GNU Linux-libre 4.19.211-gnu1
[releases.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 allocate and initialise a new transaction.  Create 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 transaction_t *
81 jbd2_get_transaction(journal_t *journal, 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         return transaction;
107 }
108
109 /*
110  * Handle management.
111  *
112  * A handle_t is an object which represents a single atomic update to a
113  * filesystem, and which tracks all of the modifications which form part
114  * of that one update.
115  */
116
117 /*
118  * Update transaction's maximum wait time, if debugging is enabled.
119  *
120  * In order for t_max_wait to be reliable, it must be protected by a
121  * lock.  But doing so will mean that start_this_handle() can not be
122  * run in parallel on SMP systems, which limits our scalability.  So
123  * unless debugging is enabled, we no longer update t_max_wait, which
124  * means that maximum wait time reported by the jbd2_run_stats
125  * tracepoint will always be zero.
126  */
127 static inline void update_t_max_wait(transaction_t *transaction,
128                                      unsigned long ts)
129 {
130 #ifdef CONFIG_JBD2_DEBUG
131         if (jbd2_journal_enable_debug &&
132             time_after(transaction->t_start, ts)) {
133                 ts = jbd2_time_diff(ts, transaction->t_start);
134                 spin_lock(&transaction->t_handle_lock);
135                 if (ts > transaction->t_max_wait)
136                         transaction->t_max_wait = ts;
137                 spin_unlock(&transaction->t_handle_lock);
138         }
139 #endif
140 }
141
142 /*
143  * Wait until running transaction passes T_LOCKED state. Also starts the commit
144  * if needed. The function expects running transaction to exist and releases
145  * j_state_lock.
146  */
147 static void wait_transaction_locked(journal_t *journal)
148         __releases(journal->j_state_lock)
149 {
150         DEFINE_WAIT(wait);
151         int need_to_start;
152         tid_t tid = journal->j_running_transaction->t_tid;
153
154         prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
155                         TASK_UNINTERRUPTIBLE);
156         need_to_start = !tid_geq(journal->j_commit_request, tid);
157         read_unlock(&journal->j_state_lock);
158         if (need_to_start)
159                 jbd2_log_start_commit(journal, tid);
160         jbd2_might_wait_for_commit(journal);
161         schedule();
162         finish_wait(&journal->j_wait_transaction_locked, &wait);
163 }
164
165 static void sub_reserved_credits(journal_t *journal, int blocks)
166 {
167         atomic_sub(blocks, &journal->j_reserved_credits);
168         wake_up(&journal->j_wait_reserved);
169 }
170
171 /*
172  * Wait until we can add credits for handle to the running transaction.  Called
173  * with j_state_lock held for reading. Returns 0 if handle joined the running
174  * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
175  * caller must retry.
176  */
177 static int add_transaction_credits(journal_t *journal, int blocks,
178                                    int rsv_blocks)
179 {
180         transaction_t *t = journal->j_running_transaction;
181         int needed;
182         int total = blocks + rsv_blocks;
183
184         /*
185          * If the current transaction is locked down for commit, wait
186          * for the lock to be released.
187          */
188         if (t->t_state == T_LOCKED) {
189                 wait_transaction_locked(journal);
190                 return 1;
191         }
192
193         /*
194          * If there is not enough space left in the log to write all
195          * potential buffers requested by this operation, we need to
196          * stall pending a log checkpoint to free some more log space.
197          */
198         needed = atomic_add_return(total, &t->t_outstanding_credits);
199         if (needed > journal->j_max_transaction_buffers) {
200                 /*
201                  * If the current transaction is already too large,
202                  * then start to commit it: we can then go back and
203                  * attach this handle to a new transaction.
204                  */
205                 atomic_sub(total, &t->t_outstanding_credits);
206
207                 /*
208                  * Is the number of reserved credits in the current transaction too
209                  * big to fit this handle? Wait until reserved credits are freed.
210                  */
211                 if (atomic_read(&journal->j_reserved_credits) + total >
212                     journal->j_max_transaction_buffers) {
213                         read_unlock(&journal->j_state_lock);
214                         jbd2_might_wait_for_commit(journal);
215                         wait_event(journal->j_wait_reserved,
216                                    atomic_read(&journal->j_reserved_credits) + total <=
217                                    journal->j_max_transaction_buffers);
218                         return 1;
219                 }
220
221                 wait_transaction_locked(journal);
222                 return 1;
223         }
224
225         /*
226          * The commit code assumes that it can get enough log space
227          * without forcing a checkpoint.  This is *critical* for
228          * correctness: a checkpoint of a buffer which is also
229          * associated with a committing transaction creates a deadlock,
230          * so commit simply cannot force through checkpoints.
231          *
232          * We must therefore ensure the necessary space in the journal
233          * *before* starting to dirty potentially checkpointed buffers
234          * in the new transaction.
235          */
236         if (jbd2_log_space_left(journal) < jbd2_space_needed(journal)) {
237                 atomic_sub(total, &t->t_outstanding_credits);
238                 read_unlock(&journal->j_state_lock);
239                 jbd2_might_wait_for_commit(journal);
240                 write_lock(&journal->j_state_lock);
241                 if (jbd2_log_space_left(journal) < jbd2_space_needed(journal))
242                         __jbd2_log_wait_for_space(journal);
243                 write_unlock(&journal->j_state_lock);
244                 return 1;
245         }
246
247         /* No reservation? We are done... */
248         if (!rsv_blocks)
249                 return 0;
250
251         needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
252         /* We allow at most half of a transaction to be reserved */
253         if (needed > journal->j_max_transaction_buffers / 2) {
254                 sub_reserved_credits(journal, rsv_blocks);
255                 atomic_sub(total, &t->t_outstanding_credits);
256                 read_unlock(&journal->j_state_lock);
257                 jbd2_might_wait_for_commit(journal);
258                 wait_event(journal->j_wait_reserved,
259                          atomic_read(&journal->j_reserved_credits) + rsv_blocks
260                          <= journal->j_max_transaction_buffers / 2);
261                 return 1;
262         }
263         return 0;
264 }
265
266 /*
267  * start_this_handle: Given a handle, deal with any locking or stalling
268  * needed to make sure that there is enough journal space for the handle
269  * to begin.  Attach the handle to a transaction and set up the
270  * transaction's buffer credits.
271  */
272
273 static int start_this_handle(journal_t *journal, handle_t *handle,
274                              gfp_t gfp_mask)
275 {
276         transaction_t   *transaction, *new_transaction = NULL;
277         int             blocks = handle->h_buffer_credits;
278         int             rsv_blocks = 0;
279         unsigned long ts = jiffies;
280
281         if (handle->h_rsv_handle)
282                 rsv_blocks = handle->h_rsv_handle->h_buffer_credits;
283
284         /*
285          * Limit the number of reserved credits to 1/2 of maximum transaction
286          * size and limit the number of total credits to not exceed maximum
287          * transaction size per operation.
288          */
289         if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
290             (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
291                 printk(KERN_ERR "JBD2: %s wants too many credits "
292                        "credits:%d rsv_credits:%d max:%d\n",
293                        current->comm, blocks, rsv_blocks,
294                        journal->j_max_transaction_buffers);
295                 WARN_ON(1);
296                 return -ENOSPC;
297         }
298
299 alloc_transaction:
300         if (!journal->j_running_transaction) {
301                 /*
302                  * If __GFP_FS is not present, then we may be being called from
303                  * inside the fs writeback layer, so we MUST NOT fail.
304                  */
305                 if ((gfp_mask & __GFP_FS) == 0)
306                         gfp_mask |= __GFP_NOFAIL;
307                 new_transaction = kmem_cache_zalloc(transaction_cache,
308                                                     gfp_mask);
309                 if (!new_transaction)
310                         return -ENOMEM;
311         }
312
313         jbd_debug(3, "New handle %p going live.\n", handle);
314
315         /*
316          * We need to hold j_state_lock until t_updates has been incremented,
317          * for proper journal barrier handling
318          */
319 repeat:
320         read_lock(&journal->j_state_lock);
321         BUG_ON(journal->j_flags & JBD2_UNMOUNT);
322         if (is_journal_aborted(journal) ||
323             (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
324                 read_unlock(&journal->j_state_lock);
325                 jbd2_journal_free_transaction(new_transaction);
326                 return -EROFS;
327         }
328
329         /*
330          * Wait on the journal's transaction barrier if necessary. Specifically
331          * we allow reserved handles to proceed because otherwise commit could
332          * deadlock on page writeback not being able to complete.
333          */
334         if (!handle->h_reserved && journal->j_barrier_count) {
335                 read_unlock(&journal->j_state_lock);
336                 wait_event(journal->j_wait_transaction_locked,
337                                 journal->j_barrier_count == 0);
338                 goto repeat;
339         }
340
341         if (!journal->j_running_transaction) {
342                 read_unlock(&journal->j_state_lock);
343                 if (!new_transaction)
344                         goto alloc_transaction;
345                 write_lock(&journal->j_state_lock);
346                 if (!journal->j_running_transaction &&
347                     (handle->h_reserved || !journal->j_barrier_count)) {
348                         jbd2_get_transaction(journal, new_transaction);
349                         new_transaction = NULL;
350                 }
351                 write_unlock(&journal->j_state_lock);
352                 goto repeat;
353         }
354
355         transaction = journal->j_running_transaction;
356
357         if (!handle->h_reserved) {
358                 /* We may have dropped j_state_lock - restart in that case */
359                 if (add_transaction_credits(journal, blocks, rsv_blocks))
360                         goto repeat;
361         } else {
362                 /*
363                  * We have handle reserved so we are allowed to join T_LOCKED
364                  * transaction and we don't have to check for transaction size
365                  * and journal space.
366                  */
367                 sub_reserved_credits(journal, blocks);
368                 handle->h_reserved = 0;
369         }
370
371         /* OK, account for the buffers that this operation expects to
372          * use and add the handle to the running transaction. 
373          */
374         update_t_max_wait(transaction, ts);
375         handle->h_transaction = transaction;
376         handle->h_requested_credits = blocks;
377         handle->h_start_jiffies = jiffies;
378         atomic_inc(&transaction->t_updates);
379         atomic_inc(&transaction->t_handle_count);
380         jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
381                   handle, blocks,
382                   atomic_read(&transaction->t_outstanding_credits),
383                   jbd2_log_space_left(journal));
384         read_unlock(&journal->j_state_lock);
385         current->journal_info = handle;
386
387         rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
388         jbd2_journal_free_transaction(new_transaction);
389         /*
390          * Ensure that no allocations done while the transaction is open are
391          * going to recurse back to the fs layer.
