ext4: clean up compiler warning in start_this_handle()
[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         if (is_journal_aborted(journal) ||
160             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
161                 read_unlock(&journal->j_state_lock);
162                 kfree(new_transaction);
163                 return -EROFS;
164         }
165
166         /* Wait on the journal's transaction barrier if necessary */
167         if (journal->j_barrier_count) {
168                 read_unlock(&journal->j_state_lock);
169                 wait_event(journal->j_wait_transaction_locked,
170                                 journal->j_barrier_count == 0);
171                 goto repeat;
172         }
173
174         if (!journal->j_running_transaction) {
175                 read_unlock(&journal->j_state_lock);
176                 if (!new_transaction)
177                         goto alloc_transaction;
178                 write_lock(&journal->j_state_lock);
179                 if (!journal->j_running_transaction) {
180                         jbd2_get_transaction(journal, new_transaction);
181                         new_transaction = NULL;
182                 }
183                 write_unlock(&journal->j_state_lock);
184                 goto repeat;
185         }
186
187         transaction = journal->j_running_transaction;
188
189         /*
190          * If the current transaction is locked down for commit, wait for the
191          * lock to be released.
192          */
193         if (transaction->t_state == T_LOCKED) {
194                 DEFINE_WAIT(wait);
195
196                 prepare_to_wait(&journal->j_wait_transaction_locked,
197                                         &wait, TASK_UNINTERRUPTIBLE);
198                 read_unlock(&journal->j_state_lock);
199                 schedule();
200                 finish_wait(&journal->j_wait_transaction_locked, &wait);
201                 goto repeat;
202         }
203
204         /*
205          * If there is not enough space left in the log to write all potential
206          * buffers requested by this operation, we need to stall pending a log
207          * checkpoint to free some more log space.
208          */
209         needed = atomic_add_return(nblocks,
210                                    &transaction->t_outstanding_credits);
211
212         if (needed > journal->j_max_transaction_buffers) {
213                 /*
214                  * If the current transaction is already too large, then start
215                  * to commit it: we can then go back and attach this handle to
216                  * a new transaction.
217                  */
218                 DEFINE_WAIT(wait);
219
220                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
221                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
222                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
223                                 TASK_UNINTERRUPTIBLE);
224                 __jbd2_log_start_commit(journal, transaction->t_tid);
225                 read_unlock(&journal->j_state_lock);
226                 schedule();
227                 finish_wait(&journal->j_wait_transaction_locked, &wait);
228                 goto repeat;
229         }
230
231         /*
232          * The commit code assumes that it can get enough log space
233          * without forcing a checkpoint.  This is *critical* for
234          * correctness: a checkpoint of a buffer which is also
235          * associated with a committing transaction creates a deadlock,
236          * so commit simply cannot force through checkpoints.
237          *
238          * We must therefore ensure the necessary space in the journal
239          * *before* starting to dirty potentially checkpointed buffers
240          * in the new transaction.
241          *
242          * The worst part is, any transaction currently committing can
243          * reduce the free space arbitrarily.  Be careful to account for
244          * those buffers when checkpointing.
245          */
246
247         /*
248          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
249          * a _lot_ of headroom: 1/4 of the journal plus the size of
250          * the committing transaction.  Really, we only need to give it
251          * committing_transaction->t_outstanding_credits plus "enough" for
252          * the log control blocks.
253          * Also, this test is inconsitent with the matching one in
254          * jbd2_journal_extend().
255          */
256         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
257                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
258                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
259                 read_unlock(&journal->j_state_lock);
260                 write_lock(&journal->j_state_lock);
261                 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
262                         __jbd2_log_wait_for_space(journal);
263                 write_unlock(&journal->j_state_lock);
264                 goto repeat;
265         }
266
267         /* OK, account for the buffers that this operation expects to
268          * use and add the handle to the running transaction. 
269          */
270         update_t_max_wait(transaction);
271         handle->h_transaction = transaction;
272         atomic_inc(&transaction->t_updates);
273         atomic_inc(&transaction->t_handle_count);
274         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
275                   handle, nblocks,
276                   atomic_read(&transaction->t_outstanding_credits),
277                   __jbd2_log_space_left(journal));
278         read_unlock(&journal->j_state_lock);
279
280         lock_map_acquire(&handle->h_lockdep_map);
281         kfree(new_transaction);
282         return 0;
283 }
284
285 static struct lock_class_key jbd2_handle_key;
286
287 /* Allocate a new handle.  This should probably be in a slab... */
288 static handle_t *new_handle(int nblocks)
289 {
290         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
291         if (!handle)
292                 return NULL;
293         memset(handle, 0, sizeof(*handle));
294         handle->h_buffer_credits = nblocks;
295         handle->h_ref = 1;
296
297         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
298                                                 &jbd2_handle_key, 0);
299
300         return handle;
301 }
302
303 /**
304  * handle_t *jbd2_journal_start() - Obtain a new handle.
305  * @journal: Journal to start transaction on.
306  * @nblocks: number of block buffer we might modify
307  *
308  * We make sure that the transaction can guarantee at least nblocks of
309  * modified buffers in the log.  We block until the log can guarantee
310  * that much space.
311  *
312  * This function is visible to journal users (like ext3fs), so is not
313  * called with the journal already locked.
314  *
315  * Return a pointer to a newly allocated handle, or NULL on failure
316  */
317 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int gfp_mask)
318 {
319         handle_t *handle = journal_current_handle();
320         int err;
321
322         if (!journal)
323                 return ERR_PTR(-EROFS);
324
325         if (handle) {
326                 J_ASSERT(handle->h_transaction->t_journal == journal);
327                 handle->h_ref++;
328                 return handle;
329         }
330
331         handle = new_handle(nblocks);
332         if (!handle)
333                 return ERR_PTR(-ENOMEM);
334
335         current->journal_info = handle;
336
337         err = start_this_handle(journal, handle, gfp_mask);
338         if (err < 0) {
339                 jbd2_free_handle(handle);
340                 current->journal_info = NULL;
341                 handle = ERR_PTR(err);
342                 goto out;
343         }
344 out:
345         return handle;
346 }
347 EXPORT_SYMBOL(jbd2__journal_start);
348
349
350 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
351 {
352         return jbd2__journal_start(journal, nblocks, GFP_NOFS);
353 }
354 EXPORT_SYMBOL(jbd2_journal_start);
355
356
357 /**
358  * int jbd2_journal_extend() - extend buffer credits.
359  * @handle:  handle to 'extend'
360  * @nblocks: nr blocks to try to extend by.
361  *
362  * Some transactions, such as large extends and truncates, can be done
363  * atomically all at once or in several stages.  The operation requests
364  * a credit for a number of buffer modications in advance, but can
365  * extend its credit if it needs more.
366  *
367  * jbd2_journal_extend tries to give the running handle more buffer credits.
368  * It does not guarantee that allocation - this is a best-effort only.
369  * The calling process MUST be able to deal cleanly with a failure to
370  * extend here.
371  *
372  * Return 0 on success, non-zero on failure.
373  *
374  * return code < 0 implies an error
375  * return code > 0 implies normal transaction-full status.
376  */
377 int jbd2_journal_extend(handle_t *handle, int nblocks)
378 {
379         transaction_t *transaction = handle->h_transaction;
380         journal_t *journal = transaction->t_journal;
381         int result;
382         int wanted;
383
384         result = -EIO;
385         if (is_handle_aborted(handle))
386                 goto out;
387
388         result = 1;
389
390         read_lock(&journal->j_state_lock);
391
392         /* Don't extend a locked-down transaction! */
393         if (handle->h_transaction->t_state != T_RUNNING) {
394                 jbd_debug(3, "denied handle %p %d blocks: "
395                           "transaction not running\n", handle, nblocks);
396                 goto error_out;
397         }
398
399         spin_lock(&transaction->t_handle_lock);
400         wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
401
402         if (wanted > journal->j_max_transaction_buffers) {
403                 jbd_debug(3, "denied handle %p %d blocks: "
404                           "transaction too large\n", handle, nblocks);
405                 goto unlock;
406         }
407
408         if (wanted > __jbd2_log_space_left(journal)) {
409                 jbd_debug(3, "denied handle %p %d blocks: "
410                           "insufficient log space\n", handle, nblocks);
411                 goto unlock;
412         }
413
414         handle->h_buffer_credits += nblocks;
415         atomic_add(nblocks, &transaction->t_outstanding_credits);
416         result = 0;
417
418         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
419 unlock:
420         spin_unlock(&transaction->t_handle_lock);
421 error_out:
422         read_unlock(&journal->j_state_lock);
423 out:
424         return result;
425 }
426
427
428 /**
429  * int jbd2_journal_restart() - restart a handle .
