Merge tag 'pstore-v4.20-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/kees...
[sfrench/cifs-2.6.git] / fs / jbd2 / checkpoint.c
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * linux/fs/jbd2/checkpoint.c
4  *
5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
6  *
7  * Copyright 1999 Red Hat Software --- All Rights Reserved
8  *
9  * Checkpoint routines for the generic filesystem journaling code.
10  * Part of the ext2fs journaling system.
11  *
12  * Checkpointing is the process of ensuring that a section of the log is
13  * committed fully to disk, so that that portion of the log can be
14  * reused.
15  */
16
17 #include <linux/time.h>
18 #include <linux/fs.h>
19 #include <linux/jbd2.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/blkdev.h>
23 #include <trace/events/jbd2.h>
24
25 /*
26  * Unlink a buffer from a transaction checkpoint list.
27  *
28  * Called with j_list_lock held.
29  */
30 static inline void __buffer_unlink_first(struct journal_head *jh)
31 {
32         transaction_t *transaction = jh->b_cp_transaction;
33
34         jh->b_cpnext->b_cpprev = jh->b_cpprev;
35         jh->b_cpprev->b_cpnext = jh->b_cpnext;
36         if (transaction->t_checkpoint_list == jh) {
37                 transaction->t_checkpoint_list = jh->b_cpnext;
38                 if (transaction->t_checkpoint_list == jh)
39                         transaction->t_checkpoint_list = NULL;
40         }
41 }
42
43 /*
44  * Unlink a buffer from a transaction checkpoint(io) list.
45  *
46  * Called with j_list_lock held.
47  */
48 static inline void __buffer_unlink(struct journal_head *jh)
49 {
50         transaction_t *transaction = jh->b_cp_transaction;
51
52         __buffer_unlink_first(jh);
53         if (transaction->t_checkpoint_io_list == jh) {
54                 transaction->t_checkpoint_io_list = jh->b_cpnext;
55                 if (transaction->t_checkpoint_io_list == jh)
56                         transaction->t_checkpoint_io_list = NULL;
57         }
58 }
59
60 /*
61  * Move a buffer from the checkpoint list to the checkpoint io list
62  *
63  * Called with j_list_lock held
64  */
65 static inline void __buffer_relink_io(struct journal_head *jh)
66 {
67         transaction_t *transaction = jh->b_cp_transaction;
68
69         __buffer_unlink_first(jh);
70
71         if (!transaction->t_checkpoint_io_list) {
72                 jh->b_cpnext = jh->b_cpprev = jh;
73         } else {
74                 jh->b_cpnext = transaction->t_checkpoint_io_list;
75                 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
76                 jh->b_cpprev->b_cpnext = jh;
77                 jh->b_cpnext->b_cpprev = jh;
78         }
79         transaction->t_checkpoint_io_list = jh;
80 }
81
82 /*
83  * Try to release a checkpointed buffer from its transaction.
84  * Returns 1 if we released it and 2 if we also released the
85  * whole transaction.
86  *
87  * Requires j_list_lock
88  */
89 static int __try_to_free_cp_buf(struct journal_head *jh)
90 {
91         int ret = 0;
92         struct buffer_head *bh = jh2bh(jh);
93
94         if (jh->b_transaction == NULL && !buffer_locked(bh) &&
95             !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
96                 JBUFFER_TRACE(jh, "remove from checkpoint list");
97                 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
98         }
99         return ret;
100 }
101
102 /*
103  * __jbd2_log_wait_for_space: wait until there is space in the journal.
104  *
105  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
106  * for a checkpoint to free up some space in the log.
107  */
108 void __jbd2_log_wait_for_space(journal_t *journal)
109 {
110         int nblocks, space_left;
111         /* assert_spin_locked(&journal->j_state_lock); */
112
113         nblocks = jbd2_space_needed(journal);
114         while (jbd2_log_space_left(journal) < nblocks) {
115                 write_unlock(&journal->j_state_lock);
116                 mutex_lock(&journal->j_checkpoint_mutex);
117
118                 /*
119                  * Test again, another process may have checkpointed while we
120                  * were waiting for the checkpoint lock. If there are no
121                  * transactions ready to be checkpointed, try to recover
122                  * journal space by calling cleanup_journal_tail(), and if
123                  * that doesn't work, by waiting for the currently committing
124                  * transaction to complete.  If there is absolutely no way
125                  * to make progress, this is either a BUG or corrupted
126                  * filesystem, so abort the journal and leave a stack
127                  * trace for forensic evidence.
