Merge branch 'smsc47b397-new-id' into release
[linux-2.6] / fs / jbd / transaction.c
1 /*
2  * linux/fs/jbd/transaction.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem transaction handling code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages transactions (compound commits managed by the
16  * journaling code) and handles (individual atomic operations by the
17  * filesystem).
18  */
19
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28
29 static void __journal_temp_unlink_buffer(struct journal_head *jh);
30
31 /*
32  * get_transaction: obtain a new transaction_t object.
33  *
34  * Simply allocate and initialise a new transaction.  Create it in
35  * RUNNING state and add it to the current journal (which should not
36  * have an existing running transaction: we only make a new transaction
37  * once we have started to commit the old one).
38  *
39  * Preconditions:
40  *      The journal MUST be locked.  We don't perform atomic mallocs on the
41  *      new transaction and we can't block without protecting against other
42  *      processes trying to touch the journal while it is in transition.
43  *
44  * Called under j_state_lock
45  */
46
47 static transaction_t *
48 get_transaction(journal_t *journal, transaction_t *transaction)
49 {
50         transaction->t_journal = journal;
51         transaction->t_state = T_RUNNING;
52         transaction->t_tid = journal->j_transaction_sequence++;
53         transaction->t_expires = jiffies + journal->j_commit_interval;
54         spin_lock_init(&transaction->t_handle_lock);
55
56         /* Set up the commit timer for the new transaction. */
57         journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
58         add_timer(&journal->j_commit_timer);
59
60         J_ASSERT(journal->j_running_transaction == NULL);
61         journal->j_running_transaction = transaction;
62
63         return transaction;
64 }
65
66 /*
67  * Handle management.
68  *
69  * A handle_t is an object which represents a single atomic update to a
70  * filesystem, and which tracks all of the modifications which form part
71  * of that one update.
72  */
73
74 /*
75  * start_this_handle: Given a handle, deal with any locking or stalling
76  * needed to make sure that there is enough journal space for the handle
77  * to begin.  Attach the handle to a transaction and set up the
78  * transaction's buffer credits.
79  */
80
81 static int start_this_handle(journal_t *journal, handle_t *handle)
82 {
83         transaction_t *transaction;
84         int needed;
85         int nblocks = handle->h_buffer_credits;
86         transaction_t *new_transaction = NULL;
87         int ret = 0;
88
89         if (nblocks > journal->j_max_transaction_buffers) {
90                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
91                        current->comm, nblocks,
92                        journal->j_max_transaction_buffers);
93                 ret = -ENOSPC;
94                 goto out;
95         }
96
97 alloc_transaction:
98         if (!journal->j_running_transaction) {
99                 new_transaction = kzalloc(sizeof(*new_transaction),
100                                                 GFP_NOFS|__GFP_NOFAIL);
101                 if (!new_transaction) {
102                         ret = -ENOMEM;
103                         goto out;
104                 }
105         }
106
107         jbd_debug(3, "New handle %p going live.\n", handle);
108
109 repeat:
110
111         /*
112          * We need to hold j_state_lock until t_updates has been incremented,
113          * for proper journal barrier handling
114          */
115         spin_lock(&journal->j_state_lock);
116 repeat_locked:
117         if (is_journal_aborted(journal) ||
118             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
119                 spin_unlock(&journal->j_state_lock);
120                 ret = -EROFS;
121                 goto out;
122         }
123
124         /* Wait on the journal's transaction barrier if necessary */
125         if (journal->j_barrier_count) {
126                 spin_unlock(&journal->j_state_lock);
127                 wait_event(journal->j_wait_transaction_locked,
128                                 journal->j_barrier_count == 0);
129                 goto repeat;
130         }
131
132         if (!journal->j_running_transaction) {
133                 if (!new_transaction) {
134                         spin_unlock(&journal->j_state_lock);
135                         goto alloc_transaction;
136                 }
137                 get_transaction(journal, new_transaction);
138                 new_transaction = NULL;
139         }
140
141         transaction = journal->j_running_transaction;
142
143         /*
144          * If the current transaction is locked down for commit, wait for the
145          * lock to be released.
146          */
147         if (transaction->t_state == T_LOCKED) {
148                 DEFINE_WAIT(wait);
149
150                 prepare_to_wait(&journal->j_wait_transaction_locked,
151                                         &wait, TASK_UNINTERRUPTIBLE);
152                 spin_unlock(&journal->j_state_lock);
153                 schedule();
154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
155                 goto repeat;
156         }
157
158         /*
159          * If there is not enough space left in the log to write all potential
160          * buffers requested by this operation, we need to stall pending a log
161          * checkpoint to free some more log space.
162          */
163         spin_lock(&transaction->t_handle_lock);
164         needed = transaction->t_outstanding_credits + nblocks;
165
166         if (needed > journal->j_max_transaction_buffers) {
167                 /*
168                  * If the current transaction is already too large, then start
169                  * to commit it: we can then go back and attach this handle to
170                  * a new transaction.
171                  */
172                 DEFINE_WAIT(wait);
173
174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
175                 spin_unlock(&transaction->t_handle_lock);
176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177                                 TASK_UNINTERRUPTIBLE);
178                 __log_start_commit(journal, transaction->t_tid);
179                 spin_unlock(&journal->j_state_lock);
180                 schedule();
181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
182                 goto repeat;
183         }
184
185         /*
186          * The commit code assumes that it can get enough log space
187          * without forcing a checkpoint.  This is *critical* for
188          * correctness: a checkpoint of a buffer which is also
189          * associated with a committing transaction creates a deadlock,
190          * so commit simply cannot force through checkpoints.
191          *
192          * We must therefore ensure the necessary space in the journal
193          * *before* starting to dirty potentially checkpointed buffers
194          * in the new transaction.
195          *
196          * The worst part is, any transaction currently committing can
197          * reduce the free space arbitrarily.  Be careful to account for
198          * those buffers when checkpointing.
199          */
200
201         /*
202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
203          * a _lot_ of headroom: 1/4 of the journal plus the size of
204          * the committing transaction.  Really, we only need to give it
205          * committing_transaction->t_outstanding_credits plus "enough" for
206          * the log control blocks.
207          * Also, this test is inconsitent with the matching one in
208          * journal_extend().
209          */
210         if (__log_space_left(journal) < jbd_space_needed(journal)) {
211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
212                 spin_unlock(&transaction->t_handle_lock);
213                 __log_wait_for_space(journal);
214                 goto repeat_locked;
215         }
216
217         /* OK, account for the buffers that this operation expects to
218          * use and add the handle to the running transaction. */
219
220         handle->h_transaction = transaction;
221         transaction->t_outstanding_credits += nblocks;
222         transaction->t_updates++;
223         transaction->t_handle_count++;
224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
225                   handle, nblocks, transaction->t_outstanding_credits,
226                   __log_space_left(journal));
227         spin_unlock(&transaction->t_handle_lock);
228         spin_unlock(&journal->j_state_lock);
229 out:
230         if (unlikely(new_transaction))          /* It's usually NULL */
231                 kfree(new_transaction);
232         return ret;
233 }
234
235 static struct lock_class_key jbd_handle_key;
236
237 /* Allocate a new handle.  This should probably be in a slab... */
238 static handle_t *new_handle(int nblocks)
239 {
240         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
241         if (!handle)
242                 return NULL;
243         memset(handle, 0, sizeof(*handle));
244         handle->h_buffer_credits = nblocks;
245         handle->h_ref = 1;
246
247         lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
248
249         return handle;
250 }
251
252 /**
253  * handle_t *journal_start() - Obtain a new handle.
254  * @journal: Journal to start transaction on.
255  * @nblocks: number of block buffer we might modify
256  *
257  * We make sure that the transaction can guarantee at least nblocks of
258  * modified buffers in the log.  We block until the log can guarantee
259  * that much space.
260  *
261  * This function is visible to journal users (like ext3fs), so is not
262  * called with the journal already locked.
263  *
264  * Return a pointer to a newly allocated handle, or NULL on failure
265  */
266 handle_t *journal_start(journal_t *journal, int nblocks)
267 {
268         handle_t *handle = journal_current_handle();
269         int err;
270
271         if (!journal)
272                 return ERR_PTR(-EROFS);
273
274         if (handle) {
275                 J_ASSERT(handle->h_transaction->t_journal == journal);
276                 handle->h_ref++;
277                 return handle;
278         }
279
280         handle = new_handle(nblocks);
281         if (!handle)
282                 return ERR_PTR(-ENOMEM);
283
284         current->journal_info = handle;
285
286         err = start_this_handle(journal, handle);
287         if (err < 0) {
288                 jbd_free_handle(handle);
289                 current->journal_info = NULL;
290                 handle = ERR_PTR(err);
291                 goto out;
292         }
293
294         lock_acquire(&handle->h_lockdep_map, 0, 0, 0, 2, _THIS_IP_);
295
296 out:
297         return handle;
298 }
299
300 /**
301  * int journal_extend() - extend buffer credits.
