Merge tag 'iio-fixes-for-5.18a' of https://git.kernel.org/pub/scm/linux/kernel/git...
[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         // waits for any t_updates to finish
488         jbd2_journal_wait_updates(journal);
489
490         commit_transaction->t_state = T_SWITCH;
491         write_unlock(&journal->j_state_lock);
492
493         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
494                         journal->j_max_transaction_buffers);
495
496         /*
497          * First thing we are allowed to do is to discard any remaining
498          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
499          * that there are no such buffers: if a large filesystem
500          * operation like a truncate needs to split itself over multiple
501          * transactions, then it may try to do a jbd2_journal_restart() while
502          * there are still BJ_Reserved buffers outstanding.  These must
503          * be released cleanly from the current transaction.
504          *
505          * In this case, the filesystem must still reserve write access
506          * again before modifying the buffer in the new transaction, but
507          * we do not require it to remember exactly which old buffers it
508          * has reserved.  This is consistent with the existing behaviour
509          * that multiple jbd2_journal_get_write_access() calls to the same
510          * buffer are perfectly permissible.
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         /*
531          * Now try to drop any written-back buffers from the journal's
532          * checkpoint lists.  We do this *before* commit because it potentially
533          * frees some memory
534          */
535         spin_lock(&journal->j_list_lock);
536         __jbd2_journal_clean_checkpoint_list(journal, false);
537         spin_unlock(&journal->j_list_lock);
538
539         jbd_debug(3, "JBD2: commit phase 1\n");
540
541         /*
542          * Clear revoked flag to reflect there is no revoked buffers
543          * in the next transaction which is going to be started.
544          */
545         jbd2_clear_buffer_revoked_flags(journal);
546
547         /*
548          * Switch to a new revoke table.
549          */
550         jbd2_journal_switch_revoke_table(journal);
551
552         /*
553          * Reserved credits cannot be claimed anymore, free them
554          */
555         atomic_sub(atomic_read(&journal->j_reserved_credits),
556                    &commit_transaction->t_outstanding_credits);
557
558         write_lock(&journal->j_state_lock);
559         trace_jbd2_commit_flushing(journal, commit_transaction);
560         stats.run.rs_flushing = jiffies;
561         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
562                                              stats.run.rs_flushing);
563
564         commit_transaction->t_state = T_FLUSH;
565         journal->j_committing_transaction = commit_transaction;
566         journal->j_running_transaction = NULL;
567         start_time = ktime_get();
568         commit_transaction->t_log_start = journal->j_head;
569         wake_up(&journal->j_wait_transaction_locked);
570         write_unlock(&journal->j_state_lock);
571
572         jbd_debug(3, "JBD2: commit phase 2a\n");
573
574         /*
575          * Now start flushing things to disk, in the order they appear
576          * on the transaction lists.  Data blocks go first.
577          */
578         err = journal_submit_data_buffers(journal, commit_transaction);
579         if (err)
580                 jbd2_journal_abort(journal, err);
581
582         blk_start_plug(&plug);
583         jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
584
585         jbd_debug(3, "JBD2: commit phase 2b\n");
586
587         /*
588          * Way to go: we have now written out all of the data for a
589          * transaction!  Now comes the tricky part: we need to write out
590          * metadata.  Loop over the transaction's entire buffer list:
591          */
592         write_lock(&journal->j_state_lock);
593         commit_transaction->t_state = T_COMMIT;
594         write_unlock(&journal->j_state_lock);
595
596         trace_jbd2_commit_logging(journal, commit_transaction);
597         stats.run.rs_logging = jiffies;
598         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
599                                                stats.run.rs_logging);
600         stats.run.rs_blocks = commit_transaction->t_nr_buffers;
601         stats.run.rs_blocks_logged = 0;
602
603         J_ASSERT(commit_transaction->t_nr_buffers <=
604                  atomic_read(&commit_transaction->t_outstanding_credits));
605
606         err = 0;
607         bufs = 0;
608         descriptor = NULL;
609         while (commit_transaction->t_buffers) {
610
611                 /* Find the next buffer to be journaled... */
612
613                 jh = commit_transaction->t_buffers;
614
615                 /* If we're in abort mode, we just un-journal the buffer and
616                    release it. */
617
618                 if (is_journal_aborted(journal)) {
619                         clear_buffer_jbddirty(jh2bh(jh));
620                         JBUFFER_TRACE(jh, "journal is aborting: refile");
621                         jbd2_buffer_abort_trigger(jh,
622                                                   jh->b_frozen_data ?