128                  */
129                 write_lock(&journal->j_state_lock);
130                 if (journal->j_flags & JBD2_ABORT) {
131                         mutex_unlock(&journal->j_checkpoint_mutex);
132                         return;
133                 }
134                 spin_lock(&journal->j_list_lock);
135                 nblocks = jbd2_space_needed(journal);
136                 space_left = jbd2_log_space_left(journal);
137                 if (space_left < nblocks) {
138                         int chkpt = journal->j_checkpoint_transactions != NULL;
139                         tid_t tid = 0;
140
141                         if (journal->j_committing_transaction)
142                                 tid = journal->j_committing_transaction->t_tid;
143                         spin_unlock(&journal->j_list_lock);
144                         write_unlock(&journal->j_state_lock);
145                         if (chkpt) {
146                                 jbd2_log_do_checkpoint(journal);
147                         } else if (jbd2_cleanup_journal_tail(journal) == 0) {
148                                 /* We were able to recover space; yay! */
149                                 ;
150                         } else if (tid) {
151                                 /*
152                                  * jbd2_journal_commit_transaction() may want
153                                  * to take the checkpoint_mutex if JBD2_FLUSHED
154                                  * is set.  So we need to temporarily drop it.
155                                  */
156                                 mutex_unlock(&journal->j_checkpoint_mutex);
157                                 jbd2_log_wait_commit(journal, tid);
158                                 write_lock(&journal->j_state_lock);
159                                 continue;
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 static void
179 __flush_batch(journal_t *journal, int *batch_count)
180 {
181         int i;
182         struct blk_plug plug;
183
184         blk_start_plug(&plug);
185         for (i = 0; i < *batch_count; i++)
186                 write_dirty_buffer(journal->j_chkpt_bhs[i], REQ_SYNC);
187         blk_finish_plug(&plug);
188
189         for (i = 0; i < *batch_count; i++) {
190                 struct buffer_head *bh = journal->j_chkpt_bhs[i];
191                 BUFFER_TRACE(bh, "brelse");
192                 __brelse(bh);
193         }
194         *batch_count = 0;
195 }
196
197 /*
198  * Perform an actual checkpoint. We take the first transaction on the
199  * list of transactions to be checkpointed and send all its buffers
200  * to disk. We submit larger chunks of data at once.
201  *
202  * The journal should be locked before calling this function.
203  * Called with j_checkpoint_mutex held.
204  */
205 int jbd2_log_do_checkpoint(journal_t *journal)
206 {
207         struct journal_head     *jh;
208         struct buffer_head      *bh;
209         transaction_t           *transaction;
210         tid_t                   this_tid;
211         int                     result, batch_count = 0;
212
213         jbd_debug(1, "Start checkpoint\n");
214
215         /*
216          * First thing: if there are any transactions in the log which
217          * don't need checkpointing, just eliminate them from the
218          * journal straight away.
219          */
220         result = jbd2_cleanup_journal_tail(journal);
221         trace_jbd2_checkpoint(journal, result);
222         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
223         if (result <= 0)
224                 return result;
225
226         /*
227          * OK, we need to start writing disk blocks.  Take one transaction
228          * and write it.
229          */
230         result = 0;
231         spin_lock(&journal->j_list_lock);
232         if (!journal->j_checkpoint_transactions)
233                 goto out;
234         transaction = journal->j_checkpoint_transactions;
235         if (transaction->t_chp_stats.cs_chp_time == 0)
236                 transaction->t_chp_stats.cs_chp_time = jiffies;
237         this_tid = transaction->t_tid;
238 restart:
239         /*
240          * If someone cleaned up this transaction while we slept, we're
241          * done (maybe it's a new transaction, but it fell at the same
242          * address).
243          */
244         if (journal->j_checkpoint_transactions != transaction ||
245             transaction->t_tid != this_tid)
246                 goto out;
247
248         /* checkpoint all of the transaction's buffers */
249         while (transaction->t_checkpoint_list) {
250                 jh = transaction->t_checkpoint_list;
251                 bh = jh2bh(jh);
252
253                 if (buffer_locked(bh)) {
254                         get_bh(bh);
255                         spin_unlock(&journal->j_list_lock);
256                         wait_on_buffer(bh);
257                         /* the journal_head may have gone by now */
258                         BUFFER_TRACE(bh, "brelse");
259                         __brelse(bh);
260                         goto retry;
261                 }
262                 if (jh->b_transaction != NULL) {
263                         transaction_t *t = jh->b_transaction;
264                         tid_t tid = t->t_tid;
265
266                         transaction->t_chp_stats.cs_forced_to_close++;
267                         spin_unlock(&journal->j_list_lock);
268                         if (unlikely(journal->j_flags & JBD2_UNMOUNT))
269                                 /*
270                                  * The journal thread is dead; so
271                                  * starting and waiting for a commit
272                                  * to finish will cause us to wait for
273                                  * a _very_ long time.
