Loading...
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
33static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
35
36static struct kmem_cache *transaction_cache;
37int __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 return 0;
47 return -ENOMEM;
48}
49
50void jbd2_journal_destroy_transaction_cache(void)
51{
52 if (transaction_cache) {
53 kmem_cache_destroy(transaction_cache);
54 transaction_cache = NULL;
55 }
56}
57
58void 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
80static transaction_t *
81jbd2_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 */
127static 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 */
147static 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
165static 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 */
177static 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
273static 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
299alloc_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 */
319repeat:
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... */
398static 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
409handle_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}
456EXPORT_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 */
478handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
479{
480 return jbd2__journal_start(journal, nblocks, 0, GFP_NOFS, 0, 0);
481}
482EXPORT_SYMBOL(jbd2_journal_start);
483
484void 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}
492EXPORT_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 */
508int 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}
543EXPORT_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 */
565int 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);
617unlock:
618 spin_unlock(&transaction->t_handle_lock);
619error_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 */
641int 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}
693EXPORT_SYMBOL(jbd2__journal_restart);
694
695
696int jbd2_journal_restart(handle_t *handle, int nblocks)
697{
698 return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
699}
700EXPORT_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 */
712void 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 */
769void 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
780static 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. */
790static 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 */
823static int
824do_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 if (is_handle_aborted(handle))
835 return -EROFS;
836 journal = transaction->t_journal;
837
838 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
839
840 JBUFFER_TRACE(jh, "entry");
841repeat:
842 bh = jh2bh(jh);
843
844 /* @@@ Need to check for errors here at some point. */
845
846 start_lock = jiffies;
847 lock_buffer(bh);
848 jbd_lock_bh_state(bh);
849
850 /* If it takes too long to lock the buffer, trace it */
851 time_lock = jbd2_time_diff(start_lock, jiffies);
852 if (time_lock > HZ/10)
853 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
854 jiffies_to_msecs(time_lock));
855
856 /* We now hold the buffer lock so it is safe to query the buffer
857 * state. Is the buffer dirty?
858 *
859 * If so, there are two possibilities. The buffer may be
860 * non-journaled, and undergoing a quite legitimate writeback.
861 * Otherwise, it is journaled, and we don't expect dirty buffers
862 * in that state (the buffers should be marked JBD_Dirty
863 * instead.) So either the IO is being done under our own
864 * control and this is a bug, or it's a third party IO such as
865 * dump(8) (which may leave the buffer scheduled for read ---
866 * ie. locked but not dirty) or tune2fs (which may actually have
867 * the buffer dirtied, ugh.) */
868
869 if (buffer_dirty(bh)) {
870 /*
871 * First question: is this buffer already part of the current
872 * transaction or the existing committing transaction?
873 */
874 if (jh->b_transaction) {
875 J_ASSERT_JH(jh,
876 jh->b_transaction == transaction ||
877 jh->b_transaction ==
878 journal->j_committing_transaction);
879 if (jh->b_next_transaction)
880 J_ASSERT_JH(jh, jh->b_next_transaction ==
881 transaction);
882 warn_dirty_buffer(bh);
883 }
884 /*
885 * In any case we need to clean the dirty flag and we must
886 * do it under the buffer lock to be sure we don't race
887 * with running write-out.
888 */
889 JBUFFER_TRACE(jh, "Journalling dirty buffer");
890 clear_buffer_dirty(bh);
891 set_buffer_jbddirty(bh);
892 }
893
894 unlock_buffer(bh);
895
896 error = -EROFS;
897 if (is_handle_aborted(handle)) {
898 jbd_unlock_bh_state(bh);
899 goto out;
900 }
901 error = 0;
902
903 /*
904 * The buffer is already part of this transaction if b_transaction or
905 * b_next_transaction points to it
906 */
907 if (jh->b_transaction == transaction ||
908 jh->b_next_transaction == transaction)
909 goto done;
910
911 /*
912 * this is the first time this transaction is touching this buffer,
913 * reset the modified flag
914 */
915 jh->b_modified = 0;
916
917 /*
918 * If the buffer is not journaled right now, we need to make sure it
919 * doesn't get written to disk before the caller actually commits the
920 * new data
921 */
922 if (!jh->b_transaction) {
923 JBUFFER_TRACE(jh, "no transaction");
924 J_ASSERT_JH(jh, !jh->b_next_transaction);
925 JBUFFER_TRACE(jh, "file as BJ_Reserved");
926 /*
927 * Make sure all stores to jh (b_modified, b_frozen_data) are
928 * visible before attaching it to the running transaction.
929 * Paired with barrier in jbd2_write_access_granted()
930 */
931 smp_wmb();
932 spin_lock(&journal->j_list_lock);
933 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
934 spin_unlock(&journal->j_list_lock);
935 goto done;
936 }
937 /*
938 * If there is already a copy-out version of this buffer, then we don't
939 * need to make another one
940 */
941 if (jh->b_frozen_data) {
942 JBUFFER_TRACE(jh, "has frozen data");
943 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
944 goto attach_next;
945 }
946
947 JBUFFER_TRACE(jh, "owned by older transaction");
948 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
949 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
950
951 /*
952 * There is one case we have to be very careful about. If the
953 * committing transaction is currently writing this buffer out to disk
954 * and has NOT made a copy-out, then we cannot modify the buffer
955 * contents at all right now. The essence of copy-out is that it is
956 * the extra copy, not the primary copy, which gets journaled. If the
957 * primary copy is already going to disk then we cannot do copy-out
958 * here.
959 */
960 if (buffer_shadow(bh)) {
961 JBUFFER_TRACE(jh, "on shadow: sleep");
962 jbd_unlock_bh_state(bh);
963 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
964 goto repeat;
965 }
966
967 /*
968 * Only do the copy if the currently-owning transaction still needs it.
969 * If buffer isn't on BJ_Metadata list, the committing transaction is
970 * past that stage (here we use the fact that BH_Shadow is set under
971 * bh_state lock together with refiling to BJ_Shadow list and at this
972 * point we know the buffer doesn't have BH_Shadow set).
973 *
974 * Subtle point, though: if this is a get_undo_access, then we will be
975 * relying on the frozen_data to contain the new value of the
976 * committed_data record after the transaction, so we HAVE to force the
977 * frozen_data copy in that case.
978 */
979 if (jh->b_jlist == BJ_Metadata || force_copy) {
980 JBUFFER_TRACE(jh, "generate frozen data");
981 if (!frozen_buffer) {
982 JBUFFER_TRACE(jh, "allocate memory for buffer");
983 jbd_unlock_bh_state(bh);
984 frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
985 GFP_NOFS | __GFP_NOFAIL);
986 goto repeat;
987 }
988 jh->b_frozen_data = frozen_buffer;
989 frozen_buffer = NULL;
990 jbd2_freeze_jh_data(jh);
991 }
992attach_next:
993 /*
994 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
995 * before attaching it to the running transaction. Paired with barrier
996 * in jbd2_write_access_granted()
997 */
998 smp_wmb();
999 jh->b_next_transaction = transaction;
1000
1001done:
1002 jbd_unlock_bh_state(bh);
1003
1004 /*
1005 * If we are about to journal a buffer, then any revoke pending on it is
1006 * no longer valid
1007 */
1008 jbd2_journal_cancel_revoke(handle, jh);
1009
1010out:
1011 if (unlikely(frozen_buffer)) /* It's usually NULL */
1012 jbd2_free(frozen_buffer, bh->b_size);
1013
1014 JBUFFER_TRACE(jh, "exit");
1015 return error;
1016}
1017
1018/* Fast check whether buffer is already attached to the required transaction */
1019static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1020 bool undo)
1021{
1022 struct journal_head *jh;
1023 bool ret = false;
1024
1025 /* Dirty buffers require special handling... */
1026 if (buffer_dirty(bh))
1027 return false;
1028
1029 /*
1030 * RCU protects us from dereferencing freed pages. So the checks we do
1031 * are guaranteed not to oops. However the jh slab object can get freed
1032 * & reallocated while we work with it. So we have to be careful. When
1033 * we see jh attached to the running transaction, we know it must stay
1034 * so until the transaction is committed. Thus jh won't be freed and
1035 * will be attached to the same bh while we run. However it can
1036 * happen jh gets freed, reallocated, and attached to the transaction
1037 * just after we get pointer to it from bh. So we have to be careful
1038 * and recheck jh still belongs to our bh before we return success.
