Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6
[linux-2.6] / 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
29 static void __journal_temp_unlink_buffer(struct journal_head *jh);
30
31 /*
32  * get_transaction: obtain a new transaction_t object.
33  *
34  * Simply allocate and initialise a new transaction.  Create it in
35  * RUNNING state and add it to the current journal (which should not
36  * have an existing running transaction: we only make a new transaction
37  * once we have started to commit the old one).
38  *
39  * Preconditions:
40  *      The journal MUST be locked.  We don't perform atomic mallocs on the
41  *      new transaction and we can't block without protecting against other
42  *      processes trying to touch the journal while it is in transition.
43  *
44  * Called under j_state_lock
45  */
46
47 static transaction_t *
48 get_transaction(journal_t *journal, transaction_t *transaction)
49 {
50         transaction->t_journal = journal;
51         transaction->t_state = T_RUNNING;
52         transaction->t_tid = journal->j_transaction_sequence++;
53         transaction->t_expires = jiffies + journal->j_commit_interval;
54         spin_lock_init(&transaction->t_handle_lock);
55
56         /* Set up the commit timer for the new transaction. */
57         journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
58         add_timer(&journal->j_commit_timer);
59
60         J_ASSERT(journal->j_running_transaction == NULL);
61         journal->j_running_transaction = transaction;
62
63         return transaction;
64 }
65
66 /*
67  * Handle management.
68  *
69  * A handle_t is an object which represents a single atomic update to a
70  * filesystem, and which tracks all of the modifications which form part
71  * of that one update.
72  */
73
74 /*
75  * start_this_handle: Given a handle, deal with any locking or stalling
76  * needed to make sure that there is enough journal space for the handle
77  * to begin.  Attach the handle to a transaction and set up the
78  * transaction's buffer credits.
79  */
80
81 static int start_this_handle(journal_t *journal, handle_t *handle)
82 {
83         transaction_t *transaction;
84         int needed;
85         int nblocks = handle->h_buffer_credits;
86         transaction_t *new_transaction = NULL;
87         int ret = 0;
88
89         if (nblocks > journal->j_max_transaction_buffers) {
90                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
91                        current->comm, nblocks,
92                        journal->j_max_transaction_buffers);
93                 ret = -ENOSPC;
94                 goto out;
95         }
96
97 alloc_transaction:
98         if (!journal->j_running_transaction) {
99                 new_transaction = kzalloc(sizeof(*new_transaction),
100                                                 GFP_NOFS|__GFP_NOFAIL);
101                 if (!new_transaction) {
102                         ret = -ENOMEM;
103                         goto out;
104                 }
105         }
106
107         jbd_debug(3, "New handle %p going live.\n", handle);
108
109 repeat:
110
111         /*
112          * We need to hold j_state_lock until t_updates has been incremented,
113          * for proper journal barrier handling
114          */
115         spin_lock(&journal->j_state_lock);
116 repeat_locked:
117         if (is_journal_aborted(journal) ||
118             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
119                 spin_unlock(&journal->j_state_lock);
120                 ret = -EROFS;
121                 goto out;
122         }
123
124         /* Wait on the journal's transaction barrier if necessary */
125         if (journal->j_barrier_count) {
126                 spin_unlock(&journal->j_state_lock);
127                 wait_event(journal->j_wait_transaction_locked,
128                                 journal->j_barrier_count == 0);
129                 goto repeat;
130         }
131
132         if (!journal->j_running_transaction) {
133                 if (!new_transaction) {
134                         spin_unlock(&journal->j_state_lock);
135                         goto alloc_transaction;
136                 }
137                 get_transaction(journal, new_transaction);
138                 new_transaction = NULL;
139         }
140
141         transaction = journal->j_running_transaction;
142
143         /*
144          * If the current transaction is locked down for commit, wait for the
145          * lock to be released.
146          */
147         if (transaction->t_state == T_LOCKED) {
148                 DEFINE_WAIT(wait);
149
150                 prepare_to_wait(&journal->j_wait_transaction_locked,
151                                         &wait, TASK_UNINTERRUPTIBLE);
152                 spin_unlock(&journal->j_state_lock);
153                 schedule();
154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
155                 goto repeat;
156         }
157
158         /*
159          * If there is not enough space left in the log to write all potential
160          * buffers requested by this operation, we need to stall pending a log
161          * checkpoint to free some more log space.
162          */
163         spin_lock(&transaction->t_handle_lock);
164         needed = transaction->t_outstanding_credits + nblocks;
165
166         if (needed > journal->j_max_transaction_buffers) {
167                 /*
168                  * If the current transaction is already too large, then start
169                  * to commit it: we can then go back and attach this handle to
170                  * a new transaction.
171                  */
172                 DEFINE_WAIT(wait);
173
174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
175                 spin_unlock(&transaction->t_handle_lock);
176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177                                 TASK_UNINTERRUPTIBLE);
178                 __log_start_commit(journal, transaction->t_tid);
179                 spin_unlock(&journal->j_state_lock);
180                 schedule();
181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
182                 goto repeat;
183         }
184
185         /*
186          * The commit code assumes that it can get enough log space
187          * without forcing a checkpoint.  This is *critical* for
188          * correctness: a checkpoint of a buffer which is also
189          * associated with a committing transaction creates a deadlock,
190          * so commit simply cannot force through checkpoints.
191          *
192          * We must therefore ensure the necessary space in the journal
193          * *before* starting to dirty potentially checkpointed buffers
194          * in the new transaction.
195          *
196          * The worst part is, any transaction currently committing can
197          * reduce the free space arbitrarily.  Be careful to account for
198          * those buffers when checkpointing.
199          */
200
201         /*
202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
203          * a _lot_ of headroom: 1/4 of the journal plus the size of
204          * the committing transaction.  Really, we only need to give it
205          * committing_transaction->t_outstanding_credits plus "enough" for
206          * the log control blocks.
207          * Also, this test is inconsitent with the matching one in
208          * journal_extend().
209          */
210         if (__log_space_left(journal) < jbd_space_needed(journal)) {
211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
212                 spin_unlock(&transaction->t_handle_lock);
213                 __log_wait_for_space(journal);
214                 goto repeat_locked;
215         }
216
217         /* OK, account for the buffers that this operation expects to
218          * use and add the handle to the running transaction. */
219
220         handle->h_transaction = transaction;
221         transaction->t_outstanding_credits += nblocks;
222         transaction->t_updates++;
223         transaction->t_handle_count++;
224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
225                   handle, nblocks, transaction->t_outstanding_credits,
226                   __log_space_left(journal));
227         spin_unlock(&transaction->t_handle_lock);
228         spin_unlock(&journal->j_state_lock);
229 out:
230         if (unlikely(new_transaction))          /* It's usually NULL */
231                 kfree(new_transaction);
232         return ret;
233 }
234
235 static struct lock_class_key jbd_handle_key;
236
237 /* Allocate a new handle.  This should probably be in a slab... */
238 static handle_t *new_handle(int nblocks)
239 {
240         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
241         if (!handle)
242                 return NULL;
243         memset(handle, 0, sizeof(*handle));
244         handle->h_buffer_credits = nblocks;
245         handle->h_ref = 1;
246
247         lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
248
249         return handle;
250 }
251
252 /**
253  * handle_t *journal_start() - Obtain a new handle.
254  * @journal: Journal to start transaction on.
255  * @nblocks: number of block buffer we might modify
256  *
257  * We make sure that the transaction can guarantee at least nblocks of
258  * modified buffers in the log.  We block until the log can guarantee
259  * that much space.
260  *
261  * This function is visible to journal users (like ext3fs), so is not
262  * called with the journal already locked.
263  *
264  * Return a pointer to a newly allocated handle, or NULL on failure
265  */
266 handle_t *journal_start(journal_t *journal, int nblocks)
267 {
268         handle_t *handle = journal_current_handle();
269         int err;
270
271         if (!journal)
272                 return ERR_PTR(-EROFS);
273
274         if (handle) {
275                 J_ASSERT(handle->h_transaction->t_journal == journal);
276                 handle->h_ref++;
277                 return handle;
278         }
279
280         handle = new_handle(nblocks);
281         if (!handle)
282                 return ERR_PTR(-ENOMEM);
283
284         current->journal_info = handle;
285
286         err = start_this_handle(journal, handle);
287         if (err < 0) {
288                 jbd_free_handle(handle);
289                 current->journal_info = NULL;
290                 handle = ERR_PTR(err);
291                 goto out;
292         }
293
294         lock_map_acquire(&handle->h_lockdep_map);
295
296 out:
297         return handle;
298 }
299
300 /**
301  * int journal_extend() - extend buffer credits.
302  * @handle:  handle to 'extend'
303  * @nblocks: nr blocks to try to extend by.
304  *
305  * Some transactions, such as large extends and truncates, can be done
306  * atomically all at once or in several stages.  The operation requests
307  * a credit for a number of buffer modications in advance, but can
308  * extend its credit if it needs more.
309  *
310  * journal_extend tries to give the running handle more buffer credits.
311  * It does not guarantee that allocation - this is a best-effort only.
312  * The calling process MUST be able to deal cleanly with a failure to
313  * extend here.
314  *
315  * Return 0 on success, non-zero on failure.
316  *
317  * return code < 0 implies an error
318  * return code > 0 implies normal transaction-full status.