430  * @handle:  handle to restart
431  * @nblocks: nr credits requested
432  *
433  * Restart a handle for a multi-transaction filesystem
434  * operation.
435  *
436  * If the jbd2_journal_extend() call above fails to grant new buffer credits
437  * to a running handle, a call to jbd2_journal_restart will commit the
438  * handle's transaction so far and reattach the handle to a new
439  * transaction capabable of guaranteeing the requested number of
440  * credits.
441  */
442 int jbd2__journal_restart(handle_t *handle, int nblocks, int gfp_mask)
443 {
444         transaction_t *transaction = handle->h_transaction;
445         journal_t *journal = transaction->t_journal;
446         int ret;
447
448         /* If we've had an abort of any type, don't even think about
449          * actually doing the restart! */
450         if (is_handle_aborted(handle))
451                 return 0;
452
453         /*
454          * First unlink the handle from its current transaction, and start the
455          * commit on that.
456          */
457         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
458         J_ASSERT(journal_current_handle() == handle);
459
460         read_lock(&journal->j_state_lock);
461         spin_lock(&transaction->t_handle_lock);
462         atomic_sub(handle->h_buffer_credits,
463                    &transaction->t_outstanding_credits);
464         if (atomic_dec_and_test(&transaction->t_updates))
465                 wake_up(&journal->j_wait_updates);
466         spin_unlock(&transaction->t_handle_lock);
467
468         jbd_debug(2, "restarting handle %p\n", handle);
469         __jbd2_log_start_commit(journal, transaction->t_tid);
470         read_unlock(&journal->j_state_lock);
471
472         lock_map_release(&handle->h_lockdep_map);
473         handle->h_buffer_credits = nblocks;
474         ret = start_this_handle(journal, handle, gfp_mask);
475         return ret;
476 }
477 EXPORT_SYMBOL(jbd2__journal_restart);
478
479
480 int jbd2_journal_restart(handle_t *handle, int nblocks)
481 {
482         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
483 }
484 EXPORT_SYMBOL(jbd2_journal_restart);
485
486 /**
487  * void jbd2_journal_lock_updates () - establish a transaction barrier.
488  * @journal:  Journal to establish a barrier on.
489  *
490  * This locks out any further updates from being started, and blocks
491  * until all existing updates have completed, returning only once the
492  * journal is in a quiescent state with no updates running.
493  *
494  * The journal lock should not be held on entry.
495  */
496 void jbd2_journal_lock_updates(journal_t *journal)
497 {
498         DEFINE_WAIT(wait);
499
500         write_lock(&journal->j_state_lock);
501         ++journal->j_barrier_count;
502
503         /* Wait until there are no running updates */
504         while (1) {
505                 transaction_t *transaction = journal->j_running_transaction;
506
507                 if (!transaction)
508                         break;
509
510                 spin_lock(&transaction->t_handle_lock);
511                 if (!atomic_read(&transaction->t_updates)) {
512                         spin_unlock(&transaction->t_handle_lock);
513                         break;
514                 }
515                 prepare_to_wait(&journal->j_wait_updates, &wait,
516                                 TASK_UNINTERRUPTIBLE);
517                 spin_unlock(&transaction->t_handle_lock);
518                 write_unlock(&journal->j_state_lock);
519                 schedule();
520                 finish_wait(&journal->j_wait_updates, &wait);
521                 write_lock(&journal->j_state_lock);
522         }
523         write_unlock(&journal->j_state_lock);
524
525         /*
526          * We have now established a barrier against other normal updates, but
527          * we also need to barrier against other jbd2_journal_lock_updates() calls
528          * to make sure that we serialise special journal-locked operations
529          * too.
530          */
531         mutex_lock(&journal->j_barrier);
532 }
533
534 /**
535  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
536  * @journal:  Journal to release the barrier on.
537  *
538  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
539  *
540  * Should be called without the journal lock held.
541  */
542 void jbd2_journal_unlock_updates (journal_t *journal)
543 {
544         J_ASSERT(journal->j_barrier_count != 0);
545
546         mutex_unlock(&journal->j_barrier);
547         write_lock(&journal->j_state_lock);
548         --journal->j_barrier_count;
549         write_unlock(&journal->j_state_lock);
550         wake_up(&journal->j_wait_transaction_locked);
551 }
552
553 static void warn_dirty_buffer(struct buffer_head *bh)
554 {
555         char b[BDEVNAME_SIZE];
556
557         printk(KERN_WARNING
558                "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
559                "There's a risk of filesystem corruption in case of system "
560                "crash.\n",
561                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
562 }
563
564 /*
565  * If the buffer is already part of the current transaction, then there
566  * is nothing we need to do.  If it is already part of a prior
567  * transaction which we are still committing to disk, then we need to
568  * make sure that we do not overwrite the old copy: we do copy-out to
569  * preserve the copy going to disk.  We also account the buffer against
570  * the handle's metadata buffer credits (unless the buffer is already
571  * part of the transaction, that is).
572  *
573  */
574 static int
575 do_get_write_access(handle_t *handle, struct journal_head *jh,
576                         int force_copy)
577 {
578         struct buffer_head *bh;
579         transaction_t *transaction;
580         journal_t *journal;
581         int error;
582         char *frozen_buffer = NULL;
583         int need_copy = 0;
584
585         if (is_handle_aborted(handle))
586                 return -EROFS;
587
588         transaction = handle->h_transaction;
589         journal = transaction->t_journal;
590
591         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
592
593         JBUFFER_TRACE(jh, "entry");
594 repeat:
595         bh = jh2bh(jh);
596
597         /* @@@ Need to check for errors here at some point. */
598
599         lock_buffer(bh);
600         jbd_lock_bh_state(bh);
601
602         /* We now hold the buffer lock so it is safe to query the buffer
603          * state.  Is the buffer dirty?
604          *
605          * If so, there are two possibilities.  The buffer may be
606          * non-journaled, and undergoing a quite legitimate writeback.
607          * Otherwise, it is journaled, and we don't expect dirty buffers
608          * in that state (the buffers should be marked JBD_Dirty
609          * instead.)  So either the IO is being done under our own
610          * control and this is a bug, or it's a third party IO such as
611          * dump(8) (which may leave the buffer scheduled for read ---
612          * ie. locked but not dirty) or tune2fs (which may actually have
613          * the buffer dirtied, ugh.)  */
614
615         if (buffer_dirty(bh)) {
616                 /*
617                  * First question: is this buffer already part of the current
618                  * transaction or the existing committing transaction?
619                  */
620                 if (jh->b_transaction) {
621                         J_ASSERT_JH(jh,
622                                 jh->b_transaction == transaction ||
623                                 jh->b_transaction ==
624                                         journal->j_committing_transaction);
625                         if (jh->b_next_transaction)
626                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
627                                                         transaction);
628                         warn_dirty_buffer(bh);
629                 }
630                 /*
631                  * In any case we need to clean the dirty flag and we must
632                  * do it under the buffer lock to be sure we don't race
633                  * with running write-out.
634                  */
635                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
636                 clear_buffer_dirty(bh);
637                 set_buffer_jbddirty(bh);
638         }
639
640         unlock_buffer(bh);
641
642         error = -EROFS;
643         if (is_handle_aborted(handle)) {
644                 jbd_unlock_bh_state(bh);
645                 goto out;
646         }
647         error = 0;
648
649         /*
650          * The buffer is already part of this transaction if b_transaction or
651          * b_next_transaction points to it
652          */
653         if (jh->b_transaction == transaction ||
654             jh->b_next_transaction == transaction)
655                 goto done;
656
657         /*
658          * this is the first time this transaction is touching this buffer,
659          * reset the modified flag
660          */
661        jh->b_modified = 0;
662
663         /*
664          * If there is already a copy-out version of this buffer, then we don't
665          * need to make another one
666          */
667         if (jh->b_frozen_data) {
668                 JBUFFER_TRACE(jh, "has frozen data");
669                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
670                 jh->b_next_transaction = transaction;
671                 goto done;
672         }
673
674         /* Is there data here we need to preserve? */
675
676         if (jh->b_transaction && jh->b_transaction != transaction) {
677                 JBUFFER_TRACE(jh, "owned by older transaction");
678                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
679                 J_ASSERT_JH(jh, jh->b_transaction ==
680                                         journal->j_committing_transaction);
681
682                 /* There is one case we have to be very careful about.