274                                  */
275                                 printk(KERN_ERR
276                 "JBD2: %s: Waiting for Godot: block %llu\n",
277                 journal->j_devname, (unsigned long long) bh->b_blocknr);
278
279                         jbd2_log_start_commit(journal, tid);
280                         jbd2_log_wait_commit(journal, tid);
281                         goto retry;
282                 }
283                 if (!buffer_dirty(bh)) {
284                         if (unlikely(buffer_write_io_error(bh)) && !result)
285                                 result = -EIO;
286                         BUFFER_TRACE(bh, "remove from checkpoint");
287                         if (__jbd2_journal_remove_checkpoint(jh))
288                                 /* The transaction was released; we're done */
289                                 goto out;
290                         continue;
291                 }
292                 /*
293                  * Important: we are about to write the buffer, and
294                  * possibly block, while still holding the journal
295                  * lock.  We cannot afford to let the transaction
296                  * logic start messing around with this buffer before
297                  * we write it to disk, as that would break
298                  * recoverability.
299                  */
300                 BUFFER_TRACE(bh, "queue");
301                 get_bh(bh);
302                 J_ASSERT_BH(bh, !buffer_jwrite(bh));
303                 journal->j_chkpt_bhs[batch_count++] = bh;
304                 __buffer_relink_io(jh);
305                 transaction->t_chp_stats.cs_written++;
306                 if ((batch_count == JBD2_NR_BATCH) ||
307                     need_resched() ||
308                     spin_needbreak(&journal->j_list_lock))
309                         goto unlock_and_flush;
310         }
311
312         if (batch_count) {
313                 unlock_and_flush:
314                         spin_unlock(&journal->j_list_lock);
315                 retry:
316                         if (batch_count)
317                                 __flush_batch(journal, &batch_count);
318                         spin_lock(&journal->j_list_lock);
319                         goto restart;
320         }
321
322         /*
323          * Now we issued all of the transaction's buffers, let's deal
324          * with the buffers that are out for I/O.
325          */
326 restart2:
327         /* Did somebody clean up the transaction in the meanwhile? */
328         if (journal->j_checkpoint_transactions != transaction ||
329             transaction->t_tid != this_tid)
330                 goto out;
331
332         while (transaction->t_checkpoint_io_list) {
333                 jh = transaction->t_checkpoint_io_list;
334                 bh = jh2bh(jh);
335                 if (buffer_locked(bh)) {
336                         get_bh(bh);
337                         spin_unlock(&journal->j_list_lock);
338                         wait_on_buffer(bh);
339                         /* the journal_head may have gone by now */
340                         BUFFER_TRACE(bh, "brelse");
341                         __brelse(bh);
342                         spin_lock(&journal->j_list_lock);
343                         goto restart2;
344                 }
345                 if (unlikely(buffer_write_io_error(bh)) && !result)
346                         result = -EIO;
347
348                 /*
349                  * Now in whatever state the buffer currently is, we
350                  * know that it has been written out and so we can
351                  * drop it from the list
352                  */
353                 if (__jbd2_journal_remove_checkpoint(jh))
354                         break;
355         }
356 out:
357         spin_unlock(&journal->j_list_lock);
358         if (result < 0)
359                 jbd2_journal_abort(journal, result);
360         else
361                 result = jbd2_cleanup_journal_tail(journal);
362
363         return (result < 0) ? result : 0;
364 }
365
366 /*
367  * Check the list of checkpoint transactions for the journal to see if
368  * we have already got rid of any since the last update of the log tail
369  * in the journal superblock.  If so, we can instantly roll the
370  * superblock forward to remove those transactions from the log.
371  *
372  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
373  *
374  * Called with the journal lock held.
375  *
376  * This is the only part of the journaling code which really needs to be
377  * aware of transaction aborts.  Checkpointing involves writing to the
378  * main filesystem area rather than to the journal, so it can proceed
379  * even in abort state, but we must not update the super block if
380  * checkpointing may have failed.  Otherwise, we would lose some metadata
381  * buffers which should be written-back to the filesystem.
