Merge branch 'next' into upstream-merge
[linux-2.6.git] / fs / jbd2 / checkpoint.c
1 /*
2  * linux/fs/jbd2/checkpoint.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
5  *
6  * Copyright 1999 Red Hat Software --- 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  * Checkpoint routines for the generic filesystem journaling code.
13  * Part of the ext2fs journaling system.
14  *
15  * Checkpointing is the process of ensuring that a section of the log is
16  * committed fully to disk, so that that portion of the log can be
17  * reused.
18  */
19
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/blkdev.h>
26 #include <trace/events/jbd2.h>
27
28 /*
29  * Unlink a buffer from a transaction checkpoint list.
30  *
31  * Called with j_list_lock held.
32  */
33 static inline void __buffer_unlink_first(struct journal_head *jh)
34 {
35         transaction_t *transaction = jh->b_cp_transaction;
36
37         jh->b_cpnext->b_cpprev = jh->b_cpprev;
38         jh->b_cpprev->b_cpnext = jh->b_cpnext;
39         if (transaction->t_checkpoint_list == jh) {
40                 transaction->t_checkpoint_list = jh->b_cpnext;
41                 if (transaction->t_checkpoint_list == jh)
42                         transaction->t_checkpoint_list = NULL;
43         }
44 }
45
46 /*
47  * Unlink a buffer from a transaction checkpoint(io) list.
48  *
49  * Called with j_list_lock held.
50  */
51 static inline void __buffer_unlink(struct journal_head *jh)
52 {
53         transaction_t *transaction = jh->b_cp_transaction;
54
55         __buffer_unlink_first(jh);
56         if (transaction->t_checkpoint_io_list == jh) {
57                 transaction->t_checkpoint_io_list = jh->b_cpnext;
58                 if (transaction->t_checkpoint_io_list == jh)
59                         transaction->t_checkpoint_io_list = NULL;
60         }
61 }
62
63 /*
64  * Move a buffer from the checkpoint list to the checkpoint io list
65  *
66  * Called with j_list_lock held
67  */
68 static inline void __buffer_relink_io(struct journal_head *jh)
69 {
70         transaction_t *transaction = jh->b_cp_transaction;
71
72         __buffer_unlink_first(jh);
73
74         if (!transaction->t_checkpoint_io_list) {
75                 jh->b_cpnext = jh->b_cpprev = jh;
76         } else {
77                 jh->b_cpnext = transaction->t_checkpoint_io_list;
78                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
79                 jh->b_cpprev->b_cpnext = jh;
80                 jh->b_cpnext->b_cpprev = jh;
81         }
82         transaction->t_checkpoint_io_list = jh;
83 }
84
85 /*
86  * Try to release a checkpointed buffer from its transaction.
87  * Returns 1 if we released it and 2 if we also released the
88  * whole transaction.
89  *
90  * Requires j_list_lock
91  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
92  */
93 static int __try_to_free_cp_buf(struct journal_head *jh)
94 {
95         int ret = 0;
96         struct buffer_head *bh = jh2bh(jh);
97
98         if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
99             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
100                 JBUFFER_TRACE(jh, "remove from checkpoint list");
101                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
102                 jbd_unlock_bh_state(bh);
103                 jbd2_journal_remove_journal_head(bh);
104                 BUFFER_TRACE(bh, "release");
105                 __brelse(bh);
106         } else {
107                 jbd_unlock_bh_state(bh);
108         }
109         return ret;
110 }
111
112 /*
113  * __jbd2_log_wait_for_space: wait until there is space in the journal.
114  *
115  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
116  * for a checkpoint to free up some space in the log.