302  * @handle:  handle to 'extend'
303  * @nblocks: nr blocks to try to extend by.
304  *
305  * Some transactions, such as large extends and truncates, can be done
306  * atomically all at once or in several stages.  The operation requests
307  * a credit for a number of buffer modications in advance, but can
308  * extend its credit if it needs more.
309  *
310  * journal_extend tries to give the running handle more buffer credits.
311  * It does not guarantee that allocation - this is a best-effort only.
312  * The calling process MUST be able to deal cleanly with a failure to
313  * extend here.
314  *
315  * Return 0 on success, non-zero on failure.
316  *
317  * return code < 0 implies an error
318  * return code > 0 implies normal transaction-full status.
319  */
320 int journal_extend(handle_t *handle, int nblocks)
321 {
322         transaction_t *transaction = handle->h_transaction;
323         journal_t *journal = transaction->t_journal;
324         int result;
325         int wanted;
326
327         result = -EIO;
328         if (is_handle_aborted(handle))
329                 goto out;
330
331         result = 1;
332
333         spin_lock(&journal->j_state_lock);
334
335         /* Don't extend a locked-down transaction! */
336         if (handle->h_transaction->t_state != T_RUNNING) {
337                 jbd_debug(3, "denied handle %p %d blocks: "
338                           "transaction not running\n", handle, nblocks);
339                 goto error_out;
340         }
341
342         spin_lock(&transaction->t_handle_lock);
343         wanted = transaction->t_outstanding_credits + nblocks;
344
345         if (wanted > journal->j_max_transaction_buffers) {
346                 jbd_debug(3, "denied handle %p %d blocks: "
347                           "transaction too large\n", handle, nblocks);
348                 goto unlock;
349         }
350
351         if (wanted > __log_space_left(journal)) {
352                 jbd_debug(3, "denied handle %p %d blocks: "
353                           "insufficient log space\n", handle, nblocks);
354                 goto unlock;
355         }
356
357         handle->h_buffer_credits += nblocks;
358         transaction->t_outstanding_credits += nblocks;
359         result = 0;
360
361         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
362 unlock:
363         spin_unlock(&transaction->t_handle_lock);
364 error_out:
365         spin_unlock(&journal->j_state_lock);
366 out:
367         return result;
368 }
369
370
371 /**
372  * int journal_restart() - restart a handle.
373  * @handle:  handle to restart
374  * @nblocks: nr credits requested
375  *
376  * Restart a handle for a multi-transaction filesystem
377  * operation.
378  *
379  * If the journal_extend() call above fails to grant new buffer credits
380  * to a running handle, a call to journal_restart will commit the
381  * handle's transaction so far and reattach the handle to a new
382  * transaction capabable of guaranteeing the requested number of
383  * credits.
384  */
385
386 int journal_restart(handle_t *handle, int nblocks)
387 {
388         transaction_t *transaction = handle->h_transaction;
389         journal_t *journal = transaction->t_journal;
390         int ret;
391
392         /* If we've had an abort of any type, don't even think about
393          * actually doing the restart! */
394         if (is_handle_aborted(handle))
395                 return 0;
396
397         /*
398          * First unlink the handle from its current transaction, and start the
399          * commit on that.
400          */
401         J_ASSERT(transaction->t_updates > 0);
402         J_ASSERT(journal_current_handle() == handle);
403
404         spin_lock(&journal->j_state_lock);
405         spin_lock(&transaction->t_handle_lock);
406         transaction->t_outstanding_credits -= handle->h_buffer_credits;
407         transaction->t_updates--;
408
409         if (!transaction->t_updates)
410                 wake_up(&journal->j_wait_updates);
411         spin_unlock(&transaction->t_handle_lock);
412
413         jbd_debug(2, "restarting handle %p\n", handle);
414         __log_start_commit(journal, transaction->t_tid);
415         spin_unlock(&journal->j_state_lock);
416
417         handle->h_buffer_credits = nblocks;
418         ret = start_this_handle(journal, handle);
419         return ret;
420 }
421
422
423 /**
424  * void journal_lock_updates () - establish a transaction barrier.
425  * @journal:  Journal to establish a barrier on.
426  *
427  * This locks out any further updates from being started, and blocks
428  * until all existing updates have completed, returning only once the
429  * journal is in a quiescent state with no updates running.
430  *
431  * The journal lock should not be held on entry.
432  */
433 void journal_lock_updates(journal_t *journal)
434 {
435         DEFINE_WAIT(wait);
436
437         spin_lock(&journal->j_state_lock);
438         ++journal->j_barrier_count;
439
440         /* Wait until there are no running updates */
441         while (1) {
442                 transaction_t *transaction = journal->j_running_transaction;
443
444                 if (!transaction)
445                         break;
446
447                 spin_lock(&transaction->t_handle_lock);
448                 if (!transaction->t_updates) {
449                         spin_unlock(&transaction->t_handle_lock);
450                         break;
451                 }
452                 prepare_to_wait(&journal->j_wait_updates, &wait,
453                                 TASK_UNINTERRUPTIBLE);
454                 spin_unlock(&transaction->t_handle_lock);
455                 spin_unlock(&journal->j_state_lock);
456                 schedule();
457                 finish_wait(&journal->j_wait_updates, &wait);
458                 spin_lock(&journal->j_state_lock);
459         }
460         spin_unlock(&journal->j_state_lock);
461
462         /*
463          * We have now established a barrier against other normal updates, but
464          * we also need to barrier against other journal_lock_updates() calls
465          * to make sure that we serialise special journal-locked operations
466          * too.
467          */
468         mutex_lock(&journal->j_barrier);
469 }
470
471 /**
472  * void journal_unlock_updates (journal_t* journal) - release barrier
473  * @journal:  Journal to release the barrier on.
474  *
475  * Release a transaction barrier obtained with journal_lock_updates().
476  *
477  * Should be called without the journal lock held.
478  */
479 void journal_unlock_updates (journal_t *journal)
480 {
481         J_ASSERT(journal->j_barrier_count != 0);
482
483         mutex_unlock(&journal->j_barrier);
484         spin_lock(&journal->j_state_lock);
485         --journal->j_barrier_count;
486         spin_unlock(&journal->j_state_lock);
487         wake_up(&journal->j_wait_transaction_locked);
488 }
489
490 /*
491  * Report any unexpected dirty buffers which turn up.  Normally those
492  * indicate an error, but they can occur if the user is running (say)
493  * tune2fs to modify the live filesystem, so we need the option of
494  * continuing as gracefully as possible.  #
495  *
496  * The caller should already hold the journal lock and
497  * j_list_lock spinlock: most callers will need those anyway
498  * in order to probe the buffer's journaling state safely.
499  */
500 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
501 {
502         int jlist;
503
504         /* If this buffer is one which might reasonably be dirty
505          * --- ie. data, or not part of this journal --- then
506          * we're OK to leave it alone, but otherwise we need to
507          * move the dirty bit to the journal's own internal
508          * JBDDirty bit. */
509         jlist = jh->b_jlist;
510
511         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
512             jlist == BJ_Shadow || jlist == BJ_Forget) {
513                 struct buffer_head *bh = jh2bh(jh);
514
515                 if (test_clear_buffer_dirty(bh))
516                         set_buffer_jbddirty(bh);
517         }
518 }
519
520 /*
521  * If the buffer is already part of the current transaction, then there
522  * is nothing we need to do.  If it is already part of a prior
523  * transaction which we are still committing to disk, then we need to
524  * make sure that we do not overwrite the old copy: we do copy-out to
525  * preserve the copy going to disk.  We also account the buffer against
526  * the handle's metadata buffer credits (unless the buffer is already
527  * part of the transaction, that is).
528  *
529  */
530 static int
531 do_get_write_access(handle_t *handle, struct journal_head *jh,
532                         int force_copy)
533 {
534         struct buffer_head *bh;
535         transaction_t *transaction;
536         journal_t *journal;
537         int error;
538         char *frozen_buffer = NULL;
539         int need_copy = 0;
540
541         if (is_handle_aborted(handle))
542                 return -EROFS;
543
544         transaction = handle->h_transaction;
545         journal = transaction->t_journal;
546
547         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
548
549         JBUFFER_TRACE(jh, "entry");
550 repeat:
551         bh = jh2bh(jh);
552
553         /* @@@ Need to check for errors here at some point. */
554
555         lock_buffer(bh);
556         jbd_lock_bh_state(bh);
557
558         /* We now hold the buffer lock so it is safe to query the buffer
559          * state.  Is the buffer dirty?
560          *
561          * If so, there are two possibilities.  The buffer may be
562          * non-journaled, and undergoing a quite legitimate writeback.
563          * Otherwise, it is journaled, and we don't expect dirty buffers
564          * in that state (the buffers should be marked JBD_Dirty
565          * instead.)  So either the IO is being done under our own
566          * control and this is a bug, or it's a third party IO such as
567          * dump(8) (which may leave the buffer scheduled for read ---
568          * ie. locked but not dirty) or tune2fs (which may actually have
569          * the buffer dirtied, ugh.)  */
570
571         if (buffer_dirty(bh)) {
572                 /*
573                  * First question: is this buffer already part of the current
574                  * transaction or the existing committing transaction?
