Merge branch 'upstream' of git://git.infradead.org/users/pcmoore/audit
[sfrench/cifs-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 #include <linux/backing-dev.h>
30 #include <linux/bug.h>
31 #include <linux/module.h>
32
33 #include <trace/events/jbd2.h>
34
35 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
36 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
37
38 static struct kmem_cache *transaction_cache;
39 int __init jbd2_journal_init_transaction_cache(void)
40 {
41         J_ASSERT(!transaction_cache);
42         transaction_cache = kmem_cache_create("jbd2_transaction_s",
43                                         sizeof(transaction_t),
44                                         0,
45                                         SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
46                                         NULL);
47         if (transaction_cache)
48                 return 0;
49         return -ENOMEM;
50 }
51
52 void jbd2_journal_destroy_transaction_cache(void)
53 {
54         if (transaction_cache) {
55                 kmem_cache_destroy(transaction_cache);
56                 transaction_cache = NULL;
57         }
58 }
59
60 void jbd2_journal_free_transaction(transaction_t *transaction)
61 {
62         if (unlikely(ZERO_OR_NULL_PTR(transaction)))
63                 return;
64         kmem_cache_free(transaction_cache, transaction);
65 }
66
67 /*
68  * jbd2_get_transaction: obtain a new transaction_t object.
69  *
70  * Simply allocate and initialise a new transaction.  Create it in
71  * RUNNING state and add it to the current journal (which should not
72  * have an existing running transaction: we only make a new transaction
73  * once we have started to commit the old one).
74  *
75  * Preconditions:
76  *      The journal MUST be locked.  We don't perform atomic mallocs on the
77  *      new transaction and we can't block without protecting against other
78  *      processes trying to touch the journal while it is in transition.
79  *
80  */
81
82 static transaction_t *
83 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
84 {
85         transaction->t_journal = journal;
86         transaction->t_state = T_RUNNING;
87         transaction->t_start_time = ktime_get();
88         transaction->t_tid = journal->j_transaction_sequence++;
89         transaction->t_expires = jiffies + journal->j_commit_interval;
90         spin_lock_init(&transaction->t_handle_lock);
91         atomic_set(&transaction->t_updates, 0);
92         atomic_set(&transaction->t_outstanding_credits,
93                    atomic_read(&journal->j_reserved_credits));
94         atomic_set(&transaction->t_handle_count, 0);
95         INIT_LIST_HEAD(&transaction->t_inode_list);
96         INIT_LIST_HEAD(&transaction->t_private_list);
97
98         /* Set up the commit timer for the new transaction. */
99         journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
100         add_timer(&journal->j_commit_timer);
101
102         J_ASSERT(journal->j_running_transaction == NULL);
103         journal->j_running_transaction = transaction;
104         transaction->t_max_wait = 0;
105         transaction->t_start = jiffies;
106         transaction->t_requested = 0;
107
108         return transaction;
109 }
110
111 /*
112  * Handle management.
113  *
114  * A handle_t is an object which represents a single atomic update to a
115  * filesystem, and which tracks all of the modifications which form part
116  * of that one update.
117  */
118
119 /*
120  * Update transaction's maximum wait time, if debugging is enabled.
121  *
122  * In order for t_max_wait to be reliable, it must be protected by a
123  * lock.  But doing so will mean that start_this_handle() can not be
124  * run in parallel on SMP systems, which limits our scalability.  So
125  * unless debugging is enabled, we no longer update t_max_wait, which
126  * means that maximum wait time reported by the jbd2_run_stats
127  * tracepoint will always be zero.
128  */
129 static inline void update_t_max_wait(transaction_t *transaction,
130                                      unsigned long ts)
131 {
132 #ifdef CONFIG_JBD2_DEBUG
133         if (jbd2_journal_enable_debug &&
134             time_after(transaction->t_start, ts)) {
135                 ts = jbd2_time_diff(ts, transaction->t_start);
136                 spin_lock(&transaction->t_handle_lock);
137                 if (ts > transaction->t_max_wait)
138                         transaction->t_max_wait = ts;
139                 spin_unlock(&transaction->t_handle_lock);
140         }
141 #endif
142 }
143
144 /*
145  * Wait until running transaction passes T_LOCKED state. Also starts the commit
146  * if needed. The function expects running transaction to exist and releases
147  * j_state_lock.
148  */
149 static void wait_transaction_locked(journal_t *journal)
150         __releases(journal->j_state_lock)
151 {
152         DEFINE_WAIT(wait);
153         int need_to_start;
154         tid_t tid = journal->j_running_transaction->t_tid;
155
156         prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
157                         TASK_UNINTERRUPTIBLE);
158         need_to_start = !tid_geq(journal->j_commit_request, tid);
159         read_unlock(&journal->j_state_lock);
160         if (need_to_start)
161                 jbd2_log_start_commit(journal, tid);
162         schedule();
163         finish_wait(&journal->j_wait_transaction_locked, &wait);
164 }
165
166 static void sub_reserved_credits(journal_t *journal, int blocks)
167 {
168         atomic_sub(blocks, &journal->j_reserved_credits);
169         wake_up(&journal->j_wait_reserved);
170 }
171
172 /*
173  * Wait until we can add credits for handle to the running transaction.  Called
174  * with j_state_lock held for reading. Returns 0 if handle joined the running
175  * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
176  * caller must retry.
177  */
178 static int add_transaction_credits(journal_t *journal, int blocks,
179                                    int rsv_blocks)
180 {
181         transaction_t *t = journal->j_running_transaction;
182         int needed;
183         int total = blocks + rsv_blocks;
184
185         /*
186          * If the current transaction is locked down for commit, wait
187          * for the lock to be released.
188          */
189         if (t->t_state == T_LOCKED) {
190                 wait_transaction_locked(journal);
191                 return 1;
192         }
193
194         /*
195          * If there is not enough space left in the log to write all
196          * potential buffers requested by this operation, we need to
197          * stall pending a log checkpoint to free some more log space.
198          */
199         needed = atomic_add_return(total, &t->t_outstanding_credits);
200         if (needed > journal->j_max_transaction_buffers) {
201                 /*
202                  * If the current transaction is already too large,
203                  * then start to commit it: we can then go back and
204                  * attach this handle to a new transaction.
205                  */
206                 atomic_sub(total, &t->t_outstanding_credits);
207                 wait_transaction_locked(journal);
208                 return 1;
209         }
210
211         /*
212          * The commit code assumes that it can get enough log space
213          * without forcing a checkpoint.  This is *critical* for
214          * correctness: a checkpoint of a buffer which is also
215          * associated with a committing transaction creates a deadlock,
216          * so commit simply cannot force through checkpoints.
217          *
218          * We must therefore ensure the necessary space in the journal
219          * *before* starting to dirty potentially checkpointed buffers
220          * in the new transaction.
221          */
222         if (jbd2_log_space_left(journal) < jbd2_space_needed(journal)) {
223                 atomic_sub(total, &t->t_outstanding_credits);
224                 read_unlock(&journal->j_state_lock);
225                 write_lock(&journal->j_state_lock);
226                 if (jbd2_log_space_left(journal) < jbd2_space_needed(journal))
227                         __jbd2_log_wait_for_space(journal);
228                 write_unlock(&journal->j_state_lock);
229                 return 1;
230         }
231
232         /* No reservation? We are done... */
233         if (!rsv_blocks)
234                 return 0;
235
236         needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
237         /* We allow at most half of a transaction to be reserved */
238         if (needed > journal->j_max_transaction_buffers / 2) {
239                 sub_reserved_credits(journal, rsv_blocks);
240                 atomic_sub(total, &t->t_outstanding_credits);
241                 read_unlock(&journal->j_state_lock);
242                 wait_event(journal->j_wait_reserved,
243                          atomic_read(&journal->j_reserved_credits) + rsv_blocks
244                          <= journal->j_max_transaction_buffers / 2);
245                 return 1;
246         }
247         return 0;
248 }
249
250 /*
251  * start_this_handle: Given a handle, deal with any locking or stalling
252  * needed to make sure that there is enough journal space for the handle
253  * to begin.  Attach the handle to a transaction and set up the
254  * transaction's buffer credits.
255  */
256
257 static int start_this_handle(journal_t *journal, handle_t *handle,
258                              gfp_t gfp_mask)
259 {
260         transaction_t   *transaction, *new_transaction = NULL;
261         int             blocks = handle->h_buffer_credits;
262         int             rsv_blocks = 0;
263         unsigned long ts = jiffies;
264
265         /*
266          * 1/2 of transaction can be reserved so we can practically handle
267          * only 1/2 of maximum transaction size per operation
268          */
269         if (WARN_ON(blocks > journal->j_max_transaction_buffers / 2)) {
270                 printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
271                        current->comm, blocks,
272                        journal->j_max_transaction_buffers / 2);
273                 return -ENOSPC;
274         }
275
276         if (handle->h_rsv_handle)
277                 rsv_blocks = handle->h_rsv_handle->h_buffer_credits;
278
279 alloc_transaction:
280         if (!journal->j_running_transaction) {
281                 new_transaction = kmem_cache_zalloc(transaction_cache,
282                                                     gfp_mask);
283                 if (!new_transaction) {
284                         /*
285                          * If __GFP_FS is not present, then we may be
286                          * being called from inside the fs writeback
287                          * layer, so we MUST NOT fail.  Since
288                          * __GFP_NOFAIL is going away, we will arrange
289                          * to retry the allocation ourselves.
290                          */
291                         if ((gfp_mask & __GFP_FS) == 0) {
292                                 congestion_wait(BLK_RW_ASYNC, HZ/50);
293                                 goto alloc_transaction;
294                         }
295                         return -ENOMEM;
296                 }
297         }
298
299         jbd_debug(3, "New handle %p going live.\n", handle);
300
301         /*
302          * We need to hold j_state_lock until t_updates has been incremented,
303          * for proper journal barrier handling
304          */
305 repeat:
306         read_lock(&journal->j_state_lock);
307         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
308         if (is_journal_aborted(journal) ||
309             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
310                 read_unlock(&journal->j_state_lock);
311                 jbd2_journal_free_transaction(new_transaction);
312                 return -EROFS;
313         }
314
315         /*
316          * Wait on the journal's transaction barrier if necessary. Specifically
317          * we allow reserved handles to proceed because otherwise commit could
318          * deadlock on page writeback not being able to complete.
