jbd2: Remove t_handle_lock from start_this_handle()
[linux-2.6.git] / fs / jbd2 / commit.c
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                        "JBD2: Disabling barriers on %s, "
154                        "not supported by device\n", journal->j_devname);
155                 write_lock(&journal->j_state_lock);
156                 journal->j_flags &= ~JBD2_BARRIER;
157                 write_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: %s: disabling barries on %s - not supported "
184                        "by device\n", __func__, journal->j_devname);
185                 write_lock(&journal->j_state_lock);
186                 journal->j_flags &= ~JBD2_BARRIER;
187                 write_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                 commit_transaction->t_flushed_data_blocks = 1;
263                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
264                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
265         }
266         spin_unlock(&journal->j_list_lock);
267         return ret;
268 }
269
270 /*
271  * Wait for data submitted for writeout, refile inodes to proper
272  * transaction if needed.
273  *
274  */
275 static int journal_finish_inode_data_buffers(journal_t *journal,
276                 transaction_t *commit_transaction)
277 {
278         struct jbd2_inode *jinode, *next_i;
279         int err, ret = 0;
280
281         /* For locking, see the comment in journal_submit_data_buffers() */
282         spin_lock(&journal->j_list_lock);
283         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
284                 jinode->i_flags |= JI_COMMIT_RUNNING;
285                 spin_unlock(&journal->j_list_lock);
286                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
287                 if (err) {
288                         /*
289                          * Because AS_EIO is cleared by
290                          * filemap_fdatawait_range(), set it again so
291                          * that user process can get -EIO from fsync().
292                          */
293                         set_bit(AS_EIO,
294                                 &jinode->i_vfs_inode->i_mapping->flags);
295
296                         if (!ret)
297                                 ret = err;
298                 }
299                 spin_lock(&journal->j_list_lock);
300                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
301                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
302         }
303
304         /* Now refile inode to proper lists */
305         list_for_each_entry_safe(jinode, next_i,
306                                  &commit_transaction->t_inode_list, i_list) {
307                 list_del(&jinode->i_list);
308                 if (jinode->i_next_transaction) {
309                         jinode->i_transaction = jinode->i_next_transaction;
310                         jinode->i_next_transaction = NULL;
311                         list_add(&jinode->i_list,
312                                 &jinode->i_transaction->t_inode_list);
313                 } else {
314                         jinode->i_transaction = NULL;
315                 }
316         }
317         spin_unlock(&journal->j_list_lock);
318
319         return ret;
320 }
321
322 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
323 {
324         struct page *page = bh->b_page;
325         char *addr;
326         __u32 checksum;
327
328         addr = kmap_atomic(page, KM_USER0);
329         checksum = crc32_be(crc32_sum,
330                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
331         kunmap_atomic(addr, KM_USER0);
332
333         return checksum;
334 }
335
336 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
337                                    unsigned long long block)
338 {
339         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
340         if (tag_bytes > JBD2_TAG_SIZE32)
341                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
342 }
343
344 /*
345  * jbd2_journal_commit_transaction
346  *
347  * The primary function for committing a transaction to the log.  This
348  * function is called by the journal thread to begin a complete commit.
349  */
350 void jbd2_journal_commit_transaction(journal_t *journal)
351 {
352         struct transaction_stats_s stats;
353         transaction_t *commit_transaction;
354         struct journal_head *jh, *new_jh, *descriptor;
355         struct buffer_head **wbuf = journal->j_wbuf;
356         int bufs;
357         int flags;
358         int err;
359         unsigned long long blocknr;
360         ktime_t start_time;
361         u64 commit_time;
362         char *tagp = NULL;
363         journal_header_t *header;
364         journal_block_tag_t *tag = NULL;
365         int space_left = 0;
366         int first_tag = 0;
367         int tag_flag;
368         int i, to_free = 0;
369         int tag_bytes = journal_tag_bytes(journal);
370         struct buffer_head *cbh = NULL; /* For transactional checksums */
371         __u32 crc32_sum = ~0;
372         int write_op = WRITE;
373
374         /*
375          * First job: lock down the current transaction and wait for
376          * all outstanding updates to complete.
