cfq: improve fsync performance for small files
[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         struct timespec now = current_kernel_time();
105
106         if (is_journal_aborted(journal))
107                 return 0;
108
109         descriptor = jbd2_journal_get_descriptor_buffer(journal);
110         if (!descriptor)
111                 return 1;
112
113         bh = jh2bh(descriptor);
114
115         tmp = (struct commit_header *)bh->b_data;
116         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
117         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
118         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
119         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
120         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
121
122         if (JBD2_HAS_COMPAT_FEATURE(journal,
123                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
124                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
125                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
126                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
127         }
128
129         JBUFFER_TRACE(descriptor, "submit commit block");
130         lock_buffer(bh);
131         clear_buffer_dirty(bh);
132         set_buffer_uptodate(bh);
133         bh->b_end_io = journal_end_buffer_io_sync;
134
135         if (journal->j_flags & JBD2_BARRIER &&
136             !JBD2_HAS_INCOMPAT_FEATURE(journal,
137                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
138                 ret = submit_bh(WRITE_SYNC_PLUG | WRITE_BARRIER, bh);
139                 if (ret == -EOPNOTSUPP) {
140                         printk(KERN_WARNING
141                                "JBD2: Disabling barriers on %s, "
142                                "not supported by device\n", journal->j_devname);
143                         write_lock(&journal->j_state_lock);
144                         journal->j_flags &= ~JBD2_BARRIER;
145                         write_unlock(&journal->j_state_lock);
146
147                         /* And try again, without the barrier */
148                         lock_buffer(bh);
149                         set_buffer_uptodate(bh);
150                         clear_buffer_dirty(bh);
151                         ret = submit_bh(WRITE_SYNC_PLUG, bh);
152                 }
153         } else {
154                 ret = submit_bh(WRITE_SYNC_PLUG, bh);
155         }
156         *cbh = bh;
157         return ret;
158 }
159
160 /*
161  * This function along with journal_submit_commit_record
162  * allows to write the commit record asynchronously.
163  */
164 static int journal_wait_on_commit_record(journal_t *journal,
165                                          struct buffer_head *bh)
166 {
167         int ret = 0;
168
169 retry:
170         clear_buffer_dirty(bh);
171         wait_on_buffer(bh);
172         if (buffer_eopnotsupp(bh) && (journal->j_flags & JBD2_BARRIER)) {
173                 printk(KERN_WARNING
174                        "JBD2: %s: disabling barries on %s - not supported "
175                        "by device\n", __func__, journal->j_devname);
176                 write_lock(&journal->j_state_lock);
177                 journal->j_flags &= ~JBD2_BARRIER;
178                 write_unlock(&journal->j_state_lock);
179
180                 lock_buffer(bh);
181                 clear_buffer_dirty(bh);
182                 set_buffer_uptodate(bh);
183                 bh->b_end_io = journal_end_buffer_io_sync;
184
185                 ret = submit_bh(WRITE_SYNC_PLUG, bh);
186                 if (ret) {
187                         unlock_buffer(bh);
188                         return ret;
189                 }
190                 goto retry;
191         }
192
193         if (unlikely(!buffer_uptodate(bh)))
194                 ret = -EIO;
195         put_bh(bh);            /* One for getblk() */
196         jbd2_journal_put_journal_head(bh2jh(bh));
197
198         return ret;
199 }
200
201 /*
202  * write the filemap data using writepage() address_space_operations.
203  * We don't do block allocation here even for delalloc. We don't
204  * use writepages() because with dealyed allocation we may be doing
205  * block allocation in writepages().
206  */
207 static int journal_submit_inode_data_buffers(struct address_space *mapping)
208 {
209         int ret;
210         struct writeback_control wbc = {
211                 .sync_mode =  WB_SYNC_ALL,
212                 .nr_to_write = mapping->nrpages * 2,
213                 .range_start = 0,
214                 .range_end = i_size_read(mapping->host),
215         };
216
217         ret = generic_writepages(mapping, &wbc);
218         return ret;
219 }
220
221 /*
222  * Submit all the data buffers of inode associated with the transaction to
223  * disk.
224  *
225  * We are in a committing transaction. Therefore no new inode can be added to
226  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
227  * operate on from being released while we write out pages.
