xref: /linux/fs/jbd2/commit.c (revision a0b54e256d513ed99e456bea6e4e188ff92e7c46)
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 <trace/events/jbd2.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 	};
224 
225 	ret = generic_writepages(mapping, &wbc);
226 	return ret;
227 }
228 
229 /*
230  * Submit all the data buffers of inode associated with the transaction to
231  * disk.
232  *
233  * We are in a committing transaction. Therefore no new inode can be added to
234  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
235  * operate on from being released while we write out pages.
236  */
237 static int journal_submit_data_buffers(journal_t *journal,
238 		transaction_t *commit_transaction)
239 {
240 	struct jbd2_inode *jinode;
241 	int err, ret = 0;
242 	struct address_space *mapping;
243 
244 	spin_lock(&journal->j_list_lock);
245 	list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
246 		mapping = jinode->i_vfs_inode->i_mapping;
247 		jinode->i_flags |= JI_COMMIT_RUNNING;
248 		spin_unlock(&journal->j_list_lock);
249 		/*
250 		 * submit the inode data buffers. We use writepage
251 		 * instead of writepages. Because writepages can do
252 		 * block allocation  with delalloc. We need to write
253 		 * only allocated blocks here.
254 		 */
255 		trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
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_jbd2_start_commit(journal, commit_transaction);
398 	jbd_debug(1, "JBD: starting commit of transaction %d\n",
399 			commit_transaction->t_tid);
400 
401 	spin_lock(&journal->j_state_lock);
402 	commit_transaction->t_state = T_LOCKED;
403 
404 	/*
405 	 * Use plugged writes here, since we want to submit several before
406 	 * we unplug the device. We don't do explicit unplugging in here,
407 	 * instead we rely on sync_buffer() doing the unplug for us.
408 	 */
409 	if (commit_transaction->t_synchronous_commit)
410 		write_op = WRITE_SYNC_PLUG;
411 	trace_jbd2_commit_locking(journal, commit_transaction);
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 	trace_jbd2_commit_flushing(journal, commit_transaction);
488 	stats.u.run.rs_flushing = jiffies;
489 	stats.u.run.rs_locked = jbd2_time_diff(stats.u.run.rs_locked,
490 					       stats.u.run.rs_flushing);
491 
492 	commit_transaction->t_state = T_FLUSH;
493 	journal->j_committing_transaction = commit_transaction;
494 	journal->j_running_transaction = NULL;
495 	start_time = ktime_get();
496 	commit_transaction->t_log_start = journal->j_head;
497 	wake_up(&journal->j_wait_transaction_locked);
498 	spin_unlock(&journal->j_state_lock);
499 
500 	jbd_debug (3, "JBD: commit phase 2\n");
501 
502 	/*
503 	 * Now start flushing things to disk, in the order they appear
504 	 * on the transaction lists.  Data blocks go first.
505 	 */
506 	err = journal_submit_data_buffers(journal, commit_transaction);
507 	if (err)
508 		jbd2_journal_abort(journal, err);
509 
510 	jbd2_journal_write_revoke_records(journal, commit_transaction,
511 					  write_op);
512 
513 	jbd_debug(3, "JBD: commit phase 2\n");
514 
515 	/*
516 	 * Way to go: we have now written out all of the data for a
517 	 * transaction!  Now comes the tricky part: we need to write out
518 	 * metadata.  Loop over the transaction's entire buffer list:
519 	 */
520 	spin_lock(&journal->j_state_lock);
521 	commit_transaction->t_state = T_COMMIT;
522 	spin_unlock(&journal->j_state_lock);
523 
524 	trace_jbd2_commit_logging(journal, commit_transaction);
525 	stats.u.run.rs_logging = jiffies;
526 	stats.u.run.rs_flushing = jbd2_time_diff(stats.u.run.rs_flushing,
527 						 stats.u.run.rs_logging);
528 	stats.u.run.rs_blocks = commit_transaction->t_outstanding_credits;
529 	stats.u.run.rs_blocks_logged = 0;
530 
531 	J_ASSERT(commit_transaction->t_nr_buffers <=
532 		 commit_transaction->t_outstanding_credits);
533 
534 	err = 0;
535 	descriptor = NULL;
536 	bufs = 0;
537 	while (commit_transaction->t_buffers) {
538 
539 		/* Find the next buffer to be journaled... */
540 
541 		jh = commit_transaction->t_buffers;
542 
543 		/* If we're in abort mode, we just un-journal the buffer and
544 		   release it. */
545 
546 		if (is_journal_aborted(journal)) {
547 			clear_buffer_jbddirty(jh2bh(jh));
548 			JBUFFER_TRACE(jh, "journal is aborting: refile");
549 			jbd2_buffer_abort_trigger(jh,
550 						  jh->b_frozen_data ?
