Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi...
[linux-2.6-microblaze.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                 /*
282                  * If __GFP_FS is not present, then we may be being called from
283                  * inside the fs writeback layer, so we MUST NOT fail.
284                  */
285                 if ((gfp_mask & __GFP_FS) == 0)
286                         gfp_mask |= __GFP_NOFAIL;
287                 new_transaction = kmem_cache_zalloc(transaction_cache,
288                                                     gfp_mask);
289                 if (!new_transaction)
290                         return -ENOMEM;
291         }
292
293         jbd_debug(3, "New handle %p going live.\n", handle);
294
295         /*
296          * We need to hold j_state_lock until t_updates has been incremented,
297          * for proper journal barrier handling
298          */
299 repeat:
300         read_lock(&journal->j_state_lock);
301         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
302         if (is_journal_aborted(journal) ||
303             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
304                 read_unlock(&journal->j_state_lock);
305                 jbd2_journal_free_transaction(new_transaction);
306                 return -EROFS;
307         }
308
309         /*
310          * Wait on the journal's transaction barrier if necessary. Specifically
311          * we allow reserved handles to proceed because otherwise commit could
312          * deadlock on page writeback not being able to complete.
313          */
314         if (!handle->h_reserved && journal->j_barrier_count) {
315                 read_unlock(&journal->j_state_lock);
316                 wait_event(journal->j_wait_transaction_locked,
317                                 journal->j_barrier_count == 0);
318                 goto repeat;
319         }
320
321         if (!journal->j_running_transaction) {
322                 read_unlock(&journal->j_state_lock);
323                 if (!new_transaction)
324                         goto alloc_transaction;
325                 write_lock(&journal->j_state_lock);
326                 if (!journal->j_running_transaction &&
327                     (handle->h_reserved || !journal->j_barrier_count)) {
328                         jbd2_get_transaction(journal, new_transaction);
329                         new_transaction = NULL;
330                 }
331                 write_unlock(&journal->j_state_lock);
332                 goto repeat;
333         }
334
335         transaction = journal->j_running_transaction;
336
337         if (!handle->h_reserved) {
338                 /* We may have dropped j_state_lock - restart in that case */
339                 if (add_transaction_credits(journal, blocks, rsv_blocks))
340                         goto repeat;
341         } else {
342                 /*
343                  * We have handle reserved so we are allowed to join T_LOCKED
344                  * transaction and we don't have to check for transaction size
345                  * and journal space.
346                  */
347                 sub_reserved_credits(journal, blocks);
348                 handle->h_reserved = 0;
349         }
350
351         /* OK, account for the buffers that this operation expects to
352          * use and add the handle to the running transaction. 
353          */
354         update_t_max_wait(transaction, ts);
355         handle->h_transaction = transaction;
356         handle->h_requested_credits = blocks;
357         handle->h_start_jiffies = jiffies;
358         atomic_inc(&transaction->t_updates);
359         atomic_inc(&transaction->t_handle_count);
360         jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
361                   handle, blocks,
362                   atomic_read(&transaction->t_outstanding_credits),
363                   jbd2_log_space_left(journal));
364         read_unlock(&journal->j_state_lock);
365         current->journal_info = handle;
366
367         lock_map_acquire(&handle->h_lockdep_map);
368         jbd2_journal_free_transaction(new_transaction);
369         return 0;
370 }
371
372 static struct lock_class_key jbd2_handle_key;
373
374 /* Allocate a new handle.  This should probably be in a slab... */
375 static handle_t *new_handle(int nblocks)
376 {
377         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
378         if (!handle)
379                 return NULL;
380         handle->h_buffer_credits = nblocks;
381         handle->h_ref = 1;
382
383         lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
384                                                 &jbd2_handle_key, 0);
385
386         return handle;
387 }
388
389 /**
390  * handle_t *jbd2_journal_start() - Obtain a new handle.
391  * @journal: Journal to start transaction on.
392  * @nblocks: number of block buffer we might modify
393  *
394  * We make sure that the transaction can guarantee at least nblocks of
395  * modified buffers in the log.  We block until the log can guarantee
396  * that much space. Additionally, if rsv_blocks > 0, we also create another
397  * handle with rsv_blocks reserved blocks in the journal. This handle is
398  * is stored in h_rsv_handle. It is not attached to any particular transaction
399  * and thus doesn't block transaction commit. If the caller uses this reserved
400  * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
401  * on the parent handle will dispose the reserved one. Reserved handle has to
402  * be converted to a normal handle using jbd2_journal_start_reserved() before
403  * it can be used.
404  *
405  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
406  * on failure.
407  */
408 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
409                               gfp_t gfp_mask, unsigned int type,
410                               unsigned int line_no)
411 {
412         handle_t *handle = journal_current_handle();
413         int err;
414
415         if (!journal)
416                 return ERR_PTR(-EROFS);
417
418         if (handle) {
419                 J_ASSERT(handle->h_transaction->t_journal == journal);
420                 handle->h_ref++;
421                 return handle;
422         }
423
424         handle = new_handle(nblocks);
425         if (!handle)
426                 return ERR_PTR(-ENOMEM);
427         if (rsv_blocks) {
428                 handle_t *rsv_handle;
429
430                 rsv_handle = new_handle(rsv_blocks);
431                 if (!rsv_handle) {
432                         jbd2_free_handle(handle);
433                         return ERR_PTR(-ENOMEM);
434                 }
435                 rsv_handle->h_reserved = 1;
436                 rsv_handle->h_journal = journal;
437                 handle->h_rsv_handle = rsv_handle;
438         }
439
440         err = start_this_handle(journal, handle, gfp_mask);
441         if (err < 0) {
442                 if (handle->h_rsv_handle)
443                         jbd2_free_handle(handle->h_rsv_handle);
444                 jbd2_free_handle(handle);
445                 return ERR_PTR(err);
446         }
447         handle->h_type = type;
448         handle->h_line_no = line_no;
449         trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
450                                 handle->h_transaction->t_tid, type,
451                                 line_no, nblocks);
452         return handle;
453 }
454 EXPORT_SYMBOL(jbd2__journal_start);
455
456
457 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
458 {
459         return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0);
460 }
461 EXPORT_SYMBOL(jbd2_journal_start);
462
463 void jbd2_journal_free_reserved(handle_t *handle)
464 {
465         journal_t *journal = handle->h_journal;
466
467         WARN_ON(!handle->h_reserved);
468         sub_reserved_credits(journal, handle->h_buffer_credits);
469         jbd2_free_handle(handle);
470 }
471 EXPORT_SYMBOL(jbd2_journal_free_reserved);
472
473 /**
474  * int jbd2_journal_start_reserved(handle_t *handle) - start reserved handle
475  * @handle: handle to start
476  *
477  * Start handle that has been previously reserved with jbd2_journal_reserve().
478  * This attaches @handle to the running transaction (or creates one if there's
479  * not transaction running). Unlike jbd2_journal_start() this function cannot
480  * block on journal commit, checkpointing, or similar stuff. It can block on
481  * memory allocation or frozen journal though.
482  *
483  * Return 0 on success, non-zero on error - handle is freed in that case.
484  */
485 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
486                                 unsigned int line_no)
487 {
488         journal_t *journal = handle->h_journal;
489         int ret = -EIO;
490
491         if (WARN_ON(!handle->h_reserved)) {
492                 /* Someone passed in normal handle? Just stop it. */
493                 jbd2_journal_stop(handle);
494                 return ret;
495         }
496         /*
497          * Usefulness of mixing of reserved and unreserved handles is
498          * questionable. So far nobody seems to need it so just error out.
499          */
500         if (WARN_ON(current->journal_info)) {
501                 jbd2_journal_free_reserved(handle);
502                 return ret;
503         }
504
505         handle->h_journal = NULL;
506         /*
507          * GFP_NOFS is here because callers are likely from writeback or
508          * similarly constrained call sites
509          */
510         ret = start_this_handle(journal, handle, GFP_NOFS);
511         if (ret < 0) {
512                 jbd2_journal_free_reserved(handle);
513                 return ret;
514         }
515         handle->h_type = type;
516         handle->h_line_no = line_no;
517         return 0;
518 }
519 EXPORT_SYMBOL(jbd2_journal_start_reserved);
520
521 /**
522  * int jbd2_journal_extend() - extend buffer credits.
523  * @handle:  handle to 'extend'
524  * @nblocks: nr blocks to try to extend by.
525  *
526  * Some transactions, such as large extends and truncates, can be done
527  * atomically all at once or in several stages.  The operation requests
528  * a credit for a number of buffer modications in advance, but can
529  * extend its credit if it needs more.
530  *
531  * jbd2_journal_extend tries to give the running handle more buffer credits.
532  * It does not guarantee that allocation - this is a best-effort only.
533  * The calling process MUST be able to deal cleanly with a failure to
534  * extend here.
535  *
536  * Return 0 on success, non-zero on failure.
537  *
538  * return code < 0 implies an error
539  * return code > 0 implies normal transaction-full status.
