xref: /linux/fs/jbd2/commit.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 dealyed allocation we may be doing
188  * block allocation in writepages().
189  */
190 static int journal_submit_inode_data_buffers(struct address_space *mapping)
191 {
192 	int ret;
193 	struct writeback_control wbc = {
194 		.sync_mode =  WB_SYNC_ALL,
195 		.nr_to_write = mapping->nrpages * 2,
196 		.range_start = 0,
197 		.range_end = i_size_read(mapping->host),
198 	};
199 
200 	ret = generic_writepages(mapping, &wbc);
201 	return ret;
202 }
203 
204 /*
205  * Submit all the data buffers of inode associated with the transaction to
206  * disk.
207  *
208  * We are in a committing transaction. Therefore no new inode can be added to
209  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
210  * operate on from being released while we write out pages.
211  */
212 static int journal_submit_data_buffers(journal_t *journal,
213 		transaction_t *commit_transaction)
214 {
215 	struct jbd2_inode *jinode;
216 	int err, ret = 0;
217 	struct address_space *mapping;
218 
219 	spin_lock(&journal->j_list_lock);
220 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
221 		if (!(jinode->i_flags & JI_WRITE_DATA))
222 			continue;
223 		mapping = jinode->i_vfs_inode->i_mapping;
224 		jinode->i_flags |= JI_COMMIT_RUNNING;
225 		spin_unlock(&journal->j_list_lock);
226 		/*
227 		 * submit the inode data buffers. We use writepage
228 		 * instead of writepages. Because writepages can do
229 		 * block allocation  with delalloc. We need to write
230 		 * only allocated blocks here.
231 		 */
232 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
233 		err = journal_submit_inode_data_buffers(mapping);
234 		if (!ret)
235 			ret = err;
236 		spin_lock(&journal->j_list_lock);
237 		J_ASSERT(jinode->i_transaction == commit_transaction);
238 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
239 		smp_mb();
240 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
241 	}
242 	spin_unlock(&journal->j_list_lock);
243 	return ret;
244 }
245 
246 /*
247  * Wait for data submitted for writeout, refile inodes to proper
248  * transaction if needed.
249  *
250  */
251 static int journal_finish_inode_data_buffers(journal_t *journal,
252 		transaction_t *commit_transaction)
253 {
254 	struct jbd2_inode *jinode, *next_i;
255 	int err, ret = 0;
256 
257 	/* For locking, see the comment in journal_submit_data_buffers() */
258 	spin_lock(&journal->j_list_lock);
259 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
260 		if (!(jinode->i_flags & JI_WAIT_DATA))
261 			continue;
262 		jinode->i_flags |= JI_COMMIT_RUNNING;
263 		spin_unlock(&journal->j_list_lock);
264 		err = filemap_fdatawait_keep_errors(
265 				jinode->i_vfs_inode->i_mapping);
266 		if (!ret)
267 			ret = err;
268 		spin_lock(&journal->j_list_lock);
269 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
270 		smp_mb();
271 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
272 	}
273 
274 	/* Now refile inode to proper lists */
275 	list_for_each_entry_safe(jinode, next_i,
276 				 &commit_transaction->t_inode_list, i_list) {
277 		list_del(&jinode->i_list);
278 		if (jinode->i_next_transaction) {
279 			jinode->i_transaction = jinode->i_next_transaction;
280 			jinode->i_next_transaction = NULL;
281 			list_add(&jinode->i_list,
282 				&jinode->i_transaction->t_inode_list);
283 		} else {
284 			jinode->i_transaction = NULL;
285 		}
286 	}
287 	spin_unlock(&journal->j_list_lock);
288 
289 	return ret;
290 }
291 
292 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
293 {
294 	struct page *page = bh->b_page;
295 	char *addr;
296 	__u32 checksum;
297 
298 	addr = kmap_atomic(page);
299 	checksum = crc32_be(crc32_sum,
300 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
301 	kunmap_atomic(addr);
302 
303 	return checksum;
304 }
305 
306 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
307 				   unsigned long long block)
308 {
309 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
310 	if (jbd2_has_feature_64bit(j))
311 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
312 }
313 
314 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
315 				    struct buffer_head *bh, __u32 sequence)
316 {
317 	journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
318 	struct page *page = bh->b_page;
319 	__u8 *addr;
320 	__u32 csum32;
321 	__be32 seq;
322 
323 	if (!jbd2_journal_has_csum_v2or3(j))
324 		return;
325 
326 	seq = cpu_to_be32(sequence);
327 	addr = kmap_atomic(page);
328 	csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
329 	csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
330 			     bh->b_size);
331 	kunmap_atomic(addr);
332 
333 	if (jbd2_has_feature_csum3(j))
334 		tag3->t_checksum = cpu_to_be32(csum32);
335 	else
336 		tag->t_checksum = cpu_to_be16(csum32);
337 }
338 /*
339  * jbd2_journal_commit_transaction
340  *
341  * The primary function for committing a transaction to the log.  This
342  * function is called by the journal thread to begin a complete commit.
