[PATCH] ext3/4: fix J_ASSERT(transaction->t_updates > 0) in journal_stop()
[linux-2.6.git] / fs / jbd2 / transaction.c
1 /*
2  * linux/fs/transaction.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem transaction handling code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages transactions (compound commits managed by the
16  * journaling code) and handles (individual atomic operations by the
17  * filesystem).
18  */
19
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/smp_lock.h>
27 #include <linux/mm.h>
28 #include <linux/highmem.h>
29
30 /*
31  * jbd2_get_transaction: obtain a new transaction_t object.
32  *
33  * Simply allocate and initialise a new transaction.  Create it in
34  * RUNNING state and add it to the current journal (which should not
35  * have an existing running transaction: we only make a new transaction
36  * once we have started to commit the old one).
37  *
38  * Preconditions:
39  *      The journal MUST be locked.  We don't perform atomic mallocs on the
40  *      new transaction and we can't block without protecting against other
41  *      processes trying to touch the journal while it is in transition.
42  *
43  * Called under j_state_lock
44  */
45
46 static transaction_t *
47 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
48 {
49         transaction->t_journal = journal;
50         transaction->t_state = T_RUNNING;
51         transaction->t_tid = journal->j_transaction_sequence++;
52         transaction->t_expires = jiffies + journal->j_commit_interval;
53         spin_lock_init(&transaction->t_handle_lock);
54
55         /* Set up the commit timer for the new transaction. */
56         journal->j_commit_timer.expires = transaction->t_expires;
57         add_timer(&journal->j_commit_timer);
58
59         J_ASSERT(journal->j_running_transaction == NULL);
60         journal->j_running_transaction = transaction;
61
62         return transaction;
63 }
64
65 /*
66  * Handle management.
67  *
68  * A handle_t is an object which represents a single atomic update to a
69  * filesystem, and which tracks all of the modifications which form part
70  * of that one update.
71  */
72
73 /*
74  * start_this_handle: Given a handle, deal with any locking or stalling
75  * needed to make sure that there is enough journal space for the handle
76  * to begin.  Attach the handle to a transaction and set up the
77  * transaction's buffer credits.
78  */
79
80 static int start_this_handle(journal_t *journal, handle_t *handle)
81 {
82         transaction_t *transaction;
83         int needed;
84         int nblocks = handle->h_buffer_credits;
85         transaction_t *new_transaction = NULL;
86         int ret = 0;
87
88         if (nblocks > journal->j_max_transaction_buffers) {
89                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
90                        current->comm, nblocks,
91                        journal->j_max_transaction_buffers);
92                 ret = -ENOSPC;
93                 goto out;
94         }
95
96 alloc_transaction:
97         if (!journal->j_running_transaction) {
98                 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
99                                                 GFP_NOFS);
100                 if (!new_transaction) {
101                         ret = -ENOMEM;
102                         goto out;
103                 }
104                 memset(new_transaction, 0, sizeof(*new_transaction));
105         }
106
107         jbd_debug(3, "New handle %p going live.\n", handle);
108
109 repeat:
110
111         /*
112          * We need to hold j_state_lock until t_updates has been incremented,
113          * for proper journal barrier handling
114          */
115         spin_lock(&journal->j_state_lock);
116 repeat_locked:
117         if (is_journal_aborted(journal) ||
118             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
119                 spin_unlock(&journal->j_state_lock);
120                 ret = -EROFS;
121                 goto out;
122         }
123
124         /* Wait on the journal's transaction barrier if necessary */
125         if (journal->j_barrier_count) {
126                 spin_unlock(&journal->j_state_lock);
127                 wait_event(journal->j_wait_transaction_locked,
128                                 journal->j_barrier_count == 0);
129                 goto repeat;
130         }
131
132         if (!journal->j_running_transaction) {
133                 if (!new_transaction) {
134                         spin_unlock(&journal->j_state_lock);
135                         goto alloc_transaction;
136                 }
137                 jbd2_get_transaction(journal, new_transaction);
138                 new_transaction = NULL;
139         }
140
141         transaction = journal->j_running_transaction;
142
143         /*
144          * If the current transaction is locked down for commit, wait for the
145          * lock to be released.
146          */
147         if (transaction->t_state == T_LOCKED) {
148                 DEFINE_WAIT(wait);
149
150                 prepare_to_wait(&journal->j_wait_transaction_locked,
151                                         &wait, TASK_UNINTERRUPTIBLE);
152                 spin_unlock(&journal->j_state_lock);
153                 schedule();
154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
155                 goto repeat;
156         }
157
158         /*
159          * If there is not enough space left in the log to write all potential
160          * buffers requested by this operation, we need to stall pending a log
161          * checkpoint to free some more log space.
162          */
163         spin_lock(&transaction->t_handle_lock);
164         needed = transaction->t_outstanding_credits + nblocks;
165
166         if (needed > journal->j_max_transaction_buffers) {
167                 /*
168                  * If the current transaction is already too large, then start
169                  * to commit it: we can then go back and attach this handle to
170                  * a new transaction.
171                  */
172                 DEFINE_WAIT(wait);
173
174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
175                 spin_unlock(&transaction->t_handle_lock);
176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177                                 TASK_UNINTERRUPTIBLE);
178                 __jbd2_log_start_commit(journal, transaction->t_tid);
179                 spin_unlock(&journal->j_state_lock);
180                 schedule();
181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
182                 goto repeat;
183         }
184
185         /*
186          * The commit code assumes that it can get enough log space
187          * without forcing a checkpoint.  This is *critical* for
188          * correctness: a checkpoint of a buffer which is also
189          * associated with a committing transaction creates a deadlock,
190          * so commit simply cannot force through checkpoints.
191          *
192          * We must therefore ensure the necessary space in the journal
193          * *before* starting to dirty potentially checkpointed buffers
194          * in the new transaction.
195          *
196          * The worst part is, any transaction currently committing can
197          * reduce the free space arbitrarily.  Be careful to account for
198          * those buffers when checkpointing.
199          */
200
201         /*
202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
203          * a _lot_ of headroom: 1/4 of the journal plus the size of
204          * the committing transaction.  Really, we only need to give it
205          * committing_transaction->t_outstanding_credits plus "enough" for
206          * the log control blocks.
207          * Also, this test is inconsitent with the matching one in
208          * jbd2_journal_extend().
209          */
210         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
212                 spin_unlock(&transaction->t_handle_lock);
213                 __jbd2_log_wait_for_space(journal);
214                 goto repeat_locked;
215         }
216
217         /* OK, account for the buffers that this operation expects to
218          * use and add the handle to the running transaction. */
219
220         handle->h_transaction = transaction;
221         transaction->t_outstanding_credits += nblocks;
222         transaction->t_updates++;
223         transaction->t_handle_count++;
224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
225                   handle, nblocks, transaction->t_outstanding_credits,
226                   __jbd2_log_space_left(journal));
227         spin_unlock(&transaction->t_handle_lock);
228         spin_unlock(&journal->j_state_lock);
229 out:
230         if (unlikely(new_transaction))          /* It's usually NULL */
231                 kfree(new_transaction);
232         return ret;
233 }
234
235 /* Allocate a new handle.  This should probably be in a slab... */
236 static handle_t *new_handle(int nblocks)
237 {
238         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
239         if (!handle)
240                 return NULL;
241         memset(handle, 0, sizeof(*handle));
242         handle->h_buffer_credits = nblocks;
243         handle->h_ref = 1;
244
245         return handle;
246 }
247
248 /**
249  * handle_t *jbd2_journal_start() - Obtain a new handle.
250  * @journal: Journal to start transaction on.
251  * @nblocks: number of block buffer we might modify
252  *
253  * We make sure that the transaction can guarantee at least nblocks of
254  * modified buffers in the log.  We block until the log can guarantee
255  * that much space.
256  *
257  * This function is visible to journal users (like ext3fs), so is not
258  * called with the journal already locked.
259  *
260  * Return a pointer to a newly allocated handle, or NULL on failure
261  */
262 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
263 {
264         handle_t *handle = journal_current_handle();
265         int err;
266
267         if (!journal)
268                 return ERR_PTR(-EROFS);
269
270         if (handle) {
271                 J_ASSERT(handle->h_transaction->t_journal == journal);
272                 handle->h_ref++;
273                 return handle;
274         }
275
276         handle = new_handle(nblocks);
277         if (!handle)
278                 return ERR_PTR(-ENOMEM);
279
280         current->journal_info = handle;
281
282         err = start_this_handle(journal, handle);
283         if (err < 0) {
284                 jbd_free_handle(handle);
285                 current->journal_info = NULL;
286                 handle = ERR_PTR(err);
287         }
288         return handle;
289 }
290
291 /**
292  * int jbd2_journal_extend() - extend buffer credits.