319          */
320         if (!handle->h_reserved && journal->j_barrier_count) {
321                 read_unlock(&journal->j_state_lock);
322                 wait_event(journal->j_wait_transaction_locked,
323                                 journal->j_barrier_count == 0);
324                 goto repeat;
325         }
326
327         if (!journal->j_running_transaction) {
328                 read_unlock(&journal->j_state_lock);
329                 if (!new_transaction)
330                         goto alloc_transaction;
331                 write_lock(&journal->j_state_lock);
332                 if (!journal->j_running_transaction &&
333                     (handle->h_reserved || !journal->j_barrier_count)) {
334                         jbd2_get_transaction(journal, new_transaction);
335                         new_transaction = NULL;
336                 }
337                 write_unlock(&journal->j_state_lock);
338                 goto repeat;
339         }
340
341         transaction = journal->j_running_transaction;
342
343         if (!handle->h_reserved) {
344                 /* We may have dropped j_state_lock - restart in that case */
345                 if (add_transaction_credits(journal, blocks, rsv_blocks))
346                         goto repeat;
347         } else {
348                 /*
349                  * We have handle reserved so we are allowed to join T_LOCKED
350                  * transaction and we don't have to check for transaction size
351                  * and journal space.
352                  */
353                 sub_reserved_credits(journal, blocks);
354                 handle->h_reserved = 0;
355         }
356
357         /* OK, account for the buffers that this operation expects to
358          * use and add the handle to the running transaction. 
359          */
360         update_t_max_wait(transaction, ts);
361         handle->h_transaction = transaction;
362         handle->h_requested_credits = blocks;
363         handle->h_start_jiffies = jiffies;
364         atomic_inc(&transaction->t_updates);
365         atomic_inc(&transaction->t_handle_count);
366         jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
367                   handle, blocks,
368                   atomic_read(&transaction->t_outstanding_credits),
369                   jbd2_log_space_left(journal));
370         read_unlock(&journal->j_state_lock);
371         current->journal_info = handle;
372
373         lock_map_acquire(&handle->h_lockdep_map);
374         jbd2_journal_free_transaction(new_transaction);
375         return 0;
376 }
377
378 static struct lock_class_key jbd2_handle_key;
379
380 /* Allocate a new handle.  This should probably be in a slab... */
381 static handle_t *new_handle(int nblocks)
382 {
383         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
384         if (!handle)
385                 return NULL;
386         handle->h_buffer_credits = nblocks;
387         handle->h_ref = 1;
388
389         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
390                                                 &jbd2_handle_key, 0);
391
392         return handle;
393 }
394
395 /**
396  * handle_t *jbd2_journal_start() - Obtain a new handle.
397  * @journal: Journal to start transaction on.
398  * @nblocks: number of block buffer we might modify
399  *
400  * We make sure that the transaction can guarantee at least nblocks of
401  * modified buffers in the log.  We block until the log can guarantee
402  * that much space. Additionally, if rsv_blocks > 0, we also create another
403  * handle with rsv_blocks reserved blocks in the journal. This handle is
404  * is stored in h_rsv_handle. It is not attached to any particular transaction
405  * and thus doesn't block transaction commit. If the caller uses this reserved
406  * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
407  * on the parent handle will dispose the reserved one. Reserved handle has to
408  * be converted to a normal handle using jbd2_journal_start_reserved() before
409  * it can be used.
410  *
411  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
412  * on failure.
413  */
414 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
415                               gfp_t gfp_mask, unsigned int type,
416                               unsigned int line_no)
417 {
418         handle_t *handle = journal_current_handle();
419         int err;
420
421         if (!journal)
422                 return ERR_PTR(-EROFS);
423
424         if (handle) {
425                 J_ASSERT(handle->h_transaction->t_journal == journal);
426                 handle->h_ref++;
427                 return handle;
428         }
429
430         handle = new_handle(nblocks);
431         if (!handle)
432                 return ERR_PTR(-ENOMEM);
433         if (rsv_blocks) {
434                 handle_t *rsv_handle;
435
436                 rsv_handle = new_handle(rsv_blocks);
437                 if (!rsv_handle) {
438                         jbd2_free_handle(handle);
439                         return ERR_PTR(-ENOMEM);
440                 }
441                 rsv_handle->h_reserved = 1;
442                 rsv_handle->h_journal = journal;
443                 handle->h_rsv_handle = rsv_handle;
444         }
445
446         err = start_this_handle(journal, handle, gfp_mask);
447         if (err < 0) {
448                 if (handle->h_rsv_handle)
449                         jbd2_free_handle(handle->h_rsv_handle);
450                 jbd2_free_handle(handle);
451                 return ERR_PTR(err);
452         }
453         handle->h_type = type;
454         handle->h_line_no = line_no;
455         trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
456                                 handle->h_transaction->t_tid, type,
457                                 line_no, nblocks);
458         return handle;
459 }
460 EXPORT_SYMBOL(jbd2__journal_start);
461
462
463 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
464 {
465         return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0);
466 }
467 EXPORT_SYMBOL(jbd2_journal_start);
468
469 void jbd2_journal_free_reserved(handle_t *handle)
470 {
471         journal_t *journal = handle->h_journal;
472
473         WARN_ON(!handle->h_reserved);
474         sub_reserved_credits(journal, handle->h_buffer_credits);
475         jbd2_free_handle(handle);
476 }
477 EXPORT_SYMBOL(jbd2_journal_free_reserved);
478
479 /**
480  * int jbd2_journal_start_reserved(handle_t *handle) - start reserved handle
481  * @handle: handle to start
482  *
483  * Start handle that has been previously reserved with jbd2_journal_reserve().
484  * This attaches @handle to the running transaction (or creates one if there's
485  * not transaction running). Unlike jbd2_journal_start() this function cannot
486  * block on journal commit, checkpointing, or similar stuff. It can block on
487  * memory allocation or frozen journal though.
488  *
489  * Return 0 on success, non-zero on error - handle is freed in that case.
490  */
491 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
492                                 unsigned int line_no)
493 {
494         journal_t *journal = handle->h_journal;
495         int ret = -EIO;
496
497         if (WARN_ON(!handle->h_reserved)) {
498                 /* Someone passed in normal handle? Just stop it. */
499                 jbd2_journal_stop(handle);
500                 return ret;
501         }
502         /*
503          * Usefulness of mixing of reserved and unreserved handles is
504          * questionable. So far nobody seems to need it so just error out.
505          */
506         if (WARN_ON(current->journal_info)) {
507                 jbd2_journal_free_reserved(handle);
508                 return ret;
509         }
510
511         handle->h_journal = NULL;
512         /*
513          * GFP_NOFS is here because callers are likely from writeback or
514          * similarly constrained call sites
515          */
516         ret = start_this_handle(journal, handle, GFP_NOFS);
517         if (ret < 0) {
518                 jbd2_journal_free_reserved(handle);
519                 return ret;
520         }
521         handle->h_type = type;
522         handle->h_line_no = line_no;
523         return 0;
524 }
525 EXPORT_SYMBOL(jbd2_journal_start_reserved);
526
527 /**
528  * int jbd2_journal_extend() - extend buffer credits.
529  * @handle:  handle to 'extend'
530  * @nblocks: nr blocks to try to extend by.
531  *
532  * Some transactions, such as large extends and truncates, can be done
533  * atomically all at once or in several stages.  The operation requests
534  * a credit for a number of buffer modications in advance, but can
535  * extend its credit if it needs more.
536  *
537  * jbd2_journal_extend tries to give the running handle more buffer credits.
538  * It does not guarantee that allocation - this is a best-effort only.
539  * The calling process MUST be able to deal cleanly with a failure to
540  * extend here.
541  *
542  * Return 0 on success, non-zero on failure.
543  *
544  * return code < 0 implies an error
545  * return code > 0 implies normal transaction-full status.
546  */
547 int jbd2_journal_extend(handle_t *handle, int nblocks)
548 {
549         transaction_t *transaction = handle->h_transaction;
550         journal_t *journal;
551         int result;
552         int wanted;
553
554         WARN_ON(!transaction);
555         if (is_handle_aborted(handle))
556                 return -EROFS;
557         journal = transaction->t_journal;
558
559         result = 1;
560
561         read_lock(&journal->j_state_lock);
562
563         /* Don't extend a locked-down transaction! */
564         if (transaction->t_state != T_RUNNING) {
565                 jbd_debug(3, "denied handle %p %d blocks: "
566                           "transaction not running\n", handle, nblocks);
567                 goto error_out;
568         }
569
570         spin_lock(&transaction->t_handle_lock);
571         wanted = atomic_add_return(nblocks,
572                                    &transaction->t_outstanding_credits);
573
574         if (wanted > journal->j_max_transaction_buffers) {
575                 jbd_debug(3, "denied handle %p %d blocks: "
576                           "transaction too large\n", handle, nblocks);
577                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
578                 goto unlock;
579         }
580
581         if (wanted + (wanted >> JBD2_CONTROL_BLOCKS_SHIFT) >
582             jbd2_log_space_left(journal)) {
583                 jbd_debug(3, "denied handle %p %d blocks: "
584                           "insufficient log space\n", handle, nblocks);
585                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
586                 goto unlock;
587         }
588
589         trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
590                                  transaction->t_tid,
591                                  handle->h_type, handle->h_line_no,
592                                  handle->h_buffer_credits,
593                                  nblocks);
594
595         handle->h_buffer_credits += nblocks;
596         handle->h_requested_credits += nblocks;
597         result = 0;
598
599         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
600 unlock:
601         spin_unlock(&transaction->t_handle_lock);
602 error_out:
603         read_unlock(&journal->j_state_lock);
604         return result;
605 }
606
607
608 /**
609  * int jbd2_journal_restart() - restart a handle .
610  * @handle:  handle to restart
611  * @nblocks: nr credits requested
612  *
613  * Restart a handle for a multi-transaction filesystem
614  * operation.
615  *
616  * If the jbd2_journal_extend() call above fails to grant new buffer credits
617  * to a running handle, a call to jbd2_journal_restart will commit the
618  * handle's transaction so far and reattach the handle to a new
619  * transaction capabable of guaranteeing the requested number of
620  * credits. We preserve reserved handle if there's any attached to the
621  * passed in handle.
622  */
623 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
624 {
625         transaction_t *transaction = handle->h_transaction;
626         journal_t *journal;
627         tid_t           tid;
628         int             need_to_start, ret;
629
630         WARN_ON(!transaction);
631         /* If we've had an abort of any type, don't even think about
632          * actually doing the restart! */
633         if (is_handle_aborted(handle))
634                 return 0;
635         journal = transaction->t_journal;
636
637         /*
638          * First unlink the handle from its current transaction, and start the
639          * commit on that.