228  */
229 static int journal_submit_data_buffers(journal_t *journal,
230                 transaction_t *commit_transaction)
231 {
232         struct jbd2_inode *jinode;
233         int err, ret = 0;
234         struct address_space *mapping;
235
236         spin_lock(&journal->j_list_lock);
237         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
238                 mapping = jinode->i_vfs_inode->i_mapping;
239                 jinode->i_flags |= JI_COMMIT_RUNNING;
240                 spin_unlock(&journal->j_list_lock);
241                 /*
242                  * submit the inode data buffers. We use writepage
243                  * instead of writepages. Because writepages can do
244                  * block allocation  with delalloc. We need to write
245                  * only allocated blocks here.
246                  */
247                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
248                 err = journal_submit_inode_data_buffers(mapping);
249                 if (!ret)
250                         ret = err;
251                 spin_lock(&journal->j_list_lock);
252                 J_ASSERT(jinode->i_transaction == commit_transaction);
253                 commit_transaction->t_flushed_data_blocks = 1;
254                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
255                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
256         }
257         spin_unlock(&journal->j_list_lock);
258         return ret;
259 }
260
261 /*
262  * Wait for data submitted for writeout, refile inodes to proper
263  * transaction if needed.
264  *
265  */
266 static int journal_finish_inode_data_buffers(journal_t *journal,
267                 transaction_t *commit_transaction)
268 {
269         struct jbd2_inode *jinode, *next_i;
270         int err, ret = 0;
271
272         /* For locking, see the comment in journal_submit_data_buffers() */
273         spin_lock(&journal->j_list_lock);
274         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
275                 jinode->i_flags |= JI_COMMIT_RUNNING;
276                 spin_unlock(&journal->j_list_lock);
277                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
278                 if (err) {
279                         /*
280                          * Because AS_EIO is cleared by
281                          * filemap_fdatawait_range(), set it again so
282                          * that user process can get -EIO from fsync().
283                          */
284                         set_bit(AS_EIO,
285                                 &jinode->i_vfs_inode->i_mapping->flags);
286
287                         if (!ret)
288                                 ret = err;
289                 }
290                 spin_lock(&journal->j_list_lock);
291                 jinode->i_flags &= ~JI_COMMIT_RUNNING;
292                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
293         }
294
295         /* Now refile inode to proper lists */
296         list_for_each_entry_safe(jinode, next_i,
297                                  &commit_transaction->t_inode_list, i_list) {
298                 list_del(&jinode->i_list);
299                 if (jinode->i_next_transaction) {
300                         jinode->i_transaction = jinode->i_next_transaction;
301                         jinode->i_next_transaction = NULL;
302                         list_add(&jinode->i_list,
303                                 &jinode->i_transaction->t_inode_list);
304                 } else {
305                         jinode->i_transaction = NULL;
306                 }
307         }
308         spin_unlock(&journal->j_list_lock);
309
310         return ret;
311 }
312
313 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
314 {
315         struct page *page = bh->b_page;
316         char *addr;
317         __u32 checksum;
318
319         addr = kmap_atomic(page, KM_USER0);
320         checksum = crc32_be(crc32_sum,
321                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
322         kunmap_atomic(addr, KM_USER0);
323
324         return checksum;
325 }
326
327 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
328                                    unsigned long long block)
329 {
330         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
331         if (tag_bytes > JBD2_TAG_SIZE32)
332                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
333 }
334
335 /*
336  * jbd2_journal_commit_transaction
337  *
338  * The primary function for committing a transaction to the log.  This
339  * function is called by the journal thread to begin a complete commit.
340  */
341 void jbd2_journal_commit_transaction(journal_t *journal)
342 {
343         struct transaction_stats_s stats;
344         transaction_t *commit_transaction;
345         struct journal_head *jh, *new_jh, *descriptor;
346         struct buffer_head **wbuf = journal->j_wbuf;
347         int bufs;
348         int flags;
349         int err;
350         unsigned long long blocknr;
351         ktime_t start_time;
352         u64 commit_time;
353         char *tagp = NULL;
354         journal_header_t *header;
355         journal_block_tag_t *tag = NULL;
356         int space_left = 0;
357         int first_tag = 0;
358         int tag_flag;
359         int i, to_free = 0;
360         int tag_bytes = journal_tag_bytes(journal);
361         struct buffer_head *cbh = NULL; /* For transactional checksums */
362         __u32 crc32_sum = ~0;
363         int write_op = WRITE_SYNC;
364
365         /*
366          * First job: lock down the current transaction and wait for
367          * all outstanding updates to complete.
