sched: Move up affinity check to mitigate useless redoing overhead
[platform/adaptation/renesas_rcar/renesas_kernel.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 <linux/bitops.h>
30 #include <trace/events/jbd2.h>
31
32 /*
33  * Default IO end handler for temporary BJ_IO buffer_heads.
34  */
35 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
36 {
37         BUFFER_TRACE(bh, "");
38         if (uptodate)
39                 set_buffer_uptodate(bh);
40         else
41                 clear_buffer_uptodate(bh);
42         unlock_buffer(bh);
43 }
44
45 /*
46  * When an ext4 file is truncated, it is possible that some pages are not
47  * successfully freed, because they are attached to a committing transaction.
48  * After the transaction commits, these pages are left on the LRU, with no
49  * ->mapping, and with attached buffers.  These pages are trivially reclaimable
50  * by the VM, but their apparent absence upsets the VM accounting, and it makes
51  * the numbers in /proc/meminfo look odd.
52  *
53  * So here, we have a buffer which has just come off the forget list.  Look to
54  * see if we can strip all buffers from the backing page.
55  *
56  * Called under lock_journal(), and possibly under journal_datalist_lock.  The
57  * caller provided us with a ref against the buffer, and we drop that here.
58  */
59 static void release_buffer_page(struct buffer_head *bh)
60 {
61         struct page *page;
62
63         if (buffer_dirty(bh))
64                 goto nope;
65         if (atomic_read(&bh->b_count) != 1)
66                 goto nope;
67         page = bh->b_page;
68         if (!page)
69                 goto nope;
70         if (page->mapping)
71                 goto nope;
72
73         /* OK, it's a truncated page */
74         if (!trylock_page(page))
75                 goto nope;
76
77         page_cache_get(page);
78         __brelse(bh);
79         try_to_free_buffers(page);
80         unlock_page(page);
81         page_cache_release(page);
82         return;
83
84 nope:
85         __brelse(bh);
86 }
87
88 static void jbd2_commit_block_csum_set(journal_t *j,
89                                        struct journal_head *descriptor)
90 {
91         struct commit_header *h;
92         __u32 csum;
93
94         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
95                 return;
96
97         h = (struct commit_header *)(jh2bh(descriptor)->b_data);
98         h->h_chksum_type = 0;
99         h->h_chksum_size = 0;
100         h->h_chksum[0] = 0;
101         csum = jbd2_chksum(j, j->j_csum_seed, jh2bh(descriptor)->b_data,
102                            j->j_blocksize);
103         h->h_chksum[0] = cpu_to_be32(csum);
104 }
105
106 /*
107  * Done it all: now submit the commit record.  We should have
108  * cleaned up our previous buffers by now, so if we are in abort
109  * mode we can now just skip the rest of the journal write
110  * entirely.
111  *
112  * Returns 1 if the journal needs to be aborted or 0 on success
113  */
114 static int journal_submit_commit_record(journal_t *journal,
115                                         transaction_t *commit_transaction,
116                                         struct buffer_head **cbh,
117                                         __u32 crc32_sum)
118 {
119         struct journal_head *descriptor;
120         struct commit_header *tmp;
121         struct buffer_head *bh;
122         int ret;
123         struct timespec now = current_kernel_time();
124
125         *cbh = NULL;
126
127         if (is_journal_aborted(journal))
128                 return 0;
129
130         descriptor = jbd2_journal_get_descriptor_buffer(journal);
131         if (!descriptor)
132                 return 1;
133
134         bh = jh2bh(descriptor);
135
136         tmp = (struct commit_header *)bh->b_data;
137         tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
138         tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
139         tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
140         tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
141         tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
142
143         if (JBD2_HAS_COMPAT_FEATURE(journal,
144                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
145                 tmp->h_chksum_type      = JBD2_CRC32_CHKSUM;
146                 tmp->h_chksum_size      = JBD2_CRC32_CHKSUM_SIZE;
147                 tmp->h_chksum[0]        = cpu_to_be32(crc32_sum);
148         }
149         jbd2_commit_block_csum_set(journal, descriptor);
150
151         JBUFFER_TRACE(descriptor, "submit commit block");
152         lock_buffer(bh);
153         clear_buffer_dirty(bh);
154         set_buffer_uptodate(bh);
155         bh->b_end_io = journal_end_buffer_io_sync;
156
157         if (journal->j_flags & JBD2_BARRIER &&
158             !JBD2_HAS_INCOMPAT_FEATURE(journal,
159                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
160                 ret = submit_bh(WRITE_SYNC | WRITE_FLUSH_FUA, bh);
161         else
162                 ret = submit_bh(WRITE_SYNC, bh);
163
164         *cbh = bh;
165         return ret;
166 }
167
168 /*
169  * This function along with journal_submit_commit_record
170  * allows to write the commit record asynchronously.
171  */
172 static int journal_wait_on_commit_record(journal_t *journal,
173                                          struct buffer_head *bh)
174 {
175         int ret = 0;
176
177         clear_buffer_dirty(bh);
178         wait_on_buffer(bh);
179
180         if (unlikely(!buffer_uptodate(bh)))
181                 ret = -EIO;
182         put_bh(bh);            /* One for getblk() */
183         jbd2_journal_put_journal_head(bh2jh(bh));
184
185         return ret;
186 }
187
188 /*
189  * write the filemap data using writepage() address_space_operations.
