xref: /linux/fs/jbd2/commit.c (revision 98b8788ae91694499d1995035625bea16a4db0c4)
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 	int barrier_done = 0;
105 	struct timespec now = current_kernel_time();
106 
107 	if (is_journal_aborted(journal))
108 		return 0;
109 
110 	descriptor = jbd2_journal_get_descriptor_buffer(journal);
111 	if (!descriptor)
112 		return 1;
113 
114 	bh = jh2bh(descriptor);
115 
116 	tmp = (struct commit_header *)bh->b_data;
117 	tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
118 	tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
119 	tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
120 	tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
121 	tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
122 
123 	if (JBD2_HAS_COMPAT_FEATURE(journal,
124 				    JBD2_FEATURE_COMPAT_CHECKSUM)) {
125 		tmp->h_chksum_type 	= JBD2_CRC32_CHKSUM;
126 		tmp->h_chksum_size 	= JBD2_CRC32_CHKSUM_SIZE;
127 		tmp->h_chksum[0] 	= cpu_to_be32(crc32_sum);
128 	}
129 
130 	JBUFFER_TRACE(descriptor, "submit commit block");
131 	lock_buffer(bh);
132 	clear_buffer_dirty(bh);
133 	set_buffer_uptodate(bh);
134 	bh->b_end_io = journal_end_buffer_io_sync;
135 
136 	if (journal->j_flags & JBD2_BARRIER &&
137 	    !JBD2_HAS_INCOMPAT_FEATURE(journal,
138 				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
139 		set_buffer_ordered(bh);
140 		barrier_done = 1;
141 	}
142 	ret = submit_bh(WRITE_SYNC_PLUG, bh);
143 	if (barrier_done)
144 		clear_buffer_ordered(bh);
145 
146 	/* is it possible for another commit to fail at roughly
147 	 * the same time as this one?  If so, we don't want to
148 	 * trust the barrier flag in the super, but instead want
149 	 * to remember if we sent a barrier request
150 	 */
151 	if (ret == -EOPNOTSUPP && barrier_done) {
152 		printk(KERN_WARNING
153 		       "JBD: barrier-based sync failed on %s - "
154 		       "disabling barriers\n", journal->j_devname);
155 		spin_lock(&journal->j_state_lock);
156 		journal->j_flags &= ~JBD2_BARRIER;
157 		spin_unlock(&journal->j_state_lock);
158 
159 		/* And try again, without the barrier */
160 		lock_buffer(bh);
161 		set_buffer_uptodate(bh);
162 		clear_buffer_dirty(bh);
163 		ret = submit_bh(WRITE_SYNC_PLUG, bh);
164 	}
165 	*cbh = bh;
166 	return ret;
167 }
168 
169 /*
170  * This function along with journal_submit_commit_record
171  * allows to write the commit record asynchronously.
172  */
173 static int journal_wait_on_commit_record(journal_t *journal,
174 					 struct buffer_head *bh)
175 {
176 	int ret = 0;
177 
178 retry:
179 	clear_buffer_dirty(bh);
180 	wait_on_buffer(bh);
181 	if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
182 		printk(KERN_WARNING
183 		       "JBD2: wait_on_commit_record: sync failed on %s - "
184 		       "disabling barriers\n", journal->j_devname);
185 		spin_lock(&journal->j_state_lock);
186 		journal->j_flags &= ~JBD2_BARRIER;
187 		spin_unlock(&journal->j_state_lock);
188 
189 		lock_buffer(bh);
190 		clear_buffer_dirty(bh);
191 		set_buffer_uptodate(bh);
192 		bh->b_end_io = journal_end_buffer_io_sync;
193 
194 		ret = submit_bh(WRITE_SYNC_PLUG, bh);
195 		if (ret) {
196 			unlock_buffer(bh);
197 			return ret;
198 		}
199 		goto retry;
200 	}
201 
202 	if (unlikely(!buffer_uptodate(bh)))
203 		ret = -EIO;
204 	put_bh(bh);            /* One for getblk() */
205 	jbd2_journal_put_journal_head(bh2jh(bh));
206 
207 	return ret;
208 }
209 
210 /*
211  * write the filemap data using writepage() address_space_operations.
212  * We don't do block allocation here even for delalloc. We don't
213  * use writepages() because with dealyed allocation we may be doing
214  * block allocation in writepages().
