006f9ad838a26aed53ae9f1d54a9065d781e3f2a
[linux-2.6.git] / fs / jbd / transaction.c
1 /*
2  * linux/fs/jbd/transaction.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem transaction handling code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages transactions (compound commits managed by the
16  * journaling code) and handles (individual atomic operations by the
17  * filesystem).
18  */
19
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29
30 static void __journal_temp_unlink_buffer(struct journal_head *jh);
31
32 /*
33  * get_transaction: obtain a new transaction_t object.
34  *
35  * Simply allocate and initialise a new transaction.  Create it in
36  * RUNNING state and add it to the current journal (which should not
37  * have an existing running transaction: we only make a new transaction
38  * once we have started to commit the old one).
39  *
40  * Preconditions:
41  *      The journal MUST be locked.  We don't perform atomic mallocs on the
42  *      new transaction and we can't block without protecting against other
43  *      processes trying to touch the journal while it is in transition.
44  *
45  * Called under j_state_lock
46  */
47
48 static transaction_t *
49 get_transaction(journal_t *journal, transaction_t *transaction)
50 {
51         transaction->t_journal = journal;
52         transaction->t_state = T_RUNNING;
53         transaction->t_start_time = ktime_get();
54         transaction->t_tid = journal->j_transaction_sequence++;
55         transaction->t_expires = jiffies + journal->j_commit_interval;
56         spin_lock_init(&transaction->t_handle_lock);
57
58         /* Set up the commit timer for the new transaction. */
59         journal->j_commit_timer.expires =
60                                 round_jiffies_up(transaction->t_expires);
61         add_timer(&journal->j_commit_timer);
62
63         J_ASSERT(journal->j_running_transaction == NULL);
64         journal->j_running_transaction = transaction;
65
66         return transaction;
67 }
68
69 /*
70  * Handle management.
71  *
72  * A handle_t is an object which represents a single atomic update to a
73  * filesystem, and which tracks all of the modifications which form part
74  * of that one update.
75  */
76
77 /*
78  * start_this_handle: Given a handle, deal with any locking or stalling
79  * needed to make sure that there is enough journal space for the handle
80  * to begin.  Attach the handle to a transaction and set up the
81  * transaction's buffer credits.
82  */
83
84 static int start_this_handle(journal_t *journal, handle_t *handle)
85 {
86         transaction_t *transaction;
87         int needed;
88         int nblocks = handle->h_buffer_credits;
89         transaction_t *new_transaction = NULL;
90         int ret = 0;
91
92         if (nblocks > journal->j_max_transaction_buffers) {
93                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
94                        current->comm, nblocks,
95                        journal->j_max_transaction_buffers);
96                 ret = -ENOSPC;
97                 goto out;
98         }
99
100 alloc_transaction:
101         if (!journal->j_running_transaction) {
102                 new_transaction = kzalloc(sizeof(*new_transaction),
103                                                 GFP_NOFS|__GFP_NOFAIL);
104                 if (!new_transaction) {
105                         ret = -ENOMEM;
106                         goto out;
107                 }
108         }
109
110         jbd_debug(3, "New handle %p going live.\n", handle);
111
112 repeat:
113
114         /*
115          * We need to hold j_state_lock until t_updates has been incremented,
116          * for proper journal barrier handling
117          */
118         spin_lock(&journal->j_state_lock);
119 repeat_locked:
120         if (is_journal_aborted(journal) ||
121             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
122                 spin_unlock(&journal->j_state_lock);
123                 ret = -EROFS;
124                 goto out;
125         }
126
127         /* Wait on the journal's transaction barrier if necessary */
128         if (journal->j_barrier_count) {
129                 spin_unlock(&journal->j_state_lock);
130                 wait_event(journal->j_wait_transaction_locked,
131                                 journal->j_barrier_count == 0);
132                 goto repeat;
133         }
134
135         if (!journal->j_running_transaction) {
136                 if (!new_transaction) {
137                         spin_unlock(&journal->j_state_lock);
138                         goto alloc_transaction;
139                 }
140                 get_transaction(journal, new_transaction);
141                 new_transaction = NULL;
142         }
143
144         transaction = journal->j_running_transaction;
145
146         /*
147          * If the current transaction is locked down for commit, wait for the
148          * lock to be released.
149          */
150         if (transaction->t_state == T_LOCKED) {
151                 DEFINE_WAIT(wait);
152
153                 prepare_to_wait(&journal->j_wait_transaction_locked,
154                                         &wait, TASK_UNINTERRUPTIBLE);
155                 spin_unlock(&journal->j_state_lock);
156                 schedule();
157                 finish_wait(&journal->j_wait_transaction_locked, &wait);
158                 goto repeat;
159         }
160
161         /*
162          * If there is not enough space left in the log to write all potential
163          * buffers requested by this operation, we need to stall pending a log
164          * checkpoint to free some more log space.
165          */
166         spin_lock(&transaction->t_handle_lock);
167         needed = transaction->t_outstanding_credits + nblocks;
168
169         if (needed > journal->j_max_transaction_buffers) {
170                 /*
171                  * If the current transaction is already too large, then start
172                  * to commit it: we can then go back and attach this handle to
173                  * a new transaction.
174                  */
175                 DEFINE_WAIT(wait);
176
177                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
178                 spin_unlock(&transaction->t_handle_lock);
179                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
180                                 TASK_UNINTERRUPTIBLE);
181                 __log_start_commit(journal, transaction->t_tid);
182                 spin_unlock(&journal->j_state_lock);
183                 schedule();
184                 finish_wait(&journal->j_wait_transaction_locked, &wait);
185                 goto repeat;
186         }
187
188         /*
189          * The commit code assumes that it can get enough log space
190          * without forcing a checkpoint.  This is *critical* for
191          * correctness: a checkpoint of a buffer which is also
192          * associated with a committing transaction creates a deadlock,
193          * so commit simply cannot force through checkpoints.
194          *
195          * We must therefore ensure the necessary space in the journal
196          * *before* starting to dirty potentially checkpointed buffers
197          * in the new transaction.
198          *
199          * The worst part is, any transaction currently committing can
200          * reduce the free space arbitrarily.  Be careful to account for
201          * those buffers when checkpointing.
202          */
203
204         /*
205          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
206          * a _lot_ of headroom: 1/4 of the journal plus the size of
207          * the committing transaction.  Really, we only need to give it
208          * committing_transaction->t_outstanding_credits plus "enough" for
209          * the log control blocks.
210          * Also, this test is inconsitent with the matching one in
211          * journal_extend().
212          */
213         if (__log_space_left(journal) < jbd_space_needed(journal)) {
214                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
215                 spin_unlock(&transaction->t_handle_lock);
216                 __log_wait_for_space(journal);
217                 goto repeat_locked;
218         }
219
220         /* OK, account for the buffers that this operation expects to
221          * use and add the handle to the running transaction. */
222
223         handle->h_transaction = transaction;
224         transaction->t_outstanding_credits += nblocks;
225         transaction->t_updates++;
226         transaction->t_handle_count++;
227         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
228                   handle, nblocks, transaction->t_outstanding_credits,
229                   __log_space_left(journal));
230         spin_unlock(&transaction->t_handle_lock);
231         spin_unlock(&journal->j_state_lock);
232
233         lock_map_acquire(&handle->h_lockdep_map);
234 out:
235         if (unlikely(new_transaction))          /* It's usually NULL */
236                 kfree(new_transaction);
237         return ret;
238 }
239
240 static struct lock_class_key jbd_handle_key;
241
242 /* Allocate a new handle.  This should probably be in a slab... */
243 static handle_t *new_handle(int nblocks)
244 {
245         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
246         if (!handle)
247                 return NULL;
248         memset(handle, 0, sizeof(*handle));
249         handle->h_buffer_credits = nblocks;
250         handle->h_ref = 1;
251
252         lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
253
254         return handle;
255 }
256
257 /**
258  * handle_t *journal_start() - Obtain a new handle.
259  * @journal: Journal to start transaction on.
260  * @nblocks: number of block buffer we might modify
261  *
262  * We make sure that the transaction can guarantee at least nblocks of
263  * modified buffers in the log.  We block until the log can guarantee
264  * that much space.
265  *
266  * This function is visible to journal users (like ext3fs), so is not
267  * called with the journal already locked.
268  *
269  * Return a pointer to a newly allocated handle, or NULL on failure
270  */
271 handle_t *journal_start(journal_t *journal, int nblocks)
272 {
273         handle_t *handle = journal_current_handle();
274         int err;
275
276         if (!journal)
277                 return ERR_PTR(-EROFS);
278
279         if (handle) {
280                 J_ASSERT(handle->h_transaction->t_journal == journal);
281                 handle->h_ref++;
282                 return handle;
283         }
284
285         handle = new_handle(nblocks);
286         if (!handle)
287                 return ERR_PTR(-ENOMEM);
288
289         current->journal_info = handle;
290
291         err = start_this_handle(journal, handle);
292         if (err < 0) {
293                 jbd_free_handle(handle);
294                 current->journal_info = NULL;
295                 handle = ERR_PTR(err);
296                 goto out;
297         }
298 out:
299         return handle;
300 }
301
302 /**
303  * int journal_extend() - extend buffer credits.
304  * @handle:  handle to 'extend'
305  * @nblocks: nr blocks to try to extend by.