190  * We don't do block allocation here even for delalloc. We don't
191  * use writepages() because with dealyed allocation we may be doing
192  * block allocation in writepages().
193  */
194 static int journal_submit_inode_data_buffers(struct address_space *mapping)
195 {
196         int ret;
197         struct writeback_control wbc = {
198                 .sync_mode =  WB_SYNC_ALL,
199                 .nr_to_write = mapping->nrpages * 2,
200                 .range_start = 0,
201                 .range_end = i_size_read(mapping->host),
202         };
203
204         ret = generic_writepages(mapping, &wbc);
205         return ret;
206 }
207
208 /*
209  * Submit all the data buffers of inode associated with the transaction to
210  * disk.
211  *
212  * We are in a committing transaction. Therefore no new inode can be added to
213  * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
214  * operate on from being released while we write out pages.
215  */
216 static int journal_submit_data_buffers(journal_t *journal,
217                 transaction_t *commit_transaction)
218 {
219         struct jbd2_inode *jinode;
220         int err, ret = 0;
221         struct address_space *mapping;
222
223         spin_lock(&journal->j_list_lock);
224         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
225                 mapping = jinode->i_vfs_inode->i_mapping;
226                 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
227                 spin_unlock(&journal->j_list_lock);
228                 /*
229                  * submit the inode data buffers. We use writepage
230                  * instead of writepages. Because writepages can do
231                  * block allocation  with delalloc. We need to write
232                  * only allocated blocks here.
233                  */
234                 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
235                 err = journal_submit_inode_data_buffers(mapping);
236                 if (!ret)
237                         ret = err;
238                 spin_lock(&journal->j_list_lock);
239                 J_ASSERT(jinode->i_transaction == commit_transaction);
240                 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
241                 smp_mb__after_clear_bit();
242                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
243         }
244         spin_unlock(&journal->j_list_lock);
245         return ret;
246 }
247
248 /*
249  * Wait for data submitted for writeout, refile inodes to proper
250  * transaction if needed.
251  *
252  */
253 static int journal_finish_inode_data_buffers(journal_t *journal,
254                 transaction_t *commit_transaction)
255 {
256         struct jbd2_inode *jinode, *next_i;
257         int err, ret = 0;
258
259         /* For locking, see the comment in journal_submit_data_buffers() */
260         spin_lock(&journal->j_list_lock);
261         list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
262                 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
263                 spin_unlock(&journal->j_list_lock);
264                 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
265                 if (err) {
266                         /*
267                          * Because AS_EIO is cleared by
268                          * filemap_fdatawait_range(), set it again so
269                          * that user process can get -EIO from fsync().
270                          */
271                         set_bit(AS_EIO,
272                                 &jinode->i_vfs_inode->i_mapping->flags);
273
274                         if (!ret)
275                                 ret = err;
276                 }
277                 spin_lock(&journal->j_list_lock);
278                 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
279                 smp_mb__after_clear_bit();
280                 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
281         }
282
283         /* Now refile inode to proper lists */
284         list_for_each_entry_safe(jinode, next_i,
285                                  &commit_transaction->t_inode_list, i_list) {
286                 list_del(&jinode->i_list);
287                 if (jinode->i_next_transaction) {
288                         jinode->i_transaction = jinode->i_next_transaction;
289                         jinode->i_next_transaction = NULL;
290                         list_add(&jinode->i_list,
291                                 &jinode->i_transaction->t_inode_list);
292                 } else {
293                         jinode->i_transaction = NULL;
294                 }
295         }
296         spin_unlock(&journal->j_list_lock);
297
298         return ret;
299 }
300
301 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
302 {
303         struct page *page = bh->b_page;
304         char *addr;
305         __u32 checksum;
306
307         addr = kmap_atomic(page);
308         checksum = crc32_be(crc32_sum,
309                 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
310         kunmap_atomic(addr);
311
312         return checksum;
313 }
314
315 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
316                                    unsigned long long block)
317 {
318         tag->t_blocknr = cpu_to_be32(block & (u32)~0);
319         if (tag_bytes > JBD2_TAG_SIZE32)
320                 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
321 }
322
323 static void jbd2_descr_block_csum_set(journal_t *j,
324                                       struct journal_head *descriptor)
325 {
326         struct jbd2_journal_block_tail *tail;
327         __u32 csum;
328
329         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
330                 return;
331
332         tail = (struct jbd2_journal_block_tail *)
333                         (jh2bh(descriptor)->b_data + j->j_blocksize -
334                         sizeof(struct jbd2_journal_block_tail));
335         tail->t_checksum = 0;
336         csum = jbd2_chksum(j, j->j_csum_seed, jh2bh(descriptor)->b_data,
337                            j->j_blocksize);
338         tail->t_checksum = cpu_to_be32(csum);
339 }
340
341 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
342                                     struct buffer_head *bh, __u32 sequence)
343 {
344         struct page *page = bh->b_page;
345         __u8 *addr;
346         __u32 csum;
347
348         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
349                 return;
350
351         sequence = cpu_to_be32(sequence);
352         addr = kmap_atomic(page);
353         csum = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&sequence,
354                           sizeof(sequence));
355         csum = jbd2_chksum(j, csum, addr + offset_in_page(bh->b_data),
356                           bh->b_size);
357         kunmap_atomic(addr);
358
359         tag->t_checksum = cpu_to_be32(csum);
360 }
361 /*
362  * jbd2_journal_commit_transaction
363  *
364  * The primary function for committing a transaction to the log.  This
365  * function is called by the journal thread to begin a complete commit.