343  */
344 void jbd2_journal_commit_transaction(journal_t *journal)
345 {
346 	struct transaction_stats_s stats;
347 	transaction_t *commit_transaction;
348 	struct journal_head *jh;
349 	struct buffer_head *descriptor;
350 	struct buffer_head **wbuf = journal->j_wbuf;
351 	int bufs;
352 	int flags;
353 	int err;
354 	unsigned long long blocknr;
355 	ktime_t start_time;
356 	u64 commit_time;
357 	char *tagp = NULL;
358 	journal_block_tag_t *tag = NULL;
359 	int space_left = 0;
360 	int first_tag = 0;
361 	int tag_flag;
362 	int i;
363 	int tag_bytes = journal_tag_bytes(journal);
364 	struct buffer_head *cbh = NULL; /* For transactional checksums */
365 	__u32 crc32_sum = ~0;
366 	struct blk_plug plug;
367 	/* Tail of the journal */
368 	unsigned long first_block;
369 	tid_t first_tid;
370 	int update_tail;
371 	int csum_size = 0;
372 	LIST_HEAD(io_bufs);
373 	LIST_HEAD(log_bufs);
374 
375 	if (jbd2_journal_has_csum_v2or3(journal))
376 		csum_size = sizeof(struct jbd2_journal_block_tail);
377 
378 	/*
379 	 * First job: lock down the current transaction and wait for
380 	 * all outstanding updates to complete.
381 	 */
382 
383 	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
384 	if (journal->j_flags & JBD2_FLUSHED) {
385 		jbd_debug(3, "super block updated\n");
386 		mutex_lock_io(&journal->j_checkpoint_mutex);
387 		/*
388 		 * We hold j_checkpoint_mutex so tail cannot change under us.
389 		 * We don't need any special data guarantees for writing sb
390 		 * since journal is empty and it is ok for write to be
391 		 * flushed only with transaction commit.
392 		 */
393 		jbd2_journal_update_sb_log_tail(journal,
394 						journal->j_tail_sequence,
395 						journal->j_tail,
396 						REQ_SYNC);
397 		mutex_unlock(&journal->j_checkpoint_mutex);
398 	} else {
399 		jbd_debug(3, "superblock not updated\n");
400 	}
401 
402 	J_ASSERT(journal->j_running_transaction != NULL);
403 	J_ASSERT(journal->j_committing_transaction == NULL);
404 
405 	commit_transaction = journal->j_running_transaction;
406 
407 	trace_jbd2_start_commit(journal, commit_transaction);
408 	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
409 			commit_transaction->t_tid);
410 
411 	write_lock(&journal->j_state_lock);
412 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
413 	commit_transaction->t_state = T_LOCKED;
414 
415 	trace_jbd2_commit_locking(journal, commit_transaction);
416 	stats.run.rs_wait = commit_transaction->t_max_wait;
417 	stats.run.rs_request_delay = 0;
418 	stats.run.rs_locked = jiffies;
419 	if (commit_transaction->t_requested)
420 		stats.run.rs_request_delay =
421 			jbd2_time_diff(commit_transaction->t_requested,
422 				       stats.run.rs_locked);
423 	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
424 					      stats.run.rs_locked);
425 
426 	spin_lock(&commit_transaction->t_handle_lock);
427 	while (atomic_read(&commit_transaction->t_updates)) {
428 		DEFINE_WAIT(wait);
429 
430 		prepare_to_wait(&journal->j_wait_updates, &wait,
431 					TASK_UNINTERRUPTIBLE);
432 		if (atomic_read(&commit_transaction->t_updates)) {
433 			spin_unlock(&commit_transaction->t_handle_lock);
434 			write_unlock(&journal->j_state_lock);
435 			schedule();
436 			write_lock(&journal->j_state_lock);
437 			spin_lock(&commit_transaction->t_handle_lock);
438 		}
439 		finish_wait(&journal->j_wait_updates, &wait);
440 	}
441 	spin_unlock(&commit_transaction->t_handle_lock);
442 	commit_transaction->t_state = T_SWITCH;
443 	write_unlock(&journal->j_state_lock);
444 
445 	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
446 			journal->j_max_transaction_buffers);
447 
448 	/*
449 	 * First thing we are allowed to do is to discard any remaining
450 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
451 	 * that there are no such buffers: if a large filesystem
452 	 * operation like a truncate needs to split itself over multiple
453 	 * transactions, then it may try to do a jbd2_journal_restart() while
454 	 * there are still BJ_Reserved buffers outstanding.  These must
455 	 * be released cleanly from the current transaction.
