xref: /linux/fs/jbd2/commit.c (revision be3de80dc2e671d9ee15e69fe9cd84d2b71e2225)
1 /*
2  * linux/fs/jbd2/commit.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Journal commit routines for the generic filesystem journaling code;
13  * part of the ext2fs journaling system.
14  */
15 
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/crc32.h>
25 #include <linux/writeback.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bio.h>
28 #include <linux/blkdev.h>
29 #include <linux/bitops.h>
30 #include <trace/events/jbd2.h>
31 #include <asm/system.h>
32 
33 /*
34  * Default IO end handler for temporary BJ_IO buffer_heads.
35  */
36 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
37 {
38 	BUFFER_TRACE(bh, "");
39 	if (uptodate)
40 		set_buffer_uptodate(bh);
41 	else
42 		clear_buffer_uptodate(bh);
43 	unlock_buffer(bh);
44 }
45 
46 /*
47  * When an ext4 file is truncated, it is possible that some pages are not
48  * successfully freed, because they are attached to a committing transaction.
49  * After the transaction commits, these pages are left on the LRU, with no
50  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
51  * by the VM, but their apparent absence upsets the VM accounting, and it makes
52  * the numbers in /proc/meminfo look odd.
53  *
54  * So here, we have a buffer which has just come off the forget list.  Look to
55  * see if we can strip all buffers from the backing page.
56  *
57  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
58  * caller provided us with a ref against the buffer, and we drop that here.
59  */
60 static void release_buffer_page(struct buffer_head *bh)
61 {
62 	struct page *page;
63 
64 	if (buffer_dirty(bh))
65 		goto nope;
66 	if (atomic_read(&bh->b_count) != 1)
67 		goto nope;
68 	page = bh->b_page;
69 	if (!page)
70 		goto nope;
71 	if (page->mapping)
72 		goto nope;
73 
74 	/* OK, it's a truncated page */
75 	if (!trylock_page(page))
76 		goto nope;
77 
78 	page_cache_get(page);
79 	__brelse(bh);
80 	try_to_free_buffers(page);
81 	unlock_page(page);
82 	page_cache_release(page);
83 	return;
84 
85 nope:
86 	__brelse(bh);
87 }
88 
89 /*
90  * Done it all: now submit the commit record.  We should have
91  * cleaned up our previous buffers by now, so if we are in abort
92  * mode we can now just skip the rest of the journal write
93  * entirely.
94  *
95  * Returns 1 if the journal needs to be aborted or 0 on success
96  */
97 static int journal_submit_commit_record(journal_t *journal,
98 					transaction_t *commit_transaction,
99 					struct buffer_head **cbh,
100 					__u32 crc32_sum)
101 {
102 	struct journal_head *descriptor;
103 	struct commit_header *tmp;
104 	struct buffer_head *bh;
105 	int ret;
106 	struct timespec now = current_kernel_time();
107 
108 	*cbh = NULL;
109 
110 	if (is_journal_aborted(journal))
111 		return 0;
112 
113 	descriptor = jbd2_journal_get_descriptor_buffer(journal);
114 	if (!descriptor)
115 		return 1;
116 
117 	bh = jh2bh(descriptor);
118 
119 	tmp = (struct commit_header *)bh->b_data;
120 	tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
121 	tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
122 	tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
123 	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
124 	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
125 
126 	if (JBD2_HAS_COMPAT_FEATURE(journal,
127 				    JBD2_FEATURE_COMPAT_CHECKSUM)) {
128 		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
129 		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
130 		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
131 	}
132 
133 	JBUFFER_TRACE(descriptor, "submit commit block");
134 	lock_buffer(bh);
135 	clear_buffer_dirty(bh);
136 	set_buffer_uptodate(bh);
137 	bh->b_end_io = journal_end_buffer_io_sync;
138 
139 	if (journal->j_flags & JBD2_BARRIER &&
140 	    !JBD2_HAS_INCOMPAT_FEATURE(journal,
141 				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
142 		ret = submit_bh(WRITE_SYNC | WRITE_FLUSH_FUA, bh);
143 	else
144 		ret = submit_bh(WRITE_SYNC, bh);
145 
146 	*cbh = bh;
147 	return ret;
148 }
149 
150 /*
151  * This function along with journal_submit_commit_record
152  * allows to write the commit record asynchronously.
153  */
154 static int journal_wait_on_commit_record(journal_t *journal,
155 					 struct buffer_head *bh)
156 {
157 	int ret = 0;
158 
159 	clear_buffer_dirty(bh);
160 	wait_on_buffer(bh);
161 
162 	if (unlikely(!buffer_uptodate(bh)))
163 		ret = -EIO;
164 	put_bh(bh);            /* One for getblk() */
165 	jbd2_journal_put_journal_head(bh2jh(bh));
166 
167 	return ret;
168 }
169 
170 /*
171  * write the filemap data using writepage() address_space_operations.
