xref: /linux/fs/jbd2/commit.c (revision cfbf1eecd70db9a7a49c42a0613c00f7a2a86dfb)
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/marker.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/crc32.h>
26 #include <linux/writeback.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bio.h>
29 
30 /*
31  * Default IO end handler for temporary BJ_IO buffer_heads.
32  */
33 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
34 {
35 	BUFFER_TRACE(bh, "");
36 	if (uptodate)
37 		set_buffer_uptodate(bh);
38 	else
39 		clear_buffer_uptodate(bh);
40 	unlock_buffer(bh);
41 }
42 
43 /*
44  * When an ext4 file is truncated, it is possible that some pages are not
45  * successfully freed, because they are attached to a committing transaction.
46  * After the transaction commits, these pages are left on the LRU, with no
47  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
48  * by the VM, but their apparent absence upsets the VM accounting, and it makes
49  * the numbers in /proc/meminfo look odd.
50  *
51  * So here, we have a buffer which has just come off the forget list.  Look to
52  * see if we can strip all buffers from the backing page.
53  *
54  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
55  * caller provided us with a ref against the buffer, and we drop that here.
56  */
57 static void release_buffer_page(struct buffer_head *bh)
58 {
59 	struct page *page;
60 
61 	if (buffer_dirty(bh))
62 		goto nope;
63 	if (atomic_read(&bh->b_count) != 1)
64 		goto nope;
65 	page = bh->b_page;
66 	if (!page)
67 		goto nope;
68 	if (page->mapping)
69 		goto nope;
70 
71 	/* OK, it's a truncated page */
72 	if (!trylock_page(page))
73 		goto nope;
74 
75 	page_cache_get(page);
76 	__brelse(bh);
77 	try_to_free_buffers(page);
78 	unlock_page(page);
79 	page_cache_release(page);
80 	return;
81 
82 nope:
83 	__brelse(bh);
84 }
85 
86 /*
87  * Done it all: now submit the commit record.  We should have
88  * cleaned up our previous buffers by now, so if we are in abort
89  * mode we can now just skip the rest of the journal write
90  * entirely.
91  *
92  * Returns 1 if the journal needs to be aborted or 0 on success
93  */
94 static int journal_submit_commit_record(journal_t *journal,
95 					transaction_t *commit_transaction,
96 					struct buffer_head **cbh,
97 					__u32 crc32_sum)
98 {
99 	struct journal_head *descriptor;
100 	struct commit_header *tmp;
101 	struct buffer_head *bh;
102 	int ret;
103 	int barrier_done = 0;
104 	struct timespec now = current_kernel_time();
105 
106 	if (is_journal_aborted(journal))
107 		return 0;
108 
109 	descriptor = jbd2_journal_get_descriptor_buffer(journal);
110 	if (!descriptor)
111 		return 1;
112 
113 	bh = jh2bh(descriptor);
114 
115 	tmp = (struct commit_header *)bh->b_data;
116 	tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
117 	tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
118 	tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
119 	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
120 	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
121 
122 	if (JBD2_HAS_COMPAT_FEATURE(journal,
123 				    JBD2_FEATURE_COMPAT_CHECKSUM)) {
124 		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
125 		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
126 		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
127 	}
128 
129 	JBUFFER_TRACE(descriptor, "submit commit block");
130 	lock_buffer(bh);
131 	clear_buffer_dirty(bh);
132 	set_buffer_uptodate(bh);
133 	bh->b_end_io = journal_end_buffer_io_sync;
134 
135 	if (journal->j_flags & JBD2_BARRIER &&
136 		!JBD2_HAS_INCOMPAT_FEATURE(journal,
137 					 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
138 		set_buffer_ordered(bh);
139 		barrier_done = 1;
140 	}
141 	ret = submit_bh(WRITE_SYNC_PLUG, bh);
142 	if (barrier_done)
143 		clear_buffer_ordered(bh);
144 
145 	/* is it possible for another commit to fail at roughly
146 	 * the same time as this one?  If so, we don't want to
147 	 * trust the barrier flag in the super, but instead want
148 	 * to remember if we sent a barrier request
149 	 */
150 	if (ret == -EOPNOTSUPP && barrier_done) {
151 		printk(KERN_WARNING
152 		       "JBD: barrier-based sync failed on %s - "
153 		       "disabling barriers\n", journal->j_devname);
154 		spin_lock(&journal->j_state_lock);
155 		journal->j_flags &= ~JBD2_BARRIER;
156 		spin_unlock(&journal->j_state_lock);
157 
158 		/* And try again, without the barrier */
159 		lock_buffer(bh);
160 		set_buffer_uptodate(bh);
161 		clear_buffer_dirty(bh);
162 		ret = submit_bh(WRITE_SYNC_PLUG, bh);
163 	}
164 	*cbh = bh;
165 	return ret;
166 }
167 
168 /*
169  * This function along with journal_submit_commit_record
170  * allows to write the commit record asynchronously.
