locking/lockdep: Remove unused @nested argument from lock_release()
[sfrench/cifs-2.6.git] / fs / jbd2 / transaction.c
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * linux/fs/jbd2/transaction.c
4  *
5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6  *
7  * Copyright 1998 Red Hat corp --- All Rights Reserved
8  *
9  * Generic filesystem transaction handling code; part of the ext2fs
10  * journaling system.
11  *
12  * This file manages transactions (compound commits managed by the
13  * journaling code) and handles (individual atomic operations by the
14  * filesystem).
15  */
16
17 #include <linux/time.h>
18 #include <linux/fs.h>
19 #include <linux/jbd2.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/timer.h>
23 #include <linux/mm.h>
24 #include <linux/highmem.h>
25 #include <linux/hrtimer.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bug.h>
28 #include <linux/module.h>
29 #include <linux/sched/mm.h>
30
31 #include <trace/events/jbd2.h>
32
33 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
35
36 static struct kmem_cache *transaction_cache;
37 int __init jbd2_journal_init_transaction_cache(void)
38 {
39         J_ASSERT(!transaction_cache);
40         transaction_cache = kmem_cache_create("jbd2_transaction_s",
41                                         sizeof(transaction_t),
42                                         0,
43                                         SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
44                                         NULL);
45         if (!transaction_cache) {
46                 pr_emerg("JBD2: failed to create transaction cache\n");
47                 return -ENOMEM;
48         }
49         return 0;
50 }
51
52 void jbd2_journal_destroy_transaction_cache(void)
53 {
54         kmem_cache_destroy(transaction_cache);
55         transaction_cache = NULL;
56 }
57
58 void jbd2_journal_free_transaction(transaction_t *transaction)
59 {
60         if (unlikely(ZERO_OR_NULL_PTR(transaction)))
61                 return;
62         kmem_cache_free(transaction_cache, transaction);
63 }
64
65 /*
66  * jbd2_get_transaction: obtain a new transaction_t object.
67  *
68  * Simply initialise a new transaction. Initialize it in
69  * RUNNING state and add it to the current journal (which should not
70  * have an existing running transaction: we only make a new transaction
71  * once we have started to commit the old one).
72  *
73  * Preconditions:
74  *      The journal MUST be locked.  We don't perform atomic mallocs on the
75  *      new transaction and we can't block without protecting against other
76  *      processes trying to touch the journal while it is in transition.
77  *
78  */
79
80 static void jbd2_get_transaction(journal_t *journal,
81                                 transaction_t *transaction)
82 {
83         transaction->t_journal = journal;
84         transaction->t_state = T_RUNNING;
85         transaction->t_start_time = ktime_get();
86         transaction->t_tid = journal->j_transaction_sequence++;
87         transaction->t_expires = jiffies + journal->j_commit_interval;
88         spin_lock_init(&transaction->t_handle_lock);
89         atomic_set(&transaction->t_updates, 0);
90         atomic_set(&transaction->t_outstanding_credits,
91                    atomic_read(&journal->j_reserved_credits));
92         atomic_set(&transaction->t_handle_count, 0);
93         INIT_LIST_HEAD(&transaction->t_inode_list);
94         INIT_LIST_HEAD(&transaction->t_private_list);
95
96         /* Set up the commit timer for the new transaction. */
97         journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
98         add_timer(&journal->j_commit_timer);
99
100         J_ASSERT(journal->j_running_transaction == NULL);
101         journal->j_running_transaction = transaction;
102         transaction->t_max_wait = 0;
103         transaction->t_start = jiffies;
104         transaction->t_requested = 0;
105 }
106
107 /*
108  * Handle management.
109  *
110  * A handle_t is an object which represents a single atomic update to a
111  * filesystem, and which tracks all of the modifications which form part
112  * of that one update.
113  */
114
115 /*
116  * Update transaction's maximum wait time, if debugging is enabled.
117  *
118  * In order for t_max_wait to be reliable, it must be protected by a
119  * lock.  But doing so will mean that start_this_handle() can not be
120  * run in parallel on SMP systems, which limits our scalability.  So
121  * unless debugging is enabled, we no longer update t_max_wait, which
122  * means that maximum wait time reported by the jbd2_run_stats
123  * tracepoint will always be zero.
124  */
125 static inline void update_t_max_wait(transaction_t *transaction,
126                                      unsigned long ts)
127 {
128 #ifdef CONFIG_JBD2_DEBUG
129         if (jbd2_journal_enable_debug &&
130             time_after(transaction->t_start, ts)) {
131                 ts = jbd2_time_diff(ts, transaction->t_start);
132                 spin_lock(&transaction->t_handle_lock);
133                 if (ts > transaction->t_max_wait)
134                         transaction->t_max_wait = ts;
135                 spin_unlock(&transaction->t_handle_lock);
136         }
137 #endif
138 }
139
140 /*
141  * Wait until running transaction passes to T_FLUSH state and new transaction
142  * can thus be started. Also starts the commit if needed. The function expects
143  * running transaction to exist and releases j_state_lock.
144  */
145 static void wait_transaction_locked(journal_t *journal)
146         __releases(journal->j_state_lock)
147 {
148         DEFINE_WAIT(wait);
149         int need_to_start;
150         tid_t tid = journal->j_running_transaction->t_tid;
151
152         prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
153                         TASK_UNINTERRUPTIBLE);
154         need_to_start = !tid_geq(journal->j_commit_request, tid);
155         read_unlock(&journal->j_state_lock);
156         if (need_to_start)
157                 jbd2_log_start_commit(journal, tid);
158         jbd2_might_wait_for_commit(journal);
159         schedule();
160         finish_wait(&journal->j_wait_transaction_locked, &wait);
161 }
162
163 /*
164  * Wait until running transaction transitions from T_SWITCH to T_FLUSH
165  * state and new transaction can thus be started. The function releases
166  * j_state_lock.
167  */
168 static void wait_transaction_switching(journal_t *journal)
169         __releases(journal->j_state_lock)
170 {
171         DEFINE_WAIT(wait);
172
173         if (WARN_ON(!journal->j_running_transaction ||
174                     journal->j_running_transaction->t_state != T_SWITCH))
175                 return;
176         prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177                         TASK_UNINTERRUPTIBLE);
178         read_unlock(&journal->j_state_lock);
179         /*
180          * We don't call jbd2_might_wait_for_commit() here as there's no
181          * waiting for outstanding handles happening anymore in T_SWITCH state
182          * and handling of reserved handles actually relies on that for
183          * correctness.
184          */
185         schedule();
186         finish_wait(&journal->j_wait_transaction_locked, &wait);
187 }
188
189 static void sub_reserved_credits(journal_t *journal, int blocks)
190 {
191         atomic_sub(blocks, &journal->j_reserved_credits);
192         wake_up(&journal->j_wait_reserved);
193 }
194
195 /*
196  * Wait until we can add credits for handle to the running transaction.  Called
197  * with j_state_lock held for reading. Returns 0 if handle joined the running
198  * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
199  * caller must retry.
200  */
201 static int add_transaction_credits(journal_t *journal, int blocks,
202                                    int rsv_blocks)
203 {
204         transaction_t *t = journal->j_running_transaction;
205         int needed;
206         int total = blocks + rsv_blocks;
207
208         /*
209          * If the current transaction is locked down for commit, wait
210          * for the lock to be released.
211          */
212         if (t->t_state != T_RUNNING) {
213                 WARN_ON_ONCE(t->t_state >= T_FLUSH);
214                 wait_transaction_locked(journal);
215                 return 1;
216         }
217
218         /*
219          * If there is not enough space left in the log to write all
220          * potential buffers requested by this operation, we need to
221          * stall pending a log checkpoint to free some more log space.
222          */
223         needed = atomic_add_return(total, &t->t_outstanding_credits);
224         if (needed > journal->j_max_transaction_buffers) {
225                 /*
226                  * If the current transaction is already too large,
227                  * then start to commit it: we can then go back and
228                  * attach this handle to a new transaction.
229                  */
230                 atomic_sub(total, &t->t_outstanding_credits);
231
232                 /*
233                  * Is the number of reserved credits in the current transaction too
234                  * big to fit this handle? Wait until reserved credits are freed.
235                  */
236                 if (atomic_read(&journal->j_reserved_credits) + total >
237                     journal->j_max_transaction_buffers) {
238                         read_unlock(&journal->j_state_lock);
239                         jbd2_might_wait_for_commit(journal);
240                         wait_event(journal->j_wait_reserved,
241                                    atomic_read(&journal->j_reserved_credits) + total <=
242                                    journal->j_max_transaction_buffers);
243                         return 1;
244                 }
245
246                 wait_transaction_locked(journal);
247                 return 1;
248         }
249
250         /*
251          * The commit code assumes that it can get enough log space
252          * without forcing a checkpoint.  This is *critical* for
253          * correctness: a checkpoint of a buffer which is also
254          * associated with a committing transaction creates a deadlock,
255          * so commit simply cannot force through checkpoints.
256          *
257          * We must therefore ensure the necessary space in the journal
258          * *before* starting to dirty potentially checkpointed buffers
259          * in the new transaction.
260          */
261         if (jbd2_log_space_left(journal) < jbd2_space_needed(journal)) {
262                 atomic_sub(total, &t->t_outstanding_credits);
263                 read_unlock(&journal->j_state_lock);
264                 jbd2_might_wait_for_commit(journal);
265                 write_lock(&journal->j_state_lock);
266                 if (jbd2_log_space_left(journal) < jbd2_space_needed(journal))
267                         __jbd2_log_wait_for_space(journal);
268                 write_unlock(&journal->j_state_lock);
269                 return 1;
270         }
271
272         /* No reservation? We are done... */
273         if (!rsv_blocks)
274                 return 0;
275
276         needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
277         /* We allow at most half of a transaction to be reserved */
278         if (needed > journal->j_max_transaction_buffers / 2) {
279                 sub_reserved_credits(journal, rsv_blocks);
280                 atomic_sub(total, &t->t_outstanding_credits);
281                 read_unlock(&journal->j_state_lock);
282                 jbd2_might_wait_for_commit(journal);
283                 wait_event(journal->j_wait_reserved,
284                          atomic_read(&journal->j_reserved_credits) + rsv_blocks
285                          <= journal->j_max_transaction_buffers / 2);
286                 return 1;
287         }
288         return 0;
289 }
290
291 /*
292  * start_this_handle: Given a handle, deal with any locking or stalling
293  * needed to make sure that there is enough journal space for the handle
294  * to begin.  Attach the handle to a transaction and set up the
295  * transaction's buffer credits.
296  */
297
298 static int start_this_handle(journal_t *journal, handle_t *handle,
299                              gfp_t gfp_mask)
300 {
301         transaction_t   *transaction, *new_transaction = NULL;
302         int             blocks = handle->h_buffer_credits;
303         int             rsv_blocks = 0;
304         unsigned long ts = jiffies;
305
306         if (handle->h_rsv_handle)
307                 rsv_blocks = handle->h_rsv_handle->h_buffer_credits;
308
309         /*
310          * Limit the number of reserved credits to 1/2 of maximum transaction
311          * size and limit the number of total credits to not exceed maximum
312          * transaction size per operation.
313          */
314         if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
315             (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
316                 printk(KERN_ERR "JBD2: %s wants too many credits "
317                        "credits:%d rsv_credits:%d max:%d\n",
318                        current->comm, blocks, rsv_blocks,
319                        journal->j_max_transaction_buffers);
320                 WARN_ON(1);
321                 return -ENOSPC;
322         }
323
324 alloc_transaction:
325         if (!journal->j_running_transaction) {
326                 /*
327                  * If __GFP_FS is not present, then we may be being called from
328                  * inside the fs writeback layer, so we MUST NOT fail.
