6f691c289c7144a5d9e107d2f5f0e34055cd8389
[kvmfornfv.git] / kernel / 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  */
92 static int __try_to_free_cp_buf(struct journal_head *jh)
93 {
94         int ret = 0;
95         struct buffer_head *bh = jh2bh(jh);
96
97         if (jh->b_transaction == NULL && !buffer_locked(bh) &&
98             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
99                 JBUFFER_TRACE(jh, "remove from checkpoint list");
100                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
101         }
102         return ret;
103 }
104
105 /*
106  * __jbd2_log_wait_for_space: wait until there is space in the journal.
107  *
108  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
109  * for a checkpoint to free up some space in the log.
110  */
111 void __jbd2_log_wait_for_space(journal_t *journal)
112 {
113         int nblocks, space_left;
114         /* assert_spin_locked(&journal->j_state_lock); */
115
116         nblocks = jbd2_space_needed(journal);
117         while (jbd2_log_space_left(journal) < nblocks) {
118                 write_unlock(&journal->j_state_lock);
119                 if (current->plug)
120                         io_schedule();
121                 mutex_lock(&journal->j_checkpoint_mutex);
122
123                 /*
124                  * Test again, another process may have checkpointed while we
125                  * were waiting for the checkpoint lock. If there are no
126                  * transactions ready to be checkpointed, try to recover
127                  * journal space by calling cleanup_journal_tail(), and if
128                  * that doesn't work, by waiting for the currently committing
129                  * transaction to complete.  If there is absolutely no way
130                  * to make progress, this is either a BUG or corrupted
131                  * filesystem, so abort the journal and leave a stack
132                  * trace for forensic evidence.
133                  */
134                 write_lock(&journal->j_state_lock);
135                 if (journal->j_flags & JBD2_ABORT) {
136                         mutex_unlock(&journal->j_checkpoint_mutex);
137                         return;
138                 }
139                 spin_lock(&journal->j_list_lock);
140                 nblocks = jbd2_space_needed(journal);
141                 space_left = jbd2_log_space_left(journal);
142                 if (space_left < nblocks) {
143                         int chkpt = journal->j_checkpoint_transactions != NULL;
144                         tid_t tid = 0;
145
146                         if (journal->j_committing_transaction)
147                                 tid = journal->j_committing_transaction->t_tid;
148                         spin_unlock(&journal->j_list_lock);
149                         write_unlock(&journal->j_state_lock);
150                         if (chkpt) {
151                                 jbd2_log_do_checkpoint(journal);
152                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
153                                 /* We were able to recover space; yay! */
154                                 ;
155                         } else if (tid) {
156                                 /*
157                                  * jbd2_journal_commit_transaction() may want
158                                  * to take the checkpoint_mutex if JBD2_FLUSHED
159                                  * is set.  So we need to temporarily drop it.
160                                  */
161                                 mutex_unlock(&journal->j_checkpoint_mutex);
162                                 jbd2_log_wait_commit(journal, tid);
163                                 write_lock(&journal->j_state_lock);
164                                 continue;
165                         } else {
166                                 printk(KERN_ERR "%s: needed %d blocks and "
167                                        "only had %d space available\n",
168                                        __func__, nblocks, space_left);
169                                 printk(KERN_ERR "%s: no way to get more "
170                                        "journal space in %s\n", __func__,
171                                        journal->j_devname);
172                                 WARN_ON(1);
173                                 jbd2_journal_abort(journal, 0);
174                         }
175                         write_lock(&journal->j_state_lock);
176                 } else {
177                         spin_unlock(&journal->j_list_lock);
178                 }
179                 mutex_unlock(&journal->j_checkpoint_mutex);
180         }
181 }
182
183 static void
184 __flush_batch(journal_t *journal, int *batch_count)
185 {
186         int i;
187         struct blk_plug plug;
188
189         blk_start_plug(&plug);
190         for (i = 0; i < *batch_count; i++)
191                 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE_SYNC);
192         blk_finish_plug(&plug);
193
194         for (i = 0; i < *batch_count; i++) {
195                 struct buffer_head *bh = journal->j_chkpt_bhs[i];
196                 BUFFER_TRACE(bh, "brelse");
197                 __brelse(bh);
198         }
199         *batch_count = 0;
200 }
201
202 /*
203  * Perform an actual checkpoint. We take the first transaction on the
204  * list of transactions to be checkpointed and send all its buffers
205  * to disk. We submit larger chunks of data at once.
206  *
207  * The journal should be locked before calling this function.
208  * Called with j_checkpoint_mutex held.
209  */
210 int jbd2_log_do_checkpoint(journal_t *journal)
211 {
212         struct journal_head     *jh;
213         struct buffer_head      *bh;
214         transaction_t           *transaction;
215         tid_t                   this_tid;
216         int                     result, batch_count = 0;
217
218         jbd_debug(1, "Start checkpoint\n");
219
220         /*
221          * First thing: if there are any transactions in the log which
222          * don't need checkpointing, just eliminate them from the
223          * journal straight away.
