ext4: avoid hang when mounting non-journal filesystems with orphan list
[~shefty/rdma-dev.git] / fs / jbd2 / transaction.c
1 /*
2  * linux/fs/jbd2/transaction.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem transaction handling code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages transactions (compound commits managed by the
16  * journaling code) and handles (individual atomic operations by the
17  * filesystem).
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/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29 #include <linux/backing-dev.h>
30 #include <linux/bug.h>
31 #include <linux/module.h>
32
33 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
35
36 static struct kmem_cache *transaction_cache;
37 int __init jbd2_journal_init_transaction_cache(void)
38 {
39         J_ASSERT(!transaction_cache);
40         transaction_cache = kmem_cache_create("jbd2_transaction_s",
41                                         sizeof(transaction_t),
42                                         0,
43                                         SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
44                                         NULL);
45         if (transaction_cache)
46                 return 0;
47         return -ENOMEM;
48 }
49
50 void jbd2_journal_destroy_transaction_cache(void)
51 {
52         if (transaction_cache) {
53                 kmem_cache_destroy(transaction_cache);
54                 transaction_cache = NULL;
55         }
56 }
57
58 void jbd2_journal_free_transaction(transaction_t *transaction)
59 {
60         if (unlikely(ZERO_OR_NULL_PTR(transaction)))
61                 return;
62         kmem_cache_free(transaction_cache, transaction);
63 }
64
65 /*
66  * jbd2_get_transaction: obtain a new transaction_t object.
67  *
68  * Simply allocate and initialise a new transaction.  Create it in
69  * RUNNING state and add it to the current journal (which should not
70  * have an existing running transaction: we only make a new transaction
71  * once we have started to commit the old one).
72  *
73  * Preconditions:
74  *      The journal MUST be locked.  We don't perform atomic mallocs on the
75  *      new transaction and we can't block without protecting against other
76  *      processes trying to touch the journal while it is in transition.
77  *
78  */
79
80 static transaction_t *
81 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
82 {
83         transaction->t_journal = journal;
84         transaction->t_state = T_RUNNING;
85         transaction->t_start_time = ktime_get();
86         transaction->t_tid = journal->j_transaction_sequence++;
87         transaction->t_expires = jiffies + journal->j_commit_interval;
88         spin_lock_init(&transaction->t_handle_lock);
89         atomic_set(&transaction->t_updates, 0);
90         atomic_set(&transaction->t_outstanding_credits, 0);
91         atomic_set(&transaction->t_handle_count, 0);
92         INIT_LIST_HEAD(&transaction->t_inode_list);
93         INIT_LIST_HEAD(&transaction->t_private_list);
94
95         /* Set up the commit timer for the new transaction. */
96         journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
97         add_timer(&journal->j_commit_timer);
98
99         J_ASSERT(journal->j_running_transaction == NULL);
100         journal->j_running_transaction = transaction;
101         transaction->t_max_wait = 0;
102         transaction->t_start = jiffies;
103
104         return transaction;
105 }
106
107 /*
108  * Handle management.
109  *
110  * A handle_t is an object which represents a single atomic update to a
111  * filesystem, and which tracks all of the modifications which form part
112  * of that one update.
113  */
114
115 /*
116  * Update transaction's maximum wait time, if debugging is enabled.
117  *
118  * In order for t_max_wait to be reliable, it must be protected by a
119  * lock.  But doing so will mean that start_this_handle() can not be
120  * run in parallel on SMP systems, which limits our scalability.  So
121  * unless debugging is enabled, we no longer update t_max_wait, which
122  * means that maximum wait time reported by the jbd2_run_stats
123  * tracepoint will always be zero.
124  */
125 static inline void update_t_max_wait(transaction_t *transaction,
126                                      unsigned long ts)
127 {
128 #ifdef CONFIG_JBD2_DEBUG
129         if (jbd2_journal_enable_debug &&
130             time_after(transaction->t_start, ts)) {
131                 ts = jbd2_time_diff(ts, transaction->t_start);
132                 spin_lock(&transaction->t_handle_lock);
133                 if (ts > transaction->t_max_wait)
134                         transaction->t_max_wait = ts;
135                 spin_unlock(&transaction->t_handle_lock);
136         }
137 #endif
138 }
139
140 /*
141  * start_this_handle: Given a handle, deal with any locking or stalling
142  * needed to make sure that there is enough journal space for the handle
143  * to begin.  Attach the handle to a transaction and set up the
144  * transaction's buffer credits.
145  */
146
147 static int start_this_handle(journal_t *journal, handle_t *handle,
148                              gfp_t gfp_mask)
149 {
150         transaction_t   *transaction, *new_transaction = NULL;
151         tid_t           tid;
152         int             needed, need_to_start;
153         int             nblocks = handle->h_buffer_credits;
154         unsigned long ts = jiffies;
155
156         if (nblocks > journal->j_max_transaction_buffers) {
157                 printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
158                        current->comm, nblocks,
159                        journal->j_max_transaction_buffers);
160                 return -ENOSPC;
161         }
162
163 alloc_transaction:
164         if (!journal->j_running_transaction) {
165                 new_transaction = kmem_cache_zalloc(transaction_cache,
166                                                     gfp_mask);
167                 if (!new_transaction) {
168                         /*
169                          * If __GFP_FS is not present, then we may be
170                          * being called from inside the fs writeback
171                          * layer, so we MUST NOT fail.  Since
172                          * __GFP_NOFAIL is going away, we will arrange
173                          * to retry the allocation ourselves.
174                          */
175                         if ((gfp_mask & __GFP_FS) == 0) {
176                                 congestion_wait(BLK_RW_ASYNC, HZ/50);
177                                 goto alloc_transaction;
178                         }
179                         return -ENOMEM;
180                 }
181         }
182
183         jbd_debug(3, "New handle %p going live.\n", handle);
184
185         /*
186          * We need to hold j_state_lock until t_updates has been incremented,
187          * for proper journal barrier handling
188          */
189 repeat:
190         read_lock(&journal->j_state_lock);
191         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
192         if (is_journal_aborted(journal) ||
193             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
194                 read_unlock(&journal->j_state_lock);
195                 jbd2_journal_free_transaction(new_transaction);
196                 return -EROFS;
197         }
198
199         /* Wait on the journal's transaction barrier if necessary */
200         if (journal->j_barrier_count) {
201                 read_unlock(&journal->j_state_lock);
202                 wait_event(journal->j_wait_transaction_locked,
203                                 journal->j_barrier_count == 0);
204                 goto repeat;
205         }
206
207         if (!journal->j_running_transaction) {
208                 read_unlock(&journal->j_state_lock);
209                 if (!new_transaction)
210                         goto alloc_transaction;
211                 write_lock(&journal->j_state_lock);
212                 if (!journal->j_running_transaction &&
213                     !journal->j_barrier_count) {
214                         jbd2_get_transaction(journal, new_transaction);
215                         new_transaction = NULL;
216                 }
217                 write_unlock(&journal->j_state_lock);
218                 goto repeat;
219         }
220
221         transaction = journal->j_running_transaction;
222
223         /*
224          * If the current transaction is locked down for commit, wait for the
225          * lock to be released.
226          */
227         if (transaction->t_state == T_LOCKED) {
228                 DEFINE_WAIT(wait);
229
230                 prepare_to_wait(&journal->j_wait_transaction_locked,
231                                         &wait, TASK_UNINTERRUPTIBLE);
232                 read_unlock(&journal->j_state_lock);
233                 schedule();
234                 finish_wait(&journal->j_wait_transaction_locked, &wait);
235                 goto repeat;
236         }
237
238         /*
239          * If there is not enough space left in the log to write all potential
240          * buffers requested by this operation, we need to stall pending a log
241          * checkpoint to free some more log space.
242          */
243         needed = atomic_add_return(nblocks,
244                                    &transaction->t_outstanding_credits);
245
246         if (needed > journal->j_max_transaction_buffers) {
247                 /*
248                  * If the current transaction is already too large, then start
249                  * to commit it: we can then go back and attach this handle to
250                  * a new transaction.
251                  */
252                 DEFINE_WAIT(wait);
253
254                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
255                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
256                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
257                                 TASK_UNINTERRUPTIBLE);
258                 tid = transaction->t_tid;
259                 need_to_start = !tid_geq(journal->j_commit_request, tid);
260                 read_unlock(&journal->j_state_lock);
261                 if (need_to_start)
262                         jbd2_log_start_commit(journal, tid);
263                 schedule();
264                 finish_wait(&journal->j_wait_transaction_locked, &wait);
265                 goto repeat;
266         }
267
268         /*
269          * The commit code assumes that it can get enough log space
270          * without forcing a checkpoint.  This is *critical* for
271          * correctness: a checkpoint of a buffer which is also
272          * associated with a committing transaction creates a deadlock,
273          * so commit simply cannot force through checkpoints.
274          *
275          * We must therefore ensure the necessary space in the journal
276          * *before* starting to dirty potentially checkpointed buffers
277          * in the new transaction.
278          *
279          * The worst part is, any transaction currently committing can
280          * reduce the free space arbitrarily.  Be careful to account for
281          * those buffers when checkpointing.
282          */
283
284         /*
285          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
286          * a _lot_ of headroom: 1/4 of the journal plus the size of
287          * the committing transaction.  Really, we only need to give it
288          * committing_transaction->t_outstanding_credits plus "enough" for
289          * the log control blocks.
290          * Also, this test is inconsistent with the matching one in
291          * jbd2_journal_extend().
292          */
293         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
294                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
295                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
296                 read_unlock(&journal->j_state_lock);
297                 write_lock(&journal->j_state_lock);
298                 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
299                         __jbd2_log_wait_for_space(journal);
300                 write_unlock(&journal->j_state_lock);
301                 goto repeat;
302         }
303
304         /* OK, account for the buffers that this operation expects to
305          * use and add the handle to the running transaction. 
