_lock);
1110                 goto restart_loop;
1111         }
1112
1113         /* Add the transaction to the checkpoint list
1114          * __journal_remove_checkpoint() can not destroy transaction
1115          * under us because it is not marked as T_FINISHED yet */
1116         if (journal->j_checkpoint_transactions == NULL) {
1117                 journal->j_checkpoint_transactions = commit_transaction;
1118                 commit_transaction->t_cpnext = commit_transaction;
1119                 commit_transaction->t_cpprev = commit_transaction;
1120         } else {
1121                 commit_transaction->t_cpnext =
1122                         journal->j_checkpoint_transactions;
1123                 commit_transaction->t_cpprev =
1124                         commit_transaction->t_cpnext->t_cpprev;
1125                 commit_transaction->t_cpnext->t_cpprev =
1126                         commit_transaction;
1127                 commit_transaction->t_cpprev->t_cpnext =
1128                                 commit_transaction;
1129         }
1130         spin_unlock(&journal->j_list_lock);
1131
1132         /* Done with this transaction! */
1133
1134         jbd_debug(3, "JBD2: commit phase 7\n");
1135
1136         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1137
1138         commit_transaction->t_start = jiffies;
1139         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1140                                               commit_transaction->t_start);
1141
1142         /*
1143          * File the transaction statistics
1144          */
1145         stats.ts_tid = commit_transaction->t_tid;
1146         stats.run.rs_handle_count =
1147                 atomic_read(&commit_transaction->t_handle_count);
1148         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1149                              commit_transaction->t_tid, &stats.run);
1150         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1151
1152         commit_transaction->t_state = T_COMMIT_CALLBACK;
1153         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1154         journal->j_commit_sequence = commit_transaction->t_tid;
1155         journal->j_committing_transaction = NULL;
1156         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1157
1158         /*
1159          * weight the commit time higher than the average time so we don't
1160          * react too strongly to vast changes in the commit time
1161          */
1162         if (likely(journal->j_average_commit_time))
1163                 journal->j_average_commit_time = (commit_time +
1164                                 journal->j_average_commit_time*3) / 4;
1165         else
1166                 journal->j_average_commit_time = commit_time;
1167
1168         write_unlock(&journal->j_state_lock);
1169
1170         if (journal->j_commit_callback)
1171                 journal->j_commit_callback(journal, commit_transaction);
1172         if (journal->j_fc_cleanup_callback)
1173                 journal->j_fc_cleanup_callback(journal, 1);
1174
1175         trace_jbd2_end_commit(journal, commit_transaction);
1176         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1177                   journal->j_commit_sequence, journal->j_tail_sequence);
1178
1179         write_lock(&journal->j_state_lock);
1180         journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1181         journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1182         spin_lock(&journal->j_list_lock);
1183         commit_transaction->t_state = T_FINISHED;
1184         /* Check if the transaction can be dropped now that we are finished */
1185         if (commit_transaction->t_checkpoint_list == NULL &&
1186             commit_transaction->t_checkpoint_io_list == NULL) {
1187                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1188                 jbd2_journal_free_transaction(commit_transaction);
1189         }
1190         spin_unlock(&journal->j_list_lock);
1191         write_unlock(&journal->j_state_lock);
1192         wake_up(&journal->j_wait_done_commit);
1193         wake_up(&journal->j_fc_wait);
1194
1195         /*
1196          * Calculate overall stats
1197          */
1198         spin_lock(&journal->j_history_lock);
1199         journal->j_stats.ts_tid++;
1200         journal->j_stats.ts_requested += stats.ts_requested;
1201         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1202         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1203         journal->j_stats.run.rs_running += stats.run.rs_running;
1204         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1205         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1206         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1207         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1208         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1209         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1210         spin_unlock(&journal->j_history_lock);
1211 }