224          */
225         result = jbd2_cleanup_journal_tail(journal);
226         trace_jbd2_checkpoint(journal, result);
227         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
228         if (result <= 0)
229                 return result;
230
231         /*
232          * OK, we need to start writing disk blocks.  Take one transaction
233          * and write it.
234          */
235         result = 0;
236         spin_lock(&journal->j_list_lock);
237         if (!journal->j_checkpoint_transactions)
238                 goto out;
239         transaction = journal->j_checkpoint_transactions;
240         if (transaction->t_chp_stats.cs_chp_time == 0)
241                 transaction->t_chp_stats.cs_chp_time = jiffies;
242         this_tid = transaction->t_tid;
243 restart:
244         /*
245          * If someone cleaned up this transaction while we slept, we're
246          * done (maybe it's a new transaction, but it fell at the same
247          * address).
248          */
249         if (journal->j_checkpoint_transactions != transaction ||
250             transaction->t_tid != this_tid)
251                 goto out;
252
253         /* checkpoint all of the transaction's buffers */
254         while (transaction->t_checkpoint_list) {
255                 jh = transaction->t_checkpoint_list;
256                 bh = jh2bh(jh);
257
258                 if (buffer_locked(bh)) {
259                         spin_unlock(&journal->j_list_lock);
260                         get_bh(bh);
261                         wait_on_buffer(bh);
262                         /* the journal_head may have gone by now */
263                         BUFFER_TRACE(bh, "brelse");
264                         __brelse(bh);
265                         goto retry;
266                 }
267                 if (jh->b_transaction != NULL) {
268                         transaction_t *t = jh->b_transaction;
269                         tid_t tid = t->t_tid;
270
271                         transaction->t_chp_stats.cs_forced_to_close++;
272                         spin_unlock(&journal->j_list_lock);
273                         if (unlikely(journal->j_flags & JBD2_UNMOUNT))
274                                 /*
275                                  * The journal thread is dead; so
276                                  * starting and waiting for a commit
277                                  * to finish will cause us to wait for
278                                  * a _very_ long time.
279                                  */
280                                 printk(KERN_ERR
281                 "JBD2: %s: Waiting for Godot: block %llu\n",
282                 journal->j_devname, (unsigned long long) bh->b_blocknr);
283
284                         jbd2_log_start_commit(journal, tid);
285                         jbd2_log_wait_commit(journal, tid);
286                         goto retry;
287                 }
288                 if (!buffer_dirty(bh)) {
289                         if (unlikely(buffer_write_io_error(bh)) && !result)
290                                 result = -EIO;
291                         BUFFER_TRACE(bh, "remove from checkpoint");
292                         if (__jbd2_journal_remove_checkpoint(jh))
293                                 /* The transaction was released; we're done */
294                                 goto out;
295                         continue;
296                 }
297                 /*
298                  * Important: we are about to write the buffer, and
299                  * possibly block, while still holding the journal
300                  * lock.  We cannot afford to let the transaction
301                  * logic start messing around with this buffer before
302                  * we write it to disk, as that would break
303                  * recoverability.
304                  */
305                 BUFFER_TRACE(bh, "queue");
306                 get_bh(bh);
307                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
308                 journal->j_chkpt_bhs[batch_count++] = bh;
309                 __buffer_relink_io(jh);
310                 transaction->t_chp_stats.cs_written++;
311                 if ((batch_count == JBD2_NR_BATCH) ||
312                     need_resched() ||
313                     spin_needbreak(&journal->j_list_lock))
314                         goto unlock_and_flush;
315         }
316
317         if (batch_count) {
318                 unlock_and_flush:
319                         spin_unlock(&journal->j_list_lock);
320                 retry:
321                         if (batch_count)
322                                 __flush_batch(journal, &batch_count);
323                         spin_lock(&journal->j_list_lock);
324                         goto restart;
325         }
326
327         /*
328          * Now we issued all of the transaction's buffers, let's deal
329          * with the buffers that are out for I/O.