329                  */
330                 if ((gfp_mask & __GFP_FS) == 0)
331                         gfp_mask |= __GFP_NOFAIL;
332                 new_transaction = kmem_cache_zalloc(transaction_cache,
333                                                     gfp_mask);
334                 if (!new_transaction)
335                         return -ENOMEM;
336         }
337
338         jbd_debug(3, "New handle %p going live.\n", handle);
339
340         /*
341          * We need to hold j_state_lock until t_updates has been incremented,
342          * for proper journal barrier handling
343          */
344 repeat:
345         read_lock(&journal->j_state_lock);
346         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
347         if (is_journal_aborted(journal) ||
348             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
349                 read_unlock(&journal->j_state_lock);
350                 jbd2_journal_free_transaction(new_transaction);
351                 return -EROFS;
352         }
353
354         /*
355          * Wait on the journal's transaction barrier if necessary. Specifically
356          * we allow reserved handles to proceed because otherwise commit could
357          * deadlock on page writeback not being able to complete.
358          */
359         if (!handle->h_reserved && journal->j_barrier_count) {
360                 read_unlock(&journal->j_state_lock);
361                 wait_event(journal->j_wait_transaction_locked,
362                                 journal->j_barrier_count == 0);
363                 goto repeat;
364         }
365
366         if (!journal->j_running_transaction) {
367                 read_unlock(&journal->j_state_lock);
368                 if (!new_transaction)
369                         goto alloc_transaction;
370                 write_lock(&journal->j_state_lock);
371                 if (!journal->j_running_transaction &&
372                     (handle->h_reserved || !journal->j_barrier_count)) {
373                         jbd2_get_transaction(journal, new_transaction);
374                         new_transaction = NULL;
375                 }
376                 write_unlock(&journal->j_state_lock);
377                 goto repeat;
378         }
379
380         transaction = journal->j_running_transaction;
381
382         if (!handle->h_reserved) {
383                 /* We may have dropped j_state_lock - restart in that case */
384                 if (add_transaction_credits(journal, blocks, rsv_blocks))
385                         goto repeat;
386         } else {
387                 /*
388                  * We have handle reserved so we are allowed to join T_LOCKED
389                  * transaction and we don't have to check for transaction size
390                  * and journal space. But we still have to wait while running
391                  * transaction is being switched to a committing one as it
392                  * won't wait for any handles anymore.
393                  */
394                 if (transaction->t_state == T_SWITCH) {
395                         wait_transaction_switching(journal);
396                         goto repeat;
397                 }
398                 sub_reserved_credits(journal, blocks);
399                 handle->h_reserved = 0;
400         }
401
402         /* OK, account for the buffers that this operation expects to
403          * use and add the handle to the running transaction. 
404          */
405         update_t_max_wait(transaction, ts);
406         handle->h_transaction = transaction;
407         handle->h_requested_credits = blocks;
408         handle->h_start_jiffies = jiffies;
409         atomic_inc(&transaction->t_updates);
410         atomic_inc(&transaction->t_handle_count);
411         jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
412                   handle, blocks,
413                   atomic_read(&transaction->t_outstanding_credits),
414                   jbd2_log_space_left(journal));
415         read_unlock(&journal->j_state_lock);
416         current->journal_info = handle;
417
418         rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
419         jbd2_journal_free_transaction(new_transaction);
420         /*
421          * Ensure that no allocations done while the transaction is open are
422          * going to recurse back to the fs layer.
423          */
424         handle->saved_alloc_context = memalloc_nofs_save();
425         return 0;
426 }
427
428 /* Allocate a new handle.  This should probably be in a slab... */
429 static handle_t *new_handle(int nblocks)
430 {
431         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
432         if (!handle)
433                 return NULL;
434         handle->h_buffer_credits = nblocks;
435         handle->h_ref = 1;
436
437         return handle;
438 }
439
440 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
441                               gfp_t gfp_mask, unsigned int type,
442                               unsigned int line_no)
443 {
444         handle_t *handle = journal_current_handle();
445         int err;
446
447         if (!journal)
448                 return ERR_PTR(-EROFS);
449
450         if (handle) {
451                 J_ASSERT(handle->h_transaction->t_journal == journal);
452                 handle->h_ref++;
453                 return handle;
454         }
455
456         handle = new_handle(nblocks);
457         if (!handle)
458                 return ERR_PTR(-ENOMEM);
459         if (rsv_blocks) {
460                 handle_t *rsv_handle;
461
462                 rsv_handle = new_handle(rsv_blocks);
463                 if (!rsv_handle) {
464                         jbd2_free_handle(handle);
465                         return ERR_PTR(-ENOMEM);
466                 }
467                 rsv_handle->h_reserved = 1;
468                 rsv_handle->h_journal = journal;
469                 handle->h_rsv_handle = rsv_handle;
470         }
471
472         err = start_this_handle(journal, handle, gfp_mask);
473         if (err < 0) {
474                 if (handle->h_rsv_handle)
475                         jbd2_free_handle(handle->h_rsv_handle);
476                 jbd2_free_handle(handle);
477                 return ERR_PTR(err);
478         }
479         handle->h_type = type;
480         handle->h_line_no = line_no;
481         trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
482                                 handle->h_transaction->t_tid, type,
483                                 line_no, nblocks);
484
485         return handle;
486 }
487 EXPORT_SYMBOL(jbd2__journal_start);
488
489
490 /**
491  * handle_t *jbd2_journal_start() - Obtain a new handle.
492  * @journal: Journal to start transaction on.
493  * @nblocks: number of block buffer we might modify
494  *
495  * We make sure that the transaction can guarantee at least nblocks of
496  * modified buffers in the log.  We block until the log can guarantee
497  * that much space. Additionally, if rsv_blocks > 0, we also create another
498  * handle with rsv_blocks reserved blocks in the journal. This handle is
499  * is stored in h_rsv_handle. It is not attached to any particular transaction
500  * and thus doesn't block transaction commit. If the caller uses this reserved
501  * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
502  * on the parent handle will dispose the reserved one. Reserved handle has to
503  * be converted to a normal handle using jbd2_journal_start_reserved() before
504  * it can be used.
505  *
506  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
507  * on failure.
508  */
509 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
510 {
511         return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0);
512 }
513 EXPORT_SYMBOL(jbd2_journal_start);
514
515 void jbd2_journal_free_reserved(handle_t *handle)
516 {
517         journal_t *journal = handle->h_journal;
518
519         WARN_ON(!handle->h_reserved);
520         sub_reserved_credits(journal, handle->h_buffer_credits);
521         jbd2_free_handle(handle);
522 }
523 EXPORT_SYMBOL(jbd2_journal_free_reserved);
524
525 /**
526  * int jbd2_journal_start_reserved() - start reserved handle
527  * @handle: handle to start
528  * @type: for handle statistics
529  * @line_no: for handle statistics
530  *
531  * Start handle that has been previously reserved with jbd2_journal_reserve().
532  * This attaches @handle to the running transaction (or creates one if there's
533  * not transaction running). Unlike jbd2_journal_start() this function cannot
534  * block on journal commit, checkpointing, or similar stuff. It can block on
535  * memory allocation or frozen journal though.
536  *
537  * Return 0 on success, non-zero on error - handle is freed in that case.
538  */
539 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
540                                 unsigned int line_no)
541 {
542         journal_t *journal = handle->h_journal;
543         int ret = -EIO;
544
545         if (WARN_ON(!handle->h_reserved)) {
546                 /* Someone passed in normal handle? Just stop it. */
547                 jbd2_journal_stop(handle);
548                 return ret;
549         }
550         /*
551          * Usefulness of mixing of reserved and unreserved handles is
552          * questionable. So far nobody seems to need it so just error out.
553          */
554         if (WARN_ON(current->journal_info)) {
555                 jbd2_journal_free_reserved(handle);
556                 return ret;
557         }
558
559         handle->h_journal = NULL;
560         /*
561          * GFP_NOFS is here because callers are likely from writeback or
562          * similarly constrained call sites
563          */
564         ret = start_this_handle(journal, handle, GFP_NOFS);
565         if (ret < 0) {
566                 handle->h_journal = journal;
567                 jbd2_journal_free_reserved(handle);
568                 return ret;
569         }
570         handle->h_type = type;
571         handle->h_line_no = line_no;
572         trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
573                                 handle->h_transaction->t_tid, type,
574                                 line_no, handle->h_buffer_credits);
575         return 0;
576 }
577 EXPORT_SYMBOL(jbd2_journal_start_reserved);
578
579 /**
580  * int jbd2_journal_extend() - extend buffer credits.
581  * @handle:  handle to 'extend'
582  * @nblocks: nr blocks to try to extend by.
583  *
584  * Some transactions, such as large extends and truncates, can be done
585  * atomically all at once or in several stages.  The operation requests
586  * a credit for a number of buffer modifications in advance, but can
587  * extend its credit if it needs more.
588  *
589  * jbd2_journal_extend tries to give the running handle more buffer credits.
590  * It does not guarantee that allocation - this is a best-effort only.
591  * The calling process MUST be able to deal cleanly with a failure to
592  * extend here.
593  *
594  * Return 0 on success, non-zero on failure.
595  *
596  * return code < 0 implies an error
597  * return code > 0 implies normal transaction-full status.
598  */
599 int jbd2_journal_extend(handle_t *handle, int nblocks)
600 {
601         transaction_t *transaction = handle->h_transaction;
602         journal_t *journal;
603         int result;
604         int wanted;
605
606         if (is_handle_aborted(handle))
607                 return -EROFS;
608         journal = transaction->t_journal;
609
610         result = 1;
611
612         read_lock(&journal->j_state_lock);
613
614         /* Don't extend a locked-down transaction! */
615         if (transaction->t_state != T_RUNNING) {
616                 jbd_debug(3, "denied handle %p %d blocks: "
617                           "transaction not running\n", handle, nblocks);
618                 goto error_out;
619         }
620
621         spin_lock(&transaction->t_handle_lock);
622         wanted = atomic_add_return(nblocks,
623                                    &transaction->t_outstanding_credits);
624
625         if (wanted > journal->j_max_transaction_buffers) {
626                 jbd_debug(3, "denied handle %p %d blocks: "
627                           "transaction too large\n", handle, nblocks);
628                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
629                 goto unlock;
630         }
631
632         if (wanted + (wanted >> JBD2_CONTROL_BLOCKS_SHIFT) >
633             jbd2_log_space_left(journal)) {
634                 jbd_debug(3, "denied handle %p %d blocks: "
635                           "insufficient log space\n", handle, nblocks);
636                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
637                 goto unlock;
638         }
639
640         trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
641                                  transaction->t_tid,
642                                  handle->h_type, handle->h_line_no,
643                                  handle->h_buffer_credits,
644                                  nblocks);
645
646         handle->h_buffer_credits += nblocks;
647         handle->h_requested_credits += nblocks;
648         result = 0;
649
650         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
651 unlock:
652         spin_unlock(&transaction->t_handle_lock);
653 error_out:
654         read_unlock(&journal->j_state_lock);
655         return result;
656 }
657
658
659 /**
660  * int jbd2_journal_restart() - restart a handle .
661  * @handle:  handle to restart
662  * @nblocks: nr credits requested
663  * @gfp_mask: memory allocation flags (for start_this_handle)
664  *
665  * Restart a handle for a multi-transaction filesystem
666  * operation.
667  *
668  * If the jbd2_journal_extend() call above fails to grant new buffer credits
669  * to a running handle, a call to jbd2_journal_restart will commit the
670  * handle's transaction so far and reattach the handle to a new
671  * transaction capable of guaranteeing the requested number of
672  * credits. We preserve reserved handle if there's any attached to the
673  * passed in handle.
674  */
675 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
676 {
677         transaction_t *transaction = handle->h_transaction;
678         journal_t *journal;
679         tid_t           tid;
680         int             need_to_start, ret;
681
682         /* If we've had an abort of any type, don't even think about
683          * actually doing the restart! */
684         if (is_handle_aborted(handle))
685                 return 0;
686         journal = transaction->t_journal;
687
688         /*
689          * First unlink the handle from its current transaction, and start the
690          * commit on that.