377          */
378
379 #ifdef COMMIT_STATS
380         spin_lock(&journal->j_list_lock);
381         summarise_journal_usage(journal);
382         spin_unlock(&journal->j_list_lock);
383 #endif
384
385         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
386         if (journal->j_flags & JBD2_FLUSHED) {
387                 jbd_debug(3, "super block updated\n");
388                 jbd2_journal_update_superblock(journal, 1);
389         } else {
390                 jbd_debug(3, "superblock not updated\n");
391         }
392
393         J_ASSERT(journal->j_running_transaction != NULL);
394         J_ASSERT(journal->j_committing_transaction == NULL);
395
396         commit_transaction = journal->j_running_transaction;
397         J_ASSERT(commit_transaction->t_state == T_RUNNING);
398
399         trace_jbd2_start_commit(journal, commit_transaction);
400         jbd_debug(1, "JBD: starting commit of transaction %d\n",
401                         commit_transaction->t_tid);
402
403         write_lock(&journal->j_state_lock);
404         commit_transaction->t_state = T_LOCKED;
405
406         /*
407          * Use plugged writes here, since we want to submit several before
408          * we unplug the device. We don't do explicit unplugging in here,
409          * instead we rely on sync_buffer() doing the unplug for us.
410          */
411         if (commit_transaction->t_synchronous_commit)
412                 write_op = WRITE_SYNC_PLUG;
413         trace_jbd2_commit_locking(journal, commit_transaction);
414         stats.run.rs_wait = commit_transaction->t_max_wait;
415         stats.run.rs_locked = jiffies;
416         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
417                                               stats.run.rs_locked);
418
419         spin_lock(&commit_transaction->t_handle_lock);
420         while (atomic_read(&commit_transaction->t_updates)) {
421                 DEFINE_WAIT(wait);
422
423                 prepare_to_wait(&journal->j_wait_updates, &wait,
424                                         TASK_UNINTERRUPTIBLE);
425                 if (atomic_read(&commit_transaction->t_updates)) {
426                         spin_unlock(&commit_transaction->t_handle_lock);
427                         write_unlock(&journal->j_state_lock);
428                         schedule();
429                         write_lock(&journal->j_state_lock);
430                         spin_lock(&commit_transaction->t_handle_lock);
431                 }
432                 finish_wait(&journal->j_wait_updates, &wait);
433         }
434         spin_unlock(&commit_transaction->t_handle_lock);
435
436         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
437                         journal->j_max_transaction_buffers);
438
439         /*
440          * First thing we are allowed to do is to discard any remaining
441          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
442          * that there are no such buffers: if a large filesystem
443          * operation like a truncate needs to split itself over multiple
444          * transactions, then it may try to do a jbd2_journal_restart() while
445          * there are still BJ_Reserved buffers outstanding.  These must
446          * be released cleanly from the current transaction.
447          *
448          * In this case, the filesystem must still reserve write access
449          * again before modifying the buffer in the new transaction, but
450          * we do not require it to remember exactly which old buffers it
451          * has reserved.  This is consistent with the existing behaviour
452          * that multiple jbd2_journal_get_write_access() calls to the same
453          * buffer are perfectly permissable.
454          */
455         while (commit_transaction->t_reserved_list) {
456                 jh = commit_transaction->t_reserved_list;
457                 JBUFFER_TRACE(jh, "reserved, unused: refile");
458                 /*
459                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
460                  * leave undo-committed data.
461                  */
462                 if (jh->b_committed_data) {
463                         struct buffer_head *bh = jh2bh(jh);
464
465                         jbd_lock_bh_state(bh);
466                         jbd2_free(jh->b_committed_data, bh->b_size);
467                         jh->b_committed_data = NULL;
468                         jbd_unlock_bh_state(bh);
469                 }
470                 jbd2_journal_refile_buffer(journal, jh);
471         }
472
473         /*
474          * Now try to drop any written-back buffers from the journal's
475          * checkpoint lists.  We do this *before* commit because it potentially
476          * frees some memory
477          */
478         spin_lock(&journal->j_list_lock);
479         __jbd2_journal_clean_checkpoint_list(journal);
480         spin_unlock(&journal->j_list_lock);
481
482         jbd_debug (3, "JBD: commit phase 1\n");
483
484         /*
485          * Switch to a new revoke table.