215  */
216 static int journal_submit_inode_data_buffers(struct address_space *mapping)
217 {
218 	int ret;
219 	struct writeback_control wbc = {
220 		.sync_mode =  WB_SYNC_ALL,
221 		.nr_to_write = mapping->nrpages * 2,
222 		.range_start = 0,
223 		.range_end = i_size_read(mapping->host),
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 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
257 		err = journal_submit_inode_data_buffers(mapping);
258 		if (!ret)
259 			ret = err;
260 		spin_lock(&journal->j_list_lock);
261 		J_ASSERT(jinode->i_transaction == commit_transaction);
262 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
263 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
264 	}
265 	spin_unlock(&journal->j_list_lock);
266 	return ret;
267 }
268 
269 /*
270  * Wait for data submitted for writeout, refile inodes to proper
271  * transaction if needed.
272  *
273  */
274 static int journal_finish_inode_data_buffers(journal_t *journal,
275 		transaction_t *commit_transaction)
276 {
277 	struct jbd2_inode *jinode, *next_i;
278 	int err, ret = 0;
279 
280 	/* For locking, see the comment in journal_submit_data_buffers() */
281 	spin_lock(&journal->j_list_lock);
282 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
283 		jinode->i_flags |= JI_COMMIT_RUNNING;
284 		spin_unlock(&journal->j_list_lock);
285 		err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
286 		if (err) {
287 			/*
288 			 * Because AS_EIO is cleared by
289 			 * filemap_fdatawait_range(), set it again so
290 			 * that user process can get -EIO from fsync().
291 			 */
292 			set_bit(AS_EIO,
293 				&jinode->i_vfs_inode->i_mapping->flags);
294 
295 			if (!ret)
296 				ret = err;
297 		}
298 		spin_lock(&journal->j_list_lock);
299 		jinode->i_flags &= ~JI_COMMIT_RUNNING;
300 		wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
301 	}
302 
303 	/* Now refile inode to proper lists */
304 	list_for_each_entry_safe(jinode, next_i,
305 				 &commit_transaction->t_inode_list, i_list) {
306 		list_del(&jinode->i_list);
307 		if (jinode->i_next_transaction) {
308 			jinode->i_transaction = jinode->i_next_transaction;
309 			jinode->i_next_transaction = NULL;
310 			list_add(&jinode->i_list,
311 				&jinode->i_transaction->t_inode_list);
312 		} else {
313 			jinode->i_transaction = NULL;
314 		}
315 	}
316 	spin_unlock(&journal->j_list_lock);
317 
318 	return ret;
319 }
320 
321 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
322 {
323 	struct page *page = bh->b_page;
324 	char *addr;
325 	__u32 checksum;
326 
327 	addr = kmap_atomic(page, KM_USER0);
328 	checksum = crc32_be(crc32_sum,
329 		(void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
330 	kunmap_atomic(addr, KM_USER0);
331 
332 	return checksum;
333 }
334 
335 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
336 				   unsigned long long block)
337 {
338 	tag->t_blocknr = cpu_to_be32(block & (u32)~0);
339 	if (tag_bytes > JBD2_TAG_SIZE32)
340 		tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
341 }
342 
343 /*
344  * jbd2_journal_commit_transaction
345  *
346  * The primary function for committing a transaction to the log.  This
347  * function is called by the journal thread to begin a complete commit.
348  */
349 void jbd2_journal_commit_transaction(journal_t *journal)
350 {
351 	struct transaction_stats_s stats;
352 	transaction_t *commit_transaction;
353 	struct journal_head *jh, *new_jh, *descriptor;
354 	struct buffer_head **wbuf = journal->j_wbuf;
355 	int bufs;
356 	int flags;
357 	int err;
358 	unsigned long long blocknr;
359 	ktime_t start_time;
360 	u64 commit_time;
361 	char *tagp = NULL;
362 	journal_header_t *header;
363 	journal_block_tag_t *tag = NULL;
364 	int space_left = 0;
365 	int first_tag = 0;
366 	int tag_flag;
367 	int i, to_free = 0;
368 	int tag_bytes = journal_tag_bytes(journal);
369 	struct buffer_head *cbh = NULL; /* For transactional checksums */
370 	__u32 crc32_sum = ~0;
371 	int write_op = WRITE;
372 
373 	/*
374 	 * First job: lock down the current transaction and wait for
375 	 * all outstanding updates to complete.