368          */
369
370 #ifdef COMMIT_STATS
371         spin_lock(&journal->j_list_lock);
372         summarise_journal_usage(journal);
373         spin_unlock(&journal->j_list_lock);
374 #endif
375
376         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
377         if (journal->j_flags & JBD2_FLUSHED) {
378                 jbd_debug(3, "super block updated\n");
379                 jbd2_journal_update_superblock(journal, 1);
380         } else {
381                 jbd_debug(3, "superblock not updated\n");
382         }
383
384         J_ASSERT(journal->j_running_transaction != NULL);
385         J_ASSERT(journal->j_committing_transaction == NULL);
386
387         commit_transaction = journal->j_running_transaction;
388         J_ASSERT(commit_transaction->t_state == T_RUNNING);
389
390         trace_jbd2_start_commit(journal, commit_transaction);
391         jbd_debug(1, "JBD: starting commit of transaction %d\n",
392                         commit_transaction->t_tid);
393
394         write_lock(&journal->j_state_lock);
395         commit_transaction->t_state = T_LOCKED;
396
397         /*
398          * Use plugged writes here, since we want to submit several before
399          * we unplug the device. We don't do explicit unplugging in here,
400          * instead we rely on sync_buffer() doing the unplug for us.
401          */
402         if (commit_transaction->t_synchronous_commit)
403                 write_op = WRITE_SYNC_PLUG;
404         trace_jbd2_commit_locking(journal, commit_transaction);
405         stats.run.rs_wait = commit_transaction->t_max_wait;
406         stats.run.rs_locked = jiffies;
407         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
408                                               stats.run.rs_locked);
409
410         spin_lock(&commit_transaction->t_handle_lock);
411         while (atomic_read(&commit_transaction->t_updates)) {
412                 DEFINE_WAIT(wait);
413
414                 prepare_to_wait(&journal->j_wait_updates, &wait,
415                                         TASK_UNINTERRUPTIBLE);
416                 if (atomic_read(&commit_transaction->t_updates)) {
417                         spin_unlock(&commit_transaction->t_handle_lock);
418                         write_unlock(&journal->j_state_lock);
419                         schedule();
420                         write_lock(&journal->j_state_lock);
421                         spin_lock(&commit_transaction->t_handle_lock);
422                 }
423                 finish_wait(&journal->j_wait_updates, &wait);
424         }
425         spin_unlock(&commit_transaction->t_handle_lock);
426
427         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
428                         journal->j_max_transaction_buffers);
429
430         /*
431          * First thing we are allowed to do is to discard any remaining
432          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
433          * that there are no such buffers: if a large filesystem
434          * operation like a truncate needs to split itself over multiple
435          * transactions, then it may try to do a jbd2_journal_restart() while
436          * there are still BJ_Reserved buffers outstanding.  These must
437          * be released cleanly from the current transaction.
438          *
439          * In this case, the filesystem must still reserve write access
440          * again before modifying the buffer in the new transaction, but
441          * we do not require it to remember exactly which old buffers it
442          * has reserved.  This is consistent with the existing behaviour
443          * that multiple jbd2_journal_get_write_access() calls to the same
444          * buffer are perfectly permissable.
445          */
446         while (commit_transaction->t_reserved_list) {
447                 jh = commit_transaction->t_reserved_list;
448                 JBUFFER_TRACE(jh, "reserved, unused: refile");
449                 /*
450                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
451                  * leave undo-committed data.
452                  */
453                 if (jh->b_committed_data) {
454                         struct buffer_head *bh = jh2bh(jh);
455
456                         jbd_lock_bh_state(bh);
457                         jbd2_free(jh->b_committed_data, bh->b_size);
458                         jh->b_committed_data = NULL;
459                         jbd_unlock_bh_state(bh);
460                 }
461                 jbd2_journal_refile_buffer(journal, jh);
462         }
463
464         /*
465          * Now try to drop any written-back buffers from the journal's
466          * checkpoint lists.  We do this *before* commit because it potentially
467          * frees some memory
468          */
469         spin_lock(&journal->j_list_lock);
470         __jbd2_journal_clean_checkpoint_list(journal);
471         spin_unlock(&journal->j_list_lock);
472
473         jbd_debug (3, "JBD: commit phase 1\n");
474
475         /*
476          * Switch to a new revoke table.