319  */
320 int journal_extend(handle_t *handle, int nblocks)
321 {
322         transaction_t *transaction = handle->h_transaction;
323         journal_t *journal = transaction->t_journal;
324         int result;
325         int wanted;
326
327         result = -EIO;
328         if (is_handle_aborted(handle))
329                 goto out;
330
331         result = 1;
332
333         spin_lock(&journal->j_state_lock);
334
335         /* Don't extend a locked-down transaction! */
336         if (handle->h_transaction->t_state != T_RUNNING) {
337                 jbd_debug(3, "denied handle %p %d blocks: "
338                           "transaction not running\n", handle, nblocks);
339                 goto error_out;
340         }
341
342         spin_lock(&transaction->t_handle_lock);
343         wanted = transaction->t_outstanding_credits + nblocks;
344
345         if (wanted > journal->j_max_transaction_buffers) {
346                 jbd_debug(3, "denied handle %p %d blocks: "
347                           "transaction too large\n", handle, nblocks);
348                 goto unlock;
349         }
350
351         if (wanted > __log_space_left(journal)) {
352                 jbd_debug(3, "denied handle %p %d blocks: "
353                           "insufficient log space\n", handle, nblocks);
354                 goto unlock;
355         }
356
357         handle->h_buffer_credits += nblocks;
358         transaction->t_outstanding_credits += nblocks;
359         result = 0;
360
361         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
362 unlock:
363         spin_unlock(&transaction->t_handle_lock);
364 error_out:
365         spin_unlock(&journal->j_state_lock);
366 out:
367         return result;
368 }
369
370
371 /**
372  * int journal_restart() - restart a handle.
373  * @handle:  handle to restart
374  * @nblocks: nr credits requested
375  *
376  * Restart a handle for a multi-transaction filesystem
377  * operation.
378  *
379  * If the journal_extend() call above fails to grant new buffer credits
380  * to a running handle, a call to journal_restart will commit the
381  * handle's transaction so far and reattach the handle to a new
382  * transaction capabable of guaranteeing the requested number of
383  * credits.
384  */
385
386 int journal_restart(handle_t *handle, int nblocks)
387 {
388         transaction_t *transaction = handle->h_transaction;
389         journal_t *journal = transaction->t_journal;
390         int ret;
391
392         /* If we've had an abort of any type, don't even think about
393          * actually doing the restart! */
394         if (is_handle_aborted(handle))
395                 return 0;
396
397         /*
398          * First unlink the handle from its current transaction, and start the
399          * commit on that.
400          */
401         J_ASSERT(transaction->t_updates > 0);
402         J_ASSERT(journal_current_handle() == handle);
403
404         spin_lock(&journal->j_state_lock);
405         spin_lock(&transaction->t_handle_lock);
406         transaction->t_outstanding_credits -= handle->h_buffer_credits;
407         transaction->t_updates--;
408
409         if (!transaction->t_updates)
410                 wake_up(&journal->j_wait_updates);
411         spin_unlock(&transaction->t_handle_lock);
412
413         jbd_debug(2, "restarting handle %p\n", handle);
414         __log_start_commit(journal, transaction->t_tid);
415         spin_unlock(&journal->j_state_lock);
416
417         handle->h_buffer_credits = nblocks;
418         ret = start_this_handle(journal, handle);
419         return ret;
420 }
421
422
423 /**
424  * void journal_lock_updates () - establish a transaction barrier.
425  * @journal:  Journal to establish a barrier on.
426  *
427  * This locks out any further updates from being started, and blocks
428  * until all existing updates have completed, returning only once the
429  * journal is in a quiescent state with no updates running.
430  *
431  * The journal lock should not be held on entry.
432  */
433 void journal_lock_updates(journal_t *journal)
434 {
435         DEFINE_WAIT(wait);
436
437         spin_lock(&journal->j_state_lock);
438         ++journal->j_barrier_count;
439
440         /* Wait until there are no running updates */
441         while (1) {
442                 transaction_t *transaction = journal->j_running_transaction;
443
444                 if (!transaction)
445                         break;
446
447                 spin_lock(&transaction->t_handle_lock);
448                 if (!transaction->t_updates) {
449                         spin_unlock(&transaction->t_handle_lock);
450                         break;
451                 }
452                 prepare_to_wait(&journal->j_wait_updates, &wait,
453                                 TASK_UNINTERRUPTIBLE);
454                 spin_unlock(&transaction->t_handle_lock);
455                 spin_unlock(&journal->j_state_lock);
456                 schedule();
457                 finish_wait(&journal->j_wait_updates, &wait);
458                 spin_lock(&journal->j_state_lock);
459         }
460         spin_unlock(&journal->j_state_lock);
461
462         /*
463          * We have now established a barrier against other normal updates, but
464          * we also need to barrier against other journal_lock_updates() calls
465          * to make sure that we serialise special journal-locked operations
466          * too.
467          */
468         mutex_lock(&journal->j_barrier);
469 }
470
471 /**
472  * void journal_unlock_updates (journal_t* journal) - release barrier
473  * @journal:  Journal to release the barrier on.
474  *
475  * Release a transaction barrier obtained with journal_lock_updates().
476  *
477  * Should be called without the journal lock held.
478  */
479 void journal_unlock_updates (journal_t *journal)
480 {
481         J_ASSERT(journal->j_barrier_count != 0);
482
483         mutex_unlock(&journal->j_barrier);
484         spin_lock(&journal->j_state_lock);
485         --journal->j_barrier_count;
486         spin_unlock(&journal->j_state_lock);
487         wake_up(&journal->j_wait_transaction_locked);
488 }
489
490 /*
491  * Report any unexpected dirty buffers which turn up.  Normally those
492  * indicate an error, but they can occur if the user is running (say)
493  * tune2fs to modify the live filesystem, so we need the option of
494  * continuing as gracefully as possible.  #
495  *
496  * The caller should already hold the journal lock and
497  * j_list_lock spinlock: most callers will need those anyway
498  * in order to probe the buffer's journaling state safely.
499  */
500 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
501 {
502         int jlist;
503
504         /* If this buffer is one which might reasonably be dirty
505          * --- ie. data, or not part of this journal --- then
506          * we're OK to leave it alone, but otherwise we need to
507          * move the dirty bit to the journal's own internal
508          * JBDDirty bit. */
509         jlist = jh->b_jlist;
510
511         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
512             jlist == BJ_Shadow || jlist == BJ_Forget) {
513                 struct buffer_head *bh = jh2bh(jh);
514
515                 if (test_clear_buffer_dirty(bh))
516                         set_buffer_jbddirty(bh);
517         }
518 }
519
520 /*
521  * If the buffer is already part of the current transaction, then there
522  * is nothing we need to do.  If it is already part of a prior
523  * transaction which we are still committing to disk, then we need to
524  * make sure that we do not overwrite the old copy: we do copy-out to
525  * preserve the copy going to disk.  We also account the buffer against
526  * the handle's metadata buffer credits (unless the buffer is already
527  * part of the transaction, that is).
528  *
529  */
530 static int
531 do_get_write_access(handle_t *handle, struct journal_head *jh,
532                         int force_copy)
533 {
534         struct buffer_head *bh;
535         transaction_t *transaction;
536         journal_t *journal;
537         int error;
538         char *frozen_buffer = NULL;
539         int need_copy = 0;
540
541         if (is_handle_aborted(handle))
542                 return -EROFS;
543
544         transaction = handle->h_transaction;
545         journal = transaction->t_journal;
546
547         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
548
549         JBUFFER_TRACE(jh, "entry");
550 repeat:
551         bh = jh2bh(jh);
552
553         /* @@@ Need to check for errors here at some point. */
554
555         lock_buffer(bh);
556         jbd_lock_bh_state(bh);
557
558         /* We now hold the buffer lock so it is safe to query the buffer
559          * state.  Is the buffer dirty?
560          *
561          * If so, there are two possibilities.  The buffer may be
562          * non-journaled, and undergoing a quite legitimate writeback.
563          * Otherwise, it is journaled, and we don't expect dirty buffers
564          * in that state (the buffers should be marked JBD_Dirty
565          * instead.)  So either the IO is being done under our own
566          * control and this is a bug, or it's a third party IO such as
567          * dump(8) (which may leave the buffer scheduled for read ---
568          * ie. locked but not dirty) or tune2fs (which may actually have
569          * the buffer dirtied, ugh.)  */
570
571         if (buffer_dirty(bh)) {
572                 /*
573                  * First question: is this buffer already part of the current
574                  * transaction or the existing committing transaction?
575                  */
576                 if (jh->b_transaction) {
577                         J_ASSERT_JH(jh,
578                                 jh->b_transaction == transaction ||
579                                 jh->b_transaction ==
580                                         journal->j_committing_transaction);
581                         if (jh->b_next_transaction)
582                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
583                                                         transaction);
584                 }
585                 /*
586                  * In any case we need to clean the dirty flag and we must
587                  * do it under the buffer lock to be sure we don't race
588                  * with running write-out.