172  * We don't do block allocation here even for delalloc. We don't
173  * use writepages() because with dealyed allocation we may be doing
174  * block allocation in writepages().
175  */
176 static int journal_submit_inode_data_buffers(struct address_space *mapping)
177 {
178 	int ret;
179 	struct writeback_control wbc = {
180 		.sync_mode =  WB_SYNC_ALL,
181 		.nr_to_write = mapping->nrpages * 2,
182 		.range_start = 0,
183 		.range_end = i_size_read(mapping->host),
184 	};
185 
186 	ret = generic_writepages(mapping, &wbc);
187 	return ret;
188 }
189 
190 /*
191  * Submit all the data buffers of inode associated with the transaction to
192  * disk.
193  *
194  * We are in a committing transaction. Therefore no new inode can be added to
195  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
196  * operate on from being released while we write out pages.
197  */
198 static int journal_submit_data_buffers(journal_t *journal,
199 		transaction_t *commit_transaction)
200 {
201 	struct jbd2_inode *jinode;
202 	int err, ret = 0;
203 	struct address_space *mapping;
204 
205 	spin_lock(&journal->j_list_lock);
206 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
207 		mapping = jinode->i_vfs_inode->i_mapping;
208 		set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
209 		spin_unlock(&journal->j_list_lock);
210 		/*
211 		 * submit the inode data buffers. We use writepage
212 		 * instead of writepages. Because writepages can do
213 		 * block allocation  with delalloc. We need to write
214 		 * only allocated blocks here.
215 		 */
216 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
217 		err = journal_submit_inode_data_buffers(mapping);
218 		if (!ret)
219 			ret = err;
220 		spin_lock(&journal->j_list_lock);
221 		J_ASSERT(jinode->i_transaction == commit_transaction);
222 		clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
223 		smp_mb__after_clear_bit();
224 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
225 	}
226 	spin_unlock(&journal->j_list_lock);
227 	return ret;
228 }
229 
230 /*
231  * Wait for data submitted for writeout, refile inodes to proper
232  * transaction if needed.
233  *
234  */
235 static int journal_finish_inode_data_buffers(journal_t *journal,
236 		transaction_t *commit_transaction)
237 {
238 	struct jbd2_inode *jinode, *next_i;
239 	int err, ret = 0;
240 
241 	/* For locking, see the comment in journal_submit_data_buffers() */
242 	spin_lock(&journal->j_list_lock);
243 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
244 		set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
245 		spin_unlock(&journal->j_list_lock);
246 		err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
247 		if (err) {
248 			/*
249 			 * Because AS_EIO is cleared by
250 			 * filemap_fdatawait_range(), set it again so
251 			 * that user process can get -EIO from fsync().
252 			 */
253 			set_bit(AS_EIO,
254 				&jinode->i_vfs_inode->i_mapping->flags);
255 
256 			if (!ret)
257 				ret = err;
258 		}
259 		spin_lock(&journal->j_list_lock);
260 		clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
261 		smp_mb__after_clear_bit();
262 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
263 	}
264 
265 	/* Now refile inode to proper lists */
266 	list_for_each_entry_safe(jinode, next_i,
267 				 &commit_transaction->t_inode_list, i_list) {
268 		list_del(&jinode->i_list);
269 		if (jinode->i_next_transaction) {
270 			jinode->i_transaction = jinode->i_next_transaction;
271 			jinode->i_next_transaction = NULL;
272 			list_add(&jinode->i_list,
273 				&jinode->i_transaction->t_inode_list);
274 		} else {
275 			jinode->i_transaction = NULL;
276 		}
277 	}
278 	spin_unlock(&journal->j_list_lock);
279 
280 	return ret;
281 }
282 
283 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
284 {
285 	struct page *page = bh->b_page;
286 	char *addr;
287 	__u32 checksum;
288 
289 	addr = kmap_atomic(page, KM_USER0);
290 	checksum = crc32_be(crc32_sum,
291 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
292 	kunmap_atomic(addr, KM_USER0);
293 
294 	return checksum;
295 }
296 
297 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
298 				   unsigned long long block)
299 {
300 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
301 	if (tag_bytes > JBD2_TAG_SIZE32)
302 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
303 }
304 
305 /*
306  * jbd2_journal_commit_transaction
307  *
308  * The primary function for committing a transaction to the log.  This
309  * function is called by the journal thread to begin a complete commit.