456 	 *
457 	 * In this case, the filesystem must still reserve write access
458 	 * again before modifying the buffer in the new transaction, but
459 	 * we do not require it to remember exactly which old buffers it
460 	 * has reserved.  This is consistent with the existing behaviour
461 	 * that multiple jbd2_journal_get_write_access() calls to the same
462 	 * buffer are perfectly permissible.
463 	 */
464 	while (commit_transaction->t_reserved_list) {
465 		jh = commit_transaction->t_reserved_list;
466 		JBUFFER_TRACE(jh, "reserved, unused: refile");
467 		/*
468 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
469 		 * leave undo-committed data.
470 		 */
471 		if (jh->b_committed_data) {
472 			struct buffer_head *bh = jh2bh(jh);
473 
474 			jbd_lock_bh_state(bh);
475 			jbd2_free(jh->b_committed_data, bh->b_size);
476 			jh->b_committed_data = NULL;
477 			jbd_unlock_bh_state(bh);
478 		}
479 		jbd2_journal_refile_buffer(journal, jh);
480 	}
481 
482 	/*
483 	 * Now try to drop any written-back buffers from the journal's
484 	 * checkpoint lists.  We do this *before* commit because it potentially
485 	 * frees some memory
486 	 */
487 	spin_lock(&journal->j_list_lock);
488 	__jbd2_journal_clean_checkpoint_list(journal, false);
489 	spin_unlock(&journal->j_list_lock);
490 
491 	jbd_debug(3, "JBD2: commit phase 1\n");
492 
493 	/*
494 	 * Clear revoked flag to reflect there is no revoked buffers
495 	 * in the next transaction which is going to be started.
496 	 */
497 	jbd2_clear_buffer_revoked_flags(journal);
498 
499 	/*
500 	 * Switch to a new revoke table.
501 	 */
502 	jbd2_journal_switch_revoke_table(journal);
503 
504 	/*
505 	 * Reserved credits cannot be claimed anymore, free them
506 	 */
507 	atomic_sub(atomic_read(&journal->j_reserved_credits),
508 		   &commit_transaction->t_outstanding_credits);
509 
510 	write_lock(&journal->j_state_lock);
511 	trace_jbd2_commit_flushing(journal, commit_transaction);
512 	stats.run.rs_flushing = jiffies;
513 	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
514 					     stats.run.rs_flushing);
515 
516 	commit_transaction->t_state = T_FLUSH;
517 	journal->j_committing_transaction = commit_transaction;
518 	journal->j_running_transaction = NULL;
519 	start_time = ktime_get();
520 	commit_transaction->t_log_start = journal->j_head;
521 	wake_up(&journal->j_wait_transaction_locked);
522 	write_unlock(&journal->j_state_lock);
523 
524 	jbd_debug(3, "JBD2: commit phase 2a\n");
525 
526 	/*
527 	 * Now start flushing things to disk, in the order they appear
528 	 * on the transaction lists.  Data blocks go first.