540  */
541 int jbd2_journal_extend(handle_t *handle, int nblocks)
542 {
543         transaction_t *transaction = handle->h_transaction;
544         journal_t *journal;
545         int result;
546         int wanted;
547
548         if (is_handle_aborted(handle))
549                 return -EROFS;
550         journal = transaction->t_journal;
551
552         result = 1;
553
554         read_lock(&journal->j_state_lock);
555
556         /* Don't extend a locked-down transaction! */
557         if (transaction->t_state != T_RUNNING) {
558                 jbd_debug(3, "denied handle %p %d blocks: "
559                           "transaction not running\n", handle, nblocks);
560                 goto error_out;
561         }
562
563         spin_lock(&transaction->t_handle_lock);
564         wanted = atomic_add_return(nblocks,
565                                    &transaction->t_outstanding_credits);
566
567         if (wanted > journal->j_max_transaction_buffers) {
568                 jbd_debug(3, "denied handle %p %d blocks: "
569                           "transaction too large\n", handle, nblocks);
570                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
571                 goto unlock;
572         }
573
574         if (wanted + (wanted >> JBD2_CONTROL_BLOCKS_SHIFT) >
575             jbd2_log_space_left(journal)) {
576                 jbd_debug(3, "denied handle %p %d blocks: "
577                           "insufficient log space\n", handle, nblocks);
578                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
579                 goto unlock;
580         }
581
582         trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
583                                  transaction->t_tid,
584                                  handle->h_type, handle->h_line_no,
585                                  handle->h_buffer_credits,
586                                  nblocks);
587
588         handle->h_buffer_credits += nblocks;
589         handle->h_requested_credits += nblocks;
590         result = 0;
591
592         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
593 unlock:
594         spin_unlock(&transaction->t_handle_lock);
595 error_out:
596         read_unlock(&journal->j_state_lock);
597         return result;
598 }
599
600
601 /**
602  * int jbd2_journal_restart() - restart a handle .
603  * @handle:  handle to restart
604  * @nblocks: nr credits requested
605  *
606  * Restart a handle for a multi-transaction filesystem
607  * operation.
608  *
609  * If the jbd2_journal_extend() call above fails to grant new buffer credits
610  * to a running handle, a call to jbd2_journal_restart will commit the
611  * handle's transaction so far and reattach the handle to a new
612  * transaction capabable of guaranteeing the requested number of
613  * credits. We preserve reserved handle if there's any attached to the
614  * passed in handle.
615  */
616 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
617 {
618         transaction_t *transaction = handle->h_transaction;
619         journal_t *journal;
620         tid_t           tid;
621         int             need_to_start, ret;
622
623         /* If we've had an abort of any type, don't even think about
624          * actually doing the restart! */
625         if (is_handle_aborted(handle))
626                 return 0;
627         journal = transaction->t_journal;
628
629         /*
630          * First unlink the handle from its current transaction, and start the
631          * commit on that.
632          */
633         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
634         J_ASSERT(journal_current_handle() == handle);
635
636         read_lock(&journal->j_state_lock);
637         spin_lock(&transaction->t_handle_lock);
638         atomic_sub(handle->h_buffer_credits,
639                    &transaction->t_outstanding_credits);
640         if (handle->h_rsv_handle) {
641                 sub_reserved_credits(journal,
642                                      handle->h_rsv_handle->h_buffer_credits);
643         }
644         if (atomic_dec_and_test(&transaction->t_updates))
645                 wake_up(&journal->j_wait_updates);
646         tid = transaction->t_tid;
647         spin_unlock(&transaction->t_handle_lock);
648         handle->h_transaction = NULL;
649         current->journal_info = NULL;
650
651         jbd_debug(2, "restarting handle %p\n", handle);
652         need_to_start = !tid_geq(journal->j_commit_request, tid);
653         read_unlock(&journal->j_state_lock);
654         if (need_to_start)
655                 jbd2_log_start_commit(journal, tid);
656
657         lock_map_release(&handle->h_lockdep_map);
658         handle->h_buffer_credits = nblocks;
659         ret = start_this_handle(journal, handle, gfp_mask);
660         return ret;
661 }
662 EXPORT_SYMBOL(jbd2__journal_restart);
663
664
665 int jbd2_journal_restart(handle_t *handle, int nblocks)
666 {
667         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
668 }
669 EXPORT_SYMBOL(jbd2_journal_restart);
670
671 /**
672  * void jbd2_journal_lock_updates () - establish a transaction barrier.
673  * @journal:  Journal to establish a barrier on.
674  *
675  * This locks out any further updates from being started, and blocks
676  * until all existing updates have completed, returning only once the
677  * journal is in a quiescent state with no updates running.
678  *
679  * The journal lock should not be held on entry.
680  */
681 void jbd2_journal_lock_updates(journal_t *journal)
682 {
683         DEFINE_WAIT(wait);
684
685         write_lock(&journal->j_state_lock);
686         ++journal->j_barrier_count;
687
688         /* Wait until there are no reserved handles */
689         if (atomic_read(&journal->j_reserved_credits)) {
690                 write_unlock(&journal->j_state_lock);
691                 wait_event(journal->j_wait_reserved,
692                            atomic_read(&journal->j_reserved_credits) == 0);
693                 write_lock(&journal->j_state_lock);
694         }
695
696         /* Wait until there are no running updates */
697         while (1) {
698                 transaction_t *transaction = journal->j_running_transaction;
699
700                 if (!transaction)
701                         break;
702
703                 spin_lock(&transaction->t_handle_lock);
704                 prepare_to_wait(&journal->j_wait_updates, &wait,
705                                 TASK_UNINTERRUPTIBLE);
706                 if (!atomic_read(&transaction->t_updates)) {
707                         spin_unlock(&transaction->t_handle_lock);
708                         finish_wait(&journal->j_wait_updates, &wait);
709                         break;
710                 }
711                 spin_unlock(&transaction->t_handle_lock);
712                 write_unlock(&journal->j_state_lock);
713                 schedule();
714                 finish_wait(&journal->j_wait_updates, &wait);
715                 write_lock(&journal->j_state_lock);
716         }
717         write_unlock(&journal->j_state_lock);
718
719         /*
720          * We have now established a barrier against other normal updates, but
721          * we also need to barrier against other jbd2_journal_lock_updates() calls
722          * to make sure that we serialise special journal-locked operations
723          * too.
724          */
725         mutex_lock(&journal->j_barrier);
726 }
727
728 /**
729  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
730  * @journal:  Journal to release the barrier on.
731  *
732  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
733  *
734  * Should be called without the journal lock held.
735  */
736 void jbd2_journal_unlock_updates (journal_t *journal)
737 {
738         J_ASSERT(journal->j_barrier_count != 0);
739
740         mutex_unlock(&journal->j_barrier);
741         write_lock(&journal->j_state_lock);
742         --journal->j_barrier_count;
743         write_unlock(&journal->j_state_lock);
744         wake_up(&journal->j_wait_transaction_locked);
745 }
746
747 static void warn_dirty_buffer(struct buffer_head *bh)
748 {
749         char b[BDEVNAME_SIZE];
750
751         printk(KERN_WARNING
752                "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
753                "There's a risk of filesystem corruption in case of system "
754                "crash.\n",
755                bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
756 }
757
758 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
759 static void jbd2_freeze_jh_data(struct journal_head *jh)
760 {
761         struct page *page;
762         int offset;
763         char *source;
764         struct buffer_head *bh = jh2bh(jh);
765
766         J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
767         page = bh->b_page;
768         offset = offset_in_page(bh->b_data);
769         source = kmap_atomic(page);
770         /* Fire data frozen trigger just before we copy the data */
771         jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
772         memcpy(jh->b_frozen_data, source + offset, bh->b_size);
773         kunmap_atomic(source);
774
775         /*
776          * Now that the frozen data is saved off, we need to store any matching
777          * triggers.
778          */
779         jh->b_frozen_triggers = jh->b_triggers;
780 }
781
782 /*
783  * If the buffer is already part of the current transaction, then there
784  * is nothing we need to do.  If it is already part of a prior
785  * transaction which we are still committing to disk, then we need to
786  * make sure that we do not overwrite the old copy: we do copy-out to
787  * preserve the copy going to disk.  We also account the buffer against
788  * the handle's metadata buffer credits (unless the buffer is already
789  * part of the transaction, that is).
790  *
791  */
792 static int
793 do_get_write_access(handle_t *handle, struct journal_head *jh,
794                         int force_copy)
795 {
796         struct buffer_head *bh;
797         transaction_t *transaction = handle->h_transaction;
798         journal_t *journal;
799         int error;
800         char *frozen_buffer = NULL;
801         unsigned long start_lock, time_lock;
802
803         if (is_handle_aborted(handle))
804                 return -EROFS;
805         journal = transaction->t_journal;
806
807         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
808
809         JBUFFER_TRACE(jh, "entry");
810 repeat:
811         bh = jh2bh(jh);
812
813         /* @@@ Need to check for errors here at some point. */
814
815         start_lock = jiffies;
816         lock_buffer(bh);
817         jbd_lock_bh_state(bh);
818
819         /* If it takes too long to lock the buffer, trace it */
820         time_lock = jbd2_time_diff(start_lock, jiffies);
821         if (time_lock > HZ/10)
822                 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
823                         jiffies_to_msecs(time_lock));
824
825         /* We now hold the buffer lock so it is safe to query the buffer
826          * state.  Is the buffer dirty?
827          *
828          * If so, there are two possibilities.  The buffer may be
829          * non-journaled, and undergoing a quite legitimate writeback.
830          * Otherwise, it is journaled, and we don't expect dirty buffers
831          * in that state (the buffers should be marked JBD_Dirty
832          * instead.)  So either the IO is being done under our own
833          * control and this is a bug, or it's a third party IO such as
834          * dump(8) (which may leave the buffer scheduled for read ---
835          * ie. locked but not dirty) or tune2fs (which may actually have
836          * the buffer dirtied, ugh.)  */
837
838         if (buffer_dirty(bh)) {
839                 /*
840                  * First question: is this buffer already part of the current
841                  * transaction or the existing committing transaction?
842                  */
843                 if (jh->b_transaction) {
844                         J_ASSERT_JH(jh,
845                                 jh->b_transaction == transaction ||
846                                 jh->b_transaction ==
847                                         journal->j_committing_transaction);
848                         if (jh->b_next_transaction)
849                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
850                                                         transaction);
851                         warn_dirty_buffer(bh);
852                 }
853                 /*
854                  * In any case we need to clean the dirty flag and we must
855                  * do it under the buffer lock to be sure we don't race
856                  * with running write-out.