640          */
641         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
642         J_ASSERT(journal_current_handle() == handle);
643
644         read_lock(&journal->j_state_lock);
645         spin_lock(&transaction->t_handle_lock);
646         atomic_sub(handle->h_buffer_credits,
647                    &transaction->t_outstanding_credits);
648         if (handle->h_rsv_handle) {
649                 sub_reserved_credits(journal,
650                                      handle->h_rsv_handle->h_buffer_credits);
651         }
652         if (atomic_dec_and_test(&transaction->t_updates))
653                 wake_up(&journal->j_wait_updates);
654         tid = transaction->t_tid;
655         spin_unlock(&transaction->t_handle_lock);
656         handle->h_transaction = NULL;
657         current->journal_info = NULL;
658
659         jbd_debug(2, "restarting handle %p\n", handle);
660         need_to_start = !tid_geq(journal->j_commit_request, tid);
661         read_unlock(&journal->j_state_lock);
662         if (need_to_start)
663                 jbd2_log_start_commit(journal, tid);
664
665         lock_map_release(&handle->h_lockdep_map);
666         handle->h_buffer_credits = nblocks;
667         ret = start_this_handle(journal, handle, gfp_mask);
668         return ret;
669 }
670 EXPORT_SYMBOL(jbd2__journal_restart);
671
672
673 int jbd2_journal_restart(handle_t *handle, int nblocks)
674 {
675         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
676 }
677 EXPORT_SYMBOL(jbd2_journal_restart);
678
679 /**
680  * void jbd2_journal_lock_updates () - establish a transaction barrier.
681  * @journal:  Journal to establish a barrier on.
682  *
683  * This locks out any further updates from being started, and blocks
684  * until all existing updates have completed, returning only once the
685  * journal is in a quiescent state with no updates running.
686  *
687  * The journal lock should not be held on entry.
688  */
689 void jbd2_journal_lock_updates(journal_t *journal)
690 {
691         DEFINE_WAIT(wait);
692
693         write_lock(&journal->j_state_lock);
694         ++journal->j_barrier_count;
695
696         /* Wait until there are no reserved handles */
697         if (atomic_read(&journal->j_reserved_credits)) {
698                 write_unlock(&journal->j_state_lock);
699                 wait_event(journal->j_wait_reserved,
700                            atomic_read(&journal->j_reserved_credits) == 0);
701                 write_lock(&journal->j_state_lock);
702         }
703
704         /* Wait until there are no running updates */
705         while (1) {
706                 transaction_t *transaction = journal->j_running_transaction;
707
708                 if (!transaction)
709                         break;
710
711                 spin_lock(&transaction->t_handle_lock);
712                 prepare_to_wait(&journal->j_wait_updates, &wait,
713                                 TASK_UNINTERRUPTIBLE);
714                 if (!atomic_read(&transaction->t_updates)) {
715                         spin_unlock(&transaction->t_handle_lock);
716                         finish_wait(&journal->j_wait_updates, &wait);
717                         break;
718                 }
719                 spin_unlock(&transaction->t_handle_lock);
720                 write_unlock(&journal->j_state_lock);
721                 schedule();
722                 finish_wait(&journal->j_wait_updates, &wait);
723                 write_lock(&journal->j_state_lock);
724         }
725         write_unlock(&journal->j_state_lock);
726
727         /*
728          * We have now established a barrier against other normal updates, but
729          * we also need to barrier against other jbd2_journal_lock_updates() calls
730          * to make sure that we serialise special journal-locked operations
731          * too.
732          */
733         mutex_lock(&journal->j_barrier);
734 }
735
736 /**
737  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
738  * @journal:  Journal to release the barrier on.
739  *
740  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
741  *
742  * Should be called without the journal lock held.
743  */
744 void jbd2_journal_unlock_updates (journal_t *journal)
745 {
746         J_ASSERT(journal->j_barrier_count != 0);
747
748         mutex_unlock(&journal->j_barrier);
749         write_lock(&journal->j_state_lock);
750         --journal->j_barrier_count;
751         write_unlock(&journal->j_state_lock);
752         wake_up(&journal->j_wait_transaction_locked);
753 }
754
755 static void warn_dirty_buffer(struct buffer_head *bh)
756 {
757         char b[BDEVNAME_SIZE];
758
759         printk(KERN_WARNING
760                "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
761                "There's a risk of filesystem corruption in case of system "
762                "crash.\n",
763                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
764 }
765
766 /*
767  * If the buffer is already part of the current transaction, then there
768  * is nothing we need to do.  If it is already part of a prior
769  * transaction which we are still committing to disk, then we need to
770  * make sure that we do not overwrite the old copy: we do copy-out to
771  * preserve the copy going to disk.  We also account the buffer against
772  * the handle's metadata buffer credits (unless the buffer is already
773  * part of the transaction, that is).
774  *
775  */
776 static int
777 do_get_write_access(handle_t *handle, struct journal_head *jh,
778                         int force_copy)
779 {
780         struct buffer_head *bh;
781         transaction_t *transaction = handle->h_transaction;
782         journal_t *journal;
783         int error;
784         char *frozen_buffer = NULL;
785         int need_copy = 0;
786         unsigned long start_lock, time_lock;
787
788         WARN_ON(!transaction);
789         if (is_handle_aborted(handle))
790                 return -EROFS;
791         journal = transaction->t_journal;
792
793         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
794
795         JBUFFER_TRACE(jh, "entry");
796 repeat:
797         bh = jh2bh(jh);
798
799         /* @@@ Need to check for errors here at some point. */
800
801         start_lock = jiffies;
802         lock_buffer(bh);
803         jbd_lock_bh_state(bh);
804
805         /* If it takes too long to lock the buffer, trace it */
806         time_lock = jbd2_time_diff(start_lock, jiffies);
807         if (time_lock > HZ/10)
808                 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
809                         jiffies_to_msecs(time_lock));
810
811         /* We now hold the buffer lock so it is safe to query the buffer
812          * state.  Is the buffer dirty?
813          *
814          * If so, there are two possibilities.  The buffer may be
815          * non-journaled, and undergoing a quite legitimate writeback.
816          * Otherwise, it is journaled, and we don't expect dirty buffers
817          * in that state (the buffers should be marked JBD_Dirty
818          * instead.)  So either the IO is being done under our own
819          * control and this is a bug, or it's a third party IO such as
820          * dump(8) (which may leave the buffer scheduled for read ---
821          * ie. locked but not dirty) or tune2fs (which may actually have
822          * the buffer dirtied, ugh.)  */
823
824         if (buffer_dirty(bh)) {
825                 /*
826                  * First question: is this buffer already part of the current
827                  * transaction or the existing committing transaction?
828                  */
829                 if (jh->b_transaction) {
830                         J_ASSERT_JH(jh,
831                                 jh->b_transaction == transaction ||
832                                 jh->b_transaction ==
833                                         journal->j_committing_transaction);
834                         if (jh->b_next_transaction)
835                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
836                                                         transaction);
837                         warn_dirty_buffer(bh);
838                 }
839                 /*
840                  * In any case we need to clean the dirty flag and we must
841                  * do it under the buffer lock to be sure we don't race
842                  * with running write-out.
843                  */
844                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
845                 clear_buffer_dirty(bh);
846                 set_buffer_jbddirty(bh);
847         }
848
849         unlock_buffer(bh);
850
851         error = -EROFS;
852         if (is_handle_aborted(handle)) {
853                 jbd_unlock_bh_state(bh);
854                 goto out;
855         }
856         error = 0;
857
858         /*
859          * The buffer is already part of this transaction if b_transaction or
860          * b_next_transaction points to it
861          */
862         if (jh->b_transaction == transaction ||
863             jh->b_next_transaction == transaction)
864                 goto done;
865
866         /*
867          * this is the first time this transaction is touching this buffer,
868          * reset the modified flag
869          */
870        jh->b_modified = 0;
871
872         /*
873          * If there is already a copy-out version of this buffer, then we don't
874          * need to make another one
875          */
876         if (jh->b_frozen_data) {
877                 JBUFFER_TRACE(jh, "has frozen data");
878                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
879                 jh->b_next_transaction = transaction;
880                 goto done;
881         }
882
883         /* Is there data here we need to preserve? */
884
885         if (jh->b_transaction && jh->b_transaction != transaction) {
886                 JBUFFER_TRACE(jh, "owned by older transaction");
887                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
888                 J_ASSERT_JH(jh, jh->b_transaction ==
889                                         journal->j_committing_transaction);
890
891                 /* There is one case we have to be very careful about.
892                  * If the committing transaction is currently writing
893                  * this buffer out to disk and has NOT made a copy-out,
894                  * then we cannot modify the buffer contents at all
895                  * right now.  The essence of copy-out is that it is the
896                  * extra copy, not the primary copy, which gets
897                  * journaled.  If the primary copy is already going to
898                  * disk then we cannot do copy-out here. */
899
900                 if (buffer_shadow(bh)) {
901                         JBUFFER_TRACE(jh, "on shadow: sleep");
902                         jbd_unlock_bh_state(bh);
903                         wait_on_bit_io(&bh->b_state, BH_Shadow,
904                                        TASK_UNINTERRUPTIBLE);
905                         goto repeat;
906                 }
907
908                 /*
909                  * Only do the copy if the currently-owning transaction still
910                  * needs it. If buffer isn't on BJ_Metadata list, the
911                  * committing transaction is past that stage (here we use the
912                  * fact that BH_Shadow is set under bh_state lock together with
913                  * refiling to BJ_Shadow list and at this point we know the
914                  * buffer doesn't have BH_Shadow set).
915                  *
916                  * Subtle point, though: if this is a get_undo_access,
917                  * then we will be relying on the frozen_data to contain
918                  * the new value of the committed_data record after the
919                  * transaction, so we HAVE to force the frozen_data copy
920                  * in that case.
921                  */
922                 if (jh->b_jlist == BJ_Metadata || force_copy) {
923                         JBUFFER_TRACE(jh, "generate frozen data");
924                         if (!frozen_buffer) {
925                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
926                                 jbd_unlock_bh_state(bh);
927                                 frozen_buffer =
928                                         jbd2_alloc(jh2bh(jh)->b_size,
929                                                          GFP_NOFS);
930                                 if (!frozen_buffer) {
931                                         printk(KERN_ERR
932                                                "%s: OOM for frozen_buffer\n",
933                                                __func__);
934                                         JBUFFER_TRACE(jh, "oom!");
935                                         error = -ENOMEM;
936                                         jbd_lock_bh_state(bh);
937                                         goto done;
938                                 }
939                                 goto repeat;
940                         }
941                         jh->b_frozen_data = frozen_buffer;
942                         frozen_buffer = NULL;
943                         need_copy = 1;
944                 }
945                 jh->b_next_transaction = transaction;
946         }
947
948
949         /*
950          * Finally, if the buffer is not journaled right now, we need to make
951          * sure it doesn't get written to disk before the caller actually
952          * commits the new data
953          */
954         if (!jh->b_transaction) {
955                 JBUFFER_TRACE(jh, "no transaction");
956                 J_ASSERT_JH(jh, !jh->b_next_transaction);
957                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
958                 spin_lock(&journal->j_list_lock);
959                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
960                 spin_unlock(&journal->j_list_lock);
961         }
962
963 done:
964         if (need_copy) {
965                 struct page *page;
966                 int offset;
967                 char *source;
968
969                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
970                             "Possible IO failure.\n");
971                 page = jh2bh(jh)->b_page;
972                 offset = offset_in_page(jh2bh(jh)->b_data);
973                 source = kmap_atomic(page);
974                 /* Fire data frozen trigger just before we copy the data */
975                 jbd2_buffer_frozen_trigger(jh, source + offset,
976                                            jh->b_triggers);
977                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
978                 kunmap_atomic(source);
979
980                 /*
981                  * Now that the frozen data is saved off, we need to store
982                  * any matching triggers.
