jbd2: Improve scalability by not taking j_state_lock in jbd2_journal_stop()
[linux-2.6.git] / 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_up(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
242         lock_map_acquire(&handle->h_lockdep_map);
243 out:
244         if (unlikely(new_transaction))          /* It's usually NULL */
245                 kfree(new_transaction);
246         return ret;
247 }
248
249 static struct lock_class_key jbd2_handle_key;
250
251 /* Allocate a new handle.  This should probably be in a slab... */
252 static handle_t *new_handle(int nblocks)
253 {
254         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
255         if (!handle)
256                 return NULL;
257         memset(handle, 0, sizeof(*handle));
258         handle->h_buffer_credits = nblocks;
259         handle->h_ref = 1;
260
261         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
262                                                 &jbd2_handle_key, 0);
263
264         return handle;
265 }
266
267 /**
268  * handle_t *jbd2_journal_start() - Obtain a new handle.
269  * @journal: Journal to start transaction on.
270  * @nblocks: number of block buffer we might modify
271  *
272  * We make sure that the transaction can guarantee at least nblocks of
273  * modified buffers in the log.  We block until the log can guarantee
274  * that much space.
275  *
276  * This function is visible to journal users (like ext3fs), so is not
277  * called with the journal already locked.
278  *
279  * Return a pointer to a newly allocated handle, or NULL on failure
280  */
281 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
282 {
283         handle_t *handle = journal_current_handle();
284         int err;
285
286         if (!journal)
287                 return ERR_PTR(-EROFS);
288
289         if (handle) {
290                 J_ASSERT(handle->h_transaction->t_journal == journal);
291                 handle->h_ref++;
292                 return handle;
293         }
294
295         handle = new_handle(nblocks);
296         if (!handle)
297                 return ERR_PTR(-ENOMEM);
298
299         current->journal_info = handle;
300
301         err = start_this_handle(journal, handle);
302         if (err < 0) {
303                 jbd2_free_handle(handle);
304                 current->journal_info = NULL;
305                 handle = ERR_PTR(err);
306                 goto out;
307         }
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         lock_map_release(&handle->h_lockdep_map);
430         handle->h_buffer_credits = nblocks;
431         ret = start_this_handle(journal, handle);
432         return ret;
433 }
434
435
436 /**
437  * void jbd2_journal_lock_updates () - establish a transaction barrier.
438  * @journal:  Journal to establish a barrier on.
439  *
440  * This locks out any further updates from being started, and blocks
441  * until all existing updates have completed, returning only once the
442  * journal is in a quiescent state with no updates running.
443  *
444  * The journal lock should not be held on entry.
445  */
446 void jbd2_journal_lock_updates(journal_t *journal)
447 {
448         DEFINE_WAIT(wait);
449
450         spin_lock(&journal->j_state_lock);
451         ++journal->j_barrier_count;
452
453         /* Wait until there are no running updates */
454         while (1) {
455                 transaction_t *transaction = journal->j_running_transaction;
456
457                 if (!transaction)
458                         break;
459
460                 spin_lock(&transaction->t_handle_lock);
461                 if (!transaction->t_updates) {
462                         spin_unlock(&transaction->t_handle_lock);
463                         break;
464                 }
465                 prepare_to_wait(&journal->j_wait_updates, &wait,
466                                 TASK_UNINTERRUPTIBLE);
467                 spin_unlock(&transaction->t_handle_lock);
468                 spin_unlock(&journal->j_state_lock);
469                 schedule();
470                 finish_wait(&journal->j_wait_updates, &wait);
471                 spin_lock(&journal->j_state_lock);
472         }
473         spin_unlock(&journal->j_state_lock);
474
475         /*
476          * We have now established a barrier against other normal updates, but
477          * we also need to barrier against other jbd2_journal_lock_updates() calls
478          * to make sure that we serialise special journal-locked operations
479          * too.
480          */
481         mutex_lock(&journal->j_barrier);
482 }
483
484 /**
485  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
486  * @journal:  Journal to release the barrier on.
487  *
488  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
489  *
490  * Should be called without the journal lock held.
491  */
492 void jbd2_journal_unlock_updates (journal_t *journal)
493 {
494         J_ASSERT(journal->j_barrier_count != 0);
495
496         mutex_unlock(&journal->j_barrier);
497         spin_lock(&journal->j_state_lock);
498         --journal->j_barrier_count;
499         spin_unlock(&journal->j_state_lock);
500         wake_up(&journal->j_wait_transaction_locked);
501 }
502
503 static void warn_dirty_buffer(struct buffer_head *bh)
504 {
505         char b[BDEVNAME_SIZE];
506
507         printk(KERN_WARNING
508                "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
509                "There's a risk of filesystem corruption in case of system "
510                "crash.\n",
511                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
512 }
513
514 /*
515  * If the buffer is already part of the current transaction, then there
516  * is nothing we need to do.  If it is already part of a prior
517  * transaction which we are still committing to disk, then we need to
518  * make sure that we do not overwrite the old copy: we do copy-out to
519  * preserve the copy going to disk.  We also account the buffer against
520  * the handle's metadata buffer credits (unless the buffer is already
521  * part of the transaction, that is).
522  *
523  */
524 static int
525 do_get_write_access(handle_t *handle, struct journal_head *jh,
526                         int force_copy)
527 {
528         struct buffer_head *bh;
529         transaction_t *transaction;
530         journal_t *journal;
531         int error;
532         char *frozen_buffer = NULL;
533         int need_copy = 0;
534
535         if (is_handle_aborted(handle))
536                 return -EROFS;
537
538         transaction = handle->h_transaction;
539         journal = transaction->t_journal;
540
541         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
542
543         JBUFFER_TRACE(jh, "entry");
544 repeat:
545         bh = jh2bh(jh);
546
547         /* @@@ Need to check for errors here at some point. */
548
549         lock_buffer(bh);
550         jbd_lock_bh_state(bh);
551
552         /* We now hold the buffer lock so it is safe to query the buffer
553          * state.  Is the buffer dirty?
554          *
555          * If so, there are two possibilities.  The buffer may be
556          * non-journaled, and undergoing a quite legitimate writeback.
557          * Otherwise, it is journaled, and we don't expect dirty buffers
558          * in that state (the buffers should be marked JBD_Dirty
559          * instead.)  So either the IO is being done under our own
560          * control and this is a bug, or it's a third party IO such as
561          * dump(8) (which may leave the buffer scheduled for read ---
562          * ie. locked but not dirty) or tune2fs (which may actually have
563          * the buffer dirtied, ugh.)  */
564
565         if (buffer_dirty(bh)) {
566                 /*
567                  * First question: is this buffer already part of the current
568                  * transaction or the existing committing transaction?
569                  */
570                 if (jh->b_transaction) {
571                         J_ASSERT_JH(jh,
572                                 jh->b_transaction == transaction ||
573                                 jh->b_transaction ==
574                                         journal->j_committing_transaction);
575                         if (jh->b_next_transaction)
576                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
577                                                         transaction);
578                         warn_dirty_buffer(bh);
579                 }
580                 /*
581                  * In any case we need to clean the dirty flag and we must
582                  * do it under the buffer lock to be sure we don't race
583                  * with running write-out.