392          */
393         handle->saved_alloc_context = memalloc_nofs_save();
394         return 0;
395 }
396
397 /* Allocate a new handle.  This should probably be in a slab... */
398 static handle_t *new_handle(int nblocks)
399 {
400         handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
401         if (!handle)
402                 return NULL;
403         handle->h_buffer_credits = nblocks;
404         handle->h_ref = 1;
405
406         return handle;
407 }
408
409 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
410                               gfp_t gfp_mask, unsigned int type,
411                               unsigned int line_no)
412 {
413         handle_t *handle = journal_current_handle();
414         int err;
415
416         if (!journal)
417                 return ERR_PTR(-EROFS);
418
419         if (handle) {
420                 J_ASSERT(handle->h_transaction->t_journal == journal);
421                 handle->h_ref++;
422                 return handle;
423         }
424
425         handle = new_handle(nblocks);
426         if (!handle)
427                 return ERR_PTR(-ENOMEM);
428         if (rsv_blocks) {
429                 handle_t *rsv_handle;
430
431                 rsv_handle = new_handle(rsv_blocks);
432                 if (!rsv_handle) {
433                         jbd2_free_handle(handle);
434                         return ERR_PTR(-ENOMEM);
435                 }
436                 rsv_handle->h_reserved = 1;
437                 rsv_handle->h_journal = journal;
438                 handle->h_rsv_handle = rsv_handle;
439         }
440
441         err = start_this_handle(journal, handle, gfp_mask);
442         if (err < 0) {
443                 if (handle->h_rsv_handle)
444                         jbd2_free_handle(handle->h_rsv_handle);
445                 jbd2_free_handle(handle);
446                 return ERR_PTR(err);
447         }
448         handle->h_type = type;
449         handle->h_line_no = line_no;
450         trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
451                                 handle->h_transaction->t_tid, type,
452                                 line_no, nblocks);
453
454         return handle;
455 }
456 EXPORT_SYMBOL(jbd2__journal_start);
457
458
459 /**
460  * handle_t *jbd2_journal_start() - Obtain a new handle.
461  * @journal: Journal to start transaction on.
462  * @nblocks: number of block buffer we might modify
463  *
464  * We make sure that the transaction can guarantee at least nblocks of
465  * modified buffers in the log.  We block until the log can guarantee
466  * that much space. Additionally, if rsv_blocks > 0, we also create another
467  * handle with rsv_blocks reserved blocks in the journal. This handle is
468  * is stored in h_rsv_handle. It is not attached to any particular transaction
469  * and thus doesn't block transaction commit. If the caller uses this reserved
470  * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
471  * on the parent handle will dispose the reserved one. Reserved handle has to
472  * be converted to a normal handle using jbd2_journal_start_reserved() before
473  * it can be used.
474  *
475  * Return a pointer to a newly allocated handle, or an ERR_PTR() value
476  * on failure.
477  */
478 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
479 {
480         return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0);
481 }
482 EXPORT_SYMBOL(jbd2_journal_start);
483
484 void jbd2_journal_free_reserved(handle_t *handle)
485 {
486         journal_t *journal = handle->h_journal;
487
488         WARN_ON(!handle->h_reserved);
489         sub_reserved_credits(journal, handle->h_buffer_credits);
490         jbd2_free_handle(handle);
491 }
492 EXPORT_SYMBOL(jbd2_journal_free_reserved);
493
494 /**
495  * int jbd2_journal_start_reserved() - start reserved handle
496  * @handle: handle to start
497  * @type: for handle statistics
498  * @line_no: for handle statistics
499  *
500  * Start handle that has been previously reserved with jbd2_journal_reserve().
501  * This attaches @handle to the running transaction (or creates one if there's
502  * not transaction running). Unlike jbd2_journal_start() this function cannot
503  * block on journal commit, checkpointing, or similar stuff. It can block on
504  * memory allocation or frozen journal though.
505  *
506  * Return 0 on success, non-zero on error - handle is freed in that case.
507  */
508 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
509                                 unsigned int line_no)
510 {
511         journal_t *journal = handle->h_journal;
512         int ret = -EIO;
513
514         if (WARN_ON(!handle->h_reserved)) {
515                 /* Someone passed in normal handle? Just stop it. */
516                 jbd2_journal_stop(handle);
517                 return ret;
518         }
519         /*
520          * Usefulness of mixing of reserved and unreserved handles is
521          * questionable. So far nobody seems to need it so just error out.
522          */
523         if (WARN_ON(current->journal_info)) {
524                 jbd2_journal_free_reserved(handle);
525                 return ret;
526         }
527
528         handle->h_journal = NULL;
529         /*
530          * GFP_NOFS is here because callers are likely from writeback or
531          * similarly constrained call sites
532          */
533         ret = start_this_handle(journal, handle, GFP_NOFS);
534         if (ret < 0) {
535                 handle->h_journal = journal;
536                 jbd2_journal_free_reserved(handle);
537                 return ret;
538         }
539         handle->h_type = type;
540         handle->h_line_no = line_no;
541         return 0;
542 }
543 EXPORT_SYMBOL(jbd2_journal_start_reserved);
544
545 /**
546  * int jbd2_journal_extend() - extend buffer credits.
547  * @handle:  handle to 'extend'
548  * @nblocks: nr blocks to try to extend by.
549  *
550  * Some transactions, such as large extends and truncates, can be done
551  * atomically all at once or in several stages.  The operation requests
552  * a credit for a number of buffer modifications in advance, but can
553  * extend its credit if it needs more.
554  *
555  * jbd2_journal_extend tries to give the running handle more buffer credits.
556  * It does not guarantee that allocation - this is a best-effort only.
557  * The calling process MUST be able to deal cleanly with a failure to
558  * extend here.
559  *
560  * Return 0 on success, non-zero on failure.
561  *
562  * return code < 0 implies an error
563  * return code > 0 implies normal transaction-full status.
564  */
565 int jbd2_journal_extend(handle_t *handle, int nblocks)
566 {
567         transaction_t *transaction = handle->h_transaction;
568         journal_t *journal;
569         int result;
570         int wanted;
571
572         if (is_handle_aborted(handle))
573                 return -EROFS;
574         journal = transaction->t_journal;
575
576         result = 1;
577
578         read_lock(&journal->j_state_lock);
579
580         /* Don't extend a locked-down transaction! */
581         if (transaction->t_state != T_RUNNING) {
582                 jbd_debug(3, "denied handle %p %d blocks: "
583                           "transaction not running\n", handle, nblocks);
584                 goto error_out;
585         }
586
587         spin_lock(&transaction->t_handle_lock);
588         wanted = atomic_add_return(nblocks,
589                                    &transaction->t_outstanding_credits);
590
591         if (wanted > journal->j_max_transaction_buffers) {
592                 jbd_debug(3, "denied handle %p %d blocks: "
593                           "transaction too large\n", handle, nblocks);
594                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
595                 goto unlock;
596         }
597
598         if (wanted + (wanted >> JBD2_CONTROL_BLOCKS_SHIFT) >
599             jbd2_log_space_left(journal)) {
600                 jbd_debug(3, "denied handle %p %d blocks: "
601                           "insufficient log space\n", handle, nblocks);
602                 atomic_sub(nblocks, &transaction->t_outstanding_credits);
603                 goto unlock;
604         }
605
606         trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
607                                  transaction->t_tid,
608                                  handle->h_type, handle->h_line_no,
609                                  handle->h_buffer_credits,
610                                  nblocks);
611
612         handle->h_buffer_credits += nblocks;
613         handle->h_requested_credits += nblocks;
614         result = 0;
615
616         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
617 unlock:
618         spin_unlock(&transaction->t_handle_lock);
619 error_out:
620         read_unlock(&journal->j_state_lock);
621         return result;
622 }
623
624
625 /**
626  * int jbd2_journal_restart() - restart a handle .
627  * @handle:  handle to restart
628  * @nblocks: nr credits requested
629  * @gfp_mask: memory allocation flags (for start_this_handle)
630  *
631  * Restart a handle for a multi-transaction filesystem
632  * operation.
633  *
634  * If the jbd2_journal_extend() call above fails to grant new buffer credits
635  * to a running handle, a call to jbd2_journal_restart will commit the
636  * handle's transaction so far and reattach the handle to a new
637  * transaction capable of guaranteeing the requested number of
638  * credits. We preserve reserved handle if there's any attached to the
639  * passed in handle.
640  */
641 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
642 {
643         transaction_t *transaction = handle->h_transaction;
644         journal_t *journal;
645         tid_t           tid;
646         int             need_to_start, ret;
647
648         /* If we've had an abort of any type, don't even think about
649          * actually doing the restart! */
650         if (is_handle_aborted(handle))
651                 return 0;
652         journal = transaction->t_journal;
653
654         /*
655          * First unlink the handle from its current transaction, and start the
656          * commit on that.
657          */
658         J_ASSERT(atomic_read(&transaction->t_updates) > 0);
659         J_ASSERT(journal_current_handle() == handle);
660
661         read_lock(&journal->j_state_lock);
662         spin_lock(&transaction->t_handle_lock);
663         atomic_sub(handle->h_buffer_credits,
664                    &transaction->t_outstanding_credits);
665         if (handle->h_rsv_handle) {
666                 sub_reserved_credits(journal,
667                                      handle->h_rsv_handle->h_buffer_credits);
668         }
669         if (atomic_dec_and_test(&transaction->t_updates))
670                 wake_up(&journal->j_wait_updates);
671         tid = transaction->t_tid;
672         spin_unlock(&transaction->t_handle_lock);
673         handle->h_transaction = NULL;
674         current->journal_info = NULL;
675
676         jbd_debug(2, "restarting handle %p\n", handle);
677         need_to_start = !tid_geq(journal->j_commit_request, tid);
678         read_unlock(&journal->j_state_lock);
679         if (need_to_start)
680                 jbd2_log_start_commit(journal, tid);
681
682         rwsem_release(&journal->j_trans_commit_map, 1, _THIS_IP_);
683         handle->h_buffer_credits = nblocks;
684         /*
685          * Restore the original nofs context because the journal restart
686          * is basically the same thing as journal stop and start.
687          * start_this_handle will start a new nofs context.
688          */
689         memalloc_nofs_restore(handle->saved_alloc_context);
690         ret = start_this_handle(journal, handle, gfp_mask);
691         return ret;
692 }
693 EXPORT_SYMBOL(jbd2__journal_restart);
694
695
696 int jbd2_journal_restart(handle_t *handle, int nblocks)
697 {
698         return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
699 }
700 EXPORT_SYMBOL(jbd2_journal_restart);
701
702 /**
703  * void jbd2_journal_lock_updates () - establish a transaction barrier.
704  * @journal:  Journal to establish a barrier on.
705  *
706  * This locks out any further updates from being started, and blocks
707  * until all existing updates have completed, returning only once the
708  * journal is in a quiescent state with no updates running.
709  *
710  * The journal lock should not be held on entry.
711  */
712 void jbd2_journal_lock_updates(journal_t *journal)
713 {
714         DEFINE_WAIT(wait);
715
716         jbd2_might_wait_for_commit(journal);
717
718         write_lock(&journal->j_state_lock);
719         ++journal->j_barrier_count;
720
721         /* Wait until there are no reserved handles */
722         if (atomic_read(&journal->j_reserved_credits)) {
723                 write_unlock(&journal->j_state_lock);
724                 wait_event(journal->j_wait_reserved,
725                            atomic_read(&journal->j_reserved_credits) == 0);
726                 write_lock(&journal->j_state_lock);
727         }
728
729         /* Wait until there are no running updates */
730         while (1) {
731                 transaction_t *transaction = journal->j_running_transaction;
732
733                 if (!transaction)
734                         break;
735
736                 spin_lock(&transaction->t_handle_lock);
737                 prepare_to_wait(&journal->j_wait_updates, &wait,
738                                 TASK_UNINTERRUPTIBLE);
739                 if (!atomic_read(&transaction->t_updates)) {
740                         spin_unlock(&transaction->t_handle_lock);
741                         finish_wait(&journal->j_wait_updates, &wait);
742                         break;
743                 }
744                 spin_unlock(&transaction->t_handle_lock);
745                 write_unlock(&journal->j_state_lock);
746                 schedule();
747                 finish_wait(&journal->j_wait_updates, &wait);
748                 write_lock(&journal->j_state_lock);
749         }
750         write_unlock(&journal->j_state_lock);
751
752         /*
753          * We have now established a barrier against other normal updates, but
754          * we also need to barrier against other jbd2_journal_lock_updates() calls
755          * to make sure that we serialise special journal-locked operations
756          * too.