983                  */
984                 jh->b_frozen_triggers = jh->b_triggers;
985         }
986         jbd_unlock_bh_state(bh);
987
988         /*
989          * If we are about to journal a buffer, then any revoke pending on it is
990          * no longer valid
991          */
992         jbd2_journal_cancel_revoke(handle, jh);
993
994 out:
995         if (unlikely(frozen_buffer))    /* It's usually NULL */
996                 jbd2_free(frozen_buffer, bh->b_size);
997
998         JBUFFER_TRACE(jh, "exit");
999         return error;
1000 }
1001
1002 /**
1003  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1004  * @handle: transaction to add buffer modifications to
1005  * @bh:     bh to be used for metadata writes
1006  *
1007  * Returns an error code or 0 on success.
1008  *
1009  * In full data journalling mode the buffer may be of type BJ_AsyncData,
1010  * because we're write()ing a buffer which is also part of a shared mapping.
1011  */
1012
1013 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1014 {
1015         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1016         int rc;
1017
1018         /* We do not want to get caught playing with fields which the
1019          * log thread also manipulates.  Make sure that the buffer
1020          * completes any outstanding IO before proceeding. */
1021         rc = do_get_write_access(handle, jh, 0);
1022         jbd2_journal_put_journal_head(jh);
1023         return rc;
1024 }
1025
1026
1027 /*
1028  * When the user wants to journal a newly created buffer_head
1029  * (ie. getblk() returned a new buffer and we are going to populate it
1030  * manually rather than reading off disk), then we need to keep the
1031  * buffer_head locked until it has been completely filled with new
1032  * data.  In this case, we should be able to make the assertion that
1033  * the bh is not already part of an existing transaction.
1034  *
1035  * The buffer should already be locked by the caller by this point.
1036  * There is no lock ranking violation: it was a newly created,
1037  * unlocked buffer beforehand. */
1038
1039 /**
1040  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1041  * @handle: transaction to new buffer to
1042  * @bh: new buffer.
1043  *
1044  * Call this if you create a new bh.
1045  */
1046 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1047 {
1048         transaction_t *transaction = handle->h_transaction;
1049         journal_t *journal;
1050         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1051         int err;
1052
1053         jbd_debug(5, "journal_head %p\n", jh);
1054         WARN_ON(!transaction);
1055         err = -EROFS;
1056         if (is_handle_aborted(handle))
1057                 goto out;
1058         journal = transaction->t_journal;
1059         err = 0;
1060
1061         JBUFFER_TRACE(jh, "entry");
1062         /*
1063          * The buffer may already belong to this transaction due to pre-zeroing
1064          * in the filesystem's new_block code.  It may also be on the previous,
1065          * committing transaction's lists, but it HAS to be in Forget state in
1066          * that case: the transaction must have deleted the buffer for it to be
1067          * reused here.
1068          */
1069         jbd_lock_bh_state(bh);
1070         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1071                 jh->b_transaction == NULL ||
1072                 (jh->b_transaction == journal->j_committing_transaction &&
1073                           jh->b_jlist == BJ_Forget)));
1074
1075         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1076         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1077
1078         if (jh->b_transaction == NULL) {
1079                 /*
1080                  * Previous jbd2_journal_forget() could have left the buffer
1081                  * with jbddirty bit set because it was being committed. When
1082                  * the commit finished, we've filed the buffer for
1083                  * checkpointing and marked it dirty. Now we are reallocating
1084                  * the buffer so the transaction freeing it must have
1085                  * committed and so it's safe to clear the dirty bit.
1086                  */
1087                 clear_buffer_dirty(jh2bh(jh));
1088                 /* first access by this transaction */
1089                 jh->b_modified = 0;
1090
1091                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1092                 spin_lock(&journal->j_list_lock);
1093                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1094         } else if (jh->b_transaction == journal->j_committing_transaction) {
1095                 /* first access by this transaction */
1096                 jh->b_modified = 0;
1097
1098                 JBUFFER_TRACE(jh, "set next transaction");
1099                 spin_lock(&journal->j_list_lock);
1100                 jh->b_next_transaction = transaction;
1101         }
1102         spin_unlock(&journal->j_list_lock);
1103         jbd_unlock_bh_state(bh);
1104
1105         /*
1106          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
1107          * blocks which contain freed but then revoked metadata.  We need
1108          * to cancel the revoke in case we end up freeing it yet again
1109          * and the reallocating as data - this would cause a second revoke,
1110          * which hits an assertion error.
1111          */
1112         JBUFFER_TRACE(jh, "cancelling revoke");
1113         jbd2_journal_cancel_revoke(handle, jh);
1114 out:
1115         jbd2_journal_put_journal_head(jh);
1116         return err;
1117 }
1118
1119 /**
1120  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
1121  *     non-rewindable consequences
1122  * @handle: transaction
1123  * @bh: buffer to undo
1124  *
1125  * Sometimes there is a need to distinguish between metadata which has
1126  * been committed to disk and that which has not.  The ext3fs code uses
1127  * this for freeing and allocating space, we have to make sure that we
1128  * do not reuse freed space until the deallocation has been committed,
1129  * since if we overwrote that space we would make the delete
1130  * un-rewindable in case of a crash.
1131  *
1132  * To deal with that, jbd2_journal_get_undo_access requests write access to a
1133  * buffer for parts of non-rewindable operations such as delete
1134  * operations on the bitmaps.  The journaling code must keep a copy of
1135  * the buffer's contents prior to the undo_access call until such time
1136  * as we know that the buffer has definitely been committed to disk.
1137  *
1138  * We never need to know which transaction the committed data is part
1139  * of, buffers touched here are guaranteed to be dirtied later and so
1140  * will be committed to a new transaction in due course, at which point
1141  * we can discard the old committed data pointer.
1142  *
1143  * Returns error number or 0 on success.
1144  */
1145 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1146 {
1147         int err;
1148         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1149         char *committed_data = NULL;
1150
1151         JBUFFER_TRACE(jh, "entry");
1152
1153         /*
1154          * Do this first --- it can drop the journal lock, so we want to
1155          * make sure that obtaining the committed_data is done
1156          * atomically wrt. completion of any outstanding commits.
1157          */
1158         err = do_get_write_access(handle, jh, 1);
1159         if (err)
1160                 goto out;
1161
1162 repeat:
1163         if (!jh->b_committed_data) {
1164                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
1165                 if (!committed_data) {
1166                         printk(KERN_ERR "%s: No memory for committed data\n",
1167                                 __func__);
1168                         err = -ENOMEM;
1169                         goto out;
1170                 }
1171         }
1172
1173         jbd_lock_bh_state(bh);
1174         if (!jh->b_committed_data) {
1175                 /* Copy out the current buffer contents into the
1176                  * preserved, committed copy. */
1177                 JBUFFER_TRACE(jh, "generate b_committed data");
1178                 if (!committed_data) {
1179                         jbd_unlock_bh_state(bh);
1180                         goto repeat;
1181                 }
1182
1183                 jh->b_committed_data = committed_data;
1184                 committed_data = NULL;
1185                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1186         }
1187         jbd_unlock_bh_state(bh);
1188 out:
1189         jbd2_journal_put_journal_head(jh);
1190         if (unlikely(committed_data))
1191                 jbd2_free(committed_data, bh->b_size);
1192         return err;
1193 }
1194
1195 /**
1196  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1197  * @bh: buffer to trigger on
1198  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1199  *
1200  * Set any triggers on this journal_head.  This is always safe, because
1201  * triggers for a committing buffer will be saved off, and triggers for
1202  * a running transaction will match the buffer in that transaction.
1203  *
1204  * Call with NULL to clear the triggers.
1205  */
1206 void jbd2_journal_set_triggers(struct buffer_head *bh,
1207                                struct jbd2_buffer_trigger_type *type)
1208 {
1209         struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1210
1211         if (WARN_ON(!jh))
1212                 return;
1213         jh->b_triggers = type;
1214         jbd2_journal_put_journal_head(jh);
1215 }
1216
1217 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1218                                 struct jbd2_buffer_trigger_type *triggers)
1219 {
1220         struct buffer_head *bh = jh2bh(jh);
1221
1222         if (!triggers || !triggers->t_frozen)
1223                 return;
1224
1225         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1226 }
1227
1228 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1229                                struct jbd2_buffer_trigger_type *triggers)
1230 {
1231         if (!triggers || !triggers->t_abort)
1232                 return;
1233
1234         triggers->t_abort(triggers, jh2bh(jh));
1235 }
1236
1237
1238
1239 /**
1240  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1241  * @handle: transaction to add buffer to.
1242  * @bh: buffer to mark
1243  *
1244  * mark dirty metadata which needs to be journaled as part of the current
1245  * transaction.
1246  *
1247  * The buffer must have previously had jbd2_journal_get_write_access()
1248  * called so that it has a valid journal_head attached to the buffer
1249  * head.
1250  *
1251  * The buffer is placed on the transaction's metadata list and is marked
1252  * as belonging to the transaction.
1253  *
1254  * Returns error number or 0 on success.
1255  *
1256  * Special care needs to be taken if the buffer already belongs to the
1257  * current committing transaction (in which case we should have frozen
1258  * data present for that commit).  In that case, we don't relink the
1259  * buffer: that only gets done when the old transaction finally
1260  * completes its commit.
1261  */
1262 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1263 {
1264         transaction_t *transaction = handle->h_transaction;
1265         journal_t *journal;
1266         struct journal_head *jh;
1267         int ret = 0;
1268
1269         WARN_ON(!transaction);
1270         if (is_handle_aborted(handle))
1271                 return -EROFS;
1272         journal = transaction->t_journal;
1273         jh = jbd2_journal_grab_journal_head(bh);
1274         if (!jh) {
1275                 ret = -EUCLEAN;
1276                 goto out;
1277         }
1278         jbd_debug(5, "journal_head %p\n", jh);
1279         JBUFFER_TRACE(jh, "entry");
1280
1281         jbd_lock_bh_state(bh);
1282
1283         if (jh->b_modified == 0) {
1284                 /*
1285                  * This buffer's got modified and becoming part
1286                  * of the transaction. This needs to be done
1287                  * once a transaction -bzzz
1288                  */
1289                 jh->b_modified = 1;
1290                 if (handle->h_buffer_credits <= 0) {
1291                         ret = -ENOSPC;
1292                         goto out_unlock_bh;
1293                 }
1294                 handle->h_buffer_credits--;
1295         }
1296
1297         /*
1298          * fastpath, to avoid expensive locking.  If this buffer is already
1299          * on the running transaction's metadata list there is nothing to do.