529 	 */
530 	err = journal_submit_data_buffers(journal, commit_transaction);
531 	if (err)
532 		jbd2_journal_abort(journal, err);
533 
534 	blk_start_plug(&plug);
535 	jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
536 
537 	jbd_debug(3, "JBD2: commit phase 2b\n");
538 
539 	/*
540 	 * Way to go: we have now written out all of the data for a
541 	 * transaction!  Now comes the tricky part: we need to write out
542 	 * metadata.  Loop over the transaction's entire buffer list:
543 	 */
544 	write_lock(&journal->j_state_lock);
545 	commit_transaction->t_state = T_COMMIT;
546 	write_unlock(&journal->j_state_lock);
547 
548 	trace_jbd2_commit_logging(journal, commit_transaction);
549 	stats.run.rs_logging = jiffies;
550 	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
551 					       stats.run.rs_logging);
552 	stats.run.rs_blocks =
553 		atomic_read(&commit_transaction->t_outstanding_credits);
554 	stats.run.rs_blocks_logged = 0;
555 
556 	J_ASSERT(commit_transaction->t_nr_buffers <=
557 		 atomic_read(&commit_transaction->t_outstanding_credits));
558 
559 	err = 0;
560 	bufs = 0;
561 	descriptor = NULL;
562 	while (commit_transaction->t_buffers) {
563 
564 		/* Find the next buffer to be journaled... */
565 
566 		jh = commit_transaction->t_buffers;
567 
568 		/* If we're in abort mode, we just un-journal the buffer and
569 		   release it. */
570 
571 		if (is_journal_aborted(journal)) {
572 			clear_buffer_jbddirty(jh2bh(jh));
573 			JBUFFER_TRACE(jh, "journal is aborting: refile");
574 			jbd2_buffer_abort_trigger(jh,
575 						  jh->b_frozen_data ?
576 						  jh->b_frozen_triggers :
577 						  jh->b_triggers);
578 			jbd2_journal_refile_buffer(journal, jh);
579 			/* If that was the last one, we need to clean up
580 			 * any descriptor buffers which may have been
581 			 * already allocated, even if we are now
582 			 * aborting. */
583 			if (!commit_transaction->t_buffers)
584 				goto start_journal_io;
585 			continue;
586 		}
587 
588 		/* Make sure we have a descriptor block in which to
589 		   record the metadata buffer. */
590 
591 		if (!descriptor) {
592 			J_ASSERT (bufs == 0);
593 
594 			jbd_debug(4, "JBD2: get descriptor\n");
595 
596 			descriptor = jbd2_journal_get_descriptor_buffer(
597 							commit_transaction,
598 							JBD2_DESCRIPTOR_BLOCK);
599 			if (!descriptor) {
600 				jbd2_journal_abort(journal, -EIO);
601 				continue;
602 			}
603 
604 			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
605 				(unsigned long long)descriptor->b_blocknr,
606 				descriptor->b_data);
607 			tagp = &descriptor->b_data[sizeof(journal_header_t)];
608 			space_left = descriptor->b_size -
609 						sizeof(journal_header_t);
610 			first_tag = 1;
611 			set_buffer_jwrite(descriptor);
612 			set_buffer_dirty(descriptor);
613 			wbuf[bufs++] = descriptor;
614 
615 			/* Record it so that we can wait for IO
616                            completion later */
617 			BUFFER_TRACE(descriptor, "ph3: file as descriptor");
618 			jbd2_file_log_bh(&log_bufs, descriptor);
619 		}
620 
621 		/* Where is the buffer to be written? */
622 
623 		err = jbd2_journal_next_log_block(journal, &blocknr);
624 		/* If the block mapping failed, just abandon the buffer
625 		   and repeat this loop: we'll fall into the
626 		   refile-on-abort condition above. */
627 		if (err) {
628 			jbd2_journal_abort(journal, err);
629 			continue;
630 		}
631 
632 		/*
633 		 * start_this_handle() uses t_outstanding_credits to determine
634 		 * the free space in the log, but this counter is changed
635 		 * by jbd2_journal_next_log_block() also.
636 		 */
637 		atomic_dec(&commit_transaction->t_outstanding_credits);
638 
639 		/* Bump b_count to prevent truncate from stumbling over
640                    the shadowed buffer!  @@@ This can go if we ever get
641                    rid of the shadow pairing of buffers. */
642 		atomic_inc(&jh2bh(jh)->b_count);
643 
644 		/*
645 		 * Make a temporary IO buffer with which to write it out
646 		 * (this will requeue the metadata buffer to BJ_Shadow).