584                  */
585                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
586                 clear_buffer_dirty(bh);
587                 set_buffer_jbddirty(bh);
588         }
589
590         unlock_buffer(bh);
591
592         error = -EROFS;
593         if (is_handle_aborted(handle)) {
594                 jbd_unlock_bh_state(bh);
595                 goto out;
596         }
597         error = 0;
598
599         /*
600          * The buffer is already part of this transaction if b_transaction or
601          * b_next_transaction points to it
602          */
603         if (jh->b_transaction == transaction ||
604             jh->b_next_transaction == transaction)
605                 goto done;
606
607         /*
608          * this is the first time this transaction is touching this buffer,
609          * reset the modified flag
610          */
611        jh->b_modified = 0;
612
613         /*
614          * If there is already a copy-out version of this buffer, then we don't
615          * need to make another one
616          */
617         if (jh->b_frozen_data) {
618                 JBUFFER_TRACE(jh, "has frozen data");
619                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
620                 jh->b_next_transaction = transaction;
621                 goto done;
622         }
623
624         /* Is there data here we need to preserve? */
625
626         if (jh->b_transaction && jh->b_transaction != transaction) {
627                 JBUFFER_TRACE(jh, "owned by older transaction");
628                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
629                 J_ASSERT_JH(jh, jh->b_transaction ==
630                                         journal->j_committing_transaction);
631
632                 /* There is one case we have to be very careful about.
633                  * If the committing transaction is currently writing
634                  * this buffer out to disk and has NOT made a copy-out,
635                  * then we cannot modify the buffer contents at all
636                  * right now.  The essence of copy-out is that it is the
637                  * extra copy, not the primary copy, which gets
638                  * journaled.  If the primary copy is already going to
639                  * disk then we cannot do copy-out here. */
640
641                 if (jh->b_jlist == BJ_Shadow) {
642                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
643                         wait_queue_head_t *wqh;
644
645                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
646
647                         JBUFFER_TRACE(jh, "on shadow: sleep");
648                         jbd_unlock_bh_state(bh);
649                         /* commit wakes up all shadow buffers after IO */
650                         for ( ; ; ) {
651                                 prepare_to_wait(wqh, &wait.wait,
652                                                 TASK_UNINTERRUPTIBLE);
653                                 if (jh->b_jlist != BJ_Shadow)
654                                         break;
655                                 schedule();
656                         }
657                         finish_wait(wqh, &wait.wait);
658                         goto repeat;
659                 }
660
661                 /* Only do the copy if the currently-owning transaction
662                  * still needs it.  If it is on the Forget list, the
663                  * committing transaction is past that stage.  The
664                  * buffer had better remain locked during the kmalloc,
665                  * but that should be true --- we hold the journal lock
666                  * still and the buffer is already on the BUF_JOURNAL
667                  * list so won't be flushed.
668                  *
669                  * Subtle point, though: if this is a get_undo_access,
670                  * then we will be relying on the frozen_data to contain
671                  * the new value of the committed_data record after the
672                  * transaction, so we HAVE to force the frozen_data copy
673                  * in that case. */
674
675                 if (jh->b_jlist != BJ_Forget || force_copy) {
676                         JBUFFER_TRACE(jh, "generate frozen data");
677                         if (!frozen_buffer) {
678                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
679                                 jbd_unlock_bh_state(bh);
680                                 frozen_buffer =
681                                         jbd2_alloc(jh2bh(jh)->b_size,
682                                                          GFP_NOFS);
683                                 if (!frozen_buffer) {
684                                         printk(KERN_EMERG
685                                                "%s: OOM for frozen_buffer\n",
686                                                __func__);
687                                         JBUFFER_TRACE(jh, "oom!");
688                                         error = -ENOMEM;
689                                         jbd_lock_bh_state(bh);
690                                         goto done;
691                                 }
692                                 goto repeat;
693                         }
694                         jh->b_frozen_data = frozen_buffer;
695                         frozen_buffer = NULL;
696                         need_copy = 1;
697                 }
698                 jh->b_next_transaction = transaction;
699         }
700
701
702         /*
703          * Finally, if the buffer is not journaled right now, we need to make
704          * sure it doesn't get written to disk before the caller actually
705          * commits the new data
706          */
707         if (!jh->b_transaction) {
708                 JBUFFER_TRACE(jh, "no transaction");
709                 J_ASSERT_JH(jh, !jh->b_next_transaction);
710                 jh->b_transaction = transaction;
711                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
712                 spin_lock(&journal->j_list_lock);
713                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
714                 spin_unlock(&journal->j_list_lock);
715         }
716
717 done:
718         if (need_copy) {
719                 struct page *page;
720                 int offset;
721                 char *source;
722
723                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
724                             "Possible IO failure.\n");
725                 page = jh2bh(jh)->b_page;
726                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
727                 source = kmap_atomic(page, KM_USER0);
728                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
729                 kunmap_atomic(source, KM_USER0);
730
731                 /*
732                  * Now that the frozen data is saved off, we need to store
733                  * any matching triggers.
734                  */
735                 jh->b_frozen_triggers = jh->b_triggers;
736         }
737         jbd_unlock_bh_state(bh);
738
739         /*
740          * If we are about to journal a buffer, then any revoke pending on it is
741          * no longer valid
742          */
743         jbd2_journal_cancel_revoke(handle, jh);
744
745 out:
746         if (unlikely(frozen_buffer))    /* It's usually NULL */
747                 jbd2_free(frozen_buffer, bh->b_size);
748
749         JBUFFER_TRACE(jh, "exit");
750         return error;
751 }
752
753 /**
754  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
755  * @handle: transaction to add buffer modifications to
756  * @bh:     bh to be used for metadata writes
757  * @credits: variable that will receive credits for the buffer
758  *
759  * Returns an error code or 0 on success.
760  *
761  * In full data journalling mode the buffer may be of type BJ_AsyncData,
762  * because we're write()ing a buffer which is also part of a shared mapping.
763  */
764
765 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
766 {
767         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
768         int rc;
769
770         /* We do not want to get caught playing with fields which the
771          * log thread also manipulates.  Make sure that the buffer
772          * completes any outstanding IO before proceeding. */
773         rc = do_get_write_access(handle, jh, 0);
774         jbd2_journal_put_journal_head(jh);
775         return rc;
776 }
777
778
779 /*
780  * When the user wants to journal a newly created buffer_head
781  * (ie. getblk() returned a new buffer and we are going to populate it
782  * manually rather than reading off disk), then we need to keep the
783  * buffer_head locked until it has been completely filled with new
784  * data.  In this case, we should be able to make the assertion that
785  * the bh is not already part of an existing transaction.
786  *
787  * The buffer should already be locked by the caller by this point.
788  * There is no lock ranking violation: it was a newly created,
789  * unlocked buffer beforehand. */
790
791 /**
792  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
793  * @handle: transaction to new buffer to
794  * @bh: new buffer.
795  *
796  * Call this if you create a new bh.
797  */
798 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
799 {
800         transaction_t *transaction = handle->h_transaction;
801         journal_t *journal = transaction->t_journal;
802         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
803         int err;
804
805         jbd_debug(5, "journal_head %p\n", jh);
806         err = -EROFS;
807         if (is_handle_aborted(handle))
808                 goto out;
809         err = 0;
810
811         JBUFFER_TRACE(jh, "entry");
812         /*
813          * The buffer may already belong to this transaction due to pre-zeroing
814          * in the filesystem's new_block code.  It may also be on the previous,
815          * committing transaction's lists, but it HAS to be in Forget state in
816          * that case: the transaction must have deleted the buffer for it to be
817          * reused here.
818          */
819         jbd_lock_bh_state(bh);
820         spin_lock(&journal->j_list_lock);
821         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
822                 jh->b_transaction == NULL ||
823                 (jh->b_transaction == journal->j_committing_transaction &&
824                           jh->b_jlist == BJ_Forget)));
825
826         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
827         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
828
829         if (jh->b_transaction == NULL) {
830                 /*
831                  * Previous jbd2_journal_forget() could have left the buffer
832                  * with jbddirty bit set because it was being committed. When
833                  * the commit finished, we've filed the buffer for
834                  * checkpointing and marked it dirty. Now we are reallocating
835                  * the buffer so the transaction freeing it must have
836                  * committed and so it's safe to clear the dirty bit.
