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 */
journal_end_buffer_io_sync(struct buffer_head * bh,int uptodate)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 */
release_buffer_page(struct buffer_head * bh)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
jbd2_commit_block_csum_set(journal_t * j,struct buffer_head * bh)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 */
journal_submit_commit_record(journal_t * journal,transaction_t * commit_transaction,struct buffer_head ** cbh,__u32 crc32_sum)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 */
journal_wait_on_commit_record(journal_t * journal,struct buffer_head * bh)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 */
jbd2_journal_submit_inode_data_buffers(struct jbd2_inode * jinode)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 */
jbd2_submit_inode_data(struct jbd2_inode * jinode)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
jbd2_wait_inode_data(journal_t * journal,struct jbd2_inode * jinode)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 */
journal_submit_data_buffers(journal_t * journal,transaction_t * commit_transaction)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
jbd2_journal_finish_inode_data_buffers(struct jbd2_inode * jinode)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 */
journal_finish_inode_data_buffers(journal_t * journal,transaction_t * commit_transaction)284 static int journal_finish_inode_data_buffers(journal_t *journal,
285 transaction_t *commit_transaction)
286 {
287 struct jbd2_inode *jinode, *next_i;
288 int err, ret = 0;
289
290 /* For locking, see the comment in journal_submit_data_buffers() */
291 spin_lock(&journal->j_list_lock);
292 list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
293 if (!(jinode->i_flags & JI_WAIT_DATA))
294 continue;
295 jinode->i_flags |= JI_COMMIT_RUNNING;
296 spin_unlock(&journal->j_list_lock);
297 /* wait for the inode data buffers writeout. */
298 if (journal->j_finish_inode_data_buffers) {
299 err = journal->j_finish_inode_data_buffers(jinode);
300 if (!ret)
301 ret = err;
302 }
303 cond_resched();
304 spin_lock(&journal->j_list_lock);
305 jinode->i_flags &= ~JI_COMMIT_RUNNING;
306 smp_mb();
307 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
308 }
309
310 /* Now refile inode to proper lists */
311 list_for_each_entry_safe(jinode, next_i,
312 &commit_transaction->t_inode_list, i_list) {
313 list_del(&jinode->i_list);
314 if (jinode->i_next_transaction) {
315 jinode->i_transaction = jinode->i_next_transaction;
316 jinode->i_next_transaction = NULL;
317 list_add(&jinode->i_list,
318 &jinode->i_transaction->t_inode_list);
319 } else {
320 jinode->i_transaction = NULL;
321 jinode->i_dirty_start = 0;
322 jinode->i_dirty_end = 0;
323 }
324 }
325 spin_unlock(&journal->j_list_lock);
326
327 return ret;
328 }
329
jbd2_checksum_data(__u32 crc32_sum,struct buffer_head * bh)330 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
331 {
332 struct page *page = bh->b_page;
333 char *addr;
334 __u32 checksum;
335
336 addr = kmap_atomic(page);
337 checksum = crc32_be(crc32_sum,
338 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
339 kunmap_atomic(addr);
340
341 return checksum;
342 }
343
write_tag_block(journal_t * j,journal_block_tag_t * tag,unsigned long long block)344 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
345 unsigned long long block)
346 {
347 tag->t_blocknr = cpu_to_be32(block & (u32)~0);
348 if (jbd2_has_feature_64bit(j))
349 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
350 }
351
jbd2_block_tag_csum_set(journal_t * j,journal_block_tag_t * tag,struct buffer_head * bh,__u32 sequence)352 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
353 struct buffer_head *bh, __u32 sequence)
354 {
355 journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
356 struct page *page = bh->b_page;
357 __u8 *addr;
358 __u32 csum32;
359 __be32 seq;
360
361 if (!jbd2_journal_has_csum_v2or3(j))
362 return;
363
364 seq = cpu_to_be32(sequence);
365 addr = kmap_atomic(page);
366 csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
367 csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
368 bh->b_size);
369 kunmap_atomic(addr);
370
371 if (jbd2_has_feature_csum3(j))
372 tag3->t_checksum = cpu_to_be32(csum32);
373 else
374 tag->t_checksum = cpu_to_be16(csum32);
375 }
376 /*
377 * jbd2_journal_commit_transaction
378 *
379 * The primary function for committing a transaction to the log. This
380 * function is called by the journal thread to begin a complete commit.