857                  */
858                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
859                 clear_buffer_dirty(bh);
860                 set_buffer_jbddirty(bh);
861         }
862
863         unlock_buffer(bh);
864
865         error = -EROFS;
866         if (is_handle_aborted(handle)) {
867                 jbd_unlock_bh_state(bh);
868                 goto out;
869         }
870         error = 0;
871
872         /*
873          * The buffer is already part of this transaction if b_transaction or
874          * b_next_transaction points to it
875          */
876         if (jh->b_transaction == transaction ||
877             jh->b_next_transaction == transaction)
878                 goto done;
879
880         /*
881          * this is the first time this transaction is touching this buffer,
882          * reset the modified flag
883          */
884        jh->b_modified = 0;
885
886         /*
887          * If the buffer is not journaled right now, we need to make sure it
888          * doesn't get written to disk before the caller actually commits the
889          * new data
890          */
891         if (!jh->b_transaction) {
892                 JBUFFER_TRACE(jh, "no transaction");
893                 J_ASSERT_JH(jh, !jh->b_next_transaction);
894                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
895                 /*
896                  * Make sure all stores to jh (b_modified, b_frozen_data) are
897                  * visible before attaching it to the running transaction.
898                  * Paired with barrier in jbd2_write_access_granted()
899                  */
900                 smp_wmb();
901                 spin_lock(&journal->j_list_lock);
902                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
903                 spin_unlock(&journal->j_list_lock);
904                 goto done;
905         }
906         /*
907          * If there is already a copy-out version of this buffer, then we don't
908          * need to make another one
909          */
910         if (jh->b_frozen_data) {
911                 JBUFFER_TRACE(jh, "has frozen data");
912                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
913                 goto attach_next;
914         }
915
916         JBUFFER_TRACE(jh, "owned by older transaction");
917         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
918         J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
919
920         /*
921          * There is one case we have to be very careful about.  If the
922          * committing transaction is currently writing this buffer out to disk
923          * and has NOT made a copy-out, then we cannot modify the buffer
924          * contents at all right now.  The essence of copy-out is that it is
925          * the extra copy, not the primary copy, which gets journaled.  If the
926          * primary copy is already going to disk then we cannot do copy-out
927          * here.
928          */
929         if (buffer_shadow(bh)) {
930                 JBUFFER_TRACE(jh, "on shadow: sleep");
931                 jbd_unlock_bh_state(bh);
932                 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
933                 goto repeat;
934         }
935
936         /*
937          * Only do the copy if the currently-owning transaction still needs it.
938          * If buffer isn't on BJ_Metadata list, the committing transaction is
939          * past that stage (here we use the fact that BH_Shadow is set under
940          * bh_state lock together with refiling to BJ_Shadow list and at this
941          * point we know the buffer doesn't have BH_Shadow set).
942          *
943          * Subtle point, though: if this is a get_undo_access, then we will be
944          * relying on the frozen_data to contain the new value of the
945          * committed_data record after the transaction, so we HAVE to force the
946          * frozen_data copy in that case.
947          */
948         if (jh->b_jlist == BJ_Metadata || force_copy) {
949                 JBUFFER_TRACE(jh, "generate frozen data");
950                 if (!frozen_buffer) {
951                         JBUFFER_TRACE(jh, "allocate memory for buffer");
952                         jbd_unlock_bh_state(bh);
953                         frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
954                         if (!frozen_buffer) {
955                                 printk(KERN_ERR "%s: OOM for frozen_buffer\n",
956                                        __func__);
957                                 JBUFFER_TRACE(jh, "oom!");
958                                 error = -ENOMEM;
959                                 goto out;
960                         }
961                         goto repeat;
962                 }
963                 jh->b_frozen_data = frozen_buffer;
964                 frozen_buffer = NULL;
965                 jbd2_freeze_jh_data(jh);
966         }
967 attach_next:
968         /*
969          * Make sure all stores to jh (b_modified, b_frozen_data) are visible
970          * before attaching it to the running transaction. Paired with barrier
971          * in jbd2_write_access_granted()
972          */
973         smp_wmb();
974         jh->b_next_transaction = transaction;
975
976 done:
977         jbd_unlock_bh_state(bh);
978
979         /*
980          * If we are about to journal a buffer, then any revoke pending on it is
981          * no longer valid
982          */
983         jbd2_journal_cancel_revoke(handle, jh);
984
985 out:
986         if (unlikely(frozen_buffer))    /* It's usually NULL */
987                 jbd2_free(frozen_buffer, bh->b_size);
988
989         JBUFFER_TRACE(jh, "exit");
990         return error;
991 }
992
993 /* Fast check whether buffer is already attached to the required transaction */
994 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh)
995 {
996         struct journal_head *jh;
997         bool ret = false;
998
999         /* Dirty buffers require special handling... */
1000         if (buffer_dirty(bh))
1001                 return false;
1002
1003         /*
1004          * RCU protects us from dereferencing freed pages. So the checks we do
1005          * are guaranteed not to oops. However the jh slab object can get freed
1006          * & reallocated while we work with it. So we have to be careful. When
1007          * we see jh attached to the running transaction, we know it must stay
1008          * so until the transaction is committed. Thus jh won't be freed and
1009          * will be attached to the same bh while we run.  However it can
1010          * happen jh gets freed, reallocated, and attached to the transaction
1011          * just after we get pointer to it from bh. So we have to be careful
1012          * and recheck jh still belongs to our bh before we return success.
1013          */
1014         rcu_read_lock();
1015         if (!buffer_jbd(bh))
1016                 goto out;
1017         /* This should be bh2jh() but that doesn't work with inline functions */
1018         jh = READ_ONCE(bh->b_private);
1019         if (!jh)
1020                 goto out;
1021         if (jh->b_transaction != handle->h_transaction &&
1022             jh->b_next_transaction != handle->h_transaction)
1023                 goto out;
1024         /*
1025          * There are two reasons for the barrier here:
1026          * 1) Make sure to fetch b_bh after we did previous checks so that we
1027          * detect when jh went through free, realloc, attach to transaction
1028          * while we were checking. Paired with implicit barrier in that path.
1029          * 2) So that access to bh done after jbd2_write_access_granted()
1030          * doesn't get reordered and see inconsistent state of concurrent
1031          * do_get_write_access().
1032          */
1033         smp_mb();
1034         if (unlikely(jh->b_bh != bh))
1035                 goto out;
1036         ret = true;
1037 out:
1038         rcu_read_unlock();
1039         return ret;
1040 }
1041
1042 /**
1043  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1044  * @handle: transaction to add buffer modifications to
1045  * @bh:     bh to be used for metadata writes
1046  *
1047  * Returns an error code or 0 on success.
1048  *
1049  * In full data journalling mode the buffer may be of type BJ_AsyncData,
1050  * because we're write()ing a buffer which is also part of a shared mapping.
1051  */
1052
1053 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1054 {
1055         struct journal_head *jh;
1056         int rc;
1057
1058         if (jbd2_write_access_granted(handle, bh))
1059                 return 0;
1060
1061         jh = jbd2_journal_add_journal_head(bh);
1062         /* We do not want to get caught playing with fields which the
1063          * log thread also manipulates.  Make sure that the buffer
1064          * completes any outstanding IO before proceeding. */
1065         rc = do_get_write_access(handle, jh, 0);
1066         jbd2_journal_put_journal_head(jh);
1067         return rc;
1068 }
1069
1070
1071 /*
1072  * When the user wants to journal a newly created buffer_head
1073  * (ie. getblk() returned a new buffer and we are going to populate it
1074  * manually rather than reading off disk), then we need to keep the
1075  * buffer_head locked until it has been completely filled with new
1076  * data.  In this case, we should be able to make the assertion that
1077  * the bh is not already part of an existing transaction.
1078  *
1079  * The buffer should already be locked by the caller by this point.
1080  * There is no lock ranking violation: it was a newly created,
1081  * unlocked buffer beforehand. */
1082
1083 /**
1084  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1085  * @handle: transaction to new buffer to
1086  * @bh: new buffer.
1087  *
1088  * Call this if you create a new bh.
1089  */
1090 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1091 {
1092         transaction_t *transaction = handle->h_transaction;
1093         journal_t *journal;
1094         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1095         int err;
1096
1097         jbd_debug(5, "journal_head %p\n", jh);
1098         err = -EROFS;
1099         if (is_handle_aborted(handle))
1100                 goto out;
1101         journal = transaction->t_journal;
1102         err = 0;
1103
1104         JBUFFER_TRACE(jh, "entry");
1105         /*
1106          * The buffer may already belong to this transaction due to pre-zeroing
1107          * in the filesystem's new_block code.  It may also be on the previous,
1108          * committing transaction's lists, but it HAS to be in Forget state in
1109          * that case: the transaction must have deleted the buffer for it to be
1110          * reused here.
1111          */
1112         jbd_lock_bh_state(bh);
1113         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1114                 jh->b_transaction == NULL ||
1115                 (jh->b_transaction == journal->j_committing_transaction &&
1116                           jh->b_jlist == BJ_Forget)));
1117
1118         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1119         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1120
1121         if (jh->b_transaction == NULL) {
1122                 /*
1123                  * Previous jbd2_journal_forget() could have left the buffer
1124                  * with jbddirty bit set because it was being committed. When
1125                  * the commit finished, we've filed the buffer for
1126                  * checkpointing and marked it dirty. Now we are reallocating
1127                  * the buffer so the transaction freeing it must have
1128                  * committed and so it's safe to clear the dirty bit.
1129                  */
1130                 clear_buffer_dirty(jh2bh(jh));
1131                 /* first access by this transaction */
1132                 jh->b_modified = 0;
1133
1134                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1135                 spin_lock(&journal->j_list_lock);
1136                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1137         } else if (jh->b_transaction == journal->j_committing_transaction) {
1138                 /* first access by this transaction */
1139                 jh->b_modified = 0;
1140
1141                 JBUFFER_TRACE(jh, "set next transaction");
1142                 spin_lock(&journal->j_list_lock);
1143                 jh->b_next_transaction = transaction;
1144         }
1145         spin_unlock(&journal->j_list_lock);
1146         jbd_unlock_bh_state(bh);
1147
1148         /*
1149          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
1150          * blocks which contain freed but then revoked metadata.  We need
1151          * to cancel the revoke in case we end up freeing it yet again
1152          * and the reallocating as data - this would cause a second revoke,
1153          * which hits an assertion error.
