Merge branch 'master' into gfs2
[pandora-kernel.git] / fs / jbd / checkpoint.c
1 /*
2  * linux/fs/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/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25
26 /*
27  * Unlink a buffer from a transaction checkpoint list.
28  *
29  * Called with j_list_lock held.
30  */
31 static inline void __buffer_unlink_first(struct journal_head *jh)
32 {
33         transaction_t *transaction = jh->b_cp_transaction;
34
35         jh->b_cpnext->b_cpprev = jh->b_cpprev;
36         jh->b_cpprev->b_cpnext = jh->b_cpnext;
37         if (transaction->t_checkpoint_list == jh) {
38                 transaction->t_checkpoint_list = jh->b_cpnext;
39                 if (transaction->t_checkpoint_list == jh)
40                         transaction->t_checkpoint_list = NULL;
41         }
42 }
43
44 /*
45  * Unlink a buffer from a transaction checkpoint(io) list.
46  *
47  * Called with j_list_lock held.
48  */
49 static inline void __buffer_unlink(struct journal_head *jh)
50 {
51         transaction_t *transaction = jh->b_cp_transaction;
52
53         __buffer_unlink_first(jh);
54         if (transaction->t_checkpoint_io_list == jh) {
55                 transaction->t_checkpoint_io_list = jh->b_cpnext;
56                 if (transaction->t_checkpoint_io_list == jh)
57                         transaction->t_checkpoint_io_list = NULL;
58         }
59 }
60
61 /*
62  * Move a buffer from the checkpoint list to the checkpoint io list
63  *
64  * Called with j_list_lock held
65  */
66 static inline void __buffer_relink_io(struct journal_head *jh)
67 {
68         transaction_t *transaction = jh->b_cp_transaction;
69
70         __buffer_unlink_first(jh);
71
72         if (!transaction->t_checkpoint_io_list) {
73                 jh->b_cpnext = jh->b_cpprev = jh;
74         } else {
75                 jh->b_cpnext = transaction->t_checkpoint_io_list;
76                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
77                 jh->b_cpprev->b_cpnext = jh;
78                 jh->b_cpnext->b_cpprev = jh;
79         }
80         transaction->t_checkpoint_io_list = jh;
81 }
82
83 /*
84  * Try to release a checkpointed buffer from its transaction.
85  * Returns 1 if we released it and 2 if we also released the
86  * whole transaction.
87  *
88  * Requires j_list_lock
89  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
90  */
91 static int __try_to_free_cp_buf(struct journal_head *jh)
92 {
93         int ret = 0;
94         struct buffer_head *bh = jh2bh(jh);
95
96         if (jh->b_jlist == BJ_None && !buffer_locked(bh) && !buffer_dirty(bh)) {
97                 JBUFFER_TRACE(jh, "remove from checkpoint list");
98                 ret = __journal_remove_checkpoint(jh) + 1;
99                 jbd_unlock_bh_state(bh);
100                 journal_remove_journal_head(bh);
101                 BUFFER_TRACE(bh, "release");
102                 __brelse(bh);
103         } else {
104                 jbd_unlock_bh_state(bh);
105         }
106         return ret;
107 }
108
109 /*
110  * __log_wait_for_space: wait until there is space in the journal.
111  *
112  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
113  * for a checkpoint to free up some space in the log.
114  */
115 void __log_wait_for_space(journal_t *journal)
116 {
117         int nblocks;
118         assert_spin_locked(&journal->j_state_lock);
119
120         nblocks = jbd_space_needed(journal);
121         while (__log_space_left(journal) < nblocks) {
122                 if (journal->j_flags & JFS_ABORT)
123                         return;
124                 spin_unlock(&journal->j_state_lock);
125                 mutex_lock(&journal->j_checkpoint_mutex);
126
127                 /*
128                  * Test again, another process may have checkpointed while we
129                  * were waiting for the checkpoint lock
130                  */
131                 spin_lock(&journal->j_state_lock);
132                 nblocks = jbd_space_needed(journal);
133                 if (__log_space_left(journal) < nblocks) {
134                         spin_unlock(&journal->j_state_lock);
135                         log_do_checkpoint(journal);
136                         spin_lock(&journal->j_state_lock);
137                 }
138                 mutex_unlock(&journal->j_checkpoint_mutex);
139         }
140 }
141
142 /*
143  * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
144  * The caller must restart a list walk.  Wait for someone else to run
145  * jbd_unlock_bh_state().
