xref: /linux/fs/jbd2/commit.c (revision 58b3bafff8257c6946df5d6aeb215b8ac839ed2a)
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 static int journal_submit_inode_data_buffers(struct address_space *mapping,
191 		loff_t dirty_start, loff_t dirty_end)
192 {
193 	int ret;
194 	struct writeback_control wbc = {
195 		.sync_mode =  WB_SYNC_ALL,
196 		.nr_to_write = mapping->nrpages * 2,
197 		.range_start = dirty_start,
198 		.range_end = dirty_end,
199 	};
200 
201 	ret = generic_writepages(mapping, &wbc);
202 	return ret;
203 }
204 
205 /*
206  * Submit all the data buffers of inode associated with the transaction to
207  * disk.
208  *
209  * We are in a committing transaction. Therefore no new inode can be added to
210  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
211  * operate on from being released while we write out pages.
212  */
213 static int journal_submit_data_buffers(journal_t *journal,
214 		transaction_t *commit_transaction)
215 {
216 	struct jbd2_inode *jinode;
217 	int err, ret = 0;
218 	struct address_space *mapping;
219 
220 	spin_lock(&journal->j_list_lock);
221 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
222 		loff_t dirty_start = jinode->i_dirty_start;
223 		loff_t dirty_end = jinode->i_dirty_end;
224 
225 		if (!(jinode->i_flags & JI_WRITE_DATA))
226 			continue;
227 		mapping = jinode->i_vfs_inode->i_mapping;
228 		jinode->i_flags |= JI_COMMIT_RUNNING;
229 		spin_unlock(&journal->j_list_lock);
230 		/*
231 		 * submit the inode data buffers. We use writepage
232 		 * instead of writepages. Because writepages can do
233 		 * block allocation  with delalloc. We need to write
234 		 * only allocated blocks here.
235 		 */
236 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
237 		err = journal_submit_inode_data_buffers(mapping, dirty_start,
238 				dirty_end);
239 		if (!ret)
240 			ret = err;
241 		spin_lock(&journal->j_list_lock);
242 		J_ASSERT(jinode->i_transaction == commit_transaction);
243 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
244 		smp_mb();
245 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
246 	}
247 	spin_unlock(&journal->j_list_lock);
248 	return ret;
249 }
250 
251 /*
252  * Wait for data submitted for writeout, refile inodes to proper
253  * transaction if needed.
254  *
255  */
256 static int journal_finish_inode_data_buffers(journal_t *journal,
257 		transaction_t *commit_transaction)
258 {
259 	struct jbd2_inode *jinode, *next_i;
260 	int err, ret = 0;
261 
262 	/* For locking, see the comment in journal_submit_data_buffers() */
263 	spin_lock(&journal->j_list_lock);
264 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
265 		loff_t dirty_start = jinode->i_dirty_start;
266 		loff_t dirty_end = jinode->i_dirty_end;
267 
268 		if (!(jinode->i_flags & JI_WAIT_DATA))
269 			continue;
270 		jinode->i_flags |= JI_COMMIT_RUNNING;
271 		spin_unlock(&journal->j_list_lock);
272 		err = filemap_fdatawait_range_keep_errors(
273 				jinode->i_vfs_inode->i_mapping, dirty_start,
274 				dirty_end);
275 		if (!ret)
276 			ret = err;
277 		spin_lock(&journal->j_list_lock);
278 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
279 		smp_mb();
280 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
281 	}
282 
283 	/* Now refile inode to proper lists */
284 	list_for_each_entry_safe(jinode, next_i,
285 				 &commit_transaction->t_inode_list, i_list) {
286 		list_del(&jinode->i_list);
287 		if (jinode->i_next_transaction) {
288 			jinode->i_transaction = jinode->i_next_transaction;
289 			jinode->i_next_transaction = NULL;
290 			list_add(&jinode->i_list,
291 				&jinode->i_transaction->t_inode_list);
292 		} else {
293 			jinode->i_transaction = NULL;
294 			jinode->i_dirty_start = 0;
295 			jinode->i_dirty_end = 0;
296 		}
297 	}
298 	spin_unlock(&journal->j_list_lock);
299 
300 	return ret;
301 }
302 
303 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
304 {
305 	struct page *page = bh->b_page;
306 	char *addr;
307 	__u32 checksum;
308 
309 	addr = kmap_atomic(page);
310 	checksum = crc32_be(crc32_sum,
311 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
312 	kunmap_atomic(addr);
313 
314 	return checksum;
315 }
316 
317 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
318 				   unsigned long long block)
319 {
320 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
321 	if (jbd2_has_feature_64bit(j))
322 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
323 }
324 
325 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
326 				    struct buffer_head *bh, __u32 sequence)
327 {
328 	journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
329 	struct page *page = bh->b_page;
330 	__u8 *addr;
331 	__u32 csum32;
332 	__be32 seq;
333 
334 	if (!jbd2_journal_has_csum_v2or3(j))
335 		return;
336 
337 	seq = cpu_to_be32(sequence);
338 	addr = kmap_atomic(page);
339 	csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
340 	csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
341 			     bh->b_size);
342 	kunmap_atomic(addr);
343 
344 	if (jbd2_has_feature_csum3(j))
345 		tag3->t_checksum = cpu_to_be32(csum32);
346 	else
347 		tag->t_checksum = cpu_to_be16(csum32);
348 }
349 /*
350  * jbd2_journal_commit_transaction
351  *
352  * The primary function for committing a transaction to the log.  This
353  * function is called by the journal thread to begin a complete commit.