310  */
311 void jbd2_journal_commit_transaction(journal_t *journal)
312 {
313 	struct transaction_stats_s stats;
314 	transaction_t *commit_transaction;
315 	struct journal_head *jh, *new_jh, *descriptor;
316 	struct buffer_head **wbuf = journal->j_wbuf;
317 	int bufs;
318 	int flags;
319 	int err;
320 	unsigned long long blocknr;
321 	ktime_t start_time;
322 	u64 commit_time;
323 	char *tagp = NULL;
324 	journal_header_t *header;
325 	journal_block_tag_t *tag = NULL;
326 	int space_left = 0;
327 	int first_tag = 0;
328 	int tag_flag;
329 	int i, to_free = 0;
330 	int tag_bytes = journal_tag_bytes(journal);
331 	struct buffer_head *cbh = NULL; /* For transactional checksums */
332 	__u32 crc32_sum = ~0;
333 	struct blk_plug plug;
334 
335 	/*
336 	 * First job: lock down the current transaction and wait for
337 	 * all outstanding updates to complete.
338 	 */
339 
340 	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
341 	if (journal->j_flags & JBD2_FLUSHED) {
342 		jbd_debug(3, "super block updated\n");
343 		jbd2_journal_update_superblock(journal, 1);
344 	} else {
345 		jbd_debug(3, "superblock not updated\n");
346 	}
347 
348 	J_ASSERT(journal->j_running_transaction != NULL);
349 	J_ASSERT(journal->j_committing_transaction == NULL);
350 
351 	commit_transaction = journal->j_running_transaction;
352 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
353 
354 	trace_jbd2_start_commit(journal, commit_transaction);
355 	jbd_debug(1, "JBD2: starting commit of transaction %d\n",
356 			commit_transaction->t_tid);
357 
358 	write_lock(&journal->j_state_lock);
359 	commit_transaction->t_state = T_LOCKED;
360 
361 	trace_jbd2_commit_locking(journal, commit_transaction);
362 	stats.run.rs_wait = commit_transaction->t_max_wait;
363 	stats.run.rs_locked = jiffies;
364 	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
365 					      stats.run.rs_locked);
366 
367 	spin_lock(&commit_transaction->t_handle_lock);
368 	while (atomic_read(&commit_transaction->t_updates)) {
369 		DEFINE_WAIT(wait);
370 
371 		prepare_to_wait(&journal->j_wait_updates, &wait,
372 					TASK_UNINTERRUPTIBLE);
373 		if (atomic_read(&commit_transaction->t_updates)) {
374 			spin_unlock(&commit_transaction->t_handle_lock);
375 			write_unlock(&journal->j_state_lock);
376 			schedule();
377 			write_lock(&journal->j_state_lock);
378 			spin_lock(&commit_transaction->t_handle_lock);
379 		}
380 		finish_wait(&journal->j_wait_updates, &wait);
381 	}
382 	spin_unlock(&commit_transaction->t_handle_lock);
383 
384 	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
385 			journal->j_max_transaction_buffers);
386 
387 	/*
388 	 * First thing we are allowed to do is to discard any remaining
389 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
390 	 * that there are no such buffers: if a large filesystem
391 	 * operation like a truncate needs to split itself over multiple
392 	 * transactions, then it may try to do a jbd2_journal_restart() while
393 	 * there are still BJ_Reserved buffers outstanding.  These must
394 	 * be released cleanly from the current transaction.
395 	 *
396 	 * In this case, the filesystem must still reserve write access
397 	 * again before modifying the buffer in the new transaction, but
398 	 * we do not require it to remember exactly which old buffers it
399 	 * has reserved.  This is consistent with the existing behaviour
400 	 * that multiple jbd2_journal_get_write_access() calls to the same
401 	 * buffer are perfectly permissible.
402 	 */
403 	while (commit_transaction->t_reserved_list) {
404 		jh = commit_transaction->t_reserved_list;
405 		JBUFFER_TRACE(jh, "reserved, unused: refile");
406 		/*
407 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
408 		 * leave undo-committed data.
409 		 */
410 		if (jh->b_committed_data) {
411 			struct buffer_head *bh = jh2bh(jh);
412 
413 			jbd_lock_bh_state(bh);
414 			jbd2_free(jh->b_committed_data, bh->b_size);
415 			jh->b_committed_data = NULL;
416 			jbd_unlock_bh_state(bh);
417 		}
418 		jbd2_journal_refile_buffer(journal, jh);
419 	}
420 
421 	/*
422 	 * Now try to drop any written-back buffers from the journal's
423 	 * checkpoint lists.  We do this *before* commit because it potentially
424 	 * frees some memory
425 	 */
426 	spin_lock(&journal->j_list_lock);
427 	__jbd2_journal_clean_checkpoint_list(journal);
428 	spin_unlock(&journal->j_list_lock);
429 
430 	jbd_debug(3, "JBD2: commit phase 1\n");
431 
432 	/*
433 	 * Switch to a new revoke table.
