Merge branch 'x86/cpu' into x86/x2apic
[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
29 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
30
31 /*
32  * jbd2_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  */
45
46 static transaction_t *
47 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
48 {
49         transaction->t_journal = journal;
50         transaction->t_state = T_RUNNING;
51         transaction->t_tid = journal->j_transaction_sequence++;
52         transaction->t_expires = jiffies + journal->j_commit_interval;
53         spin_lock_init(&transaction->t_handle_lock);
54         INIT_LIST_HEAD(&transaction->t_inode_list);
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         transaction->t_max_wait = 0;
63         transaction->t_start = jiffies;
64
65         return transaction;
66 }
67
68 /*
69  * Handle management.
70  *
71  * A handle_t is an object which represents a single atomic update to a
72  * filesystem, and which tracks all of the modifications which form part
73  * of that one update.
74  */
75
76 /*
77  * start_this_handle: Given a handle, deal with any locking or stalling
78  * needed to make sure that there is enough journal space for the handle
79  * to begin.  Attach the handle to a transaction and set up the
80  * transaction's buffer credits.
81  */
82
83 static int start_this_handle(journal_t *journal, handle_t *handle)
84 {
85         transaction_t *transaction;
86         int needed;
87         int nblocks = handle->h_buffer_credits;
88         transaction_t *new_transaction = NULL;
89         int ret = 0;
90         unsigned long ts = jiffies;
91
92         if (nblocks > journal->j_max_transaction_buffers) {
93                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
94                        current->comm, nblocks,
95                        journal->j_max_transaction_buffers);
96                 ret = -ENOSPC;
97                 goto out;
98         }
99
100 alloc_transaction:
101         if (!journal->j_running_transaction) {
102                 new_transaction = kzalloc(sizeof(*new_transaction),
103                                                 GFP_NOFS|__GFP_NOFAIL);
104                 if (!new_transaction) {
105                         ret = -ENOMEM;
106                         goto out;
107                 }
108         }
109
110         jbd_debug(3, "New handle %p going live.\n", handle);
111
112 repeat:
113
114         /*
115          * We need to hold j_state_lock until t_updates has been incremented,
116          * for proper journal barrier handling
117          */
118         spin_lock(&journal->j_state_lock);
119 repeat_locked:
120         if (is_journal_aborted(journal) ||
121             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
122                 spin_unlock(&journal->j_state_lock);
123                 ret = -EROFS;
124                 goto out;
125         }
126
127         /* Wait on the journal's transaction barrier if necessary */
128         if (journal->j_barrier_count) {
129                 spin_unlock(&journal->j_state_lock);
130                 wait_event(journal->j_wait_transaction_locked,
131                                 journal->j_barrier_count == 0);
132                 goto repeat;
133         }
134
135         if (!journal->j_running_transaction) {
136                 if (!new_transaction) {
137                         spin_unlock(&journal->j_state_lock);
138                         goto alloc_transaction;
139                 }
140                 jbd2_get_transaction(journal, new_transaction);
141                 new_transaction = NULL;
142         }
143
144         transaction = journal->j_running_transaction;
145
146         /*
147          * If the current transaction is locked down for commit, wait for the
148          * lock to be released.
149          */
150         if (transaction->t_state == T_LOCKED) {
151                 DEFINE_WAIT(wait);
152
153                 prepare_to_wait(&journal->j_wait_transaction_locked,
154                                         &wait, TASK_UNINTERRUPTIBLE);
155                 spin_unlock(&journal->j_state_lock);
156                 schedule();
157                 finish_wait(&journal->j_wait_transaction_locked, &wait);
158                 goto repeat;
159         }
160
161         /*
162          * If there is not enough space left in the log to write all potential
163          * buffers requested by this operation, we need to stall pending a log
164          * checkpoint to free some more log space.
165          */
166         spin_lock(&transaction->t_handle_lock);
167         needed = transaction->t_outstanding_credits + nblocks;
168
169         if (needed > journal->j_max_transaction_buffers) {
170                 /*
171                  * If the current transaction is already too large, then start
172                  * to commit it: we can then go back and attach this handle to
173                  * a new transaction.
174                  */
175                 DEFINE_WAIT(wait);
176
177                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
178                 spin_unlock(&transaction->t_handle_lock);
179                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
180                                 TASK_UNINTERRUPTIBLE);
181                 __jbd2_log_start_commit(journal, transaction->t_tid);
182                 spin_unlock(&journal->j_state_lock);
183                 schedule();
184                 finish_wait(&journal->j_wait_transaction_locked, &wait);
185                 goto repeat;
186         }
187
188         /*
189          * The commit code assumes that it can get enough log space
190          * without forcing a checkpoint.  This is *critical* for
191          * correctness: a checkpoint of a buffer which is also
192          * associated with a committing transaction creates a deadlock,
193          * so commit simply cannot force through checkpoints.
194          *
195          * We must therefore ensure the necessary space in the journal
196          * *before* starting to dirty potentially checkpointed buffers
197          * in the new transaction.
198          *
199          * The worst part is, any transaction currently committing can
200          * reduce the free space arbitrarily.  Be careful to account for
201          * those buffers when checkpointing.
202          */
203
204         /*
205          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
206          * a _lot_ of headroom: 1/4 of the journal plus the size of
207          * the committing transaction.  Really, we only need to give it
208          * committing_transaction->t_outstanding_credits plus "enough" for
209          * the log control blocks.
210          * Also, this test is inconsitent with the matching one in
211          * jbd2_journal_extend().
212          */
213         if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
214                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
215                 spin_unlock(&transaction->t_handle_lock);
216                 __jbd2_log_wait_for_space(journal);
217                 goto repeat_locked;
218         }
219
220         /* OK, account for the buffers that this operation expects to
221          * use and add the handle to the running transaction. */
222
223         if (time_after(transaction->t_start, ts)) {
224                 ts = jbd2_time_diff(ts, transaction->t_start);
225                 if (ts > transaction->t_max_wait)
226                         transaction->t_max_wait = ts;
227         }
228
229         handle->h_transaction = transaction;
230         transaction->t_outstanding_credits += nblocks;
231         transaction->t_updates++;
232         transaction->t_handle_count++;
233         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
234                   handle, nblocks, transaction->t_outstanding_credits,
235                   __jbd2_log_space_left(journal));
236         spin_unlock(&transaction->t_handle_lock);
237         spin_unlock(&journal->j_state_lock);
238 out:
239         if (unlikely(new_transaction))          /* It's usually NULL */
240                 kfree(new_transaction);
241         return ret;
242 }
243
244 static struct lock_class_key jbd2_handle_key;
245
246 /* Allocate a new handle.  This should probably be in a slab... */
247 static handle_t *new_handle(int nblocks)
248 {
249         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
250         if (!handle)
251                 return NULL;
252         memset(handle, 0, sizeof(*handle));
253         handle->h_buffer_credits = nblocks;
254         handle->h_ref = 1;
255
256         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
257                                                 &jbd2_handle_key, 0);
258
259         return handle;
260 }
261
262 /**
263  * handle_t *jbd2_journal_start() - Obtain a new handle.
264  * @journal: Journal to start transaction on.
265  * @nblocks: number of block buffer we might modify
266  *
267  * We make sure that the transaction can guarantee at least nblocks of
268  * modified buffers in the log.  We block until the log can guarantee
269  * that much space.
270  *
271  * This function is visible to journal users (like ext3fs), so is not
272  * called with the journal already locked.
273  *
274  * Return a pointer to a newly allocated handle, or NULL on failure
275  */
276 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
277 {
278         handle_t *handle = journal_current_handle();
279         int err;
280
281         if (!journal)
282                 return ERR_PTR(-EROFS);
283
284         if (handle) {
285                 J_ASSERT(handle->h_transaction->t_journal == journal);
286                 handle->h_ref++;
287                 return handle;
288         }
289
290         handle = new_handle(nblocks);
291         if (!handle)
292                 return ERR_PTR(-ENOMEM);
293
294         current->journal_info = handle;
295
296         err = start_this_handle(journal, handle);
297         if (err < 0) {
298                 jbd2_free_handle(handle);
299                 current->journal_info = NULL;
300                 handle = ERR_PTR(err);
301                 goto out;
302         }
303
304         lock_map_acquire(&handle->h_lockdep_map);
305 out:
306         return handle;
307 }
308
309 /**
310  * int jbd2_journal_extend() - extend buffer credits.
311  * @handle:  handle to 'extend'
312  * @nblocks: nr blocks to try to extend by.
313  *
314  * Some transactions, such as large extends and truncates, can be done
315  * atomically all at once or in several stages.  The operation requests
316  * a credit for a number of buffer modications in advance, but can
317  * extend its credit if it needs more.
