xref: /linux/fs/jbd2/commit.c (revision dd5648c9f53b5cbd9f948d752624400545f979fb)
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 <trace/events/jbd2.h>
30 
31 /*
32  * Default IO end handler for temporary BJ_IO buffer_heads.
33  */
34 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
35 {
36 	BUFFER_TRACE(bh, "");
37 	if (uptodate)
38 		set_buffer_uptodate(bh);
39 	else
40 		clear_buffer_uptodate(bh);
41 	unlock_buffer(bh);
42 }
43 
44 /*
45  * When an ext4 file is truncated, it is possible that some pages are not
46  * successfully freed, because they are attached to a committing transaction.
47  * After the transaction commits, these pages are left on the LRU, with no
48  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
49  * by the VM, but their apparent absence upsets the VM accounting, and it makes
50  * the numbers in /proc/meminfo look odd.
51  *
52  * So here, we have a buffer which has just come off the forget list.  Look to
53  * see if we can strip all buffers from the backing page.
54  *
55  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
56  * caller provided us with a ref against the buffer, and we drop that here.
57  */
58 static void release_buffer_page(struct buffer_head *bh)
59 {
60 	struct page *page;
61 
62 	if (buffer_dirty(bh))
63 		goto nope;
64 	if (atomic_read(&bh->b_count) != 1)
65 		goto nope;
66 	page = bh->b_page;
67 	if (!page)
68 		goto nope;
69 	if (page->mapping)
70 		goto nope;
71 
72 	/* OK, it's a truncated page */
73 	if (!trylock_page(page))
74 		goto nope;
75 
76 	page_cache_get(page);
77 	__brelse(bh);
78 	try_to_free_buffers(page);
79 	unlock_page(page);
80 	page_cache_release(page);
81 	return;
82 
83 nope:
84 	__brelse(bh);
85 }
86 
87 /*
88  * Done it all: now submit the commit record.  We should have
89  * cleaned up our previous buffers by now, so if we are in abort
90  * mode we can now just skip the rest of the journal write
91  * entirely.
92  *
93  * Returns 1 if the journal needs to be aborted or 0 on success
94  */
95 static int journal_submit_commit_record(journal_t *journal,
96 					transaction_t *commit_transaction,
97 					struct buffer_head **cbh,
98 					__u32 crc32_sum)
99 {
100 	struct journal_head *descriptor;
101 	struct commit_header *tmp;
102 	struct buffer_head *bh;
103 	int ret;
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 		ret = submit_bh(WRITE_SYNC_PLUG | WRITE_FLUSH_FUA, bh);
139 	else
140 		ret = submit_bh(WRITE_SYNC_PLUG, bh);
141 
142 	*cbh = bh;
143 	return ret;
144 }
145 
146 /*
147  * This function along with journal_submit_commit_record
148  * allows to write the commit record asynchronously.
149  */
150 static int journal_wait_on_commit_record(journal_t *journal,
151 					 struct buffer_head *bh)
152 {
153 	int ret = 0;
154 
155 	clear_buffer_dirty(bh);
156 	wait_on_buffer(bh);
157 
158 	if (unlikely(!buffer_uptodate(bh)))
159 		ret = -EIO;
160 	put_bh(bh);            /* One for getblk() */
161 	jbd2_journal_put_journal_head(bh2jh(bh));
162 
163 	return ret;
164 }
165 
166 /*
167  * write the filemap data using writepage() address_space_operations.
168  * We don't do block allocation here even for delalloc. We don't
169  * use writepages() because with dealyed allocation we may be doing
170  * block allocation in writepages().
171  */
172 static int journal_submit_inode_data_buffers(struct address_space *mapping)
173 {
174 	int ret;
175 	struct writeback_control wbc = {
176 		.sync_mode =  WB_SYNC_ALL,
177 		.nr_to_write = mapping->nrpages * 2,
178 		.range_start = 0,
179 		.range_end = i_size_read(mapping->host),
180 	};
181 
182 	ret = generic_writepages(mapping, &wbc);
183 	return ret;
184 }
185 
186 /*
187  * Submit all the data buffers of inode associated with the transaction to
188  * disk.
189  *
190  * We are in a committing transaction. Therefore no new inode can be added to
191  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
192  * operate on from being released while we write out pages.