381 */
jbd2_journal_commit_transaction(journal_t * journal)382 void jbd2_journal_commit_transaction(journal_t *journal)
383 {
384 struct transaction_stats_s stats;
385 transaction_t *commit_transaction;
386 struct journal_head *jh;
387 struct buffer_head *descriptor;
388 struct buffer_head **wbuf = journal->j_wbuf;
389 int bufs;
390 int flags;
391 int err;
392 unsigned long long blocknr;
393 ktime_t start_time;
394 u64 commit_time;
395 char *tagp = NULL;
396 journal_block_tag_t *tag = NULL;
397 int space_left = 0;
398 int first_tag = 0;
399 int tag_flag;
400 int i;
401 int tag_bytes = journal_tag_bytes(journal);
402 struct buffer_head *cbh = NULL; /* For transactional checksums */
403 __u32 crc32_sum = ~0;
404 struct blk_plug plug;
405 /* Tail of the journal */
406 unsigned long first_block;
407 tid_t first_tid;
408 int update_tail;
409 int csum_size = 0;
410 LIST_HEAD(io_bufs);
411 LIST_HEAD(log_bufs);
412
413 if (jbd2_journal_has_csum_v2or3(journal))
414 csum_size = sizeof(struct jbd2_journal_block_tail);
415
416 /*
417 * First job: lock down the current transaction and wait for
418 * all outstanding updates to complete.
419 */
420
421 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
422 if (journal->j_flags & JBD2_FLUSHED) {
423 jbd_debug(3, "super block updated\n");
424 mutex_lock_io(&journal->j_checkpoint_mutex);
425 /*
426 * We hold j_checkpoint_mutex so tail cannot change under us.
427 * We don't need any special data guarantees for writing sb
428 * since journal is empty and it is ok for write to be
429 * flushed only with transaction commit.
430 */
431 jbd2_journal_update_sb_log_tail(journal,
432 journal->j_tail_sequence,
433 journal->j_tail,
434 REQ_SYNC);
435 mutex_unlock(&journal->j_checkpoint_mutex);
436 } else {
437 jbd_debug(3, "superblock not updated\n");
438 }
439
440 J_ASSERT(journal->j_running_transaction != NULL);
441 J_ASSERT(journal->j_committing_transaction == NULL);
442
443 write_lock(&journal->j_state_lock);
444 journal->j_flags |= JBD2_FULL_COMMIT_ONGOING;
445 while (journal->j_flags & JBD2_FAST_COMMIT_ONGOING) {
446 DEFINE_WAIT(wait);
447
448 prepare_to_wait(&journal->j_fc_wait, &wait,
449 TASK_UNINTERRUPTIBLE);
450 write_unlock(&journal->j_state_lock);
451 schedule();
452 write_lock(&journal->j_state_lock);
453 finish_wait(&journal->j_fc_wait, &wait);
454 /*
455 * TODO: by blocking fast commits here, we are increasing
456 * fsync() latency slightly. Strictly speaking, we don't need
457 * to block fast commits until the transaction enters T_FLUSH
458 * state. So an optimization is possible where we block new fast
459 * commits here and wait for existing ones to complete
460 * just before we enter T_FLUSH. That way, the existing fast
461 * commits and this full commit can proceed parallely.
462 */
463 }
464 write_unlock(&journal->j_state_lock);
465
466 commit_transaction = journal->j_running_transaction;
467
468 trace_jbd2_start_commit(journal, commit_transaction);
469 jbd_debug(1, "JBD2: starting commit of transaction %d\n",
470 commit_transaction->t_tid);
471
472 write_lock(&journal->j_state_lock);
473 journal->j_fc_off = 0;
474 J_ASSERT(commit_transaction->t_state == T_RUNNING);
475 commit_transaction->t_state = T_LOCKED;
476
477 trace_jbd2_commit_locking(journal, commit_transaction);
478 stats.run.rs_wait = commit_transaction->t_max_wait;
479 stats.run.rs_request_delay = 0;
480 stats.run.rs_locked = jiffies;
481 if (commit_transaction->t_requested)
482 stats.run.rs_request_delay =
483 jbd2_time_diff(commit_transaction->t_requested,
484 stats.run.rs_locked);
485 stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
486 stats.run.rs_locked);
487
488 spin_lock(&commit_transaction->t_handle_lock);
489 while (atomic_read(&commit_transaction->t_updates)) {
490 DEFINE_WAIT(wait);
491
492 prepare_to_wait(&journal->j_wait_updates, &wait,
493 TASK_UNINTERRUPTIBLE);
494 if (atomic_read(&commit_transaction->t_updates)) {
495 spin_unlock(&commit_transaction->t_handle_lock);
496 write_unlock(&journal->j_state_lock);
497 schedule();
498 write_lock(&journal->j_state_lock);
499 spin_lock(&commit_transaction->t_handle_lock);
500 }
501 finish_wait(&journal->j_wait_updates, &wait);
502 }
503 spin_unlock(&commit_transaction->t_handle_lock);
504 commit_transaction->t_state = T_SWITCH;
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 * We use journal->j_state_lock here to serialize processing of
525 * t_reserved_list with eviction of buffers from journal_unmap_buffer().