376 	 */
377 
378 #ifdef COMMIT_STATS
379 	spin_lock(&journal->j_list_lock);
380 	summarise_journal_usage(journal);
381 	spin_unlock(&journal->j_list_lock);
382 #endif
383 
384 	/* Do we need to erase the effects of a prior jbd2_journal_flush? */
385 	if (journal->j_flags & JBD2_FLUSHED) {
386 		jbd_debug(3, "super block updated\n");
387 		jbd2_journal_update_superblock(journal, 1);
388 	} else {
389 		jbd_debug(3, "superblock not updated\n");
390 	}
391 
392 	J_ASSERT(journal->j_running_transaction != NULL);
393 	J_ASSERT(journal->j_committing_transaction == NULL);
394 
395 	commit_transaction = journal->j_running_transaction;
396 	J_ASSERT(commit_transaction->t_state == T_RUNNING);
397 
398 	trace_jbd2_start_commit(journal, commit_transaction);
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 	trace_jbd2_commit_locking(journal, commit_transaction);
413 	stats.run.rs_wait = commit_transaction->t_max_wait;
414 	stats.run.rs_locked = jiffies;
415 	stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
416 					      stats.run.rs_locked);
417 
418 	spin_lock(&commit_transaction->t_handle_lock);
419 	while (commit_transaction->t_updates) {
420 		DEFINE_WAIT(wait);
421 
422 		prepare_to_wait(&journal->j_wait_updates, &wait,
423 					TASK_UNINTERRUPTIBLE);
424 		if (commit_transaction->t_updates) {
425 			spin_unlock(&commit_transaction->t_handle_lock);
426 			spin_unlock(&journal->j_state_lock);
427 			schedule();
428 			spin_lock(&journal->j_state_lock);
429 			spin_lock(&commit_transaction->t_handle_lock);
430 		}
431 		finish_wait(&journal->j_wait_updates, &wait);
432 	}
433 	spin_unlock(&commit_transaction->t_handle_lock);
434 
435 	J_ASSERT (commit_transaction->t_outstanding_credits <=
436 			journal->j_max_transaction_buffers);
437 
438 	/*
439 	 * First thing we are allowed to do is to discard any remaining
440 	 * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
441 	 * that there are no such buffers: if a large filesystem
442 	 * operation like a truncate needs to split itself over multiple
443 	 * transactions, then it may try to do a jbd2_journal_restart() while
444 	 * there are still BJ_Reserved buffers outstanding.  These must
445 	 * be released cleanly from the current transaction.
446 	 *
447 	 * In this case, the filesystem must still reserve write access
448 	 * again before modifying the buffer in the new transaction, but
449 	 * we do not require it to remember exactly which old buffers it
450 	 * has reserved.  This is consistent with the existing behaviour
451 	 * that multiple jbd2_journal_get_write_access() calls to the same
452 	 * buffer are perfectly permissable.
453 	 */
454 	while (commit_transaction->t_reserved_list) {
455 		jh = commit_transaction->t_reserved_list;
456 		JBUFFER_TRACE(jh, "reserved, unused: refile");
457 		/*
458 		 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
459 		 * leave undo-committed data.
460 		 */
461 		if (jh->b_committed_data) {
462 			struct buffer_head *bh = jh2bh(jh);
463 
464 			jbd_lock_bh_state(bh);
465 			jbd2_free(jh->b_committed_data, bh->b_size);
466 			jh->b_committed_data = NULL;
467 			jbd_unlock_bh_state(bh);
468 		}
469 		jbd2_journal_refile_buffer(journal, jh);
470 	}
471 
472 	/*
473 	 * Now try to drop any written-back buffers from the journal's
474 	 * checkpoint lists.  We do this *before* commit because it potentially
475 	 * frees some memory
476 	 */
477 	spin_lock(&journal->j_list_lock);
478 	__jbd2_journal_clean_checkpoint_list(journal);
479 	spin_unlock(&journal->j_list_lock);
480 
481 	jbd_debug (3, "JBD: commit phase 1\n");
482 
483 	/*
484 	 * Switch to a new revoke table.
485 	 */
486 	jbd2_journal_switch_revoke_table(journal);
487 
488 	trace_jbd2_commit_flushing(journal, commit_transaction);
489 	stats.run.rs_flushing = jiffies;
490 	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
491 					     stats.run.rs_flushing);
492 
493 	commit_transaction->t_state = T_FLUSH;
494 	journal->j_committing_transaction = commit_transaction;
495 	journal->j_running_transaction = NULL;
496 	start_time = ktime_get();
497 	commit_transaction->t_log_start = journal->j_head;
498 	wake_up(&journal->j_wait_transaction_locked);
499 	spin_unlock(&journal->j_state_lock);
500 
501 	jbd_debug (3, "JBD: commit phase 2\n");
502 
503 	/*
504 	 * Now start flushing things to disk, in the order they appear
505 	 * on the transaction lists.  Data blocks go first.