551 						  jh->b_frozen_triggers :
552 						  jh->b_triggers);
553 			jbd2_journal_refile_buffer(journal, jh);
554 			/* If that was the last one, we need to clean up
555 			 * any descriptor buffers which may have been
556 			 * already allocated, even if we are now
557 			 * aborting. */
558 			if (!commit_transaction->t_buffers)
559 				goto start_journal_io;
560 			continue;
561 		}
562 
563 		/* Make sure we have a descriptor block in which to
564 		   record the metadata buffer. */
565 
566 		if (!descriptor) {
567 			struct buffer_head *bh;
568 
569 			J_ASSERT (bufs == 0);
570 
571 			jbd_debug(4, "JBD: get descriptor\n");
572 
573 			descriptor = jbd2_journal_get_descriptor_buffer(journal);
574 			if (!descriptor) {
575 				jbd2_journal_abort(journal, -EIO);
576 				continue;
577 			}
578 
579 			bh = jh2bh(descriptor);
580 			jbd_debug(4, "JBD: got buffer %llu (%p)\n",
581 				(unsigned long long)bh->b_blocknr, bh->b_data);
582 			header = (journal_header_t *)&bh->b_data[0];
583 			header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
584 			header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
585 			header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
586 
587 			tagp = &bh->b_data[sizeof(journal_header_t)];
588 			space_left = bh->b_size - sizeof(journal_header_t);
589 			first_tag = 1;
590 			set_buffer_jwrite(bh);
591 			set_buffer_dirty(bh);
592 			wbuf[bufs++] = bh;
593 
594 			/* Record it so that we can wait for IO
595                            completion later */
596 			BUFFER_TRACE(bh, "ph3: file as descriptor");
597 			jbd2_journal_file_buffer(descriptor, commit_transaction,
598 					BJ_LogCtl);
599 		}
600 
601 		/* Where is the buffer to be written? */
602 
603 		err = jbd2_journal_next_log_block(journal, &blocknr);
604 		/* If the block mapping failed, just abandon the buffer
605 		   and repeat this loop: we'll fall into the
606 		   refile-on-abort condition above. */
607 		if (err) {
608 			jbd2_journal_abort(journal, err);
609 			continue;
610 		}
611 
612 		/*
613 		 * start_this_handle() uses t_outstanding_credits to determine
614 		 * the free space in the log, but this counter is changed
615 		 * by jbd2_journal_next_log_block() also.
616 		 */
617 		commit_transaction->t_outstanding_credits--;
618 
619 		/* Bump b_count to prevent truncate from stumbling over
620                    the shadowed buffer!  @@@ This can go if we ever get
621                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
622 		atomic_inc(&jh2bh(jh)->b_count);
623 
624 		/* Make a temporary IO buffer with which to write it out
625                    (this will requeue both the metadata buffer and the
626                    temporary IO buffer). new_bh goes on BJ_IO*/
627 
628 		set_bit(BH_JWrite, &jh2bh(jh)->b_state);
629 		/*
630 		 * akpm: jbd2_journal_write_metadata_buffer() sets
631 		 * new_bh->b_transaction to commit_transaction.