477          */
478         jbd2_journal_switch_revoke_table(journal);
479
480         trace_jbd2_commit_flushing(journal, commit_transaction);
481         stats.run.rs_flushing = jiffies;
482         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
483                                              stats.run.rs_flushing);
484
485         commit_transaction->t_state = T_FLUSH;
486         journal->j_committing_transaction = commit_transaction;
487         journal->j_running_transaction = NULL;
488         start_time = ktime_get();
489         commit_transaction->t_log_start = journal->j_head;
490         wake_up(&journal->j_wait_transaction_locked);
491         write_unlock(&journal->j_state_lock);
492
493         jbd_debug (3, "JBD: commit phase 2\n");
494
495         /*
496          * Now start flushing things to disk, in the order they appear
497          * on the transaction lists.  Data blocks go first.
498          */
499         err = journal_submit_data_buffers(journal, commit_transaction);
500         if (err)
501                 jbd2_journal_abort(journal, err);
502
503         jbd2_journal_write_revoke_records(journal, commit_transaction,
504                                           write_op);
505
506         jbd_debug(3, "JBD: commit phase 2\n");
507
508         /*
509          * Way to go: we have now written out all of the data for a
510          * transaction!  Now comes the tricky part: we need to write out
511          * metadata.  Loop over the transaction's entire buffer list:
512          */
513         write_lock(&journal->j_state_lock);
514         commit_transaction->t_state = T_COMMIT;
515         write_unlock(&journal->j_state_lock);
516
517         trace_jbd2_commit_logging(journal, commit_transaction);
518         stats.run.rs_logging = jiffies;
519         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
520                                                stats.run.rs_logging);
521         stats.run.rs_blocks =
522                 atomic_read(&commit_transaction->t_outstanding_credits);
523         stats.run.rs_blocks_logged = 0;
524
525         J_ASSERT(commit_transaction->t_nr_buffers <=
526                  atomic_read(&commit_transaction->t_outstanding_credits));
527
528         err = 0;
529         descriptor = NULL;
530         bufs = 0;
531         while (commit_transaction->t_buffers) {
532
533                 /* Find the next buffer to be journaled... */
534
535                 jh = commit_transaction->t_buffers;
536
537                 /* If we're in abort mode, we just un-journal the buffer and
538                    release it. */
539
540                 if (is_journal_aborted(journal)) {
541                         clear_buffer_jbddirty(jh2bh(jh));
542                         JBUFFER_TRACE(jh, "journal is aborting: refile");
543                         jbd2_buffer_abort_trigger(jh,
544                                                   jh->b_frozen_data ?
545                                                   jh->b_frozen_triggers :
546                                                   jh->b_triggers);
547                         jbd2_journal_refile_buffer(journal, jh);
548                         /* If that was the last one, we need to clean up
549                          * any descriptor buffers which may have been
550                          * already allocated, even if we are now
551                          * aborting. */
552                         if (!commit_transaction->t_buffers)
553                                 goto start_journal_io;
554                         continue;
555                 }
556
557                 /* Make sure we have a descriptor block in which to
558                    record the metadata buffer. */
559
560                 if (!descriptor) {
561                         struct buffer_head *bh;
562
563                         J_ASSERT (bufs == 0);
564
565                         jbd_debug(4, "JBD: get descriptor\n");
566
567                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
568                         if (!descriptor) {
569                                 jbd2_journal_abort(journal, -EIO);
570                                 continue;
571                         }
572
573                         bh = jh2bh(descriptor);
574                         jbd_debug(4, "JBD: got buffer %llu (%p)\n",
575                                 (unsigned long long)bh->b_blocknr, bh->b_data);
576                         header = (journal_header_t *)&bh->b_data[0];
577                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
578                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
579                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
580
581                         tagp = &bh->b_data[sizeof(journal_header_t)];
582                         space_left = bh->b_size - sizeof(journal_header_t);
583                         first_tag = 1;
584                         set_buffer_jwrite(bh);
585                         set_buffer_dirty(bh);
586                         wbuf[bufs++] = bh;
587
588                         /* Record it so that we can wait for IO
589                            completion later */
590                         BUFFER_TRACE(bh, "ph3: file as descriptor");
591                         jbd2_journal_file_buffer(descriptor, commit_transaction,
592                                         BJ_LogCtl);
593                 }
594
595                 /* Where is the buffer to be written? */
596
597                 err = jbd2_journal_next_log_block(journal, &blocknr);
598                 /* If the block mapping failed, just abandon the buffer
599                    and repeat this loop: we'll fall into the
600                    refile-on-abort condition above. */
601                 if (err) {
602                         jbd2_journal_abort(journal, err);
603                         continue;
604                 }
605
606                 /*
607                  * start_this_handle() uses t_outstanding_credits to determine
608                  * the free space in the log, but this counter is changed
609                  * by jbd2_journal_next_log_block() also.