293  * @handle:  handle to 'extend'
294  * @nblocks: nr blocks to try to extend by.
295  *
296  * Some transactions, such as large extends and truncates, can be done
297  * atomically all at once or in several stages.  The operation requests
298  * a credit for a number of buffer modications in advance, but can
299  * extend its credit if it needs more.
300  *
301  * jbd2_journal_extend tries to give the running handle more buffer credits.
302  * It does not guarantee that allocation - this is a best-effort only.
303  * The calling process MUST be able to deal cleanly with a failure to
304  * extend here.
305  *
306  * Return 0 on success, non-zero on failure.
307  *
308  * return code < 0 implies an error
309  * return code > 0 implies normal transaction-full status.
310  */
311 int jbd2_journal_extend(handle_t *handle, int nblocks)
312 {
313         transaction_t *transaction = handle->h_transaction;
314         journal_t *journal = transaction->t_journal;
315         int result;
316         int wanted;
317
318         result = -EIO;
319         if (is_handle_aborted(handle))
320                 goto out;
321
322         result = 1;
323
324         spin_lock(&journal->j_state_lock);
325
326         /* Don't extend a locked-down transaction! */
327         if (handle->h_transaction->t_state != T_RUNNING) {
328                 jbd_debug(3, "denied handle %p %d blocks: "
329                           "transaction not running\n", handle, nblocks);
330                 goto error_out;
331         }
332
333         spin_lock(&transaction->t_handle_lock);
334         wanted = transaction->t_outstanding_credits + nblocks;
335
336         if (wanted > journal->j_max_transaction_buffers) {
337                 jbd_debug(3, "denied handle %p %d blocks: "
338                           "transaction too large\n", handle, nblocks);
339                 goto unlock;
340         }
341
342         if (wanted > __jbd2_log_space_left(journal)) {
343                 jbd_debug(3, "denied handle %p %d blocks: "
344                           "insufficient log space\n", handle, nblocks);
345                 goto unlock;
346         }
347
348         handle->h_buffer_credits += nblocks;
349         transaction->t_outstanding_credits += nblocks;
350         result = 0;
351
352         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
353 unlock:
354         spin_unlock(&transaction->t_handle_lock);
355 error_out:
356         spin_unlock(&journal->j_state_lock);
357 out:
358         return result;
359 }
360
361
362 /**
363  * int jbd2_journal_restart() - restart a handle .
364  * @handle:  handle to restart
365  * @nblocks: nr credits requested
366  *
367  * Restart a handle for a multi-transaction filesystem
368  * operation.
369  *
370  * If the jbd2_journal_extend() call above fails to grant new buffer credits
371  * to a running handle, a call to jbd2_journal_restart will commit the
372  * handle's transaction so far and reattach the handle to a new
373  * transaction capabable of guaranteeing the requested number of
374  * credits.
375  */
376
377 int jbd2_journal_restart(handle_t *handle, int nblocks)
378 {
379         transaction_t *transaction = handle->h_transaction;
380         journal_t *journal = transaction->t_journal;
381         int ret;
382
383         /* If we've had an abort of any type, don't even think about
384          * actually doing the restart! */
385         if (is_handle_aborted(handle))
386                 return 0;
387
388         /*
389          * First unlink the handle from its current transaction, and start the
390          * commit on that.
391          */
392         J_ASSERT(transaction->t_updates > 0);
393         J_ASSERT(journal_current_handle() == handle);
394
395         spin_lock(&journal->j_state_lock);
396         spin_lock(&transaction->t_handle_lock);
397         transaction->t_outstanding_credits -= handle->h_buffer_credits;
398         transaction->t_updates--;
399
400         if (!transaction->t_updates)
401                 wake_up(&journal->j_wait_updates);
402         spin_unlock(&transaction->t_handle_lock);
403
404         jbd_debug(2, "restarting handle %p\n", handle);
405         __jbd2_log_start_commit(journal, transaction->t_tid);
406         spin_unlock(&journal->j_state_lock);
407
408         handle->h_buffer_credits = nblocks;
409         ret = start_this_handle(journal, handle);
410         return ret;
411 }
412
413
414 /**
415  * void jbd2_journal_lock_updates () - establish a transaction barrier.
416  * @journal:  Journal to establish a barrier on.
417  *
418  * This locks out any further updates from being started, and blocks
419  * until all existing updates have completed, returning only once the
420  * journal is in a quiescent state with no updates running.
421  *
422  * The journal lock should not be held on entry.
423  */
424 void jbd2_journal_lock_updates(journal_t *journal)
425 {
426         DEFINE_WAIT(wait);
427
428         spin_lock(&journal->j_state_lock);
429         ++journal->j_barrier_count;
430
431         /* Wait until there are no running updates */
432         while (1) {
433                 transaction_t *transaction = journal->j_running_transaction;
434
435                 if (!transaction)
436                         break;
437
438                 spin_lock(&transaction->t_handle_lock);
439                 if (!transaction->t_updates) {
440                         spin_unlock(&transaction->t_handle_lock);
441                         break;
442                 }
443                 prepare_to_wait(&journal->j_wait_updates, &wait,
444                                 TASK_UNINTERRUPTIBLE);
445                 spin_unlock(&transaction->t_handle_lock);
446                 spin_unlock(&journal->j_state_lock);
447                 schedule();
448                 finish_wait(&journal->j_wait_updates, &wait);
449                 spin_lock(&journal->j_state_lock);
450         }
451         spin_unlock(&journal->j_state_lock);
452
453         /*
454          * We have now established a barrier against other normal updates, but
455          * we also need to barrier against other jbd2_journal_lock_updates() calls
456          * to make sure that we serialise special journal-locked operations
457          * too.
458          */
459         mutex_lock(&journal->j_barrier);
460 }
461
462 /**
463  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
464  * @journal:  Journal to release the barrier on.
465  *
466  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
467  *
468  * Should be called without the journal lock held.
469  */
470 void jbd2_journal_unlock_updates (journal_t *journal)
471 {
472         J_ASSERT(journal->j_barrier_count != 0);
473
474         mutex_unlock(&journal->j_barrier);
475         spin_lock(&journal->j_state_lock);
476         --journal->j_barrier_count;
477         spin_unlock(&journal->j_state_lock);
478         wake_up(&journal->j_wait_transaction_locked);
479 }
480
481 /*
482  * Report any unexpected dirty buffers which turn up.  Normally those
483  * indicate an error, but they can occur if the user is running (say)
484  * tune2fs to modify the live filesystem, so we need the option of
485  * continuing as gracefully as possible.  #
486  *
487  * The caller should already hold the journal lock and
488  * j_list_lock spinlock: most callers will need those anyway
489  * in order to probe the buffer's journaling state safely.
490  */
491 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
492 {
493         int jlist;
494
495         /* If this buffer is one which might reasonably be dirty
496          * --- ie. data, or not part of this journal --- then
497          * we're OK to leave it alone, but otherwise we need to
498          * move the dirty bit to the journal's own internal
499          * JBDDirty bit. */
500         jlist = jh->b_jlist;
501
502         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
503             jlist == BJ_Shadow || jlist == BJ_Forget) {
504                 struct buffer_head *bh = jh2bh(jh);
505
506                 if (test_clear_buffer_dirty(bh))
507                         set_buffer_jbddirty(bh);
508         }
509 }
510
511 /*
512  * If the buffer is already part of the current transaction, then there
513  * is nothing we need to do.  If it is already part of a prior
514  * transaction which we are still committing to disk, then we need to
515  * make sure that we do not overwrite the old copy: we do copy-out to
516  * preserve the copy going to disk.  We also account the buffer against
517  * the handle's metadata buffer credits (unless the buffer is already
518  * part of the transaction, that is).
519  *
520  */
521 static int
522 do_get_write_access(handle_t *handle, struct journal_head *jh,
523                         int force_copy)
524 {
525         struct buffer_head *bh;
526         transaction_t *transaction;
527         journal_t *journal;
528         int error;
529         char *frozen_buffer = NULL;
530         int need_copy = 0;
531
532         if (is_handle_aborted(handle))
533                 return -EROFS;
534
535         transaction = handle->h_transaction;
536         journal = transaction->t_journal;
537
538         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
539
540         JBUFFER_TRACE(jh, "entry");
541 repeat:
542         bh = jh2bh(jh);
543
544         /* @@@ Need to check for errors here at some point. */
545
546         lock_buffer(bh);
547         jbd_lock_bh_state(bh);
548
549         /* We now hold the buffer lock so it is safe to query the buffer
550          * state.  Is the buffer dirty?