632 		 * We need to clean this up before we release new_bh
633 		 * (which is of type BJ_IO)
634 		 */
635 		JBUFFER_TRACE(jh, "ph3: write metadata");
636 		flags = jbd2_journal_write_metadata_buffer(commit_transaction,
637 						      jh, &new_jh, blocknr);
638 		set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
639 		wbuf[bufs++] = jh2bh(new_jh);
640 
641 		/* Record the new block's tag in the current descriptor
642                    buffer */
643 
644 		tag_flag = 0;
645 		if (flags & 1)
646 			tag_flag |= JBD2_FLAG_ESCAPE;
647 		if (!first_tag)
648 			tag_flag |= JBD2_FLAG_SAME_UUID;
649 
650 		tag = (journal_block_tag_t *) tagp;
651 		write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
652 		tag->t_flags = cpu_to_be32(tag_flag);
653 		tagp += tag_bytes;
654 		space_left -= tag_bytes;
655 
656 		if (first_tag) {
657 			memcpy (tagp, journal->j_uuid, 16);
658 			tagp += 16;
659 			space_left -= 16;
660 			first_tag = 0;
661 		}
662 
663 		/* If there's no more to do, or if the descriptor is full,
664 		   let the IO rip! */
665 
666 		if (bufs == journal->j_wbufsize ||
667 		    commit_transaction->t_buffers == NULL ||
668 		    space_left < tag_bytes + 16) {
669 
670 			jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
671 
672 			/* Write an end-of-descriptor marker before
673                            submitting the IOs.  "tag" still points to
674                            the last tag we set up. */
675 
676 			tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
677 
678 start_journal_io:
679 			for (i = 0; i < bufs; i++) {
680 				struct buffer_head *bh = wbuf[i];
681 				/*
682 				 * Compute checksum.
683 				 */
684 				if (JBD2_HAS_COMPAT_FEATURE(journal,
685 					JBD2_FEATURE_COMPAT_CHECKSUM)) {
686 					crc32_sum =
687 					    jbd2_checksum_data(crc32_sum, bh);
688 				}
689 
690 				lock_buffer(bh);
691 				clear_buffer_dirty(bh);
692 				set_buffer_uptodate(bh);
693 				bh->b_end_io = journal_end_buffer_io_sync;
694 				submit_bh(write_op, bh);
695 			}
696 			cond_resched();
697 			stats.u.run.rs_blocks_logged += bufs;
698 
699 			/* Force a new descriptor to be generated next
700                            time round the loop. */
701 			descriptor = NULL;
702 			bufs = 0;
703 		}
704 	}
705 
706 	/* Done it all: now write the commit record asynchronously. */
707 
708 	if (JBD2_HAS_INCOMPAT_FEATURE(journal,
709 		JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
710 		err = journal_submit_commit_record(journal, commit_transaction,
711 						 &cbh, crc32_sum);
712 		if (err)
713 			__jbd2_journal_abort_hard(journal);
714 	}
715 
716 	/*
717 	 * This is the right place to wait for data buffers both for ASYNC
718 	 * and !ASYNC commit. If commit is ASYNC, we need to wait only after
719 	 * the commit block went to disk (which happens above). If commit is
720 	 * SYNC, we need to wait for data buffers before we start writing
721 	 * commit block, which happens below in such setting.
722 	 */
723 	err = journal_finish_inode_data_buffers(journal, commit_transaction);
724 	if (err) {
725 		printk(KERN_WARNING
726 			"JBD2: Detected IO errors while flushing file data "
727 		       "on %s\n", journal->j_devname);
728 		if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
729 			jbd2_journal_abort(journal, err);
730 		err = 0;
731 	}
732 
733 	/* Lo and behold: we have just managed to send a transaction to
734            the log.  Before we can commit it, wait for the IO so far to
735            complete.  Control buffers being written are on the
736            transaction's t_log_list queue, and metadata buffers are on
737            the t_iobuf_list queue.
738 
739 	   Wait for the buffers in reverse order.  That way we are
740 	   less likely to be woken up until all IOs have completed, and
741 	   so we incur less scheduling load.
742 	*/
743 
744 	jbd_debug(3, "JBD: commit phase 3\n");
745 
746 	/*
747 	 * akpm: these are BJ_IO, and j_list_lock is not needed.
748 	 * See __journal_try_to_free_buffer.
749 	 */
750 wait_for_iobuf:
751 	while (commit_transaction->t_iobuf_list != NULL) {
752 		struct buffer_head *bh;
753 
754 		jh = commit_transaction->t_iobuf_list->b_tprev;
755 		bh = jh2bh(jh);
756 		if (buffer_locked(bh)) {
757 			wait_on_buffer(bh);
758 			goto wait_for_iobuf;
759 		}
760 		if (cond_resched())
761 			goto wait_for_iobuf;
762 
763 		if (unlikely(!buffer_uptodate(bh)))
764 			err = -EIO;
765 
766 		clear_buffer_jwrite(bh);
767 
768 		JBUFFER_TRACE(jh, "ph4: unfile after journal write");
769 		jbd2_journal_unfile_buffer(journal, jh);
770 
771 		/*
772 		 * ->t_iobuf_list should contain only dummy buffer_heads
773 		 * which were created by jbd2_journal_write_metadata_buffer().
