Merge branch 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial
[linux-2.6.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/module.h>
31
32 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
33
34 /*
35  * jbd2_get_transaction: obtain a new transaction_t object.
36  *
37  * Simply allocate and initialise a new transaction.  Create it in
38  * RUNNING state and add it to the current journal (which should not
39  * have an existing running transaction: we only make a new transaction
40  * once we have started to commit the old one).
41  *
42  * Preconditions:
43  *      The journal MUST be locked.  We don't perform atomic mallocs on the
44  *      new transaction and we can't block without protecting against other
45  *      processes trying to touch the journal while it is in transition.
46  *
47  */
48
49 static transaction_t *
50 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
51 {
52         transaction->t_journal = journal;
53         transaction->t_state = T_RUNNING;
54         transaction->t_start_time = ktime_get();
55         transaction->t_tid = journal->j_transaction_sequence++;
56         transaction->t_expires = jiffies + journal->j_commit_interval;
57         spin_lock_init(&transaction->t_handle_lock);
58         atomic_set(&transaction->t_updates, 0);
59         atomic_set(&transaction->t_outstanding_credits, 0);
60         atomic_set(&transaction->t_handle_count, 0);
61         INIT_LIST_HEAD(&transaction->t_inode_list);
62         INIT_LIST_HEAD(&transaction->t_private_list);
63
64         /* Set up the commit timer for the new transaction. */
65         journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
66         add_timer(&journal->j_commit_timer);
67
68         J_ASSERT(journal->j_running_transaction == NULL);
69         journal->j_running_transaction = transaction;
70         transaction->t_max_wait = 0;
71         transaction->t_start = jiffies;
72
73         return transaction;
74 }
75
76 /*
77  * Handle management.
78  *
79  * A handle_t is an object which represents a single atomic update to a
80  * filesystem, and which tracks all of the modifications which form part
81  * of that one update.
82  */
83
84 /*
85  * Update transiaction's maximum wait time, if debugging is enabled.
86  *
87  * In order for t_max_wait to be reliable, it must be protected by a
88  * lock.  But doing so will mean that start_this_handle() can not be
89  * run in parallel on SMP systems, which limits our scalability.  So
90  * unless debugging is enabled, we no longer update t_max_wait, which
91  * means that maximum wait time reported by the jbd2_run_stats
92  * tracepoint will always be zero.
93  */
94 static inline void update_t_max_wait(transaction_t *transaction)
95 {
96 #ifdef CONFIG_JBD2_DEBUG
97         unsigned long ts = jiffies;
98
99         if (jbd2_journal_enable_debug &&
100             time_after(transaction->t_start, ts)) {
101                 ts = jbd2_time_diff(ts, transaction->t_start);
102                 spin_lock(&transaction->t_handle_lock);
103                 if (ts > transaction->t_max_wait)
104                         transaction->t_max_wait = ts;
105                 spin_unlock(&transaction->t_handle_lock);
106         }
107 #endif
108 }
109
110 /*
111  * start_this_handle: Given a handle, deal with any locking or stalling
112  * needed to make sure that there is enough journal space for the handle
113  * to begin.  Attach the handle to a transaction and set up the
114  * transaction's buffer credits.
115  */
116
117 static int start_this_handle(journal_t *journal, handle_t *handle,
118                              int gfp_mask)
119 {
120         transaction_t *transaction;
121         int needed;
122         int nblocks = handle->h_buffer_credits;
123         transaction_t *new_transaction = NULL;
124
125         if (nblocks > journal->j_max_transaction_buffers) {
126                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
127                        current->comm, nblocks,
128                        journal->j_max_transaction_buffers);
129                 return -ENOSPC;
130         }
131
132 alloc_transaction:
133         if (!journal->j_running_transaction) {
134                 new_transaction = kzalloc(sizeof(*new_transaction), gfp_mask);
135                 if (!new_transaction) {
136                         /*
137                          * If __GFP_FS is not present, then we may be
138                          * being called from inside the fs writeback
139                          * layer, so we MUST NOT fail.  Since
140                          * __GFP_NOFAIL is going away, we will arrange
141                          * to retry the allocation ourselves.
142                          */
143                         if ((gfp_mask & __GFP_FS) == 0) {
144                                 congestion_wait(BLK_RW_ASYNC, HZ/50);
145                                 goto alloc_transaction;
146                         }
147                         return -ENOMEM;
148                 }
149         }
150
151         jbd_debug(3, "New handle %p going live.\n", handle);
152
153         /*
154          * We need to hold j_state_lock until t_updates has been incremented,
155          * for proper journal barrier handling
156          */
157 repeat:
158         read_lock(&journal->j_state_lock);
159         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
160         if (is_journal_aborted(journal) ||
161             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
162                 read_unlock(&journal->j_state_lock);
163                 kfree(new_transaction);
164                 return -EROFS;
165         }
166
167         /* Wait on the journal's transaction barrier if necessary */
168         if (journal->j_barrier_count) {
169                 read_unlock(&journal->j_state_lock);
170                 wait_event(journal->j_wait_transaction_locked,
171                                 journal->j_barrier_count == 0);
172                 goto repeat;
173         }
174
175         if (!journal->j_running_transaction) {
176                 read_unlock(&journal->j_state_lock);
177                 if (!new_transaction)
178                         goto alloc_transaction;
179                 write_lock(&journal->j_state_lock);
180                 if (!journal->j_running_transaction) {
181                         jbd2_get_transaction(journal, new_transaction);
182                         new_transaction = NULL;
183                 }
184                 write_unlock(&journal->j_state_lock);
185                 goto repeat;
186         }
187
188         transaction = journal->j_running_transaction;
189
190         /*
191          * If the current transaction is locked down for commit, wait for the
192          * lock to be released.
193          */
194         if (transaction->t_state == T_LOCKED) {
195                 DEFINE_WAIT(wait);
196
197                 prepare_to_wait(&journal->j_wait_transaction_locked,
198                                         &wait, TASK_UNINTERRUPTIBLE);
199                 read_unlock(&journal->j_state_lock);
200                 schedule();
201                 finish_wait(&journal->j_wait_transaction_locked, &wait);
202                 goto repeat;
203         }
204
205         /*
206          * If there is not enough space left in the log to write all potential
207          * buffers requested by this operation, we need to stall pending a log
208          * checkpoint to free some more log space.
209          */
210         needed = atomic_add_return(nblocks,
211                                    &transaction->t_outstanding_credits);
212
213         if (needed > journal->j_max_transaction_buffers) {
214                 /*
215                  * If the current transaction is already too large, then start
216                  * to commit it: we can then go back and attach this handle to
217                  * a new transaction.
218                  */
219                 DEFINE_WAIT(wait);
220
221                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
222                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
223                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
224                                 TASK_UNINTERRUPTIBLE);
225                 __jbd2_log_start_commit(journal, transaction->t_tid);
226                 read_unlock(&journal->j_state_lock);
227                 schedule();
228                 finish_wait(&journal->j_wait_transaction_locked, &wait);
229                 goto repeat;
230         }
231
232         /*
233          * The commit code assumes that it can get enough log space
234          * without forcing a checkpoint.  This is *critical* for
235          * correctness: a checkpoint of a buffer which is also
236          * associated with a committing transaction creates a deadlock,
237          * so commit simply cannot force through checkpoints.
238          *
239          * We must therefore ensure the necessary space in the journal
240          * *before* starting to dirty potentially checkpointed buffers
241          * in the new transaction.
242          *
243          * The worst part is, any transaction currently committing can
244          * reduce the free space arbitrarily.  Be careful to account for
245          * those buffers when checkpointing.
246          */
247
248         /*
249          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
250          * a _lot_ of headroom: 1/4 of the journal plus the size of
251          * the committing transaction.  Really, we only need to give it
252          * committing_transaction->t_outstanding_credits plus "enough" for
253          * the log control blocks.
254          * Also, this test is inconsistent with the matching one in
255          * jbd2_journal_extend().
256          */
257         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
258                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
259                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
260                 read_unlock(&journal->j_state_lock);
261                 write_lock(&journal->j_state_lock);
262                 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
263                         __jbd2_log_wait_for_space(journal);
264                 write_unlock(&journal->j_state_lock);
265                 goto repeat;
266         }
267
268         /* OK, account for the buffers that this operation expects to
269          * use and add the handle to the running transaction. 
270          */
271         update_t_max_wait(transaction);
272         handle->h_transaction = transaction;
273         atomic_inc(&transaction->t_updates);
274         atomic_inc(&transaction->t_handle_count);
275         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
276                   handle, nblocks,
277                   atomic_read(&transaction->t_outstanding_credits),
278                   __jbd2_log_space_left(journal));
279         read_unlock(&journal->j_state_lock);
280
281         lock_map_acquire(&handle->h_lockdep_map);
282         kfree(new_transaction);
283         return 0;
284 }
285
286 static struct lock_class_key jbd2_handle_key;
287
288 /* Allocate a new handle.  This should probably be in a slab... */
289 static handle_t *new_handle(int nblocks)
290 {
291         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
292         if (!handle)
293                 return NULL;
294         memset(handle, 0, sizeof(*handle));
295         handle->h_buffer_credits = nblocks;
296         handle->h_ref = 1;
297
298         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
299                                                 &jbd2_handle_key, 0);
300
301         return handle;
302 }
303
304 /**
305  * handle_t *jbd2_journal_start() - Obtain a new handle.