683                  * If the committing transaction is currently writing
684                  * this buffer out to disk and has NOT made a copy-out,
685                  * then we cannot modify the buffer contents at all
686                  * right now.  The essence of copy-out is that it is the
687                  * extra copy, not the primary copy, which gets
688                  * journaled.  If the primary copy is already going to
689                  * disk then we cannot do copy-out here. */
690
691                 if (jh->b_jlist == BJ_Shadow) {
692                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
693                         wait_queue_head_t *wqh;
694
695                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
696
697                         JBUFFER_TRACE(jh, "on shadow: sleep");
698                         jbd_unlock_bh_state(bh);
699                         /* commit wakes up all shadow buffers after IO */
700                         for ( ; ; ) {
701                                 prepare_to_wait(wqh, &wait.wait,
702                                                 TASK_UNINTERRUPTIBLE);
703                                 if (jh->b_jlist != BJ_Shadow)
704                                         break;
705                                 schedule();
706                         }
707                         finish_wait(wqh, &wait.wait);
708                         goto repeat;
709                 }
710
711                 /* Only do the copy if the currently-owning transaction
712                  * still needs it.  If it is on the Forget list, the
713                  * committing transaction is past that stage.  The
714                  * buffer had better remain locked during the kmalloc,
715                  * but that should be true --- we hold the journal lock
716                  * still and the buffer is already on the BUF_JOURNAL
717                  * list so won't be flushed.
718                  *
719                  * Subtle point, though: if this is a get_undo_access,
720                  * then we will be relying on the frozen_data to contain
721                  * the new value of the committed_data record after the
722                  * transaction, so we HAVE to force the frozen_data copy
723                  * in that case. */
724
725                 if (jh->b_jlist != BJ_Forget || force_copy) {
726                         JBUFFER_TRACE(jh, "generate frozen data");
727                         if (!frozen_buffer) {
728                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
729                                 jbd_unlock_bh_state(bh);
730                                 frozen_buffer =
731                                         jbd2_alloc(jh2bh(jh)->b_size,
732                                                          GFP_NOFS);
733                                 if (!frozen_buffer) {
734                                         printk(KERN_EMERG
735                                                "%s: OOM for frozen_buffer\n",
736                                                __func__);
737                                         JBUFFER_TRACE(jh, "oom!");
738                                         error = -ENOMEM;
739                                         jbd_lock_bh_state(bh);
740                                         goto done;
741                                 }
742                                 goto repeat;
743                         }
744                         jh->b_frozen_data = frozen_buffer;
745                         frozen_buffer = NULL;
746                         need_copy = 1;
747                 }
748                 jh->b_next_transaction = transaction;
749         }
750
751
752         /*
753          * Finally, if the buffer is not journaled right now, we need to make
754          * sure it doesn't get written to disk before the caller actually
755          * commits the new data
756          */
757         if (!jh->b_transaction) {
758                 JBUFFER_TRACE(jh, "no transaction");
759                 J_ASSERT_JH(jh, !jh->b_next_transaction);
760                 jh->b_transaction = transaction;
761                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
762                 spin_lock(&journal->j_list_lock);
763                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
764                 spin_unlock(&journal->j_list_lock);
765         }
766
767 done:
768         if (need_copy) {
769                 struct page *page;
770                 int offset;
771                 char *source;
772
773                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
774                             "Possible IO failure.\n");
775                 page = jh2bh(jh)->b_page;
776                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
777                 source = kmap_atomic(page, KM_USER0);
778                 /* Fire data frozen trigger just before we copy the data */
779                 jbd2_buffer_frozen_trigger(jh, source + offset,
780                                            jh->b_triggers);
781                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
782                 kunmap_atomic(source, KM_USER0);
783
784                 /*
785                  * Now that the frozen data is saved off, we need to store
786                  * any matching triggers.
787                  */
788                 jh->b_frozen_triggers = jh->b_triggers;
789         }
790         jbd_unlock_bh_state(bh);
791
792         /*
793          * If we are about to journal a buffer, then any revoke pending on it is
794          * no longer valid
795          */
796         jbd2_journal_cancel_revoke(handle, jh);
797
798 out:
799         if (unlikely(frozen_buffer))    /* It's usually NULL */
800                 jbd2_free(frozen_buffer, bh->b_size);
801
802         JBUFFER_TRACE(jh, "exit");
803         return error;
804 }
805
806 /**
807  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
808  * @handle: transaction to add buffer modifications to
809  * @bh:     bh to be used for metadata writes
810  * @credits: variable that will receive credits for the buffer
811  *
812  * Returns an error code or 0 on success.
813  *
814  * In full data journalling mode the buffer may be of type BJ_AsyncData,
815  * because we're write()ing a buffer which is also part of a shared mapping.
816  */
817
818 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
819 {
820         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
821         int rc;
822
823         /* We do not want to get caught playing with fields which the
824          * log thread also manipulates.  Make sure that the buffer
825          * completes any outstanding IO before proceeding. */
826         rc = do_get_write_access(handle, jh, 0);
827         jbd2_journal_put_journal_head(jh);
828         return rc;
829 }
830
831
832 /*
833  * When the user wants to journal a newly created buffer_head
834  * (ie. getblk() returned a new buffer and we are going to populate it
835  * manually rather than reading off disk), then we need to keep the
836  * buffer_head locked until it has been completely filled with new
837  * data.  In this case, we should be able to make the assertion that
838  * the bh is not already part of an existing transaction.
839  *
840  * The buffer should already be locked by the caller by this point.
841  * There is no lock ranking violation: it was a newly created,
842  * unlocked buffer beforehand. */
843
844 /**
845  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
846  * @handle: transaction to new buffer to
847  * @bh: new buffer.
848  *
849  * Call this if you create a new bh.
850  */
851 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
852 {
853         transaction_t *transaction = handle->h_transaction;
854         journal_t *journal = transaction->t_journal;
855         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
856         int err;
857
858         jbd_debug(5, "journal_head %p\n", jh);
859         err = -EROFS;
860         if (is_handle_aborted(handle))
861                 goto out;
862         err = 0;
863
864         JBUFFER_TRACE(jh, "entry");
865         /*
866          * The buffer may already belong to this transaction due to pre-zeroing
867          * in the filesystem's new_block code.  It may also be on the previous,
868          * committing transaction's lists, but it HAS to be in Forget state in
869          * that case: the transaction must have deleted the buffer for it to be
870          * reused here.
871          */
872         jbd_lock_bh_state(bh);
873         spin_lock(&journal->j_list_lock);
874         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
875                 jh->b_transaction == NULL ||
876                 (jh->b_transaction == journal->j_committing_transaction &&
877                           jh->b_jlist == BJ_Forget)));
878
879         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
880         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
881
882         if (jh->b_transaction == NULL) {
883                 /*
884                  * Previous jbd2_journal_forget() could have left the buffer
885                  * with jbddirty bit set because it was being committed. When
886                  * the commit finished, we've filed the buffer for
887                  * checkpointing and marked it dirty. Now we are reallocating
888                  * the buffer so the transaction freeing it must have
889                  * committed and so it's safe to clear the dirty bit.
890                  */
891                 clear_buffer_dirty(jh2bh(jh));
892                 jh->b_transaction = transaction;
893
894                 /* first access by this transaction */
895                 jh->b_modified = 0;
896
897                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
898                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
899         } else if (jh->b_transaction == journal->j_committing_transaction) {
900                 /* first access by this transaction */
901                 jh->b_modified = 0;
902
903                 JBUFFER_TRACE(jh, "set next transaction");
904                 jh->b_next_transaction = transaction;
905         }
906         spin_unlock(&journal->j_list_lock);
907         jbd_unlock_bh_state(bh);
908
909         /*
910          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
911          * blocks which contain freed but then revoked metadata.  We need
912          * to cancel the revoke in case we end up freeing it yet again
913          * and the reallocating as data - this would cause a second revoke,
914          * which hits an assertion error.
915          */
916         JBUFFER_TRACE(jh, "cancelling revoke");
917         jbd2_journal_cancel_revoke(handle, jh);
918         jbd2_journal_put_journal_head(jh);
919 out:
920         return err;
921 }
922
923 /**
924  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
925  *     non-rewindable consequences
926  * @handle: transaction
927  * @bh: buffer to undo
928  * @credits: store the number of taken credits here (if not NULL)
929  *
930  * Sometimes there is a need to distinguish between metadata which has
931  * been committed to disk and that which has not.  The ext3fs code uses
932  * this for freeing and allocating space, we have to make sure that we
933  * do not reuse freed space until the deallocation has been committed,
934  * since if we overwrote that space we would make the delete
935  * un-rewindable in case of a crash.
936  *
937  * To deal with that, jbd2_journal_get_undo_access requests write access to a
938  * buffer for parts of non-rewindable operations such as delete
939  * operations on the bitmaps.  The journaling code must keep a copy of
940  * the buffer's contents prior to the undo_access call until such time
941  * as we know that the buffer has definitely been committed to disk.
942  *
943  * We never need to know which transaction the committed data is part
944  * of, buffers touched here are guaranteed to be dirtied later and so
945  * will be committed to a new transaction in due course, at which point
946  * we can discard the old committed data pointer.
947  *
948  * Returns error number or 0 on success.
