Pull sem2mutex-ioc4 into release branch
[linux-2.6] / fs / jbd / transaction.c
1 /*
2  * linux/fs/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/smp_lock.h>
27 #include <linux/mm.h>
28 #include <linux/highmem.h>
29
30 /*
31  * get_transaction: obtain a new transaction_t object.
32  *
33  * Simply allocate and initialise a new transaction.  Create it in
34  * RUNNING state and add it to the current journal (which should not
35  * have an existing running transaction: we only make a new transaction
36  * once we have started to commit the old one).
37  *
38  * Preconditions:
39  *      The journal MUST be locked.  We don't perform atomic mallocs on the
40  *      new transaction and we can't block without protecting against other
41  *      processes trying to touch the journal while it is in transition.
42  *
43  * Called under j_state_lock
44  */
45
46 static transaction_t *
47 get_transaction(journal_t *journal, transaction_t *transaction)
48 {
49         transaction->t_journal = journal;
50         transaction->t_state = T_RUNNING;
51         transaction->t_tid = journal->j_transaction_sequence++;
52         transaction->t_expires = jiffies + journal->j_commit_interval;
53         spin_lock_init(&transaction->t_handle_lock);
54
55         /* Set up the commit timer for the new transaction. */
56         journal->j_commit_timer->expires = transaction->t_expires;
57         add_timer(journal->j_commit_timer);
58
59         J_ASSERT(journal->j_running_transaction == NULL);
60         journal->j_running_transaction = transaction;
61
62         return transaction;
63 }
64
65 /*
66  * Handle management.
67  *
68  * A handle_t is an object which represents a single atomic update to a
69  * filesystem, and which tracks all of the modifications which form part
70  * of that one update.
71  */
72
73 /*
74  * start_this_handle: Given a handle, deal with any locking or stalling
75  * needed to make sure that there is enough journal space for the handle
76  * to begin.  Attach the handle to a transaction and set up the
77  * transaction's buffer credits.  
78  */
79
80 static int start_this_handle(journal_t *journal, handle_t *handle)
81 {
82         transaction_t *transaction;
83         int needed;
84         int nblocks = handle->h_buffer_credits;
85         transaction_t *new_transaction = NULL;
86         int ret = 0;
87
88         if (nblocks > journal->j_max_transaction_buffers) {
89                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
90                        current->comm, nblocks,
91                        journal->j_max_transaction_buffers);
92                 ret = -ENOSPC;
93                 goto out;
94         }
95
96 alloc_transaction:
97         if (!journal->j_running_transaction) {
98                 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
99                                                 GFP_NOFS);
100                 if (!new_transaction) {
101                         ret = -ENOMEM;
102                         goto out;
103                 }
104                 memset(new_transaction, 0, sizeof(*new_transaction));
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         kfree(new_transaction);
231         return ret;
232 }
233
234 /* Allocate a new handle.  This should probably be in a slab... */
235 static handle_t *new_handle(int nblocks)
236 {
237         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
238         if (!handle)
239                 return NULL;
240         memset(handle, 0, sizeof(*handle));
241         handle->h_buffer_credits = nblocks;
242         handle->h_ref = 1;
243
244         return handle;
245 }
246
247 /**
248  * handle_t *journal_start() - Obtain a new handle.  
249  * @journal: Journal to start transaction on.
250  * @nblocks: number of block buffer we might modify
251  *
252  * We make sure that the transaction can guarantee at least nblocks of
253  * modified buffers in the log.  We block until the log can guarantee
254  * that much space.  
255  *
256  * This function is visible to journal users (like ext3fs), so is not
257  * called with the journal already locked.
258  *
259  * Return a pointer to a newly allocated handle, or NULL on failure
260  */
261 handle_t *journal_start(journal_t *journal, int nblocks)
262 {
263         handle_t *handle = journal_current_handle();
264         int err;
265
266         if (!journal)
267                 return ERR_PTR(-EROFS);
268
269         if (handle) {
270                 J_ASSERT(handle->h_transaction->t_journal == journal);
271                 handle->h_ref++;
272                 return handle;
273         }
274
275         handle = new_handle(nblocks);
276         if (!handle)
277                 return ERR_PTR(-ENOMEM);
278
279         current->journal_info = handle;
280
281         err = start_this_handle(journal, handle);
282         if (err < 0) {
283                 jbd_free_handle(handle);
284                 current->journal_info = NULL;
285                 handle = ERR_PTR(err);
286         }
287         return handle;
288 }
289
290 /**
291  * int journal_extend() - extend buffer credits.
292  * @handle:  handle to 'extend'
293  * @nblocks: nr blocks to try to extend by.
294  * 
295  * Some transactions, such as large extends and truncates, can be done
296  * atomically all at once or in several stages.  The operation requests
297  * a credit for a number of buffer modications in advance, but can
298  * extend its credit if it needs more.  
299  *
300  * journal_extend tries to give the running handle more buffer credits.
301  * It does not guarantee that allocation - this is a best-effort only.
302  * The calling process MUST be able to deal cleanly with a failure to
303  * extend here.
304  *
305  * Return 0 on success, non-zero on failure.
306  *
307  * return code < 0 implies an error
308  * return code > 0 implies normal transaction-full status.
309  */
310 int journal_extend(handle_t *handle, int nblocks)
311 {
312         transaction_t *transaction = handle->h_transaction;
313         journal_t *journal = transaction->t_journal;
314         int result;
315         int wanted;
316
317         result = -EIO;
318         if (is_handle_aborted(handle))
319                 goto out;
320
321         result = 1;
322
323         spin_lock(&journal->j_state_lock);
324
325         /* Don't extend a locked-down transaction! */
326         if (handle->h_transaction->t_state != T_RUNNING) {
327                 jbd_debug(3, "denied handle %p %d blocks: "
328                           "transaction not running\n", handle, nblocks);
329                 goto error_out;
330         }
331
332         spin_lock(&transaction->t_handle_lock);
333         wanted = transaction->t_outstanding_credits + nblocks;
334
335         if (wanted > journal->j_max_transaction_buffers) {
336                 jbd_debug(3, "denied handle %p %d blocks: "
337                           "transaction too large\n", handle, nblocks);
338                 goto unlock;
339         }
340
341         if (wanted > __log_space_left(journal)) {
342                 jbd_debug(3, "denied handle %p %d blocks: "
343                           "insufficient log space\n", handle, nblocks);
344                 goto unlock;
345         }
346
347         handle->h_buffer_credits += nblocks;
348         transaction->t_outstanding_credits += nblocks;
349         result = 0;
350
351         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
352 unlock:
353         spin_unlock(&transaction->t_handle_lock);
354 error_out:
355         spin_unlock(&journal->j_state_lock);
356 out:
357         return result;
358 }
359
360
361 /**
362  * int journal_restart() - restart a handle .
363  * @handle:  handle to restart
364  * @nblocks: nr credits requested
365  * 
366  * Restart a handle for a multi-transaction filesystem
367  * operation.
368  *
369  * If the journal_extend() call above fails to grant new buffer credits
370  * to a running handle, a call to journal_restart will commit the
371  * handle's transaction so far and reattach the handle to a new
372  * transaction capabable of guaranteeing the requested number of
373  * credits.
374  */
375
376 int journal_restart(handle_t *handle, int nblocks)
377 {
378         transaction_t *transaction = handle->h_transaction;
379         journal_t *journal = transaction->t_journal;
380         int ret;
381
382         /* If we've had an abort of any type, don't even think about
383          * actually doing the restart! */
384         if (is_handle_aborted(handle))
385                 return 0;
386
387         /*
388          * First unlink the handle from its current transaction, and start the
389          * commit on that.
390          */
391         J_ASSERT(transaction->t_updates > 0);
392         J_ASSERT(journal_current_handle() == handle);
393
394         spin_lock(&journal->j_state_lock);
395         spin_lock(&transaction->t_handle_lock);
396         transaction->t_outstanding_credits -= handle->h_buffer_credits;
397         transaction->t_updates--;
398
399         if (!transaction->t_updates)
400                 wake_up(&journal->j_wait_updates);
401         spin_unlock(&transaction->t_handle_lock);
402
403         jbd_debug(2, "restarting handle %p\n", handle);
404         __log_start_commit(journal, transaction->t_tid);
405         spin_unlock(&journal->j_state_lock);
406
407         handle->h_buffer_credits = nblocks;
408         ret = start_this_handle(journal, handle);
409         return ret;
410 }
411
412
413 /**
414  * void journal_lock_updates () - establish a transaction barrier.