757          */
758         mutex_lock(&journal->j_barrier);
759 }
760
761 /**
762  * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
763  * @journal:  Journal to release the barrier on.
764  *
765  * Release a transaction barrier obtained with jbd2_journal_lock_updates().
766  *
767  * Should be called without the journal lock held.
768  */
769 void jbd2_journal_unlock_updates (journal_t *journal)
770 {
771         J_ASSERT(journal->j_barrier_count != 0);
772
773         mutex_unlock(&journal->j_barrier);
774         write_lock(&journal->j_state_lock);
775         --journal->j_barrier_count;
776         write_unlock(&journal->j_state_lock);
777         wake_up(&journal->j_wait_transaction_locked);
778 }
779
780 static void warn_dirty_buffer(struct buffer_head *bh)
781 {
782         printk(KERN_WARNING
783                "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
784                "There's a risk of filesystem corruption in case of system "
785                "crash.\n",
786                bh->b_bdev, (unsigned long long)bh->b_blocknr);
787 }
788
789 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
790 static void jbd2_freeze_jh_data(struct journal_head *jh)
791 {
792         struct page *page;
793         int offset;
794         char *source;
795         struct buffer_head *bh = jh2bh(jh);
796
797         J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
798         page = bh->b_page;
799         offset = offset_in_page(bh->b_data);
800         source = kmap_atomic(page);
801         /* Fire data frozen trigger just before we copy the data */
802         jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
803         memcpy(jh->b_frozen_data, source + offset, bh->b_size);
804         kunmap_atomic(source);
805
806         /*
807          * Now that the frozen data is saved off, we need to store any matching
808          * triggers.
809          */
810         jh->b_frozen_triggers = jh->b_triggers;
811 }
812
813 /*
814  * If the buffer is already part of the current transaction, then there
815  * is nothing we need to do.  If it is already part of a prior
816  * transaction which we are still committing to disk, then we need to
817  * make sure that we do not overwrite the old copy: we do copy-out to
818  * preserve the copy going to disk.  We also account the buffer against
819  * the handle's metadata buffer credits (unless the buffer is already
820  * part of the transaction, that is).
821  *
822  */
823 static int
824 do_get_write_access(handle_t *handle, struct journal_head *jh,
825                         int force_copy)
826 {
827         struct buffer_head *bh;
828         transaction_t *transaction = handle->h_transaction;
829         journal_t *journal;
830         int error;
831         char *frozen_buffer = NULL;
832         unsigned long start_lock, time_lock;
833
834         journal = transaction->t_journal;
835
836         jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
837
838         JBUFFER_TRACE(jh, "entry");
839 repeat:
840         bh = jh2bh(jh);
841
842         /* @@@ Need to check for errors here at some point. */
843
844         start_lock = jiffies;
845         lock_buffer(bh);
846         jbd_lock_bh_state(bh);
847
848         /* If it takes too long to lock the buffer, trace it */
849         time_lock = jbd2_time_diff(start_lock, jiffies);
850         if (time_lock > HZ/10)
851                 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
852                         jiffies_to_msecs(time_lock));
853
854         /* We now hold the buffer lock so it is safe to query the buffer
855          * state.  Is the buffer dirty?
856          *
857          * If so, there are two possibilities.  The buffer may be
858          * non-journaled, and undergoing a quite legitimate writeback.
859          * Otherwise, it is journaled, and we don't expect dirty buffers
860          * in that state (the buffers should be marked JBD_Dirty
861          * instead.)  So either the IO is being done under our own
862          * control and this is a bug, or it's a third party IO such as
863          * dump(8) (which may leave the buffer scheduled for read ---
864          * ie. locked but not dirty) or tune2fs (which may actually have
865          * the buffer dirtied, ugh.)  */
866
867         if (buffer_dirty(bh)) {
868                 /*
869                  * First question: is this buffer already part of the current
870                  * transaction or the existing committing transaction?
871                  */
872                 if (jh->b_transaction) {
873                         J_ASSERT_JH(jh,
874                                 jh->b_transaction == transaction ||
875                                 jh->b_transaction ==
876                                         journal->j_committing_transaction);
877                         if (jh->b_next_transaction)
878                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
879                                                         transaction);
880                         warn_dirty_buffer(bh);
881                 }
882                 /*
883                  * In any case we need to clean the dirty flag and we must
884                  * do it under the buffer lock to be sure we don't race
885                  * with running write-out.
886                  */
887                 JBUFFER_TRACE(jh, "Journalling dirty buffer");
888                 clear_buffer_dirty(bh);
889                 set_buffer_jbddirty(bh);
890         }
891
892         unlock_buffer(bh);
893
894         error = -EROFS;
895         if (is_handle_aborted(handle)) {
896                 jbd_unlock_bh_state(bh);
897                 goto out;
898         }
899         error = 0;
900
901         /*
902          * The buffer is already part of this transaction if b_transaction or
903          * b_next_transaction points to it
904          */
905         if (jh->b_transaction == transaction ||
906             jh->b_next_transaction == transaction)
907                 goto done;
908
909         /*
910          * this is the first time this transaction is touching this buffer,
911          * reset the modified flag
912          */
913        jh->b_modified = 0;
914
915         /*
916          * If the buffer is not journaled right now, we need to make sure it
917          * doesn't get written to disk before the caller actually commits the
918          * new data
919          */
920         if (!jh->b_transaction) {
921                 JBUFFER_TRACE(jh, "no transaction");
922                 J_ASSERT_JH(jh, !jh->b_next_transaction);
923                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
924                 /*
925                  * Make sure all stores to jh (b_modified, b_frozen_data) are
926                  * visible before attaching it to the running transaction.
927                  * Paired with barrier in jbd2_write_access_granted()
928                  */
929                 smp_wmb();
930                 spin_lock(&journal->j_list_lock);
931                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
932                 spin_unlock(&journal->j_list_lock);
933                 goto done;
934         }
935         /*
936          * If there is already a copy-out version of this buffer, then we don't
937          * need to make another one
938          */
939         if (jh->b_frozen_data) {
940                 JBUFFER_TRACE(jh, "has frozen data");
941                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
942                 goto attach_next;
943         }
944
945         JBUFFER_TRACE(jh, "owned by older transaction");
946         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
947         J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
948
949         /*
950          * There is one case we have to be very careful about.  If the
951          * committing transaction is currently writing this buffer out to disk
952          * and has NOT made a copy-out, then we cannot modify the buffer
953          * contents at all right now.  The essence of copy-out is that it is
954          * the extra copy, not the primary copy, which gets journaled.  If the
955          * primary copy is already going to disk then we cannot do copy-out
956          * here.
957          */
958         if (buffer_shadow(bh)) {
959                 JBUFFER_TRACE(jh, "on shadow: sleep");
960                 jbd_unlock_bh_state(bh);
961                 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
962                 goto repeat;
963         }
964
965         /*
966          * Only do the copy if the currently-owning transaction still needs it.
967          * If buffer isn't on BJ_Metadata list, the committing transaction is
968          * past that stage (here we use the fact that BH_Shadow is set under
969          * bh_state lock together with refiling to BJ_Shadow list and at this
970          * point we know the buffer doesn't have BH_Shadow set).
971          *
972          * Subtle point, though: if this is a get_undo_access, then we will be
973          * relying on the frozen_data to contain the new value of the
974          * committed_data record after the transaction, so we HAVE to force the
975          * frozen_data copy in that case.
976          */
977         if (jh->b_jlist == BJ_Metadata || force_copy) {
978                 JBUFFER_TRACE(jh, "generate frozen data");
979                 if (!frozen_buffer) {
980                         JBUFFER_TRACE(jh, "allocate memory for buffer");
981                         jbd_unlock_bh_state(bh);
982                         frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
983                                                    GFP_NOFS | __GFP_NOFAIL);
984                         goto repeat;
985                 }
986                 jh->b_frozen_data = frozen_buffer;
987                 frozen_buffer = NULL;
988                 jbd2_freeze_jh_data(jh);
989         }
990 attach_next:
991         /*
992          * Make sure all stores to jh (b_modified, b_frozen_data) are visible
993          * before attaching it to the running transaction. Paired with barrier
994          * in jbd2_write_access_granted()
995          */
996         smp_wmb();
997         jh->b_next_transaction = transaction;
998
999 done:
1000         jbd_unlock_bh_state(bh);
1001
1002         /*
1003          * If we are about to journal a buffer, then any revoke pending on it is
1004          * no longer valid
1005          */
1006         jbd2_journal_cancel_revoke(handle, jh);
1007
1008 out:
1009         if (unlikely(frozen_buffer))    /* It's usually NULL */
1010                 jbd2_free(frozen_buffer, bh->b_size);
1011
1012         JBUFFER_TRACE(jh, "exit");
1013         return error;
1014 }
1015
1016 /* Fast check whether buffer is already attached to the required transaction */
1017 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1018                                                         bool undo)
1019 {
1020         struct journal_head *jh;
1021         bool ret = false;
1022
1023         /* Dirty buffers require special handling... */
1024         if (buffer_dirty(bh))
1025                 return false;
1026
1027         /*
1028          * RCU protects us from dereferencing freed pages. So the checks we do
1029          * are guaranteed not to oops. However the jh slab object can get freed
1030          * & reallocated while we work with it. So we have to be careful. When
1031          * we see jh attached to the running transaction, we know it must stay
1032          * so until the transaction is committed. Thus jh won't be freed and
1033          * will be attached to the same bh while we run.  However it can
1034          * happen jh gets freed, reallocated, and attached to the transaction
1035          * just after we get pointer to it from bh. So we have to be careful
1036          * and recheck jh still belongs to our bh before we return success.
1037          */
1038         rcu_read_lock();
1039         if (!buffer_jbd(bh))
1040                 goto out;
1041         /* This should be bh2jh() but that doesn't work with inline functions */
1042         jh = READ_ONCE(bh->b_private);
1043         if (!jh)
1044                 goto out;
1045         /* For undo access buffer must have data copied */
1046         if (undo && !jh->b_committed_data)
1047                 goto out;
1048         if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1049             READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1050                 goto out;
1051         /*
1052          * There are two reasons for the barrier here:
1053          * 1) Make sure to fetch b_bh after we did previous checks so that we
1054          * detect when jh went through free, realloc, attach to transaction
1055          * while we were checking. Paired with implicit barrier in that path.
1056          * 2) So that access to bh done after jbd2_write_access_granted()
1057          * doesn't get reordered and see inconsistent state of concurrent
1058          * do_get_write_access().
1059          */
1060         smp_mb();
1061         if (unlikely(jh->b_bh != bh))
1062                 goto out;
1063         ret = true;
1064 out:
1065         rcu_read_unlock();
1066         return ret;
1067 }
1068
1069 /**
1070  * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1071  * @handle: transaction to add buffer modifications to
1072  * @bh:     bh to be used for metadata writes
1073  *
1074  * Returns: error code or 0 on success.