171  */
172 static int journal_wait_on_commit_record(journal_t *journal,
173 					 struct buffer_head *bh)
174 {
175 	int ret = 0;
176 
177 retry:
178 	clear_buffer_dirty(bh);
179 	wait_on_buffer(bh);
180 	if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
181 		printk(KERN_WARNING
182 		       "JBD2: wait_on_commit_record: sync failed on %s - "
183 		       "disabling barriers\n", journal->j_devname);
184 		spin_lock(&journal->j_state_lock);
185 		journal->j_flags &= ~JBD2_BARRIER;
186 		spin_unlock(&journal->j_state_lock);
187 
188 		lock_buffer(bh);
189 		clear_buffer_dirty(bh);
190 		set_buffer_uptodate(bh);
191 		bh->b_end_io = journal_end_buffer_io_sync;
192 
193 		ret = submit_bh(WRITE_SYNC_PLUG, bh);
194 		if (ret) {
195 			unlock_buffer(bh);
196 			return ret;
197 		}
198 		goto retry;
199 	}
200 
201 	if (unlikely(!buffer_uptodate(bh)))
202 		ret = -EIO;
203 	put_bh(bh);            /* One for getblk() */
204 	jbd2_journal_put_journal_head(bh2jh(bh));
205 
206 	return ret;
207 }
208 
209 /*
210  * write the filemap data using writepage() address_space_operations.
211  * We don't do block allocation here even for delalloc. We don't
212  * use writepages() because with dealyed allocation we may be doing
213  * block allocation in writepages().
214  */
215 static int journal_submit_inode_data_buffers(struct address_space *mapping)
216 {
217 	int ret;
218 	struct writeback_control wbc = {
219 		.sync_mode =  WB_SYNC_ALL,
220 		.nr_to_write = mapping->nrpages * 2,
221 		.range_start = 0,
222 		.range_end = i_size_read(mapping->host),
223 		.for_writepages = 1,
224 	};
225 
226 	ret = generic_writepages(mapping, &wbc);
227 	return ret;
228 }
229 
230 /*
231  * Submit all the data buffers of inode associated with the transaction to
232  * disk.
233  *
234  * We are in a committing transaction. Therefore no new inode can be added to
235  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
236  * operate on from being released while we write out pages.
237  */
238 static int journal_submit_data_buffers(journal_t *journal,
239 		transaction_t *commit_transaction)
240 {
241 	struct jbd2_inode *jinode;
242 	int err, ret = 0;
243 	struct address_space *mapping;
244 
245 	spin_lock(&journal->j_list_lock);
246 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
247 		mapping = jinode->i_vfs_inode->i_mapping;
248 		jinode->i_flags |= JI_COMMIT_RUNNING;
249 		spin_unlock(&journal->j_list_lock);
250 		/*
251 		 * submit the inode data buffers. We use writepage
252 		 * instead of writepages. Because writepages can do
253 		 * block allocation  with delalloc. We need to write
254 		 * only allocated blocks here.
255 		 */
256 		err = journal_submit_inode_data_buffers(mapping);
257 		if (!ret)
258 			ret = err;
259 		spin_lock(&journal->j_list_lock);
260 		J_ASSERT(jinode->i_transaction == commit_transaction);
261 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
262 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
263 	}
264 	spin_unlock(&journal->j_list_lock);
265 	return ret;
266 }
267 
268 /*
269  * Wait for data submitted for writeout, refile inodes to proper
270  * transaction if needed.
271  *
272  */
273 static int journal_finish_inode_data_buffers(journal_t *journal,
274 		transaction_t *commit_transaction)
275 {
276 	struct jbd2_inode *jinode, *next_i;
277 	int err, ret = 0;
278 
279 	/* For locking, see the comment in journal_submit_data_buffers() */
280 	spin_lock(&journal->j_list_lock);
281 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
282 		jinode->i_flags |= JI_COMMIT_RUNNING;
283 		spin_unlock(&journal->j_list_lock);
284 		err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
285 		if (err) {
286 			/*
287 			 * Because AS_EIO is cleared by
288 			 * wait_on_page_writeback_range(), set it again so
289 			 * that user process can get -EIO from fsync().