306  *
307  * Some transactions, such as large extends and truncates, can be done
308  * atomically all at once or in several stages.  The operation requests
309  * a credit for a number of buffer modications in advance, but can
310  * extend its credit if it needs more.
311  *
312  * journal_extend tries to give the running handle more buffer credits.
313  * It does not guarantee that allocation - this is a best-effort only.
314  * The calling process MUST be able to deal cleanly with a failure to
315  * extend here.
316  *
317  * Return 0 on success, non-zero on failure.
318  *
319  * return code < 0 implies an error
320  * return code > 0 implies normal transaction-full status.
321  */
322 int journal_extend(handle_t *handle, int nblocks)
323 {
324         transaction_t *transaction = handle->h_transaction;
325         journal_t *journal = transaction->t_journal;
326         int result;
327         int wanted;
328
329         result = -EIO;
330         if (is_handle_aborted(handle))
331                 goto out;
332
333         result = 1;
334
335         spin_lock(&journal->j_state_lock);
336
337         /* Don't extend a locked-down transaction! */
338         if (handle->h_transaction->t_state != T_RUNNING) {
339                 jbd_debug(3, "denied handle %p %d blocks: "
340                           "transaction not running\n", handle, nblocks);
341                 goto error_out;
342         }
343
344         spin_lock(&transaction->t_handle_lock);
345         wanted = transaction->t_outstanding_credits + nblocks;
346
347         if (wanted > journal->j_max_transaction_buffers) {
348                 jbd_debug(3, "denied handle %p %d blocks: "
349                           "transaction too large\n", handle, nblocks);
350                 goto unlock;
351         }
352
353         if (wanted > __log_space_left(journal)) {
354                 jbd_debug(3, "denied handle %p %d blocks: "
355                           "insufficient log space\n", handle, nblocks);
356                 goto unlock;
357         }
358
359         handle->h_buffer_credits += nblocks;
360         transaction->t_outstanding_credits += nblocks;
361         result = 0;
362
363         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
364 unlock:
365         spin_unlock(&transaction->t_handle_lock);
366 error_out:
367         spin_unlock(&journal->j_state_lock);
368 out:
369         return result;
370 }
371
372
373 /**
374  * int journal_restart() - restart a handle.
375  * @handle:  handle to restart
376  * @nblocks: nr credits requested
377  *
378  * Restart a handle for a multi-transaction filesystem
379  * operation.
380  *
381  * If the journal_extend() call above fails to grant new buffer credits
382  * to a running handle, a call to journal_restart will commit the
383  * handle's transaction so far and reattach the handle to a new
384  * transaction capabable of guaranteeing the requested number of
385  * credits.
386  */
387
388 int journal_restart(handle_t *handle, int nblocks)
389 {
390         transaction_t *transaction = handle->h_transaction;
391         journal_t *journal = transaction->t_journal;
392         int ret;
393
394         /* If we've had an abort of any type, don't even think about
395          * actually doing the restart! */
396         if (is_handle_aborted(handle))
397                 return 0;
398
399         /*
400          * First unlink the handle from its current transaction, and start the
401          * commit on that.
402          */
403         J_ASSERT(transaction->t_updates > 0);
404         J_ASSERT(journal_current_handle() == handle);
405
406         spin_lock(&journal->j_state_lock);
407         spin_lock(&transaction->t_handle_lock);
408         transaction->t_outstanding_credits -= handle->h_buffer_credits;
409         transaction->t_updates--;
410
411         if (!transaction->t_updates)
412                 wake_up(&journal->j_wait_updates);
413         spin_unlock(&transaction->t_handle_lock);
414
415         jbd_debug(2, "restarting handle %p\n", handle);
416         __log_start_commit(journal, transaction->t_tid);
417         spin_unlock(&journal->j_state_lock);
418
419         lock_map_release(&handle->h_lockdep_map);
420         handle->h_buffer_credits = nblocks;
421         ret = start_this_handle(journal, handle);
422         return ret;
423 }
424
425
426 /**
427  * void journal_lock_updates () - establish a transaction barrier.
428  * @journal:  Journal to establish a barrier on.
429  *
430  * This locks out any further updates from being started, and blocks
431  * until all existing updates have completed, returning only once the
432  * journal is in a quiescent state with no updates running.
433  *
434  * The journal lock should not be held on entry.
435  */
436 void journal_lock_updates(journal_t *journal)
437 {
438         DEFINE_WAIT(wait);
439
440         spin_lock(&journal->j_state_lock);
441         ++journal->j_barrier_count;
442
443         /* Wait until there are no running updates */
444         while (1) {
445                 transaction_t *transaction = journal->j_running_transaction;
446
447                 if (!transaction)
448                         break;
449
450                 spin_lock(&transaction->t_handle_lock);
451                 if (!transaction->t_updates) {
452                         spin_unlock(&transaction->t_handle_lock);
453                         break;
454                 }
455                 prepare_to_wait(&journal->j_wait_updates, &wait,
456                                 TASK_UNINTERRUPTIBLE);
457                 spin_unlock(&transaction->t_handle_lock);
458                 spin_unlock(&journal->j_state_lock);
459                 schedule();
460                 finish_wait(&journal->j_wait_updates, &wait);
461                 spin_lock(&journal->j_state_lock);
462         }
463         spin_unlock(&journal->j_state_lock);
464
465         /*
466          * We have now established a barrier against other normal updates, but
467          * we also need to barrier against other journal_lock_updates() calls
468          * to make sure that we serialise special journal-locked operations
469          * too.
470          */
471         mutex_lock(&journal->j_barrier);
472 }
473
474 /**
475  * void journal_unlock_updates (journal_t* journal) - release barrier
476  * @journal:  Journal to release the barrier on.
477  *
478  * Release a transaction barrier obtained with journal_lock_updates().
479  *
480  * Should be called without the journal lock held.
481  */
482 void journal_unlock_updates (journal_t *journal)
483 {
484         J_ASSERT(journal->j_barrier_count != 0);
485
486         mutex_unlock(&journal->j_barrier);
487         spin_lock(&journal->j_state_lock);
488         --journal->j_barrier_count;
489         spin_unlock(&journal->j_state_lock);
490         wake_up(&journal->j_wait_transaction_locked);
491 }
492
493 static void warn_dirty_buffer(struct buffer_head *bh)
494 {
495         char b[BDEVNAME_SIZE];
496
497         printk(KERN_WARNING
498                "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
499                "There's a risk of filesystem corruption in case of system "
500                "crash.\n",
501                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
502 }
503
504 /*
505  * If the buffer is already part of the current transaction, then there
506  * is nothing we need to do.  If it is already part of a prior
507  * transaction which we are still committing to disk, then we need to
508  * make sure that we do not overwrite the old copy: we do copy-out to
509  * preserve the copy going to disk.  We also account the buffer against
510  * the handle's metadata buffer credits (unless the buffer is already
511  * part of the transaction, that is).
512  *
513  */
514 static int
515 do_get_write_access(handle_t *handle, struct journal_head *jh,
516                         int force_copy)
517 {
518         struct buffer_head *bh;
519         transaction_t *transaction;
520         journal_t *journal;
521         int error;
522         char *frozen_buffer = NULL;
523         int need_copy = 0;
524
525         if (is_handle_aborted(handle))
526                 return -EROFS;
527
528         transaction = handle->h_transaction;
529         journal = transaction->t_journal;
530
531         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
532
533         JBUFFER_TRACE(jh, "entry");
534 repeat:
535         bh = jh2bh(jh);
536
537         /* @@@ Need to check for errors here at some point. */
538
539         lock_buffer(bh);
540         jbd_lock_bh_state(bh);
541
542         /* We now hold the buffer lock so it is safe to query the buffer
543          * state.  Is the buffer dirty?
544          *
545          * If so, there are two possibilities.  The buffer may be
546          * non-journaled, and undergoing a quite legitimate writeback.
547          * Otherwise, it is journaled, and we don't expect dirty buffers
548          * in that state (the buffers should be marked JBD_Dirty
549          * instead.)  So either the IO is being done under our own
550          * control and this is a bug, or it's a third party IO such as
551          * dump(8) (which may leave the buffer scheduled for read ---
552          * ie. locked but not dirty) or tune2fs (which may actually have
553          * the buffer dirtied, ugh.)  */
554
555         if (buffer_dirty(bh)) {
556                 /*
557                  * First question: is this buffer already part of the current
558                  * transaction or the existing committing transaction?
559                  */
560                 if (jh->b_transaction) {
561                         J_ASSERT_JH(jh,
562                                 jh->b_transaction == transaction ||
563                                 jh->b_transaction ==
564                                         journal->j_committing_transaction);
565                         if (jh->b_next_transaction)
566                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
567                                                         transaction);
568                         warn_dirty_buffer(bh);
569                 }
570                 /*
571                  * In any case we need to clean the dirty flag and we must
572                  * do it under the buffer lock to be sure we don't race
573                  * with running write-out.