506 	 */
507 	err = journal_submit_data_buffers(journal, commit_transaction);
508 	if (err)
509 		jbd2_journal_abort(journal, err);
510 
511 	jbd2_journal_write_revoke_records(journal, commit_transaction,
512 					  write_op);
513 
514 	jbd_debug(3, "JBD: commit phase 2\n");
515 
516 	/*
517 	 * Way to go: we have now written out all of the data for a
518 	 * transaction!  Now comes the tricky part: we need to write out
519 	 * metadata.  Loop over the transaction's entire buffer list:
520 	 */
521 	spin_lock(&journal->j_state_lock);
522 	commit_transaction->t_state = T_COMMIT;
523 	spin_unlock(&journal->j_state_lock);
524 
525 	trace_jbd2_commit_logging(journal, commit_transaction);
526 	stats.run.rs_logging = jiffies;
527 	stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
528 					       stats.run.rs_logging);
529 	stats.run.rs_blocks = commit_transaction->t_outstanding_credits;
530 	stats.run.rs_blocks_logged = 0;
531 
532 	J_ASSERT(commit_transaction->t_nr_buffers <=
533 		 commit_transaction->t_outstanding_credits);
534 
535 	err = 0;
536 	descriptor = NULL;
537 	bufs = 0;
538 	while (commit_transaction->t_buffers) {
539 
540 		/* Find the next buffer to be journaled... */
541 
542 		jh = commit_transaction->t_buffers;
543 
544 		/* If we're in abort mode, we just un-journal the buffer and
545 		   release it. */
546 
547 		if (is_journal_aborted(journal)) {
548 			clear_buffer_jbddirty(jh2bh(jh));
549 			JBUFFER_TRACE(jh, "journal is aborting: refile");
550 			jbd2_buffer_abort_trigger(jh,
551 						  jh->b_frozen_data ?
552 						  jh->b_frozen_triggers :
553 						  jh->b_triggers);
554 			jbd2_journal_refile_buffer(journal, jh);
555 			/* If that was the last one, we need to clean up
556 			 * any descriptor buffers which may have been
557 			 * already allocated, even if we are now
558 			 * aborting. */
559 			if (!commit_transaction->t_buffers)
560 				goto start_journal_io;
561 			continue;
562 		}
563 
564 		/* Make sure we have a descriptor block in which to
565 		   record the metadata buffer. */
566 
567 		if (!descriptor) {
568 			struct buffer_head *bh;
569 
570 			J_ASSERT (bufs == 0);
571 
572 			jbd_debug(4, "JBD: get descriptor\n");
573 
574 			descriptor = jbd2_journal_get_descriptor_buffer(journal);
575 			if (!descriptor) {
576 				jbd2_journal_abort(journal, -EIO);
577 				continue;
578 			}
579 
580 			bh = jh2bh(descriptor);
581 			jbd_debug(4, "JBD: got buffer %llu (%p)\n",
582 				(unsigned long long)bh->b_blocknr, bh->b_data);
583 			header = (journal_header_t *)&bh->b_data[0];
584 			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
585 			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
586 			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
587 
588 			tagp = &bh->b_data[sizeof(journal_header_t)];
589 			space_left = bh->b_size - sizeof(journal_header_t);
590 			first_tag = 1;
591 			set_buffer_jwrite(bh);
592 			set_buffer_dirty(bh);
593 			wbuf[bufs++] = bh;
594 
595 			/* Record it so that we can wait for IO
596                            completion later */
597 			BUFFER_TRACE(bh, "ph3: file as descriptor");
598 			jbd2_journal_file_buffer(descriptor, commit_transaction,
599 					BJ_LogCtl);
600 		}
601 
602 		/* Where is the buffer to be written? */
603 
604 		err = jbd2_journal_next_log_block(journal, &blocknr);
605 		/* If the block mapping failed, just abandon the buffer
606 		   and repeat this loop: we'll fall into the
607 		   refile-on-abort condition above. */
608 		if (err) {
609 			jbd2_journal_abort(journal, err);
610 			continue;
611 		}
612 
613 		/*
614 		 * start_this_handle() uses t_outstanding_credits to determine
615 		 * the free space in the log, but this counter is changed
616 		 * by jbd2_journal_next_log_block() also.