647 		 */
648 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
649 		JBUFFER_TRACE(jh, "ph3: write metadata");
650 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
651 						jh, &wbuf[bufs], blocknr);
652 		if (flags < 0) {
653 			jbd2_journal_abort(journal, flags);
654 			continue;
655 		}
656 		jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
657 
658 		/* Record the new block's tag in the current descriptor
659                    buffer */
660 
661 		tag_flag = 0;
662 		if (flags & 1)
663 			tag_flag |= JBD2_FLAG_ESCAPE;
664 		if (!first_tag)
665 			tag_flag |= JBD2_FLAG_SAME_UUID;
666 
667 		tag = (journal_block_tag_t *) tagp;
668 		write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
669 		tag->t_flags = cpu_to_be16(tag_flag);
670 		jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
671 					commit_transaction->t_tid);
672 		tagp += tag_bytes;
673 		space_left -= tag_bytes;
674 		bufs++;
675 
676 		if (first_tag) {
677 			memcpy (tagp, journal->j_uuid, 16);
678 			tagp += 16;
679 			space_left -= 16;
680 			first_tag = 0;
681 		}
682 
683 		/* If there's no more to do, or if the descriptor is full,
684 		   let the IO rip! */
685 
686 		if (bufs == journal->j_wbufsize ||
687 		    commit_transaction->t_buffers == NULL ||
688 		    space_left < tag_bytes + 16 + csum_size) {
689 
690 			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
691 
692 			/* Write an end-of-descriptor marker before
693                            submitting the IOs.  "tag" still points to
694                            the last tag we set up. */
695 
696 			tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
697 start_journal_io:
698 			if (descriptor)
699 				jbd2_descriptor_block_csum_set(journal,
700 							descriptor);
701 
702 			for (i = 0; i < bufs; i++) {
703 				struct buffer_head *bh = wbuf[i];
704 				/*
705 				 * Compute checksum.
706 				 */
707 				if (jbd2_has_feature_checksum(journal)) {
708 					crc32_sum =
709 					    jbd2_checksum_data(crc32_sum, bh);
710 				}
711 
712 				lock_buffer(bh);
713 				clear_buffer_dirty(bh);
714 				set_buffer_uptodate(bh);
715 				bh->b_end_io = journal_end_buffer_io_sync;
716 				submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
717 			}
718 			cond_resched();
719 			stats.run.rs_blocks_logged += bufs;
720 
721 			/* Force a new descriptor to be generated next
722                            time round the loop. */
723 			descriptor = NULL;
724 			bufs = 0;
725 		}
726 	}
727 
728 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
729 	if (err) {
730 		printk(KERN_WARNING
731 			"JBD2: Detected IO errors while flushing file data "
732 		       "on %s\n", journal->j_devname);
733 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
734 			jbd2_journal_abort(journal, err);
735 		err = 0;
736 	}
737 
738 	/*
739 	 * Get current oldest transaction in the log before we issue flush
740 	 * to the filesystem device. After the flush we can be sure that
741 	 * blocks of all older transactions are checkpointed to persistent
742 	 * storage and we will be safe to update journal start in the
743 	 * superblock with the numbers we get here.
744 	 */
745 	update_tail =
746 		jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
747 
748 	write_lock(&journal->j_state_lock);
749 	if (update_tail) {
750 		long freed = first_block - journal->j_tail;
751 
752 		if (first_block < journal->j_tail)
753 			freed += journal->j_last - journal->j_first;
754 		/* Update tail only if we free significant amount of space */
755 		if (freed < journal->j_maxlen / 4)
756 			update_tail = 0;
757 	}
758 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
759 	commit_transaction->t_state = T_COMMIT_DFLUSH;
760 	write_unlock(&journal->j_state_lock);
761 
762 	/*
763 	 * If the journal is not located on the file system device,
764 	 * then we must flush the file system device before we issue
765 	 * the commit record
766 	 */
767 	if (commit_transaction->t_need_data_flush &&
768 	    (journal->j_fs_dev != journal->j_dev) &&
769 	    (journal->j_flags & JBD2_BARRIER))
770 		blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
771 
772 	/* Done it all: now write the commit record asynchronously. */
773 	if (jbd2_has_feature_async_commit(journal)) {
774 		err = journal_submit_commit_record(journal, commit_transaction,
775 						 &cbh, crc32_sum);
776 		if (err)
777 			__jbd2_journal_abort_hard(journal);
778 	}
779 
780 	blk_finish_plug(&plug);
781 
782 	/* Lo and behold: we have just managed to send a transaction to
783            the log.  Before we can commit it, wait for the IO so far to
784            complete.  Control buffers being written are on the
785            transaction's t_log_list queue, and metadata buffers are on
786            the io_bufs list.
787 
788 	   Wait for the buffers in reverse order.  That way we are
789 	   less likely to be woken up until all IOs have completed, and
790 	   so we incur less scheduling load.