415  * @journal:  Journal to establish a barrier on.
416  *
417  * This locks out any further updates from being started, and blocks
418  * until all existing updates have completed, returning only once the
419  * journal is in a quiescent state with no updates running.
420  *
421  * The journal lock should not be held on entry.
422  */
423 void journal_lock_updates(journal_t *journal)
424 {
425         DEFINE_WAIT(wait);
426
427         spin_lock(&journal->j_state_lock);
428         ++journal->j_barrier_count;
429
430         /* Wait until there are no running updates */
431         while (1) {
432                 transaction_t *transaction = journal->j_running_transaction;
433
434                 if (!transaction)
435                         break;
436
437                 spin_lock(&transaction->t_handle_lock);
438                 if (!transaction->t_updates) {
439                         spin_unlock(&transaction->t_handle_lock);
440                         break;
441                 }
442                 prepare_to_wait(&journal->j_wait_updates, &wait,
443                                 TASK_UNINTERRUPTIBLE);
444                 spin_unlock(&transaction->t_handle_lock);
445                 spin_unlock(&journal->j_state_lock);
446                 schedule();
447                 finish_wait(&journal->j_wait_updates, &wait);
448                 spin_lock(&journal->j_state_lock);
449         }
450         spin_unlock(&journal->j_state_lock);
451
452         /*
453          * We have now established a barrier against other normal updates, but
454          * we also need to barrier against other journal_lock_updates() calls
455          * to make sure that we serialise special journal-locked operations
456          * too.
457          */
458         down(&journal->j_barrier);
459 }
460
461 /**
462  * void journal_unlock_updates (journal_t* journal) - release barrier
463  * @journal:  Journal to release the barrier on.
464  * 
465  * Release a transaction barrier obtained with journal_lock_updates().
466  *
467  * Should be called without the journal lock held.
468  */
469 void journal_unlock_updates (journal_t *journal)
470 {
471         J_ASSERT(journal->j_barrier_count != 0);
472
473         up(&journal->j_barrier);
474         spin_lock(&journal->j_state_lock);
475         --journal->j_barrier_count;
476         spin_unlock(&journal->j_state_lock);
477         wake_up(&journal->j_wait_transaction_locked);
478 }
479
480 /*
481  * Report any unexpected dirty buffers which turn up.  Normally those
482  * indicate an error, but they can occur if the user is running (say)
483  * tune2fs to modify the live filesystem, so we need the option of
484  * continuing as gracefully as possible.  #
485  *
486  * The caller should already hold the journal lock and
487  * j_list_lock spinlock: most callers will need those anyway
488  * in order to probe the buffer's journaling state safely.
489  */
490 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
491 {
492         int jlist;
493
494         /* If this buffer is one which might reasonably be dirty
495          * --- ie. data, or not part of this journal --- then
496          * we're OK to leave it alone, but otherwise we need to
497          * move the dirty bit to the journal's own internal
498          * JBDDirty bit. */
499         jlist = jh->b_jlist;
500
501         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
502             jlist == BJ_Shadow || jlist == BJ_Forget) {
503                 struct buffer_head *bh = jh2bh(jh);
504
505                 if (test_clear_buffer_dirty(bh))
506                         set_buffer_jbddirty(bh);
507         }
508 }
509
510 /*
511  * If the buffer is already part of the current transaction, then there
512  * is nothing we need to do.  If it is already part of a prior
513  * transaction which we are still committing to disk, then we need to
514  * make sure that we do not overwrite the old copy: we do copy-out to
515  * preserve the copy going to disk.  We also account the buffer against
516  * the handle's metadata buffer credits (unless the buffer is already
517  * part of the transaction, that is).
518  *
519  */
520 static int
521 do_get_write_access(handle_t *handle, struct journal_head *jh,
522                         int force_copy)
523 {
524         struct buffer_head *bh;
525         transaction_t *transaction;
526         journal_t *journal;
527         int error;
528         char *frozen_buffer = NULL;
529         int need_copy = 0;
530
531         if (is_handle_aborted(handle))
532                 return -EROFS;
533
534         transaction = handle->h_transaction;
535         journal = transaction->t_journal;
536
537         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
538
539         JBUFFER_TRACE(jh, "entry");
540 repeat:
541         bh = jh2bh(jh);
542
543         /* @@@ Need to check for errors here at some point. */
544
545         lock_buffer(bh);
546         jbd_lock_bh_state(bh);
547
548         /* We now hold the buffer lock so it is safe to query the buffer
549          * state.  Is the buffer dirty? 
550          * 
551          * If so, there are two possibilities.  The buffer may be
552          * non-journaled, and undergoing a quite legitimate writeback.
553          * Otherwise, it is journaled, and we don't expect dirty buffers
554          * in that state (the buffers should be marked JBD_Dirty
555          * instead.)  So either the IO is being done under our own
556          * control and this is a bug, or it's a third party IO such as
557          * dump(8) (which may leave the buffer scheduled for read ---
558          * ie. locked but not dirty) or tune2fs (which may actually have
559          * the buffer dirtied, ugh.)  */
560
561         if (buffer_dirty(bh)) {
562                 /*
563                  * First question: is this buffer already part of the current
564                  * transaction or the existing committing transaction?
565                  */
566                 if (jh->b_transaction) {
567                         J_ASSERT_JH(jh,
568                                 jh->b_transaction == transaction || 
569                                 jh->b_transaction ==
570                                         journal->j_committing_transaction);
571                         if (jh->b_next_transaction)
572                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
573                                                         transaction);
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, "Unexpected dirty buffer");
581                 jbd_unexpected_dirty_buffer(jh);
582         }
583
584         unlock_buffer(bh);
585
586         error = -EROFS;
587         if (is_handle_aborted(handle)) {
588                 jbd_unlock_bh_state(bh);
589                 goto out;
590         }
591         error = 0;
592
593         /*
594          * The buffer is already part of this transaction if b_transaction or
595          * b_next_transaction points to it
596          */
597         if (jh->b_transaction == transaction ||
598             jh->b_next_transaction == transaction)
599                 goto done;
600
601         /*
602          * If there is already a copy-out version of this buffer, then we don't
603          * need to make another one
604          */
605         if (jh->b_frozen_data) {
606                 JBUFFER_TRACE(jh, "has frozen data");
607                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
608                 jh->b_next_transaction = transaction;
609                 goto done;
610         }
611
612         /* Is there data here we need to preserve? */
613
614         if (jh->b_transaction && jh->b_transaction != transaction) {
615                 JBUFFER_TRACE(jh, "owned by older transaction");
616                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
617                 J_ASSERT_JH(jh, jh->b_transaction ==
618                                         journal->j_committing_transaction);
619
620                 /* There is one case we have to be very careful about.
621                  * If the committing transaction is currently writing
622                  * this buffer out to disk and has NOT made a copy-out,
623                  * then we cannot modify the buffer contents at all
624                  * right now.  The essence of copy-out is that it is the
625                  * extra copy, not the primary copy, which gets
626                  * journaled.  If the primary copy is already going to
627                  * disk then we cannot do copy-out here. */
628
629                 if (jh->b_jlist == BJ_Shadow) {
630                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
631                         wait_queue_head_t *wqh;
632
633                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
634
635                         JBUFFER_TRACE(jh, "on shadow: sleep");
636                         jbd_unlock_bh_state(bh);
637                         /* commit wakes up all shadow buffers after IO */
638                         for ( ; ; ) {
639                                 prepare_to_wait(wqh, &wait.wait,
640                                                 TASK_UNINTERRUPTIBLE);
641                                 if (jh->b_jlist != BJ_Shadow)
642                                         break;
643                                 schedule();
644                         }
645                         finish_wait(wqh, &wait.wait);
646                         goto repeat;
647                 }
648
649                 /* Only do the copy if the currently-owning transaction
650                  * still needs it.  If it is on the Forget list, the
651                  * committing transaction is past that stage.  The
652                  * buffer had better remain locked during the kmalloc,
653                  * but that should be true --- we hold the journal lock
654                  * still and the buffer is already on the BUF_JOURNAL
655                  * list so won't be flushed. 
