jbd: ordered data integrity fix
[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 = kzalloc(sizeof(*new_transaction),
100                                                 GFP_NOFS|__GFP_NOFAIL);
101                 if (!new_transaction) {
102                         ret = -ENOMEM;
103                         goto out;
104                 }
105         }
106
107         jbd_debug(3, "New handle %p going live.\n", handle);
108
109 repeat:
110
111         /*
112          * We need to hold j_state_lock until t_updates has been incremented,
113          * for proper journal barrier handling
114          */
115         spin_lock(&journal->j_state_lock);
116 repeat_locked:
117         if (is_journal_aborted(journal) ||
118             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
119                 spin_unlock(&journal->j_state_lock);
120                 ret = -EROFS;
121                 goto out;
122         }
123
124         /* Wait on the journal's transaction barrier if necessary */
125         if (journal->j_barrier_count) {
126                 spin_unlock(&journal->j_state_lock);
127                 wait_event(journal->j_wait_transaction_locked,
128                                 journal->j_barrier_count == 0);
129                 goto repeat;
130         }
131
132         if (!journal->j_running_transaction) {
133                 if (!new_transaction) {
134                         spin_unlock(&journal->j_state_lock);
135                         goto alloc_transaction;
136                 }
137                 get_transaction(journal, new_transaction);
138                 new_transaction = NULL;
139         }
140
141         transaction = journal->j_running_transaction;
142
143         /*
144          * If the current transaction is locked down for commit, wait for the
145          * lock to be released.
146          */
147         if (transaction->t_state == T_LOCKED) {
148                 DEFINE_WAIT(wait);
149
150                 prepare_to_wait(&journal->j_wait_transaction_locked,
151                                         &wait, TASK_UNINTERRUPTIBLE);
152                 spin_unlock(&journal->j_state_lock);
153                 schedule();
154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
155                 goto repeat;
156         }
157
158         /*
159          * If there is not enough space left in the log to write all potential
160          * buffers requested by this operation, we need to stall pending a log
161          * checkpoint to free some more log space.
162          */
163         spin_lock(&transaction->t_handle_lock);
164         needed = transaction->t_outstanding_credits + nblocks;
165
166         if (needed > journal->j_max_transaction_buffers) {
167                 /*
168                  * If the current transaction is already too large, then start
169                  * to commit it: we can then go back and attach this handle to
170                  * a new transaction.
171                  */
172                 DEFINE_WAIT(wait);
173
174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
175                 spin_unlock(&transaction->t_handle_lock);
176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177                                 TASK_UNINTERRUPTIBLE);
178                 __log_start_commit(journal, transaction->t_tid);
179                 spin_unlock(&journal->j_state_lock);
180                 schedule();
181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
182                 goto repeat;
183         }
184
185         /*
186          * The commit code assumes that it can get enough log space
187          * without forcing a checkpoint.  This is *critical* for
188          * correctness: a checkpoint of a buffer which is also
189          * associated with a committing transaction creates a deadlock,
190          * so commit simply cannot force through checkpoints.
191          *
192          * We must therefore ensure the necessary space in the journal
193          * *before* starting to dirty potentially checkpointed buffers
194          * in the new transaction.
195          *
196          * The worst part is, any transaction currently committing can
197          * reduce the free space arbitrarily.  Be careful to account for
198          * those buffers when checkpointing.
199          */
200
201         /*
202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
203          * a _lot_ of headroom: 1/4 of the journal plus the size of
204          * the committing transaction.  Really, we only need to give it
205          * committing_transaction->t_outstanding_credits plus "enough" for
206          * the log control blocks.
207          * Also, this test is inconsitent with the matching one in
208          * journal_extend().
209          */
210         if (__log_space_left(journal) < jbd_space_needed(journal)) {
211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
212                 spin_unlock(&transaction->t_handle_lock);
213                 __log_wait_for_space(journal);
214                 goto repeat_locked;
215         }
216
217         /* OK, account for the buffers that this operation expects to
218          * use and add the handle to the running transaction. */
219
220         handle->h_transaction = transaction;
221         transaction->t_outstanding_credits += nblocks;
222         transaction->t_updates++;
223         transaction->t_handle_count++;
224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
225                   handle, nblocks, transaction->t_outstanding_credits,
226                   __log_space_left(journal));
227         spin_unlock(&transaction->t_handle_lock);
228         spin_unlock(&journal->j_state_lock);
229 out:
230         if (unlikely(new_transaction))          /* It's usually NULL */
231                 kfree(new_transaction);
232         return ret;
233 }
234
235 static struct lock_class_key jbd_handle_key;
236
237 /* Allocate a new handle.  This should probably be in a slab... */
238 static handle_t *new_handle(int nblocks)
239 {
240         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
241         if (!handle)
242                 return NULL;
243         memset(handle, 0, sizeof(*handle));
244         handle->h_buffer_credits = nblocks;
245         handle->h_ref = 1;
246
247         lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
248
249         return handle;
250 }
251
252 /**
253  * handle_t *journal_start() - Obtain a new handle.
254  * @journal: Journal to start transaction on.
255  * @nblocks: number of block buffer we might modify
256  *
257  * We make sure that the transaction can guarantee at least nblocks of
258  * modified buffers in the log.  We block until the log can guarantee
259  * that much space.
260  *
261  * This function is visible to journal users (like ext3fs), so is not
262  * called with the journal already locked.
263  *
264  * Return a pointer to a newly allocated handle, or NULL on failure
265  */
266 handle_t *journal_start(journal_t *journal, int nblocks)
267 {
268         handle_t *handle = journal_current_handle();
269         int err;
270
271         if (!journal)
272                 return ERR_PTR(-EROFS);
273
274         if (handle) {
275                 J_ASSERT(handle->h_transaction->t_journal == journal);
276                 handle->h_ref++;
277                 return handle;
278         }
279
280         handle = new_handle(nblocks);
281         if (!handle)
282                 return ERR_PTR(-ENOMEM);
283
284         current->journal_info = handle;
285
286         err = start_this_handle(journal, handle);
287         if (err < 0) {
288                 jbd_free_handle(handle);
289                 current->journal_info = NULL;
290                 handle = ERR_PTR(err);
291                 goto out;
292         }
293
294         lock_map_acquire(&handle->h_lockdep_map);
295
296 out:
297         return handle;
298 }
299
300 /**
301  * int journal_extend() - extend buffer credits.
302  * @handle:  handle to 'extend'
303  * @nblocks: nr blocks to try to extend by.
304  *
305  * Some transactions, such as large extends and truncates, can be done
306  * atomically all at once or in several stages.  The operation requests
307  * a credit for a number of buffer modications in advance, but can
308  * extend its credit if it needs more.
309  *
310  * journal_extend tries to give the running handle more buffer credits.
311  * It does not guarantee that allocation - this is a best-effort only.
312  * The calling process MUST be able to deal cleanly with a failure to
313  * extend here.
314  *
315  * Return 0 on success, non-zero on failure.
316  *
317  * return code < 0 implies an error
318  * return code > 0 implies normal transaction-full status.
319  */
320 int journal_extend(handle_t *handle, int nblocks)
321 {
322         transaction_t *transaction = handle->h_transaction;
323         journal_t *journal = transaction->t_journal;
324         int result;
325         int wanted;
326
327         result = -EIO;
328         if (is_handle_aborted(handle))
329                 goto out;
330
331         result = 1;
332
333         spin_lock(&journal->j_state_lock);
334
335         /* Don't extend a locked-down transaction! */
336         if (handle->h_transaction->t_state != T_RUNNING) {
337                 jbd_debug(3, "denied handle %p %d blocks: "
338                           "transaction not running\n", handle, nblocks);
339                 goto error_out;
340         }
341
342         spin_lock(&transaction->t_handle_lock);
343         wanted = transaction->t_outstanding_credits + nblocks;
344
345         if (wanted > journal->j_max_transaction_buffers) {
346                 jbd_debug(3, "denied handle %p %d blocks: "
347                           "transaction too large\n", handle, nblocks);
348                 goto unlock;
349         }
350
351         if (wanted > __log_space_left(journal)) {
352                 jbd_debug(3, "denied handle %p %d blocks: "
353                           "insufficient log space\n", handle, nblocks);
354                 goto unlock;
355         }
356
357         handle->h_buffer_credits += nblocks;
358         transaction->t_outstanding_credits += nblocks;
359         result = 0;
360
361         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
362 unlock:
363         spin_unlock(&transaction->t_handle_lock);
364 error_out:
365         spin_unlock(&journal->j_state_lock);
366 out:
367         return result;
368 }
369
370
371 /**
372  * int journal_restart() - restart a handle.
373  * @handle:  handle to restart
374  * @nblocks: nr credits requested
375  *
376  * Restart a handle for a multi-transaction filesystem
377  * operation.
378  *
379  * If the journal_extend() call above fails to grant new buffer credits
380  * to a running handle, a call to journal_restart will commit the
381  * handle's transaction so far and reattach the handle to a new
382  * transaction capabable of guaranteeing the requested number of
383  * credits.