589                  */
590                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
591                 jbd_unexpected_dirty_buffer(jh);
592         }
593
594         unlock_buffer(bh);
595
596         error = -EROFS;
597         if (is_handle_aborted(handle)) {
598                 jbd_unlock_bh_state(bh);
599                 goto out;
600         }
601         error = 0;
602
603         /*
604          * The buffer is already part of this transaction if b_transaction or
605          * b_next_transaction points to it
606          */
607         if (jh->b_transaction == transaction ||
608             jh->b_next_transaction == transaction)
609                 goto done;
610
611         /*
612          * this is the first time this transaction is touching this buffer,
613          * reset the modified flag
614          */
615         jh->b_modified = 0;
616
617         /*
618          * If there is already a copy-out version of this buffer, then we don't
619          * need to make another one
620          */
621         if (jh->b_frozen_data) {
622                 JBUFFER_TRACE(jh, "has frozen data");
623                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
624                 jh->b_next_transaction = transaction;
625                 goto done;
626         }
627
628         /* Is there data here we need to preserve? */
629
630         if (jh->b_transaction && jh->b_transaction != transaction) {
631                 JBUFFER_TRACE(jh, "owned by older transaction");
632                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
633                 J_ASSERT_JH(jh, jh->b_transaction ==
634                                         journal->j_committing_transaction);
635
636                 /* There is one case we have to be very careful about.
637                  * If the committing transaction is currently writing
638                  * this buffer out to disk and has NOT made a copy-out,
639                  * then we cannot modify the buffer contents at all
640                  * right now.  The essence of copy-out is that it is the
641                  * extra copy, not the primary copy, which gets
642                  * journaled.  If the primary copy is already going to
643                  * disk then we cannot do copy-out here. */
644
645                 if (jh->b_jlist == BJ_Shadow) {
646                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
647                         wait_queue_head_t *wqh;
648
649                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
650
651                         JBUFFER_TRACE(jh, "on shadow: sleep");
652                         jbd_unlock_bh_state(bh);
653                         /* commit wakes up all shadow buffers after IO */
654                         for ( ; ; ) {
655                                 prepare_to_wait(wqh, &wait.wait,
656                                                 TASK_UNINTERRUPTIBLE);
657                                 if (jh->b_jlist != BJ_Shadow)
658                                         break;
659                                 schedule();
660                         }
661                         finish_wait(wqh, &wait.wait);
662                         goto repeat;
663                 }
664
665                 /* Only do the copy if the currently-owning transaction
666                  * still needs it.  If it is on the Forget list, the
667                  * committing transaction is past that stage.  The
668                  * buffer had better remain locked during the kmalloc,
669                  * but that should be true --- we hold the journal lock
670                  * still and the buffer is already on the BUF_JOURNAL
671                  * list so won't be flushed.
672                  *
673                  * Subtle point, though: if this is a get_undo_access,
674                  * then we will be relying on the frozen_data to contain
675                  * the new value of the committed_data record after the
676                  * transaction, so we HAVE to force the frozen_data copy
677                  * in that case. */
678
679                 if (jh->b_jlist != BJ_Forget || force_copy) {
680                         JBUFFER_TRACE(jh, "generate frozen data");
681                         if (!frozen_buffer) {
682                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
683                                 jbd_unlock_bh_state(bh);
684                                 frozen_buffer =
685                                         jbd_alloc(jh2bh(jh)->b_size,
686                                                          GFP_NOFS);
687                                 if (!frozen_buffer) {
688                                         printk(KERN_EMERG
689                                                "%s: OOM for frozen_buffer\n",
690                                                __func__);
691                                         JBUFFER_TRACE(jh, "oom!");
692                                         error = -ENOMEM;
693                                         jbd_lock_bh_state(bh);
694                                         goto done;
695                                 }
696                                 goto repeat;
697                         }
698                         jh->b_frozen_data = frozen_buffer;
699                         frozen_buffer = NULL;
700                         need_copy = 1;
701                 }
702                 jh->b_next_transaction = transaction;
703         }
704
705
706         /*
707          * Finally, if the buffer is not journaled right now, we need to make
708          * sure it doesn't get written to disk before the caller actually
709          * commits the new data
710          */
711         if (!jh->b_transaction) {
712                 JBUFFER_TRACE(jh, "no transaction");
713                 J_ASSERT_JH(jh, !jh->b_next_transaction);
714                 jh->b_transaction = transaction;
715                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
716                 spin_lock(&journal->j_list_lock);
717                 __journal_file_buffer(jh, transaction, BJ_Reserved);
718                 spin_unlock(&journal->j_list_lock);
719         }
720
721 done:
722         if (need_copy) {
723                 struct page *page;
724                 int offset;
725                 char *source;
726
727                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
728                             "Possible IO failure.\n");
729                 page = jh2bh(jh)->b_page;
730                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
731                 source = kmap_atomic(page, KM_USER0);
732                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
733                 kunmap_atomic(source, KM_USER0);
734         }
735         jbd_unlock_bh_state(bh);
736
737         /*
738          * If we are about to journal a buffer, then any revoke pending on it is
739          * no longer valid
740          */
741         journal_cancel_revoke(handle, jh);
742
743 out:
744         if (unlikely(frozen_buffer))    /* It's usually NULL */
745                 jbd_free(frozen_buffer, bh->b_size);
746
747         JBUFFER_TRACE(jh, "exit");
748         return error;
749 }
750
751 /**
752  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
753  * @handle: transaction to add buffer modifications to
754  * @bh:     bh to be used for metadata writes
755  * @credits: variable that will receive credits for the buffer
756  *
757  * Returns an error code or 0 on success.
758  *
759  * In full data journalling mode the buffer may be of type BJ_AsyncData,
760  * because we're write()ing a buffer which is also part of a shared mapping.
761  */
762
763 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
764 {
765         struct journal_head *jh = journal_add_journal_head(bh);
766         int rc;
767
768         /* We do not want to get caught playing with fields which the
769          * log thread also manipulates.  Make sure that the buffer
770          * completes any outstanding IO before proceeding. */
771         rc = do_get_write_access(handle, jh, 0);
772         journal_put_journal_head(jh);
773         return rc;
774 }
775
776
777 /*
778  * When the user wants to journal a newly created buffer_head
779  * (ie. getblk() returned a new buffer and we are going to populate it
780  * manually rather than reading off disk), then we need to keep the
781  * buffer_head locked until it has been completely filled with new
782  * data.  In this case, we should be able to make the assertion that
783  * the bh is not already part of an existing transaction.
784  *
785  * The buffer should already be locked by the caller by this point.
786  * There is no lock ranking violation: it was a newly created,
787  * unlocked buffer beforehand. */
788
789 /**
790  * int journal_get_create_access () - notify intent to use newly created bh
791  * @handle: transaction to new buffer to
792  * @bh: new buffer.
793  *
794  * Call this if you create a new bh.
795  */
796 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
797 {
798         transaction_t *transaction = handle->h_transaction;
799         journal_t *journal = transaction->t_journal;
800         struct journal_head *jh = journal_add_journal_head(bh);
801         int err;
802
803         jbd_debug(5, "journal_head %p\n", jh);
804         err = -EROFS;
805         if (is_handle_aborted(handle))
806                 goto out;
807         err = 0;
808
809         JBUFFER_TRACE(jh, "entry");
810         /*
811          * The buffer may already belong to this transaction due to pre-zeroing
812          * in the filesystem's new_block code.  It may also be on the previous,
813          * committing transaction's lists, but it HAS to be in Forget state in
814          * that case: the transaction must have deleted the buffer for it to be
815          * reused here.
816          */
817         jbd_lock_bh_state(bh);
818         spin_lock(&journal->j_list_lock);
819         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
820                 jh->b_transaction == NULL ||
821                 (jh->b_transaction == journal->j_committing_transaction &&
822                           jh->b_jlist == BJ_Forget)));
823
824         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
825         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
826
827         if (jh->b_transaction == NULL) {
828                 jh->b_transaction = transaction;
829
830                 /* first access by this transaction */
831                 jh->b_modified = 0;
832
833                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
834                 __journal_file_buffer(jh, transaction, BJ_Reserved);
835         } else if (jh->b_transaction == journal->j_committing_transaction) {
836                 /* first access by this transaction */
837                 jh->b_modified = 0;
838
839                 JBUFFER_TRACE(jh, "set next transaction");
840                 jh->b_next_transaction = transaction;
841         }
842         spin_unlock(&journal->j_list_lock);
843         jbd_unlock_bh_state(bh);
844
845         /*
846          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
847          * blocks which contain freed but then revoked metadata.  We need
848          * to cancel the revoke in case we end up freeing it yet again
849          * and the reallocating as data - this would cause a second revoke,
850          * which hits an assertion error.
851          */
852         JBUFFER_TRACE(jh, "cancelling revoke");
853         journal_cancel_revoke(handle, jh);
854         journal_put_journal_head(jh);
855 out:
856         return err;
857 }
858
859 /**
860  * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
861  * @handle: transaction
862  * @bh: buffer to undo
863  * @credits: store the number of taken credits here (if not NULL)
864  *
865  * Sometimes there is a need to distinguish between metadata which has
866  * been committed to disk and that which has not.  The ext3fs code uses
867  * this for freeing and allocating space, we have to make sure that we
868  * do not reuse freed space until the deallocation has been committed,
869  * since if we overwrote that space we would make the delete
870  * un-rewindable in case of a crash.