290 			 */
291 			set_bit(AS_EIO,
292 				&jinode->i_vfs_inode->i_mapping->flags);
293 
294 			if (!ret)
295 				ret = err;
296 		}
297 		spin_lock(&journal->j_list_lock);
298 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
299 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
300 	}
301 
302 	/* Now refile inode to proper lists */
303 	list_for_each_entry_safe(jinode, next_i,
304 				 &commit_transaction->t_inode_list, i_list) {
305 		list_del(&jinode->i_list);
306 		if (jinode->i_next_transaction) {
307 			jinode->i_transaction = jinode->i_next_transaction;
308 			jinode->i_next_transaction = NULL;
309 			list_add(&jinode->i_list,
310 				&jinode->i_transaction->t_inode_list);
311 		} else {
312 			jinode->i_transaction = NULL;
313 		}
314 	}
315 	spin_unlock(&journal->j_list_lock);
316 
317 	return ret;
318 }
319 
320 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
321 {
322 	struct page *page = bh->b_page;
323 	char *addr;
324 	__u32 checksum;
325 
326 	addr = kmap_atomic(page, KM_USER0);
327 	checksum = crc32_be(crc32_sum,
328 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
329 	kunmap_atomic(addr, KM_USER0);
330 
331 	return checksum;
332 }
333 
334 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
335 				   unsigned long long block)
336 {
337 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
338 	if (tag_bytes > JBD2_TAG_SIZE32)
339 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
340 }
341 
342 /*
343  * jbd2_journal_commit_transaction
344  *
345  * The primary function for committing a transaction to the log.  This
346  * function is called by the journal thread to begin a complete commit.
347  */
348 void jbd2_journal_commit_transaction(journal_t *journal)
349 {
350 	struct transaction_stats_s stats;
351 	transaction_t *commit_transaction;
352 	struct journal_head *jh, *new_jh, *descriptor;
353 	struct buffer_head **wbuf = journal->j_wbuf;
354 	int bufs;
355 	int flags;
356 	int err;
357 	unsigned long long blocknr;
358 	ktime_t start_time;
359 	u64 commit_time;
360 	char *tagp = NULL;
361 	journal_header_t *header;
362 	journal_block_tag_t *tag = NULL;
363 	int space_left = 0;
364 	int first_tag = 0;
365 	int tag_flag;
366 	int i, to_free = 0;
367 	int tag_bytes = journal_tag_bytes(journal);
368 	struct buffer_head *cbh = NULL; /* For transactional checksums */
369 	__u32 crc32_sum = ~0;
370 	int write_op = WRITE;
371 
372 	/*
373 	 * First job: lock down the current transaction and wait for
374 	 * all outstanding updates to complete.
375 	 */
376 
377 #ifdef COMMIT_STATS
378 	spin_lock(&journal->j_list_lock);
379 	summarise_journal_usage(journal);
380 	spin_unlock(&journal->j_list_lock);
381 #endif
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 		jbd2_journal_update_superblock(journal, 1);
387 	} else {
388 		jbd_debug(3, "superblock not updated\n");
389 	}
390 
391 	J_ASSERT(journal->j_running_transaction != NULL);
392 	J_ASSERT(journal->j_committing_transaction == NULL);
393 
394 	commit_transaction = journal->j_running_transaction;
395 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
396 
397 	trace_mark(jbd2_start_commit, "dev %s transaction %d",
398 		   journal->j_devname, commit_transaction->t_tid);
399 	jbd_debug(1, "JBD: starting commit of transaction %d\n",
400 			commit_transaction->t_tid);
401 
402 	spin_lock(&journal->j_state_lock);
403 	commit_transaction->t_state = T_LOCKED;
404 
405 	/*
406 	 * Use plugged writes here, since we want to submit several before
407 	 * we unplug the device. We don't do explicit unplugging in here,
408 	 * instead we rely on sync_buffer() doing the unplug for us.
409 	 */
410 	if (commit_transaction->t_synchronous_commit)
411 		write_op = WRITE_SYNC_PLUG;
412 	stats.u.run.rs_wait = commit_transaction->t_max_wait;
413 	stats.u.run.rs_locked = jiffies;
414 	stats.u.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
415 						stats.u.run.rs_locked);
416 
417 	spin_lock(&commit_transaction->t_handle_lock);
418 	while (commit_transaction->t_updates) {
419 		DEFINE_WAIT(wait);
420 
421 		prepare_to_wait(&journal->j_wait_updates, &wait,
422 					TASK_UNINTERRUPTIBLE);
423 		if (commit_transaction->t_updates) {
424 			spin_unlock(&commit_transaction->t_handle_lock);
425 			spin_unlock(&journal->j_state_lock);
426 			schedule();
427 			spin_lock(&journal->j_state_lock);
428 			spin_lock(&commit_transaction->t_handle_lock);
429 		}
430 		finish_wait(&journal->j_wait_updates, &wait);
431 	}
432 	spin_unlock(&commit_transaction->t_handle_lock);
433 
434 	J_ASSERT (commit_transaction->t_outstanding_credits <=
435 			journal->j_max_transaction_buffers);
436 
437 	/*
438 	 * First thing we are allowed to do is to discard any remaining
439 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
440 	 * that there are no such buffers: if a large filesystem
441 	 * operation like a truncate needs to split itself over multiple
442 	 * transactions, then it may try to do a jbd2_journal_restart() while
443 	 * there are still BJ_Reserved buffers outstanding.  These must
444 	 * be released cleanly from the current transaction.