354  */
355 void jbd2_journal_commit_transaction(journal_t *journal)
356 {
357 	struct transaction_stats_s stats;
358 	transaction_t *commit_transaction;
359 	struct journal_head *jh;
360 	struct buffer_head *descriptor;
361 	struct buffer_head **wbuf = journal->j_wbuf;
362 	int bufs;
363 	int flags;
364 	int err;
365 	unsigned long long blocknr;
366 	ktime_t start_time;
367 	u64 commit_time;
368 	char *tagp = NULL;
369 	journal_block_tag_t *tag = NULL;
370 	int space_left = 0;
371 	int first_tag = 0;
372 	int tag_flag;
373 	int i;
374 	int tag_bytes = journal_tag_bytes(journal);
375 	struct buffer_head *cbh = NULL; /* For transactional checksums */
376 	__u32 crc32_sum = ~0;
377 	struct blk_plug plug;
378 	/* Tail of the journal */
379 	unsigned long first_block;
380 	tid_t first_tid;
381 	int update_tail;
382 	int csum_size = 0;
383 	LIST_HEAD(io_bufs);
384 	LIST_HEAD(log_bufs);
385 
386 	if (jbd2_journal_has_csum_v2or3(journal))
387 		csum_size = sizeof(struct jbd2_journal_block_tail);
388 
389 	/*
390 	 * First job: lock down the current transaction and wait for
391 	 * all outstanding updates to complete.
392 	 */
393 
394 	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
395 	if (journal->j_flags & JBD2_FLUSHED) {
396 		jbd_debug(3, "super block updated\n");
397 		mutex_lock_io(&journal->j_checkpoint_mutex);
398 		/*
399 		 * We hold j_checkpoint_mutex so tail cannot change under us.
400 		 * We don't need any special data guarantees for writing sb
401 		 * since journal is empty and it is ok for write to be
402 		 * flushed only with transaction commit.
403 		 */
404 		jbd2_journal_update_sb_log_tail(journal,
405 						journal->j_tail_sequence,
406 						journal->j_tail,
407 						REQ_SYNC);
408 		mutex_unlock(&journal->j_checkpoint_mutex);
409 	} else {
410 		jbd_debug(3, "superblock not updated\n");
411 	}
412 
413 	J_ASSERT(journal->j_running_transaction != NULL);
414 	J_ASSERT(journal->j_committing_transaction == NULL);
415 
416 	commit_transaction = journal->j_running_transaction;
417 
418 	trace_jbd2_start_commit(journal, commit_transaction);
419 	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
420 			commit_transaction->t_tid);
421 
422 	write_lock(&journal->j_state_lock);
423 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
424 	commit_transaction->t_state = T_LOCKED;
425 
426 	trace_jbd2_commit_locking(journal, commit_transaction);
427 	stats.run.rs_wait = commit_transaction->t_max_wait;
428 	stats.run.rs_request_delay = 0;
429 	stats.run.rs_locked = jiffies;
430 	if (commit_transaction->t_requested)
431 		stats.run.rs_request_delay =
432 			jbd2_time_diff(commit_transaction->t_requested,
433 				       stats.run.rs_locked);
434 	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
435 					      stats.run.rs_locked);
436 
437 	spin_lock(&commit_transaction->t_handle_lock);
438 	while (atomic_read(&commit_transaction->t_updates)) {
439 		DEFINE_WAIT(wait);
440 
441 		prepare_to_wait(&journal->j_wait_updates, &wait,
442 					TASK_UNINTERRUPTIBLE);
443 		if (atomic_read(&commit_transaction->t_updates)) {
444 			spin_unlock(&commit_transaction->t_handle_lock);
445 			write_unlock(&journal->j_state_lock);
446 			schedule();
447 			write_lock(&journal->j_state_lock);
448 			spin_lock(&commit_transaction->t_handle_lock);
449 		}
450 		finish_wait(&journal->j_wait_updates, &wait);
451 	}
452 	spin_unlock(&commit_transaction->t_handle_lock);
453 	commit_transaction->t_state = T_SWITCH;
454 	write_unlock(&journal->j_state_lock);
455 
456 	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
457 			journal->j_max_transaction_buffers);
458 
459 	/*
460 	 * First thing we are allowed to do is to discard any remaining
461 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
462 	 * that there are no such buffers: if a large filesystem
463 	 * operation like a truncate needs to split itself over multiple
464 	 * transactions, then it may try to do a jbd2_journal_restart() while
465 	 * there are still BJ_Reserved buffers outstanding.  These must
466 	 * be released cleanly from the current transaction.