384  */
385
386 int journal_restart(handle_t *handle, int nblocks)
387 {
388         transaction_t *transaction = handle->h_transaction;
389         journal_t *journal = transaction->t_journal;
390         int ret;
391
392         /* If we've had an abort of any type, don't even think about
393          * actually doing the restart! */
394         if (is_handle_aborted(handle))
395                 return 0;
396
397         /*
398          * First unlink the handle from its current transaction, and start the
399          * commit on that.
400          */
401         J_ASSERT(transaction->t_updates > 0);
402         J_ASSERT(journal_current_handle() == handle);
403
404         spin_lock(&journal->j_state_lock);
405         spin_lock(&transaction->t_handle_lock);
406         transaction->t_outstanding_credits -= handle->h_buffer_credits;
407         transaction->t_updates--;
408
409         if (!transaction->t_updates)
410                 wake_up(&journal->j_wait_updates);
411         spin_unlock(&transaction->t_handle_lock);
412
413         jbd_debug(2, "restarting handle %p\n", handle);
414         __log_start_commit(journal, transaction->t_tid);
415         spin_unlock(&journal->j_state_lock);
416
417         handle->h_buffer_credits = nblocks;
418         ret = start_this_handle(journal, handle);
419         return ret;
420 }
421
422
423 /**
424  * void journal_lock_updates () - establish a transaction barrier.
425  * @journal:  Journal to establish a barrier on.
426  *
427  * This locks out any further updates from being started, and blocks
428  * until all existing updates have completed, returning only once the
429  * journal is in a quiescent state with no updates running.
430  *
431  * The journal lock should not be held on entry.
432  */
433 void journal_lock_updates(journal_t *journal)
434 {
435         DEFINE_WAIT(wait);
436
437         spin_lock(&journal->j_state_lock);
438         ++journal->j_barrier_count;
439
440         /* Wait until there are no running updates */
441         while (1) {
442                 transaction_t *transaction = journal->j_running_transaction;
443
444                 if (!transaction)
445                         break;
446
447                 spin_lock(&transaction->t_handle_lock);
448                 if (!transaction->t_updates) {
449                         spin_unlock(&transaction->t_handle_lock);
450                         break;
451                 }
452                 prepare_to_wait(&journal->j_wait_updates, &wait,
453                                 TASK_UNINTERRUPTIBLE);
454                 spin_unlock(&transaction->t_handle_lock);
455                 spin_unlock(&journal->j_state_lock);
456                 schedule();
457                 finish_wait(&journal->j_wait_updates, &wait);
458                 spin_lock(&journal->j_state_lock);
459         }
460         spin_unlock(&journal->j_state_lock);
461
462         /*
463          * We have now established a barrier against other normal updates, but
464          * we also need to barrier against other journal_lock_updates() calls
465          * to make sure that we serialise special journal-locked operations
466          * too.
467          */
468         mutex_lock(&journal->j_barrier);
469 }
470
471 /**
472  * void journal_unlock_updates (journal_t* journal) - release barrier
473  * @journal:  Journal to release the barrier on.
474  *
475  * Release a transaction barrier obtained with journal_lock_updates().
476  *
477  * Should be called without the journal lock held.
478  */
479 void journal_unlock_updates (journal_t *journal)
480 {
481         J_ASSERT(journal->j_barrier_count != 0);
482
483         mutex_unlock(&journal->j_barrier);
484         spin_lock(&journal->j_state_lock);
485         --journal->j_barrier_count;
486         spin_unlock(&journal->j_state_lock);
487         wake_up(&journal->j_wait_transaction_locked);
488 }
489
490 /*
491  * Report any unexpected dirty buffers which turn up.  Normally those
492  * indicate an error, but they can occur if the user is running (say)
493  * tune2fs to modify the live filesystem, so we need the option of
494  * continuing as gracefully as possible.  #
495  *
496  * The caller should already hold the journal lock and
497  * j_list_lock spinlock: most callers will need those anyway
498  * in order to probe the buffer's journaling state safely.
499  */
500 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
501 {
502         int jlist;
503
504         /* If this buffer is one which might reasonably be dirty
505          * --- ie. data, or not part of this journal --- then
506          * we're OK to leave it alone, but otherwise we need to
507          * move the dirty bit to the journal's own internal
508          * JBDDirty bit. */
509         jlist = jh->b_jlist;
510
511         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
512             jlist == BJ_Shadow || jlist == BJ_Forget) {
513                 struct buffer_head *bh = jh2bh(jh);
514
515                 if (test_clear_buffer_dirty(bh))
516                         set_buffer_jbddirty(bh);
517         }
518 }
519
520 /*
521  * If the buffer is already part of the current transaction, then there
522  * is nothing we need to do.  If it is already part of a prior
523  * transaction which we are still committing to disk, then we need to
524  * make sure that we do not overwrite the old copy: we do copy-out to
525  * preserve the copy going to disk.  We also account the buffer against
526  * the handle's metadata buffer credits (unless the buffer is already
527  * part of the transaction, that is).
528  *
529  */
530 static int
531 do_get_write_access(handle_t *handle, struct journal_head *jh,
532                         int force_copy)
533 {
534         struct buffer_head *bh;
535         transaction_t *transaction;
536         journal_t *journal;
537         int error;
538         char *frozen_buffer = NULL;
539         int need_copy = 0;
540
541         if (is_handle_aborted(handle))
542                 return -EROFS;
543
544         transaction = handle->h_transaction;
545         journal = transaction->t_journal;
546
547         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
548
549         JBUFFER_TRACE(jh, "entry");
550 repeat:
551         bh = jh2bh(jh);
552
553         /* @@@ Need to check for errors here at some point. */
554
555         lock_buffer(bh);
556         jbd_lock_bh_state(bh);
557
558         /* We now hold the buffer lock so it is safe to query the buffer
559          * state.  Is the buffer dirty?
560          *
561          * If so, there are two possibilities.  The buffer may be
562          * non-journaled, and undergoing a quite legitimate writeback.
563          * Otherwise, it is journaled, and we don't expect dirty buffers
564          * in that state (the buffers should be marked JBD_Dirty
565          * instead.)  So either the IO is being done under our own
566          * control and this is a bug, or it's a third party IO such as
567          * dump(8) (which may leave the buffer scheduled for read ---
568          * ie. locked but not dirty) or tune2fs (which may actually have
569          * the buffer dirtied, ugh.)  */
570
571         if (buffer_dirty(bh)) {
572                 /*
573                  * First question: is this buffer already part of the current
574                  * transaction or the existing committing transaction?
575                  */
576                 if (jh->b_transaction) {
577                         J_ASSERT_JH(jh,
578                                 jh->b_transaction == transaction ||
579                                 jh->b_transaction ==
580                                         journal->j_committing_transaction);
581                         if (jh->b_next_transaction)
582                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
583                                                         transaction);
584                 }
585                 /*
586                  * In any case we need to clean the dirty flag and we must
587                  * do it under the buffer lock to be sure we don't race
588                  * with running write-out.
589                  */
590                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
591                 jbd_unexpected_dirty_buffer(jh);
592         }
593
594         unlock_buffer(bh);
595
596         error = -EROFS;
597         if (is_handle_aborted(handle)) {
598                 jbd_unlock_bh_state(bh);
599                 goto out;
600         }
601         error = 0;
602
603         /*
604          * The buffer is already part of this transaction if b_transaction or
605          * b_next_transaction points to it
606          */
607         if (jh->b_transaction == transaction ||
608             jh->b_next_transaction == transaction)
609                 goto done;
610
611         /*
612          * this is the first time this transaction is touching this buffer,
613          * reset the modified flag
614          */
615         jh->b_modified = 0;
616
617         /*
618          * If there is already a copy-out version of this buffer, then we don't
619          * need to make another one
620          */
621         if (jh->b_frozen_data) {
622                 JBUFFER_TRACE(jh, "has frozen data");
623                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
624                 jh->b_next_transaction = transaction;
625                 goto done;
626         }
627
628         /* Is there data here we need to preserve? */
629
630         if (jh->b_transaction && jh->b_transaction != transaction) {
631                 JBUFFER_TRACE(jh, "owned by older transaction");
632                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
633                 J_ASSERT_JH(jh, jh->b_transaction ==
634                                         journal->j_committing_transaction);
635
636                 /* There is one case we have to be very careful about.