382  */
383
384 int jbd2_cleanup_journal_tail(journal_t *journal)
385 {
386         tid_t           first_tid;
387         unsigned long   blocknr;
388
389         if (is_journal_aborted(journal))
390                 return -EIO;
391
392         if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
393                 return 1;
394         J_ASSERT(blocknr != 0);
395
396         /*
397          * We need to make sure that any blocks that were recently written out
398          * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
399          * we drop the transactions from the journal. It's unlikely this will
400          * be necessary, especially with an appropriately sized journal, but we
401          * need this to guarantee correctness.  Fortunately
402          * jbd2_cleanup_journal_tail() doesn't get called all that often.
403          */
404         if (journal->j_flags & JBD2_BARRIER)
405                 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
406
407         return __jbd2_update_log_tail(journal, first_tid, blocknr);
408 }
409
410
411 /* Checkpoint list management */
412
413 /*
414  * journal_clean_one_cp_list
415  *
416  * Find all the written-back checkpoint buffers in the given list and
417  * release them. If 'destroy' is set, clean all buffers unconditionally.
418  *
419  * Called with j_list_lock held.
420  * Returns 1 if we freed the transaction, 0 otherwise.
421  */
422 static int journal_clean_one_cp_list(struct journal_head *jh, bool destroy)
423 {
424         struct journal_head *last_jh;
425         struct journal_head *next_jh = jh;
426         int ret;
427
428         if (!jh)
429                 return 0;
430
431         last_jh = jh->b_cpprev;
432         do {
433                 jh = next_jh;
434                 next_jh = jh->b_cpnext;
435                 if (!destroy)
436                         ret = __try_to_free_cp_buf(jh);
437                 else
438                         ret = __jbd2_journal_remove_checkpoint(jh) + 1;
439                 if (!ret)
440                         return 0;
441                 if (ret == 2)
442                         return 1;
443                 /*
444                  * This function only frees up some memory
445                  * if possible so we dont have an obligation
446                  * to finish processing. Bail out if preemption
447                  * requested:
448                  */
449                 if (need_resched())
450                         return 0;
451         } while (jh != last_jh);
452
453         return 0;
454 }
455
456 /*
457  * journal_clean_checkpoint_list
458  *
459  * Find all the written-back checkpoint buffers in the journal and release them.
460  * If 'destroy' is set, release all buffers unconditionally.
461  *
462  * Called with j_list_lock held.
463  */
464 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy)
465 {
466         transaction_t *transaction, *last_transaction, *next_transaction;
467         int ret;
468
469         transaction = journal->j_checkpoint_transactions;
470         if (!transaction)
471                 return;
472
473         last_transaction = transaction->t_cpprev;
474         next_transaction = transaction;
475         do {
476                 transaction = next_transaction;
477                 next_transaction = transaction->t_cpnext;
478                 ret = journal_clean_one_cp_list(transaction->t_checkpoint_list,
479                                                 destroy);
480                 /*
481                  * This function only frees up some memory if possible so we
482                  * dont have an obligation to finish processing. Bail out if
483                  * preemption requested:
484                  */
485                 if (need_resched())
486                         return;
487                 if (ret)
488                         continue;
489                 /*
490                  * It is essential that we are as careful as in the case of
491                  * t_checkpoint_list with removing the buffer from the list as
492                  * we can possibly see not yet submitted buffers on io_list
493                  */
494                 ret = journal_clean_one_cp_list(transaction->
495                                 t_checkpoint_io_list, destroy);
496                 if (need_resched())
497                         return;
498                 /*
499                  * Stop scanning if we couldn't free the transaction. This
500                  * avoids pointless scanning of transactions which still
501                  * weren't checkpointed.
502                  */
503                 if (!ret)
504                         return;
505         } while (transaction != last_transaction);
506 }
507
508 /*
509  * Remove buffers from all checkpoint lists as journal is aborted and we just
510  * need to free memory
511  */
512 void jbd2_journal_destroy_checkpoint(journal_t *journal)
513 {
514         /*
515          * We loop because __jbd2_journal_clean_checkpoint_list() may abort
516          * early due to a need of rescheduling.
517          */
518         while (1) {
519                 spin_lock(&journal->j_list_lock);
520                 if (!journal->j_checkpoint_transactions) {
521                         spin_unlock(&journal->j_list_lock);
522                         break;
523                 }
524                 __jbd2_journal_clean_checkpoint_list(journal, true);
525                 spin_unlock(&journal->j_list_lock);
526                 cond_resched();
527         }
528 }
529
530 /*
531  * journal_remove_checkpoint: called after a buffer has been committed
532  * to disk (either by being write-back flushed to disk, or being
533  * committed to the log).
