Fix misspellings of "whether" in comments.
[linux-3.10.git] / fs / jbd / transaction.c
1 /*
2  * linux/fs/jbd/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/jbd.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
31 static void __journal_temp_unlink_buffer(struct journal_head *jh);
32
33 /*
34  * get_transaction: obtain a new transaction_t object.
35  *
36  * Simply allocate and initialise a new transaction.  Create it in
37  * RUNNING state and add it to the current journal (which should not
38  * have an existing running transaction: we only make a new transaction
39  * once we have started to commit the old one).
40  *
41  * Preconditions:
42  *      The journal MUST be locked.  We don't perform atomic mallocs on the
43  *      new transaction and we can't block without protecting against other
44  *      processes trying to touch the journal while it is in transition.
45  *
46  * Called under j_state_lock
47  */
48
49 static transaction_t *
50 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
59         /* Set up the commit timer for the new transaction. */
60         journal->j_commit_timer.expires =
61                                 round_jiffies_up(transaction->t_expires);
62         add_timer(&journal->j_commit_timer);
63
64         J_ASSERT(journal->j_running_transaction == NULL);
65         journal->j_running_transaction = transaction;
66
67         return transaction;
68 }
69
70 /*
71  * Handle management.
72  *
73  * A handle_t is an object which represents a single atomic update to a
74  * filesystem, and which tracks all of the modifications which form part
75  * of that one update.
76  */
77
78 /*
79  * start_this_handle: Given a handle, deal with any locking or stalling
80  * needed to make sure that there is enough journal space for the handle
81  * to begin.  Attach the handle to a transaction and set up the
82  * transaction's buffer credits.
83  */
84
85 static int start_this_handle(journal_t *journal, handle_t *handle)
86 {
87         transaction_t *transaction;
88         int needed;
89         int nblocks = handle->h_buffer_credits;
90         transaction_t *new_transaction = NULL;
91         int ret = 0;
92
93         if (nblocks > journal->j_max_transaction_buffers) {
94                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
95                        current->comm, nblocks,
96                        journal->j_max_transaction_buffers);
97                 ret = -ENOSPC;
98                 goto out;
99         }
100
101 alloc_transaction:
102         if (!journal->j_running_transaction) {
103                 new_transaction = kzalloc(sizeof(*new_transaction), GFP_NOFS);
104                 if (!new_transaction) {
105                         congestion_wait(BLK_RW_ASYNC, HZ/50);
106                         goto alloc_transaction;
107                 }
108         }
109
110         jbd_debug(3, "New handle %p going live.\n", handle);
111
112 repeat:
113
114         /*
115          * We need to hold j_state_lock until t_updates has been incremented,
116          * for proper journal barrier handling
117          */
118         spin_lock(&journal->j_state_lock);
119 repeat_locked:
120         if (is_journal_aborted(journal) ||
121             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
122                 spin_unlock(&journal->j_state_lock);
123                 ret = -EROFS;
124                 goto out;
125         }
126
127         /* Wait on the journal's transaction barrier if necessary */
128         if (journal->j_barrier_count) {
129                 spin_unlock(&journal->j_state_lock);
130                 wait_event(journal->j_wait_transaction_locked,
131                                 journal->j_barrier_count == 0);
132                 goto repeat;
133         }
134
135         if (!journal->j_running_transaction) {
136                 if (!new_transaction) {
137                         spin_unlock(&journal->j_state_lock);
138                         goto alloc_transaction;
139                 }
140                 get_transaction(journal, new_transaction);
141                 new_transaction = NULL;
142         }
143
144         transaction = journal->j_running_transaction;
145
146         /*
147          * If the current transaction is locked down for commit, wait for the
148          * lock to be released.
149          */
150         if (transaction->t_state == T_LOCKED) {
151                 DEFINE_WAIT(wait);
152
153                 prepare_to_wait(&journal->j_wait_transaction_locked,
154                                         &wait, TASK_UNINTERRUPTIBLE);
155                 spin_unlock(&journal->j_state_lock);
156                 schedule();
157                 finish_wait(&journal->j_wait_transaction_locked, &wait);
158                 goto repeat;
159         }
160
161         /*
162          * If there is not enough space left in the log to write all potential
163          * buffers requested by this operation, we need to stall pending a log
164          * checkpoint to free some more log space.
165          */
166         spin_lock(&transaction->t_handle_lock);
167         needed = transaction->t_outstanding_credits + nblocks;
168
169         if (needed > journal->j_max_transaction_buffers) {
170                 /*
171                  * If the current transaction is already too large, then start
172                  * to commit it: we can then go back and attach this handle to
173                  * a new transaction.
174                  */
175                 DEFINE_WAIT(wait);
176
177                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
178                 spin_unlock(&transaction->t_handle_lock);
179                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
180                                 TASK_UNINTERRUPTIBLE);
181                 __log_start_commit(journal, transaction->t_tid);
182                 spin_unlock(&journal->j_state_lock);
183                 schedule();
184                 finish_wait(&journal->j_wait_transaction_locked, &wait);
185                 goto repeat;
186         }
187
188         /*
189          * The commit code assumes that it can get enough log space
190          * without forcing a checkpoint.  This is *critical* for
191          * correctness: a checkpoint of a buffer which is also
192          * associated with a committing transaction creates a deadlock,
193          * so commit simply cannot force through checkpoints.
194          *
195          * We must therefore ensure the necessary space in the journal
196          * *before* starting to dirty potentially checkpointed buffers
197          * in the new transaction.
198          *
199          * The worst part is, any transaction currently committing can
200          * reduce the free space arbitrarily.  Be careful to account for
201          * those buffers when checkpointing.
202          */
203
204         /*
205          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
206          * a _lot_ of headroom: 1/4 of the journal plus the size of
207          * the committing transaction.  Really, we only need to give it
208          * committing_transaction->t_outstanding_credits plus "enough" for
209          * the log control blocks.
210          * Also, this test is inconsistent with the matching one in
211          * journal_extend().
212          */
213         if (__log_space_left(journal) < jbd_space_needed(journal)) {
214                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
215                 spin_unlock(&transaction->t_handle_lock);
216                 __log_wait_for_space(journal);
217                 goto repeat_locked;
218         }
219
220         /* OK, account for the buffers that this operation expects to
221          * use and add the handle to the running transaction. */
222
223         handle->h_transaction = transaction;
224         transaction->t_outstanding_credits += nblocks;
225         transaction->t_updates++;
226         transaction->t_handle_count++;
227         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
228                   handle, nblocks, transaction->t_outstanding_credits,
229                   __log_space_left(journal));
230         spin_unlock(&transaction->t_handle_lock);
231         spin_unlock(&journal->j_state_lock);
232
233         lock_map_acquire(&handle->h_lockdep_map);
234 out:
235         if (unlikely(new_transaction))          /* It's usually NULL */
236                 kfree(new_transaction);
237         return ret;
238 }
239
240 static struct lock_class_key jbd_handle_key;
241
242 /* Allocate a new handle.  This should probably be in a slab... */
243 static handle_t *new_handle(int nblocks)
244 {
245         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
246         if (!handle)
247                 return NULL;
248         memset(handle, 0, sizeof(*handle));
249         handle->h_buffer_credits = nblocks;
250         handle->h_ref = 1;
251
252         lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
253
254         return handle;
255 }
256
257 /**
258  * handle_t *journal_start() - Obtain a new handle.
259  * @journal: Journal to start transaction on.
260  * @nblocks: number of block buffer we might modify
261  *
262  * We make sure that the transaction can guarantee at least nblocks of
263  * modified buffers in the log.  We block until the log can guarantee
264  * that much space.
265  *
266  * This function is visible to journal users (like ext3fs), so is not
267  * called with the journal already locked.
268  *
269  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
270  * on failure.
271  */
272 handle_t *journal_start(journal_t *journal, int nblocks)
273 {
274         handle_t *handle = journal_current_handle();
275         int err;
276
277         if (!journal)
278                 return ERR_PTR(-EROFS);
279
280         if (handle) {
281                 J_ASSERT(handle->h_transaction->t_journal == journal);
282                 handle->h_ref++;
283                 return handle;
284         }
285
286         handle = new_handle(nblocks);
287         if (!handle)
288                 return ERR_PTR(-ENOMEM);
289
290         current->journal_info = handle;
291
292         err = start_this_handle(journal, handle);
293         if (err < 0) {
294                 jbd_free_handle(handle);
295                 current->journal_info = NULL;
296                 handle = ERR_PTR(err);
297         }
298         return handle;
299 }
300
301 /**
302  * int journal_extend() - extend buffer credits.
303  * @handle:  handle to 'extend'
304  * @nblocks: nr blocks to try to extend by.
305  *
306  * Some transactions, such as large extends and truncates, can be done
307  * atomically all at once or in several stages.  The operation requests
308  * a credit for a number of buffer modications in advance, but can
309  * extend its credit if it needs more.
310  *
311  * journal_extend tries to give the running handle more buffer credits.
312  * It does not guarantee that allocation - this is a best-effort only.
313  * The calling process MUST be able to deal cleanly with a failure to
314  * extend here.
315  *
316  * Return 0 on success, non-zero on failure.
317  *
318  * return code < 0 implies an error
319  * return code > 0 implies normal transaction-full status.