575                  */
576                 if (jh->b_transaction) {
577                         J_ASSERT_JH(jh,
578                                 jh->b_transaction == transaction ||
579                                 jh->b_transaction ==
580                                         journal->j_committing_transaction);
581                         if (jh->b_next_transaction)
582                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
583                                                         transaction);
584                 }
585                 /*
586                  * In any case we need to clean the dirty flag and we must
587                  * do it under the buffer lock to be sure we don't race
588                  * with running write-out.
589                  */
590                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
591                 jbd_unexpected_dirty_buffer(jh);
592         }
593
594         unlock_buffer(bh);
595
596         error = -EROFS;
597         if (is_handle_aborted(handle)) {
598                 jbd_unlock_bh_state(bh);
599                 goto out;
600         }
601         error = 0;
602
603         /*
604          * The buffer is already part of this transaction if b_transaction or
605          * b_next_transaction points to it
606          */
607         if (jh->b_transaction == transaction ||
608             jh->b_next_transaction == transaction)
609                 goto done;
610
611         /*
612          * If there is already a copy-out version of this buffer, then we don't
613          * need to make another one
614          */
615         if (jh->b_frozen_data) {
616                 JBUFFER_TRACE(jh, "has frozen data");
617                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
618                 jh->b_next_transaction = transaction;
619                 goto done;
620         }
621
622         /* Is there data here we need to preserve? */
623
624         if (jh->b_transaction && jh->b_transaction != transaction) {
625                 JBUFFER_TRACE(jh, "owned by older transaction");
626                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
627                 J_ASSERT_JH(jh, jh->b_transaction ==
628                                         journal->j_committing_transaction);
629
630                 /* There is one case we have to be very careful about.
631                  * If the committing transaction is currently writing
632                  * this buffer out to disk and has NOT made a copy-out,
633                  * then we cannot modify the buffer contents at all
634                  * right now.  The essence of copy-out is that it is the
635                  * extra copy, not the primary copy, which gets
636                  * journaled.  If the primary copy is already going to
637                  * disk then we cannot do copy-out here. */
638
639                 if (jh->b_jlist == BJ_Shadow) {
640                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
641                         wait_queue_head_t *wqh;
642
643                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
644
645                         JBUFFER_TRACE(jh, "on shadow: sleep");
646                         jbd_unlock_bh_state(bh);
647                         /* commit wakes up all shadow buffers after IO */
648                         for ( ; ; ) {
649                                 prepare_to_wait(wqh, &wait.wait,
650                                                 TASK_UNINTERRUPTIBLE);
651                                 if (jh->b_jlist != BJ_Shadow)
652                                         break;
653                                 schedule();
654                         }
655                         finish_wait(wqh, &wait.wait);
656                         goto repeat;
657                 }
658
659                 /* Only do the copy if the currently-owning transaction
660                  * still needs it.  If it is on the Forget list, the
661                  * committing transaction is past that stage.  The
662                  * buffer had better remain locked during the kmalloc,
663                  * but that should be true --- we hold the journal lock
664                  * still and the buffer is already on the BUF_JOURNAL
665                  * list so won't be flushed.
666                  *
667                  * Subtle point, though: if this is a get_undo_access,
668                  * then we will be relying on the frozen_data to contain
669                  * the new value of the committed_data record after the
670                  * transaction, so we HAVE to force the frozen_data copy
671                  * in that case. */
672
673                 if (jh->b_jlist != BJ_Forget || force_copy) {
674                         JBUFFER_TRACE(jh, "generate frozen data");
675                         if (!frozen_buffer) {
676                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
677                                 jbd_unlock_bh_state(bh);
678                                 frozen_buffer =
679                                         jbd_alloc(jh2bh(jh)->b_size,
680                                                          GFP_NOFS);
681                                 if (!frozen_buffer) {
682                                         printk(KERN_EMERG
683                                                "%s: OOM for frozen_buffer\n",
684                                                __FUNCTION__);
685                                         JBUFFER_TRACE(jh, "oom!");
686                                         error = -ENOMEM;
687                                         jbd_lock_bh_state(bh);
688                                         goto done;
689                                 }
690                                 goto repeat;
691                         }
692                         jh->b_frozen_data = frozen_buffer;
693                         frozen_buffer = NULL;
694                         need_copy = 1;
695                 }
696                 jh->b_next_transaction = transaction;
697         }
698
699
700         /*
701          * Finally, if the buffer is not journaled right now, we need to make
702          * sure it doesn't get written to disk before the caller actually
703          * commits the new data
704          */
705         if (!jh->b_transaction) {
706                 JBUFFER_TRACE(jh, "no transaction");
707                 J_ASSERT_JH(jh, !jh->b_next_transaction);
708                 jh->b_transaction = transaction;
709                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
710                 spin_lock(&journal->j_list_lock);
711                 __journal_file_buffer(jh, transaction, BJ_Reserved);
712                 spin_unlock(&journal->j_list_lock);
713         }
714
715 done:
716         if (need_copy) {
717                 struct page *page;
718                 int offset;
719                 char *source;
720
721                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
722                             "Possible IO failure.\n");
723                 page = jh2bh(jh)->b_page;
724                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
725                 source = kmap_atomic(page, KM_USER0);
726                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
727                 kunmap_atomic(source, KM_USER0);
728         }
729         jbd_unlock_bh_state(bh);
730
731         /*
732          * If we are about to journal a buffer, then any revoke pending on it is
733          * no longer valid
734          */
735         journal_cancel_revoke(handle, jh);
736
737 out:
738         if (unlikely(frozen_buffer))    /* It's usually NULL */
739                 jbd_free(frozen_buffer, bh->b_size);
740
741         JBUFFER_TRACE(jh, "exit");
742         return error;
743 }
744
745 /**
746  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
747  * @handle: transaction to add buffer modifications to
748  * @bh:     bh to be used for metadata writes
749  * @credits: variable that will receive credits for the buffer
750  *
751  * Returns an error code or 0 on success.
752  *
753  * In full data journalling mode the buffer may be of type BJ_AsyncData,
754  * because we're write()ing a buffer which is also part of a shared mapping.
755  */
756
757 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
758 {
759         struct journal_head *jh = journal_add_journal_head(bh);
760         int rc;
761
762         /* We do not want to get caught playing with fields which the
763          * log thread also manipulates.  Make sure that the buffer
764          * completes any outstanding IO before proceeding. */
765         rc = do_get_write_access(handle, jh, 0);
766         journal_put_journal_head(jh);
767         return rc;
768 }
769
770
771 /*
772  * When the user wants to journal a newly created buffer_head
773  * (ie. getblk() returned a new buffer and we are going to populate it
774  * manually rather than reading off disk), then we need to keep the
775  * buffer_head locked until it has been completely filled with new
776  * data.  In this case, we should be able to make the assertion that
777  * the bh is not already part of an existing transaction.
778  *
779  * The buffer should already be locked by the caller by this point.
780  * There is no lock ranking violation: it was a newly created,
781  * unlocked buffer beforehand. */
782
783 /**
784  * int journal_get_create_access () - notify intent to use newly created bh
785  * @handle: transaction to new buffer to
786  * @bh: new buffer.
787  *
788  * Call this if you create a new bh.
789  */
790 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
791 {
792         transaction_t *transaction = handle->h_transaction;
793         journal_t *journal = transaction->t_journal;
794         struct journal_head *jh = journal_add_journal_head(bh);
795         int err;
796
797         jbd_debug(5, "journal_head %p\n", jh);
798         err = -EROFS;
799         if (is_handle_aborted(handle))
800                 goto out;
801         err = 0;
802
803         JBUFFER_TRACE(jh, "entry");
804         /*
805          * The buffer may already belong to this transaction due to pre-zeroing
806          * in the filesystem's new_block code.  It may also be on the previous,
807          * committing transaction's lists, but it HAS to be in Forget state in
808          * that case: the transaction must have deleted the buffer for it to be
809          * reused here.
810          */
811         jbd_lock_bh_state(bh);
812         spin_lock(&journal->j_list_lock);
813         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
814                 jh->b_transaction == NULL ||
815                 (jh->b_transaction == journal->j_committing_transaction &&
816                           jh->b_jlist == BJ_Forget)));
817
818         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
819         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
820
821         if (jh->b_transaction == NULL) {
822                 jh->b_transaction = transaction;
823                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
824                 __journal_file_buffer(jh, transaction, BJ_Reserved);
825         } else if (jh->b_transaction == journal->j_committing_transaction) {
826                 JBUFFER_TRACE(jh, "set next transaction");
827                 jh->b_next_transaction = transaction;
828         }
829         spin_unlock(&journal->j_list_lock);
830         jbd_unlock_bh_state(bh);
831
832         /*
833          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
834          * blocks which contain freed but then revoked metadata.  We need
835          * to cancel the revoke in case we end up freeing it yet again
836          * and the reallocating as data - this would cause a second revoke,
837          * which hits an assertion error.