318  *
319  * jbd2_journal_extend tries to give the running handle more buffer credits.
320  * It does not guarantee that allocation - this is a best-effort only.
321  * The calling process MUST be able to deal cleanly with a failure to
322  * extend here.
323  *
324  * Return 0 on success, non-zero on failure.
325  *
326  * return code < 0 implies an error
327  * return code > 0 implies normal transaction-full status.
328  */
329 int jbd2_journal_extend(handle_t *handle, int nblocks)
330 {
331         transaction_t *transaction = handle->h_transaction;
332         journal_t *journal = transaction->t_journal;
333         int result;
334         int wanted;
335
336         result = -EIO;
337         if (is_handle_aborted(handle))
338                 goto out;
339
340         result = 1;
341
342         spin_lock(&journal->j_state_lock);
343
344         /* Don't extend a locked-down transaction! */
345         if (handle->h_transaction->t_state != T_RUNNING) {
346                 jbd_debug(3, "denied handle %p %d blocks: "
347                           "transaction not running\n", handle, nblocks);
348                 goto error_out;
349         }
350
351         spin_lock(&transaction->t_handle_lock);
352         wanted = transaction->t_outstanding_credits + nblocks;
353
354         if (wanted > journal->j_max_transaction_buffers) {
355                 jbd_debug(3, "denied handle %p %d blocks: "
356                           "transaction too large\n", handle, nblocks);
357                 goto unlock;
358         }
359
360         if (wanted > __jbd2_log_space_left(journal)) {
361                 jbd_debug(3, "denied handle %p %d blocks: "
362                           "insufficient log space\n", handle, nblocks);
363                 goto unlock;
364         }
365
366         handle->h_buffer_credits += nblocks;
367         transaction->t_outstanding_credits += nblocks;
368         result = 0;
369
370         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
371 unlock:
372         spin_unlock(&transaction->t_handle_lock);
373 error_out:
374         spin_unlock(&journal->j_state_lock);
375 out:
376         return result;
377 }
378
379
380 /**
381  * int jbd2_journal_restart() - restart a handle .
382  * @handle:  handle to restart
383  * @nblocks: nr credits requested
384  *
385  * Restart a handle for a multi-transaction filesystem
386  * operation.
387  *
388  * If the jbd2_journal_extend() call above fails to grant new buffer credits
389  * to a running handle, a call to jbd2_journal_restart will commit the
390  * handle's transaction so far and reattach the handle to a new
391  * transaction capabable of guaranteeing the requested number of
392  * credits.
393  */
394
395 int jbd2_journal_restart(handle_t *handle, int nblocks)
396 {
397         transaction_t *transaction = handle->h_transaction;
398         journal_t *journal = transaction->t_journal;
399         int ret;
400
401         /* If we've had an abort of any type, don't even think about
402          * actually doing the restart! */
403         if (is_handle_aborted(handle))
404                 return 0;
405
406         /*
407          * First unlink the handle from its current transaction, and start the
408          * commit on that.
409          */
410         J_ASSERT(transaction->t_updates > 0);
411         J_ASSERT(journal_current_handle() == handle);
412
413         spin_lock(&journal->j_state_lock);
414         spin_lock(&transaction->t_handle_lock);
415         transaction->t_outstanding_credits -= handle->h_buffer_credits;
416         transaction->t_updates--;
417
418         if (!transaction->t_updates)
419                 wake_up(&journal->j_wait_updates);
420         spin_unlock(&transaction->t_handle_lock);
421
422         jbd_debug(2, "restarting handle %p\n", handle);
423         __jbd2_log_start_commit(journal, transaction->t_tid);
424         spin_unlock(&journal->j_state_lock);
425
426         handle->h_buffer_credits = nblocks;
427         ret = start_this_handle(journal, handle);
428         return ret;
429 }
430
431
432 /**
433  * void jbd2_journal_lock_updates () - establish a transaction barrier.
434  * @journal:  Journal to establish a barrier on.
435  *
436  * This locks out any further updates from being started, and blocks
437  * until all existing updates have completed, returning only once the
438  * journal is in a quiescent state with no updates running.
439  *
440  * The journal lock should not be held on entry.
441  */
442 void jbd2_journal_lock_updates(journal_t *journal)
443 {
444         DEFINE_WAIT(wait);
445
446         spin_lock(&journal->j_state_lock);
447         ++journal->j_barrier_count;
448
449         /* Wait until there are no running updates */
450         while (1) {
451                 transaction_t *transaction = journal->j_running_transaction;
452
453                 if (!transaction)
454                         break;
455
456                 spin_lock(&transaction->t_handle_lock);
457                 if (!transaction->t_updates) {
458                         spin_unlock(&transaction->t_handle_lock);
459                         break;
460                 }
461                 prepare_to_wait(&journal->j_wait_updates, &wait,
462                                 TASK_UNINTERRUPTIBLE);
463                 spin_unlock(&transaction->t_handle_lock);
464                 spin_unlock(&journal->j_state_lock);
465                 schedule();
466                 finish_wait(&journal->j_wait_updates, &wait);
467                 spin_lock(&journal->j_state_lock);
468         }
469         spin_unlock(&journal->j_state_lock);
470
471         /*
472          * We have now established a barrier against other normal updates, but
473          * we also need to barrier against other jbd2_journal_lock_updates() calls
474          * to make sure that we serialise special journal-locked operations
475          * too.
476          */
477         mutex_lock(&journal->j_barrier);
478 }
479
480 /**
481  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
482  * @journal:  Journal to release the barrier on.
483  *
484  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
485  *
486  * Should be called without the journal lock held.
487  */
488 void jbd2_journal_unlock_updates (journal_t *journal)
489 {
490         J_ASSERT(journal->j_barrier_count != 0);
491
492         mutex_unlock(&journal->j_barrier);
493         spin_lock(&journal->j_state_lock);
494         --journal->j_barrier_count;
495         spin_unlock(&journal->j_state_lock);
496         wake_up(&journal->j_wait_transaction_locked);
497 }
498
499 /*
500  * Report any unexpected dirty buffers which turn up.  Normally those
501  * indicate an error, but they can occur if the user is running (say)
502  * tune2fs to modify the live filesystem, so we need the option of
503  * continuing as gracefully as possible.  #
504  *
505  * The caller should already hold the journal lock and
506  * j_list_lock spinlock: most callers will need those anyway
507  * in order to probe the buffer's journaling state safely.
508  */
509 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
510 {
511         int jlist;
512
513         /* If this buffer is one which might reasonably be dirty
514          * --- ie. data, or not part of this journal --- then
515          * we're OK to leave it alone, but otherwise we need to
516          * move the dirty bit to the journal's own internal
517          * JBDDirty bit. */
518         jlist = jh->b_jlist;
519
520         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
521             jlist == BJ_Shadow || jlist == BJ_Forget) {
522                 struct buffer_head *bh = jh2bh(jh);
523
524                 if (test_clear_buffer_dirty(bh))
525                         set_buffer_jbddirty(bh);
526         }
527 }
528
529 /*
530  * If the buffer is already part of the current transaction, then there
531  * is nothing we need to do.  If it is already part of a prior
532  * transaction which we are still committing to disk, then we need to
533  * make sure that we do not overwrite the old copy: we do copy-out to
534  * preserve the copy going to disk.  We also account the buffer against
535  * the handle's metadata buffer credits (unless the buffer is already
536  * part of the transaction, that is).
537  *
538  */
539 static int
540 do_get_write_access(handle_t *handle, struct journal_head *jh,
541                         int force_copy)
542 {
543         struct buffer_head *bh;
544         transaction_t *transaction;
545         journal_t *journal;
546         int error;
547         char *frozen_buffer = NULL;
548         int need_copy = 0;
549
550         if (is_handle_aborted(handle))
551                 return -EROFS;
552
553         transaction = handle->h_transaction;
554         journal = transaction->t_journal;
555
556         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
557
558         JBUFFER_TRACE(jh, "entry");
559 repeat:
560         bh = jh2bh(jh);
561
562         /* @@@ Need to check for errors here at some point. */
563
564         lock_buffer(bh);
565         jbd_lock_bh_state(bh);
566
567         /* We now hold the buffer lock so it is safe to query the buffer
568          * state.  Is the buffer dirty?
569          *
570          * If so, there are two possibilities.  The buffer may be
571          * non-journaled, and undergoing a quite legitimate writeback.