366  */
367 void jbd2_journal_commit_transaction(journal_t *journal)
368 {
369         struct transaction_stats_s stats;
370         transaction_t *commit_transaction;
371         struct journal_head *jh, *new_jh, *descriptor;
372         struct buffer_head **wbuf = journal->j_wbuf;
373         int bufs;
374         int flags;
375         int err;
376         unsigned long long blocknr;
377         ktime_t start_time;
378         u64 commit_time;
379         char *tagp = NULL;
380         journal_header_t *header;
381         journal_block_tag_t *tag = NULL;
382         int space_left = 0;
383         int first_tag = 0;
384         int tag_flag;
385         int i, to_free = 0;
386         int tag_bytes = journal_tag_bytes(journal);
387         struct buffer_head *cbh = NULL; /* For transactional checksums */
388         __u32 crc32_sum = ~0;
389         struct blk_plug plug;
390         /* Tail of the journal */
391         unsigned long first_block;
392         tid_t first_tid;
393         int update_tail;
394         int csum_size = 0;
395
396         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
397                 csum_size = sizeof(struct jbd2_journal_block_tail);
398
399         /*
400          * First job: lock down the current transaction and wait for
401          * all outstanding updates to complete.
402          */
403
404         /* Do we need to erase the effects of a prior jbd2_journal_flush? */
405         if (journal->j_flags & JBD2_FLUSHED) {
406                 jbd_debug(3, "super block updated\n");
407                 mutex_lock(&journal->j_checkpoint_mutex);
408                 /*
409                  * We hold j_checkpoint_mutex so tail cannot change under us.
410                  * We don't need any special data guarantees for writing sb
411                  * since journal is empty and it is ok for write to be
412                  * flushed only with transaction commit.
413                  */
414                 jbd2_journal_update_sb_log_tail(journal,
415                                                 journal->j_tail_sequence,
416                                                 journal->j_tail,
417                                                 WRITE_SYNC);
418                 mutex_unlock(&journal->j_checkpoint_mutex);
419         } else {
420                 jbd_debug(3, "superblock not updated\n");
421         }
422
423         J_ASSERT(journal->j_running_transaction != NULL);
424         J_ASSERT(journal->j_committing_transaction == NULL);
425
426         commit_transaction = journal->j_running_transaction;
427         J_ASSERT(commit_transaction->t_state == T_RUNNING);
428
429         trace_jbd2_start_commit(journal, commit_transaction);
430         jbd_debug(1, "JBD2: starting commit of transaction %d\n",
431                         commit_transaction->t_tid);
432
433         write_lock(&journal->j_state_lock);
434         commit_transaction->t_state = T_LOCKED;
435
436         trace_jbd2_commit_locking(journal, commit_transaction);
437         stats.run.rs_wait = commit_transaction->t_max_wait;
438         stats.run.rs_request_delay = 0;
439         stats.run.rs_locked = jiffies;
440         if (commit_transaction->t_requested)
441                 stats.run.rs_request_delay =
442                         jbd2_time_diff(commit_transaction->t_requested,
443                                        stats.run.rs_locked);
444         stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
445                                               stats.run.rs_locked);
446
447         spin_lock(&commit_transaction->t_handle_lock);
448         while (atomic_read(&commit_transaction->t_updates)) {
449                 DEFINE_WAIT(wait);
450
451                 prepare_to_wait(&journal->j_wait_updates, &wait,
452                                         TASK_UNINTERRUPTIBLE);
453                 if (atomic_read(&commit_transaction->t_updates)) {
454                         spin_unlock(&commit_transaction->t_handle_lock);
455                         write_unlock(&journal->j_state_lock);
456                         schedule();
457                         write_lock(&journal->j_state_lock);
458                         spin_lock(&commit_transaction->t_handle_lock);
459                 }
460                 finish_wait(&journal->j_wait_updates, &wait);
461         }
462         spin_unlock(&commit_transaction->t_handle_lock);
463
464         J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
465                         journal->j_max_transaction_buffers);
466
467         /*
468          * First thing we are allowed to do is to discard any remaining
469          * BJ_Reserved buffers.  Note, it is _not_ permissible to assume
470          * that there are no such buffers: if a large filesystem
471          * operation like a truncate needs to split itself over multiple
472          * transactions, then it may try to do a jbd2_journal_restart() while
473          * there are still BJ_Reserved buffers outstanding.  These must
474          * be released cleanly from the current transaction.
475          *
476          * In this case, the filesystem must still reserve write access
477          * again before modifying the buffer in the new transaction, but
478          * we do not require it to remember exactly which old buffers it
479          * has reserved.  This is consistent with the existing behaviour
480          * that multiple jbd2_journal_get_write_access() calls to the same
481          * buffer are perfectly permissible.
482          */
483         while (commit_transaction->t_reserved_list) {
484                 jh = commit_transaction->t_reserved_list;
485                 JBUFFER_TRACE(jh, "reserved, unused: refile");
486                 /*
487                  * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
488                  * leave undo-committed data.