306  * @journal: Journal to start transaction on.
307  * @nblocks: number of block buffer we might modify
308  *
309  * We make sure that the transaction can guarantee at least nblocks of
310  * modified buffers in the log.  We block until the log can guarantee
311  * that much space.
312  *
313  * This function is visible to journal users (like ext3fs), so is not
314  * called with the journal already locked.
315  *
316  * Return a pointer to a newly allocated handle, or NULL on failure
317  */
318 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int gfp_mask)
319 {
320         handle_t *handle = journal_current_handle();
321         int err;
322
323         if (!journal)
324                 return ERR_PTR(-EROFS);
325
326         if (handle) {
327                 J_ASSERT(handle->h_transaction->t_journal == journal);
328                 handle->h_ref++;
329                 return handle;
330         }
331
332         handle = new_handle(nblocks);
333         if (!handle)
334                 return ERR_PTR(-ENOMEM);
335
336         current->journal_info = handle;
337
338         err = start_this_handle(journal, handle, gfp_mask);
339         if (err < 0) {
340                 jbd2_free_handle(handle);
341                 current->journal_info = NULL;
342                 handle = ERR_PTR(err);
343         }
344         return handle;
345 }
346 EXPORT_SYMBOL(jbd2__journal_start);
347
348
349 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
350 {
351         return jbd2__journal_start(journal, nblocks, GFP_NOFS);
352 }
353 EXPORT_SYMBOL(jbd2_journal_start);
354
355
356 /**
357  * int jbd2_journal_extend() - extend buffer credits.
358  * @handle:  handle to 'extend'
359  * @nblocks: nr blocks to try to extend by.
360  *
361  * Some transactions, such as large extends and truncates, can be done
362  * atomically all at once or in several stages.  The operation requests
363  * a credit for a number of buffer modications in advance, but can
364  * extend its credit if it needs more.
365  *
366  * jbd2_journal_extend tries to give the running handle more buffer credits.
367  * It does not guarantee that allocation - this is a best-effort only.
368  * The calling process MUST be able to deal cleanly with a failure to
369  * extend here.
370  *
371  * Return 0 on success, non-zero on failure.
372  *
373  * return code < 0 implies an error
374  * return code > 0 implies normal transaction-full status.
375  */
376 int jbd2_journal_extend(handle_t *handle, int nblocks)
377 {
378         transaction_t *transaction = handle->h_transaction;
379         journal_t *journal = transaction->t_journal;
380         int result;
381         int wanted;
382
383         result = -EIO;
384         if (is_handle_aborted(handle))
385                 goto out;
386
387         result = 1;
388
389         read_lock(&journal->j_state_lock);
390
391         /* Don't extend a locked-down transaction! */
392         if (handle->h_transaction->t_state != T_RUNNING) {
393                 jbd_debug(3, "denied handle %p %d blocks: "
394                           "transaction not running\n", handle, nblocks);
395                 goto error_out;
396         }
397
398         spin_lock(&transaction->t_handle_lock);
399         wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
400
401         if (wanted > journal->j_max_transaction_buffers) {
402                 jbd_debug(3, "denied handle %p %d blocks: "
403                           "transaction too large\n", handle, nblocks);
404                 goto unlock;
405         }
406
407         if (wanted > __jbd2_log_space_left(journal)) {
408                 jbd_debug(3, "denied handle %p %d blocks: "
409                           "insufficient log space\n", handle, nblocks);
410                 goto unlock;
411         }
412
413         handle->h_buffer_credits += nblocks;
414         atomic_add(nblocks, &transaction->t_outstanding_credits);
415         result = 0;
416
417         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
418 unlock:
419         spin_unlock(&transaction->t_handle_lock);
420 error_out:
421         read_unlock(&journal->j_state_lock);
422 out:
423         return result;
424 }
425
426
427 /**
428  * int jbd2_journal_restart() - restart a handle .
429  * @handle:  handle to restart
430  * @nblocks: nr credits requested
431  *
432  * Restart a handle for a multi-transaction filesystem
433  * operation.
434  *
435  * If the jbd2_journal_extend() call above fails to grant new buffer credits
436  * to a running handle, a call to jbd2_journal_restart will commit the
437  * handle's transaction so far and reattach the handle to a new
438  * transaction capabable of guaranteeing the requested number of
439  * credits.
440  */
441 int jbd2__journal_restart(handle_t *handle, int nblocks, int gfp_mask)
442 {
443         transaction_t *transaction = handle->h_transaction;
444         journal_t *journal = transaction->t_journal;
445         int ret;
446
447         /* If we've had an abort of any type, don't even think about
448          * actually doing the restart! */
449         if (is_handle_aborted(handle))
450                 return 0;
451
452         /*
453          * First unlink the handle from its current transaction, and start the
454          * commit on that.
455          */
456         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
457         J_ASSERT(journal_current_handle() == handle);
458
459         read_lock(&journal->j_state_lock);
460         spin_lock(&transaction->t_handle_lock);
461         atomic_sub(handle->h_buffer_credits,
462                    &transaction->t_outstanding_credits);
463         if (atomic_dec_and_test(&transaction->t_updates))
464                 wake_up(&journal->j_wait_updates);
465         spin_unlock(&transaction->t_handle_lock);
466
467         jbd_debug(2, "restarting handle %p\n", handle);
468         __jbd2_log_start_commit(journal, transaction->t_tid);
469         read_unlock(&journal->j_state_lock);
470
471         lock_map_release(&handle->h_lockdep_map);
472         handle->h_buffer_credits = nblocks;
473         ret = start_this_handle(journal, handle, gfp_mask);
474         return ret;
475 }
476 EXPORT_SYMBOL(jbd2__journal_restart);
477
478
479 int jbd2_journal_restart(handle_t *handle, int nblocks)
480 {
481         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
482 }
483 EXPORT_SYMBOL(jbd2_journal_restart);
484
485 /**
486  * void jbd2_journal_lock_updates () - establish a transaction barrier.
487  * @journal:  Journal to establish a barrier on.
488  *
489  * This locks out any further updates from being started, and blocks
490  * until all existing updates have completed, returning only once the
491  * journal is in a quiescent state with no updates running.
492  *
493  * The journal lock should not be held on entry.
494  */
495 void jbd2_journal_lock_updates(journal_t *journal)
496 {
497         DEFINE_WAIT(wait);
498
499         write_lock(&journal->j_state_lock);
500         ++journal->j_barrier_count;
501
502         /* Wait until there are no running updates */
503         while (1) {
504                 transaction_t *transaction = journal->j_running_transaction;
505
506                 if (!transaction)
507                         break;
508
509                 spin_lock(&transaction->t_handle_lock);
510                 if (!atomic_read(&transaction->t_updates)) {
511                         spin_unlock(&transaction->t_handle_lock);
512                         break;
513                 }
514                 prepare_to_wait(&journal->j_wait_updates, &wait,
515                                 TASK_UNINTERRUPTIBLE);
516                 spin_unlock(&transaction->t_handle_lock);
517                 write_unlock(&journal->j_state_lock);
518                 schedule();
519                 finish_wait(&journal->j_wait_updates, &wait);
520                 write_lock(&journal->j_state_lock);
521         }
522         write_unlock(&journal->j_state_lock);
523
524         /*
525          * We have now established a barrier against other normal updates, but
526          * we also need to barrier against other jbd2_journal_lock_updates() calls
527          * to make sure that we serialise special journal-locked operations
528          * too.
529          */
530         mutex_lock(&journal->j_barrier);
531 }
532
533 /**
534  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
535  * @journal:  Journal to release the barrier on.
536  *
537  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
538  *
539  * Should be called without the journal lock held.
540  */
541 void jbd2_journal_unlock_updates (journal_t *journal)
542 {
543         J_ASSERT(journal->j_barrier_count != 0);
544
545         mutex_unlock(&journal->j_barrier);
546         write_lock(&journal->j_state_lock);
547         --journal->j_barrier_count;
548         write_unlock(&journal->j_state_lock);
549         wake_up(&journal->j_wait_transaction_locked);
550 }
551
552 static void warn_dirty_buffer(struct buffer_head *bh)
553 {
554         char b[BDEVNAME_SIZE];
555
556         printk(KERN_WARNING
557                "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
558                "There's a risk of filesystem corruption in case of system "
559                "crash.\n",
560                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
561 }
562
563 /*
564  * If the buffer is already part of the current transaction, then there
565  * is nothing we need to do.  If it is already part of a prior
566  * transaction which we are still committing to disk, then we need to
567  * make sure that we do not overwrite the old copy: we do copy-out to
568  * preserve the copy going to disk.  We also account the buffer against
569  * the handle's metadata buffer credits (unless the buffer is already
570  * part of the transaction, that is).
571  *
572  */
573 static int
574 do_get_write_access(handle_t *handle, struct journal_head *jh,
575                         int force_copy)
576 {
577         struct buffer_head *bh;
578         transaction_t *transaction;
579         journal_t *journal;
580         int error;
581         char *frozen_buffer = NULL;
582         int need_copy = 0;
583
584         if (is_handle_aborted(handle))
585                 return -EROFS;
586
587         transaction = handle->h_transaction;
588         journal = transaction->t_journal;
589
590         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
591
592         JBUFFER_TRACE(jh, "entry");
593 repeat:
594         bh = jh2bh(jh);
595
596         /* @@@ Need to check for errors here at some point. */
597
598         lock_buffer(bh);
599         jbd_lock_bh_state(bh);
600
601         /* We now hold the buffer lock so it is safe to query the buffer
602          * state.  Is the buffer dirty?