656                  *
657                  * Subtle point, though: if this is a get_undo_access,
658                  * then we will be relying on the frozen_data to contain
659                  * the new value of the committed_data record after the
660                  * transaction, so we HAVE to force the frozen_data copy
661                  * in that case. */
662
663                 if (jh->b_jlist != BJ_Forget || force_copy) {
664                         JBUFFER_TRACE(jh, "generate frozen data");
665                         if (!frozen_buffer) {
666                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
667                                 jbd_unlock_bh_state(bh);
668                                 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
669                                                             GFP_NOFS);
670                                 if (!frozen_buffer) {
671                                         printk(KERN_EMERG
672                                                "%s: OOM for frozen_buffer\n",
673                                                __FUNCTION__);
674                                         JBUFFER_TRACE(jh, "oom!");
675                                         error = -ENOMEM;
676                                         jbd_lock_bh_state(bh);
677                                         goto done;
678                                 }
679                                 goto repeat;
680                         }
681                         jh->b_frozen_data = frozen_buffer;
682                         frozen_buffer = NULL;
683                         need_copy = 1;
684                 }
685                 jh->b_next_transaction = transaction;
686         }
687
688
689         /*
690          * Finally, if the buffer is not journaled right now, we need to make
691          * sure it doesn't get written to disk before the caller actually
692          * commits the new data
693          */
694         if (!jh->b_transaction) {
695                 JBUFFER_TRACE(jh, "no transaction");
696                 J_ASSERT_JH(jh, !jh->b_next_transaction);
697                 jh->b_transaction = transaction;
698                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
699                 spin_lock(&journal->j_list_lock);
700                 __journal_file_buffer(jh, transaction, BJ_Reserved);
701                 spin_unlock(&journal->j_list_lock);
702         }
703
704 done:
705         if (need_copy) {
706                 struct page *page;
707                 int offset;
708                 char *source;
709
710                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
711                             "Possible IO failure.\n");
712                 page = jh2bh(jh)->b_page;
713                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
714                 source = kmap_atomic(page, KM_USER0);
715                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
716                 kunmap_atomic(source, KM_USER0);
717         }
718         jbd_unlock_bh_state(bh);
719
720         /*
721          * If we are about to journal a buffer, then any revoke pending on it is
722          * no longer valid
723          */
724         journal_cancel_revoke(handle, jh);
725
726 out:
727         kfree(frozen_buffer);
728
729         JBUFFER_TRACE(jh, "exit");
730         return error;
731 }
732
733 /**
734  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
735  * @handle: transaction to add buffer modifications to
736  * @bh:     bh to be used for metadata writes
737  * @credits: variable that will receive credits for the buffer
738  *
739  * Returns an error code or 0 on success.
740  *
741  * In full data journalling mode the buffer may be of type BJ_AsyncData,
742  * because we're write()ing a buffer which is also part of a shared mapping.
743  */
744
745 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
746 {
747         struct journal_head *jh = journal_add_journal_head(bh);
748         int rc;
749
750         /* We do not want to get caught playing with fields which the
751          * log thread also manipulates.  Make sure that the buffer
752          * completes any outstanding IO before proceeding. */
753         rc = do_get_write_access(handle, jh, 0);
754         journal_put_journal_head(jh);
755         return rc;
756 }
757
758
759 /*
760  * When the user wants to journal a newly created buffer_head
761  * (ie. getblk() returned a new buffer and we are going to populate it
762  * manually rather than reading off disk), then we need to keep the
763  * buffer_head locked until it has been completely filled with new
764  * data.  In this case, we should be able to make the assertion that
765  * the bh is not already part of an existing transaction.  
766  * 
767  * The buffer should already be locked by the caller by this point.
768  * There is no lock ranking violation: it was a newly created,
769  * unlocked buffer beforehand. */
770
771 /**
772  * int journal_get_create_access () - notify intent to use newly created bh
773  * @handle: transaction to new buffer to
774  * @bh: new buffer.
775  *
776  * Call this if you create a new bh.
777  */
778 int journal_get_create_access(handle_t *handle, struct buffer_head *bh) 
779 {
780         transaction_t *transaction = handle->h_transaction;
781         journal_t *journal = transaction->t_journal;
782         struct journal_head *jh = journal_add_journal_head(bh);
783         int err;
784
785         jbd_debug(5, "journal_head %p\n", jh);
786         err = -EROFS;
787         if (is_handle_aborted(handle))
788                 goto out;
789         err = 0;
790
791         JBUFFER_TRACE(jh, "entry");
792         /*
793          * The buffer may already belong to this transaction due to pre-zeroing
794          * in the filesystem's new_block code.  It may also be on the previous,
795          * committing transaction's lists, but it HAS to be in Forget state in
796          * that case: the transaction must have deleted the buffer for it to be
797          * reused here.
798          */
799         jbd_lock_bh_state(bh);
800         spin_lock(&journal->j_list_lock);
801         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
802                 jh->b_transaction == NULL ||
803                 (jh->b_transaction == journal->j_committing_transaction &&
804                           jh->b_jlist == BJ_Forget)));
805
806         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
807         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
808
809         if (jh->b_transaction == NULL) {
810                 jh->b_transaction = transaction;
811                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
812                 __journal_file_buffer(jh, transaction, BJ_Reserved);
813         } else if (jh->b_transaction == journal->j_committing_transaction) {
814                 JBUFFER_TRACE(jh, "set next transaction");
815                 jh->b_next_transaction = transaction;
816         }
817         spin_unlock(&journal->j_list_lock);
818         jbd_unlock_bh_state(bh);
819
820         /*
821          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
822          * blocks which contain freed but then revoked metadata.  We need
823          * to cancel the revoke in case we end up freeing it yet again
824          * and the reallocating as data - this would cause a second revoke,
825          * which hits an assertion error.
826          */
827         JBUFFER_TRACE(jh, "cancelling revoke");
828         journal_cancel_revoke(handle, jh);
829         journal_put_journal_head(jh);
830 out:
831         return err;
832 }
833
834 /**
835  * int journal_get_undo_access() -  Notify intent to modify metadata with
836  *     non-rewindable consequences
837  * @handle: transaction
838  * @bh: buffer to undo
839  * @credits: store the number of taken credits here (if not NULL)
840  *
841  * Sometimes there is a need to distinguish between metadata which has
842  * been committed to disk and that which has not.  The ext3fs code uses
843  * this for freeing and allocating space, we have to make sure that we
844  * do not reuse freed space until the deallocation has been committed,
845  * since if we overwrote that space we would make the delete
846  * un-rewindable in case of a crash.
847  * 
848  * To deal with that, journal_get_undo_access requests write access to a
849  * buffer for parts of non-rewindable operations such as delete
850  * operations on the bitmaps.  The journaling code must keep a copy of
851  * the buffer's contents prior to the undo_access call until such time
852  * as we know that the buffer has definitely been committed to disk.
853  * 
854  * We never need to know which transaction the committed data is part
855  * of, buffers touched here are guaranteed to be dirtied later and so
856  * will be committed to a new transaction in due course, at which point
857  * we can discard the old committed data pointer.
858  *
859  * Returns error number or 0 on success.
860  */
861 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
862 {
863         int err;
864         struct journal_head *jh = journal_add_journal_head(bh);
865         char *committed_data = NULL;
866
867         JBUFFER_TRACE(jh, "entry");
868
869         /*
870          * Do this first --- it can drop the journal lock, so we want to
871          * make sure that obtaining the committed_data is done
872          * atomically wrt. completion of any outstanding commits.
873          */
874         err = do_get_write_access(handle, jh, 1);
875         if (err)
876                 goto out;
877
878 repeat:
879         if (!jh->b_committed_data) {
880                 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
881                 if (!committed_data) {
882                         printk(KERN_EMERG "%s: No memory for committed data\n",
883                                 __FUNCTION__);
884                         err = -ENOMEM;
885                         goto out;
886                 }
887         }
888
889         jbd_lock_bh_state(bh);
890         if (!jh->b_committed_data) {
891                 /* Copy out the current buffer contents into the
892                  * preserved, committed copy. */
893                 JBUFFER_TRACE(jh, "generate b_committed data");
894                 if (!committed_data) {
895                         jbd_unlock_bh_state(bh);
896                         goto repeat;
897                 }
898
899                 jh->b_committed_data = committed_data;
900                 committed_data = NULL;
901                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
902         }
903         jbd_unlock_bh_state(bh);
904 out:
905         journal_put_journal_head(jh);
906         kfree(committed_data);
907         return err;
908 }
909
910 /** 
911  * int journal_dirty_data() -  mark a buffer as containing dirty data which
912  *                             needs to be flushed before we can commit the
913  *                             current transaction.  