1154          */
1155         JBUFFER_TRACE(jh, "cancelling revoke");
1156         jbd2_journal_cancel_revoke(handle, jh);
1157 out:
1158         jbd2_journal_put_journal_head(jh);
1159         return err;
1160 }
1161
1162 /**
1163  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
1164  *     non-rewindable consequences
1165  * @handle: transaction
1166  * @bh: buffer to undo
1167  *
1168  * Sometimes there is a need to distinguish between metadata which has
1169  * been committed to disk and that which has not.  The ext3fs code uses
1170  * this for freeing and allocating space, we have to make sure that we
1171  * do not reuse freed space until the deallocation has been committed,
1172  * since if we overwrote that space we would make the delete
1173  * un-rewindable in case of a crash.
1174  *
1175  * To deal with that, jbd2_journal_get_undo_access requests write access to a
1176  * buffer for parts of non-rewindable operations such as delete
1177  * operations on the bitmaps.  The journaling code must keep a copy of
1178  * the buffer's contents prior to the undo_access call until such time
1179  * as we know that the buffer has definitely been committed to disk.
1180  *
1181  * We never need to know which transaction the committed data is part
1182  * of, buffers touched here are guaranteed to be dirtied later and so
1183  * will be committed to a new transaction in due course, at which point
1184  * we can discard the old committed data pointer.
1185  *
1186  * Returns error number or 0 on success.
1187  */
1188 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1189 {
1190         int err;
1191         struct journal_head *jh;
1192         char *committed_data = NULL;
1193
1194         JBUFFER_TRACE(jh, "entry");
1195         if (jbd2_write_access_granted(handle, bh))
1196                 return 0;
1197
1198         jh = jbd2_journal_add_journal_head(bh);
1199         /*
1200          * Do this first --- it can drop the journal lock, so we want to
1201          * make sure that obtaining the committed_data is done
1202          * atomically wrt. completion of any outstanding commits.
1203          */
1204         err = do_get_write_access(handle, jh, 1);
1205         if (err)
1206                 goto out;
1207
1208 repeat:
1209         if (!jh->b_committed_data) {
1210                 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
1211                 if (!committed_data) {
1212                         printk(KERN_ERR "%s: No memory for committed data\n",
1213                                 __func__);
1214                         err = -ENOMEM;
1215                         goto out;
1216                 }
1217         }
1218
1219         jbd_lock_bh_state(bh);
1220         if (!jh->b_committed_data) {
1221                 /* Copy out the current buffer contents into the
1222                  * preserved, committed copy. */
1223                 JBUFFER_TRACE(jh, "generate b_committed data");
1224                 if (!committed_data) {
1225                         jbd_unlock_bh_state(bh);
1226                         goto repeat;
1227                 }
1228
1229                 jh->b_committed_data = committed_data;
1230                 committed_data = NULL;
1231                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1232         }
1233         jbd_unlock_bh_state(bh);
1234 out:
1235         jbd2_journal_put_journal_head(jh);
1236         if (unlikely(committed_data))
1237                 jbd2_free(committed_data, bh->b_size);
1238         return err;
1239 }
1240
1241 /**
1242  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1243  * @bh: buffer to trigger on
1244  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1245  *
1246  * Set any triggers on this journal_head.  This is always safe, because
1247  * triggers for a committing buffer will be saved off, and triggers for
1248  * a running transaction will match the buffer in that transaction.
1249  *
1250  * Call with NULL to clear the triggers.
1251  */
1252 void jbd2_journal_set_triggers(struct buffer_head *bh,
1253                                struct jbd2_buffer_trigger_type *type)
1254 {
1255         struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1256
1257         if (WARN_ON(!jh))
1258                 return;
1259         jh->b_triggers = type;
1260         jbd2_journal_put_journal_head(jh);
1261 }
1262
1263 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1264                                 struct jbd2_buffer_trigger_type *triggers)
1265 {
1266         struct buffer_head *bh = jh2bh(jh);
1267
1268         if (!triggers || !triggers->t_frozen)
1269                 return;
1270
1271         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1272 }
1273
1274 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1275                                struct jbd2_buffer_trigger_type *triggers)
1276 {
1277         if (!triggers || !triggers->t_abort)
1278                 return;
1279
1280         triggers->t_abort(triggers, jh2bh(jh));
1281 }
1282
1283
1284
1285 /**
1286  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1287  * @handle: transaction to add buffer to.
1288  * @bh: buffer to mark
1289  *
1290  * mark dirty metadata which needs to be journaled as part of the current
1291  * transaction.
1292  *
1293  * The buffer must have previously had jbd2_journal_get_write_access()
1294  * called so that it has a valid journal_head attached to the buffer
1295  * head.
1296  *
1297  * The buffer is placed on the transaction's metadata list and is marked
1298  * as belonging to the transaction.
1299  *
1300  * Returns error number or 0 on success.
1301  *
1302  * Special care needs to be taken if the buffer already belongs to the
1303  * current committing transaction (in which case we should have frozen
1304  * data present for that commit).  In that case, we don't relink the
1305  * buffer: that only gets done when the old transaction finally
1306  * completes its commit.
1307  */
1308 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1309 {
1310         transaction_t *transaction = handle->h_transaction;
1311         journal_t *journal;
1312         struct journal_head *jh;
1313         int ret = 0;
1314
1315         if (is_handle_aborted(handle))
1316                 return -EROFS;
1317         journal = transaction->t_journal;
1318         jh = jbd2_journal_grab_journal_head(bh);
1319         if (!jh) {
1320                 ret = -EUCLEAN;
1321                 goto out;
1322         }
1323         jbd_debug(5, "journal_head %p\n", jh);
1324         JBUFFER_TRACE(jh, "entry");
1325
1326         jbd_lock_bh_state(bh);
1327
1328         if (jh->b_modified == 0) {
1329                 /*
1330                  * This buffer's got modified and becoming part
1331                  * of the transaction. This needs to be done
1332                  * once a transaction -bzzz
1333                  */
1334                 jh->b_modified = 1;
1335                 if (handle->h_buffer_credits <= 0) {
1336                         ret = -ENOSPC;
1337                         goto out_unlock_bh;
1338                 }
1339                 handle->h_buffer_credits--;
1340         }
1341
1342         /*
1343          * fastpath, to avoid expensive locking.  If this buffer is already
1344          * on the running transaction's metadata list there is nothing to do.
1345          * Nobody can take it off again because there is a handle open.
1346          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1347          * result in this test being false, so we go in and take the locks.
1348          */
1349         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1350                 JBUFFER_TRACE(jh, "fastpath");
1351                 if (unlikely(jh->b_transaction !=
1352                              journal->j_running_transaction)) {
1353                         printk(KERN_ERR "JBD2: %s: "
1354                                "jh->b_transaction (%llu, %p, %u) != "
1355                                "journal->j_running_transaction (%p, %u)\n",
1356                                journal->j_devname,
1357                                (unsigned long long) bh->b_blocknr,
1358                                jh->b_transaction,
1359                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1360                                journal->j_running_transaction,
1361                                journal->j_running_transaction ?
1362                                journal->j_running_transaction->t_tid : 0);
1363                         ret = -EINVAL;
1364                 }
1365                 goto out_unlock_bh;
1366         }
1367
1368         set_buffer_jbddirty(bh);
1369
1370         /*
1371          * Metadata already on the current transaction list doesn't
1372          * need to be filed.  Metadata on another transaction's list must
1373          * be committing, and will be refiled once the commit completes:
1374          * leave it alone for now.
1375          */
1376         if (jh->b_transaction != transaction) {
1377                 JBUFFER_TRACE(jh, "already on other transaction");
1378                 if (unlikely(((jh->b_transaction !=
1379                                journal->j_committing_transaction)) ||
1380                              (jh->b_next_transaction != transaction))) {
1381                         printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1382                                "bad jh for block %llu: "
1383                                "transaction (%p, %u), "
1384                                "jh->b_transaction (%p, %u), "
1385                                "jh->b_next_transaction (%p, %u), jlist %u\n",
1386                                journal->j_devname,
1387                                (unsigned long long) bh->b_blocknr,
1388                                transaction, transaction->t_tid,
1389                                jh->b_transaction,
1390                                jh->b_transaction ?
1391                                jh->b_transaction->t_tid : 0,
1392                                jh->b_next_transaction,
1393                                jh->b_next_transaction ?
1394                                jh->b_next_transaction->t_tid : 0,
1395                                jh->b_jlist);
1396                         WARN_ON(1);
1397                         ret = -EINVAL;
1398                 }
1399                 /* And this case is illegal: we can't reuse another
1400                  * transaction's data buffer, ever. */
1401                 goto out_unlock_bh;
1402         }
1403
1404         /* That test should have eliminated the following case: */
1405         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1406
1407         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1408         spin_lock(&journal->j_list_lock);
1409         __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1410         spin_unlock(&journal->j_list_lock);
1411 out_unlock_bh:
1412         jbd_unlock_bh_state(bh);
1413         jbd2_journal_put_journal_head(jh);
1414 out:
1415         JBUFFER_TRACE(jh, "exit");
1416         return ret;
1417 }
1418
1419 /**
1420  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1421  * @handle: transaction handle
1422  * @bh:     bh to 'forget'
1423  *
1424  * We can only do the bforget if there are no commits pending against the
1425  * buffer.  If the buffer is dirty in the current running transaction we
1426  * can safely unlink it.
1427  *
1428  * bh may not be a journalled buffer at all - it may be a non-JBD
1429  * buffer which came off the hashtable.  Check for this.
1430  *
1431  * Decrements bh->b_count by one.