467 	 *
468 	 * In this case, the filesystem must still reserve write access
469 	 * again before modifying the buffer in the new transaction, but
470 	 * we do not require it to remember exactly which old buffers it
471 	 * has reserved.  This is consistent with the existing behaviour
472 	 * that multiple jbd2_journal_get_write_access() calls to the same
473 	 * buffer are perfectly permissible.
474 	 */
475 	while (commit_transaction->t_reserved_list) {
476 		jh = commit_transaction->t_reserved_list;
477 		JBUFFER_TRACE(jh, "reserved, unused: refile");
478 		/*
479 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
480 		 * leave undo-committed data.
481 		 */
482 		if (jh->b_committed_data) {
483 			struct buffer_head *bh = jh2bh(jh);
484 
485 			spin_lock(&jh->b_state_lock);
486 			jbd2_free(jh->b_committed_data, bh->b_size);
487 			jh->b_committed_data = NULL;
488 			spin_unlock(&jh->b_state_lock);
489 		}
490 		jbd2_journal_refile_buffer(journal, jh);
491 	}
492 
493 	/*
494 	 * Now try to drop any written-back buffers from the journal's
495 	 * checkpoint lists.  We do this *before* commit because it potentially
496 	 * frees some memory
497 	 */
498 	spin_lock(&journal->j_list_lock);
499 	__jbd2_journal_clean_checkpoint_list(journal, false);
500 	spin_unlock(&journal->j_list_lock);
501 
502 	jbd_debug(3, "JBD2: commit phase 1\n");
503 
504 	/*
505 	 * Clear revoked flag to reflect there is no revoked buffers
506 	 * in the next transaction which is going to be started.
507 	 */
508 	jbd2_clear_buffer_revoked_flags(journal);
509 
510 	/*
511 	 * Switch to a new revoke table.
512 	 */
513 	jbd2_journal_switch_revoke_table(journal);
514 
515 	/*
516 	 * Reserved credits cannot be claimed anymore, free them
517 	 */
518 	atomic_sub(atomic_read(&journal->j_reserved_credits),
519 		   &commit_transaction->t_outstanding_credits);
520 
521 	write_lock(&journal->j_state_lock);
522 	trace_jbd2_commit_flushing(journal, commit_transaction);
523 	stats.run.rs_flushing = jiffies;
524 	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
525 					     stats.run.rs_flushing);
526 
527 	commit_transaction->t_state = T_FLUSH;
528 	journal->j_committing_transaction = commit_transaction;
529 	journal->j_running_transaction = NULL;
530 	start_time = ktime_get();
531 	commit_transaction->t_log_start = journal->j_head;
532 	wake_up(&journal->j_wait_transaction_locked);
533 	write_unlock(&journal->j_state_lock);
534 
535 	jbd_debug(3, "JBD2: commit phase 2a\n");
536 
537 	/*
538 	 * Now start flushing things to disk, in the order they appear
539 	 * on the transaction lists.  Data blocks go first.