306          */
307         update_t_max_wait(transaction, ts);
308         handle->h_transaction = transaction;
309         atomic_inc(&transaction->t_updates);
310         atomic_inc(&transaction->t_handle_count);
311         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
312                   handle, nblocks,
313                   atomic_read(&transaction->t_outstanding_credits),
314                   __jbd2_log_space_left(journal));
315         read_unlock(&journal->j_state_lock);
316
317         lock_map_acquire(&handle->h_lockdep_map);
318         jbd2_journal_free_transaction(new_transaction);
319         return 0;
320 }
321
322 static struct lock_class_key jbd2_handle_key;
323
324 /* Allocate a new handle.  This should probably be in a slab... */
325 static handle_t *new_handle(int nblocks)
326 {
327         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
328         if (!handle)
329                 return NULL;
330         memset(handle, 0, sizeof(*handle));
331         handle->h_buffer_credits = nblocks;
332         handle->h_ref = 1;
333
334         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
335                                                 &jbd2_handle_key, 0);
336
337         return handle;
338 }
339
340 /**
341  * handle_t *jbd2_journal_start() - Obtain a new handle.
342  * @journal: Journal to start transaction on.
343  * @nblocks: number of block buffer we might modify
344  *
345  * We make sure that the transaction can guarantee at least nblocks of
346  * modified buffers in the log.  We block until the log can guarantee
347  * that much space.
348  *
349  * This function is visible to journal users (like ext3fs), so is not
350  * called with the journal already locked.
351  *
352  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
353  * on failure.
354  */
355 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, gfp_t gfp_mask)
356 {
357         handle_t *handle = journal_current_handle();
358         int err;
359
360         if (!journal)
361                 return ERR_PTR(-EROFS);
362
363         if (handle) {
364                 J_ASSERT(handle->h_transaction->t_journal == journal);
365                 handle->h_ref++;
366                 return handle;
367         }
368
369         handle = new_handle(nblocks);
370         if (!handle)
371                 return ERR_PTR(-ENOMEM);
372
373         current->journal_info = handle;
374
375         err = start_this_handle(journal, handle, gfp_mask);
376         if (err < 0) {
377                 jbd2_free_handle(handle);
378                 current->journal_info = NULL;
379                 handle = ERR_PTR(err);
380         }
381         return handle;
382 }
383 EXPORT_SYMBOL(jbd2__journal_start);
384
385
386 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
387 {
388         return jbd2__journal_start(journal, nblocks, GFP_NOFS);
389 }
390 EXPORT_SYMBOL(jbd2_journal_start);
391
392
393 /**
394  * int jbd2_journal_extend() - extend buffer credits.
395  * @handle:  handle to 'extend'
396  * @nblocks: nr blocks to try to extend by.
397  *
398  * Some transactions, such as large extends and truncates, can be done
399  * atomically all at once or in several stages.  The operation requests
400  * a credit for a number of buffer modications in advance, but can
401  * extend its credit if it needs more.
402  *
403  * jbd2_journal_extend tries to give the running handle more buffer credits.
404  * It does not guarantee that allocation - this is a best-effort only.
405  * The calling process MUST be able to deal cleanly with a failure to
406  * extend here.
407  *
408  * Return 0 on success, non-zero on failure.
409  *
410  * return code < 0 implies an error
411  * return code > 0 implies normal transaction-full status.
412  */
413 int jbd2_journal_extend(handle_t *handle, int nblocks)
414 {
415         transaction_t *transaction = handle->h_transaction;
416         journal_t *journal = transaction->t_journal;
417         int result;
418         int wanted;
419
420         result = -EIO;
421         if (is_handle_aborted(handle))
422                 goto out;
423
424         result = 1;
425
426         read_lock(&journal->j_state_lock);
427
428         /* Don't extend a locked-down transaction! */
429         if (handle->h_transaction->t_state != T_RUNNING) {
430                 jbd_debug(3, "denied handle %p %d blocks: "
431                           "transaction not running\n", handle, nblocks);
432                 goto error_out;
433         }
434
435         spin_lock(&transaction->t_handle_lock);
436         wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
437
438         if (wanted > journal->j_max_transaction_buffers) {
439                 jbd_debug(3, "denied handle %p %d blocks: "
440                           "transaction too large\n", handle, nblocks);
441                 goto unlock;
442         }
443
444         if (wanted > __jbd2_log_space_left(journal)) {
445                 jbd_debug(3, "denied handle %p %d blocks: "
446                           "insufficient log space\n", handle, nblocks);
447                 goto unlock;
448         }
449
450         handle->h_buffer_credits += nblocks;
451         atomic_add(nblocks, &transaction->t_outstanding_credits);
452         result = 0;
453
454         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
455 unlock:
456         spin_unlock(&transaction->t_handle_lock);
457 error_out:
458         read_unlock(&journal->j_state_lock);
459 out:
460         return result;
461 }
462
463
464 /**
465  * int jbd2_journal_restart() - restart a handle .
466  * @handle:  handle to restart
467  * @nblocks: nr credits requested
468  *
469  * Restart a handle for a multi-transaction filesystem
470  * operation.
471  *
472  * If the jbd2_journal_extend() call above fails to grant new buffer credits
473  * to a running handle, a call to jbd2_journal_restart will commit the
474  * handle's transaction so far and reattach the handle to a new
475  * transaction capabable of guaranteeing the requested number of
476  * credits.
477  */
478 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
479 {
480         transaction_t *transaction = handle->h_transaction;
481         journal_t *journal = transaction->t_journal;
482         tid_t           tid;
483         int             need_to_start, ret;
484
485         /* If we've had an abort of any type, don't even think about
486          * actually doing the restart! */
487         if (is_handle_aborted(handle))
488                 return 0;
489
490         /*
491          * First unlink the handle from its current transaction, and start the
492          * commit on that.
493          */
494         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
495         J_ASSERT(journal_current_handle() == handle);
496
497         read_lock(&journal->j_state_lock);
498         spin_lock(&transaction->t_handle_lock);
499         atomic_sub(handle->h_buffer_credits,
500                    &transaction->t_outstanding_credits);
501         if (atomic_dec_and_test(&transaction->t_updates))
502                 wake_up(&journal->j_wait_updates);
503         spin_unlock(&transaction->t_handle_lock);
504
505         jbd_debug(2, "restarting handle %p\n", handle);
506         tid = transaction->t_tid;
507         need_to_start = !tid_geq(journal->j_commit_request, tid);
508         read_unlock(&journal->j_state_lock);
509         if (need_to_start)
510                 jbd2_log_start_commit(journal, tid);
511
512         lock_map_release(&handle->h_lockdep_map);
513         handle->h_buffer_credits = nblocks;
514         ret = start_this_handle(journal, handle, gfp_mask);
515         return ret;
516 }
517 EXPORT_SYMBOL(jbd2__journal_restart);
518
519
520 int jbd2_journal_restart(handle_t *handle, int nblocks)
521 {
522         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
523 }
524 EXPORT_SYMBOL(jbd2_journal_restart);
525
526 /**
527  * void jbd2_journal_lock_updates () - establish a transaction barrier.
528  * @journal:  Journal to establish a barrier on.
529  *
530  * This locks out any further updates from being started, and blocks
531  * until all existing updates have completed, returning only once the
532  * journal is in a quiescent state with no updates running.
533  *
534  * The journal lock should not be held on entry.
535  */
536 void jbd2_journal_lock_updates(journal_t *journal)
537 {
538         DEFINE_WAIT(wait);
539
540         write_lock(&journal->j_state_lock);
541         ++journal->j_barrier_count;
542
543         /* Wait until there are no running updates */
544         while (1) {
545                 transaction_t *transaction = journal->j_running_transaction;
546
547                 if (!transaction)
548                         break;
549
550                 spin_lock(&transaction->t_handle_lock);
551                 prepare_to_wait(&journal->j_wait_updates, &wait,
552                                 TASK_UNINTERRUPTIBLE);
553                 if (!atomic_read(&transaction->t_updates)) {
554                         spin_unlock(&transaction->t_handle_lock);
555                         finish_wait(&journal->j_wait_updates, &wait);
556                         break;
557                 }
558                 spin_unlock(&transaction->t_handle_lock);
559                 write_unlock(&journal->j_state_lock);
560                 schedule();
561                 finish_wait(&journal->j_wait_updates, &wait);
562                 write_lock(&journal->j_state_lock);
563         }
564         write_unlock(&journal->j_state_lock);
565
566         /*
567          * We have now established a barrier against other normal updates, but
568          * we also need to barrier against other jbd2_journal_lock_updates() calls
569          * to make sure that we serialise special journal-locked operations
570          * too.
571          */
572         mutex_lock(&journal->j_barrier);
573 }
574
575 /**
576  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
577  * @journal:  Journal to release the barrier on.
578  *
579  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
580  *
581  * Should be called without the journal lock held.
582  */
583 void jbd2_journal_unlock_updates (journal_t *journal)
584 {
585         J_ASSERT(journal->j_barrier_count != 0);
586
587         mutex_unlock(&journal->j_barrier);
588         write_lock(&journal->j_state_lock);
589         --journal->j_barrier_count;
590         write_unlock(&journal->j_state_lock);
591         wake_up(&journal->j_wait_transaction_locked);
592 }
593
594 static void warn_dirty_buffer(struct buffer_head *bh)
595 {
596         char b[BDEVNAME_SIZE];
597
598         printk(KERN_WARNING
599                "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
600                "There's a risk of filesystem corruption in case of system "
601                "crash.\n",
602                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
603 }
604
605 /*
606  * If the buffer is already part of the current transaction, then there
607  * is nothing we need to do.  If it is already part of a prior
608  * transaction which we are still committing to disk, then we need to
609  * make sure that we do not overwrite the old copy: we do copy-out to
610  * preserve the copy going to disk.  We also account the buffer against
611  * the handle's metadata buffer credits (unless the buffer is already
612  * part of the transaction, that is).