489                  */
490                 if (jh->b_committed_data) {
491                         struct buffer_head *bh = jh2bh(jh);
492
493                         jbd_lock_bh_state(bh);
494                         jbd2_free(jh->b_committed_data, bh->b_size);
495                         jh->b_committed_data = NULL;
496                         jbd_unlock_bh_state(bh);
497                 }
498                 jbd2_journal_refile_buffer(journal, jh);
499         }
500
501         /*
502          * Now try to drop any written-back buffers from the journal's
503          * checkpoint lists.  We do this *before* commit because it potentially
504          * frees some memory
505          */
506         spin_lock(&journal->j_list_lock);
507         __jbd2_journal_clean_checkpoint_list(journal);
508         spin_unlock(&journal->j_list_lock);
509
510         jbd_debug(3, "JBD2: commit phase 1\n");
511
512         /*
513          * Clear revoked flag to reflect there is no revoked buffers
514          * in the next transaction which is going to be started.
515          */
516         jbd2_clear_buffer_revoked_flags(journal);
517
518         /*
519          * Switch to a new revoke table.
520          */
521         jbd2_journal_switch_revoke_table(journal);
522
523         trace_jbd2_commit_flushing(journal, commit_transaction);
524         stats.run.rs_flushing = jiffies;
525         stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
526                                              stats.run.rs_flushing);
527
528         commit_transaction->t_state = T_FLUSH;
529         journal->j_committing_transaction = commit_transaction;
530         journal->j_running_transaction = NULL;
531         start_time = ktime_get();
532         commit_transaction->t_log_start = journal->j_head;
533         wake_up(&journal->j_wait_transaction_locked);
534         write_unlock(&journal->j_state_lock);
535
536         jbd_debug(3, "JBD2: commit phase 2\n");
537
538         /*
539          * Now start flushing things to disk, in the order they appear
540          * on the transaction lists.  Data blocks go first.
541          */
542         err = journal_submit_data_buffers(journal, commit_transaction);
543         if (err)
544                 jbd2_journal_abort(journal, err);
545
546         blk_start_plug(&plug);
547         jbd2_journal_write_revoke_records(journal, commit_transaction,
548                                           WRITE_SYNC);
549         blk_finish_plug(&plug);
550
551         jbd_debug(3, "JBD2: commit phase 2\n");
552
553         /*
554          * Way to go: we have now written out all of the data for a
555          * transaction!  Now comes the tricky part: we need to write out
556          * metadata.  Loop over the transaction's entire buffer list:
557          */
558         write_lock(&journal->j_state_lock);
559         commit_transaction->t_state = T_COMMIT;
560         write_unlock(&journal->j_state_lock);
561
562         trace_jbd2_commit_logging(journal, commit_transaction);
563         stats.run.rs_logging = jiffies;
564         stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
565                                                stats.run.rs_logging);
566         stats.run.rs_blocks =
567                 atomic_read(&commit_transaction->t_outstanding_credits);
568         stats.run.rs_blocks_logged = 0;
569
570         J_ASSERT(commit_transaction->t_nr_buffers <=
571                  atomic_read(&commit_transaction->t_outstanding_credits));
572
573         err = 0;
574         descriptor = NULL;
575         bufs = 0;
576         blk_start_plug(&plug);
577         while (commit_transaction->t_buffers) {
578
579                 /* Find the next buffer to be journaled... */
580
581                 jh = commit_transaction->t_buffers;
582
583                 /* If we're in abort mode, we just un-journal the buffer and
584                    release it. */
585
586                 if (is_journal_aborted(journal)) {
587                         clear_buffer_jbddirty(jh2bh(jh));
588                         JBUFFER_TRACE(jh, "journal is aborting: refile");
589                         jbd2_buffer_abort_trigger(jh,
590                                                   jh->b_frozen_data ?
591                                                   jh->b_frozen_triggers :
592                                                   jh->b_triggers);
593                         jbd2_journal_refile_buffer(journal, jh);
594                         /* If that was the last one, we need to clean up
595                          * any descriptor buffers which may have been
596                          * already allocated, even if we are now
597                          * aborting. */
598                         if (!commit_transaction->t_buffers)
599                                 goto start_journal_io;
600                         continue;
601                 }
602
603                 /* Make sure we have a descriptor block in which to
604                    record the metadata buffer. */
605
606                 if (!descriptor) {
607                         struct buffer_head *bh;
608
609                         J_ASSERT (bufs == 0);
610
611                         jbd_debug(4, "JBD2: get descriptor\n");
612
613                         descriptor = jbd2_journal_get_descriptor_buffer(journal);
614                         if (!descriptor) {
615                                 jbd2_journal_abort(journal, -EIO);
616                                 continue;
617                         }
618
619                         bh = jh2bh(descriptor);
620                         jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
621                                 (unsigned long long)bh->b_blocknr, bh->b_data);
622                         header = (journal_header_t *)&bh->b_data[0];
623                         header->h_magic     = cpu_to_be32(JBD2_MAGIC_NUMBER);
624                         header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
625                         header->h_sequence  = cpu_to_be32(commit_transaction->t_tid);
626
627                         tagp = &bh->b_data[sizeof(journal_header_t)];
628                         space_left = bh->b_size - sizeof(journal_header_t);
629                         first_tag = 1;
630                         set_buffer_jwrite(bh);
631                         set_buffer_dirty(bh);
632                         wbuf[bufs++] = bh;
633
634                         /* Record it so that we can wait for IO
635                            completion later */
636                         BUFFER_TRACE(bh, "ph3: file as descriptor");
637                         jbd2_journal_file_buffer(descriptor, commit_transaction,
638                                         BJ_LogCtl);
639                 }
640
641                 /* Where is the buffer to be written? */
642
643                 err = jbd2_journal_next_log_block(journal, &blocknr);
644                 /* If the block mapping failed, just abandon the buffer
645                    and repeat this loop: we'll fall into the
646                    refile-on-abort condition above. */
647                 if (err) {
648                         jbd2_journal_abort(journal, err);
649                         continue;
650                 }
651
652                 /*
653                  * start_this_handle() uses t_outstanding_credits to determine
654                  * the free space in the log, but this counter is changed
655                  * by jbd2_journal_next_log_block() also.