193  */
194 static int journal_submit_data_buffers(journal_t *journal,
195 		transaction_t *commit_transaction)
196 {
197 	struct jbd2_inode *jinode;
198 	int err, ret = 0;
199 	struct address_space *mapping;
200 
201 	spin_lock(&journal->j_list_lock);
202 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
203 		mapping = jinode->i_vfs_inode->i_mapping;
204 		jinode->i_flags |= JI_COMMIT_RUNNING;
205 		spin_unlock(&journal->j_list_lock);
206 		/*
207 		 * submit the inode data buffers. We use writepage
208 		 * instead of writepages. Because writepages can do
209 		 * block allocation  with delalloc. We need to write
210 		 * only allocated blocks here.
211 		 */
212 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
213 		err = journal_submit_inode_data_buffers(mapping);
214 		if (!ret)
215 			ret = err;
216 		spin_lock(&journal->j_list_lock);
217 		J_ASSERT(jinode->i_transaction == commit_transaction);
218 		commit_transaction->t_flushed_data_blocks = 1;
219 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
220 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
221 	}
222 	spin_unlock(&journal->j_list_lock);
223 	return ret;
224 }
225 
226 /*
227  * Wait for data submitted for writeout, refile inodes to proper
228  * transaction if needed.
229  *
230  */
231 static int journal_finish_inode_data_buffers(journal_t *journal,
232 		transaction_t *commit_transaction)
233 {
234 	struct jbd2_inode *jinode, *next_i;
235 	int err, ret = 0;
236 
237 	/* For locking, see the comment in journal_submit_data_buffers() */
238 	spin_lock(&journal->j_list_lock);
239 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
240 		jinode->i_flags |= JI_COMMIT_RUNNING;
241 		spin_unlock(&journal->j_list_lock);
242 		err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
243 		if (err) {
244 			/*
245 			 * Because AS_EIO is cleared by
246 			 * filemap_fdatawait_range(), set it again so
247 			 * that user process can get -EIO from fsync().
248 			 */
249 			set_bit(AS_EIO,
250 				&jinode->i_vfs_inode->i_mapping->flags);
251 
252 			if (!ret)
253 				ret = err;
254 		}
255 		spin_lock(&journal->j_list_lock);
256 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
257 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
258 	}
259 
260 	/* Now refile inode to proper lists */
261 	list_for_each_entry_safe(jinode, next_i,
262 				 &commit_transaction->t_inode_list, i_list) {
263 		list_del(&jinode->i_list);
264 		if (jinode->i_next_transaction) {
265 			jinode->i_transaction = jinode->i_next_transaction;
266 			jinode->i_next_transaction = NULL;
267 			list_add(&jinode->i_list,
268 				&jinode->i_transaction->t_inode_list);
269 		} else {
270 			jinode->i_transaction = NULL;
271 		}
272 	}
273 	spin_unlock(&journal->j_list_lock);
274 
275 	return ret;
276 }
277 
278 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
279 {
280 	struct page *page = bh->b_page;
281 	char *addr;
282 	__u32 checksum;
283 
284 	addr = kmap_atomic(page, KM_USER0);
285 	checksum = crc32_be(crc32_sum,
286 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
287 	kunmap_atomic(addr, KM_USER0);
288 
289 	return checksum;
290 }
291 
292 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
293 				   unsigned long long block)
294 {
295 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
296 	if (tag_bytes > JBD2_TAG_SIZE32)
297 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
298 }
299 
300 /*
301  * jbd2_journal_commit_transaction
302  *
303  * The primary function for committing a transaction to the log.  This
304  * function is called by the journal thread to begin a complete commit.
305  */
306 void jbd2_journal_commit_transaction(journal_t *journal)
307 {
308 	struct transaction_stats_s stats;
309 	transaction_t *commit_transaction;
310 	struct journal_head *jh, *new_jh, *descriptor;
311 	struct buffer_head **wbuf = journal->j_wbuf;
312 	int bufs;
313 	int flags;
314 	int err;
315 	unsigned long long blocknr;
316 	ktime_t start_time;
317 	u64 commit_time;
318 	char *tagp = NULL;
319 	journal_header_t *header;
320 	journal_block_tag_t *tag = NULL;
321 	int space_left = 0;
322 	int first_tag = 0;
323 	int tag_flag;
324 	int i, to_free = 0;
325 	int tag_bytes = journal_tag_bytes(journal);
326 	struct buffer_head *cbh = NULL; /* For transactional checksums */
327 	__u32 crc32_sum = ~0;
328 	int write_op = WRITE_SYNC;
329 
330 	/*
331 	 * First job: lock down the current transaction and wait for
332 	 * all outstanding updates to complete.