1039 */
1040 rcu_read_lock();
1041 if (!buffer_jbd(bh))
1042 goto out;
1043 /* This should be bh2jh() but that doesn't work with inline functions */
1044 jh = READ_ONCE(bh->b_private);
1045 if (!jh)
1046 goto out;
1047 /* For undo access buffer must have data copied */
1048 if (undo && !jh->b_committed_data)
1049 goto out;
1050 if (jh->b_transaction != handle->h_transaction &&
1051 jh->b_next_transaction != handle->h_transaction)
1052 goto out;
1053 /*
1054 * There are two reasons for the barrier here:
1055 * 1) Make sure to fetch b_bh after we did previous checks so that we
1056 * detect when jh went through free, realloc, attach to transaction
1057 * while we were checking. Paired with implicit barrier in that path.
1058 * 2) So that access to bh done after jbd2_write_access_granted()
1059 * doesn't get reordered and see inconsistent state of concurrent
1060 * do_get_write_access().
1061 */
1062 smp_mb();
1063 if (unlikely(jh->b_bh != bh))
1064 goto out;
1065 ret = true;
1066out:
1067 rcu_read_unlock();
1068 return ret;
1069}
1070
1071/**
1072 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1073 * @handle: transaction to add buffer modifications to
1074 * @bh: bh to be used for metadata writes
1075 *
1076 * Returns: error code or 0 on success.
1077 *
1078 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1079 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1080 */
1081
1082int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1083{
1084 struct journal_head *jh;
1085 int rc;
1086
1087 if (jbd2_write_access_granted(handle, bh, false))
1088 return 0;
1089
1090 jh = jbd2_journal_add_journal_head(bh);
1091 /* We do not want to get caught playing with fields which the
1092 * log thread also manipulates. Make sure that the buffer
1093 * completes any outstanding IO before proceeding. */
1094 rc = do_get_write_access(handle, jh, 0);
1095 jbd2_journal_put_journal_head(jh);
1096 return rc;
1097}
1098
1099
1100/*
1101 * When the user wants to journal a newly created buffer_head
1102 * (ie. getblk() returned a new buffer and we are going to populate it
1103 * manually rather than reading off disk), then we need to keep the
1104 * buffer_head locked until it has been completely filled with new
1105 * data. In this case, we should be able to make the assertion that
1106 * the bh is not already part of an existing transaction.
1107 *
1108 * The buffer should already be locked by the caller by this point.
1109 * There is no lock ranking violation: it was a newly created,
1110 * unlocked buffer beforehand. */
1111
1112/**
1113 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1114 * @handle: transaction to new buffer to
1115 * @bh: new buffer.
1116 *
1117 * Call this if you create a new bh.
1118 */
1119int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1120{
1121 transaction_t *transaction = handle->h_transaction;
1122 journal_t *journal;
1123 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1124 int err;
1125
1126 jbd_debug(5, "journal_head %p\n", jh);
1127 err = -EROFS;
1128 if (is_handle_aborted(handle))
1129 goto out;
1130 journal = transaction->t_journal;
1131 err = 0;
1132
1133 JBUFFER_TRACE(jh, "entry");
1134 /*
1135 * The buffer may already belong to this transaction due to pre-zeroing
1136 * in the filesystem's new_block code. It may also be on the previous,
1137 * committing transaction's lists, but it HAS to be in Forget state in
1138 * that case: the transaction must have deleted the buffer for it to be
1139 * reused here.
1140 */
1141 jbd_lock_bh_state(bh);
1142 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1143 jh->b_transaction == NULL ||
1144 (jh->b_transaction == journal->j_committing_transaction &&
1145 jh->b_jlist == BJ_Forget)));
1146
1147 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1148 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1149
1150 if (jh->b_transaction == NULL) {
1151 /*
1152 * Previous jbd2_journal_forget() could have left the buffer
1153 * with jbddirty bit set because it was being committed. When
1154 * the commit finished, we've filed the buffer for
1155 * checkpointing and marked it dirty. Now we are reallocating
1156 * the buffer so the transaction freeing it must have
1157 * committed and so it's safe to clear the dirty bit.
1158 */
1159 clear_buffer_dirty(jh2bh(jh));
1160 /* first access by this transaction */
1161 jh->b_modified = 0;
1162
1163 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1164 spin_lock(&journal->j_list_lock);
1165 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1166 spin_unlock(&journal->j_list_lock);
1167 } else if (jh->b_transaction == journal->j_committing_transaction) {
1168 /* first access by this transaction */
1169 jh->b_modified = 0;
1170
1171 JBUFFER_TRACE(jh, "set next transaction");
1172 spin_lock(&journal->j_list_lock);
1173 jh->b_next_transaction = transaction;
1174 spin_unlock(&journal->j_list_lock);
1175 }
1176 jbd_unlock_bh_state(bh);
1177
1178 /*
1179 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1180 * blocks which contain freed but then revoked metadata. We need
1181 * to cancel the revoke in case we end up freeing it yet again
1182 * and the reallocating as data - this would cause a second revoke,
1183 * which hits an assertion error.
1184 */
1185 JBUFFER_TRACE(jh, "cancelling revoke");
1186 jbd2_journal_cancel_revoke(handle, jh);
1187out:
1188 jbd2_journal_put_journal_head(jh);
1189 return err;
1190}
1191
1192/**
1193 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1194 * non-rewindable consequences
1195 * @handle: transaction
1196 * @bh: buffer to undo
1197 *
1198 * Sometimes there is a need to distinguish between metadata which has
1199 * been committed to disk and that which has not. The ext3fs code uses
1200 * this for freeing and allocating space, we have to make sure that we
1201 * do not reuse freed space until the deallocation has been committed,
1202 * since if we overwrote that space we would make the delete
1203 * un-rewindable in case of a crash.
1204 *
1205 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1206 * buffer for parts of non-rewindable operations such as delete
1207 * operations on the bitmaps. The journaling code must keep a copy of
1208 * the buffer's contents prior to the undo_access call until such time
1209 * as we know that the buffer has definitely been committed to disk.
1210 *
1211 * We never need to know which transaction the committed data is part
1212 * of, buffers touched here are guaranteed to be dirtied later and so
1213 * will be committed to a new transaction in due course, at which point
1214 * we can discard the old committed data pointer.
1215 *
1216 * Returns error number or 0 on success.