1432  *
1433  * Allow this call even if the handle has aborted --- it may be part of
1434  * the caller's cleanup after an abort.
1435  */
1436 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1437 {
1438         transaction_t *transaction = handle->h_transaction;
1439         journal_t *journal;
1440         struct journal_head *jh;
1441         int drop_reserve = 0;
1442         int err = 0;
1443         int was_modified = 0;
1444
1445         if (is_handle_aborted(handle))
1446                 return -EROFS;
1447         journal = transaction->t_journal;
1448
1449         BUFFER_TRACE(bh, "entry");
1450
1451         jbd_lock_bh_state(bh);
1452
1453         if (!buffer_jbd(bh))
1454                 goto not_jbd;
1455         jh = bh2jh(bh);
1456
1457         /* Critical error: attempting to delete a bitmap buffer, maybe?
1458          * Don't do any jbd operations, and return an error. */
1459         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1460                          "inconsistent data on disk")) {
1461                 err = -EIO;
1462                 goto not_jbd;
1463         }
1464
1465         /* keep track of whether or not this transaction modified us */
1466         was_modified = jh->b_modified;
1467
1468         /*
1469          * The buffer's going from the transaction, we must drop
1470          * all references -bzzz
1471          */
1472         jh->b_modified = 0;
1473
1474         if (jh->b_transaction == transaction) {
1475                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1476
1477                 /* If we are forgetting a buffer which is already part
1478                  * of this transaction, then we can just drop it from
1479                  * the transaction immediately. */
1480                 clear_buffer_dirty(bh);
1481                 clear_buffer_jbddirty(bh);
1482
1483                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1484
1485                 /*
1486                  * we only want to drop a reference if this transaction
1487                  * modified the buffer
1488                  */
1489                 if (was_modified)
1490                         drop_reserve = 1;
1491
1492                 /*
1493                  * We are no longer going to journal this buffer.
1494                  * However, the commit of this transaction is still
1495                  * important to the buffer: the delete that we are now
1496                  * processing might obsolete an old log entry, so by
1497                  * committing, we can satisfy the buffer's checkpoint.
1498                  *
1499                  * So, if we have a checkpoint on the buffer, we should
1500                  * now refile the buffer on our BJ_Forget list so that
1501                  * we know to remove the checkpoint after we commit.
1502                  */
1503
1504                 spin_lock(&journal->j_list_lock);
1505                 if (jh->b_cp_transaction) {
1506                         __jbd2_journal_temp_unlink_buffer(jh);
1507                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1508                 } else {
1509                         __jbd2_journal_unfile_buffer(jh);
1510                         if (!buffer_jbd(bh)) {
1511                                 spin_unlock(&journal->j_list_lock);
1512                                 jbd_unlock_bh_state(bh);
1513                                 __bforget(bh);
1514                                 goto drop;
1515                         }
1516                 }
1517                 spin_unlock(&journal->j_list_lock);
1518         } else if (jh->b_transaction) {
1519                 J_ASSERT_JH(jh, (jh->b_transaction ==
1520                                  journal->j_committing_transaction));
1521                 /* However, if the buffer is still owned by a prior
1522                  * (committing) transaction, we can't drop it yet... */
1523                 JBUFFER_TRACE(jh, "belongs to older transaction");
1524                 /* ... but we CAN drop it from the new transaction if we
1525                  * have also modified it since the original commit. */
1526
1527                 if (jh->b_next_transaction) {
1528                         J_ASSERT(jh->b_next_transaction == transaction);
1529                         spin_lock(&journal->j_list_lock);
1530                         jh->b_next_transaction = NULL;
1531                         spin_unlock(&journal->j_list_lock);
1532
1533                         /*
1534                          * only drop a reference if this transaction modified
1535                          * the buffer
1536                          */
1537                         if (was_modified)
1538                                 drop_reserve = 1;
1539                 }
1540         }
1541
1542 not_jbd:
1543         jbd_unlock_bh_state(bh);
1544         __brelse(bh);
1545 drop:
1546         if (drop_reserve) {
1547                 /* no need to reserve log space for this block -bzzz */
1548                 handle->h_buffer_credits++;
1549         }
1550         return err;
1551 }
1552
1553 /**
1554  * int jbd2_journal_stop() - complete a transaction
1555  * @handle: tranaction to complete.
1556  *
1557  * All done for a particular handle.
1558  *
1559  * There is not much action needed here.  We just return any remaining
1560  * buffer credits to the transaction and remove the handle.  The only
1561  * complication is that we need to start a commit operation if the
1562  * filesystem is marked for synchronous update.
1563  *
1564  * jbd2_journal_stop itself will not usually return an error, but it may
1565  * do so in unusual circumstances.  In particular, expect it to
1566  * return -EIO if a jbd2_journal_abort has been executed since the
1567  * transaction began.
1568  */
1569 int jbd2_journal_stop(handle_t *handle)
1570 {
1571         transaction_t *transaction = handle->h_transaction;
1572         journal_t *journal;
1573         int err = 0, wait_for_commit = 0;
1574         tid_t tid;
1575         pid_t pid;
1576
1577         if (!transaction) {
1578                 /*
1579                  * Handle is already detached from the transaction so
1580                  * there is nothing to do other than decrease a refcount,
1581                  * or free the handle if refcount drops to zero
1582                  */
1583                 if (--handle->h_ref > 0) {
1584                         jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1585                                                          handle->h_ref);
1586                         return err;
1587                 } else {
1588                         if (handle->h_rsv_handle)
1589                                 jbd2_free_handle(handle->h_rsv_handle);
1590                         goto free_and_exit;
1591                 }
1592         }
1593         journal = transaction->t_journal;
1594
1595         J_ASSERT(journal_current_handle() == handle);
1596
1597         if (is_handle_aborted(handle))
1598                 err = -EIO;
1599         else
1600                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1601
1602         if (--handle->h_ref > 0) {
1603                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1604                           handle->h_ref);
1605                 return err;
1606         }
1607
1608         jbd_debug(4, "Handle %p going down\n", handle);
1609         trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1610                                 transaction->t_tid,
1611                                 handle->h_type, handle->h_line_no,
1612                                 jiffies - handle->h_start_jiffies,
1613                                 handle->h_sync, handle->h_requested_credits,
1614                                 (handle->h_requested_credits -
1615                                  handle->h_buffer_credits));
1616
1617         /*
1618          * Implement synchronous transaction batching.  If the handle
1619          * was synchronous, don't force a commit immediately.  Let's
1620          * yield and let another thread piggyback onto this
1621          * transaction.  Keep doing that while new threads continue to
1622          * arrive.  It doesn't cost much - we're about to run a commit
1623          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1624          * operations by 30x or more...
1625          *
1626          * We try and optimize the sleep time against what the
1627          * underlying disk can do, instead of having a static sleep
1628          * time.  This is useful for the case where our storage is so
1629          * fast that it is more optimal to go ahead and force a flush
1630          * and wait for the transaction to be committed than it is to
1631          * wait for an arbitrary amount of time for new writers to
1632          * join the transaction.  We achieve this by measuring how
1633          * long it takes to commit a transaction, and compare it with
1634          * how long this transaction has been running, and if run time
1635          * < commit time then we sleep for the delta and commit.  This
1636          * greatly helps super fast disks that would see slowdowns as
1637          * more threads started doing fsyncs.
1638          *
1639          * But don't do this if this process was the most recent one
1640          * to perform a synchronous write.  We do this to detect the
1641          * case where a single process is doing a stream of sync
1642          * writes.  No point in waiting for joiners in that case.
1643          *
1644          * Setting max_batch_time to 0 disables this completely.
1645          */
1646         pid = current->pid;
1647         if (handle->h_sync && journal->j_last_sync_writer != pid &&
1648             journal->j_max_batch_time) {
1649                 u64 commit_time, trans_time;
1650
1651                 journal->j_last_sync_writer = pid;
1652
1653                 read_lock(&journal->j_state_lock);
1654                 commit_time = journal->j_average_commit_time;
1655                 read_unlock(&journal->j_state_lock);
1656
1657                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1658                                                    transaction->t_start_time));
1659
1660                 commit_time = max_t(u64, commit_time,
1661                                     1000*journal->j_min_batch_time);
1662                 commit_time = min_t(u64, commit_time,
1663                                     1000*journal->j_max_batch_time);
1664
1665                 if (trans_time < commit_time) {
1666                         ktime_t expires = ktime_add_ns(ktime_get(),
1667                                                        commit_time);
1668                         set_current_state(TASK_UNINTERRUPTIBLE);
1669                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1670                 }
1671         }
1672
1673         if (handle->h_sync)
1674                 transaction->t_synchronous_commit = 1;
1675         current->journal_info = NULL;
1676         atomic_sub(handle->h_buffer_credits,
1677                    &transaction->t_outstanding_credits);
1678
1679         /*
1680          * If the handle is marked SYNC, we need to set another commit
1681          * going!  We also want to force a commit if the current
1682          * transaction is occupying too much of the log, or if the
1683          * transaction is too old now.
1684          */
1685         if (handle->h_sync ||
1686             (atomic_read(&transaction->t_outstanding_credits) >
1687              journal->j_max_transaction_buffers) ||
1688             time_after_eq(jiffies, transaction->t_expires)) {
1689                 /* Do this even for aborted journals: an abort still
1690                  * completes the commit thread, it just doesn't write
1691                  * anything to disk. */
1692
1693                 jbd_debug(2, "transaction too old, requesting commit for "
1694                                         "handle %p\n", handle);
1695                 /* This is non-blocking */
1696                 jbd2_log_start_commit(journal, transaction->t_tid);
1697
1698                 /*
1699                  * Special case: JBD2_SYNC synchronous updates require us
1700                  * to wait for the commit to complete.
1701                  */
1702                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1703                         wait_for_commit = 1;
1704         }
1705
1706         /*
1707          * Once we drop t_updates, if it goes to zero the transaction
1708          * could start committing on us and eventually disappear.  So
1709          * once we do this, we must not dereference transaction
1710          * pointer again.