637                  * If the committing transaction is currently writing
638                  * this buffer out to disk and has NOT made a copy-out,
639                  * then we cannot modify the buffer contents at all
640                  * right now.  The essence of copy-out is that it is the
641                  * extra copy, not the primary copy, which gets
642                  * journaled.  If the primary copy is already going to
643                  * disk then we cannot do copy-out here. */
644
645                 if (jh->b_jlist == BJ_Shadow) {
646                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
647                         wait_queue_head_t *wqh;
648
649                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
650
651                         JBUFFER_TRACE(jh, "on shadow: sleep");
652                         jbd_unlock_bh_state(bh);
653                         /* commit wakes up all shadow buffers after IO */
654                         for ( ; ; ) {
655                                 prepare_to_wait(wqh, &wait.wait,
656                                                 TASK_UNINTERRUPTIBLE);
657                                 if (jh->b_jlist != BJ_Shadow)
658                                         break;
659                                 schedule();
660                         }
661                         finish_wait(wqh, &wait.wait);
662                         goto repeat;
663                 }
664
665                 /* Only do the copy if the currently-owning transaction
666                  * still needs it.  If it is on the Forget list, the
667                  * committing transaction is past that stage.  The
668                  * buffer had better remain locked during the kmalloc,
669                  * but that should be true --- we hold the journal lock
670                  * still and the buffer is already on the BUF_JOURNAL
671                  * list so won't be flushed.
672                  *
673                  * Subtle point, though: if this is a get_undo_access,
674                  * then we will be relying on the frozen_data to contain
675                  * the new value of the committed_data record after the
676                  * transaction, so we HAVE to force the frozen_data copy
677                  * in that case. */
678
679                 if (jh->b_jlist != BJ_Forget || force_copy) {
680                         JBUFFER_TRACE(jh, "generate frozen data");
681                         if (!frozen_buffer) {
682                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
683                                 jbd_unlock_bh_state(bh);
684                                 frozen_buffer =
685                                         jbd_alloc(jh2bh(jh)->b_size,
686                                                          GFP_NOFS);
687                                 if (!frozen_buffer) {
688                                         printk(KERN_EMERG
689                                                "%s: OOM for frozen_buffer\n",
690                                                __func__);
691                                         JBUFFER_TRACE(jh, "oom!");
692                                         error = -ENOMEM;
693                                         jbd_lock_bh_state(bh);
694                                         goto done;
695                                 }
696                                 goto repeat;
697                         }
698                         jh->b_frozen_data = frozen_buffer;
699                         frozen_buffer = NULL;
700                         need_copy = 1;
701                 }
702                 jh->b_next_transaction = transaction;
703         }
704
705
706         /*
707          * Finally, if the buffer is not journaled right now, we need to make
708          * sure it doesn't get written to disk before the caller actually
709          * commits the new data
710          */
711         if (!jh->b_transaction) {
712                 JBUFFER_TRACE(jh, "no transaction");
713                 J_ASSERT_JH(jh, !jh->b_next_transaction);
714                 jh->b_transaction = transaction;
715                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
716                 spin_lock(&journal->j_list_lock);
717                 __journal_file_buffer(jh, transaction, BJ_Reserved);
718                 spin_unlock(&journal->j_list_lock);
719         }
720
721 done:
722         if (need_copy) {
723                 struct page *page;
724                 int offset;
725                 char *source;
726
727                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
728                             "Possible IO failure.\n");
729                 page = jh2bh(jh)->b_page;
730                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
731                 source = kmap_atomic(page, KM_USER0);
732                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
733                 kunmap_atomic(source, KM_USER0);
734         }
735         jbd_unlock_bh_state(bh);
736
737         /*
738          * If we are about to journal a buffer, then any revoke pending on it is
739          * no longer valid
740          */
741         journal_cancel_revoke(handle, jh);
742
743 out:
744         if (unlikely(frozen_buffer))    /* It's usually NULL */
745                 jbd_free(frozen_buffer, bh->b_size);
746
747         JBUFFER_TRACE(jh, "exit");
748         return error;
749 }
750
751 /**
752  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
753  * @handle: transaction to add buffer modifications to
754  * @bh:     bh to be used for metadata writes
755  * @credits: variable that will receive credits for the buffer
756  *
757  * Returns an error code or 0 on success.
758  *
759  * In full data journalling mode the buffer may be of type BJ_AsyncData,
760  * because we're write()ing a buffer which is also part of a shared mapping.
761  */
762
763 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
764 {
765         struct journal_head *jh = journal_add_journal_head(bh);
766         int rc;
767
768         /* We do not want to get caught playing with fields which the
769          * log thread also manipulates.  Make sure that the buffer
770          * completes any outstanding IO before proceeding. */
771         rc = do_get_write_access(handle, jh, 0);
772         journal_put_journal_head(jh);
773         return rc;
774 }
775
776
777 /*
778  * When the user wants to journal a newly created buffer_head
779  * (ie. getblk() returned a new buffer and we are going to populate it
780  * manually rather than reading off disk), then we need to keep the
781  * buffer_head locked until it has been completely filled with new
782  * data.  In this case, we should be able to make the assertion that
783  * the bh is not already part of an existing transaction.
784  *
785  * The buffer should already be locked by the caller by this point.
786  * There is no lock ranking violation: it was a newly created,
787  * unlocked buffer beforehand. */
788
789 /**
790  * int journal_get_create_access () - notify intent to use newly created bh
791  * @handle: transaction to new buffer to
792  * @bh: new buffer.
793  *
794  * Call this if you create a new bh.
795  */
796 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
797 {
798         transaction_t *transaction = handle->h_transaction;
799         journal_t *journal = transaction->t_journal;
800         struct journal_head *jh = journal_add_journal_head(bh);
801         int err;
802
803         jbd_debug(5, "journal_head %p\n", jh);
804         err = -EROFS;
805         if (is_handle_aborted(handle))
806                 goto out;
807         err = 0;
808
809         JBUFFER_TRACE(jh, "entry");
810         /*
811          * The buffer may already belong to this transaction due to pre-zeroing
812          * in the filesystem's new_block code.  It may also be on the previous,
813          * committing transaction's lists, but it HAS to be in Forget state in
814          * that case: the transaction must have deleted the buffer for it to be
815          * reused here.
816          */
817         jbd_lock_bh_state(bh);
818         spin_lock(&journal->j_list_lock);
819         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
820                 jh->b_transaction == NULL ||
821                 (jh->b_transaction == journal->j_committing_transaction &&
822                           jh->b_jlist == BJ_Forget)));
823
824         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
825         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
826
827         if (jh->b_transaction == NULL) {
828                 jh->b_transaction = transaction;
829
830                 /* first access by this transaction */
831                 jh->b_modified = 0;
832
833                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
834                 __journal_file_buffer(jh, transaction, BJ_Reserved);
835         } else if (jh->b_transaction == journal->j_committing_transaction) {
836                 /* first access by this transaction */
837                 jh->b_modified = 0;
838
839                 JBUFFER_TRACE(jh, "set next transaction");
840                 jh->b_next_transaction = transaction;
841         }
842         spin_unlock(&journal->j_list_lock);
843         jbd_unlock_bh_state(bh);
844
845         /*
846          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
847          * blocks which contain freed but then revoked metadata.  We need
848          * to cancel the revoke in case we end up freeing it yet again
849          * and the reallocating as data - this would cause a second revoke,
850          * which hits an assertion error.
851          */
852         JBUFFER_TRACE(jh, "cancelling revoke");
853         journal_cancel_revoke(handle, jh);
854         journal_put_journal_head(jh);
855 out:
856         return err;
857 }
858
859 /**
860  * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
861  * @handle: transaction
862  * @bh: buffer to undo
863  * @credits: store the number of taken credits here (if not NULL)
864  *
865  * Sometimes there is a need to distinguish between metadata which has
866  * been committed to disk and that which has not.  The ext3fs code uses
867  * this for freeing and allocating space, we have to make sure that we
868  * do not reuse freed space until the deallocation has been committed,
869  * since if we overwrote that space we would make the delete
870  * un-rewindable in case of a crash.
871  *
872  * To deal with that, journal_get_undo_access requests write access to a
873  * buffer for parts of non-rewindable operations such as delete
874  * operations on the bitmaps.  The journaling code must keep a copy of
875  * the buffer's contents prior to the undo_access call until such time
876  * as we know that the buffer has definitely been committed to disk.
877  *
878  * We never need to know which transaction the committed data is part
879  * of, buffers touched here are guaranteed to be dirtied later and so
880  * will be committed to a new transaction in due course, at which point
881  * we can discard the old committed data pointer.
882  *
883  * Returns error number or 0 on success.
884  */
885 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
886 {
887         int err;
888         struct journal_head *jh = journal_add_journal_head(bh);
889         char *committed_data = NULL;
890
891         JBUFFER_TRACE(jh, "entry");
892
893         /*
894          * Do this first --- it can drop the journal lock, so we want to
895          * make sure that obtaining the committed_data is done
896          * atomically wrt. completion of any outstanding commits.
897          */
898         err = do_get_write_access(handle, jh, 1);
899         if (err)
900                 goto out;
901
902 repeat:
903         if (!jh->b_committed_data) {
904                 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
905                 if (!committed_data) {
906                         printk(KERN_EMERG "%s: No memory for committed data\n",
907                                 __func__);
908                         err = -ENOMEM;
909                         goto out;
910                 }
911         }
912
913         jbd_lock_bh_state(bh);
914         if (!jh->b_committed_data) {
915                 /* Copy out the current buffer contents into the
916                  * preserved, committed copy. */
917                 JBUFFER_TRACE(jh, "generate b_committed data");
918                 if (!committed_data) {
919                         jbd_unlock_bh_state(bh);
920                         goto repeat;
921                 }
922
923                 jh->b_committed_data = committed_data;
924                 committed_data = NULL;
925                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
926         }
927         jbd_unlock_bh_state(bh);
928 out:
929         journal_put_journal_head(jh);
930         if (unlikely(committed_data))
931                 jbd_free(committed_data, bh->b_size);
932         return err;
933 }
934
935 /**
936  * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
937  * @handle: transaction
938  * @bh: bufferhead to mark
939  *
940  * Description:
941  * Mark a buffer as containing dirty data which needs to be flushed before
942  * we can commit the current transaction.