486          */
487         jbd2_journal_switch_revoke_table(journal);
488
489         trace_jbd2_commit_flushing(journal, commit_transaction);
490         stats.run.rs_flushing = jiffies;
491         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
492                                              stats.run.rs_flushing);
493
494         commit_transaction->t_state = T_FLUSH;
495         journal->j_committing_transaction = commit_transaction;
496         journal->j_running_transaction = NULL;
497         start_time = ktime_get();
498         commit_transaction->t_log_start = journal->j_head;
499         wake_up(&journal->j_wait_transaction_locked);
500         write_unlock(&journal->j_state_lock);
501
502         jbd_debug (3, "JBD: commit phase 2\n");
503
504         /*
505          * Now start flushing things to disk, in the order they appear
506          * on the transaction lists.  Data blocks go first.
507          */
508         err = journal_submit_data_buffers(journal, commit_transaction);
509         if (err)
510                 jbd2_journal_abort(journal, err);
511
512         jbd2_journal_write_revoke_records(journal, commit_transaction,
513                                           write_op);
514
515         jbd_debug(3, "JBD: commit phase 2\n");
516
517         /*
518          * Way to go: we have now written out all of the data for a
519          * transaction!  Now comes the tricky part: we need to write out
520          * metadata.  Loop over the transaction's entire buffer list:
521          */
522         write_lock(&journal->j_state_lock);
523         commit_transaction->t_state = T_COMMIT;
524         write_unlock(&journal->j_state_lock);
525
526         trace_jbd2_commit_logging(journal, commit_transaction);
527         stats.run.rs_logging = jiffies;
528         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
529                                                stats.run.rs_logging);
530         stats.run.rs_blocks =
531                 atomic_read(&commit_transaction->t_outstanding_credits);
532         stats.run.rs_blocks_logged = 0;
533
534         J_ASSERT(commit_transaction->t_nr_buffers <=
535                  atomic_read(&commit_transaction->t_outstanding_credits));
536
537         err = 0;
538         descriptor = NULL;
539         bufs = 0;
540         while (commit_transaction->t_buffers) {
541
542                 /* Find the next buffer to be journaled... */
543
544                 jh = commit_transaction->t_buffers;
545
546                 /* If we're in abort mode, we just un-journal the buffer and
547                    release it. */
548
549                 if (is_journal_aborted(journal)) {
550                         clear_buffer_jbddirty(jh2bh(jh));
551                         JBUFFER_TRACE(jh, "journal is aborting: refile");
552                         jbd2_buffer_abort_trigger(jh,
553                                                   jh->b_frozen_data ?
554                                                   jh->b_frozen_triggers :
555                                                   jh->b_triggers);
556                         jbd2_journal_refile_buffer(journal, jh);
557                         /* If that was the last one, we need to clean up
558                          * any descriptor buffers which may have been
559                          * already allocated, even if we are now
560                          * aborting. */
561                         if (!commit_transaction->t_buffers)
562                                 goto start_journal_io;
563                         continue;
564                 }
565
566                 /* Make sure we have a descriptor block in which to
567                    record the metadata buffer. */
568
569                 if (!descriptor) {
570                         struct buffer_head *bh;
571
572                         J_ASSERT (bufs == 0);
573
574                         jbd_debug(4, "JBD: get descriptor\n");
575
576                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
577                         if (!descriptor) {
578                                 jbd2_journal_abort(journal, -EIO);
579                                 continue;
580                         }
581
582                         bh = jh2bh(descriptor);
583                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
584                                 (unsigned long long)bh->b_blocknr, bh->b_data);
585                         header = (journal_header_t *)&bh->b_data[0];
586                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
587                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
588                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
589
590                         tagp = &bh->b_data[sizeof(journal_header_t)];
591                         space_left = bh->b_size - sizeof(journal_header_t);
592                         first_tag = 1;
593                         set_buffer_jwrite(bh);
594                         set_buffer_dirty(bh);
595                         wbuf[bufs++] = bh;
596
597                         /* Record it so that we can wait for IO
598                            completion later */
599                         BUFFER_TRACE(bh, "ph3: file as descriptor");
600                         jbd2_journal_file_buffer(descriptor, commit_transaction,
601                                         BJ_LogCtl);
602                 }
603
604                 /* Where is the buffer to be written? */
605
606                 err = jbd2_journal_next_log_block(journal, &blocknr);
607                 /* If the block mapping failed, just abandon the buffer
608                    and repeat this loop: we'll fall into the
609                    refile-on-abort condition above. */
610                 if (err) {
611                         jbd2_journal_abort(journal, err);
612                         continue;
613                 }
614
615                 /*
616                  * start_this_handle() uses t_outstanding_credits to determine
617                  * the free space in the log, but this counter is changed
618                  * by jbd2_journal_next_log_block() also.