117  */
118 void __jbd2_log_wait_for_space(journal_t *journal)
119 {
120         int nblocks, space_left;
121         /* assert_spin_locked(&journal->j_state_lock); */
122
123         nblocks = jbd_space_needed(journal);
124         while (__jbd2_log_space_left(journal) < nblocks) {
125                 if (journal->j_flags & JBD2_ABORT)
126                         return;
127                 write_unlock(&journal->j_state_lock);
128                 mutex_lock(&journal->j_checkpoint_mutex);
129
130                 /*
131                  * Test again, another process may have checkpointed while we
132                  * were waiting for the checkpoint lock. If there are no
133                  * transactions ready to be checkpointed, try to recover
134                  * journal space by calling cleanup_journal_tail(), and if
135                  * that doesn't work, by waiting for the currently committing
136                  * transaction to complete.  If there is absolutely no way
137                  * to make progress, this is either a BUG or corrupted
138                  * filesystem, so abort the journal and leave a stack
139                  * trace for forensic evidence.
140                  */
141                 write_lock(&journal->j_state_lock);
142                 spin_lock(&journal->j_list_lock);
143                 nblocks = jbd_space_needed(journal);
144                 space_left = __jbd2_log_space_left(journal);
145                 if (space_left < nblocks) {
146                         int chkpt = journal->j_checkpoint_transactions != NULL;
147                         tid_t tid = 0;
148
149                         if (journal->j_committing_transaction)
150                                 tid = journal->j_committing_transaction->t_tid;
151                         spin_unlock(&journal->j_list_lock);
152                         write_unlock(&journal->j_state_lock);
153                         if (chkpt) {
154                                 jbd2_log_do_checkpoint(journal);
155                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
156                                 /* We were able to recover space; yay! */
157                                 ;
158                         } else if (tid) {
159                                 jbd2_log_wait_commit(journal, tid);
160                         } else {
161                                 printk(KERN_ERR "%s: needed %d blocks and "
162                                        "only had %d space available\n",
163                                        __func__, nblocks, space_left);
164                                 printk(KERN_ERR "%s: no way to get more "
165                                        "journal space in %s\n", __func__,
166                                        journal->j_devname);
167                                 WARN_ON(1);
168                                 jbd2_journal_abort(journal, 0);
169                         }
170                         write_lock(&journal->j_state_lock);
171                 } else {
172                         spin_unlock(&journal->j_list_lock);
173                 }
174                 mutex_unlock(&journal->j_checkpoint_mutex);
175         }
176 }
177
178 /*
179  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
180  * The caller must restart a list walk.  Wait for someone else to run
181  * jbd_unlock_bh_state().
182  */
183 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
184         __releases(journal->j_list_lock)
185 {
186         get_bh(bh);
187         spin_unlock(&journal->j_list_lock);
188         jbd_lock_bh_state(bh);
189         jbd_unlock_bh_state(bh);
190         put_bh(bh);
191 }
192
193 /*
194  * Clean up transaction's list of buffers submitted for io.
195  * We wait for any pending IO to complete and remove any clean
196  * buffers. Note that we take the buffers in the opposite ordering
197  * from the one in which they were submitted for IO.
198  *
199  * Return 0 on success, and return <0 if some buffers have failed
200  * to be written out.
201  *
202  * Called with j_list_lock held.