320  */
321 int journal_extend(handle_t *handle, int nblocks)
322 {
323         transaction_t *transaction = handle->h_transaction;
324         journal_t *journal = transaction->t_journal;
325         int result;
326         int wanted;
327
328         result = -EIO;
329         if (is_handle_aborted(handle))
330                 goto out;
331
332         result = 1;
333
334         spin_lock(&journal->j_state_lock);
335
336         /* Don't extend a locked-down transaction! */
337         if (handle->h_transaction->t_state != T_RUNNING) {
338                 jbd_debug(3, "denied handle %p %d blocks: "
339                           "transaction not running\n", handle, nblocks);
340                 goto error_out;
341         }
342
343         spin_lock(&transaction->t_handle_lock);
344         wanted = transaction->t_outstanding_credits + nblocks;
345
346         if (wanted > journal->j_max_transaction_buffers) {
347                 jbd_debug(3, "denied handle %p %d blocks: "
348                           "transaction too large\n", handle, nblocks);
349                 goto unlock;
350         }
351
352         if (wanted > __log_space_left(journal)) {
353                 jbd_debug(3, "denied handle %p %d blocks: "
354                           "insufficient log space\n", handle, nblocks);
355                 goto unlock;
356         }
357
358         handle->h_buffer_credits += nblocks;
359         transaction->t_outstanding_credits += nblocks;
360         result = 0;
361
362         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
363 unlock:
364         spin_unlock(&transaction->t_handle_lock);
365 error_out:
366         spin_unlock(&journal->j_state_lock);
367 out:
368         return result;
369 }
370
371
372 /**
373  * int journal_restart() - restart a handle.
374  * @handle:  handle to restart
375  * @nblocks: nr credits requested
376  *
377  * Restart a handle for a multi-transaction filesystem
378  * operation.
379  *
380  * If the journal_extend() call above fails to grant new buffer credits
381  * to a running handle, a call to journal_restart will commit the
382  * handle's transaction so far and reattach the handle to a new
383  * transaction capabable of guaranteeing the requested number of
384  * credits.
385  */
386
387 int journal_restart(handle_t *handle, int nblocks)
388 {
389         transaction_t *transaction = handle->h_transaction;
390         journal_t *journal = transaction->t_journal;
391         int ret;
392
393         /* If we've had an abort of any type, don't even think about
394          * actually doing the restart! */
395         if (is_handle_aborted(handle))
396                 return 0;
397
398         /*
399          * First unlink the handle from its current transaction, and start the
400          * commit on that.
401          */
402         J_ASSERT(transaction->t_updates > 0);
403         J_ASSERT(journal_current_handle() == handle);
404
405         spin_lock(&journal->j_state_lock);
406         spin_lock(&transaction->t_handle_lock);
407         transaction->t_outstanding_credits -= handle->h_buffer_credits;
408         transaction->t_updates--;
409
410         if (!transaction->t_updates)
411                 wake_up(&journal->j_wait_updates);
412         spin_unlock(&transaction->t_handle_lock);
413
414         jbd_debug(2, "restarting handle %p\n", handle);
415         __log_start_commit(journal, transaction->t_tid);
416         spin_unlock(&journal->j_state_lock);
417
418         lock_map_release(&handle->h_lockdep_map);
419         handle->h_buffer_credits = nblocks;
420         ret = start_this_handle(journal, handle);
421         return ret;
422 }
423
424
425 /**
426  * void journal_lock_updates () - establish a transaction barrier.
427  * @journal:  Journal to establish a barrier on.
428  *
429  * This locks out any further updates from being started, and blocks until all
430  * existing updates have completed, returning only once the journal is in a
431  * quiescent state with no updates running.
432  *
433  * We do not use simple mutex for synchronization as there are syscalls which
434  * want to return with filesystem locked and that trips up lockdep. Also
435  * hibernate needs to lock filesystem but locked mutex then blocks hibernation.
436  * Since locking filesystem is rare operation, we use simple counter and
437  * waitqueue for locking.
438  */
439 void journal_lock_updates(journal_t *journal)
440 {
441         DEFINE_WAIT(wait);
442
443 wait:
444         /* Wait for previous locked operation to finish */
445         wait_event(journal->j_wait_transaction_locked,
446                    journal->j_barrier_count == 0);
447
448         spin_lock(&journal->j_state_lock);
449         /*
450          * Check reliably under the lock whether we are the ones winning the race
451          * and locking the journal
452          */
453         if (journal->j_barrier_count > 0) {
454                 spin_unlock(&journal->j_state_lock);
455                 goto wait;
456         }
457         ++journal->j_barrier_count;
458
459         /* Wait until there are no running updates */
460         while (1) {
461                 transaction_t *transaction = journal->j_running_transaction;
462
463                 if (!transaction)
464                         break;
465
466                 spin_lock(&transaction->t_handle_lock);
467                 if (!transaction->t_updates) {
468                         spin_unlock(&transaction->t_handle_lock);
469                         break;
470                 }
471                 prepare_to_wait(&journal->j_wait_updates, &wait,
472                                 TASK_UNINTERRUPTIBLE);
473                 spin_unlock(&transaction->t_handle_lock);
474                 spin_unlock(&journal->j_state_lock);
475                 schedule();
476                 finish_wait(&journal->j_wait_updates, &wait);
477                 spin_lock(&journal->j_state_lock);
478         }
479         spin_unlock(&journal->j_state_lock);
480 }
481
482 /**
483  * void journal_unlock_updates (journal_t* journal) - release barrier
484  * @journal:  Journal to release the barrier on.
485  *
486  * Release a transaction barrier obtained with journal_lock_updates().
487  */
488 void journal_unlock_updates (journal_t *journal)
489 {
490         J_ASSERT(journal->j_barrier_count != 0);
491
492         spin_lock(&journal->j_state_lock);
493         --journal->j_barrier_count;
494         spin_unlock(&journal->j_state_lock);
495         wake_up(&journal->j_wait_transaction_locked);
496 }
497
498 static void warn_dirty_buffer(struct buffer_head *bh)
499 {
500         char b[BDEVNAME_SIZE];
501
502         printk(KERN_WARNING
503                "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
504                "There's a risk of filesystem corruption in case of system "
505                "crash.\n",
506                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
507 }
508
509 /*
510  * If the buffer is already part of the current transaction, then there
511  * is nothing we need to do.  If it is already part of a prior
512  * transaction which we are still committing to disk, then we need to
513  * make sure that we do not overwrite the old copy: we do copy-out to
514  * preserve the copy going to disk.  We also account the buffer against
515  * the handle's metadata buffer credits (unless the buffer is already
516  * part of the transaction, that is).
517  *
518  */
519 static int
520 do_get_write_access(handle_t *handle, struct journal_head *jh,
521                         int force_copy)
522 {
523         struct buffer_head *bh;
524         transaction_t *transaction;
525         journal_t *journal;
526         int error;
527         char *frozen_buffer = NULL;
528         int need_copy = 0;
529
530         if (is_handle_aborted(handle))
531                 return -EROFS;
532
533         transaction = handle->h_transaction;
534         journal = transaction->t_journal;
535
536         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
537
538         JBUFFER_TRACE(jh, "entry");
539 repeat:
540         bh = jh2bh(jh);
541
542         /* @@@ Need to check for errors here at some point. */
543
544         lock_buffer(bh);
545         jbd_lock_bh_state(bh);
546
547         /* We now hold the buffer lock so it is safe to query the buffer
548          * state.  Is the buffer dirty?
549          *
550          * If so, there are two possibilities.  The buffer may be
551          * non-journaled, and undergoing a quite legitimate writeback.
552          * Otherwise, it is journaled, and we don't expect dirty buffers
553          * in that state (the buffers should be marked JBD_Dirty
554          * instead.)  So either the IO is being done under our own
555          * control and this is a bug, or it's a third party IO such as
556          * dump(8) (which may leave the buffer scheduled for read ---
557          * ie. locked but not dirty) or tune2fs (which may actually have
558          * the buffer dirtied, ugh.)  */
559
560         if (buffer_dirty(bh)) {
561                 /*
562                  * First question: is this buffer already part of the current
563                  * transaction or the existing committing transaction?
564                  */
565                 if (jh->b_transaction) {
566                         J_ASSERT_JH(jh,
567                                 jh->b_transaction == transaction ||
568                                 jh->b_transaction ==
569                                         journal->j_committing_transaction);
570                         if (jh->b_next_transaction)
571                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
572                                                         transaction);
573                         warn_dirty_buffer(bh);
574                 }
575                 /*
576                  * In any case we need to clean the dirty flag and we must
577                  * do it under the buffer lock to be sure we don't race
578                  * with running write-out.
579                  */
580                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
581                 clear_buffer_dirty(bh);
582                 set_buffer_jbddirty(bh);
583         }
584
585         unlock_buffer(bh);
586
587         error = -EROFS;
588         if (is_handle_aborted(handle)) {
589                 jbd_unlock_bh_state(bh);
590                 goto out;
591         }
592         error = 0;
593
594         /*
595          * The buffer is already part of this transaction if b_transaction or
596          * b_next_transaction points to it
597          */
598         if (jh->b_transaction == transaction ||
599             jh->b_next_transaction == transaction)
600                 goto done;
601
602         /*
603          * this is the first time this transaction is touching this buffer,
604          * reset the modified flag
605          */
606         jh->b_modified = 0;
607
608         /*
609          * If there is already a copy-out version of this buffer, then we don't
610          * need to make another one
611          */
612         if (jh->b_frozen_data) {
613                 JBUFFER_TRACE(jh, "has frozen data");
614                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
615                 jh->b_next_transaction = transaction;
616                 goto done;
617         }
618
619         /* Is there data here we need to preserve? */
620
621         if (jh->b_transaction && jh->b_transaction != transaction) {
622                 JBUFFER_TRACE(jh, "owned by older transaction");
623                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
624                 J_ASSERT_JH(jh, jh->b_transaction ==
625                                         journal->j_committing_transaction);
626
627                 /* There is one case we have to be very careful about.