551          *
552          * If so, there are two possibilities.  The buffer may be
553          * non-journaled, and undergoing a quite legitimate writeback.
554          * Otherwise, it is journaled, and we don't expect dirty buffers
555          * in that state (the buffers should be marked JBD_Dirty
556          * instead.)  So either the IO is being done under our own
557          * control and this is a bug, or it's a third party IO such as
558          * dump(8) (which may leave the buffer scheduled for read ---
559          * ie. locked but not dirty) or tune2fs (which may actually have
560          * the buffer dirtied, ugh.)  */
561
562         if (buffer_dirty(bh)) {
563                 /*
564                  * First question: is this buffer already part of the current
565                  * transaction or the existing committing transaction?
566                  */
567                 if (jh->b_transaction) {
568                         J_ASSERT_JH(jh,
569                                 jh->b_transaction == transaction ||
570                                 jh->b_transaction ==
571                                         journal->j_committing_transaction);
572                         if (jh->b_next_transaction)
573                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
574                                                         transaction);
575                 }
576                 /*
577                  * In any case we need to clean the dirty flag and we must
578                  * do it under the buffer lock to be sure we don't race
579                  * with running write-out.
580                  */
581                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
582                 jbd_unexpected_dirty_buffer(jh);
583         }
584
585         unlock_buffer(bh);
586
587         error = -EROFS;
588         if (is_handle_aborted(handle)) {
589                 jbd_unlock_bh_state(bh);
590                 goto out;
591         }
592         error = 0;
593
594         /*
595          * The buffer is already part of this transaction if b_transaction or
596          * b_next_transaction points to it
597          */
598         if (jh->b_transaction == transaction ||
599             jh->b_next_transaction == transaction)
600                 goto done;
601
602         /*
603          * If there is already a copy-out version of this buffer, then we don't
604          * need to make another one
605          */
606         if (jh->b_frozen_data) {
607                 JBUFFER_TRACE(jh, "has frozen data");
608                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
609                 jh->b_next_transaction = transaction;
610                 goto done;
611         }
612
613         /* Is there data here we need to preserve? */
614
615         if (jh->b_transaction && jh->b_transaction != transaction) {
616                 JBUFFER_TRACE(jh, "owned by older transaction");
617                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
618                 J_ASSERT_JH(jh, jh->b_transaction ==
619                                         journal->j_committing_transaction);
620
621                 /* There is one case we have to be very careful about.
622                  * If the committing transaction is currently writing
623                  * this buffer out to disk and has NOT made a copy-out,
624                  * then we cannot modify the buffer contents at all
625                  * right now.  The essence of copy-out is that it is the
626                  * extra copy, not the primary copy, which gets
627                  * journaled.  If the primary copy is already going to
628                  * disk then we cannot do copy-out here. */
629
630                 if (jh->b_jlist == BJ_Shadow) {
631                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
632                         wait_queue_head_t *wqh;
633
634                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
635
636                         JBUFFER_TRACE(jh, "on shadow: sleep");
637                         jbd_unlock_bh_state(bh);
638                         /* commit wakes up all shadow buffers after IO */
639                         for ( ; ; ) {
640                                 prepare_to_wait(wqh, &wait.wait,
641                                                 TASK_UNINTERRUPTIBLE);
642                                 if (jh->b_jlist != BJ_Shadow)
643                                         break;
644                                 schedule();
645                         }
646                         finish_wait(wqh, &wait.wait);
647                         goto repeat;
648                 }
649
650                 /* Only do the copy if the currently-owning transaction
651                  * still needs it.  If it is on the Forget list, the
652                  * committing transaction is past that stage.  The
653                  * buffer had better remain locked during the kmalloc,
654                  * but that should be true --- we hold the journal lock
655                  * still and the buffer is already on the BUF_JOURNAL
656                  * list so won't be flushed.
657                  *
658                  * Subtle point, though: if this is a get_undo_access,
659                  * then we will be relying on the frozen_data to contain
660                  * the new value of the committed_data record after the
661                  * transaction, so we HAVE to force the frozen_data copy
662                  * in that case. */
663
664                 if (jh->b_jlist != BJ_Forget || force_copy) {
665                         JBUFFER_TRACE(jh, "generate frozen data");
666                         if (!frozen_buffer) {
667                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
668                                 jbd_unlock_bh_state(bh);
669                                 frozen_buffer =
670                                         jbd2_slab_alloc(jh2bh(jh)->b_size,
671                                                          GFP_NOFS);
672                                 if (!frozen_buffer) {
673                                         printk(KERN_EMERG
674                                                "%s: OOM for frozen_buffer\n",
675                                                __FUNCTION__);
676                                         JBUFFER_TRACE(jh, "oom!");
677                                         error = -ENOMEM;
678                                         jbd_lock_bh_state(bh);
679                                         goto done;
680                                 }
681                                 goto repeat;
682                         }
683                         jh->b_frozen_data = frozen_buffer;
684                         frozen_buffer = NULL;
685                         need_copy = 1;
686                 }
687                 jh->b_next_transaction = transaction;
688         }
689
690
691         /*
692          * Finally, if the buffer is not journaled right now, we need to make
693          * sure it doesn't get written to disk before the caller actually
694          * commits the new data
695          */
696         if (!jh->b_transaction) {
697                 JBUFFER_TRACE(jh, "no transaction");
698                 J_ASSERT_JH(jh, !jh->b_next_transaction);
699                 jh->b_transaction = transaction;
700                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
701                 spin_lock(&journal->j_list_lock);
702                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
703                 spin_unlock(&journal->j_list_lock);
704         }
705
706 done:
707         if (need_copy) {
708                 struct page *page;
709                 int offset;
710                 char *source;
711
712                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
713                             "Possible IO failure.\n");
714                 page = jh2bh(jh)->b_page;
715                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
716                 source = kmap_atomic(page, KM_USER0);
717                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
718                 kunmap_atomic(source, KM_USER0);
719         }
720         jbd_unlock_bh_state(bh);
721
722         /*
723          * If we are about to journal a buffer, then any revoke pending on it is
724          * no longer valid
725          */
726         jbd2_journal_cancel_revoke(handle, jh);
727
728 out:
729         if (unlikely(frozen_buffer))    /* It's usually NULL */
730                 jbd2_slab_free(frozen_buffer, bh->b_size);
731
732         JBUFFER_TRACE(jh, "exit");
733         return error;
734 }
735
736 /**
737  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
738  * @handle: transaction to add buffer modifications to
739  * @bh:     bh to be used for metadata writes
740  * @credits: variable that will receive credits for the buffer
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 jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
749 {
750         struct journal_head *jh = jbd2_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         jbd2_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 jbd2_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 jbd2_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 = jbd2_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                 jh->b_transaction = transaction;
814                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
815                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
816         } else if (jh->b_transaction == journal->j_committing_transaction) {
817                 JBUFFER_TRACE(jh, "set next transaction");
818                 jh->b_next_transaction = transaction;
819         }
820         spin_unlock(&journal->j_list_lock);
821         jbd_unlock_bh_state(bh);
822
823         /*
824          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
825          * blocks which contain freed but then revoked metadata.  We need
826          * to cancel the revoke in case we end up freeing it yet again
827          * and the reallocating as data - this would cause a second revoke,
828          * which hits an assertion error.
829          */
830         JBUFFER_TRACE(jh, "cancelling revoke");
831         jbd2_journal_cancel_revoke(handle, jh);
832         jbd2_journal_put_journal_head(jh);
833 out:
834         return err;
835 }
836
837 /**
838  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
839  *     non-rewindable consequences
840  * @handle: transaction
841  * @bh: buffer to undo
842  * @credits: store the number of taken credits here (if not NULL)
843  *
844  * Sometimes there is a need to distinguish between metadata which has
845  * been committed to disk and that which has not.  The ext3fs code uses
846  * this for freeing and allocating space, we have to make sure that we
847  * do not reuse freed space until the deallocation has been committed,
848  * since if we overwrote that space we would make the delete
849  * un-rewindable in case of a crash.
850  *
851  * To deal with that, jbd2_journal_get_undo_access requests write access to a
852  * buffer for parts of non-rewindable operations such as delete
853  * operations on the bitmaps.  The journaling code must keep a copy of
854  * the buffer's contents prior to the undo_access call until such time
855  * as we know that the buffer has definitely been committed to disk.
856  *
857  * We never need to know which transaction the committed data is part
858  * of, buffers touched here are guaranteed to be dirtied later and so
859  * will be committed to a new transaction in due course, at which point
860  * we can discard the old committed data pointer.