574                  */
575                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
576                 clear_buffer_dirty(bh);
577                 set_buffer_jbddirty(bh);
578         }
579
580         unlock_buffer(bh);
581
582         error = -EROFS;
583         if (is_handle_aborted(handle)) {
584                 jbd_unlock_bh_state(bh);
585                 goto out;
586         }
587         error = 0;
588
589         /*
590          * The buffer is already part of this transaction if b_transaction or
591          * b_next_transaction points to it
592          */
593         if (jh->b_transaction == transaction ||
594             jh->b_next_transaction == transaction)
595                 goto done;
596
597         /*
598          * this is the first time this transaction is touching this buffer,
599          * reset the modified flag
600          */
601         jh->b_modified = 0;
602
603         /*
604          * If there is already a copy-out version of this buffer, then we don't
605          * need to make another one
606          */
607         if (jh->b_frozen_data) {
608                 JBUFFER_TRACE(jh, "has frozen data");
609                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
610                 jh->b_next_transaction = transaction;
611                 goto done;
612         }
613
614         /* Is there data here we need to preserve? */
615
616         if (jh->b_transaction && jh->b_transaction != transaction) {
617                 JBUFFER_TRACE(jh, "owned by older transaction");
618                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
619                 J_ASSERT_JH(jh, jh->b_transaction ==
620                                         journal->j_committing_transaction);
621
622                 /* There is one case we have to be very careful about.
623                  * If the committing transaction is currently writing
624                  * this buffer out to disk and has NOT made a copy-out,
625                  * then we cannot modify the buffer contents at all
626                  * right now.  The essence of copy-out is that it is the
627                  * extra copy, not the primary copy, which gets
628                  * journaled.  If the primary copy is already going to
629                  * disk then we cannot do copy-out here. */
630
631                 if (jh->b_jlist == BJ_Shadow) {
632                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
633                         wait_queue_head_t *wqh;
634
635                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
636
637                         JBUFFER_TRACE(jh, "on shadow: sleep");
638                         jbd_unlock_bh_state(bh);
639                         /* commit wakes up all shadow buffers after IO */
640                         for ( ; ; ) {
641                                 prepare_to_wait(wqh, &wait.wait,
642                                                 TASK_UNINTERRUPTIBLE);
643                                 if (jh->b_jlist != BJ_Shadow)
644                                         break;
645                                 schedule();
646                         }
647                         finish_wait(wqh, &wait.wait);
648                         goto repeat;
649                 }
650
651                 /* Only do the copy if the currently-owning transaction
652                  * still needs it.  If it is on the Forget list, the
653                  * committing transaction is past that stage.  The
654                  * buffer had better remain locked during the kmalloc,
655                  * but that should be true --- we hold the journal lock
656                  * still and the buffer is already on the BUF_JOURNAL
657                  * list so won't be flushed.
658                  *
659                  * Subtle point, though: if this is a get_undo_access,
660                  * then we will be relying on the frozen_data to contain
661                  * the new value of the committed_data record after the
662                  * transaction, so we HAVE to force the frozen_data copy
663                  * in that case. */
664
665                 if (jh->b_jlist != BJ_Forget || force_copy) {
666                         JBUFFER_TRACE(jh, "generate frozen data");
667                         if (!frozen_buffer) {
668                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
669                                 jbd_unlock_bh_state(bh);
670                                 frozen_buffer =
671                                         jbd_alloc(jh2bh(jh)->b_size,
672                                                          GFP_NOFS);
673                                 if (!frozen_buffer) {
674                                         printk(KERN_EMERG
675                                                "%s: OOM for frozen_buffer\n",
676                                                __func__);
677                                         JBUFFER_TRACE(jh, "oom!");
678                                         error = -ENOMEM;
679                                         jbd_lock_bh_state(bh);
680                                         goto done;
681                                 }
682                                 goto repeat;
683                         }
684                         jh->b_frozen_data = frozen_buffer;
685                         frozen_buffer = NULL;
686                         need_copy = 1;
687                 }
688                 jh->b_next_transaction = transaction;
689         }
690
691
692         /*
693          * Finally, if the buffer is not journaled right now, we need to make
694          * sure it doesn't get written to disk before the caller actually
695          * commits the new data
696          */
697         if (!jh->b_transaction) {
698                 JBUFFER_TRACE(jh, "no transaction");
699                 J_ASSERT_JH(jh, !jh->b_next_transaction);
700                 jh->b_transaction = transaction;
701                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
702                 spin_lock(&journal->j_list_lock);
703                 __journal_file_buffer(jh, transaction, BJ_Reserved);
704                 spin_unlock(&journal->j_list_lock);
705         }
706
707 done:
708         if (need_copy) {
709                 struct page *page;
710                 int offset;
711                 char *source;
712
713                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
714                             "Possible IO failure.\n");
715                 page = jh2bh(jh)->b_page;
716                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
717                 source = kmap_atomic(page, KM_USER0);
718                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
719                 kunmap_atomic(source, KM_USER0);
720         }
721         jbd_unlock_bh_state(bh);
722
723         /*
724          * If we are about to journal a buffer, then any revoke pending on it is
725          * no longer valid
726          */
727         journal_cancel_revoke(handle, jh);
728
729 out:
730         if (unlikely(frozen_buffer))    /* It's usually NULL */
731                 jbd_free(frozen_buffer, bh->b_size);
732
733         JBUFFER_TRACE(jh, "exit");
734         return error;
735 }
736
737 /**
738  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
739  * @handle: transaction to add buffer modifications to
740  * @bh:     bh to be used for metadata writes
741  *
742  * Returns an error code or 0 on success.
743  *
744  * In full data journalling mode the buffer may be of type BJ_AsyncData,
745  * because we're write()ing a buffer which is also part of a shared mapping.
746  */
747
748 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
749 {
750         struct journal_head *jh = journal_add_journal_head(bh);
751         int rc;
752
753         /* We do not want to get caught playing with fields which the
754          * log thread also manipulates.  Make sure that the buffer
755          * completes any outstanding IO before proceeding. */
756         rc = do_get_write_access(handle, jh, 0);
757         journal_put_journal_head(jh);
758         return rc;
759 }
760
761
762 /*
763  * When the user wants to journal a newly created buffer_head
764  * (ie. getblk() returned a new buffer and we are going to populate it
765  * manually rather than reading off disk), then we need to keep the
766  * buffer_head locked until it has been completely filled with new
767  * data.  In this case, we should be able to make the assertion that
768  * the bh is not already part of an existing transaction.
769  *
770  * The buffer should already be locked by the caller by this point.
771  * There is no lock ranking violation: it was a newly created,
772  * unlocked buffer beforehand. */
773
774 /**
775  * int journal_get_create_access () - notify intent to use newly created bh
776  * @handle: transaction to new buffer to
777  * @bh: new buffer.
778  *
779  * Call this if you create a new bh.
780  */
781 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
782 {
783         transaction_t *transaction = handle->h_transaction;
784         journal_t *journal = transaction->t_journal;
785         struct journal_head *jh = journal_add_journal_head(bh);
786         int err;
787
788         jbd_debug(5, "journal_head %p\n", jh);
789         err = -EROFS;
790         if (is_handle_aborted(handle))
791                 goto out;
792         err = 0;
793
794         JBUFFER_TRACE(jh, "entry");
795         /*
796          * The buffer may already belong to this transaction due to pre-zeroing
797          * in the filesystem's new_block code.  It may also be on the previous,
798          * committing transaction's lists, but it HAS to be in Forget state in
799          * that case: the transaction must have deleted the buffer for it to be
800          * reused here.
801          */
802         jbd_lock_bh_state(bh);
803         spin_lock(&journal->j_list_lock);
804         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
805                 jh->b_transaction == NULL ||
806                 (jh->b_transaction == journal->j_committing_transaction &&
807                           jh->b_jlist == BJ_Forget)));
808
809         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
810         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
811
812         if (jh->b_transaction == NULL) {
813                 /*
814                  * Previous journal_forget() could have left the buffer
815                  * with jbddirty bit set because it was being committed. When
816                  * the commit finished, we've filed the buffer for
817                  * checkpointing and marked it dirty. Now we are reallocating
818                  * the buffer so the transaction freeing it must have
819                  * committed and so it's safe to clear the dirty bit.
820                  */
821                 clear_buffer_dirty(jh2bh(jh));
822                 jh->b_transaction = transaction;
823
824                 /* first access by this transaction */
825                 jh->b_modified = 0;
826
827                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
828                 __journal_file_buffer(jh, transaction, BJ_Reserved);
829         } else if (jh->b_transaction == journal->j_committing_transaction) {
830                 /* first access by this transaction */
831                 jh->b_modified = 0;
832
833                 JBUFFER_TRACE(jh, "set next transaction");
834                 jh->b_next_transaction = transaction;
835         }
836         spin_unlock(&journal->j_list_lock);
837         jbd_unlock_bh_state(bh);
838
839         /*
840          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
841          * blocks which contain freed but then revoked metadata.  We need
842          * to cancel the revoke in case we end up freeing it yet again
843          * and the reallocating as data - this would cause a second revoke,
844          * which hits an assertion error.
845          */
846         JBUFFER_TRACE(jh, "cancelling revoke");
847         journal_cancel_revoke(handle, jh);
848         journal_put_journal_head(jh);
849 out:
850         return err;
851 }
852
853 /**
854  * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
855  * @handle: transaction
856  * @bh: buffer to undo
857  *
858  * Sometimes there is a need to distinguish between metadata which has
859  * been committed to disk and that which has not.  The ext3fs code uses
860  * this for freeing and allocating space, we have to make sure that we
861  * do not reuse freed space until the deallocation has been committed,
862  * since if we overwrote that space we would make the delete
863  * un-rewindable in case of a crash.