610                  */
611                 atomic_dec(&commit_transaction->t_outstanding_credits);
612
613                 /* Bump b_count to prevent truncate from stumbling over
614                    the shadowed buffer!  @@@ This can go if we ever get
615                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
616                 atomic_inc(&jh2bh(jh)->b_count);
617
618                 /* Make a temporary IO buffer with which to write it out
619                    (this will requeue both the metadata buffer and the
620                    temporary IO buffer). new_bh goes on BJ_IO*/
621
622                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
623                 /*
624                  * akpm: jbd2_journal_write_metadata_buffer() sets
625                  * new_bh->b_transaction to commit_transaction.
626                  * We need to clean this up before we release new_bh
627                  * (which is of type BJ_IO)
628                  */
629                 JBUFFER_TRACE(jh, "ph3: write metadata");
630                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
631                                                       jh, &new_jh, blocknr);
632                 if (flags < 0) {
633                         jbd2_journal_abort(journal, flags);
634                         continue;
635                 }
636                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
637                 wbuf[bufs++] = jh2bh(new_jh);
638
639                 /* Record the new block's tag in the current descriptor
640                    buffer */
641
642                 tag_flag = 0;
643                 if (flags & 1)
644                         tag_flag |= JBD2_FLAG_ESCAPE;
645                 if (!first_tag)
646                         tag_flag |= JBD2_FLAG_SAME_UUID;
647
648                 tag = (journal_block_tag_t *) tagp;
649                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
650                 tag->t_flags = cpu_to_be32(tag_flag);
651                 tagp += tag_bytes;
652                 space_left -= tag_bytes;
653
654                 if (first_tag) {
655                         memcpy (tagp, journal->j_uuid, 16);
656                         tagp += 16;
657                         space_left -= 16;
658                         first_tag = 0;
659                 }
660
661                 /* If there's no more to do, or if the descriptor is full,
662                    let the IO rip! */
663
664                 if (bufs == journal->j_wbufsize ||
665                     commit_transaction->t_buffers == NULL ||
666                     space_left < tag_bytes + 16) {
667
668                         jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
669
670                         /* Write an end-of-descriptor marker before
671                            submitting the IOs.  "tag" still points to
672                            the last tag we set up. */
673
674                         tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
675
676 start_journal_io:
677                         for (i = 0; i < bufs; i++) {
678                                 struct buffer_head *bh = wbuf[i];
679                                 /*
680                                  * Compute checksum.
681                                  */
682                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
683                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
684                                         crc32_sum =
685                                             jbd2_checksum_data(crc32_sum, bh);
686                                 }
687
688                                 lock_buffer(bh);
689                                 clear_buffer_dirty(bh);
690                                 set_buffer_uptodate(bh);
691                                 bh->b_end_io = journal_end_buffer_io_sync;
692                                 submit_bh(write_op, bh);
693                         }
694                         cond_resched();
695                         stats.run.rs_blocks_logged += bufs;
696
697                         /* Force a new descriptor to be generated next
698                            time round the loop. */
699                         descriptor = NULL;
700                         bufs = 0;
701                 }
702         }
703
704         /* 
705          * If the journal is not located on the file system device,
706          * then we must flush the file system device before we issue
707          * the commit record
708          */
709         if (commit_transaction->t_flushed_data_blocks &&
710             (journal->j_fs_dev != journal->j_dev) &&
711             (journal->j_flags & JBD2_BARRIER))
712                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL,
713                         BLKDEV_IFL_WAIT);
714
715         /* Done it all: now write the commit record asynchronously. */
716         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
717                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
718                 err = journal_submit_commit_record(journal, commit_transaction,
719                                                  &cbh, crc32_sum);
720                 if (err)
721                         __jbd2_journal_abort_hard(journal);
722                 if (journal->j_flags & JBD2_BARRIER)
723                         blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL,
724                                 BLKDEV_IFL_WAIT);
725         }
726
727         err = journal_finish_inode_data_buffers(journal, commit_transaction);
728         if (err) {
729                 printk(KERN_WARNING
730                         "JBD2: Detected IO errors while flushing file data "
731                        "on %s\n", journal->j_devname);
732                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
733                         jbd2_journal_abort(journal, err);
734                 err = 0;
735         }
736
737         /* Lo and behold: we have just managed to send a transaction to
738            the log.  Before we can commit it, wait for the IO so far to
739            complete.  Control buffers being written are on the
740            transaction's t_log_list queue, and metadata buffers are on
741            the t_iobuf_list queue.