333 	 */
334 
335 #ifdef COMMIT_STATS
336 	spin_lock(&journal->j_list_lock);
337 	summarise_journal_usage(journal);
338 	spin_unlock(&journal->j_list_lock);
339 #endif
340 
341 	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
342 	if (journal->j_flags & JBD2_FLUSHED) {
343 		jbd_debug(3, "super block updated\n");
344 		jbd2_journal_update_superblock(journal, 1);
345 	} else {
346 		jbd_debug(3, "superblock not updated\n");
347 	}
348 
349 	J_ASSERT(journal->j_running_transaction != NULL);
350 	J_ASSERT(journal->j_committing_transaction == NULL);
351 
352 	commit_transaction = journal->j_running_transaction;
353 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
354 
355 	trace_jbd2_start_commit(journal, commit_transaction);
356 	jbd_debug(1, "JBD: starting commit of transaction %d\n",
357 			commit_transaction->t_tid);
358 
359 	write_lock(&journal->j_state_lock);
360 	commit_transaction->t_state = T_LOCKED;
361 
362 	/*
363 	 * Use plugged writes here, since we want to submit several before
364 	 * we unplug the device. We don't do explicit unplugging in here,
365 	 * instead we rely on sync_buffer() doing the unplug for us.
366 	 */
367 	if (commit_transaction->t_synchronous_commit)
368 		write_op = WRITE_SYNC_PLUG;
369 	trace_jbd2_commit_locking(journal, commit_transaction);
370 	stats.run.rs_wait = commit_transaction->t_max_wait;
371 	stats.run.rs_locked = jiffies;
372 	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
373 					      stats.run.rs_locked);
374 
375 	spin_lock(&commit_transaction->t_handle_lock);
376 	while (atomic_read(&commit_transaction->t_updates)) {
377 		DEFINE_WAIT(wait);
378 
379 		prepare_to_wait(&journal->j_wait_updates, &wait,
380 					TASK_UNINTERRUPTIBLE);
381 		if (atomic_read(&commit_transaction->t_updates)) {
382 			spin_unlock(&commit_transaction->t_handle_lock);
383 			write_unlock(&journal->j_state_lock);
384 			schedule();
385 			write_lock(&journal->j_state_lock);
386 			spin_lock(&commit_transaction->t_handle_lock);
387 		}
388 		finish_wait(&journal->j_wait_updates, &wait);
389 	}
390 	spin_unlock(&commit_transaction->t_handle_lock);
391 
392 	J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
393 			journal->j_max_transaction_buffers);
394 
395 	/*
396 	 * First thing we are allowed to do is to discard any remaining
397 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
398 	 * that there are no such buffers: if a large filesystem
399 	 * operation like a truncate needs to split itself over multiple
400 	 * transactions, then it may try to do a jbd2_journal_restart() while
401 	 * there are still BJ_Reserved buffers outstanding.  These must
402 	 * be released cleanly from the current transaction.
403 	 *
404 	 * In this case, the filesystem must still reserve write access
405 	 * again before modifying the buffer in the new transaction, but
406 	 * we do not require it to remember exactly which old buffers it
407 	 * has reserved.  This is consistent with the existing behaviour
408 	 * that multiple jbd2_journal_get_write_access() calls to the same
409 	 * buffer are perfectly permissable.
410 	 */
411 	while (commit_transaction->t_reserved_list) {
412 		jh = commit_transaction->t_reserved_list;
413 		JBUFFER_TRACE(jh, "reserved, unused: refile");
414 		/*
415 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
416 		 * leave undo-committed data.
417 		 */
418 		if (jh->b_committed_data) {
419 			struct buffer_head *bh = jh2bh(jh);
420 
421 			jbd_lock_bh_state(bh);
422 			jbd2_free(jh->b_committed_data, bh->b_size);
423 			jh->b_committed_data = NULL;
424 			jbd_unlock_bh_state(bh);
425 		}
426 		jbd2_journal_refile_buffer(journal, jh);
427 	}
428 
429 	/*
430 	 * Now try to drop any written-back buffers from the journal's
431 	 * checkpoint lists.  We do this *before* commit because it potentially
432 	 * frees some memory
433 	 */
434 	spin_lock(&journal->j_list_lock);
435 	__jbd2_journal_clean_checkpoint_list(journal);
436 	spin_unlock(&journal->j_list_lock);
437 
438 	jbd_debug (3, "JBD: commit phase 1\n");
439 
440 	/*
441 	 * Switch to a new revoke table.