613  *
614  */
615 static int
616 do_get_write_access(handle_t *handle, struct journal_head *jh,
617                         int force_copy)
618 {
619         struct buffer_head *bh;
620         transaction_t *transaction;
621         journal_t *journal;
622         int error;
623         char *frozen_buffer = NULL;
624         int need_copy = 0;
625
626         if (is_handle_aborted(handle))
627                 return -EROFS;
628
629         transaction = handle->h_transaction;
630         journal = transaction->t_journal;
631
632         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
633
634         JBUFFER_TRACE(jh, "entry");
635 repeat:
636         bh = jh2bh(jh);
637
638         /* @@@ Need to check for errors here at some point. */
639
640         lock_buffer(bh);
641         jbd_lock_bh_state(bh);
642
643         /* We now hold the buffer lock so it is safe to query the buffer
644          * state.  Is the buffer dirty?
645          *
646          * If so, there are two possibilities.  The buffer may be
647          * non-journaled, and undergoing a quite legitimate writeback.
648          * Otherwise, it is journaled, and we don't expect dirty buffers
649          * in that state (the buffers should be marked JBD_Dirty
650          * instead.)  So either the IO is being done under our own
651          * control and this is a bug, or it's a third party IO such as
652          * dump(8) (which may leave the buffer scheduled for read ---
653          * ie. locked but not dirty) or tune2fs (which may actually have
654          * the buffer dirtied, ugh.)  */
655
656         if (buffer_dirty(bh)) {
657                 /*
658                  * First question: is this buffer already part of the current
659                  * transaction or the existing committing transaction?
660                  */
661                 if (jh->b_transaction) {
662                         J_ASSERT_JH(jh,
663                                 jh->b_transaction == transaction ||
664                                 jh->b_transaction ==
665                                         journal->j_committing_transaction);
666                         if (jh->b_next_transaction)
667                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
668                                                         transaction);
669                         warn_dirty_buffer(bh);
670                 }
671                 /*
672                  * In any case we need to clean the dirty flag and we must
673                  * do it under the buffer lock to be sure we don't race
674                  * with running write-out.
675                  */
676                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
677                 clear_buffer_dirty(bh);
678                 set_buffer_jbddirty(bh);
679         }
680
681         unlock_buffer(bh);
682
683         error = -EROFS;
684         if (is_handle_aborted(handle)) {
685                 jbd_unlock_bh_state(bh);
686                 goto out;
687         }
688         error = 0;
689
690         /*
691          * The buffer is already part of this transaction if b_transaction or
692          * b_next_transaction points to it
693          */
694         if (jh->b_transaction == transaction ||
695             jh->b_next_transaction == transaction)
696                 goto done;
697
698         /*
699          * this is the first time this transaction is touching this buffer,
700          * reset the modified flag
701          */
702        jh->b_modified = 0;
703
704         /*
705          * If there is already a copy-out version of this buffer, then we don't
706          * need to make another one
707          */
708         if (jh->b_frozen_data) {
709                 JBUFFER_TRACE(jh, "has frozen data");
710                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
711                 jh->b_next_transaction = transaction;
712                 goto done;
713         }
714
715         /* Is there data here we need to preserve? */
716
717         if (jh->b_transaction && jh->b_transaction != transaction) {
718                 JBUFFER_TRACE(jh, "owned by older transaction");
719                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
720                 J_ASSERT_JH(jh, jh->b_transaction ==
721                                         journal->j_committing_transaction);
722
723                 /* There is one case we have to be very careful about.
724                  * If the committing transaction is currently writing
725                  * this buffer out to disk and has NOT made a copy-out,
726                  * then we cannot modify the buffer contents at all
727                  * right now.  The essence of copy-out is that it is the
728                  * extra copy, not the primary copy, which gets
729                  * journaled.  If the primary copy is already going to
730                  * disk then we cannot do copy-out here. */
731
732                 if (jh->b_jlist == BJ_Shadow) {
733                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
734                         wait_queue_head_t *wqh;
735
736                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
737
738                         JBUFFER_TRACE(jh, "on shadow: sleep");
739                         jbd_unlock_bh_state(bh);
740                         /* commit wakes up all shadow buffers after IO */
741                         for ( ; ; ) {
742                                 prepare_to_wait(wqh, &wait.wait,
743                                                 TASK_UNINTERRUPTIBLE);
744                                 if (jh->b_jlist != BJ_Shadow)
745                                         break;
746                                 schedule();
747                         }
748                         finish_wait(wqh, &wait.wait);
749                         goto repeat;
750                 }
751
752                 /* Only do the copy if the currently-owning transaction
753                  * still needs it.  If it is on the Forget list, the
754                  * committing transaction is past that stage.  The
755                  * buffer had better remain locked during the kmalloc,
756                  * but that should be true --- we hold the journal lock
757                  * still and the buffer is already on the BUF_JOURNAL
758                  * list so won't be flushed.
759                  *
760                  * Subtle point, though: if this is a get_undo_access,
761                  * then we will be relying on the frozen_data to contain
762                  * the new value of the committed_data record after the
763                  * transaction, so we HAVE to force the frozen_data copy
764                  * in that case. */
765
766                 if (jh->b_jlist != BJ_Forget || force_copy) {
767                         JBUFFER_TRACE(jh, "generate frozen data");
768                         if (!frozen_buffer) {
769                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
770                                 jbd_unlock_bh_state(bh);
771                                 frozen_buffer =
772                                         jbd2_alloc(jh2bh(jh)->b_size,
773                                                          GFP_NOFS);
774                                 if (!frozen_buffer) {
775                                         printk(KERN_EMERG
776                                                "%s: OOM for frozen_buffer\n",
777                                                __func__);
778                                         JBUFFER_TRACE(jh, "oom!");
779                                         error = -ENOMEM;
780                                         jbd_lock_bh_state(bh);
781                                         goto done;
782                                 }
783                                 goto repeat;
784                         }
785                         jh->b_frozen_data = frozen_buffer;
786                         frozen_buffer = NULL;
787                         need_copy = 1;
788                 }
789                 jh->b_next_transaction = transaction;
790         }
791
792
793         /*
794          * Finally, if the buffer is not journaled right now, we need to make
795          * sure it doesn't get written to disk before the caller actually
796          * commits the new data
797          */
798         if (!jh->b_transaction) {
799                 JBUFFER_TRACE(jh, "no transaction");
800                 J_ASSERT_JH(jh, !jh->b_next_transaction);
801                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
802                 spin_lock(&journal->j_list_lock);
803                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
804                 spin_unlock(&journal->j_list_lock);
805         }
806
807 done:
808         if (need_copy) {
809                 struct page *page;
810                 int offset;
811                 char *source;
812
813                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
814                             "Possible IO failure.\n");
815                 page = jh2bh(jh)->b_page;
816                 offset = offset_in_page(jh2bh(jh)->b_data);
817                 source = kmap_atomic(page);
818                 /* Fire data frozen trigger just before we copy the data */
819                 jbd2_buffer_frozen_trigger(jh, source + offset,
820                                            jh->b_triggers);
821                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
822                 kunmap_atomic(source);
823
824                 /*
825                  * Now that the frozen data is saved off, we need to store
826                  * any matching triggers.
827                  */
828                 jh->b_frozen_triggers = jh->b_triggers;
829         }
830         jbd_unlock_bh_state(bh);
831
832         /*
833          * If we are about to journal a buffer, then any revoke pending on it is
834          * no longer valid
835          */
836         jbd2_journal_cancel_revoke(handle, jh);
837
838 out:
839         if (unlikely(frozen_buffer))    /* It's usually NULL */
840                 jbd2_free(frozen_buffer, bh->b_size);
841
842         JBUFFER_TRACE(jh, "exit");
843         return error;
844 }
845
846 /**
847  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
848  * @handle: transaction to add buffer modifications to
849  * @bh:     bh to be used for metadata writes
850  *
851  * Returns an error code or 0 on success.
852  *
853  * In full data journalling mode the buffer may be of type BJ_AsyncData,
854  * because we're write()ing a buffer which is also part of a shared mapping.
855  */
856
857 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
858 {
859         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
860         int rc;
861
862         /* We do not want to get caught playing with fields which the
863          * log thread also manipulates.  Make sure that the buffer
864          * completes any outstanding IO before proceeding. */
865         rc = do_get_write_access(handle, jh, 0);
866         jbd2_journal_put_journal_head(jh);
867         return rc;
868 }
869
870
871 /*
872  * When the user wants to journal a newly created buffer_head
873  * (ie. getblk() returned a new buffer and we are going to populate it
874  * manually rather than reading off disk), then we need to keep the
875  * buffer_head locked until it has been completely filled with new
876  * data.  In this case, we should be able to make the assertion that
877  * the bh is not already part of an existing transaction.
878  *
879  * The buffer should already be locked by the caller by this point.
880  * There is no lock ranking violation: it was a newly created,
881  * unlocked buffer beforehand. */
882
883 /**
884  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
885  * @handle: transaction to new buffer to
886  * @bh: new buffer.
887  *
888  * Call this if you create a new bh.
889  */
890 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
891 {
892         transaction_t *transaction = handle->h_transaction;
893         journal_t *journal = transaction->t_journal;
894         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
895         int err;
896
897         jbd_debug(5, "journal_head %p\n", jh);
898         err = -EROFS;
899         if (is_handle_aborted(handle))
900                 goto out;
901         err = 0;
902
903         JBUFFER_TRACE(jh, "entry");
904         /*
905          * The buffer may already belong to this transaction due to pre-zeroing
906          * in the filesystem's new_block code.  It may also be on the previous,
907          * committing transaction's lists, but it HAS to be in Forget state in
908          * that case: the transaction must have deleted the buffer for it to be
909          * reused here.