572          * Otherwise, it is journaled, and we don't expect dirty buffers
573          * in that state (the buffers should be marked JBD_Dirty
574          * instead.)  So either the IO is being done under our own
575          * control and this is a bug, or it's a third party IO such as
576          * dump(8) (which may leave the buffer scheduled for read ---
577          * ie. locked but not dirty) or tune2fs (which may actually have
578          * the buffer dirtied, ugh.)  */
579
580         if (buffer_dirty(bh)) {
581                 /*
582                  * First question: is this buffer already part of the current
583                  * transaction or the existing committing transaction?
584                  */
585                 if (jh->b_transaction) {
586                         J_ASSERT_JH(jh,
587                                 jh->b_transaction == transaction ||
588                                 jh->b_transaction ==
589                                         journal->j_committing_transaction);
590                         if (jh->b_next_transaction)
591                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
592                                                         transaction);
593                 }
594                 /*
595                  * In any case we need to clean the dirty flag and we must
596                  * do it under the buffer lock to be sure we don't race
597                  * with running write-out.
598                  */
599                 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
600                 jbd_unexpected_dirty_buffer(jh);
601         }
602
603         unlock_buffer(bh);
604
605         error = -EROFS;
606         if (is_handle_aborted(handle)) {
607                 jbd_unlock_bh_state(bh);
608                 goto out;
609         }
610         error = 0;
611
612         /*
613          * The buffer is already part of this transaction if b_transaction or
614          * b_next_transaction points to it
615          */
616         if (jh->b_transaction == transaction ||
617             jh->b_next_transaction == transaction)
618                 goto done;
619
620         /*
621          * this is the first time this transaction is touching this buffer,
622          * reset the modified flag
623          */
624        jh->b_modified = 0;
625
626         /*
627          * If there is already a copy-out version of this buffer, then we don't
628          * need to make another one
629          */
630         if (jh->b_frozen_data) {
631                 JBUFFER_TRACE(jh, "has frozen data");
632                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
633                 jh->b_next_transaction = transaction;
634                 goto done;
635         }
636
637         /* Is there data here we need to preserve? */
638
639         if (jh->b_transaction && jh->b_transaction != transaction) {
640                 JBUFFER_TRACE(jh, "owned by older transaction");
641                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
642                 J_ASSERT_JH(jh, jh->b_transaction ==
643                                         journal->j_committing_transaction);
644
645                 /* There is one case we have to be very careful about.
646                  * If the committing transaction is currently writing
647                  * this buffer out to disk and has NOT made a copy-out,
648                  * then we cannot modify the buffer contents at all
649                  * right now.  The essence of copy-out is that it is the
650                  * extra copy, not the primary copy, which gets
651                  * journaled.  If the primary copy is already going to
652                  * disk then we cannot do copy-out here. */
653
654                 if (jh->b_jlist == BJ_Shadow) {
655                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
656                         wait_queue_head_t *wqh;
657
658                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
659
660                         JBUFFER_TRACE(jh, "on shadow: sleep");
661                         jbd_unlock_bh_state(bh);
662                         /* commit wakes up all shadow buffers after IO */
663                         for ( ; ; ) {
664                                 prepare_to_wait(wqh, &wait.wait,
665                                                 TASK_UNINTERRUPTIBLE);
666                                 if (jh->b_jlist != BJ_Shadow)
667                                         break;
668                                 schedule();
669                         }
670                         finish_wait(wqh, &wait.wait);
671                         goto repeat;
672                 }
673
674                 /* Only do the copy if the currently-owning transaction
675                  * still needs it.  If it is on the Forget list, the
676                  * committing transaction is past that stage.  The
677                  * buffer had better remain locked during the kmalloc,
678                  * but that should be true --- we hold the journal lock
679                  * still and the buffer is already on the BUF_JOURNAL
680                  * list so won't be flushed.
681                  *
682                  * Subtle point, though: if this is a get_undo_access,
683                  * then we will be relying on the frozen_data to contain
684                  * the new value of the committed_data record after the
685                  * transaction, so we HAVE to force the frozen_data copy
686                  * in that case. */
687
688                 if (jh->b_jlist != BJ_Forget || force_copy) {
689                         JBUFFER_TRACE(jh, "generate frozen data");
690                         if (!frozen_buffer) {
691                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
692                                 jbd_unlock_bh_state(bh);
693                                 frozen_buffer =
694                                         jbd2_alloc(jh2bh(jh)->b_size,
695                                                          GFP_NOFS);
696                                 if (!frozen_buffer) {
697                                         printk(KERN_EMERG
698                                                "%s: OOM for frozen_buffer\n",
699                                                __func__);
700                                         JBUFFER_TRACE(jh, "oom!");
701                                         error = -ENOMEM;
702                                         jbd_lock_bh_state(bh);
703                                         goto done;
704                                 }
705                                 goto repeat;
706                         }
707                         jh->b_frozen_data = frozen_buffer;
708                         frozen_buffer = NULL;
709                         need_copy = 1;
710                 }
711                 jh->b_next_transaction = transaction;
712         }
713
714
715         /*
716          * Finally, if the buffer is not journaled right now, we need to make
717          * sure it doesn't get written to disk before the caller actually
718          * commits the new data
719          */
720         if (!jh->b_transaction) {
721                 JBUFFER_TRACE(jh, "no transaction");
722                 J_ASSERT_JH(jh, !jh->b_next_transaction);
723                 jh->b_transaction = transaction;
724                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
725                 spin_lock(&journal->j_list_lock);
726                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
727                 spin_unlock(&journal->j_list_lock);
728         }
729
730 done:
731         if (need_copy) {
732                 struct page *page;
733                 int offset;
734                 char *source;
735
736                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
737                             "Possible IO failure.\n");
738                 page = jh2bh(jh)->b_page;
739                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
740                 source = kmap_atomic(page, KM_USER0);
741                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
742                 kunmap_atomic(source, KM_USER0);
743         }
744         jbd_unlock_bh_state(bh);
745
746         /*
747          * If we are about to journal a buffer, then any revoke pending on it is
748          * no longer valid
749          */
750         jbd2_journal_cancel_revoke(handle, jh);
751
752 out:
753         if (unlikely(frozen_buffer))    /* It's usually NULL */
754                 jbd2_free(frozen_buffer, bh->b_size);
755
756         JBUFFER_TRACE(jh, "exit");
757         return error;
758 }
759
760 /**
761  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
762  * @handle: transaction to add buffer modifications to
763  * @bh:     bh to be used for metadata writes
764  * @credits: variable that will receive credits for the buffer
765  *
766  * Returns an error code or 0 on success.
767  *
768  * In full data journalling mode the buffer may be of type BJ_AsyncData,
769  * because we're write()ing a buffer which is also part of a shared mapping.
770  */
771
772 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
773 {
774         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
775         int rc;
776
777         /* We do not want to get caught playing with fields which the
778          * log thread also manipulates.  Make sure that the buffer
779          * completes any outstanding IO before proceeding. */
780         rc = do_get_write_access(handle, jh, 0);
781         jbd2_journal_put_journal_head(jh);
782         return rc;
783 }
784
785
786 /*
787  * When the user wants to journal a newly created buffer_head
788  * (ie. getblk() returned a new buffer and we are going to populate it
789  * manually rather than reading off disk), then we need to keep the
790  * buffer_head locked until it has been completely filled with new
791  * data.  In this case, we should be able to make the assertion that
792  * the bh is not already part of an existing transaction.
793  *
794  * The buffer should already be locked by the caller by this point.
795  * There is no lock ranking violation: it was a newly created,
796  * unlocked buffer beforehand. */
797
798 /**
799  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
800  * @handle: transaction to new buffer to
801  * @bh: new buffer.
802  *
803  * Call this if you create a new bh.
804  */
805 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
806 {
807         transaction_t *transaction = handle->h_transaction;
808         journal_t *journal = transaction->t_journal;
809         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
810         int err;
811
812         jbd_debug(5, "journal_head %p\n", jh);
813         err = -EROFS;
814         if (is_handle_aborted(handle))
815                 goto out;
816         err = 0;
817
818         JBUFFER_TRACE(jh, "entry");
819         /*
820          * The buffer may already belong to this transaction due to pre-zeroing
821          * in the filesystem's new_block code.  It may also be on the previous,
822          * committing transaction's lists, but it HAS to be in Forget state in
823          * that case: the transaction must have deleted the buffer for it to be
824          * reused here.