442 	 */
443 	jbd2_journal_switch_revoke_table(journal);
444 
445 	trace_jbd2_commit_flushing(journal, commit_transaction);
446 	stats.run.rs_flushing = jiffies;
447 	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
448 					     stats.run.rs_flushing);
449 
450 	commit_transaction->t_state = T_FLUSH;
451 	journal->j_committing_transaction = commit_transaction;
452 	journal->j_running_transaction = NULL;
453 	start_time = ktime_get();
454 	commit_transaction->t_log_start = journal->j_head;
455 	wake_up(&journal->j_wait_transaction_locked);
456 	write_unlock(&journal->j_state_lock);
457 
458 	jbd_debug (3, "JBD: commit phase 2\n");
459 
460 	/*
461 	 * Now start flushing things to disk, in the order they appear
462 	 * on the transaction lists.  Data blocks go first.
463 	 */
464 	err = journal_submit_data_buffers(journal, commit_transaction);
465 	if (err)
466 		jbd2_journal_abort(journal, err);
467 
468 	jbd2_journal_write_revoke_records(journal, commit_transaction,
469 					  write_op);
470 
471 	jbd_debug(3, "JBD: commit phase 2\n");
472 
473 	/*
474 	 * Way to go: we have now written out all of the data for a
475 	 * transaction!  Now comes the tricky part: we need to write out
476 	 * metadata.  Loop over the transaction's entire buffer list:
477 	 */
478 	write_lock(&journal->j_state_lock);
479 	commit_transaction->t_state = T_COMMIT;
480 	write_unlock(&journal->j_state_lock);
481 
482 	trace_jbd2_commit_logging(journal, commit_transaction);
483 	stats.run.rs_logging = jiffies;
484 	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
485 					       stats.run.rs_logging);
486 	stats.run.rs_blocks =
487 		atomic_read(&commit_transaction->t_outstanding_credits);
488 	stats.run.rs_blocks_logged = 0;
489 
490 	J_ASSERT(commit_transaction->t_nr_buffers <=
491 		 atomic_read(&commit_transaction->t_outstanding_credits));
492 
493 	err = 0;
494 	descriptor = NULL;
495 	bufs = 0;
496 	while (commit_transaction->t_buffers) {
497 
498 		/* Find the next buffer to be journaled... */
499 
500 		jh = commit_transaction->t_buffers;
501 
502 		/* If we're in abort mode, we just un-journal the buffer and
503 		   release it. */
504 
505 		if (is_journal_aborted(journal)) {
506 			clear_buffer_jbddirty(jh2bh(jh));
507 			JBUFFER_TRACE(jh, "journal is aborting: refile");
508 			jbd2_buffer_abort_trigger(jh,
509 						  jh->b_frozen_data ?
510 						  jh->b_frozen_triggers :
511 						  jh->b_triggers);
512 			jbd2_journal_refile_buffer(journal, jh);
513 			/* If that was the last one, we need to clean up
514 			 * any descriptor buffers which may have been
515 			 * already allocated, even if we are now
516 			 * aborting. */
517 			if (!commit_transaction->t_buffers)
518 				goto start_journal_io;
519 			continue;
520 		}
521 
522 		/* Make sure we have a descriptor block in which to
523 		   record the metadata buffer. */
524 
525 		if (!descriptor) {
526 			struct buffer_head *bh;
527 
528 			J_ASSERT (bufs == 0);
529 
530 			jbd_debug(4, "JBD: get descriptor\n");
531 
532 			descriptor = jbd2_journal_get_descriptor_buffer(journal);
533 			if (!descriptor) {
534 				jbd2_journal_abort(journal, -EIO);
535 				continue;
536 			}
537 
538 			bh = jh2bh(descriptor);
539 			jbd_debug(4, "JBD: got buffer %llu (%p)\n",
540 				(unsigned long long)bh->b_blocknr, bh->b_data);
541 			header = (journal_header_t *)&bh->b_data[0];
542 			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
543 			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
544 			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
545 
546 			tagp = &bh->b_data[sizeof(journal_header_t)];
547 			space_left = bh->b_size - sizeof(journal_header_t);
548 			first_tag = 1;
549 			set_buffer_jwrite(bh);
550 			set_buffer_dirty(bh);
551 			wbuf[bufs++] = bh;
552 
553 			/* Record it so that we can wait for IO
554                            completion later */
555 			BUFFER_TRACE(bh, "ph3: file as descriptor");
556 			jbd2_journal_file_buffer(descriptor, commit_transaction,
557 					BJ_LogCtl);
558 		}
559 
560 		/* Where is the buffer to be written? */
561 
562 		err = jbd2_journal_next_log_block(journal, &blocknr);
563 		/* If the block mapping failed, just abandon the buffer
564 		   and repeat this loop: we'll fall into the
565 		   refile-on-abort condition above. */
566 		if (err) {
567 			jbd2_journal_abort(journal, err);
568 			continue;
569 		}
570 
571 		/*
572 		 * start_this_handle() uses t_outstanding_credits to determine
573 		 * the free space in the log, but this counter is changed
574 		 * by jbd2_journal_next_log_block() also.