864  *
865  * To deal with that, journal_get_undo_access requests write access to a
866  * buffer for parts of non-rewindable operations such as delete
867  * operations on the bitmaps.  The journaling code must keep a copy of
868  * the buffer's contents prior to the undo_access call until such time
869  * as we know that the buffer has definitely been committed to disk.
870  *
871  * We never need to know which transaction the committed data is part
872  * of, buffers touched here are guaranteed to be dirtied later and so
873  * will be committed to a new transaction in due course, at which point
874  * we can discard the old committed data pointer.
875  *
876  * Returns error number or 0 on success.
877  */
878 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
879 {
880         int err;
881         struct journal_head *jh = journal_add_journal_head(bh);
882         char *committed_data = NULL;
883
884         JBUFFER_TRACE(jh, "entry");
885
886         /*
887          * Do this first --- it can drop the journal lock, so we want to
888          * make sure that obtaining the committed_data is done
889          * atomically wrt. completion of any outstanding commits.
890          */
891         err = do_get_write_access(handle, jh, 1);
892         if (err)
893                 goto out;
894
895 repeat:
896         if (!jh->b_committed_data) {
897                 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
898                 if (!committed_data) {
899                         printk(KERN_EMERG "%s: No memory for committed data\n",
900                                 __func__);
901                         err = -ENOMEM;
902                         goto out;
903                 }
904         }
905
906         jbd_lock_bh_state(bh);
907         if (!jh->b_committed_data) {
908                 /* Copy out the current buffer contents into the
909                  * preserved, committed copy. */
910                 JBUFFER_TRACE(jh, "generate b_committed data");
911                 if (!committed_data) {
912                         jbd_unlock_bh_state(bh);
913                         goto repeat;
914                 }
915
916                 jh->b_committed_data = committed_data;
917                 committed_data = NULL;
918                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
919         }
920         jbd_unlock_bh_state(bh);
921 out:
922         journal_put_journal_head(jh);
923         if (unlikely(committed_data))
924                 jbd_free(committed_data, bh->b_size);
925         return err;
926 }
927
928 /**
929  * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
930  * @handle: transaction
931  * @bh: bufferhead to mark
932  *
933  * Description:
934  * Mark a buffer as containing dirty data which needs to be flushed before
935  * we can commit the current transaction.
936  *
937  * The buffer is placed on the transaction's data list and is marked as
938  * belonging to the transaction.
939  *
940  * Returns error number or 0 on success.
941  *
942  * journal_dirty_data() can be called via page_launder->ext3_writepage
943  * by kswapd.
944  */
945 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
946 {
947         journal_t *journal = handle->h_transaction->t_journal;
948         int need_brelse = 0;
949         struct journal_head *jh;
950         int ret = 0;
951
952         if (is_handle_aborted(handle))
953                 return ret;
954
955         jh = journal_add_journal_head(bh);
956         JBUFFER_TRACE(jh, "entry");
957
958         /*
959          * The buffer could *already* be dirty.  Writeout can start
960          * at any time.
961          */
962         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
963
964         /*
965          * What if the buffer is already part of a running transaction?
966          *
967          * There are two cases:
968          * 1) It is part of the current running transaction.  Refile it,
969          *    just in case we have allocated it as metadata, deallocated
970          *    it, then reallocated it as data.
971          * 2) It is part of the previous, still-committing transaction.
972          *    If all we want to do is to guarantee that the buffer will be
973          *    written to disk before this new transaction commits, then
974          *    being sure that the *previous* transaction has this same
975          *    property is sufficient for us!  Just leave it on its old
976          *    transaction.
977          *
978          * In case (2), the buffer must not already exist as metadata
979          * --- that would violate write ordering (a transaction is free
980          * to write its data at any point, even before the previous
981          * committing transaction has committed).  The caller must
982          * never, ever allow this to happen: there's nothing we can do
983          * about it in this layer.
984          */
985         jbd_lock_bh_state(bh);
986         spin_lock(&journal->j_list_lock);
987
988         /* Now that we have bh_state locked, are we really still mapped? */
989         if (!buffer_mapped(bh)) {
990                 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
991                 goto no_journal;
992         }
993
994         if (jh->b_transaction) {
995                 JBUFFER_TRACE(jh, "has transaction");
996                 if (jh->b_transaction != handle->h_transaction) {
997                         JBUFFER_TRACE(jh, "belongs to older transaction");
998                         J_ASSERT_JH(jh, jh->b_transaction ==
999                                         journal->j_committing_transaction);
1000
1001                         /* @@@ IS THIS TRUE  ? */
1002                         /*
1003                          * Not any more.  Scenario: someone does a write()
1004                          * in data=journal mode.  The buffer's transaction has
1005                          * moved into commit.  Then someone does another
1006                          * write() to the file.  We do the frozen data copyout
1007                          * and set b_next_transaction to point to j_running_t.
1008                          * And while we're in that state, someone does a
1009                          * writepage() in an attempt to pageout the same area
1010                          * of the file via a shared mapping.  At present that
1011                          * calls journal_dirty_data(), and we get right here.
1012                          * It may be too late to journal the data.  Simply
1013                          * falling through to the next test will suffice: the
1014                          * data will be dirty and wil be checkpointed.  The
1015                          * ordering comments in the next comment block still
1016                          * apply.
1017                          */
1018                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1019
1020                         /*
1021                          * If we're journalling data, and this buffer was
1022                          * subject to a write(), it could be metadata, forget
1023                          * or shadow against the committing transaction.  Now,
1024                          * someone has dirtied the same darn page via a mapping
1025                          * and it is being writepage()'d.
1026                          * We *could* just steal the page from commit, with some
1027                          * fancy locking there.  Instead, we just skip it -
1028                          * don't tie the page's buffers to the new transaction
1029                          * at all.
1030                          * Implication: if we crash before the writepage() data
1031                          * is written into the filesystem, recovery will replay
1032                          * the write() data.
1033                          */
1034                         if (jh->b_jlist != BJ_None &&
1035                                         jh->b_jlist != BJ_SyncData &&
1036                                         jh->b_jlist != BJ_Locked) {
1037                                 JBUFFER_TRACE(jh, "Not stealing");
1038                                 goto no_journal;
1039                         }
1040
1041                         /*
1042                          * This buffer may be undergoing writeout in commit.  We
1043                          * can't return from here and let the caller dirty it
1044                          * again because that can cause the write-out loop in
1045                          * commit to never terminate.
1046                          */
1047                         if (buffer_dirty(bh)) {
1048                                 get_bh(bh);
1049                                 spin_unlock(&journal->j_list_lock);
1050                                 jbd_unlock_bh_state(bh);
1051                                 need_brelse = 1;
1052                                 sync_dirty_buffer(bh);
1053                                 jbd_lock_bh_state(bh);
1054                                 spin_lock(&journal->j_list_lock);
1055                                 /* Since we dropped the lock... */
1056                                 if (!buffer_mapped(bh)) {
1057                                         JBUFFER_TRACE(jh, "buffer got unmapped");
1058                                         goto no_journal;
1059                                 }
1060                                 /* The buffer may become locked again at any
1061                                    time if it is redirtied */
1062                         }
1063
1064                         /*
1065                          * We cannot remove the buffer with io error from the
1066                          * committing transaction, because otherwise it would
1067                          * miss the error and the commit would not abort.
1068                          */
1069                         if (unlikely(!buffer_uptodate(bh))) {
1070                                 ret = -EIO;
1071                                 goto no_journal;
1072                         }
1073
1074                         if (jh->b_transaction != NULL) {
1075                                 JBUFFER_TRACE(jh, "unfile from commit");
1076                                 __journal_temp_unlink_buffer(jh);
1077                                 /* It still points to the committing
1078                                  * transaction; move it to this one so
1079                                  * that the refile assert checks are
1080                                  * happy. */
1081                                 jh->b_transaction = handle->h_transaction;
1082                         }
1083                         /* The buffer will be refiled below */
1084
1085                 }
1086                 /*
1087                  * Special case --- the buffer might actually have been
1088                  * allocated and then immediately deallocated in the previous,
1089                  * committing transaction, so might still be left on that
1090                  * transaction's metadata lists.
1091                  */
1092                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1093                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1094                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1095                         __journal_temp_unlink_buffer(jh);
1096                         jh->b_transaction = handle->h_transaction;
1097                         JBUFFER_TRACE(jh, "file as data");
1098                         __journal_file_buffer(jh, handle->h_transaction,
1099                                                 BJ_SyncData);
1100                 }
1101         } else {
1102                 JBUFFER_TRACE(jh, "not on a transaction");
1103                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1104         }
1105 no_journal:
1106         spin_unlock(&journal->j_list_lock);
1107         jbd_unlock_bh_state(bh);
1108         if (need_brelse) {
1109                 BUFFER_TRACE(bh, "brelse");
1110                 __brelse(bh);
1111         }
1112         JBUFFER_TRACE(jh, "exit");
1113         journal_put_journal_head(jh);
1114         return ret;
1115 }
1116
1117 /**
1118  * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1119  * @handle: transaction to add buffer to.
1120  * @bh: buffer to mark
1121  *
1122  * Mark dirty metadata which needs to be journaled as part of the current
1123  * transaction.