146  */
147 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
148         __releases(journal->j_list_lock)
149 {
150         get_bh(bh);
151         spin_unlock(&journal->j_list_lock);
152         jbd_lock_bh_state(bh);
153         jbd_unlock_bh_state(bh);
154         put_bh(bh);
155 }
156
157 /*
158  * Clean up transaction's list of buffers submitted for io.
159  * We wait for any pending IO to complete and remove any clean
160  * buffers. Note that we take the buffers in the opposite ordering
161  * from the one in which they were submitted for IO.
162  *
163  * Called with j_list_lock held.
164  */
165 static void __wait_cp_io(journal_t *journal, transaction_t *transaction)
166 {
167         struct journal_head *jh;
168         struct buffer_head *bh;
169         tid_t this_tid;
170         int released = 0;
171
172         this_tid = transaction->t_tid;
173 restart:
174         /* Did somebody clean up the transaction in the meanwhile? */
175         if (journal->j_checkpoint_transactions != transaction ||
176                         transaction->t_tid != this_tid)
177                 return;
178         while (!released && transaction->t_checkpoint_io_list) {
179                 jh = transaction->t_checkpoint_io_list;
180                 bh = jh2bh(jh);
181                 if (!jbd_trylock_bh_state(bh)) {
182                         jbd_sync_bh(journal, bh);
183                         spin_lock(&journal->j_list_lock);
184                         goto restart;
185                 }
186                 if (buffer_locked(bh)) {
187                         atomic_inc(&bh->b_count);
188                         spin_unlock(&journal->j_list_lock);
189                         jbd_unlock_bh_state(bh);
190                         wait_on_buffer(bh);
191                         /* the journal_head may have gone by now */
192                         BUFFER_TRACE(bh, "brelse");
193                         __brelse(bh);
194                         spin_lock(&journal->j_list_lock);
195                         goto restart;
196                 }
197                 /*
198                  * Now in whatever state the buffer currently is, we know that
199                  * it has been written out and so we can drop it from the list
200                  */
201                 released = __journal_remove_checkpoint(jh);
202                 jbd_unlock_bh_state(bh);
203                 journal_remove_journal_head(bh);
204                 __brelse(bh);
205         }
206 }
207
208 #define NR_BATCH        64
209
210 static void
211 __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
212 {
213         int i;
214
215         ll_rw_block(SWRITE, *batch_count, bhs);
216         for (i = 0; i < *batch_count; i++) {
217                 struct buffer_head *bh = bhs[i];
218                 clear_buffer_jwrite(bh);
219                 BUFFER_TRACE(bh, "brelse");
220                 __brelse(bh);
221         }
222         *batch_count = 0;
223 }
224
225 /*
226  * Try to flush one buffer from the checkpoint list to disk.
227  *
228  * Return 1 if something happened which requires us to abort the current
229  * scan of the checkpoint list.
230  *
231  * Called with j_list_lock held and drops it if 1 is returned
232  * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
233  */
234 static int __process_buffer(journal_t *journal, struct journal_head *jh,
235                         struct buffer_head **bhs, int *batch_count)
236 {
237         struct buffer_head *bh = jh2bh(jh);
238         int ret = 0;
239
240         if (buffer_locked(bh)) {
241                 atomic_inc(&bh->b_count);
242                 spin_unlock(&journal->j_list_lock);
243                 jbd_unlock_bh_state(bh);
244                 wait_on_buffer(bh);
245                 /* the journal_head may have gone by now */
246                 BUFFER_TRACE(bh, "brelse");
247                 __brelse(bh);
248                 ret = 1;
249         } else if (jh->b_transaction != NULL) {
250                 transaction_t *t = jh->b_transaction;
251                 tid_t tid = t->t_tid;
252
253                 spin_unlock(&journal->j_list_lock);
254                 jbd_unlock_bh_state(bh);
255                 log_start_commit(journal, tid);
256                 log_wait_commit(journal, tid);
257                 ret = 1;
258         } else if (!buffer_dirty(bh)) {
259                 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
260                 BUFFER_TRACE(bh, "remove from checkpoint");
261                 __journal_remove_checkpoint(jh);
262                 spin_unlock(&journal->j_list_lock);
263                 jbd_unlock_bh_state(bh);
264                 journal_remove_journal_head(bh);
265                 __brelse(bh);
266                 ret = 1;
267         } else {
268                 /*
269                  * Important: we are about to write the buffer, and
270                  * possibly block, while still holding the journal lock.
271                  * We cannot afford to let the transaction logic start
272                  * messing around with this buffer before we write it to
273                  * disk, as that would break recoverability.