575 		 */
576 		atomic_dec(&commit_transaction->t_outstanding_credits);
577 
578 		/* Bump b_count to prevent truncate from stumbling over
579                    the shadowed buffer!  @@@ This can go if we ever get
580                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
581 		atomic_inc(&jh2bh(jh)->b_count);
582 
583 		/* Make a temporary IO buffer with which to write it out
584                    (this will requeue both the metadata buffer and the
585                    temporary IO buffer). new_bh goes on BJ_IO*/
586 
587 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
588 		/*
589 		 * akpm: jbd2_journal_write_metadata_buffer() sets
590 		 * new_bh->b_transaction to commit_transaction.
591 		 * We need to clean this up before we release new_bh
592 		 * (which is of type BJ_IO)
593 		 */
594 		JBUFFER_TRACE(jh, "ph3: write metadata");
595 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
596 						      jh, &new_jh, blocknr);
597 		if (flags < 0) {
598 			jbd2_journal_abort(journal, flags);
599 			continue;
600 		}
601 		set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
602 		wbuf[bufs++] = jh2bh(new_jh);
603 
604 		/* Record the new block's tag in the current descriptor
605                    buffer */
606 
607 		tag_flag = 0;
608 		if (flags & 1)
609 			tag_flag |= JBD2_FLAG_ESCAPE;
610 		if (!first_tag)
611 			tag_flag |= JBD2_FLAG_SAME_UUID;
612 
613 		tag = (journal_block_tag_t *) tagp;
614 		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
615 		tag->t_flags = cpu_to_be32(tag_flag);
616 		tagp += tag_bytes;
617 		space_left -= tag_bytes;
618 
619 		if (first_tag) {
620 			memcpy (tagp, journal->j_uuid, 16);
621 			tagp += 16;
622 			space_left -= 16;
623 			first_tag = 0;
624 		}
625 
626 		/* If there's no more to do, or if the descriptor is full,
627 		   let the IO rip! */
628 
629 		if (bufs == journal->j_wbufsize ||
630 		    commit_transaction->t_buffers == NULL ||
631 		    space_left < tag_bytes + 16) {
632 
633 			jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
634 
635 			/* Write an end-of-descriptor marker before
636                            submitting the IOs.  "tag" still points to
637                            the last tag we set up. */
638 
639 			tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
640 
641 start_journal_io:
642 			for (i = 0; i < bufs; i++) {
643 				struct buffer_head *bh = wbuf[i];
644 				/*
645 				 * Compute checksum.
646 				 */
647 				if (JBD2_HAS_COMPAT_FEATURE(journal,
648 					JBD2_FEATURE_COMPAT_CHECKSUM)) {
649 					crc32_sum =
650 					    jbd2_checksum_data(crc32_sum, bh);
651 				}
652 
653 				lock_buffer(bh);
654 				clear_buffer_dirty(bh);
655 				set_buffer_uptodate(bh);
656 				bh->b_end_io = journal_end_buffer_io_sync;
657 				submit_bh(write_op, bh);
658 			}
659 			cond_resched();
660 			stats.run.rs_blocks_logged += bufs;
661 
662 			/* Force a new descriptor to be generated next
663                            time round the loop. */
664 			descriptor = NULL;
665 			bufs = 0;
666 		}
667 	}
668 
669 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
670 	if (err) {
671 		printk(KERN_WARNING
672 			"JBD2: Detected IO errors while flushing file data "
673 		       "on %s\n", journal->j_devname);
674 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
675 			jbd2_journal_abort(journal, err);
676 		err = 0;
677 	}
678 
679 	/*
680 	 * If the journal is not located on the file system device,
681 	 * then we must flush the file system device before we issue
682 	 * the commit record
683 	 */
684 	if (commit_transaction->t_flushed_data_blocks &&
685 	    (journal->j_fs_dev != journal->j_dev) &&
686 	    (journal->j_flags & JBD2_BARRIER))
687 		blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
688 
689 	/* Done it all: now write the commit record asynchronously. */
690 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
691 				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
692 		err = journal_submit_commit_record(journal, commit_transaction,
693 						 &cbh, crc32_sum);
694 		if (err)
695 			__jbd2_journal_abort_hard(journal);
696 	}
697 
698 	/* Lo and behold: we have just managed to send a transaction to
699            the log.  Before we can commit it, wait for the IO so far to
700            complete.  Control buffers being written are on the
701            transaction's t_log_list queue, and metadata buffers are on
702            the t_iobuf_list queue.