949  */
950 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
951 {
952         int err;
953         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
954         char *committed_data = NULL;
955
956         JBUFFER_TRACE(jh, "entry");
957
958         /*
959          * Do this first --- it can drop the journal lock, so we want to
960          * make sure that obtaining the committed_data is done
961          * atomically wrt. completion of any outstanding commits.
962          */
963         err = do_get_write_access(handle, jh, 1);
964         if (err)
965                 goto out;
966
967 repeat:
968         if (!jh->b_committed_data) {
969                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
970                 if (!committed_data) {
971                         printk(KERN_EMERG "%s: No memory for committed data\n",
972                                 __func__);
973                         err = -ENOMEM;
974                         goto out;
975                 }
976         }
977
978         jbd_lock_bh_state(bh);
979         if (!jh->b_committed_data) {
980                 /* Copy out the current buffer contents into the
981                  * preserved, committed copy. */
982                 JBUFFER_TRACE(jh, "generate b_committed data");
983                 if (!committed_data) {
984                         jbd_unlock_bh_state(bh);
985                         goto repeat;
986                 }
987
988                 jh->b_committed_data = committed_data;
989                 committed_data = NULL;
990                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
991         }
992         jbd_unlock_bh_state(bh);
993 out:
994         jbd2_journal_put_journal_head(jh);
995         if (unlikely(committed_data))
996                 jbd2_free(committed_data, bh->b_size);
997         return err;
998 }
999
1000 /**
1001  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1002  * @bh: buffer to trigger on
1003  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1004  *
1005  * Set any triggers on this journal_head.  This is always safe, because
1006  * triggers for a committing buffer will be saved off, and triggers for
1007  * a running transaction will match the buffer in that transaction.
1008  *
1009  * Call with NULL to clear the triggers.
1010  */
1011 void jbd2_journal_set_triggers(struct buffer_head *bh,
1012                                struct jbd2_buffer_trigger_type *type)
1013 {
1014         struct journal_head *jh = bh2jh(bh);
1015
1016         jh->b_triggers = type;
1017 }
1018
1019 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1020                                 struct jbd2_buffer_trigger_type *triggers)
1021 {
1022         struct buffer_head *bh = jh2bh(jh);
1023
1024         if (!triggers || !triggers->t_frozen)
1025                 return;
1026
1027         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1028 }
1029
1030 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1031                                struct jbd2_buffer_trigger_type *triggers)
1032 {
1033         if (!triggers || !triggers->t_abort)
1034                 return;
1035
1036         triggers->t_abort(triggers, jh2bh(jh));
1037 }
1038
1039
1040
1041 /**
1042  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1043  * @handle: transaction to add buffer to.
1044  * @bh: buffer to mark
1045  *
1046  * mark dirty metadata which needs to be journaled as part of the current
1047  * transaction.
1048  *
1049  * The buffer is placed on the transaction's metadata list and is marked
1050  * as belonging to the transaction.
1051  *
1052  * Returns error number or 0 on success.
1053  *
1054  * Special care needs to be taken if the buffer already belongs to the
1055  * current committing transaction (in which case we should have frozen
1056  * data present for that commit).  In that case, we don't relink the
1057  * buffer: that only gets done when the old transaction finally
1058  * completes its commit.
1059  */
1060 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1061 {
1062         transaction_t *transaction = handle->h_transaction;
1063         journal_t *journal = transaction->t_journal;
1064         struct journal_head *jh = bh2jh(bh);
1065
1066         jbd_debug(5, "journal_head %p\n", jh);
1067         JBUFFER_TRACE(jh, "entry");
1068         if (is_handle_aborted(handle))
1069                 goto out;
1070
1071         jbd_lock_bh_state(bh);
1072
1073         if (jh->b_modified == 0) {
1074                 /*
1075                  * This buffer's got modified and becoming part
1076                  * of the transaction. This needs to be done
1077                  * once a transaction -bzzz
1078                  */
1079                 jh->b_modified = 1;
1080                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1081                 handle->h_buffer_credits--;
1082         }
1083
1084         /*
1085          * fastpath, to avoid expensive locking.  If this buffer is already
1086          * on the running transaction's metadata list there is nothing to do.
1087          * Nobody can take it off again because there is a handle open.
1088          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1089          * result in this test being false, so we go in and take the locks.
1090          */
1091         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1092                 JBUFFER_TRACE(jh, "fastpath");
1093                 J_ASSERT_JH(jh, jh->b_transaction ==
1094                                         journal->j_running_transaction);
1095                 goto out_unlock_bh;
1096         }
1097
1098         set_buffer_jbddirty(bh);
1099
1100         /*
1101          * Metadata already on the current transaction list doesn't
1102          * need to be filed.  Metadata on another transaction's list must
1103          * be committing, and will be refiled once the commit completes:
1104          * leave it alone for now.
1105          */
1106         if (jh->b_transaction != transaction) {
1107                 JBUFFER_TRACE(jh, "already on other transaction");
1108                 J_ASSERT_JH(jh, jh->b_transaction ==
1109                                         journal->j_committing_transaction);
1110                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1111                 /* And this case is illegal: we can't reuse another
1112                  * transaction's data buffer, ever. */
1113                 goto out_unlock_bh;
1114         }
1115
1116         /* That test should have eliminated the following case: */
1117         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1118
1119         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1120         spin_lock(&journal->j_list_lock);
1121         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1122         spin_unlock(&journal->j_list_lock);
1123 out_unlock_bh:
1124         jbd_unlock_bh_state(bh);
1125 out:
1126         JBUFFER_TRACE(jh, "exit");
1127         return 0;
1128 }
1129
1130 /*
1131  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1132  * updates, if the update decided in the end that it didn't need access.
1133  *
1134  */
1135 void
1136 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1137 {
1138         BUFFER_TRACE(bh, "entry");
1139 }
1140
1141 /**
1142  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1143  * @handle: transaction handle
1144  * @bh:     bh to 'forget'
1145  *
1146  * We can only do the bforget if there are no commits pending against the
1147  * buffer.  If the buffer is dirty in the current running transaction we
1148  * can safely unlink it.
1149  *
1150  * bh may not be a journalled buffer at all - it may be a non-JBD
1151  * buffer which came off the hashtable.  Check for this.
1152  *
1153  * Decrements bh->b_count by one.
1154  *
1155  * Allow this call even if the handle has aborted --- it may be part of
1156  * the caller's cleanup after an abort.
1157  */
1158 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1159 {
1160         transaction_t *transaction = handle->h_transaction;
1161         journal_t *journal = transaction->t_journal;
1162         struct journal_head *jh;
1163         int drop_reserve = 0;
1164         int err = 0;
1165         int was_modified = 0;
1166
1167         BUFFER_TRACE(bh, "entry");
1168
1169         jbd_lock_bh_state(bh);
1170         spin_lock(&journal->j_list_lock);
1171
1172         if (!buffer_jbd(bh))
1173                 goto not_jbd;
1174         jh = bh2jh(bh);
1175
1176         /* Critical error: attempting to delete a bitmap buffer, maybe?
1177          * Don't do any jbd operations, and return an error. */
1178         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1179                          "inconsistent data on disk")) {
1180                 err = -EIO;
1181                 goto not_jbd;
1182         }
1183
1184         /* keep track of wether or not this transaction modified us */
1185         was_modified = jh->b_modified;
1186
1187         /*
1188          * The buffer's going from the transaction, we must drop
1189          * all references -bzzz
1190          */
1191         jh->b_modified = 0;
1192
1193         if (jh->b_transaction == handle->h_transaction) {
1194                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1195
1196                 /* If we are forgetting a buffer which is already part
1197                  * of this transaction, then we can just drop it from
1198                  * the transaction immediately. */
1199                 clear_buffer_dirty(bh);
1200                 clear_buffer_jbddirty(bh);
1201
1202                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1203
1204                 /*
1205                  * we only want to drop a reference if this transaction
1206                  * modified the buffer
1207                  */
1208                 if (was_modified)
1209                         drop_reserve = 1;
1210
1211                 /*
1212                  * We are no longer going to journal this buffer.
1213                  * However, the commit of this transaction is still
1214                  * important to the buffer: the delete that we are now
1215                  * processing might obsolete an old log entry, so by
1216                  * committing, we can satisfy the buffer's checkpoint.
1217                  *
1218                  * So, if we have a checkpoint on the buffer, we should
1219                  * now refile the buffer on our BJ_Forget list so that
1220                  * we know to remove the checkpoint after we commit.