445 	 *
446 	 * In this case, the filesystem must still reserve write access
447 	 * again before modifying the buffer in the new transaction, but
448 	 * we do not require it to remember exactly which old buffers it
449 	 * has reserved.  This is consistent with the existing behaviour
450 	 * that multiple jbd2_journal_get_write_access() calls to the same
451 	 * buffer are perfectly permissable.
452 	 */
453 	while (commit_transaction->t_reserved_list) {
454 		jh = commit_transaction->t_reserved_list;
455 		JBUFFER_TRACE(jh, "reserved, unused: refile");
456 		/*
457 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
458 		 * leave undo-committed data.
459 		 */
460 		if (jh->b_committed_data) {
461 			struct buffer_head *bh = jh2bh(jh);
462 
463 			jbd_lock_bh_state(bh);
464 			jbd2_free(jh->b_committed_data, bh->b_size);
465 			jh->b_committed_data = NULL;
466 			jbd_unlock_bh_state(bh);
467 		}
468 		jbd2_journal_refile_buffer(journal, jh);
469 	}
470 
471 	/*
472 	 * Now try to drop any written-back buffers from the journal's
473 	 * checkpoint lists.  We do this *before* commit because it potentially
474 	 * frees some memory
475 	 */
476 	spin_lock(&journal->j_list_lock);
477 	__jbd2_journal_clean_checkpoint_list(journal);
478 	spin_unlock(&journal->j_list_lock);
479 
480 	jbd_debug (3, "JBD: commit phase 1\n");
481 
482 	/*
483 	 * Switch to a new revoke table.
484 	 */
485 	jbd2_journal_switch_revoke_table(journal);
486 
487 	stats.u.run.rs_flushing = jiffies;
488 	stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
489 					       stats.u.run.rs_flushing);
490 
491 	commit_transaction->t_state = T_FLUSH;
492 	journal->j_committing_transaction = commit_transaction;
493 	journal->j_running_transaction = NULL;
494 	start_time = ktime_get();
495 	commit_transaction->t_log_start = journal->j_head;
496 	wake_up(&journal->j_wait_transaction_locked);
497 	spin_unlock(&journal->j_state_lock);
498 
499 	jbd_debug (3, "JBD: commit phase 2\n");
500 
501 	/*
502 	 * Now start flushing things to disk, in the order they appear
503 	 * on the transaction lists.  Data blocks go first.
504 	 */
505 	err = journal_submit_data_buffers(journal, commit_transaction);
506 	if (err)
507 		jbd2_journal_abort(journal, err);
508 
509 	jbd2_journal_write_revoke_records(journal, commit_transaction);
510 
511 	jbd_debug(3, "JBD: commit phase 2\n");
512 
513 	/*
514 	 * Way to go: we have now written out all of the data for a
515 	 * transaction!  Now comes the tricky part: we need to write out
516 	 * metadata.  Loop over the transaction's entire buffer list:
517 	 */
518 	spin_lock(&journal->j_state_lock);
519 	commit_transaction->t_state = T_COMMIT;
520 	spin_unlock(&journal->j_state_lock);
521 
522 	stats.u.run.rs_logging = jiffies;
523 	stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
524 						 stats.u.run.rs_logging);
525 	stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
526 	stats.u.run.rs_blocks_logged = 0;
527 
528 	J_ASSERT(commit_transaction->t_nr_buffers <=
529 		 commit_transaction->t_outstanding_credits);
530 
531 	err = 0;
532 	descriptor = NULL;
533 	bufs = 0;
534 	while (commit_transaction->t_buffers) {
535 
536 		/* Find the next buffer to be journaled... */
537 
538 		jh = commit_transaction->t_buffers;
539 
540 		/* If we're in abort mode, we just un-journal the buffer and
541 		   release it. */
542 
543 		if (is_journal_aborted(journal)) {
544 			clear_buffer_jbddirty(jh2bh(jh));
545 			JBUFFER_TRACE(jh, "journal is aborting: refile");
546 			jbd2_buffer_abort_trigger(jh,
547 						  jh->b_frozen_data ?