203  */
204 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
205 {
206         struct journal_head *jh;
207         struct buffer_head *bh;
208         tid_t this_tid;
209         int released = 0;
210         int ret = 0;
211
212         this_tid = transaction->t_tid;
213 restart:
214         /* Did somebody clean up the transaction in the meanwhile? */
215         if (journal->j_checkpoint_transactions != transaction ||
216                         transaction->t_tid != this_tid)
217                 return ret;
218         while (!released && transaction->t_checkpoint_io_list) {
219                 jh = transaction->t_checkpoint_io_list;
220                 bh = jh2bh(jh);
221                 if (!jbd_trylock_bh_state(bh)) {
222                         jbd_sync_bh(journal, bh);
223                         spin_lock(&journal->j_list_lock);
224                         goto restart;
225                 }
226                 if (buffer_locked(bh)) {
227                         atomic_inc(&bh->b_count);
228                         spin_unlock(&journal->j_list_lock);
229                         jbd_unlock_bh_state(bh);
230                         wait_on_buffer(bh);
231                         /* the journal_head may have gone by now */
232                         BUFFER_TRACE(bh, "brelse");
233                         __brelse(bh);
234                         spin_lock(&journal->j_list_lock);
235                         goto restart;
236                 }
237                 if (unlikely(buffer_write_io_error(bh)))
238                         ret = -EIO;
239
240                 /*
241                  * Now in whatever state the buffer currently is, we know that
242                  * it has been written out and so we can drop it from the list
243                  */
244                 released = __jbd2_journal_remove_checkpoint(jh);
245                 jbd_unlock_bh_state(bh);
246                 jbd2_journal_remove_journal_head(bh);
247                 __brelse(bh);
248         }
249
250         return ret;
251 }
252
253 static void
254 __flush_batch(journal_t *journal, int *batch_count)
255 {
256         int i;
257
258         for (i = 0; i < *batch_count; i++)
259                 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE);
260
261         for (i = 0; i < *batch_count; i++) {
262                 struct buffer_head *bh = journal->j_chkpt_bhs[i];
263                 clear_buffer_jwrite(bh);
264                 BUFFER_TRACE(bh, "brelse");
265                 __brelse(bh);
266         }
267         *batch_count = 0;
268 }
269
270 /*
271  * Try to flush one buffer from the checkpoint list to disk.
272  *
273  * Return 1 if something happened which requires us to abort the current
274  * scan of the checkpoint list.  Return <0 if the buffer has failed to
275  * be written out.
276  *
277  * Called with j_list_lock held and drops it if 1 is returned
278  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
279  */
280 static int __process_buffer(journal_t *journal, struct journal_head *jh,
281                             int *batch_count, transaction_t *transaction)
282 {
283         struct buffer_head *bh = jh2bh(jh);
284         int ret = 0;
285
286         if (buffer_locked(bh)) {
287                 atomic_inc(&bh->b_count);
288                 spin_unlock(&journal->j_list_lock);
289                 jbd_unlock_bh_state(bh);
290                 wait_on_buffer(bh);
291                 /* the journal_head may have gone by now */
292                 BUFFER_TRACE(bh, "brelse");
293                 __brelse(bh);
294                 ret = 1;
295         } else if (jh->b_transaction != NULL) {
296                 transaction_t *t = jh->b_transaction;
297                 tid_t tid = t->t_tid;
298
299                 transaction->t_chp_stats.cs_forced_to_close++;
300                 spin_unlock(&journal->j_list_lock);
301                 jbd_unlock_bh_state(bh);
302                 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
303                         /*
304                          * The journal thread is dead; so starting and
305                          * waiting for a commit to finish will cause
306                          * us to wait for a _very_ long time.
307                          */
308                         printk(KERN_ERR "JBD2: %s: "
309                                "Waiting for Godot: block %llu\n",
310                                journal->j_devname,
311                                (unsigned long long) bh->b_blocknr);
312                 jbd2_log_start_commit(journal, tid);
313                 jbd2_log_wait_commit(journal, tid);
314                 ret = 1;
315         } else if (!buffer_dirty(bh)) {
316                 ret = 1;
317                 if (unlikely(buffer_write_io_error(bh)))
318                         ret = -EIO;
319                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
320                 BUFFER_TRACE(bh, "remove from checkpoint");
321                 __jbd2_journal_remove_checkpoint(jh);
322                 spin_unlock(&journal->j_list_lock);
323                 jbd_unlock_bh_state(bh);
324                 jbd2_journal_remove_journal_head(bh);
325                 __brelse(bh);
326         } else {
327                 /*
328                  * Important: we are about to write the buffer, and
329                  * possibly block, while still holding the journal lock.