628                  * If the committing transaction is currently writing
629                  * this buffer out to disk and has NOT made a copy-out,
630                  * then we cannot modify the buffer contents at all
631                  * right now.  The essence of copy-out is that it is the
632                  * extra copy, not the primary copy, which gets
633                  * journaled.  If the primary copy is already going to
634                  * disk then we cannot do copy-out here. */
635
636                 if (jh->b_jlist == BJ_Shadow) {
637                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
638                         wait_queue_head_t *wqh;
639
640                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
641
642                         JBUFFER_TRACE(jh, "on shadow: sleep");
643                         jbd_unlock_bh_state(bh);
644                         /* commit wakes up all shadow buffers after IO */
645                         for ( ; ; ) {
646                                 prepare_to_wait(wqh, &wait.wait,
647                                                 TASK_UNINTERRUPTIBLE);
648                                 if (jh->b_jlist != BJ_Shadow)
649                                         break;
650                                 schedule();
651                         }
652                         finish_wait(wqh, &wait.wait);
653                         goto repeat;
654                 }
655
656                 /* Only do the copy if the currently-owning transaction
657                  * still needs it.  If it is on the Forget list, the
658                  * committing transaction is past that stage.  The
659                  * buffer had better remain locked during the kmalloc,
660                  * but that should be true --- we hold the journal lock
661                  * still and the buffer is already on the BUF_JOURNAL
662                  * list so won't be flushed.
663                  *
664                  * Subtle point, though: if this is a get_undo_access,
665                  * then we will be relying on the frozen_data to contain
666                  * the new value of the committed_data record after the
667                  * transaction, so we HAVE to force the frozen_data copy
668                  * in that case. */
669
670                 if (jh->b_jlist != BJ_Forget || force_copy) {
671                         JBUFFER_TRACE(jh, "generate frozen data");
672                         if (!frozen_buffer) {
673                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
674                                 jbd_unlock_bh_state(bh);
675                                 frozen_buffer =
676                                         jbd_alloc(jh2bh(jh)->b_size,
677                                                          GFP_NOFS);
678                                 if (!frozen_buffer) {
679                                         printk(KERN_EMERG
680                                                "%s: OOM for frozen_buffer\n",
681                                                __func__);
682                                         JBUFFER_TRACE(jh, "oom!");
683                                         error = -ENOMEM;
684                                         jbd_lock_bh_state(bh);
685                                         goto done;
686                                 }
687                                 goto repeat;
688                         }
689                         jh->b_frozen_data = frozen_buffer;
690                         frozen_buffer = NULL;
691                         need_copy = 1;
692                 }
693                 jh->b_next_transaction = transaction;
694         }
695
696
697         /*
698          * Finally, if the buffer is not journaled right now, we need to make
699          * sure it doesn't get written to disk before the caller actually
700          * commits the new data
701          */
702         if (!jh->b_transaction) {
703                 JBUFFER_TRACE(jh, "no transaction");
704                 J_ASSERT_JH(jh, !jh->b_next_transaction);
705                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
706                 spin_lock(&journal->j_list_lock);
707                 __journal_file_buffer(jh, transaction, BJ_Reserved);
708                 spin_unlock(&journal->j_list_lock);
709         }
710
711 done:
712         if (need_copy) {
713                 struct page *page;
714                 int offset;
715                 char *source;
716
717                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
718                             "Possible IO failure.\n");
719                 page = jh2bh(jh)->b_page;
720                 offset = offset_in_page(jh2bh(jh)->b_data);
721                 source = kmap_atomic(page);
722                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
723                 kunmap_atomic(source);
724         }
725         jbd_unlock_bh_state(bh);
726
727         /*
728          * If we are about to journal a buffer, then any revoke pending on it is
729          * no longer valid
730          */
731         journal_cancel_revoke(handle, jh);
732
733 out:
734         if (unlikely(frozen_buffer))    /* It's usually NULL */
735                 jbd_free(frozen_buffer, bh->b_size);
736
737         JBUFFER_TRACE(jh, "exit");
738         return error;
739 }
740
741 /**
742  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
743  * @handle: transaction to add buffer modifications to
744  * @bh:     bh to be used for metadata writes
745  *
746  * Returns an error code or 0 on success.
747  *
748  * In full data journalling mode the buffer may be of type BJ_AsyncData,
749  * because we're write()ing a buffer which is also part of a shared mapping.
750  */
751
752 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
753 {
754         struct journal_head *jh = journal_add_journal_head(bh);
755         int rc;
756
757         /* We do not want to get caught playing with fields which the
758          * log thread also manipulates.  Make sure that the buffer
759          * completes any outstanding IO before proceeding. */
760         rc = do_get_write_access(handle, jh, 0);
761         journal_put_journal_head(jh);
762         return rc;
763 }
764
765
766 /*
767  * When the user wants to journal a newly created buffer_head
768  * (ie. getblk() returned a new buffer and we are going to populate it
769  * manually rather than reading off disk), then we need to keep the
770  * buffer_head locked until it has been completely filled with new
771  * data.  In this case, we should be able to make the assertion that
772  * the bh is not already part of an existing transaction.
773  *
774  * The buffer should already be locked by the caller by this point.
775  * There is no lock ranking violation: it was a newly created,
776  * unlocked buffer beforehand. */
777
778 /**
779  * int journal_get_create_access () - notify intent to use newly created bh
780  * @handle: transaction to new buffer to
781  * @bh: new buffer.
782  *
783  * Call this if you create a new bh.
784  */
785 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
786 {
787         transaction_t *transaction = handle->h_transaction;
788         journal_t *journal = transaction->t_journal;
789         struct journal_head *jh = journal_add_journal_head(bh);
790         int err;
791
792         jbd_debug(5, "journal_head %p\n", jh);
793         err = -EROFS;
794         if (is_handle_aborted(handle))
795                 goto out;
796         err = 0;
797
798         JBUFFER_TRACE(jh, "entry");
799         /*
800          * The buffer may already belong to this transaction due to pre-zeroing
801          * in the filesystem's new_block code.  It may also be on the previous,
802          * committing transaction's lists, but it HAS to be in Forget state in
803          * that case: the transaction must have deleted the buffer for it to be
804          * reused here.
805          */
806         jbd_lock_bh_state(bh);
807         spin_lock(&journal->j_list_lock);
808         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
809                 jh->b_transaction == NULL ||
810                 (jh->b_transaction == journal->j_committing_transaction &&
811                           jh->b_jlist == BJ_Forget)));
812
813         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
814         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
815
816         if (jh->b_transaction == NULL) {
817                 /*
818                  * Previous journal_forget() could have left the buffer
819                  * with jbddirty bit set because it was being committed. When
820                  * the commit finished, we've filed the buffer for
821                  * checkpointing and marked it dirty. Now we are reallocating
822                  * the buffer so the transaction freeing it must have
823                  * committed and so it's safe to clear the dirty bit.
824                  */
825                 clear_buffer_dirty(jh2bh(jh));
826
827                 /* first access by this transaction */
828                 jh->b_modified = 0;
829
830                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
831                 __journal_file_buffer(jh, transaction, BJ_Reserved);
832         } else if (jh->b_transaction == journal->j_committing_transaction) {
833                 /* first access by this transaction */
834                 jh->b_modified = 0;
835
836                 JBUFFER_TRACE(jh, "set next transaction");
837                 jh->b_next_transaction = transaction;
838         }
839         spin_unlock(&journal->j_list_lock);
840         jbd_unlock_bh_state(bh);
841
842         /*
843          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
844          * blocks which contain freed but then revoked metadata.  We need
845          * to cancel the revoke in case we end up freeing it yet again
846          * and the reallocating as data - this would cause a second revoke,
847          * which hits an assertion error.
848          */
849         JBUFFER_TRACE(jh, "cancelling revoke");
850         journal_cancel_revoke(handle, jh);
851 out:
852         journal_put_journal_head(jh);
853         return err;
854 }
855
856 /**
857  * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
858  * @handle: transaction
859  * @bh: buffer to undo
860  *
861  * Sometimes there is a need to distinguish between metadata which has
862  * been committed to disk and that which has not.  The ext3fs code uses
863  * this for freeing and allocating space, we have to make sure that we
864  * do not reuse freed space until the deallocation has been committed,
865  * since if we overwrote that space we would make the delete
866  * un-rewindable in case of a crash.
867  *
868  * To deal with that, journal_get_undo_access requests write access to a
869  * buffer for parts of non-rewindable operations such as delete
870  * operations on the bitmaps.  The journaling code must keep a copy of
871  * the buffer's contents prior to the undo_access call until such time
872  * as we know that the buffer has definitely been committed to disk.
873  *
874  * We never need to know which transaction the committed data is part
875  * of, buffers touched here are guaranteed to be dirtied later and so
876  * will be committed to a new transaction in due course, at which point
877  * we can discard the old committed data pointer.
878  *
879  * Returns error number or 0 on success.
880  */
881 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
882 {
883         int err;
884         struct journal_head *jh = journal_add_journal_head(bh);
885         char *committed_data = NULL;
886
887         JBUFFER_TRACE(jh, "entry");
888
889         /*
890          * Do this first --- it can drop the journal lock, so we want to
891          * make sure that obtaining the committed_data is done
892          * atomically wrt. completion of any outstanding commits.