548 						  jh->b_frozen_triggers :
549 						  jh->b_triggers);
550 			jbd2_journal_refile_buffer(journal, jh);
551 			/* If that was the last one, we need to clean up
552 			 * any descriptor buffers which may have been
553 			 * already allocated, even if we are now
554 			 * aborting. */
555 			if (!commit_transaction->t_buffers)
556 				goto start_journal_io;
557 			continue;
558 		}
559 
560 		/* Make sure we have a descriptor block in which to
561 		   record the metadata buffer. */
562 
563 		if (!descriptor) {
564 			struct buffer_head *bh;
565 
566 			J_ASSERT (bufs == 0);
567 
568 			jbd_debug(4, "JBD: get descriptor\n");
569 
570 			descriptor = jbd2_journal_get_descriptor_buffer(journal);
571 			if (!descriptor) {
572 				jbd2_journal_abort(journal, -EIO);
573 				continue;
574 			}
575 
576 			bh = jh2bh(descriptor);
577 			jbd_debug(4, "JBD: got buffer %llu (%p)\n",
578 				(unsigned long long)bh->b_blocknr, bh->b_data);
579 			header = (journal_header_t *)&bh->b_data[0];
580 			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
581 			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
582 			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
583 
584 			tagp = &bh->b_data[sizeof(journal_header_t)];
585 			space_left = bh->b_size - sizeof(journal_header_t);
586 			first_tag = 1;
587 			set_buffer_jwrite(bh);
588 			set_buffer_dirty(bh);
589 			wbuf[bufs++] = bh;
590 
591 			/* Record it so that we can wait for IO
592                            completion later */
593 			BUFFER_TRACE(bh, "ph3: file as descriptor");
594 			jbd2_journal_file_buffer(descriptor, commit_transaction,
595 					BJ_LogCtl);
596 		}
597 
598 		/* Where is the buffer to be written? */
599 
600 		err = jbd2_journal_next_log_block(journal, &blocknr);
601 		/* If the block mapping failed, just abandon the buffer
602 		   and repeat this loop: we'll fall into the
603 		   refile-on-abort condition above. */
604 		if (err) {
605 			jbd2_journal_abort(journal, err);
606 			continue;
607 		}
608 
609 		/*
610 		 * start_this_handle() uses t_outstanding_credits to determine
611 		 * the free space in the log, but this counter is changed
612 		 * by jbd2_journal_next_log_block() also.
613 		 */
614 		commit_transaction->t_outstanding_credits--;
615 
616 		/* Bump b_count to prevent truncate from stumbling over
617                    the shadowed buffer!  @@@ This can go if we ever get
618                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
619 		atomic_inc(&jh2bh(jh)->b_count);
620 
621 		/* Make a temporary IO buffer with which to write it out
622                    (this will requeue both the metadata buffer and the
623                    temporary IO buffer). new_bh goes on BJ_IO*/
624 
625 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
626 		/*
627 		 * akpm: jbd2_journal_write_metadata_buffer() sets
628 		 * new_bh->b_transaction to commit_transaction.
629 		 * We need to clean this up before we release new_bh
630 		 * (which is of type BJ_IO)
631 		 */
632 		JBUFFER_TRACE(jh, "ph3: write metadata");
633 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
634 						      jh, &new_jh, blocknr);
635 		set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
636 		wbuf[bufs++] = jh2bh(new_jh);
637 
638 		/* Record the new block's tag in the current descriptor
639                    buffer */
640 
641 		tag_flag = 0;
642 		if (flags & 1)
643 			tag_flag |= JBD2_FLAG_ESCAPE;
644 		if (!first_tag)
645 			tag_flag |= JBD2_FLAG_SAME_UUID;
646 
647 		tag = (journal_block_tag_t *) tagp;
648 		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
649 		tag->t_flags = cpu_to_be32(tag_flag);
650 		tagp += tag_bytes;
651 		space_left -= tag_bytes;
652 
653 		if (first_tag) {
654 			memcpy (tagp, journal->j_uuid, 16);
655 			tagp += 16;
656 			space_left -= 16;
657 			first_tag = 0;
658 		}
659 
660 		/* If there's no more to do, or if the descriptor is full,
661 		   let the IO rip! */
662 
663 		if (bufs == journal->j_wbufsize ||
664 		    commit_transaction->t_buffers == NULL ||
665 		    space_left < tag_bytes + 16) {
666 
667 			jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
668 
669 			/* Write an end-of-descriptor marker before
670                            submitting the IOs.  "tag" still points to
671                            the last tag we set up. */
672 
673 			tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
674 
675 start_journal_io:
676 			for (i = 0; i < bufs; i++) {
677 				struct buffer_head *bh = wbuf[i];
678 				/*
679 				 * Compute checksum.