861  *
862  * Returns error number or 0 on success.
863  */
864 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
865 {
866         int err;
867         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
868         char *committed_data = NULL;
869
870         JBUFFER_TRACE(jh, "entry");
871
872         /*
873          * Do this first --- it can drop the journal lock, so we want to
874          * make sure that obtaining the committed_data is done
875          * atomically wrt. completion of any outstanding commits.
876          */
877         err = do_get_write_access(handle, jh, 1);
878         if (err)
879                 goto out;
880
881 repeat:
882         if (!jh->b_committed_data) {
883                 committed_data = jbd2_slab_alloc(jh2bh(jh)->b_size, GFP_NOFS);
884                 if (!committed_data) {
885                         printk(KERN_EMERG "%s: No memory for committed data\n",
886                                 __FUNCTION__);
887                         err = -ENOMEM;
888                         goto out;
889                 }
890         }
891
892         jbd_lock_bh_state(bh);
893         if (!jh->b_committed_data) {
894                 /* Copy out the current buffer contents into the
895                  * preserved, committed copy. */
896                 JBUFFER_TRACE(jh, "generate b_committed data");
897                 if (!committed_data) {
898                         jbd_unlock_bh_state(bh);
899                         goto repeat;
900                 }
901
902                 jh->b_committed_data = committed_data;
903                 committed_data = NULL;
904                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
905         }
906         jbd_unlock_bh_state(bh);
907 out:
908         jbd2_journal_put_journal_head(jh);
909         if (unlikely(committed_data))
910                 jbd2_slab_free(committed_data, bh->b_size);
911         return err;
912 }
913
914 /**
915  * int jbd2_journal_dirty_data() -  mark a buffer as containing dirty data which
916  *                             needs to be flushed before we can commit the
917  *                             current transaction.
918  * @handle: transaction
919  * @bh: bufferhead to mark
920  *
921  * The buffer is placed on the transaction's data list and is marked as
922  * belonging to the transaction.
923  *
924  * Returns error number or 0 on success.
925  *
926  * jbd2_journal_dirty_data() can be called via page_launder->ext3_writepage
927  * by kswapd.
928  */
929 int jbd2_journal_dirty_data(handle_t *handle, struct buffer_head *bh)
930 {
931         journal_t *journal = handle->h_transaction->t_journal;
932         int need_brelse = 0;
933         struct journal_head *jh;
934
935         if (is_handle_aborted(handle))
936                 return 0;
937
938         jh = jbd2_journal_add_journal_head(bh);
939         JBUFFER_TRACE(jh, "entry");
940
941         /*
942          * The buffer could *already* be dirty.  Writeout can start
943          * at any time.
944          */
945         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
946
947         /*
948          * What if the buffer is already part of a running transaction?
949          *
950          * There are two cases:
951          * 1) It is part of the current running transaction.  Refile it,
952          *    just in case we have allocated it as metadata, deallocated
953          *    it, then reallocated it as data.
954          * 2) It is part of the previous, still-committing transaction.
955          *    If all we want to do is to guarantee that the buffer will be
956          *    written to disk before this new transaction commits, then
957          *    being sure that the *previous* transaction has this same
958          *    property is sufficient for us!  Just leave it on its old
959          *    transaction.
960          *
961          * In case (2), the buffer must not already exist as metadata
962          * --- that would violate write ordering (a transaction is free
963          * to write its data at any point, even before the previous
964          * committing transaction has committed).  The caller must
965          * never, ever allow this to happen: there's nothing we can do
966          * about it in this layer.
967          */
968         jbd_lock_bh_state(bh);
969         spin_lock(&journal->j_list_lock);
970         if (jh->b_transaction) {
971                 JBUFFER_TRACE(jh, "has transaction");
972                 if (jh->b_transaction != handle->h_transaction) {
973                         JBUFFER_TRACE(jh, "belongs to older transaction");
974                         J_ASSERT_JH(jh, jh->b_transaction ==
975                                         journal->j_committing_transaction);
976
977                         /* @@@ IS THIS TRUE  ? */
978                         /*
979                          * Not any more.  Scenario: someone does a write()
980                          * in data=journal mode.  The buffer's transaction has
981                          * moved into commit.  Then someone does another
982                          * write() to the file.  We do the frozen data copyout
983                          * and set b_next_transaction to point to j_running_t.
984                          * And while we're in that state, someone does a
985                          * writepage() in an attempt to pageout the same area
986                          * of the file via a shared mapping.  At present that
987                          * calls jbd2_journal_dirty_data(), and we get right here.
988                          * It may be too late to journal the data.  Simply
989                          * falling through to the next test will suffice: the
990                          * data will be dirty and wil be checkpointed.  The
991                          * ordering comments in the next comment block still
992                          * apply.
993                          */
994                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
995
996                         /*
997                          * If we're journalling data, and this buffer was
998                          * subject to a write(), it could be metadata, forget
999                          * or shadow against the committing transaction.  Now,
1000                          * someone has dirtied the same darn page via a mapping
1001                          * and it is being writepage()'d.
1002                          * We *could* just steal the page from commit, with some
1003                          * fancy locking there.  Instead, we just skip it -
1004                          * don't tie the page's buffers to the new transaction
1005                          * at all.
1006                          * Implication: if we crash before the writepage() data
1007                          * is written into the filesystem, recovery will replay
1008                          * the write() data.
1009                          */
1010                         if (jh->b_jlist != BJ_None &&
1011                                         jh->b_jlist != BJ_SyncData &&
1012                                         jh->b_jlist != BJ_Locked) {
1013                                 JBUFFER_TRACE(jh, "Not stealing");
1014                                 goto no_journal;
1015                         }
1016
1017                         /*
1018                          * This buffer may be undergoing writeout in commit.  We
1019                          * can't return from here and let the caller dirty it
1020                          * again because that can cause the write-out loop in
1021                          * commit to never terminate.
1022                          */
1023                         if (buffer_dirty(bh)) {
1024                                 get_bh(bh);
1025                                 spin_unlock(&journal->j_list_lock);
1026                                 jbd_unlock_bh_state(bh);
1027                                 need_brelse = 1;
1028                                 sync_dirty_buffer(bh);
1029                                 jbd_lock_bh_state(bh);
1030                                 spin_lock(&journal->j_list_lock);
1031                                 /* The buffer may become locked again at any
1032                                    time if it is redirtied */
1033                         }
1034
1035                         /* journal_clean_data_list() may have got there first */
1036                         if (jh->b_transaction != NULL) {
1037                                 JBUFFER_TRACE(jh, "unfile from commit");
1038                                 __jbd2_journal_temp_unlink_buffer(jh);
1039                                 /* It still points to the committing
1040                                  * transaction; move it to this one so
1041                                  * that the refile assert checks are
1042                                  * happy. */
1043                                 jh->b_transaction = handle->h_transaction;
1044                         }
1045                         /* The buffer will be refiled below */
1046
1047                 }
1048                 /*
1049                  * Special case --- the buffer might actually have been
1050                  * allocated and then immediately deallocated in the previous,
1051                  * committing transaction, so might still be left on that
1052                  * transaction's metadata lists.
1053                  */
1054                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1055                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1056                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1057                         __jbd2_journal_temp_unlink_buffer(jh);
1058                         jh->b_transaction = handle->h_transaction;
1059                         JBUFFER_TRACE(jh, "file as data");
1060                         __jbd2_journal_file_buffer(jh, handle->h_transaction,
1061                                                 BJ_SyncData);
1062                 }
1063         } else {
1064                 JBUFFER_TRACE(jh, "not on a transaction");
1065                 __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1066         }
1067 no_journal:
1068         spin_unlock(&journal->j_list_lock);
1069         jbd_unlock_bh_state(bh);
1070         if (need_brelse) {
1071                 BUFFER_TRACE(bh, "brelse");
1072                 __brelse(bh);
1073         }
1074         JBUFFER_TRACE(jh, "exit");
1075         jbd2_journal_put_journal_head(jh);
1076         return 0;
1077 }
1078
1079 /**
1080  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1081  * @handle: transaction to add buffer to.
1082  * @bh: buffer to mark
1083  *
1084  * mark dirty metadata which needs to be journaled as part of the current
1085  * transaction.
1086  *
1087  * The buffer is placed on the transaction's metadata list and is marked
1088  * as belonging to the transaction.
1089  *
1090  * Returns error number or 0 on success.
1091  *
1092  * Special care needs to be taken if the buffer already belongs to the
1093  * current committing transaction (in which case we should have frozen
1094  * data present for that commit).  In that case, we don't relink the
1095  * buffer: that only gets done when the old transaction finally
1096  * completes its commit.