526 */
527 while (commit_transaction->t_reserved_list) {
528 jh = commit_transaction->t_reserved_list;
529 JBUFFER_TRACE(jh, "reserved, unused: refile");
530 /*
531 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
532 * leave undo-committed data.
533 */
534 if (jh->b_committed_data) {
535 struct buffer_head *bh = jh2bh(jh);
536
537 spin_lock(&jh->b_state_lock);
538 jbd2_free(jh->b_committed_data, bh->b_size);
539 jh->b_committed_data = NULL;
540 spin_unlock(&jh->b_state_lock);
541 }
542 jbd2_journal_refile_buffer(journal, jh);
543 }
544
545 write_unlock(&journal->j_state_lock);
546 /*
547 * Now try to drop any written-back buffers from the journal's
548 * checkpoint lists. We do this *before* commit because it potentially
549 * frees some memory
550 */
551 spin_lock(&journal->j_list_lock);
552 __jbd2_journal_clean_checkpoint_list(journal, false);
553 spin_unlock(&journal->j_list_lock);
554
555 jbd_debug(3, "JBD2: commit phase 1\n");
556
557 /*
558 * Clear revoked flag to reflect there is no revoked buffers
559 * in the next transaction which is going to be started.
560 */
561 jbd2_clear_buffer_revoked_flags(journal);
562
563 /*
564 * Switch to a new revoke table.
565 */
566 jbd2_journal_switch_revoke_table(journal);
567
568 write_lock(&journal->j_state_lock);
569 /*
570 * Reserved credits cannot be claimed anymore, free them
571 */
572 atomic_sub(atomic_read(&journal->j_reserved_credits),
573 &commit_transaction->t_outstanding_credits);
574
575 trace_jbd2_commit_flushing(journal, commit_transaction);
576 stats.run.rs_flushing = jiffies;
577 stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
578 stats.run.rs_flushing);
579
580 commit_transaction->t_state = T_FLUSH;
581 journal->j_committing_transaction = commit_transaction;
582 journal->j_running_transaction = NULL;
583 start_time = ktime_get();
584 commit_transaction->t_log_start = journal->j_head;
585 wake_up_all(&journal->j_wait_transaction_locked);
586 write_unlock(&journal->j_state_lock);
587
588 jbd_debug(3, "JBD2: commit phase 2a\n");
589
590 /*
591 * Now start flushing things to disk, in the order they appear
592 * on the transaction lists. Data blocks go first.
593 */
594 err = journal_submit_data_buffers(journal, commit_transaction);
595 if (err)
596 jbd2_journal_abort(journal, err);
597
598 blk_start_plug(&plug);
599 jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
600
601 jbd_debug(3, "JBD2: commit phase 2b\n");
602
603 /*
604 * Way to go: we have now written out all of the data for a
605 * transaction! Now comes the tricky part: we need to write out
606 * metadata. Loop over the transaction's entire buffer list:
607 */
608 write_lock(&journal->j_state_lock);
609 commit_transaction->t_state = T_COMMIT;
610 write_unlock(&journal->j_state_lock);
611
612 trace_jbd2_commit_logging(journal, commit_transaction);
613 stats.run.rs_logging = jiffies;
614 stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
615 stats.run.rs_logging);
616 stats.run.rs_blocks = commit_transaction->t_nr_buffers;
617 stats.run.rs_blocks_logged = 0;
618
619 J_ASSERT(commit_transaction->t_nr_buffers <=
620 atomic_read(&commit_transaction->t_outstanding_credits));
621
622 err = 0;
623 bufs = 0;
624 descriptor = NULL;
625 while (commit_transaction->t_buffers) {
626
627 /* Find the next buffer to be journaled... */
628
629 jh = commit_transaction->t_buffers;
630
631 /* If we're in abort mode, we just un-journal the buffer and
632 release it. */
633
634 if (is_journal_aborted(journal)) {
635 clear_buffer_jbddirty(jh2bh(jh));
636 JBUFFER_TRACE(jh, "journal is aborting: refile");
637 jbd2_buffer_abort_trigger(jh,
638 jh->b_frozen_data ?