619                  */
620                 atomic_dec(&commit_transaction->t_outstanding_credits);
621
622                 /* Bump b_count to prevent truncate from stumbling over
623                    the shadowed buffer!  @@@ This can go if we ever get
624                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
625                 atomic_inc(&jh2bh(jh)->b_count);
626
627                 /* Make a temporary IO buffer with which to write it out
628                    (this will requeue both the metadata buffer and the
629                    temporary IO buffer). new_bh goes on BJ_IO*/
630
631                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
632                 /*
633                  * akpm: jbd2_journal_write_metadata_buffer() sets
634                  * new_bh->b_transaction to commit_transaction.
635                  * We need to clean this up before we release new_bh
636                  * (which is of type BJ_IO)
637                  */
638                 JBUFFER_TRACE(jh, "ph3: write metadata");
639                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
640                                                       jh, &new_jh, blocknr);
641                 if (flags < 0) {
642                         jbd2_journal_abort(journal, flags);
643                         continue;
644                 }
645                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
646                 wbuf[bufs++] = jh2bh(new_jh);
647
648                 /* Record the new block's tag in the current descriptor
649                    buffer */
650
651                 tag_flag = 0;
652                 if (flags & 1)
653                         tag_flag |= JBD2_FLAG_ESCAPE;
654                 if (!first_tag)
655                         tag_flag |= JBD2_FLAG_SAME_UUID;
656
657                 tag = (journal_block_tag_t *) tagp;
658                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
659                 tag->t_flags = cpu_to_be32(tag_flag);
660                 tagp += tag_bytes;
661                 space_left -= tag_bytes;
662
663                 if (first_tag) {
664                         memcpy (tagp, journal->j_uuid, 16);
665                         tagp += 16;
666                         space_left -= 16;
667                         first_tag = 0;
668                 }
669
670                 /* If there's no more to do, or if the descriptor is full,
671                    let the IO rip! */
672
673                 if (bufs == journal->j_wbufsize ||
674                     commit_transaction->t_buffers == NULL ||
675                     space_left < tag_bytes + 16) {
676
677                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
678
679                         /* Write an end-of-descriptor marker before
680                            submitting the IOs.  "tag" still points to
681                            the last tag we set up. */
682
683                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
684
685 start_journal_io:
686                         for (i = 0; i < bufs; i++) {
687                                 struct buffer_head *bh = wbuf[i];
688                                 /*
689                                  * Compute checksum.
690                                  */
691                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
692                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
693                                         crc32_sum =
694                                             jbd2_checksum_data(crc32_sum, bh);
695                                 }
696
697                                 lock_buffer(bh);
698                                 clear_buffer_dirty(bh);
699                                 set_buffer_uptodate(bh);
700                                 bh->b_end_io = journal_end_buffer_io_sync;
701                                 submit_bh(write_op, bh);
702                         }
703                         cond_resched();
704                         stats.run.rs_blocks_logged += bufs;
705
706                         /* Force a new descriptor to be generated next
707                            time round the loop. */
708                         descriptor = NULL;
709                         bufs = 0;
710                 }
711         }
712
713         /* 
714          * If the journal is not located on the file system device,
715          * then we must flush the file system device before we issue
716          * the commit record
717          */
718         if (commit_transaction->t_flushed_data_blocks &&
719             (journal->j_fs_dev != journal->j_dev) &&
720             (journal->j_flags & JBD2_BARRIER))
721                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL,
722                         BLKDEV_IFL_WAIT);
723
724         /* Done it all: now write the commit record asynchronously. */
725         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
726                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
727                 err = journal_submit_commit_record(journal, commit_transaction,
728                                                  &cbh, crc32_sum);
729                 if (err)
730                         __jbd2_journal_abort_hard(journal);
731                 if (journal->j_flags & JBD2_BARRIER)
732                         blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL,
733                                 BLKDEV_IFL_WAIT);
734         }
735
736         err = journal_finish_inode_data_buffers(journal, commit_transaction);
737         if (err) {
738                 printk(KERN_WARNING
739                         "JBD2: Detected IO errors while flushing file data "
740                        "on %s\n", journal->j_devname);
741                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
742                         jbd2_journal_abort(journal, err);
743                 err = 0;
744         }
745
746         /* Lo and behold: we have just managed to send a transaction to
747            the log.  Before we can commit it, wait for the IO so far to
748            complete.  Control buffers being written are on the
749            transaction's t_log_list queue, and metadata buffers are on
750            the t_iobuf_list queue.