623                                                   jh->b_frozen_triggers :
624                                                   jh->b_triggers);
625                         jbd2_journal_refile_buffer(journal, jh);
626                         /* If that was the last one, we need to clean up
627                          * any descriptor buffers which may have been
628                          * already allocated, even if we are now
629                          * aborting. */
630                         if (!commit_transaction->t_buffers)
631                                 goto start_journal_io;
632                         continue;
633                 }
634
635                 /* Make sure we have a descriptor block in which to
636                    record the metadata buffer. */
637
638                 if (!descriptor) {
639                         J_ASSERT (bufs == 0);
640
641                         jbd_debug(4, "JBD2: get descriptor\n");
642
643                         descriptor = jbd2_journal_get_descriptor_buffer(
644                                                         commit_transaction,
645                                                         JBD2_DESCRIPTOR_BLOCK);
646                         if (!descriptor) {
647                                 jbd2_journal_abort(journal, -EIO);
648                                 continue;
649                         }
650
651                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
652                                 (unsigned long long)descriptor->b_blocknr,
653                                 descriptor->b_data);
654                         tagp = &descriptor->b_data[sizeof(journal_header_t)];
655                         space_left = descriptor->b_size -
656                                                 sizeof(journal_header_t);
657                         first_tag = 1;
658                         set_buffer_jwrite(descriptor);
659                         set_buffer_dirty(descriptor);
660                         wbuf[bufs++] = descriptor;
661
662                         /* Record it so that we can wait for IO
663                            completion later */
664                         BUFFER_TRACE(descriptor, "ph3: file as descriptor");
665                         jbd2_file_log_bh(&log_bufs, descriptor);
666                 }
667
668                 /* Where is the buffer to be written? */
669
670                 err = jbd2_journal_next_log_block(journal, &blocknr);
671                 /* If the block mapping failed, just abandon the buffer
672                    and repeat this loop: we'll fall into the
673                    refile-on-abort condition above. */
674                 if (err) {
675                         jbd2_journal_abort(journal, err);
676                         continue;
677                 }
678
679                 /*
680                  * start_this_handle() uses t_outstanding_credits to determine
681                  * the free space in the log.
682                  */
683                 atomic_dec(&commit_transaction->t_outstanding_credits);
684
685                 /* Bump b_count to prevent truncate from stumbling over
686                    the shadowed buffer!  @@@ This can go if we ever get
687                    rid of the shadow pairing of buffers. */
688                 atomic_inc(&jh2bh(jh)->b_count);
689
690                 /*
691                  * Make a temporary IO buffer with which to write it out
692                  * (this will requeue the metadata buffer to BJ_Shadow).
693                  */
694                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
695                 JBUFFER_TRACE(jh, "ph3: write metadata");
696                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
697                                                 jh, &wbuf[bufs], blocknr);
698                 if (flags < 0) {
699                         jbd2_journal_abort(journal, flags);
700                         continue;
701                 }
702                 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
703
704                 /* Record the new block's tag in the current descriptor
705                    buffer */
706
707                 tag_flag = 0;
708                 if (flags & 1)
709                         tag_flag |= JBD2_FLAG_ESCAPE;
710                 if (!first_tag)
711                         tag_flag |= JBD2_FLAG_SAME_UUID;
712
713                 tag = (journal_block_tag_t *) tagp;
714                 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
715                 tag->t_flags = cpu_to_be16(tag_flag);
716                 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
717                                         commit_transaction->t_tid);
718                 tagp += tag_bytes;
719                 space_left -= tag_bytes;
720                 bufs++;
721
722                 if (first_tag) {
723                         memcpy (tagp, journal->j_uuid, 16);
724                         tagp += 16;
725                         space_left -= 16;
726                         first_tag = 0;
727                 }
728
729                 /* If there's no more to do, or if the descriptor is full,
730                    let the IO rip! */
731
732                 if (bufs == journal->j_wbufsize ||
733                     commit_transaction->t_buffers == NULL ||
734                     space_left < tag_bytes + 16 + csum_size) {
735
736                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
737
738                         /* Write an end-of-descriptor marker before
739                            submitting the IOs.  "tag" still points to
740                            the last tag we set up. */
741
742                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
743 start_journal_io:
744                         if (descriptor)
745                                 jbd2_descriptor_block_csum_set(journal,
746                                                         descriptor);
747
748                         for (i = 0; i < bufs; i++) {
749                                 struct buffer_head *bh = wbuf[i];
750                                 /*
751                                  * Compute checksum.