330                  * We cannot afford to let the transaction logic start
331                  * messing around with this buffer before we write it to
332                  * disk, as that would break recoverability.
333                  */
334                 BUFFER_TRACE(bh, "queue");
335                 get_bh(bh);
336                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
337                 set_buffer_jwrite(bh);
338                 journal->j_chkpt_bhs[*batch_count] = bh;
339                 __buffer_relink_io(jh);
340                 jbd_unlock_bh_state(bh);
341                 transaction->t_chp_stats.cs_written++;
342                 (*batch_count)++;
343                 if (*batch_count == JBD2_NR_BATCH) {
344                         spin_unlock(&journal->j_list_lock);
345                         __flush_batch(journal, batch_count);
346                         ret = 1;
347                 }
348         }
349         return ret;
350 }
351
352 /*
353  * Perform an actual checkpoint. We take the first transaction on the
354  * list of transactions to be checkpointed and send all its buffers
355  * to disk. We submit larger chunks of data at once.
356  *
357  * The journal should be locked before calling this function.
358  * Called with j_checkpoint_mutex held.
359  */
360 int jbd2_log_do_checkpoint(journal_t *journal)
361 {
362         transaction_t *transaction;
363         tid_t this_tid;
364         int result;
365
366         jbd_debug(1, "Start checkpoint\n");
367
368         /*
369          * First thing: if there are any transactions in the log which
370          * don't need checkpointing, just eliminate them from the
371          * journal straight away.
372          */
373         result = jbd2_cleanup_journal_tail(journal);
374         trace_jbd2_checkpoint(journal, result);
375         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
376         if (result <= 0)
377                 return result;
378
379         /*
380          * OK, we need to start writing disk blocks.  Take one transaction
381          * and write it.
382          */
383         result = 0;
384         spin_lock(&journal->j_list_lock);
385         if (!journal->j_checkpoint_transactions)
386                 goto out;
387         transaction = journal->j_checkpoint_transactions;
388         if (transaction->t_chp_stats.cs_chp_time == 0)
389                 transaction->t_chp_stats.cs_chp_time = jiffies;
390         this_tid = transaction->t_tid;
391 restart:
392         /*
393          * If someone cleaned up this transaction while we slept, we're
394          * done (maybe it's a new transaction, but it fell at the same
395          * address).
396          */
397         if (journal->j_checkpoint_transactions == transaction &&
398                         transaction->t_tid == this_tid) {
399                 int batch_count = 0;
400                 struct journal_head *jh;
401                 int retry = 0, err;
402
403                 while (!retry && transaction->t_checkpoint_list) {
404                         struct buffer_head *bh;
405
406                         jh = transaction->t_checkpoint_list;
407                         bh = jh2bh(jh);
408                         if (!jbd_trylock_bh_state(bh)) {
409                                 jbd_sync_bh(journal, bh);
410                                 retry = 1;
411                                 break;
412                         }
413                         retry = __process_buffer(journal, jh, &batch_count,
414                                                  transaction);
415                         if (retry < 0 && !result)
416                                 result = retry;
417                         if (!retry && (need_resched() ||
418                                 spin_needbreak(&journal->j_list_lock))) {
419                                 spin_unlock(&journal->j_list_lock);
420                                 retry = 1;
421                                 break;
422                         }
423                 }
424
425                 if (batch_count) {
426                         if (!retry) {
427                                 spin_unlock(&journal->j_list_lock);
428                                 retry = 1;
429                         }
430                         __flush_batch(journal, &batch_count);
431                 }
432
433                 if (retry) {
434                         spin_lock(&journal->j_list_lock);
435                         goto restart;
436                 }
437                 /*
438                  * Now we have cleaned up the first transaction's checkpoint
439                  * list. Let's clean up the second one
440                  */
441                 err = __wait_cp_io(journal, transaction);
442                 if (!result)
443                         result = err;
444         }
445 out:
446         spin_unlock(&journal->j_list_lock);
447         if (result < 0)
448                 jbd2_journal_abort(journal, result);
449         else
450                 result = jbd2_cleanup_journal_tail(journal);
451
452         return (result < 0) ? result : 0;
453 }
454
455 /*
456  * Check the list of checkpoint transactions for the journal to see if
457  * we have already got rid of any since the last update of the log tail
458  * in the journal superblock.  If so, we can instantly roll the
459  * superblock forward to remove those transactions from the log.