837                  */
838                 clear_buffer_dirty(jh2bh(jh));
839                 jh->b_transaction = transaction;
840
841                 /* first access by this transaction */
842                 jh->b_modified = 0;
843
844                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
845                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
846         } else if (jh->b_transaction == journal->j_committing_transaction) {
847                 /* first access by this transaction */
848                 jh->b_modified = 0;
849
850                 JBUFFER_TRACE(jh, "set next transaction");
851                 jh->b_next_transaction = transaction;
852         }
853         spin_unlock(&journal->j_list_lock);
854         jbd_unlock_bh_state(bh);
855
856         /*
857          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
858          * blocks which contain freed but then revoked metadata.  We need
859          * to cancel the revoke in case we end up freeing it yet again
860          * and the reallocating as data - this would cause a second revoke,
861          * which hits an assertion error.
862          */
863         JBUFFER_TRACE(jh, "cancelling revoke");
864         jbd2_journal_cancel_revoke(handle, jh);
865         jbd2_journal_put_journal_head(jh);
866 out:
867         return err;
868 }
869
870 /**
871  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
872  *     non-rewindable consequences
873  * @handle: transaction
874  * @bh: buffer to undo
875  * @credits: store the number of taken credits here (if not NULL)
876  *
877  * Sometimes there is a need to distinguish between metadata which has
878  * been committed to disk and that which has not.  The ext3fs code uses
879  * this for freeing and allocating space, we have to make sure that we
880  * do not reuse freed space until the deallocation has been committed,
881  * since if we overwrote that space we would make the delete
882  * un-rewindable in case of a crash.
883  *
884  * To deal with that, jbd2_journal_get_undo_access requests write access to a
885  * buffer for parts of non-rewindable operations such as delete
886  * operations on the bitmaps.  The journaling code must keep a copy of
887  * the buffer's contents prior to the undo_access call until such time
888  * as we know that the buffer has definitely been committed to disk.
889  *
890  * We never need to know which transaction the committed data is part
891  * of, buffers touched here are guaranteed to be dirtied later and so
892  * will be committed to a new transaction in due course, at which point
893  * we can discard the old committed data pointer.
894  *
895  * Returns error number or 0 on success.
896  */
897 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
898 {
899         int err;
900         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
901         char *committed_data = NULL;
902
903         JBUFFER_TRACE(jh, "entry");
904
905         /*
906          * Do this first --- it can drop the journal lock, so we want to
907          * make sure that obtaining the committed_data is done
908          * atomically wrt. completion of any outstanding commits.
909          */
910         err = do_get_write_access(handle, jh, 1);
911         if (err)
912                 goto out;
913
914 repeat:
915         if (!jh->b_committed_data) {
916                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
917                 if (!committed_data) {
918                         printk(KERN_EMERG "%s: No memory for committed data\n",
919                                 __func__);
920                         err = -ENOMEM;
921                         goto out;
922                 }
923         }
924
925         jbd_lock_bh_state(bh);
926         if (!jh->b_committed_data) {
927                 /* Copy out the current buffer contents into the
928                  * preserved, committed copy. */
929                 JBUFFER_TRACE(jh, "generate b_committed data");
930                 if (!committed_data) {
931                         jbd_unlock_bh_state(bh);
932                         goto repeat;
933                 }
934
935                 jh->b_committed_data = committed_data;
936                 committed_data = NULL;
937                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
938         }
939         jbd_unlock_bh_state(bh);
940 out:
941         jbd2_journal_put_journal_head(jh);
942         if (unlikely(committed_data))
943                 jbd2_free(committed_data, bh->b_size);
944         return err;
945 }
946
947 /**
948  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
949  * @bh: buffer to trigger on
950  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
951  *
952  * Set any triggers on this journal_head.  This is always safe, because
953  * triggers for a committing buffer will be saved off, and triggers for
954  * a running transaction will match the buffer in that transaction.
955  *
956  * Call with NULL to clear the triggers.
957  */
958 void jbd2_journal_set_triggers(struct buffer_head *bh,
959                                struct jbd2_buffer_trigger_type *type)
960 {
961         struct journal_head *jh = bh2jh(bh);
962
963         jh->b_triggers = type;
964 }
965
966 void jbd2_buffer_commit_trigger(struct journal_head *jh, void *mapped_data,
967                                 struct jbd2_buffer_trigger_type *triggers)
968 {
969         struct buffer_head *bh = jh2bh(jh);
970
971         if (!triggers || !triggers->t_commit)
972                 return;
973
974         triggers->t_commit(triggers, bh, mapped_data, bh->b_size);
975 }
976
977 void jbd2_buffer_abort_trigger(struct journal_head *jh,
978                                struct jbd2_buffer_trigger_type *triggers)
979 {
980         if (!triggers || !triggers->t_abort)
981                 return;
982
983         triggers->t_abort(triggers, jh2bh(jh));
984 }
985
986
987
988 /**
989  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
990  * @handle: transaction to add buffer to.
991  * @bh: buffer to mark
992  *
993  * mark dirty metadata which needs to be journaled as part of the current
994  * transaction.
995  *
996  * The buffer is placed on the transaction's metadata list and is marked
997  * as belonging to the transaction.
998  *
999  * Returns error number or 0 on success.
1000  *
1001  * Special care needs to be taken if the buffer already belongs to the
1002  * current committing transaction (in which case we should have frozen
1003  * data present for that commit).  In that case, we don't relink the
1004  * buffer: that only gets done when the old transaction finally
1005  * completes its commit.
1006  */
1007 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1008 {
1009         transaction_t *transaction = handle->h_transaction;
1010         journal_t *journal = transaction->t_journal;
1011         struct journal_head *jh = bh2jh(bh);
1012
1013         jbd_debug(5, "journal_head %p\n", jh);
1014         JBUFFER_TRACE(jh, "entry");
1015         if (is_handle_aborted(handle))
1016                 goto out;
1017
1018         jbd_lock_bh_state(bh);
1019
1020         if (jh->b_modified == 0) {
1021                 /*
1022                  * This buffer's got modified and becoming part
1023                  * of the transaction. This needs to be done
1024                  * once a transaction -bzzz
1025                  */
1026                 jh->b_modified = 1;
1027                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1028                 handle->h_buffer_credits--;
1029         }
1030
1031         /*
1032          * fastpath, to avoid expensive locking.  If this buffer is already
1033          * on the running transaction's metadata list there is nothing to do.
1034          * Nobody can take it off again because there is a handle open.
1035          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1036          * result in this test being false, so we go in and take the locks.
1037          */
1038         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1039                 JBUFFER_TRACE(jh, "fastpath");
1040                 J_ASSERT_JH(jh, jh->b_transaction ==
1041                                         journal->j_running_transaction);
1042                 goto out_unlock_bh;
1043         }
1044
1045         set_buffer_jbddirty(bh);
1046
1047         /*
1048          * Metadata already on the current transaction list doesn't
1049          * need to be filed.  Metadata on another transaction's list must
1050          * be committing, and will be refiled once the commit completes:
1051          * leave it alone for now.
1052          */
1053         if (jh->b_transaction != transaction) {
1054                 JBUFFER_TRACE(jh, "already on other transaction");
1055                 J_ASSERT_JH(jh, jh->b_transaction ==
1056                                         journal->j_committing_transaction);
1057                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1058                 /* And this case is illegal: we can't reuse another
1059                  * transaction's data buffer, ever. */
1060                 goto out_unlock_bh;
1061         }
1062
1063         /* That test should have eliminated the following case: */
1064         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1065
1066         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1067         spin_lock(&journal->j_list_lock);
1068         __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1069         spin_unlock(&journal->j_list_lock);
1070 out_unlock_bh:
1071         jbd_unlock_bh_state(bh);
1072 out:
1073         JBUFFER_TRACE(jh, "exit");
1074         return 0;
1075 }
1076
1077 /*
1078  * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1079  * updates, if the update decided in the end that it didn't need access.