330          */
331 restart2:
332         /* Did somebody clean up the transaction in the meanwhile? */
333         if (journal->j_checkpoint_transactions != transaction ||
334             transaction->t_tid != this_tid)
335                 goto out;
336
337         while (transaction->t_checkpoint_io_list) {
338                 jh = transaction->t_checkpoint_io_list;
339                 bh = jh2bh(jh);
340                 if (buffer_locked(bh)) {
341                         spin_unlock(&journal->j_list_lock);
342                         get_bh(bh);
343                         wait_on_buffer(bh);
344                         /* the journal_head may have gone by now */
345                         BUFFER_TRACE(bh, "brelse");
346                         __brelse(bh);
347                         spin_lock(&journal->j_list_lock);
348                         goto restart2;
349                 }
350                 if (unlikely(buffer_write_io_error(bh)) && !result)
351                         result = -EIO;
352
353                 /*
354                  * Now in whatever state the buffer currently is, we
355                  * know that it has been written out and so we can
356                  * drop it from the list
357                  */
358                 if (__jbd2_journal_remove_checkpoint(jh))
359                         break;
360         }
361 out:
362         spin_unlock(&journal->j_list_lock);
363         if (result < 0)
364                 jbd2_journal_abort(journal, result);
365         else
366                 result = jbd2_cleanup_journal_tail(journal);
367
368         return (result < 0) ? result : 0;
369 }
370
371 /*
372  * Check the list of checkpoint transactions for the journal to see if
373  * we have already got rid of any since the last update of the log tail
374  * in the journal superblock.  If so, we can instantly roll the
375  * superblock forward to remove those transactions from the log.
376  *
377  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
378  *
379  * Called with the journal lock held.
380  *
381  * This is the only part of the journaling code which really needs to be
382  * aware of transaction aborts.  Checkpointing involves writing to the
383  * main filesystem area rather than to the journal, so it can proceed
384  * even in abort state, but we must not update the super block if
385  * checkpointing may have failed.  Otherwise, we would lose some metadata
386  * buffers which should be written-back to the filesystem.
387  */
388
389 int jbd2_cleanup_journal_tail(journal_t *journal)
390 {
391         tid_t           first_tid;
392         unsigned long   blocknr;
393
394         if (is_journal_aborted(journal))
395                 return 1;
396
397         if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
398                 return 1;
399         J_ASSERT(blocknr != 0);
400
401         /*
402          * We need to make sure that any blocks that were recently written out
403          * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
404          * we drop the transactions from the journal. It's unlikely this will
405          * be necessary, especially with an appropriately sized journal, but we
406          * need this to guarantee correctness.  Fortunately
407          * jbd2_cleanup_journal_tail() doesn't get called all that often.
408          */
409         if (journal->j_flags & JBD2_BARRIER)
410                 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
411
412         __jbd2_update_log_tail(journal, first_tid, blocknr);
413         return 0;
414 }
415
416
417 /* Checkpoint list management */
418
419 /*
420  * journal_clean_one_cp_list
421  *
422  * Find all the written-back checkpoint buffers in the given list and
423  * release them.
424  *
425  * Called with j_list_lock held.
426  * Returns 1 if we freed the transaction, 0 otherwise.
427  */
428 static int journal_clean_one_cp_list(struct journal_head *jh)
429 {
430         struct journal_head *last_jh;
431         struct journal_head *next_jh = jh;
432         int ret;
433         int freed = 0;
434
435         if (!jh)
436                 return 0;
437
438         last_jh = jh->b_cpprev;
439         do {
440                 jh = next_jh;
441                 next_jh = jh->b_cpnext;
442                 ret = __try_to_free_cp_buf(jh);
443                 if (!ret)
444                         return freed;
445                 if (ret == 2)
446                         return 1;
447                 freed = 1;
448                 /*
449                  * This function only frees up some memory
450                  * if possible so we dont have an obligation
451                  * to finish processing. Bail out if preemption
452                  * requested:
453                  */
454                 if (need_resched())
455                         return freed;
456         } while (jh != last_jh);
457
458         return freed;
459 }
460
461 /*
462  * journal_clean_checkpoint_list
463  *
464  * Find all the written-back checkpoint buffers in the journal and release them.
465  *
466  * Called with j_list_lock held.
467  */
468 void __jbd2_journal_clean_checkpoint_list(journal_t *journal)
469 {
470         transaction_t *transaction, *last_transaction, *next_transaction;
471         int ret;
472
473         transaction = journal->j_checkpoint_transactions;
474         if (!transaction)
475                 return;
476
477         last_transaction = transaction->t_cpprev;
478         next_transaction = transaction;
479         do {
480                 transaction = next_transaction;
481                 next_transaction = transaction->t_cpnext;
482                 ret = journal_clean_one_cp_list(transaction->t_checkpoint_list);
483                 /*
484                  * This function only frees up some memory if possible so we
485                  * dont have an obligation to finish processing. Bail out if
486                  * preemption requested:
487                  */
488                 if (need_resched())
489                         return;
490                 if (ret)
491                         continue;
492                 /*
493                  * It is essential that we are as careful as in the case of
494                  * t_checkpoint_list with removing the buffer from the list as
495                  * we can possibly see not yet submitted buffers on io_list
496                  */
497                 ret = journal_clean_one_cp_list(transaction->
498                                 t_checkpoint_io_list);
499                 if (need_resched())
500                         return;
501                 /*
502                  * Stop scanning if we couldn't free the transaction. This
503                  * avoids pointless scanning of transactions which still
504                  * weren't checkpointed.