914  * @handle: transaction
915  * @bh: bufferhead to mark
916  * 
917  * The buffer is placed on the transaction's data list and is marked as
918  * belonging to the transaction.
919  *
920  * Returns error number or 0 on success.
921  *
922  * journal_dirty_data() can be called via page_launder->ext3_writepage
923  * by kswapd.
924  */
925 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
926 {
927         journal_t *journal = handle->h_transaction->t_journal;
928         int need_brelse = 0;
929         struct journal_head *jh;
930
931         if (is_handle_aborted(handle))
932                 return 0;
933
934         jh = journal_add_journal_head(bh);
935         JBUFFER_TRACE(jh, "entry");
936
937         /*
938          * The buffer could *already* be dirty.  Writeout can start
939          * at any time.
940          */
941         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
942
943         /*
944          * What if the buffer is already part of a running transaction?
945          * 
946          * There are two cases:
947          * 1) It is part of the current running transaction.  Refile it,
948          *    just in case we have allocated it as metadata, deallocated
949          *    it, then reallocated it as data. 
950          * 2) It is part of the previous, still-committing transaction.
951          *    If all we want to do is to guarantee that the buffer will be
952          *    written to disk before this new transaction commits, then
953          *    being sure that the *previous* transaction has this same 
954          *    property is sufficient for us!  Just leave it on its old
955          *    transaction.
956          *
957          * In case (2), the buffer must not already exist as metadata
958          * --- that would violate write ordering (a transaction is free
959          * to write its data at any point, even before the previous
960          * committing transaction has committed).  The caller must
961          * never, ever allow this to happen: there's nothing we can do
962          * about it in this layer.
963          */
964         jbd_lock_bh_state(bh);
965         spin_lock(&journal->j_list_lock);
966         if (jh->b_transaction) {
967                 JBUFFER_TRACE(jh, "has transaction");
968                 if (jh->b_transaction != handle->h_transaction) {
969                         JBUFFER_TRACE(jh, "belongs to older transaction");
970                         J_ASSERT_JH(jh, jh->b_transaction ==
971                                         journal->j_committing_transaction);
972
973                         /* @@@ IS THIS TRUE  ? */
974                         /*
975                          * Not any more.  Scenario: someone does a write()
976                          * in data=journal mode.  The buffer's transaction has
977                          * moved into commit.  Then someone does another
978                          * write() to the file.  We do the frozen data copyout
979                          * and set b_next_transaction to point to j_running_t.
980                          * And while we're in that state, someone does a
981                          * writepage() in an attempt to pageout the same area
982                          * of the file via a shared mapping.  At present that
983                          * calls journal_dirty_data(), and we get right here.
984                          * It may be too late to journal the data.  Simply
985                          * falling through to the next test will suffice: the
986                          * data will be dirty and wil be checkpointed.  The
987                          * ordering comments in the next comment block still
988                          * apply.
989                          */
990                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
991
992                         /*
993                          * If we're journalling data, and this buffer was
994                          * subject to a write(), it could be metadata, forget
995                          * or shadow against the committing transaction.  Now,
996                          * someone has dirtied the same darn page via a mapping
997                          * and it is being writepage()'d.
998                          * We *could* just steal the page from commit, with some
999                          * fancy locking there.  Instead, we just skip it -
1000                          * don't tie the page's buffers to the new transaction
1001                          * at all.
1002                          * Implication: if we crash before the writepage() data
1003                          * is written into the filesystem, recovery will replay
1004                          * the write() data.
1005                          */
1006                         if (jh->b_jlist != BJ_None &&
1007                                         jh->b_jlist != BJ_SyncData &&
1008                                         jh->b_jlist != BJ_Locked) {
1009                                 JBUFFER_TRACE(jh, "Not stealing");
1010                                 goto no_journal;
1011                         }
1012
1013                         /*
1014                          * This buffer may be undergoing writeout in commit.  We
1015                          * can't return from here and let the caller dirty it
1016                          * again because that can cause the write-out loop in
1017                          * commit to never terminate.
1018                          */
1019                         if (buffer_dirty(bh)) {
1020                                 get_bh(bh);
1021                                 spin_unlock(&journal->j_list_lock);
1022                                 jbd_unlock_bh_state(bh);
1023                                 need_brelse = 1;
1024                                 sync_dirty_buffer(bh);
1025                                 jbd_lock_bh_state(bh);
1026                                 spin_lock(&journal->j_list_lock);
1027                                 /* The buffer may become locked again at any
1028                                    time if it is redirtied */
1029                         }
1030
1031                         /* journal_clean_data_list() may have got there first */
1032                         if (jh->b_transaction != NULL) {
1033                                 JBUFFER_TRACE(jh, "unfile from commit");
1034                                 __journal_temp_unlink_buffer(jh);
1035                                 /* It still points to the committing
1036                                  * transaction; move it to this one so
1037                                  * that the refile assert checks are
1038                                  * happy. */
1039                                 jh->b_transaction = handle->h_transaction;
1040                         }
1041                         /* The buffer will be refiled below */
1042
1043                 }
1044                 /*
1045                  * Special case --- the buffer might actually have been
1046                  * allocated and then immediately deallocated in the previous,
1047                  * committing transaction, so might still be left on that
1048                  * transaction's metadata lists.
1049                  */
1050                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1051                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1052                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1053                         __journal_temp_unlink_buffer(jh);
1054                         jh->b_transaction = handle->h_transaction;
1055                         JBUFFER_TRACE(jh, "file as data");
1056                         __journal_file_buffer(jh, handle->h_transaction,
1057                                                 BJ_SyncData);
1058                 }
1059         } else {
1060                 JBUFFER_TRACE(jh, "not on a transaction");
1061                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1062         }
1063 no_journal:
1064         spin_unlock(&journal->j_list_lock);
1065         jbd_unlock_bh_state(bh);
1066         if (need_brelse) {
1067                 BUFFER_TRACE(bh, "brelse");
1068                 __brelse(bh);
1069         }
1070         JBUFFER_TRACE(jh, "exit");
1071         journal_put_journal_head(jh);
1072         return 0;
1073 }
1074
1075 /** 
1076  * int journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1077  * @handle: transaction to add buffer to.
1078  * @bh: buffer to mark 
1079  * 
1080  * mark dirty metadata which needs to be journaled as part of the current
1081  * transaction.
1082  *
1083  * The buffer is placed on the transaction's metadata list and is marked
1084  * as belonging to the transaction.  
1085  *
1086  * Returns error number or 0 on success.  
1087  *
1088  * Special care needs to be taken if the buffer already belongs to the
1089  * current committing transaction (in which case we should have frozen
1090  * data present for that commit).  In that case, we don't relink the
1091  * buffer: that only gets done when the old transaction finally
1092  * completes its commit.
1093  */
1094 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1095 {
1096         transaction_t *transaction = handle->h_transaction;
1097         journal_t *journal = transaction->t_journal;
1098         struct journal_head *jh = bh2jh(bh);
1099
1100         jbd_debug(5, "journal_head %p\n", jh);
1101         JBUFFER_TRACE(jh, "entry");
1102         if (is_handle_aborted(handle))
1103                 goto out;
1104
1105         jbd_lock_bh_state(bh);
1106
1107         if (jh->b_modified == 0) {
1108                 /*
1109                  * This buffer's got modified and becoming part
1110                  * of the transaction. This needs to be done
1111                  * once a transaction -bzzz
1112                  */
1113                 jh->b_modified = 1;
1114                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1115                 handle->h_buffer_credits--;
1116         }
1117
1118         /*
1119          * fastpath, to avoid expensive locking.  If this buffer is already
1120          * on the running transaction's metadata list there is nothing to do.
1121          * Nobody can take it off again because there is a handle open.
1122          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1123          * result in this test being false, so we go in and take the locks.
1124          */
1125         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1126                 JBUFFER_TRACE(jh, "fastpath");
1127                 J_ASSERT_JH(jh, jh->b_transaction ==
1128                                         journal->j_running_transaction);
1129                 goto out_unlock_bh;
1130         }
1131
1132         set_buffer_jbddirty(bh);
1133
1134         /* 
1135          * Metadata already on the current transaction list doesn't
1136          * need to be filed.  Metadata on another transaction's list must
1137          * be committing, and will be refiled once the commit completes:
1138          * leave it alone for now. 