838          */
839         JBUFFER_TRACE(jh, "cancelling revoke");
840         journal_cancel_revoke(handle, jh);
841         journal_put_journal_head(jh);
842 out:
843         return err;
844 }
845
846 /**
847  * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
848  * @handle: transaction
849  * @bh: buffer to undo
850  * @credits: store the number of taken credits here (if not NULL)
851  *
852  * Sometimes there is a need to distinguish between metadata which has
853  * been committed to disk and that which has not.  The ext3fs code uses
854  * this for freeing and allocating space, we have to make sure that we
855  * do not reuse freed space until the deallocation has been committed,
856  * since if we overwrote that space we would make the delete
857  * un-rewindable in case of a crash.
858  *
859  * To deal with that, journal_get_undo_access requests write access to a
860  * buffer for parts of non-rewindable operations such as delete
861  * operations on the bitmaps.  The journaling code must keep a copy of
862  * the buffer's contents prior to the undo_access call until such time
863  * as we know that the buffer has definitely been committed to disk.
864  *
865  * We never need to know which transaction the committed data is part
866  * of, buffers touched here are guaranteed to be dirtied later and so
867  * will be committed to a new transaction in due course, at which point
868  * we can discard the old committed data pointer.
869  *
870  * Returns error number or 0 on success.
871  */
872 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
873 {
874         int err;
875         struct journal_head *jh = journal_add_journal_head(bh);
876         char *committed_data = NULL;
877
878         JBUFFER_TRACE(jh, "entry");
879
880         /*
881          * Do this first --- it can drop the journal lock, so we want to
882          * make sure that obtaining the committed_data is done
883          * atomically wrt. completion of any outstanding commits.
884          */
885         err = do_get_write_access(handle, jh, 1);
886         if (err)
887                 goto out;
888
889 repeat:
890         if (!jh->b_committed_data) {
891                 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
892                 if (!committed_data) {
893                         printk(KERN_EMERG "%s: No memory for committed data\n",
894                                 __FUNCTION__);
895                         err = -ENOMEM;
896                         goto out;
897                 }
898         }
899
900         jbd_lock_bh_state(bh);
901         if (!jh->b_committed_data) {
902                 /* Copy out the current buffer contents into the
903                  * preserved, committed copy. */
904                 JBUFFER_TRACE(jh, "generate b_committed data");
905                 if (!committed_data) {
906                         jbd_unlock_bh_state(bh);
907                         goto repeat;
908                 }
909
910                 jh->b_committed_data = committed_data;
911                 committed_data = NULL;
912                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
913         }
914         jbd_unlock_bh_state(bh);
915 out:
916         journal_put_journal_head(jh);
917         if (unlikely(committed_data))
918                 jbd_free(committed_data, bh->b_size);
919         return err;
920 }
921
922 /**
923  * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
924  * @handle: transaction
925  * @bh: bufferhead to mark
926  *
927  * Description:
928  * Mark a buffer as containing dirty data which needs to be flushed before
929  * we can commit the current transaction.
930  *
931  * The buffer is placed on the transaction's data list and is marked as
932  * belonging to the transaction.
933  *
934  * Returns error number or 0 on success.
935  *
936  * journal_dirty_data() can be called via page_launder->ext3_writepage
937  * by kswapd.
938  */
939 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
940 {
941         journal_t *journal = handle->h_transaction->t_journal;
942         int need_brelse = 0;
943         struct journal_head *jh;
944
945         if (is_handle_aborted(handle))
946                 return 0;
947
948         jh = journal_add_journal_head(bh);
949         JBUFFER_TRACE(jh, "entry");
950
951         /*
952          * The buffer could *already* be dirty.  Writeout can start
953          * at any time.
954          */
955         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
956
957         /*
958          * What if the buffer is already part of a running transaction?
959          *
960          * There are two cases:
961          * 1) It is part of the current running transaction.  Refile it,
962          *    just in case we have allocated it as metadata, deallocated
963          *    it, then reallocated it as data.
964          * 2) It is part of the previous, still-committing transaction.
965          *    If all we want to do is to guarantee that the buffer will be
966          *    written to disk before this new transaction commits, then
967          *    being sure that the *previous* transaction has this same
968          *    property is sufficient for us!  Just leave it on its old
969          *    transaction.
970          *
971          * In case (2), the buffer must not already exist as metadata
972          * --- that would violate write ordering (a transaction is free
973          * to write its data at any point, even before the previous
974          * committing transaction has committed).  The caller must
975          * never, ever allow this to happen: there's nothing we can do
976          * about it in this layer.
977          */
978         jbd_lock_bh_state(bh);
979         spin_lock(&journal->j_list_lock);
980
981         /* Now that we have bh_state locked, are we really still mapped? */
982         if (!buffer_mapped(bh)) {
983                 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
984                 goto no_journal;
985         }
986
987         if (jh->b_transaction) {
988                 JBUFFER_TRACE(jh, "has transaction");
989                 if (jh->b_transaction != handle->h_transaction) {
990                         JBUFFER_TRACE(jh, "belongs to older transaction");
991                         J_ASSERT_JH(jh, jh->b_transaction ==
992                                         journal->j_committing_transaction);
993
994                         /* @@@ IS THIS TRUE  ? */
995                         /*
996                          * Not any more.  Scenario: someone does a write()
997                          * in data=journal mode.  The buffer's transaction has
998                          * moved into commit.  Then someone does another
999                          * write() to the file.  We do the frozen data copyout
1000                          * and set b_next_transaction to point to j_running_t.
1001                          * And while we're in that state, someone does a
1002                          * writepage() in an attempt to pageout the same area
1003                          * of the file via a shared mapping.  At present that
1004                          * calls journal_dirty_data(), and we get right here.
1005                          * It may be too late to journal the data.  Simply
1006                          * falling through to the next test will suffice: the
1007                          * data will be dirty and wil be checkpointed.  The
1008                          * ordering comments in the next comment block still
1009                          * apply.
1010                          */
1011                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1012
1013                         /*
1014                          * If we're journalling data, and this buffer was
1015                          * subject to a write(), it could be metadata, forget
1016                          * or shadow against the committing transaction.  Now,
1017                          * someone has dirtied the same darn page via a mapping
1018                          * and it is being writepage()'d.
1019                          * We *could* just steal the page from commit, with some
1020                          * fancy locking there.  Instead, we just skip it -
1021                          * don't tie the page's buffers to the new transaction
1022                          * at all.
1023                          * Implication: if we crash before the writepage() data
1024                          * is written into the filesystem, recovery will replay
1025                          * the write() data.
1026                          */
1027                         if (jh->b_jlist != BJ_None &&
1028                                         jh->b_jlist != BJ_SyncData &&
1029                                         jh->b_jlist != BJ_Locked) {
1030                                 JBUFFER_TRACE(jh, "Not stealing");
1031                                 goto no_journal;
1032                         }
1033
1034                         /*
1035                          * This buffer may be undergoing writeout in commit.  We
1036                          * can't return from here and let the caller dirty it
1037                          * again because that can cause the write-out loop in
1038                          * commit to never terminate.
1039                          */
1040                         if (buffer_dirty(bh)) {
1041                                 get_bh(bh);
1042                                 spin_unlock(&journal->j_list_lock);
1043                                 jbd_unlock_bh_state(bh);
1044                                 need_brelse = 1;
1045                                 sync_dirty_buffer(bh);
1046                                 jbd_lock_bh_state(bh);
1047                                 spin_lock(&journal->j_list_lock);
1048                                 /* Since we dropped the lock... */
1049                                 if (!buffer_mapped(bh)) {
1050                                         JBUFFER_TRACE(jh, "buffer got unmapped");
1051                                         goto no_journal;
1052                                 }
1053                                 /* The buffer may become locked again at any
1054                                    time if it is redirtied */
1055                         }
1056
1057                         /* journal_clean_data_list() may have got there first */
1058                         if (jh->b_transaction != NULL) {
1059                                 JBUFFER_TRACE(jh, "unfile from commit");
1060                                 __journal_temp_unlink_buffer(jh);
1061                                 /* It still points to the committing
1062                                  * transaction; move it to this one so
1063                                  * that the refile assert checks are
1064                                  * happy. */
1065                                 jh->b_transaction = handle->h_transaction;
1066                         }
1067                         /* The buffer will be refiled below */
1068
1069                 }
1070                 /*
1071                  * Special case --- the buffer might actually have been
1072                  * allocated and then immediately deallocated in the previous,
1073                  * committing transaction, so might still be left on that
1074                  * transaction's metadata lists.