540 	 */
541 	err = journal_submit_data_buffers(journal, commit_transaction);
542 	if (err)
543 		jbd2_journal_abort(journal, err);
544 
545 	blk_start_plug(&plug);
546 	jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
547 
548 	jbd_debug(3, "JBD2: commit phase 2b\n");
549 
550 	/*
551 	 * Way to go: we have now written out all of the data for a
552 	 * transaction!  Now comes the tricky part: we need to write out
553 	 * metadata.  Loop over the transaction's entire buffer list:
554 	 */
555 	write_lock(&journal->j_state_lock);
556 	commit_transaction->t_state = T_COMMIT;
557 	write_unlock(&journal->j_state_lock);
558 
559 	trace_jbd2_commit_logging(journal, commit_transaction);
560 	stats.run.rs_logging = jiffies;
561 	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
562 					       stats.run.rs_logging);
563 	stats.run.rs_blocks = commit_transaction->t_nr_buffers;
564 	stats.run.rs_blocks_logged = 0;
565 
566 	J_ASSERT(commit_transaction->t_nr_buffers <=
567 		 atomic_read(&commit_transaction->t_outstanding_credits));
568 
569 	err = 0;
570 	bufs = 0;
571 	descriptor = NULL;
572 	while (commit_transaction->t_buffers) {
573 
574 		/* Find the next buffer to be journaled... */
575 
576 		jh = commit_transaction->t_buffers;
577 
578 		/* If we're in abort mode, we just un-journal the buffer and
579 		   release it. */
580 
581 		if (is_journal_aborted(journal)) {
582 			clear_buffer_jbddirty(jh2bh(jh));
583 			JBUFFER_TRACE(jh, "journal is aborting: refile");
584 			jbd2_buffer_abort_trigger(jh,
585 						  jh->b_frozen_data ?
586 						  jh->b_frozen_triggers :
587 						  jh->b_triggers);
588 			jbd2_journal_refile_buffer(journal, jh);
589 			/* If that was the last one, we need to clean up
590 			 * any descriptor buffers which may have been
591 			 * already allocated, even if we are now
592 			 * aborting. */
593 			if (!commit_transaction->t_buffers)
594 				goto start_journal_io;
595 			continue;
596 		}
597 
598 		/* Make sure we have a descriptor block in which to
599 		   record the metadata buffer. */
600 
601 		if (!descriptor) {
602 			J_ASSERT (bufs == 0);
603 
604 			jbd_debug(4, "JBD2: get descriptor\n");
605 
606 			descriptor = jbd2_journal_get_descriptor_buffer(
607 							commit_transaction,
608 							JBD2_DESCRIPTOR_BLOCK);
609 			if (!descriptor) {
610 				jbd2_journal_abort(journal, -EIO);
611 				continue;
612 			}
613 
614 			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
615 				(unsigned long long)descriptor->b_blocknr,
616 				descriptor->b_data);
617 			tagp = &descriptor->b_data[sizeof(journal_header_t)];
618 			space_left = descriptor->b_size -
619 						sizeof(journal_header_t);
620 			first_tag = 1;
621 			set_buffer_jwrite(descriptor);
622 			set_buffer_dirty(descriptor);
623 			wbuf[bufs++] = descriptor;
624 
625 			/* Record it so that we can wait for IO
626                            completion later */
627 			BUFFER_TRACE(descriptor, "ph3: file as descriptor");
628 			jbd2_file_log_bh(&log_bufs, descriptor);
629 		}
630 
631 		/* Where is the buffer to be written? */
632 
633 		err = jbd2_journal_next_log_block(journal, &blocknr);
634 		/* If the block mapping failed, just abandon the buffer
635 		   and repeat this loop: we'll fall into the
636 		   refile-on-abort condition above. */
637 		if (err) {
638 			jbd2_journal_abort(journal, err);
639 			continue;
640 		}
641 
642 		/*
643 		 * start_this_handle() uses t_outstanding_credits to determine
644 		 * the free space in the log.
645 		 */
646 		atomic_dec(&commit_transaction->t_outstanding_credits);
647 
648 		/* Bump b_count to prevent truncate from stumbling over
649                    the shadowed buffer!  @@@ This can go if we ever get
650                    rid of the shadow pairing of buffers. */
651 		atomic_inc(&jh2bh(jh)->b_count);
652 
653 		/*
654 		 * Make a temporary IO buffer with which to write it out
655 		 * (this will requeue the metadata buffer to BJ_Shadow).