893          */
894         err = do_get_write_access(handle, jh, 1);
895         if (err)
896                 goto out;
897
898 repeat:
899         if (!jh->b_committed_data) {
900                 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
901                 if (!committed_data) {
902                         printk(KERN_EMERG "%s: No memory for committed data\n",
903                                 __func__);
904                         err = -ENOMEM;
905                         goto out;
906                 }
907         }
908
909         jbd_lock_bh_state(bh);
910         if (!jh->b_committed_data) {
911                 /* Copy out the current buffer contents into the
912                  * preserved, committed copy. */
913                 JBUFFER_TRACE(jh, "generate b_committed data");
914                 if (!committed_data) {
915                         jbd_unlock_bh_state(bh);
916                         goto repeat;
917                 }
918
919                 jh->b_committed_data = committed_data;
920                 committed_data = NULL;
921                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
922         }
923         jbd_unlock_bh_state(bh);
924 out:
925         journal_put_journal_head(jh);
926         if (unlikely(committed_data))
927                 jbd_free(committed_data, bh->b_size);
928         return err;
929 }
930
931 /**
932  * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
933  * @handle: transaction
934  * @bh: bufferhead to mark
935  *
936  * Description:
937  * Mark a buffer as containing dirty data which needs to be flushed before
938  * we can commit the current transaction.
939  *
940  * The buffer is placed on the transaction's data list and is marked as
941  * belonging to the transaction.
942  *
943  * Returns error number or 0 on success.
944  *
945  * journal_dirty_data() can be called via page_launder->ext3_writepage
946  * by kswapd.
947  */
948 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
949 {
950         journal_t *journal = handle->h_transaction->t_journal;
951         int need_brelse = 0;
952         struct journal_head *jh;
953         int ret = 0;
954
955         if (is_handle_aborted(handle))
956                 return ret;
957
958         jh = journal_add_journal_head(bh);
959         JBUFFER_TRACE(jh, "entry");
960
961         /*
962          * The buffer could *already* be dirty.  Writeout can start
963          * at any time.
964          */
965         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
966
967         /*
968          * What if the buffer is already part of a running transaction?
969          *
970          * There are two cases:
971          * 1) It is part of the current running transaction.  Refile it,
972          *    just in case we have allocated it as metadata, deallocated
973          *    it, then reallocated it as data.
974          * 2) It is part of the previous, still-committing transaction.
975          *    If all we want to do is to guarantee that the buffer will be
976          *    written to disk before this new transaction commits, then
977          *    being sure that the *previous* transaction has this same
978          *    property is sufficient for us!  Just leave it on its old
979          *    transaction.
980          *
981          * In case (2), the buffer must not already exist as metadata
982          * --- that would violate write ordering (a transaction is free
983          * to write its data at any point, even before the previous
984          * committing transaction has committed).  The caller must
985          * never, ever allow this to happen: there's nothing we can do
986          * about it in this layer.
987          */
988         jbd_lock_bh_state(bh);
989         spin_lock(&journal->j_list_lock);
990
991         /* Now that we have bh_state locked, are we really still mapped? */
992         if (!buffer_mapped(bh)) {
993                 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
994                 goto no_journal;
995         }
996
997         if (jh->b_transaction) {
998                 JBUFFER_TRACE(jh, "has transaction");
999                 if (jh->b_transaction != handle->h_transaction) {
1000                         JBUFFER_TRACE(jh, "belongs to older transaction");
1001                         J_ASSERT_JH(jh, jh->b_transaction ==
1002                                         journal->j_committing_transaction);
1003
1004                         /* @@@ IS THIS TRUE  ? */
1005                         /*
1006                          * Not any more.  Scenario: someone does a write()
1007                          * in data=journal mode.  The buffer's transaction has
1008                          * moved into commit.  Then someone does another
1009                          * write() to the file.  We do the frozen data copyout
1010                          * and set b_next_transaction to point to j_running_t.
1011                          * And while we're in that state, someone does a
1012                          * writepage() in an attempt to pageout the same area
1013                          * of the file via a shared mapping.  At present that
1014                          * calls journal_dirty_data(), and we get right here.
1015                          * It may be too late to journal the data.  Simply
1016                          * falling through to the next test will suffice: the
1017                          * data will be dirty and wil be checkpointed.  The
1018                          * ordering comments in the next comment block still
1019                          * apply.
1020                          */
1021                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1022
1023                         /*
1024                          * If we're journalling data, and this buffer was
1025                          * subject to a write(), it could be metadata, forget
1026                          * or shadow against the committing transaction.  Now,
1027                          * someone has dirtied the same darn page via a mapping
1028                          * and it is being writepage()'d.
1029                          * We *could* just steal the page from commit, with some
1030                          * fancy locking there.  Instead, we just skip it -
1031                          * don't tie the page's buffers to the new transaction
1032                          * at all.
1033                          * Implication: if we crash before the writepage() data
1034                          * is written into the filesystem, recovery will replay
1035                          * the write() data.
1036                          */
1037                         if (jh->b_jlist != BJ_None &&
1038                                         jh->b_jlist != BJ_SyncData &&
1039                                         jh->b_jlist != BJ_Locked) {
1040                                 JBUFFER_TRACE(jh, "Not stealing");
1041                                 goto no_journal;
1042                         }
1043
1044                         /*
1045                          * This buffer may be undergoing writeout in commit.  We
1046                          * can't return from here and let the caller dirty it
1047                          * again because that can cause the write-out loop in
1048                          * commit to never terminate.
1049                          */
1050                         if (buffer_dirty(bh)) {
1051                                 get_bh(bh);
1052                                 spin_unlock(&journal->j_list_lock);
1053                                 jbd_unlock_bh_state(bh);
1054                                 need_brelse = 1;
1055                                 sync_dirty_buffer(bh);
1056                                 jbd_lock_bh_state(bh);
1057                                 spin_lock(&journal->j_list_lock);
1058                                 /* Since we dropped the lock... */
1059                                 if (!buffer_mapped(bh)) {
1060                                         JBUFFER_TRACE(jh, "buffer got unmapped");
1061                                         goto no_journal;
1062                                 }
1063                                 /* The buffer may become locked again at any
1064                                    time if it is redirtied */
1065                         }
1066
1067                         /*
1068                          * We cannot remove the buffer with io error from the
1069                          * committing transaction, because otherwise it would
1070                          * miss the error and the commit would not abort.
1071                          */
1072                         if (unlikely(!buffer_uptodate(bh))) {
1073                                 ret = -EIO;
1074                                 goto no_journal;
1075                         }
1076                         /* We might have slept so buffer could be refiled now */
1077                         if (jh->b_transaction != NULL &&
1078                             jh->b_transaction != handle->h_transaction) {
1079                                 JBUFFER_TRACE(jh, "unfile from commit");
1080                                 __journal_temp_unlink_buffer(jh);
1081                                 /* It still points to the committing
1082                                  * transaction; move it to this one so
1083                                  * that the refile assert checks are
1084                                  * happy. */
1085                                 jh->b_transaction = handle->h_transaction;
1086                         }
1087                         /* The buffer will be refiled below */
1088
1089                 }
1090                 /*
1091                  * Special case --- the buffer might actually have been
1092                  * allocated and then immediately deallocated in the previous,
1093                  * committing transaction, so might still be left on that
1094                  * transaction's metadata lists.
1095                  */
1096                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1097                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1098                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1099                         JBUFFER_TRACE(jh, "file as data");
1100                         __journal_file_buffer(jh, handle->h_transaction,
1101                                                 BJ_SyncData);
1102                 }
1103         } else {
1104                 JBUFFER_TRACE(jh, "not on a transaction");
1105                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1106         }
1107 no_journal:
1108         spin_unlock(&journal->j_list_lock);
1109         jbd_unlock_bh_state(bh);
1110         if (need_brelse) {
1111                 BUFFER_TRACE(bh, "brelse");
1112                 __brelse(bh);
1113         }
1114         JBUFFER_TRACE(jh, "exit");
1115         journal_put_journal_head(jh);
1116         return ret;
1117 }
1118
1119 /**
1120  * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1121  * @handle: transaction to add buffer to.
1122  * @bh: buffer to mark
1123  *
1124  * Mark dirty metadata which needs to be journaled as part of the current
1125  * transaction.
1126  *
1127  * The buffer is placed on the transaction's metadata list and is marked
1128  * as belonging to the transaction.
1129  *
1130  * Returns error number or 0 on success.
1131  *
1132  * Special care needs to be taken if the buffer already belongs to the
1133  * current committing transaction (in which case we should have frozen
1134  * data present for that commit).  In that case, we don't relink the
1135  * buffer: that only gets done when the old transaction finally
1136  * completes its commit.
1137  */
1138 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1139 {
1140         transaction_t *transaction = handle->h_transaction;
1141         journal_t *journal = transaction->t_journal;
1142         struct journal_head *jh = bh2jh(bh);
1143
1144         jbd_debug(5, "journal_head %p\n", jh);
1145         JBUFFER_TRACE(jh, "entry");
1146         if (is_handle_aborted(handle))
1147                 goto out;
1148
1149         jbd_lock_bh_state(bh);
1150
1151         if (jh->b_modified == 0) {
1152                 /*
1153                  * This buffer's got modified and becoming part
1154                  * of the transaction. This needs to be done
1155                  * once a transaction -bzzz
1156                  */
1157                 jh->b_modified = 1;
1158                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1159                 handle->h_buffer_credits--;
1160         }
1161
1162         /*
1163          * fastpath, to avoid expensive locking.  If this buffer is already
1164          * on the running transaction's metadata list there is nothing to do.