434 	 */
435 	jbd2_journal_switch_revoke_table(journal);
436 
437 	trace_jbd2_commit_flushing(journal, commit_transaction);
438 	stats.run.rs_flushing = jiffies;
439 	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
440 					     stats.run.rs_flushing);
441 
442 	commit_transaction->t_state = T_FLUSH;
443 	journal->j_committing_transaction = commit_transaction;
444 	journal->j_running_transaction = NULL;
445 	start_time = ktime_get();
446 	commit_transaction->t_log_start = journal->j_head;
447 	wake_up(&journal->j_wait_transaction_locked);
448 	write_unlock(&journal->j_state_lock);
449 
450 	jbd_debug(3, "JBD2: commit phase 2\n");
451 
452 	/*
453 	 * Now start flushing things to disk, in the order they appear
454 	 * on the transaction lists.  Data blocks go first.
455 	 */
456 	err = journal_submit_data_buffers(journal, commit_transaction);
457 	if (err)
458 		jbd2_journal_abort(journal, err);
459 
460 	blk_start_plug(&plug);
461 	jbd2_journal_write_revoke_records(journal, commit_transaction,
462 					  WRITE_SYNC);
463 	blk_finish_plug(&plug);
464 
465 	jbd_debug(3, "JBD2: commit phase 2\n");
466 
467 	/*
468 	 * Way to go: we have now written out all of the data for a
469 	 * transaction!  Now comes the tricky part: we need to write out
470 	 * metadata.  Loop over the transaction's entire buffer list:
471 	 */
472 	write_lock(&journal->j_state_lock);
473 	commit_transaction->t_state = T_COMMIT;
474 	write_unlock(&journal->j_state_lock);
475 
476 	trace_jbd2_commit_logging(journal, commit_transaction);
477 	stats.run.rs_logging = jiffies;
478 	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
479 					       stats.run.rs_logging);
480 	stats.run.rs_blocks =
481 		atomic_read(&commit_transaction->t_outstanding_credits);
482 	stats.run.rs_blocks_logged = 0;
483 
484 	J_ASSERT(commit_transaction->t_nr_buffers <=
485 		 atomic_read(&commit_transaction->t_outstanding_credits));
486 
487 	err = 0;
488 	descriptor = NULL;
489 	bufs = 0;
490 	blk_start_plug(&plug);
491 	while (commit_transaction->t_buffers) {
492 
493 		/* Find the next buffer to be journaled... */
494 
495 		jh = commit_transaction->t_buffers;
496 
497 		/* If we're in abort mode, we just un-journal the buffer and
498 		   release it. */
499 
500 		if (is_journal_aborted(journal)) {
501 			clear_buffer_jbddirty(jh2bh(jh));
502 			JBUFFER_TRACE(jh, "journal is aborting: refile");
503 			jbd2_buffer_abort_trigger(jh,
504 						  jh->b_frozen_data ?
505 						  jh->b_frozen_triggers :
506 						  jh->b_triggers);
507 			jbd2_journal_refile_buffer(journal, jh);
508 			/* If that was the last one, we need to clean up
509 			 * any descriptor buffers which may have been
510 			 * already allocated, even if we are now
511 			 * aborting. */
512 			if (!commit_transaction->t_buffers)
513 				goto start_journal_io;
514 			continue;
515 		}
516 
517 		/* Make sure we have a descriptor block in which to
518 		   record the metadata buffer. */
519 
520 		if (!descriptor) {
521 			struct buffer_head *bh;
522 
523 			J_ASSERT (bufs == 0);
524 
525 			jbd_debug(4, "JBD2: get descriptor\n");
526 
527 			descriptor = jbd2_journal_get_descriptor_buffer(journal);
528 			if (!descriptor) {
529 				jbd2_journal_abort(journal, -EIO);
530 				continue;
531 			}
532 
533 			bh = jh2bh(descriptor);
534 			jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
535 				(unsigned long long)bh->b_blocknr, bh->b_data);
536 			header = (journal_header_t *)&bh->b_data[0];
537 			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
538 			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
539 			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
540 
541 			tagp = &bh->b_data[sizeof(journal_header_t)];
542 			space_left = bh->b_size - sizeof(journal_header_t);
543 			first_tag = 1;
544 			set_buffer_jwrite(bh);
545 			set_buffer_dirty(bh);
546 			wbuf[bufs++] = bh;
547 
548 			/* Record it so that we can wait for IO
549                            completion later */
550 			BUFFER_TRACE(bh, "ph3: file as descriptor");
551 			jbd2_journal_file_buffer(descriptor, commit_transaction,
552 					BJ_LogCtl);
553 		}
554 
555 		/* Where is the buffer to be written? */
556 
557 		err = jbd2_journal_next_log_block(journal, &blocknr);
558 		/* If the block mapping failed, just abandon the buffer
559 		   and repeat this loop: we'll fall into the
560 		   refile-on-abort condition above. */
561 		if (err) {
562 			jbd2_journal_abort(journal, err);
563 			continue;
564 		}
565 
566 		/*
567 		 * start_this_handle() uses t_outstanding_credits to determine
568 		 * the free space in the log, but this counter is changed
569 		 * by jbd2_journal_next_log_block() also.