1075                  */
1076                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1077                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1078                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1079                         __journal_temp_unlink_buffer(jh);
1080                         jh->b_transaction = handle->h_transaction;
1081                         JBUFFER_TRACE(jh, "file as data");
1082                         __journal_file_buffer(jh, handle->h_transaction,
1083                                                 BJ_SyncData);
1084                 }
1085         } else {
1086                 JBUFFER_TRACE(jh, "not on a transaction");
1087                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1088         }
1089 no_journal:
1090         spin_unlock(&journal->j_list_lock);
1091         jbd_unlock_bh_state(bh);
1092         if (need_brelse) {
1093                 BUFFER_TRACE(bh, "brelse");
1094                 __brelse(bh);
1095         }
1096         JBUFFER_TRACE(jh, "exit");
1097         journal_put_journal_head(jh);
1098         return 0;
1099 }
1100
1101 /**
1102  * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1103  * @handle: transaction to add buffer to.
1104  * @bh: buffer to mark
1105  *
1106  * Mark dirty metadata which needs to be journaled as part of the current
1107  * transaction.
1108  *
1109  * The buffer is placed on the transaction's metadata list and is marked
1110  * as belonging to the transaction.
1111  *
1112  * Returns error number or 0 on success.
1113  *
1114  * Special care needs to be taken if the buffer already belongs to the
1115  * current committing transaction (in which case we should have frozen
1116  * data present for that commit).  In that case, we don't relink the
1117  * buffer: that only gets done when the old transaction finally
1118  * completes its commit.
1119  */
1120 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1121 {
1122         transaction_t *transaction = handle->h_transaction;
1123         journal_t *journal = transaction->t_journal;
1124         struct journal_head *jh = bh2jh(bh);
1125
1126         jbd_debug(5, "journal_head %p\n", jh);
1127         JBUFFER_TRACE(jh, "entry");
1128         if (is_handle_aborted(handle))
1129                 goto out;
1130
1131         jbd_lock_bh_state(bh);
1132
1133         if (jh->b_modified == 0) {
1134                 /*
1135                  * This buffer's got modified and becoming part
1136                  * of the transaction. This needs to be done
1137                  * once a transaction -bzzz
1138                  */
1139                 jh->b_modified = 1;
1140                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1141                 handle->h_buffer_credits--;
1142         }
1143
1144         /*
1145          * fastpath, to avoid expensive locking.  If this buffer is already
1146          * on the running transaction's metadata list there is nothing to do.
1147          * Nobody can take it off again because there is a handle open.
1148          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1149          * result in this test being false, so we go in and take the locks.
1150          */
1151         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1152                 JBUFFER_TRACE(jh, "fastpath");
1153                 J_ASSERT_JH(jh, jh->b_transaction ==
1154                                         journal->j_running_transaction);
1155                 goto out_unlock_bh;
1156         }
1157
1158         set_buffer_jbddirty(bh);
1159
1160         /*
1161          * Metadata already on the current transaction list doesn't
1162          * need to be filed.  Metadata on another transaction's list must
1163          * be committing, and will be refiled once the commit completes:
1164          * leave it alone for now.
1165          */
1166         if (jh->b_transaction != transaction) {
1167                 JBUFFER_TRACE(jh, "already on other transaction");
1168                 J_ASSERT_JH(jh, jh->b_transaction ==
1169                                         journal->j_committing_transaction);
1170                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1171                 /* And this case is illegal: we can't reuse another
1172                  * transaction's data buffer, ever. */
1173                 goto out_unlock_bh;
1174         }
1175
1176         /* That test should have eliminated the following case: */
1177         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1178
1179         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1180         spin_lock(&journal->j_list_lock);
1181         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1182         spin_unlock(&journal->j_list_lock);
1183 out_unlock_bh:
1184         jbd_unlock_bh_state(bh);
1185 out:
1186         JBUFFER_TRACE(jh, "exit");
1187         return 0;
1188 }
1189
1190 /*
1191  * journal_release_buffer: undo a get_write_access without any buffer
1192  * updates, if the update decided in the end that it didn't need access.
1193  *
1194  */
1195 void
1196 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1197 {
1198         BUFFER_TRACE(bh, "entry");
1199 }
1200
1201 /**
1202  * void journal_forget() - bforget() for potentially-journaled buffers.
1203  * @handle: transaction handle
1204  * @bh:     bh to 'forget'
1205  *
1206  * We can only do the bforget if there are no commits pending against the
1207  * buffer.  If the buffer is dirty in the current running transaction we
1208  * can safely unlink it.
1209  *
1210  * bh may not be a journalled buffer at all - it may be a non-JBD
1211  * buffer which came off the hashtable.  Check for this.
1212  *
1213  * Decrements bh->b_count by one.
1214  *
1215  * Allow this call even if the handle has aborted --- it may be part of
1216  * the caller's cleanup after an abort.
1217  */
1218 int journal_forget (handle_t *handle, struct buffer_head *bh)
1219 {
1220         transaction_t *transaction = handle->h_transaction;
1221         journal_t *journal = transaction->t_journal;
1222         struct journal_head *jh;
1223         int drop_reserve = 0;
1224         int err = 0;
1225
1226         BUFFER_TRACE(bh, "entry");
1227
1228         jbd_lock_bh_state(bh);
1229         spin_lock(&journal->j_list_lock);
1230
1231         if (!buffer_jbd(bh))
1232                 goto not_jbd;
1233         jh = bh2jh(bh);
1234
1235         /* Critical error: attempting to delete a bitmap buffer, maybe?
1236          * Don't do any jbd operations, and return an error. */
1237         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1238                          "inconsistent data on disk")) {
1239                 err = -EIO;
1240                 goto not_jbd;
1241         }
1242
1243         /*
1244          * The buffer's going from the transaction, we must drop
1245          * all references -bzzz
1246          */
1247         jh->b_modified = 0;
1248
1249         if (jh->b_transaction == handle->h_transaction) {
1250                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1251
1252                 /* If we are forgetting a buffer which is already part
1253                  * of this transaction, then we can just drop it from
1254                  * the transaction immediately. */
1255                 clear_buffer_dirty(bh);
1256                 clear_buffer_jbddirty(bh);
1257
1258                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1259
1260                 drop_reserve = 1;
1261
1262                 /*
1263                  * We are no longer going to journal this buffer.
1264                  * However, the commit of this transaction is still
1265                  * important to the buffer: the delete that we are now
1266                  * processing might obsolete an old log entry, so by
1267                  * committing, we can satisfy the buffer's checkpoint.
1268                  *
1269                  * So, if we have a checkpoint on the buffer, we should
1270                  * now refile the buffer on our BJ_Forget list so that
1271                  * we know to remove the checkpoint after we commit.
1272                  */
1273
1274                 if (jh->b_cp_transaction) {
1275                         __journal_temp_unlink_buffer(jh);
1276                         __journal_file_buffer(jh, transaction, BJ_Forget);
1277                 } else {
1278                         __journal_unfile_buffer(jh);
1279                         journal_remove_journal_head(bh);
1280                         __brelse(bh);
1281                         if (!buffer_jbd(bh)) {
1282                                 spin_unlock(&journal->j_list_lock);
1283                                 jbd_unlock_bh_state(bh);
1284                                 __bforget(bh);
1285                                 goto drop;
1286                         }
1287                 }
1288         } else if (jh->b_transaction) {
1289                 J_ASSERT_JH(jh, (jh->b_transaction ==
1290                                  journal->j_committing_transaction));
1291                 /* However, if the buffer is still owned by a prior
1292                  * (committing) transaction, we can't drop it yet... */
1293                 JBUFFER_TRACE(jh, "belongs to older transaction");
1294                 /* ... but we CAN drop it from the new transaction if we
1295                  * have also modified it since the original commit. */
1296
1297                 if (jh->b_next_transaction) {
1298                         J_ASSERT(jh->b_next_transaction == transaction);
1299                         jh->b_next_transaction = NULL;
1300                         drop_reserve = 1;
1301                 }
1302         }
1303
1304 not_jbd:
1305         spin_unlock(&journal->j_list_lock);
1306         jbd_unlock_bh_state(bh);
1307         __brelse(bh);
1308 drop:
1309         if (drop_reserve) {
1310                 /* no need to reserve log space for this block -bzzz */
1311                 handle->h_buffer_credits++;
1312         }
1313         return err;
1314 }
1315
1316 /**
1317  * int journal_stop() - complete a transaction
1318  * @handle: tranaction to complete.
1319  *
1320  * All done for a particular handle.
1321  *
1322  * There is not much action needed here.  We just return any remaining
1323  * buffer credits to the transaction and remove the handle.  The only
1324  * complication is that we need to start a commit operation if the
1325  * filesystem is marked for synchronous update.
1326  *
1327  * journal_stop itself will not usually return an error, but it may
1328  * do so in unusual circumstances.  In particular, expect it to
1329  * return -EIO if a journal_abort has been executed since the
1330  * transaction began.