1080  *
1081  */
1082 void
1083 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1084 {
1085         BUFFER_TRACE(bh, "entry");
1086 }
1087
1088 /**
1089  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1090  * @handle: transaction handle
1091  * @bh:     bh to 'forget'
1092  *
1093  * We can only do the bforget if there are no commits pending against the
1094  * buffer.  If the buffer is dirty in the current running transaction we
1095  * can safely unlink it.
1096  *
1097  * bh may not be a journalled buffer at all - it may be a non-JBD
1098  * buffer which came off the hashtable.  Check for this.
1099  *
1100  * Decrements bh->b_count by one.
1101  *
1102  * Allow this call even if the handle has aborted --- it may be part of
1103  * the caller's cleanup after an abort.
1104  */
1105 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1106 {
1107         transaction_t *transaction = handle->h_transaction;
1108         journal_t *journal = transaction->t_journal;
1109         struct journal_head *jh;
1110         int drop_reserve = 0;
1111         int err = 0;
1112         int was_modified = 0;
1113
1114         BUFFER_TRACE(bh, "entry");
1115
1116         jbd_lock_bh_state(bh);
1117         spin_lock(&journal->j_list_lock);
1118
1119         if (!buffer_jbd(bh))
1120                 goto not_jbd;
1121         jh = bh2jh(bh);
1122
1123         /* Critical error: attempting to delete a bitmap buffer, maybe?
1124          * Don't do any jbd operations, and return an error. */
1125         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1126                          "inconsistent data on disk")) {
1127                 err = -EIO;
1128                 goto not_jbd;
1129         }
1130
1131         /* keep track of wether or not this transaction modified us */
1132         was_modified = jh->b_modified;
1133
1134         /*
1135          * The buffer's going from the transaction, we must drop
1136          * all references -bzzz
1137          */
1138         jh->b_modified = 0;
1139
1140         if (jh->b_transaction == handle->h_transaction) {
1141                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1142
1143                 /* If we are forgetting a buffer which is already part
1144                  * of this transaction, then we can just drop it from
1145                  * the transaction immediately. */
1146                 clear_buffer_dirty(bh);
1147                 clear_buffer_jbddirty(bh);
1148
1149                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1150
1151                 /*
1152                  * we only want to drop a reference if this transaction
1153                  * modified the buffer
1154                  */
1155                 if (was_modified)
1156                         drop_reserve = 1;
1157
1158                 /*
1159                  * We are no longer going to journal this buffer.
1160                  * However, the commit of this transaction is still
1161                  * important to the buffer: the delete that we are now
1162                  * processing might obsolete an old log entry, so by
1163                  * committing, we can satisfy the buffer's checkpoint.
1164                  *
1165                  * So, if we have a checkpoint on the buffer, we should
1166                  * now refile the buffer on our BJ_Forget list so that
1167                  * we know to remove the checkpoint after we commit.
1168                  */
1169
1170                 if (jh->b_cp_transaction) {
1171                         __jbd2_journal_temp_unlink_buffer(jh);
1172                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1173                 } else {
1174                         __jbd2_journal_unfile_buffer(jh);
1175                         jbd2_journal_remove_journal_head(bh);
1176                         __brelse(bh);
1177                         if (!buffer_jbd(bh)) {
1178                                 spin_unlock(&journal->j_list_lock);
1179                                 jbd_unlock_bh_state(bh);
1180                                 __bforget(bh);
1181                                 goto drop;
1182                         }
1183                 }
1184         } else if (jh->b_transaction) {
1185                 J_ASSERT_JH(jh, (jh->b_transaction ==
1186                                  journal->j_committing_transaction));
1187                 /* However, if the buffer is still owned by a prior
1188                  * (committing) transaction, we can't drop it yet... */
1189                 JBUFFER_TRACE(jh, "belongs to older transaction");
1190                 /* ... but we CAN drop it from the new transaction if we
1191                  * have also modified it since the original commit. */
1192
1193                 if (jh->b_next_transaction) {
1194                         J_ASSERT(jh->b_next_transaction == transaction);
1195                         jh->b_next_transaction = NULL;
1196
1197                         /*
1198                          * only drop a reference if this transaction modified
1199                          * the buffer
1200                          */
1201                         if (was_modified)
1202                                 drop_reserve = 1;
1203                 }
1204         }
1205
1206 not_jbd:
1207         spin_unlock(&journal->j_list_lock);
1208         jbd_unlock_bh_state(bh);
1209         __brelse(bh);
1210 drop:
1211         if (drop_reserve) {
1212                 /* no need to reserve log space for this block -bzzz */
1213                 handle->h_buffer_credits++;
1214         }
1215         return err;
1216 }
1217
1218 /**
1219  * int jbd2_journal_stop() - complete a transaction
1220  * @handle: tranaction to complete.
1221  *
1222  * All done for a particular handle.
1223  *
1224  * There is not much action needed here.  We just return any remaining
1225  * buffer credits to the transaction and remove the handle.  The only
1226  * complication is that we need to start a commit operation if the
1227  * filesystem is marked for synchronous update.
1228  *
1229  * jbd2_journal_stop itself will not usually return an error, but it may
1230  * do so in unusual circumstances.  In particular, expect it to
1231  * return -EIO if a jbd2_journal_abort has been executed since the
1232  * transaction began.
1233  */
1234 int jbd2_journal_stop(handle_t *handle)
1235 {
1236         transaction_t *transaction = handle->h_transaction;
1237         journal_t *journal = transaction->t_journal;
1238         int err;
1239         pid_t pid;
1240
1241         J_ASSERT(journal_current_handle() == handle);
1242
1243         if (is_handle_aborted(handle))
1244                 err = -EIO;
1245         else {
1246                 J_ASSERT(transaction->t_updates > 0);
1247                 err = 0;
1248         }
1249
1250         if (--handle->h_ref > 0) {
1251                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1252                           handle->h_ref);
1253                 return err;
1254         }
1255
1256         jbd_debug(4, "Handle %p going down\n", handle);
1257
1258         /*
1259          * Implement synchronous transaction batching.  If the handle
1260          * was synchronous, don't force a commit immediately.  Let's
1261          * yield and let another thread piggyback onto this
1262          * transaction.  Keep doing that while new threads continue to
1263          * arrive.  It doesn't cost much - we're about to run a commit
1264          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1265          * operations by 30x or more...
1266          *
1267          * We try and optimize the sleep time against what the
1268          * underlying disk can do, instead of having a static sleep
1269          * time.  This is useful for the case where our storage is so
1270          * fast that it is more optimal to go ahead and force a flush
1271          * and wait for the transaction to be committed than it is to
1272          * wait for an arbitrary amount of time for new writers to
1273          * join the transaction.  We achieve this by measuring how
1274          * long it takes to commit a transaction, and compare it with
1275          * how long this transaction has been running, and if run time
1276          * < commit time then we sleep for the delta and commit.  This
1277          * greatly helps super fast disks that would see slowdowns as
1278          * more threads started doing fsyncs.
1279          *
1280          * But don't do this if this process was the most recent one
1281          * to perform a synchronous write.  We do this to detect the
1282          * case where a single process is doing a stream of sync
1283          * writes.  No point in waiting for joiners in that case.