943  *
944  * The buffer is placed on the transaction's data list and is marked as
945  * belonging to the transaction.
946  *
947  * Returns error number or 0 on success.
948  *
949  * journal_dirty_data() can be called via page_launder->ext3_writepage
950  * by kswapd.
951  */
952 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
953 {
954         journal_t *journal = handle->h_transaction->t_journal;
955         int need_brelse = 0;
956         struct journal_head *jh;
957         int ret = 0;
958
959         if (is_handle_aborted(handle))
960                 return ret;
961
962         jh = journal_add_journal_head(bh);
963         JBUFFER_TRACE(jh, "entry");
964
965         /*
966          * The buffer could *already* be dirty.  Writeout can start
967          * at any time.
968          */
969         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
970
971         /*
972          * What if the buffer is already part of a running transaction?
973          *
974          * There are two cases:
975          * 1) It is part of the current running transaction.  Refile it,
976          *    just in case we have allocated it as metadata, deallocated
977          *    it, then reallocated it as data.
978          * 2) It is part of the previous, still-committing transaction.
979          *    If all we want to do is to guarantee that the buffer will be
980          *    written to disk before this new transaction commits, then
981          *    being sure that the *previous* transaction has this same
982          *    property is sufficient for us!  Just leave it on its old
983          *    transaction.
984          *
985          * In case (2), the buffer must not already exist as metadata
986          * --- that would violate write ordering (a transaction is free
987          * to write its data at any point, even before the previous
988          * committing transaction has committed).  The caller must
989          * never, ever allow this to happen: there's nothing we can do
990          * about it in this layer.
991          */
992         jbd_lock_bh_state(bh);
993         spin_lock(&journal->j_list_lock);
994
995         /* Now that we have bh_state locked, are we really still mapped? */
996         if (!buffer_mapped(bh)) {
997                 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
998                 goto no_journal;
999         }
1000
1001         if (jh->b_transaction) {
1002                 JBUFFER_TRACE(jh, "has transaction");
1003                 if (jh->b_transaction != handle->h_transaction) {
1004                         JBUFFER_TRACE(jh, "belongs to older transaction");
1005                         J_ASSERT_JH(jh, jh->b_transaction ==
1006                                         journal->j_committing_transaction);
1007
1008                         /* @@@ IS THIS TRUE  ? */
1009                         /*
1010                          * Not any more.  Scenario: someone does a write()
1011                          * in data=journal mode.  The buffer's transaction has
1012                          * moved into commit.  Then someone does another
1013                          * write() to the file.  We do the frozen data copyout
1014                          * and set b_next_transaction to point to j_running_t.
1015                          * And while we're in that state, someone does a
1016                          * writepage() in an attempt to pageout the same area
1017                          * of the file via a shared mapping.  At present that
1018                          * calls journal_dirty_data(), and we get right here.
1019                          * It may be too late to journal the data.  Simply
1020                          * falling through to the next test will suffice: the
1021                          * data will be dirty and wil be checkpointed.  The
1022                          * ordering comments in the next comment block still
1023                          * apply.
1024                          */
1025                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1026
1027                         /*
1028                          * If we're journalling data, and this buffer was
1029                          * subject to a write(), it could be metadata, forget
1030                          * or shadow against the committing transaction.  Now,
1031                          * someone has dirtied the same darn page via a mapping
1032                          * and it is being writepage()'d.
1033                          * We *could* just steal the page from commit, with some
1034                          * fancy locking there.  Instead, we just skip it -
1035                          * don't tie the page's buffers to the new transaction
1036                          * at all.
1037                          * Implication: if we crash before the writepage() data
1038                          * is written into the filesystem, recovery will replay
1039                          * the write() data.
1040                          */
1041                         if (jh->b_jlist != BJ_None &&
1042                                         jh->b_jlist != BJ_SyncData &&
1043                                         jh->b_jlist != BJ_Locked) {
1044                                 JBUFFER_TRACE(jh, "Not stealing");
1045                                 goto no_journal;
1046                         }
1047
1048                         /*
1049                          * This buffer may be undergoing writeout in commit.  We
1050                          * can't return from here and let the caller dirty it
1051                          * again because that can cause the write-out loop in
1052                          * commit to never terminate.
1053                          */
1054                         if (buffer_dirty(bh)) {
1055                                 get_bh(bh);
1056                                 spin_unlock(&journal->j_list_lock);
1057                                 jbd_unlock_bh_state(bh);
1058                                 need_brelse = 1;
1059                                 sync_dirty_buffer(bh);
1060                                 jbd_lock_bh_state(bh);
1061                                 spin_lock(&journal->j_list_lock);
1062                                 /* Since we dropped the lock... */
1063                                 if (!buffer_mapped(bh)) {
1064                                         JBUFFER_TRACE(jh, "buffer got unmapped");
1065                                         goto no_journal;
1066                                 }
1067                                 /* The buffer may become locked again at any
1068                                    time if it is redirtied */
1069                         }
1070
1071                         /*
1072                          * We cannot remove the buffer with io error from the
1073                          * committing transaction, because otherwise it would
1074                          * miss the error and the commit would not abort.
1075                          */
1076                         if (unlikely(!buffer_uptodate(bh))) {
1077                                 ret = -EIO;
1078                                 goto no_journal;
1079                         }
1080
1081                         if (jh->b_transaction != NULL) {
1082                                 JBUFFER_TRACE(jh, "unfile from commit");
1083                                 __journal_temp_unlink_buffer(jh);
1084                                 /* It still points to the committing
1085                                  * transaction; move it to this one so
1086                                  * that the refile assert checks are
1087                                  * happy. */
1088                                 jh->b_transaction = handle->h_transaction;
1089                         }
1090                         /* The buffer will be refiled below */
1091
1092                 }
1093                 /*
1094                  * Special case --- the buffer might actually have been
1095                  * allocated and then immediately deallocated in the previous,
1096                  * committing transaction, so might still be left on that
1097                  * transaction's metadata lists.
1098                  */
1099                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1100                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1101                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1102                         __journal_temp_unlink_buffer(jh);
1103                         jh->b_transaction = handle->h_transaction;
1104                         JBUFFER_TRACE(jh, "file as data");
1105                         __journal_file_buffer(jh, handle->h_transaction,
1106                                                 BJ_SyncData);
1107                 }
1108         } else {
1109                 JBUFFER_TRACE(jh, "not on a transaction");
1110                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1111         }
1112 no_journal:
1113         spin_unlock(&journal->j_list_lock);
1114         jbd_unlock_bh_state(bh);
1115         if (need_brelse) {
1116                 BUFFER_TRACE(bh, "brelse");
1117                 __brelse(bh);
1118         }
1119         JBUFFER_TRACE(jh, "exit");
1120         journal_put_journal_head(jh);
1121         return ret;
1122 }
1123
1124 /**
1125  * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1126  * @handle: transaction to add buffer to.
1127  * @bh: buffer to mark
1128  *
1129  * Mark dirty metadata which needs to be journaled as part of the current
1130  * transaction.
1131  *
1132  * The buffer is placed on the transaction's metadata list and is marked
1133  * as belonging to the transaction.
1134  *
1135  * Returns error number or 0 on success.
1136  *
1137  * Special care needs to be taken if the buffer already belongs to the
1138  * current committing transaction (in which case we should have frozen
1139  * data present for that commit).  In that case, we don't relink the
1140  * buffer: that only gets done when the old transaction finally
1141  * completes its commit.
1142  */
1143 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1144 {
1145         transaction_t *transaction = handle->h_transaction;
1146         journal_t *journal = transaction->t_journal;
1147         struct journal_head *jh = bh2jh(bh);
1148
1149         jbd_debug(5, "journal_head %p\n", jh);
1150         JBUFFER_TRACE(jh, "entry");
1151         if (is_handle_aborted(handle))
1152                 goto out;
1153
1154         jbd_lock_bh_state(bh);
1155
1156         if (jh->b_modified == 0) {
1157                 /*
1158                  * This buffer's got modified and becoming part
1159                  * of the transaction. This needs to be done
1160                  * once a transaction -bzzz
1161                  */
1162                 jh->b_modified = 1;
1163                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1164                 handle->h_buffer_credits--;
1165         }
1166
1167         /*
1168          * fastpath, to avoid expensive locking.  If this buffer is already
1169          * on the running transaction's metadata list there is nothing to do.
1170          * Nobody can take it off again because there is a handle open.