1165          * Nobody can take it off again because there is a handle open.
1166          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1167          * result in this test being false, so we go in and take the locks.
1168          */
1169         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1170                 JBUFFER_TRACE(jh, "fastpath");
1171                 J_ASSERT_JH(jh, jh->b_transaction ==
1172                                         journal->j_running_transaction);
1173                 goto out_unlock_bh;
1174         }
1175
1176         set_buffer_jbddirty(bh);
1177
1178         /*
1179          * Metadata already on the current transaction list doesn't
1180          * need to be filed.  Metadata on another transaction's list must
1181          * be committing, and will be refiled once the commit completes:
1182          * leave it alone for now.
1183          */
1184         if (jh->b_transaction != transaction) {
1185                 JBUFFER_TRACE(jh, "already on other transaction");
1186                 J_ASSERT_JH(jh, jh->b_transaction ==
1187                                         journal->j_committing_transaction);
1188                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1189                 /* And this case is illegal: we can't reuse another
1190                  * transaction's data buffer, ever. */
1191                 goto out_unlock_bh;
1192         }
1193
1194         /* That test should have eliminated the following case: */
1195         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1196
1197         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1198         spin_lock(&journal->j_list_lock);
1199         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1200         spin_unlock(&journal->j_list_lock);
1201 out_unlock_bh:
1202         jbd_unlock_bh_state(bh);
1203 out:
1204         JBUFFER_TRACE(jh, "exit");
1205         return 0;
1206 }
1207
1208 /*
1209  * journal_release_buffer: undo a get_write_access without any buffer
1210  * updates, if the update decided in the end that it didn't need access.
1211  *
1212  */
1213 void
1214 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1215 {
1216         BUFFER_TRACE(bh, "entry");
1217 }
1218
1219 /**
1220  * void journal_forget() - bforget() for potentially-journaled buffers.
1221  * @handle: transaction handle
1222  * @bh:     bh to 'forget'
1223  *
1224  * We can only do the bforget if there are no commits pending against the
1225  * buffer.  If the buffer is dirty in the current running transaction we
1226  * can safely unlink it.
1227  *
1228  * bh may not be a journalled buffer at all - it may be a non-JBD
1229  * buffer which came off the hashtable.  Check for this.
1230  *
1231  * Decrements bh->b_count by one.
1232  *
1233  * Allow this call even if the handle has aborted --- it may be part of
1234  * the caller's cleanup after an abort.
1235  */
1236 int journal_forget (handle_t *handle, struct buffer_head *bh)
1237 {
1238         transaction_t *transaction = handle->h_transaction;
1239         journal_t *journal = transaction->t_journal;
1240         struct journal_head *jh;
1241         int drop_reserve = 0;
1242         int err = 0;
1243         int was_modified = 0;
1244
1245         BUFFER_TRACE(bh, "entry");
1246
1247         jbd_lock_bh_state(bh);
1248         spin_lock(&journal->j_list_lock);
1249
1250         if (!buffer_jbd(bh))
1251                 goto not_jbd;
1252         jh = bh2jh(bh);
1253
1254         /* Critical error: attempting to delete a bitmap buffer, maybe?
1255          * Don't do any jbd operations, and return an error. */
1256         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1257                          "inconsistent data on disk")) {
1258                 err = -EIO;
1259                 goto not_jbd;
1260         }
1261
1262         /* keep track of whether or not this transaction modified us */
1263         was_modified = jh->b_modified;
1264
1265         /*
1266          * The buffer's going from the transaction, we must drop
1267          * all references -bzzz
1268          */
1269         jh->b_modified = 0;
1270
1271         if (jh->b_transaction == handle->h_transaction) {
1272                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1273
1274                 /* If we are forgetting a buffer which is already part
1275                  * of this transaction, then we can just drop it from
1276                  * the transaction immediately. */
1277                 clear_buffer_dirty(bh);
1278                 clear_buffer_jbddirty(bh);
1279
1280                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1281
1282                 /*
1283                  * we only want to drop a reference if this transaction
1284                  * modified the buffer
1285                  */
1286                 if (was_modified)
1287                         drop_reserve = 1;
1288
1289                 /*
1290                  * We are no longer going to journal this buffer.
1291                  * However, the commit of this transaction is still
1292                  * important to the buffer: the delete that we are now
1293                  * processing might obsolete an old log entry, so by
1294                  * committing, we can satisfy the buffer's checkpoint.
1295                  *
1296                  * So, if we have a checkpoint on the buffer, we should
1297                  * now refile the buffer on our BJ_Forget list so that
1298                  * we know to remove the checkpoint after we commit.
1299                  */
1300
1301                 if (jh->b_cp_transaction) {
1302                         __journal_temp_unlink_buffer(jh);
1303                         __journal_file_buffer(jh, transaction, BJ_Forget);
1304                 } else {
1305                         __journal_unfile_buffer(jh);
1306                         if (!buffer_jbd(bh)) {
1307                                 spin_unlock(&journal->j_list_lock);
1308                                 jbd_unlock_bh_state(bh);
1309                                 __bforget(bh);
1310                                 goto drop;
1311                         }
1312                 }
1313         } else if (jh->b_transaction) {
1314                 J_ASSERT_JH(jh, (jh->b_transaction ==
1315                                  journal->j_committing_transaction));
1316                 /* However, if the buffer is still owned by a prior
1317                  * (committing) transaction, we can't drop it yet... */
1318                 JBUFFER_TRACE(jh, "belongs to older transaction");
1319                 /* ... but we CAN drop it from the new transaction if we
1320                  * have also modified it since the original commit. */
1321
1322                 if (jh->b_next_transaction) {
1323                         J_ASSERT(jh->b_next_transaction == transaction);
1324                         jh->b_next_transaction = NULL;
1325
1326                         /*
1327                          * only drop a reference if this transaction modified
1328                          * the buffer
1329                          */
1330                         if (was_modified)
1331                                 drop_reserve = 1;
1332                 }
1333         }
1334
1335 not_jbd:
1336         spin_unlock(&journal->j_list_lock);
1337         jbd_unlock_bh_state(bh);
1338         __brelse(bh);
1339 drop:
1340         if (drop_reserve) {
1341                 /* no need to reserve log space for this block -bzzz */
1342                 handle->h_buffer_credits++;
1343         }
1344         return err;
1345 }
1346
1347 /**
1348  * int journal_stop() - complete a transaction
1349  * @handle: tranaction to complete.
1350  *
1351  * All done for a particular handle.
1352  *
1353  * There is not much action needed here.  We just return any remaining
1354  * buffer credits to the transaction and remove the handle.  The only
1355  * complication is that we need to start a commit operation if the
1356  * filesystem is marked for synchronous update.
1357  *
1358  * journal_stop itself will not usually return an error, but it may
1359  * do so in unusual circumstances.  In particular, expect it to
1360  * return -EIO if a journal_abort has been executed since the
1361  * transaction began.
1362  */
1363 int journal_stop(handle_t *handle)
1364 {
1365         transaction_t *transaction = handle->h_transaction;
1366         journal_t *journal = transaction->t_journal;
1367         int err;
1368         pid_t pid;
1369
1370         J_ASSERT(journal_current_handle() == handle);
1371
1372         if (is_handle_aborted(handle))
1373                 err = -EIO;
1374         else {
1375                 J_ASSERT(transaction->t_updates > 0);
1376                 err = 0;
1377         }
1378
1379         if (--handle->h_ref > 0) {
1380                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1381                           handle->h_ref);
1382                 return err;
1383         }
1384
1385         jbd_debug(4, "Handle %p going down\n", handle);
1386
1387         /*
1388          * Implement synchronous transaction batching.  If the handle
1389          * was synchronous, don't force a commit immediately.  Let's
1390          * yield and let another thread piggyback onto this transaction.
1391          * Keep doing that while new threads continue to arrive.
1392          * It doesn't cost much - we're about to run a commit and sleep
1393          * on IO anyway.  Speeds up many-threaded, many-dir operations
1394          * by 30x or more...
1395          *
1396          * We try and optimize the sleep time against what the underlying disk
1397          * can do, instead of having a static sleep time.  This is useful for
1398          * the case where our storage is so fast that it is more optimal to go
1399          * ahead and force a flush and wait for the transaction to be committed
1400          * than it is to wait for an arbitrary amount of time for new writers to
1401          * join the transaction.  We achieve this by measuring how long it takes
1402          * to commit a transaction, and compare it with how long this
1403          * transaction has been running, and if run time < commit time then we
1404          * sleep for the delta and commit.  This greatly helps super fast disks
1405          * that would see slowdowns as more threads started doing fsyncs.
1406          *
1407          * But don't do this if this process was the most recent one to
1408          * perform a synchronous write.  We do this to detect the case where a
1409          * single process is doing a stream of sync writes.  No point in waiting
1410          * for joiners in that case.