791 	*/
792 
793 	jbd_debug(3, "JBD2: commit phase 3\n");
794 
795 	while (!list_empty(&io_bufs)) {
796 		struct buffer_head *bh = list_entry(io_bufs.prev,
797 						    struct buffer_head,
798 						    b_assoc_buffers);
799 
800 		wait_on_buffer(bh);
801 		cond_resched();
802 
803 		if (unlikely(!buffer_uptodate(bh)))
804 			err = -EIO;
805 		jbd2_unfile_log_bh(bh);
806 
807 		/*
808 		 * The list contains temporary buffer heads created by
809 		 * jbd2_journal_write_metadata_buffer().
810 		 */
811 		BUFFER_TRACE(bh, "dumping temporary bh");
812 		__brelse(bh);
813 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
814 		free_buffer_head(bh);
815 
816 		/* We also have to refile the corresponding shadowed buffer */
817 		jh = commit_transaction->t_shadow_list->b_tprev;
818 		bh = jh2bh(jh);
819 		clear_buffer_jwrite(bh);
820 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
821 		J_ASSERT_BH(bh, !buffer_shadow(bh));
822 
823 		/* The metadata is now released for reuse, but we need
824                    to remember it against this transaction so that when
825                    we finally commit, we can do any checkpointing
826                    required. */
827 		JBUFFER_TRACE(jh, "file as BJ_Forget");
828 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
829 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
830 		__brelse(bh);
831 	}
832 
833 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
834 
835 	jbd_debug(3, "JBD2: commit phase 4\n");
836 
837 	/* Here we wait for the revoke record and descriptor record buffers */
838 	while (!list_empty(&log_bufs)) {
839 		struct buffer_head *bh;
840 
841 		bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
842 		wait_on_buffer(bh);
843 		cond_resched();
844 
845 		if (unlikely(!buffer_uptodate(bh)))
846 			err = -EIO;
847 
848 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
849 		clear_buffer_jwrite(bh);
850 		jbd2_unfile_log_bh(bh);
851 		__brelse(bh);		/* One for getblk */
852 		/* AKPM: bforget here */
853 	}
854 
855 	if (err)
856 		jbd2_journal_abort(journal, err);
857 
858 	jbd_debug(3, "JBD2: commit phase 5\n");
859 	write_lock(&journal->j_state_lock);
860 	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
861 	commit_transaction->t_state = T_COMMIT_JFLUSH;
862 	write_unlock(&journal->j_state_lock);
863 
864 	if (!jbd2_has_feature_async_commit(journal)) {
865 		err = journal_submit_commit_record(journal, commit_transaction,
866 						&cbh, crc32_sum);
867 		if (err)
868 			__jbd2_journal_abort_hard(journal);
869 	}
870 	if (cbh)
871 		err = journal_wait_on_commit_record(journal, cbh);
872 	if (jbd2_has_feature_async_commit(journal) &&
873 	    journal->j_flags & JBD2_BARRIER) {
874 		blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
875 	}
876 
877 	if (err)
878 		jbd2_journal_abort(journal, err);
879 
880 	/*
881 	 * Now disk caches for filesystem device are flushed so we are safe to
882 	 * erase checkpointed transactions from the log by updating journal
883 	 * superblock.
884 	 */
885 	if (update_tail)
886 		jbd2_update_log_tail(journal, first_tid, first_block);
887 
888 	/* End of a transaction!  Finally, we can do checkpoint
889            processing: any buffers committed as a result of this
890            transaction can be removed from any checkpoint list it was on
891            before. */
892 
893 	jbd_debug(3, "JBD2: commit phase 6\n");
894 
895 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
896 	J_ASSERT(commit_transaction->t_buffers == NULL);
897 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
898 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
899 
900 restart_loop:
901 	/*
902 	 * As there are other places (journal_unmap_buffer()) adding buffers
903 	 * to this list we have to be careful and hold the j_list_lock.
904 	 */
905 	spin_lock(&journal->j_list_lock);
906 	while (commit_transaction->t_forget) {
907 		transaction_t *cp_transaction;
908 		struct buffer_head *bh;
909 		int try_to_free = 0;
910 
911 		jh = commit_transaction->t_forget;
912 		spin_unlock(&journal->j_list_lock);
913 		bh = jh2bh(jh);
914 		/*
915 		 * Get a reference so that bh cannot be freed before we are
916 		 * done with it.