691          */
692         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
693         J_ASSERT(journal_current_handle() == handle);
694
695         read_lock(&journal->j_state_lock);
696         spin_lock(&transaction->t_handle_lock);
697         atomic_sub(handle->h_buffer_credits,
698                    &transaction->t_outstanding_credits);
699         if (handle->h_rsv_handle) {
700                 sub_reserved_credits(journal,
701                                      handle->h_rsv_handle->h_buffer_credits);
702         }
703         if (atomic_dec_and_test(&transaction->t_updates))
704                 wake_up(&journal->j_wait_updates);
705         tid = transaction->t_tid;
706         spin_unlock(&transaction->t_handle_lock);
707         handle->h_transaction = NULL;
708         current->journal_info = NULL;
709
710         jbd_debug(2, "restarting handle %p\n", handle);
711         need_to_start = !tid_geq(journal->j_commit_request, tid);
712         read_unlock(&journal->j_state_lock);
713         if (need_to_start)
714                 jbd2_log_start_commit(journal, tid);
715
716         rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
717         handle->h_buffer_credits = nblocks;
718         /*
719          * Restore the original nofs context because the journal restart
720          * is basically the same thing as journal stop and start.
721          * start_this_handle will start a new nofs context.
722          */
723         memalloc_nofs_restore(handle->saved_alloc_context);
724         ret = start_this_handle(journal, handle, gfp_mask);
725         return ret;
726 }
727 EXPORT_SYMBOL(jbd2__journal_restart);
728
729
730 int jbd2_journal_restart(handle_t *handle, int nblocks)
731 {
732         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
733 }
734 EXPORT_SYMBOL(jbd2_journal_restart);
735
736 /**
737  * void jbd2_journal_lock_updates () - establish a transaction barrier.
738  * @journal:  Journal to establish a barrier on.
739  *
740  * This locks out any further updates from being started, and blocks
741  * until all existing updates have completed, returning only once the
742  * journal is in a quiescent state with no updates running.
743  *
744  * The journal lock should not be held on entry.
745  */
746 void jbd2_journal_lock_updates(journal_t *journal)
747 {
748         DEFINE_WAIT(wait);
749
750         jbd2_might_wait_for_commit(journal);
751
752         write_lock(&journal->j_state_lock);
753         ++journal->j_barrier_count;
754
755         /* Wait until there are no reserved handles */
756         if (atomic_read(&journal->j_reserved_credits)) {
757                 write_unlock(&journal->j_state_lock);
758                 wait_event(journal->j_wait_reserved,
759                            atomic_read(&journal->j_reserved_credits) == 0);
760                 write_lock(&journal->j_state_lock);
761         }
762
763         /* Wait until there are no running updates */
764         while (1) {
765                 transaction_t *transaction = journal->j_running_transaction;
766
767                 if (!transaction)
768                         break;
769
770                 spin_lock(&transaction->t_handle_lock);
771                 prepare_to_wait(&journal->j_wait_updates, &wait,
772                                 TASK_UNINTERRUPTIBLE);
773                 if (!atomic_read(&transaction->t_updates)) {
774                         spin_unlock(&transaction->t_handle_lock);
775                         finish_wait(&journal->j_wait_updates, &wait);
776                         break;
777                 }
778                 spin_unlock(&transaction->t_handle_lock);
779                 write_unlock(&journal->j_state_lock);
780                 schedule();
781                 finish_wait(&journal->j_wait_updates, &wait);
782                 write_lock(&journal->j_state_lock);
783         }
784         write_unlock(&journal->j_state_lock);
785
786         /*
787          * We have now established a barrier against other normal updates, but
788          * we also need to barrier against other jbd2_journal_lock_updates() calls
789          * to make sure that we serialise special journal-locked operations
790          * too.
791          */
792         mutex_lock(&journal->j_barrier);
793 }
794
795 /**
796  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
797  * @journal:  Journal to release the barrier on.
798  *
799  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
800  *
801  * Should be called without the journal lock held.
802  */
803 void jbd2_journal_unlock_updates (journal_t *journal)
804 {
805         J_ASSERT(journal->j_barrier_count != 0);
806
807         mutex_unlock(&journal->j_barrier);
808         write_lock(&journal->j_state_lock);
809         --journal->j_barrier_count;
810         write_unlock(&journal->j_state_lock);
811         wake_up(&journal->j_wait_transaction_locked);
812 }
813
814 static void warn_dirty_buffer(struct buffer_head *bh)
815 {
816         printk(KERN_WARNING
817                "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
818                "There's a risk of filesystem corruption in case of system "
819                "crash.\n",
820                bh->b_bdev, (unsigned long long)bh->b_blocknr);
821 }
822
823 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
824 static void jbd2_freeze_jh_data(struct journal_head *jh)
825 {
826         struct page *page;
827         int offset;
828         char *source;
829         struct buffer_head *bh = jh2bh(jh);
830
831         J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
832         page = bh->b_page;
833         offset = offset_in_page(bh->b_data);
834         source = kmap_atomic(page);
835         /* Fire data frozen trigger just before we copy the data */
836         jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
837         memcpy(jh->b_frozen_data, source + offset, bh->b_size);
838         kunmap_atomic(source);
839
840         /*
841          * Now that the frozen data is saved off, we need to store any matching
842          * triggers.
843          */
844         jh->b_frozen_triggers = jh->b_triggers;
845 }
846
847 /*
848  * If the buffer is already part of the current transaction, then there
849  * is nothing we need to do.  If it is already part of a prior
850  * transaction which we are still committing to disk, then we need to
851  * make sure that we do not overwrite the old copy: we do copy-out to
852  * preserve the copy going to disk.  We also account the buffer against
853  * the handle's metadata buffer credits (unless the buffer is already
854  * part of the transaction, that is).
855  *
856  */
857 static int
858 do_get_write_access(handle_t *handle, struct journal_head *jh,
859                         int force_copy)
860 {
861         struct buffer_head *bh;
862         transaction_t *transaction = handle->h_transaction;
863         journal_t *journal;
864         int error;
865         char *frozen_buffer = NULL;
866         unsigned long start_lock, time_lock;
867
868         if (is_handle_aborted(handle))
869                 return -EROFS;
870         journal = transaction->t_journal;
871
872         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
873
874         JBUFFER_TRACE(jh, "entry");
875 repeat:
876         bh = jh2bh(jh);
877
878         /* @@@ Need to check for errors here at some point. */
879
880         start_lock = jiffies;
881         lock_buffer(bh);
882         jbd_lock_bh_state(bh);
883
884         /* If it takes too long to lock the buffer, trace it */
885         time_lock = jbd2_time_diff(start_lock, jiffies);
886         if (time_lock > HZ/10)
887                 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
888                         jiffies_to_msecs(time_lock));
889
890         /* We now hold the buffer lock so it is safe to query the buffer
891          * state.  Is the buffer dirty?
892          *
893          * If so, there are two possibilities.  The buffer may be
894          * non-journaled, and undergoing a quite legitimate writeback.
895          * Otherwise, it is journaled, and we don't expect dirty buffers
896          * in that state (the buffers should be marked JBD_Dirty
897          * instead.)  So either the IO is being done under our own
898          * control and this is a bug, or it's a third party IO such as
899          * dump(8) (which may leave the buffer scheduled for read ---
900          * ie. locked but not dirty) or tune2fs (which may actually have
901          * the buffer dirtied, ugh.)  */
902
903         if (buffer_dirty(bh)) {
904                 /*
905                  * First question: is this buffer already part of the current
906                  * transaction or the existing committing transaction?
907                  */
908                 if (jh->b_transaction) {
909                         J_ASSERT_JH(jh,
910                                 jh->b_transaction == transaction ||
911                                 jh->b_transaction ==
912                                         journal->j_committing_transaction);
913                         if (jh->b_next_transaction)
914                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
915                                                         transaction);
916                         warn_dirty_buffer(bh);
917                 }
918                 /*
919                  * In any case we need to clean the dirty flag and we must
920                  * do it under the buffer lock to be sure we don't race
921                  * with running write-out.
922                  */
923                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
924                 clear_buffer_dirty(bh);
925                 set_buffer_jbddirty(bh);
926         }
927
928         unlock_buffer(bh);
929
930         error = -EROFS;
931         if (is_handle_aborted(handle)) {
932                 jbd_unlock_bh_state(bh);
933                 goto out;
934         }
935         error = 0;
936
937         /*
938          * The buffer is already part of this transaction if b_transaction or
939          * b_next_transaction points to it
940          */
941         if (jh->b_transaction == transaction ||
942             jh->b_next_transaction == transaction)
943                 goto done;
944
945         /*
946          * this is the first time this transaction is touching this buffer,
947          * reset the modified flag
948          */
949         jh->b_modified = 0;
950
951         /*
952          * If the buffer is not journaled right now, we need to make sure it
953          * doesn't get written to disk before the caller actually commits the
954          * new data
955          */
956         if (!jh->b_transaction) {
957                 JBUFFER_TRACE(jh, "no transaction");
958                 J_ASSERT_JH(jh, !jh->b_next_transaction);
959                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
960                 /*
961                  * Make sure all stores to jh (b_modified, b_frozen_data) are
962                  * visible before attaching it to the running transaction.
963                  * Paired with barrier in jbd2_write_access_granted()
964                  */
965                 smp_wmb();
966                 spin_lock(&journal->j_list_lock);
967                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
968                 spin_unlock(&journal->j_list_lock);
969                 goto done;
970         }
971         /*
972          * If there is already a copy-out version of this buffer, then we don't
973          * need to make another one
974          */
975         if (jh->b_frozen_data) {
976                 JBUFFER_TRACE(jh, "has frozen data");
977                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
978                 goto attach_next;
979         }
980
981         JBUFFER_TRACE(jh, "owned by older transaction");
982         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
983         J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
984
985         /*
986          * There is one case we have to be very careful about.  If the
987          * committing transaction is currently writing this buffer out to disk
988          * and has NOT made a copy-out, then we cannot modify the buffer
989          * contents at all right now.  The essence of copy-out is that it is
990          * the extra copy, not the primary copy, which gets journaled.  If the
991          * primary copy is already going to disk then we cannot do copy-out
992          * here.
993          */
994         if (buffer_shadow(bh)) {
995                 JBUFFER_TRACE(jh, "on shadow: sleep");
996                 jbd_unlock_bh_state(bh);
997                 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
998                 goto repeat;
999         }
1000
1001         /*
1002          * Only do the copy if the currently-owning transaction still needs it.
1003          * If buffer isn't on BJ_Metadata list, the committing transaction is
1004          * past that stage (here we use the fact that BH_Shadow is set under
1005          * bh_state lock together with refiling to BJ_Shadow list and at this
1006          * point we know the buffer doesn't have BH_Shadow set).
1007          *
1008          * Subtle point, though: if this is a get_undo_access, then we will be
1009          * relying on the frozen_data to contain the new value of the
1010          * committed_data record after the transaction, so we HAVE to force the
1011          * frozen_data copy in that case.