1221                  */
1222
1223                 if (jh->b_cp_transaction) {
1224                         __jbd2_journal_temp_unlink_buffer(jh);
1225                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1226                 } else {
1227                         __jbd2_journal_unfile_buffer(jh);
1228                         jbd2_journal_remove_journal_head(bh);
1229                         __brelse(bh);
1230                         if (!buffer_jbd(bh)) {
1231                                 spin_unlock(&journal->j_list_lock);
1232                                 jbd_unlock_bh_state(bh);
1233                                 __bforget(bh);
1234                                 goto drop;
1235                         }
1236                 }
1237         } else if (jh->b_transaction) {
1238                 J_ASSERT_JH(jh, (jh->b_transaction ==
1239                                  journal->j_committing_transaction));
1240                 /* However, if the buffer is still owned by a prior
1241                  * (committing) transaction, we can't drop it yet... */
1242                 JBUFFER_TRACE(jh, "belongs to older transaction");
1243                 /* ... but we CAN drop it from the new transaction if we
1244                  * have also modified it since the original commit. */
1245
1246                 if (jh->b_next_transaction) {
1247                         J_ASSERT(jh->b_next_transaction == transaction);
1248                         jh->b_next_transaction = NULL;
1249
1250                         /*
1251                          * only drop a reference if this transaction modified
1252                          * the buffer
1253                          */
1254                         if (was_modified)
1255                                 drop_reserve = 1;
1256                 }
1257         }
1258
1259 not_jbd:
1260         spin_unlock(&journal->j_list_lock);
1261         jbd_unlock_bh_state(bh);
1262         __brelse(bh);
1263 drop:
1264         if (drop_reserve) {
1265                 /* no need to reserve log space for this block -bzzz */
1266                 handle->h_buffer_credits++;
1267         }
1268         return err;
1269 }
1270
1271 /**
1272  * int jbd2_journal_stop() - complete a transaction
1273  * @handle: tranaction to complete.
1274  *
1275  * All done for a particular handle.
1276  *
1277  * There is not much action needed here.  We just return any remaining
1278  * buffer credits to the transaction and remove the handle.  The only
1279  * complication is that we need to start a commit operation if the
1280  * filesystem is marked for synchronous update.
1281  *
1282  * jbd2_journal_stop itself will not usually return an error, but it may
1283  * do so in unusual circumstances.  In particular, expect it to
1284  * return -EIO if a jbd2_journal_abort has been executed since the
1285  * transaction began.
1286  */
1287 int jbd2_journal_stop(handle_t *handle)
1288 {
1289         transaction_t *transaction = handle->h_transaction;
1290         journal_t *journal = transaction->t_journal;
1291         int err, wait_for_commit = 0;
1292         tid_t tid;
1293         pid_t pid;
1294
1295         J_ASSERT(journal_current_handle() == handle);
1296
1297         if (is_handle_aborted(handle))
1298                 err = -EIO;
1299         else {
1300                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1301                 err = 0;
1302         }
1303
1304         if (--handle->h_ref > 0) {
1305                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1306                           handle->h_ref);
1307                 return err;
1308         }
1309
1310         jbd_debug(4, "Handle %p going down\n", handle);
1311
1312         /*
1313          * Implement synchronous transaction batching.  If the handle
1314          * was synchronous, don't force a commit immediately.  Let's
1315          * yield and let another thread piggyback onto this
1316          * transaction.  Keep doing that while new threads continue to
1317          * arrive.  It doesn't cost much - we're about to run a commit
1318          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1319          * operations by 30x or more...
1320          *
1321          * We try and optimize the sleep time against what the
1322          * underlying disk can do, instead of having a static sleep
1323          * time.  This is useful for the case where our storage is so
1324          * fast that it is more optimal to go ahead and force a flush
1325          * and wait for the transaction to be committed than it is to
1326          * wait for an arbitrary amount of time for new writers to
1327          * join the transaction.  We achieve this by measuring how
1328          * long it takes to commit a transaction, and compare it with
1329          * how long this transaction has been running, and if run time
1330          * < commit time then we sleep for the delta and commit.  This
1331          * greatly helps super fast disks that would see slowdowns as
1332          * more threads started doing fsyncs.
1333          *
1334          * But don't do this if this process was the most recent one
1335          * to perform a synchronous write.  We do this to detect the
1336          * case where a single process is doing a stream of sync
1337          * writes.  No point in waiting for joiners in that case.
1338          */
1339         pid = current->pid;
1340         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1341                 u64 commit_time, trans_time;
1342
1343                 journal->j_last_sync_writer = pid;
1344
1345                 read_lock(&journal->j_state_lock);
1346                 commit_time = journal->j_average_commit_time;
1347                 read_unlock(&journal->j_state_lock);
1348
1349                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1350                                                    transaction->t_start_time));
1351
1352                 commit_time = max_t(u64, commit_time,
1353                                     1000*journal->j_min_batch_time);
1354                 commit_time = min_t(u64, commit_time,
1355                                     1000*journal->j_max_batch_time);
1356
1357                 if (trans_time < commit_time) {
1358                         ktime_t expires = ktime_add_ns(ktime_get(),
1359                                                        commit_time);
1360                         set_current_state(TASK_UNINTERRUPTIBLE);
1361                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1362                 }
1363         }
1364
1365         if (handle->h_sync)
1366                 transaction->t_synchronous_commit = 1;
1367         current->journal_info = NULL;
1368         atomic_sub(handle->h_buffer_credits,
1369                    &transaction->t_outstanding_credits);
1370
1371         /*
1372          * If the handle is marked SYNC, we need to set another commit
1373          * going!  We also want to force a commit if the current
1374          * transaction is occupying too much of the log, or if the
1375          * transaction is too old now.
1376          */
1377         if (handle->h_sync ||
1378             (atomic_read(&transaction->t_outstanding_credits) >
1379              journal->j_max_transaction_buffers) ||
1380             time_after_eq(jiffies, transaction->t_expires)) {
1381                 /* Do this even for aborted journals: an abort still
1382                  * completes the commit thread, it just doesn't write
1383                  * anything to disk. */
1384
1385                 jbd_debug(2, "transaction too old, requesting commit for "
1386                                         "handle %p\n", handle);
1387                 /* This is non-blocking */
1388                 jbd2_log_start_commit(journal, transaction->t_tid);
1389
1390                 /*
1391                  * Special case: JBD2_SYNC synchronous updates require us
1392                  * to wait for the commit to complete.
1393                  */
1394                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1395                         wait_for_commit = 1;
1396         }
1397
1398         /*
1399          * Once we drop t_updates, if it goes to zero the transaction
1400          * could start commiting on us and eventually disappear.  So
1401          * once we do this, we must not dereference transaction
1402          * pointer again.
1403          */
1404         tid = transaction->t_tid;
1405         if (atomic_dec_and_test(&transaction->t_updates)) {
1406                 wake_up(&journal->j_wait_updates);
1407                 if (journal->j_barrier_count)
1408                         wake_up(&journal->j_wait_transaction_locked);
1409         }
1410
1411         if (wait_for_commit)
1412                 err = jbd2_log_wait_commit(journal, tid);
1413
1414         lock_map_release(&handle->h_lockdep_map);
1415
1416         jbd2_free_handle(handle);
1417         return err;
1418 }
1419
1420 /**
1421  * int jbd2_journal_force_commit() - force any uncommitted transactions
1422  * @journal: journal to force
1423  *
1424  * For synchronous operations: force any uncommitted transactions
1425  * to disk.  May seem kludgy, but it reuses all the handle batching
1426  * code in a very simple manner.
1427  */
1428 int jbd2_journal_force_commit(journal_t *journal)
1429 {
1430         handle_t *handle;
1431         int ret;
1432
1433         handle = jbd2_journal_start(journal, 1);
1434         if (IS_ERR(handle)) {
1435                 ret = PTR_ERR(handle);
1436         } else {
1437                 handle->h_sync = 1;
1438                 ret = jbd2_journal_stop(handle);
1439         }
1440         return ret;
1441 }
1442
1443 /*
1444  *
1445  * List management code snippets: various functions for manipulating the
1446  * transaction buffer lists.
1447  *
1448  */
1449
1450 /*
1451  * Append a buffer to a transaction list, given the transaction's list head
1452  * pointer.
1453  *
1454  * j_list_lock is held.
1455  *
1456  * jbd_lock_bh_state(jh2bh(jh)) is held.
1457  */
1458
1459 static inline void
1460 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1461 {
1462         if (!*list) {
1463                 jh->b_tnext = jh->b_tprev = jh;
1464                 *list = jh;
1465         } else {
1466                 /* Insert at the tail of the list to preserve order */
1467                 struct journal_head *first = *list, *last = first->b_tprev;
1468                 jh->b_tprev = last;
1469                 jh->b_tnext = first;
1470                 last->b_tnext = first->b_tprev = jh;
1471         }
1472 }
1473
1474 /*
1475  * Remove a buffer from a transaction list, given the transaction's list
1476  * head pointer.