656 		 */
657 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
658 		JBUFFER_TRACE(jh, "ph3: write metadata");
659 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
660 						jh, &wbuf[bufs], blocknr);
661 		if (flags < 0) {
662 			jbd2_journal_abort(journal, flags);
663 			continue;
664 		}
665 		jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
666 
667 		/* Record the new block's tag in the current descriptor
668                    buffer */
669 
670 		tag_flag = 0;
671 		if (flags & 1)
672 			tag_flag |= JBD2_FLAG_ESCAPE;
673 		if (!first_tag)
674 			tag_flag |= JBD2_FLAG_SAME_UUID;
675 
676 		tag = (journal_block_tag_t *) tagp;
677 		write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
678 		tag->t_flags = cpu_to_be16(tag_flag);
679 		jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
680 					commit_transaction->t_tid);
681 		tagp += tag_bytes;
682 		space_left -= tag_bytes;
683 		bufs++;
684 
685 		if (first_tag) {
686 			memcpy (tagp, journal->j_uuid, 16);
687 			tagp += 16;
688 			space_left -= 16;
689 			first_tag = 0;
690 		}
691 
692 		/* If there's no more to do, or if the descriptor is full,
693 		   let the IO rip! */
694 
695 		if (bufs == journal->j_wbufsize ||
696 		    commit_transaction->t_buffers == NULL ||
697 		    space_left < tag_bytes + 16 + csum_size) {
698 
699 			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
700 
701 			/* Write an end-of-descriptor marker before
702                            submitting the IOs.  "tag" still points to
703                            the last tag we set up. */
704 
705 			tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
706 start_journal_io:
707 			if (descriptor)
708 				jbd2_descriptor_block_csum_set(journal,
709 							descriptor);
710 
711 			for (i = 0; i < bufs; i++) {
712 				struct buffer_head *bh = wbuf[i];
713 				/*
714 				 * Compute checksum.
715 				 */
716 				if (jbd2_has_feature_checksum(journal)) {
717 					crc32_sum =
718 					    jbd2_checksum_data(crc32_sum, bh);
719 				}
720 
721 				lock_buffer(bh);
722 				clear_buffer_dirty(bh);
723 				set_buffer_uptodate(bh);
724 				bh->b_end_io = journal_end_buffer_io_sync;
725 				submit_bh(REQ_OP_WRITE, REQ_SYNC, bh);
726 			}
727 			cond_resched();
728 
729 			/* Force a new descriptor to be generated next
730                            time round the loop. */
731 			descriptor = NULL;
732 			bufs = 0;
733 		}
734 	}
735 
736 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
737 	if (err) {
738 		printk(KERN_WARNING
739 			"JBD2: Detected IO errors while flushing file data "
740 		       "on %s\n", journal->j_devname);
741 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
742 			jbd2_journal_abort(journal, err);
743 		err = 0;
744 	}
745 
746 	/*
747 	 * Get current oldest transaction in the log before we issue flush
748 	 * to the filesystem device. After the flush we can be sure that
749 	 * blocks of all older transactions are checkpointed to persistent
750 	 * storage and we will be safe to update journal start in the
751 	 * superblock with the numbers we get here.
752 	 */
753 	update_tail =
754 		jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
755 
756 	write_lock(&journal->j_state_lock);
757 	if (update_tail) {
758 		long freed = first_block - journal->j_tail;
759 
760 		if (first_block < journal->j_tail)
761 			freed += journal->j_last - journal->j_first;
762 		/* Update tail only if we free significant amount of space */
763 		if (freed < journal->j_maxlen / 4)
764 			update_tail = 0;
765 	}
766 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
767 	commit_transaction->t_state = T_COMMIT_DFLUSH;
768 	write_unlock(&journal->j_state_lock);
769 
770 	/*
771 	 * If the journal is not located on the file system device,
772 	 * then we must flush the file system device before we issue
773 	 * the commit record
774 	 */
775 	if (commit_transaction->t_need_data_flush &&
776 	    (journal->j_fs_dev != journal->j_dev) &&
777 	    (journal->j_flags & JBD2_BARRIER))
778 		blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
779 
780 	/* Done it all: now write the commit record asynchronously. */
781 	if (jbd2_has_feature_async_commit(journal)) {
782 		err = journal_submit_commit_record(journal, commit_transaction,
783 						 &cbh, crc32_sum);
784 		if (err)
785 			__jbd2_journal_abort_hard(journal);
786 	}
787 
788 	blk_finish_plug(&plug);
789 
790 	/* Lo and behold: we have just managed to send a transaction to
791            the log.  Before we can commit it, wait for the IO so far to
792            complete.  Control buffers being written are on the
793            transaction's t_log_list queue, and metadata buffers are on
794            the io_bufs list.
795 
796 	   Wait for the buffers in reverse order.  That way we are
797 	   less likely to be woken up until all IOs have completed, and
798 	   so we incur less scheduling load.