751
752            Wait for the buffers in reverse order.  That way we are
753            less likely to be woken up until all IOs have completed, and
754            so we incur less scheduling load.
755         */
756
757         jbd_debug(3, "JBD: commit phase 3\n");
758
759         /*
760          * akpm: these are BJ_IO, and j_list_lock is not needed.
761          * See __journal_try_to_free_buffer.
762          */
763 wait_for_iobuf:
764         while (commit_transaction->t_iobuf_list != NULL) {
765                 struct buffer_head *bh;
766
767                 jh = commit_transaction->t_iobuf_list->b_tprev;
768                 bh = jh2bh(jh);
769                 if (buffer_locked(bh)) {
770                         wait_on_buffer(bh);
771                         goto wait_for_iobuf;
772                 }
773                 if (cond_resched())
774                         goto wait_for_iobuf;
775
776                 if (unlikely(!buffer_uptodate(bh)))
777                         err = -EIO;
778
779                 clear_buffer_jwrite(bh);
780
781                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
782                 jbd2_journal_unfile_buffer(journal, jh);
783
784                 /*
785                  * ->t_iobuf_list should contain only dummy buffer_heads
786                  * which were created by jbd2_journal_write_metadata_buffer().
787                  */
788                 BUFFER_TRACE(bh, "dumping temporary bh");
789                 jbd2_journal_put_journal_head(jh);
790                 __brelse(bh);
791                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
792                 free_buffer_head(bh);
793
794                 /* We also have to unlock and free the corresponding
795                    shadowed buffer */
796                 jh = commit_transaction->t_shadow_list->b_tprev;
797                 bh = jh2bh(jh);
798                 clear_bit(BH_JWrite, &bh->b_state);
799                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
800
801                 /* The metadata is now released for reuse, but we need
802                    to remember it against this transaction so that when
803                    we finally commit, we can do any checkpointing
804                    required. */
805                 JBUFFER_TRACE(jh, "file as BJ_Forget");
806                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
807                 /* Wake up any transactions which were waiting for this
808                    IO to complete */
809                 wake_up_bit(&bh->b_state, BH_Unshadow);
810                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
811                 __brelse(bh);
812         }
813
814         J_ASSERT (commit_transaction->t_shadow_list == NULL);
815
816         jbd_debug(3, "JBD: commit phase 4\n");
817
818         /* Here we wait for the revoke record and descriptor record buffers */
819  wait_for_ctlbuf:
820         while (commit_transaction->t_log_list != NULL) {
821                 struct buffer_head *bh;
822
823                 jh = commit_transaction->t_log_list->b_tprev;
824                 bh = jh2bh(jh);
825                 if (buffer_locked(bh)) {
826                         wait_on_buffer(bh);
827                         goto wait_for_ctlbuf;
828                 }
829                 if (cond_resched())
830                         goto wait_for_ctlbuf;
831
832                 if (unlikely(!buffer_uptodate(bh)))
833                         err = -EIO;
834
835                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
836                 clear_buffer_jwrite(bh);
837                 jbd2_journal_unfile_buffer(journal, jh);
838                 jbd2_journal_put_journal_head(jh);
839                 __brelse(bh);           /* One for getblk */
840                 /* AKPM: bforget here */
841         }
842
843         if (err)
844                 jbd2_journal_abort(journal, err);
845
846         jbd_debug(3, "JBD: commit phase 5\n");
847
848         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
849                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
850                 err = journal_submit_commit_record(journal, commit_transaction,
851                                                 &cbh, crc32_sum);
852                 if (err)
853                         __jbd2_journal_abort_hard(journal);
854         }
855         if (!err && !is_journal_aborted(journal))
856                 err = journal_wait_on_commit_record(journal, cbh);
857
858         if (err)
859                 jbd2_journal_abort(journal, err);
860
861         /* End of a transaction!  Finally, we can do checkpoint
862            processing: any buffers committed as a result of this
863            transaction can be removed from any checkpoint list it was on
864            before. */
865
866         jbd_debug(3, "JBD: commit phase 6\n");
867
868         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
869         J_ASSERT(commit_transaction->t_buffers == NULL);
870         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
871         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
872         J_ASSERT(commit_transaction->t_shadow_list == NULL);
873         J_ASSERT(commit_transaction->t_log_list == NULL);
874
875 restart_loop:
876         /*
877          * As there are other places (journal_unmap_buffer()) adding buffers
878          * to this list we have to be careful and hold the j_list_lock.