825          */
826         jbd_lock_bh_state(bh);
827         spin_lock(&journal->j_list_lock);
828         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
829                 jh->b_transaction == NULL ||
830                 (jh->b_transaction == journal->j_committing_transaction &&
831                           jh->b_jlist == BJ_Forget)));
832
833         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
834         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
835
836         if (jh->b_transaction == NULL) {
837                 jh->b_transaction = transaction;
838
839                 /* first access by this transaction */
840                 jh->b_modified = 0;
841
842                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
843                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
844         } else if (jh->b_transaction == journal->j_committing_transaction) {
845                 /* first access by this transaction */
846                 jh->b_modified = 0;
847
848                 JBUFFER_TRACE(jh, "set next transaction");
849                 jh->b_next_transaction = transaction;
850         }
851         spin_unlock(&journal->j_list_lock);
852         jbd_unlock_bh_state(bh);
853
854         /*
855          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
856          * blocks which contain freed but then revoked metadata.  We need
857          * to cancel the revoke in case we end up freeing it yet again
858          * and the reallocating as data - this would cause a second revoke,
859          * which hits an assertion error.
860          */
861         JBUFFER_TRACE(jh, "cancelling revoke");
862         jbd2_journal_cancel_revoke(handle, jh);
863         jbd2_journal_put_journal_head(jh);
864 out:
865         return err;
866 }
867
868 /**
869  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
870  *     non-rewindable consequences
871  * @handle: transaction
872  * @bh: buffer to undo
873  * @credits: store the number of taken credits here (if not NULL)
874  *
875  * Sometimes there is a need to distinguish between metadata which has
876  * been committed to disk and that which has not.  The ext3fs code uses
877  * this for freeing and allocating space, we have to make sure that we
878  * do not reuse freed space until the deallocation has been committed,
879  * since if we overwrote that space we would make the delete
880  * un-rewindable in case of a crash.
881  *
882  * To deal with that, jbd2_journal_get_undo_access requests write access to a
883  * buffer for parts of non-rewindable operations such as delete
884  * operations on the bitmaps.  The journaling code must keep a copy of
885  * the buffer's contents prior to the undo_access call until such time
886  * as we know that the buffer has definitely been committed to disk.
887  *
888  * We never need to know which transaction the committed data is part
889  * of, buffers touched here are guaranteed to be dirtied later and so
890  * will be committed to a new transaction in due course, at which point
891  * we can discard the old committed data pointer.
892  *
893  * Returns error number or 0 on success.
894  */
895 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
896 {
897         int err;
898         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
899         char *committed_data = NULL;
900
901         JBUFFER_TRACE(jh, "entry");
902
903         /*
904          * Do this first --- it can drop the journal lock, so we want to
905          * make sure that obtaining the committed_data is done
906          * atomically wrt. completion of any outstanding commits.
907          */
908         err = do_get_write_access(handle, jh, 1);
909         if (err)
910                 goto out;
911
912 repeat:
913         if (!jh->b_committed_data) {
914                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
915                 if (!committed_data) {
916                         printk(KERN_EMERG "%s: No memory for committed data\n",
917                                 __func__);
918                         err = -ENOMEM;
919                         goto out;
920                 }
921         }
922
923         jbd_lock_bh_state(bh);
924         if (!jh->b_committed_data) {
925                 /* Copy out the current buffer contents into the
926                  * preserved, committed copy. */
927                 JBUFFER_TRACE(jh, "generate b_committed data");
928                 if (!committed_data) {
929                         jbd_unlock_bh_state(bh);
930                         goto repeat;
931                 }
932
933                 jh->b_committed_data = committed_data;
934                 committed_data = NULL;
935                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
936         }
937         jbd_unlock_bh_state(bh);
938 out:
939         jbd2_journal_put_journal_head(jh);
940         if (unlikely(committed_data))
941                 jbd2_free(committed_data, bh->b_size);
942         return err;
943 }
944
945 /**
946  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
947  * @handle: transaction to add buffer to.
948  * @bh: buffer to mark
949  *
950  * mark dirty metadata which needs to be journaled as part of the current
951  * transaction.
952  *
953  * The buffer is placed on the transaction's metadata list and is marked
954  * as belonging to the transaction.
955  *
956  * Returns error number or 0 on success.
957  *
958  * Special care needs to be taken if the buffer already belongs to the
959  * current committing transaction (in which case we should have frozen
960  * data present for that commit).  In that case, we don't relink the
961  * buffer: that only gets done when the old transaction finally
962  * completes its commit.
963  */
964 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
965 {
966         transaction_t *transaction = handle->h_transaction;
967         journal_t *journal = transaction->t_journal;
968         struct journal_head *jh = bh2jh(bh);
969
970         jbd_debug(5, "journal_head %p\n", jh);
971         JBUFFER_TRACE(jh, "entry");
972         if (is_handle_aborted(handle))
973                 goto out;
974
975         jbd_lock_bh_state(bh);
976
977         if (jh->b_modified == 0) {
978                 /*
979                  * This buffer's got modified and becoming part
980                  * of the transaction. This needs to be done
981                  * once a transaction -bzzz
982                  */
983                 jh->b_modified = 1;
984                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
985                 handle->h_buffer_credits--;
986         }
987
988         /*
989          * fastpath, to avoid expensive locking.  If this buffer is already
990          * on the running transaction's metadata list there is nothing to do.
991          * Nobody can take it off again because there is a handle open.
992          * I _think_ we're OK here with SMP barriers - a mistaken decision will
993          * result in this test being false, so we go in and take the locks.
994          */
995         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
996                 JBUFFER_TRACE(jh, "fastpath");
997                 J_ASSERT_JH(jh, jh->b_transaction ==
998                                         journal->j_running_transaction);
999                 goto out_unlock_bh;
1000         }
1001
1002         set_buffer_jbddirty(bh);
1003
1004         /*
1005          * Metadata already on the current transaction list doesn't
1006          * need to be filed.  Metadata on another transaction's list must
1007          * be committing, and will be refiled once the commit completes:
1008          * leave it alone for now.
1009          */
1010         if (jh->b_transaction != transaction) {
1011                 JBUFFER_TRACE(jh, "already on other transaction");
1012                 J_ASSERT_JH(jh, jh->b_transaction ==
1013                                         journal->j_committing_transaction);
1014                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1015                 /* And this case is illegal: we can't reuse another
1016                  * transaction's data buffer, ever. */
1017                 goto out_unlock_bh;
1018         }
1019
1020         /* That test should have eliminated the following case: */
1021         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1022
1023         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1024         spin_lock(&journal->j_list_lock);
1025         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1026         spin_unlock(&journal->j_list_lock);
1027 out_unlock_bh:
1028         jbd_unlock_bh_state(bh);
1029 out:
1030         JBUFFER_TRACE(jh, "exit");
1031         return 0;
1032 }
1033
1034 /*
1035  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1036  * updates, if the update decided in the end that it didn't need access.
1037  *
1038  */
1039 void
1040 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1041 {
1042         BUFFER_TRACE(bh, "entry");
1043 }
1044
1045 /**
1046  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1047  * @handle: transaction handle
1048  * @bh:     bh to 'forget'
1049  *
1050  * We can only do the bforget if there are no commits pending against the
1051  * buffer.  If the buffer is dirty in the current running transaction we
1052  * can safely unlink it.
1053  *
1054  * bh may not be a journalled buffer at all - it may be a non-JBD
1055  * buffer which came off the hashtable.  Check for this.
1056  *
1057  * Decrements bh->b_count by one.
1058  *
1059  * Allow this call even if the handle has aborted --- it may be part of
1060  * the caller's cleanup after an abort.
1061  */
1062 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1063 {
1064         transaction_t *transaction = handle->h_transaction;
1065         journal_t *journal = transaction->t_journal;
1066         struct journal_head *jh;
1067         int drop_reserve = 0;
1068         int err = 0;
1069         int was_modified = 0;
1070
1071         BUFFER_TRACE(bh, "entry");
1072
1073         jbd_lock_bh_state(bh);
1074         spin_lock(&journal->j_list_lock);
1075
1076         if (!buffer_jbd(bh))
1077                 goto not_jbd;
1078         jh = bh2jh(bh);
1079
1080         /* Critical error: attempting to delete a bitmap buffer, maybe?
1081          * Don't do any jbd operations, and return an error. */
1082         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1083                          "inconsistent data on disk")) {
1084                 err = -EIO;
1085                 goto not_jbd;
1086         }
1087
1088         /* keep track of wether or not this transaction modified us */
1089         was_modified = jh->b_modified;
1090
1091         /*
1092          * The buffer's going from the transaction, we must drop
1093          * all references -bzzz
1094          */
1095         jh->b_modified = 0;
1096
1097         if (jh->b_transaction == handle->h_transaction) {
1098                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1099
1100                 /* If we are forgetting a buffer which is already part
1101                  * of this transaction, then we can just drop it from
1102                  * the transaction immediately. */
1103                 clear_buffer_dirty(bh);
1104                 clear_buffer_jbddirty(bh);
1105
1106                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1107
1108                 /*
1109                  * we only want to drop a reference if this transaction
1110                  * modified the buffer
1111                  */
1112                 if (was_modified)
1113                         drop_reserve = 1;
1114
1115                 /*
1116                  * We are no longer going to journal this buffer.