799 	*/
800 
801 	jbd_debug(3, "JBD2: commit phase 3\n");
802 
803 	while (!list_empty(&io_bufs)) {
804 		struct buffer_head *bh = list_entry(io_bufs.prev,
805 						    struct buffer_head,
806 						    b_assoc_buffers);
807 
808 		wait_on_buffer(bh);
809 		cond_resched();
810 
811 		if (unlikely(!buffer_uptodate(bh)))
812 			err = -EIO;
813 		jbd2_unfile_log_bh(bh);
814 		stats.run.rs_blocks_logged++;
815 
816 		/*
817 		 * The list contains temporary buffer heads created by
818 		 * jbd2_journal_write_metadata_buffer().
819 		 */
820 		BUFFER_TRACE(bh, "dumping temporary bh");
821 		__brelse(bh);
822 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
823 		free_buffer_head(bh);
824 
825 		/* We also have to refile the corresponding shadowed buffer */
826 		jh = commit_transaction->t_shadow_list->b_tprev;
827 		bh = jh2bh(jh);
828 		clear_buffer_jwrite(bh);
829 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
830 		J_ASSERT_BH(bh, !buffer_shadow(bh));
831 
832 		/* The metadata is now released for reuse, but we need
833                    to remember it against this transaction so that when
834                    we finally commit, we can do any checkpointing
835                    required. */
836 		JBUFFER_TRACE(jh, "file as BJ_Forget");
837 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
838 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
839 		__brelse(bh);
840 	}
841 
842 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
843 
844 	jbd_debug(3, "JBD2: commit phase 4\n");
845 
846 	/* Here we wait for the revoke record and descriptor record buffers */
847 	while (!list_empty(&log_bufs)) {
848 		struct buffer_head *bh;
849 
850 		bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
851 		wait_on_buffer(bh);
852 		cond_resched();
853 
854 		if (unlikely(!buffer_uptodate(bh)))
855 			err = -EIO;
856 
857 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
858 		clear_buffer_jwrite(bh);
859 		jbd2_unfile_log_bh(bh);
860 		stats.run.rs_blocks_logged++;
861 		__brelse(bh);		/* One for getblk */
862 		/* AKPM: bforget here */
863 	}
864 
865 	if (err)
866 		jbd2_journal_abort(journal, err);
867 
868 	jbd_debug(3, "JBD2: commit phase 5\n");
869 	write_lock(&journal->j_state_lock);
870 	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
871 	commit_transaction->t_state = T_COMMIT_JFLUSH;
872 	write_unlock(&journal->j_state_lock);
873 
874 	if (!jbd2_has_feature_async_commit(journal)) {
875 		err = journal_submit_commit_record(journal, commit_transaction,
876 						&cbh, crc32_sum);
877 		if (err)
878 			__jbd2_journal_abort_hard(journal);
879 	}
880 	if (cbh)
881 		err = journal_wait_on_commit_record(journal, cbh);
882 	stats.run.rs_blocks_logged++;
883 	if (jbd2_has_feature_async_commit(journal) &&
884 	    journal->j_flags & JBD2_BARRIER) {
885 		blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
886 	}
887 
888 	if (err)
889 		jbd2_journal_abort(journal, err);
890 
891 	WARN_ON_ONCE(
892 		atomic_read(&commit_transaction->t_outstanding_credits) < 0);
893 
894 	/*
895 	 * Now disk caches for filesystem device are flushed so we are safe to
896 	 * erase checkpointed transactions from the log by updating journal
897 	 * superblock.
898 	 */
899 	if (update_tail)
900 		jbd2_update_log_tail(journal, first_tid, first_block);
901 
902 	/* End of a transaction!  Finally, we can do checkpoint
903            processing: any buffers committed as a result of this
904            transaction can be removed from any checkpoint list it was on
905            before. */
906 
907 	jbd_debug(3, "JBD2: commit phase 6\n");
908 
909 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
910 	J_ASSERT(commit_transaction->t_buffers == NULL);
911 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
912 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
913 
914 restart_loop:
915 	/*
916 	 * As there are other places (journal_unmap_buffer()) adding buffers
917 	 * to this list we have to be careful and hold the j_list_lock.
918 	 */
919 	spin_lock(&journal->j_list_lock);
920 	while (commit_transaction->t_forget) {
921 		transaction_t *cp_transaction;
922 		struct buffer_head *bh;
923 		int try_to_free = 0;
924 		bool drop_ref;
925 
926 		jh = commit_transaction->t_forget;
927 		spin_unlock(&journal->j_list_lock);
928 		bh = jh2bh(jh);
929 		/*
930 		 * Get a reference so that bh cannot be freed before we are
931 		 * done with it.