603          *
604          * If so, there are two possibilities.  The buffer may be
605          * non-journaled, and undergoing a quite legitimate writeback.
606          * Otherwise, it is journaled, and we don't expect dirty buffers
607          * in that state (the buffers should be marked JBD_Dirty
608          * instead.)  So either the IO is being done under our own
609          * control and this is a bug, or it's a third party IO such as
610          * dump(8) (which may leave the buffer scheduled for read ---
611          * ie. locked but not dirty) or tune2fs (which may actually have
612          * the buffer dirtied, ugh.)  */
613
614         if (buffer_dirty(bh)) {
615                 /*
616                  * First question: is this buffer already part of the current
617                  * transaction or the existing committing transaction?
618                  */
619                 if (jh->b_transaction) {
620                         J_ASSERT_JH(jh,
621                                 jh->b_transaction == transaction ||
622                                 jh->b_transaction ==
623                                         journal->j_committing_transaction);
624                         if (jh->b_next_transaction)
625                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
626                                                         transaction);
627                         warn_dirty_buffer(bh);
628                 }
629                 /*
630                  * In any case we need to clean the dirty flag and we must
631                  * do it under the buffer lock to be sure we don't race
632                  * with running write-out.
633                  */
634                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
635                 clear_buffer_dirty(bh);
636                 set_buffer_jbddirty(bh);
637         }
638
639         unlock_buffer(bh);
640
641         error = -EROFS;
642         if (is_handle_aborted(handle)) {
643                 jbd_unlock_bh_state(bh);
644                 goto out;
645         }
646         error = 0;
647
648         /*
649          * The buffer is already part of this transaction if b_transaction or
650          * b_next_transaction points to it
651          */
652         if (jh->b_transaction == transaction ||
653             jh->b_next_transaction == transaction)
654                 goto done;
655
656         /*
657          * this is the first time this transaction is touching this buffer,
658          * reset the modified flag
659          */
660        jh->b_modified = 0;
661
662         /*
663          * If there is already a copy-out version of this buffer, then we don't
664          * need to make another one
665          */
666         if (jh->b_frozen_data) {
667                 JBUFFER_TRACE(jh, "has frozen data");
668                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
669                 jh->b_next_transaction = transaction;
670                 goto done;
671         }
672
673         /* Is there data here we need to preserve? */
674
675         if (jh->b_transaction && jh->b_transaction != transaction) {
676                 JBUFFER_TRACE(jh, "owned by older transaction");
677                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
678                 J_ASSERT_JH(jh, jh->b_transaction ==
679                                         journal->j_committing_transaction);
680
681                 /* There is one case we have to be very careful about.
682                  * If the committing transaction is currently writing
683                  * this buffer out to disk and has NOT made a copy-out,
684                  * then we cannot modify the buffer contents at all
685                  * right now.  The essence of copy-out is that it is the
686                  * extra copy, not the primary copy, which gets
687                  * journaled.  If the primary copy is already going to
688                  * disk then we cannot do copy-out here. */
689
690                 if (jh->b_jlist == BJ_Shadow) {
691                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
692                         wait_queue_head_t *wqh;
693
694                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
695
696                         JBUFFER_TRACE(jh, "on shadow: sleep");
697                         jbd_unlock_bh_state(bh);
698                         /* commit wakes up all shadow buffers after IO */
699                         for ( ; ; ) {
700                                 prepare_to_wait(wqh, &wait.wait,
701                                                 TASK_UNINTERRUPTIBLE);
702                                 if (jh->b_jlist != BJ_Shadow)
703                                         break;
704                                 schedule();
705                         }
706                         finish_wait(wqh, &wait.wait);
707                         goto repeat;
708                 }
709
710                 /* Only do the copy if the currently-owning transaction
711                  * still needs it.  If it is on the Forget list, the
712                  * committing transaction is past that stage.  The
713                  * buffer had better remain locked during the kmalloc,
714                  * but that should be true --- we hold the journal lock
715                  * still and the buffer is already on the BUF_JOURNAL
716                  * list so won't be flushed.
717                  *
718                  * Subtle point, though: if this is a get_undo_access,
719                  * then we will be relying on the frozen_data to contain
720                  * the new value of the committed_data record after the
721                  * transaction, so we HAVE to force the frozen_data copy
722                  * in that case. */
723
724                 if (jh->b_jlist != BJ_Forget || force_copy) {
725                         JBUFFER_TRACE(jh, "generate frozen data");
726                         if (!frozen_buffer) {
727                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
728                                 jbd_unlock_bh_state(bh);
729                                 frozen_buffer =
730                                         jbd2_alloc(jh2bh(jh)->b_size,
731                                                          GFP_NOFS);
732                                 if (!frozen_buffer) {
733                                         printk(KERN_EMERG
734                                                "%s: OOM for frozen_buffer\n",
735                                                __func__);
736                                         JBUFFER_TRACE(jh, "oom!");
737                                         error = -ENOMEM;
738                                         jbd_lock_bh_state(bh);
739                                         goto done;
740                                 }
741                                 goto repeat;
742                         }
743                         jh->b_frozen_data = frozen_buffer;
744                         frozen_buffer = NULL;
745                         need_copy = 1;
746                 }
747                 jh->b_next_transaction = transaction;
748         }
749
750
751         /*
752          * Finally, if the buffer is not journaled right now, we need to make
753          * sure it doesn't get written to disk before the caller actually
754          * commits the new data
755          */
756         if (!jh->b_transaction) {
757                 JBUFFER_TRACE(jh, "no transaction");
758                 J_ASSERT_JH(jh, !jh->b_next_transaction);
759                 jh->b_transaction = transaction;
760                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
761                 spin_lock(&journal->j_list_lock);
762                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
763                 spin_unlock(&journal->j_list_lock);
764         }
765
766 done:
767         if (need_copy) {
768                 struct page *page;
769                 int offset;
770                 char *source;
771
772                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
773                             "Possible IO failure.\n");
774                 page = jh2bh(jh)->b_page;
775                 offset = offset_in_page(jh2bh(jh)->b_data);
776                 source = kmap_atomic(page, KM_USER0);
777                 /* Fire data frozen trigger just before we copy the data */
778                 jbd2_buffer_frozen_trigger(jh, source + offset,
779                                            jh->b_triggers);
780                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
781                 kunmap_atomic(source, KM_USER0);
782
783                 /*
784                  * Now that the frozen data is saved off, we need to store
785                  * any matching triggers.
786                  */
787                 jh->b_frozen_triggers = jh->b_triggers;
788         }
789         jbd_unlock_bh_state(bh);
790
791         /*
792          * If we are about to journal a buffer, then any revoke pending on it is
793          * no longer valid
794          */
795         jbd2_journal_cancel_revoke(handle, jh);
796
797 out:
798         if (unlikely(frozen_buffer))    /* It's usually NULL */
799                 jbd2_free(frozen_buffer, bh->b_size);
800
801         JBUFFER_TRACE(jh, "exit");
802         return error;
803 }
804
805 /**
806  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
807  * @handle: transaction to add buffer modifications to
808  * @bh:     bh to be used for metadata writes
809  * @credits: variable that will receive credits for the buffer
810  *
811  * Returns an error code or 0 on success.
812  *
813  * In full data journalling mode the buffer may be of type BJ_AsyncData,
814  * because we're write()ing a buffer which is also part of a shared mapping.
815  */
816
817 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
818 {
819         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
820         int rc;
821
822         /* We do not want to get caught playing with fields which the
823          * log thread also manipulates.  Make sure that the buffer
824          * completes any outstanding IO before proceeding. */
825         rc = do_get_write_access(handle, jh, 0);
826         jbd2_journal_put_journal_head(jh);
827         return rc;
828 }
829
830
831 /*
832  * When the user wants to journal a newly created buffer_head
833  * (ie. getblk() returned a new buffer and we are going to populate it
834  * manually rather than reading off disk), then we need to keep the
835  * buffer_head locked until it has been completely filled with new
836  * data.  In this case, we should be able to make the assertion that
837  * the bh is not already part of an existing transaction.
838  *
839  * The buffer should already be locked by the caller by this point.
840  * There is no lock ranking violation: it was a newly created,
841  * unlocked buffer beforehand. */
842
843 /**
844  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
845  * @handle: transaction to new buffer to
846  * @bh: new buffer.
847  *
848  * Call this if you create a new bh.
849  */
850 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
851 {
852         transaction_t *transaction = handle->h_transaction;
853         journal_t *journal = transaction->t_journal;
854         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
855         int err;
856
857         jbd_debug(5, "journal_head %p\n", jh);
858         err = -EROFS;
859         if (is_handle_aborted(handle))
860                 goto out;
861         err = 0;
862
863         JBUFFER_TRACE(jh, "entry");
864         /*
865          * The buffer may already belong to this transaction due to pre-zeroing
866          * in the filesystem's new_block code.  It may also be on the previous,
867          * committing transaction's lists, but it HAS to be in Forget state in
868          * that case: the transaction must have deleted the buffer for it to be
869          * reused here.