617 		 */
618 		commit_transaction->t_outstanding_credits--;
619 
620 		/* Bump b_count to prevent truncate from stumbling over
621                    the shadowed buffer!  @@@ This can go if we ever get
622                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
623 		atomic_inc(&jh2bh(jh)->b_count);
624 
625 		/* Make a temporary IO buffer with which to write it out
626                    (this will requeue both the metadata buffer and the
627                    temporary IO buffer). new_bh goes on BJ_IO*/
628 
629 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
630 		/*
631 		 * akpm: jbd2_journal_write_metadata_buffer() sets
632 		 * new_bh->b_transaction to commit_transaction.
633 		 * We need to clean this up before we release new_bh
634 		 * (which is of type BJ_IO)
635 		 */
636 		JBUFFER_TRACE(jh, "ph3: write metadata");
637 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
638 						      jh, &new_jh, blocknr);
639 		if (flags < 0) {
640 			jbd2_journal_abort(journal, flags);
641 			continue;
642 		}
643 		set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
644 		wbuf[bufs++] = jh2bh(new_jh);
645 
646 		/* Record the new block's tag in the current descriptor
647                    buffer */
648 
649 		tag_flag = 0;
650 		if (flags & 1)
651 			tag_flag |= JBD2_FLAG_ESCAPE;
652 		if (!first_tag)
653 			tag_flag |= JBD2_FLAG_SAME_UUID;
654 
655 		tag = (journal_block_tag_t *) tagp;
656 		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
657 		tag->t_flags = cpu_to_be32(tag_flag);
658 		tagp += tag_bytes;
659 		space_left -= tag_bytes;
660 
661 		if (first_tag) {
662 			memcpy (tagp, journal->j_uuid, 16);
663 			tagp += 16;
664 			space_left -= 16;
665 			first_tag = 0;
666 		}
667 
668 		/* If there's no more to do, or if the descriptor is full,
669 		   let the IO rip! */
670 
671 		if (bufs == journal->j_wbufsize ||
672 		    commit_transaction->t_buffers == NULL ||
673 		    space_left < tag_bytes + 16) {
674 
675 			jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
676 
677 			/* Write an end-of-descriptor marker before
678                            submitting the IOs.  "tag" still points to
679                            the last tag we set up. */
680 
681 			tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
682 
683 start_journal_io:
684 			for (i = 0; i < bufs; i++) {
685 				struct buffer_head *bh = wbuf[i];
686 				/*
687 				 * Compute checksum.
688 				 */
689 				if (JBD2_HAS_COMPAT_FEATURE(journal,
690 					JBD2_FEATURE_COMPAT_CHECKSUM)) {
691 					crc32_sum =
692 					    jbd2_checksum_data(crc32_sum, bh);
693 				}
694 
695 				lock_buffer(bh);
696 				clear_buffer_dirty(bh);
697 				set_buffer_uptodate(bh);
698 				bh->b_end_io = journal_end_buffer_io_sync;
699 				submit_bh(write_op, bh);
700 			}
701 			cond_resched();
702 			stats.run.rs_blocks_logged += bufs;
703 
704 			/* Force a new descriptor to be generated next
705                            time round the loop. */
706 			descriptor = NULL;
707 			bufs = 0;
708 		}
709 	}
710 
711 	/* Done it all: now write the commit record asynchronously. */
712 
713 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
714 				      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
715 		err = journal_submit_commit_record(journal, commit_transaction,
716 						 &cbh, crc32_sum);
717 		if (err)
718 			__jbd2_journal_abort_hard(journal);
719 		if (journal->j_flags & JBD2_BARRIER)
720 			blkdev_issue_flush(journal->j_dev, NULL);
721 	}
722 
723 	/*
724 	 * This is the right place to wait for data buffers both for ASYNC
725 	 * and !ASYNC commit. If commit is ASYNC, we need to wait only after
726 	 * the commit block went to disk (which happens above). If commit is
727 	 * SYNC, we need to wait for data buffers before we start writing
728 	 * commit block, which happens below in such setting.