871  *
872  * To deal with that, journal_get_undo_access requests write access to a
873  * buffer for parts of non-rewindable operations such as delete
874  * operations on the bitmaps.  The journaling code must keep a copy of
875  * the buffer's contents prior to the undo_access call until such time
876  * as we know that the buffer has definitely been committed to disk.
877  *
878  * We never need to know which transaction the committed data is part
879  * of, buffers touched here are guaranteed to be dirtied later and so
880  * will be committed to a new transaction in due course, at which point
881  * we can discard the old committed data pointer.
882  *
883  * Returns error number or 0 on success.
884  */
885 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
886 {
887         int err;
888         struct journal_head *jh = journal_add_journal_head(bh);
889         char *committed_data = NULL;
890
891         JBUFFER_TRACE(jh, "entry");
892
893         /*
894          * Do this first --- it can drop the journal lock, so we want to
895          * make sure that obtaining the committed_data is done
896          * atomically wrt. completion of any outstanding commits.
897          */
898         err = do_get_write_access(handle, jh, 1);
899         if (err)
900                 goto out;
901
902 repeat:
903         if (!jh->b_committed_data) {
904                 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
905                 if (!committed_data) {
906                         printk(KERN_EMERG "%s: No memory for committed data\n",
907                                 __func__);
908                         err = -ENOMEM;
909                         goto out;
910                 }
911         }
912
913         jbd_lock_bh_state(bh);
914         if (!jh->b_committed_data) {
915                 /* Copy out the current buffer contents into the
916                  * preserved, committed copy. */
917                 JBUFFER_TRACE(jh, "generate b_committed data");
918                 if (!committed_data) {
919                         jbd_unlock_bh_state(bh);
920                         goto repeat;
921                 }
922
923                 jh->b_committed_data = committed_data;
924                 committed_data = NULL;
925                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
926         }
927         jbd_unlock_bh_state(bh);
928 out:
929         journal_put_journal_head(jh);
930         if (unlikely(committed_data))
931                 jbd_free(committed_data, bh->b_size);
932         return err;
933 }
934
935 /**
936  * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
937  * @handle: transaction
938  * @bh: bufferhead to mark
939  *
940  * Description:
941  * Mark a buffer as containing dirty data which needs to be flushed before
942  * we can commit the current transaction.
943  *
944  * The buffer is placed on the transaction's data list and is marked as
945  * belonging to the transaction.
946  *
947  * Returns error number or 0 on success.
948  *
949  * journal_dirty_data() can be called via page_launder->ext3_writepage
950  * by kswapd.
951  */
952 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
953 {
954         journal_t *journal = handle->h_transaction->t_journal;
955         int need_brelse = 0;
956         struct journal_head *jh;
957
958         if (is_handle_aborted(handle))
959                 return 0;
960
961         jh = journal_add_journal_head(bh);
962         JBUFFER_TRACE(jh, "entry");
963
964         /*
965          * The buffer could *already* be dirty.  Writeout can start
966          * at any time.
967          */
968         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
969
970         /*
971          * What if the buffer is already part of a running transaction?
972          *
973          * There are two cases:
974          * 1) It is part of the current running transaction.  Refile it,
975          *    just in case we have allocated it as metadata, deallocated
976          *    it, then reallocated it as data.
977          * 2) It is part of the previous, still-committing transaction.
978          *    If all we want to do is to guarantee that the buffer will be
979          *    written to disk before this new transaction commits, then
980          *    being sure that the *previous* transaction has this same
981          *    property is sufficient for us!  Just leave it on its old
982          *    transaction.
983          *
984          * In case (2), the buffer must not already exist as metadata
985          * --- that would violate write ordering (a transaction is free
986          * to write its data at any point, even before the previous
987          * committing transaction has committed).  The caller must
988          * never, ever allow this to happen: there's nothing we can do
989          * about it in this layer.
990          */
991         jbd_lock_bh_state(bh);
992         spin_lock(&journal->j_list_lock);
993
994         /* Now that we have bh_state locked, are we really still mapped? */
995         if (!buffer_mapped(bh)) {
996                 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
997                 goto no_journal;
998         }
999
1000         if (jh->b_transaction) {
1001                 JBUFFER_TRACE(jh, "has transaction");
1002                 if (jh->b_transaction != handle->h_transaction) {
1003                         JBUFFER_TRACE(jh, "belongs to older transaction");
1004                         J_ASSERT_JH(jh, jh->b_transaction ==
1005                                         journal->j_committing_transaction);
1006
1007                         /* @@@ IS THIS TRUE  ? */
1008                         /*
1009                          * Not any more.  Scenario: someone does a write()
1010                          * in data=journal mode.  The buffer's transaction has
1011                          * moved into commit.  Then someone does another
1012                          * write() to the file.  We do the frozen data copyout
1013                          * and set b_next_transaction to point to j_running_t.
1014                          * And while we're in that state, someone does a
1015                          * writepage() in an attempt to pageout the same area
1016                          * of the file via a shared mapping.  At present that
1017                          * calls journal_dirty_data(), and we get right here.
1018                          * It may be too late to journal the data.  Simply
1019                          * falling through to the next test will suffice: the
1020                          * data will be dirty and wil be checkpointed.  The
1021                          * ordering comments in the next comment block still
1022                          * apply.
1023                          */
1024                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1025
1026                         /*
1027                          * If we're journalling data, and this buffer was
1028                          * subject to a write(), it could be metadata, forget
1029                          * or shadow against the committing transaction.  Now,
1030                          * someone has dirtied the same darn page via a mapping
1031                          * and it is being writepage()'d.
1032                          * We *could* just steal the page from commit, with some
1033                          * fancy locking there.  Instead, we just skip it -
1034                          * don't tie the page's buffers to the new transaction
1035                          * at all.
1036                          * Implication: if we crash before the writepage() data
1037                          * is written into the filesystem, recovery will replay
1038                          * the write() data.
1039                          */
1040                         if (jh->b_jlist != BJ_None &&
1041                                         jh->b_jlist != BJ_SyncData &&
1042                                         jh->b_jlist != BJ_Locked) {
1043                                 JBUFFER_TRACE(jh, "Not stealing");
1044                                 goto no_journal;
1045                         }
1046
1047                         /*
1048                          * This buffer may be undergoing writeout in commit.  We
1049                          * can't return from here and let the caller dirty it
1050                          * again because that can cause the write-out loop in
1051                          * commit to never terminate.
1052                          */
1053                         if (buffer_dirty(bh)) {
1054                                 get_bh(bh);
1055                                 spin_unlock(&journal->j_list_lock);
1056                                 jbd_unlock_bh_state(bh);
1057                                 need_brelse = 1;
1058                                 sync_dirty_buffer(bh);
1059                                 jbd_lock_bh_state(bh);
1060                                 spin_lock(&journal->j_list_lock);
1061                                 /* Since we dropped the lock... */
1062                                 if (!buffer_mapped(bh)) {
1063                                         JBUFFER_TRACE(jh, "buffer got unmapped");
1064                                         goto no_journal;
1065                                 }
1066                                 /* The buffer may become locked again at any
1067                                    time if it is redirtied */
1068                         }
1069
1070                         /* journal_clean_data_list() may have got there first */
1071                         if (jh->b_transaction != NULL) {
1072                                 JBUFFER_TRACE(jh, "unfile from commit");
1073                                 __journal_temp_unlink_buffer(jh);
1074                                 /* It still points to the committing
1075                                  * transaction; move it to this one so
1076                                  * that the refile assert checks are
1077                                  * happy. */
1078                                 jh->b_transaction = handle->h_transaction;
1079                         }
1080                         /* The buffer will be refiled below */
1081
1082                 }
1083                 /*
1084                  * Special case --- the buffer might actually have been
1085                  * allocated and then immediately deallocated in the previous,
1086                  * committing transaction, so might still be left on that
1087                  * transaction's metadata lists.
1088                  */
1089                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1090                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1091                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1092                         __journal_temp_unlink_buffer(jh);
1093                         jh->b_transaction = handle->h_transaction;
1094                         JBUFFER_TRACE(jh, "file as data");
1095                         __journal_file_buffer(jh, handle->h_transaction,
1096                                                 BJ_SyncData);
1097                 }
1098         } else {
1099                 JBUFFER_TRACE(jh, "not on a transaction");
1100                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1101         }
1102 no_journal:
1103         spin_unlock(&journal->j_list_lock);
1104         jbd_unlock_bh_state(bh);
1105         if (need_brelse) {
1106                 BUFFER_TRACE(bh, "brelse");
1107                 __brelse(bh);
1108         }
1109         JBUFFER_TRACE(jh, "exit");
1110         journal_put_journal_head(jh);
1111         return 0;
1112 }
1113
1114 /**
1115  * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1116  * @handle: transaction to add buffer to.
1117  * @bh: buffer to mark
1118  *
1119  * Mark dirty metadata which needs to be journaled as part of the current
1120  * transaction.
1121  *
1122  * The buffer is placed on the transaction's metadata list and is marked
1123  * as belonging to the transaction.
1124  *
1125  * Returns error number or 0 on success.
1126  *
1127  * Special care needs to be taken if the buffer already belongs to the
1128  * current committing transaction (in which case we should have frozen
1129  * data present for that commit).  In that case, we don't relink the
1130  * buffer: that only gets done when the old transaction finally
1131  * completes its commit.