774 		 */
775 		BUFFER_TRACE(bh, "dumping temporary bh");
776 		jbd2_journal_put_journal_head(jh);
777 		__brelse(bh);
778 		J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
779 		free_buffer_head(bh);
780 
781 		/* We also have to unlock and free the corresponding
782                    shadowed buffer */
783 		jh = commit_transaction->t_shadow_list->b_tprev;
784 		bh = jh2bh(jh);
785 		clear_bit(BH_JWrite, &bh->b_state);
786 		J_ASSERT_BH(bh, buffer_jbddirty(bh));
787 
788 		/* The metadata is now released for reuse, but we need
789                    to remember it against this transaction so that when
790                    we finally commit, we can do any checkpointing
791                    required. */
792 		JBUFFER_TRACE(jh, "file as BJ_Forget");
793 		jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
794 		/* Wake up any transactions which were waiting for this
795 		   IO to complete */
796 		wake_up_bit(&bh->b_state, BH_Unshadow);
797 		JBUFFER_TRACE(jh, "brelse shadowed buffer");
798 		__brelse(bh);
799 	}
800 
801 	J_ASSERT (commit_transaction->t_shadow_list == NULL);
802 
803 	jbd_debug(3, "JBD: commit phase 4\n");
804 
805 	/* Here we wait for the revoke record and descriptor record buffers */
806  wait_for_ctlbuf:
807 	while (commit_transaction->t_log_list != NULL) {
808 		struct buffer_head *bh;
809 
810 		jh = commit_transaction->t_log_list->b_tprev;
811 		bh = jh2bh(jh);
812 		if (buffer_locked(bh)) {
813 			wait_on_buffer(bh);
814 			goto wait_for_ctlbuf;
815 		}
816 		if (cond_resched())
817 			goto wait_for_ctlbuf;
818 
819 		if (unlikely(!buffer_uptodate(bh)))
820 			err = -EIO;
821 
822 		BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
823 		clear_buffer_jwrite(bh);
824 		jbd2_journal_unfile_buffer(journal, jh);
825 		jbd2_journal_put_journal_head(jh);
826 		__brelse(bh);		/* One for getblk */
827 		/* AKPM: bforget here */
828 	}
829 
830 	if (err)
831 		jbd2_journal_abort(journal, err);
832 
833 	jbd_debug(3, "JBD: commit phase 5\n");
834 
835 	if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
836 		JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
837 		err = journal_submit_commit_record(journal, commit_transaction,
838 						&cbh, crc32_sum);
839 		if (err)
840 			__jbd2_journal_abort_hard(journal);
841 	}
842 	if (!err && !is_journal_aborted(journal))
843 		err = journal_wait_on_commit_record(journal, cbh);
844 
845 	if (err)
846 		jbd2_journal_abort(journal, err);
847 
848 	/* End of a transaction!  Finally, we can do checkpoint
849            processing: any buffers committed as a result of this
850            transaction can be removed from any checkpoint list it was on
851            before. */
852 
853 	jbd_debug(3, "JBD: commit phase 6\n");
854 
855 	J_ASSERT(list_empty(&commit_transaction->t_inode_list));
856 	J_ASSERT(commit_transaction->t_buffers == NULL);
857 	J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
858 	J_ASSERT(commit_transaction->t_iobuf_list == NULL);
859 	J_ASSERT(commit_transaction->t_shadow_list == NULL);
860 	J_ASSERT(commit_transaction->t_log_list == NULL);
861 
862 restart_loop:
863 	/*
864 	 * As there are other places (journal_unmap_buffer()) adding buffers
865 	 * to this list we have to be careful and hold the j_list_lock.
866 	 */
867 	spin_lock(&journal->j_list_lock);
868 	while (commit_transaction->t_forget) {
869 		transaction_t *cp_transaction;
870 		struct buffer_head *bh;
871 
872 		jh = commit_transaction->t_forget;
873 		spin_unlock(&journal->j_list_lock);
874 		bh = jh2bh(jh);
875 		jbd_lock_bh_state(bh);
876 		J_ASSERT_JH(jh,	jh->b_transaction == commit_transaction ||
877 			jh->b_transaction == journal->j_running_transaction);
878 
879 		/*
880 		 * If there is undo-protected committed data against
881 		 * this buffer, then we can remove it now.  If it is a
882 		 * buffer needing such protection, the old frozen_data
883 		 * field now points to a committed version of the
884 		 * buffer, so rotate that field to the new committed
885 		 * data.