917 		 */
918 		get_bh(bh);
919 		jbd_lock_bh_state(bh);
920 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
921 
922 		/*
923 		 * If there is undo-protected committed data against
924 		 * this buffer, then we can remove it now.  If it is a
925 		 * buffer needing such protection, the old frozen_data
926 		 * field now points to a committed version of the
927 		 * buffer, so rotate that field to the new committed
928 		 * data.
929 		 *
930 		 * Otherwise, we can just throw away the frozen data now.
931 		 *
932 		 * We also know that the frozen data has already fired
933 		 * its triggers if they exist, so we can clear that too.
934 		 */
935 		if (jh->b_committed_data) {
936 			jbd2_free(jh->b_committed_data, bh->b_size);
937 			jh->b_committed_data = NULL;
938 			if (jh->b_frozen_data) {
939 				jh->b_committed_data = jh->b_frozen_data;
940 				jh->b_frozen_data = NULL;
941 				jh->b_frozen_triggers = NULL;
942 			}
943 		} else if (jh->b_frozen_data) {
944 			jbd2_free(jh->b_frozen_data, bh->b_size);
945 			jh->b_frozen_data = NULL;
946 			jh->b_frozen_triggers = NULL;
947 		}
948 
949 		spin_lock(&journal->j_list_lock);
950 		cp_transaction = jh->b_cp_transaction;
951 		if (cp_transaction) {
952 			JBUFFER_TRACE(jh, "remove from old cp transaction");
953 			cp_transaction->t_chp_stats.cs_dropped++;
954 			__jbd2_journal_remove_checkpoint(jh);
955 		}
956 
957 		/* Only re-checkpoint the buffer_head if it is marked
958 		 * dirty.  If the buffer was added to the BJ_Forget list
959 		 * by jbd2_journal_forget, it may no longer be dirty and
960 		 * there's no point in keeping a checkpoint record for
961 		 * it. */
962 
963 		/*
964 		* A buffer which has been freed while still being journaled by
965 		* a previous transaction.
966 		*/
967 		if (buffer_freed(bh)) {
968 			/*
969 			 * If the running transaction is the one containing
970 			 * "add to orphan" operation (b_next_transaction !=
971 			 * NULL), we have to wait for that transaction to
972 			 * commit before we can really get rid of the buffer.
973 			 * So just clear b_modified to not confuse transaction
974 			 * credit accounting and refile the buffer to
975 			 * BJ_Forget of the running transaction. If the just
976 			 * committed transaction contains "add to orphan"
977 			 * operation, we can completely invalidate the buffer
978 			 * now. We are rather through in that since the
979 			 * buffer may be still accessible when blocksize <
980 			 * pagesize and it is attached to the last partial
981 			 * page.
982 			 */
983 			jh->b_modified = 0;
984 			if (!jh->b_next_transaction) {
985 				clear_buffer_freed(bh);
986 				clear_buffer_jbddirty(bh);
987 				clear_buffer_mapped(bh);
988 				clear_buffer_new(bh);
989 				clear_buffer_req(bh);
990 				bh->b_bdev = NULL;
991 			}
992 		}
993 
994 		if (buffer_jbddirty(bh)) {
995 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
996 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
997 			if (is_journal_aborted(journal))
998 				clear_buffer_jbddirty(bh);
999 		} else {
1000 			J_ASSERT_BH(bh, !buffer_dirty(bh));
1001 			/*
1002 			 * The buffer on BJ_Forget list and not jbddirty means
1003 			 * it has been freed by this transaction and hence it
1004 			 * could not have been reallocated until this
1005 			 * transaction has committed. *BUT* it could be
1006 			 * reallocated once we have written all the data to
1007 			 * disk and before we process the buffer on BJ_Forget
1008 			 * list.
1009 			 */
1010 			if (!jh->b_next_transaction)
1011 				try_to_free = 1;
1012 		}
1013 		JBUFFER_TRACE(jh, "refile or unfile buffer");
1014 		__jbd2_journal_refile_buffer(jh);
1015 		jbd_unlock_bh_state(bh);
1016 		if (try_to_free)
1017 			release_buffer_page(bh);	/* Drops bh reference */
1018 		else
1019 			__brelse(bh);
1020 		cond_resched_lock(&journal->j_list_lock);
1021 	}
1022 	spin_unlock(&journal->j_list_lock);
1023 	/*
1024 	 * This is a bit sleazy.  We use j_list_lock to protect transition
1025 	 * of a transaction into T_FINISHED state and calling
1026 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1027 	 * other checkpointing code processing the transaction...