703 
704 	   Wait for the buffers in reverse order.  That way we are
705 	   less likely to be woken up until all IOs have completed, and
706 	   so we incur less scheduling load.
707 	*/
708 
709 	jbd_debug(3, "JBD: commit phase 3\n");
710 
711 	/*
712 	 * akpm: these are BJ_IO, and j_list_lock is not needed.
713 	 * See __journal_try_to_free_buffer.
714 	 */
715 wait_for_iobuf:
716 	while (commit_transaction->t_iobuf_list != NULL) {
717 		struct buffer_head *bh;
718 
719 		jh = commit_transaction->t_iobuf_list->b_tprev;
720 		bh = jh2bh(jh);
721 		if (buffer_locked(bh)) {
722 			wait_on_buffer(bh);
723 			goto wait_for_iobuf;
724 		}
725 		if (cond_resched())
726 			goto wait_for_iobuf;
727 
728 		if (unlikely(!buffer_uptodate(bh)))
729 			err = -EIO;
730 
731 		clear_buffer_jwrite(bh);
732 
733 		JBUFFER_TRACE(jh, "ph4: unfile after journal write");
734 		jbd2_journal_unfile_buffer(journal, jh);
735 
736 		/*
737 		 * ->t_iobuf_list should contain only dummy buffer_heads
738 		 * which were created by jbd2_journal_write_metadata_buffer().
739 		 */
740 		BUFFER_TRACE(bh, "dumping temporary bh");
741 		jbd2_journal_put_journal_head(jh);
742 		__brelse(bh);
743 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
744 		free_buffer_head(bh);
745 
746 		/* We also have to unlock and free the corresponding
747                    shadowed buffer */
748 		jh = commit_transaction->t_shadow_list->b_tprev;
749 		bh = jh2bh(jh);
750 		clear_bit(BH_JWrite, &bh->b_state);
751 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
752 
753 		/* The metadata is now released for reuse, but we need
754                    to remember it against this transaction so that when
755                    we finally commit, we can do any checkpointing
756                    required. */
757 		JBUFFER_TRACE(jh, "file as BJ_Forget");
758 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
759 		/* Wake up any transactions which were waiting for this
760 		   IO to complete */
761 		wake_up_bit(&bh->b_state, BH_Unshadow);
762 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
763 		__brelse(bh);
764 	}
765 
766 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
767 
768 	jbd_debug(3, "JBD: commit phase 4\n");
769 
770 	/* Here we wait for the revoke record and descriptor record buffers */
771  wait_for_ctlbuf:
772 	while (commit_transaction->t_log_list != NULL) {
773 		struct buffer_head *bh;
774 
775 		jh = commit_transaction->t_log_list->b_tprev;
776 		bh = jh2bh(jh);
777 		if (buffer_locked(bh)) {
778 			wait_on_buffer(bh);
779 			goto wait_for_ctlbuf;
780 		}
781 		if (cond_resched())
782 			goto wait_for_ctlbuf;
783 
784 		if (unlikely(!buffer_uptodate(bh)))
785 			err = -EIO;
786 
787 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
788 		clear_buffer_jwrite(bh);
789 		jbd2_journal_unfile_buffer(journal, jh);
790 		jbd2_journal_put_journal_head(jh);
791 		__brelse(bh);		/* One for getblk */
792 		/* AKPM: bforget here */
793 	}
794 
795 	if (err)
796 		jbd2_journal_abort(journal, err);
797 
798 	jbd_debug(3, "JBD: commit phase 5\n");
799 
800 	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
801 				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
802 		err = journal_submit_commit_record(journal, commit_transaction,
803 						&cbh, crc32_sum);
804 		if (err)
805 			__jbd2_journal_abort_hard(journal);
806 	}
807 	if (!err && !is_journal_aborted(journal))
808 		err = journal_wait_on_commit_record(journal, cbh);
809 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
810 				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
811 	    journal->j_flags & JBD2_BARRIER) {
812 		blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL);
813 	}
814 
815 	if (err)
816 		jbd2_journal_abort(journal, err);
817 
818 	/* End of a transaction!  Finally, we can do checkpoint
819            processing: any buffers committed as a result of this
820            transaction can be removed from any checkpoint list it was on
821            before. */
822 
823 	jbd_debug(3, "JBD: commit phase 6\n");
824 
825 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
826 	J_ASSERT(commit_transaction->t_buffers == NULL);
827 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
828 	J_ASSERT(commit_transaction->t_iobuf_list == NULL);
829 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
830 	J_ASSERT(commit_transaction->t_log_list == NULL);
831 
832 restart_loop:
833 	/*
834 	 * As there are other places (journal_unmap_buffer()) adding buffers
835 	 * to this list we have to be careful and hold the j_list_lock.