1117                  * However, the commit of this transaction is still
1118                  * important to the buffer: the delete that we are now
1119                  * processing might obsolete an old log entry, so by
1120                  * committing, we can satisfy the buffer's checkpoint.
1121                  *
1122                  * So, if we have a checkpoint on the buffer, we should
1123                  * now refile the buffer on our BJ_Forget list so that
1124                  * we know to remove the checkpoint after we commit.
1125                  */
1126
1127                 if (jh->b_cp_transaction) {
1128                         __jbd2_journal_temp_unlink_buffer(jh);
1129                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1130                 } else {
1131                         __jbd2_journal_unfile_buffer(jh);
1132                         jbd2_journal_remove_journal_head(bh);
1133                         __brelse(bh);
1134                         if (!buffer_jbd(bh)) {
1135                                 spin_unlock(&journal->j_list_lock);
1136                                 jbd_unlock_bh_state(bh);
1137                                 __bforget(bh);
1138                                 goto drop;
1139                         }
1140                 }
1141         } else if (jh->b_transaction) {
1142                 J_ASSERT_JH(jh, (jh->b_transaction ==
1143                                  journal->j_committing_transaction));
1144                 /* However, if the buffer is still owned by a prior
1145                  * (committing) transaction, we can't drop it yet... */
1146                 JBUFFER_TRACE(jh, "belongs to older transaction");
1147                 /* ... but we CAN drop it from the new transaction if we
1148                  * have also modified it since the original commit. */
1149
1150                 if (jh->b_next_transaction) {
1151                         J_ASSERT(jh->b_next_transaction == transaction);
1152                         jh->b_next_transaction = NULL;
1153
1154                         /*
1155                          * only drop a reference if this transaction modified
1156                          * the buffer
1157                          */
1158                         if (was_modified)
1159                                 drop_reserve = 1;
1160                 }
1161         }
1162
1163 not_jbd:
1164         spin_unlock(&journal->j_list_lock);
1165         jbd_unlock_bh_state(bh);
1166         __brelse(bh);
1167 drop:
1168         if (drop_reserve) {
1169                 /* no need to reserve log space for this block -bzzz */
1170                 handle->h_buffer_credits++;
1171         }
1172         return err;
1173 }
1174
1175 /**
1176  * int jbd2_journal_stop() - complete a transaction
1177  * @handle: tranaction to complete.
1178  *
1179  * All done for a particular handle.
1180  *
1181  * There is not much action needed here.  We just return any remaining
1182  * buffer credits to the transaction and remove the handle.  The only
1183  * complication is that we need to start a commit operation if the
1184  * filesystem is marked for synchronous update.
1185  *
1186  * jbd2_journal_stop itself will not usually return an error, but it may
1187  * do so in unusual circumstances.  In particular, expect it to
1188  * return -EIO if a jbd2_journal_abort has been executed since the
1189  * transaction began.
1190  */
1191 int jbd2_journal_stop(handle_t *handle)
1192 {
1193         transaction_t *transaction = handle->h_transaction;
1194         journal_t *journal = transaction->t_journal;
1195         int old_handle_count, err;
1196         pid_t pid;
1197
1198         J_ASSERT(journal_current_handle() == handle);
1199
1200         if (is_handle_aborted(handle))
1201                 err = -EIO;
1202         else {
1203                 J_ASSERT(transaction->t_updates > 0);
1204                 err = 0;
1205         }
1206
1207         if (--handle->h_ref > 0) {
1208                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1209                           handle->h_ref);
1210                 return err;
1211         }
1212
1213         jbd_debug(4, "Handle %p going down\n", handle);
1214
1215         /*
1216          * Implement synchronous transaction batching.  If the handle
1217          * was synchronous, don't force a commit immediately.  Let's
1218          * yield and let another thread piggyback onto this transaction.
1219          * Keep doing that while new threads continue to arrive.
1220          * It doesn't cost much - we're about to run a commit and sleep
1221          * on IO anyway.  Speeds up many-threaded, many-dir operations
1222          * by 30x or more...
1223          *
1224          * But don't do this if this process was the most recent one to
1225          * perform a synchronous write.  We do this to detect the case where a
1226          * single process is doing a stream of sync writes.  No point in waiting
1227          * for joiners in that case.
1228          */
1229         pid = current->pid;
1230         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1231                 journal->j_last_sync_writer = pid;
1232                 do {
1233                         old_handle_count = transaction->t_handle_count;
1234                         schedule_timeout_uninterruptible(1);
1235                 } while (old_handle_count != transaction->t_handle_count);
1236         }
1237
1238         current->journal_info = NULL;
1239         spin_lock(&journal->j_state_lock);
1240         spin_lock(&transaction->t_handle_lock);
1241         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1242         transaction->t_updates--;
1243         if (!transaction->t_updates) {
1244                 wake_up(&journal->j_wait_updates);
1245                 if (journal->j_barrier_count)
1246                         wake_up(&journal->j_wait_transaction_locked);
1247         }
1248
1249         /*
1250          * If the handle is marked SYNC, we need to set another commit
1251          * going!  We also want to force a commit if the current
1252          * transaction is occupying too much of the log, or if the
1253          * transaction is too old now.
1254          */
1255         if (handle->h_sync ||
1256                         transaction->t_outstanding_credits >
1257                                 journal->j_max_transaction_buffers ||
1258                         time_after_eq(jiffies, transaction->t_expires)) {
1259                 /* Do this even for aborted journals: an abort still
1260                  * completes the commit thread, it just doesn't write
1261                  * anything to disk. */
1262                 tid_t tid = transaction->t_tid;
1263
1264                 spin_unlock(&transaction->t_handle_lock);
1265                 jbd_debug(2, "transaction too old, requesting commit for "
1266                                         "handle %p\n", handle);
1267                 /* This is non-blocking */
1268                 __jbd2_log_start_commit(journal, transaction->t_tid);
1269                 spin_unlock(&journal->j_state_lock);
1270
1271                 /*
1272                  * Special case: JBD2_SYNC synchronous updates require us
1273                  * to wait for the commit to complete.
1274                  */
1275                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1276                         err = jbd2_log_wait_commit(journal, tid);
1277         } else {
1278                 spin_unlock(&transaction->t_handle_lock);
1279                 spin_unlock(&journal->j_state_lock);
1280         }
1281
1282         lock_map_release(&handle->h_lockdep_map);
1283
1284         jbd2_free_handle(handle);
1285         return err;
1286 }
1287
1288 /**
1289  * int jbd2_journal_force_commit() - force any uncommitted transactions
1290  * @journal: journal to force
1291  *
1292  * For synchronous operations: force any uncommitted transactions
1293  * to disk.  May seem kludgy, but it reuses all the handle batching
1294  * code in a very simple manner.
1295  */
1296 int jbd2_journal_force_commit(journal_t *journal)
1297 {
1298         handle_t *handle;
1299         int ret;
1300
1301         handle = jbd2_journal_start(journal, 1);
1302         if (IS_ERR(handle)) {
1303                 ret = PTR_ERR(handle);
1304         } else {
1305                 handle->h_sync = 1;
1306                 ret = jbd2_journal_stop(handle);
1307         }
1308         return ret;
1309 }
1310
1311 /*
1312  *
1313  * List management code snippets: various functions for manipulating the
1314  * transaction buffer lists.
1315  *
1316  */
1317
1318 /*
1319  * Append a buffer to a transaction list, given the transaction's list head
1320  * pointer.
1321  *
1322  * j_list_lock is held.
1323  *
1324  * jbd_lock_bh_state(jh2bh(jh)) is held.
1325  */
1326
1327 static inline void
1328 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1329 {
1330         if (!*list) {
1331                 jh->b_tnext = jh->b_tprev = jh;
1332                 *list = jh;
1333         } else {
1334                 /* Insert at the tail of the list to preserve order */
1335                 struct journal_head *first = *list, *last = first->b_tprev;
1336                 jh->b_tprev = last;
1337                 jh->b_tnext = first;
1338                 last->b_tnext = first->b_tprev = jh;
1339         }
1340 }
1341
1342 /*
1343  * Remove a buffer from a transaction list, given the transaction's list
1344  * head pointer.
1345  *
1346  * Called with j_list_lock held, and the journal may not be locked.
1347  *
1348  * jbd_lock_bh_state(jh2bh(jh)) is held.