1284          */
1285         pid = current->pid;
1286         if (handle->h_sync && journal->j_last_sync_writer != pid) {
1287                 u64 commit_time, trans_time;
1288
1289                 journal->j_last_sync_writer = pid;
1290
1291                 spin_lock(&journal->j_state_lock);
1292                 commit_time = journal->j_average_commit_time;
1293                 spin_unlock(&journal->j_state_lock);
1294
1295                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1296                                                    transaction->t_start_time));
1297
1298                 commit_time = max_t(u64, commit_time,
1299                                     1000*journal->j_min_batch_time);
1300                 commit_time = min_t(u64, commit_time,
1301                                     1000*journal->j_max_batch_time);
1302
1303                 if (trans_time < commit_time) {
1304                         ktime_t expires = ktime_add_ns(ktime_get(),
1305                                                        commit_time);
1306                         set_current_state(TASK_UNINTERRUPTIBLE);
1307                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1308                 }
1309         }
1310
1311         if (handle->h_sync)
1312                 transaction->t_synchronous_commit = 1;
1313         current->journal_info = NULL;
1314         spin_lock(&transaction->t_handle_lock);
1315         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1316         transaction->t_updates--;
1317         if (!transaction->t_updates) {
1318                 wake_up(&journal->j_wait_updates);
1319                 if (journal->j_barrier_count)
1320                         wake_up(&journal->j_wait_transaction_locked);
1321         }
1322
1323         /*
1324          * If the handle is marked SYNC, we need to set another commit
1325          * going!  We also want to force a commit if the current
1326          * transaction is occupying too much of the log, or if the
1327          * transaction is too old now.
1328          */
1329         if (handle->h_sync ||
1330                         transaction->t_outstanding_credits >
1331                                 journal->j_max_transaction_buffers ||
1332                         time_after_eq(jiffies, transaction->t_expires)) {
1333                 /* Do this even for aborted journals: an abort still
1334                  * completes the commit thread, it just doesn't write
1335                  * anything to disk. */
1336                 tid_t tid = transaction->t_tid;
1337
1338                 spin_unlock(&transaction->t_handle_lock);
1339                 jbd_debug(2, "transaction too old, requesting commit for "
1340                                         "handle %p\n", handle);
1341                 /* This is non-blocking */
1342                 jbd2_log_start_commit(journal, transaction->t_tid);
1343
1344                 /*
1345                  * Special case: JBD2_SYNC synchronous updates require us
1346                  * to wait for the commit to complete.
1347                  */
1348                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1349                         err = jbd2_log_wait_commit(journal, tid);
1350         } else {
1351                 spin_unlock(&transaction->t_handle_lock);
1352         }
1353
1354         lock_map_release(&handle->h_lockdep_map);
1355
1356         jbd2_free_handle(handle);
1357         return err;
1358 }
1359
1360 /**
1361  * int jbd2_journal_force_commit() - force any uncommitted transactions
1362  * @journal: journal to force
1363  *
1364  * For synchronous operations: force any uncommitted transactions
1365  * to disk.  May seem kludgy, but it reuses all the handle batching
1366  * code in a very simple manner.
1367  */
1368 int jbd2_journal_force_commit(journal_t *journal)
1369 {
1370         handle_t *handle;
1371         int ret;
1372
1373         handle = jbd2_journal_start(journal, 1);
1374         if (IS_ERR(handle)) {
1375                 ret = PTR_ERR(handle);
1376         } else {
1377                 handle->h_sync = 1;
1378                 ret = jbd2_journal_stop(handle);
1379         }
1380         return ret;
1381 }
1382
1383 /*
1384  *
1385  * List management code snippets: various functions for manipulating the
1386  * transaction buffer lists.
1387  *
1388  */
1389
1390 /*
1391  * Append a buffer to a transaction list, given the transaction's list head
1392  * pointer.
1393  *
1394  * j_list_lock is held.
1395  *
1396  * jbd_lock_bh_state(jh2bh(jh)) is held.
1397  */
1398
1399 static inline void
1400 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1401 {
1402         if (!*list) {
1403                 jh->b_tnext = jh->b_tprev = jh;
1404                 *list = jh;
1405         } else {
1406                 /* Insert at the tail of the list to preserve order */
1407                 struct journal_head *first = *list, *last = first->b_tprev;
1408                 jh->b_tprev = last;
1409                 jh->b_tnext = first;
1410                 last->b_tnext = first->b_tprev = jh;
1411         }
1412 }
1413
1414 /*
1415  * Remove a buffer from a transaction list, given the transaction's list
1416  * head pointer.
1417  *
1418  * Called with j_list_lock held, and the journal may not be locked.
1419  *
1420  * jbd_lock_bh_state(jh2bh(jh)) is held.
1421  */
1422
1423 static inline void
1424 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1425 {
1426         if (*list == jh) {
1427                 *list = jh->b_tnext;
1428                 if (*list == jh)
1429                         *list = NULL;
1430         }
1431         jh->b_tprev->b_tnext = jh->b_tnext;
1432         jh->b_tnext->b_tprev = jh->b_tprev;
1433 }
1434
1435 /*
1436  * Remove a buffer from the appropriate transaction list.
1437  *
1438  * Note that this function can *change* the value of
1439  * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1440  * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1441  * of these pointers, it could go bad.  Generally the caller needs to re-read
1442  * the pointer from the transaction_t.
1443  *
1444  * Called under j_list_lock.  The journal may not be locked.
1445  */
1446 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1447 {
1448         struct journal_head **list = NULL;
1449         transaction_t *transaction;
1450         struct buffer_head *bh = jh2bh(jh);
1451
1452         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1453         transaction = jh->b_transaction;
1454         if (transaction)
1455                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1456
1457         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1458         if (jh->b_jlist != BJ_None)
1459                 J_ASSERT_JH(jh, transaction != NULL);
1460
1461         switch (jh->b_jlist) {
1462         case BJ_None:
1463                 return;
1464         case BJ_Metadata:
1465                 transaction->t_nr_buffers--;
1466                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1467                 list = &transaction->t_buffers;
1468                 break;
1469         case BJ_Forget:
1470                 list = &transaction->t_forget;
1471                 break;
1472         case BJ_IO:
1473                 list = &transaction->t_iobuf_list;
1474                 break;
1475         case BJ_Shadow:
1476                 list = &transaction->t_shadow_list;
1477                 break;
1478         case BJ_LogCtl:
1479                 list = &transaction->t_log_list;
1480                 break;
1481         case BJ_Reserved:
1482                 list = &transaction->t_reserved_list;
1483                 break;
1484         }
1485
1486         __blist_del_buffer(list, jh);
1487         jh->b_jlist = BJ_None;
1488         if (test_clear_buffer_jbddirty(bh))
1489                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1490 }
1491
1492 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1493 {
1494         __jbd2_journal_temp_unlink_buffer(jh);
1495         jh->b_transaction = NULL;
1496 }
1497
1498 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1499 {
1500         jbd_lock_bh_state(jh2bh(jh));
1501         spin_lock(&journal->j_list_lock);
1502         __jbd2_journal_unfile_buffer(jh);
1503         spin_unlock(&journal->j_list_lock);
1504         jbd_unlock_bh_state(jh2bh(jh));
1505 }
1506
1507 /*
1508  * Called from jbd2_journal_try_to_free_buffers().
1509  *
1510  * Called under jbd_lock_bh_state(bh)
1511  */
1512 static void
1513 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1514 {
1515         struct journal_head *jh;
1516
1517         jh = bh2jh(bh);
1518
1519         if (buffer_locked(bh) || buffer_dirty(bh))
1520                 goto out;
1521
1522         if (jh->b_next_transaction != NULL)
1523                 goto out;
1524
1525         spin_lock(&journal->j_list_lock);
1526         if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1527                 /* written-back checkpointed metadata buffer */
1528                 if (jh->b_jlist == BJ_None) {
1529                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1530                         __jbd2_journal_remove_checkpoint(jh);
1531                         jbd2_journal_remove_journal_head(bh);
1532                         __brelse(bh);
1533                 }
1534         }
1535         spin_unlock(&journal->j_list_lock);
1536 out:
1537         return;
1538 }
1539
1540 /**
1541  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1542  * @journal: journal for operation
1543  * @page: to try and free
1544  * @gfp_mask: we use the mask to detect how hard should we try to release
1545  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1546  * release the buffers.