910          */
911         jbd_lock_bh_state(bh);
912         spin_lock(&journal->j_list_lock);
913         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
914                 jh->b_transaction == NULL ||
915                 (jh->b_transaction == journal->j_committing_transaction &&
916                           jh->b_jlist == BJ_Forget)));
917
918         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
919         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
920
921         if (jh->b_transaction == NULL) {
922                 /*
923                  * Previous jbd2_journal_forget() could have left the buffer
924                  * with jbddirty bit set because it was being committed. When
925                  * the commit finished, we've filed the buffer for
926                  * checkpointing and marked it dirty. Now we are reallocating
927                  * the buffer so the transaction freeing it must have
928                  * committed and so it's safe to clear the dirty bit.
929                  */
930                 clear_buffer_dirty(jh2bh(jh));
931                 /* first access by this transaction */
932                 jh->b_modified = 0;
933
934                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
935                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
936         } else if (jh->b_transaction == journal->j_committing_transaction) {
937                 /* first access by this transaction */
938                 jh->b_modified = 0;
939
940                 JBUFFER_TRACE(jh, "set next transaction");
941                 jh->b_next_transaction = transaction;
942         }
943         spin_unlock(&journal->j_list_lock);
944         jbd_unlock_bh_state(bh);
945
946         /*
947          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
948          * blocks which contain freed but then revoked metadata.  We need
949          * to cancel the revoke in case we end up freeing it yet again
950          * and the reallocating as data - this would cause a second revoke,
951          * which hits an assertion error.
952          */
953         JBUFFER_TRACE(jh, "cancelling revoke");
954         jbd2_journal_cancel_revoke(handle, jh);
955 out:
956         jbd2_journal_put_journal_head(jh);
957         return err;
958 }
959
960 /**
961  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
962  *     non-rewindable consequences
963  * @handle: transaction
964  * @bh: buffer to undo
965  *
966  * Sometimes there is a need to distinguish between metadata which has
967  * been committed to disk and that which has not.  The ext3fs code uses
968  * this for freeing and allocating space, we have to make sure that we
969  * do not reuse freed space until the deallocation has been committed,
970  * since if we overwrote that space we would make the delete
971  * un-rewindable in case of a crash.
972  *
973  * To deal with that, jbd2_journal_get_undo_access requests write access to a
974  * buffer for parts of non-rewindable operations such as delete
975  * operations on the bitmaps.  The journaling code must keep a copy of
976  * the buffer's contents prior to the undo_access call until such time
977  * as we know that the buffer has definitely been committed to disk.
978  *
979  * We never need to know which transaction the committed data is part
980  * of, buffers touched here are guaranteed to be dirtied later and so
981  * will be committed to a new transaction in due course, at which point
982  * we can discard the old committed data pointer.
983  *
984  * Returns error number or 0 on success.
985  */
986 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
987 {
988         int err;
989         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
990         char *committed_data = NULL;
991
992         JBUFFER_TRACE(jh, "entry");
993
994         /*
995          * Do this first --- it can drop the journal lock, so we want to
996          * make sure that obtaining the committed_data is done
997          * atomically wrt. completion of any outstanding commits.
998          */
999         err = do_get_write_access(handle, jh, 1);
1000         if (err)
1001                 goto out;
1002
1003 repeat:
1004         if (!jh->b_committed_data) {
1005                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
1006                 if (!committed_data) {
1007                         printk(KERN_EMERG "%s: No memory for committed data\n",
1008                                 __func__);
1009                         err = -ENOMEM;
1010                         goto out;
1011                 }
1012         }
1013
1014         jbd_lock_bh_state(bh);
1015         if (!jh->b_committed_data) {
1016                 /* Copy out the current buffer contents into the
1017                  * preserved, committed copy. */
1018                 JBUFFER_TRACE(jh, "generate b_committed data");
1019                 if (!committed_data) {
1020                         jbd_unlock_bh_state(bh);
1021                         goto repeat;
1022                 }
1023
1024                 jh->b_committed_data = committed_data;
1025                 committed_data = NULL;
1026                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1027         }
1028         jbd_unlock_bh_state(bh);
1029 out:
1030         jbd2_journal_put_journal_head(jh);
1031         if (unlikely(committed_data))
1032                 jbd2_free(committed_data, bh->b_size);
1033         return err;
1034 }
1035
1036 /**
1037  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1038  * @bh: buffer to trigger on
1039  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1040  *
1041  * Set any triggers on this journal_head.  This is always safe, because
1042  * triggers for a committing buffer will be saved off, and triggers for
1043  * a running transaction will match the buffer in that transaction.
1044  *
1045  * Call with NULL to clear the triggers.
1046  */
1047 void jbd2_journal_set_triggers(struct buffer_head *bh,
1048                                struct jbd2_buffer_trigger_type *type)
1049 {
1050         struct journal_head *jh = bh2jh(bh);
1051
1052         jh->b_triggers = type;
1053 }
1054
1055 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1056                                 struct jbd2_buffer_trigger_type *triggers)
1057 {
1058         struct buffer_head *bh = jh2bh(jh);
1059
1060         if (!triggers || !triggers->t_frozen)
1061                 return;
1062
1063         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1064 }
1065
1066 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1067                                struct jbd2_buffer_trigger_type *triggers)
1068 {
1069         if (!triggers || !triggers->t_abort)
1070                 return;
1071
1072         triggers->t_abort(triggers, jh2bh(jh));
1073 }
1074
1075
1076
1077 /**
1078  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1079  * @handle: transaction to add buffer to.
1080  * @bh: buffer to mark
1081  *
1082  * mark dirty metadata which needs to be journaled as part of the current
1083  * transaction.
1084  *
1085  * The buffer must have previously had jbd2_journal_get_write_access()
1086  * called so that it has a valid journal_head attached to the buffer
1087  * head.
1088  *
1089  * The buffer is placed on the transaction's metadata list and is marked
1090  * as belonging to the transaction.
1091  *
1092  * Returns error number or 0 on success.
1093  *
1094  * Special care needs to be taken if the buffer already belongs to the
1095  * current committing transaction (in which case we should have frozen
1096  * data present for that commit).  In that case, we don't relink the
1097  * buffer: that only gets done when the old transaction finally
1098  * completes its commit.
1099  */
1100 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1101 {
1102         transaction_t *transaction = handle->h_transaction;
1103         journal_t *journal = transaction->t_journal;
1104         struct journal_head *jh = bh2jh(bh);
1105         int ret = 0;
1106
1107         jbd_debug(5, "journal_head %p\n", jh);
1108         JBUFFER_TRACE(jh, "entry");
1109         if (is_handle_aborted(handle))
1110                 goto out;
1111         if (!buffer_jbd(bh)) {
1112                 ret = -EUCLEAN;
1113                 goto out;
1114         }
1115
1116         jbd_lock_bh_state(bh);
1117
1118         if (jh->b_modified == 0) {
1119                 /*
1120                  * This buffer's got modified and becoming part
1121                  * of the transaction. This needs to be done
1122                  * once a transaction -bzzz
1123                  */
1124                 jh->b_modified = 1;
1125                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1126                 handle->h_buffer_credits--;
1127         }
1128
1129         /*
1130          * fastpath, to avoid expensive locking.  If this buffer is already
1131          * on the running transaction's metadata list there is nothing to do.
1132          * Nobody can take it off again because there is a handle open.
1133          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1134          * result in this test being false, so we go in and take the locks.
1135          */
1136         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1137                 JBUFFER_TRACE(jh, "fastpath");
1138                 if (unlikely(jh->b_transaction !=
1139                              journal->j_running_transaction)) {
1140                         printk(KERN_EMERG "JBD: %s: "
1141                                "jh->b_transaction (%llu, %p, %u) != "
1142                                "journal->j_running_transaction (%p, %u)",
1143                                journal->j_devname,
1144                                (unsigned long long) bh->b_blocknr,
1145                                jh->b_transaction,
1146                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1147                                journal->j_running_transaction,
1148                                journal->j_running_transaction ?
1149                                journal->j_running_transaction->t_tid : 0);
1150                         ret = -EINVAL;
1151                 }
1152                 goto out_unlock_bh;
1153         }
1154
1155         set_buffer_jbddirty(bh);
1156
1157         /*
1158          * Metadata already on the current transaction list doesn't
1159          * need to be filed.  Metadata on another transaction's list must
1160          * be committing, and will be refiled once the commit completes:
1161          * leave it alone for now.
1162          */
1163         if (jh->b_transaction != transaction) {
1164                 JBUFFER_TRACE(jh, "already on other transaction");
1165                 if (unlikely(jh->b_transaction !=
1166                              journal->j_committing_transaction)) {
1167                         printk(KERN_EMERG "JBD: %s: "
1168                                "jh->b_transaction (%llu, %p, %u) != "
1169                                "journal->j_committing_transaction (%p, %u)",
1170                                journal->j_devname,
1171                                (unsigned long long) bh->b_blocknr,
1172                                jh->b_transaction,
1173                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1174                                journal->j_committing_transaction,
1175                                journal->j_committing_transaction ?
1176                                journal->j_committing_transaction->t_tid : 0);
1177                         ret = -EINVAL;
1178                 }
1179                 if (unlikely(jh->b_next_transaction != transaction)) {
1180                         printk(KERN_EMERG "JBD: %s: "
1181                                "jh->b_next_transaction (%llu, %p, %u) != "
1182                                "transaction (%p, %u)",
1183                                journal->j_devname,
1184                                (unsigned long long) bh->b_blocknr,
1185                                jh->b_next_transaction,
1186                                jh->b_next_transaction ?
1187                                jh->b_next_transaction->t_tid : 0,
1188                                transaction, transaction->t_tid);
1189                         ret = -EINVAL;
1190                 }
1191                 /* And this case is illegal: we can't reuse another
1192                  * transaction's data buffer, ever. */
1193                 goto out_unlock_bh;
1194         }
1195
1196         /* That test should have eliminated the following case: */
1197         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1198
1199         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1200         spin_lock(&journal->j_list_lock);
1201         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1202         spin_unlock(&journal->j_list_lock);
1203 out_unlock_bh:
1204         jbd_unlock_bh_state(bh);
1205 out:
1206         JBUFFER_TRACE(jh, "exit");
1207         WARN_ON(ret);   /* All errors are bugs, so dump the stack */
1208         return ret;
1209 }
1210
1211 /**
1212  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1213  * @handle: transaction handle
1214  * @bh:     bh to 'forget'
1215  *
1216  * We can only do the bforget if there are no commits pending against the
1217  * buffer.  If the buffer is dirty in the current running transaction we
1218  * can safely unlink it.