836 	 */
837 	spin_lock(&journal->j_list_lock);
838 	while (commit_transaction->t_forget) {
839 		transaction_t *cp_transaction;
840 		struct buffer_head *bh;
841 
842 		jh = commit_transaction->t_forget;
843 		spin_unlock(&journal->j_list_lock);
844 		bh = jh2bh(jh);
845 		jbd_lock_bh_state(bh);
846 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction);
847 
848 		/*
849 		 * If there is undo-protected committed data against
850 		 * this buffer, then we can remove it now.  If it is a
851 		 * buffer needing such protection, the old frozen_data
852 		 * field now points to a committed version of the
853 		 * buffer, so rotate that field to the new committed
854 		 * data.
855 		 *
856 		 * Otherwise, we can just throw away the frozen data now.
857 		 *
858 		 * We also know that the frozen data has already fired
859 		 * its triggers if they exist, so we can clear that too.
860 		 */
861 		if (jh->b_committed_data) {
862 			jbd2_free(jh->b_committed_data, bh->b_size);
863 			jh->b_committed_data = NULL;
864 			if (jh->b_frozen_data) {
865 				jh->b_committed_data = jh->b_frozen_data;
866 				jh->b_frozen_data = NULL;
867 				jh->b_frozen_triggers = NULL;
868 			}
869 		} else if (jh->b_frozen_data) {
870 			jbd2_free(jh->b_frozen_data, bh->b_size);
871 			jh->b_frozen_data = NULL;
872 			jh->b_frozen_triggers = NULL;
873 		}
874 
875 		spin_lock(&journal->j_list_lock);
876 		cp_transaction = jh->b_cp_transaction;
877 		if (cp_transaction) {
878 			JBUFFER_TRACE(jh, "remove from old cp transaction");
879 			cp_transaction->t_chp_stats.cs_dropped++;
880 			__jbd2_journal_remove_checkpoint(jh);
881 		}
882 
883 		/* Only re-checkpoint the buffer_head if it is marked
884 		 * dirty.  If the buffer was added to the BJ_Forget list
885 		 * by jbd2_journal_forget, it may no longer be dirty and
886 		 * there's no point in keeping a checkpoint record for
887 		 * it. */
888 
889 		/* A buffer which has been freed while still being
890 		 * journaled by a previous transaction may end up still
891 		 * being dirty here, but we want to avoid writing back
892 		 * that buffer in the future after the "add to orphan"
893 		 * operation been committed,  That's not only a performance
894 		 * gain, it also stops aliasing problems if the buffer is
895 		 * left behind for writeback and gets reallocated for another
896 		 * use in a different page. */
897 		if (buffer_freed(bh) && !jh->b_next_transaction) {
898 			clear_buffer_freed(bh);
899 			clear_buffer_jbddirty(bh);
900 		}
901 
902 		if (buffer_jbddirty(bh)) {
903 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
904 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
905 			if (is_journal_aborted(journal))
906 				clear_buffer_jbddirty(bh);
907 			JBUFFER_TRACE(jh, "refile for checkpoint writeback");
908 			__jbd2_journal_refile_buffer(jh);
909 			jbd_unlock_bh_state(bh);
910 		} else {
911 			J_ASSERT_BH(bh, !buffer_dirty(bh));
912 			/* The buffer on BJ_Forget list and not jbddirty means
913 			 * it has been freed by this transaction and hence it
914 			 * could not have been reallocated until this
915 			 * transaction has committed. *BUT* it could be
916 			 * reallocated once we have written all the data to
917 			 * disk and before we process the buffer on BJ_Forget
918 			 * list. */
919 			JBUFFER_TRACE(jh, "refile or unfile freed buffer");
920 			__jbd2_journal_refile_buffer(jh);
921 			if (!jh->b_transaction) {
922 				jbd_unlock_bh_state(bh);
923 				 /* needs a brelse */
924 				jbd2_journal_remove_journal_head(bh);
925 				release_buffer_page(bh);
926 			} else
927 				jbd_unlock_bh_state(bh);
928 		}
929 		cond_resched_lock(&journal->j_list_lock);
930 	}
931 	spin_unlock(&journal->j_list_lock);
932 	/*
933 	 * This is a bit sleazy.  We use j_list_lock to protect transition
934 	 * of a transaction into T_FINISHED state and calling
935 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
936 	 * other checkpointing code processing the transaction...