460  *
461  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
462  *
463  * Called with the journal lock held.
464  *
465  * This is the only part of the journaling code which really needs to be
466  * aware of transaction aborts.  Checkpointing involves writing to the
467  * main filesystem area rather than to the journal, so it can proceed
468  * even in abort state, but we must not update the super block if
469  * checkpointing may have failed.  Otherwise, we would lose some metadata
470  * buffers which should be written-back to the filesystem.
471  */
472
473 int jbd2_cleanup_journal_tail(journal_t *journal)
474 {
475         transaction_t * transaction;
476         tid_t           first_tid;
477         unsigned long   blocknr, freed;
478
479         if (is_journal_aborted(journal))
480                 return 1;
481
482         /* OK, work out the oldest transaction remaining in the log, and
483          * the log block it starts at.
484          *
485          * If the log is now empty, we need to work out which is the
486          * next transaction ID we will write, and where it will
487          * start. */
488
489         write_lock(&journal->j_state_lock);
490         spin_lock(&journal->j_list_lock);
491         transaction = journal->j_checkpoint_transactions;
492         if (transaction) {
493                 first_tid = transaction->t_tid;
494                 blocknr = transaction->t_log_start;
495         } else if ((transaction = journal->j_committing_transaction) != NULL) {
496                 first_tid = transaction->t_tid;
497                 blocknr = transaction->t_log_start;
498         } else if ((transaction = journal->j_running_transaction) != NULL) {
499                 first_tid = transaction->t_tid;
500                 blocknr = journal->j_head;
501         } else {
502                 first_tid = journal->j_transaction_sequence;
503                 blocknr = journal->j_head;
504         }
505         spin_unlock(&journal->j_list_lock);
506         J_ASSERT(blocknr != 0);
507
508         /* If the oldest pinned transaction is at the tail of the log
509            already then there's not much we can do right now. */
510         if (journal->j_tail_sequence == first_tid) {
511                 write_unlock(&journal->j_state_lock);
512                 return 1;
513         }
514
515         /* OK, update the superblock to recover the freed space.
516          * Physical blocks come first: have we wrapped beyond the end of
517          * the log?  */
518         freed = blocknr - journal->j_tail;
519         if (blocknr < journal->j_tail)
520                 freed = freed + journal->j_last - journal->j_first;
521
522         trace_jbd2_cleanup_journal_tail(journal, first_tid, blocknr, freed);
523         jbd_debug(1,
524                   "Cleaning journal tail from %d to %d (offset %lu), "
525                   "freeing %lu\n",
526                   journal->j_tail_sequence, first_tid, blocknr, freed);
527
528         journal->j_free += freed;
529         journal->j_tail_sequence = first_tid;
530         journal->j_tail = blocknr;
531         write_unlock(&journal->j_state_lock);
532
533         /*
534          * If there is an external journal, we need to make sure that
535          * any data blocks that were recently written out --- perhaps
536          * by jbd2_log_do_checkpoint() --- are flushed out before we
537          * drop the transactions from the external journal.  It's
538          * unlikely this will be necessary, especially with a
539          * appropriately sized journal, but we need this to guarantee
540          * correctness.  Fortunately jbd2_cleanup_journal_tail()
541          * doesn't get called all that often.