870          */
871         jbd_lock_bh_state(bh);
872         spin_lock(&journal->j_list_lock);
873         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
874                 jh->b_transaction == NULL ||
875                 (jh->b_transaction == journal->j_committing_transaction &&
876                           jh->b_jlist == BJ_Forget)));
877
878         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
879         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
880
881         if (jh->b_transaction == NULL) {
882                 /*
883                  * Previous jbd2_journal_forget() could have left the buffer
884                  * with jbddirty bit set because it was being committed. When
885                  * the commit finished, we've filed the buffer for
886                  * checkpointing and marked it dirty. Now we are reallocating
887                  * the buffer so the transaction freeing it must have
888                  * committed and so it's safe to clear the dirty bit.
889                  */
890                 clear_buffer_dirty(jh2bh(jh));
891                 jh->b_transaction = transaction;
892
893                 /* first access by this transaction */
894                 jh->b_modified = 0;
895
896                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
897                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
898         } else if (jh->b_transaction == journal->j_committing_transaction) {
899                 /* first access by this transaction */
900                 jh->b_modified = 0;
901
902                 JBUFFER_TRACE(jh, "set next transaction");
903                 jh->b_next_transaction = transaction;
904         }
905         spin_unlock(&journal->j_list_lock);
906         jbd_unlock_bh_state(bh);
907
908         /*
909          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
910          * blocks which contain freed but then revoked metadata.  We need
911          * to cancel the revoke in case we end up freeing it yet again
912          * and the reallocating as data - this would cause a second revoke,
913          * which hits an assertion error.
914          */
915         JBUFFER_TRACE(jh, "cancelling revoke");
916         jbd2_journal_cancel_revoke(handle, jh);
917         jbd2_journal_put_journal_head(jh);
918 out:
919         return err;
920 }
921
922 /**
923  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
924  *     non-rewindable consequences
925  * @handle: transaction
926  * @bh: buffer to undo
927  * @credits: store the number of taken credits here (if not NULL)
928  *
929  * Sometimes there is a need to distinguish between metadata which has
930  * been committed to disk and that which has not.  The ext3fs code uses
931  * this for freeing and allocating space, we have to make sure that we
932  * do not reuse freed space until the deallocation has been committed,
933  * since if we overwrote that space we would make the delete
934  * un-rewindable in case of a crash.
935  *
936  * To deal with that, jbd2_journal_get_undo_access requests write access to a
937  * buffer for parts of non-rewindable operations such as delete
938  * operations on the bitmaps.  The journaling code must keep a copy of
939  * the buffer's contents prior to the undo_access call until such time
940  * as we know that the buffer has definitely been committed to disk.
941  *
942  * We never need to know which transaction the committed data is part
943  * of, buffers touched here are guaranteed to be dirtied later and so
944  * will be committed to a new transaction in due course, at which point
945  * we can discard the old committed data pointer.
946  *
947  * Returns error number or 0 on success.
948  */
949 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
950 {
951         int err;
952         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
953         char *committed_data = NULL;
954
955         JBUFFER_TRACE(jh, "entry");
956
957         /*
958          * Do this first --- it can drop the journal lock, so we want to
959          * make sure that obtaining the committed_data is done
960          * atomically wrt. completion of any outstanding commits.
961          */
962         err = do_get_write_access(handle, jh, 1);
963         if (err)
964                 goto out;
965
966 repeat:
967         if (!jh->b_committed_data) {
968                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
969                 if (!committed_data) {
970                         printk(KERN_EMERG "%s: No memory for committed data\n",
971                                 __func__);
972                         err = -ENOMEM;
973                         goto out;
974                 }
975         }
976
977         jbd_lock_bh_state(bh);
978         if (!jh->b_committed_data) {
979                 /* Copy out the current buffer contents into the
980                  * preserved, committed copy. */
981                 JBUFFER_TRACE(jh, "generate b_committed data");
982                 if (!committed_data) {
983                         jbd_unlock_bh_state(bh);
984                         goto repeat;
985                 }
986
987                 jh->b_committed_data = committed_data;
988                 committed_data = NULL;
989                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
990         }
991         jbd_unlock_bh_state(bh);
992 out:
993         jbd2_journal_put_journal_head(jh);
994         if (unlikely(committed_data))
995                 jbd2_free(committed_data, bh->b_size);
996         return err;
997 }
998
999 /**
1000  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1001  * @bh: buffer to trigger on
1002  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1003  *
1004  * Set any triggers on this journal_head.  This is always safe, because
1005  * triggers for a committing buffer will be saved off, and triggers for
1006  * a running transaction will match the buffer in that transaction.
1007  *
1008  * Call with NULL to clear the triggers.
1009  */
1010 void jbd2_journal_set_triggers(struct buffer_head *bh,
1011                                struct jbd2_buffer_trigger_type *type)
1012 {
1013         struct journal_head *jh = bh2jh(bh);
1014
1015         jh->b_triggers = type;
1016 }
1017
1018 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1019                                 struct jbd2_buffer_trigger_type *triggers)
1020 {
1021         struct buffer_head *bh = jh2bh(jh);
1022
1023         if (!triggers || !triggers->t_frozen)
1024                 return;
1025
1026         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1027 }
1028
1029 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1030                                struct jbd2_buffer_trigger_type *triggers)
1031 {
1032         if (!triggers || !triggers->t_abort)
1033                 return;
1034
1035         triggers->t_abort(triggers, jh2bh(jh));
1036 }
1037
1038
1039
1040 /**
1041  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1042  * @handle: transaction to add buffer to.
1043  * @bh: buffer to mark
1044  *
1045  * mark dirty metadata which needs to be journaled as part of the current
1046  * transaction.
1047  *
1048  * The buffer is placed on the transaction's metadata list and is marked
1049  * as belonging to the transaction.
1050  *
1051  * Returns error number or 0 on success.
1052  *
1053  * Special care needs to be taken if the buffer already belongs to the
1054  * current committing transaction (in which case we should have frozen
1055  * data present for that commit).  In that case, we don't relink the
1056  * buffer: that only gets done when the old transaction finally
1057  * completes its commit.
1058  */
1059 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1060 {
1061         transaction_t *transaction = handle->h_transaction;
1062         journal_t *journal = transaction->t_journal;
1063         struct journal_head *jh = bh2jh(bh);
1064
1065         jbd_debug(5, "journal_head %p\n", jh);
1066         JBUFFER_TRACE(jh, "entry");
1067         if (is_handle_aborted(handle))
1068                 goto out;
1069
1070         jbd_lock_bh_state(bh);
1071
1072         if (jh->b_modified == 0) {
1073                 /*
1074                  * This buffer's got modified and becoming part
1075                  * of the transaction. This needs to be done
1076                  * once a transaction -bzzz
1077                  */
1078                 jh->b_modified = 1;
1079                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1080                 handle->h_buffer_credits--;
1081         }
1082
1083         /*
1084          * fastpath, to avoid expensive locking.  If this buffer is already
1085          * on the running transaction's metadata list there is nothing to do.
1086          * Nobody can take it off again because there is a handle open.
1087          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1088          * result in this test being false, so we go in and take the locks.
1089          */
1090         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1091                 JBUFFER_TRACE(jh, "fastpath");
1092                 J_ASSERT_JH(jh, jh->b_transaction ==
1093                                         journal->j_running_transaction);
1094                 goto out_unlock_bh;
1095         }
1096
1097         set_buffer_jbddirty(bh);
1098
1099         /*
1100          * Metadata already on the current transaction list doesn't
1101          * need to be filed.  Metadata on another transaction's list must
1102          * be committing, and will be refiled once the commit completes:
1103          * leave it alone for now.
1104          */
1105         if (jh->b_transaction != transaction) {
1106                 JBUFFER_TRACE(jh, "already on other transaction");
1107                 J_ASSERT_JH(jh, jh->b_transaction ==
1108                                         journal->j_committing_transaction);
1109                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1110                 /* And this case is illegal: we can't reuse another
1111                  * transaction's data buffer, ever. */
1112                 goto out_unlock_bh;
1113         }
1114
1115         /* That test should have eliminated the following case: */
1116         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1117
1118         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1119         spin_lock(&journal->j_list_lock);
1120         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1121         spin_unlock(&journal->j_list_lock);
1122 out_unlock_bh:
1123         jbd_unlock_bh_state(bh);
1124 out:
1125         JBUFFER_TRACE(jh, "exit");
1126         return 0;
1127 }
1128
1129 /*
1130  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1131  * updates, if the update decided in the end that it didn't need access.
1132  *
1133  */
1134 void
1135 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1136 {
1137         BUFFER_TRACE(bh, "entry");
1138 }
1139
1140 /**
1141  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1142  * @handle: transaction handle
1143  * @bh:     bh to 'forget'
1144  *
1145  * We can only do the bforget if there are no commits pending against the
1146  * buffer.  If the buffer is dirty in the current running transaction we
1147  * can safely unlink it.
1148  *
1149  * bh may not be a journalled buffer at all - it may be a non-JBD
1150  * buffer which came off the hashtable.  Check for this.
1151  *
1152  * Decrements bh->b_count by one.
1153  *
1154  * Allow this call even if the handle has aborted --- it may be part of
1155  * the caller's cleanup after an abort.
