Merge branch 'topic/snd_card_new-err' into for-linus
[linux-2.6] / fs / jbd2 / transaction.c
1 /*
2  * linux/fs/jbd2/transaction.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem transaction handling code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages transactions (compound commits managed by the
16  * journaling code) and handles (individual atomic operations by the
17  * filesystem).
18  */
19
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29
30 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
31
32 /*
33  * jbd2_get_transaction: obtain a new transaction_t object.
34  *
35  * Simply allocate and initialise a new transaction.  Create it in
36  * RUNNING state and add it to the current journal (which should not
37  * have an existing running transaction: we only make a new transaction
38  * once we have started to commit the old one).
39  *
40  * Preconditions:
41  *      The journal MUST be locked.  We don't perform atomic mallocs on the
42  *      new transaction and we can't block without protecting against other
43  *      processes trying to touch the journal while it is in transition.
44  *
45  */
46
47 static transaction_t *
48 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
49 {
50         transaction->t_journal = journal;
51         transaction->t_state = T_RUNNING;
52         transaction->t_start_time = ktime_get();
53         transaction->t_tid = journal->j_transaction_sequence++;
54         transaction->t_expires = jiffies + journal->j_commit_interval;
55         spin_lock_init(&transaction->t_handle_lock);
56         INIT_LIST_HEAD(&transaction->t_inode_list);
57         INIT_LIST_HEAD(&transaction->t_private_list);
58
59         /* Set up the commit timer for the new transaction. */
60         journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
61         add_timer(&journal->j_commit_timer);
62
63         J_ASSERT(journal->j_running_transaction == NULL);
64         journal->j_running_transaction = transaction;
65         transaction->t_max_wait = 0;
66         transaction->t_start = jiffies;
67
68         return transaction;
69 }
70
71 /*
72  * Handle management.
73  *
74  * A handle_t is an object which represents a single atomic update to a
75  * filesystem, and which tracks all of the modifications which form part
76  * of that one update.
77  */
78
79 /*
80  * start_this_handle: Given a handle, deal with any locking or stalling
81  * needed to make sure that there is enough journal space for the handle
82  * to begin.  Attach the handle to a transaction and set up the
83  * transaction's buffer credits.
84  */
85
86 static int start_this_handle(journal_t *journal, handle_t *handle)
87 {
88         transaction_t *transaction;
89         int needed;
90         int nblocks = handle->h_buffer_credits;
91         transaction_t *new_transaction = NULL;
92         int ret = 0;
93         unsigned long ts = jiffies;
94
95         if (nblocks > journal->j_max_transaction_buffers) {
96                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
97                        current->comm, nblocks,
98                        journal->j_max_transaction_buffers);
99                 ret = -ENOSPC;
100                 goto out;
101         }
102
103 alloc_transaction:
104         if (!journal->j_running_transaction) {
105                 new_transaction = kzalloc(sizeof(*new_transaction),
106                                                 GFP_NOFS|__GFP_NOFAIL);
107                 if (!new_transaction) {
108                         ret = -ENOMEM;
109                         goto out;
110                 }
111         }
112
113         jbd_debug(3, "New handle %p going live.\n", handle);
114
115 repeat:
116
117         /*
118          * We need to hold j_state_lock until t_updates has been incremented,
119          * for proper journal barrier handling
120          */
121         spin_lock(&journal->j_state_lock);
122 repeat_locked:
123         if (is_journal_aborted(journal) ||
124             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
125                 spin_unlock(&journal->j_state_lock);
126                 ret = -EROFS;
127                 goto out;
128         }
129
130         /* Wait on the journal's transaction barrier if necessary */
131         if (journal->j_barrier_count) {
132                 spin_unlock(&journal->j_state_lock);
133                 wait_event(journal->j_wait_transaction_locked,
134                                 journal->j_barrier_count == 0);
135                 goto repeat;
136         }
137
138         if (!journal->j_running_transaction) {
139                 if (!new_transaction) {
140                         spin_unlock(&journal->j_state_lock);
141                         goto alloc_transaction;
142                 }
143                 jbd2_get_transaction(journal, new_transaction);
144                 new_transaction = NULL;
145         }
146
147         transaction = journal->j_running_transaction;
148
149         /*
150          * If the current transaction is locked down for commit, wait for the
151          * lock to be released.
152          */
153         if (transaction->t_state == T_LOCKED) {
154                 DEFINE_WAIT(wait);
155
156                 prepare_to_wait(&journal->j_wait_transaction_locked,
157                                         &wait, TASK_UNINTERRUPTIBLE);
158                 spin_unlock(&journal->j_state_lock);
159                 schedule();
160                 finish_wait(&journal->j_wait_transaction_locked, &wait);
161                 goto repeat;
162         }
163
164         /*
165          * If there is not enough space left in the log to write all potential
166          * buffers requested by this operation, we need to stall pending a log
167          * checkpoint to free some more log space.
168          */
169         spin_lock(&transaction->t_handle_lock);
170         needed = transaction->t_outstanding_credits + nblocks;
171
172         if (needed > journal->j_max_transaction_buffers) {
173                 /*
174                  * If the current transaction is already too large, then start
175                  * to commit it: we can then go back and attach this handle to
176                  * a new transaction.
177                  */
178                 DEFINE_WAIT(wait);
179
180                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
181                 spin_unlock(&transaction->t_handle_lock);
182                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
183                                 TASK_UNINTERRUPTIBLE);
184                 __jbd2_log_start_commit(journal, transaction->t_tid);
185                 spin_unlock(&journal->j_state_lock);
186                 schedule();
187                 finish_wait(&journal->j_wait_transaction_locked, &wait);
188                 goto repeat;
189         }
190
191         /*
192          * The commit code assumes that it can get enough log space
193          * without forcing a checkpoint.  This is *critical* for
194          * correctness: a checkpoint of a buffer which is also
195          * associated with a committing transaction creates a deadlock,
196          * so commit simply cannot force through checkpoints.
197          *
198          * We must therefore ensure the necessary space in the journal
199          * *before* starting to dirty potentially checkpointed buffers
200          * in the new transaction.
201          *
202          * The worst part is, any transaction currently committing can
203          * reduce the free space arbitrarily.  Be careful to account for
204          * those buffers when checkpointing.
205          */
206
207         /*
208          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
209          * a _lot_ of headroom: 1/4 of the journal plus the size of
210          * the committing transaction.  Really, we only need to give it
211          * committing_transaction->t_outstanding_credits plus "enough" for
212          * the log control blocks.
213          * Also, this test is inconsitent with the matching one in
214          * jbd2_journal_extend().
215          */
216         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
217                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
218                 spin_unlock(&transaction->t_handle_lock);
219                 __jbd2_log_wait_for_space(journal);
220                 goto repeat_locked;
221         }
222
223         /* OK, account for the buffers that this operation expects to
224          * use and add the handle to the running transaction. */
225
226         if (time_after(transaction->t_start, ts)) {
227                 ts = jbd2_time_diff(ts, transaction->t_start);
228                 if (ts > transaction->t_max_wait)
229                         transaction->t_max_wait = ts;
230         }
231
232         handle->h_transaction = transaction;
233         transaction->t_outstanding_credits += nblocks;
234         transaction->t_updates++;
235         transaction->t_handle_count++;
236         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
237                   handle, nblocks, transaction->t_outstanding_credits,
238                   __jbd2_log_space_left(journal));
239         spin_unlock(&transaction->t_handle_lock);
240         spin_unlock(&journal->j_state_lock);
241 out:
242         if (unlikely(new_transaction))          /* It's usually NULL */
243                 kfree(new_transaction);
244         return ret;
245 }
246
247 static struct lock_class_key jbd2_handle_key;
248
249 /* Allocate a new handle.  This should probably be in a slab... */
250 static handle_t *new_handle(int nblocks)
251 {
252         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
253         if (!handle)
254                 return NULL;
255         memset(handle, 0, sizeof(*handle));
256         handle->h_buffer_credits = nblocks;
257         handle->h_ref = 1;
258
259         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
260                                                 &jbd2_handle_key, 0);
261
262         return handle;
263 }
264
265 /**
266  * handle_t *jbd2_journal_start() - Obtain a new handle.
267  * @journal: Journal to start transaction on.
268  * @nblocks: number of block buffer we might modify
269  *
270  * We make sure that the transaction can guarantee at least nblocks of
271  * modified buffers in the log.  We block until the log can guarantee
272  * that much space.
273  *
274  * This function is visible to journal users (like ext3fs), so is not
275  * called with the journal already locked.
276  *
277  * Return a pointer to a newly allocated handle, or NULL on failure
278  */
279 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
280 {
281         handle_t *handle = journal_current_handle();
282         int err;
283
284         if (!journal)
285                 return ERR_PTR(-EROFS);
286
287         if (handle) {
288                 J_ASSERT(handle->h_transaction->t_journal == journal);
289                 handle->h_ref++;
290                 return handle;
291         }
292
293         handle = new_handle(nblocks);
294         if (!handle)
295                 return ERR_PTR(-ENOMEM);
296
297         current->journal_info = handle;
298
299         err = start_this_handle(journal, handle);
300         if (err < 0) {
301                 jbd2_free_handle(handle);
302                 current->journal_info = NULL;
303                 handle = ERR_PTR(err);
304                 goto out;
305         }
306
307         lock_map_acquire(&handle->h_lockdep_map);
308 out:
309         return handle;
310 }
311
312 /**
313  * int jbd2_journal_extend() - extend buffer credits.