1331  */
1332 int journal_stop(handle_t *handle)
1333 {
1334         transaction_t *transaction = handle->h_transaction;
1335         journal_t *journal = transaction->t_journal;
1336         int old_handle_count, err;
1337         pid_t pid;
1338
1339         J_ASSERT(journal_current_handle() == handle);
1340
1341         if (is_handle_aborted(handle))
1342                 err = -EIO;
1343         else {
1344                 J_ASSERT(transaction->t_updates > 0);
1345                 err = 0;
1346         }
1347
1348         if (--handle->h_ref > 0) {
1349                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1350                           handle->h_ref);
1351                 return err;
1352         }
1353
1354         jbd_debug(4, "Handle %p going down\n", handle);
1355
1356         /*
1357          * Implement synchronous transaction batching.  If the handle
1358          * was synchronous, don't force a commit immediately.  Let's
1359          * yield and let another thread piggyback onto this transaction.
1360          * Keep doing that while new threads continue to arrive.
1361          * It doesn't cost much - we're about to run a commit and sleep
1362          * on IO anyway.  Speeds up many-threaded, many-dir operations
1363          * by 30x or more...
1364          *
1365          * But don't do this if this process was the most recent one to
1366          * perform a synchronous write.  We do this to detect the case where a
1367          * single process is doing a stream of sync writes.  No point in waiting
1368          * for joiners in that case.
1369          */
1370         pid = current->pid;
1371         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1372                 journal->j_last_sync_writer = pid;
1373                 do {
1374                         old_handle_count = transaction->t_handle_count;
1375                         schedule_timeout_uninterruptible(1);
1376                 } while (old_handle_count != transaction->t_handle_count);
1377         }
1378
1379         current->journal_info = NULL;
1380         spin_lock(&journal->j_state_lock);
1381         spin_lock(&transaction->t_handle_lock);
1382         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1383         transaction->t_updates--;
1384         if (!transaction->t_updates) {
1385                 wake_up(&journal->j_wait_updates);
1386                 if (journal->j_barrier_count)
1387                         wake_up(&journal->j_wait_transaction_locked);
1388         }
1389
1390         /*
1391          * If the handle is marked SYNC, we need to set another commit
1392          * going!  We also want to force a commit if the current
1393          * transaction is occupying too much of the log, or if the
1394          * transaction is too old now.
1395          */
1396         if (handle->h_sync ||
1397                         transaction->t_outstanding_credits >
1398                                 journal->j_max_transaction_buffers ||
1399                         time_after_eq(jiffies, transaction->t_expires)) {
1400                 /* Do this even for aborted journals: an abort still
1401                  * completes the commit thread, it just doesn't write
1402                  * anything to disk. */
1403                 tid_t tid = transaction->t_tid;
1404
1405                 spin_unlock(&transaction->t_handle_lock);
1406                 jbd_debug(2, "transaction too old, requesting commit for "
1407                                         "handle %p\n", handle);
1408                 /* This is non-blocking */
1409                 __log_start_commit(journal, transaction->t_tid);
1410                 spin_unlock(&journal->j_state_lock);
1411
1412                 /*
1413                  * Special case: JFS_SYNC synchronous updates require us
1414                  * to wait for the commit to complete.
1415                  */
1416                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1417                         err = log_wait_commit(journal, tid);
1418         } else {
1419                 spin_unlock(&transaction->t_handle_lock);
1420                 spin_unlock(&journal->j_state_lock);
1421         }
1422
1423         lock_release(&handle->h_lockdep_map, 1, _THIS_IP_);
1424
1425         jbd_free_handle(handle);
1426         return err;
1427 }
1428
1429 /**
1430  * int journal_force_commit() - force any uncommitted transactions
1431  * @journal: journal to force
1432  *
1433  * For synchronous operations: force any uncommitted transactions
1434  * to disk.  May seem kludgy, but it reuses all the handle batching
1435  * code in a very simple manner.
1436  */
1437 int journal_force_commit(journal_t *journal)
1438 {
1439         handle_t *handle;
1440         int ret;
1441
1442         handle = journal_start(journal, 1);
1443         if (IS_ERR(handle)) {
1444                 ret = PTR_ERR(handle);
1445         } else {
1446                 handle->h_sync = 1;
1447                 ret = journal_stop(handle);
1448         }
1449         return ret;
1450 }
1451
1452 /*
1453  *
1454  * List management code snippets: various functions for manipulating the
1455  * transaction buffer lists.
1456  *
1457  */
1458
1459 /*
1460  * Append a buffer to a transaction list, given the transaction's list head
1461  * pointer.
1462  *
1463  * j_list_lock is held.
1464  *
1465  * jbd_lock_bh_state(jh2bh(jh)) is held.
1466  */
1467
1468 static inline void
1469 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1470 {
1471         if (!*list) {
1472                 jh->b_tnext = jh->b_tprev = jh;
1473                 *list = jh;
1474         } else {
1475                 /* Insert at the tail of the list to preserve order */
1476                 struct journal_head *first = *list, *last = first->b_tprev;
1477                 jh->b_tprev = last;
1478                 jh->b_tnext = first;
1479                 last->b_tnext = first->b_tprev = jh;
1480         }
1481 }
1482
1483 /*
1484  * Remove a buffer from a transaction list, given the transaction's list
1485  * head pointer.
1486  *
1487  * Called with j_list_lock held, and the journal may not be locked.
1488  *
1489  * jbd_lock_bh_state(jh2bh(jh)) is held.
1490  */
1491
1492 static inline void
1493 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1494 {
1495         if (*list == jh) {
1496                 *list = jh->b_tnext;
1497                 if (*list == jh)
1498                         *list = NULL;
1499         }
1500         jh->b_tprev->b_tnext = jh->b_tnext;
1501         jh->b_tnext->b_tprev = jh->b_tprev;
1502 }
1503
1504 /*
1505  * Remove a buffer from the appropriate transaction list.
1506  *
1507  * Note that this function can *change* the value of
1508  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1509  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1510  * is holding onto a copy of one of thee pointers, it could go bad.
1511  * Generally the caller needs to re-read the pointer from the transaction_t.
1512  *
1513  * Called under j_list_lock.  The journal may not be locked.
1514  */
1515 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1516 {
1517         struct journal_head **list = NULL;
1518         transaction_t *transaction;
1519         struct buffer_head *bh = jh2bh(jh);
1520
1521         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1522         transaction = jh->b_transaction;
1523         if (transaction)
1524                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1525
1526         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1527         if (jh->b_jlist != BJ_None)
1528                 J_ASSERT_JH(jh, transaction != NULL);
1529
1530         switch (jh->b_jlist) {
1531         case BJ_None:
1532                 return;
1533         case BJ_SyncData:
1534                 list = &transaction->t_sync_datalist;
1535                 break;
1536         case BJ_Metadata:
1537                 transaction->t_nr_buffers--;
1538                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1539                 list = &transaction->t_buffers;
1540                 break;
1541         case BJ_Forget:
1542                 list = &transaction->t_forget;
1543                 break;
1544         case BJ_IO:
1545                 list = &transaction->t_iobuf_list;
1546                 break;
1547         case BJ_Shadow:
1548                 list = &transaction->t_shadow_list;
1549                 break;
1550         case BJ_LogCtl:
1551                 list = &transaction->t_log_list;
1552                 break;
1553         case BJ_Reserved:
1554                 list = &transaction->t_reserved_list;
1555                 break;
1556         case BJ_Locked:
1557                 list = &transaction->t_locked_list;
1558                 break;
1559         }
1560
1561         __blist_del_buffer(list, jh);
1562         jh->b_jlist = BJ_None;
1563         if (test_clear_buffer_jbddirty(bh))
1564                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1565 }
1566
1567 void __journal_unfile_buffer(struct journal_head *jh)
1568 {
1569         __journal_temp_unlink_buffer(jh);
1570         jh->b_transaction = NULL;
1571 }
1572
1573 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1574 {
1575         jbd_lock_bh_state(jh2bh(jh));
1576         spin_lock(&journal->j_list_lock);
1577         __journal_unfile_buffer(jh);
1578         spin_unlock(&journal->j_list_lock);
1579         jbd_unlock_bh_state(jh2bh(jh));
1580 }
1581
1582 /*
1583  * Called from journal_try_to_free_buffers().
1584  *
1585  * Called under jbd_lock_bh_state(bh)
1586  */
1587 static void
1588 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1589 {
1590         struct journal_head *jh;
1591
1592         jh = bh2jh(bh);
1593
1594         if (buffer_locked(bh) || buffer_dirty(bh))
1595                 goto out;
1596
1597         if (jh->b_next_transaction != NULL)
1598                 goto out;
1599
1600         spin_lock(&journal->j_list_lock);
1601         if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1602                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1603                         /* A written-back ordered data buffer */
1604                         JBUFFER_TRACE(jh, "release data");
1605                         __journal_unfile_buffer(jh);
1606                         journal_remove_journal_head(bh);
1607                         __brelse(bh);
1608                 }
1609         } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1610                 /* written-back checkpointed metadata buffer */
1611                 if (jh->b_jlist == BJ_None) {
1612                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1613                         __journal_remove_checkpoint(jh);
1614                         journal_remove_journal_head(bh);
1615                         __brelse(bh);
1616                 }
1617         }
1618         spin_unlock(&journal->j_list_lock);
1619 out:
1620         return;
1621 }
1622
1623
1624 /**
1625  * int journal_try_to_free_buffers() - try to free page buffers.