1124  *
1125  * The buffer is placed on the transaction's metadata list and is marked
1126  * as belonging to the transaction.
1127  *
1128  * Returns error number or 0 on success.
1129  *
1130  * Special care needs to be taken if the buffer already belongs to the
1131  * current committing transaction (in which case we should have frozen
1132  * data present for that commit).  In that case, we don't relink the
1133  * buffer: that only gets done when the old transaction finally
1134  * completes its commit.
1135  */
1136 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1137 {
1138         transaction_t *transaction = handle->h_transaction;
1139         journal_t *journal = transaction->t_journal;
1140         struct journal_head *jh = bh2jh(bh);
1141
1142         jbd_debug(5, "journal_head %p\n", jh);
1143         JBUFFER_TRACE(jh, "entry");
1144         if (is_handle_aborted(handle))
1145                 goto out;
1146
1147         jbd_lock_bh_state(bh);
1148
1149         if (jh->b_modified == 0) {
1150                 /*
1151                  * This buffer's got modified and becoming part
1152                  * of the transaction. This needs to be done
1153                  * once a transaction -bzzz
1154                  */
1155                 jh->b_modified = 1;
1156                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1157                 handle->h_buffer_credits--;
1158         }
1159
1160         /*
1161          * fastpath, to avoid expensive locking.  If this buffer is already
1162          * on the running transaction's metadata list there is nothing to do.
1163          * Nobody can take it off again because there is a handle open.
1164          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1165          * result in this test being false, so we go in and take the locks.
1166          */
1167         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1168                 JBUFFER_TRACE(jh, "fastpath");
1169                 J_ASSERT_JH(jh, jh->b_transaction ==
1170                                         journal->j_running_transaction);
1171                 goto out_unlock_bh;
1172         }
1173
1174         set_buffer_jbddirty(bh);
1175
1176         /*
1177          * Metadata already on the current transaction list doesn't
1178          * need to be filed.  Metadata on another transaction's list must
1179          * be committing, and will be refiled once the commit completes:
1180          * leave it alone for now.
1181          */
1182         if (jh->b_transaction != transaction) {
1183                 JBUFFER_TRACE(jh, "already on other transaction");
1184                 J_ASSERT_JH(jh, jh->b_transaction ==
1185                                         journal->j_committing_transaction);
1186                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1187                 /* And this case is illegal: we can't reuse another
1188                  * transaction's data buffer, ever. */
1189                 goto out_unlock_bh;
1190         }
1191
1192         /* That test should have eliminated the following case: */
1193         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1194
1195         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1196         spin_lock(&journal->j_list_lock);
1197         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1198         spin_unlock(&journal->j_list_lock);
1199 out_unlock_bh:
1200         jbd_unlock_bh_state(bh);
1201 out:
1202         JBUFFER_TRACE(jh, "exit");
1203         return 0;
1204 }
1205
1206 /*
1207  * journal_release_buffer: undo a get_write_access without any buffer
1208  * updates, if the update decided in the end that it didn't need access.
1209  *
1210  */
1211 void
1212 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1213 {
1214         BUFFER_TRACE(bh, "entry");
1215 }
1216
1217 /**
1218  * void journal_forget() - bforget() for potentially-journaled buffers.
1219  * @handle: transaction handle
1220  * @bh:     bh to 'forget'
1221  *
1222  * We can only do the bforget if there are no commits pending against the
1223  * buffer.  If the buffer is dirty in the current running transaction we
1224  * can safely unlink it.
1225  *
1226  * bh may not be a journalled buffer at all - it may be a non-JBD
1227  * buffer which came off the hashtable.  Check for this.
1228  *
1229  * Decrements bh->b_count by one.
1230  *
1231  * Allow this call even if the handle has aborted --- it may be part of
1232  * the caller's cleanup after an abort.
1233  */
1234 int journal_forget (handle_t *handle, struct buffer_head *bh)
1235 {
1236         transaction_t *transaction = handle->h_transaction;
1237         journal_t *journal = transaction->t_journal;
1238         struct journal_head *jh;
1239         int drop_reserve = 0;
1240         int err = 0;
1241         int was_modified = 0;
1242
1243         BUFFER_TRACE(bh, "entry");
1244
1245         jbd_lock_bh_state(bh);
1246         spin_lock(&journal->j_list_lock);
1247
1248         if (!buffer_jbd(bh))
1249                 goto not_jbd;
1250         jh = bh2jh(bh);
1251
1252         /* Critical error: attempting to delete a bitmap buffer, maybe?
1253          * Don't do any jbd operations, and return an error. */
1254         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1255                          "inconsistent data on disk")) {
1256                 err = -EIO;
1257                 goto not_jbd;
1258         }
1259
1260         /* keep track of wether or not this transaction modified us */
1261         was_modified = jh->b_modified;
1262
1263         /*
1264          * The buffer's going from the transaction, we must drop
1265          * all references -bzzz
1266          */
1267         jh->b_modified = 0;
1268
1269         if (jh->b_transaction == handle->h_transaction) {
1270                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1271
1272                 /* If we are forgetting a buffer which is already part
1273                  * of this transaction, then we can just drop it from
1274                  * the transaction immediately. */
1275                 clear_buffer_dirty(bh);
1276                 clear_buffer_jbddirty(bh);
1277
1278                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1279
1280                 /*
1281                  * we only want to drop a reference if this transaction
1282                  * modified the buffer
1283                  */
1284                 if (was_modified)
1285                         drop_reserve = 1;
1286
1287                 /*
1288                  * We are no longer going to journal this buffer.
1289                  * However, the commit of this transaction is still
1290                  * important to the buffer: the delete that we are now
1291                  * processing might obsolete an old log entry, so by
1292                  * committing, we can satisfy the buffer's checkpoint.
1293                  *
1294                  * So, if we have a checkpoint on the buffer, we should
1295                  * now refile the buffer on our BJ_Forget list so that
1296                  * we know to remove the checkpoint after we commit.
1297                  */
1298
1299                 if (jh->b_cp_transaction) {
1300                         __journal_temp_unlink_buffer(jh);
1301                         __journal_file_buffer(jh, transaction, BJ_Forget);
1302                 } else {
1303                         __journal_unfile_buffer(jh);
1304                         journal_remove_journal_head(bh);
1305                         __brelse(bh);
1306                         if (!buffer_jbd(bh)) {
1307                                 spin_unlock(&journal->j_list_lock);
1308                                 jbd_unlock_bh_state(bh);
1309                                 __bforget(bh);
1310                                 goto drop;
1311                         }
1312                 }
1313         } else if (jh->b_transaction) {
1314                 J_ASSERT_JH(jh, (jh->b_transaction ==
1315                                  journal->j_committing_transaction));
1316                 /* However, if the buffer is still owned by a prior
1317                  * (committing) transaction, we can't drop it yet... */
1318                 JBUFFER_TRACE(jh, "belongs to older transaction");
1319                 /* ... but we CAN drop it from the new transaction if we
1320                  * have also modified it since the original commit. */
1321
1322                 if (jh->b_next_transaction) {
1323                         J_ASSERT(jh->b_next_transaction == transaction);
1324                         jh->b_next_transaction = NULL;
1325
1326                         /*
1327                          * only drop a reference if this transaction modified
1328                          * the buffer
1329                          */
1330                         if (was_modified)
1331                                 drop_reserve = 1;
1332                 }
1333         }
1334
1335 not_jbd:
1336         spin_unlock(&journal->j_list_lock);
1337         jbd_unlock_bh_state(bh);
1338         __brelse(bh);
1339 drop:
1340         if (drop_reserve) {
1341                 /* no need to reserve log space for this block -bzzz */
1342                 handle->h_buffer_credits++;
1343         }
1344         return err;
1345 }
1346
1347 /**
1348  * int journal_stop() - complete a transaction
1349  * @handle: tranaction to complete.
1350  *
1351  * All done for a particular handle.
1352  *
1353  * There is not much action needed here.  We just return any remaining
1354  * buffer credits to the transaction and remove the handle.  The only
1355  * complication is that we need to start a commit operation if the
1356  * filesystem is marked for synchronous update.
1357  *
1358  * journal_stop itself will not usually return an error, but it may
1359  * do so in unusual circumstances.  In particular, expect it to
1360  * return -EIO if a journal_abort has been executed since the
1361  * transaction began.
1362  */
1363 int journal_stop(handle_t *handle)
1364 {
1365         transaction_t *transaction = handle->h_transaction;
1366         journal_t *journal = transaction->t_journal;
1367         int err;
1368         pid_t pid;
1369
1370         J_ASSERT(journal_current_handle() == handle);
1371
1372         if (is_handle_aborted(handle))
1373                 err = -EIO;
1374         else {
1375                 J_ASSERT(transaction->t_updates > 0);
1376                 err = 0;
1377         }
1378
1379         if (--handle->h_ref > 0) {
1380                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1381                           handle->h_ref);
1382                 return err;
1383         }
1384
1385         jbd_debug(4, "Handle %p going down\n", handle);
1386
1387         /*
1388          * Implement synchronous transaction batching.  If the handle
1389          * was synchronous, don't force a commit immediately.  Let's
1390          * yield and let another thread piggyback onto this transaction.
1391          * Keep doing that while new threads continue to arrive.