879          */
880         spin_lock(&journal->j_list_lock);
881         while (commit_transaction->t_forget) {
882                 transaction_t *cp_transaction;
883                 struct buffer_head *bh;
884
885                 jh = commit_transaction->t_forget;
886                 spin_unlock(&journal->j_list_lock);
887                 bh = jh2bh(jh);
888                 jbd_lock_bh_state(bh);
889                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
890
891                 /*
892                  * If there is undo-protected committed data against
893                  * this buffer, then we can remove it now.  If it is a
894                  * buffer needing such protection, the old frozen_data
895                  * field now points to a committed version of the
896                  * buffer, so rotate that field to the new committed
897                  * data.
898                  *
899                  * Otherwise, we can just throw away the frozen data now.
900                  *
901                  * We also know that the frozen data has already fired
902                  * its triggers if they exist, so we can clear that too.
903                  */
904                 if (jh->b_committed_data) {
905                         jbd2_free(jh->b_committed_data, bh->b_size);
906                         jh->b_committed_data = NULL;
907                         if (jh->b_frozen_data) {
908                                 jh->b_committed_data = jh->b_frozen_data;
909                                 jh->b_frozen_data = NULL;
910                                 jh->b_frozen_triggers = NULL;
911                         }
912                 } else if (jh->b_frozen_data) {
913                         jbd2_free(jh->b_frozen_data, bh->b_size);
914                         jh->b_frozen_data = NULL;
915                         jh->b_frozen_triggers = NULL;
916                 }
917
918                 spin_lock(&journal->j_list_lock);
919                 cp_transaction = jh->b_cp_transaction;
920                 if (cp_transaction) {
921                         JBUFFER_TRACE(jh, "remove from old cp transaction");
922                         cp_transaction->t_chp_stats.cs_dropped++;
923                         __jbd2_journal_remove_checkpoint(jh);
924                 }
925
926                 /* Only re-checkpoint the buffer_head if it is marked
927                  * dirty.  If the buffer was added to the BJ_Forget list
928                  * by jbd2_journal_forget, it may no longer be dirty and
929                  * there's no point in keeping a checkpoint record for
930                  * it. */
931
932                 /* A buffer which has been freed while still being
933                  * journaled by a previous transaction may end up still
934                  * being dirty here, but we want to avoid writing back
935                  * that buffer in the future after the "add to orphan"
936                  * operation been committed,  That's not only a performance
937                  * gain, it also stops aliasing problems if the buffer is
938                  * left behind for writeback and gets reallocated for another
939                  * use in a different page. */
940                 if (buffer_freed(bh) && !jh->b_next_transaction) {
941                         clear_buffer_freed(bh);
942                         clear_buffer_jbddirty(bh);
943                 }
944
945                 if (buffer_jbddirty(bh)) {
946                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
947                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
948                         if (is_journal_aborted(journal))
949                                 clear_buffer_jbddirty(bh);
950                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
951                         __jbd2_journal_refile_buffer(jh);
952                         jbd_unlock_bh_state(bh);
953                 } else {
954                         J_ASSERT_BH(bh, !buffer_dirty(bh));
955                         /* The buffer on BJ_Forget list and not jbddirty means
956                          * it has been freed by this transaction and hence it
957                          * could not have been reallocated until this
958                          * transaction has committed. *BUT* it could be
959                          * reallocated once we have written all the data to
960                          * disk and before we process the buffer on BJ_Forget
961                          * list. */
962                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
963                         __jbd2_journal_refile_buffer(jh);
964                         if (!jh->b_transaction) {
965                                 jbd_unlock_bh_state(bh);
966                                  /* needs a brelse */
967                                 jbd2_journal_remove_journal_head(bh);
968                                 release_buffer_page(bh);
969                         } else
970                                 jbd_unlock_bh_state(bh);
971                 }
972                 cond_resched_lock(&journal->j_list_lock);
973         }
974         spin_unlock(&journal->j_list_lock);
975         /*
976          * This is a bit sleazy.  We use j_list_lock to protect transition
977          * of a transaction into T_FINISHED state and calling
978          * __jbd2_journal_drop_transaction(). Otherwise we could race with
979          * other checkpointing code processing the transaction...