314  * @handle:  handle to 'extend'
315  * @nblocks: nr blocks to try to extend by.
316  *
317  * Some transactions, such as large extends and truncates, can be done
318  * atomically all at once or in several stages.  The operation requests
319  * a credit for a number of buffer modications in advance, but can
320  * extend its credit if it needs more.
321  *
322  * jbd2_journal_extend tries to give the running handle more buffer credits.
323  * It does not guarantee that allocation - this is a best-effort only.
324  * The calling process MUST be able to deal cleanly with a failure to
325  * extend here.
326  *
327  * Return 0 on success, non-zero on failure.
328  *
329  * return code < 0 implies an error
330  * return code > 0 implies normal transaction-full status.
331  */
332 int jbd2_journal_extend(handle_t *handle, int nblocks)
333 {
334         transaction_t *transaction = handle->h_transaction;
335         journal_t *journal = transaction->t_journal;
336         int result;
337         int wanted;
338
339         result = -EIO;
340         if (is_handle_aborted(handle))
341                 goto out;
342
343         result = 1;
344
345         spin_lock(&journal->j_state_lock);
346
347         /* Don't extend a locked-down transaction! */
348         if (handle->h_transaction->t_state != T_RUNNING) {
349                 jbd_debug(3, "denied handle %p %d blocks: "
350                           "transaction not running\n", handle, nblocks);
351                 goto error_out;
352         }
353
354         spin_lock(&transaction->t_handle_lock);
355         wanted = transaction->t_outstanding_credits + nblocks;
356
357         if (wanted > journal->j_max_transaction_buffers) {
358                 jbd_debug(3, "denied handle %p %d blocks: "
359                           "transaction too large\n", handle, nblocks);
360                 goto unlock;
361         }
362
363         if (wanted > __jbd2_log_space_left(journal)) {
364                 jbd_debug(3, "denied handle %p %d blocks: "
365                           "insufficient log space\n", handle, nblocks);
366                 goto unlock;
367         }
368
369         handle->h_buffer_credits += nblocks;
370         transaction->t_outstanding_credits += nblocks;
371         result = 0;
372
373         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
374 unlock:
375         spin_unlock(&transaction->t_handle_lock);
376 error_out:
377         spin_unlock(&journal->j_state_lock);
378 out:
379         return result;
380 }
381
382
383 /**
384  * int jbd2_journal_restart() - restart a handle .
385  * @handle:  handle to restart
386  * @nblocks: nr credits requested
387  *
388  * Restart a handle for a multi-transaction filesystem
389  * operation.
390  *
391  * If the jbd2_journal_extend() call above fails to grant new buffer credits
392  * to a running handle, a call to jbd2_journal_restart will commit the
393  * handle's transaction so far and reattach the handle to a new
394  * transaction capabable of guaranteeing the requested number of
395  * credits.
396  */
397
398 int jbd2_journal_restart(handle_t *handle, int nblocks)
399 {
400         transaction_t *transaction = handle->h_transaction;
401         journal_t *journal = transaction->t_journal;
402         int ret;
403
404         /* If we've had an abort of any type, don't even think about
405          * actually doing the restart! */
406         if (is_handle_aborted(handle))
407                 return 0;
408
409         /*
410          * First unlink the handle from its current transaction, and start the
411          * commit on that.
412          */
413         J_ASSERT(transaction->t_updates > 0);
414         J_ASSERT(journal_current_handle() == handle);
415
416         spin_lock(&journal->j_state_lock);
417         spin_lock(&transaction->t_handle_lock);
418         transaction->t_outstanding_credits -= handle->h_buffer_credits;
419         transaction->t_updates--;
420
421         if (!transaction->t_updates)
422                 wake_up(&journal->j_wait_updates);
423         spin_unlock(&transaction->t_handle_lock);
424
425         jbd_debug(2, "restarting handle %p\n", handle);
426         __jbd2_log_start_commit(journal, transaction->t_tid);
427         spin_unlock(&journal->j_state_lock);
428
429         handle->h_buffer_credits = nblocks;
430         ret = start_this_handle(journal, handle);
431         return ret;
432 }
433
434
435 /**
436  * void jbd2_journal_lock_updates () - establish a transaction barrier.
437  * @journal:  Journal to establish a barrier on.
438  *
439  * This locks out any further updates from being started, and blocks
440  * until all existing updates have completed, returning only once the
441  * journal is in a quiescent state with no updates running.
442  *
443  * The journal lock should not be held on entry.
444  */
445 void jbd2_journal_lock_updates(journal_t *journal)
446 {
447         DEFINE_WAIT(wait);
448
449         spin_lock(&journal->j_state_lock);
450         ++journal->j_barrier_count;
451
452         /* Wait until there are no running updates */
453         while (1) {
454                 transaction_t *transaction = journal->j_running_transaction;
455
456                 if (!transaction)
457                         break;
458
459                 spin_lock(&transaction->t_handle_lock);
460                 if (!transaction->t_updates) {
461                         spin_unlock(&transaction->t_handle_lock);
462                         break;
463                 }
464                 prepare_to_wait(&journal->j_wait_updates, &wait,
465                                 TASK_UNINTERRUPTIBLE);
466                 spin_unlock(&transaction->t_handle_lock);
467                 spin_unlock(&journal->j_state_lock);
468                 schedule();
469                 finish_wait(&journal->j_wait_updates, &wait);
470                 spin_lock(&journal->j_state_lock);
471         }
472         spin_unlock(&journal->j_state_lock);
473
474         /*
475          * We have now established a barrier against other normal updates, but
476          * we also need to barrier against other jbd2_journal_lock_updates() calls
477          * to make sure that we serialise special journal-locked operations
478          * too.
479          */
480         mutex_lock(&journal->j_barrier);
481 }
482
483 /**
484  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
485  * @journal:  Journal to release the barrier on.
486  *
487  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
488  *
489  * Should be called without the journal lock held.
490  */
491 void jbd2_journal_unlock_updates (journal_t *journal)
492 {
493         J_ASSERT(journal->j_barrier_count != 0);
494
495         mutex_unlock(&journal->j_barrier);
496         spin_lock(&journal->j_state_lock);
497         --journal->j_barrier_count;
498         spin_unlock(&journal->j_state_lock);
499         wake_up(&journal->j_wait_transaction_locked);
500 }
501
502 /*
503  * Report any unexpected dirty buffers which turn up.  Normally those
504  * indicate an error, but they can occur if the user is running (say)
505  * tune2fs to modify the live filesystem, so we need the option of
506  * continuing as gracefully as possible.  #
507  *
508  * The caller should already hold the journal lock and
509  * j_list_lock spinlock: most callers will need those anyway
510  * in order to probe the buffer's journaling state safely.
511  */
512 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
513 {
514         int jlist;
515
516         /* If this buffer is one which might reasonably be dirty
517          * --- ie. data, or not part of this journal --- then
518          * we're OK to leave it alone, but otherwise we need to
519          * move the dirty bit to the journal's own internal
520          * JBDDirty bit. */
521         jlist = jh->b_jlist;
522
523         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
524             jlist == BJ_Shadow || jlist == BJ_Forget) {
525                 struct buffer_head *bh = jh2bh(jh);
526
527                 if (test_clear_buffer_dirty(bh))
528                         set_buffer_jbddirty(bh);
529         }
530 }
531
532 /*
533  * If the buffer is already part of the current transaction, then there
534  * is nothing we need to do.  If it is already part of a prior
535  * transaction which we are still committing to disk, then we need to
536  * make sure that we do not overwrite the old copy: we do copy-out to
537  * preserve the copy going to disk.  We also account the buffer against
538  * the handle's metadata buffer credits (unless the buffer is already
539  * part of the transaction, that is).
540  *
541  */
542 static int
543 do_get_write_access(handle_t *handle, struct journal_head *jh,
544                         int force_copy)
545 {
546         struct buffer_head *bh;
547         transaction_t *transaction;
548         journal_t *journal;
549         int error;
550         char *frozen_buffer = NULL;
551         int need_copy = 0;
552
553         if (is_handle_aborted(handle))
554                 return -EROFS;
555
556         transaction = handle->h_transaction;
557         journal = transaction->t_journal;
558
559         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
560
561         JBUFFER_TRACE(jh, "entry");
562 repeat:
563         bh = jh2bh(jh);
564
565         /* @@@ Need to check for errors here at some point. */
566
567         lock_buffer(bh);
568         jbd_lock_bh_state(bh);
569
570         /* We now hold the buffer lock so it is safe to query the buffer
571          * state.  Is the buffer dirty?
572          *
573          * If so, there are two possibilities.  The buffer may be
574          * non-journaled, and undergoing a quite legitimate writeback.