1392          * It doesn't cost much - we're about to run a commit and sleep
1393          * on IO anyway.  Speeds up many-threaded, many-dir operations
1394          * by 30x or more...
1395          *
1396          * We try and optimize the sleep time against what the underlying disk
1397          * can do, instead of having a static sleep time.  This is usefull for
1398          * the case where our storage is so fast that it is more optimal to go
1399          * ahead and force a flush and wait for the transaction to be committed
1400          * than it is to wait for an arbitrary amount of time for new writers to
1401          * join the transaction.  We acheive this by measuring how long it takes
1402          * to commit a transaction, and compare it with how long this
1403          * transaction has been running, and if run time < commit time then we
1404          * sleep for the delta and commit.  This greatly helps super fast disks
1405          * that would see slowdowns as more threads started doing fsyncs.
1406          *
1407          * But don't do this if this process was the most recent one to
1408          * perform a synchronous write.  We do this to detect the case where a
1409          * single process is doing a stream of sync writes.  No point in waiting
1410          * for joiners in that case.
1411          */
1412         pid = current->pid;
1413         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1414                 u64 commit_time, trans_time;
1415
1416                 journal->j_last_sync_writer = pid;
1417
1418                 spin_lock(&journal->j_state_lock);
1419                 commit_time = journal->j_average_commit_time;
1420                 spin_unlock(&journal->j_state_lock);
1421
1422                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1423                                                    transaction->t_start_time));
1424
1425                 commit_time = min_t(u64, commit_time,
1426                                     1000*jiffies_to_usecs(1));
1427
1428                 if (trans_time < commit_time) {
1429                         ktime_t expires = ktime_add_ns(ktime_get(),
1430                                                        commit_time);
1431                         set_current_state(TASK_UNINTERRUPTIBLE);
1432                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1433                 }
1434         }
1435
1436         if (handle->h_sync)
1437                 transaction->t_synchronous_commit = 1;
1438         current->journal_info = NULL;
1439         spin_lock(&journal->j_state_lock);
1440         spin_lock(&transaction->t_handle_lock);
1441         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1442         transaction->t_updates--;
1443         if (!transaction->t_updates) {
1444                 wake_up(&journal->j_wait_updates);
1445                 if (journal->j_barrier_count)
1446                         wake_up(&journal->j_wait_transaction_locked);
1447         }
1448
1449         /*
1450          * If the handle is marked SYNC, we need to set another commit
1451          * going!  We also want to force a commit if the current
1452          * transaction is occupying too much of the log, or if the
1453          * transaction is too old now.
1454          */
1455         if (handle->h_sync ||
1456                         transaction->t_outstanding_credits >
1457                                 journal->j_max_transaction_buffers ||
1458                         time_after_eq(jiffies, transaction->t_expires)) {
1459                 /* Do this even for aborted journals: an abort still
1460                  * completes the commit thread, it just doesn't write
1461                  * anything to disk. */
1462                 tid_t tid = transaction->t_tid;
1463
1464                 spin_unlock(&transaction->t_handle_lock);
1465                 jbd_debug(2, "transaction too old, requesting commit for "
1466                                         "handle %p\n", handle);
1467                 /* This is non-blocking */
1468                 __log_start_commit(journal, transaction->t_tid);
1469                 spin_unlock(&journal->j_state_lock);
1470
1471                 /*
1472                  * Special case: JFS_SYNC synchronous updates require us
1473                  * to wait for the commit to complete.
1474                  */
1475                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1476                         err = log_wait_commit(journal, tid);
1477         } else {
1478                 spin_unlock(&transaction->t_handle_lock);
1479                 spin_unlock(&journal->j_state_lock);
1480         }
1481
1482         lock_map_release(&handle->h_lockdep_map);
1483
1484         jbd_free_handle(handle);
1485         return err;
1486 }
1487
1488 /**
1489  * int journal_force_commit() - force any uncommitted transactions
1490  * @journal: journal to force
1491  *
1492  * For synchronous operations: force any uncommitted transactions
1493  * to disk.  May seem kludgy, but it reuses all the handle batching
1494  * code in a very simple manner.
1495  */
1496 int journal_force_commit(journal_t *journal)
1497 {
1498         handle_t *handle;
1499         int ret;
1500
1501         handle = journal_start(journal, 1);
1502         if (IS_ERR(handle)) {
1503                 ret = PTR_ERR(handle);
1504         } else {
1505                 handle->h_sync = 1;
1506                 ret = journal_stop(handle);
1507         }
1508         return ret;
1509 }
1510
1511 /*
1512  *
1513  * List management code snippets: various functions for manipulating the
1514  * transaction buffer lists.
1515  *
1516  */
1517
1518 /*
1519  * Append a buffer to a transaction list, given the transaction's list head
1520  * pointer.
1521  *
1522  * j_list_lock is held.
1523  *
1524  * jbd_lock_bh_state(jh2bh(jh)) is held.
1525  */
1526
1527 static inline void
1528 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1529 {
1530         if (!*list) {
1531                 jh->b_tnext = jh->b_tprev = jh;
1532                 *list = jh;
1533         } else {
1534                 /* Insert at the tail of the list to preserve order */
1535                 struct journal_head *first = *list, *last = first->b_tprev;
1536                 jh->b_tprev = last;
1537                 jh->b_tnext = first;
1538                 last->b_tnext = first->b_tprev = jh;
1539         }
1540 }
1541
1542 /*
1543  * Remove a buffer from a transaction list, given the transaction's list
1544  * head pointer.
1545  *
1546  * Called with j_list_lock held, and the journal may not be locked.
1547  *
1548  * jbd_lock_bh_state(jh2bh(jh)) is held.
1549  */
1550
1551 static inline void
1552 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1553 {
1554         if (*list == jh) {
1555                 *list = jh->b_tnext;
1556                 if (*list == jh)
1557                         *list = NULL;
1558         }
1559         jh->b_tprev->b_tnext = jh->b_tnext;
1560         jh->b_tnext->b_tprev = jh->b_tprev;
1561 }
1562
1563 /*
1564  * Remove a buffer from the appropriate transaction list.
1565  *
1566  * Note that this function can *change* the value of
1567  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1568  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1569  * is holding onto a copy of one of thee pointers, it could go bad.
1570  * Generally the caller needs to re-read the pointer from the transaction_t.
1571  *
1572  * Called under j_list_lock.  The journal may not be locked.
1573  */
1574 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1575 {
1576         struct journal_head **list = NULL;
1577         transaction_t *transaction;
1578         struct buffer_head *bh = jh2bh(jh);
1579
1580         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1581         transaction = jh->b_transaction;
1582         if (transaction)
1583                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1584
1585         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1586         if (jh->b_jlist != BJ_None)
1587                 J_ASSERT_JH(jh, transaction != NULL);
1588
1589         switch (jh->b_jlist) {
1590         case BJ_None:
1591                 return;
1592         case BJ_SyncData:
1593                 list = &transaction->t_sync_datalist;
1594                 break;
1595         case BJ_Metadata:
1596                 transaction->t_nr_buffers--;
1597                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1598                 list = &transaction->t_buffers;
1599                 break;
1600         case BJ_Forget:
1601                 list = &transaction->t_forget;
1602                 break;
1603         case BJ_IO:
1604                 list = &transaction->t_iobuf_list;
1605                 break;
1606         case BJ_Shadow:
1607                 list = &transaction->t_shadow_list;
1608                 break;
1609         case BJ_LogCtl:
1610                 list = &transaction->t_log_list;
1611                 break;
1612         case BJ_Reserved:
1613                 list = &transaction->t_reserved_list;
1614                 break;
1615         case BJ_Locked:
1616                 list = &transaction->t_locked_list;
1617                 break;
1618         }
1619
1620         __blist_del_buffer(list, jh);
1621         jh->b_jlist = BJ_None;
1622         if (test_clear_buffer_jbddirty(bh))
1623                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1624 }
1625
1626 void __journal_unfile_buffer(struct journal_head *jh)
1627 {
1628         __journal_temp_unlink_buffer(jh);
1629         jh->b_transaction = NULL;
1630 }
1631
1632 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1633 {
1634         jbd_lock_bh_state(jh2bh(jh));
1635         spin_lock(&journal->j_list_lock);
1636         __journal_unfile_buffer(jh);
1637         spin_unlock(&journal->j_list_lock);
1638         jbd_unlock_bh_state(jh2bh(jh));
1639 }
1640
1641 /*
1642  * Called from journal_try_to_free_buffers().
1643  *
1644  * Called under jbd_lock_bh_state(bh)
1645  */
1646 static void
1647 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1648 {
1649         struct journal_head *jh;
1650
1651         jh = bh2jh(bh);
1652
1653         if (buffer_locked(bh) || buffer_dirty(bh))
1654                 goto out;
1655
1656         if (jh->b_next_transaction != NULL)
1657                 goto out;
1658
1659         spin_lock(&journal->j_list_lock);
1660         if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1661                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1662                         /* A written-back ordered data buffer */
1663                         JBUFFER_TRACE(jh, "release data");
1664                         __journal_unfile_buffer(jh);
1665                         journal_remove_journal_head(bh);
1666                         __brelse(bh);
1667                 }
1668         } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1669                 /* written-back checkpointed metadata buffer */
1670                 if (jh->b_jlist == BJ_None) {
1671                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1672                         __journal_remove_checkpoint(jh);
1673                         journal_remove_journal_head(bh);
1674                         __brelse(bh);
1675                 }
1676         }
1677         spin_unlock(&journal->j_list_lock);
1678 out:
1679         return;
1680 }
1681
1682 /**
1683  * int journal_try_to_free_buffers() - try to free page buffers.