1075  *
1076  * In full data journalling mode the buffer may be of type BJ_AsyncData,
1077  * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1078  */
1079
1080 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1081 {
1082         struct journal_head *jh;
1083         int rc;
1084
1085         if (is_handle_aborted(handle))
1086                 return -EROFS;
1087
1088         if (jbd2_write_access_granted(handle, bh, false))
1089                 return 0;
1090
1091         jh = jbd2_journal_add_journal_head(bh);
1092         /* We do not want to get caught playing with fields which the
1093          * log thread also manipulates.  Make sure that the buffer
1094          * completes any outstanding IO before proceeding. */
1095         rc = do_get_write_access(handle, jh, 0);
1096         jbd2_journal_put_journal_head(jh);
1097         return rc;
1098 }
1099
1100
1101 /*
1102  * When the user wants to journal a newly created buffer_head
1103  * (ie. getblk() returned a new buffer and we are going to populate it
1104  * manually rather than reading off disk), then we need to keep the
1105  * buffer_head locked until it has been completely filled with new
1106  * data.  In this case, we should be able to make the assertion that
1107  * the bh is not already part of an existing transaction.
1108  *
1109  * The buffer should already be locked by the caller by this point.
1110  * There is no lock ranking violation: it was a newly created,
1111  * unlocked buffer beforehand. */
1112
1113 /**
1114  * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1115  * @handle: transaction to new buffer to
1116  * @bh: new buffer.
1117  *
1118  * Call this if you create a new bh.
1119  */
1120 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1121 {
1122         transaction_t *transaction = handle->h_transaction;
1123         journal_t *journal;
1124         struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1125         int err;
1126
1127         jbd_debug(5, "journal_head %p\n", jh);
1128         err = -EROFS;
1129         if (is_handle_aborted(handle))
1130                 goto out;
1131         journal = transaction->t_journal;
1132         err = 0;
1133
1134         JBUFFER_TRACE(jh, "entry");
1135         /*
1136          * The buffer may already belong to this transaction due to pre-zeroing
1137          * in the filesystem's new_block code.  It may also be on the previous,
1138          * committing transaction's lists, but it HAS to be in Forget state in
1139          * that case: the transaction must have deleted the buffer for it to be
1140          * reused here.
1141          */
1142         jbd_lock_bh_state(bh);
1143         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1144                 jh->b_transaction == NULL ||
1145                 (jh->b_transaction == journal->j_committing_transaction &&
1146                           jh->b_jlist == BJ_Forget)));
1147
1148         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1149         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1150
1151         if (jh->b_transaction == NULL) {
1152                 /*
1153                  * Previous jbd2_journal_forget() could have left the buffer
1154                  * with jbddirty bit set because it was being committed. When
1155                  * the commit finished, we've filed the buffer for
1156                  * checkpointing and marked it dirty. Now we are reallocating
1157                  * the buffer so the transaction freeing it must have
1158                  * committed and so it's safe to clear the dirty bit.
1159                  */
1160                 clear_buffer_dirty(jh2bh(jh));
1161                 /* first access by this transaction */
1162                 jh->b_modified = 0;
1163
1164                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1165                 spin_lock(&journal->j_list_lock);
1166                 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1167                 spin_unlock(&journal->j_list_lock);
1168         } else if (jh->b_transaction == journal->j_committing_transaction) {
1169                 /* first access by this transaction */
1170                 jh->b_modified = 0;
1171
1172                 JBUFFER_TRACE(jh, "set next transaction");
1173                 spin_lock(&journal->j_list_lock);
1174                 jh->b_next_transaction = transaction;
1175                 spin_unlock(&journal->j_list_lock);
1176         }
1177         jbd_unlock_bh_state(bh);
1178
1179         /*
1180          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
1181          * blocks which contain freed but then revoked metadata.  We need
1182          * to cancel the revoke in case we end up freeing it yet again
1183          * and the reallocating as data - this would cause a second revoke,
1184          * which hits an assertion error.
1185          */
1186         JBUFFER_TRACE(jh, "cancelling revoke");
1187         jbd2_journal_cancel_revoke(handle, jh);
1188 out:
1189         jbd2_journal_put_journal_head(jh);
1190         return err;
1191 }
1192
1193 /**
1194  * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
1195  *     non-rewindable consequences
1196  * @handle: transaction
1197  * @bh: buffer to undo
1198  *
1199  * Sometimes there is a need to distinguish between metadata which has
1200  * been committed to disk and that which has not.  The ext3fs code uses
1201  * this for freeing and allocating space, we have to make sure that we
1202  * do not reuse freed space until the deallocation has been committed,
1203  * since if we overwrote that space we would make the delete
1204  * un-rewindable in case of a crash.
1205  *
1206  * To deal with that, jbd2_journal_get_undo_access requests write access to a
1207  * buffer for parts of non-rewindable operations such as delete
1208  * operations on the bitmaps.  The journaling code must keep a copy of
1209  * the buffer's contents prior to the undo_access call until such time
1210  * as we know that the buffer has definitely been committed to disk.
1211  *
1212  * We never need to know which transaction the committed data is part
1213  * of, buffers touched here are guaranteed to be dirtied later and so
1214  * will be committed to a new transaction in due course, at which point
1215  * we can discard the old committed data pointer.
1216  *
1217  * Returns error number or 0 on success.
1218  */
1219 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1220 {
1221         int err;
1222         struct journal_head *jh;
1223         char *committed_data = NULL;
1224
1225         if (is_handle_aborted(handle))
1226                 return -EROFS;
1227
1228         if (jbd2_write_access_granted(handle, bh, true))
1229                 return 0;
1230
1231         jh = jbd2_journal_add_journal_head(bh);
1232         JBUFFER_TRACE(jh, "entry");
1233
1234         /*
1235          * Do this first --- it can drop the journal lock, so we want to
1236          * make sure that obtaining the committed_data is done
1237          * atomically wrt. completion of any outstanding commits.
1238          */
1239         err = do_get_write_access(handle, jh, 1);
1240         if (err)
1241                 goto out;
1242
1243 repeat:
1244         if (!jh->b_committed_data)
1245                 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1246                                             GFP_NOFS|__GFP_NOFAIL);
1247
1248         jbd_lock_bh_state(bh);
1249         if (!jh->b_committed_data) {
1250                 /* Copy out the current buffer contents into the
1251                  * preserved, committed copy. */
1252                 JBUFFER_TRACE(jh, "generate b_committed data");
1253                 if (!committed_data) {
1254                         jbd_unlock_bh_state(bh);
1255                         goto repeat;
1256                 }
1257
1258                 jh->b_committed_data = committed_data;
1259                 committed_data = NULL;
1260                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1261         }
1262         jbd_unlock_bh_state(bh);
1263 out:
1264         jbd2_journal_put_journal_head(jh);
1265         if (unlikely(committed_data))
1266                 jbd2_free(committed_data, bh->b_size);
1267         return err;
1268 }
1269
1270 /**
1271  * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1272  * @bh: buffer to trigger on
1273  * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1274  *
1275  * Set any triggers on this journal_head.  This is always safe, because
1276  * triggers for a committing buffer will be saved off, and triggers for
1277  * a running transaction will match the buffer in that transaction.
1278  *
1279  * Call with NULL to clear the triggers.
1280  */
1281 void jbd2_journal_set_triggers(struct buffer_head *bh,
1282                                struct jbd2_buffer_trigger_type *type)
1283 {
1284         struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1285
1286         if (WARN_ON(!jh))
1287                 return;
1288         jh->b_triggers = type;
1289         jbd2_journal_put_journal_head(jh);
1290 }
1291
1292 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1293                                 struct jbd2_buffer_trigger_type *triggers)
1294 {
1295         struct buffer_head *bh = jh2bh(jh);
1296
1297         if (!triggers || !triggers->t_frozen)
1298                 return;
1299
1300         triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1301 }
1302
1303 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1304                                struct jbd2_buffer_trigger_type *triggers)
1305 {
1306         if (!triggers || !triggers->t_abort)
1307                 return;
1308
1309         triggers->t_abort(triggers, jh2bh(jh));
1310 }
1311
1312 /**
1313  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1314  * @handle: transaction to add buffer to.
1315  * @bh: buffer to mark
1316  *
1317  * mark dirty metadata which needs to be journaled as part of the current
1318  * transaction.
1319  *
1320  * The buffer must have previously had jbd2_journal_get_write_access()
1321  * called so that it has a valid journal_head attached to the buffer
1322  * head.
1323  *
1324  * The buffer is placed on the transaction's metadata list and is marked
1325  * as belonging to the transaction.
1326  *
1327  * Returns error number or 0 on success.
1328  *
1329  * Special care needs to be taken if the buffer already belongs to the
1330  * current committing transaction (in which case we should have frozen
1331  * data present for that commit).  In that case, we don't relink the
1332  * buffer: that only gets done when the old transaction finally
1333  * completes its commit.
1334  */
1335 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1336 {
1337         transaction_t *transaction = handle->h_transaction;
1338         journal_t *journal;
1339         struct journal_head *jh;
1340         int ret = 0;
1341
1342         if (is_handle_aborted(handle))
1343                 return -EROFS;
1344         if (!buffer_jbd(bh))
1345                 return -EUCLEAN;
1346
1347         /*
1348          * We don't grab jh reference here since the buffer must be part
1349          * of the running transaction.
1350          */
1351         jh = bh2jh(bh);
1352         jbd_debug(5, "journal_head %p\n", jh);
1353         JBUFFER_TRACE(jh, "entry");
1354
1355         /*
1356          * This and the following assertions are unreliable since we may see jh
1357          * in inconsistent state unless we grab bh_state lock. But this is
1358          * crucial to catch bugs so let's do a reliable check until the
1359          * lockless handling is fully proven.
1360          */
1361         if (jh->b_transaction != transaction &&
1362             jh->b_next_transaction != transaction) {
1363                 jbd_lock_bh_state(bh);
1364                 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1365                                 jh->b_next_transaction == transaction);
1366                 jbd_unlock_bh_state(bh);
1367         }
1368         if (jh->b_modified == 1) {
1369                 /* If it's in our transaction it must be in BJ_Metadata list. */
1370                 if (jh->b_transaction == transaction &&
1371                     jh->b_jlist != BJ_Metadata) {
1372                         jbd_lock_bh_state(bh);
1373                         if (jh->b_transaction == transaction &&
1374                             jh->b_jlist != BJ_Metadata)
1375                                 pr_err("JBD2: assertion failure: h_type=%u "
1376                                        "h_line_no=%u block_no=%llu jlist=%u\n",
1377                                        handle->h_type, handle->h_line_no,
1378                                        (unsigned long long) bh->b_blocknr,
1379                                        jh->b_jlist);
1380                         J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1381                                         jh->b_jlist == BJ_Metadata);
1382                         jbd_unlock_bh_state(bh);
1383                 }
1384                 goto out;
1385         }
1386
1387         journal = transaction->t_journal;
1388         jbd_lock_bh_state(bh);
1389
1390         if (jh->b_modified == 0) {
1391                 /*
1392                  * This buffer's got modified and becoming part
1393                  * of the transaction. This needs to be done
1394                  * once a transaction -bzzz
1395                  */
1396                 if (handle->h_buffer_credits <= 0) {
1397                         ret = -ENOSPC;
1398                         goto out_unlock_bh;
1399                 }
1400                 jh->b_modified = 1;
1401                 handle->h_buffer_credits--;
1402         }
1403
1404         /*
1405          * fastpath, to avoid expensive locking.  If this buffer is already
1406          * on the running transaction's metadata list there is nothing to do.