1171          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1172          * result in this test being false, so we go in and take the locks.
1173          */
1174         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1175                 JBUFFER_TRACE(jh, "fastpath");
1176                 J_ASSERT_JH(jh, jh->b_transaction ==
1177                                         journal->j_running_transaction);
1178                 goto out_unlock_bh;
1179         }
1180
1181         set_buffer_jbddirty(bh);
1182
1183         /*
1184          * Metadata already on the current transaction list doesn't
1185          * need to be filed.  Metadata on another transaction's list must
1186          * be committing, and will be refiled once the commit completes:
1187          * leave it alone for now.
1188          */
1189         if (jh->b_transaction != transaction) {
1190                 JBUFFER_TRACE(jh, "already on other transaction");
1191                 J_ASSERT_JH(jh, jh->b_transaction ==
1192                                         journal->j_committing_transaction);
1193                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1194                 /* And this case is illegal: we can't reuse another
1195                  * transaction's data buffer, ever. */
1196                 goto out_unlock_bh;
1197         }
1198
1199         /* That test should have eliminated the following case: */
1200         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1201
1202         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1203         spin_lock(&journal->j_list_lock);
1204         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1205         spin_unlock(&journal->j_list_lock);
1206 out_unlock_bh:
1207         jbd_unlock_bh_state(bh);
1208 out:
1209         JBUFFER_TRACE(jh, "exit");
1210         return 0;
1211 }
1212
1213 /*
1214  * journal_release_buffer: undo a get_write_access without any buffer
1215  * updates, if the update decided in the end that it didn't need access.
1216  *
1217  */
1218 void
1219 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1220 {
1221         BUFFER_TRACE(bh, "entry");
1222 }
1223
1224 /**
1225  * void journal_forget() - bforget() for potentially-journaled buffers.
1226  * @handle: transaction handle
1227  * @bh:     bh to 'forget'
1228  *
1229  * We can only do the bforget if there are no commits pending against the
1230  * buffer.  If the buffer is dirty in the current running transaction we
1231  * can safely unlink it.
1232  *
1233  * bh may not be a journalled buffer at all - it may be a non-JBD
1234  * buffer which came off the hashtable.  Check for this.
1235  *
1236  * Decrements bh->b_count by one.
1237  *
1238  * Allow this call even if the handle has aborted --- it may be part of
1239  * the caller's cleanup after an abort.
1240  */
1241 int journal_forget (handle_t *handle, struct buffer_head *bh)
1242 {
1243         transaction_t *transaction = handle->h_transaction;
1244         journal_t *journal = transaction->t_journal;
1245         struct journal_head *jh;
1246         int drop_reserve = 0;
1247         int err = 0;
1248         int was_modified = 0;
1249
1250         BUFFER_TRACE(bh, "entry");
1251
1252         jbd_lock_bh_state(bh);
1253         spin_lock(&journal->j_list_lock);
1254
1255         if (!buffer_jbd(bh))
1256                 goto not_jbd;
1257         jh = bh2jh(bh);
1258
1259         /* Critical error: attempting to delete a bitmap buffer, maybe?
1260          * Don't do any jbd operations, and return an error. */
1261         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1262                          "inconsistent data on disk")) {
1263                 err = -EIO;
1264                 goto not_jbd;
1265         }
1266
1267         /* keep track of wether or not this transaction modified us */
1268         was_modified = jh->b_modified;
1269
1270         /*
1271          * The buffer's going from the transaction, we must drop
1272          * all references -bzzz
1273          */
1274         jh->b_modified = 0;
1275
1276         if (jh->b_transaction == handle->h_transaction) {
1277                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1278
1279                 /* If we are forgetting a buffer which is already part
1280                  * of this transaction, then we can just drop it from
1281                  * the transaction immediately. */
1282                 clear_buffer_dirty(bh);
1283                 clear_buffer_jbddirty(bh);
1284
1285                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1286
1287                 /*
1288                  * we only want to drop a reference if this transaction
1289                  * modified the buffer
1290                  */
1291                 if (was_modified)
1292                         drop_reserve = 1;
1293
1294                 /*
1295                  * We are no longer going to journal this buffer.
1296                  * However, the commit of this transaction is still
1297                  * important to the buffer: the delete that we are now
1298                  * processing might obsolete an old log entry, so by
1299                  * committing, we can satisfy the buffer's checkpoint.
1300                  *
1301                  * So, if we have a checkpoint on the buffer, we should
1302                  * now refile the buffer on our BJ_Forget list so that
1303                  * we know to remove the checkpoint after we commit.
1304                  */
1305
1306                 if (jh->b_cp_transaction) {
1307                         __journal_temp_unlink_buffer(jh);
1308                         __journal_file_buffer(jh, transaction, BJ_Forget);
1309                 } else {
1310                         __journal_unfile_buffer(jh);
1311                         journal_remove_journal_head(bh);
1312                         __brelse(bh);
1313                         if (!buffer_jbd(bh)) {
1314                                 spin_unlock(&journal->j_list_lock);
1315                                 jbd_unlock_bh_state(bh);
1316                                 __bforget(bh);
1317                                 goto drop;
1318                         }
1319                 }
1320         } else if (jh->b_transaction) {
1321                 J_ASSERT_JH(jh, (jh->b_transaction ==
1322                                  journal->j_committing_transaction));
1323                 /* However, if the buffer is still owned by a prior
1324                  * (committing) transaction, we can't drop it yet... */
1325                 JBUFFER_TRACE(jh, "belongs to older transaction");
1326                 /* ... but we CAN drop it from the new transaction if we
1327                  * have also modified it since the original commit. */
1328
1329                 if (jh->b_next_transaction) {
1330                         J_ASSERT(jh->b_next_transaction == transaction);
1331                         jh->b_next_transaction = NULL;
1332
1333                         /*
1334                          * only drop a reference if this transaction modified
1335                          * the buffer
1336                          */
1337                         if (was_modified)
1338                                 drop_reserve = 1;
1339                 }
1340         }
1341
1342 not_jbd:
1343         spin_unlock(&journal->j_list_lock);
1344         jbd_unlock_bh_state(bh);
1345         __brelse(bh);
1346 drop:
1347         if (drop_reserve) {
1348                 /* no need to reserve log space for this block -bzzz */
1349                 handle->h_buffer_credits++;
1350         }
1351         return err;
1352 }
1353
1354 /**
1355  * int journal_stop() - complete a transaction
1356  * @handle: tranaction to complete.
1357  *
1358  * All done for a particular handle.
1359  *
1360  * There is not much action needed here.  We just return any remaining
1361  * buffer credits to the transaction and remove the handle.  The only
1362  * complication is that we need to start a commit operation if the
1363  * filesystem is marked for synchronous update.
1364  *
1365  * journal_stop itself will not usually return an error, but it may
1366  * do so in unusual circumstances.  In particular, expect it to
1367  * return -EIO if a journal_abort has been executed since the
1368  * transaction began.
1369  */
1370 int journal_stop(handle_t *handle)
1371 {
1372         transaction_t *transaction = handle->h_transaction;
1373         journal_t *journal = transaction->t_journal;
1374         int old_handle_count, err;
1375         pid_t pid;
1376
1377         J_ASSERT(journal_current_handle() == handle);
1378
1379         if (is_handle_aborted(handle))
1380                 err = -EIO;
1381         else {
1382                 J_ASSERT(transaction->t_updates > 0);
1383                 err = 0;
1384         }
1385
1386         if (--handle->h_ref > 0) {
1387                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1388                           handle->h_ref);
1389                 return err;
1390         }
1391
1392         jbd_debug(4, "Handle %p going down\n", handle);
1393
1394         /*
1395          * Implement synchronous transaction batching.  If the handle
1396          * was synchronous, don't force a commit immediately.  Let's
1397          * yield and let another thread piggyback onto this transaction.
1398          * Keep doing that while new threads continue to arrive.
1399          * It doesn't cost much - we're about to run a commit and sleep
1400          * on IO anyway.  Speeds up many-threaded, many-dir operations
1401          * by 30x or more...
1402          *
1403          * But don't do this if this process was the most recent one to
1404          * perform a synchronous write.  We do this to detect the case where a
1405          * single process is doing a stream of sync writes.  No point in waiting
1406          * for joiners in that case.
1407          */
1408         pid = current->pid;
1409         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1410                 journal->j_last_sync_writer = pid;
1411                 do {
1412                         old_handle_count = transaction->t_handle_count;
1413                         schedule_timeout_uninterruptible(1);
1414                 } while (old_handle_count != transaction->t_handle_count);
1415         }
1416
1417         current->journal_info = NULL;
1418         spin_lock(&journal->j_state_lock);
1419         spin_lock(&transaction->t_handle_lock);
1420         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1421         transaction->t_updates--;
1422         if (!transaction->t_updates) {
1423                 wake_up(&journal->j_wait_updates);
1424                 if (journal->j_barrier_count)
1425                         wake_up(&journal->j_wait_transaction_locked);
1426         }
1427
1428         /*
1429          * If the handle is marked SYNC, we need to set another commit
1430          * going!  We also want to force a commit if the current
1431          * transaction is occupying too much of the log, or if the
1432          * transaction is too old now.