639 jh->b_frozen_triggers :
640 jh->b_triggers);
641 jbd2_journal_refile_buffer(journal, jh);
642 /* If that was the last one, we need to clean up
643 * any descriptor buffers which may have been
644 * already allocated, even if we are now
645 * aborting. */
646 if (!commit_transaction->t_buffers)
647 goto start_journal_io;
648 continue;
649 }
650
651 /* Make sure we have a descriptor block in which to
652 record the metadata buffer. */
653
654 if (!descriptor) {
655 J_ASSERT (bufs == 0);
656
657 jbd_debug(4, "JBD2: get descriptor\n");
658
659 descriptor = jbd2_journal_get_descriptor_buffer(
660 commit_transaction,
661 JBD2_DESCRIPTOR_BLOCK);
662 if (!descriptor) {
663 jbd2_journal_abort(journal, -EIO);
664 continue;
665 }
666
667 jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
668 (unsigned long long)descriptor->b_blocknr,
669 descriptor->b_data);
670 tagp = &descriptor->b_data[sizeof(journal_header_t)];
671 space_left = descriptor->b_size -
672 sizeof(journal_header_t);
673 first_tag = 1;
674 set_buffer_jwrite(descriptor);
675 set_buffer_dirty(descriptor);
676 wbuf[bufs++] = descriptor;
677
678 /* Record it so that we can wait for IO
679 completion later */
680 BUFFER_TRACE(descriptor, "ph3: file as descriptor");
681 jbd2_file_log_bh(&log_bufs, descriptor);
682 }
683
684 /* Where is the buffer to be written? */
685
686 err = jbd2_journal_next_log_block(journal, &blocknr);
687 /* If the block mapping failed, just abandon the buffer
688 and repeat this loop: we'll fall into the
689 refile-on-abort condition above. */
690 if (err) {
691 jbd2_journal_abort(journal, err);
692 continue;
693 }
694
695 /*
696 * start_this_handle() uses t_outstanding_credits to determine
697 * the free space in the log.
698 */
699 atomic_dec(&commit_transaction->t_outstanding_credits);
700
701 /* Bump b_count to prevent truncate from stumbling over
702 the shadowed buffer! @@@ This can go if we ever get
703 rid of the shadow pairing of buffers. */
704 atomic_inc(&jh2bh(jh)->b_count);
705
706 /*
707 * Make a temporary IO buffer with which to write it out
708 * (this will requeue the metadata buffer to BJ_Shadow).
709 */
710 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
711 JBUFFER_TRACE(jh, "ph3: write metadata");
712 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
713 jh, &wbuf[bufs], blocknr);
714 if (flags < 0) {
715 jbd2_journal_abort(journal, flags);
716 continue;
717 }
718 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
719
720 /* Record the new block's tag in the current descriptor
721 buffer */
722
723 tag_flag = 0;
724 if (flags & 1)
725 tag_flag |= JBD2_FLAG_ESCAPE;
726 if (!first_tag)
727 tag_flag |= JBD2_FLAG_SAME_UUID;
728
729 tag = (journal_block_tag_t *) tagp;
730 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
731 tag->t_flags = cpu_to_be16(tag_flag);
732 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
733 commit_transaction->t_tid);
734 tagp += tag_bytes;
735 space_left -= tag_bytes;
736 bufs++;
737
738 if (first_tag) {
739 memcpy (tagp, journal->j_uuid, 16);
740 tagp += 16;
741 space_left -= 16;
742 first_tag = 0;
743 }
744
745 /* If there's no more to do, or if the descriptor is full,
746 let the IO rip! */
747
748 if (bufs == journal->j_wbufsize ||
749 commit_transaction->t_buffers == NULL ||
750 space_left < tag_bytes + 16 + csum_size) {
751
752 jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
753
754 /* Write an end-of-descriptor marker before
755 submitting the IOs. "tag" still points to
756 the last tag we set up. */
757
758 tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
759 start_journal_io:
760 if (descriptor)
761 jbd2_descriptor_block_csum_set(journal,
762 descriptor);
763
764 for (i = 0; i < bufs; i++) {
765 struct buffer_head *bh = wbuf[i];
766 /*
767 * Compute checksum.