1139          */
1140         if (jh->b_transaction != transaction) {
1141                 JBUFFER_TRACE(jh, "already on other transaction");
1142                 J_ASSERT_JH(jh, jh->b_transaction ==
1143                                         journal->j_committing_transaction);
1144                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1145                 /* And this case is illegal: we can't reuse another
1146                  * transaction's data buffer, ever. */
1147                 goto out_unlock_bh;
1148         }
1149
1150         /* That test should have eliminated the following case: */
1151         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1152
1153         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1154         spin_lock(&journal->j_list_lock);
1155         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1156         spin_unlock(&journal->j_list_lock);
1157 out_unlock_bh:
1158         jbd_unlock_bh_state(bh);
1159 out:
1160         JBUFFER_TRACE(jh, "exit");
1161         return 0;
1162 }
1163
1164 /* 
1165  * journal_release_buffer: undo a get_write_access without any buffer
1166  * updates, if the update decided in the end that it didn't need access.
1167  *
1168  */
1169 void
1170 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1171 {
1172         BUFFER_TRACE(bh, "entry");
1173 }
1174
1175 /** 
1176  * void journal_forget() - bforget() for potentially-journaled buffers.
1177  * @handle: transaction handle
1178  * @bh:     bh to 'forget'
1179  *
1180  * We can only do the bforget if there are no commits pending against the
1181  * buffer.  If the buffer is dirty in the current running transaction we
1182  * can safely unlink it. 
1183  *
1184  * bh may not be a journalled buffer at all - it may be a non-JBD
1185  * buffer which came off the hashtable.  Check for this.
1186  *
1187  * Decrements bh->b_count by one.
1188  * 
1189  * Allow this call even if the handle has aborted --- it may be part of
1190  * the caller's cleanup after an abort.
1191  */
1192 int journal_forget (handle_t *handle, struct buffer_head *bh)
1193 {
1194         transaction_t *transaction = handle->h_transaction;
1195         journal_t *journal = transaction->t_journal;
1196         struct journal_head *jh;
1197         int drop_reserve = 0;
1198         int err = 0;
1199
1200         BUFFER_TRACE(bh, "entry");
1201
1202         jbd_lock_bh_state(bh);
1203         spin_lock(&journal->j_list_lock);
1204
1205         if (!buffer_jbd(bh))
1206                 goto not_jbd;
1207         jh = bh2jh(bh);
1208
1209         /* Critical error: attempting to delete a bitmap buffer, maybe?
1210          * Don't do any jbd operations, and return an error. */
1211         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1212                          "inconsistent data on disk")) {
1213                 err = -EIO;
1214                 goto not_jbd;
1215         }
1216
1217         /*
1218          * The buffer's going from the transaction, we must drop
1219          * all references -bzzz
1220          */
1221         jh->b_modified = 0;
1222
1223         if (jh->b_transaction == handle->h_transaction) {
1224                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1225
1226                 /* If we are forgetting a buffer which is already part
1227                  * of this transaction, then we can just drop it from
1228                  * the transaction immediately. */
1229                 clear_buffer_dirty(bh);
1230                 clear_buffer_jbddirty(bh);
1231
1232                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1233
1234                 drop_reserve = 1;
1235
1236                 /* 
1237                  * We are no longer going to journal this buffer.
1238                  * However, the commit of this transaction is still
1239                  * important to the buffer: the delete that we are now
1240                  * processing might obsolete an old log entry, so by
1241                  * committing, we can satisfy the buffer's checkpoint.
1242                  *
1243                  * So, if we have a checkpoint on the buffer, we should
1244                  * now refile the buffer on our BJ_Forget list so that
1245                  * we know to remove the checkpoint after we commit. 
1246                  */
1247
1248                 if (jh->b_cp_transaction) {
1249                         __journal_temp_unlink_buffer(jh);
1250                         __journal_file_buffer(jh, transaction, BJ_Forget);
1251                 } else {
1252                         __journal_unfile_buffer(jh);
1253                         journal_remove_journal_head(bh);
1254                         __brelse(bh);
1255                         if (!buffer_jbd(bh)) {
1256                                 spin_unlock(&journal->j_list_lock);
1257                                 jbd_unlock_bh_state(bh);
1258                                 __bforget(bh);
1259                                 goto drop;
1260                         }
1261                 }
1262         } else if (jh->b_transaction) {
1263                 J_ASSERT_JH(jh, (jh->b_transaction == 
1264                                  journal->j_committing_transaction));
1265                 /* However, if the buffer is still owned by a prior
1266                  * (committing) transaction, we can't drop it yet... */
1267                 JBUFFER_TRACE(jh, "belongs to older transaction");
1268                 /* ... but we CAN drop it from the new transaction if we
1269                  * have also modified it since the original commit. */
1270
1271                 if (jh->b_next_transaction) {
1272                         J_ASSERT(jh->b_next_transaction == transaction);
1273                         jh->b_next_transaction = NULL;
1274                         drop_reserve = 1;
1275                 }
1276         }
1277
1278 not_jbd:
1279         spin_unlock(&journal->j_list_lock);
1280         jbd_unlock_bh_state(bh);
1281         __brelse(bh);
1282 drop:
1283         if (drop_reserve) {
1284                 /* no need to reserve log space for this block -bzzz */
1285                 handle->h_buffer_credits++;
1286         }
1287         return err;
1288 }
1289
1290 /**
1291  * int journal_stop() - complete a transaction
1292  * @handle: tranaction to complete.
1293  * 
1294  * All done for a particular handle.
1295  *
1296  * There is not much action needed here.  We just return any remaining
1297  * buffer credits to the transaction and remove the handle.  The only
1298  * complication is that we need to start a commit operation if the
1299  * filesystem is marked for synchronous update.
1300  *
1301  * journal_stop itself will not usually return an error, but it may
1302  * do so in unusual circumstances.  In particular, expect it to 
1303  * return -EIO if a journal_abort has been executed since the
1304  * transaction began.
1305  */
1306 int journal_stop(handle_t *handle)
1307 {
1308         transaction_t *transaction = handle->h_transaction;
1309         journal_t *journal = transaction->t_journal;
1310         int old_handle_count, err;
1311         pid_t pid;
1312
1313         J_ASSERT(transaction->t_updates > 0);
1314         J_ASSERT(journal_current_handle() == handle);
1315
1316         if (is_handle_aborted(handle))
1317                 err = -EIO;
1318         else
1319                 err = 0;
1320
1321         if (--handle->h_ref > 0) {
1322                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1323                           handle->h_ref);
1324                 return err;
1325         }
1326
1327         jbd_debug(4, "Handle %p going down\n", handle);
1328
1329         /*
1330          * Implement synchronous transaction batching.  If the handle
1331          * was synchronous, don't force a commit immediately.  Let's
1332          * yield and let another thread piggyback onto this transaction.
1333          * Keep doing that while new threads continue to arrive.
1334          * It doesn't cost much - we're about to run a commit and sleep
1335          * on IO anyway.  Speeds up many-threaded, many-dir operations
1336          * by 30x or more...
1337          *
1338          * But don't do this if this process was the most recent one to
1339          * perform a synchronous write.  We do this to detect the case where a
1340          * single process is doing a stream of sync writes.  No point in waiting
1341          * for joiners in that case.
1342          */
1343         pid = current->pid;
1344         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1345                 journal->j_last_sync_writer = pid;
1346                 do {
1347                         old_handle_count = transaction->t_handle_count;
1348                         schedule_timeout_uninterruptible(1);
1349                 } while (old_handle_count != transaction->t_handle_count);
1350         }
1351
1352         current->journal_info = NULL;
1353         spin_lock(&journal->j_state_lock);
1354         spin_lock(&transaction->t_handle_lock);
1355         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1356         transaction->t_updates--;
1357         if (!transaction->t_updates) {
1358                 wake_up(&journal->j_wait_updates);
1359                 if (journal->j_barrier_count)
1360                         wake_up(&journal->j_wait_transaction_locked);
1361         }
1362
1363         /*
1364          * If the handle is marked SYNC, we need to set another commit
1365          * going!  We also want to force a commit if the current
1366          * transaction is occupying too much of the log, or if the
1367          * transaction is too old now.