752                                  */
753                                 if (jbd2_has_feature_checksum(journal)) {
754                                         crc32_sum =
755                                             jbd2_checksum_data(crc32_sum, bh);
756                                 }
757
758                                 lock_buffer(bh);
759                                 clear_buffer_dirty(bh);
760                                 set_buffer_uptodate(bh);
761                                 bh->b_end_io = journal_end_buffer_io_sync;
762                                 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
763                         }
764                         cond_resched();
765
766                         /* Force a new descriptor to be generated next
767                            time round the loop. */
768                         descriptor = NULL;
769                         bufs = 0;
770                 }
771         }
772
773         err = journal_finish_inode_data_buffers(journal, commit_transaction);
774         if (err) {
775                 printk(KERN_WARNING
776                         "JBD2: Detected IO errors while flushing file data "
777                        "on %s\n", journal->j_devname);
778                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
779                         jbd2_journal_abort(journal, err);
780                 err = 0;
781         }
782
783         /*
784          * Get current oldest transaction in the log before we issue flush
785          * to the filesystem device. After the flush we can be sure that
786          * blocks of all older transactions are checkpointed to persistent
787          * storage and we will be safe to update journal start in the
788          * superblock with the numbers we get here.
789          */
790         update_tail =
791                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
792
793         write_lock(&journal->j_state_lock);
794         if (update_tail) {
795                 long freed = first_block - journal->j_tail;
796
797                 if (first_block < journal->j_tail)
798                         freed += journal->j_last - journal->j_first;
799                 /* Update tail only if we free significant amount of space */
800                 if (freed < jbd2_journal_get_max_txn_bufs(journal))
801                         update_tail = 0;
802         }
803         J_ASSERT(commit_transaction->t_state == T_COMMIT);
804         commit_transaction->t_state = T_COMMIT_DFLUSH;
805         write_unlock(&journal->j_state_lock);
806
807         /*
808          * If the journal is not located on the file system device,
809          * then we must flush the file system device before we issue
810          * the commit record
811          */
812         if (commit_transaction->t_need_data_flush &&
813             (journal->j_fs_dev != journal->j_dev) &&
814             (journal->j_flags & JBD2_BARRIER))
815                 blkdev_issue_flush(journal->j_fs_dev);
816
817         /* Done it all: now write the commit record asynchronously. */
818         if (jbd2_has_feature_async_commit(journal)) {
819                 err = journal_submit_commit_record(journal, commit_transaction,
820                                                  &cbh, crc32_sum);
821                 if (err)
822                         jbd2_journal_abort(journal, err);
823         }
824
825         blk_finish_plug(&plug);
826
827         /* Lo and behold: we have just managed to send a transaction to
828            the log.  Before we can commit it, wait for the IO so far to
829            complete.  Control buffers being written are on the
830            transaction's t_log_list queue, and metadata buffers are on
831            the io_bufs list.
832
833            Wait for the buffers in reverse order.  That way we are
834            less likely to be woken up until all IOs have completed, and
835            so we incur less scheduling load.
836         */
837
838         jbd_debug(3, "JBD2: commit phase 3\n");
839
840         while (!list_empty(&io_bufs)) {
841                 struct buffer_head *bh = list_entry(io_bufs.prev,
842                                                     struct buffer_head,
843                                                     b_assoc_buffers);
844
845                 wait_on_buffer(bh);
846                 cond_resched();
847
848                 if (unlikely(!buffer_uptodate(bh)))
849                         err = -EIO;
850                 jbd2_unfile_log_bh(bh);
851                 stats.run.rs_blocks_logged++;
852
853                 /*
854                  * The list contains temporary buffer heads created by
855                  * jbd2_journal_write_metadata_buffer().