656                  */
657                 atomic_dec(&commit_transaction->t_outstanding_credits);
658
659                 /* Bump b_count to prevent truncate from stumbling over
660                    the shadowed buffer!  @@@ This can go if we ever get
661                    rid of the BJ_IO/BJ_Shadow pairing of buffers. */
662                 atomic_inc(&jh2bh(jh)->b_count);
663
664                 /* Make a temporary IO buffer with which to write it out
665                    (this will requeue both the metadata buffer and the
666                    temporary IO buffer). new_bh goes on BJ_IO*/
667
668                 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
669                 /*
670                  * akpm: jbd2_journal_write_metadata_buffer() sets
671                  * new_bh->b_transaction to commit_transaction.
672                  * We need to clean this up before we release new_bh
673                  * (which is of type BJ_IO)
674                  */
675                 JBUFFER_TRACE(jh, "ph3: write metadata");
676                 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
677                                                       jh, &new_jh, blocknr);
678                 if (flags < 0) {
679                         jbd2_journal_abort(journal, flags);
680                         continue;
681                 }
682                 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
683                 wbuf[bufs++] = jh2bh(new_jh);
684
685                 /* Record the new block's tag in the current descriptor
686                    buffer */
687
688                 tag_flag = 0;
689                 if (flags & 1)
690                         tag_flag |= JBD2_FLAG_ESCAPE;
691                 if (!first_tag)
692                         tag_flag |= JBD2_FLAG_SAME_UUID;
693
694                 tag = (journal_block_tag_t *) tagp;
695                 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
696                 tag->t_flags = cpu_to_be16(tag_flag);
697                 jbd2_block_tag_csum_set(journal, tag, jh2bh(new_jh),
698                                         commit_transaction->t_tid);
699                 tagp += tag_bytes;
700                 space_left -= tag_bytes;
701
702                 if (first_tag) {
703                         memcpy (tagp, journal->j_uuid, 16);
704                         tagp += 16;
705                         space_left -= 16;
706                         first_tag = 0;
707                 }
708
709                 /* If there's no more to do, or if the descriptor is full,
710                    let the IO rip! */
711
712                 if (bufs == journal->j_wbufsize ||
713                     commit_transaction->t_buffers == NULL ||
714                     space_left < tag_bytes + 16 + csum_size) {
715
716                         jbd_debug(4, "JBD2: Submit %d IOs\n", bufs);
717
718                         /* Write an end-of-descriptor marker before
719                            submitting the IOs.  "tag" still points to
720                            the last tag we set up. */
721
722                         tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
723
724                         jbd2_descr_block_csum_set(journal, descriptor);
725 start_journal_io:
726                         for (i = 0; i < bufs; i++) {
727                                 struct buffer_head *bh = wbuf[i];
728                                 /*
729                                  * Compute checksum.
730                                  */
731                                 if (JBD2_HAS_COMPAT_FEATURE(journal,
732                                         JBD2_FEATURE_COMPAT_CHECKSUM)) {
733                                         crc32_sum =
734                                             jbd2_checksum_data(crc32_sum, bh);
735                                 }
736
737                                 lock_buffer(bh);
738                                 clear_buffer_dirty(bh);
739                                 set_buffer_uptodate(bh);
740                                 bh->b_end_io = journal_end_buffer_io_sync;
741                                 submit_bh(WRITE_SYNC, bh);
742                         }
743                         cond_resched();
744                         stats.run.rs_blocks_logged += bufs;
745
746                         /* Force a new descriptor to be generated next
747                            time round the loop. */
748                         descriptor = NULL;
749                         bufs = 0;
750                 }
751         }
752
753         err = journal_finish_inode_data_buffers(journal, commit_transaction);
754         if (err) {
755                 printk(KERN_WARNING
756                         "JBD2: Detected IO errors while flushing file data "
757                        "on %s\n", journal->j_devname);
758                 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
759                         jbd2_journal_abort(journal, err);
760                 err = 0;
761         }
762
763         /*
764          * Get current oldest transaction in the log before we issue flush
765          * to the filesystem device. After the flush we can be sure that
766          * blocks of all older transactions are checkpointed to persistent
767          * storage and we will be safe to update journal start in the
768          * superblock with the numbers we get here.