575          * Otherwise, it is journaled, and we don't expect dirty buffers
576          * in that state (the buffers should be marked JBD_Dirty
577          * instead.)  So either the IO is being done under our own
578          * control and this is a bug, or it's a third party IO such as
579          * dump(8) (which may leave the buffer scheduled for read ---
580          * ie. locked but not dirty) or tune2fs (which may actually have
581          * the buffer dirtied, ugh.)  */
582
583         if (buffer_dirty(bh)) {
584                 /*
585                  * First question: is this buffer already part of the current
586                  * transaction or the existing committing transaction?
587                  */
588                 if (jh->b_transaction) {
589                         J_ASSERT_JH(jh,
590                                 jh->b_transaction == transaction ||
591                                 jh->b_transaction ==
592                                         journal->j_committing_transaction);
593                         if (jh->b_next_transaction)
594                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
595                                                         transaction);
596                 }
597                 /*
598                  * In any case we need to clean the dirty flag and we must
599                  * do it under the buffer lock to be sure we don't race
600                  * with running write-out.
601                  */
602                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
603                 jbd_unexpected_dirty_buffer(jh);
604         }
605
606         unlock_buffer(bh);
607
608         error = -EROFS;
609         if (is_handle_aborted(handle)) {
610                 jbd_unlock_bh_state(bh);
611                 goto out;
612         }
613         error = 0;
614
615         /*
616          * The buffer is already part of this transaction if b_transaction or
617          * b_next_transaction points to it
618          */
619         if (jh->b_transaction == transaction ||
620             jh->b_next_transaction == transaction)
621                 goto done;
622
623         /*
624          * this is the first time this transaction is touching this buffer,
625          * reset the modified flag
626          */
627        jh->b_modified = 0;
628
629         /*
630          * If there is already a copy-out version of this buffer, then we don't
631          * need to make another one
632          */
633         if (jh->b_frozen_data) {
634                 JBUFFER_TRACE(jh, "has frozen data");
635                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
636                 jh->b_next_transaction = transaction;
637                 goto done;
638         }
639
640         /* Is there data here we need to preserve? */
641
642         if (jh->b_transaction && jh->b_transaction != transaction) {
643                 JBUFFER_TRACE(jh, "owned by older transaction");
644                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
645                 J_ASSERT_JH(jh, jh->b_transaction ==
646                                         journal->j_committing_transaction);
647
648                 /* There is one case we have to be very careful about.
649                  * If the committing transaction is currently writing
650                  * this buffer out to disk and has NOT made a copy-out,
651                  * then we cannot modify the buffer contents at all
652                  * right now.  The essence of copy-out is that it is the
653                  * extra copy, not the primary copy, which gets
654                  * journaled.  If the primary copy is already going to
655                  * disk then we cannot do copy-out here. */
656
657                 if (jh->b_jlist == BJ_Shadow) {
658                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
659                         wait_queue_head_t *wqh;
660
661                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
662
663                         JBUFFER_TRACE(jh, "on shadow: sleep");
664                         jbd_unlock_bh_state(bh);
665                         /* commit wakes up all shadow buffers after IO */
666                         for ( ; ; ) {
667                                 prepare_to_wait(wqh, &wait.wait,
668                                                 TASK_UNINTERRUPTIBLE);
669                                 if (jh->b_jlist != BJ_Shadow)
670                                         break;
671                                 schedule();
672                         }
673                         finish_wait(wqh, &wait.wait);
674                         goto repeat;
675                 }
676
677                 /* Only do the copy if the currently-owning transaction
678                  * still needs it.  If it is on the Forget list, the
679                  * committing transaction is past that stage.  The
680                  * buffer had better remain locked during the kmalloc,
681                  * but that should be true --- we hold the journal lock
682                  * still and the buffer is already on the BUF_JOURNAL
683                  * list so won't be flushed.
684                  *
685                  * Subtle point, though: if this is a get_undo_access,
686                  * then we will be relying on the frozen_data to contain
687                  * the new value of the committed_data record after the
688                  * transaction, so we HAVE to force the frozen_data copy
689                  * in that case. */
690
691                 if (jh->b_jlist != BJ_Forget || force_copy) {
692                         JBUFFER_TRACE(jh, "generate frozen data");
693                         if (!frozen_buffer) {
694                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
695                                 jbd_unlock_bh_state(bh);
696                                 frozen_buffer =
697                                         jbd2_alloc(jh2bh(jh)->b_size,
698                                                          GFP_NOFS);
699                                 if (!frozen_buffer) {
700                                         printk(KERN_EMERG
701                                                "%s: OOM for frozen_buffer\n",
702                                                __func__);
703                                         JBUFFER_TRACE(jh, "oom!");
704                                         error = -ENOMEM;
705                                         jbd_lock_bh_state(bh);
706                                         goto done;
707                                 }
708                                 goto repeat;
709                         }
710                         jh->b_frozen_data = frozen_buffer;
711                         frozen_buffer = NULL;
712                         need_copy = 1;
713                 }
714                 jh->b_next_transaction = transaction;
715         }
716
717
718         /*
719          * Finally, if the buffer is not journaled right now, we need to make
720          * sure it doesn't get written to disk before the caller actually
721          * commits the new data
722          */
723         if (!jh->b_transaction) {
724                 JBUFFER_TRACE(jh, "no transaction");
725                 J_ASSERT_JH(jh, !jh->b_next_transaction);
726                 jh->b_transaction = transaction;
727                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
728                 spin_lock(&journal->j_list_lock);
729                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
730                 spin_unlock(&journal->j_list_lock);
731         }
732
733 done:
734         if (need_copy) {
735                 struct page *page;
736                 int offset;
737                 char *source;
738
739                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
740                             "Possible IO failure.\n");
741                 page = jh2bh(jh)->b_page;
742                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
743                 source = kmap_atomic(page, KM_USER0);
744                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
745                 kunmap_atomic(source, KM_USER0);
746
747                 /*
748                  * Now that the frozen data is saved off, we need to store
749                  * any matching triggers.
750                  */
751                 jh->b_frozen_triggers = jh->b_triggers;
752         }
753         jbd_unlock_bh_state(bh);
754
755         /*
756          * If we are about to journal a buffer, then any revoke pending on it is
757          * no longer valid
758          */
759         jbd2_journal_cancel_revoke(handle, jh);
760
761 out:
762         if (unlikely(frozen_buffer))    /* It's usually NULL */
763                 jbd2_free(frozen_buffer, bh->b_size);
764
765         JBUFFER_TRACE(jh, "exit");
766         return error;
767 }
768
769 /**
770  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
771  * @handle: transaction to add buffer modifications to
772  * @bh:     bh to be used for metadata writes
773  * @credits: variable that will receive credits for the buffer
774  *
775  * Returns an error code or 0 on success.
776  *
777  * In full data journalling mode the buffer may be of type BJ_AsyncData,
778  * because we're write()ing a buffer which is also part of a shared mapping.
779  */
780
781 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
782 {
783         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
784         int rc;
785
786         /* We do not want to get caught playing with fields which the
787          * log thread also manipulates.  Make sure that the buffer
788          * completes any outstanding IO before proceeding. */
789         rc = do_get_write_access(handle, jh, 0);
790         jbd2_journal_put_journal_head(jh);
791         return rc;
792 }
793
794
795 /*
796  * When the user wants to journal a newly created buffer_head
797  * (ie. getblk() returned a new buffer and we are going to populate it
798  * manually rather than reading off disk), then we need to keep the
799  * buffer_head locked until it has been completely filled with new
800  * data.  In this case, we should be able to make the assertion that
801  * the bh is not already part of an existing transaction.
802  *
803  * The buffer should already be locked by the caller by this point.
804  * There is no lock ranking violation: it was a newly created,
805  * unlocked buffer beforehand. */
806
807 /**
808  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
809  * @handle: transaction to new buffer to
810  * @bh: new buffer.
811  *
812  * Call this if you create a new bh.
813  */
814 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
815 {
816         transaction_t *transaction = handle->h_transaction;
817         journal_t *journal = transaction->t_journal;
818         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
819         int err;
820
821         jbd_debug(5, "journal_head %p\n", jh);
822         err = -EROFS;
823         if (is_handle_aborted(handle))
824                 goto out;
825         err = 0;
826
827         JBUFFER_TRACE(jh, "entry");
828         /*
829          * The buffer may already belong to this transaction due to pre-zeroing
830          * in the filesystem's new_block code.  It may also be on the previous,
831          * committing transaction's lists, but it HAS to be in Forget state in
832          * that case: the transaction must have deleted the buffer for it to be
833          * reused here.
834          */
835         jbd_lock_bh_state(bh);
836         spin_lock(&journal->j_list_lock);
837         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
838                 jh->b_transaction == NULL ||
839                 (jh->b_transaction == journal->j_committing_transaction &&
840                           jh->b_jlist == BJ_Forget)));
841
842         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
843         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
844
845         if (jh->b_transaction == NULL) {
846                 jh->b_transaction = transaction;
847
848                 /* first access by this transaction */
849                 jh->b_modified = 0;
850
851                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
852                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
853         } else if (jh->b_transaction == journal->j_committing_transaction) {
854                 /* first access by this transaction */
855                 jh->b_modified = 0;
856
857                 JBUFFER_TRACE(jh, "set next transaction");
858                 jh->b_next_transaction = transaction;
859         }
860         spin_unlock(&journal->j_list_lock);
861         jbd_unlock_bh_state(bh);
862
863         /*
864          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
865          * blocks which contain freed but then revoked metadata.  We need
866          * to cancel the revoke in case we end up freeing it yet again
867          * and the reallocating as data - this would cause a second revoke,
868          * which hits an assertion error.