932 		 */
933 		get_bh(bh);
934 		spin_lock(&jh->b_state_lock);
935 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
936 
937 		/*
938 		 * If there is undo-protected committed data against
939 		 * this buffer, then we can remove it now.  If it is a
940 		 * buffer needing such protection, the old frozen_data
941 		 * field now points to a committed version of the
942 		 * buffer, so rotate that field to the new committed
943 		 * data.
944 		 *
945 		 * Otherwise, we can just throw away the frozen data now.
946 		 *
947 		 * We also know that the frozen data has already fired
948 		 * its triggers if they exist, so we can clear that too.
949 		 */
950 		if (jh->b_committed_data) {
951 			jbd2_free(jh->b_committed_data, bh->b_size);
952 			jh->b_committed_data = NULL;
953 			if (jh->b_frozen_data) {
954 				jh->b_committed_data = jh->b_frozen_data;
955 				jh->b_frozen_data = NULL;
956 				jh->b_frozen_triggers = NULL;
957 			}
958 		} else if (jh->b_frozen_data) {
959 			jbd2_free(jh->b_frozen_data, bh->b_size);
960 			jh->b_frozen_data = NULL;
961 			jh->b_frozen_triggers = NULL;
962 		}
963 
964 		spin_lock(&journal->j_list_lock);
965 		cp_transaction = jh->b_cp_transaction;
966 		if (cp_transaction) {
967 			JBUFFER_TRACE(jh, "remove from old cp transaction");
968 			cp_transaction->t_chp_stats.cs_dropped++;
969 			__jbd2_journal_remove_checkpoint(jh);
970 		}
971 
972 		/* Only re-checkpoint the buffer_head if it is marked
973 		 * dirty.  If the buffer was added to the BJ_Forget list
974 		 * by jbd2_journal_forget, it may no longer be dirty and
975 		 * there's no point in keeping a checkpoint record for
976 		 * it. */
977 
978 		/*
979 		* A buffer which has been freed while still being journaled by
980 		* a previous transaction.
981 		*/
982 		if (buffer_freed(bh)) {
983 			/*
984 			 * If the running transaction is the one containing
985 			 * "add to orphan" operation (b_next_transaction !=
986 			 * NULL), we have to wait for that transaction to
987 			 * commit before we can really get rid of the buffer.
988 			 * So just clear b_modified to not confuse transaction
989 			 * credit accounting and refile the buffer to
990 			 * BJ_Forget of the running transaction. If the just
991 			 * committed transaction contains "add to orphan"
992 			 * operation, we can completely invalidate the buffer
993 			 * now. We are rather through in that since the
994 			 * buffer may be still accessible when blocksize <
995 			 * pagesize and it is attached to the last partial
996 			 * page.
997 			 */
998 			jh->b_modified = 0;
999 			if (!jh->b_next_transaction) {
1000 				clear_buffer_freed(bh);
1001 				clear_buffer_jbddirty(bh);
1002 				clear_buffer_mapped(bh);
1003 				clear_buffer_new(bh);
1004 				clear_buffer_req(bh);
1005 				bh->b_bdev = NULL;
1006 			}
1007 		}
1008 
1009 		if (buffer_jbddirty(bh)) {
1010 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
1011 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
1012 			if (is_journal_aborted(journal))
1013 				clear_buffer_jbddirty(bh);
1014 		} else {
1015 			J_ASSERT_BH(bh, !buffer_dirty(bh));
1016 			/*
1017 			 * The buffer on BJ_Forget list and not jbddirty means
1018 			 * it has been freed by this transaction and hence it
1019 			 * could not have been reallocated until this
1020 			 * transaction has committed. *BUT* it could be
1021 			 * reallocated once we have written all the data to
1022 			 * disk and before we process the buffer on BJ_Forget
1023 			 * list.
1024 			 */
1025 			if (!jh->b_next_transaction)
1026 				try_to_free = 1;
1027 		}
1028 		JBUFFER_TRACE(jh, "refile or unfile buffer");
1029 		drop_ref = __jbd2_journal_refile_buffer(jh);
1030 		spin_unlock(&jh->b_state_lock);
1031 		if (drop_ref)
1032 			jbd2_journal_put_journal_head(jh);
1033 		if (try_to_free)
1034 			release_buffer_page(bh);	/* Drops bh reference */
1035 		else
1036 			__brelse(bh);
1037 		cond_resched_lock(&journal->j_list_lock);
1038 	}
1039 	spin_unlock(&journal->j_list_lock);
1040 	/*
1041 	 * This is a bit sleazy.  We use j_list_lock to protect transition
1042 	 * of a transaction into T_FINISHED state and calling
1043 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1044 	 * other checkpointing code processing the transaction...