274                  */
275                 BUFFER_TRACE(bh, "queue");
276                 get_bh(bh);
277                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
278                 set_buffer_jwrite(bh);
279                 bhs[*batch_count] = bh;
280                 __buffer_relink_io(jh);
281                 jbd_unlock_bh_state(bh);
282                 (*batch_count)++;
283                 if (*batch_count == NR_BATCH) {
284                         spin_unlock(&journal->j_list_lock);
285                         __flush_batch(journal, bhs, batch_count);
286                         ret = 1;
287                 }
288         }
289         return ret;
290 }
291
292 /*
293  * Perform an actual checkpoint. We take the first transaction on the
294  * list of transactions to be checkpointed and send all its buffers
295  * to disk. We submit larger chunks of data at once.
296  *
297  * The journal should be locked before calling this function.
298  */
299 int log_do_checkpoint(journal_t *journal)
300 {
301         transaction_t *transaction;
302         tid_t this_tid;
303         int result;
304
305         jbd_debug(1, "Start checkpoint\n");
306
307         /*
308          * First thing: if there are any transactions in the log which
309          * don't need checkpointing, just eliminate them from the
310          * journal straight away.
311          */
312         result = cleanup_journal_tail(journal);
313         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
314         if (result <= 0)
315                 return result;
316
317         /*
318          * OK, we need to start writing disk blocks.  Take one transaction
319          * and write it.
320          */
321         spin_lock(&journal->j_list_lock);
322         if (!journal->j_checkpoint_transactions)
323                 goto out;
324         transaction = journal->j_checkpoint_transactions;
325         this_tid = transaction->t_tid;
326 restart:
327         /*
328          * If someone cleaned up this transaction while we slept, we're
329          * done (maybe it's a new transaction, but it fell at the same
330          * address).
331          */
332         if (journal->j_checkpoint_transactions == transaction &&
333                         transaction->t_tid == this_tid) {
334                 int batch_count = 0;
335                 struct buffer_head *bhs[NR_BATCH];
336                 struct journal_head *jh;
337                 int retry = 0;
338
339                 while (!retry && transaction->t_checkpoint_list) {
340                         struct buffer_head *bh;
341
342                         jh = transaction->t_checkpoint_list;
343                         bh = jh2bh(jh);
344                         if (!jbd_trylock_bh_state(bh)) {
345                                 jbd_sync_bh(journal, bh);
346                                 retry = 1;
347                                 break;
348                         }
349                         retry = __process_buffer(journal, jh, bhs,&batch_count);
350                         if (!retry && lock_need_resched(&journal->j_list_lock)){
351                                 spin_unlock(&journal->j_list_lock);
352                                 retry = 1;
353                                 break;
354                         }
355                 }
356
357                 if (batch_count) {
358                         if (!retry) {
359                                 spin_unlock(&journal->j_list_lock);
360                                 retry = 1;
361                         }
362                         __flush_batch(journal, bhs, &batch_count);
363                 }
364
365                 if (retry) {
366                         spin_lock(&journal->j_list_lock);
367                         goto restart;
368                 }
369                 /*
370                  * Now we have cleaned up the first transaction's checkpoint
371                  * list. Let's clean up the second one
372                  */
373                 __wait_cp_io(journal, transaction);
374         }
375 out:
376         spin_unlock(&journal->j_list_lock);
377         result = cleanup_journal_tail(journal);
378         if (result < 0)
379                 return result;
380         return 0;
381 }
382
383 /*
384  * Check the list of checkpoint transactions for the journal to see if
385  * we have already got rid of any since the last update of the log tail
386  * in the journal superblock.  If so, we can instantly roll the
387  * superblock forward to remove those transactions from the log.
388  *
389  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
390  *
391  * Called with the journal lock held.
392  *
393  * This is the only part of the journaling code which really needs to be
394  * aware of transaction aborts.  Checkpointing involves writing to the
395  * main filesystem area rather than to the journal, so it can proceed
396  * even in abort state, but we must not update the journal superblock if
397  * we have an abort error outstanding.
398  */
399
400 int cleanup_journal_tail(journal_t *journal)
401 {
402         transaction_t * transaction;
403         tid_t           first_tid;
404         unsigned long   blocknr, freed;
405
406         /* OK, work out the oldest transaction remaining in the log, and
407          * the log block it starts at.