1156  */
1157 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1158 {
1159         transaction_t *transaction = handle->h_transaction;
1160         journal_t *journal = transaction->t_journal;
1161         struct journal_head *jh;
1162         int drop_reserve = 0;
1163         int err = 0;
1164         int was_modified = 0;
1165
1166         BUFFER_TRACE(bh, "entry");
1167
1168         jbd_lock_bh_state(bh);
1169         spin_lock(&journal->j_list_lock);
1170
1171         if (!buffer_jbd(bh))
1172                 goto not_jbd;
1173         jh = bh2jh(bh);
1174
1175         /* Critical error: attempting to delete a bitmap buffer, maybe?
1176          * Don't do any jbd operations, and return an error. */
1177         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1178                          "inconsistent data on disk")) {
1179                 err = -EIO;
1180                 goto not_jbd;
1181         }
1182
1183         /* keep track of wether or not this transaction modified us */
1184         was_modified = jh->b_modified;
1185
1186         /*
1187          * The buffer's going from the transaction, we must drop
1188          * all references -bzzz
1189          */
1190         jh->b_modified = 0;
1191
1192         if (jh->b_transaction == handle->h_transaction) {
1193                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1194
1195                 /* If we are forgetting a buffer which is already part
1196                  * of this transaction, then we can just drop it from
1197                  * the transaction immediately. */
1198                 clear_buffer_dirty(bh);
1199                 clear_buffer_jbddirty(bh);
1200
1201                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1202
1203                 /*
1204                  * we only want to drop a reference if this transaction
1205                  * modified the buffer
1206                  */
1207                 if (was_modified)
1208                         drop_reserve = 1;
1209
1210                 /*
1211                  * We are no longer going to journal this buffer.
1212                  * However, the commit of this transaction is still
1213                  * important to the buffer: the delete that we are now
1214                  * processing might obsolete an old log entry, so by
1215                  * committing, we can satisfy the buffer's checkpoint.
1216                  *
1217                  * So, if we have a checkpoint on the buffer, we should
1218                  * now refile the buffer on our BJ_Forget list so that
1219                  * we know to remove the checkpoint after we commit.
1220                  */
1221
1222                 if (jh->b_cp_transaction) {
1223                         __jbd2_journal_temp_unlink_buffer(jh);
1224                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1225                 } else {
1226                         __jbd2_journal_unfile_buffer(jh);
1227                         jbd2_journal_remove_journal_head(bh);
1228                         __brelse(bh);
1229                         if (!buffer_jbd(bh)) {
1230                                 spin_unlock(&journal->j_list_lock);
1231                                 jbd_unlock_bh_state(bh);
1232                                 __bforget(bh);
1233                                 goto drop;
1234                         }
1235                 }
1236         } else if (jh->b_transaction) {
1237                 J_ASSERT_JH(jh, (jh->b_transaction ==
1238                                  journal->j_committing_transaction));
1239                 /* However, if the buffer is still owned by a prior
1240                  * (committing) transaction, we can't drop it yet... */
1241                 JBUFFER_TRACE(jh, "belongs to older transaction");
1242                 /* ... but we CAN drop it from the new transaction if we
1243                  * have also modified it since the original commit. */
1244
1245                 if (jh->b_next_transaction) {
1246                         J_ASSERT(jh->b_next_transaction == transaction);
1247                         jh->b_next_transaction = NULL;
1248
1249                         /*
1250                          * only drop a reference if this transaction modified
1251                          * the buffer
1252                          */
1253                         if (was_modified)
1254                                 drop_reserve = 1;
1255                 }
1256         }
1257
1258 not_jbd:
1259         spin_unlock(&journal->j_list_lock);
1260         jbd_unlock_bh_state(bh);
1261         __brelse(bh);
1262 drop:
1263         if (drop_reserve) {
1264                 /* no need to reserve log space for this block -bzzz */
1265                 handle->h_buffer_credits++;
1266         }
1267         return err;
1268 }
1269
1270 /**
1271  * int jbd2_journal_stop() - complete a transaction
1272  * @handle: tranaction to complete.
1273  *
1274  * All done for a particular handle.
1275  *
1276  * There is not much action needed here.  We just return any remaining
1277  * buffer credits to the transaction and remove the handle.  The only
1278  * complication is that we need to start a commit operation if the
1279  * filesystem is marked for synchronous update.
1280  *
1281  * jbd2_journal_stop itself will not usually return an error, but it may
1282  * do so in unusual circumstances.  In particular, expect it to
1283  * return -EIO if a jbd2_journal_abort has been executed since the
1284  * transaction began.
1285  */
1286 int jbd2_journal_stop(handle_t *handle)
1287 {
1288         transaction_t *transaction = handle->h_transaction;
1289         journal_t *journal = transaction->t_journal;
1290         int err, wait_for_commit = 0;
1291         tid_t tid;
1292         pid_t pid;
1293
1294         J_ASSERT(journal_current_handle() == handle);
1295
1296         if (is_handle_aborted(handle))
1297                 err = -EIO;
1298         else {
1299                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1300                 err = 0;
1301         }
1302
1303         if (--handle->h_ref > 0) {
1304                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1305                           handle->h_ref);
1306                 return err;
1307         }
1308
1309         jbd_debug(4, "Handle %p going down\n", handle);
1310
1311         /*
1312          * Implement synchronous transaction batching.  If the handle
1313          * was synchronous, don't force a commit immediately.  Let's
1314          * yield and let another thread piggyback onto this
1315          * transaction.  Keep doing that while new threads continue to
1316          * arrive.  It doesn't cost much - we're about to run a commit
1317          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1318          * operations by 30x or more...
1319          *
1320          * We try and optimize the sleep time against what the
1321          * underlying disk can do, instead of having a static sleep
1322          * time.  This is useful for the case where our storage is so
1323          * fast that it is more optimal to go ahead and force a flush
1324          * and wait for the transaction to be committed than it is to
1325          * wait for an arbitrary amount of time for new writers to
1326          * join the transaction.  We achieve this by measuring how
1327          * long it takes to commit a transaction, and compare it with
1328          * how long this transaction has been running, and if run time
1329          * < commit time then we sleep for the delta and commit.  This
1330          * greatly helps super fast disks that would see slowdowns as
1331          * more threads started doing fsyncs.
1332          *
1333          * But don't do this if this process was the most recent one
1334          * to perform a synchronous write.  We do this to detect the
1335          * case where a single process is doing a stream of sync
1336          * writes.  No point in waiting for joiners in that case.
1337          */
1338         pid = current->pid;
1339         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1340                 u64 commit_time, trans_time;
1341
1342                 journal->j_last_sync_writer = pid;
1343
1344                 read_lock(&journal->j_state_lock);
1345                 commit_time = journal->j_average_commit_time;
1346                 read_unlock(&journal->j_state_lock);
1347
1348                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1349                                                    transaction->t_start_time));
1350
1351                 commit_time = max_t(u64, commit_time,
1352                                     1000*journal->j_min_batch_time);
1353                 commit_time = min_t(u64, commit_time,
1354                                     1000*journal->j_max_batch_time);
1355
1356                 if (trans_time < commit_time) {
1357                         ktime_t expires = ktime_add_ns(ktime_get(),
1358                                                        commit_time);
1359                         set_current_state(TASK_UNINTERRUPTIBLE);
1360                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1361                 }
1362         }
1363
1364         if (handle->h_sync)
1365                 transaction->t_synchronous_commit = 1;
1366         current->journal_info = NULL;
1367         atomic_sub(handle->h_buffer_credits,
1368                    &transaction->t_outstanding_credits);
1369
1370         /*
1371          * If the handle is marked SYNC, we need to set another commit
1372          * going!  We also want to force a commit if the current
1373          * transaction is occupying too much of the log, or if the
1374          * transaction is too old now.
1375          */
1376         if (handle->h_sync ||
1377             (atomic_read(&transaction->t_outstanding_credits) >
1378              journal->j_max_transaction_buffers) ||
1379             time_after_eq(jiffies, transaction->t_expires)) {
1380                 /* Do this even for aborted journals: an abort still
1381                  * completes the commit thread, it just doesn't write
1382                  * anything to disk. */
1383
1384                 jbd_debug(2, "transaction too old, requesting commit for "
1385                                         "handle %p\n", handle);
1386                 /* This is non-blocking */
1387                 jbd2_log_start_commit(journal, transaction->t_tid);
1388
1389                 /*
1390                  * Special case: JBD2_SYNC synchronous updates require us
1391                  * to wait for the commit to complete.
1392                  */
1393                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1394                         wait_for_commit = 1;
1395         }
1396
1397         /*
1398          * Once we drop t_updates, if it goes to zero the transaction
1399          * could start commiting on us and eventually disappear.  So
1400          * once we do this, we must not dereference transaction
1401          * pointer again.
1402          */
1403         tid = transaction->t_tid;
1404         if (atomic_dec_and_test(&transaction->t_updates)) {
1405                 wake_up(&journal->j_wait_updates);
1406                 if (journal->j_barrier_count)
1407                         wake_up(&journal->j_wait_transaction_locked);
1408         }
1409
1410         if (wait_for_commit)
1411                 err = jbd2_log_wait_commit(journal, tid);
1412
1413         lock_map_release(&handle->h_lockdep_map);
1414
1415         jbd2_free_handle(handle);
1416         return err;
1417 }
1418
1419 /**
1420  * int jbd2_journal_force_commit() - force any uncommitted transactions
1421  * @journal: journal to force
1422  *
1423  * For synchronous operations: force any uncommitted transactions
1424  * to disk.  May seem kludgy, but it reuses all the handle batching
1425  * code in a very simple manner.