1477  *
1478  * Called with j_list_lock held, and the journal may not be locked.
1479  *
1480  * jbd_lock_bh_state(jh2bh(jh)) is held.
1481  */
1482
1483 static inline void
1484 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1485 {
1486         if (*list == jh) {
1487                 *list = jh->b_tnext;
1488                 if (*list == jh)
1489                         *list = NULL;
1490         }
1491         jh->b_tprev->b_tnext = jh->b_tnext;
1492         jh->b_tnext->b_tprev = jh->b_tprev;
1493 }
1494
1495 /*
1496  * Remove a buffer from the appropriate transaction list.
1497  *
1498  * Note that this function can *change* the value of
1499  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1500  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1501  * of these pointers, it could go bad.  Generally the caller needs to re-read
1502  * the pointer from the transaction_t.
1503  *
1504  * Called under j_list_lock.  The journal may not be locked.
1505  */
1506 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1507 {
1508         struct journal_head **list = NULL;
1509         transaction_t *transaction;
1510         struct buffer_head *bh = jh2bh(jh);
1511
1512         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1513         transaction = jh->b_transaction;
1514         if (transaction)
1515                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1516
1517         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1518         if (jh->b_jlist != BJ_None)
1519                 J_ASSERT_JH(jh, transaction != NULL);
1520
1521         switch (jh->b_jlist) {
1522         case BJ_None:
1523                 return;
1524         case BJ_Metadata:
1525                 transaction->t_nr_buffers--;
1526                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1527                 list = &transaction->t_buffers;
1528                 break;
1529         case BJ_Forget:
1530                 list = &transaction->t_forget;
1531                 break;
1532         case BJ_IO:
1533                 list = &transaction->t_iobuf_list;
1534                 break;
1535         case BJ_Shadow:
1536                 list = &transaction->t_shadow_list;
1537                 break;
1538         case BJ_LogCtl:
1539                 list = &transaction->t_log_list;
1540                 break;
1541         case BJ_Reserved:
1542                 list = &transaction->t_reserved_list;
1543                 break;
1544         }
1545
1546         __blist_del_buffer(list, jh);
1547         jh->b_jlist = BJ_None;
1548         if (test_clear_buffer_jbddirty(bh))
1549                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1550 }
1551
1552 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1553 {
1554         __jbd2_journal_temp_unlink_buffer(jh);
1555         jh->b_transaction = NULL;
1556 }
1557
1558 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1559 {
1560         jbd_lock_bh_state(jh2bh(jh));
1561         spin_lock(&journal->j_list_lock);
1562         __jbd2_journal_unfile_buffer(jh);
1563         spin_unlock(&journal->j_list_lock);
1564         jbd_unlock_bh_state(jh2bh(jh));
1565 }
1566
1567 /*
1568  * Called from jbd2_journal_try_to_free_buffers().
1569  *
1570  * Called under jbd_lock_bh_state(bh)
1571  */
1572 static void
1573 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1574 {
1575         struct journal_head *jh;
1576
1577         jh = bh2jh(bh);
1578
1579         if (buffer_locked(bh) || buffer_dirty(bh))
1580                 goto out;
1581
1582         if (jh->b_next_transaction != NULL)
1583                 goto out;
1584
1585         spin_lock(&journal->j_list_lock);
1586         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1587                 /* written-back checkpointed metadata buffer */
1588                 if (jh->b_jlist == BJ_None) {
1589                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1590                         __jbd2_journal_remove_checkpoint(jh);
1591                         jbd2_journal_remove_journal_head(bh);
1592                         __brelse(bh);
1593                 }
1594         }
1595         spin_unlock(&journal->j_list_lock);
1596 out:
1597         return;
1598 }
1599
1600 /**
1601  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1602  * @journal: journal for operation
1603  * @page: to try and free
1604  * @gfp_mask: we use the mask to detect how hard should we try to release
1605  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1606  * release the buffers.
1607  *
1608  *
1609  * For all the buffers on this page,
1610  * if they are fully written out ordered data, move them onto BUF_CLEAN
1611  * so try_to_free_buffers() can reap them.
1612  *
1613  * This function returns non-zero if we wish try_to_free_buffers()
1614  * to be called. We do this if the page is releasable by try_to_free_buffers().
1615  * We also do it if the page has locked or dirty buffers and the caller wants
1616  * us to perform sync or async writeout.
1617  *
1618  * This complicates JBD locking somewhat.  We aren't protected by the
1619  * BKL here.  We wish to remove the buffer from its committing or
1620  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1621  *
1622  * This may *change* the value of transaction_t->t_datalist, so anyone
1623  * who looks at t_datalist needs to lock against this function.
1624  *
1625  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1626  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1627  * will come out of the lock with the buffer dirty, which makes it
1628  * ineligible for release here.
1629  *
1630  * Who else is affected by this?  hmm...  Really the only contender
1631  * is do_get_write_access() - it could be looking at the buffer while
1632  * journal_try_to_free_buffer() is changing its state.  But that
1633  * cannot happen because we never reallocate freed data as metadata
1634  * while the data is part of a transaction.  Yes?
1635  *
1636  * Return 0 on failure, 1 on success
1637  */
1638 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1639                                 struct page *page, gfp_t gfp_mask)
1640 {
1641         struct buffer_head *head;
1642         struct buffer_head *bh;
1643         int ret = 0;
1644
1645         J_ASSERT(PageLocked(page));
1646
1647         head = page_buffers(page);
1648         bh = head;
1649         do {
1650                 struct journal_head *jh;
1651
1652                 /*
1653                  * We take our own ref against the journal_head here to avoid
1654                  * having to add tons of locking around each instance of
1655                  * jbd2_journal_remove_journal_head() and
1656                  * jbd2_journal_put_journal_head().
1657                  */
1658                 jh = jbd2_journal_grab_journal_head(bh);
1659                 if (!jh)
1660                         continue;
1661
1662                 jbd_lock_bh_state(bh);
1663                 __journal_try_to_free_buffer(journal, bh);
1664                 jbd2_journal_put_journal_head(jh);
1665                 jbd_unlock_bh_state(bh);
1666                 if (buffer_jbd(bh))
1667                         goto busy;
1668         } while ((bh = bh->b_this_page) != head);
1669
1670         ret = try_to_free_buffers(page);
1671
1672 busy:
1673         return ret;
1674 }
1675
1676 /*
1677  * This buffer is no longer needed.  If it is on an older transaction's
1678  * checkpoint list we need to record it on this transaction's forget list
1679  * to pin this buffer (and hence its checkpointing transaction) down until
1680  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1681  * release it.
1682  * Returns non-zero if JBD no longer has an interest in the buffer.
1683  *
1684  * Called under j_list_lock.
1685  *
1686  * Called under jbd_lock_bh_state(bh).
1687  */
1688 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1689 {
1690         int may_free = 1;
1691         struct buffer_head *bh = jh2bh(jh);
1692
1693         __jbd2_journal_unfile_buffer(jh);
1694
1695         if (jh->b_cp_transaction) {
1696                 JBUFFER_TRACE(jh, "on running+cp transaction");
1697                 /*
1698                  * We don't want to write the buffer anymore, clear the
1699                  * bit so that we don't confuse checks in
1700                  * __journal_file_buffer
1701                  */
1702                 clear_buffer_dirty(bh);
1703                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1704                 may_free = 0;
1705         } else {
1706                 JBUFFER_TRACE(jh, "on running transaction");
1707                 jbd2_journal_remove_journal_head(bh);
1708                 __brelse(bh);
1709         }
1710         return may_free;
1711 }
1712
1713 /*
1714  * jbd2_journal_invalidatepage
1715  *
1716  * This code is tricky.  It has a number of cases to deal with.
1717  *
1718  * There are two invariants which this code relies on:
1719  *
1720  * i_size must be updated on disk before we start calling invalidatepage on the
1721  * data.
1722  *
1723  *  This is done in ext3 by defining an ext3_setattr method which
1724  *  updates i_size before truncate gets going.  By maintaining this
1725  *  invariant, we can be sure that it is safe to throw away any buffers
1726  *  attached to the current transaction: once the transaction commits,
1727  *  we know that the data will not be needed.
1728  *
1729  *  Note however that we can *not* throw away data belonging to the
1730  *  previous, committing transaction!
1731  *
1732  * Any disk blocks which *are* part of the previous, committing
1733  * transaction (and which therefore cannot be discarded immediately) are
1734  * not going to be reused in the new running transaction
1735  *
1736  *  The bitmap committed_data images guarantee this: any block which is
1737  *  allocated in one transaction and removed in the next will be marked
1738  *  as in-use in the committed_data bitmap, so cannot be reused until
1739  *  the next transaction to delete the block commits.  This means that
1740  *  leaving committing buffers dirty is quite safe: the disk blocks
1741  *  cannot be reallocated to a different file and so buffer aliasing is
1742  *  not possible.