742
743            Wait for the buffers in reverse order.  That way we are
744            less likely to be woken up until all IOs have completed, and
745            so we incur less scheduling load.
746         */
747
748         jbd_debug(3, "JBD: commit phase 3\n");
749
750         /*
751          * akpm: these are BJ_IO, and j_list_lock is not needed.
752          * See __journal_try_to_free_buffer.
753          */
754 wait_for_iobuf:
755         while (commit_transaction->t_iobuf_list != NULL) {
756                 struct buffer_head *bh;
757
758                 jh = commit_transaction->t_iobuf_list->b_tprev;
759                 bh = jh2bh(jh);
760                 if (buffer_locked(bh)) {
761                         wait_on_buffer(bh);
762                         goto wait_for_iobuf;
763                 }
764                 if (cond_resched())
765                         goto wait_for_iobuf;
766
767                 if (unlikely(!buffer_uptodate(bh)))
768                         err = -EIO;
769
770                 clear_buffer_jwrite(bh);
771
772                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
773                 jbd2_journal_unfile_buffer(journal, jh);
774
775                 /*
776                  * ->t_iobuf_list should contain only dummy buffer_heads
777                  * which were created by jbd2_journal_write_metadata_buffer().
778                  */
779                 BUFFER_TRACE(bh, "dumping temporary bh");
780                 jbd2_journal_put_journal_head(jh);
781                 __brelse(bh);
782                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
783                 free_buffer_head(bh);
784
785                 /* We also have to unlock and free the corresponding
786                    shadowed buffer */
787                 jh = commit_transaction->t_shadow_list->b_tprev;
788                 bh = jh2bh(jh);
789                 clear_bit(BH_JWrite, &bh->b_state);
790                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
791
792                 /* The metadata is now released for reuse, but we need
793                    to remember it against this transaction so that when
794                    we finally commit, we can do any checkpointing
795                    required. */
796                 JBUFFER_TRACE(jh, "file as BJ_Forget");
797                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
798                 /* Wake up any transactions which were waiting for this
799                    IO to complete */
800                 wake_up_bit(&bh->b_state, BH_Unshadow);
801                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
802                 __brelse(bh);
803         }
804
805         J_ASSERT (commit_transaction->t_shadow_list == NULL);
806
807         jbd_debug(3, "JBD: commit phase 4\n");
808
809         /* Here we wait for the revoke record and descriptor record buffers */
810  wait_for_ctlbuf:
811         while (commit_transaction->t_log_list != NULL) {
812                 struct buffer_head *bh;
813
814                 jh = commit_transaction->t_log_list->b_tprev;
815                 bh = jh2bh(jh);
816                 if (buffer_locked(bh)) {
817                         wait_on_buffer(bh);
818                         goto wait_for_ctlbuf;
819                 }
820                 if (cond_resched())
821                         goto wait_for_ctlbuf;
822
823                 if (unlikely(!buffer_uptodate(bh)))
824                         err = -EIO;
825
826                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
827                 clear_buffer_jwrite(bh);
828                 jbd2_journal_unfile_buffer(journal, jh);
829                 jbd2_journal_put_journal_head(jh);
830                 __brelse(bh);           /* One for getblk */
831                 /* AKPM: bforget here */
832         }
833
834         if (err)
835                 jbd2_journal_abort(journal, err);
836
837         jbd_debug(3, "JBD: commit phase 5\n");
838
839         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
840                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
841                 err = journal_submit_commit_record(journal, commit_transaction,
842                                                 &cbh, crc32_sum);
843                 if (err)
844                         __jbd2_journal_abort_hard(journal);
845         }
846         if (!err && !is_journal_aborted(journal))
847                 err = journal_wait_on_commit_record(journal, cbh);
848
849         if (err)
850                 jbd2_journal_abort(journal, err);
851
852         /* End of a transaction!  Finally, we can do checkpoint
853            processing: any buffers committed as a result of this
854            transaction can be removed from any checkpoint list it was on
855            before. */
856
857         jbd_debug(3, "JBD: commit phase 6\n");
858
859         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
860         J_ASSERT(commit_transaction->t_buffers == NULL);
861         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
862         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
863         J_ASSERT(commit_transaction->t_shadow_list == NULL);
864         J_ASSERT(commit_transaction->t_log_list == NULL);
865
866 restart_loop:
867         /*
868          * As there are other places (journal_unmap_buffer()) adding buffers
869          * to this list we have to be careful and hold the j_list_lock.