1219  *
1220  * bh may not be a journalled buffer at all - it may be a non-JBD
1221  * buffer which came off the hashtable.  Check for this.
1222  *
1223  * Decrements bh->b_count by one.
1224  *
1225  * Allow this call even if the handle has aborted --- it may be part of
1226  * the caller's cleanup after an abort.
1227  */
1228 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1229 {
1230         transaction_t *transaction = handle->h_transaction;
1231         journal_t *journal = transaction->t_journal;
1232         struct journal_head *jh;
1233         int drop_reserve = 0;
1234         int err = 0;
1235         int was_modified = 0;
1236
1237         BUFFER_TRACE(bh, "entry");
1238
1239         jbd_lock_bh_state(bh);
1240         spin_lock(&journal->j_list_lock);
1241
1242         if (!buffer_jbd(bh))
1243                 goto not_jbd;
1244         jh = bh2jh(bh);
1245
1246         /* Critical error: attempting to delete a bitmap buffer, maybe?
1247          * Don't do any jbd operations, and return an error. */
1248         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1249                          "inconsistent data on disk")) {
1250                 err = -EIO;
1251                 goto not_jbd;
1252         }
1253
1254         /* keep track of wether or not this transaction modified us */
1255         was_modified = jh->b_modified;
1256
1257         /*
1258          * The buffer's going from the transaction, we must drop
1259          * all references -bzzz
1260          */
1261         jh->b_modified = 0;
1262
1263         if (jh->b_transaction == handle->h_transaction) {
1264                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1265
1266                 /* If we are forgetting a buffer which is already part
1267                  * of this transaction, then we can just drop it from
1268                  * the transaction immediately. */
1269                 clear_buffer_dirty(bh);
1270                 clear_buffer_jbddirty(bh);
1271
1272                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1273
1274                 /*
1275                  * we only want to drop a reference if this transaction
1276                  * modified the buffer
1277                  */
1278                 if (was_modified)
1279                         drop_reserve = 1;
1280
1281                 /*
1282                  * We are no longer going to journal this buffer.
1283                  * However, the commit of this transaction is still
1284                  * important to the buffer: the delete that we are now
1285                  * processing might obsolete an old log entry, so by
1286                  * committing, we can satisfy the buffer's checkpoint.
1287                  *
1288                  * So, if we have a checkpoint on the buffer, we should
1289                  * now refile the buffer on our BJ_Forget list so that
1290                  * we know to remove the checkpoint after we commit.
1291                  */
1292
1293                 if (jh->b_cp_transaction) {
1294                         __jbd2_journal_temp_unlink_buffer(jh);
1295                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1296                 } else {
1297                         __jbd2_journal_unfile_buffer(jh);
1298                         if (!buffer_jbd(bh)) {
1299                                 spin_unlock(&journal->j_list_lock);
1300                                 jbd_unlock_bh_state(bh);
1301                                 __bforget(bh);
1302                                 goto drop;
1303                         }
1304                 }
1305         } else if (jh->b_transaction) {
1306                 J_ASSERT_JH(jh, (jh->b_transaction ==
1307                                  journal->j_committing_transaction));
1308                 /* However, if the buffer is still owned by a prior
1309                  * (committing) transaction, we can't drop it yet... */
1310                 JBUFFER_TRACE(jh, "belongs to older transaction");
1311                 /* ... but we CAN drop it from the new transaction if we
1312                  * have also modified it since the original commit. */
1313
1314                 if (jh->b_next_transaction) {
1315                         J_ASSERT(jh->b_next_transaction == transaction);
1316                         jh->b_next_transaction = NULL;
1317
1318                         /*
1319                          * only drop a reference if this transaction modified
1320                          * the buffer
1321                          */
1322                         if (was_modified)
1323                                 drop_reserve = 1;
1324                 }
1325         }
1326
1327 not_jbd:
1328         spin_unlock(&journal->j_list_lock);
1329         jbd_unlock_bh_state(bh);
1330         __brelse(bh);
1331 drop:
1332         if (drop_reserve) {
1333                 /* no need to reserve log space for this block -bzzz */
1334                 handle->h_buffer_credits++;
1335         }
1336         return err;
1337 }
1338
1339 /**
1340  * int jbd2_journal_stop() - complete a transaction
1341  * @handle: tranaction to complete.
1342  *
1343  * All done for a particular handle.
1344  *
1345  * There is not much action needed here.  We just return any remaining
1346  * buffer credits to the transaction and remove the handle.  The only
1347  * complication is that we need to start a commit operation if the
1348  * filesystem is marked for synchronous update.
1349  *
1350  * jbd2_journal_stop itself will not usually return an error, but it may
1351  * do so in unusual circumstances.  In particular, expect it to
1352  * return -EIO if a jbd2_journal_abort has been executed since the
1353  * transaction began.
1354  */
1355 int jbd2_journal_stop(handle_t *handle)
1356 {
1357         transaction_t *transaction = handle->h_transaction;
1358         journal_t *journal = transaction->t_journal;
1359         int err, wait_for_commit = 0;
1360         tid_t tid;
1361         pid_t pid;
1362
1363         J_ASSERT(journal_current_handle() == handle);
1364
1365         if (is_handle_aborted(handle))
1366                 err = -EIO;
1367         else {
1368                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1369                 err = 0;
1370         }
1371
1372         if (--handle->h_ref > 0) {
1373                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1374                           handle->h_ref);
1375                 return err;
1376         }
1377
1378         jbd_debug(4, "Handle %p going down\n", handle);
1379
1380         /*
1381          * Implement synchronous transaction batching.  If the handle
1382          * was synchronous, don't force a commit immediately.  Let's
1383          * yield and let another thread piggyback onto this
1384          * transaction.  Keep doing that while new threads continue to
1385          * arrive.  It doesn't cost much - we're about to run a commit
1386          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1387          * operations by 30x or more...
1388          *
1389          * We try and optimize the sleep time against what the
1390          * underlying disk can do, instead of having a static sleep
1391          * time.  This is useful for the case where our storage is so
1392          * fast that it is more optimal to go ahead and force a flush
1393          * and wait for the transaction to be committed than it is to
1394          * wait for an arbitrary amount of time for new writers to
1395          * join the transaction.  We achieve this by measuring how
1396          * long it takes to commit a transaction, and compare it with
1397          * how long this transaction has been running, and if run time
1398          * < commit time then we sleep for the delta and commit.  This
1399          * greatly helps super fast disks that would see slowdowns as
1400          * more threads started doing fsyncs.
1401          *
1402          * But don't do this if this process was the most recent one
1403          * to perform a synchronous write.  We do this to detect the
1404          * case where a single process is doing a stream of sync
1405          * writes.  No point in waiting for joiners in that case.
1406          */
1407         pid = current->pid;
1408         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1409                 u64 commit_time, trans_time;
1410
1411                 journal->j_last_sync_writer = pid;
1412
1413                 read_lock(&journal->j_state_lock);
1414                 commit_time = journal->j_average_commit_time;
1415                 read_unlock(&journal->j_state_lock);
1416
1417                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1418                                                    transaction->t_start_time));
1419
1420                 commit_time = max_t(u64, commit_time,
1421                                     1000*journal->j_min_batch_time);
1422                 commit_time = min_t(u64, commit_time,
1423                                     1000*journal->j_max_batch_time);
1424
1425                 if (trans_time < commit_time) {
1426                         ktime_t expires = ktime_add_ns(ktime_get(),
1427                                                        commit_time);
1428                         set_current_state(TASK_UNINTERRUPTIBLE);
1429                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1430                 }
1431         }
1432
1433         if (handle->h_sync)
1434                 transaction->t_synchronous_commit = 1;
1435         current->journal_info = NULL;
1436         atomic_sub(handle->h_buffer_credits,
1437                    &transaction->t_outstanding_credits);
1438
1439         /*
1440          * If the handle is marked SYNC, we need to set another commit
1441          * going!  We also want to force a commit if the current
1442          * transaction is occupying too much of the log, or if the
1443          * transaction is too old now.
1444          */
1445         if (handle->h_sync ||
1446             (atomic_read(&transaction->t_outstanding_credits) >
1447              journal->j_max_transaction_buffers) ||
1448             time_after_eq(jiffies, transaction->t_expires)) {
1449                 /* Do this even for aborted journals: an abort still
1450                  * completes the commit thread, it just doesn't write
1451                  * anything to disk. */
1452
1453                 jbd_debug(2, "transaction too old, requesting commit for "
1454                                         "handle %p\n", handle);
1455                 /* This is non-blocking */
1456                 jbd2_log_start_commit(journal, transaction->t_tid);
1457
1458                 /*
1459                  * Special case: JBD2_SYNC synchronous updates require us
1460                  * to wait for the commit to complete.
1461                  */
1462                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1463                         wait_for_commit = 1;
1464         }
1465
1466         /*
1467          * Once we drop t_updates, if it goes to zero the transaction
1468          * could start committing on us and eventually disappear.  So
1469          * once we do this, we must not dereference transaction
1470          * pointer again.