729 	 */
730 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
731 	if (err) {
732 		printk(KERN_WARNING
733 			"JBD2: Detected IO errors while flushing file data "
734 		       "on %s\n", journal->j_devname);
735 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
736 			jbd2_journal_abort(journal, err);
737 		err = 0;
738 	}
739 
740 	/* Lo and behold: we have just managed to send a transaction to
741            the log.  Before we can commit it, wait for the IO so far to
742            complete.  Control buffers being written are on the
743            transaction's t_log_list queue, and metadata buffers are on
744            the t_iobuf_list queue.
745 
746 	   Wait for the buffers in reverse order.  That way we are
747 	   less likely to be woken up until all IOs have completed, and
748 	   so we incur less scheduling load.
749 	*/
750 
751 	jbd_debug(3, "JBD: commit phase 3\n");
752 
753 	/*
754 	 * akpm: these are BJ_IO, and j_list_lock is not needed.
755 	 * See __journal_try_to_free_buffer.
756 	 */
757 wait_for_iobuf:
758 	while (commit_transaction->t_iobuf_list != NULL) {
759 		struct buffer_head *bh;
760 
761 		jh = commit_transaction->t_iobuf_list->b_tprev;
762 		bh = jh2bh(jh);
763 		if (buffer_locked(bh)) {
764 			wait_on_buffer(bh);
765 			goto wait_for_iobuf;
766 		}
767 		if (cond_resched())
768 			goto wait_for_iobuf;
769 
770 		if (unlikely(!buffer_uptodate(bh)))
771 			err = -EIO;
772 
773 		clear_buffer_jwrite(bh);
774 
775 		JBUFFER_TRACE(jh, "ph4: unfile after journal write");
776 		jbd2_journal_unfile_buffer(journal, jh);
777 
778 		/*
779 		 * ->t_iobuf_list should contain only dummy buffer_heads
780 		 * which were created by jbd2_journal_write_metadata_buffer().
781 		 */
782 		BUFFER_TRACE(bh, "dumping temporary bh");
783 		jbd2_journal_put_journal_head(jh);
784 		__brelse(bh);
785 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
786 		free_buffer_head(bh);
787 
788 		/* We also have to unlock and free the corresponding
789                    shadowed buffer */
790 		jh = commit_transaction->t_shadow_list->b_tprev;
791 		bh = jh2bh(jh);
792 		clear_bit(BH_JWrite, &bh->b_state);
793 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
794 
795 		/* The metadata is now released for reuse, but we need
796                    to remember it against this transaction so that when
797                    we finally commit, we can do any checkpointing
798                    required. */
799 		JBUFFER_TRACE(jh, "file as BJ_Forget");
800 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
801 		/* Wake up any transactions which were waiting for this
802 		   IO to complete */
803 		wake_up_bit(&bh->b_state, BH_Unshadow);
804 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
805 		__brelse(bh);
806 	}
807 
808 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
809 
810 	jbd_debug(3, "JBD: commit phase 4\n");
811 
812 	/* Here we wait for the revoke record and descriptor record buffers */
813  wait_for_ctlbuf:
814 	while (commit_transaction->t_log_list != NULL) {
815 		struct buffer_head *bh;
816 
817 		jh = commit_transaction->t_log_list->b_tprev;
818 		bh = jh2bh(jh);
819 		if (buffer_locked(bh)) {
820 			wait_on_buffer(bh);
821 			goto wait_for_ctlbuf;
822 		}
823 		if (cond_resched())
824 			goto wait_for_ctlbuf;
825 
826 		if (unlikely(!buffer_uptodate(bh)))
827 			err = -EIO;
828 
829 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
830 		clear_buffer_jwrite(bh);
831 		jbd2_journal_unfile_buffer(journal, jh);
832 		jbd2_journal_put_journal_head(jh);
833 		__brelse(bh);		/* One for getblk */
834 		/* AKPM: bforget here */
835 	}
836 
837 	if (err)
838 		jbd2_journal_abort(journal, err);
839 
840 	jbd_debug(3, "JBD: commit phase 5\n");
841 
842 	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
843 				       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
844 		err = journal_submit_commit_record(journal, commit_transaction,
845 						&cbh, crc32_sum);
846 		if (err)
847 			__jbd2_journal_abort_hard(journal);
848 	}
849 	if (!err && !is_journal_aborted(journal))
850 		err = journal_wait_on_commit_record(journal, cbh);
851 
852 	if (err)
853 		jbd2_journal_abort(journal, err);
854 
855 	/* End of a transaction!  Finally, we can do checkpoint
856            processing: any buffers committed as a result of this
857            transaction can be removed from any checkpoint list it was on
858            before. */
859 
860 	jbd_debug(3, "JBD: commit phase 6\n");
861 
862 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
863 	J_ASSERT(commit_transaction->t_buffers == NULL);
864 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
865 	J_ASSERT(commit_transaction->t_iobuf_list == NULL);
866 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
867 	J_ASSERT(commit_transaction->t_log_list == NULL);
868 
869 restart_loop:
870 	/*
871 	 * As there are other places (journal_unmap_buffer()) adding buffers
872 	 * to this list we have to be careful and hold the j_list_lock.