1012          */
1013         if (jh->b_jlist == BJ_Metadata || force_copy) {
1014                 JBUFFER_TRACE(jh, "generate frozen data");
1015                 if (!frozen_buffer) {
1016                         JBUFFER_TRACE(jh, "allocate memory for buffer");
1017                         jbd_unlock_bh_state(bh);
1018                         frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
1019                                                    GFP_NOFS | __GFP_NOFAIL);
1020                         goto repeat;
1021                 }
1022                 jh->b_frozen_data = frozen_buffer;
1023                 frozen_buffer = NULL;
1024                 jbd2_freeze_jh_data(jh);
1025         }
1026 attach_next:
1027         /*
1028          * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1029          * before attaching it to the running transaction. Paired with barrier
1030          * in jbd2_write_access_granted()
1031          */
1032         smp_wmb();
1033         jh->b_next_transaction = transaction;
1034
1035 done:
1036         jbd_unlock_bh_state(bh);
1037
1038         /*
1039          * If we are about to journal a buffer, then any revoke pending on it is
1040          * no longer valid
1041          */
1042         jbd2_journal_cancel_revoke(handle, jh);
1043
1044 out:
1045         if (unlikely(frozen_buffer))    /* It's usually NULL */
1046                 jbd2_free(frozen_buffer, bh->b_size);
1047
1048         JBUFFER_TRACE(jh, "exit");
1049         return error;
1050 }
1051
1052 /* Fast check whether buffer is already attached to the required transaction */
1053 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1054                                                         bool undo)
1055 {
1056         struct journal_head *jh;
1057         bool ret = false;
1058
1059         /* Dirty buffers require special handling... */
1060         if (buffer_dirty(bh))
1061                 return false;
1062
1063         /*
1064          * RCU protects us from dereferencing freed pages. So the checks we do
1065          * are guaranteed not to oops. However the jh slab object can get freed
1066          * & reallocated while we work with it. So we have to be careful. When
1067          * we see jh attached to the running transaction, we know it must stay
1068          * so until the transaction is committed. Thus jh won't be freed and
1069          * will be attached to the same bh while we run.  However it can
1070          * happen jh gets freed, reallocated, and attached to the transaction
1071          * just after we get pointer to it from bh. So we have to be careful
1072          * and recheck jh still belongs to our bh before we return success.
1073          */
1074         rcu_read_lock();
1075         if (!buffer_jbd(bh))
1076                 goto out;
1077         /* This should be bh2jh() but that doesn't work with inline functions */
1078         jh = READ_ONCE(bh->b_private);
1079         if (!jh)
1080                 goto out;
1081         /* For undo access buffer must have data copied */
1082         if (undo && !jh->b_committed_data)
1083                 goto out;
1084         if (jh->b_transaction != handle->h_transaction &&
1085             jh->b_next_transaction != handle->h_transaction)
1086                 goto out;
1087         /*
1088          * There are two reasons for the barrier here:
1089          * 1) Make sure to fetch b_bh after we did previous checks so that we
1090          * detect when jh went through free, realloc, attach to transaction
1091          * while we were checking. Paired with implicit barrier in that path.
1092          * 2) So that access to bh done after jbd2_write_access_granted()
1093          * doesn't get reordered and see inconsistent state of concurrent
1094          * do_get_write_access().
1095          */
1096         smp_mb();
1097         if (unlikely(jh->b_bh != bh))
1098                 goto out;
1099         ret = true;
1100 out:
1101         rcu_read_unlock();
1102         return ret;
1103 }
1104
1105 /**
1106  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1107  * @handle: transaction to add buffer modifications to
1108  * @bh:     bh to be used for metadata writes
1109  *
1110  * Returns: error code or 0 on success.
1111  *
1112  * In full data journalling mode the buffer may be of type BJ_AsyncData,
1113  * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1114  */
1115
1116 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1117 {
1118         struct journal_head *jh;
1119         int rc;
1120
1121         if (jbd2_write_access_granted(handle, bh, false))
1122                 return 0;
1123
1124         jh = jbd2_journal_add_journal_head(bh);
1125         /* We do not want to get caught playing with fields which the
1126          * log thread also manipulates.  Make sure that the buffer
1127          * completes any outstanding IO before proceeding. */
1128         rc = do_get_write_access(handle, jh, 0);
1129         jbd2_journal_put_journal_head(jh);
1130         return rc;
1131 }
1132
1133
1134 /*
1135  * When the user wants to journal a newly created buffer_head
1136  * (ie. getblk() returned a new buffer and we are going to populate it
1137  * manually rather than reading off disk), then we need to keep the
1138  * buffer_head locked until it has been completely filled with new
1139  * data.  In this case, we should be able to make the assertion that
1140  * the bh is not already part of an existing transaction.
1141  *
1142  * The buffer should already be locked by the caller by this point.
1143  * There is no lock ranking violation: it was a newly created,
1144  * unlocked buffer beforehand. */
1145
1146 /**
1147  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1148  * @handle: transaction to new buffer to
1149  * @bh: new buffer.
1150  *
1151  * Call this if you create a new bh.
1152  */
1153 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1154 {
1155         transaction_t *transaction = handle->h_transaction;
1156         journal_t *journal;
1157         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1158         int err;
1159
1160         jbd_debug(5, "journal_head %p\n", jh);
1161         err = -EROFS;
1162         if (is_handle_aborted(handle))
1163                 goto out;
1164         journal = transaction->t_journal;
1165         err = 0;
1166
1167         JBUFFER_TRACE(jh, "entry");
1168         /*
1169          * The buffer may already belong to this transaction due to pre-zeroing
1170          * in the filesystem's new_block code.  It may also be on the previous,
1171          * committing transaction's lists, but it HAS to be in Forget state in
1172          * that case: the transaction must have deleted the buffer for it to be
1173          * reused here.
1174          */
1175         jbd_lock_bh_state(bh);
1176         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1177                 jh->b_transaction == NULL ||
1178                 (jh->b_transaction == journal->j_committing_transaction &&
1179                           jh->b_jlist == BJ_Forget)));
1180
1181         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1182         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1183
1184         if (jh->b_transaction == NULL) {
1185                 /*
1186                  * Previous jbd2_journal_forget() could have left the buffer
1187                  * with jbddirty bit set because it was being committed. When
1188                  * the commit finished, we've filed the buffer for
1189                  * checkpointing and marked it dirty. Now we are reallocating
1190                  * the buffer so the transaction freeing it must have
1191                  * committed and so it's safe to clear the dirty bit.
1192                  */
1193                 clear_buffer_dirty(jh2bh(jh));
1194                 /* first access by this transaction */
1195                 jh->b_modified = 0;
1196
1197                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1198                 spin_lock(&journal->j_list_lock);
1199                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1200                 spin_unlock(&journal->j_list_lock);
1201         } else if (jh->b_transaction == journal->j_committing_transaction) {
1202                 /* first access by this transaction */
1203                 jh->b_modified = 0;
1204
1205                 JBUFFER_TRACE(jh, "set next transaction");
1206                 spin_lock(&journal->j_list_lock);
1207                 jh->b_next_transaction = transaction;
1208                 spin_unlock(&journal->j_list_lock);
1209         }
1210         jbd_unlock_bh_state(bh);
1211
1212         /*
1213          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
1214          * blocks which contain freed but then revoked metadata.  We need
1215          * to cancel the revoke in case we end up freeing it yet again
1216          * and the reallocating as data - this would cause a second revoke,
1217          * which hits an assertion error.
1218          */
1219         JBUFFER_TRACE(jh, "cancelling revoke");
1220         jbd2_journal_cancel_revoke(handle, jh);
1221 out:
1222         jbd2_journal_put_journal_head(jh);
1223         return err;
1224 }
1225
1226 /**
1227  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
1228  *     non-rewindable consequences
1229  * @handle: transaction
1230  * @bh: buffer to undo
1231  *
1232  * Sometimes there is a need to distinguish between metadata which has
1233  * been committed to disk and that which has not.  The ext3fs code uses
1234  * this for freeing and allocating space, we have to make sure that we
1235  * do not reuse freed space until the deallocation has been committed,
1236  * since if we overwrote that space we would make the delete
1237  * un-rewindable in case of a crash.
1238  *
1239  * To deal with that, jbd2_journal_get_undo_access requests write access to a
1240  * buffer for parts of non-rewindable operations such as delete
1241  * operations on the bitmaps.  The journaling code must keep a copy of
1242  * the buffer's contents prior to the undo_access call until such time
1243  * as we know that the buffer has definitely been committed to disk.
1244  *
1245  * We never need to know which transaction the committed data is part
1246  * of, buffers touched here are guaranteed to be dirtied later and so
1247  * will be committed to a new transaction in due course, at which point
1248  * we can discard the old committed data pointer.
1249  *
1250  * Returns error number or 0 on success.
1251  */
1252 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1253 {
1254         int err;
1255         struct journal_head *jh;
1256         char *committed_data = NULL;
1257
1258         if (jbd2_write_access_granted(handle, bh, true))
1259                 return 0;
1260
1261         jh = jbd2_journal_add_journal_head(bh);
1262         JBUFFER_TRACE(jh, "entry");
1263
1264         /*
1265          * Do this first --- it can drop the journal lock, so we want to
1266          * make sure that obtaining the committed_data is done
1267          * atomically wrt. completion of any outstanding commits.
1268          */
1269         err = do_get_write_access(handle, jh, 1);
1270         if (err)
1271                 goto out;
1272
1273 repeat:
1274         if (!jh->b_committed_data)
1275                 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1276                                             GFP_NOFS|__GFP_NOFAIL);
1277
1278         jbd_lock_bh_state(bh);
1279         if (!jh->b_committed_data) {
1280                 /* Copy out the current buffer contents into the
1281                  * preserved, committed copy. */
1282                 JBUFFER_TRACE(jh, "generate b_committed data");
1283                 if (!committed_data) {
1284                         jbd_unlock_bh_state(bh);
1285                         goto repeat;
1286                 }
1287
1288                 jh->b_committed_data = committed_data;
1289                 committed_data = NULL;
1290                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1291         }
1292         jbd_unlock_bh_state(bh);
1293 out:
1294         jbd2_journal_put_journal_head(jh);
1295         if (unlikely(committed_data))
1296                 jbd2_free(committed_data, bh->b_size);
1297         return err;
1298 }
1299
1300 /**
1301  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1302  * @bh: buffer to trigger on
1303  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1304  *
1305  * Set any triggers on this journal_head.  This is always safe, because
1306  * triggers for a committing buffer will be saved off, and triggers for
1307  * a running transaction will match the buffer in that transaction.
1308  *
1309  * Call with NULL to clear the triggers.
1310  */
1311 void jbd2_journal_set_triggers(struct buffer_head *bh,
1312                                struct jbd2_buffer_trigger_type *type)
1313 {
1314         struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1315
1316         if (WARN_ON(!jh))
1317                 return;
1318         jh->b_triggers = type;
1319         jbd2_journal_put_journal_head(jh);
1320 }
1321
1322 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1323                                 struct jbd2_buffer_trigger_type *triggers)
1324 {
1325         struct buffer_head *bh = jh2bh(jh);
1326
1327         if (!triggers || !triggers->t_frozen)
1328                 return;
1329
1330         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1331 }
1332
1333 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1334                                struct jbd2_buffer_trigger_type *triggers)
1335 {
1336         if (!triggers || !triggers->t_abort)
1337                 return;
1338
1339         triggers->t_abort(triggers, jh2bh(jh));
1340 }
1341
1342 /**
1343  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1344  * @handle: transaction to add buffer to.
1345  * @bh: buffer to mark
1346  *
1347  * mark dirty metadata which needs to be journaled as part of the current
1348  * transaction.
1349  *
1350  * The buffer must have previously had jbd2_journal_get_write_access()
1351  * called so that it has a valid journal_head attached to the buffer
1352  * head.
1353  *
1354  * The buffer is placed on the transaction's metadata list and is marked
1355  * as belonging to the transaction.
1356  *
1357  * Returns error number or 0 on success.
1358  *
1359  * Special care needs to be taken if the buffer already belongs to the
1360  * current committing transaction (in which case we should have frozen
1361  * data present for that commit).  In that case, we don't relink the
1362  * buffer: that only gets done when the old transaction finally
1363  * completes its commit.