1471          */
1472         tid = transaction->t_tid;
1473         if (atomic_dec_and_test(&transaction->t_updates)) {
1474                 wake_up(&journal->j_wait_updates);
1475                 if (journal->j_barrier_count)
1476                         wake_up(&journal->j_wait_transaction_locked);
1477         }
1478
1479         if (wait_for_commit)
1480                 err = jbd2_log_wait_commit(journal, tid);
1481
1482         lock_map_release(&handle->h_lockdep_map);
1483
1484         jbd2_free_handle(handle);
1485         return err;
1486 }
1487
1488 /**
1489  * int jbd2_journal_force_commit() - force any uncommitted transactions
1490  * @journal: journal to force
1491  *
1492  * For synchronous operations: force any uncommitted transactions
1493  * to disk.  May seem kludgy, but it reuses all the handle batching
1494  * code in a very simple manner.
1495  */
1496 int jbd2_journal_force_commit(journal_t *journal)
1497 {
1498         handle_t *handle;
1499         int ret;
1500
1501         handle = jbd2_journal_start(journal, 1);
1502         if (IS_ERR(handle)) {
1503                 ret = PTR_ERR(handle);
1504         } else {
1505                 handle->h_sync = 1;
1506                 ret = jbd2_journal_stop(handle);
1507         }
1508         return ret;
1509 }
1510
1511 /*
1512  *
1513  * List management code snippets: various functions for manipulating the
1514  * transaction buffer lists.
1515  *
1516  */
1517
1518 /*
1519  * Append a buffer to a transaction list, given the transaction's list head
1520  * pointer.
1521  *
1522  * j_list_lock is held.
1523  *
1524  * jbd_lock_bh_state(jh2bh(jh)) is held.
1525  */
1526
1527 static inline void
1528 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1529 {
1530         if (!*list) {
1531                 jh->b_tnext = jh->b_tprev = jh;
1532                 *list = jh;
1533         } else {
1534                 /* Insert at the tail of the list to preserve order */
1535                 struct journal_head *first = *list, *last = first->b_tprev;
1536                 jh->b_tprev = last;
1537                 jh->b_tnext = first;
1538                 last->b_tnext = first->b_tprev = jh;
1539         }
1540 }
1541
1542 /*
1543  * Remove a buffer from a transaction list, given the transaction's list
1544  * head pointer.
1545  *
1546  * Called with j_list_lock held, and the journal may not be locked.
1547  *
1548  * jbd_lock_bh_state(jh2bh(jh)) is held.
1549  */
1550
1551 static inline void
1552 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1553 {
1554         if (*list == jh) {
1555                 *list = jh->b_tnext;
1556                 if (*list == jh)
1557                         *list = NULL;
1558         }
1559         jh->b_tprev->b_tnext = jh->b_tnext;
1560         jh->b_tnext->b_tprev = jh->b_tprev;
1561 }
1562
1563 /*
1564  * Remove a buffer from the appropriate transaction list.
1565  *
1566  * Note that this function can *change* the value of
1567  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1568  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1569  * of these pointers, it could go bad.  Generally the caller needs to re-read
1570  * the pointer from the transaction_t.
1571  *
1572  * Called under j_list_lock.
1573  */
1574 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1575 {
1576         struct journal_head **list = NULL;
1577         transaction_t *transaction;
1578         struct buffer_head *bh = jh2bh(jh);
1579
1580         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1581         transaction = jh->b_transaction;
1582         if (transaction)
1583                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1584
1585         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1586         if (jh->b_jlist != BJ_None)
1587                 J_ASSERT_JH(jh, transaction != NULL);
1588
1589         switch (jh->b_jlist) {
1590         case BJ_None:
1591                 return;
1592         case BJ_Metadata:
1593                 transaction->t_nr_buffers--;
1594                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1595                 list = &transaction->t_buffers;
1596                 break;
1597         case BJ_Forget:
1598                 list = &transaction->t_forget;
1599                 break;
1600         case BJ_IO:
1601                 list = &transaction->t_iobuf_list;
1602                 break;
1603         case BJ_Shadow:
1604                 list = &transaction->t_shadow_list;
1605                 break;
1606         case BJ_LogCtl:
1607                 list = &transaction->t_log_list;
1608                 break;
1609         case BJ_Reserved:
1610                 list = &transaction->t_reserved_list;
1611                 break;
1612         }
1613
1614         __blist_del_buffer(list, jh);
1615         jh->b_jlist = BJ_None;
1616         if (test_clear_buffer_jbddirty(bh))
1617                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1618 }
1619
1620 /*
1621  * Remove buffer from all transactions.
1622  *
1623  * Called with bh_state lock and j_list_lock
1624  *
1625  * jh and bh may be already freed when this function returns.
1626  */
1627 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1628 {
1629         __jbd2_journal_temp_unlink_buffer(jh);
1630         jh->b_transaction = NULL;
1631         jbd2_journal_put_journal_head(jh);
1632 }
1633
1634 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1635 {
1636         struct buffer_head *bh = jh2bh(jh);
1637
1638         /* Get reference so that buffer cannot be freed before we unlock it */
1639         get_bh(bh);
1640         jbd_lock_bh_state(bh);
1641         spin_lock(&journal->j_list_lock);
1642         __jbd2_journal_unfile_buffer(jh);
1643         spin_unlock(&journal->j_list_lock);
1644         jbd_unlock_bh_state(bh);
1645         __brelse(bh);
1646 }
1647
1648 /*
1649  * Called from jbd2_journal_try_to_free_buffers().
1650  *
1651  * Called under jbd_lock_bh_state(bh)
1652  */
1653 static void
1654 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1655 {
1656         struct journal_head *jh;
1657
1658         jh = bh2jh(bh);
1659
1660         if (buffer_locked(bh) || buffer_dirty(bh))
1661                 goto out;
1662
1663         if (jh->b_next_transaction != NULL)
1664                 goto out;
1665
1666         spin_lock(&journal->j_list_lock);
1667         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1668                 /* written-back checkpointed metadata buffer */
1669                 JBUFFER_TRACE(jh, "remove from checkpoint list");
1670                 __jbd2_journal_remove_checkpoint(jh);
1671         }
1672         spin_unlock(&journal->j_list_lock);
1673 out:
1674         return;
1675 }
1676
1677 /**
1678  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1679  * @journal: journal for operation
1680  * @page: to try and free
1681  * @gfp_mask: we use the mask to detect how hard should we try to release
1682  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1683  * release the buffers.
1684  *
1685  *
1686  * For all the buffers on this page,
1687  * if they are fully written out ordered data, move them onto BUF_CLEAN
1688  * so try_to_free_buffers() can reap them.
1689  *
1690  * This function returns non-zero if we wish try_to_free_buffers()
1691  * to be called. We do this if the page is releasable by try_to_free_buffers().
1692  * We also do it if the page has locked or dirty buffers and the caller wants
1693  * us to perform sync or async writeout.
1694  *
1695  * This complicates JBD locking somewhat.  We aren't protected by the
1696  * BKL here.  We wish to remove the buffer from its committing or
1697  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1698  *
1699  * This may *change* the value of transaction_t->t_datalist, so anyone
1700  * who looks at t_datalist needs to lock against this function.
1701  *
1702  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1703  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1704  * will come out of the lock with the buffer dirty, which makes it
1705  * ineligible for release here.
1706  *
1707  * Who else is affected by this?  hmm...  Really the only contender
1708  * is do_get_write_access() - it could be looking at the buffer while
1709  * journal_try_to_free_buffer() is changing its state.  But that
1710  * cannot happen because we never reallocate freed data as metadata
1711  * while the data is part of a transaction.  Yes?
1712  *
1713  * Return 0 on failure, 1 on success
1714  */
1715 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1716                                 struct page *page, gfp_t gfp_mask)
1717 {
1718         struct buffer_head *head;
1719         struct buffer_head *bh;
1720         int ret = 0;
1721
1722         J_ASSERT(PageLocked(page));
1723
1724         head = page_buffers(page);
1725         bh = head;
1726         do {
1727                 struct journal_head *jh;
1728
1729                 /*
1730                  * We take our own ref against the journal_head here to avoid
1731                  * having to add tons of locking around each instance of
1732                  * jbd2_journal_put_journal_head().
1733                  */
1734                 jh = jbd2_journal_grab_journal_head(bh);
1735                 if (!jh)
1736                         continue;
1737
1738                 jbd_lock_bh_state(bh);
1739                 __journal_try_to_free_buffer(journal, bh);
1740                 jbd2_journal_put_journal_head(jh);
1741                 jbd_unlock_bh_state(bh);
1742                 if (buffer_jbd(bh))
1743                         goto busy;
1744         } while ((bh = bh->b_this_page) != head);
1745
1746         ret = try_to_free_buffers(page);
1747
1748 busy:
1749         return ret;
1750 }
1751
1752 /*
1753  * This buffer is no longer needed.  If it is on an older transaction's
1754  * checkpoint list we need to record it on this transaction's forget list
1755  * to pin this buffer (and hence its checkpointing transaction) down until
1756  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1757  * release it.
1758  * Returns non-zero if JBD no longer has an interest in the buffer.
1759  *
1760  * Called under j_list_lock.
1761  *
1762  * Called under jbd_lock_bh_state(bh).
1763  */
1764 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1765 {
1766         int may_free = 1;
1767         struct buffer_head *bh = jh2bh(jh);
1768
1769         if (jh->b_cp_transaction) {
1770                 JBUFFER_TRACE(jh, "on running+cp transaction");
1771                 __jbd2_journal_temp_unlink_buffer(jh);
1772                 /*
1773                  * We don't want to write the buffer anymore, clear the
1774                  * bit so that we don't confuse checks in
1775                  * __journal_file_buffer
1776                  */
1777                 clear_buffer_dirty(bh);
1778                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1779                 may_free = 0;
1780         } else {
1781                 JBUFFER_TRACE(jh, "on running transaction");
1782                 __jbd2_journal_unfile_buffer(jh);
1783         }
1784         return may_free;
1785 }
1786
1787 /*
1788  * jbd2_journal_invalidatepage
1789  *
1790  * This code is tricky.  It has a number of cases to deal with.
1791  *
1792  * There are two invariants which this code relies on:
1793  *
1794  * i_size must be updated on disk before we start calling invalidatepage on the
1795  * data.