1411          */
1412         pid = current->pid;
1413         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1414                 u64 commit_time, trans_time;
1415
1416                 journal->j_last_sync_writer = pid;
1417
1418                 spin_lock(&journal->j_state_lock);
1419                 commit_time = journal->j_average_commit_time;
1420                 spin_unlock(&journal->j_state_lock);
1421
1422                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1423                                                    transaction->t_start_time));
1424
1425                 commit_time = min_t(u64, commit_time,
1426                                     1000*jiffies_to_usecs(1));
1427
1428                 if (trans_time < commit_time) {
1429                         ktime_t expires = ktime_add_ns(ktime_get(),
1430                                                        commit_time);
1431                         set_current_state(TASK_UNINTERRUPTIBLE);
1432                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1433                 }
1434         }
1435
1436         current->journal_info = NULL;
1437         spin_lock(&journal->j_state_lock);
1438         spin_lock(&transaction->t_handle_lock);
1439         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1440         transaction->t_updates--;
1441         if (!transaction->t_updates) {
1442                 wake_up(&journal->j_wait_updates);
1443                 if (journal->j_barrier_count)
1444                         wake_up(&journal->j_wait_transaction_locked);
1445         }
1446
1447         /*
1448          * If the handle is marked SYNC, we need to set another commit
1449          * going!  We also want to force a commit if the current
1450          * transaction is occupying too much of the log, or if the
1451          * transaction is too old now.
1452          */
1453         if (handle->h_sync ||
1454                         transaction->t_outstanding_credits >
1455                                 journal->j_max_transaction_buffers ||
1456                         time_after_eq(jiffies, transaction->t_expires)) {
1457                 /* Do this even for aborted journals: an abort still
1458                  * completes the commit thread, it just doesn't write
1459                  * anything to disk. */
1460                 tid_t tid = transaction->t_tid;
1461
1462                 spin_unlock(&transaction->t_handle_lock);
1463                 jbd_debug(2, "transaction too old, requesting commit for "
1464                                         "handle %p\n", handle);
1465                 /* This is non-blocking */
1466                 __log_start_commit(journal, transaction->t_tid);
1467                 spin_unlock(&journal->j_state_lock);
1468
1469                 /*
1470                  * Special case: JFS_SYNC synchronous updates require us
1471                  * to wait for the commit to complete.
1472                  */
1473                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1474                         err = log_wait_commit(journal, tid);
1475         } else {
1476                 spin_unlock(&transaction->t_handle_lock);
1477                 spin_unlock(&journal->j_state_lock);
1478         }
1479
1480         lock_map_release(&handle->h_lockdep_map);
1481
1482         jbd_free_handle(handle);
1483         return err;
1484 }
1485
1486 /**
1487  * int journal_force_commit() - force any uncommitted transactions
1488  * @journal: journal to force
1489  *
1490  * For synchronous operations: force any uncommitted transactions
1491  * to disk.  May seem kludgy, but it reuses all the handle batching
1492  * code in a very simple manner.
1493  */
1494 int journal_force_commit(journal_t *journal)
1495 {
1496         handle_t *handle;
1497         int ret;
1498
1499         handle = journal_start(journal, 1);
1500         if (IS_ERR(handle)) {
1501                 ret = PTR_ERR(handle);
1502         } else {
1503                 handle->h_sync = 1;
1504                 ret = journal_stop(handle);
1505         }
1506         return ret;
1507 }
1508
1509 /*
1510  *
1511  * List management code snippets: various functions for manipulating the
1512  * transaction buffer lists.
1513  *
1514  */
1515
1516 /*
1517  * Append a buffer to a transaction list, given the transaction's list head
1518  * pointer.
1519  *
1520  * j_list_lock is held.
1521  *
1522  * jbd_lock_bh_state(jh2bh(jh)) is held.
1523  */
1524
1525 static inline void
1526 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1527 {
1528         if (!*list) {
1529                 jh->b_tnext = jh->b_tprev = jh;
1530                 *list = jh;
1531         } else {
1532                 /* Insert at the tail of the list to preserve order */
1533                 struct journal_head *first = *list, *last = first->b_tprev;
1534                 jh->b_tprev = last;
1535                 jh->b_tnext = first;
1536                 last->b_tnext = first->b_tprev = jh;
1537         }
1538 }
1539
1540 /*
1541  * Remove a buffer from a transaction list, given the transaction's list
1542  * head pointer.
1543  *
1544  * Called with j_list_lock held, and the journal may not be locked.
1545  *
1546  * jbd_lock_bh_state(jh2bh(jh)) is held.
1547  */
1548
1549 static inline void
1550 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1551 {
1552         if (*list == jh) {
1553                 *list = jh->b_tnext;
1554                 if (*list == jh)
1555                         *list = NULL;
1556         }
1557         jh->b_tprev->b_tnext = jh->b_tnext;
1558         jh->b_tnext->b_tprev = jh->b_tprev;
1559 }
1560
1561 /*
1562  * Remove a buffer from the appropriate transaction list.
1563  *
1564  * Note that this function can *change* the value of
1565  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1566  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1567  * is holding onto a copy of one of thee pointers, it could go bad.
1568  * Generally the caller needs to re-read the pointer from the transaction_t.
1569  *
1570  * Called under j_list_lock.  The journal may not be locked.
1571  */
1572 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1573 {
1574         struct journal_head **list = NULL;
1575         transaction_t *transaction;
1576         struct buffer_head *bh = jh2bh(jh);
1577
1578         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1579         transaction = jh->b_transaction;
1580         if (transaction)
1581                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1582
1583         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1584         if (jh->b_jlist != BJ_None)
1585                 J_ASSERT_JH(jh, transaction != NULL);
1586
1587         switch (jh->b_jlist) {
1588         case BJ_None:
1589                 return;
1590         case BJ_SyncData:
1591                 list = &transaction->t_sync_datalist;
1592                 break;
1593         case BJ_Metadata:
1594                 transaction->t_nr_buffers--;
1595                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1596                 list = &transaction->t_buffers;
1597                 break;
1598         case BJ_Forget:
1599                 list = &transaction->t_forget;
1600                 break;
1601         case BJ_IO:
1602                 list = &transaction->t_iobuf_list;
1603                 break;
1604         case BJ_Shadow:
1605                 list = &transaction->t_shadow_list;
1606                 break;
1607         case BJ_LogCtl:
1608                 list = &transaction->t_log_list;
1609                 break;
1610         case BJ_Reserved:
1611                 list = &transaction->t_reserved_list;
1612                 break;
1613         case BJ_Locked:
1614                 list = &transaction->t_locked_list;
1615                 break;
1616         }
1617
1618         __blist_del_buffer(list, jh);
1619         jh->b_jlist = BJ_None;
1620         if (test_clear_buffer_jbddirty(bh))
1621                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1622 }
1623
1624 /*
1625  * Remove buffer from all transactions.
1626  *
1627  * Called with bh_state lock and j_list_lock
1628  *
1629  * jh and bh may be already freed when this function returns.
1630  */
1631 void __journal_unfile_buffer(struct journal_head *jh)
1632 {
1633         __journal_temp_unlink_buffer(jh);
1634         jh->b_transaction = NULL;
1635         journal_put_journal_head(jh);
1636 }
1637
1638 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1639 {
1640         struct buffer_head *bh = jh2bh(jh);
1641
1642         /* Get reference so that buffer cannot be freed before we unlock it */
1643         get_bh(bh);
1644         jbd_lock_bh_state(bh);
1645         spin_lock(&journal->j_list_lock);
1646         __journal_unfile_buffer(jh);
1647         spin_unlock(&journal->j_list_lock);
1648         jbd_unlock_bh_state(bh);
1649         __brelse(bh);
1650 }
1651
1652 /*
1653  * Called from journal_try_to_free_buffers().
1654  *
1655  * Called under jbd_lock_bh_state(bh)
1656  */
1657 static void
1658 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1659 {
1660         struct journal_head *jh;
1661
1662         jh = bh2jh(bh);
1663
1664         if (buffer_locked(bh) || buffer_dirty(bh))
1665                 goto out;
1666
1667         if (jh->b_next_transaction != NULL)
1668                 goto out;
1669
1670         spin_lock(&journal->j_list_lock);
1671         if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1672                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1673                         /* A written-back ordered data buffer */
1674                         JBUFFER_TRACE(jh, "release data");
1675                         __journal_unfile_buffer(jh);
1676                 }
1677         } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1678                 /* written-back checkpointed metadata buffer */
1679                 if (jh->b_jlist == BJ_None) {
1680                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1681                         __journal_remove_checkpoint(jh);
1682                 }
1683         }
1684         spin_unlock(&journal->j_list_lock);
1685 out:
1686         return;
1687 }
1688
1689 /**
1690  * int journal_try_to_free_buffers() - try to free page buffers.
1691  * @journal: journal for operation
1692  * @page: to try and free
1693  * @gfp_mask: we use the mask to detect how hard should we try to release
1694  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1695  * release the buffers.
1696  *
1697  *
1698  * For all the buffers on this page,
1699  * if they are fully written out ordered data, move them onto BUF_CLEAN
1700  * so try_to_free_buffers() can reap them.
1701  *
1702  * This function returns non-zero if we wish try_to_free_buffers()
1703  * to be called. We do this if the page is releasable by try_to_free_buffers().
1704  * We also do it if the page has locked or dirty buffers and the caller wants
1705  * us to perform sync or async writeout.
1706  *
1707  * This complicates JBD locking somewhat.  We aren't protected by the
1708  * BKL here.  We wish to remove the buffer from its committing or
1709  * running transaction's ->t_datalist via __journal_unfile_buffer.
1710  *
1711  * This may *change* the value of transaction_t->t_datalist, so anyone
1712  * who looks at t_datalist needs to lock against this function.
1713  *
1714  * Even worse, someone may be doing a journal_dirty_data on this
1715  * buffer.  So we need to lock against that.  journal_dirty_data()
1716  * will come out of the lock with the buffer dirty, which makes it
1717  * ineligible for release here.