873 	 */
874 	spin_lock(&journal->j_list_lock);
875 	while (commit_transaction->t_forget) {
876 		transaction_t *cp_transaction;
877 		struct buffer_head *bh;
878 
879 		jh = commit_transaction->t_forget;
880 		spin_unlock(&journal->j_list_lock);
881 		bh = jh2bh(jh);
882 		jbd_lock_bh_state(bh);
883 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction ||
884 			jh->b_transaction == journal->j_running_transaction);
885 
886 		/*
887 		 * If there is undo-protected committed data against
888 		 * this buffer, then we can remove it now.  If it is a
889 		 * buffer needing such protection, the old frozen_data
890 		 * field now points to a committed version of the
891 		 * buffer, so rotate that field to the new committed
892 		 * data.
893 		 *
894 		 * Otherwise, we can just throw away the frozen data now.
895 		 *
896 		 * We also know that the frozen data has already fired
897 		 * its triggers if they exist, so we can clear that too.
898 		 */
899 		if (jh->b_committed_data) {
900 			jbd2_free(jh->b_committed_data, bh->b_size);
901 			jh->b_committed_data = NULL;
902 			if (jh->b_frozen_data) {
903 				jh->b_committed_data = jh->b_frozen_data;
904 				jh->b_frozen_data = NULL;
905 				jh->b_frozen_triggers = NULL;
906 			}
907 		} else if (jh->b_frozen_data) {
908 			jbd2_free(jh->b_frozen_data, bh->b_size);
909 			jh->b_frozen_data = NULL;
910 			jh->b_frozen_triggers = NULL;
911 		}
912 
913 		spin_lock(&journal->j_list_lock);
914 		cp_transaction = jh->b_cp_transaction;
915 		if (cp_transaction) {
916 			JBUFFER_TRACE(jh, "remove from old cp transaction");
917 			cp_transaction->t_chp_stats.cs_dropped++;
918 			__jbd2_journal_remove_checkpoint(jh);
919 		}
920 
921 		/* Only re-checkpoint the buffer_head if it is marked
922 		 * dirty.  If the buffer was added to the BJ_Forget list
923 		 * by jbd2_journal_forget, it may no longer be dirty and
924 		 * there's no point in keeping a checkpoint record for
925 		 * it. */
926 
927 		/* A buffer which has been freed while still being
928 		 * journaled by a previous transaction may end up still
929 		 * being dirty here, but we want to avoid writing back
930 		 * that buffer in the future now that the last use has
931 		 * been committed.  That's not only a performance gain,
932 		 * it also stops aliasing problems if the buffer is left
933 		 * behind for writeback and gets reallocated for another
934 		 * use in a different page. */
935 		if (buffer_freed(bh)) {
936 			clear_buffer_freed(bh);
937 			clear_buffer_jbddirty(bh);
938 		}
939 
940 		if (buffer_jbddirty(bh)) {
941 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
942 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
943 			if (is_journal_aborted(journal))
944 				clear_buffer_jbddirty(bh);
945 			JBUFFER_TRACE(jh, "refile for checkpoint writeback");
946 			__jbd2_journal_refile_buffer(jh);
947 			jbd_unlock_bh_state(bh);
948 		} else {
949 			J_ASSERT_BH(bh, !buffer_dirty(bh));
950 			/* The buffer on BJ_Forget list and not jbddirty means
951 			 * it has been freed by this transaction and hence it
952 			 * could not have been reallocated until this
953 			 * transaction has committed. *BUT* it could be
954 			 * reallocated once we have written all the data to
955 			 * disk and before we process the buffer on BJ_Forget
956 			 * list. */
957 			JBUFFER_TRACE(jh, "refile or unfile freed buffer");
958 			__jbd2_journal_refile_buffer(jh);
959 			if (!jh->b_transaction) {
960 				jbd_unlock_bh_state(bh);
961 				 /* needs a brelse */
962 				jbd2_journal_remove_journal_head(bh);
963 				release_buffer_page(bh);
964 			} else
965 				jbd_unlock_bh_state(bh);
966 		}
967 		cond_resched_lock(&journal->j_list_lock);
968 	}
969 	spin_unlock(&journal->j_list_lock);
970 	/*
971 	 * This is a bit sleazy.  We use j_list_lock to protect transition
972 	 * of a transaction into T_FINISHED state and calling
973 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
974 	 * other checkpointing code processing the transaction...