1426  */
1427 int jbd2_journal_force_commit(journal_t *journal)
1428 {
1429         handle_t *handle;
1430         int ret;
1431
1432         handle = jbd2_journal_start(journal, 1);
1433         if (IS_ERR(handle)) {
1434                 ret = PTR_ERR(handle);
1435         } else {
1436                 handle->h_sync = 1;
1437                 ret = jbd2_journal_stop(handle);
1438         }
1439         return ret;
1440 }
1441
1442 /*
1443  *
1444  * List management code snippets: various functions for manipulating the
1445  * transaction buffer lists.
1446  *
1447  */
1448
1449 /*
1450  * Append a buffer to a transaction list, given the transaction's list head
1451  * pointer.
1452  *
1453  * j_list_lock is held.
1454  *
1455  * jbd_lock_bh_state(jh2bh(jh)) is held.
1456  */
1457
1458 static inline void
1459 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1460 {
1461         if (!*list) {
1462                 jh->b_tnext = jh->b_tprev = jh;
1463                 *list = jh;
1464         } else {
1465                 /* Insert at the tail of the list to preserve order */
1466                 struct journal_head *first = *list, *last = first->b_tprev;
1467                 jh->b_tprev = last;
1468                 jh->b_tnext = first;
1469                 last->b_tnext = first->b_tprev = jh;
1470         }
1471 }
1472
1473 /*
1474  * Remove a buffer from a transaction list, given the transaction's list
1475  * head pointer.
1476  *
1477  * Called with j_list_lock held, and the journal may not be locked.
1478  *
1479  * jbd_lock_bh_state(jh2bh(jh)) is held.
1480  */
1481
1482 static inline void
1483 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1484 {
1485         if (*list == jh) {
1486                 *list = jh->b_tnext;
1487                 if (*list == jh)
1488                         *list = NULL;
1489         }
1490         jh->b_tprev->b_tnext = jh->b_tnext;
1491         jh->b_tnext->b_tprev = jh->b_tprev;
1492 }
1493
1494 /*
1495  * Remove a buffer from the appropriate transaction list.
1496  *
1497  * Note that this function can *change* the value of
1498  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1499  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1500  * of these pointers, it could go bad.  Generally the caller needs to re-read
1501  * the pointer from the transaction_t.
1502  *
1503  * Called under j_list_lock.  The journal may not be locked.
1504  */
1505 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1506 {
1507         struct journal_head **list = NULL;
1508         transaction_t *transaction;
1509         struct buffer_head *bh = jh2bh(jh);
1510
1511         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1512         transaction = jh->b_transaction;
1513         if (transaction)
1514                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1515
1516         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1517         if (jh->b_jlist != BJ_None)
1518                 J_ASSERT_JH(jh, transaction != NULL);
1519
1520         switch (jh->b_jlist) {
1521         case BJ_None:
1522                 return;
1523         case BJ_Metadata:
1524                 transaction->t_nr_buffers--;
1525                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1526                 list = &transaction->t_buffers;
1527                 break;
1528         case BJ_Forget:
1529                 list = &transaction->t_forget;
1530                 break;
1531         case BJ_IO:
1532                 list = &transaction->t_iobuf_list;
1533                 break;
1534         case BJ_Shadow:
1535                 list = &transaction->t_shadow_list;
1536                 break;
1537         case BJ_LogCtl:
1538                 list = &transaction->t_log_list;
1539                 break;
1540         case BJ_Reserved:
1541                 list = &transaction->t_reserved_list;
1542                 break;
1543         }
1544
1545         __blist_del_buffer(list, jh);
1546         jh->b_jlist = BJ_None;
1547         if (test_clear_buffer_jbddirty(bh))
1548                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1549 }
1550
1551 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1552 {
1553         __jbd2_journal_temp_unlink_buffer(jh);
1554         jh->b_transaction = NULL;
1555 }
1556
1557 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1558 {
1559         jbd_lock_bh_state(jh2bh(jh));
1560         spin_lock(&journal->j_list_lock);
1561         __jbd2_journal_unfile_buffer(jh);
1562         spin_unlock(&journal->j_list_lock);
1563         jbd_unlock_bh_state(jh2bh(jh));
1564 }
1565
1566 /*
1567  * Called from jbd2_journal_try_to_free_buffers().
1568  *
1569  * Called under jbd_lock_bh_state(bh)
1570  */
1571 static void
1572 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1573 {
1574         struct journal_head *jh;
1575
1576         jh = bh2jh(bh);
1577
1578         if (buffer_locked(bh) || buffer_dirty(bh))
1579                 goto out;
1580
1581         if (jh->b_next_transaction != NULL)
1582                 goto out;
1583
1584         spin_lock(&journal->j_list_lock);
1585         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1586                 /* written-back checkpointed metadata buffer */
1587                 if (jh->b_jlist == BJ_None) {
1588                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1589                         __jbd2_journal_remove_checkpoint(jh);
1590                         jbd2_journal_remove_journal_head(bh);
1591                         __brelse(bh);
1592                 }
1593         }
1594         spin_unlock(&journal->j_list_lock);
1595 out:
1596         return;
1597 }
1598
1599 /**
1600  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1601  * @journal: journal for operation
1602  * @page: to try and free
1603  * @gfp_mask: we use the mask to detect how hard should we try to release
1604  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1605  * release the buffers.
1606  *
1607  *
1608  * For all the buffers on this page,
1609  * if they are fully written out ordered data, move them onto BUF_CLEAN
1610  * so try_to_free_buffers() can reap them.
1611  *
1612  * This function returns non-zero if we wish try_to_free_buffers()
1613  * to be called. We do this if the page is releasable by try_to_free_buffers().
1614  * We also do it if the page has locked or dirty buffers and the caller wants
1615  * us to perform sync or async writeout.
1616  *
1617  * This complicates JBD locking somewhat.  We aren't protected by the
1618  * BKL here.  We wish to remove the buffer from its committing or
1619  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1620  *
1621  * This may *change* the value of transaction_t->t_datalist, so anyone
1622  * who looks at t_datalist needs to lock against this function.
1623  *
1624  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1625  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1626  * will come out of the lock with the buffer dirty, which makes it
1627  * ineligible for release here.
1628  *
1629  * Who else is affected by this?  hmm...  Really the only contender
1630  * is do_get_write_access() - it could be looking at the buffer while
1631  * journal_try_to_free_buffer() is changing its state.  But that
1632  * cannot happen because we never reallocate freed data as metadata
1633  * while the data is part of a transaction.  Yes?
1634  *
1635  * Return 0 on failure, 1 on success
1636  */
1637 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1638                                 struct page *page, gfp_t gfp_mask)
1639 {
1640         struct buffer_head *head;
1641         struct buffer_head *bh;
1642         int ret = 0;
1643
1644         J_ASSERT(PageLocked(page));
1645
1646         head = page_buffers(page);
1647         bh = head;
1648         do {
1649                 struct journal_head *jh;
1650
1651                 /*
1652                  * We take our own ref against the journal_head here to avoid
1653                  * having to add tons of locking around each instance of
1654                  * jbd2_journal_remove_journal_head() and
1655                  * jbd2_journal_put_journal_head().
1656                  */
1657                 jh = jbd2_journal_grab_journal_head(bh);
1658                 if (!jh)
1659                         continue;
1660
1661                 jbd_lock_bh_state(bh);
1662                 __journal_try_to_free_buffer(journal, bh);
1663                 jbd2_journal_put_journal_head(jh);
1664                 jbd_unlock_bh_state(bh);
1665                 if (buffer_jbd(bh))
1666                         goto busy;
1667         } while ((bh = bh->b_this_page) != head);
1668
1669         ret = try_to_free_buffers(page);
1670
1671 busy:
1672         return ret;
1673 }
1674
1675 /*
1676  * This buffer is no longer needed.  If it is on an older transaction's
1677  * checkpoint list we need to record it on this transaction's forget list
1678  * to pin this buffer (and hence its checkpointing transaction) down until
1679  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1680  * release it.
1681  * Returns non-zero if JBD no longer has an interest in the buffer.
1682  *
1683  * Called under j_list_lock.
1684  *
1685  * Called under jbd_lock_bh_state(bh).
1686  */
1687 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1688 {
1689         int may_free = 1;
1690         struct buffer_head *bh = jh2bh(jh);
1691
1692         __jbd2_journal_unfile_buffer(jh);
1693
1694         if (jh->b_cp_transaction) {
1695                 JBUFFER_TRACE(jh, "on running+cp transaction");
1696                 /*
1697                  * We don't want to write the buffer anymore, clear the
1698                  * bit so that we don't confuse checks in
1699                  * __journal_file_buffer
1700                  */
1701                 clear_buffer_dirty(bh);
1702                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1703                 may_free = 0;
1704         } else {
1705                 JBUFFER_TRACE(jh, "on running transaction");
1706                 jbd2_journal_remove_journal_head(bh);
1707                 __brelse(bh);
1708         }
1709         return may_free;
1710 }
1711
1712 /*
1713  * jbd2_journal_invalidatepage
1714  *
1715  * This code is tricky.  It has a number of cases to deal with.