1368          */
1369         if (handle->h_sync ||
1370                         transaction->t_outstanding_credits >
1371                                 journal->j_max_transaction_buffers ||
1372                         time_after_eq(jiffies, transaction->t_expires)) {
1373                 /* Do this even for aborted journals: an abort still
1374                  * completes the commit thread, it just doesn't write
1375                  * anything to disk. */
1376                 tid_t tid = transaction->t_tid;
1377
1378                 spin_unlock(&transaction->t_handle_lock);
1379                 jbd_debug(2, "transaction too old, requesting commit for "
1380                                         "handle %p\n", handle);
1381                 /* This is non-blocking */
1382                 __log_start_commit(journal, transaction->t_tid);
1383                 spin_unlock(&journal->j_state_lock);
1384
1385                 /*
1386                  * Special case: JFS_SYNC synchronous updates require us
1387                  * to wait for the commit to complete.  
1388                  */
1389                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1390                         err = log_wait_commit(journal, tid);
1391         } else {
1392                 spin_unlock(&transaction->t_handle_lock);
1393                 spin_unlock(&journal->j_state_lock);
1394         }
1395
1396         jbd_free_handle(handle);
1397         return err;
1398 }
1399
1400 /**int journal_force_commit() - force any uncommitted transactions
1401  * @journal: journal to force
1402  *
1403  * For synchronous operations: force any uncommitted transactions
1404  * to disk.  May seem kludgy, but it reuses all the handle batching
1405  * code in a very simple manner.
1406  */
1407 int journal_force_commit(journal_t *journal)
1408 {
1409         handle_t *handle;
1410         int ret;
1411
1412         handle = journal_start(journal, 1);
1413         if (IS_ERR(handle)) {
1414                 ret = PTR_ERR(handle);
1415         } else {
1416                 handle->h_sync = 1;
1417                 ret = journal_stop(handle);
1418         }
1419         return ret;
1420 }
1421
1422 /*
1423  *
1424  * List management code snippets: various functions for manipulating the
1425  * transaction buffer lists.
1426  *
1427  */
1428
1429 /*
1430  * Append a buffer to a transaction list, given the transaction's list head
1431  * pointer.
1432  *
1433  * j_list_lock is held.
1434  *
1435  * jbd_lock_bh_state(jh2bh(jh)) is held.
1436  */
1437
1438 static inline void 
1439 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1440 {
1441         if (!*list) {
1442                 jh->b_tnext = jh->b_tprev = jh;
1443                 *list = jh;
1444         } else {
1445                 /* Insert at the tail of the list to preserve order */
1446                 struct journal_head *first = *list, *last = first->b_tprev;
1447                 jh->b_tprev = last;
1448                 jh->b_tnext = first;
1449                 last->b_tnext = first->b_tprev = jh;
1450         }
1451 }
1452
1453 /* 
1454  * Remove a buffer from a transaction list, given the transaction's list
1455  * head pointer.
1456  *
1457  * Called with j_list_lock held, and the journal may not be locked.
1458  *
1459  * jbd_lock_bh_state(jh2bh(jh)) is held.
1460  */
1461
1462 static inline void
1463 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1464 {
1465         if (*list == jh) {
1466                 *list = jh->b_tnext;
1467                 if (*list == jh)
1468                         *list = NULL;
1469         }
1470         jh->b_tprev->b_tnext = jh->b_tnext;
1471         jh->b_tnext->b_tprev = jh->b_tprev;
1472 }
1473
1474 /* 
1475  * Remove a buffer from the appropriate transaction list.
1476  *
1477  * Note that this function can *change* the value of
1478  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1479  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1480  * is holding onto a copy of one of thee pointers, it could go bad.
1481  * Generally the caller needs to re-read the pointer from the transaction_t.
1482  *
1483  * Called under j_list_lock.  The journal may not be locked.
1484  */
1485 void __journal_temp_unlink_buffer(struct journal_head *jh)
1486 {
1487         struct journal_head **list = NULL;
1488         transaction_t *transaction;
1489         struct buffer_head *bh = jh2bh(jh);
1490
1491         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1492         transaction = jh->b_transaction;
1493         if (transaction)
1494                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1495
1496         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1497         if (jh->b_jlist != BJ_None)
1498                 J_ASSERT_JH(jh, transaction != 0);
1499
1500         switch (jh->b_jlist) {
1501         case BJ_None:
1502                 return;
1503         case BJ_SyncData:
1504                 list = &transaction->t_sync_datalist;
1505                 break;
1506         case BJ_Metadata:
1507                 transaction->t_nr_buffers--;
1508                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1509                 list = &transaction->t_buffers;
1510                 break;
1511         case BJ_Forget:
1512                 list = &transaction->t_forget;
1513                 break;
1514         case BJ_IO:
1515                 list = &transaction->t_iobuf_list;
1516                 break;
1517         case BJ_Shadow:
1518                 list = &transaction->t_shadow_list;
1519                 break;
1520         case BJ_LogCtl:
1521                 list = &transaction->t_log_list;
1522                 break;
1523         case BJ_Reserved:
1524                 list = &transaction->t_reserved_list;
1525                 break;
1526         case BJ_Locked:
1527                 list = &transaction->t_locked_list;
1528                 break;
1529         }
1530
1531         __blist_del_buffer(list, jh);
1532         jh->b_jlist = BJ_None;
1533         if (test_clear_buffer_jbddirty(bh))
1534                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1535 }
1536
1537 void __journal_unfile_buffer(struct journal_head *jh)
1538 {
1539         __journal_temp_unlink_buffer(jh);
1540         jh->b_transaction = NULL;
1541 }
1542
1543 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1544 {
1545         jbd_lock_bh_state(jh2bh(jh));
1546         spin_lock(&journal->j_list_lock);
1547         __journal_unfile_buffer(jh);
1548         spin_unlock(&journal->j_list_lock);
1549         jbd_unlock_bh_state(jh2bh(jh));
1550 }
1551
1552 /*
1553  * Called from journal_try_to_free_buffers().
1554  *
1555  * Called under jbd_lock_bh_state(bh)
1556  */
1557 static void
1558 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1559 {
1560         struct journal_head *jh;
1561
1562         jh = bh2jh(bh);
1563
1564         if (buffer_locked(bh) || buffer_dirty(bh))
1565                 goto out;
1566
1567         if (jh->b_next_transaction != 0)
1568                 goto out;
1569
1570         spin_lock(&journal->j_list_lock);
1571         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1572                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1573                         /* A written-back ordered data buffer */
1574                         JBUFFER_TRACE(jh, "release data");
1575                         __journal_unfile_buffer(jh);
1576                         journal_remove_journal_head(bh);
1577                         __brelse(bh);
1578                 }
1579         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1580                 /* written-back checkpointed metadata buffer */
1581                 if (jh->b_jlist == BJ_None) {
1582                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1583                         __journal_remove_checkpoint(jh);
1584                         journal_remove_journal_head(bh);
1585                         __brelse(bh);
1586                 }
1587         }
1588         spin_unlock(&journal->j_list_lock);
1589 out:
1590         return;
1591 }
1592
1593
1594 /** 
1595  * int journal_try_to_free_buffers() - try to free page buffers.
1596  * @journal: journal for operation
1597  * @page: to try and free
1598  * @unused_gfp_mask: unused
1599  *
1600  * 
1601  * For all the buffers on this page,
1602  * if they are fully written out ordered data, move them onto BUF_CLEAN
1603  * so try_to_free_buffers() can reap them.
1604  * 
1605  * This function returns non-zero if we wish try_to_free_buffers()
1606  * to be called. We do this if the page is releasable by try_to_free_buffers().
1607  * We also do it if the page has locked or dirty buffers and the caller wants
1608  * us to perform sync or async writeout.
1609  *
1610  * This complicates JBD locking somewhat.  We aren't protected by the
1611  * BKL here.  We wish to remove the buffer from its committing or
1612  * running transaction's ->t_datalist via __journal_unfile_buffer.
1613  *
1614  * This may *change* the value of transaction_t->t_datalist, so anyone
1615  * who looks at t_datalist needs to lock against this function.
1616  *
1617  * Even worse, someone may be doing a journal_dirty_data on this
1618  * buffer.  So we need to lock against that.  journal_dirty_data()
1619  * will come out of the lock with the buffer dirty, which makes it
1620  * ineligible for release here.