1132  */
1133 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1134 {
1135         transaction_t *transaction = handle->h_transaction;
1136         journal_t *journal = transaction->t_journal;
1137         struct journal_head *jh = bh2jh(bh);
1138
1139         jbd_debug(5, "journal_head %p\n", jh);
1140         JBUFFER_TRACE(jh, "entry");
1141         if (is_handle_aborted(handle))
1142                 goto out;
1143
1144         jbd_lock_bh_state(bh);
1145
1146         if (jh->b_modified == 0) {
1147                 /*
1148                  * This buffer's got modified and becoming part
1149                  * of the transaction. This needs to be done
1150                  * once a transaction -bzzz
1151                  */
1152                 jh->b_modified = 1;
1153                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1154                 handle->h_buffer_credits--;
1155         }
1156
1157         /*
1158          * fastpath, to avoid expensive locking.  If this buffer is already
1159          * on the running transaction's metadata list there is nothing to do.
1160          * Nobody can take it off again because there is a handle open.
1161          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1162          * result in this test being false, so we go in and take the locks.
1163          */
1164         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1165                 JBUFFER_TRACE(jh, "fastpath");
1166                 J_ASSERT_JH(jh, jh->b_transaction ==
1167                                         journal->j_running_transaction);
1168                 goto out_unlock_bh;
1169         }
1170
1171         set_buffer_jbddirty(bh);
1172
1173         /*
1174          * Metadata already on the current transaction list doesn't
1175          * need to be filed.  Metadata on another transaction's list must
1176          * be committing, and will be refiled once the commit completes:
1177          * leave it alone for now.
1178          */
1179         if (jh->b_transaction != transaction) {
1180                 JBUFFER_TRACE(jh, "already on other transaction");
1181                 J_ASSERT_JH(jh, jh->b_transaction ==
1182                                         journal->j_committing_transaction);
1183                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1184                 /* And this case is illegal: we can't reuse another
1185                  * transaction's data buffer, ever. */
1186                 goto out_unlock_bh;
1187         }
1188
1189         /* That test should have eliminated the following case: */
1190         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1191
1192         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1193         spin_lock(&journal->j_list_lock);
1194         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1195         spin_unlock(&journal->j_list_lock);
1196 out_unlock_bh:
1197         jbd_unlock_bh_state(bh);
1198 out:
1199         JBUFFER_TRACE(jh, "exit");
1200         return 0;
1201 }
1202
1203 /*
1204  * journal_release_buffer: undo a get_write_access without any buffer
1205  * updates, if the update decided in the end that it didn't need access.
1206  *
1207  */
1208 void
1209 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1210 {
1211         BUFFER_TRACE(bh, "entry");
1212 }
1213
1214 /**
1215  * void journal_forget() - bforget() for potentially-journaled buffers.
1216  * @handle: transaction handle
1217  * @bh:     bh to 'forget'
1218  *
1219  * We can only do the bforget if there are no commits pending against the
1220  * buffer.  If the buffer is dirty in the current running transaction we
1221  * can safely unlink it.
1222  *
1223  * bh may not be a journalled buffer at all - it may be a non-JBD
1224  * buffer which came off the hashtable.  Check for this.
1225  *
1226  * Decrements bh->b_count by one.
1227  *
1228  * Allow this call even if the handle has aborted --- it may be part of
1229  * the caller's cleanup after an abort.
1230  */
1231 int journal_forget (handle_t *handle, struct buffer_head *bh)
1232 {
1233         transaction_t *transaction = handle->h_transaction;
1234         journal_t *journal = transaction->t_journal;
1235         struct journal_head *jh;
1236         int drop_reserve = 0;
1237         int err = 0;
1238         int was_modified = 0;
1239
1240         BUFFER_TRACE(bh, "entry");
1241
1242         jbd_lock_bh_state(bh);
1243         spin_lock(&journal->j_list_lock);
1244
1245         if (!buffer_jbd(bh))
1246                 goto not_jbd;
1247         jh = bh2jh(bh);
1248
1249         /* Critical error: attempting to delete a bitmap buffer, maybe?
1250          * Don't do any jbd operations, and return an error. */
1251         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1252                          "inconsistent data on disk")) {
1253                 err = -EIO;
1254                 goto not_jbd;
1255         }
1256
1257         /* keep track of wether or not this transaction modified us */
1258         was_modified = jh->b_modified;
1259
1260         /*
1261          * The buffer's going from the transaction, we must drop
1262          * all references -bzzz
1263          */
1264         jh->b_modified = 0;
1265
1266         if (jh->b_transaction == handle->h_transaction) {
1267                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1268
1269                 /* If we are forgetting a buffer which is already part
1270                  * of this transaction, then we can just drop it from
1271                  * the transaction immediately. */
1272                 clear_buffer_dirty(bh);
1273                 clear_buffer_jbddirty(bh);
1274
1275                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1276
1277                 /*
1278                  * we only want to drop a reference if this transaction
1279                  * modified the buffer
1280                  */
1281                 if (was_modified)
1282                         drop_reserve = 1;
1283
1284                 /*
1285                  * We are no longer going to journal this buffer.
1286                  * However, the commit of this transaction is still
1287                  * important to the buffer: the delete that we are now
1288                  * processing might obsolete an old log entry, so by
1289                  * committing, we can satisfy the buffer's checkpoint.
1290                  *
1291                  * So, if we have a checkpoint on the buffer, we should
1292                  * now refile the buffer on our BJ_Forget list so that
1293                  * we know to remove the checkpoint after we commit.
1294                  */
1295
1296                 if (jh->b_cp_transaction) {
1297                         __journal_temp_unlink_buffer(jh);
1298                         __journal_file_buffer(jh, transaction, BJ_Forget);
1299                 } else {
1300                         __journal_unfile_buffer(jh);
1301                         journal_remove_journal_head(bh);
1302                         __brelse(bh);
1303                         if (!buffer_jbd(bh)) {
1304                                 spin_unlock(&journal->j_list_lock);
1305                                 jbd_unlock_bh_state(bh);
1306                                 __bforget(bh);
1307                                 goto drop;
1308                         }
1309                 }
1310         } else if (jh->b_transaction) {
1311                 J_ASSERT_JH(jh, (jh->b_transaction ==
1312                                  journal->j_committing_transaction));
1313                 /* However, if the buffer is still owned by a prior
1314                  * (committing) transaction, we can't drop it yet... */
1315                 JBUFFER_TRACE(jh, "belongs to older transaction");
1316                 /* ... but we CAN drop it from the new transaction if we
1317                  * have also modified it since the original commit. */
1318
1319                 if (jh->b_next_transaction) {
1320                         J_ASSERT(jh->b_next_transaction == transaction);
1321                         jh->b_next_transaction = NULL;
1322
1323                         /*
1324                          * only drop a reference if this transaction modified
1325                          * the buffer
1326                          */
1327                         if (was_modified)
1328                                 drop_reserve = 1;
1329                 }
1330         }
1331
1332 not_jbd:
1333         spin_unlock(&journal->j_list_lock);
1334         jbd_unlock_bh_state(bh);
1335         __brelse(bh);
1336 drop:
1337         if (drop_reserve) {
1338                 /* no need to reserve log space for this block -bzzz */
1339                 handle->h_buffer_credits++;
1340         }
1341         return err;
1342 }
1343
1344 /**
1345  * int journal_stop() - complete a transaction
1346  * @handle: tranaction to complete.
1347  *
1348  * All done for a particular handle.
1349  *
1350  * There is not much action needed here.  We just return any remaining
1351  * buffer credits to the transaction and remove the handle.  The only
1352  * complication is that we need to start a commit operation if the
1353  * filesystem is marked for synchronous update.
1354  *
1355  * journal_stop itself will not usually return an error, but it may
1356  * do so in unusual circumstances.  In particular, expect it to
1357  * return -EIO if a journal_abort has been executed since the
1358  * transaction began.
1359  */
1360 int journal_stop(handle_t *handle)
1361 {
1362         transaction_t *transaction = handle->h_transaction;
1363         journal_t *journal = transaction->t_journal;
1364         int old_handle_count, err;
1365         pid_t pid;
1366
1367         J_ASSERT(journal_current_handle() == handle);
1368
1369         if (is_handle_aborted(handle))
1370                 err = -EIO;
1371         else {
1372                 J_ASSERT(transaction->t_updates > 0);
1373                 err = 0;
1374         }
1375
1376         if (--handle->h_ref > 0) {
1377                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1378                           handle->h_ref);
1379                 return err;
1380         }
1381
1382         jbd_debug(4, "Handle %p going down\n", handle);
1383
1384         /*
1385          * Implement synchronous transaction batching.  If the handle
1386          * was synchronous, don't force a commit immediately.  Let's
1387          * yield and let another thread piggyback onto this transaction.
1388          * Keep doing that while new threads continue to arrive.
1389          * It doesn't cost much - we're about to run a commit and sleep
1390          * on IO anyway.  Speeds up many-threaded, many-dir operations
1391          * by 30x or more...
1392          *
1393          * But don't do this if this process was the most recent one to
1394          * perform a synchronous write.  We do this to detect the case where a
1395          * single process is doing a stream of sync writes.  No point in waiting
1396          * for joiners in that case.