1349  */
1350
1351 static inline void
1352 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1353 {
1354         if (*list == jh) {
1355                 *list = jh->b_tnext;
1356                 if (*list == jh)
1357                         *list = NULL;
1358         }
1359         jh->b_tprev->b_tnext = jh->b_tnext;
1360         jh->b_tnext->b_tprev = jh->b_tprev;
1361 }
1362
1363 /*
1364  * Remove a buffer from the appropriate transaction list.
1365  *
1366  * Note that this function can *change* the value of
1367  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1368  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1369  * of these pointers, it could go bad.  Generally the caller needs to re-read
1370  * the pointer from the transaction_t.
1371  *
1372  * Called under j_list_lock.  The journal may not be locked.
1373  */
1374 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1375 {
1376         struct journal_head **list = NULL;
1377         transaction_t *transaction;
1378         struct buffer_head *bh = jh2bh(jh);
1379
1380         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1381         transaction = jh->b_transaction;
1382         if (transaction)
1383                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1384
1385         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1386         if (jh->b_jlist != BJ_None)
1387                 J_ASSERT_JH(jh, transaction != NULL);
1388
1389         switch (jh->b_jlist) {
1390         case BJ_None:
1391                 return;
1392         case BJ_Metadata:
1393                 transaction->t_nr_buffers--;
1394                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1395                 list = &transaction->t_buffers;
1396                 break;
1397         case BJ_Forget:
1398                 list = &transaction->t_forget;
1399                 break;
1400         case BJ_IO:
1401                 list = &transaction->t_iobuf_list;
1402                 break;
1403         case BJ_Shadow:
1404                 list = &transaction->t_shadow_list;
1405                 break;
1406         case BJ_LogCtl:
1407                 list = &transaction->t_log_list;
1408                 break;
1409         case BJ_Reserved:
1410                 list = &transaction->t_reserved_list;
1411                 break;
1412         }
1413
1414         __blist_del_buffer(list, jh);
1415         jh->b_jlist = BJ_None;
1416         if (test_clear_buffer_jbddirty(bh))
1417                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1418 }
1419
1420 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1421 {
1422         __jbd2_journal_temp_unlink_buffer(jh);
1423         jh->b_transaction = NULL;
1424 }
1425
1426 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1427 {
1428         jbd_lock_bh_state(jh2bh(jh));
1429         spin_lock(&journal->j_list_lock);
1430         __jbd2_journal_unfile_buffer(jh);
1431         spin_unlock(&journal->j_list_lock);
1432         jbd_unlock_bh_state(jh2bh(jh));
1433 }
1434
1435 /*
1436  * Called from jbd2_journal_try_to_free_buffers().
1437  *
1438  * Called under jbd_lock_bh_state(bh)
1439  */
1440 static void
1441 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1442 {
1443         struct journal_head *jh;
1444
1445         jh = bh2jh(bh);
1446
1447         if (buffer_locked(bh) || buffer_dirty(bh))
1448                 goto out;
1449
1450         if (jh->b_next_transaction != NULL)
1451                 goto out;
1452
1453         spin_lock(&journal->j_list_lock);
1454         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1455                 /* written-back checkpointed metadata buffer */
1456                 if (jh->b_jlist == BJ_None) {
1457                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1458                         __jbd2_journal_remove_checkpoint(jh);
1459                         jbd2_journal_remove_journal_head(bh);
1460                         __brelse(bh);
1461                 }
1462         }
1463         spin_unlock(&journal->j_list_lock);
1464 out:
1465         return;
1466 }
1467
1468 /*
1469  * jbd2_journal_try_to_free_buffers() could race with
1470  * jbd2_journal_commit_transaction(). The later might still hold the
1471  * reference count to the buffers when inspecting them on
1472  * t_syncdata_list or t_locked_list.
1473  *
1474  * jbd2_journal_try_to_free_buffers() will call this function to
1475  * wait for the current transaction to finish syncing data buffers, before
1476  * try to free that buffer.
1477  *
1478  * Called with journal->j_state_lock hold.
1479  */
1480 static void jbd2_journal_wait_for_transaction_sync_data(journal_t *journal)
1481 {
1482         transaction_t *transaction;
1483         tid_t tid;
1484
1485         spin_lock(&journal->j_state_lock);
1486         transaction = journal->j_committing_transaction;
1487
1488         if (!transaction) {
1489                 spin_unlock(&journal->j_state_lock);
1490                 return;
1491         }
1492
1493         tid = transaction->t_tid;
1494         spin_unlock(&journal->j_state_lock);
1495         jbd2_log_wait_commit(journal, tid);
1496 }
1497
1498 /**
1499  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1500  * @journal: journal for operation
1501  * @page: to try and free
1502  * @gfp_mask: we use the mask to detect how hard should we try to release
1503  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1504  * release the buffers.
1505  *
1506  *
1507  * For all the buffers on this page,
1508  * if they are fully written out ordered data, move them onto BUF_CLEAN
1509  * so try_to_free_buffers() can reap them.
1510  *
1511  * This function returns non-zero if we wish try_to_free_buffers()
1512  * to be called. We do this if the page is releasable by try_to_free_buffers().
1513  * We also do it if the page has locked or dirty buffers and the caller wants
1514  * us to perform sync or async writeout.
1515  *
1516  * This complicates JBD locking somewhat.  We aren't protected by the
1517  * BKL here.  We wish to remove the buffer from its committing or
1518  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1519  *
1520  * This may *change* the value of transaction_t->t_datalist, so anyone
1521  * who looks at t_datalist needs to lock against this function.
1522  *
1523  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1524  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1525  * will come out of the lock with the buffer dirty, which makes it
1526  * ineligible for release here.
1527  *
1528  * Who else is affected by this?  hmm...  Really the only contender
1529  * is do_get_write_access() - it could be looking at the buffer while
1530  * journal_try_to_free_buffer() is changing its state.  But that
1531  * cannot happen because we never reallocate freed data as metadata
1532  * while the data is part of a transaction.  Yes?
1533  *
1534  * Return 0 on failure, 1 on success
1535  */
1536 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1537                                 struct page *page, gfp_t gfp_mask)
1538 {
1539         struct buffer_head *head;
1540         struct buffer_head *bh;
1541         int ret = 0;
1542
1543         J_ASSERT(PageLocked(page));
1544
1545         head = page_buffers(page);
1546         bh = head;
1547         do {
1548                 struct journal_head *jh;
1549
1550                 /*
1551                  * We take our own ref against the journal_head here to avoid
1552                  * having to add tons of locking around each instance of
1553                  * jbd2_journal_remove_journal_head() and
1554                  * jbd2_journal_put_journal_head().
1555                  */
1556                 jh = jbd2_journal_grab_journal_head(bh);
1557                 if (!jh)
1558                         continue;
1559
1560                 jbd_lock_bh_state(bh);
1561                 __journal_try_to_free_buffer(journal, bh);
1562                 jbd2_journal_put_journal_head(jh);
1563                 jbd_unlock_bh_state(bh);
1564                 if (buffer_jbd(bh))
1565                         goto busy;
1566         } while ((bh = bh->b_this_page) != head);
1567
1568         ret = try_to_free_buffers(page);
1569
1570         /*
1571          * There are a number of places where jbd2_journal_try_to_free_buffers()
1572          * could race with jbd2_journal_commit_transaction(), the later still
1573          * holds the reference to the buffers to free while processing them.
1574          * try_to_free_buffers() failed to free those buffers. Some of the
1575          * caller of releasepage() request page buffers to be dropped, otherwise
1576          * treat the fail-to-free as errors (such as generic_file_direct_IO())
1577          *
1578          * So, if the caller of try_to_release_page() wants the synchronous
1579          * behaviour(i.e make sure buffers are dropped upon return),
1580          * let's wait for the current transaction to finish flush of
1581          * dirty data buffers, then try to free those buffers again,
1582          * with the journal locked.
1583          */
1584         if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
1585                 jbd2_journal_wait_for_transaction_sync_data(journal);
1586                 ret = try_to_free_buffers(page);
1587         }
1588
1589 busy:
1590         return ret;
1591 }
1592
1593 /*
1594  * This buffer is no longer needed.  If it is on an older transaction's
1595  * checkpoint list we need to record it on this transaction's forget list
1596  * to pin this buffer (and hence its checkpointing transaction) down until
1597  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1598  * release it.
1599  * Returns non-zero if JBD no longer has an interest in the buffer.
1600  *
1601  * Called under j_list_lock.
1602  *
1603  * Called under jbd_lock_bh_state(bh).