870          */
871         spin_lock(&journal->j_list_lock);
872         while (commit_transaction->t_forget) {
873                 transaction_t *cp_transaction;
874                 struct buffer_head *bh;
875
876                 jh = commit_transaction->t_forget;
877                 spin_unlock(&journal->j_list_lock);
878                 bh = jh2bh(jh);
879                 jbd_lock_bh_state(bh);
880                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
881
882                 /*
883                  * If there is undo-protected committed data against
884                  * this buffer, then we can remove it now.  If it is a
885                  * buffer needing such protection, the old frozen_data
886                  * field now points to a committed version of the
887                  * buffer, so rotate that field to the new committed
888                  * data.
889                  *
890                  * Otherwise, we can just throw away the frozen data now.
891                  *
892                  * We also know that the frozen data has already fired
893                  * its triggers if they exist, so we can clear that too.
894                  */
895                 if (jh->b_committed_data) {
896                         jbd2_free(jh->b_committed_data, bh->b_size);
897                         jh->b_committed_data = NULL;
898                         if (jh->b_frozen_data) {
899                                 jh->b_committed_data = jh->b_frozen_data;
900                                 jh->b_frozen_data = NULL;
901                                 jh->b_frozen_triggers = NULL;
902                         }
903                 } else if (jh->b_frozen_data) {
904                         jbd2_free(jh->b_frozen_data, bh->b_size);
905                         jh->b_frozen_data = NULL;
906                         jh->b_frozen_triggers = NULL;
907                 }
908
909                 spin_lock(&journal->j_list_lock);
910                 cp_transaction = jh->b_cp_transaction;
911                 if (cp_transaction) {
912                         JBUFFER_TRACE(jh, "remove from old cp transaction");
913                         cp_transaction->t_chp_stats.cs_dropped++;
914                         __jbd2_journal_remove_checkpoint(jh);
915                 }
916
917                 /* Only re-checkpoint the buffer_head if it is marked
918                  * dirty.  If the buffer was added to the BJ_Forget list
919                  * by jbd2_journal_forget, it may no longer be dirty and
920                  * there's no point in keeping a checkpoint record for
921                  * it. */
922
923                 /* A buffer which has been freed while still being
924                  * journaled by a previous transaction may end up still
925                  * being dirty here, but we want to avoid writing back
926                  * that buffer in the future after the "add to orphan"
927                  * operation been committed,  That's not only a performance
928                  * gain, it also stops aliasing problems if the buffer is
929                  * left behind for writeback and gets reallocated for another
930                  * use in a different page. */
931                 if (buffer_freed(bh) && !jh->b_next_transaction) {
932                         clear_buffer_freed(bh);
933                         clear_buffer_jbddirty(bh);
934                 }
935
936                 if (buffer_jbddirty(bh)) {
937                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
938                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
939                         if (is_journal_aborted(journal))
940                                 clear_buffer_jbddirty(bh);
941                         JBUFFER_TRACE(jh, "refile for checkpoint writeback");
942                         __jbd2_journal_refile_buffer(jh);
943                         jbd_unlock_bh_state(bh);
944                 } else {
945                         J_ASSERT_BH(bh, !buffer_dirty(bh));
946                         /* The buffer on BJ_Forget list and not jbddirty means
947                          * it has been freed by this transaction and hence it
948                          * could not have been reallocated until this
949                          * transaction has committed. *BUT* it could be
950                          * reallocated once we have written all the data to
951                          * disk and before we process the buffer on BJ_Forget
952                          * list. */
953                         JBUFFER_TRACE(jh, "refile or unfile freed buffer");
954                         __jbd2_journal_refile_buffer(jh);
955                         if (!jh->b_transaction) {
956                                 jbd_unlock_bh_state(bh);
957                                  /* needs a brelse */
958                                 jbd2_journal_remove_journal_head(bh);
959                                 release_buffer_page(bh);
960                         } else
961                                 jbd_unlock_bh_state(bh);
962                 }
963                 cond_resched_lock(&journal->j_list_lock);
964         }
965         spin_unlock(&journal->j_list_lock);
966         /*
967          * This is a bit sleazy.  We use j_list_lock to protect transition
968          * of a transaction into T_FINISHED state and calling
969          * __jbd2_journal_drop_transaction(). Otherwise we could race with
970          * other checkpointing code processing the transaction...