1217 */
1218int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1219{
1220 int err;
1221 struct journal_head *jh;
1222 char *committed_data = NULL;
1223
1224 JBUFFER_TRACE(jh, "entry");
1225 if (jbd2_write_access_granted(handle, bh, true))
1226 return 0;
1227
1228 jh = jbd2_journal_add_journal_head(bh);
1229 /*
1230 * Do this first --- it can drop the journal lock, so we want to
1231 * make sure that obtaining the committed_data is done
1232 * atomically wrt. completion of any outstanding commits.
1233 */
1234 err = do_get_write_access(handle, jh, 1);
1235 if (err)
1236 goto out;
1237
1238repeat:
1239 if (!jh->b_committed_data)
1240 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1241 GFP_NOFS|__GFP_NOFAIL);
1242
1243 jbd_lock_bh_state(bh);
1244 if (!jh->b_committed_data) {
1245 /* Copy out the current buffer contents into the
1246 * preserved, committed copy. */
1247 JBUFFER_TRACE(jh, "generate b_committed data");
1248 if (!committed_data) {
1249 jbd_unlock_bh_state(bh);
1250 goto repeat;
1251 }
1252
1253 jh->b_committed_data = committed_data;
1254 committed_data = NULL;
1255 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1256 }
1257 jbd_unlock_bh_state(bh);
1258out:
1259 jbd2_journal_put_journal_head(jh);
1260 if (unlikely(committed_data))
1261 jbd2_free(committed_data, bh->b_size);
1262 return err;
1263}
1264
1265/**
1266 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1267 * @bh: buffer to trigger on
1268 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1269 *
1270 * Set any triggers on this journal_head. This is always safe, because
1271 * triggers for a committing buffer will be saved off, and triggers for
1272 * a running transaction will match the buffer in that transaction.
1273 *
1274 * Call with NULL to clear the triggers.
1275 */
1276void jbd2_journal_set_triggers(struct buffer_head *bh,
1277 struct jbd2_buffer_trigger_type *type)
1278{
1279 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1280
1281 if (WARN_ON(!jh))
1282 return;
1283 jh->b_triggers = type;
1284 jbd2_journal_put_journal_head(jh);
1285}
1286
1287void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1288 struct jbd2_buffer_trigger_type *triggers)
1289{
1290 struct buffer_head *bh = jh2bh(jh);
1291
1292 if (!triggers || !triggers->t_frozen)
1293 return;
1294
1295 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1296}
1297
1298void jbd2_buffer_abort_trigger(struct journal_head *jh,
1299 struct jbd2_buffer_trigger_type *triggers)
1300{
1301 if (!triggers || !triggers->t_abort)
1302 return;
1303
1304 triggers->t_abort(triggers, jh2bh(jh));
1305}
1306
1307/**
1308 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1309 * @handle: transaction to add buffer to.
1310 * @bh: buffer to mark
1311 *
1312 * mark dirty metadata which needs to be journaled as part of the current
1313 * transaction.
1314 *
1315 * The buffer must have previously had jbd2_journal_get_write_access()
1316 * called so that it has a valid journal_head attached to the buffer
1317 * head.
1318 *
1319 * The buffer is placed on the transaction's metadata list and is marked
1320 * as belonging to the transaction.
1321 *
1322 * Returns error number or 0 on success.
1323 *
1324 * Special care needs to be taken if the buffer already belongs to the
1325 * current committing transaction (in which case we should have frozen
1326 * data present for that commit). In that case, we don't relink the
1327 * buffer: that only gets done when the old transaction finally
1328 * completes its commit.
1329 */
1330int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1331{
1332 transaction_t *transaction = handle->h_transaction;
1333 journal_t *journal;
1334 struct journal_head *jh;
1335 int ret = 0;
1336
1337 if (is_handle_aborted(handle))
1338 return -EROFS;
1339 if (!buffer_jbd(bh)) {
1340 ret = -EUCLEAN;
1341 goto out;
1342 }
1343 /*
1344 * We don't grab jh reference here since the buffer must be part
1345 * of the running transaction.
1346 */
1347 jh = bh2jh(bh);
1348 /*
1349 * This and the following assertions are unreliable since we may see jh
1350 * in inconsistent state unless we grab bh_state lock. But this is
1351 * crucial to catch bugs so let's do a reliable check until the
1352 * lockless handling is fully proven.
1353 */
1354 if (jh->b_transaction != transaction &&
1355 jh->b_next_transaction != transaction) {
1356 jbd_lock_bh_state(bh);
1357 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1358 jh->b_next_transaction == transaction);
1359 jbd_unlock_bh_state(bh);
1360 }
1361 if (jh->b_modified == 1) {
1362 /* If it's in our transaction it must be in BJ_Metadata list. */
1363 if (jh->b_transaction == transaction &&
1364 jh->b_jlist != BJ_Metadata) {
1365 jbd_lock_bh_state(bh);
1366 J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1367 jh->b_jlist == BJ_Metadata);
1368 jbd_unlock_bh_state(bh);
1369 }
1370 goto out;
1371 }
1372
1373 journal = transaction->t_journal;
1374 jbd_debug(5, "journal_head %p\n", jh);
1375 JBUFFER_TRACE(jh, "entry");
1376
1377 jbd_lock_bh_state(bh);
1378
1379 if (jh->b_modified == 0) {
1380 /*
1381 * This buffer's got modified and becoming part
1382 * of the transaction. This needs to be done
1383 * once a transaction -bzzz
1384 */
1385 jh->b_modified = 1;
1386 if (handle->h_buffer_credits <= 0) {
1387 ret = -ENOSPC;
1388 goto out_unlock_bh;
1389 }
1390 handle->h_buffer_credits--;
1391 }
1392
1393 /*
1394 * fastpath, to avoid expensive locking. If this buffer is already
1395 * on the running transaction's metadata list there is nothing to do.
1396 * Nobody can take it off again because there is a handle open.
1397 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1398 * result in this test being false, so we go in and take the locks.
1399 */
1400 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1401 JBUFFER_TRACE(jh, "fastpath");
1402 if (unlikely(jh->b_transaction !=
1403 journal->j_running_transaction)) {
1404 printk(KERN_ERR "JBD2: %s: "
1405 "jh->b_transaction (%llu, %p, %u) != "
1406 "journal->j_running_transaction (%p, %u)\n",
1407 journal->j_devname,
1408 (unsigned long long) bh->b_blocknr,
1409 jh->b_transaction,
1410 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1411 journal->j_running_transaction,
1412 journal->j_running_transaction ?
1413 journal->j_running_transaction->t_tid : 0);
1414 ret = -EINVAL;
1415 }
1416 goto out_unlock_bh;
1417 }
1418
1419 set_buffer_jbddirty(bh);
1420
1421 /*
1422 * Metadata already on the current transaction list doesn't
1423 * need to be filed. Metadata on another transaction's list must
1424 * be committing, and will be refiled once the commit completes:
1425 * leave it alone for now.
1426 */
1427 if (jh->b_transaction != transaction) {
1428 JBUFFER_TRACE(jh, "already on other transaction");
1429 if (unlikely(((jh->b_transaction !=
1430 journal->j_committing_transaction)) ||
1431 (jh->b_next_transaction != transaction))) {
1432 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1433 "bad jh for block %llu: "
1434 "transaction (%p, %u), "
1435 "jh->b_transaction (%p, %u), "
1436 "jh->b_next_transaction (%p, %u), jlist %u\n",
1437 journal->j_devname,
1438 (unsigned long long) bh->b_blocknr,
1439 transaction, transaction->t_tid,
1440 jh->b_transaction,
1441 jh->b_transaction ?