1718  *
1719  * Who else is affected by this?  hmm...  Really the only contender
1720  * is do_get_write_access() - it could be looking at the buffer while
1721  * journal_try_to_free_buffer() is changing its state.  But that
1722  * cannot happen because we never reallocate freed data as metadata
1723  * while the data is part of a transaction.  Yes?
1724  *
1725  * Return 0 on failure, 1 on success
1726  */
1727 int journal_try_to_free_buffers(journal_t *journal,
1728                                 struct page *page, gfp_t gfp_mask)
1729 {
1730         struct buffer_head *head;
1731         struct buffer_head *bh;
1732         int ret = 0;
1733
1734         J_ASSERT(PageLocked(page));
1735
1736         head = page_buffers(page);
1737         bh = head;
1738         do {
1739                 struct journal_head *jh;
1740
1741                 /*
1742                  * We take our own ref against the journal_head here to avoid
1743                  * having to add tons of locking around each instance of
1744                  * journal_put_journal_head().
1745                  */
1746                 jh = journal_grab_journal_head(bh);
1747                 if (!jh)
1748                         continue;
1749
1750                 jbd_lock_bh_state(bh);
1751                 __journal_try_to_free_buffer(journal, bh);
1752                 journal_put_journal_head(jh);
1753                 jbd_unlock_bh_state(bh);
1754                 if (buffer_jbd(bh))
1755                         goto busy;
1756         } while ((bh = bh->b_this_page) != head);
1757
1758         ret = try_to_free_buffers(page);
1759
1760 busy:
1761         return ret;
1762 }
1763
1764 /*
1765  * This buffer is no longer needed.  If it is on an older transaction's
1766  * checkpoint list we need to record it on this transaction's forget list
1767  * to pin this buffer (and hence its checkpointing transaction) down until
1768  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1769  * release it.
1770  * Returns non-zero if JBD no longer has an interest in the buffer.
1771  *
1772  * Called under j_list_lock.
1773  *
1774  * Called under jbd_lock_bh_state(bh).
1775  */
1776 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1777 {
1778         int may_free = 1;
1779         struct buffer_head *bh = jh2bh(jh);
1780
1781         if (jh->b_cp_transaction) {
1782                 JBUFFER_TRACE(jh, "on running+cp transaction");
1783                 __journal_temp_unlink_buffer(jh);
1784                 /*
1785                  * We don't want to write the buffer anymore, clear the
1786                  * bit so that we don't confuse checks in
1787                  * __journal_file_buffer
1788                  */
1789                 clear_buffer_dirty(bh);
1790                 __journal_file_buffer(jh, transaction, BJ_Forget);
1791                 may_free = 0;
1792         } else {
1793                 JBUFFER_TRACE(jh, "on running transaction");
1794                 __journal_unfile_buffer(jh);
1795         }
1796         return may_free;
1797 }
1798
1799 /*
1800  * journal_invalidatepage
1801  *
1802  * This code is tricky.  It has a number of cases to deal with.
1803  *
1804  * There are two invariants which this code relies on:
1805  *
1806  * i_size must be updated on disk before we start calling invalidatepage on the
1807  * data.
1808  *
1809  *  This is done in ext3 by defining an ext3_setattr method which
1810  *  updates i_size before truncate gets going.  By maintaining this
1811  *  invariant, we can be sure that it is safe to throw away any buffers
1812  *  attached to the current transaction: once the transaction commits,
1813  *  we know that the data will not be needed.
1814  *
1815  *  Note however that we can *not* throw away data belonging to the
1816  *  previous, committing transaction!
1817  *
1818  * Any disk blocks which *are* part of the previous, committing
1819  * transaction (and which therefore cannot be discarded immediately) are
1820  * not going to be reused in the new running transaction
1821  *
1822  *  The bitmap committed_data images guarantee this: any block which is
1823  *  allocated in one transaction and removed in the next will be marked
1824  *  as in-use in the committed_data bitmap, so cannot be reused until
1825  *  the next transaction to delete the block commits.  This means that
1826  *  leaving committing buffers dirty is quite safe: the disk blocks
1827  *  cannot be reallocated to a different file and so buffer aliasing is
1828  *  not possible.
1829  *
1830  *
1831  * The above applies mainly to ordered data mode.  In writeback mode we
1832  * don't make guarantees about the order in which data hits disk --- in
1833  * particular we don't guarantee that new dirty data is flushed before
1834  * transaction commit --- so it is always safe just to discard data
1835  * immediately in that mode.  --sct
1836  */
1837
1838 /*
1839  * The journal_unmap_buffer helper function returns zero if the buffer
1840  * concerned remains pinned as an anonymous buffer belonging to an older
1841  * transaction.
1842  *
1843  * We're outside-transaction here.  Either or both of j_running_transaction
1844  * and j_committing_transaction may be NULL.
1845  */
1846 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1847                                 int partial_page)
1848 {
1849         transaction_t *transaction;
1850         struct journal_head *jh;
1851         int may_free = 1;
1852
1853         BUFFER_TRACE(bh, "entry");
1854
1855 retry:
1856         /*
1857          * It is safe to proceed here without the j_list_lock because the
1858          * buffers cannot be stolen by try_to_free_buffers as long as we are
1859          * holding the page lock. --sct
1860          */
1861
1862         if (!buffer_jbd(bh))
1863                 goto zap_buffer_unlocked;
1864
1865         spin_lock(&journal->j_state_lock);
1866         jbd_lock_bh_state(bh);
1867         spin_lock(&journal->j_list_lock);
1868
1869         jh = journal_grab_journal_head(bh);
1870         if (!jh)
1871                 goto zap_buffer_no_jh;
1872
1873         /*
1874          * We cannot remove the buffer from checkpoint lists until the
1875          * transaction adding inode to orphan list (let's call it T)
1876          * is committed.  Otherwise if the transaction changing the
1877          * buffer would be cleaned from the journal before T is
1878          * committed, a crash will cause that the correct contents of
1879          * the buffer will be lost.  On the other hand we have to
1880          * clear the buffer dirty bit at latest at the moment when the
1881          * transaction marking the buffer as freed in the filesystem
1882          * structures is committed because from that moment on the
1883          * block can be reallocated and used by a different page.
1884          * Since the block hasn't been freed yet but the inode has
1885          * already been added to orphan list, it is safe for us to add
1886          * the buffer to BJ_Forget list of the newest transaction.
1887          *
1888          * Also we have to clear buffer_mapped flag of a truncated buffer
1889          * because the buffer_head may be attached to the page straddling
1890          * i_size (can happen only when blocksize < pagesize) and thus the
1891          * buffer_head can be reused when the file is extended again. So we end
1892          * up keeping around invalidated buffers attached to transactions'
1893          * BJ_Forget list just to stop checkpointing code from cleaning up
1894          * the transaction this buffer was modified in.
1895          */
1896         transaction = jh->b_transaction;
1897         if (transaction == NULL) {
1898                 /* First case: not on any transaction.  If it
1899                  * has no checkpoint link, then we can zap it:
1900                  * it's a writeback-mode buffer so we don't care
1901                  * if it hits disk safely. */
1902                 if (!jh->b_cp_transaction) {
1903                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1904                         goto zap_buffer;
1905                 }
1906
1907                 if (!buffer_dirty(bh)) {
1908                         /* bdflush has written it.  We can drop it now */
1909                         goto zap_buffer;
1910                 }
1911
1912                 /* OK, it must be in the journal but still not
1913                  * written fully to disk: it's metadata or
1914                  * journaled data... */
1915
1916                 if (journal->j_running_transaction) {
1917                         /* ... and once the current transaction has
1918                          * committed, the buffer won't be needed any
1919                          * longer. */
1920                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1921                         may_free = __dispose_buffer(jh,
1922                                         journal->j_running_transaction);
1923                         goto zap_buffer;
1924                 } else {
1925                         /* There is no currently-running transaction. So the
1926                          * orphan record which we wrote for this file must have
1927                          * passed into commit.  We must attach this buffer to
1928                          * the committing transaction, if it exists. */
1929                         if (journal->j_committing_transaction) {
1930                                 JBUFFER_TRACE(jh, "give to committing trans");
1931                                 may_free = __dispose_buffer(jh,
1932                                         journal->j_committing_transaction);
1933                                 goto zap_buffer;
1934                         } else {
1935                                 /* The orphan record's transaction has
1936                                  * committed.  We can cleanse this buffer */
1937                                 clear_buffer_jbddirty(bh);
1938                                 goto zap_buffer;
1939                         }
1940                 }
1941         } else if (transaction == journal->j_committing_transaction) {
1942                 JBUFFER_TRACE(jh, "on committing transaction");
1943                 if (jh->b_jlist == BJ_Locked) {
1944                         /*
1945                          * The buffer is on the committing transaction's locked
1946                          * list.  We have the buffer locked, so I/O has
1947                          * completed.  So we can nail the buffer now.
1948                          */
1949                         may_free = __dispose_buffer(jh, transaction);
1950                         goto zap_buffer;
1951                 }
1952                 /*
1953                  * The buffer is committing, we simply cannot touch
1954                  * it. If the page is straddling i_size we have to wait
1955                  * for commit and try again.