1626  * @journal: journal for operation
1627  * @page: to try and free
1628  * @unused_gfp_mask: unused
1629  *
1630  *
1631  * For all the buffers on this page,
1632  * if they are fully written out ordered data, move them onto BUF_CLEAN
1633  * so try_to_free_buffers() can reap them.
1634  *
1635  * This function returns non-zero if we wish try_to_free_buffers()
1636  * to be called. We do this if the page is releasable by try_to_free_buffers().
1637  * We also do it if the page has locked or dirty buffers and the caller wants
1638  * us to perform sync or async writeout.
1639  *
1640  * This complicates JBD locking somewhat.  We aren't protected by the
1641  * BKL here.  We wish to remove the buffer from its committing or
1642  * running transaction's ->t_datalist via __journal_unfile_buffer.
1643  *
1644  * This may *change* the value of transaction_t->t_datalist, so anyone
1645  * who looks at t_datalist needs to lock against this function.
1646  *
1647  * Even worse, someone may be doing a journal_dirty_data on this
1648  * buffer.  So we need to lock against that.  journal_dirty_data()
1649  * will come out of the lock with the buffer dirty, which makes it
1650  * ineligible for release here.
1651  *
1652  * Who else is affected by this?  hmm...  Really the only contender
1653  * is do_get_write_access() - it could be looking at the buffer while
1654  * journal_try_to_free_buffer() is changing its state.  But that
1655  * cannot happen because we never reallocate freed data as metadata
1656  * while the data is part of a transaction.  Yes?
1657  */
1658 int journal_try_to_free_buffers(journal_t *journal,
1659                                 struct page *page, gfp_t unused_gfp_mask)
1660 {
1661         struct buffer_head *head;
1662         struct buffer_head *bh;
1663         int ret = 0;
1664
1665         J_ASSERT(PageLocked(page));
1666
1667         head = page_buffers(page);
1668         bh = head;
1669         do {
1670                 struct journal_head *jh;
1671
1672                 /*
1673                  * We take our own ref against the journal_head here to avoid
1674                  * having to add tons of locking around each instance of
1675                  * journal_remove_journal_head() and journal_put_journal_head().
1676                  */
1677                 jh = journal_grab_journal_head(bh);
1678                 if (!jh)
1679                         continue;
1680
1681                 jbd_lock_bh_state(bh);
1682                 __journal_try_to_free_buffer(journal, bh);
1683                 journal_put_journal_head(jh);
1684                 jbd_unlock_bh_state(bh);
1685                 if (buffer_jbd(bh))
1686                         goto busy;
1687         } while ((bh = bh->b_this_page) != head);
1688         ret = try_to_free_buffers(page);
1689 busy:
1690         return ret;
1691 }
1692
1693 /*
1694  * This buffer is no longer needed.  If it is on an older transaction's
1695  * checkpoint list we need to record it on this transaction's forget list
1696  * to pin this buffer (and hence its checkpointing transaction) down until
1697  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1698  * release it.
1699  * Returns non-zero if JBD no longer has an interest in the buffer.
1700  *
1701  * Called under j_list_lock.
1702  *
1703  * Called under jbd_lock_bh_state(bh).
1704  */
1705 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1706 {
1707         int may_free = 1;
1708         struct buffer_head *bh = jh2bh(jh);
1709
1710         __journal_unfile_buffer(jh);
1711
1712         if (jh->b_cp_transaction) {
1713                 JBUFFER_TRACE(jh, "on running+cp transaction");
1714                 __journal_file_buffer(jh, transaction, BJ_Forget);
1715                 clear_buffer_jbddirty(bh);
1716                 may_free = 0;
1717         } else {
1718                 JBUFFER_TRACE(jh, "on running transaction");
1719                 journal_remove_journal_head(bh);
1720                 __brelse(bh);
1721         }
1722         return may_free;
1723 }
1724
1725 /*
1726  * journal_invalidatepage
1727  *
1728  * This code is tricky.  It has a number of cases to deal with.
1729  *
1730  * There are two invariants which this code relies on:
1731  *
1732  * i_size must be updated on disk before we start calling invalidatepage on the
1733  * data.
1734  *
1735  *  This is done in ext3 by defining an ext3_setattr method which
1736  *  updates i_size before truncate gets going.  By maintaining this
1737  *  invariant, we can be sure that it is safe to throw away any buffers
1738  *  attached to the current transaction: once the transaction commits,
1739  *  we know that the data will not be needed.
1740  *
1741  *  Note however that we can *not* throw away data belonging to the
1742  *  previous, committing transaction!
1743  *
1744  * Any disk blocks which *are* part of the previous, committing
1745  * transaction (and which therefore cannot be discarded immediately) are
1746  * not going to be reused in the new running transaction
1747  *
1748  *  The bitmap committed_data images guarantee this: any block which is
1749  *  allocated in one transaction and removed in the next will be marked
1750  *  as in-use in the committed_data bitmap, so cannot be reused until
1751  *  the next transaction to delete the block commits.  This means that
1752  *  leaving committing buffers dirty is quite safe: the disk blocks
1753  *  cannot be reallocated to a different file and so buffer aliasing is
1754  *  not possible.
1755  *
1756  *
1757  * The above applies mainly to ordered data mode.  In writeback mode we
1758  * don't make guarantees about the order in which data hits disk --- in
1759  * particular we don't guarantee that new dirty data is flushed before
1760  * transaction commit --- so it is always safe just to discard data
1761  * immediately in that mode.  --sct
1762  */
1763
1764 /*
1765  * The journal_unmap_buffer helper function returns zero if the buffer
1766  * concerned remains pinned as an anonymous buffer belonging to an older
1767  * transaction.
1768  *
1769  * We're outside-transaction here.  Either or both of j_running_transaction
1770  * and j_committing_transaction may be NULL.
1771  */
1772 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1773 {
1774         transaction_t *transaction;
1775         struct journal_head *jh;
1776         int may_free = 1;
1777         int ret;
1778
1779         BUFFER_TRACE(bh, "entry");
1780
1781         /*
1782          * It is safe to proceed here without the j_list_lock because the
1783          * buffers cannot be stolen by try_to_free_buffers as long as we are
1784          * holding the page lock. --sct
1785          */
1786
1787         if (!buffer_jbd(bh))
1788                 goto zap_buffer_unlocked;
1789
1790         spin_lock(&journal->j_state_lock);
1791         jbd_lock_bh_state(bh);
1792         spin_lock(&journal->j_list_lock);
1793
1794         jh = journal_grab_journal_head(bh);
1795         if (!jh)
1796                 goto zap_buffer_no_jh;
1797
1798         transaction = jh->b_transaction;
1799         if (transaction == NULL) {
1800                 /* First case: not on any transaction.  If it
1801                  * has no checkpoint link, then we can zap it:
1802                  * it's a writeback-mode buffer so we don't care
1803                  * if it hits disk safely. */
1804                 if (!jh->b_cp_transaction) {
1805                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1806                         goto zap_buffer;
1807                 }
1808
1809                 if (!buffer_dirty(bh)) {
1810                         /* bdflush has written it.  We can drop it now */
1811                         goto zap_buffer;
1812                 }
1813
1814                 /* OK, it must be in the journal but still not
1815                  * written fully to disk: it's metadata or
1816                  * journaled data... */
1817
1818                 if (journal->j_running_transaction) {
1819                         /* ... and once the current transaction has
1820                          * committed, the buffer won't be needed any
1821                          * longer. */
1822                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1823                         ret = __dispose_buffer(jh,
1824                                         journal->j_running_transaction);
1825                         journal_put_journal_head(jh);
1826                         spin_unlock(&journal->j_list_lock);
1827                         jbd_unlock_bh_state(bh);
1828                         spin_unlock(&journal->j_state_lock);
1829                         return ret;
1830                 } else {
1831                         /* There is no currently-running transaction. So the
1832                          * orphan record which we wrote for this file must have
1833                          * passed into commit.  We must attach this buffer to
1834                          * the committing transaction, if it exists. */
1835                         if (journal->j_committing_transaction) {
1836                                 JBUFFER_TRACE(jh, "give to committing trans");
1837                                 ret = __dispose_buffer(jh,
1838                                         journal->j_committing_transaction);
1839                                 journal_put_journal_head(jh);
1840                                 spin_unlock(&journal->j_list_lock);
1841                                 jbd_unlock_bh_state(bh);
1842                                 spin_unlock(&journal->j_state_lock);
1843                                 return ret;
1844                         } else {
1845                                 /* The orphan record's transaction has
1846                                  * committed.  We can cleanse this buffer */
1847                                 clear_buffer_jbddirty(bh);
1848                                 goto zap_buffer;
1849                         }
1850                 }
1851         } else if (transaction == journal->j_committing_transaction) {
1852                 JBUFFER_TRACE(jh, "on committing transaction");
1853                 if (jh->b_jlist == BJ_Locked) {
1854                         /*
1855                          * The buffer is on the committing transaction's locked
1856                          * list.  We have the buffer locked, so I/O has
1857                          * completed.  So we can nail the buffer now.