1097  */
1098 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1099 {
1100         transaction_t *transaction = handle->h_transaction;
1101         journal_t *journal = transaction->t_journal;
1102         struct journal_head *jh = bh2jh(bh);
1103
1104         jbd_debug(5, "journal_head %p\n", jh);
1105         JBUFFER_TRACE(jh, "entry");
1106         if (is_handle_aborted(handle))
1107                 goto out;
1108
1109         jbd_lock_bh_state(bh);
1110
1111         if (jh->b_modified == 0) {
1112                 /*
1113                  * This buffer's got modified and becoming part
1114                  * of the transaction. This needs to be done
1115                  * once a transaction -bzzz
1116                  */
1117                 jh->b_modified = 1;
1118                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1119                 handle->h_buffer_credits--;
1120         }
1121
1122         /*
1123          * fastpath, to avoid expensive locking.  If this buffer is already
1124          * on the running transaction's metadata list there is nothing to do.
1125          * Nobody can take it off again because there is a handle open.
1126          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1127          * result in this test being false, so we go in and take the locks.
1128          */
1129         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1130                 JBUFFER_TRACE(jh, "fastpath");
1131                 J_ASSERT_JH(jh, jh->b_transaction ==
1132                                         journal->j_running_transaction);
1133                 goto out_unlock_bh;
1134         }
1135
1136         set_buffer_jbddirty(bh);
1137
1138         /*
1139          * Metadata already on the current transaction list doesn't
1140          * need to be filed.  Metadata on another transaction's list must
1141          * be committing, and will be refiled once the commit completes:
1142          * leave it alone for now.
1143          */
1144         if (jh->b_transaction != transaction) {
1145                 JBUFFER_TRACE(jh, "already on other transaction");
1146                 J_ASSERT_JH(jh, jh->b_transaction ==
1147                                         journal->j_committing_transaction);
1148                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1149                 /* And this case is illegal: we can't reuse another
1150                  * transaction's data buffer, ever. */
1151                 goto out_unlock_bh;
1152         }
1153
1154         /* That test should have eliminated the following case: */
1155         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1156
1157         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1158         spin_lock(&journal->j_list_lock);
1159         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1160         spin_unlock(&journal->j_list_lock);
1161 out_unlock_bh:
1162         jbd_unlock_bh_state(bh);
1163 out:
1164         JBUFFER_TRACE(jh, "exit");
1165         return 0;
1166 }
1167
1168 /*
1169  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1170  * updates, if the update decided in the end that it didn't need access.
1171  *
1172  */
1173 void
1174 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1175 {
1176         BUFFER_TRACE(bh, "entry");
1177 }
1178
1179 /**
1180  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1181  * @handle: transaction handle
1182  * @bh:     bh to 'forget'
1183  *
1184  * We can only do the bforget if there are no commits pending against the
1185  * buffer.  If the buffer is dirty in the current running transaction we
1186  * can safely unlink it.
1187  *
1188  * bh may not be a journalled buffer at all - it may be a non-JBD
1189  * buffer which came off the hashtable.  Check for this.
1190  *
1191  * Decrements bh->b_count by one.
1192  *
1193  * Allow this call even if the handle has aborted --- it may be part of
1194  * the caller's cleanup after an abort.
1195  */
1196 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1197 {
1198         transaction_t *transaction = handle->h_transaction;
1199         journal_t *journal = transaction->t_journal;
1200         struct journal_head *jh;
1201         int drop_reserve = 0;
1202         int err = 0;
1203
1204         BUFFER_TRACE(bh, "entry");
1205
1206         jbd_lock_bh_state(bh);
1207         spin_lock(&journal->j_list_lock);
1208
1209         if (!buffer_jbd(bh))
1210                 goto not_jbd;
1211         jh = bh2jh(bh);
1212
1213         /* Critical error: attempting to delete a bitmap buffer, maybe?
1214          * Don't do any jbd operations, and return an error. */
1215         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1216                          "inconsistent data on disk")) {
1217                 err = -EIO;
1218                 goto not_jbd;
1219         }
1220
1221         /*
1222          * The buffer's going from the transaction, we must drop
1223          * all references -bzzz
1224          */
1225         jh->b_modified = 0;
1226
1227         if (jh->b_transaction == handle->h_transaction) {
1228                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1229
1230                 /* If we are forgetting a buffer which is already part
1231                  * of this transaction, then we can just drop it from
1232                  * the transaction immediately. */
1233                 clear_buffer_dirty(bh);
1234                 clear_buffer_jbddirty(bh);
1235
1236                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1237
1238                 drop_reserve = 1;
1239
1240                 /*
1241                  * We are no longer going to journal this buffer.
1242                  * However, the commit of this transaction is still
1243                  * important to the buffer: the delete that we are now
1244                  * processing might obsolete an old log entry, so by
1245                  * committing, we can satisfy the buffer's checkpoint.
1246                  *
1247                  * So, if we have a checkpoint on the buffer, we should
1248                  * now refile the buffer on our BJ_Forget list so that
1249                  * we know to remove the checkpoint after we commit.
1250                  */
1251
1252                 if (jh->b_cp_transaction) {
1253                         __jbd2_journal_temp_unlink_buffer(jh);
1254                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1255                 } else {
1256                         __jbd2_journal_unfile_buffer(jh);
1257                         jbd2_journal_remove_journal_head(bh);
1258                         __brelse(bh);
1259                         if (!buffer_jbd(bh)) {
1260                                 spin_unlock(&journal->j_list_lock);
1261                                 jbd_unlock_bh_state(bh);
1262                                 __bforget(bh);
1263                                 goto drop;
1264                         }
1265                 }
1266         } else if (jh->b_transaction) {
1267                 J_ASSERT_JH(jh, (jh->b_transaction ==
1268                                  journal->j_committing_transaction));
1269                 /* However, if the buffer is still owned by a prior
1270                  * (committing) transaction, we can't drop it yet... */
1271                 JBUFFER_TRACE(jh, "belongs to older transaction");
1272                 /* ... but we CAN drop it from the new transaction if we
1273                  * have also modified it since the original commit. */
1274
1275                 if (jh->b_next_transaction) {
1276                         J_ASSERT(jh->b_next_transaction == transaction);
1277                         jh->b_next_transaction = NULL;
1278                         drop_reserve = 1;
1279                 }
1280         }
1281
1282 not_jbd:
1283         spin_unlock(&journal->j_list_lock);
1284         jbd_unlock_bh_state(bh);
1285         __brelse(bh);
1286 drop:
1287         if (drop_reserve) {
1288                 /* no need to reserve log space for this block -bzzz */
1289                 handle->h_buffer_credits++;
1290         }
1291         return err;
1292 }
1293
1294 /**
1295  * int jbd2_journal_stop() - complete a transaction
1296  * @handle: tranaction to complete.
1297  *
1298  * All done for a particular handle.
1299  *
1300  * There is not much action needed here.  We just return any remaining
1301  * buffer credits to the transaction and remove the handle.  The only
1302  * complication is that we need to start a commit operation if the
1303  * filesystem is marked for synchronous update.
1304  *
1305  * jbd2_journal_stop itself will not usually return an error, but it may
1306  * do so in unusual circumstances.  In particular, expect it to
1307  * return -EIO if a jbd2_journal_abort has been executed since the
1308  * transaction began.
1309  */
1310 int jbd2_journal_stop(handle_t *handle)
1311 {
1312         transaction_t *transaction = handle->h_transaction;
1313         journal_t *journal = transaction->t_journal;
1314         int old_handle_count, err;
1315         pid_t pid;
1316
1317         J_ASSERT(journal_current_handle() == handle);
1318
1319         if (is_handle_aborted(handle))
1320                 err = -EIO;
1321         else {
1322                 J_ASSERT(transaction->t_updates > 0);
1323                 err = 0;
1324         }
1325
1326         if (--handle->h_ref > 0) {
1327                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1328                           handle->h_ref);
1329                 return err;
1330         }
1331
1332         jbd_debug(4, "Handle %p going down\n", handle);
1333
1334         /*
1335          * Implement synchronous transaction batching.  If the handle
1336          * was synchronous, don't force a commit immediately.  Let's
1337          * yield and let another thread piggyback onto this transaction.
1338          * Keep doing that while new threads continue to arrive.
1339          * It doesn't cost much - we're about to run a commit and sleep
1340          * on IO anyway.  Speeds up many-threaded, many-dir operations
1341          * by 30x or more...