869          */
870         JBUFFER_TRACE(jh, "cancelling revoke");
871         jbd2_journal_cancel_revoke(handle, jh);
872         jbd2_journal_put_journal_head(jh);
873 out:
874         return err;
875 }
876
877 /**
878  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
879  *     non-rewindable consequences
880  * @handle: transaction
881  * @bh: buffer to undo
882  * @credits: store the number of taken credits here (if not NULL)
883  *
884  * Sometimes there is a need to distinguish between metadata which has
885  * been committed to disk and that which has not.  The ext3fs code uses
886  * this for freeing and allocating space, we have to make sure that we
887  * do not reuse freed space until the deallocation has been committed,
888  * since if we overwrote that space we would make the delete
889  * un-rewindable in case of a crash.
890  *
891  * To deal with that, jbd2_journal_get_undo_access requests write access to a
892  * buffer for parts of non-rewindable operations such as delete
893  * operations on the bitmaps.  The journaling code must keep a copy of
894  * the buffer's contents prior to the undo_access call until such time
895  * as we know that the buffer has definitely been committed to disk.
896  *
897  * We never need to know which transaction the committed data is part
898  * of, buffers touched here are guaranteed to be dirtied later and so
899  * will be committed to a new transaction in due course, at which point
900  * we can discard the old committed data pointer.
901  *
902  * Returns error number or 0 on success.
903  */
904 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
905 {
906         int err;
907         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
908         char *committed_data = NULL;
909
910         JBUFFER_TRACE(jh, "entry");
911
912         /*
913          * Do this first --- it can drop the journal lock, so we want to
914          * make sure that obtaining the committed_data is done
915          * atomically wrt. completion of any outstanding commits.
916          */
917         err = do_get_write_access(handle, jh, 1);
918         if (err)
919                 goto out;
920
921 repeat:
922         if (!jh->b_committed_data) {
923                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
924                 if (!committed_data) {
925                         printk(KERN_EMERG "%s: No memory for committed data\n",
926                                 __func__);
927                         err = -ENOMEM;
928                         goto out;
929                 }
930         }
931
932         jbd_lock_bh_state(bh);
933         if (!jh->b_committed_data) {
934                 /* Copy out the current buffer contents into the
935                  * preserved, committed copy. */
936                 JBUFFER_TRACE(jh, "generate b_committed data");
937                 if (!committed_data) {
938                         jbd_unlock_bh_state(bh);
939                         goto repeat;
940                 }
941
942                 jh->b_committed_data = committed_data;
943                 committed_data = NULL;
944                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
945         }
946         jbd_unlock_bh_state(bh);
947 out:
948         jbd2_journal_put_journal_head(jh);
949         if (unlikely(committed_data))
950                 jbd2_free(committed_data, bh->b_size);
951         return err;
952 }
953
954 /**
955  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
956  * @bh: buffer to trigger on
957  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
958  *
959  * Set any triggers on this journal_head.  This is always safe, because
960  * triggers for a committing buffer will be saved off, and triggers for
961  * a running transaction will match the buffer in that transaction.
962  *
963  * Call with NULL to clear the triggers.
964  */
965 void jbd2_journal_set_triggers(struct buffer_head *bh,
966                                struct jbd2_buffer_trigger_type *type)
967 {
968         struct journal_head *jh = bh2jh(bh);
969
970         jh->b_triggers = type;
971 }
972
973 void jbd2_buffer_commit_trigger(struct journal_head *jh, void *mapped_data,
974                                 struct jbd2_buffer_trigger_type *triggers)
975 {
976         struct buffer_head *bh = jh2bh(jh);
977
978         if (!triggers || !triggers->t_commit)
979                 return;
980
981         triggers->t_commit(triggers, bh, mapped_data, bh->b_size);
982 }
983
984 void jbd2_buffer_abort_trigger(struct journal_head *jh,
985                                struct jbd2_buffer_trigger_type *triggers)
986 {
987         if (!triggers || !triggers->t_abort)
988                 return;
989
990         triggers->t_abort(triggers, jh2bh(jh));
991 }
992
993
994
995 /**
996  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
997  * @handle: transaction to add buffer to.
998  * @bh: buffer to mark
999  *
1000  * mark dirty metadata which needs to be journaled as part of the current
1001  * transaction.
1002  *
1003  * The buffer is placed on the transaction's metadata list and is marked
1004  * as belonging to the transaction.
1005  *
1006  * Returns error number or 0 on success.
1007  *
1008  * Special care needs to be taken if the buffer already belongs to the
1009  * current committing transaction (in which case we should have frozen
1010  * data present for that commit).  In that case, we don't relink the
1011  * buffer: that only gets done when the old transaction finally
1012  * completes its commit.
1013  */
1014 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1015 {
1016         transaction_t *transaction = handle->h_transaction;
1017         journal_t *journal = transaction->t_journal;
1018         struct journal_head *jh = bh2jh(bh);
1019
1020         jbd_debug(5, "journal_head %p\n", jh);
1021         JBUFFER_TRACE(jh, "entry");
1022         if (is_handle_aborted(handle))
1023                 goto out;
1024
1025         jbd_lock_bh_state(bh);
1026
1027         if (jh->b_modified == 0) {
1028                 /*
1029                  * This buffer's got modified and becoming part
1030                  * of the transaction. This needs to be done
1031                  * once a transaction -bzzz
1032                  */
1033                 jh->b_modified = 1;
1034                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1035                 handle->h_buffer_credits--;
1036         }
1037
1038         /*
1039          * fastpath, to avoid expensive locking.  If this buffer is already
1040          * on the running transaction's metadata list there is nothing to do.
1041          * Nobody can take it off again because there is a handle open.
1042          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1043          * result in this test being false, so we go in and take the locks.
1044          */
1045         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1046                 JBUFFER_TRACE(jh, "fastpath");
1047                 J_ASSERT_JH(jh, jh->b_transaction ==
1048                                         journal->j_running_transaction);
1049                 goto out_unlock_bh;
1050         }
1051
1052         set_buffer_jbddirty(bh);
1053
1054         /*
1055          * Metadata already on the current transaction list doesn't
1056          * need to be filed.  Metadata on another transaction's list must
1057          * be committing, and will be refiled once the commit completes:
1058          * leave it alone for now.
1059          */
1060         if (jh->b_transaction != transaction) {
1061                 JBUFFER_TRACE(jh, "already on other transaction");
1062                 J_ASSERT_JH(jh, jh->b_transaction ==
1063                                         journal->j_committing_transaction);
1064                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1065                 /* And this case is illegal: we can't reuse another
1066                  * transaction's data buffer, ever. */
1067                 goto out_unlock_bh;
1068         }
1069
1070         /* That test should have eliminated the following case: */
1071         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1072
1073         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1074         spin_lock(&journal->j_list_lock);
1075         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1076         spin_unlock(&journal->j_list_lock);
1077 out_unlock_bh:
1078         jbd_unlock_bh_state(bh);
1079 out:
1080         JBUFFER_TRACE(jh, "exit");
1081         return 0;
1082 }
1083
1084 /*
1085  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1086  * updates, if the update decided in the end that it didn't need access.
1087  *
1088  */
1089 void
1090 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1091 {
1092         BUFFER_TRACE(bh, "entry");
1093 }
1094
1095 /**
1096  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1097  * @handle: transaction handle
1098  * @bh:     bh to 'forget'
1099  *
1100  * We can only do the bforget if there are no commits pending against the
1101  * buffer.  If the buffer is dirty in the current running transaction we
1102  * can safely unlink it.
1103  *
1104  * bh may not be a journalled buffer at all - it may be a non-JBD
1105  * buffer which came off the hashtable.  Check for this.
1106  *
1107  * Decrements bh->b_count by one.
1108  *
1109  * Allow this call even if the handle has aborted --- it may be part of
1110  * the caller's cleanup after an abort.
1111  */
1112 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1113 {
1114         transaction_t *transaction = handle->h_transaction;
1115         journal_t *journal = transaction->t_journal;
1116         struct journal_head *jh;
1117         int drop_reserve = 0;
1118         int err = 0;
1119         int was_modified = 0;
1120
1121         BUFFER_TRACE(bh, "entry");
1122
1123         jbd_lock_bh_state(bh);
1124         spin_lock(&journal->j_list_lock);
1125
1126         if (!buffer_jbd(bh))
1127                 goto not_jbd;
1128         jh = bh2jh(bh);
1129
1130         /* Critical error: attempting to delete a bitmap buffer, maybe?