1364  */
1365 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1366 {
1367         transaction_t *transaction = handle->h_transaction;
1368         journal_t *journal;
1369         struct journal_head *jh;
1370         int ret = 0;
1371
1372         if (is_handle_aborted(handle))
1373                 return -EROFS;
1374         if (!buffer_jbd(bh))
1375                 return -EUCLEAN;
1376
1377         /*
1378          * We don't grab jh reference here since the buffer must be part
1379          * of the running transaction.
1380          */
1381         jh = bh2jh(bh);
1382         jbd_debug(5, "journal_head %p\n", jh);
1383         JBUFFER_TRACE(jh, "entry");
1384
1385         /*
1386          * This and the following assertions are unreliable since we may see jh
1387          * in inconsistent state unless we grab bh_state lock. But this is
1388          * crucial to catch bugs so let's do a reliable check until the
1389          * lockless handling is fully proven.
1390          */
1391         if (jh->b_transaction != transaction &&
1392             jh->b_next_transaction != transaction) {
1393                 jbd_lock_bh_state(bh);
1394                 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1395                                 jh->b_next_transaction == transaction);
1396                 jbd_unlock_bh_state(bh);
1397         }
1398         if (jh->b_modified == 1) {
1399                 /* If it's in our transaction it must be in BJ_Metadata list. */
1400                 if (jh->b_transaction == transaction &&
1401                     jh->b_jlist != BJ_Metadata) {
1402                         jbd_lock_bh_state(bh);
1403                         if (jh->b_transaction == transaction &&
1404                             jh->b_jlist != BJ_Metadata)
1405                                 pr_err("JBD2: assertion failure: h_type=%u "
1406                                        "h_line_no=%u block_no=%llu jlist=%u\n",
1407                                        handle->h_type, handle->h_line_no,
1408                                        (unsigned long long) bh->b_blocknr,
1409                                        jh->b_jlist);
1410                         J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1411                                         jh->b_jlist == BJ_Metadata);
1412                         jbd_unlock_bh_state(bh);
1413                 }
1414                 goto out;
1415         }
1416
1417         journal = transaction->t_journal;
1418         jbd_lock_bh_state(bh);
1419
1420         if (jh->b_modified == 0) {
1421                 /*
1422                  * This buffer's got modified and becoming part
1423                  * of the transaction. This needs to be done
1424                  * once a transaction -bzzz
1425                  */
1426                 if (handle->h_buffer_credits <= 0) {
1427                         ret = -ENOSPC;
1428                         goto out_unlock_bh;
1429                 }
1430                 jh->b_modified = 1;
1431                 handle->h_buffer_credits--;
1432         }
1433
1434         /*
1435          * fastpath, to avoid expensive locking.  If this buffer is already
1436          * on the running transaction's metadata list there is nothing to do.
1437          * Nobody can take it off again because there is a handle open.
1438          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1439          * result in this test being false, so we go in and take the locks.
1440          */
1441         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1442                 JBUFFER_TRACE(jh, "fastpath");
1443                 if (unlikely(jh->b_transaction !=
1444                              journal->j_running_transaction)) {
1445                         printk(KERN_ERR "JBD2: %s: "
1446                                "jh->b_transaction (%llu, %p, %u) != "
1447                                "journal->j_running_transaction (%p, %u)\n",
1448                                journal->j_devname,
1449                                (unsigned long long) bh->b_blocknr,
1450                                jh->b_transaction,
1451                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1452                                journal->j_running_transaction,
1453                                journal->j_running_transaction ?
1454                                journal->j_running_transaction->t_tid : 0);
1455                         ret = -EINVAL;
1456                 }
1457                 goto out_unlock_bh;
1458         }
1459
1460         set_buffer_jbddirty(bh);
1461
1462         /*
1463          * Metadata already on the current transaction list doesn't
1464          * need to be filed.  Metadata on another transaction's list must
1465          * be committing, and will be refiled once the commit completes:
1466          * leave it alone for now.
1467          */
1468         if (jh->b_transaction != transaction) {
1469                 JBUFFER_TRACE(jh, "already on other transaction");
1470                 if (unlikely(((jh->b_transaction !=
1471                                journal->j_committing_transaction)) ||
1472                              (jh->b_next_transaction != transaction))) {
1473                         printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1474                                "bad jh for block %llu: "
1475                                "transaction (%p, %u), "
1476                                "jh->b_transaction (%p, %u), "
1477                                "jh->b_next_transaction (%p, %u), jlist %u\n",
1478                                journal->j_devname,
1479                                (unsigned long long) bh->b_blocknr,
1480                                transaction, transaction->t_tid,
1481                                jh->b_transaction,
1482                                jh->b_transaction ?
1483                                jh->b_transaction->t_tid : 0,
1484                                jh->b_next_transaction,
1485                                jh->b_next_transaction ?
1486                                jh->b_next_transaction->t_tid : 0,
1487                                jh->b_jlist);
1488                         WARN_ON(1);
1489                         ret = -EINVAL;
1490                 }
1491                 /* And this case is illegal: we can't reuse another
1492                  * transaction's data buffer, ever. */
1493                 goto out_unlock_bh;
1494         }
1495
1496         /* That test should have eliminated the following case: */
1497         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1498
1499         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1500         spin_lock(&journal->j_list_lock);
1501         __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1502         spin_unlock(&journal->j_list_lock);
1503 out_unlock_bh:
1504         jbd_unlock_bh_state(bh);
1505 out:
1506         JBUFFER_TRACE(jh, "exit");
1507         return ret;
1508 }
1509
1510 /**
1511  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1512  * @handle: transaction handle
1513  * @bh:     bh to 'forget'
1514  *
1515  * We can only do the bforget if there are no commits pending against the
1516  * buffer.  If the buffer is dirty in the current running transaction we
1517  * can safely unlink it.
1518  *
1519  * bh may not be a journalled buffer at all - it may be a non-JBD
1520  * buffer which came off the hashtable.  Check for this.
1521  *
1522  * Decrements bh->b_count by one.
1523  *
1524  * Allow this call even if the handle has aborted --- it may be part of
1525  * the caller's cleanup after an abort.
1526  */
1527 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1528 {
1529         transaction_t *transaction = handle->h_transaction;
1530         journal_t *journal;
1531         struct journal_head *jh;
1532         int drop_reserve = 0;
1533         int err = 0;
1534         int was_modified = 0;
1535
1536         if (is_handle_aborted(handle))
1537                 return -EROFS;
1538         journal = transaction->t_journal;
1539
1540         BUFFER_TRACE(bh, "entry");
1541
1542         jbd_lock_bh_state(bh);
1543
1544         if (!buffer_jbd(bh))
1545                 goto not_jbd;
1546         jh = bh2jh(bh);
1547
1548         /* Critical error: attempting to delete a bitmap buffer, maybe?
1549          * Don't do any jbd operations, and return an error. */
1550         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1551                          "inconsistent data on disk")) {
1552                 err = -EIO;
1553                 goto not_jbd;
1554         }
1555
1556         /* keep track of whether or not this transaction modified us */
1557         was_modified = jh->b_modified;
1558
1559         /*
1560          * The buffer's going from the transaction, we must drop
1561          * all references -bzzz
1562          */
1563         jh->b_modified = 0;
1564
1565         if (jh->b_transaction == transaction) {
1566                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1567
1568                 /* If we are forgetting a buffer which is already part
1569                  * of this transaction, then we can just drop it from
1570                  * the transaction immediately. */
1571                 clear_buffer_dirty(bh);
1572                 clear_buffer_jbddirty(bh);
1573
1574                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1575
1576                 /*
1577                  * we only want to drop a reference if this transaction
1578                  * modified the buffer
1579                  */
1580                 if (was_modified)
1581                         drop_reserve = 1;
1582
1583                 /*
1584                  * We are no longer going to journal this buffer.
1585                  * However, the commit of this transaction is still
1586                  * important to the buffer: the delete that we are now
1587                  * processing might obsolete an old log entry, so by
1588                  * committing, we can satisfy the buffer's checkpoint.
1589                  *
1590                  * So, if we have a checkpoint on the buffer, we should
1591                  * now refile the buffer on our BJ_Forget list so that
1592                  * we know to remove the checkpoint after we commit.
1593                  */
1594
1595                 spin_lock(&journal->j_list_lock);
1596                 if (jh->b_cp_transaction) {
1597                         __jbd2_journal_temp_unlink_buffer(jh);
1598                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1599                 } else {
1600                         __jbd2_journal_unfile_buffer(jh);
1601                         if (!buffer_jbd(bh)) {
1602                                 spin_unlock(&journal->j_list_lock);
1603                                 goto not_jbd;
1604                         }
1605                 }
1606                 spin_unlock(&journal->j_list_lock);
1607         } else if (jh->b_transaction) {
1608                 J_ASSERT_JH(jh, (jh->b_transaction ==
1609                                  journal->j_committing_transaction));
1610                 /* However, if the buffer is still owned by a prior
1611                  * (committing) transaction, we can't drop it yet... */
1612                 JBUFFER_TRACE(jh, "belongs to older transaction");
1613                 /* ... but we CAN drop it from the new transaction through
1614                  * marking the buffer as freed and set j_next_transaction to
1615                  * the new transaction, so that not only the commit code
1616                  * knows it should clear dirty bits when it is done with the
1617                  * buffer, but also the buffer can be checkpointed only
1618                  * after the new transaction commits. */
1619
1620                 set_buffer_freed(bh);
1621
1622                 if (!jh->b_next_transaction) {
1623                         spin_lock(&journal->j_list_lock);
1624                         jh->b_next_transaction = transaction;
1625                         spin_unlock(&journal->j_list_lock);
1626                 } else {
1627                         J_ASSERT(jh->b_next_transaction == transaction);
1628
1629                         /*
1630                          * only drop a reference if this transaction modified
1631                          * the buffer
1632                          */
1633                         if (was_modified)
1634                                 drop_reserve = 1;
1635                 }
1636         } else {
1637                 /*
1638                  * Finally, if the buffer is not belongs to any
1639                  * transaction, we can just drop it now if it has no
1640                  * checkpoint.
1641                  */
1642                 spin_lock(&journal->j_list_lock);
1643                 if (!jh->b_cp_transaction) {
1644                         JBUFFER_TRACE(jh, "belongs to none transaction");
1645                         spin_unlock(&journal->j_list_lock);
1646                         goto not_jbd;
1647                 }
1648
1649                 /*
1650                  * Otherwise, if the buffer has been written to disk,
1651                  * it is safe to remove the checkpoint and drop it.
1652                  */
1653                 if (!buffer_dirty(bh)) {
1654                         __jbd2_journal_remove_checkpoint(jh);
1655                         spin_unlock(&journal->j_list_lock);
1656                         goto not_jbd;
1657                 }
1658
1659                 /*
1660                  * The buffer is still not written to disk, we should
1661                  * attach this buffer to current transaction so that the
1662                  * buffer can be checkpointed only after the current
1663                  * transaction commits.
1664                  */
1665                 clear_buffer_dirty(bh);
1666                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1667                 spin_unlock(&journal->j_list_lock);
1668         }
1669
1670         jbd_unlock_bh_state(bh);
1671         __brelse(bh);
1672 drop:
1673         if (drop_reserve) {
1674                 /* no need to reserve log space for this block -bzzz */
1675                 handle->h_buffer_credits++;
1676         }
1677         return err;
1678
1679 not_jbd:
1680         jbd_unlock_bh_state(bh);
1681         __bforget(bh);
1682         goto drop;
1683 }
1684
1685 /**
1686  * int jbd2_journal_stop() - complete a transaction
1687  * @handle: transaction to complete.
1688  *
1689  * All done for a particular handle.
1690  *
1691  * There is not much action needed here.  We just return any remaining
1692  * buffer credits to the transaction and remove the handle.  The only
1693  * complication is that we need to start a commit operation if the
1694  * filesystem is marked for synchronous update.
1695  *
1696  * jbd2_journal_stop itself will not usually return an error, but it may
1697  * do so in unusual circumstances.  In particular, expect it to
1698  * return -EIO if a jbd2_journal_abort has been executed since the
1699  * transaction began.