1300          * Nobody can take it off again because there is a handle open.
1301          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1302          * result in this test being false, so we go in and take the locks.
1303          */
1304         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1305                 JBUFFER_TRACE(jh, "fastpath");
1306                 if (unlikely(jh->b_transaction !=
1307                              journal->j_running_transaction)) {
1308                         printk(KERN_ERR "JBD2: %s: "
1309                                "jh->b_transaction (%llu, %p, %u) != "
1310                                "journal->j_running_transaction (%p, %u)\n",
1311                                journal->j_devname,
1312                                (unsigned long long) bh->b_blocknr,
1313                                jh->b_transaction,
1314                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1315                                journal->j_running_transaction,
1316                                journal->j_running_transaction ?
1317                                journal->j_running_transaction->t_tid : 0);
1318                         ret = -EINVAL;
1319                 }
1320                 goto out_unlock_bh;
1321         }
1322
1323         set_buffer_jbddirty(bh);
1324
1325         /*
1326          * Metadata already on the current transaction list doesn't
1327          * need to be filed.  Metadata on another transaction's list must
1328          * be committing, and will be refiled once the commit completes:
1329          * leave it alone for now.
1330          */
1331         if (jh->b_transaction != transaction) {
1332                 JBUFFER_TRACE(jh, "already on other transaction");
1333                 if (unlikely(((jh->b_transaction !=
1334                                journal->j_committing_transaction)) ||
1335                              (jh->b_next_transaction != transaction))) {
1336                         printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1337                                "bad jh for block %llu: "
1338                                "transaction (%p, %u), "
1339                                "jh->b_transaction (%p, %u), "
1340                                "jh->b_next_transaction (%p, %u), jlist %u\n",
1341                                journal->j_devname,
1342                                (unsigned long long) bh->b_blocknr,
1343                                transaction, transaction->t_tid,
1344                                jh->b_transaction,
1345                                jh->b_transaction ?
1346                                jh->b_transaction->t_tid : 0,
1347                                jh->b_next_transaction,
1348                                jh->b_next_transaction ?
1349                                jh->b_next_transaction->t_tid : 0,
1350                                jh->b_jlist);
1351                         WARN_ON(1);
1352                         ret = -EINVAL;
1353                 }
1354                 /* And this case is illegal: we can't reuse another
1355                  * transaction's data buffer, ever. */
1356                 goto out_unlock_bh;
1357         }
1358
1359         /* That test should have eliminated the following case: */
1360         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1361
1362         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1363         spin_lock(&journal->j_list_lock);
1364         __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1365         spin_unlock(&journal->j_list_lock);
1366 out_unlock_bh:
1367         jbd_unlock_bh_state(bh);
1368         jbd2_journal_put_journal_head(jh);
1369 out:
1370         JBUFFER_TRACE(jh, "exit");
1371         return ret;
1372 }
1373
1374 /**
1375  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1376  * @handle: transaction handle
1377  * @bh:     bh to 'forget'
1378  *
1379  * We can only do the bforget if there are no commits pending against the
1380  * buffer.  If the buffer is dirty in the current running transaction we
1381  * can safely unlink it.
1382  *
1383  * bh may not be a journalled buffer at all - it may be a non-JBD
1384  * buffer which came off the hashtable.  Check for this.
1385  *
1386  * Decrements bh->b_count by one.
1387  *
1388  * Allow this call even if the handle has aborted --- it may be part of
1389  * the caller's cleanup after an abort.
1390  */
1391 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1392 {
1393         transaction_t *transaction = handle->h_transaction;
1394         journal_t *journal;
1395         struct journal_head *jh;
1396         int drop_reserve = 0;
1397         int err = 0;
1398         int was_modified = 0;
1399
1400         WARN_ON(!transaction);
1401         if (is_handle_aborted(handle))
1402                 return -EROFS;
1403         journal = transaction->t_journal;
1404
1405         BUFFER_TRACE(bh, "entry");
1406
1407         jbd_lock_bh_state(bh);
1408
1409         if (!buffer_jbd(bh))
1410                 goto not_jbd;
1411         jh = bh2jh(bh);
1412
1413         /* Critical error: attempting to delete a bitmap buffer, maybe?
1414          * Don't do any jbd operations, and return an error. */
1415         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1416                          "inconsistent data on disk")) {
1417                 err = -EIO;
1418                 goto not_jbd;
1419         }
1420
1421         /* keep track of whether or not this transaction modified us */
1422         was_modified = jh->b_modified;
1423
1424         /*
1425          * The buffer's going from the transaction, we must drop
1426          * all references -bzzz
1427          */
1428         jh->b_modified = 0;
1429
1430         if (jh->b_transaction == transaction) {
1431                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1432
1433                 /* If we are forgetting a buffer which is already part
1434                  * of this transaction, then we can just drop it from
1435                  * the transaction immediately. */
1436                 clear_buffer_dirty(bh);
1437                 clear_buffer_jbddirty(bh);
1438
1439                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1440
1441                 /*
1442                  * we only want to drop a reference if this transaction
1443                  * modified the buffer
1444                  */
1445                 if (was_modified)
1446                         drop_reserve = 1;
1447
1448                 /*
1449                  * We are no longer going to journal this buffer.
1450                  * However, the commit of this transaction is still
1451                  * important to the buffer: the delete that we are now
1452                  * processing might obsolete an old log entry, so by
1453                  * committing, we can satisfy the buffer's checkpoint.
1454                  *
1455                  * So, if we have a checkpoint on the buffer, we should
1456                  * now refile the buffer on our BJ_Forget list so that
1457                  * we know to remove the checkpoint after we commit.
1458                  */
1459
1460                 spin_lock(&journal->j_list_lock);
1461                 if (jh->b_cp_transaction) {
1462                         __jbd2_journal_temp_unlink_buffer(jh);
1463                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1464                 } else {
1465                         __jbd2_journal_unfile_buffer(jh);
1466                         if (!buffer_jbd(bh)) {
1467                                 spin_unlock(&journal->j_list_lock);
1468                                 jbd_unlock_bh_state(bh);
1469                                 __bforget(bh);
1470                                 goto drop;
1471                         }
1472                 }
1473                 spin_unlock(&journal->j_list_lock);
1474         } else if (jh->b_transaction) {
1475                 J_ASSERT_JH(jh, (jh->b_transaction ==
1476                                  journal->j_committing_transaction));
1477                 /* However, if the buffer is still owned by a prior
1478                  * (committing) transaction, we can't drop it yet... */
1479                 JBUFFER_TRACE(jh, "belongs to older transaction");
1480                 /* ... but we CAN drop it from the new transaction if we
1481                  * have also modified it since the original commit. */
1482
1483                 if (jh->b_next_transaction) {
1484                         J_ASSERT(jh->b_next_transaction == transaction);
1485                         spin_lock(&journal->j_list_lock);
1486                         jh->b_next_transaction = NULL;
1487                         spin_unlock(&journal->j_list_lock);
1488
1489                         /*
1490                          * only drop a reference if this transaction modified
1491                          * the buffer
1492                          */
1493                         if (was_modified)
1494                                 drop_reserve = 1;
1495                 }
1496         }
1497
1498 not_jbd:
1499         jbd_unlock_bh_state(bh);
1500         __brelse(bh);
1501 drop:
1502         if (drop_reserve) {
1503                 /* no need to reserve log space for this block -bzzz */
1504                 handle->h_buffer_credits++;
1505         }
1506         return err;
1507 }
1508
1509 /**
1510  * int jbd2_journal_stop() - complete a transaction
1511  * @handle: tranaction to complete.
1512  *
1513  * All done for a particular handle.
1514  *
1515  * There is not much action needed here.  We just return any remaining
1516  * buffer credits to the transaction and remove the handle.  The only
1517  * complication is that we need to start a commit operation if the
1518  * filesystem is marked for synchronous update.
1519  *
1520  * jbd2_journal_stop itself will not usually return an error, but it may
1521  * do so in unusual circumstances.  In particular, expect it to
1522  * return -EIO if a jbd2_journal_abort has been executed since the
1523  * transaction began.
1524  */
1525 int jbd2_journal_stop(handle_t *handle)
1526 {
1527         transaction_t *transaction = handle->h_transaction;
1528         journal_t *journal;
1529         int err = 0, wait_for_commit = 0;
1530         tid_t tid;
1531         pid_t pid;
1532
1533         if (!transaction)
1534                 goto free_and_exit;
1535         journal = transaction->t_journal;
1536
1537         J_ASSERT(journal_current_handle() == handle);
1538
1539         if (is_handle_aborted(handle))
1540                 err = -EIO;
1541         else
1542                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1543
1544         if (--handle->h_ref > 0) {
1545                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1546                           handle->h_ref);
1547                 return err;
1548         }
1549
1550         jbd_debug(4, "Handle %p going down\n", handle);
1551         trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1552                                 transaction->t_tid,
1553                                 handle->h_type, handle->h_line_no,
1554                                 jiffies - handle->h_start_jiffies,
1555                                 handle->h_sync, handle->h_requested_credits,
1556                                 (handle->h_requested_credits -
1557                                  handle->h_buffer_credits));
1558
1559         /*
1560          * Implement synchronous transaction batching.  If the handle
1561          * was synchronous, don't force a commit immediately.  Let's
1562          * yield and let another thread piggyback onto this
1563          * transaction.  Keep doing that while new threads continue to
1564          * arrive.  It doesn't cost much - we're about to run a commit
1565          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1566          * operations by 30x or more...
1567          *
1568          * We try and optimize the sleep time against what the
1569          * underlying disk can do, instead of having a static sleep
1570          * time.  This is useful for the case where our storage is so
1571          * fast that it is more optimal to go ahead and force a flush
1572          * and wait for the transaction to be committed than it is to
1573          * wait for an arbitrary amount of time for new writers to
1574          * join the transaction.  We achieve this by measuring how
1575          * long it takes to commit a transaction, and compare it with
1576          * how long this transaction has been running, and if run time
1577          * < commit time then we sleep for the delta and commit.  This
1578          * greatly helps super fast disks that would see slowdowns as
1579          * more threads started doing fsyncs.