1131          * Don't do any jbd operations, and return an error. */
1132         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1133                          "inconsistent data on disk")) {
1134                 err = -EIO;
1135                 goto not_jbd;
1136         }
1137
1138         /* keep track of wether or not this transaction modified us */
1139         was_modified = jh->b_modified;
1140
1141         /*
1142          * The buffer's going from the transaction, we must drop
1143          * all references -bzzz
1144          */
1145         jh->b_modified = 0;
1146
1147         if (jh->b_transaction == handle->h_transaction) {
1148                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1149
1150                 /* If we are forgetting a buffer which is already part
1151                  * of this transaction, then we can just drop it from
1152                  * the transaction immediately. */
1153                 clear_buffer_dirty(bh);
1154                 clear_buffer_jbddirty(bh);
1155
1156                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1157
1158                 /*
1159                  * we only want to drop a reference if this transaction
1160                  * modified the buffer
1161                  */
1162                 if (was_modified)
1163                         drop_reserve = 1;
1164
1165                 /*
1166                  * We are no longer going to journal this buffer.
1167                  * However, the commit of this transaction is still
1168                  * important to the buffer: the delete that we are now
1169                  * processing might obsolete an old log entry, so by
1170                  * committing, we can satisfy the buffer's checkpoint.
1171                  *
1172                  * So, if we have a checkpoint on the buffer, we should
1173                  * now refile the buffer on our BJ_Forget list so that
1174                  * we know to remove the checkpoint after we commit.
1175                  */
1176
1177                 if (jh->b_cp_transaction) {
1178                         __jbd2_journal_temp_unlink_buffer(jh);
1179                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1180                 } else {
1181                         __jbd2_journal_unfile_buffer(jh);
1182                         jbd2_journal_remove_journal_head(bh);
1183                         __brelse(bh);
1184                         if (!buffer_jbd(bh)) {
1185                                 spin_unlock(&journal->j_list_lock);
1186                                 jbd_unlock_bh_state(bh);
1187                                 __bforget(bh);
1188                                 goto drop;
1189                         }
1190                 }
1191         } else if (jh->b_transaction) {
1192                 J_ASSERT_JH(jh, (jh->b_transaction ==
1193                                  journal->j_committing_transaction));
1194                 /* However, if the buffer is still owned by a prior
1195                  * (committing) transaction, we can't drop it yet... */
1196                 JBUFFER_TRACE(jh, "belongs to older transaction");
1197                 /* ... but we CAN drop it from the new transaction if we
1198                  * have also modified it since the original commit. */
1199
1200                 if (jh->b_next_transaction) {
1201                         J_ASSERT(jh->b_next_transaction == transaction);
1202                         jh->b_next_transaction = NULL;
1203
1204                         /*
1205                          * only drop a reference if this transaction modified
1206                          * the buffer
1207                          */
1208                         if (was_modified)
1209                                 drop_reserve = 1;
1210                 }
1211         }
1212
1213 not_jbd:
1214         spin_unlock(&journal->j_list_lock);
1215         jbd_unlock_bh_state(bh);
1216         __brelse(bh);
1217 drop:
1218         if (drop_reserve) {
1219                 /* no need to reserve log space for this block -bzzz */
1220                 handle->h_buffer_credits++;
1221         }
1222         return err;
1223 }
1224
1225 /**
1226  * int jbd2_journal_stop() - complete a transaction
1227  * @handle: tranaction to complete.
1228  *
1229  * All done for a particular handle.
1230  *
1231  * There is not much action needed here.  We just return any remaining
1232  * buffer credits to the transaction and remove the handle.  The only
1233  * complication is that we need to start a commit operation if the
1234  * filesystem is marked for synchronous update.
1235  *
1236  * jbd2_journal_stop itself will not usually return an error, but it may
1237  * do so in unusual circumstances.  In particular, expect it to
1238  * return -EIO if a jbd2_journal_abort has been executed since the
1239  * transaction began.
1240  */
1241 int jbd2_journal_stop(handle_t *handle)
1242 {
1243         transaction_t *transaction = handle->h_transaction;
1244         journal_t *journal = transaction->t_journal;
1245         int err;
1246         pid_t pid;
1247
1248         J_ASSERT(journal_current_handle() == handle);
1249
1250         if (is_handle_aborted(handle))
1251                 err = -EIO;
1252         else {
1253                 J_ASSERT(transaction->t_updates > 0);
1254                 err = 0;
1255         }
1256
1257         if (--handle->h_ref > 0) {
1258                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1259                           handle->h_ref);
1260                 return err;
1261         }
1262
1263         jbd_debug(4, "Handle %p going down\n", handle);
1264
1265         /*
1266          * Implement synchronous transaction batching.  If the handle
1267          * was synchronous, don't force a commit immediately.  Let's
1268          * yield and let another thread piggyback onto this
1269          * transaction.  Keep doing that while new threads continue to
1270          * arrive.  It doesn't cost much - we're about to run a commit
1271          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1272          * operations by 30x or more...
1273          *
1274          * We try and optimize the sleep time against what the
1275          * underlying disk can do, instead of having a static sleep
1276          * time.  This is useful for the case where our storage is so
1277          * fast that it is more optimal to go ahead and force a flush
1278          * and wait for the transaction to be committed than it is to
1279          * wait for an arbitrary amount of time for new writers to
1280          * join the transaction.  We achieve this by measuring how
1281          * long it takes to commit a transaction, and compare it with
1282          * how long this transaction has been running, and if run time
1283          * < commit time then we sleep for the delta and commit.  This
1284          * greatly helps super fast disks that would see slowdowns as
1285          * more threads started doing fsyncs.
1286          *
1287          * But don't do this if this process was the most recent one
1288          * to perform a synchronous write.  We do this to detect the
1289          * case where a single process is doing a stream of sync
1290          * writes.  No point in waiting for joiners in that case.
1291          */
1292         pid = current->pid;
1293         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1294                 u64 commit_time, trans_time;
1295
1296                 journal->j_last_sync_writer = pid;
1297
1298                 spin_lock(&journal->j_state_lock);
1299                 commit_time = journal->j_average_commit_time;
1300                 spin_unlock(&journal->j_state_lock);
1301
1302                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1303                                                    transaction->t_start_time));
1304
1305                 commit_time = max_t(u64, commit_time,
1306                                     1000*journal->j_min_batch_time);
1307                 commit_time = min_t(u64, commit_time,
1308                                     1000*journal->j_max_batch_time);
1309
1310                 if (trans_time < commit_time) {
1311                         ktime_t expires = ktime_add_ns(ktime_get(),
1312                                                        commit_time);
1313                         set_current_state(TASK_UNINTERRUPTIBLE);
1314                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1315                 }
1316         }
1317
1318         current->journal_info = NULL;
1319         spin_lock(&journal->j_state_lock);
1320         spin_lock(&transaction->t_handle_lock);
1321         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1322         transaction->t_updates--;
1323         if (!transaction->t_updates) {
1324                 wake_up(&journal->j_wait_updates);
1325                 if (journal->j_barrier_count)
1326                         wake_up(&journal->j_wait_transaction_locked);
1327         }
1328
1329         /*
1330          * If the handle is marked SYNC, we need to set another commit
1331          * going!  We also want to force a commit if the current
1332          * transaction is occupying too much of the log, or if the
1333          * transaction is too old now.
1334          */
1335         if (handle->h_sync ||
1336                         transaction->t_outstanding_credits >
1337                                 journal->j_max_transaction_buffers ||
1338                         time_after_eq(jiffies, transaction->t_expires)) {
1339                 /* Do this even for aborted journals: an abort still
1340                  * completes the commit thread, it just doesn't write
1341                  * anything to disk. */
1342                 tid_t tid = transaction->t_tid;
1343
1344                 spin_unlock(&transaction->t_handle_lock);
1345                 jbd_debug(2, "transaction too old, requesting commit for "
1346                                         "handle %p\n", handle);
1347                 /* This is non-blocking */
1348                 __jbd2_log_start_commit(journal, transaction->t_tid);
1349                 spin_unlock(&journal->j_state_lock);
1350
1351                 /*
1352                  * Special case: JBD2_SYNC synchronous updates require us
1353                  * to wait for the commit to complete.
1354                  */
1355                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1356                         err = jbd2_log_wait_commit(journal, tid);
1357         } else {
1358                 spin_unlock(&transaction->t_handle_lock);
1359                 spin_unlock(&journal->j_state_lock);
1360         }
1361
1362         lock_map_release(&handle->h_lockdep_map);
1363
1364         jbd2_free_handle(handle);
1365         return err;
1366 }
1367
1368 /**
1369  * int jbd2_journal_force_commit() - force any uncommitted transactions
1370  * @journal: journal to force
1371  *
1372  * For synchronous operations: force any uncommitted transactions
1373  * to disk.  May seem kludgy, but it reuses all the handle batching
1374  * code in a very simple manner.
1375  */
1376 int jbd2_journal_force_commit(journal_t *journal)
1377 {
1378         handle_t *handle;
1379         int ret;
1380
1381         handle = jbd2_journal_start(journal, 1);
1382         if (IS_ERR(handle)) {
1383                 ret = PTR_ERR(handle);
1384         } else {
1385                 handle->h_sync = 1;
1386                 ret = jbd2_journal_stop(handle);
1387         }
1388         return ret;
1389 }
1390
1391 /*
1392  *
1393  * List management code snippets: various functions for manipulating the
1394  * transaction buffer lists.
1395  *
1396  */
1397
1398 /*
1399  * Append a buffer to a transaction list, given the transaction's list head
1400  * pointer.
1401  *
1402  * j_list_lock is held.