1604  */
1605 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1606 {
1607         int may_free = 1;
1608         struct buffer_head *bh = jh2bh(jh);
1609
1610         __jbd2_journal_unfile_buffer(jh);
1611
1612         if (jh->b_cp_transaction) {
1613                 JBUFFER_TRACE(jh, "on running+cp transaction");
1614                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1615                 clear_buffer_jbddirty(bh);
1616                 may_free = 0;
1617         } else {
1618                 JBUFFER_TRACE(jh, "on running transaction");
1619                 jbd2_journal_remove_journal_head(bh);
1620                 __brelse(bh);
1621         }
1622         return may_free;
1623 }
1624
1625 /*
1626  * jbd2_journal_invalidatepage
1627  *
1628  * This code is tricky.  It has a number of cases to deal with.
1629  *
1630  * There are two invariants which this code relies on:
1631  *
1632  * i_size must be updated on disk before we start calling invalidatepage on the
1633  * data.
1634  *
1635  *  This is done in ext3 by defining an ext3_setattr method which
1636  *  updates i_size before truncate gets going.  By maintaining this
1637  *  invariant, we can be sure that it is safe to throw away any buffers
1638  *  attached to the current transaction: once the transaction commits,
1639  *  we know that the data will not be needed.
1640  *
1641  *  Note however that we can *not* throw away data belonging to the
1642  *  previous, committing transaction!
1643  *
1644  * Any disk blocks which *are* part of the previous, committing
1645  * transaction (and which therefore cannot be discarded immediately) are
1646  * not going to be reused in the new running transaction
1647  *
1648  *  The bitmap committed_data images guarantee this: any block which is
1649  *  allocated in one transaction and removed in the next will be marked
1650  *  as in-use in the committed_data bitmap, so cannot be reused until
1651  *  the next transaction to delete the block commits.  This means that
1652  *  leaving committing buffers dirty is quite safe: the disk blocks
1653  *  cannot be reallocated to a different file and so buffer aliasing is
1654  *  not possible.
1655  *
1656  *
1657  * The above applies mainly to ordered data mode.  In writeback mode we
1658  * don't make guarantees about the order in which data hits disk --- in
1659  * particular we don't guarantee that new dirty data is flushed before
1660  * transaction commit --- so it is always safe just to discard data
1661  * immediately in that mode.  --sct
1662  */
1663
1664 /*
1665  * The journal_unmap_buffer helper function returns zero if the buffer
1666  * concerned remains pinned as an anonymous buffer belonging to an older
1667  * transaction.
1668  *
1669  * We're outside-transaction here.  Either or both of j_running_transaction
1670  * and j_committing_transaction may be NULL.
1671  */
1672 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1673 {
1674         transaction_t *transaction;
1675         struct journal_head *jh;
1676         int may_free = 1;
1677         int ret;
1678
1679         BUFFER_TRACE(bh, "entry");
1680
1681         /*
1682          * It is safe to proceed here without the j_list_lock because the
1683          * buffers cannot be stolen by try_to_free_buffers as long as we are
1684          * holding the page lock. --sct
1685          */
1686
1687         if (!buffer_jbd(bh))
1688                 goto zap_buffer_unlocked;
1689
1690         /* OK, we have data buffer in journaled mode */
1691         spin_lock(&journal->j_state_lock);
1692         jbd_lock_bh_state(bh);
1693         spin_lock(&journal->j_list_lock);
1694
1695         jh = jbd2_journal_grab_journal_head(bh);
1696         if (!jh)
1697                 goto zap_buffer_no_jh;
1698
1699         transaction = jh->b_transaction;
1700         if (transaction == NULL) {
1701                 /* First case: not on any transaction.  If it
1702                  * has no checkpoint link, then we can zap it:
1703                  * it's a writeback-mode buffer so we don't care
1704                  * if it hits disk safely. */
1705                 if (!jh->b_cp_transaction) {
1706                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1707                         goto zap_buffer;
1708                 }
1709
1710                 if (!buffer_dirty(bh)) {
1711                         /* bdflush has written it.  We can drop it now */
1712                         goto zap_buffer;
1713                 }
1714
1715                 /* OK, it must be in the journal but still not
1716                  * written fully to disk: it's metadata or
1717                  * journaled data... */
1718
1719                 if (journal->j_running_transaction) {
1720                         /* ... and once the current transaction has
1721                          * committed, the buffer won't be needed any
1722                          * longer. */
1723                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1724                         ret = __dispose_buffer(jh,
1725                                         journal->j_running_transaction);
1726                         jbd2_journal_put_journal_head(jh);
1727                         spin_unlock(&journal->j_list_lock);
1728                         jbd_unlock_bh_state(bh);
1729                         spin_unlock(&journal->j_state_lock);
1730                         return ret;
1731                 } else {
1732                         /* There is no currently-running transaction. So the
1733                          * orphan record which we wrote for this file must have
1734                          * passed into commit.  We must attach this buffer to
1735                          * the committing transaction, if it exists. */
1736                         if (journal->j_committing_transaction) {
1737                                 JBUFFER_TRACE(jh, "give to committing trans");
1738                                 ret = __dispose_buffer(jh,
1739                                         journal->j_committing_transaction);
1740                                 jbd2_journal_put_journal_head(jh);
1741                                 spin_unlock(&journal->j_list_lock);
1742                                 jbd_unlock_bh_state(bh);
1743                                 spin_unlock(&journal->j_state_lock);
1744                                 return ret;
1745                         } else {
1746                                 /* The orphan record's transaction has
1747                                  * committed.  We can cleanse this buffer */
1748                                 clear_buffer_jbddirty(bh);
1749                                 goto zap_buffer;
1750                         }
1751                 }
1752         } else if (transaction == journal->j_committing_transaction) {
1753                 JBUFFER_TRACE(jh, "on committing transaction");
1754                 /*
1755                  * If it is committing, we simply cannot touch it.  We
1756                  * can remove it's next_transaction pointer from the
1757                  * running transaction if that is set, but nothing
1758                  * else. */
1759                 set_buffer_freed(bh);
1760                 if (jh->b_next_transaction) {
1761                         J_ASSERT(jh->b_next_transaction ==
1762                                         journal->j_running_transaction);
1763                         jh->b_next_transaction = NULL;
1764                 }
1765                 jbd2_journal_put_journal_head(jh);
1766                 spin_unlock(&journal->j_list_lock);
1767                 jbd_unlock_bh_state(bh);
1768                 spin_unlock(&journal->j_state_lock);
1769                 return 0;
1770         } else {
1771                 /* Good, the buffer belongs to the running transaction.
1772                  * We are writing our own transaction's data, not any
1773                  * previous one's, so it is safe to throw it away
1774                  * (remember that we expect the filesystem to have set
1775                  * i_size already for this truncate so recovery will not
1776                  * expose the disk blocks we are discarding here.) */
1777                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1778                 JBUFFER_TRACE(jh, "on running transaction");
1779                 may_free = __dispose_buffer(jh, transaction);
1780         }
1781
1782 zap_buffer:
1783         jbd2_journal_put_journal_head(jh);
1784 zap_buffer_no_jh:
1785         spin_unlock(&journal->j_list_lock);
1786         jbd_unlock_bh_state(bh);
1787         spin_unlock(&journal->j_state_lock);
1788 zap_buffer_unlocked:
1789         clear_buffer_dirty(bh);
1790         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1791         clear_buffer_mapped(bh);
1792         clear_buffer_req(bh);
1793         clear_buffer_new(bh);
1794         bh->b_bdev = NULL;
1795         return may_free;
1796 }
1797
1798 /**
1799  * void jbd2_journal_invalidatepage()
1800  * @journal: journal to use for flush...
1801  * @page:    page to flush
1802  * @offset:  length of page to invalidate.
1803  *
1804  * Reap page buffers containing data after offset in page.
1805  *
1806  */
1807 void jbd2_journal_invalidatepage(journal_t *journal,
1808                       struct page *page,
1809                       unsigned long offset)
1810 {
1811         struct buffer_head *head, *bh, *next;
1812         unsigned int curr_off = 0;
1813         int may_free = 1;
1814
1815         if (!PageLocked(page))
1816                 BUG();
1817         if (!page_has_buffers(page))
1818                 return;
1819
1820         /* We will potentially be playing with lists other than just the
1821          * data lists (especially for journaled data mode), so be
1822          * cautious in our locking. */
1823
1824         head = bh = page_buffers(page);
1825         do {
1826                 unsigned int next_off = curr_off + bh->b_size;
1827                 next = bh->b_this_page;
1828
1829                 if (offset <= curr_off) {
1830                         /* This block is wholly outside the truncation point */
1831                         lock_buffer(bh);
1832                         may_free &= journal_unmap_buffer(journal, bh);
1833                         unlock_buffer(bh);
1834                 }
1835                 curr_off = next_off;
1836                 bh = next;
1837
1838         } while (bh != head);
1839
1840         if (!offset) {
1841                 if (may_free && try_to_free_buffers(page))
1842                         J_ASSERT(!page_has_buffers(page));
1843         }
1844 }
1845
1846 /*
1847  * File a buffer on the given transaction list.