1342          *
1343          * But don't do this if this process was the most recent one to
1344          * perform a synchronous write.  We do this to detect the case where a
1345          * single process is doing a stream of sync writes.  No point in waiting
1346          * for joiners in that case.
1347          */
1348         pid = current->pid;
1349         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1350                 journal->j_last_sync_writer = pid;
1351                 do {
1352                         old_handle_count = transaction->t_handle_count;
1353                         schedule_timeout_uninterruptible(1);
1354                 } while (old_handle_count != transaction->t_handle_count);
1355         }
1356
1357         current->journal_info = NULL;
1358         spin_lock(&journal->j_state_lock);
1359         spin_lock(&transaction->t_handle_lock);
1360         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1361         transaction->t_updates--;
1362         if (!transaction->t_updates) {
1363                 wake_up(&journal->j_wait_updates);
1364                 if (journal->j_barrier_count)
1365                         wake_up(&journal->j_wait_transaction_locked);
1366         }
1367
1368         /*
1369          * If the handle is marked SYNC, we need to set another commit
1370          * going!  We also want to force a commit if the current
1371          * transaction is occupying too much of the log, or if the
1372          * transaction is too old now.
1373          */
1374         if (handle->h_sync ||
1375                         transaction->t_outstanding_credits >
1376                                 journal->j_max_transaction_buffers ||
1377                         time_after_eq(jiffies, transaction->t_expires)) {
1378                 /* Do this even for aborted journals: an abort still
1379                  * completes the commit thread, it just doesn't write
1380                  * anything to disk. */
1381                 tid_t tid = transaction->t_tid;
1382
1383                 spin_unlock(&transaction->t_handle_lock);
1384                 jbd_debug(2, "transaction too old, requesting commit for "
1385                                         "handle %p\n", handle);
1386                 /* This is non-blocking */
1387                 __jbd2_log_start_commit(journal, transaction->t_tid);
1388                 spin_unlock(&journal->j_state_lock);
1389
1390                 /*
1391                  * Special case: JBD2_SYNC synchronous updates require us
1392                  * to wait for the commit to complete.
1393                  */
1394                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1395                         err = jbd2_log_wait_commit(journal, tid);
1396         } else {
1397                 spin_unlock(&transaction->t_handle_lock);
1398                 spin_unlock(&journal->j_state_lock);
1399         }
1400
1401         jbd_free_handle(handle);
1402         return err;
1403 }
1404
1405 /**int jbd2_journal_force_commit() - force any uncommitted transactions
1406  * @journal: journal to force
1407  *
1408  * For synchronous operations: force any uncommitted transactions
1409  * to disk.  May seem kludgy, but it reuses all the handle batching
1410  * code in a very simple manner.
1411  */
1412 int jbd2_journal_force_commit(journal_t *journal)
1413 {
1414         handle_t *handle;
1415         int ret;
1416
1417         handle = jbd2_journal_start(journal, 1);
1418         if (IS_ERR(handle)) {
1419                 ret = PTR_ERR(handle);
1420         } else {
1421                 handle->h_sync = 1;
1422                 ret = jbd2_journal_stop(handle);
1423         }
1424         return ret;
1425 }
1426
1427 /*
1428  *
1429  * List management code snippets: various functions for manipulating the
1430  * transaction buffer lists.
1431  *
1432  */
1433
1434 /*
1435  * Append a buffer to a transaction list, given the transaction's list head
1436  * pointer.
1437  *
1438  * j_list_lock is held.
1439  *
1440  * jbd_lock_bh_state(jh2bh(jh)) is held.
1441  */
1442
1443 static inline void
1444 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1445 {
1446         if (!*list) {
1447                 jh->b_tnext = jh->b_tprev = jh;
1448                 *list = jh;
1449         } else {
1450                 /* Insert at the tail of the list to preserve order */
1451                 struct journal_head *first = *list, *last = first->b_tprev;
1452                 jh->b_tprev = last;
1453                 jh->b_tnext = first;
1454                 last->b_tnext = first->b_tprev = jh;
1455         }
1456 }
1457
1458 /*
1459  * Remove a buffer from a transaction list, given the transaction's list
1460  * head pointer.
1461  *
1462  * Called with j_list_lock held, and the journal may not be locked.
1463  *
1464  * jbd_lock_bh_state(jh2bh(jh)) is held.
1465  */
1466
1467 static inline void
1468 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1469 {
1470         if (*list == jh) {
1471                 *list = jh->b_tnext;
1472                 if (*list == jh)
1473                         *list = NULL;
1474         }
1475         jh->b_tprev->b_tnext = jh->b_tnext;
1476         jh->b_tnext->b_tprev = jh->b_tprev;
1477 }
1478
1479 /*
1480  * Remove a buffer from the appropriate transaction list.
1481  *
1482  * Note that this function can *change* the value of
1483  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1484  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1485  * is holding onto a copy of one of thee pointers, it could go bad.
1486  * Generally the caller needs to re-read the pointer from the transaction_t.
1487  *
1488  * Called under j_list_lock.  The journal may not be locked.
1489  */
1490 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1491 {
1492         struct journal_head **list = NULL;
1493         transaction_t *transaction;
1494         struct buffer_head *bh = jh2bh(jh);
1495
1496         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1497         transaction = jh->b_transaction;
1498         if (transaction)
1499                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1500
1501         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1502         if (jh->b_jlist != BJ_None)
1503                 J_ASSERT_JH(jh, transaction != 0);
1504
1505         switch (jh->b_jlist) {
1506         case BJ_None:
1507                 return;
1508         case BJ_SyncData:
1509                 list = &transaction->t_sync_datalist;
1510                 break;
1511         case BJ_Metadata:
1512                 transaction->t_nr_buffers--;
1513                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1514                 list = &transaction->t_buffers;
1515                 break;
1516         case BJ_Forget:
1517                 list = &transaction->t_forget;
1518                 break;
1519         case BJ_IO:
1520                 list = &transaction->t_iobuf_list;
1521                 break;
1522         case BJ_Shadow:
1523                 list = &transaction->t_shadow_list;
1524                 break;
1525         case BJ_LogCtl:
1526                 list = &transaction->t_log_list;
1527                 break;
1528         case BJ_Reserved:
1529                 list = &transaction->t_reserved_list;
1530                 break;
1531         case BJ_Locked:
1532                 list = &transaction->t_locked_list;
1533                 break;
1534         }
1535
1536         __blist_del_buffer(list, jh);
1537         jh->b_jlist = BJ_None;
1538         if (test_clear_buffer_jbddirty(bh))
1539                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1540 }
1541
1542 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1543 {
1544         __jbd2_journal_temp_unlink_buffer(jh);
1545         jh->b_transaction = NULL;
1546 }
1547
1548 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1549 {
1550         jbd_lock_bh_state(jh2bh(jh));
1551         spin_lock(&journal->j_list_lock);
1552         __jbd2_journal_unfile_buffer(jh);
1553         spin_unlock(&journal->j_list_lock);
1554         jbd_unlock_bh_state(jh2bh(jh));
1555 }
1556
1557 /*
1558  * Called from jbd2_journal_try_to_free_buffers().
1559  *
1560  * Called under jbd_lock_bh_state(bh)
1561  */
1562 static void
1563 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1564 {
1565         struct journal_head *jh;
1566
1567         jh = bh2jh(bh);
1568
1569         if (buffer_locked(bh) || buffer_dirty(bh))
1570                 goto out;
1571
1572         if (jh->b_next_transaction != 0)
1573                 goto out;
1574
1575         spin_lock(&journal->j_list_lock);
1576         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1577                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1578                         /* A written-back ordered data buffer */
1579                         JBUFFER_TRACE(jh, "release data");
1580                         __jbd2_journal_unfile_buffer(jh);
1581                         jbd2_journal_remove_journal_head(bh);
1582                         __brelse(bh);
1583                 }
1584         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1585                 /* written-back checkpointed metadata buffer */
1586                 if (jh->b_jlist == BJ_None) {
1587                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1588                         __jbd2_journal_remove_checkpoint(jh);
1589                         jbd2_journal_remove_journal_head(bh);
1590                         __brelse(bh);
1591                 }
1592         }
1593         spin_unlock(&journal->j_list_lock);
1594 out:
1595         return;
1596 }
1597
1598
1599 /**
1600  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1601  * @journal: journal for operation
1602  * @page: to try and free
1603  * @unused_gfp_mask: unused
1604  *
1605  *
1606  * For all the buffers on this page,
1607  * if they are fully written out ordered data, move them onto BUF_CLEAN
1608  * so try_to_free_buffers() can reap them.