769          */
770         update_tail =
771                 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
772
773         write_lock(&journal->j_state_lock);
774         if (update_tail) {
775                 long freed = first_block - journal->j_tail;
776
777                 if (first_block < journal->j_tail)
778                         freed += journal->j_last - journal->j_first;
779                 /* Update tail only if we free significant amount of space */
780                 if (freed < journal->j_maxlen / 4)
781                         update_tail = 0;
782         }
783         J_ASSERT(commit_transaction->t_state == T_COMMIT);
784         commit_transaction->t_state = T_COMMIT_DFLUSH;
785         write_unlock(&journal->j_state_lock);
786
787         /* 
788          * If the journal is not located on the file system device,
789          * then we must flush the file system device before we issue
790          * the commit record
791          */
792         if (commit_transaction->t_need_data_flush &&
793             (journal->j_fs_dev != journal->j_dev) &&
794             (journal->j_flags & JBD2_BARRIER))
795                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
796
797         /* Done it all: now write the commit record asynchronously. */
798         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
799                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
800                 err = journal_submit_commit_record(journal, commit_transaction,
801                                                  &cbh, crc32_sum);
802                 if (err)
803                         __jbd2_journal_abort_hard(journal);
804         }
805
806         blk_finish_plug(&plug);
807
808         /* Lo and behold: we have just managed to send a transaction to
809            the log.  Before we can commit it, wait for the IO so far to
810            complete.  Control buffers being written are on the
811            transaction's t_log_list queue, and metadata buffers are on
812            the t_iobuf_list queue.
813
814            Wait for the buffers in reverse order.  That way we are
815            less likely to be woken up until all IOs have completed, and
816            so we incur less scheduling load.
817         */
818
819         jbd_debug(3, "JBD2: commit phase 3\n");
820
821         /*
822          * akpm: these are BJ_IO, and j_list_lock is not needed.
823          * See __journal_try_to_free_buffer.
824          */
825 wait_for_iobuf:
826         while (commit_transaction->t_iobuf_list != NULL) {
827                 struct buffer_head *bh;
828
829                 jh = commit_transaction->t_iobuf_list->b_tprev;
830                 bh = jh2bh(jh);
831                 if (buffer_locked(bh)) {
832                         wait_on_buffer(bh);
833                         goto wait_for_iobuf;
834                 }
835                 if (cond_resched())
836                         goto wait_for_iobuf;
837
838                 if (unlikely(!buffer_uptodate(bh)))
839                         err = -EIO;
840
841                 clear_buffer_jwrite(bh);
842
843                 JBUFFER_TRACE(jh, "ph4: unfile after journal write");
844                 jbd2_journal_unfile_buffer(journal, jh);
845
846                 /*
847                  * ->t_iobuf_list should contain only dummy buffer_heads
848                  * which were created by jbd2_journal_write_metadata_buffer().
849                  */
850                 BUFFER_TRACE(bh, "dumping temporary bh");
851                 jbd2_journal_put_journal_head(jh);
852                 __brelse(bh);
853                 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
854                 free_buffer_head(bh);
855
856                 /* We also have to unlock and free the corresponding
857                    shadowed buffer */
858                 jh = commit_transaction->t_shadow_list->b_tprev;
859                 bh = jh2bh(jh);
860                 clear_bit(BH_JWrite, &bh->b_state);
861                 J_ASSERT_BH(bh, buffer_jbddirty(bh));
862
863                 /* The metadata is now released for reuse, but we need
864                    to remember it against this transaction so that when
865                    we finally commit, we can do any checkpointing
866                    required. */
867                 JBUFFER_TRACE(jh, "file as BJ_Forget");
868                 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
869                 /*
870                  * Wake up any transactions which were waiting for this IO to
871                  * complete. The barrier must be here so that changes by
872                  * jbd2_journal_file_buffer() take effect before wake_up_bit()
873                  * does the waitqueue check.
874                  */
875                 smp_mb();
876                 wake_up_bit(&bh->b_state, BH_Unshadow);
877                 JBUFFER_TRACE(jh, "brelse shadowed buffer");
878                 __brelse(bh);
879         }
880
881         J_ASSERT (commit_transaction->t_shadow_list == NULL);
882
883         jbd_debug(3, "JBD2: commit phase 4\n");
884
885         /* Here we wait for the revoke record and descriptor record buffers */
886  wait_for_ctlbuf:
887         while (commit_transaction->t_log_list != NULL) {
888                 struct buffer_head *bh;
889
890                 jh = commit_transaction->t_log_list->b_tprev;
891                 bh = jh2bh(jh);
892                 if (buffer_locked(bh)) {
893                         wait_on_buffer(bh);
894                         goto wait_for_ctlbuf;
895                 }
896                 if (cond_resched())
897                         goto wait_for_ctlbuf;
898
899                 if (unlikely(!buffer_uptodate(bh)))
900                         err = -EIO;
901
902                 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
903                 clear_buffer_jwrite(bh);
904                 jbd2_journal_unfile_buffer(journal, jh);
905                 jbd2_journal_put_journal_head(jh);
906                 __brelse(bh);           /* One for getblk */
907                 /* AKPM: bforget here */
908         }
909
910         if (err)
911                 jbd2_journal_abort(journal, err);
912
913         jbd_debug(3, "JBD2: commit phase 5\n");
914         write_lock(&journal->j_state_lock);
915         J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
916         commit_transaction->t_state = T_COMMIT_JFLUSH;
917         write_unlock(&journal->j_state_lock);
918
919         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
920                                        JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
921                 err = journal_submit_commit_record(journal, commit_transaction,
922                                                 &cbh, crc32_sum);
923                 if (err)
924                         __jbd2_journal_abort_hard(journal);
925         }
926         if (cbh)
927                 err = journal_wait_on_commit_record(journal, cbh);
928         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
929                                       JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
930             journal->j_flags & JBD2_BARRIER) {
931                 blkdev_issue_flush(journal->j_dev, GFP_NOFS, NULL);
932         }
933
934         if (err)
935                 jbd2_journal_abort(journal, err);
936
937         /*
938          * Now disk caches for filesystem device are flushed so we are safe to
939          * erase checkpointed transactions from the log by updating journal
940          * superblock.