1433          */
1434         if (handle->h_sync ||
1435                         transaction->t_outstanding_credits >
1436                                 journal->j_max_transaction_buffers ||
1437                         time_after_eq(jiffies, transaction->t_expires)) {
1438                 /* Do this even for aborted journals: an abort still
1439                  * completes the commit thread, it just doesn't write
1440                  * anything to disk. */
1441                 tid_t tid = transaction->t_tid;
1442
1443                 spin_unlock(&transaction->t_handle_lock);
1444                 jbd_debug(2, "transaction too old, requesting commit for "
1445                                         "handle %p\n", handle);
1446                 /* This is non-blocking */
1447                 __log_start_commit(journal, transaction->t_tid);
1448                 spin_unlock(&journal->j_state_lock);
1449
1450                 /*
1451                  * Special case: JFS_SYNC synchronous updates require us
1452                  * to wait for the commit to complete.
1453                  */
1454                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1455                         err = log_wait_commit(journal, tid);
1456         } else {
1457                 spin_unlock(&transaction->t_handle_lock);
1458                 spin_unlock(&journal->j_state_lock);
1459         }
1460
1461         lock_map_release(&handle->h_lockdep_map);
1462
1463         jbd_free_handle(handle);
1464         return err;
1465 }
1466
1467 /**
1468  * int journal_force_commit() - force any uncommitted transactions
1469  * @journal: journal to force
1470  *
1471  * For synchronous operations: force any uncommitted transactions
1472  * to disk.  May seem kludgy, but it reuses all the handle batching
1473  * code in a very simple manner.
1474  */
1475 int journal_force_commit(journal_t *journal)
1476 {
1477         handle_t *handle;
1478         int ret;
1479
1480         handle = journal_start(journal, 1);
1481         if (IS_ERR(handle)) {
1482                 ret = PTR_ERR(handle);
1483         } else {
1484                 handle->h_sync = 1;
1485                 ret = journal_stop(handle);
1486         }
1487         return ret;
1488 }
1489
1490 /*
1491  *
1492  * List management code snippets: various functions for manipulating the
1493  * transaction buffer lists.
1494  *
1495  */
1496
1497 /*
1498  * Append a buffer to a transaction list, given the transaction's list head
1499  * pointer.
1500  *
1501  * j_list_lock is held.
1502  *
1503  * jbd_lock_bh_state(jh2bh(jh)) is held.
1504  */
1505
1506 static inline void
1507 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1508 {
1509         if (!*list) {
1510                 jh->b_tnext = jh->b_tprev = jh;
1511                 *list = jh;
1512         } else {
1513                 /* Insert at the tail of the list to preserve order */
1514                 struct journal_head *first = *list, *last = first->b_tprev;
1515                 jh->b_tprev = last;
1516                 jh->b_tnext = first;
1517                 last->b_tnext = first->b_tprev = jh;
1518         }
1519 }
1520
1521 /*
1522  * Remove a buffer from a transaction list, given the transaction's list
1523  * head pointer.
1524  *
1525  * Called with j_list_lock held, and the journal may not be locked.
1526  *
1527  * jbd_lock_bh_state(jh2bh(jh)) is held.
1528  */
1529
1530 static inline void
1531 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1532 {
1533         if (*list == jh) {
1534                 *list = jh->b_tnext;
1535                 if (*list == jh)
1536                         *list = NULL;
1537         }
1538         jh->b_tprev->b_tnext = jh->b_tnext;
1539         jh->b_tnext->b_tprev = jh->b_tprev;
1540 }
1541
1542 /*
1543  * Remove a buffer from the appropriate transaction list.
1544  *
1545  * Note that this function can *change* the value of
1546  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1547  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1548  * is holding onto a copy of one of thee pointers, it could go bad.
1549  * Generally the caller needs to re-read the pointer from the transaction_t.
1550  *
1551  * Called under j_list_lock.  The journal may not be locked.
1552  */
1553 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1554 {
1555         struct journal_head **list = NULL;
1556         transaction_t *transaction;
1557         struct buffer_head *bh = jh2bh(jh);
1558
1559         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1560         transaction = jh->b_transaction;
1561         if (transaction)
1562                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1563
1564         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1565         if (jh->b_jlist != BJ_None)
1566                 J_ASSERT_JH(jh, transaction != NULL);
1567
1568         switch (jh->b_jlist) {
1569         case BJ_None:
1570                 return;
1571         case BJ_SyncData:
1572                 list = &transaction->t_sync_datalist;
1573                 break;
1574         case BJ_Metadata:
1575                 transaction->t_nr_buffers--;
1576                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1577                 list = &transaction->t_buffers;
1578                 break;
1579         case BJ_Forget:
1580                 list = &transaction->t_forget;
1581                 break;
1582         case BJ_IO:
1583                 list = &transaction->t_iobuf_list;
1584                 break;
1585         case BJ_Shadow:
1586                 list = &transaction->t_shadow_list;
1587                 break;
1588         case BJ_LogCtl:
1589                 list = &transaction->t_log_list;
1590                 break;
1591         case BJ_Reserved:
1592                 list = &transaction->t_reserved_list;
1593                 break;
1594         case BJ_Locked:
1595                 list = &transaction->t_locked_list;
1596                 break;
1597         }
1598
1599         __blist_del_buffer(list, jh);
1600         jh->b_jlist = BJ_None;
1601         if (test_clear_buffer_jbddirty(bh))
1602                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1603 }
1604
1605 void __journal_unfile_buffer(struct journal_head *jh)
1606 {
1607         __journal_temp_unlink_buffer(jh);
1608         jh->b_transaction = NULL;
1609 }
1610
1611 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1612 {
1613         jbd_lock_bh_state(jh2bh(jh));
1614         spin_lock(&journal->j_list_lock);
1615         __journal_unfile_buffer(jh);
1616         spin_unlock(&journal->j_list_lock);
1617         jbd_unlock_bh_state(jh2bh(jh));
1618 }
1619
1620 /*
1621  * Called from journal_try_to_free_buffers().
1622  *
1623  * Called under jbd_lock_bh_state(bh)
1624  */
1625 static void
1626 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1627 {
1628         struct journal_head *jh;
1629
1630         jh = bh2jh(bh);
1631
1632         if (buffer_locked(bh) || buffer_dirty(bh))
1633                 goto out;
1634
1635         if (jh->b_next_transaction != NULL)
1636                 goto out;
1637
1638         spin_lock(&journal->j_list_lock);
1639         if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1640                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1641                         /* A written-back ordered data buffer */
1642                         JBUFFER_TRACE(jh, "release data");
1643                         __journal_unfile_buffer(jh);
1644                         journal_remove_journal_head(bh);
1645                         __brelse(bh);
1646                 }
1647         } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1648                 /* written-back checkpointed metadata buffer */
1649                 if (jh->b_jlist == BJ_None) {
1650                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1651                         __journal_remove_checkpoint(jh);
1652                         journal_remove_journal_head(bh);
1653                         __brelse(bh);
1654                 }
1655         }
1656         spin_unlock(&journal->j_list_lock);
1657 out:
1658         return;
1659 }
1660
1661 /*
1662  * journal_try_to_free_buffers() could race with journal_commit_transaction()
1663  * The latter might still hold the a count on buffers when inspecting
1664  * them on t_syncdata_list or t_locked_list.
1665  *
1666  * journal_try_to_free_buffers() will call this function to
1667  * wait for the current transaction to finish syncing data buffers, before
1668  * tryinf to free that buffer.
1669  *
1670  * Called with journal->j_state_lock held.
1671  */
1672 static void journal_wait_for_transaction_sync_data(journal_t *journal)
1673 {
1674         transaction_t *transaction = NULL;
1675         tid_t tid;
1676
1677         spin_lock(&journal->j_state_lock);
1678         transaction = journal->j_committing_transaction;
1679
1680         if (!transaction) {
1681                 spin_unlock(&journal->j_state_lock);
1682                 return;
1683         }
1684
1685         tid = transaction->t_tid;
1686         spin_unlock(&journal->j_state_lock);
1687         log_wait_commit(journal, tid);
1688 }
1689
1690 /**
1691  * int journal_try_to_free_buffers() - try to free page buffers.
1692  * @journal: journal for operation
1693  * @page: to try and free
1694  * @gfp_mask: we use the mask to detect how hard should we try to release
1695  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1696  * release the buffers.
1697  *
1698  *
1699  * For all the buffers on this page,
1700  * if they are fully written out ordered data, move them onto BUF_CLEAN
1701  * so try_to_free_buffers() can reap them.
1702  *
1703  * This function returns non-zero if we wish try_to_free_buffers()
1704  * to be called. We do this if the page is releasable by try_to_free_buffers().
1705  * We also do it if the page has locked or dirty buffers and the caller wants
1706  * us to perform sync or async writeout.