971          */
972         write_lock(&journal->j_state_lock);
973         spin_lock(&journal->j_list_lock);
974         /*
975          * Now recheck if some buffers did not get attached to the transaction
976          * while the lock was dropped...
977          */
978         if (commit_transaction->t_forget) {
979                 spin_unlock(&journal->j_list_lock);
980                 write_unlock(&journal->j_state_lock);
981                 goto restart_loop;
982         }
983
984         /* Done with this transaction! */
985
986         jbd_debug(3, "JBD: commit phase 7\n");
987
988         J_ASSERT(commit_transaction->t_state == T_COMMIT);
989
990         commit_transaction->t_start = jiffies;
991         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
992                                               commit_transaction->t_start);
993
994         /*
995          * File the transaction statistics
996          */
997         stats.ts_tid = commit_transaction->t_tid;
998         stats.run.rs_handle_count =
999                 atomic_read(&commit_transaction->t_handle_count);
1000         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1001                              commit_transaction->t_tid, &stats.run);
1002
1003         /*
1004          * Calculate overall stats
1005          */
1006         spin_lock(&journal->j_history_lock);
1007         journal->j_stats.ts_tid++;
1008         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1009         journal->j_stats.run.rs_running += stats.run.rs_running;
1010         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1011         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1012         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1013         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1014         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1015         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1016         spin_unlock(&journal->j_history_lock);
1017
1018         commit_transaction->t_state = T_FINISHED;
1019         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1020         journal->j_commit_sequence = commit_transaction->t_tid;
1021         journal->j_committing_transaction = NULL;
1022         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1023
1024         /*
1025          * weight the commit time higher than the average time so we don't
1026          * react too strongly to vast changes in the commit time
1027          */
1028         if (likely(journal->j_average_commit_time))
1029                 journal->j_average_commit_time = (commit_time +
1030                                 journal->j_average_commit_time*3) / 4;
1031         else
1032                 journal->j_average_commit_time = commit_time;
1033         write_unlock(&journal->j_state_lock);
1034
1035         if (commit_transaction->t_checkpoint_list == NULL &&
1036             commit_transaction->t_checkpoint_io_list == NULL) {
1037                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1038                 to_free = 1;
1039         } else {
1040                 if (journal->j_checkpoint_transactions == NULL) {
1041                         journal->j_checkpoint_transactions = commit_transaction;
1042                         commit_transaction->t_cpnext = commit_transaction;
1043                         commit_transaction->t_cpprev = commit_transaction;
1044                 } else {
1045                         commit_transaction->t_cpnext =
1046                                 journal->j_checkpoint_transactions;
1047                         commit_transaction->t_cpprev =
1048                                 commit_transaction->t_cpnext->t_cpprev;
1049                         commit_transaction->t_cpnext->t_cpprev =
1050                                 commit_transaction;
1051                         commit_transaction->t_cpprev->t_cpnext =
1052                                 commit_transaction;
1053                 }
1054         }
1055         spin_unlock(&journal->j_list_lock);
1056
1057         if (journal->j_commit_callback)
1058                 journal->j_commit_callback(journal, commit_transaction);
1059
1060         trace_jbd2_end_commit(journal, commit_transaction);
1061         jbd_debug(1, "JBD: commit %d complete, head %d\n",
1062                   journal->j_commit_sequence, journal->j_tail_sequence);
1063         if (to_free)
1064                 kfree(commit_transaction);
1065
1066         wake_up(&journal->j_wait_done_commit);
1067 }