1711          */
1712         tid = transaction->t_tid;
1713         if (atomic_dec_and_test(&transaction->t_updates)) {
1714                 wake_up(&journal->j_wait_updates);
1715                 if (journal->j_barrier_count)
1716                         wake_up(&journal->j_wait_transaction_locked);
1717         }
1718
1719         if (wait_for_commit)
1720                 err = jbd2_log_wait_commit(journal, tid);
1721
1722         lock_map_release(&handle->h_lockdep_map);
1723
1724         if (handle->h_rsv_handle)
1725                 jbd2_journal_free_reserved(handle->h_rsv_handle);
1726 free_and_exit:
1727         jbd2_free_handle(handle);
1728         return err;
1729 }
1730
1731 /*
1732  *
1733  * List management code snippets: various functions for manipulating the
1734  * transaction buffer lists.
1735  *
1736  */
1737
1738 /*
1739  * Append a buffer to a transaction list, given the transaction's list head
1740  * pointer.
1741  *
1742  * j_list_lock is held.
1743  *
1744  * jbd_lock_bh_state(jh2bh(jh)) is held.
1745  */
1746
1747 static inline void
1748 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1749 {
1750         if (!*list) {
1751                 jh->b_tnext = jh->b_tprev = jh;
1752                 *list = jh;
1753         } else {
1754                 /* Insert at the tail of the list to preserve order */
1755                 struct journal_head *first = *list, *last = first->b_tprev;
1756                 jh->b_tprev = last;
1757                 jh->b_tnext = first;
1758                 last->b_tnext = first->b_tprev = jh;
1759         }
1760 }
1761
1762 /*
1763  * Remove a buffer from a transaction list, given the transaction's list
1764  * head pointer.
1765  *
1766  * Called with j_list_lock held, and the journal may not be locked.
1767  *
1768  * jbd_lock_bh_state(jh2bh(jh)) is held.
1769  */
1770
1771 static inline void
1772 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1773 {
1774         if (*list == jh) {
1775                 *list = jh->b_tnext;
1776                 if (*list == jh)
1777                         *list = NULL;
1778         }
1779         jh->b_tprev->b_tnext = jh->b_tnext;
1780         jh->b_tnext->b_tprev = jh->b_tprev;
1781 }
1782
1783 /*
1784  * Remove a buffer from the appropriate transaction list.
1785  *
1786  * Note that this function can *change* the value of
1787  * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1788  * t_reserved_list.  If the caller is holding onto a copy of one of these
1789  * pointers, it could go bad.  Generally the caller needs to re-read the
1790  * pointer from the transaction_t.
1791  *
1792  * Called under j_list_lock.
1793  */
1794 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1795 {
1796         struct journal_head **list = NULL;
1797         transaction_t *transaction;
1798         struct buffer_head *bh = jh2bh(jh);
1799
1800         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1801         transaction = jh->b_transaction;
1802         if (transaction)
1803                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1804
1805         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1806         if (jh->b_jlist != BJ_None)
1807                 J_ASSERT_JH(jh, transaction != NULL);
1808
1809         switch (jh->b_jlist) {
1810         case BJ_None:
1811                 return;
1812         case BJ_Metadata:
1813                 transaction->t_nr_buffers--;
1814                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1815                 list = &transaction->t_buffers;
1816                 break;
1817         case BJ_Forget:
1818                 list = &transaction->t_forget;
1819                 break;
1820         case BJ_Shadow:
1821                 list = &transaction->t_shadow_list;
1822                 break;
1823         case BJ_Reserved:
1824                 list = &transaction->t_reserved_list;
1825                 break;
1826         }
1827
1828         __blist_del_buffer(list, jh);
1829         jh->b_jlist = BJ_None;
1830         if (test_clear_buffer_jbddirty(bh))
1831                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1832 }
1833
1834 /*
1835  * Remove buffer from all transactions.
1836  *
1837  * Called with bh_state lock and j_list_lock
1838  *
1839  * jh and bh may be already freed when this function returns.
1840  */
1841 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1842 {
1843         __jbd2_journal_temp_unlink_buffer(jh);
1844         jh->b_transaction = NULL;
1845         jbd2_journal_put_journal_head(jh);
1846 }
1847
1848 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1849 {
1850         struct buffer_head *bh = jh2bh(jh);
1851
1852         /* Get reference so that buffer cannot be freed before we unlock it */
1853         get_bh(bh);
1854         jbd_lock_bh_state(bh);
1855         spin_lock(&journal->j_list_lock);
1856         __jbd2_journal_unfile_buffer(jh);
1857         spin_unlock(&journal->j_list_lock);
1858         jbd_unlock_bh_state(bh);
1859         __brelse(bh);
1860 }
1861
1862 /*
1863  * Called from jbd2_journal_try_to_free_buffers().
1864  *
1865  * Called under jbd_lock_bh_state(bh)
1866  */
1867 static void
1868 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1869 {
1870         struct journal_head *jh;
1871
1872         jh = bh2jh(bh);
1873
1874         if (buffer_locked(bh) || buffer_dirty(bh))
1875                 goto out;
1876
1877         if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
1878                 goto out;
1879
1880         spin_lock(&journal->j_list_lock);
1881         if (jh->b_cp_transaction != NULL) {
1882                 /* written-back checkpointed metadata buffer */
1883                 JBUFFER_TRACE(jh, "remove from checkpoint list");
1884                 __jbd2_journal_remove_checkpoint(jh);
1885         }
1886         spin_unlock(&journal->j_list_lock);
1887 out:
1888         return;
1889 }
1890
1891 /**
1892  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1893  * @journal: journal for operation
1894  * @page: to try and free
1895  * @gfp_mask: we use the mask to detect how hard should we try to release
1896  * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1897  * release the buffers.
1898  *
1899  *
1900  * For all the buffers on this page,
1901  * if they are fully written out ordered data, move them onto BUF_CLEAN
1902  * so try_to_free_buffers() can reap them.
1903  *
1904  * This function returns non-zero if we wish try_to_free_buffers()
1905  * to be called. We do this if the page is releasable by try_to_free_buffers().
1906  * We also do it if the page has locked or dirty buffers and the caller wants
1907  * us to perform sync or async writeout.
1908  *
1909  * This complicates JBD locking somewhat.  We aren't protected by the
1910  * BKL here.  We wish to remove the buffer from its committing or
1911  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1912  *
1913  * This may *change* the value of transaction_t->t_datalist, so anyone
1914  * who looks at t_datalist needs to lock against this function.
1915  *
1916  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1917  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1918  * will come out of the lock with the buffer dirty, which makes it
1919  * ineligible for release here.
1920  *
1921  * Who else is affected by this?  hmm...  Really the only contender
1922  * is do_get_write_access() - it could be looking at the buffer while
1923  * journal_try_to_free_buffer() is changing its state.  But that
1924  * cannot happen because we never reallocate freed data as metadata
1925  * while the data is part of a transaction.  Yes?
1926  *
1927  * Return 0 on failure, 1 on success
1928  */
1929 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1930                                 struct page *page, gfp_t gfp_mask)
1931 {
1932         struct buffer_head *head;
1933         struct buffer_head *bh;
1934         int ret = 0;
1935
1936         J_ASSERT(PageLocked(page));
1937
1938         head = page_buffers(page);
1939         bh = head;
1940         do {
1941                 struct journal_head *jh;
1942
1943                 /*
1944                  * We take our own ref against the journal_head here to avoid
1945                  * having to add tons of locking around each instance of
1946                  * jbd2_journal_put_journal_head().
1947                  */
1948                 jh = jbd2_journal_grab_journal_head(bh);
1949                 if (!jh)
1950                         continue;
1951
1952                 jbd_lock_bh_state(bh);
1953                 __journal_try_to_free_buffer(journal, bh);
1954                 jbd2_journal_put_journal_head(jh);
1955                 jbd_unlock_bh_state(bh);
1956                 if (buffer_jbd(bh))
1957                         goto busy;
1958         } while ((bh = bh->b_this_page) != head);
1959
1960         ret = try_to_free_buffers(page);
1961
1962 busy:
1963         return ret;
1964 }
1965
1966 /*
1967  * This buffer is no longer needed.  If it is on an older transaction's
1968  * checkpoint list we need to record it on this transaction's forget list
1969  * to pin this buffer (and hence its checkpointing transaction) down until
1970  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1971  * release it.
1972  * Returns non-zero if JBD no longer has an interest in the buffer.
1973  *
1974  * Called under j_list_lock.
1975  *
1976  * Called under jbd_lock_bh_state(bh).
1977  */
1978 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1979 {
1980         int may_free = 1;
1981         struct buffer_head *bh = jh2bh(jh);
1982
1983         if (jh->b_cp_transaction) {
1984                 JBUFFER_TRACE(jh, "on running+cp transaction");
1985                 __jbd2_journal_temp_unlink_buffer(jh);
1986                 /*
1987                  * We don't want to write the buffer anymore, clear the
1988                  * bit so that we don't confuse checks in
1989                  * __journal_file_buffer
1990                  */
1991                 clear_buffer_dirty(bh);
1992                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1993                 may_free = 0;
1994         } else {
1995                 JBUFFER_TRACE(jh, "on running transaction");
1996                 __jbd2_journal_unfile_buffer(jh);
1997         }
1998         return may_free;
1999 }
2000
2001 /*
2002  * jbd2_journal_invalidatepage
2003  *
2004  * This code is tricky.  It has a number of cases to deal with.
2005  *
2006  * There are two invariants which this code relies on:
2007  *
2008  * i_size must be updated on disk before we start calling invalidatepage on the
2009  * data.
2010  *
2011  *  This is done in ext3 by defining an ext3_setattr method which
2012  *  updates i_size before truncate gets going.  By maintaining this
2013  *  invariant, we can be sure that it is safe to throw away any buffers
2014  *  attached to the current transaction: once the transaction commits,
2015  *  we know that the data will not be needed.