856                  */
857                 BUFFER_TRACE(bh, "dumping temporary bh");
858                 __brelse(bh);
859                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
860                 free_buffer_head(bh);
861
862                 /* We also have to refile the corresponding shadowed buffer */
863                 jh = commit_transaction->t_shadow_list->b_tprev;
864                 bh = jh2bh(jh);
865                 clear_buffer_jwrite(bh);
866                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
867                 J_ASSERT_BH(bh, !buffer_shadow(bh));
868
869                 /* The metadata is now released for reuse, but we need
870                    to remember it against this transaction so that when
871                    we finally commit, we can do any checkpointing
872                    required. */
873                 JBUFFER_TRACE(jh, "file as BJ_Forget");
874                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
875                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
876                 __brelse(bh);
877         }
878
879         J_ASSERT (commit_transaction->t_shadow_list == NULL);
880
881         jbd_debug(3, "JBD2: commit phase 4\n");
882
883         /* Here we wait for the revoke record and descriptor record buffers */
884         while (!list_empty(&log_bufs)) {
885                 struct buffer_head *bh;
886
887                 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
888                 wait_on_buffer(bh);
889                 cond_resched();
890
891                 if (unlikely(!buffer_uptodate(bh)))
892                         err = -EIO;
893
894                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
895                 clear_buffer_jwrite(bh);
896                 jbd2_unfile_log_bh(bh);
897                 stats.run.rs_blocks_logged++;
898                 __brelse(bh);           /* One for getblk */
899                 /* AKPM: bforget here */
900         }
901
902         if (err)
903                 jbd2_journal_abort(journal, err);
904
905         jbd_debug(3, "JBD2: commit phase 5\n");
906         write_lock(&journal->j_state_lock);
907         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
908         commit_transaction->t_state = T_COMMIT_JFLUSH;
909         write_unlock(&journal->j_state_lock);
910
911         if (!jbd2_has_feature_async_commit(journal)) {
912                 err = journal_submit_commit_record(journal, commit_transaction,
913                                                 &cbh, crc32_sum);
914                 if (err)
915                         jbd2_journal_abort(journal, err);
916         }
917         if (cbh)
918                 err = journal_wait_on_commit_record(journal, cbh);
919         stats.run.rs_blocks_logged++;
920         if (jbd2_has_feature_async_commit(journal) &&
921             journal->j_flags & JBD2_BARRIER) {
922                 blkdev_issue_flush(journal->j_dev);
923         }
924
925         if (err)
926                 jbd2_journal_abort(journal, err);
927
928         WARN_ON_ONCE(
929                 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
930
931         /*
932          * Now disk caches for filesystem device are flushed so we are safe to
933          * erase checkpointed transactions from the log by updating journal
934          * superblock.
935          */
936         if (update_tail)
937                 jbd2_update_log_tail(journal, first_tid, first_block);
938
939         /* End of a transaction!  Finally, we can do checkpoint
940            processing: any buffers committed as a result of this
941            transaction can be removed from any checkpoint list it was on
942            before. */
943
944         jbd_debug(3, "JBD2: commit phase 6\n");
945
946         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
947         J_ASSERT(commit_transaction->t_buffers == NULL);
948         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
949         J_ASSERT(commit_transaction->t_shadow_list == NULL);
950
951 restart_loop:
952         /*
953          * As there are other places (journal_unmap_buffer()) adding buffers
954          * to this list we have to be careful and hold the j_list_lock.
955          */
956         spin_lock(&journal->j_list_lock);
957         while (commit_transaction->t_forget) {
958                 transaction_t *cp_transaction;
959                 struct buffer_head *bh;
960                 int try_to_free = 0;
961                 bool drop_ref;
962
963                 jh = commit_transaction->t_forget;
964                 spin_unlock(&journal->j_list_lock);
965                 bh = jh2bh(jh);
966                 /*
967                  * Get a reference so that bh cannot be freed before we are
968                  * done with it.
969                  */
970                 get_bh(bh);
971                 spin_lock(&jh->b_state_lock);
972                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
973
974                 /*
975                  * If there is undo-protected committed data against
976                  * this buffer, then we can remove it now.  If it is a
977                  * buffer needing such protection, the old frozen_data
978                  * field now points to a committed version of the
979                  * buffer, so rotate that field to the new committed
980                  * data.
981                  *
982                  * Otherwise, we can just throw away the frozen data now.
