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