941          */
942         if (update_tail)
943                 jbd2_update_log_tail(journal, first_tid, first_block);
944
945         /* End of a transaction!  Finally, we can do checkpoint
946            processing: any buffers committed as a result of this
947            transaction can be removed from any checkpoint list it was on
948            before. */
949
950         jbd_debug(3, "JBD2: commit phase 6\n");
951
952         J_ASSERT(list_empty(&commit_transaction->t_inode_list));
953         J_ASSERT(commit_transaction->t_buffers == NULL);
954         J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
955         J_ASSERT(commit_transaction->t_iobuf_list == NULL);
956         J_ASSERT(commit_transaction->t_shadow_list == NULL);
957         J_ASSERT(commit_transaction->t_log_list == NULL);
958
959 restart_loop:
960         /*
961          * As there are other places (journal_unmap_buffer()) adding buffers
962          * to this list we have to be careful and hold the j_list_lock.
963          */
964         spin_lock(&journal->j_list_lock);
965         while (commit_transaction->t_forget) {
966                 transaction_t *cp_transaction;
967                 struct buffer_head *bh;
968                 int try_to_free = 0;
969
970                 jh = commit_transaction->t_forget;
971                 spin_unlock(&journal->j_list_lock);
972                 bh = jh2bh(jh);
973                 /*
974                  * Get a reference so that bh cannot be freed before we are
975                  * done with it.
976                  */
977                 get_bh(bh);
978                 jbd_lock_bh_state(bh);
979                 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
980
981                 /*
982                  * If there is undo-protected committed data against
983                  * this buffer, then we can remove it now.  If it is a
984                  * buffer needing such protection, the old frozen_data
985                  * field now points to a committed version of the
986                  * buffer, so rotate that field to the new committed
987                  * data.
988                  *
989                  * Otherwise, we can just throw away the frozen data now.
990                  *
991                  * We also know that the frozen data has already fired
992                  * its triggers if they exist, so we can clear that too.
993                  */
994                 if (jh->b_committed_data) {
995                         jbd2_free(jh->b_committed_data, bh->b_size);
996                         jh->b_committed_data = NULL;
997                         if (jh->b_frozen_data) {
998                                 jh->b_committed_data = jh->b_frozen_data;
999                                 jh->b_frozen_data = NULL;
1000                                 jh->b_frozen_triggers = NULL;
1001                         }
1002                 } else if (jh->b_frozen_data) {
1003                         jbd2_free(jh->b_frozen_data, bh->b_size);
1004                         jh->b_frozen_data = NULL;
1005                         jh->b_frozen_triggers = NULL;
1006                 }
1007
1008                 spin_lock(&journal->j_list_lock);
1009                 cp_transaction = jh->b_cp_transaction;
1010                 if (cp_transaction) {
1011                         JBUFFER_TRACE(jh, "remove from old cp transaction");
1012                         cp_transaction->t_chp_stats.cs_dropped++;
1013                         __jbd2_journal_remove_checkpoint(jh);
1014                 }
1015
1016                 /* Only re-checkpoint the buffer_head if it is marked
1017                  * dirty.  If the buffer was added to the BJ_Forget list
1018                  * by jbd2_journal_forget, it may no longer be dirty and
1019                  * there's no point in keeping a checkpoint record for
1020                  * it. */
1021
1022                 /*
1023                 * A buffer which has been freed while still being journaled by
1024                 * a previous transaction.
1025                 */
1026                 if (buffer_freed(bh)) {
1027                         /*
1028                          * If the running transaction is the one containing
1029                          * "add to orphan" operation (b_next_transaction !=
1030                          * NULL), we have to wait for that transaction to
1031                          * commit before we can really get rid of the buffer.
1032                          * So just clear b_modified to not confuse transaction
1033                          * credit accounting and refile the buffer to
1034                          * BJ_Forget of the running transaction. If the just
1035                          * committed transaction contains "add to orphan"
1036                          * operation, we can completely invalidate the buffer
1037                          * now. We are rather through in that since the
1038                          * buffer may be still accessible when blocksize <
1039                          * pagesize and it is attached to the last partial
1040                          * page.
1041                          */
1042                         jh->b_modified = 0;
1043                         if (!jh->b_next_transaction) {
1044                                 clear_buffer_freed(bh);
1045                                 clear_buffer_jbddirty(bh);
1046                                 clear_buffer_mapped(bh);
1047                                 clear_buffer_new(bh);
1048                                 clear_buffer_req(bh);
1049                                 bh->b_bdev = NULL;
1050                         }
1051                 }
1052
1053                 if (buffer_jbddirty(bh)) {
1054                         JBUFFER_TRACE(jh, "add to new checkpointing trans");
1055                         __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1056                         if (is_journal_aborted(journal))
1057                                 clear_buffer_jbddirty(bh);
1058                 } else {
1059                         J_ASSERT_BH(bh, !buffer_dirty(bh));
1060                         /*
1061                          * The buffer on BJ_Forget list and not jbddirty means
1062                          * it has been freed by this transaction and hence it
1063                          * could not have been reallocated until this
1064                          * transaction has committed. *BUT* it could be
1065                          * reallocated once we have written all the data to
1066                          * disk and before we process the buffer on BJ_Forget
1067                          * list.