1442 jh->b_transaction->t_tid : 0,
1443 jh->b_next_transaction,
1444 jh->b_next_transaction ?
1445 jh->b_next_transaction->t_tid : 0,
1446 jh->b_jlist);
1447 WARN_ON(1);
1448 ret = -EINVAL;
1449 }
1450 /* And this case is illegal: we can't reuse another
1451 * transaction's data buffer, ever. */
1452 goto out_unlock_bh;
1453 }
1454
1455 /* That test should have eliminated the following case: */
1456 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1457
1458 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1459 spin_lock(&journal->j_list_lock);
1460 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1461 spin_unlock(&journal->j_list_lock);
1462out_unlock_bh:
1463 jbd_unlock_bh_state(bh);
1464out:
1465 JBUFFER_TRACE(jh, "exit");
1466 return ret;
1467}
1468
1469/**
1470 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1471 * @handle: transaction handle
1472 * @bh: bh to 'forget'
1473 *
1474 * We can only do the bforget if there are no commits pending against the
1475 * buffer. If the buffer is dirty in the current running transaction we
1476 * can safely unlink it.
1477 *
1478 * bh may not be a journalled buffer at all - it may be a non-JBD
1479 * buffer which came off the hashtable. Check for this.
1480 *
1481 * Decrements bh->b_count by one.
1482 *
1483 * Allow this call even if the handle has aborted --- it may be part of
1484 * the caller's cleanup after an abort.
1485 */
1486int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1487{
1488 transaction_t *transaction = handle->h_transaction;
1489 journal_t *journal;
1490 struct journal_head *jh;
1491 int drop_reserve = 0;
1492 int err = 0;
1493 int was_modified = 0;
1494
1495 if (is_handle_aborted(handle))
1496 return -EROFS;
1497 journal = transaction->t_journal;
1498
1499 BUFFER_TRACE(bh, "entry");
1500
1501 jbd_lock_bh_state(bh);
1502
1503 if (!buffer_jbd(bh))
1504 goto not_jbd;
1505 jh = bh2jh(bh);
1506
1507 /* Critical error: attempting to delete a bitmap buffer, maybe?
1508 * Don't do any jbd operations, and return an error. */
1509 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1510 "inconsistent data on disk")) {
1511 err = -EIO;
1512 goto not_jbd;
1513 }
1514
1515 /* keep track of whether or not this transaction modified us */
1516 was_modified = jh->b_modified;
1517
1518 /*
1519 * The buffer's going from the transaction, we must drop
1520 * all references -bzzz
1521 */
1522 jh->b_modified = 0;
1523
1524 if (jh->b_transaction == transaction) {
1525 J_ASSERT_JH(jh, !jh->b_frozen_data);
1526
1527 /* If we are forgetting a buffer which is already part
1528 * of this transaction, then we can just drop it from
1529 * the transaction immediately. */
1530 clear_buffer_dirty(bh);
1531 clear_buffer_jbddirty(bh);
1532
1533 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1534
1535 /*
1536 * we only want to drop a reference if this transaction
1537 * modified the buffer
1538 */
1539 if (was_modified)
1540 drop_reserve = 1;
1541
1542 /*
1543 * We are no longer going to journal this buffer.
1544 * However, the commit of this transaction is still
1545 * important to the buffer: the delete that we are now
1546 * processing might obsolete an old log entry, so by
1547 * committing, we can satisfy the buffer's checkpoint.
1548 *
1549 * So, if we have a checkpoint on the buffer, we should
1550 * now refile the buffer on our BJ_Forget list so that
1551 * we know to remove the checkpoint after we commit.
1552 */
1553
1554 spin_lock(&journal->j_list_lock);
1555 if (jh->b_cp_transaction) {
1556 __jbd2_journal_temp_unlink_buffer(jh);
1557 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1558 } else {
1559 __jbd2_journal_unfile_buffer(jh);
1560 if (!buffer_jbd(bh)) {
1561 spin_unlock(&journal->j_list_lock);
1562 jbd_unlock_bh_state(bh);
1563 __bforget(bh);
1564 goto drop;
1565 }
1566 }
1567 spin_unlock(&journal->j_list_lock);
1568 } else if (jh->b_transaction) {
1569 J_ASSERT_JH(jh, (jh->b_transaction ==
1570 journal->j_committing_transaction));
1571 /* However, if the buffer is still owned by a prior
1572 * (committing) transaction, we can't drop it yet... */
1573 JBUFFER_TRACE(jh, "belongs to older transaction");
1574 /* ... but we CAN drop it from the new transaction if we
1575 * have also modified it since the original commit. */
1576
1577 if (jh->b_next_transaction) {
1578 J_ASSERT(jh->b_next_transaction == transaction);
1579 spin_lock(&journal->j_list_lock);
1580 jh->b_next_transaction = NULL;
1581 spin_unlock(&journal->j_list_lock);
1582
1583 /*
1584 * only drop a reference if this transaction modified
1585 * the buffer
1586 */
1587 if (was_modified)
1588 drop_reserve = 1;
1589 }
1590 }
1591
1592not_jbd:
1593 jbd_unlock_bh_state(bh);
1594 __brelse(bh);
1595drop:
1596 if (drop_reserve) {
1597 /* no need to reserve log space for this block -bzzz */
1598 handle->h_buffer_credits++;
1599 }
1600 return err;
1601}
1602
1603/**
1604 * int jbd2_journal_stop() - complete a transaction
1605 * @handle: transaction to complete.
1606 *
1607 * All done for a particular handle.
1608 *
1609 * There is not much action needed here. We just return any remaining
1610 * buffer credits to the transaction and remove the handle. The only
1611 * complication is that we need to start a commit operation if the
1612 * filesystem is marked for synchronous update.
1613 *
1614 * jbd2_journal_stop itself will not usually return an error, but it may
1615 * do so in unusual circumstances. In particular, expect it to
1616 * return -EIO if a jbd2_journal_abort has been executed since the
1617 * transaction began.
1618 */
1619int jbd2_journal_stop(handle_t *handle)
1620{
1621 transaction_t *transaction = handle->h_transaction;
1622 journal_t *journal;
1623 int err = 0, wait_for_commit = 0;
1624 tid_t tid;
1625 pid_t pid;
1626
1627 if (!transaction) {
1628 /*
1629 * Handle is already detached from the transaction so
1630 * there is nothing to do other than decrease a refcount,
1631 * or free the handle if refcount drops to zero
1632 */
1633 if (--handle->h_ref > 0) {
1634 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1635 handle->h_ref);
1636 return err;
1637 } else {
1638 if (handle->h_rsv_handle)
1639 jbd2_free_handle(handle->h_rsv_handle);
1640 goto free_and_exit;
1641 }
1642 }
1643 journal = transaction->t_journal;
1644
1645 J_ASSERT(journal_current_handle() == handle);
1646
1647 if (is_handle_aborted(handle))
1648 err = -EIO;
1649 else
1650 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1651
1652 if (--handle->h_ref > 0) {
1653 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1654 handle->h_ref);
1655 return err;
1656 }
1657
1658 jbd_debug(4, "Handle %p going down\n", handle);
1659 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1660 transaction->t_tid,
1661 handle->h_type, handle->h_line_no,
1662 jiffies - handle->h_start_jiffies,
1663 handle->h_sync, handle->h_requested_credits,
1664 (handle->h_requested_credits -
1665 handle->h_buffer_credits));
1666
1667 /*
1668 * Implement synchronous transaction batching. If the handle
1669 * was synchronous, don't force a commit immediately. Let's
1670 * yield and let another thread piggyback onto this
1671 * transaction. Keep doing that while new threads continue to
1672 * arrive. It doesn't cost much - we're about to run a commit
1673 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1674 * operations by 30x or more...