570 		 */
571 		atomic_dec(&commit_transaction->t_outstanding_credits);
572 
573 		/* Bump b_count to prevent truncate from stumbling over
574                    the shadowed buffer!  @@@ This can go if we ever get
575                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
576 		atomic_inc(&jh2bh(jh)->b_count);
577 
578 		/* Make a temporary IO buffer with which to write it out
579                    (this will requeue both the metadata buffer and the
580                    temporary IO buffer). new_bh goes on BJ_IO*/
581 
582 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
583 		/*
584 		 * akpm: jbd2_journal_write_metadata_buffer() sets
585 		 * new_bh->b_transaction to commit_transaction.
586 		 * We need to clean this up before we release new_bh
587 		 * (which is of type BJ_IO)
588 		 */
589 		JBUFFER_TRACE(jh, "ph3: write metadata");
590 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
591 						      jh, &new_jh, blocknr);
592 		if (flags < 0) {
593 			jbd2_journal_abort(journal, flags);
594 			continue;
595 		}
596 		set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
597 		wbuf[bufs++] = jh2bh(new_jh);
598 
599 		/* Record the new block's tag in the current descriptor
600                    buffer */
601 
602 		tag_flag = 0;
603 		if (flags & 1)
604 			tag_flag |= JBD2_FLAG_ESCAPE;
605 		if (!first_tag)
606 			tag_flag |= JBD2_FLAG_SAME_UUID;
607 
608 		tag = (journal_block_tag_t *) tagp;
609 		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
610 		tag->t_flags = cpu_to_be32(tag_flag);
611 		tagp += tag_bytes;
612 		space_left -= tag_bytes;
613 
614 		if (first_tag) {
615 			memcpy (tagp, journal->j_uuid, 16);
616 			tagp += 16;
617 			space_left -= 16;
618 			first_tag = 0;
619 		}
620 
621 		/* If there's no more to do, or if the descriptor is full,
622 		   let the IO rip! */
623 
624 		if (bufs == journal->j_wbufsize ||
625 		    commit_transaction->t_buffers == NULL ||
626 		    space_left < tag_bytes + 16) {
627 
628 			jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
629 
630 			/* Write an end-of-descriptor marker before
631                            submitting the IOs.  "tag" still points to
632                            the last tag we set up. */
633 
634 			tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
635 
636 start_journal_io:
637 			for (i = 0; i < bufs; i++) {
638 				struct buffer_head *bh = wbuf[i];
639 				/*
640 				 * Compute checksum.
641 				 */
642 				if (JBD2_HAS_COMPAT_FEATURE(journal,
643 					JBD2_FEATURE_COMPAT_CHECKSUM)) {
644 					crc32_sum =
645 					    jbd2_checksum_data(crc32_sum, bh);
646 				}
647 
648 				lock_buffer(bh);
649 				clear_buffer_dirty(bh);
650 				set_buffer_uptodate(bh);
651 				bh->b_end_io = journal_end_buffer_io_sync;
652 				submit_bh(WRITE_SYNC, bh);
653 			}
654 			cond_resched();
655 			stats.run.rs_blocks_logged += bufs;
656 
657 			/* Force a new descriptor to be generated next
658                            time round the loop. */
659 			descriptor = NULL;
660 			bufs = 0;
661 		}
662 	}
663 
664 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
665 	if (err) {
666 		printk(KERN_WARNING
667 			"JBD2: Detected IO errors while flushing file data "
668 		       "on %s\n", journal->j_devname);
669 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
670 			jbd2_journal_abort(journal, err);
671 		err = 0;
672 	}
673 
674 	write_lock(&journal->j_state_lock);
675 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
676 	commit_transaction->t_state = T_COMMIT_DFLUSH;
677 	write_unlock(&journal->j_state_lock);
678 	/*
679 	 * If the journal is not located on the file system device,
680 	 * then we must flush the file system device before we issue
681 	 * the commit record
682 	 */
683 	if (commit_transaction->t_need_data_flush &&
684 	    (journal->j_fs_dev != journal->j_dev) &&
685 	    (journal->j_flags & JBD2_BARRIER))
686 		blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
687 
688 	/* Done it all: now write the commit record asynchronously. */
689 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
690 				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
691 		err = journal_submit_commit_record(journal, commit_transaction,
692 						 &cbh, crc32_sum);
693 		if (err)
694 			__jbd2_journal_abort_hard(journal);
695 	}
696 
697 	blk_finish_plug(&plug);
698 
699 	/* Lo and behold: we have just managed to send a transaction to
700            the log.  Before we can commit it, wait for the IO so far to
701            complete.  Control buffers being written are on the
702            transaction's t_log_list queue, and metadata buffers are on
703            the t_iobuf_list queue.