408          *
409          * If the log is now empty, we need to work out which is the
410          * next transaction ID we will write, and where it will
411          * start. */
412
413         spin_lock(&journal->j_state_lock);
414         spin_lock(&journal->j_list_lock);
415         transaction = journal->j_checkpoint_transactions;
416         if (transaction) {
417                 first_tid = transaction->t_tid;
418                 blocknr = transaction->t_log_start;
419         } else if ((transaction = journal->j_committing_transaction) != NULL) {
420                 first_tid = transaction->t_tid;
421                 blocknr = transaction->t_log_start;
422         } else if ((transaction = journal->j_running_transaction) != NULL) {
423                 first_tid = transaction->t_tid;
424                 blocknr = journal->j_head;
425         } else {
426                 first_tid = journal->j_transaction_sequence;
427                 blocknr = journal->j_head;
428         }
429         spin_unlock(&journal->j_list_lock);
430         J_ASSERT(blocknr != 0);
431
432         /* If the oldest pinned transaction is at the tail of the log
433            already then there's not much we can do right now. */
434         if (journal->j_tail_sequence == first_tid) {
435                 spin_unlock(&journal->j_state_lock);
436                 return 1;
437         }
438
439         /* OK, update the superblock to recover the freed space.
440          * Physical blocks come first: have we wrapped beyond the end of
441          * the log?  */
442         freed = blocknr - journal->j_tail;
443         if (blocknr < journal->j_tail)
444                 freed = freed + journal->j_last - journal->j_first;
445
446         jbd_debug(1,
447                   "Cleaning journal tail from %d to %d (offset %lu), "
448                   "freeing %lu\n",
449                   journal->j_tail_sequence, first_tid, blocknr, freed);
450
451         journal->j_free += freed;
452         journal->j_tail_sequence = first_tid;
453         journal->j_tail = blocknr;
454         spin_unlock(&journal->j_state_lock);
455         if (!(journal->j_flags & JFS_ABORT))
456                 journal_update_superblock(journal, 1);
457         return 0;
458 }
459
460
461 /* Checkpoint list management */
462
463 /*
464  * journal_clean_one_cp_list
465  *
466  * Find all the written-back checkpoint buffers in the given list and release them.
467  *
468  * Called with the journal locked.
469  * Called with j_list_lock held.
470  * Returns number of bufers reaped (for debug)
471  */
472
473 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
474 {
475         struct journal_head *last_jh;
476         struct journal_head *next_jh = jh;
477         int ret, freed = 0;
478
479         *released = 0;
480         if (!jh)
481                 return 0;
482
483         last_jh = jh->b_cpprev;
484         do {
485                 jh = next_jh;
486                 next_jh = jh->b_cpnext;
487                 /* Use trylock because of the ranking */
488                 if (jbd_trylock_bh_state(jh2bh(jh))) {
489                         ret = __try_to_free_cp_buf(jh);
490                         if (ret) {
491                                 freed++;
492                                 if (ret == 2) {
493                                         *released = 1;
494                                         return freed;
495                                 }
496                         }
497                 }
498                 /*
499                  * This function only frees up some memory
500                  * if possible so we dont have an obligation
501                  * to finish processing. Bail out if preemption
502                  * requested:
503                  */
504                 if (need_resched())
505                         return freed;
506         } while (jh != last_jh);
507
508         return freed;
509 }
510
511 /*
512  * journal_clean_checkpoint_list
513  *
514  * Find all the written-back checkpoint buffers in the journal and release them.
515  *
516  * Called with the journal locked.
517  * Called with j_list_lock held.
518  * Returns number of buffers reaped (for debug)
519  */
520
521 int __journal_clean_checkpoint_list(journal_t *journal)
522 {
523         transaction_t *transaction, *last_transaction, *next_transaction;
524         int ret = 0;
525         int released;
526
527         transaction = journal->j_checkpoint_transactions;
528         if (!transaction)
529                 goto out;
530
531         last_transaction = transaction->t_cpprev;
532         next_transaction = transaction;
533         do {
534                 transaction = next_transaction;
535                 next_transaction = transaction->t_cpnext;
536                 ret += journal_clean_one_cp_list(transaction->
537                                 t_checkpoint_list, &released);
538                 /*
539                  * This function only frees up some memory if possible so we
540                  * dont have an obligation to finish processing. Bail out if
541                  * preemption requested:
542                  */
543                 if (need_resched())
544                         goto out;
545                 if (released)
546                         continue;
547                 /*
548                  * It is essential that we are as careful as in the case of
549                  * t_checkpoint_list with removing the buffer from the list as
550                  * we can possibly see not yet submitted buffers on io_list
551                  */
552                 ret += journal_clean_one_cp_list(transaction->
553                                 t_checkpoint_io_list, &released);
554                 if (need_resched())
555                         goto out;
556         } while (transaction != last_transaction);
557 out:
558         return ret;
559 }
560
561 /*
562  * journal_remove_checkpoint: called after a buffer has been committed
563  * to disk (either by being write-back flushed to disk, or being
564  * committed to the log).