1580          *
1581          * But don't do this if this process was the most recent one
1582          * to perform a synchronous write.  We do this to detect the
1583          * case where a single process is doing a stream of sync
1584          * writes.  No point in waiting for joiners in that case.
1585          *
1586          * Setting max_batch_time to 0 disables this completely.
1587          */
1588         pid = current->pid;
1589         if (handle->h_sync && journal->j_last_sync_writer != pid &&
1590             journal->j_max_batch_time) {
1591                 u64 commit_time, trans_time;
1592
1593                 journal->j_last_sync_writer = pid;
1594
1595                 read_lock(&journal->j_state_lock);
1596                 commit_time = journal->j_average_commit_time;
1597                 read_unlock(&journal->j_state_lock);
1598
1599                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1600                                                    transaction->t_start_time));
1601
1602                 commit_time = max_t(u64, commit_time,
1603                                     1000*journal->j_min_batch_time);
1604                 commit_time = min_t(u64, commit_time,
1605                                     1000*journal->j_max_batch_time);
1606
1607                 if (trans_time < commit_time) {
1608                         ktime_t expires = ktime_add_ns(ktime_get(),
1609                                                        commit_time);
1610                         set_current_state(TASK_UNINTERRUPTIBLE);
1611                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1612                 }
1613         }
1614
1615         if (handle->h_sync)
1616                 transaction->t_synchronous_commit = 1;
1617         current->journal_info = NULL;
1618         atomic_sub(handle->h_buffer_credits,
1619                    &transaction->t_outstanding_credits);
1620
1621         /*
1622          * If the handle is marked SYNC, we need to set another commit
1623          * going!  We also want to force a commit if the current
1624          * transaction is occupying too much of the log, or if the
1625          * transaction is too old now.
1626          */
1627         if (handle->h_sync ||
1628             (atomic_read(&transaction->t_outstanding_credits) >
1629              journal->j_max_transaction_buffers) ||
1630             time_after_eq(jiffies, transaction->t_expires)) {
1631                 /* Do this even for aborted journals: an abort still
1632                  * completes the commit thread, it just doesn't write
1633                  * anything to disk. */
1634
1635                 jbd_debug(2, "transaction too old, requesting commit for "
1636                                         "handle %p\n", handle);
1637                 /* This is non-blocking */
1638                 jbd2_log_start_commit(journal, transaction->t_tid);
1639
1640                 /*
1641                  * Special case: JBD2_SYNC synchronous updates require us
1642                  * to wait for the commit to complete.
1643                  */
1644                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1645                         wait_for_commit = 1;
1646         }
1647
1648         /*
1649          * Once we drop t_updates, if it goes to zero the transaction
1650          * could start committing on us and eventually disappear.  So
1651          * once we do this, we must not dereference transaction
1652          * pointer again.
1653          */
1654         tid = transaction->t_tid;
1655         if (atomic_dec_and_test(&transaction->t_updates)) {
1656                 wake_up(&journal->j_wait_updates);
1657                 if (journal->j_barrier_count)
1658                         wake_up(&journal->j_wait_transaction_locked);
1659         }
1660
1661         if (wait_for_commit)
1662                 err = jbd2_log_wait_commit(journal, tid);
1663
1664         lock_map_release(&handle->h_lockdep_map);
1665
1666         if (handle->h_rsv_handle)
1667                 jbd2_journal_free_reserved(handle->h_rsv_handle);
1668 free_and_exit:
1669         jbd2_free_handle(handle);
1670         return err;
1671 }
1672
1673 /*
1674  *
1675  * List management code snippets: various functions for manipulating the
1676  * transaction buffer lists.
1677  *
1678  */
1679
1680 /*
1681  * Append a buffer to a transaction list, given the transaction's list head
1682  * pointer.
1683  *
1684  * j_list_lock is held.
1685  *
1686  * jbd_lock_bh_state(jh2bh(jh)) is held.
1687  */
1688
1689 static inline void
1690 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1691 {
1692         if (!*list) {
1693                 jh->b_tnext = jh->b_tprev = jh;
1694                 *list = jh;
1695         } else {
1696                 /* Insert at the tail of the list to preserve order */
1697                 struct journal_head *first = *list, *last = first->b_tprev;
1698                 jh->b_tprev = last;
1699                 jh->b_tnext = first;
1700                 last->b_tnext = first->b_tprev = jh;
1701         }
1702 }
1703
1704 /*
1705  * Remove a buffer from a transaction list, given the transaction's list
1706  * head pointer.
1707  *
1708  * Called with j_list_lock held, and the journal may not be locked.
1709  *
1710  * jbd_lock_bh_state(jh2bh(jh)) is held.
1711  */
1712
1713 static inline void
1714 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1715 {
1716         if (*list == jh) {
1717                 *list = jh->b_tnext;
1718                 if (*list == jh)
1719                         *list = NULL;
1720         }
1721         jh->b_tprev->b_tnext = jh->b_tnext;
1722         jh->b_tnext->b_tprev = jh->b_tprev;
1723 }
1724
1725 /*
1726  * Remove a buffer from the appropriate transaction list.
1727  *
1728  * Note that this function can *change* the value of
1729  * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1730  * t_reserved_list.  If the caller is holding onto a copy of one of these
1731  * pointers, it could go bad.  Generally the caller needs to re-read the
1732  * pointer from the transaction_t.
1733  *
1734  * Called under j_list_lock.
1735  */
1736 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1737 {
1738         struct journal_head **list = NULL;
1739         transaction_t *transaction;
1740         struct buffer_head *bh = jh2bh(jh);
1741
1742         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1743         transaction = jh->b_transaction;
1744         if (transaction)
1745                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1746
1747         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1748         if (jh->b_jlist != BJ_None)
1749                 J_ASSERT_JH(jh, transaction != NULL);
1750
1751         switch (jh->b_jlist) {
1752         case BJ_None:
1753                 return;
1754         case BJ_Metadata:
1755                 transaction->t_nr_buffers--;
1756                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1757                 list = &transaction->t_buffers;
1758                 break;
1759         case BJ_Forget:
1760                 list = &transaction->t_forget;
1761                 break;
1762         case BJ_Shadow:
1763                 list = &transaction->t_shadow_list;
1764                 break;
1765         case BJ_Reserved:
1766                 list = &transaction->t_reserved_list;
1767                 break;
1768         }
1769
1770         __blist_del_buffer(list, jh);
1771         jh->b_jlist = BJ_None;
1772         if (test_clear_buffer_jbddirty(bh))
1773                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1774 }
1775
1776 /*
1777  * Remove buffer from all transactions.
1778  *
1779  * Called with bh_state lock and j_list_lock
1780  *
1781  * jh and bh may be already freed when this function returns.
1782  */
1783 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1784 {
1785         __jbd2_journal_temp_unlink_buffer(jh);
1786         jh->b_transaction = NULL;
1787         jbd2_journal_put_journal_head(jh);
1788 }
1789
1790 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1791 {
1792         struct buffer_head *bh = jh2bh(jh);
1793
1794         /* Get reference so that buffer cannot be freed before we unlock it */
1795         get_bh(bh);
1796         jbd_lock_bh_state(bh);
1797         spin_lock(&journal->j_list_lock);
1798         __jbd2_journal_unfile_buffer(jh);
1799         spin_unlock(&journal->j_list_lock);
1800         jbd_unlock_bh_state(bh);
1801         __brelse(bh);
1802 }
1803
1804 /*
1805  * Called from jbd2_journal_try_to_free_buffers().
1806  *
1807  * Called under jbd_lock_bh_state(bh)
1808  */
1809 static void
1810 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1811 {
1812         struct journal_head *jh;
1813
1814         jh = bh2jh(bh);
1815
1816         if (buffer_locked(bh) || buffer_dirty(bh))
1817                 goto out;
1818
1819         if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
1820                 goto out;
1821
1822         spin_lock(&journal->j_list_lock);
1823         if (jh->b_cp_transaction != NULL) {
1824                 /* written-back checkpointed metadata buffer */
1825                 JBUFFER_TRACE(jh, "remove from checkpoint list");
1826                 __jbd2_journal_remove_checkpoint(jh);
1827         }
1828         spin_unlock(&journal->j_list_lock);
1829 out:
1830         return;
1831 }
1832
1833 /**
1834  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1835  * @journal: journal for operation
1836  * @page: to try and free
1837  * @gfp_mask: we use the mask to detect how hard should we try to release
1838  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1839  * release the buffers.
1840  *
1841  *
1842  * For all the buffers on this page,
1843  * if they are fully written out ordered data, move them onto BUF_CLEAN
1844  * so try_to_free_buffers() can reap them.
1845  *
1846  * This function returns non-zero if we wish try_to_free_buffers()
1847  * to be called. We do this if the page is releasable by try_to_free_buffers().
1848  * We also do it if the page has locked or dirty buffers and the caller wants
1849  * us to perform sync or async writeout.
1850  *
1851  * This complicates JBD locking somewhat.  We aren't protected by the
1852  * BKL here.  We wish to remove the buffer from its committing or
1853  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1854  *
1855  * This may *change* the value of transaction_t->t_datalist, so anyone
1856  * who looks at t_datalist needs to lock against this function.
1857  *
1858  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1859  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1860  * will come out of the lock with the buffer dirty, which makes it
1861  * ineligible for release here.
1862  *
1863  * Who else is affected by this?  hmm...  Really the only contender
1864  * is do_get_write_access() - it could be looking at the buffer while
1865  * journal_try_to_free_buffer() is changing its state.  But that
1866  * cannot happen because we never reallocate freed data as metadata
1867  * while the data is part of a transaction.  Yes?
1868  *
1869  * Return 0 on failure, 1 on success
1870  */
1871 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1872                                 struct page *page, gfp_t gfp_mask)
1873 {
1874         struct buffer_head *head;
1875         struct buffer_head *bh;
1876         int ret = 0;
1877
1878         J_ASSERT(PageLocked(page));
1879
1880         head = page_buffers(page);
1881         bh = head;
1882         do {
1883                 struct journal_head *jh;
1884
1885                 /*
1886                  * We take our own ref against the journal_head here to avoid
1887                  * having to add tons of locking around each instance of
1888                  * jbd2_journal_put_journal_head().
1889                  */
1890                 jh = jbd2_journal_grab_journal_head(bh);
1891                 if (!jh)
1892                         continue;
1893
1894                 jbd_lock_bh_state(bh);
1895                 __journal_try_to_free_buffer(journal, bh);
1896                 jbd2_journal_put_journal_head(jh);
1897                 jbd_unlock_bh_state(bh);
1898                 if (buffer_jbd(bh))
1899                         goto busy;
1900         } while ((bh = bh->b_this_page) != head);
1901
1902         ret = try_to_free_buffers(page);
1903
1904 busy:
1905         return ret;
1906 }
1907
1908 /*
1909  * This buffer is no longer needed.  If it is on an older transaction's
1910  * checkpoint list we need to record it on this transaction's forget list
1911  * to pin this buffer (and hence its checkpointing transaction) down until
1912  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1913  * release it.