1407          * Nobody can take it off again because there is a handle open.
1408          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1409          * result in this test being false, so we go in and take the locks.
1410          */
1411         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1412                 JBUFFER_TRACE(jh, "fastpath");
1413                 if (unlikely(jh->b_transaction !=
1414                              journal->j_running_transaction)) {
1415                         printk(KERN_ERR "JBD2: %s: "
1416                                "jh->b_transaction (%llu, %p, %u) != "
1417                                "journal->j_running_transaction (%p, %u)\n",
1418                                journal->j_devname,
1419                                (unsigned long long) bh->b_blocknr,
1420                                jh->b_transaction,
1421                                jh->b_transaction ? jh->b_transaction->t_tid : 0,
1422                                journal->j_running_transaction,
1423                                journal->j_running_transaction ?
1424                                journal->j_running_transaction->t_tid : 0);
1425                         ret = -EINVAL;
1426                 }
1427                 goto out_unlock_bh;
1428         }
1429
1430         set_buffer_jbddirty(bh);
1431
1432         /*
1433          * Metadata already on the current transaction list doesn't
1434          * need to be filed.  Metadata on another transaction's list must
1435          * be committing, and will be refiled once the commit completes:
1436          * leave it alone for now.
1437          */
1438         if (jh->b_transaction != transaction) {
1439                 JBUFFER_TRACE(jh, "already on other transaction");
1440                 if (unlikely(((jh->b_transaction !=
1441                                journal->j_committing_transaction)) ||
1442                              (jh->b_next_transaction != transaction))) {
1443                         printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1444                                "bad jh for block %llu: "
1445                                "transaction (%p, %u), "
1446                                "jh->b_transaction (%p, %u), "
1447                                "jh->b_next_transaction (%p, %u), jlist %u\n",
1448                                journal->j_devname,
1449                                (unsigned long long) bh->b_blocknr,
1450                                transaction, transaction->t_tid,
1451                                jh->b_transaction,
1452                                jh->b_transaction ?
1453                                jh->b_transaction->t_tid : 0,
1454                                jh->b_next_transaction,
1455                                jh->b_next_transaction ?
1456                                jh->b_next_transaction->t_tid : 0,
1457                                jh->b_jlist);
1458                         WARN_ON(1);
1459                         ret = -EINVAL;
1460                 }
1461                 /* And this case is illegal: we can't reuse another
1462                  * transaction's data buffer, ever. */
1463                 goto out_unlock_bh;
1464         }
1465
1466         /* That test should have eliminated the following case: */
1467         J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1468
1469         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1470         spin_lock(&journal->j_list_lock);
1471         __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1472         spin_unlock(&journal->j_list_lock);
1473 out_unlock_bh:
1474         jbd_unlock_bh_state(bh);
1475 out:
1476         JBUFFER_TRACE(jh, "exit");
1477         return ret;
1478 }
1479
1480 /**
1481  * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1482  * @handle: transaction handle
1483  * @bh:     bh to 'forget'
1484  *
1485  * We can only do the bforget if there are no commits pending against the
1486  * buffer.  If the buffer is dirty in the current running transaction we
1487  * can safely unlink it.
1488  *
1489  * bh may not be a journalled buffer at all - it may be a non-JBD
1490  * buffer which came off the hashtable.  Check for this.
1491  *
1492  * Decrements bh->b_count by one.
1493  *
1494  * Allow this call even if the handle has aborted --- it may be part of
1495  * the caller's cleanup after an abort.
1496  */
1497 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1498 {
1499         transaction_t *transaction = handle->h_transaction;
1500         journal_t *journal;
1501         struct journal_head *jh;
1502         int drop_reserve = 0;
1503         int err = 0;
1504         int was_modified = 0;
1505
1506         if (is_handle_aborted(handle))
1507                 return -EROFS;
1508         journal = transaction->t_journal;
1509
1510         BUFFER_TRACE(bh, "entry");
1511
1512         jbd_lock_bh_state(bh);
1513
1514         if (!buffer_jbd(bh))
1515                 goto not_jbd;
1516         jh = bh2jh(bh);
1517
1518         /* Critical error: attempting to delete a bitmap buffer, maybe?
1519          * Don't do any jbd operations, and return an error. */
1520         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1521                          "inconsistent data on disk")) {
1522                 err = -EIO;
1523                 goto not_jbd;
1524         }
1525
1526         /* keep track of whether or not this transaction modified us */
1527         was_modified = jh->b_modified;
1528
1529         /*
1530          * The buffer's going from the transaction, we must drop
1531          * all references -bzzz
1532          */
1533         jh->b_modified = 0;
1534
1535         if (jh->b_transaction == transaction) {
1536                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1537
1538                 /* If we are forgetting a buffer which is already part
1539                  * of this transaction, then we can just drop it from
1540                  * the transaction immediately. */
1541                 clear_buffer_dirty(bh);
1542                 clear_buffer_jbddirty(bh);
1543
1544                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1545
1546                 /*
1547                  * we only want to drop a reference if this transaction
1548                  * modified the buffer
1549                  */
1550                 if (was_modified)
1551                         drop_reserve = 1;
1552
1553                 /*
1554                  * We are no longer going to journal this buffer.
1555                  * However, the commit of this transaction is still
1556                  * important to the buffer: the delete that we are now
1557                  * processing might obsolete an old log entry, so by
1558                  * committing, we can satisfy the buffer's checkpoint.
1559                  *
1560                  * So, if we have a checkpoint on the buffer, we should
1561                  * now refile the buffer on our BJ_Forget list so that
1562                  * we know to remove the checkpoint after we commit.
1563                  */
1564
1565                 spin_lock(&journal->j_list_lock);
1566                 if (jh->b_cp_transaction) {
1567                         __jbd2_journal_temp_unlink_buffer(jh);
1568                         __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1569                 } else {
1570                         __jbd2_journal_unfile_buffer(jh);
1571                         if (!buffer_jbd(bh)) {
1572                                 spin_unlock(&journal->j_list_lock);
1573                                 jbd_unlock_bh_state(bh);
1574                                 __bforget(bh);
1575                                 goto drop;
1576                         }
1577                 }
1578                 spin_unlock(&journal->j_list_lock);
1579         } else if (jh->b_transaction) {
1580                 J_ASSERT_JH(jh, (jh->b_transaction ==
1581                                  journal->j_committing_transaction));
1582                 /* However, if the buffer is still owned by a prior
1583                  * (committing) transaction, we can't drop it yet... */
1584                 JBUFFER_TRACE(jh, "belongs to older transaction");
1585                 /* ... but we CAN drop it from the new transaction through
1586                  * marking the buffer as freed and set j_next_transaction to
1587                  * the new transaction, so that not only the commit code
1588                  * knows it should clear dirty bits when it is done with the
1589                  * buffer, but also the buffer can be checkpointed only
1590                  * after the new transaction commits. */
1591
1592                 set_buffer_freed(bh);
1593
1594                 if (!jh->b_next_transaction) {
1595                         spin_lock(&journal->j_list_lock);
1596                         jh->b_next_transaction = transaction;
1597                         spin_unlock(&journal->j_list_lock);
1598                 } else {
1599                         J_ASSERT(jh->b_next_transaction == transaction);
1600
1601                         /*
1602                          * only drop a reference if this transaction modified
1603                          * the buffer
1604                          */
1605                         if (was_modified)
1606                                 drop_reserve = 1;
1607                 }
1608         }
1609
1610 not_jbd:
1611         jbd_unlock_bh_state(bh);
1612         __brelse(bh);
1613 drop:
1614         if (drop_reserve) {
1615                 /* no need to reserve log space for this block -bzzz */
1616                 handle->h_buffer_credits++;
1617         }
1618         return err;
1619 }
1620
1621 /**
1622  * int jbd2_journal_stop() - complete a transaction
1623  * @handle: transaction to complete.
1624  *
1625  * All done for a particular handle.
1626  *
1627  * There is not much action needed here.  We just return any remaining
1628  * buffer credits to the transaction and remove the handle.  The only
1629  * complication is that we need to start a commit operation if the
1630  * filesystem is marked for synchronous update.
1631  *
1632  * jbd2_journal_stop itself will not usually return an error, but it may
1633  * do so in unusual circumstances.  In particular, expect it to
1634  * return -EIO if a jbd2_journal_abort has been executed since the
1635  * transaction began.
1636  */
1637 int jbd2_journal_stop(handle_t *handle)
1638 {
1639         transaction_t *transaction = handle->h_transaction;
1640         journal_t *journal;
1641         int err = 0, wait_for_commit = 0;
1642         tid_t tid;
1643         pid_t pid;
1644
1645         if (!transaction) {
1646                 /*
1647                  * Handle is already detached from the transaction so
1648                  * there is nothing to do other than decrease a refcount,
1649                  * or free the handle if refcount drops to zero
1650                  */
1651                 if (--handle->h_ref > 0) {
1652                         jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1653                                                          handle->h_ref);
1654                         return err;
1655                 } else {
1656                         if (handle->h_rsv_handle)
1657                                 jbd2_free_handle(handle->h_rsv_handle);
1658                         goto free_and_exit;
1659                 }
1660         }
1661         journal = transaction->t_journal;
1662
1663         J_ASSERT(journal_current_handle() == handle);
1664
1665         if (is_handle_aborted(handle))
1666                 err = -EIO;
1667         else
1668                 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1669
1670         if (--handle->h_ref > 0) {
1671                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1672                           handle->h_ref);
1673                 return err;
1674         }
1675
1676         jbd_debug(4, "Handle %p going down\n", handle);
1677         trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1678                                 transaction->t_tid,
1679                                 handle->h_type, handle->h_line_no,
1680                                 jiffies - handle->h_start_jiffies,
1681                                 handle->h_sync, handle->h_requested_credits,
1682                                 (handle->h_requested_credits -
1683                                  handle->h_buffer_credits));
1684
1685         /*
1686          * Implement synchronous transaction batching.  If the handle
1687          * was synchronous, don't force a commit immediately.  Let's
1688          * yield and let another thread piggyback onto this
1689          * transaction.  Keep doing that while new threads continue to
1690          * arrive.  It doesn't cost much - we're about to run a commit
1691          * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1692          * operations by 30x or more...
1693          *
1694          * We try and optimize the sleep time against what the
1695          * underlying disk can do, instead of having a static sleep
1696          * time.  This is useful for the case where our storage is so
1697          * fast that it is more optimal to go ahead and force a flush
1698          * and wait for the transaction to be committed than it is to
1699          * wait for an arbitrary amount of time for new writers to
1700          * join the transaction.  We achieve this by measuring how
1701          * long it takes to commit a transaction, and compare it with
1702          * how long this transaction has been running, and if run time
1703          * < commit time then we sleep for the delta and commit.  This
1704          * greatly helps super fast disks that would see slowdowns as
1705          * more threads started doing fsyncs.
1706          *
1707          * But don't do this if this process was the most recent one
1708          * to perform a synchronous write.  We do this to detect the
1709          * case where a single process is doing a stream of sync
1710          * writes.  No point in waiting for joiners in that case.
1711          *
1712          * Setting max_batch_time to 0 disables this completely.