704 
705 	   Wait for the buffers in reverse order.  That way we are
706 	   less likely to be woken up until all IOs have completed, and
707 	   so we incur less scheduling load.
708 	*/
709 
710 	jbd_debug(3, "JBD2: commit phase 3\n");
711 
712 	/*
713 	 * akpm: these are BJ_IO, and j_list_lock is not needed.
714 	 * See __journal_try_to_free_buffer.
715 	 */
716 wait_for_iobuf:
717 	while (commit_transaction->t_iobuf_list != NULL) {
718 		struct buffer_head *bh;
719 
720 		jh = commit_transaction->t_iobuf_list->b_tprev;
721 		bh = jh2bh(jh);
722 		if (buffer_locked(bh)) {
723 			wait_on_buffer(bh);
724 			goto wait_for_iobuf;
725 		}
726 		if (cond_resched())
727 			goto wait_for_iobuf;
728 
729 		if (unlikely(!buffer_uptodate(bh)))
730 			err = -EIO;
731 
732 		clear_buffer_jwrite(bh);
733 
734 		JBUFFER_TRACE(jh, "ph4: unfile after journal write");
735 		jbd2_journal_unfile_buffer(journal, jh);
736 
737 		/*
738 		 * ->t_iobuf_list should contain only dummy buffer_heads
739 		 * which were created by jbd2_journal_write_metadata_buffer().
740 		 */
741 		BUFFER_TRACE(bh, "dumping temporary bh");
742 		jbd2_journal_put_journal_head(jh);
743 		__brelse(bh);
744 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
745 		free_buffer_head(bh);
746 
747 		/* We also have to unlock and free the corresponding
748                    shadowed buffer */
749 		jh = commit_transaction->t_shadow_list->b_tprev;
750 		bh = jh2bh(jh);
751 		clear_bit(BH_JWrite, &bh->b_state);
752 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
753 
754 		/* The metadata is now released for reuse, but we need
755                    to remember it against this transaction so that when
756                    we finally commit, we can do any checkpointing
757                    required. */
758 		JBUFFER_TRACE(jh, "file as BJ_Forget");
759 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
760 		/*
761 		 * Wake up any transactions which were waiting for this IO to
762 		 * complete. The barrier must be here so that changes by
763 		 * jbd2_journal_file_buffer() take effect before wake_up_bit()
764 		 * does the waitqueue check.
765 		 */
766 		smp_mb();
767 		wake_up_bit(&bh->b_state, BH_Unshadow);
768 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
769 		__brelse(bh);
770 	}
771 
772 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
773 
774 	jbd_debug(3, "JBD2: commit phase 4\n");
775 
776 	/* Here we wait for the revoke record and descriptor record buffers */
777  wait_for_ctlbuf:
778 	while (commit_transaction->t_log_list != NULL) {
779 		struct buffer_head *bh;
780 
781 		jh = commit_transaction->t_log_list->b_tprev;
782 		bh = jh2bh(jh);
783 		if (buffer_locked(bh)) {
784 			wait_on_buffer(bh);
785 			goto wait_for_ctlbuf;
786 		}
787 		if (cond_resched())
788 			goto wait_for_ctlbuf;
789 
790 		if (unlikely(!buffer_uptodate(bh)))
791 			err = -EIO;
792 
793 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
794 		clear_buffer_jwrite(bh);
795 		jbd2_journal_unfile_buffer(journal, jh);
796 		jbd2_journal_put_journal_head(jh);
797 		__brelse(bh);		/* One for getblk */
798 		/* AKPM: bforget here */
799 	}
800 
801 	if (err)
802 		jbd2_journal_abort(journal, err);
803 
804 	jbd_debug(3, "JBD2: commit phase 5\n");
805 	write_lock(&journal->j_state_lock);
806 	J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
807 	commit_transaction->t_state = T_COMMIT_JFLUSH;
808 	write_unlock(&journal->j_state_lock);
809 
810 	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
811 				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
812 		err = journal_submit_commit_record(journal, commit_transaction,
813 						&cbh, crc32_sum);
814 		if (err)
815 			__jbd2_journal_abort_hard(journal);
816 	}
817 	if (cbh)
818 		err = journal_wait_on_commit_record(journal, cbh);
819 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
820 				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
821 	    journal->j_flags & JBD2_BARRIER) {
822 		blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL);
823 	}
824 
825 	if (err)
826 		jbd2_journal_abort(journal, err);
827 
828 	/* End of a transaction!  Finally, we can do checkpoint
829            processing: any buffers committed as a result of this
830            transaction can be removed from any checkpoint list it was on
831            before. */
832 
833 	jbd_debug(3, "JBD2: commit phase 6\n");
834 
835 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
836 	J_ASSERT(commit_transaction->t_buffers == NULL);
837 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
838 	J_ASSERT(commit_transaction->t_iobuf_list == NULL);
839 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
840 	J_ASSERT(commit_transaction->t_log_list == NULL);
841 
842 restart_loop:
843 	/*
844 	 * As there are other places (journal_unmap_buffer()) adding buffers
845 	 * to this list we have to be careful and hold the j_list_lock.