983                  *
984                  * We also know that the frozen data has already fired
985                  * its triggers if they exist, so we can clear that too.
986                  */
987                 if (jh->b_committed_data) {
988                         jbd2_free(jh->b_committed_data, bh->b_size);
989                         jh->b_committed_data = NULL;
990                         if (jh->b_frozen_data) {
991                                 jh->b_committed_data = jh->b_frozen_data;
992                                 jh->b_frozen_data = NULL;
993                                 jh->b_frozen_triggers = NULL;
994                         }
995                 } else if (jh->b_frozen_data) {
996                         jbd2_free(jh->b_frozen_data, bh->b_size);
997                         jh->b_frozen_data = NULL;
998                         jh->b_frozen_triggers = NULL;
999                 }
1000
1001                 spin_lock(&journal->j_list_lock);
1002                 cp_transaction = jh->b_cp_transaction;
1003                 if (cp_transaction) {
1004                         JBUFFER_TRACE(jh, "remove from old cp transaction");
1005                         cp_transaction->t_chp_stats.cs_dropped++;
1006                         __jbd2_journal_remove_checkpoint(jh);
1007                 }
1008
1009                 /* Only re-checkpoint the buffer_head if it is marked
1010                  * dirty.  If the buffer was added to the BJ_Forget list
1011                  * by jbd2_journal_forget, it may no longer be dirty and
1012                  * there's no point in keeping a checkpoint record for
1013                  * it. */
1014
1015                 /*
1016                  * A buffer which has been freed while still being journaled
1017                  * by a previous transaction, refile the buffer to BJ_Forget of
1018                  * the running transaction. If the just committed transaction
1019                  * contains "add to orphan" operation, we can completely
1020                  * invalidate the buffer now. We are rather through in that
1021                  * since the buffer may be still accessible when blocksize <
1022                  * pagesize and it is attached to the last partial page.
1023                  */
1024                 if (buffer_freed(bh) && !jh->b_next_transaction) {
1025                         struct address_space *mapping;
1026
1027                         clear_buffer_freed(bh);
1028                         clear_buffer_jbddirty(bh);
1029
1030                         /*
1031                          * Block device buffers need to stay mapped all the
1032                          * time, so it is enough to clear buffer_jbddirty and
1033                          * buffer_freed bits. For the file mapping buffers (i.e.
1034                          * journalled data) we need to unmap buffer and clear
1035                          * more bits. We also need to be careful about the check
1036                          * because the data page mapping can get cleared under
1037                          * our hands. Note that if mapping == NULL, we don't
1038                          * need to make buffer unmapped because the page is
1039                          * already detached from the mapping and buffers cannot
1040                          * get reused.
1041                          */
1042                         mapping = READ_ONCE(bh->b_page->mapping);
1043                         if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1044                                 clear_buffer_mapped(bh);
1045                                 clear_buffer_new(bh);
1046                                 clear_buffer_req(bh);
1047                                 bh->b_bdev = NULL;
1048                         }
1049                 }
1050
1051                 if (buffer_jbddirty(bh)) {
1052                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1053                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1054                         if (is_journal_aborted(journal))
1055                                 clear_buffer_jbddirty(bh);
1056                 } else {
1057                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1058                         /*
1059                          * The buffer on BJ_Forget list and not jbddirty means
1060                          * it has been freed by this transaction and hence it
1061                          * could not have been reallocated until this
1062                          * transaction has committed. *BUT* it could be
1063                          * reallocated once we have written all the data to
1064                          * disk and before we process the buffer on BJ_Forget
1065                          * list.
1066                          */
1067                         if (!jh->b_next_transaction)
1068                                 try_to_free = 1;
1069                 }
1070                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1071                 drop_ref = __jbd2_journal_refile_buffer(jh);
1072                 spin_unlock(&jh->b_state_lock);
1073                 if (drop_ref)
1074                         jbd2_journal_put_journal_head(jh);
1075                 if (try_to_free)
1076                         release_buffer_page(bh);        /* Drops bh reference */
1077                 else
1078                         __brelse(bh);
1079                 cond_resched_lock(&journal->j_list_lock);
1080         }
1081         spin_unlock(&journal->j_list_lock);
1082         /*
1083          * This is a bit sleazy.  We use j_list_lock to protect transition
1084          * of a transaction into T_FINISHED state and calling
1085          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1086          * other checkpointing code processing the transaction...