680 				 */
681 				if (JBD2_HAS_COMPAT_FEATURE(journal,
682 					JBD2_FEATURE_COMPAT_CHECKSUM)) {
683 					crc32_sum =
684 					    jbd2_checksum_data(crc32_sum, bh);
685 				}
686 
687 				lock_buffer(bh);
688 				clear_buffer_dirty(bh);
689 				set_buffer_uptodate(bh);
690 				bh->b_end_io = journal_end_buffer_io_sync;
691 				submit_bh(write_op, bh);
692 			}
693 			cond_resched();
694 			stats.u.run.rs_blocks_logged += bufs;
695 
696 			/* Force a new descriptor to be generated next
697                            time round the loop. */
698 			descriptor = NULL;
699 			bufs = 0;
700 		}
701 	}
702 
703 	/* Done it all: now write the commit record asynchronously. */
704 
705 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
706 		JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
707 		err = journal_submit_commit_record(journal, commit_transaction,
708 						 &cbh, crc32_sum);
709 		if (err)
710 			__jbd2_journal_abort_hard(journal);
711 	}
712 
713 	/*
714 	 * This is the right place to wait for data buffers both for ASYNC
715 	 * and !ASYNC commit. If commit is ASYNC, we need to wait only after
716 	 * the commit block went to disk (which happens above). If commit is
717 	 * SYNC, we need to wait for data buffers before we start writing
718 	 * commit block, which happens below in such setting.
719 	 */
720 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
721 	if (err) {
722 		printk(KERN_WARNING
723 			"JBD2: Detected IO errors while flushing file data "
724 		       "on %s\n", journal->j_devname);
725 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
726 			jbd2_journal_abort(journal, err);
727 		err = 0;
728 	}
729 
730 	/* Lo and behold: we have just managed to send a transaction to
731            the log.  Before we can commit it, wait for the IO so far to
732            complete.  Control buffers being written are on the
733            transaction's t_log_list queue, and metadata buffers are on
734            the t_iobuf_list queue.
735 
736 	   Wait for the buffers in reverse order.  That way we are
737 	   less likely to be woken up until all IOs have completed, and
738 	   so we incur less scheduling load.
739 	*/
740 
741 	jbd_debug(3, "JBD: commit phase 3\n");
742 
743 	/*
744 	 * akpm: these are BJ_IO, and j_list_lock is not needed.
745 	 * See __journal_try_to_free_buffer.
746 	 */
747 wait_for_iobuf:
748 	while (commit_transaction->t_iobuf_list != NULL) {
749 		struct buffer_head *bh;
750 
751 		jh = commit_transaction->t_iobuf_list->b_tprev;
752 		bh = jh2bh(jh);
753 		if (buffer_locked(bh)) {
754 			wait_on_buffer(bh);
755 			goto wait_for_iobuf;
756 		}
757 		if (cond_resched())
758 			goto wait_for_iobuf;
759 
760 		if (unlikely(!buffer_uptodate(bh)))
761 			err = -EIO;
762 
763 		clear_buffer_jwrite(bh);
764 
765 		JBUFFER_TRACE(jh, "ph4: unfile after journal write");
766 		jbd2_journal_unfile_buffer(journal, jh);
767 
768 		/*
769 		 * ->t_iobuf_list should contain only dummy buffer_heads
770 		 * which were created by jbd2_journal_write_metadata_buffer().
771 		 */
772 		BUFFER_TRACE(bh, "dumping temporary bh");
773 		jbd2_journal_put_journal_head(jh);
774 		__brelse(bh);
775 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
776 		free_buffer_head(bh);
777 
778 		/* We also have to unlock and free the corresponding
779                    shadowed buffer */
780 		jh = commit_transaction->t_shadow_list->b_tprev;
781 		bh = jh2bh(jh);
782 		clear_bit(BH_JWrite, &bh->b_state);
783 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
784 
785 		/* The metadata is now released for reuse, but we need
786                    to remember it against this transaction so that when
787                    we finally commit, we can do any checkpointing
788                    required. */
789 		JBUFFER_TRACE(jh, "file as BJ_Forget");
790 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
791 		/* Wake up any transactions which were waiting for this
792 		   IO to complete */
793 		wake_up_bit(&bh->b_state, BH_Unshadow);
794 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
795 		__brelse(bh);
796 	}
797 
798 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
799 
800 	jbd_debug(3, "JBD: commit phase 4\n");
801 
802 	/* Here we wait for the revoke record and descriptor record buffers */
803  wait_for_ctlbuf:
804 	while (commit_transaction->t_log_list != NULL) {
805 		struct buffer_head *bh;
806 
807 		jh = commit_transaction->t_log_list->b_tprev;
808 		bh = jh2bh(jh);
809 		if (buffer_locked(bh)) {
810 			wait_on_buffer(bh);
811 			goto wait_for_ctlbuf;
812 		}
813 		if (cond_resched())
814 			goto wait_for_ctlbuf;
815 
816 		if (unlikely(!buffer_uptodate(bh)))
817 			err = -EIO;
818 
819 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
820 		clear_buffer_jwrite(bh);
821 		jbd2_journal_unfile_buffer(journal, jh);
822 		jbd2_journal_put_journal_head(jh);
823 		__brelse(bh);		/* One for getblk */
824 		/* AKPM: bforget here */
825 	}
826 
827 	if (err)
828 		jbd2_journal_abort(journal, err);
829 
830 	jbd_debug(3, "JBD: commit phase 5\n");
831 
832 	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
833 		JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
834 		err = journal_submit_commit_record(journal, commit_transaction,
835 						&cbh, crc32_sum);
836 		if (err)
837 			__jbd2_journal_abort_hard(journal);
838 	}
839 	if (!err && !is_journal_aborted(journal))
840 		err = journal_wait_on_commit_record(journal, cbh);
841 
842 	if (err)
843 		jbd2_journal_abort(journal, err);
844 
845 	/* End of a transaction!  Finally, we can do checkpoint
846            processing: any buffers committed as a result of this
847            transaction can be removed from any checkpoint list it was on
848            before. */
849 
850 	jbd_debug(3, "JBD: commit phase 6\n");
851 
852 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
853 	J_ASSERT(commit_transaction->t_buffers == NULL);
854 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
855 	J_ASSERT(commit_transaction->t_iobuf_list == NULL);
856 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
857 	J_ASSERT(commit_transaction->t_log_list == NULL);
858 
859 restart_loop:
860 	/*
861 	 * As there are other places (journal_unmap_buffer()) adding buffers
862 	 * to this list we have to be careful and hold the j_list_lock.