1547  *
1548  *
1549  * For all the buffers on this page,
1550  * if they are fully written out ordered data, move them onto BUF_CLEAN
1551  * so try_to_free_buffers() can reap them.
1552  *
1553  * This function returns non-zero if we wish try_to_free_buffers()
1554  * to be called. We do this if the page is releasable by try_to_free_buffers().
1555  * We also do it if the page has locked or dirty buffers and the caller wants
1556  * us to perform sync or async writeout.
1557  *
1558  * This complicates JBD locking somewhat.  We aren't protected by the
1559  * BKL here.  We wish to remove the buffer from its committing or
1560  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1561  *
1562  * This may *change* the value of transaction_t->t_datalist, so anyone
1563  * who looks at t_datalist needs to lock against this function.
1564  *
1565  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1566  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1567  * will come out of the lock with the buffer dirty, which makes it
1568  * ineligible for release here.
1569  *
1570  * Who else is affected by this?  hmm...  Really the only contender
1571  * is do_get_write_access() - it could be looking at the buffer while
1572  * journal_try_to_free_buffer() is changing its state.  But that
1573  * cannot happen because we never reallocate freed data as metadata
1574  * while the data is part of a transaction.  Yes?
1575  *
1576  * Return 0 on failure, 1 on success
1577  */
1578 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1579                                 struct page *page, gfp_t gfp_mask)
1580 {
1581         struct buffer_head *head;
1582         struct buffer_head *bh;
1583         int ret = 0;
1584
1585         J_ASSERT(PageLocked(page));
1586
1587         head = page_buffers(page);
1588         bh = head;
1589         do {
1590                 struct journal_head *jh;
1591
1592                 /*
1593                  * We take our own ref against the journal_head here to avoid
1594                  * having to add tons of locking around each instance of
1595                  * jbd2_journal_remove_journal_head() and
1596                  * jbd2_journal_put_journal_head().
1597                  */
1598                 jh = jbd2_journal_grab_journal_head(bh);
1599                 if (!jh)
1600                         continue;
1601
1602                 jbd_lock_bh_state(bh);
1603                 __journal_try_to_free_buffer(journal, bh);
1604                 jbd2_journal_put_journal_head(jh);
1605                 jbd_unlock_bh_state(bh);
1606                 if (buffer_jbd(bh))
1607                         goto busy;
1608         } while ((bh = bh->b_this_page) != head);
1609
1610         ret = try_to_free_buffers(page);
1611
1612 busy:
1613         return ret;
1614 }
1615
1616 /*
1617  * This buffer is no longer needed.  If it is on an older transaction's
1618  * checkpoint list we need to record it on this transaction's forget list
1619  * to pin this buffer (and hence its checkpointing transaction) down until
1620  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1621  * release it.
1622  * Returns non-zero if JBD no longer has an interest in the buffer.
1623  *
1624  * Called under j_list_lock.
1625  *
1626  * Called under jbd_lock_bh_state(bh).
1627  */
1628 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1629 {
1630         int may_free = 1;
1631         struct buffer_head *bh = jh2bh(jh);
1632
1633         __jbd2_journal_unfile_buffer(jh);
1634
1635         if (jh->b_cp_transaction) {
1636                 JBUFFER_TRACE(jh, "on running+cp transaction");
1637                 /*
1638                  * We don't want to write the buffer anymore, clear the
1639                  * bit so that we don't confuse checks in
1640                  * __journal_file_buffer
1641                  */
1642                 clear_buffer_dirty(bh);
1643                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1644                 may_free = 0;
1645         } else {
1646                 JBUFFER_TRACE(jh, "on running transaction");
1647                 jbd2_journal_remove_journal_head(bh);
1648                 __brelse(bh);
1649         }
1650         return may_free;
1651 }
1652
1653 /*
1654  * jbd2_journal_invalidatepage
1655  *
1656  * This code is tricky.  It has a number of cases to deal with.
1657  *
1658  * There are two invariants which this code relies on:
1659  *
1660  * i_size must be updated on disk before we start calling invalidatepage on the
1661  * data.
1662  *
1663  *  This is done in ext3 by defining an ext3_setattr method which
1664  *  updates i_size before truncate gets going.  By maintaining this
1665  *  invariant, we can be sure that it is safe to throw away any buffers
1666  *  attached to the current transaction: once the transaction commits,
1667  *  we know that the data will not be needed.
1668  *
1669  *  Note however that we can *not* throw away data belonging to the
1670  *  previous, committing transaction!
1671  *
1672  * Any disk blocks which *are* part of the previous, committing
1673  * transaction (and which therefore cannot be discarded immediately) are
1674  * not going to be reused in the new running transaction
1675  *
1676  *  The bitmap committed_data images guarantee this: any block which is
1677  *  allocated in one transaction and removed in the next will be marked
1678  *  as in-use in the committed_data bitmap, so cannot be reused until
1679  *  the next transaction to delete the block commits.  This means that
1680  *  leaving committing buffers dirty is quite safe: the disk blocks
1681  *  cannot be reallocated to a different file and so buffer aliasing is
1682  *  not possible.
1683  *
1684  *
1685  * The above applies mainly to ordered data mode.  In writeback mode we
1686  * don't make guarantees about the order in which data hits disk --- in
1687  * particular we don't guarantee that new dirty data is flushed before
1688  * transaction commit --- so it is always safe just to discard data
1689  * immediately in that mode.  --sct
1690  */
1691
1692 /*
1693  * The journal_unmap_buffer helper function returns zero if the buffer
1694  * concerned remains pinned as an anonymous buffer belonging to an older
1695  * transaction.
1696  *
1697  * We're outside-transaction here.  Either or both of j_running_transaction
1698  * and j_committing_transaction may be NULL.
1699  */
1700 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1701 {
1702         transaction_t *transaction;
1703         struct journal_head *jh;
1704         int may_free = 1;
1705         int ret;
1706
1707         BUFFER_TRACE(bh, "entry");
1708
1709         /*
1710          * It is safe to proceed here without the j_list_lock because the
1711          * buffers cannot be stolen by try_to_free_buffers as long as we are
1712          * holding the page lock. --sct
1713          */
1714
1715         if (!buffer_jbd(bh))
1716                 goto zap_buffer_unlocked;
1717
1718         /* OK, we have data buffer in journaled mode */
1719         spin_lock(&journal->j_state_lock);
1720         jbd_lock_bh_state(bh);
1721         spin_lock(&journal->j_list_lock);
1722
1723         jh = jbd2_journal_grab_journal_head(bh);
1724         if (!jh)
1725                 goto zap_buffer_no_jh;
1726
1727         /*
1728          * We cannot remove the buffer from checkpoint lists until the
1729          * transaction adding inode to orphan list (let's call it T)
1730          * is committed.  Otherwise if the transaction changing the
1731          * buffer would be cleaned from the journal before T is
1732          * committed, a crash will cause that the correct contents of
1733          * the buffer will be lost.  On the other hand we have to
1734          * clear the buffer dirty bit at latest at the moment when the
1735          * transaction marking the buffer as freed in the filesystem
1736          * structures is committed because from that moment on the
1737          * buffer can be reallocated and used by a different page.