534  *
535  * We cannot safely clean a transaction out of the log until all of the
536  * buffer updates committed in that transaction have safely been stored
537  * elsewhere on disk.  To achieve this, all of the buffers in a
538  * transaction need to be maintained on the transaction's checkpoint
539  * lists until they have been rewritten, at which point this function is
540  * called to remove the buffer from the existing transaction's
541  * checkpoint lists.
542  *
543  * The function returns 1 if it frees the transaction, 0 otherwise.
544  * The function can free jh and bh.
545  *
546  * This function is called with j_list_lock held.
547  */
548 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
549 {
550         struct transaction_chp_stats_s *stats;
551         transaction_t *transaction;
552         journal_t *journal;
553         int ret = 0;
554
555         JBUFFER_TRACE(jh, "entry");
556
557         if ((transaction = jh->b_cp_transaction) == NULL) {
558                 JBUFFER_TRACE(jh, "not on transaction");
559                 goto out;
560         }
561         journal = transaction->t_journal;
562
563         JBUFFER_TRACE(jh, "removing from transaction");
564         __buffer_unlink(jh);
565         jh->b_cp_transaction = NULL;
566         jbd2_journal_put_journal_head(jh);
567
568         if (transaction->t_checkpoint_list != NULL ||
569             transaction->t_checkpoint_io_list != NULL)
570                 goto out;
571
572         /*
573          * There is one special case to worry about: if we have just pulled the
574          * buffer off a running or committing transaction's checkpoing list,
575          * then even if the checkpoint list is empty, the transaction obviously
576          * cannot be dropped!
577          *
578          * The locking here around t_state is a bit sleazy.
579          * See the comment at the end of jbd2_journal_commit_transaction().
580          */
581         if (transaction->t_state != T_FINISHED)
582                 goto out;
583
584         /* OK, that was the last buffer for the transaction: we can now
585            safely remove this transaction from the log */
586         stats = &transaction->t_chp_stats;
587         if (stats->cs_chp_time)
588                 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
589                                                     jiffies);
590         trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
591                                     transaction->t_tid, stats);
592
593         __jbd2_journal_drop_transaction(journal, transaction);
594         jbd2_journal_free_transaction(transaction);
595         ret = 1;
596 out:
597         return ret;
598 }
599
600 /*
601  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
602  * list so that we know when it is safe to clean the transaction out of
603  * the log.
604  *
605  * Called with the journal locked.
606  * Called with j_list_lock held.
607  */
608 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
609                                transaction_t *transaction)
610 {
611         JBUFFER_TRACE(jh, "entry");
612         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
613         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
614
615         /* Get reference for checkpointing transaction */
616         jbd2_journal_grab_journal_head(jh2bh(jh));
617         jh->b_cp_transaction = transaction;
618
619         if (!transaction->t_checkpoint_list) {
620                 jh->b_cpnext = jh->b_cpprev = jh;
621         } else {
622                 jh->b_cpnext = transaction->t_checkpoint_list;
623                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
624                 jh->b_cpprev->b_cpnext = jh;
625                 jh->b_cpnext->b_cpprev = jh;
626         }
627         transaction->t_checkpoint_list = jh;
628 }
629
630 /*
631  * We've finished with this transaction structure: adios...
632  *
633  * The transaction must have no links except for the checkpoint by this
634  * point.
635  *
636  * Called with the journal locked.
637  * Called with j_list_lock held.
638  */
639
640 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
641 {
642         assert_spin_locked(&journal->j_list_lock);
643         if (transaction->t_cpnext) {
644                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
645                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
646                 if (journal->j_checkpoint_transactions == transaction)
647                         journal->j_checkpoint_transactions =
648                                 transaction->t_cpnext;
649                 if (journal->j_checkpoint_transactions == transaction)
650                         journal->j_checkpoint_transactions = NULL;
651         }
652
653         J_ASSERT(transaction->t_state == T_FINISHED);
654         J_ASSERT(transaction->t_buffers == NULL);
655         J_ASSERT(transaction->t_forget == NULL);
656         J_ASSERT(transaction->t_shadow_list == NULL);
657         J_ASSERT(transaction->t_checkpoint_list == NULL);
658         J_ASSERT(transaction->t_checkpoint_io_list == NULL);
659         J_ASSERT(atomic_read(&transaction->t_updates) == 0);
660         J_ASSERT(journal->j_committing_transaction != transaction);
661         J_ASSERT(journal->j_running_transaction != transaction);
662
663         trace_jbd2_drop_transaction(journal, transaction);
664
665         jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
666 }