846 	 */
847 	spin_lock(&journal->j_list_lock);
848 	while (commit_transaction->t_forget) {
849 		transaction_t *cp_transaction;
850 		struct buffer_head *bh;
851 		int try_to_free = 0;
852 
853 		jh = commit_transaction->t_forget;
854 		spin_unlock(&journal->j_list_lock);
855 		bh = jh2bh(jh);
856 		/*
857 		 * Get a reference so that bh cannot be freed before we are
858 		 * done with it.
859 		 */
860 		get_bh(bh);
861 		jbd_lock_bh_state(bh);
862 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
863 
864 		/*
865 		 * If there is undo-protected committed data against
866 		 * this buffer, then we can remove it now.  If it is a
867 		 * buffer needing such protection, the old frozen_data
868 		 * field now points to a committed version of the
869 		 * buffer, so rotate that field to the new committed
870 		 * data.
871 		 *
872 		 * Otherwise, we can just throw away the frozen data now.
873 		 *
874 		 * We also know that the frozen data has already fired
875 		 * its triggers if they exist, so we can clear that too.
876 		 */
877 		if (jh->b_committed_data) {
878 			jbd2_free(jh->b_committed_data, bh->b_size);
879 			jh->b_committed_data = NULL;
880 			if (jh->b_frozen_data) {
881 				jh->b_committed_data = jh->b_frozen_data;
882 				jh->b_frozen_data = NULL;
883 				jh->b_frozen_triggers = NULL;
884 			}
885 		} else if (jh->b_frozen_data) {
886 			jbd2_free(jh->b_frozen_data, bh->b_size);
887 			jh->b_frozen_data = NULL;
888 			jh->b_frozen_triggers = NULL;
889 		}
890 
891 		spin_lock(&journal->j_list_lock);
892 		cp_transaction = jh->b_cp_transaction;
893 		if (cp_transaction) {
894 			JBUFFER_TRACE(jh, "remove from old cp transaction");
895 			cp_transaction->t_chp_stats.cs_dropped++;
896 			__jbd2_journal_remove_checkpoint(jh);
897 		}
898 
899 		/* Only re-checkpoint the buffer_head if it is marked
900 		 * dirty.  If the buffer was added to the BJ_Forget list
901 		 * by jbd2_journal_forget, it may no longer be dirty and
902 		 * there's no point in keeping a checkpoint record for
903 		 * it. */
904 
905 		/* A buffer which has been freed while still being
906 		 * journaled by a previous transaction may end up still
907 		 * being dirty here, but we want to avoid writing back
908 		 * that buffer in the future after the "add to orphan"
909 		 * operation been committed,  That's not only a performance
910 		 * gain, it also stops aliasing problems if the buffer is
911 		 * left behind for writeback and gets reallocated for another
912 		 * use in a different page. */
913 		if (buffer_freed(bh) && !jh->b_next_transaction) {
914 			clear_buffer_freed(bh);
915 			clear_buffer_jbddirty(bh);
916 		}
917 
918 		if (buffer_jbddirty(bh)) {
919 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
920 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
921 			if (is_journal_aborted(journal))
922 				clear_buffer_jbddirty(bh);
923 		} else {
924 			J_ASSERT_BH(bh, !buffer_dirty(bh));
925 			/*
926 			 * The buffer on BJ_Forget list and not jbddirty means
927 			 * it has been freed by this transaction and hence it
928 			 * could not have been reallocated until this
929 			 * transaction has committed. *BUT* it could be
930 			 * reallocated once we have written all the data to
931 			 * disk and before we process the buffer on BJ_Forget
932 			 * list.