1675 *
1676 * We try and optimize the sleep time against what the
1677 * underlying disk can do, instead of having a static sleep
1678 * time. This is useful for the case where our storage is so
1679 * fast that it is more optimal to go ahead and force a flush
1680 * and wait for the transaction to be committed than it is to
1681 * wait for an arbitrary amount of time for new writers to
1682 * join the transaction. We achieve this by measuring how
1683 * long it takes to commit a transaction, and compare it with
1684 * how long this transaction has been running, and if run time
1685 * < commit time then we sleep for the delta and commit. This
1686 * greatly helps super fast disks that would see slowdowns as
1687 * more threads started doing fsyncs.
1688 *
1689 * But don't do this if this process was the most recent one
1690 * to perform a synchronous write. We do this to detect the
1691 * case where a single process is doing a stream of sync
1692 * writes. No point in waiting for joiners in that case.
1693 *
1694 * Setting max_batch_time to 0 disables this completely.
1695 */
1696 pid = current->pid;
1697 if (handle->h_sync && journal->j_last_sync_writer != pid &&
1698 journal->j_max_batch_time) {
1699 u64 commit_time, trans_time;
1700
1701 journal->j_last_sync_writer = pid;
1702
1703 read_lock(&journal->j_state_lock);
1704 commit_time = journal->j_average_commit_time;
1705 read_unlock(&journal->j_state_lock);
1706
1707 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1708 transaction->t_start_time));
1709
1710 commit_time = max_t(u64, commit_time,
1711 1000*journal->j_min_batch_time);
1712 commit_time = min_t(u64, commit_time,
1713 1000*journal->j_max_batch_time);
1714
1715 if (trans_time < commit_time) {
1716 ktime_t expires = ktime_add_ns(ktime_get(),
1717 commit_time);
1718 set_current_state(TASK_UNINTERRUPTIBLE);
1719 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1720 }
1721 }
1722
1723 if (handle->h_sync)
1724 transaction->t_synchronous_commit = 1;
1725 current->journal_info = NULL;
1726 atomic_sub(handle->h_buffer_credits,
1727 &transaction->t_outstanding_credits);
1728
1729 /*
1730 * If the handle is marked SYNC, we need to set another commit
1731 * going! We also want to force a commit if the current
1732 * transaction is occupying too much of the log, or if the
1733 * transaction is too old now.
1734 */
1735 if (handle->h_sync ||
1736 (atomic_read(&transaction->t_outstanding_credits) >
1737 journal->j_max_transaction_buffers) ||
1738 time_after_eq(jiffies, transaction->t_expires)) {
1739 /* Do this even for aborted journals: an abort still
1740 * completes the commit thread, it just doesn't write
1741 * anything to disk. */
1742
1743 jbd_debug(2, "transaction too old, requesting commit for "
1744 "handle %p\n", handle);
1745 /* This is non-blocking */
1746 jbd2_log_start_commit(journal, transaction->t_tid);
1747
1748 /*
1749 * Special case: JBD2_SYNC synchronous updates require us
1750 * to wait for the commit to complete.
1751 */
1752 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1753 wait_for_commit = 1;
1754 }
1755
1756 /*
1757 * Once we drop t_updates, if it goes to zero the transaction
1758 * could start committing on us and eventually disappear. So
1759 * once we do this, we must not dereference transaction
1760 * pointer again.
1761 */
1762 tid = transaction->t_tid;
1763 if (atomic_dec_and_test(&transaction->t_updates)) {
1764 wake_up(&journal->j_wait_updates);
1765 if (journal->j_barrier_count)
1766 wake_up(&journal->j_wait_transaction_locked);
1767 }
1768
1769 rwsem_release(&journal->j_trans_commit_map, 1, _THIS_IP_);
1770
1771 if (wait_for_commit)
1772 err = jbd2_log_wait_commit(journal, tid);
1773
1774 if (handle->h_rsv_handle)
1775 jbd2_journal_free_reserved(handle->h_rsv_handle);
1776free_and_exit:
1777 /*
1778 * Scope of the GFP_NOFS context is over here and so we can restore the
1779 * original alloc context.
1780 */
1781 memalloc_nofs_restore(handle->saved_alloc_context);
1782 jbd2_free_handle(handle);
1783 return err;
1784}
1785
1786/*
1787 *
1788 * List management code snippets: various functions for manipulating the
1789 * transaction buffer lists.
1790 *
1791 */
1792
1793/*
1794 * Append a buffer to a transaction list, given the transaction's list head
1795 * pointer.
1796 *
1797 * j_list_lock is held.
1798 *
1799 * jbd_lock_bh_state(jh2bh(jh)) is held.
1800 */
1801
1802static inline void
1803__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1804{
1805 if (!*list) {
1806 jh->b_tnext = jh->b_tprev = jh;
1807 *list = jh;
1808 } else {
1809 /* Insert at the tail of the list to preserve order */
1810 struct journal_head *first = *list, *last = first->b_tprev;
1811 jh->b_tprev = last;
1812 jh->b_tnext = first;
1813 last->b_tnext = first->b_tprev = jh;
1814 }
1815}
1816
1817/*
1818 * Remove a buffer from a transaction list, given the transaction's list
1819 * head pointer.
1820 *
1821 * Called with j_list_lock held, and the journal may not be locked.
1822 *
1823 * jbd_lock_bh_state(jh2bh(jh)) is held.
1824 */
1825
1826static inline void
1827__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1828{
1829 if (*list == jh) {
1830 *list = jh->b_tnext;
1831 if (*list == jh)
1832 *list = NULL;
1833 }
1834 jh->b_tprev->b_tnext = jh->b_tnext;
1835 jh->b_tnext->b_tprev = jh->b_tprev;
1836}
1837
1838/*
1839 * Remove a buffer from the appropriate transaction list.
1840 *
1841 * Note that this function can *change* the value of
1842 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1843 * t_reserved_list. If the caller is holding onto a copy of one of these
1844 * pointers, it could go bad. Generally the caller needs to re-read the
1845 * pointer from the transaction_t.
1846 *
1847 * Called under j_list_lock.