1956                  */
1957                 if (partial_page) {
1958                         tid_t tid = journal->j_committing_transaction->t_tid;
1959
1960                         journal_put_journal_head(jh);
1961                         spin_unlock(&journal->j_list_lock);
1962                         jbd_unlock_bh_state(bh);
1963                         spin_unlock(&journal->j_state_lock);
1964                         log_wait_commit(journal, tid);
1965                         goto retry;
1966                 }
1967                 /*
1968                  * OK, buffer won't be reachable after truncate. We just set
1969                  * j_next_transaction to the running transaction (if there is
1970                  * one) and mark buffer as freed so that commit code knows it
1971                  * should clear dirty bits when it is done with the buffer.
1972                  */
1973                 set_buffer_freed(bh);
1974                 if (journal->j_running_transaction && buffer_jbddirty(bh))
1975                         jh->b_next_transaction = journal->j_running_transaction;
1976                 journal_put_journal_head(jh);
1977                 spin_unlock(&journal->j_list_lock);
1978                 jbd_unlock_bh_state(bh);
1979                 spin_unlock(&journal->j_state_lock);
1980                 return 0;
1981         } else {
1982                 /* Good, the buffer belongs to the running transaction.
1983                  * We are writing our own transaction's data, not any
1984                  * previous one's, so it is safe to throw it away
1985                  * (remember that we expect the filesystem to have set
1986                  * i_size already for this truncate so recovery will not
1987                  * expose the disk blocks we are discarding here.) */
1988                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1989                 JBUFFER_TRACE(jh, "on running transaction");
1990                 may_free = __dispose_buffer(jh, transaction);
1991         }
1992
1993 zap_buffer:
1994         /*
1995          * This is tricky. Although the buffer is truncated, it may be reused
1996          * if blocksize < pagesize and it is attached to the page straddling
1997          * EOF. Since the buffer might have been added to BJ_Forget list of the
1998          * running transaction, journal_get_write_access() won't clear
1999          * b_modified and credit accounting gets confused. So clear b_modified
2000          * here. */
2001         jh->b_modified = 0;
2002         journal_put_journal_head(jh);
2003 zap_buffer_no_jh:
2004         spin_unlock(&journal->j_list_lock);
2005         jbd_unlock_bh_state(bh);
2006         spin_unlock(&journal->j_state_lock);
2007 zap_buffer_unlocked:
2008         clear_buffer_dirty(bh);
2009         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2010         clear_buffer_mapped(bh);
2011         clear_buffer_req(bh);
2012         clear_buffer_new(bh);
2013         bh->b_bdev = NULL;
2014         return may_free;
2015 }
2016
2017 /**
2018  * void journal_invalidatepage() - invalidate a journal page
2019  * @journal: journal to use for flush
2020  * @page:    page to flush
2021  * @offset:  length of page to invalidate.
2022  *
2023  * Reap page buffers containing data after offset in page.
2024  */
2025 void journal_invalidatepage(journal_t *journal,
2026                       struct page *page,
2027                       unsigned long offset)
2028 {
2029         struct buffer_head *head, *bh, *next;
2030         unsigned int curr_off = 0;
2031         int may_free = 1;
2032
2033         if (!PageLocked(page))
2034                 BUG();
2035         if (!page_has_buffers(page))
2036                 return;
2037
2038         /* We will potentially be playing with lists other than just the
2039          * data lists (especially for journaled data mode), so be
2040          * cautious in our locking. */
2041
2042         head = bh = page_buffers(page);
2043         do {
2044                 unsigned int next_off = curr_off + bh->b_size;
2045                 next = bh->b_this_page;
2046
2047                 if (offset <= curr_off) {
2048                         /* This block is wholly outside the truncation point */
2049                         lock_buffer(bh);
2050                         may_free &= journal_unmap_buffer(journal, bh,
2051                                                          offset > 0);
2052                         unlock_buffer(bh);
2053                 }
2054                 curr_off = next_off;
2055                 bh = next;
2056
2057         } while (bh != head);
2058
2059         if (!offset) {
2060                 if (may_free && try_to_free_buffers(page))
2061                         J_ASSERT(!page_has_buffers(page));
2062         }
2063 }
2064
2065 /*
2066  * File a buffer on the given transaction list.
2067  */
2068 void __journal_file_buffer(struct journal_head *jh,
2069                         transaction_t *transaction, int jlist)
2070 {
2071         struct journal_head **list = NULL;
2072         int was_dirty = 0;
2073         struct buffer_head *bh = jh2bh(jh);
2074
2075         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2076         assert_spin_locked(&transaction->t_journal->j_list_lock);
2077
2078         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2079         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2080                                 jh->b_transaction == NULL);
2081
2082         if (jh->b_transaction && jh->b_jlist == jlist)
2083                 return;
2084
2085         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2086             jlist == BJ_Shadow || jlist == BJ_Forget) {
2087                 /*
2088                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2089                  * instead of buffer_dirty. We should not see a dirty bit set
2090                  * here because we clear it in do_get_write_access but e.g.
2091                  * tune2fs can modify the sb and set the dirty bit at any time
2092                  * so we try to gracefully handle that.
2093                  */
2094                 if (buffer_dirty(bh))
2095                         warn_dirty_buffer(bh);
2096                 if (test_clear_buffer_dirty(bh) ||
2097                     test_clear_buffer_jbddirty(bh))
2098                         was_dirty = 1;
2099         }
2100
2101         if (jh->b_transaction)
2102                 __journal_temp_unlink_buffer(jh);
2103         else
2104                 journal_grab_journal_head(bh);
2105         jh->b_transaction = transaction;
2106
2107         switch (jlist) {
2108         case BJ_None:
2109                 J_ASSERT_JH(jh, !jh->b_committed_data);
2110                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2111                 return;
2112         case BJ_SyncData:
2113                 list = &transaction->t_sync_datalist;
2114                 break;
2115         case BJ_Metadata:
2116                 transaction->t_nr_buffers++;
2117                 list = &transaction->t_buffers;
2118                 break;
2119         case BJ_Forget:
2120                 list = &transaction->t_forget;
2121                 break;
2122         case BJ_IO:
2123                 list = &transaction->t_iobuf_list;
2124                 break;
2125         case BJ_Shadow:
2126                 list = &transaction->t_shadow_list;
2127                 break;
2128         case BJ_LogCtl:
2129                 list = &transaction->t_log_list;
2130                 break;
2131         case BJ_Reserved:
2132                 list = &transaction->t_reserved_list;
2133                 break;
2134         case BJ_Locked:
2135                 list =  &transaction->t_locked_list;
2136                 break;
2137         }
2138
2139         __blist_add_buffer(list, jh);
2140         jh->b_jlist = jlist;
2141
2142         if (was_dirty)
2143                 set_buffer_jbddirty(bh);
2144 }
2145
2146 void journal_file_buffer(struct journal_head *jh,
2147                                 transaction_t *transaction, int jlist)
2148 {
2149         jbd_lock_bh_state(jh2bh(jh));
2150         spin_lock(&transaction->t_journal->j_list_lock);
2151         __journal_file_buffer(jh, transaction, jlist);
2152         spin_unlock(&transaction->t_journal->j_list_lock);
2153         jbd_unlock_bh_state(jh2bh(jh));
2154 }
2155
2156 /*
2157  * Remove a buffer from its current buffer list in preparation for
2158  * dropping it from its current transaction entirely.  If the buffer has
2159  * already started to be used by a subsequent transaction, refile the
2160  * buffer on that transaction's metadata list.
2161  *
2162  * Called under j_list_lock
2163  * Called under jbd_lock_bh_state(jh2bh(jh))
2164  *
2165  * jh and bh may be already free when this function returns
2166  */
2167 void __journal_refile_buffer(struct journal_head *jh)
2168 {
2169         int was_dirty, jlist;
2170         struct buffer_head *bh = jh2bh(jh);
2171
2172         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2173         if (jh->b_transaction)
2174                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2175
2176         /* If the buffer is now unused, just drop it. */
2177         if (jh->b_next_transaction == NULL) {
2178                 __journal_unfile_buffer(jh);
2179                 return;
2180         }
2181
2182         /*
2183          * It has been modified by a later transaction: add it to the new
2184          * transaction's metadata list.
2185          */
2186
2187         was_dirty = test_clear_buffer_jbddirty(bh);
2188         __journal_temp_unlink_buffer(jh);
2189         /*
2190          * We set b_transaction here because b_next_transaction will inherit
2191          * our jh reference and thus __journal_file_buffer() must not take a
2192          * new one.
2193          */
2194         jh->b_transaction = jh->b_next_transaction;
2195         jh->b_next_transaction = NULL;
2196         if (buffer_freed(bh))
2197                 jlist = BJ_Forget;
2198         else if (jh->b_modified)
2199                 jlist = BJ_Metadata;
2200         else
2201                 jlist = BJ_Reserved;
2202         __journal_file_buffer(jh, jh->b_transaction, jlist);
2203         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2204
2205         if (was_dirty)
2206                 set_buffer_jbddirty(bh);
2207 }
2208
2209 /*
2210  * __journal_refile_buffer() with necessary locking added. We take our bh
2211  * reference so that we can safely unlock bh.
2212  *
2213  * The jh and bh may be freed by this call.
2214  */
2215 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2216 {
2217         struct buffer_head *bh = jh2bh(jh);
2218
2219         /* Get reference so that buffer cannot be freed before we unlock it */
2220         get_bh(bh);
2221         jbd_lock_bh_state(bh);
2222         spin_lock(&journal->j_list_lock);
2223         __journal_refile_buffer(jh);
2224         jbd_unlock_bh_state(bh);
2225         spin_unlock(&journal->j_list_lock);
2226         __brelse(bh);
2227 }