768 */
769 if (jbd2_has_feature_checksum(journal)) {
770 crc32_sum =
771 jbd2_checksum_data(crc32_sum, bh);
772 }
773
774 lock_buffer(bh);
775 clear_buffer_dirty(bh);
776 set_buffer_uptodate(bh);
777 bh->b_end_io = journal_end_buffer_io_sync;
778 submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
779 }
780 cond_resched();
781
782 /* Force a new descriptor to be generated next
783 time round the loop. */
784 descriptor = NULL;
785 bufs = 0;
786 }
787 }
788
789 err = journal_finish_inode_data_buffers(journal, commit_transaction);
790 if (err) {
791 printk(KERN_WARNING
792 "JBD2: Detected IO errors while flushing file data "
793 "on %s\n", journal->j_devname);
794 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
795 jbd2_journal_abort(journal, err);
796 err = 0;
797 }
798
799 /*
800 * Get current oldest transaction in the log before we issue flush
801 * to the filesystem device. After the flush we can be sure that
802 * blocks of all older transactions are checkpointed to persistent
803 * storage and we will be safe to update journal start in the
804 * superblock with the numbers we get here.
805 */
806 update_tail =
807 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
808
809 write_lock(&journal->j_state_lock);
810 if (update_tail) {
811 long freed = first_block - journal->j_tail;
812
813 if (first_block < journal->j_tail)
814 freed += journal->j_last - journal->j_first;
815 /* Update tail only if we free significant amount of space */
816 if (freed < jbd2_journal_get_max_txn_bufs(journal))
817 update_tail = 0;
818 }
819 J_ASSERT(commit_transaction->t_state == T_COMMIT);
820 commit_transaction->t_state = T_COMMIT_DFLUSH;
821 write_unlock(&journal->j_state_lock);
822
823 /*
824 * If the journal is not located on the file system device,
825 * then we must flush the file system device before we issue
826 * the commit record
827 */
828 if (commit_transaction->t_need_data_flush &&
829 (journal->j_fs_dev != journal->j_dev) &&
830 (journal->j_flags & JBD2_BARRIER))
831 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS);
832
833 /* Done it all: now write the commit record asynchronously. */
834 if (jbd2_has_feature_async_commit(journal)) {
835 err = journal_submit_commit_record(journal, commit_transaction,
836 &cbh, crc32_sum);
837 if (err)
838 jbd2_journal_abort(journal, err);
839 }
840
841 blk_finish_plug(&plug);
842
843 /* Lo and behold: we have just managed to send a transaction to
844 the log. Before we can commit it, wait for the IO so far to
845 complete. Control buffers being written are on the
846 transaction's t_log_list queue, and metadata buffers are on
847 the io_bufs list.
848
849 Wait for the buffers in reverse order. That way we are
850 less likely to be woken up until all IOs have completed, and
851 so we incur less scheduling load.
852 */
853
854 jbd_debug(3, "JBD2: commit phase 3\n");
855
856 while (!list_empty(&io_bufs)) {
857 struct buffer_head *bh = list_entry(io_bufs.prev,
858 struct buffer_head,
859 b_assoc_buffers);
860
861 wait_on_buffer(bh);
862 cond_resched();
863
864 if (unlikely(!buffer_uptodate(bh)))
865 err = -EIO;
866 jbd2_unfile_log_bh(bh);
867 stats.run.rs_blocks_logged++;
868
869 /*
870 * The list contains temporary buffer heads created by
871 * jbd2_journal_write_metadata_buffer().