937 	 */
938 	write_lock(&journal->j_state_lock);
939 	spin_lock(&journal->j_list_lock);
940 	/*
941 	 * Now recheck if some buffers did not get attached to the transaction
942 	 * while the lock was dropped...
943 	 */
944 	if (commit_transaction->t_forget) {
945 		spin_unlock(&journal->j_list_lock);
946 		write_unlock(&journal->j_state_lock);
947 		goto restart_loop;
948 	}
949 
950 	/* Done with this transaction! */
951 
952 	jbd_debug(3, "JBD: commit phase 7\n");
953 
954 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
955 
956 	commit_transaction->t_start = jiffies;
957 	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
958 					      commit_transaction->t_start);
959 
960 	/*
961 	 * File the transaction statistics
962 	 */
963 	stats.ts_tid = commit_transaction->t_tid;
964 	stats.run.rs_handle_count =
965 		atomic_read(&commit_transaction->t_handle_count);
966 	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
967 			     commit_transaction->t_tid, &stats.run);
968 
969 	/*
970 	 * Calculate overall stats
971 	 */
972 	spin_lock(&journal->j_history_lock);
973 	journal->j_stats.ts_tid++;
974 	journal->j_stats.run.rs_wait += stats.run.rs_wait;
975 	journal->j_stats.run.rs_running += stats.run.rs_running;
976 	journal->j_stats.run.rs_locked += stats.run.rs_locked;
977 	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
978 	journal->j_stats.run.rs_logging += stats.run.rs_logging;
979 	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
980 	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
981 	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
982 	spin_unlock(&journal->j_history_lock);
983 
984 	commit_transaction->t_state = T_FINISHED;
985 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
986 	journal->j_commit_sequence = commit_transaction->t_tid;
987 	journal->j_committing_transaction = NULL;
988 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
989 
990 	/*
991 	 * weight the commit time higher than the average time so we don't
992 	 * react too strongly to vast changes in the commit time
993 	 */
994 	if (likely(journal->j_average_commit_time))
995 		journal->j_average_commit_time = (commit_time +
996 				journal->j_average_commit_time*3) / 4;
997 	else
998 		journal->j_average_commit_time = commit_time;
999 	write_unlock(&journal->j_state_lock);
1000 
1001 	if (commit_transaction->t_checkpoint_list == NULL &&
1002 	    commit_transaction->t_checkpoint_io_list == NULL) {
1003 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1004 		to_free = 1;
1005 	} else {
1006 		if (journal->j_checkpoint_transactions == NULL) {
1007 			journal->j_checkpoint_transactions = commit_transaction;
1008 			commit_transaction->t_cpnext = commit_transaction;
1009 			commit_transaction->t_cpprev = commit_transaction;
1010 		} else {
1011 			commit_transaction->t_cpnext =
1012 				journal->j_checkpoint_transactions;
1013 			commit_transaction->t_cpprev =
1014 				commit_transaction->t_cpnext->t_cpprev;
1015 			commit_transaction->t_cpnext->t_cpprev =
1016 				commit_transaction;
1017 			commit_transaction->t_cpprev->t_cpnext =
1018 				commit_transaction;
1019 		}
1020 	}
1021 	spin_unlock(&journal->j_list_lock);
1022 
1023 	if (journal->j_commit_callback)
1024 		journal->j_commit_callback(journal, commit_transaction);
1025 
1026 	trace_jbd2_end_commit(journal, commit_transaction);
1027 	jbd_debug(1, "JBD: commit %d complete, head %d\n",
1028 		  journal->j_commit_sequence, journal->j_tail_sequence);
1029 	if (to_free)
1030 		kfree(commit_transaction);
1031 
1032 	wake_up(&journal->j_wait_done_commit);
1033 }
1034