1403  *
1404  * jbd_lock_bh_state(jh2bh(jh)) is held.
1405  */
1406
1407 static inline void
1408 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1409 {
1410         if (!*list) {
1411                 jh->b_tnext = jh->b_tprev = jh;
1412                 *list = jh;
1413         } else {
1414                 /* Insert at the tail of the list to preserve order */
1415                 struct journal_head *first = *list, *last = first->b_tprev;
1416                 jh->b_tprev = last;
1417                 jh->b_tnext = first;
1418                 last->b_tnext = first->b_tprev = jh;
1419         }
1420 }
1421
1422 /*
1423  * Remove a buffer from a transaction list, given the transaction's list
1424  * head pointer.
1425  *
1426  * Called with j_list_lock held, and the journal may not be locked.
1427  *
1428  * jbd_lock_bh_state(jh2bh(jh)) is held.
1429  */
1430
1431 static inline void
1432 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1433 {
1434         if (*list == jh) {
1435                 *list = jh->b_tnext;
1436                 if (*list == jh)
1437                         *list = NULL;
1438         }
1439         jh->b_tprev->b_tnext = jh->b_tnext;
1440         jh->b_tnext->b_tprev = jh->b_tprev;
1441 }
1442
1443 /*
1444  * Remove a buffer from the appropriate transaction list.
1445  *
1446  * Note that this function can *change* the value of
1447  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1448  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1449  * of these pointers, it could go bad.  Generally the caller needs to re-read
1450  * the pointer from the transaction_t.
1451  *
1452  * Called under j_list_lock.  The journal may not be locked.
1453  */
1454 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1455 {
1456         struct journal_head **list = NULL;
1457         transaction_t *transaction;
1458         struct buffer_head *bh = jh2bh(jh);
1459
1460         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1461         transaction = jh->b_transaction;
1462         if (transaction)
1463                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1464
1465         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1466         if (jh->b_jlist != BJ_None)
1467                 J_ASSERT_JH(jh, transaction != NULL);
1468
1469         switch (jh->b_jlist) {
1470         case BJ_None:
1471                 return;
1472         case BJ_Metadata:
1473                 transaction->t_nr_buffers--;
1474                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1475                 list = &transaction->t_buffers;
1476                 break;
1477         case BJ_Forget:
1478                 list = &transaction->t_forget;
1479                 break;
1480         case BJ_IO:
1481                 list = &transaction->t_iobuf_list;
1482                 break;
1483         case BJ_Shadow:
1484                 list = &transaction->t_shadow_list;
1485                 break;
1486         case BJ_LogCtl:
1487                 list = &transaction->t_log_list;
1488                 break;
1489         case BJ_Reserved:
1490                 list = &transaction->t_reserved_list;
1491                 break;
1492         }
1493
1494         __blist_del_buffer(list, jh);
1495         jh->b_jlist = BJ_None;
1496         if (test_clear_buffer_jbddirty(bh))
1497                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1498 }
1499
1500 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1501 {
1502         __jbd2_journal_temp_unlink_buffer(jh);
1503         jh->b_transaction = NULL;
1504 }
1505
1506 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1507 {
1508         jbd_lock_bh_state(jh2bh(jh));
1509         spin_lock(&journal->j_list_lock);
1510         __jbd2_journal_unfile_buffer(jh);
1511         spin_unlock(&journal->j_list_lock);
1512         jbd_unlock_bh_state(jh2bh(jh));
1513 }
1514
1515 /*
1516  * Called from jbd2_journal_try_to_free_buffers().
1517  *
1518  * Called under jbd_lock_bh_state(bh)
1519  */
1520 static void
1521 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1522 {
1523         struct journal_head *jh;
1524
1525         jh = bh2jh(bh);
1526
1527         if (buffer_locked(bh) || buffer_dirty(bh))
1528                 goto out;
1529
1530         if (jh->b_next_transaction != NULL)
1531                 goto out;
1532
1533         spin_lock(&journal->j_list_lock);
1534         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1535                 /* written-back checkpointed metadata buffer */
1536                 if (jh->b_jlist == BJ_None) {
1537                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1538                         __jbd2_journal_remove_checkpoint(jh);
1539                         jbd2_journal_remove_journal_head(bh);
1540                         __brelse(bh);
1541                 }
1542         }
1543         spin_unlock(&journal->j_list_lock);
1544 out:
1545         return;
1546 }
1547
1548 /*
1549  * jbd2_journal_try_to_free_buffers() could race with
1550  * jbd2_journal_commit_transaction(). The later might still hold the
1551  * reference count to the buffers when inspecting them on
1552  * t_syncdata_list or t_locked_list.
1553  *
1554  * jbd2_journal_try_to_free_buffers() will call this function to
1555  * wait for the current transaction to finish syncing data buffers, before
1556  * try to free that buffer.
1557  *
1558  * Called with journal->j_state_lock hold.
1559  */
1560 static void jbd2_journal_wait_for_transaction_sync_data(journal_t *journal)
1561 {
1562         transaction_t *transaction;
1563         tid_t tid;
1564
1565         spin_lock(&journal->j_state_lock);
1566         transaction = journal->j_committing_transaction;
1567
1568         if (!transaction) {
1569                 spin_unlock(&journal->j_state_lock);
1570                 return;
1571         }
1572
1573         tid = transaction->t_tid;
1574         spin_unlock(&journal->j_state_lock);
1575         jbd2_log_wait_commit(journal, tid);
1576 }
1577
1578 /**
1579  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1580  * @journal: journal for operation
1581  * @page: to try and free
1582  * @gfp_mask: we use the mask to detect how hard should we try to release
1583  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1584  * release the buffers.
1585  *
1586  *
1587  * For all the buffers on this page,
1588  * if they are fully written out ordered data, move them onto BUF_CLEAN
1589  * so try_to_free_buffers() can reap them.
1590  *
1591  * This function returns non-zero if we wish try_to_free_buffers()
1592  * to be called. We do this if the page is releasable by try_to_free_buffers().
1593  * We also do it if the page has locked or dirty buffers and the caller wants
1594  * us to perform sync or async writeout.
1595  *
1596  * This complicates JBD locking somewhat.  We aren't protected by the
1597  * BKL here.  We wish to remove the buffer from its committing or
1598  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1599  *
1600  * This may *change* the value of transaction_t->t_datalist, so anyone
1601  * who looks at t_datalist needs to lock against this function.
1602  *
1603  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1604  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1605  * will come out of the lock with the buffer dirty, which makes it
1606  * ineligible for release here.
1607  *
1608  * Who else is affected by this?  hmm...  Really the only contender
1609  * is do_get_write_access() - it could be looking at the buffer while
1610  * journal_try_to_free_buffer() is changing its state.  But that
1611  * cannot happen because we never reallocate freed data as metadata
1612  * while the data is part of a transaction.  Yes?
1613  *
1614  * Return 0 on failure, 1 on success
1615  */
1616 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1617                                 struct page *page, gfp_t gfp_mask)
1618 {
1619         struct buffer_head *head;
1620         struct buffer_head *bh;
1621         int ret = 0;
1622
1623         J_ASSERT(PageLocked(page));
1624
1625         head = page_buffers(page);
1626         bh = head;
1627         do {
1628                 struct journal_head *jh;
1629
1630                 /*
1631                  * We take our own ref against the journal_head here to avoid
1632                  * having to add tons of locking around each instance of
1633                  * jbd2_journal_remove_journal_head() and
1634                  * jbd2_journal_put_journal_head().
1635                  */
1636                 jh = jbd2_journal_grab_journal_head(bh);
1637                 if (!jh)
1638                         continue;
1639
1640                 jbd_lock_bh_state(bh);
1641                 __journal_try_to_free_buffer(journal, bh);
1642                 jbd2_journal_put_journal_head(jh);
1643                 jbd_unlock_bh_state(bh);
1644                 if (buffer_jbd(bh))
1645                         goto busy;
1646         } while ((bh = bh->b_this_page) != head);
1647
1648         ret = try_to_free_buffers(page);
1649
1650         /*
1651          * There are a number of places where jbd2_journal_try_to_free_buffers()
1652          * could race with jbd2_journal_commit_transaction(), the later still
1653          * holds the reference to the buffers to free while processing them.
1654          * try_to_free_buffers() failed to free those buffers. Some of the
1655          * caller of releasepage() request page buffers to be dropped, otherwise
1656          * treat the fail-to-free as errors (such as generic_file_direct_IO())
1657          *
1658          * So, if the caller of try_to_release_page() wants the synchronous
1659          * behaviour(i.e make sure buffers are dropped upon return),
1660          * let's wait for the current transaction to finish flush of
1661          * dirty data buffers, then try to free those buffers again,
1662          * with the journal locked.
1663          */
1664         if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
1665                 jbd2_journal_wait_for_transaction_sync_data(journal);
1666                 ret = try_to_free_buffers(page);
1667         }
1668
1669 busy:
1670         return ret;
1671 }
1672
1673 /*
1674  * This buffer is no longer needed.  If it is on an older transaction's
1675  * checkpoint list we need to record it on this transaction's forget list
1676  * to pin this buffer (and hence its checkpointing transaction) down until
1677  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1678  * release it.