886 		 *
887 		 * Otherwise, we can just throw away the frozen data now.
888 		 *
889 		 * We also know that the frozen data has already fired
890 		 * its triggers if they exist, so we can clear that too.
891 		 */
892 		if (jh->b_committed_data) {
893 			jbd2_free(jh->b_committed_data, bh->b_size);
894 			jh->b_committed_data = NULL;
895 			if (jh->b_frozen_data) {
896 				jh->b_committed_data = jh->b_frozen_data;
897 				jh->b_frozen_data = NULL;
898 				jh->b_frozen_triggers = NULL;
899 			}
900 		} else if (jh->b_frozen_data) {
901 			jbd2_free(jh->b_frozen_data, bh->b_size);
902 			jh->b_frozen_data = NULL;
903 			jh->b_frozen_triggers = NULL;
904 		}
905 
906 		spin_lock(&journal->j_list_lock);
907 		cp_transaction = jh->b_cp_transaction;
908 		if (cp_transaction) {
909 			JBUFFER_TRACE(jh, "remove from old cp transaction");
910 			cp_transaction->t_chp_stats.cs_dropped++;
911 			__jbd2_journal_remove_checkpoint(jh);
912 		}
913 
914 		/* Only re-checkpoint the buffer_head if it is marked
915 		 * dirty.  If the buffer was added to the BJ_Forget list
916 		 * by jbd2_journal_forget, it may no longer be dirty and
917 		 * there's no point in keeping a checkpoint record for
918 		 * it. */
919 
920 		/* A buffer which has been freed while still being
921 		 * journaled by a previous transaction may end up still
922 		 * being dirty here, but we want to avoid writing back
923 		 * that buffer in the future now that the last use has
924 		 * been committed.  That's not only a performance gain,
925 		 * it also stops aliasing problems if the buffer is left
926 		 * behind for writeback and gets reallocated for another
927 		 * use in a different page. */
928 		if (buffer_freed(bh)) {
929 			clear_buffer_freed(bh);
930 			clear_buffer_jbddirty(bh);
931 		}
932 
933 		if (buffer_jbddirty(bh)) {
934 			JBUFFER_TRACE(jh, "add to new checkpointing trans");
935 			__jbd2_journal_insert_checkpoint(jh, commit_transaction);
936 			if (is_journal_aborted(journal))
937 				clear_buffer_jbddirty(bh);
938 			JBUFFER_TRACE(jh, "refile for checkpoint writeback");
939 			__jbd2_journal_refile_buffer(jh);
940 			jbd_unlock_bh_state(bh);
941 		} else {
942 			J_ASSERT_BH(bh, !buffer_dirty(bh));
943 			/* The buffer on BJ_Forget list and not jbddirty means
944 			 * it has been freed by this transaction and hence it
945 			 * could not have been reallocated until this
946 			 * transaction has committed. *BUT* it could be
947 			 * reallocated once we have written all the data to
948 			 * disk and before we process the buffer on BJ_Forget
949 			 * list. */
950 			JBUFFER_TRACE(jh, "refile or unfile freed buffer");
951 			__jbd2_journal_refile_buffer(jh);
952 			if (!jh->b_transaction) {
953 				jbd_unlock_bh_state(bh);
954 				 /* needs a brelse */
955 				jbd2_journal_remove_journal_head(bh);
956 				release_buffer_page(bh);
957 			} else
958 				jbd_unlock_bh_state(bh);
959 		}
960 		cond_resched_lock(&journal->j_list_lock);
961 	}
962 	spin_unlock(&journal->j_list_lock);
963 	/*
964 	 * This is a bit sleazy.  We use j_list_lock to protect transition
965 	 * of a transaction into T_FINISHED state and calling
966 	 * __jbd2_journal_drop_transaction(). Otherwise we could race with
967 	 * other checkpointing code processing the transaction...
968 	 */
969 	spin_lock(&journal->j_state_lock);
970 	spin_lock(&journal->j_list_lock);
971 	/*
972 	 * Now recheck if some buffers did not get attached to the transaction
973 	 * while the lock was dropped...