1700  */
1701 int jbd2_journal_stop(handle_t *handle)
1702 {
1703         transaction_t *transaction = handle->h_transaction;
1704         journal_t *journal;
1705         int err = 0, wait_for_commit = 0;
1706         tid_t tid;
1707         pid_t pid;
1708
1709         if (!transaction) {
1710                 /*
1711                  * Handle is already detached from the transaction so
1712                  * there is nothing to do other than decrease a refcount,
1713                  * or free the handle if refcount drops to zero
1714                  */
1715                 if (--handle->h_ref > 0) {
1716                         jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1717                                                          handle->h_ref);
1718                         return err;
1719                 } else {
1720                         if (handle->h_rsv_handle)
1721                                 jbd2_free_handle(handle->h_rsv_handle);
1722                         goto free_and_exit;
1723                 }
1724         }
1725         journal = transaction->t_journal;
1726
1727         J_ASSERT(journal_current_handle() == handle);
1728
1729         if (is_handle_aborted(handle))
1730                 err = -EIO;
1731         else
1732                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1733
1734         if (--handle->h_ref > 0) {
1735                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1736                           handle->h_ref);
1737                 return err;
1738         }
1739
1740         jbd_debug(4, "Handle %p going down\n", handle);
1741         trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1742                                 transaction->t_tid,
1743                                 handle->h_type, handle->h_line_no,
1744                                 jiffies - handle->h_start_jiffies,
1745                                 handle->h_sync, handle->h_requested_credits,
1746                                 (handle->h_requested_credits -
1747                                  handle->h_buffer_credits));
1748
1749         /*
1750          * Implement synchronous transaction batching.  If the handle
1751          * was synchronous, don't force a commit immediately.  Let's
1752          * yield and let another thread piggyback onto this
1753          * transaction.  Keep doing that while new threads continue to
1754          * arrive.  It doesn't cost much - we're about to run a commit
1755          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1756          * operations by 30x or more...
1757          *
1758          * We try and optimize the sleep time against what the
1759          * underlying disk can do, instead of having a static sleep
1760          * time.  This is useful for the case where our storage is so
1761          * fast that it is more optimal to go ahead and force a flush
1762          * and wait for the transaction to be committed than it is to
1763          * wait for an arbitrary amount of time for new writers to
1764          * join the transaction.  We achieve this by measuring how
1765          * long it takes to commit a transaction, and compare it with
1766          * how long this transaction has been running, and if run time
1767          * < commit time then we sleep for the delta and commit.  This
1768          * greatly helps super fast disks that would see slowdowns as
1769          * more threads started doing fsyncs.
1770          *
1771          * But don't do this if this process was the most recent one
1772          * to perform a synchronous write.  We do this to detect the
1773          * case where a single process is doing a stream of sync
1774          * writes.  No point in waiting for joiners in that case.
1775          *
1776          * Setting max_batch_time to 0 disables this completely.
1777          */
1778         pid = current->pid;
1779         if (handle->h_sync && journal->j_last_sync_writer != pid &&
1780             journal->j_max_batch_time) {
1781                 u64 commit_time, trans_time;
1782
1783                 journal->j_last_sync_writer = pid;
1784
1785                 read_lock(&journal->j_state_lock);
1786                 commit_time = journal->j_average_commit_time;
1787                 read_unlock(&journal->j_state_lock);
1788
1789                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1790                                                    transaction->t_start_time));
1791
1792                 commit_time = max_t(u64, commit_time,
1793                                     1000*journal->j_min_batch_time);
1794                 commit_time = min_t(u64, commit_time,
1795                                     1000*journal->j_max_batch_time);
1796
1797                 if (trans_time < commit_time) {
1798                         ktime_t expires = ktime_add_ns(ktime_get(),
1799                                                        commit_time);
1800                         set_current_state(TASK_UNINTERRUPTIBLE);
1801                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1802                 }
1803         }
1804
1805         if (handle->h_sync)
1806                 transaction->t_synchronous_commit = 1;
1807         current->journal_info = NULL;
1808         atomic_sub(handle->h_buffer_credits,
1809                    &transaction->t_outstanding_credits);
1810
1811         /*
1812          * If the handle is marked SYNC, we need to set another commit
1813          * going!  We also want to force a commit if the current
1814          * transaction is occupying too much of the log, or if the
1815          * transaction is too old now.
1816          */
1817         if (handle->h_sync ||
1818             (atomic_read(&transaction->t_outstanding_credits) >
1819              journal->j_max_transaction_buffers) ||
1820             time_after_eq(jiffies, transaction->t_expires)) {
1821                 /* Do this even for aborted journals: an abort still
1822                  * completes the commit thread, it just doesn't write
1823                  * anything to disk. */
1824
1825                 jbd_debug(2, "transaction too old, requesting commit for "
1826                                         "handle %p\n", handle);
1827                 /* This is non-blocking */
1828                 jbd2_log_start_commit(journal, transaction->t_tid);
1829
1830                 /*
1831                  * Special case: JBD2_SYNC synchronous updates require us
1832                  * to wait for the commit to complete.
1833                  */
1834                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1835                         wait_for_commit = 1;
1836         }
1837
1838         /*
1839          * Once we drop t_updates, if it goes to zero the transaction
1840          * could start committing on us and eventually disappear.  So
1841          * once we do this, we must not dereference transaction
1842          * pointer again.
1843          */
1844         tid = transaction->t_tid;
1845         if (atomic_dec_and_test(&transaction->t_updates)) {
1846                 wake_up(&journal->j_wait_updates);
1847                 if (journal->j_barrier_count)
1848                         wake_up(&journal->j_wait_transaction_locked);
1849         }
1850
1851         rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
1852
1853         if (wait_for_commit)
1854                 err = jbd2_log_wait_commit(journal, tid);
1855
1856         if (handle->h_rsv_handle)
1857                 jbd2_journal_free_reserved(handle->h_rsv_handle);
1858 free_and_exit:
1859         /*
1860          * Scope of the GFP_NOFS context is over here and so we can restore the
1861          * original alloc context.
1862          */
1863         memalloc_nofs_restore(handle->saved_alloc_context);
1864         jbd2_free_handle(handle);
1865         return err;
1866 }
1867
1868 /*
1869  *
1870  * List management code snippets: various functions for manipulating the
1871  * transaction buffer lists.
1872  *
1873  */
1874
1875 /*
1876  * Append a buffer to a transaction list, given the transaction's list head
1877  * pointer.
1878  *
1879  * j_list_lock is held.
1880  *
1881  * jbd_lock_bh_state(jh2bh(jh)) is held.
1882  */
1883
1884 static inline void
1885 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1886 {
1887         if (!*list) {
1888                 jh->b_tnext = jh->b_tprev = jh;
1889                 *list = jh;
1890         } else {
1891                 /* Insert at the tail of the list to preserve order */
1892                 struct journal_head *first = *list, *last = first->b_tprev;
1893                 jh->b_tprev = last;
1894                 jh->b_tnext = first;
1895                 last->b_tnext = first->b_tprev = jh;
1896         }
1897 }
1898
1899 /*
1900  * Remove a buffer from a transaction list, given the transaction's list
1901  * head pointer.
1902  *
1903  * Called with j_list_lock held, and the journal may not be locked.
1904  *
1905  * jbd_lock_bh_state(jh2bh(jh)) is held.
1906  */
1907
1908 static inline void
1909 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1910 {
1911         if (*list == jh) {
1912                 *list = jh->b_tnext;
1913                 if (*list == jh)
1914                         *list = NULL;
1915         }
1916         jh->b_tprev->b_tnext = jh->b_tnext;
1917         jh->b_tnext->b_tprev = jh->b_tprev;
1918 }
1919
1920 /*
1921  * Remove a buffer from the appropriate transaction list.
1922  *
1923  * Note that this function can *change* the value of
1924  * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1925  * t_reserved_list.  If the caller is holding onto a copy of one of these
1926  * pointers, it could go bad.  Generally the caller needs to re-read the
1927  * pointer from the transaction_t.
1928  *
1929  * Called under j_list_lock.
1930  */
1931 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1932 {
1933         struct journal_head **list = NULL;
1934         transaction_t *transaction;
1935         struct buffer_head *bh = jh2bh(jh);
1936
1937         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1938         transaction = jh->b_transaction;
1939         if (transaction)
1940                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1941
1942         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1943         if (jh->b_jlist != BJ_None)
1944                 J_ASSERT_JH(jh, transaction != NULL);
1945
1946         switch (jh->b_jlist) {
1947         case BJ_None:
1948                 return;
1949         case BJ_Metadata:
1950                 transaction->t_nr_buffers--;
1951                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1952                 list = &transaction->t_buffers;
1953                 break;
1954         case BJ_Forget:
1955                 list = &transaction->t_forget;
1956                 break;
1957         case BJ_Shadow:
1958                 list = &transaction->t_shadow_list;
1959                 break;
1960         case BJ_Reserved:
1961                 list = &transaction->t_reserved_list;
1962                 break;
1963         }
1964
1965         __blist_del_buffer(list, jh);
1966         jh->b_jlist = BJ_None;
1967         if (transaction && is_journal_aborted(transaction->t_journal))
1968                 clear_buffer_jbddirty(bh);
1969         else if (test_clear_buffer_jbddirty(bh))
1970                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1971 }
1972
1973 /*
1974  * Remove buffer from all transactions.
1975  *
1976  * Called with bh_state lock and j_list_lock
1977  *
1978  * jh and bh may be already freed when this function returns.
1979  */
1980 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1981 {
1982         __jbd2_journal_temp_unlink_buffer(jh);
1983         jh->b_transaction = NULL;
1984         jbd2_journal_put_journal_head(jh);
1985 }
1986
1987 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1988 {
1989         struct buffer_head *bh = jh2bh(jh);
1990
1991         /* Get reference so that buffer cannot be freed before we unlock it */
1992         get_bh(bh);
1993         jbd_lock_bh_state(bh);
1994         spin_lock(&journal->j_list_lock);
1995         __jbd2_journal_unfile_buffer(jh);
1996         spin_unlock(&journal->j_list_lock);
1997         jbd_unlock_bh_state(bh);
1998         __brelse(bh);
1999 }
2000
2001 /*
2002  * Called from jbd2_journal_try_to_free_buffers().
2003  *
2004  * Called under jbd_lock_bh_state(bh)
2005  */
2006 static void
2007 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
2008 {
2009         struct journal_head *jh;
2010
2011         jh = bh2jh(bh);
2012
2013         if (buffer_locked(bh) || buffer_dirty(bh))
2014                 goto out;
2015
2016         if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
2017                 goto out;
2018
2019         spin_lock(&journal->j_list_lock);
2020         if (jh->b_cp_transaction != NULL) {
2021                 /* written-back checkpointed metadata buffer */
2022                 JBUFFER_TRACE(jh, "remove from checkpoint list");
2023                 __jbd2_journal_remove_checkpoint(jh);
2024         }
2025         spin_unlock(&journal->j_list_lock);
2026 out:
2027         return;
2028 }
2029
2030 /**
2031  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
2032  * @journal: journal for operation
2033  * @page: to try and free
2034  * @gfp_mask: we use the mask to detect how hard should we try to release
2035  * buffers. If __GFP_DIRECT_RECLAIM and __GFP_FS is set, we wait for commit
2036  * code to release the buffers.
2037  *
2038  *
2039  * For all the buffers on this page,
2040  * if they are fully written out ordered data, move them onto BUF_CLEAN
2041  * so try_to_free_buffers() can reap them.
2042  *
2043  * This function returns non-zero if we wish try_to_free_buffers()
2044  * to be called. We do this if the page is releasable by try_to_free_buffers().
2045  * We also do it if the page has locked or dirty buffers and the caller wants
2046  * us to perform sync or async writeout.