1397          */
1398         pid = current->pid;
1399         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1400                 journal->j_last_sync_writer = pid;
1401                 do {
1402                         old_handle_count = transaction->t_handle_count;
1403                         schedule_timeout_uninterruptible(1);
1404                 } while (old_handle_count != transaction->t_handle_count);
1405         }
1406
1407         current->journal_info = NULL;
1408         spin_lock(&journal->j_state_lock);
1409         spin_lock(&transaction->t_handle_lock);
1410         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1411         transaction->t_updates--;
1412         if (!transaction->t_updates) {
1413                 wake_up(&journal->j_wait_updates);
1414                 if (journal->j_barrier_count)
1415                         wake_up(&journal->j_wait_transaction_locked);
1416         }
1417
1418         /*
1419          * If the handle is marked SYNC, we need to set another commit
1420          * going!  We also want to force a commit if the current
1421          * transaction is occupying too much of the log, or if the
1422          * transaction is too old now.
1423          */
1424         if (handle->h_sync ||
1425                         transaction->t_outstanding_credits >
1426                                 journal->j_max_transaction_buffers ||
1427                         time_after_eq(jiffies, transaction->t_expires)) {
1428                 /* Do this even for aborted journals: an abort still
1429                  * completes the commit thread, it just doesn't write
1430                  * anything to disk. */
1431                 tid_t tid = transaction->t_tid;
1432
1433                 spin_unlock(&transaction->t_handle_lock);
1434                 jbd_debug(2, "transaction too old, requesting commit for "
1435                                         "handle %p\n", handle);
1436                 /* This is non-blocking */
1437                 __log_start_commit(journal, transaction->t_tid);
1438                 spin_unlock(&journal->j_state_lock);
1439
1440                 /*
1441                  * Special case: JFS_SYNC synchronous updates require us
1442                  * to wait for the commit to complete.
1443                  */
1444                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1445                         err = log_wait_commit(journal, tid);
1446         } else {
1447                 spin_unlock(&transaction->t_handle_lock);
1448                 spin_unlock(&journal->j_state_lock);
1449         }
1450
1451         lock_map_release(&handle->h_lockdep_map);
1452
1453         jbd_free_handle(handle);
1454         return err;
1455 }
1456
1457 /**
1458  * int journal_force_commit() - force any uncommitted transactions
1459  * @journal: journal to force
1460  *
1461  * For synchronous operations: force any uncommitted transactions
1462  * to disk.  May seem kludgy, but it reuses all the handle batching
1463  * code in a very simple manner.
1464  */
1465 int journal_force_commit(journal_t *journal)
1466 {
1467         handle_t *handle;
1468         int ret;
1469
1470         handle = journal_start(journal, 1);
1471         if (IS_ERR(handle)) {
1472                 ret = PTR_ERR(handle);
1473         } else {
1474                 handle->h_sync = 1;
1475                 ret = journal_stop(handle);
1476         }
1477         return ret;
1478 }
1479
1480 /*
1481  *
1482  * List management code snippets: various functions for manipulating the
1483  * transaction buffer lists.
1484  *
1485  */
1486
1487 /*
1488  * Append a buffer to a transaction list, given the transaction's list head
1489  * pointer.
1490  *
1491  * j_list_lock is held.
1492  *
1493  * jbd_lock_bh_state(jh2bh(jh)) is held.
1494  */
1495
1496 static inline void
1497 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1498 {
1499         if (!*list) {
1500                 jh->b_tnext = jh->b_tprev = jh;
1501                 *list = jh;
1502         } else {
1503                 /* Insert at the tail of the list to preserve order */
1504                 struct journal_head *first = *list, *last = first->b_tprev;
1505                 jh->b_tprev = last;
1506                 jh->b_tnext = first;
1507                 last->b_tnext = first->b_tprev = jh;
1508         }
1509 }
1510
1511 /*
1512  * Remove a buffer from a transaction list, given the transaction's list
1513  * head pointer.
1514  *
1515  * Called with j_list_lock held, and the journal may not be locked.
1516  *
1517  * jbd_lock_bh_state(jh2bh(jh)) is held.
1518  */
1519
1520 static inline void
1521 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1522 {
1523         if (*list == jh) {
1524                 *list = jh->b_tnext;
1525                 if (*list == jh)
1526                         *list = NULL;
1527         }
1528         jh->b_tprev->b_tnext = jh->b_tnext;
1529         jh->b_tnext->b_tprev = jh->b_tprev;
1530 }
1531
1532 /*
1533  * Remove a buffer from the appropriate transaction list.
1534  *
1535  * Note that this function can *change* the value of
1536  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1537  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1538  * is holding onto a copy of one of thee pointers, it could go bad.
1539  * Generally the caller needs to re-read the pointer from the transaction_t.
1540  *
1541  * Called under j_list_lock.  The journal may not be locked.
1542  */
1543 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1544 {
1545         struct journal_head **list = NULL;
1546         transaction_t *transaction;
1547         struct buffer_head *bh = jh2bh(jh);
1548
1549         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1550         transaction = jh->b_transaction;
1551         if (transaction)
1552                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1553
1554         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1555         if (jh->b_jlist != BJ_None)
1556                 J_ASSERT_JH(jh, transaction != NULL);
1557
1558         switch (jh->b_jlist) {
1559         case BJ_None:
1560                 return;
1561         case BJ_SyncData:
1562                 list = &transaction->t_sync_datalist;
1563                 break;
1564         case BJ_Metadata:
1565                 transaction->t_nr_buffers--;
1566                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1567                 list = &transaction->t_buffers;
1568                 break;
1569         case BJ_Forget:
1570                 list = &transaction->t_forget;
1571                 break;
1572         case BJ_IO:
1573                 list = &transaction->t_iobuf_list;
1574                 break;
1575         case BJ_Shadow:
1576                 list = &transaction->t_shadow_list;
1577                 break;
1578         case BJ_LogCtl:
1579                 list = &transaction->t_log_list;
1580                 break;
1581         case BJ_Reserved:
1582                 list = &transaction->t_reserved_list;
1583                 break;
1584         case BJ_Locked:
1585                 list = &transaction->t_locked_list;
1586                 break;
1587         }
1588
1589         __blist_del_buffer(list, jh);
1590         jh->b_jlist = BJ_None;
1591         if (test_clear_buffer_jbddirty(bh))
1592                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1593 }
1594
1595 void __journal_unfile_buffer(struct journal_head *jh)
1596 {
1597         __journal_temp_unlink_buffer(jh);
1598         jh->b_transaction = NULL;
1599 }
1600
1601 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1602 {
1603         jbd_lock_bh_state(jh2bh(jh));
1604         spin_lock(&journal->j_list_lock);
1605         __journal_unfile_buffer(jh);
1606         spin_unlock(&journal->j_list_lock);
1607         jbd_unlock_bh_state(jh2bh(jh));
1608 }
1609
1610 /*
1611  * Called from journal_try_to_free_buffers().
1612  *
1613  * Called under jbd_lock_bh_state(bh)
1614  */
1615 static void
1616 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1617 {
1618         struct journal_head *jh;
1619
1620         jh = bh2jh(bh);
1621
1622         if (buffer_locked(bh) || buffer_dirty(bh))
1623                 goto out;
1624
1625         if (jh->b_next_transaction != NULL)
1626                 goto out;
1627
1628         spin_lock(&journal->j_list_lock);
1629         if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1630                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1631                         /* A written-back ordered data buffer */
1632                         JBUFFER_TRACE(jh, "release data");
1633                         __journal_unfile_buffer(jh);
1634                         journal_remove_journal_head(bh);
1635                         __brelse(bh);
1636                 }
1637         } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1638                 /* written-back checkpointed metadata buffer */
1639                 if (jh->b_jlist == BJ_None) {
1640                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1641                         __journal_remove_checkpoint(jh);
1642                         journal_remove_journal_head(bh);
1643                         __brelse(bh);
1644                 }
1645         }
1646         spin_unlock(&journal->j_list_lock);
1647 out:
1648         return;
1649 }
1650
1651 /*
1652  * journal_try_to_free_buffers() could race with journal_commit_transaction()
1653  * The latter might still hold the a count on buffers when inspecting
1654  * them on t_syncdata_list or t_locked_list.
1655  *
1656  * journal_try_to_free_buffers() will call this function to
1657  * wait for the current transaction to finish syncing data buffers, before
1658  * tryinf to free that buffer.
1659  *
1660  * Called with journal->j_state_lock held.
1661  */
1662 static void journal_wait_for_transaction_sync_data(journal_t *journal)
1663 {
1664         transaction_t *transaction = NULL;
1665         tid_t tid;
1666
1667         spin_lock(&journal->j_state_lock);
1668         transaction = journal->j_committing_transaction;
1669
1670         if (!transaction) {
1671                 spin_unlock(&journal->j_state_lock);
1672                 return;
1673         }
1674
1675         tid = transaction->t_tid;
1676         spin_unlock(&journal->j_state_lock);
1677         log_wait_commit(journal, tid);
1678 }
1679
1680 /**
1681  * int journal_try_to_free_buffers() - try to free page buffers.
1682  * @journal: journal for operation
1683  * @page: to try and free
1684  * @gfp_mask: we use the mask to detect how hard should we try to release
1685  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1686  * release the buffers.