872 */
873 BUFFER_TRACE(bh, "dumping temporary bh");
874 __brelse(bh);
875 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
876 free_buffer_head(bh);
877
878 /* We also have to refile the corresponding shadowed buffer */
879 jh = commit_transaction->t_shadow_list->b_tprev;
880 bh = jh2bh(jh);
881 clear_buffer_jwrite(bh);
882 J_ASSERT_BH(bh, buffer_jbddirty(bh));
883 J_ASSERT_BH(bh, !buffer_shadow(bh));
884
885 /* The metadata is now released for reuse, but we need
886 to remember it against this transaction so that when
887 we finally commit, we can do any checkpointing
888 required. */
889 JBUFFER_TRACE(jh, "file as BJ_Forget");
890 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
891 JBUFFER_TRACE(jh, "brelse shadowed buffer");
892 __brelse(bh);
893 }
894
895 J_ASSERT (commit_transaction->t_shadow_list == NULL);
896
897 jbd_debug(3, "JBD2: commit phase 4\n");
898
899 /* Here we wait for the revoke record and descriptor record buffers */
900 while (!list_empty(&log_bufs)) {
901 struct buffer_head *bh;
902
903 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
904 wait_on_buffer(bh);
905 cond_resched();
906
907 if (unlikely(!buffer_uptodate(bh)))
908 err = -EIO;
909
910 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
911 clear_buffer_jwrite(bh);
912 jbd2_unfile_log_bh(bh);
913 stats.run.rs_blocks_logged++;
914 __brelse(bh); /* One for getblk */
915 /* AKPM: bforget here */
916 }
917
918 if (err)
919 jbd2_journal_abort(journal, err);
920
921 jbd_debug(3, "JBD2: commit phase 5\n");
922 write_lock(&journal->j_state_lock);
923 J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
924 commit_transaction->t_state = T_COMMIT_JFLUSH;
925 write_unlock(&journal->j_state_lock);
926
927 if (!jbd2_has_feature_async_commit(journal)) {
928 err = journal_submit_commit_record(journal, commit_transaction,
929 &cbh, crc32_sum);
930 if (err)
931 jbd2_journal_abort(journal, err);
932 }
933 if (cbh)
934 err = journal_wait_on_commit_record(journal, cbh);
935 stats.run.rs_blocks_logged++;
936 if (jbd2_has_feature_async_commit(journal) &&
937 journal->j_flags & JBD2_BARRIER) {
938 blkdev_issue_flush(journal->j_dev, GFP_NOFS);
939 }
940
941 if (err)
942 jbd2_journal_abort(journal, err);
943
944 WARN_ON_ONCE(
945 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
946
947 /*
948 * Now disk caches for filesystem device are flushed so we are safe to
949 * erase checkpointed transactions from the log by updating journal
950 * superblock.
951 */
952 if (update_tail)
953 jbd2_update_log_tail(journal, first_tid, first_block);
954
955 /* End of a transaction! Finally, we can do checkpoint
956 processing: any buffers committed as a result of this
957 transaction can be removed from any checkpoint list it was on
958 before. */
959
960 jbd_debug(3, "JBD2: commit phase 6\n");
961
962 J_ASSERT(list_empty(&commit_transaction->t_inode_list));
963 J_ASSERT(commit_transaction->t_buffers == NULL);
964 J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
965 J_ASSERT(commit_transaction->t_shadow_list == NULL);
966
967 restart_loop:
968 /*
969 * As there are other places (journal_unmap_buffer()) adding buffers
970 * to this list we have to be careful and hold the j_list_lock.
971 */
972 spin_lock(&journal->j_list_lock);
973 while (commit_transaction->t_forget) {
974 transaction_t *cp_transaction;
975 struct buffer_head *bh;
976 int try_to_free = 0;
977 bool drop_ref;
978
979 jh = commit_transaction->t_forget;
980 spin_unlock(&journal->j_list_lock);
981 bh = jh2bh(jh);
982 /*
983 * Get a reference so that bh cannot be freed before we are
984 * done with it.
985 */
986 get_bh(bh);
987 spin_lock(&jh->b_state_lock);
988 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
989
990 /*
991 * If there is undo-protected committed data against
992 * this buffer, then we can remove it now. If it is a
993 * buffer needing such protection, the old frozen_data
994 * field now points to a committed version of the
995 * buffer, so rotate that field to the new committed
996 * data.
997 *
998 * Otherwise, we can just throw away the frozen data now.
999 *
1000 * We also know that the frozen data has already fired
1001 * its triggers if they exist, so we can clear that too.