1028 	 */
1029 	write_lock(&journal->j_state_lock);
1030 	spin_lock(&journal->j_list_lock);
1031 	/*
1032 	 * Now recheck if some buffers did not get attached to the transaction
1033 	 * while the lock was dropped...
1034 	 */
1035 	if (commit_transaction->t_forget) {
1036 		spin_unlock(&journal->j_list_lock);
1037 		write_unlock(&journal->j_state_lock);
1038 		goto restart_loop;
1039 	}
1040 
1041 	/* Add the transaction to the checkpoint list
1042 	 * __journal_remove_checkpoint() can not destroy transaction
1043 	 * under us because it is not marked as T_FINISHED yet */
1044 	if (journal->j_checkpoint_transactions == NULL) {
1045 		journal->j_checkpoint_transactions = commit_transaction;
1046 		commit_transaction->t_cpnext = commit_transaction;
1047 		commit_transaction->t_cpprev = commit_transaction;
1048 	} else {
1049 		commit_transaction->t_cpnext =
1050 			journal->j_checkpoint_transactions;
1051 		commit_transaction->t_cpprev =
1052 			commit_transaction->t_cpnext->t_cpprev;
1053 		commit_transaction->t_cpnext->t_cpprev =
1054 			commit_transaction;
1055 		commit_transaction->t_cpprev->t_cpnext =
1056 				commit_transaction;
1057 	}
1058 	spin_unlock(&journal->j_list_lock);
1059 
1060 	/* Done with this transaction! */
1061 
1062 	jbd_debug(3, "JBD2: commit phase 7\n");
1063 
1064 	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1065 
1066 	commit_transaction->t_start = jiffies;
1067 	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1068 					      commit_transaction->t_start);
1069 
1070 	/*
1071 	 * File the transaction statistics
1072 	 */
1073 	stats.ts_tid = commit_transaction->t_tid;
1074 	stats.run.rs_handle_count =
1075 		atomic_read(&commit_transaction->t_handle_count);
1076 	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1077 			     commit_transaction->t_tid, &stats.run);
1078 	stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1079 
1080 	commit_transaction->t_state = T_COMMIT_CALLBACK;
1081 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1082 	journal->j_commit_sequence = commit_transaction->t_tid;
1083 	journal->j_committing_transaction = NULL;
1084 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1085 
1086 	/*
1087 	 * weight the commit time higher than the average time so we don't
1088 	 * react too strongly to vast changes in the commit time
1089 	 */
1090 	if (likely(journal->j_average_commit_time))
1091 		journal->j_average_commit_time = (commit_time +
1092 				journal->j_average_commit_time*3) / 4;
1093 	else
1094 		journal->j_average_commit_time = commit_time;
1095 
1096 	write_unlock(&journal->j_state_lock);
1097 
1098 	if (journal->j_commit_callback)
1099 		journal->j_commit_callback(journal, commit_transaction);
1100 
1101 	trace_jbd2_end_commit(journal, commit_transaction);
1102 	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1103 		  journal->j_commit_sequence, journal->j_tail_sequence);
1104 
1105 	write_lock(&journal->j_state_lock);
1106 	spin_lock(&journal->j_list_lock);
1107 	commit_transaction->t_state = T_FINISHED;
1108 	/* Check if the transaction can be dropped now that we are finished */
1109 	if (commit_transaction->t_checkpoint_list == NULL &&
1110 	    commit_transaction->t_checkpoint_io_list == NULL) {
1111 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1112 		jbd2_journal_free_transaction(commit_transaction);
1113 	}
1114 	spin_unlock(&journal->j_list_lock);
1115 	write_unlock(&journal->j_state_lock);
1116 	wake_up(&journal->j_wait_done_commit);
1117 
1118 	/*
1119 	 * Calculate overall stats
1120 	 */
1121 	spin_lock(&journal->j_history_lock);
1122 	journal->j_stats.ts_tid++;
1123 	journal->j_stats.ts_requested += stats.ts_requested;
1124 	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1125 	journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1126 	journal->j_stats.run.rs_running += stats.run.rs_running;
1127 	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1128 	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1129 	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1130 	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1131 	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1132 	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1133 	spin_unlock(&journal->j_history_lock);
1134 }
1135