1738          * Since the block hasn't been freed yet but the inode has
1739          * already been added to orphan list, it is safe for us to add
1740          * the buffer to BJ_Forget list of the newest transaction.
1741          */
1742         transaction = jh->b_transaction;
1743         if (transaction == NULL) {
1744                 /* First case: not on any transaction.  If it
1745                  * has no checkpoint link, then we can zap it:
1746                  * it's a writeback-mode buffer so we don't care
1747                  * if it hits disk safely. */
1748                 if (!jh->b_cp_transaction) {
1749                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1750                         goto zap_buffer;
1751                 }
1752
1753                 if (!buffer_dirty(bh)) {
1754                         /* bdflush has written it.  We can drop it now */
1755                         goto zap_buffer;
1756                 }
1757
1758                 /* OK, it must be in the journal but still not
1759                  * written fully to disk: it's metadata or
1760                  * journaled data... */
1761
1762                 if (journal->j_running_transaction) {
1763                         /* ... and once the current transaction has
1764                          * committed, the buffer won't be needed any
1765                          * longer. */
1766                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1767                         ret = __dispose_buffer(jh,
1768                                         journal->j_running_transaction);
1769                         jbd2_journal_put_journal_head(jh);
1770                         spin_unlock(&journal->j_list_lock);
1771                         jbd_unlock_bh_state(bh);
1772                         spin_unlock(&journal->j_state_lock);
1773                         return ret;
1774                 } else {
1775                         /* There is no currently-running transaction. So the
1776                          * orphan record which we wrote for this file must have
1777                          * passed into commit.  We must attach this buffer to
1778                          * the committing transaction, if it exists. */
1779                         if (journal->j_committing_transaction) {
1780                                 JBUFFER_TRACE(jh, "give to committing trans");
1781                                 ret = __dispose_buffer(jh,
1782                                         journal->j_committing_transaction);
1783                                 jbd2_journal_put_journal_head(jh);
1784                                 spin_unlock(&journal->j_list_lock);
1785                                 jbd_unlock_bh_state(bh);
1786                                 spin_unlock(&journal->j_state_lock);
1787                                 return ret;
1788                         } else {
1789                                 /* The orphan record's transaction has
1790                                  * committed.  We can cleanse this buffer */
1791                                 clear_buffer_jbddirty(bh);
1792                                 goto zap_buffer;
1793                         }
1794                 }
1795         } else if (transaction == journal->j_committing_transaction) {
1796                 JBUFFER_TRACE(jh, "on committing transaction");
1797                 /*
1798                  * The buffer is committing, we simply cannot touch
1799                  * it. So we just set j_next_transaction to the
1800                  * running transaction (if there is one) and mark
1801                  * buffer as freed so that commit code knows it should
1802                  * clear dirty bits when it is done with the buffer.
1803                  */
1804                 set_buffer_freed(bh);
1805                 if (journal->j_running_transaction && buffer_jbddirty(bh))
1806                         jh->b_next_transaction = journal->j_running_transaction;
1807                 jbd2_journal_put_journal_head(jh);
1808                 spin_unlock(&journal->j_list_lock);
1809                 jbd_unlock_bh_state(bh);
1810                 spin_unlock(&journal->j_state_lock);
1811                 return 0;
1812         } else {
1813                 /* Good, the buffer belongs to the running transaction.
1814                  * We are writing our own transaction's data, not any
1815                  * previous one's, so it is safe to throw it away
1816                  * (remember that we expect the filesystem to have set
1817                  * i_size already for this truncate so recovery will not
1818                  * expose the disk blocks we are discarding here.) */
1819                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1820                 JBUFFER_TRACE(jh, "on running transaction");
1821                 may_free = __dispose_buffer(jh, transaction);
1822         }
1823
1824 zap_buffer:
1825         jbd2_journal_put_journal_head(jh);
1826 zap_buffer_no_jh:
1827         spin_unlock(&journal->j_list_lock);
1828         jbd_unlock_bh_state(bh);
1829         spin_unlock(&journal->j_state_lock);
1830 zap_buffer_unlocked:
1831         clear_buffer_dirty(bh);
1832         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1833         clear_buffer_mapped(bh);
1834         clear_buffer_req(bh);
1835         clear_buffer_new(bh);
1836         bh->b_bdev = NULL;
1837         return may_free;
1838 }
1839
1840 /**
1841  * void jbd2_journal_invalidatepage()
1842  * @journal: journal to use for flush...
1843  * @page:    page to flush
1844  * @offset:  length of page to invalidate.
1845  *
1846  * Reap page buffers containing data after offset in page.
1847  *
1848  */
1849 void jbd2_journal_invalidatepage(journal_t *journal,
1850                       struct page *page,
1851                       unsigned long offset)
1852 {
1853         struct buffer_head *head, *bh, *next;
1854         unsigned int curr_off = 0;
1855         int may_free = 1;
1856
1857         if (!PageLocked(page))
1858                 BUG();
1859         if (!page_has_buffers(page))
1860                 return;
1861
1862         /* We will potentially be playing with lists other than just the
1863          * data lists (especially for journaled data mode), so be
1864          * cautious in our locking. */
1865
1866         head = bh = page_buffers(page);
1867         do {
1868                 unsigned int next_off = curr_off + bh->b_size;
1869                 next = bh->b_this_page;
1870
1871                 if (offset <= curr_off) {
1872                         /* This block is wholly outside the truncation point */
1873                         lock_buffer(bh);
1874                         may_free &= journal_unmap_buffer(journal, bh);
1875                         unlock_buffer(bh);
1876                 }
1877                 curr_off = next_off;
1878                 bh = next;
1879
1880         } while (bh != head);
1881
1882         if (!offset) {
1883                 if (may_free && try_to_free_buffers(page))
1884                         J_ASSERT(!page_has_buffers(page));
1885         }
1886 }
1887
1888 /*
1889  * File a buffer on the given transaction list.
1890  */
1891 void __jbd2_journal_file_buffer(struct journal_head *jh,
1892                         transaction_t *transaction, int jlist)
1893 {
1894         struct journal_head **list = NULL;
1895         int was_dirty = 0;
1896         struct buffer_head *bh = jh2bh(jh);
1897
1898         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1899         assert_spin_locked(&transaction->t_journal->j_list_lock);
1900
1901         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1902         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1903                                 jh->b_transaction == NULL);
1904
1905         if (jh->b_transaction && jh->b_jlist == jlist)
1906                 return;
1907
1908         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1909             jlist == BJ_Shadow || jlist == BJ_Forget) {
1910                 /*
1911                  * For metadata buffers, we track dirty bit in buffer_jbddirty
1912                  * instead of buffer_dirty. We should not see a dirty bit set
1913                  * here because we clear it in do_get_write_access but e.g.
1914                  * tune2fs can modify the sb and set the dirty bit at any time
1915                  * so we try to gracefully handle that.