1858                          */
1859                         may_free = __dispose_buffer(jh, transaction);
1860                         goto zap_buffer;
1861                 }
1862                 /*
1863                  * If it is committing, we simply cannot touch it.  We
1864                  * can remove it's next_transaction pointer from the
1865                  * running transaction if that is set, but nothing
1866                  * else. */
1867                 set_buffer_freed(bh);
1868                 if (jh->b_next_transaction) {
1869                         J_ASSERT(jh->b_next_transaction ==
1870                                         journal->j_running_transaction);
1871                         jh->b_next_transaction = NULL;
1872                 }
1873                 journal_put_journal_head(jh);
1874                 spin_unlock(&journal->j_list_lock);
1875                 jbd_unlock_bh_state(bh);
1876                 spin_unlock(&journal->j_state_lock);
1877                 return 0;
1878         } else {
1879                 /* Good, the buffer belongs to the running transaction.
1880                  * We are writing our own transaction's data, not any
1881                  * previous one's, so it is safe to throw it away
1882                  * (remember that we expect the filesystem to have set
1883                  * i_size already for this truncate so recovery will not
1884                  * expose the disk blocks we are discarding here.) */
1885                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1886                 JBUFFER_TRACE(jh, "on running transaction");
1887                 may_free = __dispose_buffer(jh, transaction);
1888         }
1889
1890 zap_buffer:
1891         journal_put_journal_head(jh);
1892 zap_buffer_no_jh:
1893         spin_unlock(&journal->j_list_lock);
1894         jbd_unlock_bh_state(bh);
1895         spin_unlock(&journal->j_state_lock);
1896 zap_buffer_unlocked:
1897         clear_buffer_dirty(bh);
1898         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1899         clear_buffer_mapped(bh);
1900         clear_buffer_req(bh);
1901         clear_buffer_new(bh);
1902         bh->b_bdev = NULL;
1903         return may_free;
1904 }
1905
1906 /**
1907  * void journal_invalidatepage() - invalidate a journal page
1908  * @journal: journal to use for flush
1909  * @page:    page to flush
1910  * @offset:  length of page to invalidate.
1911  *
1912  * Reap page buffers containing data after offset in page.
1913  */
1914 void journal_invalidatepage(journal_t *journal,
1915                       struct page *page,
1916                       unsigned long offset)
1917 {
1918         struct buffer_head *head, *bh, *next;
1919         unsigned int curr_off = 0;
1920         int may_free = 1;
1921
1922         if (!PageLocked(page))
1923                 BUG();
1924         if (!page_has_buffers(page))
1925                 return;
1926
1927         /* We will potentially be playing with lists other than just the
1928          * data lists (especially for journaled data mode), so be
1929          * cautious in our locking. */
1930
1931         head = bh = page_buffers(page);
1932         do {
1933                 unsigned int next_off = curr_off + bh->b_size;
1934                 next = bh->b_this_page;
1935
1936                 if (offset <= curr_off) {
1937                         /* This block is wholly outside the truncation point */
1938                         lock_buffer(bh);
1939                         may_free &= journal_unmap_buffer(journal, bh);
1940                         unlock_buffer(bh);
1941                 }
1942                 curr_off = next_off;
1943                 bh = next;
1944
1945         } while (bh != head);
1946
1947         if (!offset) {
1948                 if (may_free && try_to_free_buffers(page))
1949                         J_ASSERT(!page_has_buffers(page));
1950         }
1951 }
1952
1953 /*
1954  * File a buffer on the given transaction list.
1955  */
1956 void __journal_file_buffer(struct journal_head *jh,
1957                         transaction_t *transaction, int jlist)
1958 {
1959         struct journal_head **list = NULL;
1960         int was_dirty = 0;
1961         struct buffer_head *bh = jh2bh(jh);
1962
1963         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1964         assert_spin_locked(&transaction->t_journal->j_list_lock);
1965
1966         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1967         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1968                                 jh->b_transaction == NULL);
1969
1970         if (jh->b_transaction && jh->b_jlist == jlist)
1971                 return;
1972
1973         /* The following list of buffer states needs to be consistent
1974          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1975          * state. */
1976
1977         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1978             jlist == BJ_Shadow || jlist == BJ_Forget) {
1979                 if (test_clear_buffer_dirty(bh) ||
1980                     test_clear_buffer_jbddirty(bh))
1981                         was_dirty = 1;
1982         }
1983
1984         if (jh->b_transaction)
1985                 __journal_temp_unlink_buffer(jh);
1986         jh->b_transaction = transaction;
1987
1988         switch (jlist) {
1989         case BJ_None:
1990                 J_ASSERT_JH(jh, !jh->b_committed_data);
1991                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1992                 return;
1993         case BJ_SyncData:
1994                 list = &transaction->t_sync_datalist;
1995                 break;
1996         case BJ_Metadata:
1997                 transaction->t_nr_buffers++;
1998                 list = &transaction->t_buffers;
1999                 break;
2000         case BJ_Forget:
2001                 list = &transaction->t_forget;
2002                 break;
2003         case BJ_IO:
2004                 list = &transaction->t_iobuf_list;
2005                 break;
2006         case BJ_Shadow:
2007                 list = &transaction->t_shadow_list;
2008                 break;
2009         case BJ_LogCtl:
2010                 list = &transaction->t_log_list;
2011                 break;
2012         case BJ_Reserved:
2013                 list = &transaction->t_reserved_list;
2014                 break;
2015         case BJ_Locked:
2016                 list =  &transaction->t_locked_list;
2017                 break;
2018         }
2019
2020         __blist_add_buffer(list, jh);
2021         jh->b_jlist = jlist;
2022
2023         if (was_dirty)
2024                 set_buffer_jbddirty(bh);
2025 }
2026
2027 void journal_file_buffer(struct journal_head *jh,
2028                                 transaction_t *transaction, int jlist)
2029 {
2030         jbd_lock_bh_state(jh2bh(jh));
2031         spin_lock(&transaction->t_journal->j_list_lock);
2032         __journal_file_buffer(jh, transaction, jlist);
2033         spin_unlock(&transaction->t_journal->j_list_lock);
2034         jbd_unlock_bh_state(jh2bh(jh));
2035 }
2036
2037 /*
2038  * Remove a buffer from its current buffer list in preparation for
2039  * dropping it from its current transaction entirely.  If the buffer has
2040  * already started to be used by a subsequent transaction, refile the
2041  * buffer on that transaction's metadata list.
2042  *
2043  * Called under journal->j_list_lock
2044  *
2045  * Called under jbd_lock_bh_state(jh2bh(jh))
2046  */
2047 void __journal_refile_buffer(struct journal_head *jh)
2048 {
2049         int was_dirty;
2050         struct buffer_head *bh = jh2bh(jh);
2051
2052         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2053         if (jh->b_transaction)
2054                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2055
2056         /* If the buffer is now unused, just drop it. */
2057         if (jh->b_next_transaction == NULL) {
2058                 __journal_unfile_buffer(jh);
2059                 return;
2060         }
2061
2062         /*
2063          * It has been modified by a later transaction: add it to the new
2064          * transaction's metadata list.
2065          */
2066
2067         was_dirty = test_clear_buffer_jbddirty(bh);
2068         __journal_temp_unlink_buffer(jh);
2069         jh->b_transaction = jh->b_next_transaction;
2070         jh->b_next_transaction = NULL;
2071         __journal_file_buffer(jh, jh->b_transaction,
2072                                 was_dirty ? BJ_Metadata : BJ_Reserved);
2073         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2074
2075         if (was_dirty)
2076                 set_buffer_jbddirty(bh);
2077 }
2078
2079 /*
2080  * For the unlocked version of this call, also make sure that any
2081  * hanging journal_head is cleaned up if necessary.
2082  *
2083  * __journal_refile_buffer is usually called as part of a single locked
2084  * operation on a buffer_head, in which the caller is probably going to
2085  * be hooking the journal_head onto other lists.  In that case it is up
2086  * to the caller to remove the journal_head if necessary.  For the
2087  * unlocked journal_refile_buffer call, the caller isn't going to be
2088  * doing anything else to the buffer so we need to do the cleanup
2089  * ourselves to avoid a jh leak.
2090  *
2091  * *** The journal_head may be freed by this call! ***
2092  */
2093 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2094 {
2095         struct buffer_head *bh = jh2bh(jh);
2096
2097         jbd_lock_bh_state(bh);
2098         spin_lock(&journal->j_list_lock);
2099
2100         __journal_refile_buffer(jh);
2101         jbd_unlock_bh_state(bh);
2102         journal_remove_journal_head(bh);
2103
2104         spin_unlock(&journal->j_list_lock);
2105         __brelse(bh);
2106 }