1707  *
1708  * This complicates JBD locking somewhat.  We aren't protected by the
1709  * BKL here.  We wish to remove the buffer from its committing or
1710  * running transaction's ->t_datalist via __journal_unfile_buffer.
1711  *
1712  * This may *change* the value of transaction_t->t_datalist, so anyone
1713  * who looks at t_datalist needs to lock against this function.
1714  *
1715  * Even worse, someone may be doing a journal_dirty_data on this
1716  * buffer.  So we need to lock against that.  journal_dirty_data()
1717  * will come out of the lock with the buffer dirty, which makes it
1718  * ineligible for release here.
1719  *
1720  * Who else is affected by this?  hmm...  Really the only contender
1721  * is do_get_write_access() - it could be looking at the buffer while
1722  * journal_try_to_free_buffer() is changing its state.  But that
1723  * cannot happen because we never reallocate freed data as metadata
1724  * while the data is part of a transaction.  Yes?
1725  *
1726  * Return 0 on failure, 1 on success
1727  */
1728 int journal_try_to_free_buffers(journal_t *journal,
1729                                 struct page *page, gfp_t gfp_mask)
1730 {
1731         struct buffer_head *head;
1732         struct buffer_head *bh;
1733         int ret = 0;
1734
1735         J_ASSERT(PageLocked(page));
1736
1737         head = page_buffers(page);
1738         bh = head;
1739         do {
1740                 struct journal_head *jh;
1741
1742                 /*
1743                  * We take our own ref against the journal_head here to avoid
1744                  * having to add tons of locking around each instance of
1745                  * journal_remove_journal_head() and journal_put_journal_head().
1746                  */
1747                 jh = journal_grab_journal_head(bh);
1748                 if (!jh)
1749                         continue;
1750
1751                 jbd_lock_bh_state(bh);
1752                 __journal_try_to_free_buffer(journal, bh);
1753                 journal_put_journal_head(jh);
1754                 jbd_unlock_bh_state(bh);
1755                 if (buffer_jbd(bh))
1756                         goto busy;
1757         } while ((bh = bh->b_this_page) != head);
1758
1759         ret = try_to_free_buffers(page);
1760
1761         /*
1762          * There are a number of places where journal_try_to_free_buffers()
1763          * could race with journal_commit_transaction(), the later still
1764          * holds the reference to the buffers to free while processing them.
1765          * try_to_free_buffers() failed to free those buffers. Some of the
1766          * caller of releasepage() request page buffers to be dropped, otherwise
1767          * treat the fail-to-free as errors (such as generic_file_direct_IO())
1768          *
1769          * So, if the caller of try_to_release_page() wants the synchronous
1770          * behaviour(i.e make sure buffers are dropped upon return),
1771          * let's wait for the current transaction to finish flush of
1772          * dirty data buffers, then try to free those buffers again,
1773          * with the journal locked.
1774          */
1775         if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
1776                 journal_wait_for_transaction_sync_data(journal);
1777                 ret = try_to_free_buffers(page);
1778         }
1779
1780 busy:
1781         return ret;
1782 }
1783
1784 /*
1785  * This buffer is no longer needed.  If it is on an older transaction's
1786  * checkpoint list we need to record it on this transaction's forget list
1787  * to pin this buffer (and hence its checkpointing transaction) down until
1788  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1789  * release it.
1790  * Returns non-zero if JBD no longer has an interest in the buffer.
1791  *
1792  * Called under j_list_lock.
1793  *
1794  * Called under jbd_lock_bh_state(bh).
1795  */
1796 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1797 {
1798         int may_free = 1;
1799         struct buffer_head *bh = jh2bh(jh);
1800
1801         __journal_unfile_buffer(jh);
1802
1803         if (jh->b_cp_transaction) {
1804                 JBUFFER_TRACE(jh, "on running+cp transaction");
1805                 __journal_file_buffer(jh, transaction, BJ_Forget);
1806                 clear_buffer_jbddirty(bh);
1807                 may_free = 0;
1808         } else {
1809                 JBUFFER_TRACE(jh, "on running transaction");
1810                 journal_remove_journal_head(bh);
1811                 __brelse(bh);
1812         }
1813         return may_free;
1814 }
1815
1816 /*
1817  * journal_invalidatepage
1818  *
1819  * This code is tricky.  It has a number of cases to deal with.
1820  *
1821  * There are two invariants which this code relies on:
1822  *
1823  * i_size must be updated on disk before we start calling invalidatepage on the
1824  * data.
1825  *
1826  *  This is done in ext3 by defining an ext3_setattr method which
1827  *  updates i_size before truncate gets going.  By maintaining this
1828  *  invariant, we can be sure that it is safe to throw away any buffers
1829  *  attached to the current transaction: once the transaction commits,
1830  *  we know that the data will not be needed.
1831  *
1832  *  Note however that we can *not* throw away data belonging to the
1833  *  previous, committing transaction!
1834  *
1835  * Any disk blocks which *are* part of the previous, committing
1836  * transaction (and which therefore cannot be discarded immediately) are
1837  * not going to be reused in the new running transaction
1838  *
1839  *  The bitmap committed_data images guarantee this: any block which is
1840  *  allocated in one transaction and removed in the next will be marked
1841  *  as in-use in the committed_data bitmap, so cannot be reused until
1842  *  the next transaction to delete the block commits.  This means that
1843  *  leaving committing buffers dirty is quite safe: the disk blocks
1844  *  cannot be reallocated to a different file and so buffer aliasing is
1845  *  not possible.
1846  *
1847  *
1848  * The above applies mainly to ordered data mode.  In writeback mode we
1849  * don't make guarantees about the order in which data hits disk --- in
1850  * particular we don't guarantee that new dirty data is flushed before
1851  * transaction commit --- so it is always safe just to discard data
1852  * immediately in that mode.  --sct
1853  */
1854
1855 /*
1856  * The journal_unmap_buffer helper function returns zero if the buffer
1857  * concerned remains pinned as an anonymous buffer belonging to an older
1858  * transaction.
1859  *
1860  * We're outside-transaction here.  Either or both of j_running_transaction
1861  * and j_committing_transaction may be NULL.
1862  */
1863 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1864 {
1865         transaction_t *transaction;
1866         struct journal_head *jh;
1867         int may_free = 1;
1868         int ret;
1869
1870         BUFFER_TRACE(bh, "entry");
1871
1872         /*
1873          * It is safe to proceed here without the j_list_lock because the
1874          * buffers cannot be stolen by try_to_free_buffers as long as we are
1875          * holding the page lock. --sct
1876          */
1877
1878         if (!buffer_jbd(bh))
1879                 goto zap_buffer_unlocked;
1880
1881         spin_lock(&journal->j_state_lock);
1882         jbd_lock_bh_state(bh);
1883         spin_lock(&journal->j_list_lock);
1884
1885         jh = journal_grab_journal_head(bh);
1886         if (!jh)
1887                 goto zap_buffer_no_jh;
1888
1889         transaction = jh->b_transaction;
1890         if (transaction == NULL) {
1891                 /* First case: not on any transaction.  If it
1892                  * has no checkpoint link, then we can zap it:
1893                  * it's a writeback-mode buffer so we don't care
1894                  * if it hits disk safely. */
1895                 if (!jh->b_cp_transaction) {
1896                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1897                         goto zap_buffer;
1898                 }
1899
1900                 if (!buffer_dirty(bh)) {
1901                         /* bdflush has written it.  We can drop it now */
1902                         goto zap_buffer;
1903                 }
1904
1905                 /* OK, it must be in the journal but still not
1906                  * written fully to disk: it's metadata or
1907                  * journaled data... */
1908
1909                 if (journal->j_running_transaction) {
1910                         /* ... and once the current transaction has
1911                          * committed, the buffer won't be needed any
1912                          * longer. */
1913                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1914                         ret = __dispose_buffer(jh,
1915                                         journal->j_running_transaction);
1916                         journal_put_journal_head(jh);
1917                         spin_unlock(&journal->j_list_lock);
1918                         jbd_unlock_bh_state(bh);
1919                         spin_unlock(&journal->j_state_lock);
1920                         return ret;
1921                 } else {
1922                         /* There is no currently-running transaction. So the
1923                          * orphan record which we wrote for this file must have
1924                          * passed into commit.  We must attach this buffer to
1925                          * the committing transaction, if it exists. */
1926                         if (journal->j_committing_transaction) {
1927                                 JBUFFER_TRACE(jh, "give to committing trans");
1928                                 ret = __dispose_buffer(jh,
1929                                         journal->j_committing_transaction);
1930                                 journal_put_journal_head(jh);
1931                                 spin_unlock(&journal->j_list_lock);
1932                                 jbd_unlock_bh_state(bh);
1933                                 spin_unlock(&journal->j_state_lock);
1934                                 return ret;
1935                         } else {
1936                                 /* The orphan record's transaction has
1937                                  * committed.  We can cleanse this buffer */
1938                                 clear_buffer_jbddirty(bh);
1939                                 goto zap_buffer;
1940                         }
1941                 }
1942         } else if (transaction == journal->j_committing_transaction) {
1943                 JBUFFER_TRACE(jh, "on committing transaction");
1944                 if (jh->b_jlist == BJ_Locked) {
1945                         /*
1946                          * The buffer is on the committing transaction's locked
1947                          * list.  We have the buffer locked, so I/O has
1948                          * completed.  So we can nail the buffer now.
