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