542          */
543         if ((journal->j_fs_dev != journal->j_dev) &&
544             (journal->j_flags & JBD2_BARRIER))
545                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
546         if (!(journal->j_flags & JBD2_ABORT))
547                 jbd2_journal_update_superblock(journal, 1);
548         return 0;
549 }
550
551
552 /* Checkpoint list management */
553
554 /*
555  * journal_clean_one_cp_list
556  *
557  * Find all the written-back checkpoint buffers in the given list and release them.
558  *
559  * Called with the journal locked.
560  * Called with j_list_lock held.
561  * Returns number of bufers reaped (for debug)
562  */
563
564 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
565 {
566         struct journal_head *last_jh;
567         struct journal_head *next_jh = jh;
568         int ret, freed = 0;
569
570         *released = 0;
571         if (!jh)
572                 return 0;
573
574         last_jh = jh->b_cpprev;
575         do {
576                 jh = next_jh;
577                 next_jh = jh->b_cpnext;
578                 /* Use trylock because of the ranking */
579                 if (jbd_trylock_bh_state(jh2bh(jh))) {
580                         ret = __try_to_free_cp_buf(jh);
581                         if (ret) {
582                                 freed++;
583                                 if (ret == 2) {
584                                         *released = 1;
585                                         return freed;
586                                 }
587                         }
588                 }
589                 /*
590                  * This function only frees up some memory
591                  * if possible so we dont have an obligation
592                  * to finish processing. Bail out if preemption
593                  * requested:
594                  */
595                 if (need_resched())
596                         return freed;
597         } while (jh != last_jh);
598
599         return freed;
600 }
601
602 /*
603  * journal_clean_checkpoint_list
604  *
605  * Find all the written-back checkpoint buffers in the journal and release them.
606  *
607  * Called with the journal locked.
608  * Called with j_list_lock held.
609  * Returns number of buffers reaped (for debug)
610  */
611
612 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
613 {
614         transaction_t *transaction, *last_transaction, *next_transaction;
615         int ret = 0;
616         int released;
617
618         transaction = journal->j_checkpoint_transactions;
619         if (!transaction)
620                 goto out;
621
622         last_transaction = transaction->t_cpprev;
623         next_transaction = transaction;
624         do {
625                 transaction = next_transaction;
626                 next_transaction = transaction->t_cpnext;
627                 ret += journal_clean_one_cp_list(transaction->
628                                 t_checkpoint_list, &released);
629                 /*
630                  * This function only frees up some memory if possible so we
631                  * dont have an obligation to finish processing. Bail out if
632                  * preemption requested:
633                  */
634                 if (need_resched())
635                         goto out;
636                 if (released)
637                         continue;
638                 /*
639                  * It is essential that we are as careful as in the case of
640                  * t_checkpoint_list with removing the buffer from the list as
641                  * we can possibly see not yet submitted buffers on io_list
642                  */
643                 ret += journal_clean_one_cp_list(transaction->
644                                 t_checkpoint_io_list, &released);
645                 if (need_resched())
646                         goto out;
647         } while (transaction != last_transaction);
648 out:
649         return ret;
650 }
651
652 /*
653  * journal_remove_checkpoint: called after a buffer has been committed
654  * to disk (either by being write-back flushed to disk, or being
655  * committed to the log).
656  *
657  * We cannot safely clean a transaction out of the log until all of the
658  * buffer updates committed in that transaction have safely been stored
659  * elsewhere on disk.  To achieve this, all of the buffers in a
660  * transaction need to be maintained on the transaction's checkpoint
661  * lists until they have been rewritten, at which point this function is
662  * called to remove the buffer from the existing transaction's
663  * checkpoint lists.
664  *
665  * The function returns 1 if it frees the transaction, 0 otherwise.
666  *
667  * This function is called with the journal locked.
668  * This function is called with j_list_lock held.