1684  * @journal: journal for operation
1685  * @page: to try and free
1686  * @gfp_mask: we use the mask to detect how hard should we try to release
1687  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1688  * release the buffers.
1689  *
1690  *
1691  * For all the buffers on this page,
1692  * if they are fully written out ordered data, move them onto BUF_CLEAN
1693  * so try_to_free_buffers() can reap them.
1694  *
1695  * This function returns non-zero if we wish try_to_free_buffers()
1696  * to be called. We do this if the page is releasable by try_to_free_buffers().
1697  * We also do it if the page has locked or dirty buffers and the caller wants
1698  * us to perform sync or async writeout.
1699  *
1700  * This complicates JBD locking somewhat.  We aren't protected by the
1701  * BKL here.  We wish to remove the buffer from its committing or
1702  * running transaction's ->t_datalist via __journal_unfile_buffer.
1703  *
1704  * This may *change* the value of transaction_t->t_datalist, so anyone
1705  * who looks at t_datalist needs to lock against this function.
1706  *
1707  * Even worse, someone may be doing a journal_dirty_data on this
1708  * buffer.  So we need to lock against that.  journal_dirty_data()
1709  * will come out of the lock with the buffer dirty, which makes it
1710  * ineligible for release here.
1711  *
1712  * Who else is affected by this?  hmm...  Really the only contender
1713  * is do_get_write_access() - it could be looking at the buffer while
1714  * journal_try_to_free_buffer() is changing its state.  But that
1715  * cannot happen because we never reallocate freed data as metadata
1716  * while the data is part of a transaction.  Yes?
1717  *
1718  * Return 0 on failure, 1 on success
1719  */
1720 int journal_try_to_free_buffers(journal_t *journal,
1721                                 struct page *page, gfp_t gfp_mask)
1722 {
1723         struct buffer_head *head;
1724         struct buffer_head *bh;
1725         int ret = 0;
1726
1727         J_ASSERT(PageLocked(page));
1728
1729         head = page_buffers(page);
1730         bh = head;
1731         do {
1732                 struct journal_head *jh;
1733
1734                 /*
1735                  * We take our own ref against the journal_head here to avoid
1736                  * having to add tons of locking around each instance of
1737                  * journal_remove_journal_head() and journal_put_journal_head().
1738                  */
1739                 jh = journal_grab_journal_head(bh);
1740                 if (!jh)
1741                         continue;
1742
1743                 jbd_lock_bh_state(bh);
1744                 __journal_try_to_free_buffer(journal, bh);
1745                 journal_put_journal_head(jh);
1746                 jbd_unlock_bh_state(bh);
1747                 if (buffer_jbd(bh))
1748                         goto busy;
1749         } while ((bh = bh->b_this_page) != head);
1750
1751         ret = try_to_free_buffers(page);
1752
1753 busy:
1754         return ret;
1755 }
1756
1757 /*
1758  * This buffer is no longer needed.  If it is on an older transaction's
1759  * checkpoint list we need to record it on this transaction's forget list
1760  * to pin this buffer (and hence its checkpointing transaction) down until
1761  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1762  * release it.
1763  * Returns non-zero if JBD no longer has an interest in the buffer.
1764  *
1765  * Called under j_list_lock.
1766  *
1767  * Called under jbd_lock_bh_state(bh).
1768  */
1769 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1770 {
1771         int may_free = 1;
1772         struct buffer_head *bh = jh2bh(jh);
1773
1774         __journal_unfile_buffer(jh);
1775
1776         if (jh->b_cp_transaction) {
1777                 JBUFFER_TRACE(jh, "on running+cp transaction");
1778                 /*
1779                  * We don't want to write the buffer anymore, clear the
1780                  * bit so that we don't confuse checks in
1781                  * __journal_file_buffer
1782                  */
1783                 clear_buffer_dirty(bh);
1784                 __journal_file_buffer(jh, transaction, BJ_Forget);
1785                 may_free = 0;
1786         } else {
1787                 JBUFFER_TRACE(jh, "on running transaction");
1788                 journal_remove_journal_head(bh);
1789                 __brelse(bh);
1790         }
1791         return may_free;
1792 }
1793
1794 /*
1795  * journal_invalidatepage
1796  *
1797  * This code is tricky.  It has a number of cases to deal with.
1798  *
1799  * There are two invariants which this code relies on:
1800  *
1801  * i_size must be updated on disk before we start calling invalidatepage on the
1802  * data.
1803  *
1804  *  This is done in ext3 by defining an ext3_setattr method which
1805  *  updates i_size before truncate gets going.  By maintaining this
1806  *  invariant, we can be sure that it is safe to throw away any buffers
1807  *  attached to the current transaction: once the transaction commits,
1808  *  we know that the data will not be needed.
1809  *
1810  *  Note however that we can *not* throw away data belonging to the
1811  *  previous, committing transaction!
1812  *
1813  * Any disk blocks which *are* part of the previous, committing
1814  * transaction (and which therefore cannot be discarded immediately) are
1815  * not going to be reused in the new running transaction
1816  *
1817  *  The bitmap committed_data images guarantee this: any block which is
1818  *  allocated in one transaction and removed in the next will be marked
1819  *  as in-use in the committed_data bitmap, so cannot be reused until
1820  *  the next transaction to delete the block commits.  This means that
1821  *  leaving committing buffers dirty is quite safe: the disk blocks
1822  *  cannot be reallocated to a different file and so buffer aliasing is
1823  *  not possible.
1824  *
1825  *
1826  * The above applies mainly to ordered data mode.  In writeback mode we
1827  * don't make guarantees about the order in which data hits disk --- in
1828  * particular we don't guarantee that new dirty data is flushed before
1829  * transaction commit --- so it is always safe just to discard data
1830  * immediately in that mode.  --sct
1831  */
1832
1833 /*
1834  * The journal_unmap_buffer helper function returns zero if the buffer
1835  * concerned remains pinned as an anonymous buffer belonging to an older
1836  * transaction.
1837  *
1838  * We're outside-transaction here.  Either or both of j_running_transaction
1839  * and j_committing_transaction may be NULL.
1840  */
1841 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1842 {
1843         transaction_t *transaction;
1844         struct journal_head *jh;
1845         int may_free = 1;
1846         int ret;
1847
1848         BUFFER_TRACE(bh, "entry");
1849
1850         /*
1851          * It is safe to proceed here without the j_list_lock because the
1852          * buffers cannot be stolen by try_to_free_buffers as long as we are
1853          * holding the page lock. --sct
1854          */
1855
1856         if (!buffer_jbd(bh))
1857                 goto zap_buffer_unlocked;
1858
1859         spin_lock(&journal->j_state_lock);
1860         jbd_lock_bh_state(bh);
1861         spin_lock(&journal->j_list_lock);
1862
1863         jh = journal_grab_journal_head(bh);
1864         if (!jh)
1865                 goto zap_buffer_no_jh;
1866
1867         transaction = jh->b_transaction;
1868         if (transaction == NULL) {
1869                 /* First case: not on any transaction.  If it
1870                  * has no checkpoint link, then we can zap it:
1871                  * it's a writeback-mode buffer so we don't care
1872                  * if it hits disk safely. */
1873                 if (!jh->b_cp_transaction) {
1874                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1875                         goto zap_buffer;
1876                 }
1877
1878                 if (!buffer_dirty(bh)) {
1879                         /* bdflush has written it.  We can drop it now */
1880                         goto zap_buffer;
1881                 }
1882
1883                 /* OK, it must be in the journal but still not
1884                  * written fully to disk: it's metadata or
1885                  * journaled data... */
1886
1887                 if (journal->j_running_transaction) {
1888                         /* ... and once the current transaction has
1889                          * committed, the buffer won't be needed any
1890                          * longer. */
1891                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1892                         ret = __dispose_buffer(jh,
1893                                         journal->j_running_transaction);
1894                         journal_put_journal_head(jh);
1895                         spin_unlock(&journal->j_list_lock);
1896                         jbd_unlock_bh_state(bh);
1897                         spin_unlock(&journal->j_state_lock);
1898                         return ret;
1899                 } else {
1900                         /* There is no currently-running transaction. So the
1901                          * orphan record which we wrote for this file must have
1902                          * passed into commit.  We must attach this buffer to
1903                          * the committing transaction, if it exists. */
1904                         if (journal->j_committing_transaction) {
1905                                 JBUFFER_TRACE(jh, "give to committing trans");
1906                                 ret = __dispose_buffer(jh,
1907                                         journal->j_committing_transaction);
1908                                 journal_put_journal_head(jh);
1909                                 spin_unlock(&journal->j_list_lock);
1910                                 jbd_unlock_bh_state(bh);
1911                                 spin_unlock(&journal->j_state_lock);
1912                                 return ret;
1913                         } else {
1914                                 /* The orphan record's transaction has
1915                                  * committed.  We can cleanse this buffer */
1916                                 clear_buffer_jbddirty(bh);
1917                                 goto zap_buffer;
1918                         }
1919                 }
1920         } else if (transaction == journal->j_committing_transaction) {
1921                 JBUFFER_TRACE(jh, "on committing transaction");
1922                 if (jh->b_jlist == BJ_Locked) {
1923                         /*
1924                          * The buffer is on the committing transaction's locked
1925                          * list.  We have the buffer locked, so I/O has
1926                          * completed.  So we can nail the buffer now.