1914  * Returns non-zero if JBD no longer has an interest in the buffer.
1915  *
1916  * Called under j_list_lock.
1917  *
1918  * Called under jbd_lock_bh_state(bh).
1919  */
1920 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1921 {
1922         int may_free = 1;
1923         struct buffer_head *bh = jh2bh(jh);
1924
1925         if (jh->b_cp_transaction) {
1926                 JBUFFER_TRACE(jh, "on running+cp transaction");
1927                 __jbd2_journal_temp_unlink_buffer(jh);
1928                 /*
1929                  * We don't want to write the buffer anymore, clear the
1930                  * bit so that we don't confuse checks in
1931                  * __journal_file_buffer
1932                  */
1933                 clear_buffer_dirty(bh);
1934                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1935                 may_free = 0;
1936         } else {
1937                 JBUFFER_TRACE(jh, "on running transaction");
1938                 __jbd2_journal_unfile_buffer(jh);
1939         }
1940         return may_free;
1941 }
1942
1943 /*
1944  * jbd2_journal_invalidatepage
1945  *
1946  * This code is tricky.  It has a number of cases to deal with.
1947  *
1948  * There are two invariants which this code relies on:
1949  *
1950  * i_size must be updated on disk before we start calling invalidatepage on the
1951  * data.
1952  *
1953  *  This is done in ext3 by defining an ext3_setattr method which
1954  *  updates i_size before truncate gets going.  By maintaining this
1955  *  invariant, we can be sure that it is safe to throw away any buffers
1956  *  attached to the current transaction: once the transaction commits,
1957  *  we know that the data will not be needed.
1958  *
1959  *  Note however that we can *not* throw away data belonging to the
1960  *  previous, committing transaction!
1961  *
1962  * Any disk blocks which *are* part of the previous, committing
1963  * transaction (and which therefore cannot be discarded immediately) are
1964  * not going to be reused in the new running transaction
1965  *
1966  *  The bitmap committed_data images guarantee this: any block which is
1967  *  allocated in one transaction and removed in the next will be marked
1968  *  as in-use in the committed_data bitmap, so cannot be reused until
1969  *  the next transaction to delete the block commits.  This means that
1970  *  leaving committing buffers dirty is quite safe: the disk blocks
1971  *  cannot be reallocated to a different file and so buffer aliasing is
1972  *  not possible.
1973  *
1974  *
1975  * The above applies mainly to ordered data mode.  In writeback mode we
1976  * don't make guarantees about the order in which data hits disk --- in
1977  * particular we don't guarantee that new dirty data is flushed before
1978  * transaction commit --- so it is always safe just to discard data
1979  * immediately in that mode.  --sct
1980  */
1981
1982 /*
1983  * The journal_unmap_buffer helper function returns zero if the buffer
1984  * concerned remains pinned as an anonymous buffer belonging to an older
1985  * transaction.
1986  *
1987  * We're outside-transaction here.  Either or both of j_running_transaction
1988  * and j_committing_transaction may be NULL.
1989  */
1990 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1991                                 int partial_page)
1992 {
1993         transaction_t *transaction;
1994         struct journal_head *jh;
1995         int may_free = 1;
1996
1997         BUFFER_TRACE(bh, "entry");
1998
1999         /*
2000          * It is safe to proceed here without the j_list_lock because the
2001          * buffers cannot be stolen by try_to_free_buffers as long as we are
2002          * holding the page lock. --sct
2003          */
2004
2005         if (!buffer_jbd(bh))
2006                 goto zap_buffer_unlocked;
2007
2008         /* OK, we have data buffer in journaled mode */
2009         write_lock(&journal->j_state_lock);
2010         jbd_lock_bh_state(bh);
2011         spin_lock(&journal->j_list_lock);
2012
2013         jh = jbd2_journal_grab_journal_head(bh);
2014         if (!jh)
2015                 goto zap_buffer_no_jh;
2016
2017         /*
2018          * We cannot remove the buffer from checkpoint lists until the
2019          * transaction adding inode to orphan list (let's call it T)
2020          * is committed.  Otherwise if the transaction changing the
2021          * buffer would be cleaned from the journal before T is
2022          * committed, a crash will cause that the correct contents of
2023          * the buffer will be lost.  On the other hand we have to
2024          * clear the buffer dirty bit at latest at the moment when the
2025          * transaction marking the buffer as freed in the filesystem
2026          * structures is committed because from that moment on the
2027          * block can be reallocated and used by a different page.
2028          * Since the block hasn't been freed yet but the inode has
2029          * already been added to orphan list, it is safe for us to add
2030          * the buffer to BJ_Forget list of the newest transaction.
2031          *
2032          * Also we have to clear buffer_mapped flag of a truncated buffer
2033          * because the buffer_head may be attached to the page straddling
2034          * i_size (can happen only when blocksize < pagesize) and thus the
2035          * buffer_head can be reused when the file is extended again. So we end
2036          * up keeping around invalidated buffers attached to transactions'
2037          * BJ_Forget list just to stop checkpointing code from cleaning up
2038          * the transaction this buffer was modified in.
2039          */
2040         transaction = jh->b_transaction;
2041         if (transaction == NULL) {
2042                 /* First case: not on any transaction.  If it
2043                  * has no checkpoint link, then we can zap it:
2044                  * it's a writeback-mode buffer so we don't care
2045                  * if it hits disk safely. */
2046                 if (!jh->b_cp_transaction) {
2047                         JBUFFER_TRACE(jh, "not on any transaction: zap");
2048                         goto zap_buffer;
2049                 }
2050
2051                 if (!buffer_dirty(bh)) {
2052                         /* bdflush has written it.  We can drop it now */
2053                         goto zap_buffer;
2054                 }
2055
2056                 /* OK, it must be in the journal but still not
2057                  * written fully to disk: it's metadata or
2058                  * journaled data... */
2059
2060                 if (journal->j_running_transaction) {
2061                         /* ... and once the current transaction has
2062                          * committed, the buffer won't be needed any
2063                          * longer. */
2064                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2065                         may_free = __dispose_buffer(jh,
2066                                         journal->j_running_transaction);
2067                         goto zap_buffer;
2068                 } else {
2069                         /* There is no currently-running transaction. So the
2070                          * orphan record which we wrote for this file must have
2071                          * passed into commit.  We must attach this buffer to
2072                          * the committing transaction, if it exists. */
2073                         if (journal->j_committing_transaction) {
2074                                 JBUFFER_TRACE(jh, "give to committing trans");
2075                                 may_free = __dispose_buffer(jh,
2076                                         journal->j_committing_transaction);
2077                                 goto zap_buffer;
2078                         } else {
2079                                 /* The orphan record's transaction has
2080                                  * committed.  We can cleanse this buffer */
2081                                 clear_buffer_jbddirty(bh);
2082                                 goto zap_buffer;
2083                         }
2084                 }
2085         } else if (transaction == journal->j_committing_transaction) {
2086                 JBUFFER_TRACE(jh, "on committing transaction");
2087                 /*
2088                  * The buffer is committing, we simply cannot touch
2089                  * it. If the page is straddling i_size we have to wait
2090                  * for commit and try again.
2091                  */
2092                 if (partial_page) {
2093                         jbd2_journal_put_journal_head(jh);
2094                         spin_unlock(&journal->j_list_lock);
2095                         jbd_unlock_bh_state(bh);
2096                         write_unlock(&journal->j_state_lock);
2097                         return -EBUSY;
2098                 }
2099                 /*
2100                  * OK, buffer won't be reachable after truncate. We just set
2101                  * j_next_transaction to the running transaction (if there is
2102                  * one) and mark buffer as freed so that commit code knows it
2103                  * should clear dirty bits when it is done with the buffer.
2104                  */
2105                 set_buffer_freed(bh);
2106                 if (journal->j_running_transaction && buffer_jbddirty(bh))
2107                         jh->b_next_transaction = journal->j_running_transaction;
2108                 jbd2_journal_put_journal_head(jh);
2109                 spin_unlock(&journal->j_list_lock);
2110                 jbd_unlock_bh_state(bh);
2111                 write_unlock(&journal->j_state_lock);
2112                 return 0;
2113         } else {
2114                 /* Good, the buffer belongs to the running transaction.
2115                  * We are writing our own transaction's data, not any
2116                  * previous one's, so it is safe to throw it away
2117                  * (remember that we expect the filesystem to have set
2118                  * i_size already for this truncate so recovery will not
2119                  * expose the disk blocks we are discarding here.) */
2120                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2121                 JBUFFER_TRACE(jh, "on running transaction");
2122                 may_free = __dispose_buffer(jh, transaction);
2123         }
2124
2125 zap_buffer:
2126         /*
2127          * This is tricky. Although the buffer is truncated, it may be reused
2128          * if blocksize < pagesize and it is attached to the page straddling
2129          * EOF. Since the buffer might have been added to BJ_Forget list of the
2130          * running transaction, journal_get_write_access() won't clear
2131          * b_modified and credit accounting gets confused. So clear b_modified
2132          * here.
2133          */
2134         jh->b_modified = 0;
2135         jbd2_journal_put_journal_head(jh);
2136 zap_buffer_no_jh:
2137         spin_unlock(&journal->j_list_lock);
2138         jbd_unlock_bh_state(bh);
2139         write_unlock(&journal->j_state_lock);
2140 zap_buffer_unlocked:
2141         clear_buffer_dirty(bh);
2142         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2143         clear_buffer_mapped(bh);
2144         clear_buffer_req(bh);
2145         clear_buffer_new(bh);
2146         clear_buffer_delay(bh);
2147         clear_buffer_unwritten(bh);
2148         bh->b_bdev = NULL;
2149         return may_free;
2150 }
2151
2152 /**
2153  * void jbd2_journal_invalidatepage()
2154  * @journal: journal to use for flush...
2155  * @page:    page to flush
2156  * @offset:  start of the range to invalidate
2157  * @length:  length of the range to invalidate
2158  *
2159  * Reap page buffers containing data after in the specified range in page.
2160  * Can return -EBUSY if buffers are part of the committing transaction and
2161  * the page is straddling i_size. Caller then has to wait for current commit
2162  * and try again.