1796  *
1797  *  This is done in ext3 by defining an ext3_setattr method which
1798  *  updates i_size before truncate gets going.  By maintaining this
1799  *  invariant, we can be sure that it is safe to throw away any buffers
1800  *  attached to the current transaction: once the transaction commits,
1801  *  we know that the data will not be needed.
1802  *
1803  *  Note however that we can *not* throw away data belonging to the
1804  *  previous, committing transaction!
1805  *
1806  * Any disk blocks which *are* part of the previous, committing
1807  * transaction (and which therefore cannot be discarded immediately) are
1808  * not going to be reused in the new running transaction
1809  *
1810  *  The bitmap committed_data images guarantee this: any block which is
1811  *  allocated in one transaction and removed in the next will be marked
1812  *  as in-use in the committed_data bitmap, so cannot be reused until
1813  *  the next transaction to delete the block commits.  This means that
1814  *  leaving committing buffers dirty is quite safe: the disk blocks
1815  *  cannot be reallocated to a different file and so buffer aliasing is
1816  *  not possible.
1817  *
1818  *
1819  * The above applies mainly to ordered data mode.  In writeback mode we
1820  * don't make guarantees about the order in which data hits disk --- in
1821  * particular we don't guarantee that new dirty data is flushed before
1822  * transaction commit --- so it is always safe just to discard data
1823  * immediately in that mode.  --sct
1824  */
1825
1826 /*
1827  * The journal_unmap_buffer helper function returns zero if the buffer
1828  * concerned remains pinned as an anonymous buffer belonging to an older
1829  * transaction.
1830  *
1831  * We're outside-transaction here.  Either or both of j_running_transaction
1832  * and j_committing_transaction may be NULL.
1833  */
1834 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1835                                 int partial_page)
1836 {
1837         transaction_t *transaction;
1838         struct journal_head *jh;
1839         int may_free = 1;
1840
1841         BUFFER_TRACE(bh, "entry");
1842
1843         /*
1844          * It is safe to proceed here without the j_list_lock because the
1845          * buffers cannot be stolen by try_to_free_buffers as long as we are
1846          * holding the page lock. --sct
1847          */
1848
1849         if (!buffer_jbd(bh))
1850                 goto zap_buffer_unlocked;
1851
1852         /* OK, we have data buffer in journaled mode */
1853         write_lock(&journal->j_state_lock);
1854         jbd_lock_bh_state(bh);
1855         spin_lock(&journal->j_list_lock);
1856
1857         jh = jbd2_journal_grab_journal_head(bh);
1858         if (!jh)
1859                 goto zap_buffer_no_jh;
1860
1861         /*
1862          * We cannot remove the buffer from checkpoint lists until the
1863          * transaction adding inode to orphan list (let's call it T)
1864          * is committed.  Otherwise if the transaction changing the
1865          * buffer would be cleaned from the journal before T is
1866          * committed, a crash will cause that the correct contents of
1867          * the buffer will be lost.  On the other hand we have to
1868          * clear the buffer dirty bit at latest at the moment when the
1869          * transaction marking the buffer as freed in the filesystem
1870          * structures is committed because from that moment on the
1871          * block can be reallocated and used by a different page.
1872          * Since the block hasn't been freed yet but the inode has
1873          * already been added to orphan list, it is safe for us to add
1874          * the buffer to BJ_Forget list of the newest transaction.
1875          *
1876          * Also we have to clear buffer_mapped flag of a truncated buffer
1877          * because the buffer_head may be attached to the page straddling
1878          * i_size (can happen only when blocksize < pagesize) and thus the
1879          * buffer_head can be reused when the file is extended again. So we end
1880          * up keeping around invalidated buffers attached to transactions'
1881          * BJ_Forget list just to stop checkpointing code from cleaning up
1882          * the transaction this buffer was modified in.
1883          */
1884         transaction = jh->b_transaction;
1885         if (transaction == NULL) {
1886                 /* First case: not on any transaction.  If it
1887                  * has no checkpoint link, then we can zap it:
1888                  * it's a writeback-mode buffer so we don't care
1889                  * if it hits disk safely. */
1890                 if (!jh->b_cp_transaction) {
1891                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1892                         goto zap_buffer;
1893                 }
1894
1895                 if (!buffer_dirty(bh)) {
1896                         /* bdflush has written it.  We can drop it now */
1897                         goto zap_buffer;
1898                 }
1899
1900                 /* OK, it must be in the journal but still not
1901                  * written fully to disk: it's metadata or
1902                  * journaled data... */
1903
1904                 if (journal->j_running_transaction) {
1905                         /* ... and once the current transaction has
1906                          * committed, the buffer won't be needed any
1907                          * longer. */
1908                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1909                         may_free = __dispose_buffer(jh,
1910                                         journal->j_running_transaction);
1911                         goto zap_buffer;
1912                 } else {
1913                         /* There is no currently-running transaction. So the
1914                          * orphan record which we wrote for this file must have
1915                          * passed into commit.  We must attach this buffer to
1916                          * the committing transaction, if it exists. */
1917                         if (journal->j_committing_transaction) {
1918                                 JBUFFER_TRACE(jh, "give to committing trans");
1919                                 may_free = __dispose_buffer(jh,
1920                                         journal->j_committing_transaction);
1921                                 goto zap_buffer;
1922                         } else {
1923                                 /* The orphan record's transaction has
1924                                  * committed.  We can cleanse this buffer */
1925                                 clear_buffer_jbddirty(bh);
1926                                 goto zap_buffer;
1927                         }
1928                 }
1929         } else if (transaction == journal->j_committing_transaction) {
1930                 JBUFFER_TRACE(jh, "on committing transaction");
1931                 /*
1932                  * The buffer is committing, we simply cannot touch
1933                  * it. If the page is straddling i_size we have to wait
1934                  * for commit and try again.
1935                  */
1936                 if (partial_page) {
1937                         jbd2_journal_put_journal_head(jh);
1938                         spin_unlock(&journal->j_list_lock);
1939                         jbd_unlock_bh_state(bh);
1940                         write_unlock(&journal->j_state_lock);
1941                         return -EBUSY;
1942                 }
1943                 /*
1944                  * OK, buffer won't be reachable after truncate. We just set
1945                  * j_next_transaction to the running transaction (if there is
1946                  * one) and mark buffer as freed so that commit code knows it
1947                  * should clear dirty bits when it is done with the buffer.
1948                  */
1949                 set_buffer_freed(bh);
1950                 if (journal->j_running_transaction && buffer_jbddirty(bh))
1951                         jh->b_next_transaction = journal->j_running_transaction;
1952                 jbd2_journal_put_journal_head(jh);
1953                 spin_unlock(&journal->j_list_lock);
1954                 jbd_unlock_bh_state(bh);
1955                 write_unlock(&journal->j_state_lock);
1956                 return 0;
1957         } else {
1958                 /* Good, the buffer belongs to the running transaction.
1959                  * We are writing our own transaction's data, not any
1960                  * previous one's, so it is safe to throw it away
1961                  * (remember that we expect the filesystem to have set
1962                  * i_size already for this truncate so recovery will not
1963                  * expose the disk blocks we are discarding here.) */
1964                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1965                 JBUFFER_TRACE(jh, "on running transaction");
1966                 may_free = __dispose_buffer(jh, transaction);
1967         }
1968
1969 zap_buffer:
1970         /*
1971          * This is tricky. Although the buffer is truncated, it may be reused
1972          * if blocksize < pagesize and it is attached to the page straddling
1973          * EOF. Since the buffer might have been added to BJ_Forget list of the
1974          * running transaction, journal_get_write_access() won't clear
1975          * b_modified and credit accounting gets confused. So clear b_modified
1976          * here.
1977          */
1978         jh->b_modified = 0;
1979         jbd2_journal_put_journal_head(jh);
1980 zap_buffer_no_jh:
1981         spin_unlock(&journal->j_list_lock);
1982         jbd_unlock_bh_state(bh);
1983         write_unlock(&journal->j_state_lock);
1984 zap_buffer_unlocked:
1985         clear_buffer_dirty(bh);
1986         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1987         clear_buffer_mapped(bh);
1988         clear_buffer_req(bh);
1989         clear_buffer_new(bh);
1990         clear_buffer_delay(bh);
1991         clear_buffer_unwritten(bh);
1992         bh->b_bdev = NULL;
1993         return may_free;
1994 }
1995
1996 /**
1997  * void jbd2_journal_invalidatepage()
1998  * @journal: journal to use for flush...
1999  * @page:    page to flush
2000  * @offset:  length of page to invalidate.
2001  *
2002  * Reap page buffers containing data after offset in page. Can return -EBUSY
2003  * if buffers are part of the committing transaction and the page is straddling
2004  * i_size. Caller then has to wait for current commit and try again.
2005  */
2006 int jbd2_journal_invalidatepage(journal_t *journal,
2007                                 struct page *page,
2008                                 unsigned long offset)
2009 {
2010         struct buffer_head *head, *bh, *next;
2011         unsigned int curr_off = 0;
2012         int may_free = 1;
2013         int ret = 0;
2014
2015         if (!PageLocked(page))
2016                 BUG();
2017         if (!page_has_buffers(page))
2018                 return 0;
2019
2020         /* We will potentially be playing with lists other than just the
2021          * data lists (especially for journaled data mode), so be
2022          * cautious in our locking. */
2023
2024         head = bh = page_buffers(page);
2025         do {
2026                 unsigned int next_off = curr_off + bh->b_size;
2027                 next = bh->b_this_page;
2028
2029                 if (offset <= curr_off) {
2030                         /* This block is wholly outside the truncation point */
2031                         lock_buffer(bh);
2032                         ret = journal_unmap_buffer(journal, bh, offset > 0);
2033                         unlock_buffer(bh);
2034                         if (ret < 0)
2035                                 return ret;
2036                         may_free &= ret;
2037                 }
2038                 curr_off = next_off;
2039                 bh = next;
2040
2041         } while (bh != head);
2042
2043         if (!offset) {
2044                 if (may_free && try_to_free_buffers(page))
2045                         J_ASSERT(!page_has_buffers(page));
2046         }
2047         return 0;
2048 }
2049
2050 /*
2051  * File a buffer on the given transaction list.