1743  *
1744  *
1745  * The above applies mainly to ordered data mode.  In writeback mode we
1746  * don't make guarantees about the order in which data hits disk --- in
1747  * particular we don't guarantee that new dirty data is flushed before
1748  * transaction commit --- so it is always safe just to discard data
1749  * immediately in that mode.  --sct
1750  */
1751
1752 /*
1753  * The journal_unmap_buffer helper function returns zero if the buffer
1754  * concerned remains pinned as an anonymous buffer belonging to an older
1755  * transaction.
1756  *
1757  * We're outside-transaction here.  Either or both of j_running_transaction
1758  * and j_committing_transaction may be NULL.
1759  */
1760 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1761 {
1762         transaction_t *transaction;
1763         struct journal_head *jh;
1764         int may_free = 1;
1765         int ret;
1766
1767         BUFFER_TRACE(bh, "entry");
1768
1769         /*
1770          * It is safe to proceed here without the j_list_lock because the
1771          * buffers cannot be stolen by try_to_free_buffers as long as we are
1772          * holding the page lock. --sct
1773          */
1774
1775         if (!buffer_jbd(bh))
1776                 goto zap_buffer_unlocked;
1777
1778         /* OK, we have data buffer in journaled mode */
1779         write_lock(&journal->j_state_lock);
1780         jbd_lock_bh_state(bh);
1781         spin_lock(&journal->j_list_lock);
1782
1783         jh = jbd2_journal_grab_journal_head(bh);
1784         if (!jh)
1785                 goto zap_buffer_no_jh;
1786
1787         /*
1788          * We cannot remove the buffer from checkpoint lists until the
1789          * transaction adding inode to orphan list (let's call it T)
1790          * is committed.  Otherwise if the transaction changing the
1791          * buffer would be cleaned from the journal before T is
1792          * committed, a crash will cause that the correct contents of
1793          * the buffer will be lost.  On the other hand we have to
1794          * clear the buffer dirty bit at latest at the moment when the
1795          * transaction marking the buffer as freed in the filesystem
1796          * structures is committed because from that moment on the
1797          * buffer can be reallocated and used by a different page.
1798          * Since the block hasn't been freed yet but the inode has
1799          * already been added to orphan list, it is safe for us to add
1800          * the buffer to BJ_Forget list of the newest transaction.
1801          */
1802         transaction = jh->b_transaction;
1803         if (transaction == NULL) {
1804                 /* First case: not on any transaction.  If it
1805                  * has no checkpoint link, then we can zap it:
1806                  * it's a writeback-mode buffer so we don't care
1807                  * if it hits disk safely. */
1808                 if (!jh->b_cp_transaction) {
1809                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1810                         goto zap_buffer;
1811                 }
1812
1813                 if (!buffer_dirty(bh)) {
1814                         /* bdflush has written it.  We can drop it now */
1815                         goto zap_buffer;
1816                 }
1817
1818                 /* OK, it must be in the journal but still not
1819                  * written fully to disk: it's metadata or
1820                  * journaled data... */
1821
1822                 if (journal->j_running_transaction) {
1823                         /* ... and once the current transaction has
1824                          * committed, the buffer won't be needed any
1825                          * longer. */
1826                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1827                         ret = __dispose_buffer(jh,
1828                                         journal->j_running_transaction);
1829                         jbd2_journal_put_journal_head(jh);
1830                         spin_unlock(&journal->j_list_lock);
1831                         jbd_unlock_bh_state(bh);
1832                         write_unlock(&journal->j_state_lock);
1833                         return ret;
1834                 } else {
1835                         /* There is no currently-running transaction. So the
1836                          * orphan record which we wrote for this file must have
1837                          * passed into commit.  We must attach this buffer to
1838                          * the committing transaction, if it exists. */
1839                         if (journal->j_committing_transaction) {
1840                                 JBUFFER_TRACE(jh, "give to committing trans");
1841                                 ret = __dispose_buffer(jh,
1842                                         journal->j_committing_transaction);
1843                                 jbd2_journal_put_journal_head(jh);
1844                                 spin_unlock(&journal->j_list_lock);
1845                                 jbd_unlock_bh_state(bh);
1846                                 write_unlock(&journal->j_state_lock);
1847                                 return ret;
1848                         } else {
1849                                 /* The orphan record's transaction has
1850                                  * committed.  We can cleanse this buffer */
1851                                 clear_buffer_jbddirty(bh);
1852                                 goto zap_buffer;
1853                         }
1854                 }
1855         } else if (transaction == journal->j_committing_transaction) {
1856                 JBUFFER_TRACE(jh, "on committing transaction");
1857                 /*
1858                  * The buffer is committing, we simply cannot touch
1859                  * it. So we just set j_next_transaction to the
1860                  * running transaction (if there is one) and mark
1861                  * buffer as freed so that commit code knows it should
1862                  * clear dirty bits when it is done with the buffer.
1863                  */
1864                 set_buffer_freed(bh);
1865                 if (journal->j_running_transaction && buffer_jbddirty(bh))
1866                         jh->b_next_transaction = journal->j_running_transaction;
1867                 jbd2_journal_put_journal_head(jh);
1868                 spin_unlock(&journal->j_list_lock);
1869                 jbd_unlock_bh_state(bh);
1870                 write_unlock(&journal->j_state_lock);
1871                 return 0;
1872         } else {
1873                 /* Good, the buffer belongs to the running transaction.
1874                  * We are writing our own transaction's data, not any
1875                  * previous one's, so it is safe to throw it away
1876                  * (remember that we expect the filesystem to have set
1877                  * i_size already for this truncate so recovery will not
1878                  * expose the disk blocks we are discarding here.) */
1879                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1880                 JBUFFER_TRACE(jh, "on running transaction");
1881                 may_free = __dispose_buffer(jh, transaction);
1882         }
1883
1884 zap_buffer:
1885         jbd2_journal_put_journal_head(jh);
1886 zap_buffer_no_jh:
1887         spin_unlock(&journal->j_list_lock);
1888         jbd_unlock_bh_state(bh);
1889         write_unlock(&journal->j_state_lock);
1890 zap_buffer_unlocked:
1891         clear_buffer_dirty(bh);
1892         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1893         clear_buffer_mapped(bh);
1894         clear_buffer_req(bh);
1895         clear_buffer_new(bh);
1896         bh->b_bdev = NULL;
1897         return may_free;
1898 }
1899
1900 /**
1901  * void jbd2_journal_invalidatepage()
1902  * @journal: journal to use for flush...
1903  * @page:    page to flush
1904  * @offset:  length of page to invalidate.
1905  *
1906  * Reap page buffers containing data after offset in page.
1907  *
1908  */
1909 void jbd2_journal_invalidatepage(journal_t *journal,
1910                       struct page *page,
1911                       unsigned long offset)
1912 {
1913         struct buffer_head *head, *bh, *next;
1914         unsigned int curr_off = 0;
1915         int may_free = 1;
1916
1917         if (!PageLocked(page))
1918                 BUG();
1919         if (!page_has_buffers(page))
1920                 return;
1921
1922         /* We will potentially be playing with lists other than just the
1923          * data lists (especially for journaled data mode), so be
1924          * cautious in our locking. */
1925
1926         head = bh = page_buffers(page);
1927         do {
1928                 unsigned int next_off = curr_off + bh->b_size;
1929                 next = bh->b_this_page;
1930
1931                 if (offset <= curr_off) {
1932                         /* This block is wholly outside the truncation point */
1933                         lock_buffer(bh);
1934                         may_free &= journal_unmap_buffer(journal, bh);
1935                         unlock_buffer(bh);
1936                 }
1937                 curr_off = next_off;
1938                 bh = next;
1939
1940         } while (bh != head);
1941
1942         if (!offset) {
1943                 if (may_free && try_to_free_buffers(page))
1944                         J_ASSERT(!page_has_buffers(page));
1945         }
1946 }
1947
1948 /*
1949  * File a buffer on the given transaction list.
1950  */
1951 void __jbd2_journal_file_buffer(struct journal_head *jh,
1952                         transaction_t *transaction, int jlist)
1953 {
1954         struct journal_head **list = NULL;
1955         int was_dirty = 0;
1956         struct buffer_head *bh = jh2bh(jh);
1957
1958         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1959         assert_spin_locked(&transaction->t_journal->j_list_lock);
1960
1961         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1962         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1963                                 jh->b_transaction == NULL);
1964
1965         if (jh->b_transaction && jh->b_jlist == jlist)
1966                 return;
1967
1968         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1969             jlist == BJ_Shadow || jlist == BJ_Forget) {
1970                 /*
1971                  * For metadata buffers, we track dirty bit in buffer_jbddirty
1972                  * instead of buffer_dirty. We should not see a dirty bit set
1973                  * here because we clear it in do_get_write_access but e.g.