2016  *
2017  *  Note however that we can *not* throw away data belonging to the
2018  *  previous, committing transaction!
2019  *
2020  * Any disk blocks which *are* part of the previous, committing
2021  * transaction (and which therefore cannot be discarded immediately) are
2022  * not going to be reused in the new running transaction
2023  *
2024  *  The bitmap committed_data images guarantee this: any block which is
2025  *  allocated in one transaction and removed in the next will be marked
2026  *  as in-use in the committed_data bitmap, so cannot be reused until
2027  *  the next transaction to delete the block commits.  This means that
2028  *  leaving committing buffers dirty is quite safe: the disk blocks
2029  *  cannot be reallocated to a different file and so buffer aliasing is
2030  *  not possible.
2031  *
2032  *
2033  * The above applies mainly to ordered data mode.  In writeback mode we
2034  * don't make guarantees about the order in which data hits disk --- in
2035  * particular we don't guarantee that new dirty data is flushed before
2036  * transaction commit --- so it is always safe just to discard data
2037  * immediately in that mode.  --sct
2038  */
2039
2040 /*
2041  * The journal_unmap_buffer helper function returns zero if the buffer
2042  * concerned remains pinned as an anonymous buffer belonging to an older
2043  * transaction.
2044  *
2045  * We're outside-transaction here.  Either or both of j_running_transaction
2046  * and j_committing_transaction may be NULL.
2047  */
2048 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2049                                 int partial_page)
2050 {
2051         transaction_t *transaction;
2052         struct journal_head *jh;
2053         int may_free = 1;
2054
2055         BUFFER_TRACE(bh, "entry");
2056
2057         /*
2058          * It is safe to proceed here without the j_list_lock because the
2059          * buffers cannot be stolen by try_to_free_buffers as long as we are
2060          * holding the page lock. --sct
2061          */
2062
2063         if (!buffer_jbd(bh))
2064                 goto zap_buffer_unlocked;
2065
2066         /* OK, we have data buffer in journaled mode */
2067         write_lock(&journal->j_state_lock);
2068         jbd_lock_bh_state(bh);
2069         spin_lock(&journal->j_list_lock);
2070
2071         jh = jbd2_journal_grab_journal_head(bh);
2072         if (!jh)
2073                 goto zap_buffer_no_jh;
2074
2075         /*
2076          * We cannot remove the buffer from checkpoint lists until the
2077          * transaction adding inode to orphan list (let's call it T)
2078          * is committed.  Otherwise if the transaction changing the
2079          * buffer would be cleaned from the journal before T is
2080          * committed, a crash will cause that the correct contents of
2081          * the buffer will be lost.  On the other hand we have to
2082          * clear the buffer dirty bit at latest at the moment when the
2083          * transaction marking the buffer as freed in the filesystem
2084          * structures is committed because from that moment on the
2085          * block can be reallocated and used by a different page.
2086          * Since the block hasn't been freed yet but the inode has
2087          * already been added to orphan list, it is safe for us to add
2088          * the buffer to BJ_Forget list of the newest transaction.
2089          *
2090          * Also we have to clear buffer_mapped flag of a truncated buffer
2091          * because the buffer_head may be attached to the page straddling
2092          * i_size (can happen only when blocksize < pagesize) and thus the
2093          * buffer_head can be reused when the file is extended again. So we end
2094          * up keeping around invalidated buffers attached to transactions'
2095          * BJ_Forget list just to stop checkpointing code from cleaning up
2096          * the transaction this buffer was modified in.
2097          */
2098         transaction = jh->b_transaction;
2099         if (transaction == NULL) {
2100                 /* First case: not on any transaction.  If it
2101                  * has no checkpoint link, then we can zap it:
2102                  * it's a writeback-mode buffer so we don't care
2103                  * if it hits disk safely. */
2104                 if (!jh->b_cp_transaction) {
2105                         JBUFFER_TRACE(jh, "not on any transaction: zap");
2106                         goto zap_buffer;
2107                 }
2108
2109                 if (!buffer_dirty(bh)) {
2110                         /* bdflush has written it.  We can drop it now */
2111                         goto zap_buffer;
2112                 }
2113
2114                 /* OK, it must be in the journal but still not
2115                  * written fully to disk: it's metadata or
2116                  * journaled data... */
2117
2118                 if (journal->j_running_transaction) {
2119                         /* ... and once the current transaction has
2120                          * committed, the buffer won't be needed any
2121                          * longer. */
2122                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2123                         may_free = __dispose_buffer(jh,
2124                                         journal->j_running_transaction);
2125                         goto zap_buffer;
2126                 } else {
2127                         /* There is no currently-running transaction. So the
2128                          * orphan record which we wrote for this file must have
2129                          * passed into commit.  We must attach this buffer to
2130                          * the committing transaction, if it exists. */
2131                         if (journal->j_committing_transaction) {
2132                                 JBUFFER_TRACE(jh, "give to committing trans");
2133                                 may_free = __dispose_buffer(jh,
2134                                         journal->j_committing_transaction);
2135                                 goto zap_buffer;
2136                         } else {
2137                                 /* The orphan record's transaction has
2138                                  * committed.  We can cleanse this buffer */
2139                                 clear_buffer_jbddirty(bh);
2140                                 goto zap_buffer;
2141                         }
2142                 }
2143         } else if (transaction == journal->j_committing_transaction) {
2144                 JBUFFER_TRACE(jh, "on committing transaction");
2145                 /*
2146                  * The buffer is committing, we simply cannot touch
2147                  * it. If the page is straddling i_size we have to wait
2148                  * for commit and try again.
2149                  */
2150                 if (partial_page) {
2151                         jbd2_journal_put_journal_head(jh);
2152                         spin_unlock(&journal->j_list_lock);
2153                         jbd_unlock_bh_state(bh);
2154                         write_unlock(&journal->j_state_lock);
2155                         return -EBUSY;
2156                 }
2157                 /*
2158                  * OK, buffer won't be reachable after truncate. We just set
2159                  * j_next_transaction to the running transaction (if there is
2160                  * one) and mark buffer as freed so that commit code knows it
2161                  * should clear dirty bits when it is done with the buffer.
2162                  */
2163                 set_buffer_freed(bh);
2164                 if (journal->j_running_transaction && buffer_jbddirty(bh))
2165                         jh->b_next_transaction = journal->j_running_transaction;
2166                 jbd2_journal_put_journal_head(jh);
2167                 spin_unlock(&journal->j_list_lock);
2168                 jbd_unlock_bh_state(bh);
2169                 write_unlock(&journal->j_state_lock);
2170                 return 0;
2171         } else {
2172                 /* Good, the buffer belongs to the running transaction.
2173                  * We are writing our own transaction's data, not any
2174                  * previous one's, so it is safe to throw it away
2175                  * (remember that we expect the filesystem to have set
2176                  * i_size already for this truncate so recovery will not
2177                  * expose the disk blocks we are discarding here.) */
2178                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2179                 JBUFFER_TRACE(jh, "on running transaction");
2180                 may_free = __dispose_buffer(jh, transaction);
2181         }
2182
2183 zap_buffer:
2184         /*
2185          * This is tricky. Although the buffer is truncated, it may be reused
2186          * if blocksize < pagesize and it is attached to the page straddling
2187          * EOF. Since the buffer might have been added to BJ_Forget list of the
2188          * running transaction, journal_get_write_access() won't clear
2189          * b_modified and credit accounting gets confused. So clear b_modified
2190          * here.
2191          */
2192         jh->b_modified = 0;
2193         jbd2_journal_put_journal_head(jh);
2194 zap_buffer_no_jh:
2195         spin_unlock(&journal->j_list_lock);
2196         jbd_unlock_bh_state(bh);
2197         write_unlock(&journal->j_state_lock);
2198 zap_buffer_unlocked:
2199         clear_buffer_dirty(bh);
2200         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2201         clear_buffer_mapped(bh);
2202         clear_buffer_req(bh);
2203         clear_buffer_new(bh);
2204         clear_buffer_delay(bh);
2205         clear_buffer_unwritten(bh);
2206         bh->b_bdev = NULL;
2207         return may_free;
2208 }
2209
2210 /**
2211  * void jbd2_journal_invalidatepage()
2212  * @journal: journal to use for flush...
2213  * @page:    page to flush
2214  * @offset:  start of the range to invalidate
2215  * @length:  length of the range to invalidate
2216  *
2217  * Reap page buffers containing data after in the specified range in page.
2218  * Can return -EBUSY if buffers are part of the committing transaction and
2219  * the page is straddling i_size. Caller then has to wait for current commit
2220  * and try again.
2221  */
2222 int jbd2_journal_invalidatepage(journal_t *journal,
2223                                 struct page *page,
2224                                 unsigned int offset,
2225                                 unsigned int length)
2226 {
2227         struct buffer_head *head, *bh, *next;
2228         unsigned int stop = offset + length;
2229         unsigned int curr_off = 0;
2230         int partial_page = (offset || length < PAGE_CACHE_SIZE);
2231         int may_free = 1;
2232         int ret = 0;
2233
2234         if (!PageLocked(page))
2235                 BUG();
2236         if (!page_has_buffers(page))
2237                 return 0;
2238
2239         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
2240
2241         /* We will potentially be playing with lists other than just the
2242          * data lists (especially for journaled data mode), so be
2243          * cautious in our locking. */
2244
2245         head = bh = page_buffers(page);
2246         do {
2247                 unsigned int next_off = curr_off + bh->b_size;
2248                 next = bh->b_this_page;
2249
2250                 if (next_off > stop)
2251                         return 0;
2252
2253                 if (offset <= curr_off) {
2254                         /* This block is wholly outside the truncation point */
2255                         lock_buffer(bh);
2256                         ret = journal_unmap_buffer(journal, bh, partial_page);
2257                         unlock_buffer(bh);
2258                         if (ret < 0)
2259                                 return ret;
2260                         may_free &= ret;
2261                 }
2262                 curr_off = next_off;
2263                 bh = next;
2264
2265         } while (bh != head);
2266
2267         if (!partial_page) {
2268                 if (may_free && try_to_free_buffers(page))
2269                         J_ASSERT(!page_has_buffers(page));
2270         }
2271         return 0;
2272 }
2273
2274 /*
2275  * File a buffer on the given transaction list.