1848 */
1849static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1850{
1851 struct journal_head **list = NULL;
1852 transaction_t *transaction;
1853 struct buffer_head *bh = jh2bh(jh);
1854
1855 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1856 transaction = jh->b_transaction;
1857 if (transaction)
1858 assert_spin_locked(&transaction->t_journal->j_list_lock);
1859
1860 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1861 if (jh->b_jlist != BJ_None)
1862 J_ASSERT_JH(jh, transaction != NULL);
1863
1864 switch (jh->b_jlist) {
1865 case BJ_None:
1866 return;
1867 case BJ_Metadata:
1868 transaction->t_nr_buffers--;
1869 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1870 list = &transaction->t_buffers;
1871 break;
1872 case BJ_Forget:
1873 list = &transaction->t_forget;
1874 break;
1875 case BJ_Shadow:
1876 list = &transaction->t_shadow_list;
1877 break;
1878 case BJ_Reserved:
1879 list = &transaction->t_reserved_list;
1880 break;
1881 }
1882
1883 __blist_del_buffer(list, jh);
1884 jh->b_jlist = BJ_None;
1885 if (transaction && is_journal_aborted(transaction->t_journal))
1886 clear_buffer_jbddirty(bh);
1887 else if (test_clear_buffer_jbddirty(bh))
1888 mark_buffer_dirty(bh); /* Expose it to the VM */
1889}
1890
1891/*
1892 * Remove buffer from all transactions.
1893 *
1894 * Called with bh_state lock and j_list_lock
1895 *
1896 * jh and bh may be already freed when this function returns.
1897 */
1898static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1899{
1900 __jbd2_journal_temp_unlink_buffer(jh);
1901 jh->b_transaction = NULL;
1902 jbd2_journal_put_journal_head(jh);
1903}
1904
1905void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1906{
1907 struct buffer_head *bh = jh2bh(jh);
1908
1909 /* Get reference so that buffer cannot be freed before we unlock it */
1910 get_bh(bh);
1911 jbd_lock_bh_state(bh);
1912 spin_lock(&journal->j_list_lock);
1913 __jbd2_journal_unfile_buffer(jh);
1914 spin_unlock(&journal->j_list_lock);
1915 jbd_unlock_bh_state(bh);
1916 __brelse(bh);
1917}
1918
1919/*
1920 * Called from jbd2_journal_try_to_free_buffers().
1921 *
1922 * Called under jbd_lock_bh_state(bh)
1923 */
1924static void
1925__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1926{
1927 struct journal_head *jh;
1928
1929 jh = bh2jh(bh);
1930
1931 if (buffer_locked(bh) || buffer_dirty(bh))
1932 goto out;
1933
1934 if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
1935 goto out;
1936
1937 spin_lock(&journal->j_list_lock);
1938 if (jh->b_cp_transaction != NULL) {
1939 /* written-back checkpointed metadata buffer */
1940 JBUFFER_TRACE(jh, "remove from checkpoint list");
1941 __jbd2_journal_remove_checkpoint(jh);
1942 }
1943 spin_unlock(&journal->j_list_lock);
1944out:
1945 return;
1946}
1947
1948/**
1949 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1950 * @journal: journal for operation
1951 * @page: to try and free
1952 * @gfp_mask: we use the mask to detect how hard should we try to release
1953 * buffers. If __GFP_DIRECT_RECLAIM and __GFP_FS is set, we wait for commit
1954 * code to release the buffers.
1955 *
1956 *
1957 * For all the buffers on this page,
1958 * if they are fully written out ordered data, move them onto BUF_CLEAN
1959 * so try_to_free_buffers() can reap them.
1960 *
1961 * This function returns non-zero if we wish try_to_free_buffers()
1962 * to be called. We do this if the page is releasable by try_to_free_buffers().
1963 * We also do it if the page has locked or dirty buffers and the caller wants
1964 * us to perform sync or async writeout.
1965 *
1966 * This complicates JBD locking somewhat. We aren't protected by the
1967 * BKL here. We wish to remove the buffer from its committing or
1968 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1969 *
1970 * This may *change* the value of transaction_t->t_datalist, so anyone
1971 * who looks at t_datalist needs to lock against this function.
1972 *
1973 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1974 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
1975 * will come out of the lock with the buffer dirty, which makes it
1976 * ineligible for release here.
1977 *
1978 * Who else is affected by this? hmm... Really the only contender
1979 * is do_get_write_access() - it could be looking at the buffer while
1980 * journal_try_to_free_buffer() is changing its state. But that
1981 * cannot happen because we never reallocate freed data as metadata
1982 * while the data is part of a transaction. Yes?
1983 *
1984 * Return 0 on failure, 1 on success
1985 */
1986int jbd2_journal_try_to_free_buffers(journal_t *journal,
1987 struct page *page, gfp_t gfp_mask)
1988{
1989 struct buffer_head *head;
1990 struct buffer_head *bh;
1991 int ret = 0;
1992
1993 J_ASSERT(PageLocked(page));
1994
1995 head = page_buffers(page);
1996 bh = head;
1997 do {
1998 struct journal_head *jh;
1999
2000 /*
2001 * We take our own ref against the journal_head here to avoid
2002 * having to add tons of locking around each instance of
2003 * jbd2_journal_put_journal_head().
2004 */
2005 jh = jbd2_journal_grab_journal_head(bh);
2006 if (!jh)
2007 continue;
2008
2009 jbd_lock_bh_state(bh);
2010 __journal_try_to_free_buffer(journal, bh);
2011 jbd2_journal_put_journal_head(jh);
2012 jbd_unlock_bh_state(bh);
2013 if (buffer_jbd(bh))
2014 goto busy;
2015 } while ((bh = bh->b_this_page) != head);
2016
2017 ret = try_to_free_buffers(page);
2018
2019busy:
2020 return ret;
2021}
2022
2023/*
2024 * This buffer is no longer needed. If it is on an older transaction's
2025 * checkpoint list we need to record it on this transaction's forget list
2026 * to pin this buffer (and hence its checkpointing transaction) down until
2027 * this transaction commits. If the buffer isn't on a checkpoint list, we
2028 * release it.
2029 * Returns non-zero if JBD no longer has an interest in the buffer.
2030 *
2031 * Called under j_list_lock.
2032 *
2033 * Called under jbd_lock_bh_state(bh).
2034 */
2035static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2036{
2037 int may_free = 1;
2038 struct buffer_head *bh = jh2bh(jh);
2039
2040 if (jh->b_cp_transaction) {
2041 JBUFFER_TRACE(jh, "on running+cp transaction");
2042 __jbd2_journal_temp_unlink_buffer(jh);
2043 /*
2044 * We don't want to write the buffer anymore, clear the
2045 * bit so that we don't confuse checks in
2046 * __journal_file_buffer
2047 */
2048 clear_buffer_dirty(bh);
2049 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2050 may_free = 0;
2051 } else {
2052 JBUFFER_TRACE(jh, "on running transaction");
2053 __jbd2_journal_unfile_buffer(jh);
2054 }
2055 return may_free;
2056}
2057
2058/*
2059 * jbd2_journal_invalidatepage
2060 *
2061 * This code is tricky. It has a number of cases to deal with.
2062 *
2063 * There are two invariants which this code relies on:
2064 *
2065 * i_size must be updated on disk before we start calling invalidatepage on the
2066 * data.
2067 *
2068 * This is done in ext3 by defining an ext3_setattr method which
2069 * updates i_size before truncate gets going. By maintaining this
2070 * invariant, we can be sure that it is safe to throw away any buffers
2071 * attached to the current transaction: once the transaction commits,
2072 * we know that the data will not be needed.
2073 *
2074 * Note however that we can *not* throw away data belonging to the
2075 * previous, committing transaction!