1716  *
1717  * There are two invariants which this code relies on:
1718  *
1719  * i_size must be updated on disk before we start calling invalidatepage on the
1720  * data.
1721  *
1722  *  This is done in ext3 by defining an ext3_setattr method which
1723  *  updates i_size before truncate gets going.  By maintaining this
1724  *  invariant, we can be sure that it is safe to throw away any buffers
1725  *  attached to the current transaction: once the transaction commits,
1726  *  we know that the data will not be needed.
1727  *
1728  *  Note however that we can *not* throw away data belonging to the
1729  *  previous, committing transaction!
1730  *
1731  * Any disk blocks which *are* part of the previous, committing
1732  * transaction (and which therefore cannot be discarded immediately) are
1733  * not going to be reused in the new running transaction
1734  *
1735  *  The bitmap committed_data images guarantee this: any block which is
1736  *  allocated in one transaction and removed in the next will be marked
1737  *  as in-use in the committed_data bitmap, so cannot be reused until
1738  *  the next transaction to delete the block commits.  This means that
1739  *  leaving committing buffers dirty is quite safe: the disk blocks
1740  *  cannot be reallocated to a different file and so buffer aliasing is
1741  *  not possible.
1742  *
1743  *
1744  * The above applies mainly to ordered data mode.  In writeback mode we
1745  * don't make guarantees about the order in which data hits disk --- in
1746  * particular we don't guarantee that new dirty data is flushed before
1747  * transaction commit --- so it is always safe just to discard data
1748  * immediately in that mode.  --sct
1749  */
1750
1751 /*
1752  * The journal_unmap_buffer helper function returns zero if the buffer
1753  * concerned remains pinned as an anonymous buffer belonging to an older
1754  * transaction.
1755  *
1756  * We're outside-transaction here.  Either or both of j_running_transaction
1757  * and j_committing_transaction may be NULL.
1758  */
1759 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1760 {
1761         transaction_t *transaction;
1762         struct journal_head *jh;
1763         int may_free = 1;
1764         int ret;
1765
1766         BUFFER_TRACE(bh, "entry");
1767
1768         /*
1769          * It is safe to proceed here without the j_list_lock because the
1770          * buffers cannot be stolen by try_to_free_buffers as long as we are
1771          * holding the page lock. --sct
1772          */
1773
1774         if (!buffer_jbd(bh))
1775                 goto zap_buffer_unlocked;
1776
1777         /* OK, we have data buffer in journaled mode */
1778         write_lock(&journal->j_state_lock);
1779         jbd_lock_bh_state(bh);
1780         spin_lock(&journal->j_list_lock);
1781
1782         jh = jbd2_journal_grab_journal_head(bh);
1783         if (!jh)
1784                 goto zap_buffer_no_jh;
1785
1786         /*
1787          * We cannot remove the buffer from checkpoint lists until the
1788          * transaction adding inode to orphan list (let's call it T)
1789          * is committed.  Otherwise if the transaction changing the
1790          * buffer would be cleaned from the journal before T is
1791          * committed, a crash will cause that the correct contents of
1792          * the buffer will be lost.  On the other hand we have to
1793          * clear the buffer dirty bit at latest at the moment when the
1794          * transaction marking the buffer as freed in the filesystem
1795          * structures is committed because from that moment on the
1796          * buffer can be reallocated and used by a different page.
1797          * Since the block hasn't been freed yet but the inode has
1798          * already been added to orphan list, it is safe for us to add
1799          * the buffer to BJ_Forget list of the newest transaction.
1800          */
1801         transaction = jh->b_transaction;
1802         if (transaction == NULL) {
1803                 /* First case: not on any transaction.  If it
1804                  * has no checkpoint link, then we can zap it:
1805                  * it's a writeback-mode buffer so we don't care
1806                  * if it hits disk safely. */
1807                 if (!jh->b_cp_transaction) {
1808                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1809                         goto zap_buffer;
1810                 }
1811
1812                 if (!buffer_dirty(bh)) {
1813                         /* bdflush has written it.  We can drop it now */
1814                         goto zap_buffer;
1815                 }
1816
1817                 /* OK, it must be in the journal but still not
1818                  * written fully to disk: it's metadata or
1819                  * journaled data... */
1820
1821                 if (journal->j_running_transaction) {
1822                         /* ... and once the current transaction has
1823                          * committed, the buffer won't be needed any
1824                          * longer. */
1825                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1826                         ret = __dispose_buffer(jh,
1827                                         journal->j_running_transaction);
1828                         jbd2_journal_put_journal_head(jh);
1829                         spin_unlock(&journal->j_list_lock);
1830                         jbd_unlock_bh_state(bh);
1831                         write_unlock(&journal->j_state_lock);
1832                         return ret;
1833                 } else {
1834                         /* There is no currently-running transaction. So the
1835                          * orphan record which we wrote for this file must have
1836                          * passed into commit.  We must attach this buffer to
1837                          * the committing transaction, if it exists. */
1838                         if (journal->j_committing_transaction) {
1839                                 JBUFFER_TRACE(jh, "give to committing trans");
1840                                 ret = __dispose_buffer(jh,
1841                                         journal->j_committing_transaction);
1842                                 jbd2_journal_put_journal_head(jh);
1843                                 spin_unlock(&journal->j_list_lock);
1844                                 jbd_unlock_bh_state(bh);
1845                                 write_unlock(&journal->j_state_lock);
1846                                 return ret;
1847                         } else {
1848                                 /* The orphan record's transaction has
1849                                  * committed.  We can cleanse this buffer */
1850                                 clear_buffer_jbddirty(bh);
1851                                 goto zap_buffer;
1852                         }
1853                 }
1854         } else if (transaction == journal->j_committing_transaction) {
1855                 JBUFFER_TRACE(jh, "on committing transaction");
1856                 /*
1857                  * The buffer is committing, we simply cannot touch
1858                  * it. So we just set j_next_transaction to the
1859                  * running transaction (if there is one) and mark
1860                  * buffer as freed so that commit code knows it should
1861                  * clear dirty bits when it is done with the buffer.
1862                  */
1863                 set_buffer_freed(bh);
1864                 if (journal->j_running_transaction && buffer_jbddirty(bh))
1865                         jh->b_next_transaction = journal->j_running_transaction;
1866                 jbd2_journal_put_journal_head(jh);
1867                 spin_unlock(&journal->j_list_lock);
1868                 jbd_unlock_bh_state(bh);
1869                 write_unlock(&journal->j_state_lock);
1870                 return 0;
1871         } else {
1872                 /* Good, the buffer belongs to the running transaction.
1873                  * We are writing our own transaction's data, not any
1874                  * previous one's, so it is safe to throw it away
1875                  * (remember that we expect the filesystem to have set
1876                  * i_size already for this truncate so recovery will not
1877                  * expose the disk blocks we are discarding here.) */
1878                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1879                 JBUFFER_TRACE(jh, "on running transaction");
1880                 may_free = __dispose_buffer(jh, transaction);
1881         }
1882
1883 zap_buffer:
1884         jbd2_journal_put_journal_head(jh);
1885 zap_buffer_no_jh:
1886         spin_unlock(&journal->j_list_lock);
1887         jbd_unlock_bh_state(bh);
1888         write_unlock(&journal->j_state_lock);
1889 zap_buffer_unlocked:
1890         clear_buffer_dirty(bh);
1891         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1892         clear_buffer_mapped(bh);
1893         clear_buffer_req(bh);
1894         clear_buffer_new(bh);
1895         bh->b_bdev = NULL;
1896         return may_free;
1897 }
1898
1899 /**
1900  * void jbd2_journal_invalidatepage()
1901  * @journal: journal to use for flush...
1902  * @page:    page to flush
1903  * @offset:  length of page to invalidate.
1904  *
1905  * Reap page buffers containing data after offset in page.
1906  *
1907  */
1908 void jbd2_journal_invalidatepage(journal_t *journal,
1909                       struct page *page,
1910                       unsigned long offset)
1911 {
1912         struct buffer_head *head, *bh, *next;
1913         unsigned int curr_off = 0;
1914         int may_free = 1;
1915
1916         if (!PageLocked(page))
1917                 BUG();
1918         if (!page_has_buffers(page))
1919                 return;
1920
1921         /* We will potentially be playing with lists other than just the
1922          * data lists (especially for journaled data mode), so be
1923          * cautious in our locking. */
1924
1925         head = bh = page_buffers(page);
1926         do {
1927                 unsigned int next_off = curr_off + bh->b_size;
1928                 next = bh->b_this_page;
1929
1930                 if (offset <= curr_off) {
1931                         /* This block is wholly outside the truncation point */
1932                         lock_buffer(bh);
1933                         may_free &= journal_unmap_buffer(journal, bh);
1934                         unlock_buffer(bh);
1935                 }
1936                 curr_off = next_off;
1937                 bh = next;
1938
1939         } while (bh != head);
1940
1941         if (!offset) {
1942                 if (may_free && try_to_free_buffers(page))
1943                         J_ASSERT(!page_has_buffers(page));
1944         }
1945 }
1946
1947 /*
1948  * File a buffer on the given transaction list.