2052  */
2053 void __jbd2_journal_file_buffer(struct journal_head *jh,
2054                         transaction_t *transaction, int jlist)
2055 {
2056         struct journal_head **list = NULL;
2057         int was_dirty = 0;
2058         struct buffer_head *bh = jh2bh(jh);
2059
2060         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2061         assert_spin_locked(&transaction->t_journal->j_list_lock);
2062
2063         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2064         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2065                                 jh->b_transaction == NULL);
2066
2067         if (jh->b_transaction && jh->b_jlist == jlist)
2068                 return;
2069
2070         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2071             jlist == BJ_Shadow || jlist == BJ_Forget) {
2072                 /*
2073                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2074                  * instead of buffer_dirty. We should not see a dirty bit set
2075                  * here because we clear it in do_get_write_access but e.g.
2076                  * tune2fs can modify the sb and set the dirty bit at any time
2077                  * so we try to gracefully handle that.
2078                  */
2079                 if (buffer_dirty(bh))
2080                         warn_dirty_buffer(bh);
2081                 if (test_clear_buffer_dirty(bh) ||
2082                     test_clear_buffer_jbddirty(bh))
2083                         was_dirty = 1;
2084         }
2085
2086         if (jh->b_transaction)
2087                 __jbd2_journal_temp_unlink_buffer(jh);
2088         else
2089                 jbd2_journal_grab_journal_head(bh);
2090         jh->b_transaction = transaction;
2091
2092         switch (jlist) {
2093         case BJ_None:
2094                 J_ASSERT_JH(jh, !jh->b_committed_data);
2095                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2096                 return;
2097         case BJ_Metadata:
2098                 transaction->t_nr_buffers++;
2099                 list = &transaction->t_buffers;
2100                 break;
2101         case BJ_Forget:
2102                 list = &transaction->t_forget;
2103                 break;
2104         case BJ_IO:
2105                 list = &transaction->t_iobuf_list;
2106                 break;
2107         case BJ_Shadow:
2108                 list = &transaction->t_shadow_list;
2109                 break;
2110         case BJ_LogCtl:
2111                 list = &transaction->t_log_list;
2112                 break;
2113         case BJ_Reserved:
2114                 list = &transaction->t_reserved_list;
2115                 break;
2116         }
2117
2118         __blist_add_buffer(list, jh);
2119         jh->b_jlist = jlist;
2120
2121         if (was_dirty)
2122                 set_buffer_jbddirty(bh);
2123 }
2124
2125 void jbd2_journal_file_buffer(struct journal_head *jh,
2126                                 transaction_t *transaction, int jlist)
2127 {
2128         jbd_lock_bh_state(jh2bh(jh));
2129         spin_lock(&transaction->t_journal->j_list_lock);
2130         __jbd2_journal_file_buffer(jh, transaction, jlist);
2131         spin_unlock(&transaction->t_journal->j_list_lock);
2132         jbd_unlock_bh_state(jh2bh(jh));
2133 }
2134
2135 /*
2136  * Remove a buffer from its current buffer list in preparation for
2137  * dropping it from its current transaction entirely.  If the buffer has
2138  * already started to be used by a subsequent transaction, refile the
2139  * buffer on that transaction's metadata list.
2140  *
2141  * Called under j_list_lock
2142  * Called under jbd_lock_bh_state(jh2bh(jh))
2143  *
2144  * jh and bh may be already free when this function returns
2145  */
2146 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2147 {
2148         int was_dirty, jlist;
2149         struct buffer_head *bh = jh2bh(jh);
2150
2151         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2152         if (jh->b_transaction)
2153                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2154
2155         /* If the buffer is now unused, just drop it. */
2156         if (jh->b_next_transaction == NULL) {
2157                 __jbd2_journal_unfile_buffer(jh);
2158                 return;
2159         }
2160
2161         /*
2162          * It has been modified by a later transaction: add it to the new
2163          * transaction's metadata list.
2164          */
2165
2166         was_dirty = test_clear_buffer_jbddirty(bh);
2167         __jbd2_journal_temp_unlink_buffer(jh);
2168         /*
2169          * We set b_transaction here because b_next_transaction will inherit
2170          * our jh reference and thus __jbd2_journal_file_buffer() must not
2171          * take a new one.
2172          */
2173         jh->b_transaction = jh->b_next_transaction;
2174         jh->b_next_transaction = NULL;
2175         if (buffer_freed(bh))
2176                 jlist = BJ_Forget;
2177         else if (jh->b_modified)
2178                 jlist = BJ_Metadata;
2179         else
2180                 jlist = BJ_Reserved;
2181         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2182         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2183
2184         if (was_dirty)
2185                 set_buffer_jbddirty(bh);
2186 }
2187
2188 /*
2189  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2190  * bh reference so that we can safely unlock bh.
2191  *
2192  * The jh and bh may be freed by this call.
2193  */
2194 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2195 {
2196         struct buffer_head *bh = jh2bh(jh);
2197
2198         /* Get reference so that buffer cannot be freed before we unlock it */
2199         get_bh(bh);
2200         jbd_lock_bh_state(bh);
2201         spin_lock(&journal->j_list_lock);
2202         __jbd2_journal_refile_buffer(jh);
2203         jbd_unlock_bh_state(bh);
2204         spin_unlock(&journal->j_list_lock);
2205         __brelse(bh);
2206 }
2207
2208 /*
2209  * File inode in the inode list of the handle's transaction
2210  */
2211 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2212 {
2213         transaction_t *transaction = handle->h_transaction;
2214         journal_t *journal = transaction->t_journal;
2215
2216         if (is_handle_aborted(handle))
2217                 return -EIO;
2218
2219         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2220                         transaction->t_tid);
2221
2222         /*
2223          * First check whether inode isn't already on the transaction's
2224          * lists without taking the lock. Note that this check is safe
2225          * without the lock as we cannot race with somebody removing inode
2226          * from the transaction. The reason is that we remove inode from the
2227          * transaction only in journal_release_jbd_inode() and when we commit
2228          * the transaction. We are guarded from the first case by holding
2229          * a reference to the inode. We are safe against the second case
2230          * because if jinode->i_transaction == transaction, commit code
2231          * cannot touch the transaction because we hold reference to it,
2232          * and if jinode->i_next_transaction == transaction, commit code
2233          * will only file the inode where we want it.
2234          */
2235         if (jinode->i_transaction == transaction ||
2236             jinode->i_next_transaction == transaction)
2237                 return 0;
2238
2239         spin_lock(&journal->j_list_lock);
2240
2241         if (jinode->i_transaction == transaction ||
2242             jinode->i_next_transaction == transaction)
2243                 goto done;
2244
2245         /*
2246          * We only ever set this variable to 1 so the test is safe. Since
2247          * t_need_data_flush is likely to be set, we do the test to save some
2248          * cacheline bouncing
2249          */
2250         if (!transaction->t_need_data_flush)
2251                 transaction->t_need_data_flush = 1;
2252         /* On some different transaction's list - should be
2253          * the committing one */
2254         if (jinode->i_transaction) {
2255                 J_ASSERT(jinode->i_next_transaction == NULL);
2256                 J_ASSERT(jinode->i_transaction ==
2257                                         journal->j_committing_transaction);
2258                 jinode->i_next_transaction = transaction;
2259                 goto done;
2260         }
2261         /* Not on any transaction list... */
2262         J_ASSERT(!jinode->i_next_transaction);
2263         jinode->i_transaction = transaction;
2264         list_add(&jinode->i_list, &transaction->t_inode_list);
2265 done:
2266         spin_unlock(&journal->j_list_lock);
2267
2268         return 0;
2269 }
2270
2271 /*
2272  * File truncate and transaction commit interact with each other in a
2273  * non-trivial way.  If a transaction writing data block A is
2274  * committing, we cannot discard the data by truncate until we have
2275  * written them.  Otherwise if we crashed after the transaction with
2276  * write has committed but before the transaction with truncate has
2277  * committed, we could see stale data in block A.  This function is a
2278  * helper to solve this problem.  It starts writeout of the truncated
2279  * part in case it is in the committing transaction.
2280  *
2281  * Filesystem code must call this function when inode is journaled in
2282  * ordered mode before truncation happens and after the inode has been
2283  * placed on orphan list with the new inode size. The second condition
2284  * avoids the race that someone writes new data and we start
2285  * committing the transaction after this function has been called but
2286  * before a transaction for truncate is started (and furthermore it
2287  * allows us to optimize the case where the addition to orphan list
2288  * happens in the same transaction as write --- we don't have to write
2289  * any data in such case).
2290  */
2291 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2292                                         struct jbd2_inode *jinode,
2293                                         loff_t new_size)
2294 {
2295         transaction_t *inode_trans, *commit_trans;
2296         int ret = 0;
2297
2298         /* This is a quick check to avoid locking if not necessary */
2299         if (!jinode->i_transaction)
2300                 goto out;
2301         /* Locks are here just to force reading of recent values, it is
2302          * enough that the transaction was not committing before we started
2303          * a transaction adding the inode to orphan list */
2304         read_lock(&journal->j_state_lock);
2305         commit_trans = journal->j_committing_transaction;
2306         read_unlock(&journal->j_state_lock);
2307         spin_lock(&journal->j_list_lock);
2308         inode_trans = jinode->i_transaction;
2309         spin_unlock(&journal->j_list_lock);
2310         if (inode_trans == commit_trans) {
2311                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2312                         new_size, LLONG_MAX);
2313                 if (ret)
2314                         jbd2_journal_abort(journal, ret);
2315         }
2316 out:
2317         return ret;
2318 }