1621  *
1622  * Who else is affected by this?  hmm...  Really the only contender
1623  * is do_get_write_access() - it could be looking at the buffer while
1624  * journal_try_to_free_buffer() is changing its state.  But that
1625  * cannot happen because we never reallocate freed data as metadata
1626  * while the data is part of a transaction.  Yes?
1627  */
1628 int journal_try_to_free_buffers(journal_t *journal, 
1629                                 struct page *page, gfp_t unused_gfp_mask)
1630 {
1631         struct buffer_head *head;
1632         struct buffer_head *bh;
1633         int ret = 0;
1634
1635         J_ASSERT(PageLocked(page));
1636
1637         head = page_buffers(page);
1638         bh = head;
1639         do {
1640                 struct journal_head *jh;
1641
1642                 /*
1643                  * We take our own ref against the journal_head here to avoid
1644                  * having to add tons of locking around each instance of
1645                  * journal_remove_journal_head() and journal_put_journal_head().
1646                  */
1647                 jh = journal_grab_journal_head(bh);
1648                 if (!jh)
1649                         continue;
1650
1651                 jbd_lock_bh_state(bh);
1652                 __journal_try_to_free_buffer(journal, bh);
1653                 journal_put_journal_head(jh);
1654                 jbd_unlock_bh_state(bh);
1655                 if (buffer_jbd(bh))
1656                         goto busy;
1657         } while ((bh = bh->b_this_page) != head);
1658         ret = try_to_free_buffers(page);
1659 busy:
1660         return ret;
1661 }
1662
1663 /*
1664  * This buffer is no longer needed.  If it is on an older transaction's
1665  * checkpoint list we need to record it on this transaction's forget list
1666  * to pin this buffer (and hence its checkpointing transaction) down until
1667  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1668  * release it.
1669  * Returns non-zero if JBD no longer has an interest in the buffer.
1670  *
1671  * Called under j_list_lock.
1672  *
1673  * Called under jbd_lock_bh_state(bh).
1674  */
1675 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1676 {
1677         int may_free = 1;
1678         struct buffer_head *bh = jh2bh(jh);
1679
1680         __journal_unfile_buffer(jh);
1681
1682         if (jh->b_cp_transaction) {
1683                 JBUFFER_TRACE(jh, "on running+cp transaction");
1684                 __journal_file_buffer(jh, transaction, BJ_Forget);
1685                 clear_buffer_jbddirty(bh);
1686                 may_free = 0;
1687         } else {
1688                 JBUFFER_TRACE(jh, "on running transaction");
1689                 journal_remove_journal_head(bh);
1690                 __brelse(bh);
1691         }
1692         return may_free;
1693 }
1694
1695 /*
1696  * journal_invalidatepage 
1697  *
1698  * This code is tricky.  It has a number of cases to deal with.
1699  *
1700  * There are two invariants which this code relies on:
1701  *
1702  * i_size must be updated on disk before we start calling invalidatepage on the
1703  * data.
1704  * 
1705  *  This is done in ext3 by defining an ext3_setattr method which
1706  *  updates i_size before truncate gets going.  By maintaining this
1707  *  invariant, we can be sure that it is safe to throw away any buffers
1708  *  attached to the current transaction: once the transaction commits,
1709  *  we know that the data will not be needed.
1710  * 
1711  *  Note however that we can *not* throw away data belonging to the
1712  *  previous, committing transaction!  
1713  *
1714  * Any disk blocks which *are* part of the previous, committing
1715  * transaction (and which therefore cannot be discarded immediately) are
1716  * not going to be reused in the new running transaction
1717  *
1718  *  The bitmap committed_data images guarantee this: any block which is
1719  *  allocated in one transaction and removed in the next will be marked
1720  *  as in-use in the committed_data bitmap, so cannot be reused until
1721  *  the next transaction to delete the block commits.  This means that
1722  *  leaving committing buffers dirty is quite safe: the disk blocks
1723  *  cannot be reallocated to a different file and so buffer aliasing is
1724  *  not possible.
1725  *
1726  *
1727  * The above applies mainly to ordered data mode.  In writeback mode we
1728  * don't make guarantees about the order in which data hits disk --- in
1729  * particular we don't guarantee that new dirty data is flushed before
1730  * transaction commit --- so it is always safe just to discard data
1731  * immediately in that mode.  --sct 
1732  */
1733
1734 /*
1735  * The journal_unmap_buffer helper function returns zero if the buffer
1736  * concerned remains pinned as an anonymous buffer belonging to an older
1737  * transaction.
1738  *
1739  * We're outside-transaction here.  Either or both of j_running_transaction
1740  * and j_committing_transaction may be NULL.
1741  */
1742 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1743 {
1744         transaction_t *transaction;
1745         struct journal_head *jh;
1746         int may_free = 1;
1747         int ret;
1748
1749         BUFFER_TRACE(bh, "entry");
1750
1751         /*
1752          * It is safe to proceed here without the j_list_lock because the
1753          * buffers cannot be stolen by try_to_free_buffers as long as we are
1754          * holding the page lock. --sct
1755          */
1756
1757         if (!buffer_jbd(bh))
1758                 goto zap_buffer_unlocked;
1759
1760         spin_lock(&journal->j_state_lock);
1761         jbd_lock_bh_state(bh);
1762         spin_lock(&journal->j_list_lock);
1763
1764         jh = journal_grab_journal_head(bh);
1765         if (!jh)
1766                 goto zap_buffer_no_jh;
1767
1768         transaction = jh->b_transaction;
1769         if (transaction == NULL) {
1770                 /* First case: not on any transaction.  If it
1771                  * has no checkpoint link, then we can zap it:
1772                  * it's a writeback-mode buffer so we don't care
1773                  * if it hits disk safely. */
1774                 if (!jh->b_cp_transaction) {
1775                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1776                         goto zap_buffer;
1777                 }
1778
1779                 if (!buffer_dirty(bh)) {
1780                         /* bdflush has written it.  We can drop it now */
1781                         goto zap_buffer;
1782                 }
1783
1784                 /* OK, it must be in the journal but still not
1785                  * written fully to disk: it's metadata or
1786                  * journaled data... */
1787
1788                 if (journal->j_running_transaction) {
1789                         /* ... and once the current transaction has
1790                          * committed, the buffer won't be needed any
1791                          * longer. */
1792                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1793                         ret = __dispose_buffer(jh,
1794                                         journal->j_running_transaction);
1795                         journal_put_journal_head(jh);
1796                         spin_unlock(&journal->j_list_lock);
1797                         jbd_unlock_bh_state(bh);
1798                         spin_unlock(&journal->j_state_lock);
1799                         return ret;
1800                 } else {
1801                         /* There is no currently-running transaction. So the
1802                          * orphan record which we wrote for this file must have
1803                          * passed into commit.  We must attach this buffer to
1804                          * the committing transaction, if it exists. */
1805                         if (journal->j_committing_transaction) {
1806                                 JBUFFER_TRACE(jh, "give to committing trans");
1807                                 ret = __dispose_buffer(jh,
1808                                         journal->j_committing_transaction);
1809                                 journal_put_journal_head(jh);
1810                                 spin_unlock(&journal->j_list_lock);
1811                                 jbd_unlock_bh_state(bh);
1812                                 spin_unlock(&journal->j_state_lock);
1813                                 return ret;
1814                         } else {
1815                                 /* The orphan record's transaction has
1816                                  * committed.  We can cleanse this buffer */
1817                                 clear_buffer_jbddirty(bh);
1818                                 goto zap_buffer;
1819                         }
1820                 }
1821         } else if (transaction == journal->j_committing_transaction) {
1822                 if (jh->b_jlist == BJ_Locked) {
1823                         /*
1824                          * The buffer is on the committing transaction's locked
1825                          * list.  We have the buffer locked, so I/O has
1826                          * completed.  So we can nail the buffer now.