1687  *
1688  *
1689  * For all the buffers on this page,
1690  * if they are fully written out ordered data, move them onto BUF_CLEAN
1691  * so try_to_free_buffers() can reap them.
1692  *
1693  * This function returns non-zero if we wish try_to_free_buffers()
1694  * to be called. We do this if the page is releasable by try_to_free_buffers().
1695  * We also do it if the page has locked or dirty buffers and the caller wants
1696  * us to perform sync or async writeout.
1697  *
1698  * This complicates JBD locking somewhat.  We aren't protected by the
1699  * BKL here.  We wish to remove the buffer from its committing or
1700  * running transaction's ->t_datalist via __journal_unfile_buffer.
1701  *
1702  * This may *change* the value of transaction_t->t_datalist, so anyone
1703  * who looks at t_datalist needs to lock against this function.
1704  *
1705  * Even worse, someone may be doing a journal_dirty_data on this
1706  * buffer.  So we need to lock against that.  journal_dirty_data()
1707  * will come out of the lock with the buffer dirty, which makes it
1708  * ineligible for release here.
1709  *
1710  * Who else is affected by this?  hmm...  Really the only contender
1711  * is do_get_write_access() - it could be looking at the buffer while
1712  * journal_try_to_free_buffer() is changing its state.  But that
1713  * cannot happen because we never reallocate freed data as metadata
1714  * while the data is part of a transaction.  Yes?
1715  *
1716  * Return 0 on failure, 1 on success
1717  */
1718 int journal_try_to_free_buffers(journal_t *journal,
1719                                 struct page *page, gfp_t gfp_mask)
1720 {
1721         struct buffer_head *head;
1722         struct buffer_head *bh;
1723         int ret = 0;
1724
1725         J_ASSERT(PageLocked(page));
1726
1727         head = page_buffers(page);
1728         bh = head;
1729         do {
1730                 struct journal_head *jh;
1731
1732                 /*
1733                  * We take our own ref against the journal_head here to avoid
1734                  * having to add tons of locking around each instance of
1735                  * journal_remove_journal_head() and journal_put_journal_head().
1736                  */
1737                 jh = journal_grab_journal_head(bh);
1738                 if (!jh)
1739                         continue;
1740
1741                 jbd_lock_bh_state(bh);
1742                 __journal_try_to_free_buffer(journal, bh);
1743                 journal_put_journal_head(jh);
1744                 jbd_unlock_bh_state(bh);
1745                 if (buffer_jbd(bh))
1746                         goto busy;
1747         } while ((bh = bh->b_this_page) != head);
1748
1749         ret = try_to_free_buffers(page);
1750
1751         /*
1752          * There are a number of places where journal_try_to_free_buffers()
1753          * could race with journal_commit_transaction(), the later still
1754          * holds the reference to the buffers to free while processing them.
1755          * try_to_free_buffers() failed to free those buffers. Some of the
1756          * caller of releasepage() request page buffers to be dropped, otherwise
1757          * treat the fail-to-free as errors (such as generic_file_direct_IO())
1758          *
1759          * So, if the caller of try_to_release_page() wants the synchronous
1760          * behaviour(i.e make sure buffers are dropped upon return),
1761          * let's wait for the current transaction to finish flush of
1762          * dirty data buffers, then try to free those buffers again,
1763          * with the journal locked.
1764          */
1765         if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
1766                 journal_wait_for_transaction_sync_data(journal);
1767                 ret = try_to_free_buffers(page);
1768         }
1769
1770 busy:
1771         return ret;
1772 }
1773
1774 /*
1775  * This buffer is no longer needed.  If it is on an older transaction's
1776  * checkpoint list we need to record it on this transaction's forget list
1777  * to pin this buffer (and hence its checkpointing transaction) down until
1778  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1779  * release it.
1780  * Returns non-zero if JBD no longer has an interest in the buffer.
1781  *
1782  * Called under j_list_lock.
1783  *
1784  * Called under jbd_lock_bh_state(bh).
1785  */
1786 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1787 {
1788         int may_free = 1;
1789         struct buffer_head *bh = jh2bh(jh);
1790
1791         __journal_unfile_buffer(jh);
1792
1793         if (jh->b_cp_transaction) {
1794                 JBUFFER_TRACE(jh, "on running+cp transaction");
1795                 __journal_file_buffer(jh, transaction, BJ_Forget);
1796                 clear_buffer_jbddirty(bh);
1797                 may_free = 0;
1798         } else {
1799                 JBUFFER_TRACE(jh, "on running transaction");
1800                 journal_remove_journal_head(bh);
1801                 __brelse(bh);
1802         }
1803         return may_free;
1804 }
1805
1806 /*
1807  * journal_invalidatepage
1808  *
1809  * This code is tricky.  It has a number of cases to deal with.
1810  *
1811  * There are two invariants which this code relies on:
1812  *
1813  * i_size must be updated on disk before we start calling invalidatepage on the
1814  * data.
1815  *
1816  *  This is done in ext3 by defining an ext3_setattr method which
1817  *  updates i_size before truncate gets going.  By maintaining this
1818  *  invariant, we can be sure that it is safe to throw away any buffers
1819  *  attached to the current transaction: once the transaction commits,
1820  *  we know that the data will not be needed.
1821  *
1822  *  Note however that we can *not* throw away data belonging to the
1823  *  previous, committing transaction!
1824  *
1825  * Any disk blocks which *are* part of the previous, committing
1826  * transaction (and which therefore cannot be discarded immediately) are
1827  * not going to be reused in the new running transaction
1828  *
1829  *  The bitmap committed_data images guarantee this: any block which is
1830  *  allocated in one transaction and removed in the next will be marked
1831  *  as in-use in the committed_data bitmap, so cannot be reused until
1832  *  the next transaction to delete the block commits.  This means that
1833  *  leaving committing buffers dirty is quite safe: the disk blocks
1834  *  cannot be reallocated to a different file and so buffer aliasing is
1835  *  not possible.
1836  *
1837  *
1838  * The above applies mainly to ordered data mode.  In writeback mode we
1839  * don't make guarantees about the order in which data hits disk --- in
1840  * particular we don't guarantee that new dirty data is flushed before
1841  * transaction commit --- so it is always safe just to discard data
1842  * immediately in that mode.  --sct
1843  */
1844
1845 /*
1846  * The journal_unmap_buffer helper function returns zero if the buffer
1847  * concerned remains pinned as an anonymous buffer belonging to an older
1848  * transaction.
1849  *
1850  * We're outside-transaction here.  Either or both of j_running_transaction
1851  * and j_committing_transaction may be NULL.
1852  */
1853 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1854 {
1855         transaction_t *transaction;
1856         struct journal_head *jh;
1857         int may_free = 1;
1858         int ret;
1859
1860         BUFFER_TRACE(bh, "entry");
1861
1862         /*
1863          * It is safe to proceed here without the j_list_lock because the
1864          * buffers cannot be stolen by try_to_free_buffers as long as we are
1865          * holding the page lock. --sct
1866          */
1867
1868         if (!buffer_jbd(bh))
1869                 goto zap_buffer_unlocked;
1870
1871         spin_lock(&journal->j_state_lock);
1872         jbd_lock_bh_state(bh);
1873         spin_lock(&journal->j_list_lock);
1874
1875         jh = journal_grab_journal_head(bh);
1876         if (!jh)
1877                 goto zap_buffer_no_jh;
1878
1879         transaction = jh->b_transaction;
1880         if (transaction == NULL) {
1881                 /* First case: not on any transaction.  If it
1882                  * has no checkpoint link, then we can zap it:
1883                  * it's a writeback-mode buffer so we don't care
1884                  * if it hits disk safely. */
1885                 if (!jh->b_cp_transaction) {
1886                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1887                         goto zap_buffer;
1888                 }
1889
1890                 if (!buffer_dirty(bh)) {
1891                         /* bdflush has written it.  We can drop it now */
1892                         goto zap_buffer;
1893                 }
1894
1895                 /* OK, it must be in the journal but still not
1896                  * written fully to disk: it's metadata or
1897                  * journaled data... */
1898
1899                 if (journal->j_running_transaction) {
1900                         /* ... and once the current transaction has
1901                          * committed, the buffer won't be needed any
1902                          * longer. */
1903                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1904                         ret = __dispose_buffer(jh,
1905                                         journal->j_running_transaction);
1906                         journal_put_journal_head(jh);
1907                         spin_unlock(&journal->j_list_lock);
1908                         jbd_unlock_bh_state(bh);
1909                         spin_unlock(&journal->j_state_lock);
1910                         return ret;
1911                 } else {
1912                         /* There is no currently-running transaction. So the
1913                          * orphan record which we wrote for this file must have
1914                          * passed into commit.  We must attach this buffer to
1915                          * the committing transaction, if it exists. */
1916                         if (journal->j_committing_transaction) {
1917                                 JBUFFER_TRACE(jh, "give to committing trans");
1918                                 ret = __dispose_buffer(jh,
1919                                         journal->j_committing_transaction);
1920                                 journal_put_journal_head(jh);
1921                                 spin_unlock(&journal->j_list_lock);
1922                                 jbd_unlock_bh_state(bh);
1923                                 spin_unlock(&journal->j_state_lock);
1924                                 return ret;
1925                         } else {
1926                                 /* The orphan record's transaction has
1927                                  * committed.  We can cleanse this buffer */
1928                                 clear_buffer_jbddirty(bh);
1929                                 goto zap_buffer;
1930                         }
1931                 }
1932         } else if (transaction == journal->j_committing_transaction) {
1933                 JBUFFER_TRACE(jh, "on committing transaction");
1934                 if (jh->b_jlist == BJ_Locked) {
1935                         /*
1936                          * The buffer is on the committing transaction's locked
1937                          * list.  We have the buffer locked, so I/O has
1938                          * completed.  So we can nail the buffer now.