975 	 */
976 	spin_lock(&journal->j_state_lock);
977 	spin_lock(&journal->j_list_lock);
978 	/*
979 	 * Now recheck if some buffers did not get attached to the transaction
980 	 * while the lock was dropped...
981 	 */
982 	if (commit_transaction->t_forget) {
983 		spin_unlock(&journal->j_list_lock);
984 		spin_unlock(&journal->j_state_lock);
985 		goto restart_loop;
986 	}
987 
988 	/* Done with this transaction! */
989 
990 	jbd_debug(3, "JBD: commit phase 7\n");
991 
992 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
993 
994 	commit_transaction->t_start = jiffies;
995 	stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
996 					      commit_transaction->t_start);
997 
998 	/*
999 	 * File the transaction statistics
1000 	 */
1001 	stats.ts_tid = commit_transaction->t_tid;
1002 	stats.run.rs_handle_count = commit_transaction->t_handle_count;
1003 	trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1004 			     commit_transaction->t_tid, &stats.run);
1005 
1006 	/*
1007 	 * Calculate overall stats
1008 	 */
1009 	spin_lock(&journal->j_history_lock);
1010 	journal->j_stats.ts_tid++;
1011 	journal->j_stats.run.rs_wait += stats.run.rs_wait;
1012 	journal->j_stats.run.rs_running += stats.run.rs_running;
1013 	journal->j_stats.run.rs_locked += stats.run.rs_locked;
1014 	journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1015 	journal->j_stats.run.rs_logging += stats.run.rs_logging;
1016 	journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1017 	journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1018 	journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1019 	spin_unlock(&journal->j_history_lock);
1020 
1021 	commit_transaction->t_state = T_FINISHED;
1022 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1023 	journal->j_commit_sequence = commit_transaction->t_tid;
1024 	journal->j_committing_transaction = NULL;
1025 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1026 
1027 	/*
1028 	 * weight the commit time higher than the average time so we don't
1029 	 * react too strongly to vast changes in the commit time
1030 	 */
1031 	if (likely(journal->j_average_commit_time))
1032 		journal->j_average_commit_time = (commit_time +
1033 				journal->j_average_commit_time*3) / 4;
1034 	else
1035 		journal->j_average_commit_time = commit_time;
1036 	spin_unlock(&journal->j_state_lock);
1037 
1038 	if (commit_transaction->t_checkpoint_list == NULL &&
1039 	    commit_transaction->t_checkpoint_io_list == NULL) {
1040 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1041 		to_free = 1;
1042 	} else {
1043 		if (journal->j_checkpoint_transactions == NULL) {
1044 			journal->j_checkpoint_transactions = commit_transaction;
1045 			commit_transaction->t_cpnext = commit_transaction;
1046 			commit_transaction->t_cpprev = commit_transaction;
1047 		} else {
1048 			commit_transaction->t_cpnext =
1049 				journal->j_checkpoint_transactions;
1050 			commit_transaction->t_cpprev =
1051 				commit_transaction->t_cpnext->t_cpprev;
1052 			commit_transaction->t_cpnext->t_cpprev =
1053 				commit_transaction;
1054 			commit_transaction->t_cpprev->t_cpnext =
1055 				commit_transaction;
1056 		}
1057 	}
1058 	spin_unlock(&journal->j_list_lock);
1059 
1060 	if (journal->j_commit_callback)
1061 		journal->j_commit_callback(journal, commit_transaction);
1062 
1063 	trace_jbd2_end_commit(journal, commit_transaction);
1064 	jbd_debug(1, "JBD: commit %d complete, head %d\n",
1065 		  journal->j_commit_sequence, journal->j_tail_sequence);
1066 	if (to_free)
1067 		kfree(commit_transaction);
1068 
1069 	wake_up(&journal->j_wait_done_commit);
1070 }
1071