974 	 */
975 	if (commit_transaction->t_forget) {
976 		spin_unlock(&journal->j_list_lock);
977 		spin_unlock(&journal->j_state_lock);
978 		goto restart_loop;
979 	}
980 
981 	/* Done with this transaction! */
982 
983 	jbd_debug(3, "JBD: commit phase 7\n");
984 
985 	J_ASSERT(commit_transaction->t_state == T_COMMIT);
986 
987 	commit_transaction->t_start = jiffies;
988 	stats.u.run.rs_logging = jbd2_time_diff(stats.u.run.rs_logging,
989 						commit_transaction->t_start);
990 
991 	/*
992 	 * File the transaction for history
993 	 */
994 	stats.ts_type = JBD2_STATS_RUN;
995 	stats.ts_tid = commit_transaction->t_tid;
996 	stats.u.run.rs_handle_count = commit_transaction->t_handle_count;
997 	spin_lock(&journal->j_history_lock);
998 	memcpy(journal->j_history + journal->j_history_cur, &stats,
999 			sizeof(stats));
1000 	if (++journal->j_history_cur == journal->j_history_max)
1001 		journal->j_history_cur = 0;
1002 
1003 	/*
1004 	 * Calculate overall stats
1005 	 */
1006 	journal->j_stats.ts_tid++;
1007 	journal->j_stats.u.run.rs_wait += stats.u.run.rs_wait;
1008 	journal->j_stats.u.run.rs_running += stats.u.run.rs_running;
1009 	journal->j_stats.u.run.rs_locked += stats.u.run.rs_locked;
1010 	journal->j_stats.u.run.rs_flushing += stats.u.run.rs_flushing;
1011 	journal->j_stats.u.run.rs_logging += stats.u.run.rs_logging;
1012 	journal->j_stats.u.run.rs_handle_count += stats.u.run.rs_handle_count;
1013 	journal->j_stats.u.run.rs_blocks += stats.u.run.rs_blocks;
1014 	journal->j_stats.u.run.rs_blocks_logged += stats.u.run.rs_blocks_logged;
1015 	spin_unlock(&journal->j_history_lock);
1016 
1017 	commit_transaction->t_state = T_FINISHED;
1018 	J_ASSERT(commit_transaction == journal->j_committing_transaction);
1019 	journal->j_commit_sequence = commit_transaction->t_tid;
1020 	journal->j_committing_transaction = NULL;
1021 	commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1022 
1023 	/*
1024 	 * weight the commit time higher than the average time so we don't
1025 	 * react too strongly to vast changes in the commit time
1026 	 */
1027 	if (likely(journal->j_average_commit_time))
1028 		journal->j_average_commit_time = (commit_time +
1029 				journal->j_average_commit_time*3) / 4;
1030 	else
1031 		journal->j_average_commit_time = commit_time;
1032 	spin_unlock(&journal->j_state_lock);
1033 
1034 	if (commit_transaction->t_checkpoint_list == NULL &&
1035 	    commit_transaction->t_checkpoint_io_list == NULL) {
1036 		__jbd2_journal_drop_transaction(journal, commit_transaction);
1037 		to_free = 1;
1038 	} else {
1039 		if (journal->j_checkpoint_transactions == NULL) {
1040 			journal->j_checkpoint_transactions = commit_transaction;
1041 			commit_transaction->t_cpnext = commit_transaction;
1042 			commit_transaction->t_cpprev = commit_transaction;
1043 		} else {
1044 			commit_transaction->t_cpnext =
1045 				journal->j_checkpoint_transactions;
1046 			commit_transaction->t_cpprev =
1047 				commit_transaction->t_cpnext->t_cpprev;
1048 			commit_transaction->t_cpnext->t_cpprev =
1049 				commit_transaction;
1050 			commit_transaction->t_cpprev->t_cpnext =
1051 				commit_transaction;
1052 		}
1053 	}
1054 	spin_unlock(&journal->j_list_lock);
1055 
1056 	if (journal->j_commit_callback)
1057 		journal->j_commit_callback(journal, commit_transaction);
1058 
1059 	trace_jbd2_end_commit(journal, commit_transaction);
1060 	jbd_debug(1, "JBD: commit %d complete, head %d\n",
1061 		  journal->j_commit_sequence, journal->j_tail_sequence);
1062 	if (to_free)
1063 		kfree(commit_transaction);
1064 
1065 	wake_up(&journal->j_wait_done_commit);
1066 }
1067