1949                          */
1950                         may_free = __dispose_buffer(jh, transaction);
1951                         goto zap_buffer;
1952                 }
1953                 /*
1954                  * If it is committing, we simply cannot touch it.  We
1955                  * can remove it's next_transaction pointer from the
1956                  * running transaction if that is set, but nothing
1957                  * else. */
1958                 set_buffer_freed(bh);
1959                 if (jh->b_next_transaction) {
1960                         J_ASSERT(jh->b_next_transaction ==
1961                                         journal->j_running_transaction);
1962                         jh->b_next_transaction = NULL;
1963                 }
1964                 journal_put_journal_head(jh);
1965                 spin_unlock(&journal->j_list_lock);
1966                 jbd_unlock_bh_state(bh);
1967                 spin_unlock(&journal->j_state_lock);
1968                 return 0;
1969         } else {
1970                 /* Good, the buffer belongs to the running transaction.
1971                  * We are writing our own transaction's data, not any
1972                  * previous one's, so it is safe to throw it away
1973                  * (remember that we expect the filesystem to have set
1974                  * i_size already for this truncate so recovery will not
1975                  * expose the disk blocks we are discarding here.) */
1976                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1977                 JBUFFER_TRACE(jh, "on running transaction");
1978                 may_free = __dispose_buffer(jh, transaction);
1979         }
1980
1981 zap_buffer:
1982         journal_put_journal_head(jh);
1983 zap_buffer_no_jh:
1984         spin_unlock(&journal->j_list_lock);
1985         jbd_unlock_bh_state(bh);
1986         spin_unlock(&journal->j_state_lock);
1987 zap_buffer_unlocked:
1988         clear_buffer_dirty(bh);
1989         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1990         clear_buffer_mapped(bh);
1991         clear_buffer_req(bh);
1992         clear_buffer_new(bh);
1993         bh->b_bdev = NULL;
1994         return may_free;
1995 }
1996
1997 /**
1998  * void journal_invalidatepage() - invalidate a journal page
1999  * @journal: journal to use for flush
2000  * @page:    page to flush
2001  * @offset:  length of page to invalidate.
2002  *
2003  * Reap page buffers containing data after offset in page.
2004  */
2005 void journal_invalidatepage(journal_t *journal,
2006                       struct page *page,
2007                       unsigned long offset)
2008 {
2009         struct buffer_head *head, *bh, *next;
2010         unsigned int curr_off = 0;
2011         int may_free = 1;
2012
2013         if (!PageLocked(page))
2014                 BUG();
2015         if (!page_has_buffers(page))
2016                 return;
2017
2018         /* We will potentially be playing with lists other than just the
2019          * data lists (especially for journaled data mode), so be
2020          * cautious in our locking. */
2021
2022         head = bh = page_buffers(page);
2023         do {
2024                 unsigned int next_off = curr_off + bh->b_size;
2025                 next = bh->b_this_page;
2026
2027                 if (offset <= curr_off) {
2028                         /* This block is wholly outside the truncation point */
2029                         lock_buffer(bh);
2030                         may_free &= journal_unmap_buffer(journal, bh);
2031                         unlock_buffer(bh);
2032                 }
2033                 curr_off = next_off;
2034                 bh = next;
2035
2036         } while (bh != head);
2037
2038         if (!offset) {
2039                 if (may_free && try_to_free_buffers(page))
2040                         J_ASSERT(!page_has_buffers(page));
2041         }
2042 }
2043
2044 /*
2045  * File a buffer on the given transaction list.
2046  */
2047 void __journal_file_buffer(struct journal_head *jh,
2048                         transaction_t *transaction, int jlist)
2049 {
2050         struct journal_head **list = NULL;
2051         int was_dirty = 0;
2052         struct buffer_head *bh = jh2bh(jh);
2053
2054         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2055         assert_spin_locked(&transaction->t_journal->j_list_lock);
2056
2057         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2058         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2059                                 jh->b_transaction == NULL);
2060
2061         if (jh->b_transaction && jh->b_jlist == jlist)
2062                 return;
2063
2064         /* The following list of buffer states needs to be consistent
2065          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
2066          * state. */
2067
2068         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2069             jlist == BJ_Shadow || jlist == BJ_Forget) {
2070                 if (test_clear_buffer_dirty(bh) ||
2071                     test_clear_buffer_jbddirty(bh))
2072                         was_dirty = 1;
2073         }
2074
2075         if (jh->b_transaction)
2076                 __journal_temp_unlink_buffer(jh);
2077         jh->b_transaction = transaction;
2078
2079         switch (jlist) {
2080         case BJ_None:
2081                 J_ASSERT_JH(jh, !jh->b_committed_data);
2082                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2083                 return;
2084         case BJ_SyncData:
2085                 list = &transaction->t_sync_datalist;
2086                 break;
2087         case BJ_Metadata:
2088                 transaction->t_nr_buffers++;
2089                 list = &transaction->t_buffers;
2090                 break;
2091         case BJ_Forget:
2092                 list = &transaction->t_forget;
2093                 break;
2094         case BJ_IO:
2095                 list = &transaction->t_iobuf_list;
2096                 break;
2097         case BJ_Shadow:
2098                 list = &transaction->t_shadow_list;
2099                 break;
2100         case BJ_LogCtl:
2101                 list = &transaction->t_log_list;
2102                 break;
2103         case BJ_Reserved:
2104                 list = &transaction->t_reserved_list;
2105                 break;
2106         case BJ_Locked:
2107                 list =  &transaction->t_locked_list;
2108                 break;
2109         }
2110
2111         __blist_add_buffer(list, jh);
2112         jh->b_jlist = jlist;
2113
2114         if (was_dirty)
2115                 set_buffer_jbddirty(bh);
2116 }
2117
2118 void journal_file_buffer(struct journal_head *jh,
2119                                 transaction_t *transaction, int jlist)
2120 {
2121         jbd_lock_bh_state(jh2bh(jh));
2122         spin_lock(&transaction->t_journal->j_list_lock);
2123         __journal_file_buffer(jh, transaction, jlist);
2124         spin_unlock(&transaction->t_journal->j_list_lock);
2125         jbd_unlock_bh_state(jh2bh(jh));
2126 }
2127
2128 /*
2129  * Remove a buffer from its current buffer list in preparation for
2130  * dropping it from its current transaction entirely.  If the buffer has
2131  * already started to be used by a subsequent transaction, refile the
2132  * buffer on that transaction's metadata list.
2133  *
2134  * Called under journal->j_list_lock
2135  *
2136  * Called under jbd_lock_bh_state(jh2bh(jh))
2137  */
2138 void __journal_refile_buffer(struct journal_head *jh)
2139 {
2140         int was_dirty;
2141         struct buffer_head *bh = jh2bh(jh);
2142
2143         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2144         if (jh->b_transaction)
2145                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2146
2147         /* If the buffer is now unused, just drop it. */
2148         if (jh->b_next_transaction == NULL) {
2149                 __journal_unfile_buffer(jh);
2150                 return;
2151         }
2152
2153         /*
2154          * It has been modified by a later transaction: add it to the new
2155          * transaction's metadata list.
2156          */
2157
2158         was_dirty = test_clear_buffer_jbddirty(bh);
2159         __journal_temp_unlink_buffer(jh);
2160         jh->b_transaction = jh->b_next_transaction;
2161         jh->b_next_transaction = NULL;
2162         __journal_file_buffer(jh, jh->b_transaction,
2163                                 jh->b_modified ? BJ_Metadata : BJ_Reserved);
2164         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2165
2166         if (was_dirty)
2167                 set_buffer_jbddirty(bh);
2168 }
2169
2170 /*
2171  * For the unlocked version of this call, also make sure that any
2172  * hanging journal_head is cleaned up if necessary.
2173  *
2174  * __journal_refile_buffer is usually called as part of a single locked
2175  * operation on a buffer_head, in which the caller is probably going to
2176  * be hooking the journal_head onto other lists.  In that case it is up
2177  * to the caller to remove the journal_head if necessary.  For the
2178  * unlocked journal_refile_buffer call, the caller isn't going to be
2179  * doing anything else to the buffer so we need to do the cleanup
2180  * ourselves to avoid a jh leak.
2181  *
2182  * *** The journal_head may be freed by this call! ***
2183  */
2184 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2185 {
2186         struct buffer_head *bh = jh2bh(jh);
2187
2188         jbd_lock_bh_state(bh);
2189         spin_lock(&journal->j_list_lock);
2190
2191         __journal_refile_buffer(jh);
2192         jbd_unlock_bh_state(bh);
2193         journal_remove_journal_head(bh);
2194
2195         spin_unlock(&journal->j_list_lock);
2196         __brelse(bh);
2197 }