565  *
566  * We cannot safely clean a transaction out of the log until all of the
567  * buffer updates committed in that transaction have safely been stored
568  * elsewhere on disk.  To achieve this, all of the buffers in a
569  * transaction need to be maintained on the transaction's checkpoint
570  * lists until they have been rewritten, at which point this function is
571  * called to remove the buffer from the existing transaction's
572  * checkpoint lists.
573  *
574  * The function returns 1 if it frees the transaction, 0 otherwise.
575  *
576  * This function is called with the journal locked.
577  * This function is called with j_list_lock held.
578  * This function is called with jbd_lock_bh_state(jh2bh(jh))
579  */
580
581 int __journal_remove_checkpoint(struct journal_head *jh)
582 {
583         transaction_t *transaction;
584         journal_t *journal;
585         int ret = 0;
586
587         JBUFFER_TRACE(jh, "entry");
588
589         if ((transaction = jh->b_cp_transaction) == NULL) {
590                 JBUFFER_TRACE(jh, "not on transaction");
591                 goto out;
592         }
593         journal = transaction->t_journal;
594
595         __buffer_unlink(jh);
596         jh->b_cp_transaction = NULL;
597
598         if (transaction->t_checkpoint_list != NULL ||
599             transaction->t_checkpoint_io_list != NULL)
600                 goto out;
601         JBUFFER_TRACE(jh, "transaction has no more buffers");
602
603         /*
604          * There is one special case to worry about: if we have just pulled the
605          * buffer off a committing transaction's forget list, then even if the
606          * checkpoint list is empty, the transaction obviously cannot be
607          * dropped!
608          *
609          * The locking here around j_committing_transaction is a bit sleazy.
610          * See the comment at the end of journal_commit_transaction().
611          */
612         if (transaction == journal->j_committing_transaction) {
613                 JBUFFER_TRACE(jh, "belongs to committing transaction");
614                 goto out;
615         }
616
617         /* OK, that was the last buffer for the transaction: we can now
618            safely remove this transaction from the log */
619
620         __journal_drop_transaction(journal, transaction);
621
622         /* Just in case anybody was waiting for more transactions to be
623            checkpointed... */
624         wake_up(&journal->j_wait_logspace);
625         ret = 1;
626 out:
627         JBUFFER_TRACE(jh, "exit");
628         return ret;
629 }
630
631 /*
632  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
633  * list so that we know when it is safe to clean the transaction out of
634  * the log.
635  *
636  * Called with the journal locked.
637  * Called with j_list_lock held.
638  */
639 void __journal_insert_checkpoint(struct journal_head *jh,
640                                transaction_t *transaction)
641 {
642         JBUFFER_TRACE(jh, "entry");
643         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
644         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
645
646         jh->b_cp_transaction = transaction;
647
648         if (!transaction->t_checkpoint_list) {
649                 jh->b_cpnext = jh->b_cpprev = jh;
650         } else {
651                 jh->b_cpnext = transaction->t_checkpoint_list;
652                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
653                 jh->b_cpprev->b_cpnext = jh;
654                 jh->b_cpnext->b_cpprev = jh;
655         }
656         transaction->t_checkpoint_list = jh;
657 }
658
659 /*
660  * We've finished with this transaction structure: adios...
661  *
662  * The transaction must have no links except for the checkpoint by this
663  * point.
664  *
665  * Called with the journal locked.
666  * Called with j_list_lock held.
667  */
668
669 void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
670 {
671         assert_spin_locked(&journal->j_list_lock);
672         if (transaction->t_cpnext) {
673                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
674                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
675                 if (journal->j_checkpoint_transactions == transaction)
676                         journal->j_checkpoint_transactions =
677                                 transaction->t_cpnext;
678                 if (journal->j_checkpoint_transactions == transaction)
679                         journal->j_checkpoint_transactions = NULL;
680         }
681
682         J_ASSERT(transaction->t_state == T_FINISHED);
683         J_ASSERT(transaction->t_buffers == NULL);
684         J_ASSERT(transaction->t_sync_datalist == NULL);
685         J_ASSERT(transaction->t_forget == NULL);
686         J_ASSERT(transaction->t_iobuf_list == NULL);
687         J_ASSERT(transaction->t_shadow_list == NULL);
688         J_ASSERT(transaction->t_log_list == NULL);
689         J_ASSERT(transaction->t_checkpoint_list == NULL);
690         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
691         J_ASSERT(transaction->t_updates == 0);
692         J_ASSERT(journal->j_committing_transaction != transaction);
693         J_ASSERT(journal->j_running_transaction != transaction);
694
695         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
696         kfree(transaction);
697 }