2163  */
2164 int jbd2_journal_invalidatepage(journal_t *journal,
2165                                 struct page *page,
2166                                 unsigned int offset,
2167                                 unsigned int length)
2168 {
2169         struct buffer_head *head, *bh, *next;
2170         unsigned int stop = offset + length;
2171         unsigned int curr_off = 0;
2172         int partial_page = (offset || length < PAGE_CACHE_SIZE);
2173         int may_free = 1;
2174         int ret = 0;
2175
2176         if (!PageLocked(page))
2177                 BUG();
2178         if (!page_has_buffers(page))
2179                 return 0;
2180
2181         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
2182
2183         /* We will potentially be playing with lists other than just the
2184          * data lists (especially for journaled data mode), so be
2185          * cautious in our locking. */
2186
2187         head = bh = page_buffers(page);
2188         do {
2189                 unsigned int next_off = curr_off + bh->b_size;
2190                 next = bh->b_this_page;
2191
2192                 if (next_off > stop)
2193                         return 0;
2194
2195                 if (offset <= curr_off) {
2196                         /* This block is wholly outside the truncation point */
2197                         lock_buffer(bh);
2198                         ret = journal_unmap_buffer(journal, bh, partial_page);
2199                         unlock_buffer(bh);
2200                         if (ret < 0)
2201                                 return ret;
2202                         may_free &= ret;
2203                 }
2204                 curr_off = next_off;
2205                 bh = next;
2206
2207         } while (bh != head);
2208
2209         if (!partial_page) {
2210                 if (may_free && try_to_free_buffers(page))
2211                         J_ASSERT(!page_has_buffers(page));
2212         }
2213         return 0;
2214 }
2215
2216 /*
2217  * File a buffer on the given transaction list.
2218  */
2219 void __jbd2_journal_file_buffer(struct journal_head *jh,
2220                         transaction_t *transaction, int jlist)
2221 {
2222         struct journal_head **list = NULL;
2223         int was_dirty = 0;
2224         struct buffer_head *bh = jh2bh(jh);
2225
2226         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2227         assert_spin_locked(&transaction->t_journal->j_list_lock);
2228
2229         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2230         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2231                                 jh->b_transaction == NULL);
2232
2233         if (jh->b_transaction && jh->b_jlist == jlist)
2234                 return;
2235
2236         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2237             jlist == BJ_Shadow || jlist == BJ_Forget) {
2238                 /*
2239                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2240                  * instead of buffer_dirty. We should not see a dirty bit set
2241                  * here because we clear it in do_get_write_access but e.g.
2242                  * tune2fs can modify the sb and set the dirty bit at any time
2243                  * so we try to gracefully handle that.
2244                  */
2245                 if (buffer_dirty(bh))
2246                         warn_dirty_buffer(bh);
2247                 if (test_clear_buffer_dirty(bh) ||
2248                     test_clear_buffer_jbddirty(bh))
2249                         was_dirty = 1;
2250         }
2251
2252         if (jh->b_transaction)
2253                 __jbd2_journal_temp_unlink_buffer(jh);
2254         else
2255                 jbd2_journal_grab_journal_head(bh);
2256         jh->b_transaction = transaction;
2257
2258         switch (jlist) {
2259         case BJ_None:
2260                 J_ASSERT_JH(jh, !jh->b_committed_data);
2261                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2262                 return;
2263         case BJ_Metadata:
2264                 transaction->t_nr_buffers++;
2265                 list = &transaction->t_buffers;
2266                 break;
2267         case BJ_Forget:
2268                 list = &transaction->t_forget;
2269                 break;
2270         case BJ_Shadow:
2271                 list = &transaction->t_shadow_list;
2272                 break;
2273         case BJ_Reserved:
2274                 list = &transaction->t_reserved_list;
2275                 break;
2276         }
2277
2278         __blist_add_buffer(list, jh);
2279         jh->b_jlist = jlist;
2280
2281         if (was_dirty)
2282                 set_buffer_jbddirty(bh);
2283 }
2284
2285 void jbd2_journal_file_buffer(struct journal_head *jh,
2286                                 transaction_t *transaction, int jlist)
2287 {
2288         jbd_lock_bh_state(jh2bh(jh));
2289         spin_lock(&transaction->t_journal->j_list_lock);
2290         __jbd2_journal_file_buffer(jh, transaction, jlist);
2291         spin_unlock(&transaction->t_journal->j_list_lock);
2292         jbd_unlock_bh_state(jh2bh(jh));
2293 }
2294
2295 /*
2296  * Remove a buffer from its current buffer list in preparation for
2297  * dropping it from its current transaction entirely.  If the buffer has
2298  * already started to be used by a subsequent transaction, refile the
2299  * buffer on that transaction's metadata list.
2300  *
2301  * Called under j_list_lock
2302  * Called under jbd_lock_bh_state(jh2bh(jh))
2303  *
2304  * jh and bh may be already free when this function returns
2305  */
2306 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2307 {
2308         int was_dirty, jlist;
2309         struct buffer_head *bh = jh2bh(jh);
2310
2311         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2312         if (jh->b_transaction)
2313                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2314
2315         /* If the buffer is now unused, just drop it. */
2316         if (jh->b_next_transaction == NULL) {
2317                 __jbd2_journal_unfile_buffer(jh);
2318                 return;
2319         }
2320
2321         /*
2322          * It has been modified by a later transaction: add it to the new
2323          * transaction's metadata list.
2324          */
2325
2326         was_dirty = test_clear_buffer_jbddirty(bh);
2327         __jbd2_journal_temp_unlink_buffer(jh);
2328         /*
2329          * We set b_transaction here because b_next_transaction will inherit
2330          * our jh reference and thus __jbd2_journal_file_buffer() must not
2331          * take a new one.
2332          */
2333         jh->b_transaction = jh->b_next_transaction;
2334         jh->b_next_transaction = NULL;
2335         if (buffer_freed(bh))
2336                 jlist = BJ_Forget;
2337         else if (jh->b_modified)
2338                 jlist = BJ_Metadata;
2339         else
2340                 jlist = BJ_Reserved;
2341         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2342         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2343
2344         if (was_dirty)
2345                 set_buffer_jbddirty(bh);
2346 }
2347
2348 /*
2349  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2350  * bh reference so that we can safely unlock bh.
2351  *
2352  * The jh and bh may be freed by this call.
2353  */
2354 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2355 {
2356         struct buffer_head *bh = jh2bh(jh);
2357
2358         /* Get reference so that buffer cannot be freed before we unlock it */
2359         get_bh(bh);
2360         jbd_lock_bh_state(bh);
2361         spin_lock(&journal->j_list_lock);
2362         __jbd2_journal_refile_buffer(jh);
2363         jbd_unlock_bh_state(bh);
2364         spin_unlock(&journal->j_list_lock);
2365         __brelse(bh);
2366 }
2367
2368 /*
2369  * File inode in the inode list of the handle's transaction
2370  */
2371 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2372 {
2373         transaction_t *transaction = handle->h_transaction;
2374         journal_t *journal;
2375
2376         WARN_ON(!transaction);
2377         if (is_handle_aborted(handle))
2378                 return -EROFS;
2379         journal = transaction->t_journal;
2380
2381         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2382                         transaction->t_tid);
2383
2384         /*
2385          * First check whether inode isn't already on the transaction's
2386          * lists without taking the lock. Note that this check is safe
2387          * without the lock as we cannot race with somebody removing inode
2388          * from the transaction. The reason is that we remove inode from the
2389          * transaction only in journal_release_jbd_inode() and when we commit
2390          * the transaction. We are guarded from the first case by holding
2391          * a reference to the inode. We are safe against the second case
2392          * because if jinode->i_transaction == transaction, commit code
2393          * cannot touch the transaction because we hold reference to it,
2394          * and if jinode->i_next_transaction == transaction, commit code
2395          * will only file the inode where we want it.
2396          */
2397         if (jinode->i_transaction == transaction ||
2398             jinode->i_next_transaction == transaction)
2399                 return 0;
2400
2401         spin_lock(&journal->j_list_lock);
2402
2403         if (jinode->i_transaction == transaction ||
2404             jinode->i_next_transaction == transaction)
2405                 goto done;
2406
2407         /*
2408          * We only ever set this variable to 1 so the test is safe. Since
2409          * t_need_data_flush is likely to be set, we do the test to save some
2410          * cacheline bouncing
2411          */
2412         if (!transaction->t_need_data_flush)
2413                 transaction->t_need_data_flush = 1;
2414         /* On some different transaction's list - should be
2415          * the committing one */
2416         if (jinode->i_transaction) {
2417                 J_ASSERT(jinode->i_next_transaction == NULL);
2418                 J_ASSERT(jinode->i_transaction ==
2419                                         journal->j_committing_transaction);
2420                 jinode->i_next_transaction = transaction;
2421                 goto done;
2422         }
2423         /* Not on any transaction list... */
2424         J_ASSERT(!jinode->i_next_transaction);
2425         jinode->i_transaction = transaction;
2426         list_add(&jinode->i_list, &transaction->t_inode_list);
2427 done:
2428         spin_unlock(&journal->j_list_lock);
2429
2430         return 0;
2431 }
2432
2433 /*
2434  * File truncate and transaction commit interact with each other in a
2435  * non-trivial way.  If a transaction writing data block A is
2436  * committing, we cannot discard the data by truncate until we have
2437  * written them.  Otherwise if we crashed after the transaction with
2438  * write has committed but before the transaction with truncate has
2439  * committed, we could see stale data in block A.  This function is a
2440  * helper to solve this problem.  It starts writeout of the truncated
2441  * part in case it is in the committing transaction.
2442  *
2443  * Filesystem code must call this function when inode is journaled in
2444  * ordered mode before truncation happens and after the inode has been
2445  * placed on orphan list with the new inode size. The second condition
2446  * avoids the race that someone writes new data and we start
2447  * committing the transaction after this function has been called but
2448  * before a transaction for truncate is started (and furthermore it
2449  * allows us to optimize the case where the addition to orphan list
2450  * happens in the same transaction as write --- we don't have to write
2451  * any data in such case).
2452  */
2453 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2454                                         struct jbd2_inode *jinode,
2455                                         loff_t new_size)
2456 {
2457         transaction_t *inode_trans, *commit_trans;
2458         int ret = 0;
2459
2460         /* This is a quick check to avoid locking if not necessary */
2461         if (!jinode->i_transaction)
2462                 goto out;
2463         /* Locks are here just to force reading of recent values, it is
2464          * enough that the transaction was not committing before we started
2465          * a transaction adding the inode to orphan list */
2466         read_lock(&journal->j_state_lock);
2467         commit_trans = journal->j_committing_transaction;
2468         read_unlock(&journal->j_state_lock);
2469         spin_lock(&journal->j_list_lock);
2470         inode_trans = jinode->i_transaction;
2471         spin_unlock(&journal->j_list_lock);
2472         if (inode_trans == commit_trans) {
2473                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2474                         new_size, LLONG_MAX);
2475                 if (ret)
2476                         jbd2_journal_abort(journal, ret);
2477         }
2478 out:
2479         return ret;
2480 }