1939                          */
1940                         may_free = __dispose_buffer(jh, transaction);
1941                         goto zap_buffer;
1942                 }
1943                 /*
1944                  * If it is committing, we simply cannot touch it.  We
1945                  * can remove it's next_transaction pointer from the
1946                  * running transaction if that is set, but nothing
1947                  * else. */
1948                 set_buffer_freed(bh);
1949                 if (jh->b_next_transaction) {
1950                         J_ASSERT(jh->b_next_transaction ==
1951                                         journal->j_running_transaction);
1952                         jh->b_next_transaction = NULL;
1953                 }
1954                 journal_put_journal_head(jh);
1955                 spin_unlock(&journal->j_list_lock);
1956                 jbd_unlock_bh_state(bh);
1957                 spin_unlock(&journal->j_state_lock);
1958                 return 0;
1959         } else {
1960                 /* Good, the buffer belongs to the running transaction.
1961                  * We are writing our own transaction's data, not any
1962                  * previous one's, so it is safe to throw it away
1963                  * (remember that we expect the filesystem to have set
1964                  * i_size already for this truncate so recovery will not
1965                  * expose the disk blocks we are discarding here.) */
1966                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1967                 JBUFFER_TRACE(jh, "on running transaction");
1968                 may_free = __dispose_buffer(jh, transaction);
1969         }
1970
1971 zap_buffer:
1972         journal_put_journal_head(jh);
1973 zap_buffer_no_jh:
1974         spin_unlock(&journal->j_list_lock);
1975         jbd_unlock_bh_state(bh);
1976         spin_unlock(&journal->j_state_lock);
1977 zap_buffer_unlocked:
1978         clear_buffer_dirty(bh);
1979         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1980         clear_buffer_mapped(bh);
1981         clear_buffer_req(bh);
1982         clear_buffer_new(bh);
1983         bh->b_bdev = NULL;
1984         return may_free;
1985 }
1986
1987 /**
1988  * void journal_invalidatepage() - invalidate a journal page
1989  * @journal: journal to use for flush
1990  * @page:    page to flush
1991  * @offset:  length of page to invalidate.
1992  *
1993  * Reap page buffers containing data after offset in page.
1994  */
1995 void journal_invalidatepage(journal_t *journal,
1996                       struct page *page,
1997                       unsigned long offset)
1998 {
1999         struct buffer_head *head, *bh, *next;
2000         unsigned int curr_off = 0;
2001         int may_free = 1;
2002
2003         if (!PageLocked(page))
2004                 BUG();
2005         if (!page_has_buffers(page))
2006                 return;
2007
2008         /* We will potentially be playing with lists other than just the
2009          * data lists (especially for journaled data mode), so be
2010          * cautious in our locking. */
2011
2012         head = bh = page_buffers(page);
2013         do {
2014                 unsigned int next_off = curr_off + bh->b_size;
2015                 next = bh->b_this_page;
2016
2017                 if (offset <= curr_off) {
2018                         /* This block is wholly outside the truncation point */
2019                         lock_buffer(bh);
2020                         may_free &= journal_unmap_buffer(journal, bh);
2021                         unlock_buffer(bh);
2022                 }
2023                 curr_off = next_off;
2024                 bh = next;
2025
2026         } while (bh != head);
2027
2028         if (!offset) {
2029                 if (may_free && try_to_free_buffers(page))
2030                         J_ASSERT(!page_has_buffers(page));
2031         }
2032 }
2033
2034 /*
2035  * File a buffer on the given transaction list.
2036  */
2037 void __journal_file_buffer(struct journal_head *jh,
2038                         transaction_t *transaction, int jlist)
2039 {
2040         struct journal_head **list = NULL;
2041         int was_dirty = 0;
2042         struct buffer_head *bh = jh2bh(jh);
2043
2044         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2045         assert_spin_locked(&transaction->t_journal->j_list_lock);
2046
2047         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2048         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2049                                 jh->b_transaction == NULL);
2050
2051         if (jh->b_transaction && jh->b_jlist == jlist)
2052                 return;
2053
2054         /* The following list of buffer states needs to be consistent
2055          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
2056          * state. */
2057
2058         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2059             jlist == BJ_Shadow || jlist == BJ_Forget) {
2060                 if (test_clear_buffer_dirty(bh) ||
2061                     test_clear_buffer_jbddirty(bh))
2062                         was_dirty = 1;
2063         }
2064
2065         if (jh->b_transaction)
2066                 __journal_temp_unlink_buffer(jh);
2067         jh->b_transaction = transaction;
2068
2069         switch (jlist) {
2070         case BJ_None:
2071                 J_ASSERT_JH(jh, !jh->b_committed_data);
2072                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2073                 return;
2074         case BJ_SyncData:
2075                 list = &transaction->t_sync_datalist;
2076                 break;
2077         case BJ_Metadata:
2078                 transaction->t_nr_buffers++;
2079                 list = &transaction->t_buffers;
2080                 break;
2081         case BJ_Forget:
2082                 list = &transaction->t_forget;
2083                 break;
2084         case BJ_IO:
2085                 list = &transaction->t_iobuf_list;
2086                 break;
2087         case BJ_Shadow:
2088                 list = &transaction->t_shadow_list;
2089                 break;
2090         case BJ_LogCtl:
2091                 list = &transaction->t_log_list;
2092                 break;
2093         case BJ_Reserved:
2094                 list = &transaction->t_reserved_list;
2095                 break;
2096         case BJ_Locked:
2097                 list =  &transaction->t_locked_list;
2098                 break;
2099         }
2100
2101         __blist_add_buffer(list, jh);
2102         jh->b_jlist = jlist;
2103
2104         if (was_dirty)
2105                 set_buffer_jbddirty(bh);
2106 }
2107
2108 void journal_file_buffer(struct journal_head *jh,
2109                                 transaction_t *transaction, int jlist)
2110 {
2111         jbd_lock_bh_state(jh2bh(jh));
2112         spin_lock(&transaction->t_journal->j_list_lock);
2113         __journal_file_buffer(jh, transaction, jlist);
2114         spin_unlock(&transaction->t_journal->j_list_lock);
2115         jbd_unlock_bh_state(jh2bh(jh));
2116 }
2117
2118 /*
2119  * Remove a buffer from its current buffer list in preparation for
2120  * dropping it from its current transaction entirely.  If the buffer has
2121  * already started to be used by a subsequent transaction, refile the
2122  * buffer on that transaction's metadata list.
2123  *
2124  * Called under journal->j_list_lock
2125  *
2126  * Called under jbd_lock_bh_state(jh2bh(jh))
2127  */
2128 void __journal_refile_buffer(struct journal_head *jh)
2129 {
2130         int was_dirty;
2131         struct buffer_head *bh = jh2bh(jh);
2132
2133         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2134         if (jh->b_transaction)
2135                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2136
2137         /* If the buffer is now unused, just drop it. */
2138         if (jh->b_next_transaction == NULL) {
2139                 __journal_unfile_buffer(jh);
2140                 return;
2141         }
2142
2143         /*
2144          * It has been modified by a later transaction: add it to the new
2145          * transaction's metadata list.
2146          */
2147
2148         was_dirty = test_clear_buffer_jbddirty(bh);
2149         __journal_temp_unlink_buffer(jh);
2150         jh->b_transaction = jh->b_next_transaction;
2151         jh->b_next_transaction = NULL;
2152         __journal_file_buffer(jh, jh->b_transaction,
2153                                 jh->b_modified ? BJ_Metadata : BJ_Reserved);
2154         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2155
2156         if (was_dirty)
2157                 set_buffer_jbddirty(bh);
2158 }
2159
2160 /*
2161  * For the unlocked version of this call, also make sure that any
2162  * hanging journal_head is cleaned up if necessary.
2163  *
2164  * __journal_refile_buffer is usually called as part of a single locked
2165  * operation on a buffer_head, in which the caller is probably going to
2166  * be hooking the journal_head onto other lists.  In that case it is up
2167  * to the caller to remove the journal_head if necessary.  For the
2168  * unlocked journal_refile_buffer call, the caller isn't going to be
2169  * doing anything else to the buffer so we need to do the cleanup
2170  * ourselves to avoid a jh leak.
2171  *
2172  * *** The journal_head may be freed by this call! ***
2173  */
2174 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2175 {
2176         struct buffer_head *bh = jh2bh(jh);
2177
2178         jbd_lock_bh_state(bh);
2179         spin_lock(&journal->j_list_lock);
2180
2181         __journal_refile_buffer(jh);
2182         jbd_unlock_bh_state(bh);
2183         journal_remove_journal_head(bh);
2184
2185         spin_unlock(&journal->j_list_lock);
2186         __brelse(bh);
2187 }