2076 *
2077 * Any disk blocks which *are* part of the previous, committing
2078 * transaction (and which therefore cannot be discarded immediately) are
2079 * not going to be reused in the new running transaction
2080 *
2081 * The bitmap committed_data images guarantee this: any block which is
2082 * allocated in one transaction and removed in the next will be marked
2083 * as in-use in the committed_data bitmap, so cannot be reused until
2084 * the next transaction to delete the block commits. This means that
2085 * leaving committing buffers dirty is quite safe: the disk blocks
2086 * cannot be reallocated to a different file and so buffer aliasing is
2087 * not possible.
2088 *
2089 *
2090 * The above applies mainly to ordered data mode. In writeback mode we
2091 * don't make guarantees about the order in which data hits disk --- in
2092 * particular we don't guarantee that new dirty data is flushed before
2093 * transaction commit --- so it is always safe just to discard data
2094 * immediately in that mode. --sct
2095 */
2096
2097/*
2098 * The journal_unmap_buffer helper function returns zero if the buffer
2099 * concerned remains pinned as an anonymous buffer belonging to an older
2100 * transaction.
2101 *
2102 * We're outside-transaction here. Either or both of j_running_transaction
2103 * and j_committing_transaction may be NULL.
2104 */
2105static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2106 int partial_page)
2107{
2108 transaction_t *transaction;
2109 struct journal_head *jh;
2110 int may_free = 1;
2111
2112 BUFFER_TRACE(bh, "entry");
2113
2114 /*
2115 * It is safe to proceed here without the j_list_lock because the
2116 * buffers cannot be stolen by try_to_free_buffers as long as we are
2117 * holding the page lock. --sct
2118 */
2119
2120 if (!buffer_jbd(bh))
2121 goto zap_buffer_unlocked;
2122
2123 /* OK, we have data buffer in journaled mode */
2124 write_lock(&journal->j_state_lock);
2125 jbd_lock_bh_state(bh);
2126 spin_lock(&journal->j_list_lock);
2127
2128 jh = jbd2_journal_grab_journal_head(bh);
2129 if (!jh)
2130 goto zap_buffer_no_jh;
2131
2132 /*
2133 * We cannot remove the buffer from checkpoint lists until the
2134 * transaction adding inode to orphan list (let's call it T)
2135 * is committed. Otherwise if the transaction changing the
2136 * buffer would be cleaned from the journal before T is
2137 * committed, a crash will cause that the correct contents of
2138 * the buffer will be lost. On the other hand we have to
2139 * clear the buffer dirty bit at latest at the moment when the
2140 * transaction marking the buffer as freed in the filesystem
2141 * structures is committed because from that moment on the
2142 * block can be reallocated and used by a different page.
2143 * Since the block hasn't been freed yet but the inode has
2144 * already been added to orphan list, it is safe for us to add
2145 * the buffer to BJ_Forget list of the newest transaction.
2146 *
2147 * Also we have to clear buffer_mapped flag of a truncated buffer
2148 * because the buffer_head may be attached to the page straddling
2149 * i_size (can happen only when blocksize < pagesize) and thus the
2150 * buffer_head can be reused when the file is extended again. So we end
2151 * up keeping around invalidated buffers attached to transactions'
2152 * BJ_Forget list just to stop checkpointing code from cleaning up
2153 * the transaction this buffer was modified in.
2154 */
2155 transaction = jh->b_transaction;
2156 if (transaction == NULL) {
2157 /* First case: not on any transaction. If it
2158 * has no checkpoint link, then we can zap it:
2159 * it's a writeback-mode buffer so we don't care
2160 * if it hits disk safely. */
2161 if (!jh->b_cp_transaction) {
2162 JBUFFER_TRACE(jh, "not on any transaction: zap");
2163 goto zap_buffer;
2164 }
2165
2166 if (!buffer_dirty(bh)) {
2167 /* bdflush has written it. We can drop it now */
2168 __jbd2_journal_remove_checkpoint(jh);
2169 goto zap_buffer;
2170 }
2171
2172 /* OK, it must be in the journal but still not
2173 * written fully to disk: it's metadata or
2174 * journaled data... */
2175
2176 if (journal->j_running_transaction) {
2177 /* ... and once the current transaction has
2178 * committed, the buffer won't be needed any
2179 * longer. */
2180 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2181 may_free = __dispose_buffer(jh,
2182 journal->j_running_transaction);
2183 goto zap_buffer;
2184 } else {
2185 /* There is no currently-running transaction. So the
2186 * orphan record which we wrote for this file must have
2187 * passed into commit. We must attach this buffer to
2188 * the committing transaction, if it exists. */
2189 if (journal->j_committing_transaction) {
2190 JBUFFER_TRACE(jh, "give to committing trans");
2191 may_free = __dispose_buffer(jh,
2192 journal->j_committing_transaction);
2193 goto zap_buffer;
2194 } else {
2195 /* The orphan record's transaction has
2196 * committed. We can cleanse this buffer */
2197 clear_buffer_jbddirty(bh);
2198 __jbd2_journal_remove_checkpoint(jh);
2199 goto zap_buffer;
2200 }
2201 }
2202 } else if (transaction == journal->j_committing_transaction) {
2203 JBUFFER_TRACE(jh, "on committing transaction");
2204 /*
2205 * The buffer is committing, we simply cannot touch
2206 * it. If the page is straddling i_size we have to wait
2207 * for commit and try again.
2208 */
2209 if (partial_page) {
2210 jbd2_journal_put_journal_head(jh);
2211 spin_unlock(&journal->j_list_lock);
2212 jbd_unlock_bh_state(bh);
2213 write_unlock(&journal->j_state_lock);
2214 return -EBUSY;
2215 }
2216 /*
2217 * OK, buffer won't be reachable after truncate. We just set
2218 * j_next_transaction to the running transaction (if there is
2219 * one) and mark buffer as freed so that commit code knows it
2220 * should clear dirty bits when it is done with the buffer.
2221 */
2222 set_buffer_freed(bh);
2223 if (journal->j_running_transaction && buffer_jbddirty(bh))
2224 jh->b_next_transaction = journal->j_running_transaction;
2225 jbd2_journal_put_journal_head(jh);
2226 spin_unlock(&journal->j_list_lock);
2227 jbd_unlock_bh_state(bh);
2228 write_unlock(&journal->j_state_lock);
2229 return 0;
2230 } else {
2231 /* Good, the buffer belongs to the running transaction.
2232 * We are writing our own transaction's data, not any
2233 * previous one's, so it is safe to throw it away
2234 * (remember that we expect the filesystem to have set
2235 * i_size already for this truncate so recovery will not
2236 * expose the disk blocks we are discarding here.) */
2237 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2238 JBUFFER_TRACE(jh, "on running transaction");
2239 may_free = __dispose_buffer(jh, transaction);
2240 }
2241
2242zap_buffer:
2243 /*
2244 * This is tricky. Although the buffer is truncated, it may be reused
2245 * if blocksize < pagesize and it is attached to the page straddling
2246 * EOF. Since the buffer might have been added to BJ_Forget list of the
2247 * running transaction, journal_get_write_access() won't clear
2248 * b_modified and credit accounting gets confused. So clear b_modified
2249 * here.