2276  */
2277 void __jbd2_journal_file_buffer(struct journal_head *jh,
2278                         transaction_t *transaction, int jlist)
2279 {
2280         struct journal_head **list = NULL;
2281         int was_dirty = 0;
2282         struct buffer_head *bh = jh2bh(jh);
2283
2284         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2285         assert_spin_locked(&transaction->t_journal->j_list_lock);
2286
2287         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2288         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2289                                 jh->b_transaction == NULL);
2290
2291         if (jh->b_transaction && jh->b_jlist == jlist)
2292                 return;
2293
2294         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2295             jlist == BJ_Shadow || jlist == BJ_Forget) {
2296                 /*
2297                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2298                  * instead of buffer_dirty. We should not see a dirty bit set
2299                  * here because we clear it in do_get_write_access but e.g.
2300                  * tune2fs can modify the sb and set the dirty bit at any time
2301                  * so we try to gracefully handle that.
2302                  */
2303                 if (buffer_dirty(bh))
2304                         warn_dirty_buffer(bh);
2305                 if (test_clear_buffer_dirty(bh) ||
2306                     test_clear_buffer_jbddirty(bh))
2307                         was_dirty = 1;
2308         }
2309
2310         if (jh->b_transaction)
2311                 __jbd2_journal_temp_unlink_buffer(jh);
2312         else
2313                 jbd2_journal_grab_journal_head(bh);
2314         jh->b_transaction = transaction;
2315
2316         switch (jlist) {
2317         case BJ_None:
2318                 J_ASSERT_JH(jh, !jh->b_committed_data);
2319                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2320                 return;
2321         case BJ_Metadata:
2322                 transaction->t_nr_buffers++;
2323                 list = &transaction->t_buffers;
2324                 break;
2325         case BJ_Forget:
2326                 list = &transaction->t_forget;
2327                 break;
2328         case BJ_Shadow:
2329                 list = &transaction->t_shadow_list;
2330                 break;
2331         case BJ_Reserved:
2332                 list = &transaction->t_reserved_list;
2333                 break;
2334         }
2335
2336         __blist_add_buffer(list, jh);
2337         jh->b_jlist = jlist;
2338
2339         if (was_dirty)
2340                 set_buffer_jbddirty(bh);
2341 }
2342
2343 void jbd2_journal_file_buffer(struct journal_head *jh,
2344                                 transaction_t *transaction, int jlist)
2345 {
2346         jbd_lock_bh_state(jh2bh(jh));
2347         spin_lock(&transaction->t_journal->j_list_lock);
2348         __jbd2_journal_file_buffer(jh, transaction, jlist);
2349         spin_unlock(&transaction->t_journal->j_list_lock);
2350         jbd_unlock_bh_state(jh2bh(jh));
2351 }
2352
2353 /*
2354  * Remove a buffer from its current buffer list in preparation for
2355  * dropping it from its current transaction entirely.  If the buffer has
2356  * already started to be used by a subsequent transaction, refile the
2357  * buffer on that transaction's metadata list.
2358  *
2359  * Called under j_list_lock
2360  * Called under jbd_lock_bh_state(jh2bh(jh))
2361  *
2362  * jh and bh may be already free when this function returns
2363  */
2364 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2365 {
2366         int was_dirty, jlist;
2367         struct buffer_head *bh = jh2bh(jh);
2368
2369         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2370         if (jh->b_transaction)
2371                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2372
2373         /* If the buffer is now unused, just drop it. */
2374         if (jh->b_next_transaction == NULL) {
2375                 __jbd2_journal_unfile_buffer(jh);
2376                 return;
2377         }
2378
2379         /*
2380          * It has been modified by a later transaction: add it to the new
2381          * transaction's metadata list.
2382          */
2383
2384         was_dirty = test_clear_buffer_jbddirty(bh);
2385         __jbd2_journal_temp_unlink_buffer(jh);
2386         /*
2387          * We set b_transaction here because b_next_transaction will inherit
2388          * our jh reference and thus __jbd2_journal_file_buffer() must not
2389          * take a new one.
2390          */
2391         jh->b_transaction = jh->b_next_transaction;
2392         jh->b_next_transaction = NULL;
2393         if (buffer_freed(bh))
2394                 jlist = BJ_Forget;
2395         else if (jh->b_modified)
2396                 jlist = BJ_Metadata;
2397         else
2398                 jlist = BJ_Reserved;
2399         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2400         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2401
2402         if (was_dirty)
2403                 set_buffer_jbddirty(bh);
2404 }
2405
2406 /*
2407  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2408  * bh reference so that we can safely unlock bh.
2409  *
2410  * The jh and bh may be freed by this call.
2411  */
2412 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2413 {
2414         struct buffer_head *bh = jh2bh(jh);
2415
2416         /* Get reference so that buffer cannot be freed before we unlock it */
2417         get_bh(bh);
2418         jbd_lock_bh_state(bh);
2419         spin_lock(&journal->j_list_lock);
2420         __jbd2_journal_refile_buffer(jh);
2421         jbd_unlock_bh_state(bh);
2422         spin_unlock(&journal->j_list_lock);
2423         __brelse(bh);
2424 }
2425
2426 /*
2427  * File inode in the inode list of the handle's transaction
2428  */
2429 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2430 {
2431         transaction_t *transaction = handle->h_transaction;
2432         journal_t *journal;
2433
2434         if (is_handle_aborted(handle))
2435                 return -EROFS;
2436         journal = transaction->t_journal;
2437
2438         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2439                         transaction->t_tid);
2440
2441         /*
2442          * First check whether inode isn't already on the transaction's
2443          * lists without taking the lock. Note that this check is safe
2444          * without the lock as we cannot race with somebody removing inode
2445          * from the transaction. The reason is that we remove inode from the
2446          * transaction only in journal_release_jbd_inode() and when we commit
2447          * the transaction. We are guarded from the first case by holding
2448          * a reference to the inode. We are safe against the second case
2449          * because if jinode->i_transaction == transaction, commit code
2450          * cannot touch the transaction because we hold reference to it,
2451          * and if jinode->i_next_transaction == transaction, commit code
2452          * will only file the inode where we want it.
2453          */
2454         if (jinode->i_transaction == transaction ||
2455             jinode->i_next_transaction == transaction)
2456                 return 0;
2457
2458         spin_lock(&journal->j_list_lock);
2459
2460         if (jinode->i_transaction == transaction ||
2461             jinode->i_next_transaction == transaction)
2462                 goto done;
2463
2464         /*
2465          * We only ever set this variable to 1 so the test is safe. Since
2466          * t_need_data_flush is likely to be set, we do the test to save some
2467          * cacheline bouncing
2468          */
2469         if (!transaction->t_need_data_flush)
2470                 transaction->t_need_data_flush = 1;
2471         /* On some different transaction's list - should be
2472          * the committing one */
2473         if (jinode->i_transaction) {
2474                 J_ASSERT(jinode->i_next_transaction == NULL);
2475                 J_ASSERT(jinode->i_transaction ==
2476                                         journal->j_committing_transaction);
2477                 jinode->i_next_transaction = transaction;
2478                 goto done;
2479         }
2480         /* Not on any transaction list... */
2481         J_ASSERT(!jinode->i_next_transaction);
2482         jinode->i_transaction = transaction;
2483         list_add(&jinode->i_list, &transaction->t_inode_list);
2484 done:
2485         spin_unlock(&journal->j_list_lock);
2486
2487         return 0;
2488 }
2489
2490 /*
2491  * File truncate and transaction commit interact with each other in a
2492  * non-trivial way.  If a transaction writing data block A is
2493  * committing, we cannot discard the data by truncate until we have
2494  * written them.  Otherwise if we crashed after the transaction with
2495  * write has committed but before the transaction with truncate has
2496  * committed, we could see stale data in block A.  This function is a
2497  * helper to solve this problem.  It starts writeout of the truncated
2498  * part in case it is in the committing transaction.
2499  *
2500  * Filesystem code must call this function when inode is journaled in
2501  * ordered mode before truncation happens and after the inode has been
2502  * placed on orphan list with the new inode size. The second condition
2503  * avoids the race that someone writes new data and we start
2504  * committing the transaction after this function has been called but
2505  * before a transaction for truncate is started (and furthermore it
2506  * allows us to optimize the case where the addition to orphan list
2507  * happens in the same transaction as write --- we don't have to write
2508  * any data in such case).
2509  */
2510 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2511                                         struct jbd2_inode *jinode,
2512                                         loff_t new_size)
2513 {
2514         transaction_t *inode_trans, *commit_trans;
2515         int ret = 0;
2516
2517         /* This is a quick check to avoid locking if not necessary */
2518         if (!jinode->i_transaction)
2519                 goto out;
2520         /* Locks are here just to force reading of recent values, it is
2521          * enough that the transaction was not committing before we started
2522          * a transaction adding the inode to orphan list */
2523         read_lock(&journal->j_state_lock);
2524         commit_trans = journal->j_committing_transaction;
2525         read_unlock(&journal->j_state_lock);
2526         spin_lock(&journal->j_list_lock);
2527         inode_trans = jinode->i_transaction;
2528         spin_unlock(&journal->j_list_lock);
2529         if (inode_trans == commit_trans) {
2530                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2531                         new_size, LLONG_MAX);
2532                 if (ret)
2533                         jbd2_journal_abort(journal, ret);
2534         }
2535 out:
2536         return ret;
2537 }