1713          */
1714         pid = current->pid;
1715         if (handle->h_sync && journal->j_last_sync_writer != pid &&
1716             journal->j_max_batch_time) {
1717                 u64 commit_time, trans_time;
1718
1719                 journal->j_last_sync_writer = pid;
1720
1721                 read_lock(&journal->j_state_lock);
1722                 commit_time = journal->j_average_commit_time;
1723                 read_unlock(&journal->j_state_lock);
1724
1725                 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1726                                                    transaction->t_start_time));
1727
1728                 commit_time = max_t(u64, commit_time,
1729                                     1000*journal->j_min_batch_time);
1730                 commit_time = min_t(u64, commit_time,
1731                                     1000*journal->j_max_batch_time);
1732
1733                 if (trans_time < commit_time) {
1734                         ktime_t expires = ktime_add_ns(ktime_get(),
1735                                                        commit_time);
1736                         set_current_state(TASK_UNINTERRUPTIBLE);
1737                         schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1738                 }
1739         }
1740
1741         if (handle->h_sync)
1742                 transaction->t_synchronous_commit = 1;
1743         current->journal_info = NULL;
1744         atomic_sub(handle->h_buffer_credits,
1745                    &transaction->t_outstanding_credits);
1746
1747         /*
1748          * If the handle is marked SYNC, we need to set another commit
1749          * going!  We also want to force a commit if the current
1750          * transaction is occupying too much of the log, or if the
1751          * transaction is too old now.
1752          */
1753         if (handle->h_sync ||
1754             (atomic_read(&transaction->t_outstanding_credits) >
1755              journal->j_max_transaction_buffers) ||
1756             time_after_eq(jiffies, transaction->t_expires)) {
1757                 /* Do this even for aborted journals: an abort still
1758                  * completes the commit thread, it just doesn't write
1759                  * anything to disk. */
1760
1761                 jbd_debug(2, "transaction too old, requesting commit for "
1762                                         "handle %p\n", handle);
1763                 /* This is non-blocking */
1764                 jbd2_log_start_commit(journal, transaction->t_tid);
1765
1766                 /*
1767                  * Special case: JBD2_SYNC synchronous updates require us
1768                  * to wait for the commit to complete.
1769                  */
1770                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1771                         wait_for_commit = 1;
1772         }
1773
1774         /*
1775          * Once we drop t_updates, if it goes to zero the transaction
1776          * could start committing on us and eventually disappear.  So
1777          * once we do this, we must not dereference transaction
1778          * pointer again.
1779          */
1780         tid = transaction->t_tid;
1781         if (atomic_dec_and_test(&transaction->t_updates)) {
1782                 wake_up(&journal->j_wait_updates);
1783                 if (journal->j_barrier_count)
1784                         wake_up(&journal->j_wait_transaction_locked);
1785         }
1786
1787         rwsem_release(&journal->j_trans_commit_map, 1, _THIS_IP_);
1788
1789         if (wait_for_commit)
1790                 err = jbd2_log_wait_commit(journal, tid);
1791
1792         if (handle->h_rsv_handle)
1793                 jbd2_journal_free_reserved(handle->h_rsv_handle);
1794 free_and_exit:
1795         /*
1796          * Scope of the GFP_NOFS context is over here and so we can restore the
1797          * original alloc context.
1798          */
1799         memalloc_nofs_restore(handle->saved_alloc_context);
1800         jbd2_free_handle(handle);
1801         return err;
1802 }
1803
1804 /*
1805  *
1806  * List management code snippets: various functions for manipulating the
1807  * transaction buffer lists.
1808  *
1809  */
1810
1811 /*
1812  * Append a buffer to a transaction list, given the transaction's list head
1813  * pointer.
1814  *
1815  * j_list_lock is held.
1816  *
1817  * jbd_lock_bh_state(jh2bh(jh)) is held.
1818  */
1819
1820 static inline void
1821 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1822 {
1823         if (!*list) {
1824                 jh->b_tnext = jh->b_tprev = jh;
1825                 *list = jh;
1826         } else {
1827                 /* Insert at the tail of the list to preserve order */
1828                 struct journal_head *first = *list, *last = first->b_tprev;
1829                 jh->b_tprev = last;
1830                 jh->b_tnext = first;
1831                 last->b_tnext = first->b_tprev = jh;
1832         }
1833 }
1834
1835 /*
1836  * Remove a buffer from a transaction list, given the transaction's list
1837  * head pointer.
1838  *
1839  * Called with j_list_lock held, and the journal may not be locked.
1840  *
1841  * jbd_lock_bh_state(jh2bh(jh)) is held.
1842  */
1843
1844 static inline void
1845 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1846 {
1847         if (*list == jh) {
1848                 *list = jh->b_tnext;
1849                 if (*list == jh)
1850                         *list = NULL;
1851         }
1852         jh->b_tprev->b_tnext = jh->b_tnext;
1853         jh->b_tnext->b_tprev = jh->b_tprev;
1854 }
1855
1856 /*
1857  * Remove a buffer from the appropriate transaction list.
1858  *
1859  * Note that this function can *change* the value of
1860  * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1861  * t_reserved_list.  If the caller is holding onto a copy of one of these
1862  * pointers, it could go bad.  Generally the caller needs to re-read the
1863  * pointer from the transaction_t.
1864  *
1865  * Called under j_list_lock.
1866  */
1867 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1868 {
1869         struct journal_head **list = NULL;
1870         transaction_t *transaction;
1871         struct buffer_head *bh = jh2bh(jh);
1872
1873         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1874         transaction = jh->b_transaction;
1875         if (transaction)
1876                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1877
1878         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1879         if (jh->b_jlist != BJ_None)
1880                 J_ASSERT_JH(jh, transaction != NULL);
1881
1882         switch (jh->b_jlist) {
1883         case BJ_None:
1884                 return;
1885         case BJ_Metadata:
1886                 transaction->t_nr_buffers--;
1887                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1888                 list = &transaction->t_buffers;
1889                 break;
1890         case BJ_Forget:
1891                 list = &transaction->t_forget;
1892                 break;
1893         case BJ_Shadow:
1894                 list = &transaction->t_shadow_list;
1895                 break;
1896         case BJ_Reserved:
1897                 list = &transaction->t_reserved_list;
1898                 break;
1899         }
1900
1901         __blist_del_buffer(list, jh);
1902         jh->b_jlist = BJ_None;
1903         if (transaction && is_journal_aborted(transaction->t_journal))
1904                 clear_buffer_jbddirty(bh);
1905         else if (test_clear_buffer_jbddirty(bh))
1906                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1907 }
1908
1909 /*
1910  * Remove buffer from all transactions.
1911  *
1912  * Called with bh_state lock and j_list_lock
1913  *
1914  * jh and bh may be already freed when this function returns.
1915  */
1916 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1917 {
1918         J_ASSERT_JH(jh, jh->b_transaction != NULL);
1919         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1920
1921         __jbd2_journal_temp_unlink_buffer(jh);
1922         jh->b_transaction = NULL;
1923         jbd2_journal_put_journal_head(jh);
1924 }
1925
1926 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1927 {
1928         struct buffer_head *bh = jh2bh(jh);
1929
1930         /* Get reference so that buffer cannot be freed before we unlock it */
1931         get_bh(bh);
1932         jbd_lock_bh_state(bh);
1933         spin_lock(&journal->j_list_lock);
1934         __jbd2_journal_unfile_buffer(jh);
1935         spin_unlock(&journal->j_list_lock);
1936         jbd_unlock_bh_state(bh);
1937         __brelse(bh);
1938 }
1939
1940 /*
1941  * Called from jbd2_journal_try_to_free_buffers().
1942  *
1943  * Called under jbd_lock_bh_state(bh)
1944  */
1945 static void
1946 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1947 {
1948         struct journal_head *jh;
1949
1950         jh = bh2jh(bh);
1951
1952         if (buffer_locked(bh) || buffer_dirty(bh))
1953                 goto out;
1954
1955         if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
1956                 goto out;
1957
1958         spin_lock(&journal->j_list_lock);
1959         if (jh->b_cp_transaction != NULL) {
1960                 /* written-back checkpointed metadata buffer */
1961                 JBUFFER_TRACE(jh, "remove from checkpoint list");
1962                 __jbd2_journal_remove_checkpoint(jh);
1963         }
1964         spin_unlock(&journal->j_list_lock);
1965 out:
1966         return;
1967 }
1968
1969 /**
1970  * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1971  * @journal: journal for operation
1972  * @page: to try and free
1973  * @gfp_mask: we use the mask to detect how hard should we try to release
1974  * buffers. If __GFP_DIRECT_RECLAIM and __GFP_FS is set, we wait for commit
1975  * code to release the buffers.
1976  *
1977  *
1978  * For all the buffers on this page,
1979  * if they are fully written out ordered data, move them onto BUF_CLEAN
1980  * so try_to_free_buffers() can reap them.
1981  *
1982  * This function returns non-zero if we wish try_to_free_buffers()
1983  * to be called. We do this if the page is releasable by try_to_free_buffers().
1984  * We also do it if the page has locked or dirty buffers and the caller wants
1985  * us to perform sync or async writeout.
1986  *
1987  * This complicates JBD locking somewhat.  We aren't protected by the
1988  * BKL here.  We wish to remove the buffer from its committing or
1989  * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1990  *
1991  * This may *change* the value of transaction_t->t_datalist, so anyone
1992  * who looks at t_datalist needs to lock against this function.
1993  *
1994  * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1995  * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1996  * will come out of the lock with the buffer dirty, which makes it
1997  * ineligible for release here.
1998  *
1999  * Who else is affected by this?  hmm...  Really the only contender
2000  * is do_get_write_access() - it could be looking at the buffer while
2001  * journal_try_to_free_buffer() is changing its state.  But that
2002  * cannot happen because we never reallocate freed data as metadata
2003  * while the data is part of a transaction.  Yes?
2004  *
2005  * Return 0 on failure, 1 on success
2006  */
2007 int jbd2_journal_try_to_free_buffers(journal_t *journal,
2008                                 struct page *page, gfp_t gfp_mask)
2009 {
2010         struct buffer_head *head;
2011         struct buffer_head *bh;
2012         bool has_write_io_error = false;
2013         int ret = 0;
2014
2015         J_ASSERT(PageLocked(page));
2016
2017         head = page_buffers(page);
2018         bh = head;
2019         do {
2020                 struct journal_head *jh;
2021
2022                 /*
2023                  * We take our own ref against the journal_head here to avoid
2024                  * having to add tons of locking around each instance of
2025                  * jbd2_journal_put_journal_head().
2026                  */
2027                 jh = jbd2_journal_grab_journal_head(bh);
2028                 if (!jh)
2029                         continue;
2030
2031                 jbd_lock_bh_state(bh);
2032                 __journal_try_to_free_buffer(journal, bh);
2033                 jbd2_journal_put_journal_head(jh);
2034                 jbd_unlock_bh_state(bh);
2035                 if (buffer_jbd(bh))
2036                         goto busy;
2037
2038                 /*
2039                  * If we free a metadata buffer which has been failed to
2040                  * write out, the jbd2 checkpoint procedure will not detect
2041                  * this failure and may lead to filesystem inconsistency
2042                  * after cleanup journal tail.