1927                          */
1928                         may_free = __dispose_buffer(jh, transaction);
1929                         goto zap_buffer;
1930                 }
1931                 /*
1932                  * If it is committing, we simply cannot touch it.  We
1933                  * can remove it's next_transaction pointer from the
1934                  * running transaction if that is set, but nothing
1935                  * else. */
1936                 set_buffer_freed(bh);
1937                 if (jh->b_next_transaction) {
1938                         J_ASSERT(jh->b_next_transaction ==
1939                                         journal->j_running_transaction);
1940                         jh->b_next_transaction = NULL;
1941                 }
1942                 journal_put_journal_head(jh);
1943                 spin_unlock(&journal->j_list_lock);
1944                 jbd_unlock_bh_state(bh);
1945                 spin_unlock(&journal->j_state_lock);
1946                 return 0;
1947         } else {
1948                 /* Good, the buffer belongs to the running transaction.
1949                  * We are writing our own transaction's data, not any
1950                  * previous one's, so it is safe to throw it away
1951                  * (remember that we expect the filesystem to have set
1952                  * i_size already for this truncate so recovery will not
1953                  * expose the disk blocks we are discarding here.) */
1954                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1955                 JBUFFER_TRACE(jh, "on running transaction");
1956                 may_free = __dispose_buffer(jh, transaction);
1957         }
1958
1959 zap_buffer:
1960         journal_put_journal_head(jh);
1961 zap_buffer_no_jh:
1962         spin_unlock(&journal->j_list_lock);
1963         jbd_unlock_bh_state(bh);
1964         spin_unlock(&journal->j_state_lock);
1965 zap_buffer_unlocked:
1966         clear_buffer_dirty(bh);
1967         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1968         clear_buffer_mapped(bh);
1969         clear_buffer_req(bh);
1970         clear_buffer_new(bh);
1971         bh->b_bdev = NULL;
1972         return may_free;
1973 }
1974
1975 /**
1976  * void journal_invalidatepage() - invalidate a journal page
1977  * @journal: journal to use for flush
1978  * @page:    page to flush
1979  * @offset:  length of page to invalidate.
1980  *
1981  * Reap page buffers containing data after offset in page.
1982  */
1983 void journal_invalidatepage(journal_t *journal,
1984                       struct page *page,
1985                       unsigned long offset)
1986 {
1987         struct buffer_head *head, *bh, *next;
1988         unsigned int curr_off = 0;
1989         int may_free = 1;
1990
1991         if (!PageLocked(page))
1992                 BUG();
1993         if (!page_has_buffers(page))
1994                 return;
1995
1996         /* We will potentially be playing with lists other than just the
1997          * data lists (especially for journaled data mode), so be
1998          * cautious in our locking. */
1999
2000         head = bh = page_buffers(page);
2001         do {
2002                 unsigned int next_off = curr_off + bh->b_size;
2003                 next = bh->b_this_page;
2004
2005                 if (offset <= curr_off) {
2006                         /* This block is wholly outside the truncation point */
2007                         lock_buffer(bh);
2008                         may_free &= journal_unmap_buffer(journal, bh);
2009                         unlock_buffer(bh);
2010                 }
2011                 curr_off = next_off;
2012                 bh = next;
2013
2014         } while (bh != head);
2015
2016         if (!offset) {
2017                 if (may_free && try_to_free_buffers(page))
2018                         J_ASSERT(!page_has_buffers(page));
2019         }
2020 }
2021
2022 /*
2023  * File a buffer on the given transaction list.
2024  */
2025 void __journal_file_buffer(struct journal_head *jh,
2026                         transaction_t *transaction, int jlist)
2027 {
2028         struct journal_head **list = NULL;
2029         int was_dirty = 0;
2030         struct buffer_head *bh = jh2bh(jh);
2031
2032         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2033         assert_spin_locked(&transaction->t_journal->j_list_lock);
2034
2035         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2036         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2037                                 jh->b_transaction == NULL);
2038
2039         if (jh->b_transaction && jh->b_jlist == jlist)
2040                 return;
2041
2042         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2043             jlist == BJ_Shadow || jlist == BJ_Forget) {
2044                 /*
2045                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2046                  * instead of buffer_dirty. We should not see a dirty bit set
2047                  * here because we clear it in do_get_write_access but e.g.
2048                  * tune2fs can modify the sb and set the dirty bit at any time
2049                  * so we try to gracefully handle that.
2050                  */
2051                 if (buffer_dirty(bh))
2052                         warn_dirty_buffer(bh);
2053                 if (test_clear_buffer_dirty(bh) ||
2054                     test_clear_buffer_jbddirty(bh))
2055                         was_dirty = 1;
2056         }
2057
2058         if (jh->b_transaction)
2059                 __journal_temp_unlink_buffer(jh);
2060         jh->b_transaction = transaction;
2061
2062         switch (jlist) {
2063         case BJ_None:
2064                 J_ASSERT_JH(jh, !jh->b_committed_data);
2065                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2066                 return;
2067         case BJ_SyncData:
2068                 list = &transaction->t_sync_datalist;
2069                 break;
2070         case BJ_Metadata:
2071                 transaction->t_nr_buffers++;
2072                 list = &transaction->t_buffers;
2073                 break;
2074         case BJ_Forget:
2075                 list = &transaction->t_forget;
2076                 break;
2077         case BJ_IO:
2078                 list = &transaction->t_iobuf_list;
2079                 break;
2080         case BJ_Shadow:
2081                 list = &transaction->t_shadow_list;
2082                 break;
2083         case BJ_LogCtl:
2084                 list = &transaction->t_log_list;
2085                 break;
2086         case BJ_Reserved:
2087                 list = &transaction->t_reserved_list;
2088                 break;
2089         case BJ_Locked:
2090                 list =  &transaction->t_locked_list;
2091                 break;
2092         }
2093
2094         __blist_add_buffer(list, jh);
2095         jh->b_jlist = jlist;
2096
2097         if (was_dirty)
2098                 set_buffer_jbddirty(bh);
2099 }
2100
2101 void journal_file_buffer(struct journal_head *jh,
2102                                 transaction_t *transaction, int jlist)
2103 {
2104         jbd_lock_bh_state(jh2bh(jh));
2105         spin_lock(&transaction->t_journal->j_list_lock);
2106         __journal_file_buffer(jh, transaction, jlist);
2107         spin_unlock(&transaction->t_journal->j_list_lock);
2108         jbd_unlock_bh_state(jh2bh(jh));
2109 }
2110
2111 /*
2112  * Remove a buffer from its current buffer list in preparation for
2113  * dropping it from its current transaction entirely.  If the buffer has
2114  * already started to be used by a subsequent transaction, refile the
2115  * buffer on that transaction's metadata list.
2116  *
2117  * Called under journal->j_list_lock
2118  *
2119  * Called under jbd_lock_bh_state(jh2bh(jh))
2120  */
2121 void __journal_refile_buffer(struct journal_head *jh)
2122 {
2123         int was_dirty;
2124         struct buffer_head *bh = jh2bh(jh);
2125
2126         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2127         if (jh->b_transaction)
2128                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2129
2130         /* If the buffer is now unused, just drop it. */
2131         if (jh->b_next_transaction == NULL) {
2132                 __journal_unfile_buffer(jh);
2133                 return;
2134         }
2135
2136         /*
2137          * It has been modified by a later transaction: add it to the new
2138          * transaction's metadata list.
2139          */
2140
2141         was_dirty = test_clear_buffer_jbddirty(bh);
2142         __journal_temp_unlink_buffer(jh);
2143         jh->b_transaction = jh->b_next_transaction;
2144         jh->b_next_transaction = NULL;
2145         __journal_file_buffer(jh, jh->b_transaction,
2146                                 jh->b_modified ? BJ_Metadata : BJ_Reserved);
2147         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2148
2149         if (was_dirty)
2150                 set_buffer_jbddirty(bh);
2151 }
2152
2153 /*
2154  * For the unlocked version of this call, also make sure that any
2155  * hanging journal_head is cleaned up if necessary.
2156  *
2157  * __journal_refile_buffer is usually called as part of a single locked
2158  * operation on a buffer_head, in which the caller is probably going to
2159  * be hooking the journal_head onto other lists.  In that case it is up
2160  * to the caller to remove the journal_head if necessary.  For the
2161  * unlocked journal_refile_buffer call, the caller isn't going to be
2162  * doing anything else to the buffer so we need to do the cleanup
2163  * ourselves to avoid a jh leak.
2164  *
2165  * *** The journal_head may be freed by this call! ***
2166  */
2167 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2168 {
2169         struct buffer_head *bh = jh2bh(jh);
2170
2171         jbd_lock_bh_state(bh);
2172         spin_lock(&journal->j_list_lock);
2173
2174         __journal_refile_buffer(jh);
2175         jbd_unlock_bh_state(bh);
2176         journal_remove_journal_head(bh);
2177
2178         spin_unlock(&journal->j_list_lock);
2179         __brelse(bh);
2180 }