1002 */
1003 if (jh->b_committed_data) {
1004 jbd2_free(jh->b_committed_data, bh->b_size);
1005 jh->b_committed_data = NULL;
1006 if (jh->b_frozen_data) {
1007 jh->b_committed_data = jh->b_frozen_data;
1008 jh->b_frozen_data = NULL;
1009 jh->b_frozen_triggers = NULL;
1010 }
1011 } else if (jh->b_frozen_data) {
1012 jbd2_free(jh->b_frozen_data, bh->b_size);
1013 jh->b_frozen_data = NULL;
1014 jh->b_frozen_triggers = NULL;
1015 }
1016
1017 spin_lock(&journal->j_list_lock);
1018 cp_transaction = jh->b_cp_transaction;
1019 if (cp_transaction) {
1020 JBUFFER_TRACE(jh, "remove from old cp transaction");
1021 cp_transaction->t_chp_stats.cs_dropped++;
1022 __jbd2_journal_remove_checkpoint(jh);
1023 }
1024
1025 /* Only re-checkpoint the buffer_head if it is marked
1026 * dirty. If the buffer was added to the BJ_Forget list
1027 * by jbd2_journal_forget, it may no longer be dirty and
1028 * there's no point in keeping a checkpoint record for
1029 * it. */
1030
1031 /*
1032 * A buffer which has been freed while still being journaled
1033 * by a previous transaction, refile the buffer to BJ_Forget of
1034 * the running transaction. If the just committed transaction
1035 * contains "add to orphan" operation, we can completely
1036 * invalidate the buffer now. We are rather through in that
1037 * since the buffer may be still accessible when blocksize <
1038 * pagesize and it is attached to the last partial page.
1039 */
1040 if (buffer_freed(bh) && !jh->b_next_transaction) {
1041 struct address_space *mapping;
1042
1043 clear_buffer_freed(bh);
1044 clear_buffer_jbddirty(bh);
1045
1046 /*
1047 * Block device buffers need to stay mapped all the
1048 * time, so it is enough to clear buffer_jbddirty and
1049 * buffer_freed bits. For the file mapping buffers (i.e.
1050 * journalled data) we need to unmap buffer and clear
1051 * more bits. We also need to be careful about the check
1052 * because the data page mapping can get cleared under
1053 * our hands. Note that if mapping == NULL, we don't
1054 * need to make buffer unmapped because the page is
1055 * already detached from the mapping and buffers cannot
1056 * get reused.
1057 */
1058 mapping = READ_ONCE(bh->b_page->mapping);
1059 if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1060 clear_buffer_mapped(bh);
1061 clear_buffer_new(bh);
1062 clear_buffer_req(bh);
1063 bh->b_bdev = NULL;
1064 }
1065 }
1066
1067 if (buffer_jbddirty(bh)) {
1068 JBUFFER_TRACE(jh, "add to new checkpointing trans");
1069 __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1070 if (is_journal_aborted(journal))
1071 clear_buffer_jbddirty(bh);
1072 } else {
1073 J_ASSERT_BH(bh, !buffer_dirty(bh));
1074 /*
1075 * The buffer on BJ_Forget list and not jbddirty means
1076 * it has been freed by this transaction and hence it
1077 * could not have been reallocated until this
1078 * transaction has committed. *BUT* it could be
1079 * reallocated once we have written all the data to
1080 * disk and before we process the buffer on BJ_Forget
1081 * list.
1082 */
1083 if (!jh->b_next_transaction)
1084 try_to_free = 1;
1085 }
1086 JBUFFER_TRACE(jh, "refile or unfile buffer");
1087 drop_ref = __jbd2_journal_refile_buffer(jh);
1088 spin_unlock(&jh->b_state_lock);
1089 if (drop_ref)
1090 jbd2_journal_put_journal_head(jh);
1091 if (try_to_free)
1092 release_buffer_page(bh); /* Drops bh reference */
1093 else
1094 __brelse(bh);
1095 cond_resched_lock(&journal->j_list_lock);
1096 }
1097 spin_unlock(&journal->j_list_lock);
1098 /*
1099 * This is a bit sleazy. We use j_list_lock to protect transition
1100 * of a transaction into T_FINISHED state and calling
1101 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1102 * other checkpointing code processing the transaction...
1103 */
1104 write_lock(&journal->j_state_lock);
1105 spin_lock(&journal->j_list_lock);
1106 /*
1107 * Now recheck if some buffers did not get attached to the transaction
1108 * while the lock was dropped...