933 			 */
934 			if (!jh->b_next_transaction)
935 				try_to_free = 1;
936 		}
937 		JBUFFER_TRACE(jh, "refile or unfile buffer");
938 		__jbd2_journal_refile_buffer(jh);
939 		jbd_unlock_bh_state(bh);
940 		if (try_to_free)
941 			release_buffer_page(bh);	/* Drops bh reference */
942 		else
943 			__brelse(bh);
944 		cond_resched_lock(&journal->j_list_lock);
945 	}
946 	spin_unlock(&journal->j_list_lock);
947 	/*
948 	 * This is a bit sleazy.  We use j_list_lock to protect transition
949 	 * of a transaction into T_FINISHED state and calling
950 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
951 	 * other checkpointing code processing the transaction...
952 	 */
953 	write_lock(&journal->j_state_lock);
954 	spin_lock(&journal->j_list_lock);
955 	/*
956 	 * Now recheck if some buffers did not get attached to the transaction
957 	 * while the lock was dropped...
958 	 */
959 	if (commit_transaction->t_forget) {
960 		spin_unlock(&journal->j_list_lock);
961 		write_unlock(&journal->j_state_lock);
962 		goto restart_loop;
963 	}
964 
965 	/* Done with this transaction! */
966 
967 	jbd_debug(3, "JBD2: commit phase 7\n");
968 
969 	J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
970 
971 	commit_transaction->t_start = jiffies;
972 	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
973 					      commit_transaction->t_start);
974 
975 	/*
976 	 * File the transaction statistics
977 	 */
978 	stats.ts_tid = commit_transaction->t_tid;
979 	stats.run.rs_handle_count =
980 		atomic_read(&commit_transaction->t_handle_count);
981 	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
982 			     commit_transaction->t_tid, &stats.run);
983 
984 	/*
985 	 * Calculate overall stats
986 	 */
987 	spin_lock(&journal->j_history_lock);
988 	journal->j_stats.ts_tid++;
989 	journal->j_stats.run.rs_wait += stats.run.rs_wait;
990 	journal->j_stats.run.rs_running += stats.run.rs_running;
991 	journal->j_stats.run.rs_locked += stats.run.rs_locked;
992 	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
993 	journal->j_stats.run.rs_logging += stats.run.rs_logging;
994 	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
995 	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
996 	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
997 	spin_unlock(&journal->j_history_lock);
998 
999 	commit_transaction->t_state = T_FINISHED;
1000 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1001 	journal->j_commit_sequence = commit_transaction->t_tid;
1002 	journal->j_committing_transaction = NULL;
1003 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1004 
1005 	/*
1006 	 * weight the commit time higher than the average time so we don't
1007 	 * react too strongly to vast changes in the commit time
1008 	 */
1009 	if (likely(journal->j_average_commit_time))
1010 		journal->j_average_commit_time = (commit_time +
1011 				journal->j_average_commit_time*3) / 4;
1012 	else
1013 		journal->j_average_commit_time = commit_time;
1014 	write_unlock(&journal->j_state_lock);
1015 
1016 	if (commit_transaction->t_checkpoint_list == NULL &&
1017 	    commit_transaction->t_checkpoint_io_list == NULL) {
1018 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1019 		to_free = 1;
1020 	} else {
1021 		if (journal->j_checkpoint_transactions == NULL) {
1022 			journal->j_checkpoint_transactions = commit_transaction;
1023 			commit_transaction->t_cpnext = commit_transaction;
1024 			commit_transaction->t_cpprev = commit_transaction;
1025 		} else {
1026 			commit_transaction->t_cpnext =
1027 				journal->j_checkpoint_transactions;
1028 			commit_transaction->t_cpprev =
1029 				commit_transaction->t_cpnext->t_cpprev;
1030 			commit_transaction->t_cpnext->t_cpprev =
1031 				commit_transaction;
1032 			commit_transaction->t_cpprev->t_cpnext =
1033 				commit_transaction;
1034 		}
1035 	}
1036 	spin_unlock(&journal->j_list_lock);
1037 
1038 	if (journal->j_commit_callback)
1039 		journal->j_commit_callback(journal, commit_transaction);
1040 
1041 	trace_jbd2_end_commit(journal, commit_transaction);
1042 	jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1043 		  journal->j_commit_sequence, journal->j_tail_sequence);
1044 	if (to_free)
1045 		kfree(commit_transaction);
1046 
1047 	wake_up(&journal->j_wait_done_commit);
1048 }
1049