980          */
981         write_lock(&journal->j_state_lock);
982         spin_lock(&journal->j_list_lock);
983         /*
984          * Now recheck if some buffers did not get attached to the transaction
985          * while the lock was dropped...
986          */
987         if (commit_transaction->t_forget) {
988                 spin_unlock(&journal->j_list_lock);
989                 write_unlock(&journal->j_state_lock);
990                 goto restart_loop;
991         }
992
993         /* Done with this transaction! */
994
995         jbd_debug(3, "JBD: commit phase 7\n");
996
997         J_ASSERT(commit_transaction->t_state == T_COMMIT);
998
999         commit_transaction->t_start = jiffies;
1000         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1001                                               commit_transaction->t_start);
1002
1003         /*
1004          * File the transaction statistics
1005          */
1006         stats.ts_tid = commit_transaction->t_tid;
1007         stats.run.rs_handle_count =
1008                 atomic_read(&commit_transaction->t_handle_count);
1009         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1010                              commit_transaction->t_tid, &stats.run);
1011
1012         /*
1013          * Calculate overall stats
1014          */
1015         spin_lock(&journal->j_history_lock);
1016         journal->j_stats.ts_tid++;
1017         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1018         journal->j_stats.run.rs_running += stats.run.rs_running;
1019         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1020         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1021         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1022         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1023         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1024         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1025         spin_unlock(&journal->j_history_lock);
1026
1027         commit_transaction->t_state = T_FINISHED;
1028         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1029         journal->j_commit_sequence = commit_transaction->t_tid;
1030         journal->j_committing_transaction = NULL;
1031         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1032
1033         /*
1034          * weight the commit time higher than the average time so we don't
1035          * react too strongly to vast changes in the commit time
1036          */
1037         if (likely(journal->j_average_commit_time))
1038                 journal->j_average_commit_time = (commit_time +
1039                                 journal->j_average_commit_time*3) / 4;
1040         else
1041                 journal->j_average_commit_time = commit_time;
1042         write_unlock(&journal->j_state_lock);
1043
1044         if (commit_transaction->t_checkpoint_list == NULL &&
1045             commit_transaction->t_checkpoint_io_list == NULL) {
1046                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1047                 to_free = 1;
1048         } else {
1049                 if (journal->j_checkpoint_transactions == NULL) {
1050                         journal->j_checkpoint_transactions = commit_transaction;
1051                         commit_transaction->t_cpnext = commit_transaction;
1052                         commit_transaction->t_cpprev = commit_transaction;
1053                 } else {
1054                         commit_transaction->t_cpnext =
1055                                 journal->j_checkpoint_transactions;
1056                         commit_transaction->t_cpprev =
1057                                 commit_transaction->t_cpnext->t_cpprev;
1058                         commit_transaction->t_cpnext->t_cpprev =
1059                                 commit_transaction;
1060                         commit_transaction->t_cpprev->t_cpnext =
1061                                 commit_transaction;
1062                 }
1063         }
1064         spin_unlock(&journal->j_list_lock);
1065
1066         if (journal->j_commit_callback)
1067                 journal->j_commit_callback(journal, commit_transaction);
1068
1069         trace_jbd2_end_commit(journal, commit_transaction);
1070         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1071                   journal->j_commit_sequence, journal->j_tail_sequence);
1072         if (to_free)
1073                 kfree(commit_transaction);
1074
1075         wake_up(&journal->j_wait_done_commit);
1076 }