863 	 */
864 	spin_lock(&journal->j_list_lock);
865 	while (commit_transaction->t_forget) {
866 		transaction_t *cp_transaction;
867 		struct buffer_head *bh;
868 
869 		jh = commit_transaction->t_forget;
870 		spin_unlock(&journal->j_list_lock);
871 		bh = jh2bh(jh);
872 		jbd_lock_bh_state(bh);
873 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction ||
874 			jh->b_transaction == journal->j_running_transaction);
875 
876 		/*
877 		 * If there is undo-protected committed data against
878 		 * this buffer, then we can remove it now.  If it is a
879 		 * buffer needing such protection, the old frozen_data
880 		 * field now points to a committed version of the
881 		 * buffer, so rotate that field to the new committed
882 		 * data.
883 		 *
884 		 * Otherwise, we can just throw away the frozen data now.
885 		 *
886 		 * We also know that the frozen data has already fired
887 		 * its triggers if they exist, so we can clear that too.
888 		 */
889 		if (jh->b_committed_data) {
890 			jbd2_free(jh->b_committed_data, bh->b_size);
891 			jh->b_committed_data = NULL;
892 			if (jh->b_frozen_data) {
893 				jh->b_committed_data = jh->b_frozen_data;
894 				jh->b_frozen_data = NULL;
895 				jh->b_frozen_triggers = NULL;
896 			}
897 		} else if (jh->b_frozen_data) {
898 			jbd2_free(jh->b_frozen_data, bh->b_size);
899 			jh->b_frozen_data = NULL;
900 			jh->b_frozen_triggers = NULL;
901 		}
902 
903 		spin_lock(&journal->j_list_lock);
904 		cp_transaction = jh->b_cp_transaction;
905 		if (cp_transaction) {
906 			JBUFFER_TRACE(jh, "remove from old cp transaction");
907 			cp_transaction->t_chp_stats.cs_dropped++;
908 			__jbd2_journal_remove_checkpoint(jh);
909 		}
910 
911 		/* Only re-checkpoint the buffer_head if it is marked
912 		 * dirty.  If the buffer was added to the BJ_Forget list
913 		 * by jbd2_journal_forget, it may no longer be dirty and
914 		 * there's no point in keeping a checkpoint record for
915 		 * it. */
916 
917 		/* A buffer which has been freed while still being
918 		 * journaled by a previous transaction may end up still
919 		 * being dirty here, but we want to avoid writing back
920 		 * that buffer in the future now that the last use has
921 		 * been committed.  That's not only a performance gain,
922 		 * it also stops aliasing problems if the buffer is left
923 		 * behind for writeback and gets reallocated for another
924 		 * use in a different page. */
925 		if (buffer_freed(bh)) {
926 			clear_buffer_freed(bh);
927 			clear_buffer_jbddirty(bh);
928 		}
929 
930 		if (buffer_jbddirty(bh)) {
931 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
932 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
933 			if (is_journal_aborted(journal))
934 				clear_buffer_jbddirty(bh);
935 			JBUFFER_TRACE(jh, "refile for checkpoint writeback");
936 			__jbd2_journal_refile_buffer(jh);
937 			jbd_unlock_bh_state(bh);
938 		} else {
939 			J_ASSERT_BH(bh, !buffer_dirty(bh));
940 			/* The buffer on BJ_Forget list and not jbddirty means
941 			 * it has been freed by this transaction and hence it
942 			 * could not have been reallocated until this
943 			 * transaction has committed. *BUT* it could be
944 			 * reallocated once we have written all the data to
945 			 * disk and before we process the buffer on BJ_Forget
946 			 * list. */
947 			JBUFFER_TRACE(jh, "refile or unfile freed buffer");
948 			__jbd2_journal_refile_buffer(jh);
949 			if (!jh->b_transaction) {
950 				jbd_unlock_bh_state(bh);
951 				 /* needs a brelse */
952 				jbd2_journal_remove_journal_head(bh);
953 				release_buffer_page(bh);
954 			} else
955 				jbd_unlock_bh_state(bh);
956 		}
957 		cond_resched_lock(&journal->j_list_lock);
958 	}
959 	spin_unlock(&journal->j_list_lock);
960 	/*
961 	 * This is a bit sleazy.  We use j_list_lock to protect transition
962 	 * of a transaction into T_FINISHED state and calling
963 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
964 	 * other checkpointing code processing the transaction...