1916                  */
1917                 if (buffer_dirty(bh))
1918                         warn_dirty_buffer(bh);
1919                 if (test_clear_buffer_dirty(bh) ||
1920                     test_clear_buffer_jbddirty(bh))
1921                         was_dirty = 1;
1922         }
1923
1924         if (jh->b_transaction)
1925                 __jbd2_journal_temp_unlink_buffer(jh);
1926         jh->b_transaction = transaction;
1927
1928         switch (jlist) {
1929         case BJ_None:
1930                 J_ASSERT_JH(jh, !jh->b_committed_data);
1931                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1932                 return;
1933         case BJ_Metadata:
1934                 transaction->t_nr_buffers++;
1935                 list = &transaction->t_buffers;
1936                 break;
1937         case BJ_Forget:
1938                 list = &transaction->t_forget;
1939                 break;
1940         case BJ_IO:
1941                 list = &transaction->t_iobuf_list;
1942                 break;
1943         case BJ_Shadow:
1944                 list = &transaction->t_shadow_list;
1945                 break;
1946         case BJ_LogCtl:
1947                 list = &transaction->t_log_list;
1948                 break;
1949         case BJ_Reserved:
1950                 list = &transaction->t_reserved_list;
1951                 break;
1952         }
1953
1954         __blist_add_buffer(list, jh);
1955         jh->b_jlist = jlist;
1956
1957         if (was_dirty)
1958                 set_buffer_jbddirty(bh);
1959 }
1960
1961 void jbd2_journal_file_buffer(struct journal_head *jh,
1962                                 transaction_t *transaction, int jlist)
1963 {
1964         jbd_lock_bh_state(jh2bh(jh));
1965         spin_lock(&transaction->t_journal->j_list_lock);
1966         __jbd2_journal_file_buffer(jh, transaction, jlist);
1967         spin_unlock(&transaction->t_journal->j_list_lock);
1968         jbd_unlock_bh_state(jh2bh(jh));
1969 }
1970
1971 /*
1972  * Remove a buffer from its current buffer list in preparation for
1973  * dropping it from its current transaction entirely.  If the buffer has
1974  * already started to be used by a subsequent transaction, refile the
1975  * buffer on that transaction's metadata list.
1976  *
1977  * Called under journal->j_list_lock
1978  *
1979  * Called under jbd_lock_bh_state(jh2bh(jh))
1980  */
1981 void __jbd2_journal_refile_buffer(struct journal_head *jh)
1982 {
1983         int was_dirty, jlist;
1984         struct buffer_head *bh = jh2bh(jh);
1985
1986         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1987         if (jh->b_transaction)
1988                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
1989
1990         /* If the buffer is now unused, just drop it. */
1991         if (jh->b_next_transaction == NULL) {
1992                 __jbd2_journal_unfile_buffer(jh);
1993                 return;
1994         }
1995
1996         /*
1997          * It has been modified by a later transaction: add it to the new
1998          * transaction's metadata list.
1999          */
2000
2001         was_dirty = test_clear_buffer_jbddirty(bh);
2002         __jbd2_journal_temp_unlink_buffer(jh);
2003         jh->b_transaction = jh->b_next_transaction;
2004         jh->b_next_transaction = NULL;
2005         if (buffer_freed(bh))
2006                 jlist = BJ_Forget;
2007         else if (jh->b_modified)
2008                 jlist = BJ_Metadata;
2009         else
2010                 jlist = BJ_Reserved;
2011         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2012         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2013
2014         if (was_dirty)
2015                 set_buffer_jbddirty(bh);
2016 }
2017
2018 /*
2019  * For the unlocked version of this call, also make sure that any
2020  * hanging journal_head is cleaned up if necessary.
2021  *
2022  * __jbd2_journal_refile_buffer is usually called as part of a single locked
2023  * operation on a buffer_head, in which the caller is probably going to
2024  * be hooking the journal_head onto other lists.  In that case it is up
2025  * to the caller to remove the journal_head if necessary.  For the
2026  * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2027  * doing anything else to the buffer so we need to do the cleanup
2028  * ourselves to avoid a jh leak.
2029  *
2030  * *** The journal_head may be freed by this call! ***
2031  */
2032 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2033 {
2034         struct buffer_head *bh = jh2bh(jh);
2035
2036         jbd_lock_bh_state(bh);
2037         spin_lock(&journal->j_list_lock);
2038
2039         __jbd2_journal_refile_buffer(jh);
2040         jbd_unlock_bh_state(bh);
2041         jbd2_journal_remove_journal_head(bh);
2042
2043         spin_unlock(&journal->j_list_lock);
2044         __brelse(bh);
2045 }
2046
2047 /*
2048  * File inode in the inode list of the handle's transaction
2049  */
2050 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2051 {
2052         transaction_t *transaction = handle->h_transaction;
2053         journal_t *journal = transaction->t_journal;
2054
2055         if (is_handle_aborted(handle))
2056                 return -EIO;
2057
2058         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2059                         transaction->t_tid);
2060
2061         /*
2062          * First check whether inode isn't already on the transaction's
2063          * lists without taking the lock. Note that this check is safe
2064          * without the lock as we cannot race with somebody removing inode
2065          * from the transaction. The reason is that we remove inode from the
2066          * transaction only in journal_release_jbd_inode() and when we commit
2067          * the transaction. We are guarded from the first case by holding
2068          * a reference to the inode. We are safe against the second case
2069          * because if jinode->i_transaction == transaction, commit code
2070          * cannot touch the transaction because we hold reference to it,
2071          * and if jinode->i_next_transaction == transaction, commit code
2072          * will only file the inode where we want it.
2073          */
2074         if (jinode->i_transaction == transaction ||
2075             jinode->i_next_transaction == transaction)
2076                 return 0;
2077
2078         spin_lock(&journal->j_list_lock);
2079
2080         if (jinode->i_transaction == transaction ||
2081             jinode->i_next_transaction == transaction)
2082                 goto done;
2083
2084         /* On some different transaction's list - should be
2085          * the committing one */
2086         if (jinode->i_transaction) {
2087                 J_ASSERT(jinode->i_next_transaction == NULL);
2088                 J_ASSERT(jinode->i_transaction ==
2089                                         journal->j_committing_transaction);
2090                 jinode->i_next_transaction = transaction;
2091                 goto done;
2092         }
2093         /* Not on any transaction list... */
2094         J_ASSERT(!jinode->i_next_transaction);
2095         jinode->i_transaction = transaction;
2096         list_add(&jinode->i_list, &transaction->t_inode_list);
2097 done:
2098         spin_unlock(&journal->j_list_lock);
2099
2100         return 0;
2101 }
2102
2103 /*
2104  * File truncate and transaction commit interact with each other in a
2105  * non-trivial way.  If a transaction writing data block A is
2106  * committing, we cannot discard the data by truncate until we have
2107  * written them.  Otherwise if we crashed after the transaction with
2108  * write has committed but before the transaction with truncate has
2109  * committed, we could see stale data in block A.  This function is a
2110  * helper to solve this problem.  It starts writeout of the truncated
2111  * part in case it is in the committing transaction.
2112  *
2113  * Filesystem code must call this function when inode is journaled in
2114  * ordered mode before truncation happens and after the inode has been
2115  * placed on orphan list with the new inode size. The second condition
2116  * avoids the race that someone writes new data and we start
2117  * committing the transaction after this function has been called but
2118  * before a transaction for truncate is started (and furthermore it
2119  * allows us to optimize the case where the addition to orphan list
2120  * happens in the same transaction as write --- we don't have to write
2121  * any data in such case).
2122  */
2123 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2124                                         struct jbd2_inode *jinode,
2125                                         loff_t new_size)
2126 {
2127         transaction_t *inode_trans, *commit_trans;
2128         int ret = 0;
2129
2130         /* This is a quick check to avoid locking if not necessary */
2131         if (!jinode->i_transaction)
2132                 goto out;
2133         /* Locks are here just to force reading of recent values, it is
2134          * enough that the transaction was not committing before we started
2135          * a transaction adding the inode to orphan list */
2136         spin_lock(&journal->j_state_lock);
2137         commit_trans = journal->j_committing_transaction;
2138         spin_unlock(&journal->j_state_lock);
2139         spin_lock(&journal->j_list_lock);
2140         inode_trans = jinode->i_transaction;
2141         spin_unlock(&journal->j_list_lock);
2142         if (inode_trans == commit_trans) {
2143                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2144                         new_size, LLONG_MAX);
2145                 if (ret)
2146                         jbd2_journal_abort(journal, ret);
2147         }
2148 out:
2149         return ret;
2150 }