1827                          */
1828                         may_free = __dispose_buffer(jh, transaction);
1829                         goto zap_buffer;
1830                 }
1831                 /*
1832                  * If it is committing, we simply cannot touch it.  We
1833                  * can remove it's next_transaction pointer from the
1834                  * running transaction if that is set, but nothing
1835                  * else. */
1836                 JBUFFER_TRACE(jh, "on committing transaction");
1837                 set_buffer_freed(bh);
1838                 if (jh->b_next_transaction) {
1839                         J_ASSERT(jh->b_next_transaction ==
1840                                         journal->j_running_transaction);
1841                         jh->b_next_transaction = NULL;
1842                 }
1843                 journal_put_journal_head(jh);
1844                 spin_unlock(&journal->j_list_lock);
1845                 jbd_unlock_bh_state(bh);
1846                 spin_unlock(&journal->j_state_lock);
1847                 return 0;
1848         } else {
1849                 /* Good, the buffer belongs to the running transaction.
1850                  * We are writing our own transaction's data, not any
1851                  * previous one's, so it is safe to throw it away
1852                  * (remember that we expect the filesystem to have set
1853                  * i_size already for this truncate so recovery will not
1854                  * expose the disk blocks we are discarding here.) */
1855                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1856                 may_free = __dispose_buffer(jh, transaction);
1857         }
1858
1859 zap_buffer:
1860         journal_put_journal_head(jh);
1861 zap_buffer_no_jh:
1862         spin_unlock(&journal->j_list_lock);
1863         jbd_unlock_bh_state(bh);
1864         spin_unlock(&journal->j_state_lock);
1865 zap_buffer_unlocked:
1866         clear_buffer_dirty(bh);
1867         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1868         clear_buffer_mapped(bh);
1869         clear_buffer_req(bh);
1870         clear_buffer_new(bh);
1871         bh->b_bdev = NULL;
1872         return may_free;
1873 }
1874
1875 /** 
1876  * int journal_invalidatepage() 
1877  * @journal: journal to use for flush... 
1878  * @page:    page to flush
1879  * @offset:  length of page to invalidate.
1880  *
1881  * Reap page buffers containing data after offset in page.
1882  *
1883  * Return non-zero if the page's buffers were successfully reaped.
1884  */
1885 int journal_invalidatepage(journal_t *journal, 
1886                       struct page *page, 
1887                       unsigned long offset)
1888 {
1889         struct buffer_head *head, *bh, *next;
1890         unsigned int curr_off = 0;
1891         int may_free = 1;
1892
1893         if (!PageLocked(page))
1894                 BUG();
1895         if (!page_has_buffers(page))
1896                 return 1;
1897
1898         /* We will potentially be playing with lists other than just the
1899          * data lists (especially for journaled data mode), so be
1900          * cautious in our locking. */
1901
1902         head = bh = page_buffers(page);
1903         do {
1904                 unsigned int next_off = curr_off + bh->b_size;
1905                 next = bh->b_this_page;
1906
1907                 if (offset <= curr_off) {
1908                         /* This block is wholly outside the truncation point */
1909                         lock_buffer(bh);
1910                         may_free &= journal_unmap_buffer(journal, bh);
1911                         unlock_buffer(bh);
1912                 }
1913                 curr_off = next_off;
1914                 bh = next;
1915
1916         } while (bh != head);
1917
1918         if (!offset) {
1919                 if (!may_free || !try_to_free_buffers(page))
1920                         return 0;
1921                 J_ASSERT(!page_has_buffers(page));
1922         }
1923         return 1;
1924 }
1925
1926 /* 
1927  * File a buffer on the given transaction list. 
1928  */
1929 void __journal_file_buffer(struct journal_head *jh,
1930                         transaction_t *transaction, int jlist)
1931 {
1932         struct journal_head **list = NULL;
1933         int was_dirty = 0;
1934         struct buffer_head *bh = jh2bh(jh);
1935
1936         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1937         assert_spin_locked(&transaction->t_journal->j_list_lock);
1938
1939         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1940         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1941                                 jh->b_transaction == 0);
1942
1943         if (jh->b_transaction && jh->b_jlist == jlist)
1944                 return;
1945
1946         /* The following list of buffer states needs to be consistent
1947          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1948          * state. */
1949
1950         if (jlist == BJ_Metadata || jlist == BJ_Reserved || 
1951             jlist == BJ_Shadow || jlist == BJ_Forget) {
1952                 if (test_clear_buffer_dirty(bh) ||
1953                     test_clear_buffer_jbddirty(bh))
1954                         was_dirty = 1;
1955         }
1956
1957         if (jh->b_transaction)
1958                 __journal_temp_unlink_buffer(jh);
1959         jh->b_transaction = transaction;
1960
1961         switch (jlist) {
1962         case BJ_None:
1963                 J_ASSERT_JH(jh, !jh->b_committed_data);
1964                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1965                 return;
1966         case BJ_SyncData:
1967                 list = &transaction->t_sync_datalist;
1968                 break;
1969         case BJ_Metadata:
1970                 transaction->t_nr_buffers++;
1971                 list = &transaction->t_buffers;
1972                 break;
1973         case BJ_Forget:
1974                 list = &transaction->t_forget;
1975                 break;
1976         case BJ_IO:
1977                 list = &transaction->t_iobuf_list;
1978                 break;
1979         case BJ_Shadow:
1980                 list = &transaction->t_shadow_list;
1981                 break;
1982         case BJ_LogCtl:
1983                 list = &transaction->t_log_list;
1984                 break;
1985         case BJ_Reserved:
1986                 list = &transaction->t_reserved_list;
1987                 break;
1988         case BJ_Locked:
1989                 list =  &transaction->t_locked_list;
1990                 break;
1991         }
1992
1993         __blist_add_buffer(list, jh);
1994         jh->b_jlist = jlist;
1995
1996         if (was_dirty)
1997                 set_buffer_jbddirty(bh);
1998 }
1999
2000 void journal_file_buffer(struct journal_head *jh,
2001                                 transaction_t *transaction, int jlist)
2002 {
2003         jbd_lock_bh_state(jh2bh(jh));
2004         spin_lock(&transaction->t_journal->j_list_lock);
2005         __journal_file_buffer(jh, transaction, jlist);
2006         spin_unlock(&transaction->t_journal->j_list_lock);
2007         jbd_unlock_bh_state(jh2bh(jh));
2008 }
2009
2010 /* 
2011  * Remove a buffer from its current buffer list in preparation for
2012  * dropping it from its current transaction entirely.  If the buffer has
2013  * already started to be used by a subsequent transaction, refile the
2014  * buffer on that transaction's metadata list.
2015  *
2016  * Called under journal->j_list_lock
2017  *
2018  * Called under jbd_lock_bh_state(jh2bh(jh))
2019  */
2020 void __journal_refile_buffer(struct journal_head *jh)
2021 {
2022         int was_dirty;
2023         struct buffer_head *bh = jh2bh(jh);
2024
2025         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2026         if (jh->b_transaction)
2027                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2028
2029         /* If the buffer is now unused, just drop it. */
2030         if (jh->b_next_transaction == NULL) {
2031                 __journal_unfile_buffer(jh);
2032                 return;
2033         }
2034
2035         /*
2036          * It has been modified by a later transaction: add it to the new
2037          * transaction's metadata list.
2038          */
2039
2040         was_dirty = test_clear_buffer_jbddirty(bh);
2041         __journal_temp_unlink_buffer(jh);
2042         jh->b_transaction = jh->b_next_transaction;
2043         jh->b_next_transaction = NULL;
2044         __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2045         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2046
2047         if (was_dirty)
2048                 set_buffer_jbddirty(bh);
2049 }
2050
2051 /*
2052  * For the unlocked version of this call, also make sure that any
2053  * hanging journal_head is cleaned up if necessary.
2054  *
2055  * __journal_refile_buffer is usually called as part of a single locked
2056  * operation on a buffer_head, in which the caller is probably going to
2057  * be hooking the journal_head onto other lists.  In that case it is up
2058  * to the caller to remove the journal_head if necessary.  For the
2059  * unlocked journal_refile_buffer call, the caller isn't going to be
2060  * doing anything else to the buffer so we need to do the cleanup
2061  * ourselves to avoid a jh leak. 
2062  *
2063  * *** The journal_head may be freed by this call! ***
2064  */
2065 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2066 {
2067         struct buffer_head *bh = jh2bh(jh);
2068
2069         jbd_lock_bh_state(bh);
2070         spin_lock(&journal->j_list_lock);
2071
2072         __journal_refile_buffer(jh);
2073         jbd_unlock_bh_state(bh);
2074         journal_remove_journal_head(bh);
2075
2076         spin_unlock(&journal->j_list_lock);
2077         __brelse(bh);
2078 }