965 	 */
966 	spin_lock(&journal->j_state_lock);
967 	spin_lock(&journal->j_list_lock);
968 	/*
969 	 * Now recheck if some buffers did not get attached to the transaction
970 	 * while the lock was dropped...
971 	 */
972 	if (commit_transaction->t_forget) {
973 		spin_unlock(&journal->j_list_lock);
974 		spin_unlock(&journal->j_state_lock);
975 		goto restart_loop;
976 	}
977 
978 	/* Done with this transaction! */
979 
980 	jbd_debug(3, "JBD: commit phase 7\n");
981 
982 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
983 
984 	commit_transaction->t_start = jiffies;
985 	stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
986 						commit_transaction->t_start);
987 
988 	/*
989 	 * File the transaction for history
990 	 */
991 	stats.ts_type = JBD2_STATS_RUN;
992 	stats.ts_tid = commit_transaction->t_tid;
993 	stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
994 	spin_lock(&journal->j_history_lock);
995 	memcpy(journal->j_history + journal->j_history_cur, &stats,
996 			sizeof(stats));
997 	if (++journal->j_history_cur == journal->j_history_max)
998 		journal->j_history_cur = 0;
999 
1000 	/*
1001 	 * Calculate overall stats
1002 	 */
1003 	journal->j_stats.ts_tid++;
1004 	journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
1005 	journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
1006 	journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
1007 	journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
1008 	journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
1009 	journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
1010 	journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
1011 	journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
1012 	spin_unlock(&journal->j_history_lock);
1013 
1014 	commit_transaction->t_state = T_FINISHED;
1015 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1016 	journal->j_commit_sequence = commit_transaction->t_tid;
1017 	journal->j_committing_transaction = NULL;
1018 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1019 
1020 	/*
1021 	 * weight the commit time higher than the average time so we don't
1022 	 * react too strongly to vast changes in the commit time
1023 	 */
1024 	if (likely(journal->j_average_commit_time))
1025 		journal->j_average_commit_time = (commit_time +
1026 				journal->j_average_commit_time*3) / 4;
1027 	else
1028 		journal->j_average_commit_time = commit_time;
1029 	spin_unlock(&journal->j_state_lock);
1030 
1031 	if (commit_transaction->t_checkpoint_list == NULL &&
1032 	    commit_transaction->t_checkpoint_io_list == NULL) {
1033 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1034 		to_free = 1;
1035 	} else {
1036 		if (journal->j_checkpoint_transactions == NULL) {
1037 			journal->j_checkpoint_transactions = commit_transaction;
1038 			commit_transaction->t_cpnext = commit_transaction;
1039 			commit_transaction->t_cpprev = commit_transaction;
1040 		} else {
1041 			commit_transaction->t_cpnext =
1042 				journal->j_checkpoint_transactions;
1043 			commit_transaction->t_cpprev =
1044 				commit_transaction->t_cpnext->t_cpprev;
1045 			commit_transaction->t_cpnext->t_cpprev =
1046 				commit_transaction;
1047 			commit_transaction->t_cpprev->t_cpnext =
1048 				commit_transaction;
1049 		}
1050 	}
1051 	spin_unlock(&journal->j_list_lock);
1052 
1053 	if (journal->j_commit_callback)
1054 		journal->j_commit_callback(journal, commit_transaction);
1055 
1056 	trace_mark(jbd2_end_commit, "dev %s transaction %d head %d",
1057 		   journal->j_devname, commit_transaction->t_tid,
1058 		   journal->j_tail_sequence);
1059 	jbd_debug(1, "JBD: commit %d complete, head %d\n",
1060 		  journal->j_commit_sequence, journal->j_tail_sequence);
1061 	if (to_free)
1062 		kfree(commit_transaction);
1063 
1064 	wake_up(&journal->j_wait_done_commit);
1065 }
1066