1068                          */
1069                         if (!jh->b_next_transaction)
1070                                 try_to_free = 1;
1071                 }
1072                 JBUFFER_TRACE(jh, "refile or unfile buffer");
1073                 __jbd2_journal_refile_buffer(jh);
1074                 jbd_unlock_bh_state(bh);
1075                 if (try_to_free)
1076                         release_buffer_page(bh);        /* Drops bh reference */
1077                 else
1078                         __brelse(bh);
1079                 cond_resched_lock(&journal->j_list_lock);
1080         }
1081         spin_unlock(&journal->j_list_lock);
1082         /*
1083          * This is a bit sleazy.  We use j_list_lock to protect transition
1084          * of a transaction into T_FINISHED state and calling
1085          * __jbd2_journal_drop_transaction(). Otherwise we could race with
1086          * other checkpointing code processing the transaction...
1087          */
1088         write_lock(&journal->j_state_lock);
1089         spin_lock(&journal->j_list_lock);
1090         /*
1091          * Now recheck if some buffers did not get attached to the transaction
1092          * while the lock was dropped...
1093          */
1094         if (commit_transaction->t_forget) {
1095                 spin_unlock(&journal->j_list_lock);
1096                 write_unlock(&journal->j_state_lock);
1097                 goto restart_loop;
1098         }
1099
1100         /* Done with this transaction! */
1101
1102         jbd_debug(3, "JBD2: commit phase 7\n");
1103
1104         J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1105
1106         commit_transaction->t_start = jiffies;
1107         stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1108                                               commit_transaction->t_start);
1109
1110         /*
1111          * File the transaction statistics
1112          */
1113         stats.ts_tid = commit_transaction->t_tid;
1114         stats.run.rs_handle_count =
1115                 atomic_read(&commit_transaction->t_handle_count);
1116         trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1117                              commit_transaction->t_tid, &stats.run);
1118
1119         /*
1120          * Calculate overall stats
1121          */
1122         spin_lock(&journal->j_history_lock);
1123         journal->j_stats.ts_tid++;
1124         if (commit_transaction->t_requested)
1125                 journal->j_stats.ts_requested++;
1126         journal->j_stats.run.rs_wait += stats.run.rs_wait;
1127         journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1128         journal->j_stats.run.rs_running += stats.run.rs_running;
1129         journal->j_stats.run.rs_locked += stats.run.rs_locked;
1130         journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1131         journal->j_stats.run.rs_logging += stats.run.rs_logging;
1132         journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1133         journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1134         journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1135         spin_unlock(&journal->j_history_lock);
1136
1137         commit_transaction->t_state = T_FINISHED;
1138         J_ASSERT(commit_transaction == journal->j_committing_transaction);
1139         journal->j_commit_sequence = commit_transaction->t_tid;
1140         journal->j_committing_transaction = NULL;
1141         commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1142
1143         /*
1144          * weight the commit time higher than the average time so we don't
1145          * react too strongly to vast changes in the commit time
1146          */
1147         if (likely(journal->j_average_commit_time))
1148                 journal->j_average_commit_time = (commit_time +
1149                                 journal->j_average_commit_time*3) / 4;
1150         else
1151                 journal->j_average_commit_time = commit_time;
1152         write_unlock(&journal->j_state_lock);
1153
1154         if (commit_transaction->t_checkpoint_list == NULL &&
1155             commit_transaction->t_checkpoint_io_list == NULL) {
1156                 __jbd2_journal_drop_transaction(journal, commit_transaction);
1157                 to_free = 1;
1158         } else {
1159                 if (journal->j_checkpoint_transactions == NULL) {
1160                         journal->j_checkpoint_transactions = commit_transaction;
1161                         commit_transaction->t_cpnext = commit_transaction;
1162                         commit_transaction->t_cpprev = commit_transaction;
1163                 } else {
1164                         commit_transaction->t_cpnext =
1165                                 journal->j_checkpoint_transactions;
1166                         commit_transaction->t_cpprev =
1167                                 commit_transaction->t_cpnext->t_cpprev;
1168                         commit_transaction->t_cpnext->t_cpprev =
1169                                 commit_transaction;
1170                         commit_transaction->t_cpprev->t_cpnext =
1171                                 commit_transaction;
1172                 }
1173         }
1174         spin_unlock(&journal->j_list_lock);
1175
1176         if (journal->j_commit_callback)
1177                 journal->j_commit_callback(journal, commit_transaction);
1178
1179         trace_jbd2_end_commit(journal, commit_transaction);
1180         jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1181                   journal->j_commit_sequence, journal->j_tail_sequence);
1182         if (to_free)
1183                 jbd2_journal_free_transaction(commit_transaction);
1184
1185         wake_up(&journal->j_wait_done_commit);
1186 }