1974                  * tune2fs can modify the sb and set the dirty bit at any time
1975                  * so we try to gracefully handle that.
1976                  */
1977                 if (buffer_dirty(bh))
1978                         warn_dirty_buffer(bh);
1979                 if (test_clear_buffer_dirty(bh) ||
1980                     test_clear_buffer_jbddirty(bh))
1981                         was_dirty = 1;
1982         }
1983
1984         if (jh->b_transaction)
1985                 __jbd2_journal_temp_unlink_buffer(jh);
1986         jh->b_transaction = transaction;
1987
1988         switch (jlist) {
1989         case BJ_None:
1990                 J_ASSERT_JH(jh, !jh->b_committed_data);
1991                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1992                 return;
1993         case BJ_Metadata:
1994                 transaction->t_nr_buffers++;
1995                 list = &transaction->t_buffers;
1996                 break;
1997         case BJ_Forget:
1998                 list = &transaction->t_forget;
1999                 break;
2000         case BJ_IO:
2001                 list = &transaction->t_iobuf_list;
2002                 break;
2003         case BJ_Shadow:
2004                 list = &transaction->t_shadow_list;
2005                 break;
2006         case BJ_LogCtl:
2007                 list = &transaction->t_log_list;
2008                 break;
2009         case BJ_Reserved:
2010                 list = &transaction->t_reserved_list;
2011                 break;
2012         }
2013
2014         __blist_add_buffer(list, jh);
2015         jh->b_jlist = jlist;
2016
2017         if (was_dirty)
2018                 set_buffer_jbddirty(bh);
2019 }
2020
2021 void jbd2_journal_file_buffer(struct journal_head *jh,
2022                                 transaction_t *transaction, int jlist)
2023 {
2024         jbd_lock_bh_state(jh2bh(jh));
2025         spin_lock(&transaction->t_journal->j_list_lock);
2026         __jbd2_journal_file_buffer(jh, transaction, jlist);
2027         spin_unlock(&transaction->t_journal->j_list_lock);
2028         jbd_unlock_bh_state(jh2bh(jh));
2029 }
2030
2031 /*
2032  * Remove a buffer from its current buffer list in preparation for
2033  * dropping it from its current transaction entirely.  If the buffer has
2034  * already started to be used by a subsequent transaction, refile the
2035  * buffer on that transaction's metadata list.
2036  *
2037  * Called under journal->j_list_lock
2038  *
2039  * Called under jbd_lock_bh_state(jh2bh(jh))
2040  */
2041 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2042 {
2043         int was_dirty, jlist;
2044         struct buffer_head *bh = jh2bh(jh);
2045
2046         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2047         if (jh->b_transaction)
2048                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2049
2050         /* If the buffer is now unused, just drop it. */
2051         if (jh->b_next_transaction == NULL) {
2052                 __jbd2_journal_unfile_buffer(jh);
2053                 return;
2054         }
2055
2056         /*
2057          * It has been modified by a later transaction: add it to the new
2058          * transaction's metadata list.
2059          */
2060
2061         was_dirty = test_clear_buffer_jbddirty(bh);
2062         __jbd2_journal_temp_unlink_buffer(jh);
2063         jh->b_transaction = jh->b_next_transaction;
2064         jh->b_next_transaction = NULL;
2065         if (buffer_freed(bh))
2066                 jlist = BJ_Forget;
2067         else if (jh->b_modified)
2068                 jlist = BJ_Metadata;
2069         else
2070                 jlist = BJ_Reserved;
2071         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2072         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2073
2074         if (was_dirty)
2075                 set_buffer_jbddirty(bh);
2076 }
2077
2078 /*
2079  * For the unlocked version of this call, also make sure that any
2080  * hanging journal_head is cleaned up if necessary.
2081  *
2082  * __jbd2_journal_refile_buffer is usually called as part of a single locked
2083  * operation on a buffer_head, in which the caller is probably going to
2084  * be hooking the journal_head onto other lists.  In that case it is up
2085  * to the caller to remove the journal_head if necessary.  For the
2086  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2087  * doing anything else to the buffer so we need to do the cleanup
2088  * ourselves to avoid a jh leak.
2089  *
2090  * *** The journal_head may be freed by this call! ***
2091  */
2092 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2093 {
2094         struct buffer_head *bh = jh2bh(jh);
2095
2096         jbd_lock_bh_state(bh);
2097         spin_lock(&journal->j_list_lock);
2098
2099         __jbd2_journal_refile_buffer(jh);
2100         jbd_unlock_bh_state(bh);
2101         jbd2_journal_remove_journal_head(bh);
2102
2103         spin_unlock(&journal->j_list_lock);
2104         __brelse(bh);
2105 }
2106
2107 /*
2108  * File inode in the inode list of the handle's transaction
2109  */
2110 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2111 {
2112         transaction_t *transaction = handle->h_transaction;
2113         journal_t *journal = transaction->t_journal;
2114
2115         if (is_handle_aborted(handle))
2116                 return -EIO;
2117
2118         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2119                         transaction->t_tid);
2120
2121         /*
2122          * First check whether inode isn't already on the transaction's
2123          * lists without taking the lock. Note that this check is safe
2124          * without the lock as we cannot race with somebody removing inode
2125          * from the transaction. The reason is that we remove inode from the
2126          * transaction only in journal_release_jbd_inode() and when we commit
2127          * the transaction. We are guarded from the first case by holding
2128          * a reference to the inode. We are safe against the second case
2129          * because if jinode->i_transaction == transaction, commit code
2130          * cannot touch the transaction because we hold reference to it,
2131          * and if jinode->i_next_transaction == transaction, commit code
2132          * will only file the inode where we want it.
2133          */
2134         if (jinode->i_transaction == transaction ||
2135             jinode->i_next_transaction == transaction)
2136                 return 0;
2137
2138         spin_lock(&journal->j_list_lock);
2139
2140         if (jinode->i_transaction == transaction ||
2141             jinode->i_next_transaction == transaction)
2142                 goto done;
2143
2144         /* On some different transaction's list - should be
2145          * the committing one */
2146         if (jinode->i_transaction) {
2147                 J_ASSERT(jinode->i_next_transaction == NULL);
2148                 J_ASSERT(jinode->i_transaction ==
2149                                         journal->j_committing_transaction);
2150                 jinode->i_next_transaction = transaction;
2151                 goto done;
2152         }
2153         /* Not on any transaction list... */
2154         J_ASSERT(!jinode->i_next_transaction);
2155         jinode->i_transaction = transaction;
2156         list_add(&jinode->i_list, &transaction->t_inode_list);
2157 done:
2158         spin_unlock(&journal->j_list_lock);
2159
2160         return 0;
2161 }
2162
2163 /*
2164  * File truncate and transaction commit interact with each other in a
2165  * non-trivial way.  If a transaction writing data block A is
2166  * committing, we cannot discard the data by truncate until we have
2167  * written them.  Otherwise if we crashed after the transaction with
2168  * write has committed but before the transaction with truncate has
2169  * committed, we could see stale data in block A.  This function is a
2170  * helper to solve this problem.  It starts writeout of the truncated
2171  * part in case it is in the committing transaction.
2172  *
2173  * Filesystem code must call this function when inode is journaled in
2174  * ordered mode before truncation happens and after the inode has been
2175  * placed on orphan list with the new inode size. The second condition
2176  * avoids the race that someone writes new data and we start
2177  * committing the transaction after this function has been called but
2178  * before a transaction for truncate is started (and furthermore it
2179  * allows us to optimize the case where the addition to orphan list
2180  * happens in the same transaction as write --- we don't have to write
2181  * any data in such case).
2182  */
2183 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2184                                         struct jbd2_inode *jinode,
2185                                         loff_t new_size)
2186 {
2187         transaction_t *inode_trans, *commit_trans;
2188         int ret = 0;
2189
2190         /* This is a quick check to avoid locking if not necessary */
2191         if (!jinode->i_transaction)
2192                 goto out;
2193         /* Locks are here just to force reading of recent values, it is
2194          * enough that the transaction was not committing before we started
2195          * a transaction adding the inode to orphan list */
2196         read_lock(&journal->j_state_lock);
2197         commit_trans = journal->j_committing_transaction;
2198         read_unlock(&journal->j_state_lock);
2199         spin_lock(&journal->j_list_lock);
2200         inode_trans = jinode->i_transaction;
2201         spin_unlock(&journal->j_list_lock);
2202         if (inode_trans == commit_trans) {
2203                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2204                         new_size, LLONG_MAX);
2205                 if (ret)
2206                         jbd2_journal_abort(journal, ret);
2207         }
2208 out:
2209         return ret;
2210 }