2043                  */
2044                 if (buffer_write_io_error(bh)) {
2045                         pr_err("JBD2: Error while async write back metadata bh %llu.",
2046                                (unsigned long long)bh->b_blocknr);
2047                         has_write_io_error = true;
2048                 }
2049         } while ((bh = bh->b_this_page) != head);
2050
2051         ret = try_to_free_buffers(page);
2052
2053 busy:
2054         if (has_write_io_error)
2055                 jbd2_journal_abort(journal, -EIO);
2056
2057         return ret;
2058 }
2059
2060 /*
2061  * This buffer is no longer needed.  If it is on an older transaction's
2062  * checkpoint list we need to record it on this transaction's forget list
2063  * to pin this buffer (and hence its checkpointing transaction) down until
2064  * this transaction commits.  If the buffer isn't on a checkpoint list, we
2065  * release it.
2066  * Returns non-zero if JBD no longer has an interest in the buffer.
2067  *
2068  * Called under j_list_lock.
2069  *
2070  * Called under jbd_lock_bh_state(bh).
2071  */
2072 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2073 {
2074         int may_free = 1;
2075         struct buffer_head *bh = jh2bh(jh);
2076
2077         if (jh->b_cp_transaction) {
2078                 JBUFFER_TRACE(jh, "on running+cp transaction");
2079                 __jbd2_journal_temp_unlink_buffer(jh);
2080                 /*
2081                  * We don't want to write the buffer anymore, clear the
2082                  * bit so that we don't confuse checks in
2083                  * __journal_file_buffer
2084                  */
2085                 clear_buffer_dirty(bh);
2086                 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2087                 may_free = 0;
2088         } else {
2089                 JBUFFER_TRACE(jh, "on running transaction");
2090                 __jbd2_journal_unfile_buffer(jh);
2091         }
2092         return may_free;
2093 }
2094
2095 /*
2096  * jbd2_journal_invalidatepage
2097  *
2098  * This code is tricky.  It has a number of cases to deal with.
2099  *
2100  * There are two invariants which this code relies on:
2101  *
2102  * i_size must be updated on disk before we start calling invalidatepage on the
2103  * data.
2104  *
2105  *  This is done in ext3 by defining an ext3_setattr method which
2106  *  updates i_size before truncate gets going.  By maintaining this
2107  *  invariant, we can be sure that it is safe to throw away any buffers
2108  *  attached to the current transaction: once the transaction commits,
2109  *  we know that the data will not be needed.
2110  *
2111  *  Note however that we can *not* throw away data belonging to the
2112  *  previous, committing transaction!
2113  *
2114  * Any disk blocks which *are* part of the previous, committing
2115  * transaction (and which therefore cannot be discarded immediately) are
2116  * not going to be reused in the new running transaction
2117  *
2118  *  The bitmap committed_data images guarantee this: any block which is
2119  *  allocated in one transaction and removed in the next will be marked
2120  *  as in-use in the committed_data bitmap, so cannot be reused until
2121  *  the next transaction to delete the block commits.  This means that
2122  *  leaving committing buffers dirty is quite safe: the disk blocks
2123  *  cannot be reallocated to a different file and so buffer aliasing is
2124  *  not possible.
2125  *
2126  *
2127  * The above applies mainly to ordered data mode.  In writeback mode we
2128  * don't make guarantees about the order in which data hits disk --- in
2129  * particular we don't guarantee that new dirty data is flushed before
2130  * transaction commit --- so it is always safe just to discard data
2131  * immediately in that mode.  --sct
2132  */
2133
2134 /*
2135  * The journal_unmap_buffer helper function returns zero if the buffer
2136  * concerned remains pinned as an anonymous buffer belonging to an older
2137  * transaction.
2138  *
2139  * We're outside-transaction here.  Either or both of j_running_transaction
2140  * and j_committing_transaction may be NULL.
2141  */
2142 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2143                                 int partial_page)
2144 {
2145         transaction_t *transaction;
2146         struct journal_head *jh;
2147         int may_free = 1;
2148
2149         BUFFER_TRACE(bh, "entry");
2150
2151         /*
2152          * It is safe to proceed here without the j_list_lock because the
2153          * buffers cannot be stolen by try_to_free_buffers as long as we are
2154          * holding the page lock. --sct
2155          */
2156
2157         if (!buffer_jbd(bh))
2158                 goto zap_buffer_unlocked;
2159
2160         /* OK, we have data buffer in journaled mode */
2161         write_lock(&journal->j_state_lock);
2162         jbd_lock_bh_state(bh);
2163         spin_lock(&journal->j_list_lock);
2164
2165         jh = jbd2_journal_grab_journal_head(bh);
2166         if (!jh)
2167                 goto zap_buffer_no_jh;
2168
2169         /*
2170          * We cannot remove the buffer from checkpoint lists until the
2171          * transaction adding inode to orphan list (let's call it T)
2172          * is committed.  Otherwise if the transaction changing the
2173          * buffer would be cleaned from the journal before T is
2174          * committed, a crash will cause that the correct contents of
2175          * the buffer will be lost.  On the other hand we have to
2176          * clear the buffer dirty bit at latest at the moment when the
2177          * transaction marking the buffer as freed in the filesystem
2178          * structures is committed because from that moment on the
2179          * block can be reallocated and used by a different page.
2180          * Since the block hasn't been freed yet but the inode has
2181          * already been added to orphan list, it is safe for us to add
2182          * the buffer to BJ_Forget list of the newest transaction.
2183          *
2184          * Also we have to clear buffer_mapped flag of a truncated buffer
2185          * because the buffer_head may be attached to the page straddling
2186          * i_size (can happen only when blocksize < pagesize) and thus the
2187          * buffer_head can be reused when the file is extended again. So we end
2188          * up keeping around invalidated buffers attached to transactions'
2189          * BJ_Forget list just to stop checkpointing code from cleaning up
2190          * the transaction this buffer was modified in.
2191          */
2192         transaction = jh->b_transaction;
2193         if (transaction == NULL) {
2194                 /* First case: not on any transaction.  If it
2195                  * has no checkpoint link, then we can zap it:
2196                  * it's a writeback-mode buffer so we don't care
2197                  * if it hits disk safely. */
2198                 if (!jh->b_cp_transaction) {
2199                         JBUFFER_TRACE(jh, "not on any transaction: zap");
2200                         goto zap_buffer;
2201                 }
2202
2203                 if (!buffer_dirty(bh)) {
2204                         /* bdflush has written it.  We can drop it now */
2205                         __jbd2_journal_remove_checkpoint(jh);
2206                         goto zap_buffer;
2207                 }
2208
2209                 /* OK, it must be in the journal but still not
2210                  * written fully to disk: it's metadata or
2211                  * journaled data... */
2212
2213                 if (journal->j_running_transaction) {
2214                         /* ... and once the current transaction has
2215                          * committed, the buffer won't be needed any
2216                          * longer. */
2217                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2218                         may_free = __dispose_buffer(jh,
2219                                         journal->j_running_transaction);
2220                         goto zap_buffer;
2221                 } else {
2222                         /* There is no currently-running transaction. So the
2223                          * orphan record which we wrote for this file must have
2224                          * passed into commit.  We must attach this buffer to
2225                          * the committing transaction, if it exists. */
2226                         if (journal->j_committing_transaction) {
2227                                 JBUFFER_TRACE(jh, "give to committing trans");
2228                                 may_free = __dispose_buffer(jh,
2229                                         journal->j_committing_transaction);
2230                                 goto zap_buffer;
2231                         } else {
2232                                 /* The orphan record's transaction has
2233                                  * committed.  We can cleanse this buffer */
2234                                 clear_buffer_jbddirty(bh);
2235                                 __jbd2_journal_remove_checkpoint(jh);
2236                                 goto zap_buffer;
2237                         }
2238                 }
2239         } else if (transaction == journal->j_committing_transaction) {
2240                 JBUFFER_TRACE(jh, "on committing transaction");
2241                 /*
2242                  * The buffer is committing, we simply cannot touch
2243                  * it. If the page is straddling i_size we have to wait
2244                  * for commit and try again.
2245                  */
2246                 if (partial_page) {
2247                         jbd2_journal_put_journal_head(jh);
2248                         spin_unlock(&journal->j_list_lock);
2249                         jbd_unlock_bh_state(bh);
2250                         write_unlock(&journal->j_state_lock);
2251                         return -EBUSY;
2252                 }
2253                 /*
2254                  * OK, buffer won't be reachable after truncate. We just clear
2255                  * b_modified to not confuse transaction credit accounting, and
2256                  * set j_next_transaction to the running transaction (if there
2257                  * is one) and mark buffer as freed so that commit code knows
2258                  * it should clear dirty bits when it is done with the buffer.
2259                  */
2260                 set_buffer_freed(bh);
2261                 if (journal->j_running_transaction && buffer_jbddirty(bh))
2262                         jh->b_next_transaction = journal->j_running_transaction;
2263                 jh->b_modified = 0;
2264                 jbd2_journal_put_journal_head(jh);
2265                 spin_unlock(&journal->j_list_lock);
2266                 jbd_unlock_bh_state(bh);
2267                 write_unlock(&journal->j_state_lock);
2268                 return 0;
2269         } else {
2270                 /* Good, the buffer belongs to the running transaction.
2271                  * We are writing our own transaction's data, not any
2272                  * previous one's, so it is safe to throw it away
2273                  * (remember that we expect the filesystem to have set
2274                  * i_size already for this truncate so recovery will not
2275                  * expose the disk blocks we are discarding here.) */
2276                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2277                 JBUFFER_TRACE(jh, "on running transaction");
2278                 may_free = __dispose_buffer(jh, transaction);
2279         }
2280
2281 zap_buffer:
2282         /*
2283          * This is tricky. Although the buffer is truncated, it may be reused
2284          * if blocksize < pagesize and it is attached to the page straddling
2285          * EOF. Since the buffer might have been added to BJ_Forget list of the
2286          * running transaction, journal_get_write_access() won't clear
2287          * b_modified and credit accounting gets confused. So clear b_modified
2288          * here.
2289          */
2290         jh->b_modified = 0;
2291         jbd2_journal_put_journal_head(jh);
2292 zap_buffer_no_jh:
2293         spin_unlock(&journal->j_list_lock);
2294         jbd_unlock_bh_state(bh);
2295         write_unlock(&journal->j_state_lock);
2296 zap_buffer_unlocked:
2297         clear_buffer_dirty(bh);
2298         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2299         clear_buffer_mapped(bh);
2300         clear_buffer_req(bh);
2301         clear_buffer_new(bh);
2302         clear_buffer_delay(bh);
2303         clear_buffer_unwritten(bh);
2304         bh->b_bdev = NULL;
2305         return may_free;
2306 }
2307
2308 /**
2309  * void jbd2_journal_invalidatepage()
2310  * @journal: journal to use for flush...
2311  * @page:    page to flush
2312  * @offset:  start of the range to invalidate
2313  * @length:  length of the range to invalidate
2314  *
2315  * Reap page buffers containing data after in the specified range in page.
2316  * Can return -EBUSY if buffers are part of the committing transaction and
2317  * the page is straddling i_size. Caller then has to wait for current commit
2318  * and try again.