1679  * Returns non-zero if JBD no longer has an interest in the buffer.
1680  *
1681  * Called under j_list_lock.
1682  *
1683  * Called under jbd_lock_bh_state(bh).
1684  */
1685 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1686 {
1687         int may_free = 1;
1688         struct buffer_head *bh = jh2bh(jh);
1689
1690         __jbd2_journal_unfile_buffer(jh);
1691
1692         if (jh->b_cp_transaction) {
1693                 JBUFFER_TRACE(jh, "on running+cp transaction");
1694                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1695                 clear_buffer_jbddirty(bh);
1696                 may_free = 0;
1697         } else {
1698                 JBUFFER_TRACE(jh, "on running transaction");
1699                 jbd2_journal_remove_journal_head(bh);
1700                 __brelse(bh);
1701         }
1702         return may_free;
1703 }
1704
1705 /*
1706  * jbd2_journal_invalidatepage
1707  *
1708  * This code is tricky.  It has a number of cases to deal with.
1709  *
1710  * There are two invariants which this code relies on:
1711  *
1712  * i_size must be updated on disk before we start calling invalidatepage on the
1713  * data.
1714  *
1715  *  This is done in ext3 by defining an ext3_setattr method which
1716  *  updates i_size before truncate gets going.  By maintaining this
1717  *  invariant, we can be sure that it is safe to throw away any buffers
1718  *  attached to the current transaction: once the transaction commits,
1719  *  we know that the data will not be needed.
1720  *
1721  *  Note however that we can *not* throw away data belonging to the
1722  *  previous, committing transaction!
1723  *
1724  * Any disk blocks which *are* part of the previous, committing
1725  * transaction (and which therefore cannot be discarded immediately) are
1726  * not going to be reused in the new running transaction
1727  *
1728  *  The bitmap committed_data images guarantee this: any block which is
1729  *  allocated in one transaction and removed in the next will be marked
1730  *  as in-use in the committed_data bitmap, so cannot be reused until
1731  *  the next transaction to delete the block commits.  This means that
1732  *  leaving committing buffers dirty is quite safe: the disk blocks
1733  *  cannot be reallocated to a different file and so buffer aliasing is
1734  *  not possible.
1735  *
1736  *
1737  * The above applies mainly to ordered data mode.  In writeback mode we
1738  * don't make guarantees about the order in which data hits disk --- in
1739  * particular we don't guarantee that new dirty data is flushed before
1740  * transaction commit --- so it is always safe just to discard data
1741  * immediately in that mode.  --sct
1742  */
1743
1744 /*
1745  * The journal_unmap_buffer helper function returns zero if the buffer
1746  * concerned remains pinned as an anonymous buffer belonging to an older
1747  * transaction.
1748  *
1749  * We're outside-transaction here.  Either or both of j_running_transaction
1750  * and j_committing_transaction may be NULL.
1751  */
1752 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1753 {
1754         transaction_t *transaction;
1755         struct journal_head *jh;
1756         int may_free = 1;
1757         int ret;
1758
1759         BUFFER_TRACE(bh, "entry");
1760
1761         /*
1762          * It is safe to proceed here without the j_list_lock because the
1763          * buffers cannot be stolen by try_to_free_buffers as long as we are
1764          * holding the page lock. --sct
1765          */
1766
1767         if (!buffer_jbd(bh))
1768                 goto zap_buffer_unlocked;
1769
1770         /* OK, we have data buffer in journaled mode */
1771         spin_lock(&journal->j_state_lock);
1772         jbd_lock_bh_state(bh);
1773         spin_lock(&journal->j_list_lock);
1774
1775         jh = jbd2_journal_grab_journal_head(bh);
1776         if (!jh)
1777                 goto zap_buffer_no_jh;
1778
1779         transaction = jh->b_transaction;
1780         if (transaction == NULL) {
1781                 /* First case: not on any transaction.  If it
1782                  * has no checkpoint link, then we can zap it:
1783                  * it's a writeback-mode buffer so we don't care
1784                  * if it hits disk safely. */
1785                 if (!jh->b_cp_transaction) {
1786                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1787                         goto zap_buffer;
1788                 }
1789
1790                 if (!buffer_dirty(bh)) {
1791                         /* bdflush has written it.  We can drop it now */
1792                         goto zap_buffer;
1793                 }
1794
1795                 /* OK, it must be in the journal but still not
1796                  * written fully to disk: it's metadata or
1797                  * journaled data... */
1798
1799                 if (journal->j_running_transaction) {
1800                         /* ... and once the current transaction has
1801                          * committed, the buffer won't be needed any
1802                          * longer. */
1803                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1804                         ret = __dispose_buffer(jh,
1805                                         journal->j_running_transaction);
1806                         jbd2_journal_put_journal_head(jh);
1807                         spin_unlock(&journal->j_list_lock);
1808                         jbd_unlock_bh_state(bh);
1809                         spin_unlock(&journal->j_state_lock);
1810                         return ret;
1811                 } else {
1812                         /* There is no currently-running transaction. So the
1813                          * orphan record which we wrote for this file must have
1814                          * passed into commit.  We must attach this buffer to
1815                          * the committing transaction, if it exists. */
1816                         if (journal->j_committing_transaction) {
1817                                 JBUFFER_TRACE(jh, "give to committing trans");
1818                                 ret = __dispose_buffer(jh,
1819                                         journal->j_committing_transaction);
1820                                 jbd2_journal_put_journal_head(jh);
1821                                 spin_unlock(&journal->j_list_lock);
1822                                 jbd_unlock_bh_state(bh);
1823                                 spin_unlock(&journal->j_state_lock);
1824                                 return ret;
1825                         } else {
1826                                 /* The orphan record's transaction has
1827                                  * committed.  We can cleanse this buffer */
1828                                 clear_buffer_jbddirty(bh);
1829                                 goto zap_buffer;
1830                         }
1831                 }
1832         } else if (transaction == journal->j_committing_transaction) {
1833                 JBUFFER_TRACE(jh, "on committing transaction");
1834                 /*
1835                  * If it is committing, we simply cannot touch it.  We
1836                  * can remove it's next_transaction pointer from the
1837                  * running transaction if that is set, but nothing
1838                  * else. */
1839                 set_buffer_freed(bh);
1840                 if (jh->b_next_transaction) {
1841                         J_ASSERT(jh->b_next_transaction ==
1842                                         journal->j_running_transaction);
1843                         jh->b_next_transaction = NULL;
1844                 }
1845                 jbd2_journal_put_journal_head(jh);
1846                 spin_unlock(&journal->j_list_lock);
1847                 jbd_unlock_bh_state(bh);
1848                 spin_unlock(&journal->j_state_lock);
1849                 return 0;
1850         } else {
1851                 /* Good, the buffer belongs to the running transaction.
1852                  * We are writing our own transaction's data, not any
1853                  * previous one's, so it is safe to throw it away
1854                  * (remember that we expect the filesystem to have set
1855                  * i_size already for this truncate so recovery will not
1856                  * expose the disk blocks we are discarding here.) */
1857                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1858                 JBUFFER_TRACE(jh, "on running transaction");
1859                 may_free = __dispose_buffer(jh, transaction);
1860         }
1861
1862 zap_buffer:
1863         jbd2_journal_put_journal_head(jh);
1864 zap_buffer_no_jh:
1865         spin_unlock(&journal->j_list_lock);
1866         jbd_unlock_bh_state(bh);
1867         spin_unlock(&journal->j_state_lock);
1868 zap_buffer_unlocked:
1869         clear_buffer_dirty(bh);
1870         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1871         clear_buffer_mapped(bh);
1872         clear_buffer_req(bh);
1873         clear_buffer_new(bh);
1874         bh->b_bdev = NULL;
1875         return may_free;
1876 }
1877
1878 /**
1879  * void jbd2_journal_invalidatepage()
1880  * @journal: journal to use for flush...
1881  * @page:    page to flush
1882  * @offset:  length of page to invalidate.
1883  *
1884  * Reap page buffers containing data after offset in page.
1885  *
1886  */
1887 void jbd2_journal_invalidatepage(journal_t *journal,
1888                       struct page *page,
1889                       unsigned long offset)
1890 {
1891         struct buffer_head *head, *bh, *next;
1892         unsigned int curr_off = 0;
1893         int may_free = 1;
1894
1895         if (!PageLocked(page))
1896                 BUG();
1897         if (!page_has_buffers(page))
1898                 return;
1899
1900         /* We will potentially be playing with lists other than just the
1901          * data lists (especially for journaled data mode), so be
1902          * cautious in our locking. */
1903
1904         head = bh = page_buffers(page);
1905         do {
1906                 unsigned int next_off = curr_off + bh->b_size;
1907                 next = bh->b_this_page;
1908
1909                 if (offset <= curr_off) {
1910                         /* This block is wholly outside the truncation point */
1911                         lock_buffer(bh);
1912                         may_free &= journal_unmap_buffer(journal, bh);
1913                         unlock_buffer(bh);
1914                 }
1915                 curr_off = next_off;
1916                 bh = next;
1917
1918         } while (bh != head);
1919
1920         if (!offset) {
1921                 if (may_free && try_to_free_buffers(page))
1922                         J_ASSERT(!page_has_buffers(page));
1923         }
1924 }
1925
1926 /*
1927  * File a buffer on the given transaction list.