1087          */
1088         write_lock(&journal->j_state_lock);
1089         spin_lock(&journal->j_list_lock);
1090         /*
1091          * Now recheck if some buffers did not get attached to the transaction
1092          * while the lock was dropped...
1093          */
1094         if (commit_transaction->t_forget) {
1095                 spin_unlock(&journal->j_list_lock);
1096                 write_unlock(&journal->j_state_lock);
1097                 goto restart_loop;
1098         }
1099
1100         /* Add the transaction to the checkpoint list
1101          * __journal_remove_checkpoint() can not destroy transaction
1102          * under us because it is not marked as T_FINISHED yet */
1103         if (journal->j_checkpoint_transactions == NULL) {
1104                 journal->j_checkpoint_transactions = commit_transaction;
1105                 commit_transaction->t_cpnext = commit_transaction;
1106                 commit_transaction->t_cpprev = commit_transaction;
1107         } else {
1108                 commit_transaction->t_cpnext =
1109                         journal->j_checkpoint_transactions;
1110                 commit_transaction->t_cpprev =
1111                         commit_transaction->t_cpnext->t_cpprev;
1112                 commit_transaction->t_cpnext->t_cpprev =
1113                         commit_transaction;
1114                 commit_transaction->t_cpprev->t_cpnext =
1115                                 commit_transaction;
1116         }
1117         spin_unlock(&journal->j_list_lock);
1118
1119         /* Done with this transaction! */
1120
1121         jbd_debug(3, "JBD2: commit phase 7\n");
1122
1123         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1124
1125         commit_transaction->t_start = jiffies;
1126         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1127                                               commit_transaction->t_start);
1128
1129         /*
1130          * File the transaction statistics
1131          */
1132         stats.ts_tid = commit_transaction->t_tid;
1133         stats.run.rs_handle_count =
1134                 atomic_read(&commit_transaction->t_handle_count);
1135         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1136                              commit_transaction->t_tid, &stats.run);
1137         stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1138
1139         commit_transaction->t_state = T_COMMIT_CALLBACK;
1140         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1141         journal->j_commit_sequence = commit_transaction->t_tid;
1142         journal->j_committing_transaction = NULL;
1143         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1144
1145         /*
1146          * weight the commit time higher than the average time so we don't
1147          * react too strongly to vast changes in the commit time
1148          */
1149         if (likely(journal->j_average_commit_time))
1150                 journal->j_average_commit_time = (commit_time +
1151                                 journal->j_average_commit_time*3) / 4;
1152         else
1153                 journal->j_average_commit_time = commit_time;
1154
1155         write_unlock(&journal->j_state_lock);
1156
1157         if (journal->j_commit_callback)
1158                 journal->j_commit_callback(journal, commit_transaction);
1159         if (journal->j_fc_cleanup_callback)
1160                 journal->j_fc_cleanup_callback(journal, 1, commit_transaction->t_tid);
1161
1162         trace_jbd2_end_commit(journal, commit_transaction);
1163         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1164                   journal->j_commit_sequence, journal->j_tail_sequence);
1165
1166         write_lock(&journal->j_state_lock);
1167         journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1168         journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1169         spin_lock(&journal->j_list_lock);
1170         commit_transaction->t_state = T_FINISHED;
1171         /* Check if the transaction can be dropped now that we are finished */
1172         if (commit_transaction->t_checkpoint_list == NULL &&
1173             commit_transaction->t_checkpoint_io_list == NULL) {
1174                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1175                 jbd2_journal_free_transaction(commit_transaction);
1176         }
1177         spin_unlock(&journal->j_list_lock);
1178         write_unlock(&journal->j_state_lock);
1179         wake_up(&journal->j_wait_done_commit);
1180         wake_up(&journal->j_fc_wait);
1181
1182         /*
1183          * Calculate overall stats
1184          */
1185         spin_lock(&journal->j_history_lock);
1186         journal->j_stats.ts_tid++;
1187         journal->j_stats.ts_requested += stats.ts_requested;
1188         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1189         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1190         journal->j_stats.run.rs_running += stats.run.rs_running;
1191         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1192         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1193         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1194         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1195         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1196         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1197         spin_unlock(&journal->j_history_lock);
1198 }