1848  */
1849 void __jbd2_journal_file_buffer(struct journal_head *jh,
1850                         transaction_t *transaction, int jlist)
1851 {
1852         struct journal_head **list = NULL;
1853         int was_dirty = 0;
1854         struct buffer_head *bh = jh2bh(jh);
1855
1856         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1857         assert_spin_locked(&transaction->t_journal->j_list_lock);
1858
1859         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1860         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1861                                 jh->b_transaction == NULL);
1862
1863         if (jh->b_transaction && jh->b_jlist == jlist)
1864                 return;
1865
1866         /* The following list of buffer states needs to be consistent
1867          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1868          * state. */
1869
1870         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1871             jlist == BJ_Shadow || jlist == BJ_Forget) {
1872                 if (test_clear_buffer_dirty(bh) ||
1873                     test_clear_buffer_jbddirty(bh))
1874                         was_dirty = 1;
1875         }
1876
1877         if (jh->b_transaction)
1878                 __jbd2_journal_temp_unlink_buffer(jh);
1879         jh->b_transaction = transaction;
1880
1881         switch (jlist) {
1882         case BJ_None:
1883                 J_ASSERT_JH(jh, !jh->b_committed_data);
1884                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1885                 return;
1886         case BJ_Metadata:
1887                 transaction->t_nr_buffers++;
1888                 list = &transaction->t_buffers;
1889                 break;
1890         case BJ_Forget:
1891                 list = &transaction->t_forget;
1892                 break;
1893         case BJ_IO:
1894                 list = &transaction->t_iobuf_list;
1895                 break;
1896         case BJ_Shadow:
1897                 list = &transaction->t_shadow_list;
1898                 break;
1899         case BJ_LogCtl:
1900                 list = &transaction->t_log_list;
1901                 break;
1902         case BJ_Reserved:
1903                 list = &transaction->t_reserved_list;
1904                 break;
1905         }
1906
1907         __blist_add_buffer(list, jh);
1908         jh->b_jlist = jlist;
1909
1910         if (was_dirty)
1911                 set_buffer_jbddirty(bh);
1912 }
1913
1914 void jbd2_journal_file_buffer(struct journal_head *jh,
1915                                 transaction_t *transaction, int jlist)
1916 {
1917         jbd_lock_bh_state(jh2bh(jh));
1918         spin_lock(&transaction->t_journal->j_list_lock);
1919         __jbd2_journal_file_buffer(jh, transaction, jlist);
1920         spin_unlock(&transaction->t_journal->j_list_lock);
1921         jbd_unlock_bh_state(jh2bh(jh));
1922 }
1923
1924 /*
1925  * Remove a buffer from its current buffer list in preparation for
1926  * dropping it from its current transaction entirely.  If the buffer has
1927  * already started to be used by a subsequent transaction, refile the
1928  * buffer on that transaction's metadata list.
1929  *
1930  * Called under journal->j_list_lock
1931  *
1932  * Called under jbd_lock_bh_state(jh2bh(jh))
1933  */
1934 void __jbd2_journal_refile_buffer(struct journal_head *jh)
1935 {
1936         int was_dirty;
1937         struct buffer_head *bh = jh2bh(jh);
1938
1939         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1940         if (jh->b_transaction)
1941                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
1942
1943         /* If the buffer is now unused, just drop it. */
1944         if (jh->b_next_transaction == NULL) {
1945                 __jbd2_journal_unfile_buffer(jh);
1946                 return;
1947         }
1948
1949         /*
1950          * It has been modified by a later transaction: add it to the new
1951          * transaction's metadata list.
1952          */
1953
1954         was_dirty = test_clear_buffer_jbddirty(bh);
1955         __jbd2_journal_temp_unlink_buffer(jh);
1956         jh->b_transaction = jh->b_next_transaction;
1957         jh->b_next_transaction = NULL;
1958         __jbd2_journal_file_buffer(jh, jh->b_transaction,
1959                                 jh->b_modified ? BJ_Metadata : BJ_Reserved);
1960         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
1961
1962         if (was_dirty)
1963                 set_buffer_jbddirty(bh);
1964 }
1965
1966 /*
1967  * For the unlocked version of this call, also make sure that any
1968  * hanging journal_head is cleaned up if necessary.
1969  *
1970  * __jbd2_journal_refile_buffer is usually called as part of a single locked
1971  * operation on a buffer_head, in which the caller is probably going to
1972  * be hooking the journal_head onto other lists.  In that case it is up
1973  * to the caller to remove the journal_head if necessary.  For the
1974  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
1975  * doing anything else to the buffer so we need to do the cleanup
1976  * ourselves to avoid a jh leak.
1977  *
1978  * *** The journal_head may be freed by this call! ***
1979  */
1980 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
1981 {
1982         struct buffer_head *bh = jh2bh(jh);
1983
1984         jbd_lock_bh_state(bh);
1985         spin_lock(&journal->j_list_lock);
1986
1987         __jbd2_journal_refile_buffer(jh);
1988         jbd_unlock_bh_state(bh);
1989         jbd2_journal_remove_journal_head(bh);
1990
1991         spin_unlock(&journal->j_list_lock);
1992         __brelse(bh);
1993 }
1994
1995 /*
1996  * File inode in the inode list of the handle's transaction
1997  */
1998 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
1999 {
2000         transaction_t *transaction = handle->h_transaction;
2001         journal_t *journal = transaction->t_journal;
2002
2003         if (is_handle_aborted(handle))
2004                 return -EIO;
2005
2006         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2007                         transaction->t_tid);
2008
2009         /*
2010          * First check whether inode isn't already on the transaction's
2011          * lists without taking the lock. Note that this check is safe
2012          * without the lock as we cannot race with somebody removing inode
2013          * from the transaction. The reason is that we remove inode from the
2014          * transaction only in journal_release_jbd_inode() and when we commit
2015          * the transaction. We are guarded from the first case by holding
2016          * a reference to the inode. We are safe against the second case
2017          * because if jinode->i_transaction == transaction, commit code
2018          * cannot touch the transaction because we hold reference to it,
2019          * and if jinode->i_next_transaction == transaction, commit code
2020          * will only file the inode where we want it.
2021          */
2022         if (jinode->i_transaction == transaction ||
2023             jinode->i_next_transaction == transaction)
2024                 return 0;
2025
2026         spin_lock(&journal->j_list_lock);
2027
2028         if (jinode->i_transaction == transaction ||
2029             jinode->i_next_transaction == transaction)
2030                 goto done;
2031
2032         /* On some different transaction's list - should be
2033          * the committing one */
2034         if (jinode->i_transaction) {
2035                 J_ASSERT(jinode->i_next_transaction == NULL);
2036                 J_ASSERT(jinode->i_transaction ==
2037                                         journal->j_committing_transaction);
2038                 jinode->i_next_transaction = transaction;
2039                 goto done;
2040         }
2041         /* Not on any transaction list... */
2042         J_ASSERT(!jinode->i_next_transaction);
2043         jinode->i_transaction = transaction;
2044         list_add(&jinode->i_list, &transaction->t_inode_list);
2045 done:
2046         spin_unlock(&journal->j_list_lock);
2047
2048         return 0;
2049 }
2050
2051 /*
2052  * This function must be called when inode is journaled in ordered mode
2053  * before truncation happens. It starts writeout of truncated part in
2054  * case it is in the committing transaction so that we stand to ordered
2055  * mode consistency guarantees.
2056  */
2057 int jbd2_journal_begin_ordered_truncate(struct jbd2_inode *inode,
2058                                         loff_t new_size)
2059 {
2060         journal_t *journal;
2061         transaction_t *commit_trans;
2062         int ret = 0;
2063
2064         if (!inode->i_transaction && !inode->i_next_transaction)
2065                 goto out;
2066         journal = inode->i_transaction->t_journal;
2067         spin_lock(&journal->j_state_lock);
2068         commit_trans = journal->j_committing_transaction;
2069         spin_unlock(&journal->j_state_lock);
2070         if (inode->i_transaction == commit_trans) {
2071                 ret = filemap_fdatawrite_range(inode->i_vfs_inode->i_mapping,
2072                         new_size, LLONG_MAX);
2073                 if (ret)
2074                         jbd2_journal_abort(journal, ret);
2075         }
2076 out:
2077         return ret;
2078 }