1609  *
1610  * This function returns non-zero if we wish try_to_free_buffers()
1611  * to be called. We do this if the page is releasable by try_to_free_buffers().
1612  * We also do it if the page has locked or dirty buffers and the caller wants
1613  * us to perform sync or async writeout.
1614  *
1615  * This complicates JBD locking somewhat.  We aren't protected by the
1616  * BKL here.  We wish to remove the buffer from its committing or
1617  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1618  *
1619  * This may *change* the value of transaction_t->t_datalist, so anyone
1620  * who looks at t_datalist needs to lock against this function.
1621  *
1622  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1623  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1624  * will come out of the lock with the buffer dirty, which makes it
1625  * ineligible for release here.
1626  *
1627  * Who else is affected by this?  hmm...  Really the only contender
1628  * is do_get_write_access() - it could be looking at the buffer while
1629  * journal_try_to_free_buffer() is changing its state.  But that
1630  * cannot happen because we never reallocate freed data as metadata
1631  * while the data is part of a transaction.  Yes?
1632  */
1633 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1634                                 struct page *page, gfp_t unused_gfp_mask)
1635 {
1636         struct buffer_head *head;
1637         struct buffer_head *bh;
1638         int ret = 0;
1639
1640         J_ASSERT(PageLocked(page));
1641
1642         head = page_buffers(page);
1643         bh = head;
1644         do {
1645                 struct journal_head *jh;
1646
1647                 /*
1648                  * We take our own ref against the journal_head here to avoid
1649                  * having to add tons of locking around each instance of
1650                  * jbd2_journal_remove_journal_head() and jbd2_journal_put_journal_head().
1651                  */
1652                 jh = jbd2_journal_grab_journal_head(bh);
1653                 if (!jh)
1654                         continue;
1655
1656                 jbd_lock_bh_state(bh);
1657                 __journal_try_to_free_buffer(journal, bh);
1658                 jbd2_journal_put_journal_head(jh);
1659                 jbd_unlock_bh_state(bh);
1660                 if (buffer_jbd(bh))
1661                         goto busy;
1662         } while ((bh = bh->b_this_page) != head);
1663         ret = try_to_free_buffers(page);
1664 busy:
1665         return ret;
1666 }
1667
1668 /*
1669  * This buffer is no longer needed.  If it is on an older transaction's
1670  * checkpoint list we need to record it on this transaction's forget list
1671  * to pin this buffer (and hence its checkpointing transaction) down until
1672  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1673  * release it.
1674  * Returns non-zero if JBD no longer has an interest in the buffer.
1675  *
1676  * Called under j_list_lock.
1677  *
1678  * Called under jbd_lock_bh_state(bh).
1679  */
1680 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1681 {
1682         int may_free = 1;
1683         struct buffer_head *bh = jh2bh(jh);
1684
1685         __jbd2_journal_unfile_buffer(jh);
1686
1687         if (jh->b_cp_transaction) {
1688                 JBUFFER_TRACE(jh, "on running+cp transaction");
1689                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1690                 clear_buffer_jbddirty(bh);
1691                 may_free = 0;
1692         } else {
1693                 JBUFFER_TRACE(jh, "on running transaction");
1694                 jbd2_journal_remove_journal_head(bh);
1695                 __brelse(bh);
1696         }
1697         return may_free;
1698 }
1699
1700 /*
1701  * jbd2_journal_invalidatepage
1702  *
1703  * This code is tricky.  It has a number of cases to deal with.
1704  *
1705  * There are two invariants which this code relies on:
1706  *
1707  * i_size must be updated on disk before we start calling invalidatepage on the
1708  * data.
1709  *
1710  *  This is done in ext3 by defining an ext3_setattr method which
1711  *  updates i_size before truncate gets going.  By maintaining this
1712  *  invariant, we can be sure that it is safe to throw away any buffers
1713  *  attached to the current transaction: once the transaction commits,
1714  *  we know that the data will not be needed.
1715  *
1716  *  Note however that we can *not* throw away data belonging to the
1717  *  previous, committing transaction!
1718  *
1719  * Any disk blocks which *are* part of the previous, committing
1720  * transaction (and which therefore cannot be discarded immediately) are
1721  * not going to be reused in the new running transaction
1722  *
1723  *  The bitmap committed_data images guarantee this: any block which is
1724  *  allocated in one transaction and removed in the next will be marked
1725  *  as in-use in the committed_data bitmap, so cannot be reused until
1726  *  the next transaction to delete the block commits.  This means that
1727  *  leaving committing buffers dirty is quite safe: the disk blocks
1728  *  cannot be reallocated to a different file and so buffer aliasing is
1729  *  not possible.
1730  *
1731  *
1732  * The above applies mainly to ordered data mode.  In writeback mode we
1733  * don't make guarantees about the order in which data hits disk --- in
1734  * particular we don't guarantee that new dirty data is flushed before
1735  * transaction commit --- so it is always safe just to discard data
1736  * immediately in that mode.  --sct
1737  */
1738
1739 /*
1740  * The journal_unmap_buffer helper function returns zero if the buffer
1741  * concerned remains pinned as an anonymous buffer belonging to an older
1742  * transaction.
1743  *
1744  * We're outside-transaction here.  Either or both of j_running_transaction
1745  * and j_committing_transaction may be NULL.
1746  */
1747 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1748 {
1749         transaction_t *transaction;
1750         struct journal_head *jh;
1751         int may_free = 1;
1752         int ret;
1753
1754         BUFFER_TRACE(bh, "entry");
1755
1756         /*
1757          * It is safe to proceed here without the j_list_lock because the
1758          * buffers cannot be stolen by try_to_free_buffers as long as we are
1759          * holding the page lock. --sct
1760          */
1761
1762         if (!buffer_jbd(bh))
1763                 goto zap_buffer_unlocked;
1764
1765         spin_lock(&journal->j_state_lock);
1766         jbd_lock_bh_state(bh);
1767         spin_lock(&journal->j_list_lock);
1768
1769         jh = jbd2_journal_grab_journal_head(bh);
1770         if (!jh)
1771                 goto zap_buffer_no_jh;
1772
1773         transaction = jh->b_transaction;
1774         if (transaction == NULL) {
1775                 /* First case: not on any transaction.  If it
1776                  * has no checkpoint link, then we can zap it:
1777                  * it's a writeback-mode buffer so we don't care
1778                  * if it hits disk safely. */
1779                 if (!jh->b_cp_transaction) {
1780                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1781                         goto zap_buffer;
1782                 }
1783
1784                 if (!buffer_dirty(bh)) {
1785                         /* bdflush has written it.  We can drop it now */
1786                         goto zap_buffer;
1787                 }
1788
1789                 /* OK, it must be in the journal but still not
1790                  * written fully to disk: it's metadata or
1791                  * journaled data... */
1792
1793                 if (journal->j_running_transaction) {
1794                         /* ... and once the current transaction has
1795                          * committed, the buffer won't be needed any
1796                          * longer. */
1797                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1798                         ret = __dispose_buffer(jh,
1799                                         journal->j_running_transaction);
1800                         jbd2_journal_put_journal_head(jh);
1801                         spin_unlock(&journal->j_list_lock);
1802                         jbd_unlock_bh_state(bh);
1803                         spin_unlock(&journal->j_state_lock);
1804                         return ret;
1805                 } else {
1806                         /* There is no currently-running transaction. So the
1807                          * orphan record which we wrote for this file must have
1808                          * passed into commit.  We must attach this buffer to
1809                          * the committing transaction, if it exists. */
1810                         if (journal->j_committing_transaction) {
1811                                 JBUFFER_TRACE(jh, "give to committing trans");
1812                                 ret = __dispose_buffer(jh,
1813                                         journal->j_committing_transaction);
1814                                 jbd2_journal_put_journal_head(jh);
1815                                 spin_unlock(&journal->j_list_lock);
1816                                 jbd_unlock_bh_state(bh);
1817                                 spin_unlock(&journal->j_state_lock);
1818                                 return ret;
1819                         } else {
1820                                 /* The orphan record's transaction has
1821                                  * committed.  We can cleanse this buffer */
1822                                 clear_buffer_jbddirty(bh);
1823                                 goto zap_buffer;
1824                         }
1825                 }
1826         } else if (transaction == journal->j_committing_transaction) {
1827                 if (jh->b_jlist == BJ_Locked) {
1828                         /*
1829                          * The buffer is on the committing transaction's locked
1830                          * list.  We have the buffer locked, so I/O has
1831                          * completed.  So we can nail the buffer now.