1949  */
1950 void __jbd2_journal_file_buffer(struct journal_head *jh,
1951                         transaction_t *transaction, int jlist)
1952 {
1953         struct journal_head **list = NULL;
1954         int was_dirty = 0;
1955         struct buffer_head *bh = jh2bh(jh);
1956
1957         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1958         assert_spin_locked(&transaction->t_journal->j_list_lock);
1959
1960         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1961         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1962                                 jh->b_transaction == NULL);
1963
1964         if (jh->b_transaction && jh->b_jlist == jlist)
1965                 return;
1966
1967         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1968             jlist == BJ_Shadow || jlist == BJ_Forget) {
1969                 /*
1970                  * For metadata buffers, we track dirty bit in buffer_jbddirty
1971                  * instead of buffer_dirty. We should not see a dirty bit set
1972                  * here because we clear it in do_get_write_access but e.g.
1973                  * tune2fs can modify the sb and set the dirty bit at any time
1974                  * so we try to gracefully handle that.
1975                  */
1976                 if (buffer_dirty(bh))
1977                         warn_dirty_buffer(bh);
1978                 if (test_clear_buffer_dirty(bh) ||
1979                     test_clear_buffer_jbddirty(bh))
1980                         was_dirty = 1;
1981         }
1982
1983         if (jh->b_transaction)
1984                 __jbd2_journal_temp_unlink_buffer(jh);
1985         jh->b_transaction = transaction;
1986
1987         switch (jlist) {
1988         case BJ_None:
1989                 J_ASSERT_JH(jh, !jh->b_committed_data);
1990                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1991                 return;
1992         case BJ_Metadata:
1993                 transaction->t_nr_buffers++;
1994                 list = &transaction->t_buffers;
1995                 break;
1996         case BJ_Forget:
1997                 list = &transaction->t_forget;
1998                 break;
1999         case BJ_IO:
2000                 list = &transaction->t_iobuf_list;
2001                 break;
2002         case BJ_Shadow:
2003                 list = &transaction->t_shadow_list;
2004                 break;
2005         case BJ_LogCtl:
2006                 list = &transaction->t_log_list;
2007                 break;
2008         case BJ_Reserved:
2009                 list = &transaction->t_reserved_list;
2010                 break;
2011         }
2012
2013         __blist_add_buffer(list, jh);
2014         jh->b_jlist = jlist;
2015
2016         if (was_dirty)
2017                 set_buffer_jbddirty(bh);
2018 }
2019
2020 void jbd2_journal_file_buffer(struct journal_head *jh,
2021                                 transaction_t *transaction, int jlist)
2022 {
2023         jbd_lock_bh_state(jh2bh(jh));
2024         spin_lock(&transaction->t_journal->j_list_lock);
2025         __jbd2_journal_file_buffer(jh, transaction, jlist);
2026         spin_unlock(&transaction->t_journal->j_list_lock);
2027         jbd_unlock_bh_state(jh2bh(jh));
2028 }
2029
2030 /*
2031  * Remove a buffer from its current buffer list in preparation for
2032  * dropping it from its current transaction entirely.  If the buffer has
2033  * already started to be used by a subsequent transaction, refile the
2034  * buffer on that transaction's metadata list.
2035  *
2036  * Called under journal->j_list_lock
2037  *
2038  * Called under jbd_lock_bh_state(jh2bh(jh))
2039  */
2040 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2041 {
2042         int was_dirty, jlist;
2043         struct buffer_head *bh = jh2bh(jh);
2044
2045         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2046         if (jh->b_transaction)
2047                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2048
2049         /* If the buffer is now unused, just drop it. */
2050         if (jh->b_next_transaction == NULL) {
2051                 __jbd2_journal_unfile_buffer(jh);
2052                 return;
2053         }
2054
2055         /*
2056          * It has been modified by a later transaction: add it to the new
2057          * transaction's metadata list.
2058          */
2059
2060         was_dirty = test_clear_buffer_jbddirty(bh);
2061         __jbd2_journal_temp_unlink_buffer(jh);
2062         jh->b_transaction = jh->b_next_transaction;
2063         jh->b_next_transaction = NULL;
2064         if (buffer_freed(bh))
2065                 jlist = BJ_Forget;
2066         else if (jh->b_modified)
2067                 jlist = BJ_Metadata;
2068         else
2069                 jlist = BJ_Reserved;
2070         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2071         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2072
2073         if (was_dirty)
2074                 set_buffer_jbddirty(bh);
2075 }
2076
2077 /*
2078  * For the unlocked version of this call, also make sure that any
2079  * hanging journal_head is cleaned up if necessary.
2080  *
2081  * __jbd2_journal_refile_buffer is usually called as part of a single locked
2082  * operation on a buffer_head, in which the caller is probably going to
2083  * be hooking the journal_head onto other lists.  In that case it is up
2084  * to the caller to remove the journal_head if necessary.  For the
2085  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2086  * doing anything else to the buffer so we need to do the cleanup
2087  * ourselves to avoid a jh leak.
2088  *
2089  * *** The journal_head may be freed by this call! ***
2090  */
2091 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2092 {
2093         struct buffer_head *bh = jh2bh(jh);
2094
2095         jbd_lock_bh_state(bh);
2096         spin_lock(&journal->j_list_lock);
2097
2098         __jbd2_journal_refile_buffer(jh);
2099         jbd_unlock_bh_state(bh);
2100         jbd2_journal_remove_journal_head(bh);
2101
2102         spin_unlock(&journal->j_list_lock);
2103         __brelse(bh);
2104 }
2105
2106 /*
2107  * File inode in the inode list of the handle's transaction
2108  */
2109 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2110 {
2111         transaction_t *transaction = handle->h_transaction;
2112         journal_t *journal = transaction->t_journal;
2113
2114         if (is_handle_aborted(handle))
2115                 return -EIO;
2116
2117         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2118                         transaction->t_tid);
2119
2120         /*
2121          * First check whether inode isn't already on the transaction's
2122          * lists without taking the lock. Note that this check is safe
2123          * without the lock as we cannot race with somebody removing inode
2124          * from the transaction. The reason is that we remove inode from the
2125          * transaction only in journal_release_jbd_inode() and when we commit
2126          * the transaction. We are guarded from the first case by holding
2127          * a reference to the inode. We are safe against the second case
2128          * because if jinode->i_transaction == transaction, commit code
2129          * cannot touch the transaction because we hold reference to it,
2130          * and if jinode->i_next_transaction == transaction, commit code
2131          * will only file the inode where we want it.
2132          */
2133         if (jinode->i_transaction == transaction ||
2134             jinode->i_next_transaction == transaction)
2135                 return 0;
2136
2137         spin_lock(&journal->j_list_lock);
2138
2139         if (jinode->i_transaction == transaction ||
2140             jinode->i_next_transaction == transaction)
2141                 goto done;
2142
2143         /* On some different transaction's list - should be
2144          * the committing one */
2145         if (jinode->i_transaction) {
2146                 J_ASSERT(jinode->i_next_transaction == NULL);
2147                 J_ASSERT(jinode->i_transaction ==
2148                                         journal->j_committing_transaction);
2149                 jinode->i_next_transaction = transaction;
2150                 goto done;
2151         }
2152         /* Not on any transaction list... */
2153         J_ASSERT(!jinode->i_next_transaction);
2154         jinode->i_transaction = transaction;
2155         list_add(&jinode->i_list, &transaction->t_inode_list);
2156 done:
2157         spin_unlock(&journal->j_list_lock);
2158
2159         return 0;
2160 }
2161
2162 /*
2163  * File truncate and transaction commit interact with each other in a
2164  * non-trivial way.  If a transaction writing data block A is
2165  * committing, we cannot discard the data by truncate until we have
2166  * written them.  Otherwise if we crashed after the transaction with
2167  * write has committed but before the transaction with truncate has
2168  * committed, we could see stale data in block A.  This function is a
2169  * helper to solve this problem.  It starts writeout of the truncated
2170  * part in case it is in the committing transaction.
2171  *
2172  * Filesystem code must call this function when inode is journaled in
2173  * ordered mode before truncation happens and after the inode has been
2174  * placed on orphan list with the new inode size. The second condition
2175  * avoids the race that someone writes new data and we start
2176  * committing the transaction after this function has been called but
2177  * before a transaction for truncate is started (and furthermore it
2178  * allows us to optimize the case where the addition to orphan list
2179  * happens in the same transaction as write --- we don't have to write
2180  * any data in such case).
2181  */
2182 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2183                                         struct jbd2_inode *jinode,
2184                                         loff_t new_size)
2185 {
2186         transaction_t *inode_trans, *commit_trans;
2187         int ret = 0;
2188
2189         /* This is a quick check to avoid locking if not necessary */
2190         if (!jinode->i_transaction)
2191                 goto out;
2192         /* Locks are here just to force reading of recent values, it is
2193          * enough that the transaction was not committing before we started
2194          * a transaction adding the inode to orphan list */
2195         read_lock(&journal->j_state_lock);
2196         commit_trans = journal->j_committing_transaction;
2197         read_unlock(&journal->j_state_lock);
2198         spin_lock(&journal->j_list_lock);
2199         inode_trans = jinode->i_transaction;
2200         spin_unlock(&journal->j_list_lock);
2201         if (inode_trans == commit_trans) {
2202                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2203                         new_size, LLONG_MAX);
2204                 if (ret)
2205                         jbd2_journal_abort(journal, ret);
2206         }
2207 out:
2208         return ret;
2209 }