669  * This function is called with jbd_lock_bh_state(jh2bh(jh))
670  */
671
672 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
673 {
674         struct transaction_chp_stats_s *stats;
675         transaction_t *transaction;
676         journal_t *journal;
677         int ret = 0;
678
679         JBUFFER_TRACE(jh, "entry");
680
681         if ((transaction = jh->b_cp_transaction) == NULL) {
682                 JBUFFER_TRACE(jh, "not on transaction");
683                 goto out;
684         }
685         journal = transaction->t_journal;
686
687         __buffer_unlink(jh);
688         jh->b_cp_transaction = NULL;
689
690         if (transaction->t_checkpoint_list != NULL ||
691             transaction->t_checkpoint_io_list != NULL)
692                 goto out;
693         JBUFFER_TRACE(jh, "transaction has no more buffers");
694
695         /*
696          * There is one special case to worry about: if we have just pulled the
697          * buffer off a running or committing transaction's checkpoing list,
698          * then even if the checkpoint list is empty, the transaction obviously
699          * cannot be dropped!
700          *
701          * The locking here around t_state is a bit sleazy.
702          * See the comment at the end of jbd2_journal_commit_transaction().
703          */
704         if (transaction->t_state != T_FINISHED) {
705                 JBUFFER_TRACE(jh, "belongs to running/committing transaction");
706                 goto out;
707         }
708
709         /* OK, that was the last buffer for the transaction: we can now
710            safely remove this transaction from the log */
711         stats = &transaction->t_chp_stats;
712         if (stats->cs_chp_time)
713                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
714                                                     jiffies);
715         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
716                                     transaction->t_tid, stats);
717
718         __jbd2_journal_drop_transaction(journal, transaction);
719         kfree(transaction);
720
721         /* Just in case anybody was waiting for more transactions to be
722            checkpointed... */
723         wake_up(&journal->j_wait_logspace);
724         ret = 1;
725 out:
726         JBUFFER_TRACE(jh, "exit");
727         return ret;
728 }
729
730 /*
731  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
732  * list so that we know when it is safe to clean the transaction out of
733  * the log.
734  *
735  * Called with the journal locked.
736  * Called with j_list_lock held.
737  */
738 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
739                                transaction_t *transaction)
740 {
741         JBUFFER_TRACE(jh, "entry");
742         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
743         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
744
745         jh->b_cp_transaction = transaction;
746
747         if (!transaction->t_checkpoint_list) {
748                 jh->b_cpnext = jh->b_cpprev = jh;
749         } else {
750                 jh->b_cpnext = transaction->t_checkpoint_list;
751                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
752                 jh->b_cpprev->b_cpnext = jh;
753                 jh->b_cpnext->b_cpprev = jh;
754         }
755         transaction->t_checkpoint_list = jh;
756 }
757
758 /*
759  * We've finished with this transaction structure: adios...
760  *
761  * The transaction must have no links except for the checkpoint by this
762  * point.
763  *
764  * Called with the journal locked.
765  * Called with j_list_lock held.
766  */
767
768 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
769 {
770         assert_spin_locked(&journal->j_list_lock);
771         if (transaction->t_cpnext) {
772                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
773                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
774                 if (journal->j_checkpoint_transactions == transaction)
775                         journal->j_checkpoint_transactions =
776                                 transaction->t_cpnext;
777                 if (journal->j_checkpoint_transactions == transaction)
778                         journal->j_checkpoint_transactions = NULL;
779         }
780
781         J_ASSERT(transaction->t_state == T_FINISHED);
782         J_ASSERT(transaction->t_buffers == NULL);
783         J_ASSERT(transaction->t_forget == NULL);
784         J_ASSERT(transaction->t_iobuf_list == NULL);
785         J_ASSERT(transaction->t_shadow_list == NULL);
786         J_ASSERT(transaction->t_log_list == NULL);
787         J_ASSERT(transaction->t_checkpoint_list == NULL);
788         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
789         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
790         J_ASSERT(journal->j_committing_transaction != transaction);
791         J_ASSERT(journal->j_running_transaction != transaction);
792
793         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
794 }