2047  *
2048  * This complicates JBD locking somewhat.  We aren't protected by the
2049  * BKL here.  We wish to remove the buffer from its committing or
2050  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2051  *
2052  * This may *change* the value of transaction_t->t_datalist, so anyone
2053  * who looks at t_datalist needs to lock against this function.
2054  *
2055  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2056  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
2057  * will come out of the lock with the buffer dirty, which makes it
2058  * ineligible for release here.
2059  *
2060  * Who else is affected by this?  hmm...  Really the only contender
2061  * is do_get_write_access() - it could be looking at the buffer while
2062  * journal_try_to_free_buffer() is changing its state.  But that
2063  * cannot happen because we never reallocate freed data as metadata
2064  * while the data is part of a transaction.  Yes?
2065  *
2066  * Return 0 on failure, 1 on success
2067  */
2068 int jbd2_journal_try_to_free_buffers(journal_t *journal,
2069                                 struct page *page, gfp_t gfp_mask)
2070 {
2071         struct buffer_head *head;
2072         struct buffer_head *bh;
2073         int ret = 0;
2074
2075         J_ASSERT(PageLocked(page));
2076
2077         head = page_buffers(page);
2078         bh = head;
2079         do {
2080                 struct journal_head *jh;
2081
2082                 /*
2083                  * We take our own ref against the journal_head here to avoid
2084                  * having to add tons of locking around each instance of
2085                  * jbd2_journal_put_journal_head().
2086                  */
2087                 jh = jbd2_journal_grab_journal_head(bh);
2088                 if (!jh)
2089                         continue;
2090
2091                 jbd_lock_bh_state(bh);
2092                 __journal_try_to_free_buffer(journal, bh);
2093                 jbd2_journal_put_journal_head(jh);
2094                 jbd_unlock_bh_state(bh);
2095                 if (buffer_jbd(bh))
2096                         goto busy;
2097         } while ((bh = bh->b_this_page) != head);
2098
2099         ret = try_to_free_buffers(page);
2100
2101 busy:
2102         return ret;
2103 }
2104
2105 /*
2106  * This buffer is no longer needed.  If it is on an older transaction's
2107  * checkpoint list we need to record it on this transaction's forget list
2108  * to pin this buffer (and hence its checkpointing transaction) down until
2109  * this transaction commits.  If the buffer isn't on a checkpoint list, we
2110  * release it.
2111  * Returns non-zero if JBD no longer has an interest in the buffer.
2112  *
2113  * Called under j_list_lock.
2114  *
2115  * Called under jbd_lock_bh_state(bh).
2116  */
2117 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2118 {
2119         int may_free = 1;
2120         struct buffer_head *bh = jh2bh(jh);
2121
2122         if (jh->b_cp_transaction) {
2123                 JBUFFER_TRACE(jh, "on running+cp transaction");
2124                 __jbd2_journal_temp_unlink_buffer(jh);
2125                 /*
2126                  * We don't want to write the buffer anymore, clear the
2127                  * bit so that we don't confuse checks in
2128                  * __journal_file_buffer
2129                  */
2130                 clear_buffer_dirty(bh);
2131                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2132                 may_free = 0;
2133         } else {
2134                 JBUFFER_TRACE(jh, "on running transaction");
2135                 __jbd2_journal_unfile_buffer(jh);
2136         }
2137         return may_free;
2138 }
2139
2140 /*
2141  * jbd2_journal_invalidatepage
2142  *
2143  * This code is tricky.  It has a number of cases to deal with.
2144  *
2145  * There are two invariants which this code relies on:
2146  *
2147  * i_size must be updated on disk before we start calling invalidatepage on the
2148  * data.
2149  *
2150  *  This is done in ext3 by defining an ext3_setattr method which
2151  *  updates i_size before truncate gets going.  By maintaining this
2152  *  invariant, we can be sure that it is safe to throw away any buffers
2153  *  attached to the current transaction: once the transaction commits,
2154  *  we know that the data will not be needed.
2155  *
2156  *  Note however that we can *not* throw away data belonging to the
2157  *  previous, committing transaction!
2158  *
2159  * Any disk blocks which *are* part of the previous, committing
2160  * transaction (and which therefore cannot be discarded immediately) are
2161  * not going to be reused in the new running transaction
2162  *
2163  *  The bitmap committed_data images guarantee this: any block which is
2164  *  allocated in one transaction and removed in the next will be marked
2165  *  as in-use in the committed_data bitmap, so cannot be reused until
2166  *  the next transaction to delete the block commits.  This means that
2167  *  leaving committing buffers dirty is quite safe: the disk blocks
2168  *  cannot be reallocated to a different file and so buffer aliasing is
2169  *  not possible.
2170  *
2171  *
2172  * The above applies mainly to ordered data mode.  In writeback mode we
2173  * don't make guarantees about the order in which data hits disk --- in
2174  * particular we don't guarantee that new dirty data is flushed before
2175  * transaction commit --- so it is always safe just to discard data
2176  * immediately in that mode.  --sct
2177  */
2178
2179 /*
2180  * The journal_unmap_buffer helper function returns zero if the buffer
2181  * concerned remains pinned as an anonymous buffer belonging to an older
2182  * transaction.
2183  *
2184  * We're outside-transaction here.  Either or both of j_running_transaction
2185  * and j_committing_transaction may be NULL.
2186  */
2187 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2188                                 int partial_page)
2189 {
2190         transaction_t *transaction;
2191         struct journal_head *jh;
2192         int may_free = 1;
2193
2194         BUFFER_TRACE(bh, "entry");
2195
2196         /*
2197          * It is safe to proceed here without the j_list_lock because the
2198          * buffers cannot be stolen by try_to_free_buffers as long as we are
2199          * holding the page lock. --sct
2200          */
2201
2202         if (!buffer_jbd(bh))
2203                 goto zap_buffer_unlocked;
2204
2205         /* OK, we have data buffer in journaled mode */
2206         write_lock(&journal->j_state_lock);
2207         jbd_lock_bh_state(bh);
2208         spin_lock(&journal->j_list_lock);
2209
2210         jh = jbd2_journal_grab_journal_head(bh);
2211         if (!jh)
2212                 goto zap_buffer_no_jh;
2213
2214         /*
2215          * We cannot remove the buffer from checkpoint lists until the
2216          * transaction adding inode to orphan list (let's call it T)
2217          * is committed.  Otherwise if the transaction changing the
2218          * buffer would be cleaned from the journal before T is
2219          * committed, a crash will cause that the correct contents of
2220          * the buffer will be lost.  On the other hand we have to
2221          * clear the buffer dirty bit at latest at the moment when the
2222          * transaction marking the buffer as freed in the filesystem
2223          * structures is committed because from that moment on the
2224          * block can be reallocated and used by a different page.
2225          * Since the block hasn't been freed yet but the inode has
2226          * already been added to orphan list, it is safe for us to add
2227          * the buffer to BJ_Forget list of the newest transaction.
2228          *
2229          * Also we have to clear buffer_mapped flag of a truncated buffer
2230          * because the buffer_head may be attached to the page straddling
2231          * i_size (can happen only when blocksize < pagesize) and thus the
2232          * buffer_head can be reused when the file is extended again. So we end
2233          * up keeping around invalidated buffers attached to transactions'
2234          * BJ_Forget list just to stop checkpointing code from cleaning up
2235          * the transaction this buffer was modified in.
2236          */
2237         transaction = jh->b_transaction;
2238         if (transaction == NULL) {
2239                 /* First case: not on any transaction.  If it
2240                  * has no checkpoint link, then we can zap it:
2241                  * it's a writeback-mode buffer so we don't care
2242                  * if it hits disk safely. */
2243                 if (!jh->b_cp_transaction) {
2244                         JBUFFER_TRACE(jh, "not on any transaction: zap");
2245                         goto zap_buffer;
2246                 }
2247
2248                 if (!buffer_dirty(bh)) {
2249                         /* bdflush has written it.  We can drop it now */
2250                         __jbd2_journal_remove_checkpoint(jh);
2251                         goto zap_buffer;
2252                 }
2253
2254                 /* OK, it must be in the journal but still not
2255                  * written fully to disk: it's metadata or
2256                  * journaled data... */
2257
2258                 if (journal->j_running_transaction) {
2259                         /* ... and once the current transaction has
2260                          * committed, the buffer won't be needed any
2261                          * longer. */
2262                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2263                         may_free = __dispose_buffer(jh,
2264                                         journal->j_running_transaction);
2265                         goto zap_buffer;
2266                 } else {
2267                         /* There is no currently-running transaction. So the
2268                          * orphan record which we wrote for this file must have
2269                          * passed into commit.  We must attach this buffer to
2270                          * the committing transaction, if it exists. */
2271                         if (journal->j_committing_transaction) {
2272                                 JBUFFER_TRACE(jh, "give to committing trans");
2273                                 may_free = __dispose_buffer(jh,
2274                                         journal->j_committing_transaction);
2275                                 goto zap_buffer;
2276                         } else {
2277                                 /* The orphan record's transaction has
2278                                  * committed.  We can cleanse this buffer */
2279                                 clear_buffer_jbddirty(bh);
2280                                 __jbd2_journal_remove_checkpoint(jh);
2281                                 goto zap_buffer;
2282                         }
2283                 }
2284         } else if (transaction == journal->j_committing_transaction) {
2285                 JBUFFER_TRACE(jh, "on committing transaction");
2286                 /*
2287                  * The buffer is committing, we simply cannot touch
2288                  * it. If the page is straddling i_size we have to wait
2289                  * for commit and try again.
2290                  */
2291                 if (partial_page) {
2292                         jbd2_journal_put_journal_head(jh);
2293                         spin_unlock(&journal->j_list_lock);
2294                         jbd_unlock_bh_state(bh);
2295                         write_unlock(&journal->j_state_lock);
2296                         return -EBUSY;
2297                 }
2298                 /*
2299                  * OK, buffer won't be reachable after truncate. We just set
2300                  * j_next_transaction to the running transaction (if there is
2301                  * one) and mark buffer as freed so that commit code knows it
2302                  * should clear dirty bits when it is done with the buffer.
2303                  */
2304                 set_buffer_freed(bh);
2305                 if (journal->j_running_transaction && buffer_jbddirty(bh))
2306                         jh->b_next_transaction = journal->j_running_transaction;
2307                 jbd2_journal_put_journal_head(jh);
2308                 spin_unlock(&journal->j_list_lock);
2309                 jbd_unlock_bh_state(bh);
2310                 write_unlock(&journal->j_state_lock);
2311                 return 0;
2312         } else {
2313                 /* Good, the buffer belongs to the running transaction.
2314                  * We are writing our own transaction's data, not any
2315                  * previous one's, so it is safe to throw it away
2316                  * (remember that we expect the filesystem to have set
2317                  * i_size already for this truncate so recovery will not
2318                  * expose the disk blocks we are discarding here.) */
2319                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2320                 JBUFFER_TRACE(jh, "on running transaction");
2321                 may_free = __dispose_buffer(jh, transaction);
2322         }
2323
2324 zap_buffer:
2325         /*
2326          * This is tricky. Although the buffer is truncated, it may be reused
2327          * if blocksize < pagesize and it is attached to the page straddling
2328          * EOF. Since the buffer might have been added to BJ_Forget list of the
2329          * running transaction, journal_get_write_access() won't clear
2330          * b_modified and credit accounting gets confused. So clear b_modified
2331          * here.
2332          */
2333         jh->b_modified = 0;
2334         jbd2_journal_put_journal_head(jh);
2335 zap_buffer_no_jh:
2336         spin_unlock(&journal->j_list_lock);
2337         jbd_unlock_bh_state(bh);
2338         write_unlock(&journal->j_state_lock);
2339 zap_buffer_unlocked:
2340         clear_buffer_dirty(bh);
2341         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2342         clear_buffer_mapped(bh);
2343         clear_buffer_req(bh);
2344         clear_buffer_new(bh);
2345         clear_buffer_delay(bh);
2346         clear_buffer_unwritten(bh);
2347         bh->b_bdev = NULL;
2348         return may_free;
2349 }
2350
2351 /**
2352  * void jbd2_journal_invalidatepage()
2353  * @journal: journal to use for flush...