2319  */
2320 int jbd2_journal_invalidatepage(journal_t *journal,
2321                                 struct page *page,
2322                                 unsigned int offset,
2323                                 unsigned int length)
2324 {
2325         struct buffer_head *head, *bh, *next;
2326         unsigned int stop = offset + length;
2327         unsigned int curr_off = 0;
2328         int partial_page = (offset || length < PAGE_SIZE);
2329         int may_free = 1;
2330         int ret = 0;
2331
2332         if (!PageLocked(page))
2333                 BUG();
2334         if (!page_has_buffers(page))
2335                 return 0;
2336
2337         BUG_ON(stop > PAGE_SIZE || stop < length);
2338
2339         /* We will potentially be playing with lists other than just the
2340          * data lists (especially for journaled data mode), so be
2341          * cautious in our locking. */
2342
2343         head = bh = page_buffers(page);
2344         do {
2345                 unsigned int next_off = curr_off + bh->b_size;
2346                 next = bh->b_this_page;
2347
2348                 if (next_off > stop)
2349                         return 0;
2350
2351                 if (offset <= curr_off) {
2352                         /* This block is wholly outside the truncation point */
2353                         lock_buffer(bh);
2354                         ret = journal_unmap_buffer(journal, bh, partial_page);
2355                         unlock_buffer(bh);
2356                         if (ret < 0)
2357                                 return ret;
2358                         may_free &= ret;
2359                 }
2360                 curr_off = next_off;
2361                 bh = next;
2362
2363         } while (bh != head);
2364
2365         if (!partial_page) {
2366                 if (may_free && try_to_free_buffers(page))
2367                         J_ASSERT(!page_has_buffers(page));
2368         }
2369         return 0;
2370 }
2371
2372 /*
2373  * File a buffer on the given transaction list.
2374  */
2375 void __jbd2_journal_file_buffer(struct journal_head *jh,
2376                         transaction_t *transaction, int jlist)
2377 {
2378         struct journal_head **list = NULL;
2379         int was_dirty = 0;
2380         struct buffer_head *bh = jh2bh(jh);
2381
2382         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2383         assert_spin_locked(&transaction->t_journal->j_list_lock);
2384
2385         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2386         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2387                                 jh->b_transaction == NULL);
2388
2389         if (jh->b_transaction && jh->b_jlist == jlist)
2390                 return;
2391
2392         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2393             jlist == BJ_Shadow || jlist == BJ_Forget) {
2394                 /*
2395                  * For metadata buffers, we track dirty bit in buffer_jbddirty
2396                  * instead of buffer_dirty. We should not see a dirty bit set
2397                  * here because we clear it in do_get_write_access but e.g.
2398                  * tune2fs can modify the sb and set the dirty bit at any time
2399                  * so we try to gracefully handle that.
2400                  */
2401                 if (buffer_dirty(bh))
2402                         warn_dirty_buffer(bh);
2403                 if (test_clear_buffer_dirty(bh) ||
2404                     test_clear_buffer_jbddirty(bh))
2405                         was_dirty = 1;
2406         }
2407
2408         if (jh->b_transaction)
2409                 __jbd2_journal_temp_unlink_buffer(jh);
2410         else
2411                 jbd2_journal_grab_journal_head(bh);
2412         jh->b_transaction = transaction;
2413
2414         switch (jlist) {
2415         case BJ_None:
2416                 J_ASSERT_JH(jh, !jh->b_committed_data);
2417                 J_ASSERT_JH(jh, !jh->b_frozen_data);
2418                 return;
2419         case BJ_Metadata:
2420                 transaction->t_nr_buffers++;
2421                 list = &transaction->t_buffers;
2422                 break;
2423         case BJ_Forget:
2424                 list = &transaction->t_forget;
2425                 break;
2426         case BJ_Shadow:
2427                 list = &transaction->t_shadow_list;
2428                 break;
2429         case BJ_Reserved:
2430                 list = &transaction->t_reserved_list;
2431                 break;
2432         }
2433
2434         __blist_add_buffer(list, jh);
2435         jh->b_jlist = jlist;
2436
2437         if (was_dirty)
2438                 set_buffer_jbddirty(bh);
2439 }
2440
2441 void jbd2_journal_file_buffer(struct journal_head *jh,
2442                                 transaction_t *transaction, int jlist)
2443 {
2444         jbd_lock_bh_state(jh2bh(jh));
2445         spin_lock(&transaction->t_journal->j_list_lock);
2446         __jbd2_journal_file_buffer(jh, transaction, jlist);
2447         spin_unlock(&transaction->t_journal->j_list_lock);
2448         jbd_unlock_bh_state(jh2bh(jh));
2449 }
2450
2451 /*
2452  * Remove a buffer from its current buffer list in preparation for
2453  * dropping it from its current transaction entirely.  If the buffer has
2454  * already started to be used by a subsequent transaction, refile the
2455  * buffer on that transaction's metadata list.
2456  *
2457  * Called under j_list_lock
2458  * Called under jbd_lock_bh_state(jh2bh(jh))
2459  *
2460  * jh and bh may be already free when this function returns
2461  */
2462 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2463 {
2464         int was_dirty, jlist;
2465         struct buffer_head *bh = jh2bh(jh);
2466
2467         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2468         if (jh->b_transaction)
2469                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2470
2471         /* If the buffer is now unused, just drop it. */
2472         if (jh->b_next_transaction == NULL) {
2473                 __jbd2_journal_unfile_buffer(jh);
2474                 return;
2475         }
2476
2477         /*
2478          * It has been modified by a later transaction: add it to the new
2479          * transaction's metadata list.
2480          */
2481
2482         was_dirty = test_clear_buffer_jbddirty(bh);
2483         __jbd2_journal_temp_unlink_buffer(jh);
2484
2485         /*
2486          * b_transaction must be set, otherwise the new b_transaction won't
2487          * be holding jh reference
2488          */
2489         J_ASSERT_JH(jh, jh->b_transaction != NULL);
2490
2491         /*
2492          * We set b_transaction here because b_next_transaction will inherit
2493          * our jh reference and thus __jbd2_journal_file_buffer() must not
2494          * take a new one.
2495          */
2496         WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2497         WRITE_ONCE(jh->b_next_transaction, NULL);
2498         if (buffer_freed(bh))
2499                 jlist = BJ_Forget;
2500         else if (jh->b_modified)
2501                 jlist = BJ_Metadata;
2502         else
2503                 jlist = BJ_Reserved;
2504         __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2505         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2506
2507         if (was_dirty)
2508                 set_buffer_jbddirty(bh);
2509 }
2510
2511 /*
2512  * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2513  * bh reference so that we can safely unlock bh.
2514  *
2515  * The jh and bh may be freed by this call.
2516  */
2517 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2518 {
2519         struct buffer_head *bh = jh2bh(jh);
2520
2521         /* Get reference so that buffer cannot be freed before we unlock it */
2522         get_bh(bh);
2523         jbd_lock_bh_state(bh);
2524         spin_lock(&journal->j_list_lock);
2525         __jbd2_journal_refile_buffer(jh);
2526         jbd_unlock_bh_state(bh);
2527         spin_unlock(&journal->j_list_lock);
2528         __brelse(bh);
2529 }
2530
2531 /*
2532  * File inode in the inode list of the handle's transaction
2533  */
2534 static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2535                 unsigned long flags, loff_t start_byte, loff_t end_byte)
2536 {
2537         transaction_t *transaction = handle->h_transaction;
2538         journal_t *journal;
2539
2540         if (is_handle_aborted(handle))
2541                 return -EROFS;
2542         journal = transaction->t_journal;
2543
2544         jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2545                         transaction->t_tid);
2546
2547         spin_lock(&journal->j_list_lock);
2548         jinode->i_flags |= flags;
2549
2550         if (jinode->i_dirty_end) {
2551                 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2552                 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2553         } else {
2554                 jinode->i_dirty_start = start_byte;
2555                 jinode->i_dirty_end = end_byte;
2556         }
2557
2558         /* Is inode already attached where we need it? */
2559         if (jinode->i_transaction == transaction ||
2560             jinode->i_next_transaction == transaction)
2561                 goto done;
2562
2563         /*
2564          * We only ever set this variable to 1 so the test is safe. Since
2565          * t_need_data_flush is likely to be set, we do the test to save some
2566          * cacheline bouncing
2567          */
2568         if (!transaction->t_need_data_flush)
2569                 transaction->t_need_data_flush = 1;
2570         /* On some different transaction's list - should be
2571          * the committing one */
2572         if (jinode->i_transaction) {
2573                 J_ASSERT(jinode->i_next_transaction == NULL);
2574                 J_ASSERT(jinode->i_transaction ==
2575                                         journal->j_committing_transaction);
2576                 jinode->i_next_transaction = transaction;
2577                 goto done;
2578         }
2579         /* Not on any transaction list... */
2580         J_ASSERT(!jinode->i_next_transaction);
2581         jinode->i_transaction = transaction;
2582         list_add(&jinode->i_list, &transaction->t_inode_list);
2583 done:
2584         spin_unlock(&journal->j_list_lock);
2585
2586         return 0;
2587 }
2588
2589 int jbd2_journal_inode_add_write(handle_t *handle, struct jbd2_inode *jinode)
2590 {
2591         return jbd2_journal_file_inode(handle, jinode,
2592                         JI_WRITE_DATA | JI_WAIT_DATA, 0, LLONG_MAX);
2593 }
2594
2595 int jbd2_journal_inode_add_wait(handle_t *handle, struct jbd2_inode *jinode)
2596 {
2597         return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA, 0,
2598                         LLONG_MAX);
2599 }
2600
2601 int jbd2_journal_inode_ranged_write(handle_t *handle,
2602                 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2603 {
2604         return jbd2_journal_file_inode(handle, jinode,
2605                         JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2606                         start_byte + length - 1);
2607 }
2608
2609 int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2610                 loff_t start_byte, loff_t length)
2611 {
2612         return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2613                         start_byte, start_byte + length - 1);
2614 }
2615
2616 /*
2617  * File truncate and transaction commit interact with each other in a
2618  * non-trivial way.  If a transaction writing data block A is
2619  * committing, we cannot discard the data by truncate until we have
2620  * written them.  Otherwise if we crashed after the transaction with
2621  * write has committed but before the transaction with truncate has
2622  * committed, we could see stale data in block A.  This function is a
2623  * helper to solve this problem.  It starts writeout of the truncated
2624  * part in case it is in the committing transaction.
2625  *
2626  * Filesystem code must call this function when inode is journaled in
2627  * ordered mode before truncation happens and after the inode has been
2628  * placed on orphan list with the new inode size. The second condition
2629  * avoids the race that someone writes new data and we start
2630  * committing the transaction after this function has been called but
2631  * before a transaction for truncate is started (and furthermore it
2632  * allows us to optimize the case where the addition to orphan list
2633  * happens in the same transaction as write --- we don't have to write
2634  * any data in such case).
2635  */
2636 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2637                                         struct jbd2_inode *jinode,
2638                                         loff_t new_size)
2639 {
2640         transaction_t *inode_trans, *commit_trans;
2641         int ret = 0;
2642
2643         /* This is a quick check to avoid locking if not necessary */
2644         if (!jinode->i_transaction)
2645                 goto out;
2646         /* Locks are here just to force reading of recent values, it is
2647          * enough that the transaction was not committing before we started
2648          * a transaction adding the inode to orphan list */
2649         read_lock(&journal->j_state_lock);
2650         commit_trans = journal->j_committing_transaction;
2651         read_unlock(&journal->j_state_lock);
2652         spin_lock(&journal->j_list_lock);
2653         inode_trans = jinode->i_transaction;
2654         spin_unlock(&journal->j_list_lock);
2655         if (inode_trans == commit_trans) {
2656                 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2657                         new_size, LLONG_MAX);
2658                 if (ret)
2659                         jbd2_journal_abort(journal, ret);
2660         }
2661 out:
2662         return ret;
2663 }