1832                          */
1833                         may_free = __dispose_buffer(jh, transaction);
1834                         goto zap_buffer;
1835                 }
1836                 /*
1837                  * If it is committing, we simply cannot touch it.  We
1838                  * can remove it's next_transaction pointer from the
1839                  * running transaction if that is set, but nothing
1840                  * else. */
1841                 JBUFFER_TRACE(jh, "on committing transaction");
1842                 set_buffer_freed(bh);
1843                 if (jh->b_next_transaction) {
1844                         J_ASSERT(jh->b_next_transaction ==
1845                                         journal->j_running_transaction);
1846                         jh->b_next_transaction = NULL;
1847                 }
1848                 jbd2_journal_put_journal_head(jh);
1849                 spin_unlock(&journal->j_list_lock);
1850                 jbd_unlock_bh_state(bh);
1851                 spin_unlock(&journal->j_state_lock);
1852                 return 0;
1853         } else {
1854                 /* Good, the buffer belongs to the running transaction.
1855                  * We are writing our own transaction's data, not any
1856                  * previous one's, so it is safe to throw it away
1857                  * (remember that we expect the filesystem to have set
1858                  * i_size already for this truncate so recovery will not
1859                  * expose the disk blocks we are discarding here.) */
1860                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1861                 may_free = __dispose_buffer(jh, transaction);
1862         }
1863
1864 zap_buffer:
1865         jbd2_journal_put_journal_head(jh);
1866 zap_buffer_no_jh:
1867         spin_unlock(&journal->j_list_lock);
1868         jbd_unlock_bh_state(bh);
1869         spin_unlock(&journal->j_state_lock);
1870 zap_buffer_unlocked:
1871         clear_buffer_dirty(bh);
1872         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1873         clear_buffer_mapped(bh);
1874         clear_buffer_req(bh);
1875         clear_buffer_new(bh);
1876         bh->b_bdev = NULL;
1877         return may_free;
1878 }
1879
1880 /**
1881  * void jbd2_journal_invalidatepage()
1882  * @journal: journal to use for flush...
1883  * @page:    page to flush
1884  * @offset:  length of page to invalidate.
1885  *
1886  * Reap page buffers containing data after offset in page.
1887  *
1888  */
1889 void jbd2_journal_invalidatepage(journal_t *journal,
1890                       struct page *page,
1891                       unsigned long offset)
1892 {
1893         struct buffer_head *head, *bh, *next;
1894         unsigned int curr_off = 0;
1895         int may_free = 1;
1896
1897         if (!PageLocked(page))
1898                 BUG();
1899         if (!page_has_buffers(page))
1900                 return;
1901
1902         /* We will potentially be playing with lists other than just the
1903          * data lists (especially for journaled data mode), so be
1904          * cautious in our locking. */
1905
1906         head = bh = page_buffers(page);
1907         do {
1908                 unsigned int next_off = curr_off + bh->b_size;
1909                 next = bh->b_this_page;
1910
1911                 if (offset <= curr_off) {
1912                         /* This block is wholly outside the truncation point */
1913                         lock_buffer(bh);
1914                         may_free &= journal_unmap_buffer(journal, bh);
1915                         unlock_buffer(bh);
1916                 }
1917                 curr_off = next_off;
1918                 bh = next;
1919
1920         } while (bh != head);
1921
1922         if (!offset) {
1923                 if (may_free && try_to_free_buffers(page))
1924                         J_ASSERT(!page_has_buffers(page));
1925         }
1926 }
1927
1928 /*
1929  * File a buffer on the given transaction list.
1930  */
1931 void __jbd2_journal_file_buffer(struct journal_head *jh,
1932                         transaction_t *transaction, int jlist)
1933 {
1934         struct journal_head **list = NULL;
1935         int was_dirty = 0;
1936         struct buffer_head *bh = jh2bh(jh);
1937
1938         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1939         assert_spin_locked(&transaction->t_journal->j_list_lock);
1940
1941         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1942         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1943                                 jh->b_transaction == 0);
1944
1945         if (jh->b_transaction && jh->b_jlist == jlist)
1946                 return;
1947
1948         /* The following list of buffer states needs to be consistent
1949          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1950          * state. */
1951
1952         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1953             jlist == BJ_Shadow || jlist == BJ_Forget) {
1954                 if (test_clear_buffer_dirty(bh) ||
1955                     test_clear_buffer_jbddirty(bh))
1956                         was_dirty = 1;
1957         }
1958
1959         if (jh->b_transaction)
1960                 __jbd2_journal_temp_unlink_buffer(jh);
1961         jh->b_transaction = transaction;
1962
1963         switch (jlist) {
1964         case BJ_None:
1965                 J_ASSERT_JH(jh, !jh->b_committed_data);
1966                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1967                 return;
1968         case BJ_SyncData:
1969                 list = &transaction->t_sync_datalist;
1970                 break;
1971         case BJ_Metadata:
1972                 transaction->t_nr_buffers++;
1973                 list = &transaction->t_buffers;
1974                 break;
1975         case BJ_Forget:
1976                 list = &transaction->t_forget;
1977                 break;
1978         case BJ_IO:
1979                 list = &transaction->t_iobuf_list;
1980                 break;
1981         case BJ_Shadow:
1982                 list = &transaction->t_shadow_list;
1983                 break;
1984         case BJ_LogCtl:
1985                 list = &transaction->t_log_list;
1986                 break;
1987         case BJ_Reserved:
1988                 list = &transaction->t_reserved_list;
1989                 break;
1990         case BJ_Locked:
1991                 list =  &transaction->t_locked_list;
1992                 break;
1993         }
1994
1995         __blist_add_buffer(list, jh);
1996         jh->b_jlist = jlist;
1997
1998         if (was_dirty)
1999                 set_buffer_jbddirty(bh);
2000 }
2001
2002 void jbd2_journal_file_buffer(struct journal_head *jh,
2003                                 transaction_t *transaction, int jlist)
2004 {
2005         jbd_lock_bh_state(jh2bh(jh));
2006         spin_lock(&transaction->t_journal->j_list_lock);
2007         __jbd2_journal_file_buffer(jh, transaction, jlist);
2008         spin_unlock(&transaction->t_journal->j_list_lock);
2009         jbd_unlock_bh_state(jh2bh(jh));
2010 }
2011
2012 /*
2013  * Remove a buffer from its current buffer list in preparation for
2014  * dropping it from its current transaction entirely.  If the buffer has
2015  * already started to be used by a subsequent transaction, refile the
2016  * buffer on that transaction's metadata list.
2017  *
2018  * Called under journal->j_list_lock
2019  *
2020  * Called under jbd_lock_bh_state(jh2bh(jh))
2021  */
2022 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2023 {
2024         int was_dirty;
2025         struct buffer_head *bh = jh2bh(jh);
2026
2027         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2028         if (jh->b_transaction)
2029                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2030
2031         /* If the buffer is now unused, just drop it. */
2032         if (jh->b_next_transaction == NULL) {
2033                 __jbd2_journal_unfile_buffer(jh);
2034                 return;
2035         }
2036
2037         /*
2038          * It has been modified by a later transaction: add it to the new
2039          * transaction's metadata list.
2040          */
2041
2042         was_dirty = test_clear_buffer_jbddirty(bh);
2043         __jbd2_journal_temp_unlink_buffer(jh);
2044         jh->b_transaction = jh->b_next_transaction;
2045         jh->b_next_transaction = NULL;
2046         __jbd2_journal_file_buffer(jh, jh->b_transaction,
2047                                 was_dirty ? BJ_Metadata : BJ_Reserved);
2048         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2049
2050         if (was_dirty)
2051                 set_buffer_jbddirty(bh);
2052 }
2053
2054 /*
2055  * For the unlocked version of this call, also make sure that any
2056  * hanging journal_head is cleaned up if necessary.
2057  *
2058  * __jbd2_journal_refile_buffer is usually called as part of a single locked
2059  * operation on a buffer_head, in which the caller is probably going to
2060  * be hooking the journal_head onto other lists.  In that case it is up
2061  * to the caller to remove the journal_head if necessary.  For the
2062  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2063  * doing anything else to the buffer so we need to do the cleanup
2064  * ourselves to avoid a jh leak.
2065  *
2066  * *** The journal_head may be freed by this call! ***
2067  */
2068 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2069 {
2070         struct buffer_head *bh = jh2bh(jh);
2071
2072         jbd_lock_bh_state(bh);
2073         spin_lock(&journal->j_list_lock);
2074
2075         __jbd2_journal_refile_buffer(jh);
2076         jbd_unlock_bh_state(bh);
2077         jbd2_journal_remove_journal_head(bh);
2078
2079         spin_unlock(&journal->j_list_lock);
2080         __brelse(bh);
2081 }