2354  * @page:    page to flush
2355  * @offset:  start of the range to invalidate
2356  * @length:  length of the range to invalidate
2357  *
2358  * Reap page buffers containing data after in the specified range in page.
2359  * Can return -EBUSY if buffers are part of the committing transaction and
2360  * the page is straddling i_size. Caller then has to wait for current commit
2361  * and try again.
2362  */
2363 int jbd2_journal_invalidatepage(journal_t *journal,
2364                                 struct page *page,
2365                                 unsigned int offset,
2366                                 unsigned int length)
2367 {
2368         struct buffer_head *head, *bh, *next;
2369         unsigned int stop = offset + length;
2370         unsigned int curr_off = 0;
2371         int partial_page = (offset || length < PAGE_SIZE);
2372         int may_free = 1;
2373         int ret = 0;
2374
2375         if (!PageLocked(page))
2376                 BUG();
2377         if (!page_has_buffers(page))
2378                 return 0;
2379
2380         BUG_ON(stop > PAGE_SIZE || stop < length);
2381
2382         /* We will potentially be playing with lists other than just the
2383          * data lists (especially for journaled data mode), so be
2384          * cautious in our locking. */
2385
2386         head = bh = page_buffers(page);
2387         do {
2388                 unsigned int next_off = curr_off + bh->b_size;
2389                 next = bh->b_this_page;
2390
2391                 if (next_off > stop)
2392                         return 0;
2393
2394                 if (offset <= curr_off) {
2395                         /* This block is wholly outside the truncation point */
2396                         lock_buffer(bh);
2397                         ret = journal_unmap_buffer(journal, bh, partial_page);
2398                         unlock_buffer(bh);
2399                         if (ret < 0)
2400                                 return ret;
2401                         may_free &= ret;
2402                 }
2403                 curr_off = next_off;
2404                 bh = next;
2405
2406         } while (bh != head);
2407
2408         if (!partial_page) {
2409                 if (may_free && try_to_free_buffers(page))
2410                         J_ASSERT(!page_has_buffers(page));
2411         }
2412         return 0;
2413 }
2414
2415 /*
2416  * File a buffer on the given transaction list.
2417  */
2418 void __jbd2_journal_file_buffer(struct journal_head *jh,
2419                         transaction_t *transaction, int jlist)
2420 {
2421         struct journal_head **list = NULL;
2422         int was_dirty = 0;
2423         struct buffer_head *bh = jh2bh(jh);
2424
2425         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2426         assert_spin_locked(&transaction->t_journal->j_list_lock);
2427
2428         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2429         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2430                                 jh->b_transaction == NULL);
2431
2432         if (jh->b_transaction && jh->b_jlist == jlist)
2433                 return;
2434
2435         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2436             jlist == BJ_Shadow || jlist == BJ_Forget) {
2437                 /*
2438                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2439                  * instead of buffer_dirty. We should not see a dirty bit set
2440                  * here because we clear it in do_get_write_access but e.g.
2441                  * tune2fs can modify the sb and set the dirty bit at any time
2442                  * so we try to gracefully handle that.
2443                  */
2444                 if (buffer_dirty(bh))
2445                         warn_dirty_buffer(bh);
2446                 if (test_clear_buffer_dirty(bh) ||
2447                     test_clear_buffer_jbddirty(bh))
2448                         was_dirty = 1;
2449         }
2450
2451         if (jh->b_transaction)
2452                 __jbd2_journal_temp_unlink_buffer(jh);
2453         else
2454                 jbd2_journal_grab_journal_head(bh);
2455         jh->b_transaction = transaction;
2456
2457         switch (jlist) {
2458         case BJ_None:
2459                 J_ASSERT_JH(jh, !jh->b_committed_data);
2460                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2461                 return;
2462         case BJ_Metadata:
2463                 transaction->t_nr_buffers++;
2464                 list = &transaction->t_buffers;
2465                 break;
2466         case BJ_Forget:
2467                 list = &transaction->t_forget;
2468                 break;
2469         case BJ_Shadow:
2470                 list = &transaction->t_shadow_list;
2471                 break;
2472         case BJ_Reserved:
2473                 list = &transaction->t_reserved_list;
2474                 break;
2475         }
2476
2477         __blist_add_buffer(list, jh);
2478         jh->b_jlist = jlist;
2479
2480         if (was_dirty)
2481                 set_buffer_jbddirty(bh);
2482 }
2483
2484 void jbd2_journal_file_buffer(struct journal_head *jh,
2485                                 transaction_t *transaction, int jlist)
2486 {
2487         jbd_lock_bh_state(jh2bh(jh));
2488         spin_lock(&transaction->t_journal->j_list_lock);
2489         __jbd2_journal_file_buffer(jh, transaction, jlist);
2490         spin_unlock(&transaction->t_journal->j_list_lock);
2491         jbd_unlock_bh_state(jh2bh(jh));
2492 }
2493
2494 /*
2495  * Remove a buffer from its current buffer list in preparation for
2496  * dropping it from its current transaction entirely.  If the buffer has
2497  * already started to be used by a subsequent transaction, refile the
2498  * buffer on that transaction's metadata list.
2499  *
2500  * Called under j_list_lock
2501  * Called under jbd_lock_bh_state(jh2bh(jh))
2502  *
2503  * jh and bh may be already free when this function returns
2504  */
2505 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2506 {
2507         int was_dirty, jlist;
2508         struct buffer_head *bh = jh2bh(jh);
2509
2510         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2511         if (jh->b_transaction)
2512                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2513
2514         /* If the buffer is now unused, just drop it. */
2515         if (jh->b_next_transaction == NULL) {
2516                 __jbd2_journal_unfile_buffer(jh);
2517                 return;
2518         }
2519
2520         /*
2521          * It has been modified by a later transaction: add it to the new
2522          * transaction's metadata list.
2523          */
2524
2525         was_dirty = test_clear_buffer_jbddirty(bh);
2526         __jbd2_journal_temp_unlink_buffer(jh);
2527         /*
2528          * We set b_transaction here because b_next_transaction will inherit
2529          * our jh reference and thus __jbd2_journal_file_buffer() must not
2530          * take a new one.
2531          */
2532         jh->b_transaction = jh->b_next_transaction;
2533         jh->b_next_transaction = NULL;
2534         if (buffer_freed(bh))
2535                 jlist = BJ_Forget;
2536         else if (jh->b_modified)
2537                 jlist = BJ_Metadata;
2538         else
2539                 jlist = BJ_Reserved;
2540         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2541         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2542
2543         if (was_dirty)
2544                 set_buffer_jbddirty(bh);
2545 }
2546
2547 /*
2548  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2549  * bh reference so that we can safely unlock bh.
2550  *
2551  * The jh and bh may be freed by this call.
2552  */
2553 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2554 {
2555         struct buffer_head *bh = jh2bh(jh);
2556
2557         /* Get reference so that buffer cannot be freed before we unlock it */
2558         get_bh(bh);
2559         jbd_lock_bh_state(bh);
2560         spin_lock(&journal->j_list_lock);
2561         __jbd2_journal_refile_buffer(jh);
2562         jbd_unlock_bh_state(bh);
2563         spin_unlock(&journal->j_list_lock);
2564         __brelse(bh);
2565 }
2566
2567 /*
2568  * File inode in the inode list of the handle's transaction
2569  */
2570 static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2571                 unsigned long flags, loff_t start_byte, loff_t end_byte)
2572 {
2573         transaction_t *transaction = handle->h_transaction;
2574         journal_t *journal;
2575
2576         if (is_handle_aborted(handle))
2577                 return -EROFS;
2578         journal = transaction->t_journal;
2579
2580         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2581                         transaction->t_tid);
2582
2583         spin_lock(&journal->j_list_lock);
2584         jinode->i_flags |= flags;
2585
2586         if (jinode->i_dirty_end) {
2587                 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2588                 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2589         } else {
2590                 jinode->i_dirty_start = start_byte;
2591                 jinode->i_dirty_end = end_byte;
2592         }
2593
2594         /* Is inode already attached where we need it? */
2595         if (jinode->i_transaction == transaction ||
2596             jinode->i_next_transaction == transaction)
2597                 goto done;
2598
2599         /*
2600          * We only ever set this variable to 1 so the test is safe. Since
2601          * t_need_data_flush is likely to be set, we do the test to save some
2602          * cacheline bouncing
2603          */
2604         if (!transaction->t_need_data_flush)
2605                 transaction->t_need_data_flush = 1;
2606         /* On some different transaction's list - should be
2607          * the committing one */
2608         if (jinode->i_transaction) {
2609                 J_ASSERT(jinode->i_next_transaction == NULL);
2610                 J_ASSERT(jinode->i_transaction ==
2611                                         journal->j_committing_transaction);
2612                 jinode->i_next_transaction = transaction;
2613                 goto done;
2614         }
2615         /* Not on any transaction list... */
2616         J_ASSERT(!jinode->i_next_transaction);
2617         jinode->i_transaction = transaction;
2618         list_add(&jinode->i_list, &transaction->t_inode_list);
2619 done:
2620         spin_unlock(&journal->j_list_lock);
2621
2622         return 0;
2623 }
2624
2625 int jbd2_journal_inode_ranged_write(handle_t *handle,
2626                 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2627 {
2628         return jbd2_journal_file_inode(handle, jinode,
2629                         JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2630                         start_byte + length - 1);
2631 }
2632
2633 int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2634                 loff_t start_byte, loff_t length)
2635 {
2636         return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2637                         start_byte, start_byte + length - 1);
2638 }
2639
2640 /*
2641  * File truncate and transaction commit interact with each other in a
2642  * non-trivial way.  If a transaction writing data block A is
2643  * committing, we cannot discard the data by truncate until we have
2644  * written them.  Otherwise if we crashed after the transaction with
2645  * write has committed but before the transaction with truncate has
2646  * committed, we could see stale data in block A.  This function is a
2647  * helper to solve this problem.  It starts writeout of the truncated
2648  * part in case it is in the committing transaction.
2649  *
2650  * Filesystem code must call this function when inode is journaled in
2651  * ordered mode before truncation happens and after the inode has been
2652  * placed on orphan list with the new inode size. The second condition
2653  * avoids the race that someone writes new data and we start
2654  * committing the transaction after this function has been called but
2655  * before a transaction for truncate is started (and furthermore it
2656  * allows us to optimize the case where the addition to orphan list
2657  * happens in the same transaction as write --- we don't have to write
2658  * any data in such case).
2659  */
2660 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2661                                         struct jbd2_inode *jinode,
2662                                         loff_t new_size)
2663 {
2664         transaction_t *inode_trans, *commit_trans;
2665         int ret = 0;
2666
2667         /* This is a quick check to avoid locking if not necessary */
2668         if (!jinode->i_transaction)
2669                 goto out;
2670         /* Locks are here just to force reading of recent values, it is
2671          * enough that the transaction was not committing before we started
2672          * a transaction adding the inode to orphan list */
2673         read_lock(&journal->j_state_lock);
2674         commit_trans = journal->j_committing_transaction;
2675         read_unlock(&journal->j_state_lock);
2676         spin_lock(&journal->j_list_lock);
2677         inode_trans = jinode->i_transaction;
2678         spin_unlock(&journal->j_list_lock);
2679         if (inode_trans == commit_trans) {
2680                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2681                         new_size, LLONG_MAX);
2682                 if (ret)
2683                         jbd2_journal_abort(journal, ret);
2684         }
2685 out:
2686         return ret;
2687 }