1109 */
1110 if (commit_transaction->t_forget) {
1111 spin_unlock(&journal->j_list_lock);
1112 write_unlock(&journal->j_state_lock);
1113 goto restart_loop;
1114 }
1115
1116 /* Add the transaction to the checkpoint list
1117 * __journal_remove_checkpoint() can not destroy transaction
1118 * under us because it is not marked as T_FINISHED yet */
1119 if (journal->j_checkpoint_transactions == NULL) {
1120 journal->j_checkpoint_transactions = commit_transaction;
1121 commit_transaction->t_cpnext = commit_transaction;
1122 commit_transaction->t_cpprev = commit_transaction;
1123 } else {
1124 commit_transaction->t_cpnext =
1125 journal->j_checkpoint_transactions;
1126 commit_transaction->t_cpprev =
1127 commit_transaction->t_cpnext->t_cpprev;
1128 commit_transaction->t_cpnext->t_cpprev =
1129 commit_transaction;
1130 commit_transaction->t_cpprev->t_cpnext =
1131 commit_transaction;
1132 }
1133 spin_unlock(&journal->j_list_lock);
1134
1135 /* Done with this transaction! */
1136
1137 jbd_debug(3, "JBD2: commit phase 7\n");
1138
1139 J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1140
1141 commit_transaction->t_start = jiffies;
1142 stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1143 commit_transaction->t_start);
1144
1145 /*
1146 * File the transaction statistics
1147 */
1148 stats.ts_tid = commit_transaction->t_tid;
1149 stats.run.rs_handle_count =
1150 atomic_read(&commit_transaction->t_handle_count);
1151 trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1152 commit_transaction->t_tid, &stats.run);
1153 stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1154
1155 commit_transaction->t_state = T_COMMIT_CALLBACK;
1156 J_ASSERT(commit_transaction == journal->j_committing_transaction);
1157 journal->j_commit_sequence = commit_transaction->t_tid;
1158 journal->j_committing_transaction = NULL;
1159 commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1160
1161 /*
1162 * weight the commit time higher than the average time so we don't
1163 * react too strongly to vast changes in the commit time
1164 */
1165 if (likely(journal->j_average_commit_time))
1166 journal->j_average_commit_time = (commit_time +
1167 journal->j_average_commit_time*3) / 4;
1168 else
1169 journal->j_average_commit_time = commit_time;
1170
1171 write_unlock(&journal->j_state_lock);
1172
1173 if (journal->j_commit_callback)
1174 journal->j_commit_callback(journal, commit_transaction);
1175 if (journal->j_fc_cleanup_callback)
1176 journal->j_fc_cleanup_callback(journal, 1);
1177
1178 trace_jbd2_end_commit(journal, commit_transaction);
1179 jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1180 journal->j_commit_sequence, journal->j_tail_sequence);
1181
1182 write_lock(&journal->j_state_lock);
1183 journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1184 journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1185 spin_lock(&journal->j_list_lock);
1186 commit_transaction->t_state = T_FINISHED;
1187 /* Check if the transaction can be dropped now that we are finished */
1188 if (commit_transaction->t_checkpoint_list == NULL &&
1189 commit_transaction->t_checkpoint_io_list == NULL) {
1190 __jbd2_journal_drop_transaction(journal, commit_transaction);
1191 jbd2_journal_free_transaction(commit_transaction);
1192 }
1193 spin_unlock(&journal->j_list_lock);
1194 write_unlock(&journal->j_state_lock);
1195 wake_up(&journal->j_wait_done_commit);
1196 wake_up(&journal->j_fc_wait);
1197
1198 /*
1199 * Calculate overall stats
1200 */
1201 spin_lock(&journal->j_history_lock);
1202 journal->j_stats.ts_tid++;
1203 journal->j_stats.ts_requested += stats.ts_requested;
1204 journal->j_stats.run.rs_wait += stats.run.rs_wait;
1205 journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1206 journal->j_stats.run.rs_running += stats.run.rs_running;
1207 journal->j_stats.run.rs_locked += stats.run.rs_locked;
1208 journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1209 journal->j_stats.run.rs_logging += stats.run.rs_logging;
1210 journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1211 journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1212 journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1213 spin_unlock(&journal->j_history_lock);
1214 }
1215