1928  */
1929 void __jbd2_journal_file_buffer(struct journal_head *jh,
1930                         transaction_t *transaction, int jlist)
1931 {
1932         struct journal_head **list = NULL;
1933         int was_dirty = 0;
1934         struct buffer_head *bh = jh2bh(jh);
1935
1936         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1937         assert_spin_locked(&transaction->t_journal->j_list_lock);
1938
1939         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1940         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1941                                 jh->b_transaction == NULL);
1942
1943         if (jh->b_transaction && jh->b_jlist == jlist)
1944                 return;
1945
1946         /* The following list of buffer states needs to be consistent
1947          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1948          * state. */
1949
1950         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1951             jlist == BJ_Shadow || jlist == BJ_Forget) {
1952                 if (test_clear_buffer_dirty(bh) ||
1953                     test_clear_buffer_jbddirty(bh))
1954                         was_dirty = 1;
1955         }
1956
1957         if (jh->b_transaction)
1958                 __jbd2_journal_temp_unlink_buffer(jh);
1959         jh->b_transaction = transaction;
1960
1961         switch (jlist) {
1962         case BJ_None:
1963                 J_ASSERT_JH(jh, !jh->b_committed_data);
1964                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1965                 return;
1966         case BJ_Metadata:
1967                 transaction->t_nr_buffers++;
1968                 list = &transaction->t_buffers;
1969                 break;
1970         case BJ_Forget:
1971                 list = &transaction->t_forget;
1972                 break;
1973         case BJ_IO:
1974                 list = &transaction->t_iobuf_list;
1975                 break;
1976         case BJ_Shadow:
1977                 list = &transaction->t_shadow_list;
1978                 break;
1979         case BJ_LogCtl:
1980                 list = &transaction->t_log_list;
1981                 break;
1982         case BJ_Reserved:
1983                 list = &transaction->t_reserved_list;
1984                 break;
1985         }
1986
1987         __blist_add_buffer(list, jh);
1988         jh->b_jlist = jlist;
1989
1990         if (was_dirty)
1991                 set_buffer_jbddirty(bh);
1992 }
1993
1994 void jbd2_journal_file_buffer(struct journal_head *jh,
1995                                 transaction_t *transaction, int jlist)
1996 {
1997         jbd_lock_bh_state(jh2bh(jh));
1998         spin_lock(&transaction->t_journal->j_list_lock);
1999         __jbd2_journal_file_buffer(jh, transaction, jlist);
2000         spin_unlock(&transaction->t_journal->j_list_lock);
2001         jbd_unlock_bh_state(jh2bh(jh));
2002 }
2003
2004 /*
2005  * Remove a buffer from its current buffer list in preparation for
2006  * dropping it from its current transaction entirely.  If the buffer has
2007  * already started to be used by a subsequent transaction, refile the
2008  * buffer on that transaction's metadata list.
2009  *
2010  * Called under journal->j_list_lock
2011  *
2012  * Called under jbd_lock_bh_state(jh2bh(jh))
2013  */
2014 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2015 {
2016         int was_dirty;
2017         struct buffer_head *bh = jh2bh(jh);
2018
2019         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2020         if (jh->b_transaction)
2021                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2022
2023         /* If the buffer is now unused, just drop it. */
2024         if (jh->b_next_transaction == NULL) {
2025                 __jbd2_journal_unfile_buffer(jh);
2026                 return;
2027         }
2028
2029         /*
2030          * It has been modified by a later transaction: add it to the new
2031          * transaction's metadata list.
2032          */
2033
2034         was_dirty = test_clear_buffer_jbddirty(bh);
2035         __jbd2_journal_temp_unlink_buffer(jh);
2036         jh->b_transaction = jh->b_next_transaction;
2037         jh->b_next_transaction = NULL;
2038         __jbd2_journal_file_buffer(jh, jh->b_transaction,
2039                                 jh->b_modified ? BJ_Metadata : BJ_Reserved);
2040         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2041
2042         if (was_dirty)
2043                 set_buffer_jbddirty(bh);
2044 }
2045
2046 /*
2047  * For the unlocked version of this call, also make sure that any
2048  * hanging journal_head is cleaned up if necessary.
2049  *
2050  * __jbd2_journal_refile_buffer is usually called as part of a single locked
2051  * operation on a buffer_head, in which the caller is probably going to
2052  * be hooking the journal_head onto other lists.  In that case it is up
2053  * to the caller to remove the journal_head if necessary.  For the
2054  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2055  * doing anything else to the buffer so we need to do the cleanup
2056  * ourselves to avoid a jh leak.
2057  *
2058  * *** The journal_head may be freed by this call! ***
2059  */
2060 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2061 {
2062         struct buffer_head *bh = jh2bh(jh);
2063
2064         jbd_lock_bh_state(bh);
2065         spin_lock(&journal->j_list_lock);
2066
2067         __jbd2_journal_refile_buffer(jh);
2068         jbd_unlock_bh_state(bh);
2069         jbd2_journal_remove_journal_head(bh);
2070
2071         spin_unlock(&journal->j_list_lock);
2072         __brelse(bh);
2073 }
2074
2075 /*
2076  * File inode in the inode list of the handle's transaction
2077  */
2078 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2079 {
2080         transaction_t *transaction = handle->h_transaction;
2081         journal_t *journal = transaction->t_journal;
2082
2083         if (is_handle_aborted(handle))
2084                 return -EIO;
2085
2086         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2087                         transaction->t_tid);
2088
2089         /*
2090          * First check whether inode isn't already on the transaction's
2091          * lists without taking the lock. Note that this check is safe
2092          * without the lock as we cannot race with somebody removing inode
2093          * from the transaction. The reason is that we remove inode from the
2094          * transaction only in journal_release_jbd_inode() and when we commit
2095          * the transaction. We are guarded from the first case by holding
2096          * a reference to the inode. We are safe against the second case
2097          * because if jinode->i_transaction == transaction, commit code
2098          * cannot touch the transaction because we hold reference to it,
2099          * and if jinode->i_next_transaction == transaction, commit code
2100          * will only file the inode where we want it.
2101          */
2102         if (jinode->i_transaction == transaction ||
2103             jinode->i_next_transaction == transaction)
2104                 return 0;
2105
2106         spin_lock(&journal->j_list_lock);
2107
2108         if (jinode->i_transaction == transaction ||
2109             jinode->i_next_transaction == transaction)
2110                 goto done;
2111
2112         /* On some different transaction's list - should be
2113          * the committing one */
2114         if (jinode->i_transaction) {
2115                 J_ASSERT(jinode->i_next_transaction == NULL);
2116                 J_ASSERT(jinode->i_transaction ==
2117                                         journal->j_committing_transaction);
2118                 jinode->i_next_transaction = transaction;
2119                 goto done;
2120         }
2121         /* Not on any transaction list... */
2122         J_ASSERT(!jinode->i_next_transaction);
2123         jinode->i_transaction = transaction;
2124         list_add(&jinode->i_list, &transaction->t_inode_list);
2125 done:
2126         spin_unlock(&journal->j_list_lock);
2127
2128         return 0;
2129 }
2130
2131 /*
2132  * File truncate and transaction commit interact with each other in a
2133  * non-trivial way.  If a transaction writing data block A is
2134  * committing, we cannot discard the data by truncate until we have
2135  * written them.  Otherwise if we crashed after the transaction with
2136  * write has committed but before the transaction with truncate has
2137  * committed, we could see stale data in block A.  This function is a
2138  * helper to solve this problem.  It starts writeout of the truncated
2139  * part in case it is in the committing transaction.
2140  *
2141  * Filesystem code must call this function when inode is journaled in
2142  * ordered mode before truncation happens and after the inode has been
2143  * placed on orphan list with the new inode size. The second condition
2144  * avoids the race that someone writes new data and we start
2145  * committing the transaction after this function has been called but
2146  * before a transaction for truncate is started (and furthermore it
2147  * allows us to optimize the case where the addition to orphan list
2148  * happens in the same transaction as write --- we don't have to write
2149  * any data in such case).
2150  */
2151 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2152                                         struct jbd2_inode *jinode,
2153                                         loff_t new_size)
2154 {
2155         transaction_t *inode_trans, *commit_trans;
2156         int ret = 0;
2157
2158         /* This is a quick check to avoid locking if not necessary */
2159         if (!jinode->i_transaction)
2160                 goto out;
2161         /* Locks are here just to force reading of recent values, it is
2162          * enough that the transaction was not committing before we started
2163          * a transaction adding the inode to orphan list */
2164         spin_lock(&journal->j_state_lock);
2165         commit_trans = journal->j_committing_transaction;
2166         spin_unlock(&journal->j_state_lock);
2167         spin_lock(&journal->j_list_lock);
2168         inode_trans = jinode->i_transaction;
2169         spin_unlock(&journal->j_list_lock);
2170         if (inode_trans == commit_trans) {
2171                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2172                         new_size, LLONG_MAX);
2173                 if (ret)
2174                         jbd2_journal_abort(journal, ret);
2175         }
2176 out:
2177         return ret;
2178 }