505                  */
506                 if (!ret)
507                         return;
508         } while (transaction != last_transaction);
509 }
510
511 /*
512  * journal_remove_checkpoint: called after a buffer has been committed
513  * to disk (either by being write-back flushed to disk, or being
514  * committed to the log).
515  *
516  * We cannot safely clean a transaction out of the log until all of the
517  * buffer updates committed in that transaction have safely been stored
518  * elsewhere on disk.  To achieve this, all of the buffers in a
519  * transaction need to be maintained on the transaction's checkpoint
520  * lists until they have been rewritten, at which point this function is
521  * called to remove the buffer from the existing transaction's
522  * checkpoint lists.
523  *
524  * The function returns 1 if it frees the transaction, 0 otherwise.
525  * The function can free jh and bh.
526  *
527  * This function is called with j_list_lock held.
528  */
529 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
530 {
531         struct transaction_chp_stats_s *stats;
532         transaction_t *transaction;
533         journal_t *journal;
534         int ret = 0;
535
536         JBUFFER_TRACE(jh, "entry");
537
538         if ((transaction = jh->b_cp_transaction) == NULL) {
539                 JBUFFER_TRACE(jh, "not on transaction");
540                 goto out;
541         }
542         journal = transaction->t_journal;
543
544         JBUFFER_TRACE(jh, "removing from transaction");
545         __buffer_unlink(jh);
546         jh->b_cp_transaction = NULL;
547         jbd2_journal_put_journal_head(jh);
548
549         if (transaction->t_checkpoint_list != NULL ||
550             transaction->t_checkpoint_io_list != NULL)
551                 goto out;
552
553         /*
554          * There is one special case to worry about: if we have just pulled the
555          * buffer off a running or committing transaction's checkpoing list,
556          * then even if the checkpoint list is empty, the transaction obviously
557          * cannot be dropped!
558          *
559          * The locking here around t_state is a bit sleazy.
560          * See the comment at the end of jbd2_journal_commit_transaction().
561          */
562         if (transaction->t_state != T_FINISHED)
563                 goto out;
564
565         /* OK, that was the last buffer for the transaction: we can now
566            safely remove this transaction from the log */
567         stats = &transaction->t_chp_stats;
568         if (stats->cs_chp_time)
569                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
570                                                     jiffies);
571         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
572                                     transaction->t_tid, stats);
573
574         __jbd2_journal_drop_transaction(journal, transaction);
575         jbd2_journal_free_transaction(transaction);
576         ret = 1;
577 out:
578         return ret;
579 }
580
581 /*
582  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
583  * list so that we know when it is safe to clean the transaction out of
584  * the log.
585  *
586  * Called with the journal locked.
587  * Called with j_list_lock held.
588  */
589 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
590                                transaction_t *transaction)
591 {
592         JBUFFER_TRACE(jh, "entry");
593         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
594         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
595
596         /* Get reference for checkpointing transaction */
597         jbd2_journal_grab_journal_head(jh2bh(jh));
598         jh->b_cp_transaction = transaction;
599
600         if (!transaction->t_checkpoint_list) {
601                 jh->b_cpnext = jh->b_cpprev = jh;
602         } else {
603                 jh->b_cpnext = transaction->t_checkpoint_list;
604                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
605                 jh->b_cpprev->b_cpnext = jh;
606                 jh->b_cpnext->b_cpprev = jh;
607         }
608         transaction->t_checkpoint_list = jh;
609 }
610
611 /*
612  * We've finished with this transaction structure: adios...
613  *
614  * The transaction must have no links except for the checkpoint by this
615  * point.
616  *
617  * Called with the journal locked.
618  * Called with j_list_lock held.
619  */
620
621 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
622 {
623         assert_spin_locked(&journal->j_list_lock);
624         if (transaction->t_cpnext) {
625                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
626                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
627                 if (journal->j_checkpoint_transactions == transaction)
628                         journal->j_checkpoint_transactions =
629                                 transaction->t_cpnext;
630                 if (journal->j_checkpoint_transactions == transaction)
631                         journal->j_checkpoint_transactions = NULL;
632         }
633
634         J_ASSERT(transaction->t_state == T_FINISHED);
635         J_ASSERT(transaction->t_buffers == NULL);
636         J_ASSERT(transaction->t_forget == NULL);
637         J_ASSERT(transaction->t_shadow_list == NULL);
638         J_ASSERT(transaction->t_checkpoint_list == NULL);
639         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
640         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
641         J_ASSERT(journal->j_committing_transaction != transaction);
642         J_ASSERT(journal->j_running_transaction != transaction);
643
644         trace_jbd2_drop_transaction(journal, transaction);
645
646         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
647 }