1045 	 */
1046 	write_lock(&journal->j_state_lock);
1047 	spin_lock(&journal->j_list_lock);
1048 	/*
1049 	 * Now recheck if some buffers did not get attached to the transaction
1050 	 * while the lock was dropped...
1051 	 */
1052 	if (commit_transaction->t_forget) {
1053 		spin_unlock(&journal->j_list_lock);
1054 		write_unlock(&journal->j_state_lock);
1055 		goto restart_loop;
1056 	}
1057 
1058 	/* Add the transaction to the checkpoint list
1059 	 * __journal_remove_checkpoint() can not destroy transaction
1060 	 * under us because it is not marked as T_FINISHED yet */
1061 	if (journal->j_checkpoint_transactions == NULL) {
1062 		journal->j_checkpoint_transactions = commit_transaction;
1063 		commit_transaction->t_cpnext = commit_transaction;
1064 		commit_transaction->t_cpprev = commit_transaction;
1065 	} else {
1066 		commit_transaction->t_cpnext =
1067 			journal->j_checkpoint_transactions;
1068 		commit_transaction->t_cpprev =
1069 			commit_transaction->t_cpnext->t_cpprev;
1070 		commit_transaction->t_cpnext->t_cpprev =
1071 			commit_transaction;
1072 		commit_transaction->t_cpprev->t_cpnext =
1073 				commit_transaction;
1074 	}
1075 	spin_unlock(&journal->j_list_lock);
1076 
1077 	/* Done with this transaction! */
1078 
1079 	jbd_debug(3, "JBD2: commit phase 7\n");
1080 
1081 	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1082 
1083 	commit_transaction->t_start = jiffies;
1084 	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1085 					      commit_transaction->t_start);
1086 
1087 	/*
1088 	 * File the transaction statistics
1089 	 */
1090 	stats.ts_tid = commit_transaction->t_tid;
1091 	stats.run.rs_handle_count =
1092 		atomic_read(&commit_transaction->t_handle_count);
1093 	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1094 			     commit_transaction->t_tid, &stats.run);
1095 	stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1096 
1097 	commit_transaction->t_state = T_COMMIT_CALLBACK;
1098 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1099 	journal->j_commit_sequence = commit_transaction->t_tid;
1100 	journal->j_committing_transaction = NULL;
1101 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1102 
1103 	/*
1104 	 * weight the commit time higher than the average time so we don't
1105 	 * react too strongly to vast changes in the commit time
1106 	 */
1107 	if (likely(journal->j_average_commit_time))
1108 		journal->j_average_commit_time = (commit_time +
1109 				journal->j_average_commit_time*3) / 4;
1110 	else
1111 		journal->j_average_commit_time = commit_time;
1112 
1113 	write_unlock(&journal->j_state_lock);
1114 
1115 	if (journal->j_commit_callback)
1116 		journal->j_commit_callback(journal, commit_transaction);
1117 
1118 	trace_jbd2_end_commit(journal, commit_transaction);
1119 	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1120 		  journal->j_commit_sequence, journal->j_tail_sequence);
1121 
1122 	write_lock(&journal->j_state_lock);
1123 	spin_lock(&journal->j_list_lock);
1124 	commit_transaction->t_state = T_FINISHED;
1125 	/* Check if the transaction can be dropped now that we are finished */
1126 	if (commit_transaction->t_checkpoint_list == NULL &&
1127 	    commit_transaction->t_checkpoint_io_list == NULL) {
1128 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1129 		jbd2_journal_free_transaction(commit_transaction);
1130 	}
1131 	spin_unlock(&journal->j_list_lock);
1132 	write_unlock(&journal->j_state_lock);
1133 	wake_up(&journal->j_wait_done_commit);
1134 
1135 	/*
1136 	 * Calculate overall stats
1137 	 */
1138 	spin_lock(&journal->j_history_lock);
1139 	journal->j_stats.ts_tid++;
1140 	journal->j_stats.ts_requested += stats.ts_requested;
1141 	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1142 	journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1143 	journal->j_stats.run.rs_running += stats.run.rs_running;
1144 	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1145 	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1146 	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1147 	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1148 	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1149 	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1150 	spin_unlock(&journal->j_history_lock);
1151 }
1152