2250 */
2251 jh->b_modified = 0;
2252 jbd2_journal_put_journal_head(jh);
2253zap_buffer_no_jh:
2254 spin_unlock(&journal->j_list_lock);
2255 jbd_unlock_bh_state(bh);
2256 write_unlock(&journal->j_state_lock);
2257zap_buffer_unlocked:
2258 clear_buffer_dirty(bh);
2259 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2260 clear_buffer_mapped(bh);
2261 clear_buffer_req(bh);
2262 clear_buffer_new(bh);
2263 clear_buffer_delay(bh);
2264 clear_buffer_unwritten(bh);
2265 bh->b_bdev = NULL;
2266 return may_free;
2267}
2268
2269/**
2270 * void jbd2_journal_invalidatepage()
2271 * @journal: journal to use for flush...
2272 * @page: page to flush
2273 * @offset: start of the range to invalidate
2274 * @length: length of the range to invalidate
2275 *
2276 * Reap page buffers containing data after in the specified range in page.
2277 * Can return -EBUSY if buffers are part of the committing transaction and
2278 * the page is straddling i_size. Caller then has to wait for current commit
2279 * and try again.
2280 */
2281int jbd2_journal_invalidatepage(journal_t *journal,
2282 struct page *page,
2283 unsigned int offset,
2284 unsigned int length)
2285{
2286 struct buffer_head *head, *bh, *next;
2287 unsigned int stop = offset + length;
2288 unsigned int curr_off = 0;
2289 int partial_page = (offset || length < PAGE_SIZE);
2290 int may_free = 1;
2291 int ret = 0;
2292
2293 if (!PageLocked(page))
2294 BUG();
2295 if (!page_has_buffers(page))
2296 return 0;
2297
2298 BUG_ON(stop > PAGE_SIZE || stop < length);
2299
2300 /* We will potentially be playing with lists other than just the
2301 * data lists (especially for journaled data mode), so be
2302 * cautious in our locking. */
2303
2304 head = bh = page_buffers(page);
2305 do {
2306 unsigned int next_off = curr_off + bh->b_size;
2307 next = bh->b_this_page;
2308
2309 if (next_off > stop)
2310 return 0;
2311
2312 if (offset <= curr_off) {
2313 /* This block is wholly outside the truncation point */
2314 lock_buffer(bh);
2315 ret = journal_unmap_buffer(journal, bh, partial_page);
2316 unlock_buffer(bh);
2317 if (ret < 0)
2318 return ret;
2319 may_free &= ret;
2320 }
2321 curr_off = next_off;
2322 bh = next;
2323
2324 } while (bh != head);
2325
2326 if (!partial_page) {
2327 if (may_free && try_to_free_buffers(page))
2328 J_ASSERT(!page_has_buffers(page));
2329 }
2330 return 0;
2331}
2332
2333/*
2334 * File a buffer on the given transaction list.
2335 */
2336void __jbd2_journal_file_buffer(struct journal_head *jh,
2337 transaction_t *transaction, int jlist)
2338{
2339 struct journal_head **list = NULL;
2340 int was_dirty = 0;
2341 struct buffer_head *bh = jh2bh(jh);
2342
2343 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2344 assert_spin_locked(&transaction->t_journal->j_list_lock);
2345
2346 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2347 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2348 jh->b_transaction == NULL);
2349
2350 if (jh->b_transaction && jh->b_jlist == jlist)
2351 return;
2352
2353 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2354 jlist == BJ_Shadow || jlist == BJ_Forget) {
2355 /*
2356 * For metadata buffers, we track dirty bit in buffer_jbddirty
2357 * instead of buffer_dirty. We should not see a dirty bit set
2358 * here because we clear it in do_get_write_access but e.g.
2359 * tune2fs can modify the sb and set the dirty bit at any time
2360 * so we try to gracefully handle that.
2361 */
2362 if (buffer_dirty(bh))
2363 warn_dirty_buffer(bh);
2364 if (test_clear_buffer_dirty(bh) ||
2365 test_clear_buffer_jbddirty(bh))
2366 was_dirty = 1;
2367 }
2368
2369 if (jh->b_transaction)
2370 __jbd2_journal_temp_unlink_buffer(jh);
2371 else
2372 jbd2_journal_grab_journal_head(bh);
2373 jh->b_transaction = transaction;
2374
2375 switch (jlist) {
2376 case BJ_None:
2377 J_ASSERT_JH(jh, !jh->b_committed_data);
2378 J_ASSERT_JH(jh, !jh->b_frozen_data);
2379 return;
2380 case BJ_Metadata:
2381 transaction->t_nr_buffers++;
2382 list = &transaction->t_buffers;
2383 break;
2384 case BJ_Forget:
2385 list = &transaction->t_forget;
2386 break;
2387 case BJ_Shadow:
2388 list = &transaction->t_shadow_list;
2389 break;
2390 case BJ_Reserved:
2391 list = &transaction->t_reserved_list;
2392 break;
2393 }
2394
2395 __blist_add_buffer(list, jh);
2396 jh->b_jlist = jlist;
2397
2398 if (was_dirty)
2399 set_buffer_jbddirty(bh);
2400}
2401
2402void jbd2_journal_file_buffer(struct journal_head *jh,
2403 transaction_t *transaction, int jlist)
2404{
2405 jbd_lock_bh_state(jh2bh(jh));
2406 spin_lock(&transaction->t_journal->j_list_lock);
2407 __jbd2_journal_file_buffer(jh, transaction, jlist);
2408 spin_unlock(&transaction->t_journal->j_list_lock);
2409 jbd_unlock_bh_state(jh2bh(jh));
2410}
2411
2412/*
2413 * Remove a buffer from its current buffer list in preparation for
2414 * dropping it from its current transaction entirely. If the buffer has
2415 * already started to be used by a subsequent transaction, refile the
2416 * buffer on that transaction's metadata list.
2417 *
2418 * Called under j_list_lock
2419 * Called under jbd_lock_bh_state(jh2bh(jh))
2420 *
2421 * jh and bh may be already free when this function returns
2422 */
2423void __jbd2_journal_refile_buffer(struct journal_head *jh)
2424{
2425 int was_dirty, jlist;
2426 struct buffer_head *bh = jh2bh(jh);
2427
2428 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2429 if (jh->b_transaction)
2430 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2431
2432 /* If the buffer is now unused, just drop it. */
2433 if (jh->b_next_transaction == NULL) {
2434 __jbd2_journal_unfile_buffer(jh);
2435 return;
2436 }
2437
2438 /*
2439 * It has been modified by a later transaction: add it to the new
2440 * transaction's metadata list.
2441 */
2442
2443 was_dirty = test_clear_buffer_jbddirty(bh);
2444 __jbd2_journal_temp_unlink_buffer(jh);
2445 /*
2446 * We set b_transaction here because b_next_transaction will inherit
2447 * our jh reference and thus __jbd2_journal_file_buffer() must not
2448 * take a new one.
2449 */
2450 jh->b_transaction = jh->b_next_transaction;
2451 jh->b_next_transaction = NULL;
2452 if (buffer_freed(bh))
2453 jlist = BJ_Forget;
2454 else if (jh->b_modified)
2455 jlist = BJ_Metadata;
2456 else
2457 jlist = BJ_Reserved;
2458 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2459 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2460
2461 if (was_dirty)
2462 set_buffer_jbddirty(bh);
2463}
2464
2465/*
2466 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2467 * bh reference so that we can safely unlock bh.
2468 *
2469 * The jh and bh may be freed by this call.
2470 */
2471void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2472{
2473 struct buffer_head *bh = jh2bh(jh);
2474
2475 /* Get reference so that buffer cannot be freed before we unlock it */
2476 get_bh(bh);
2477 jbd_lock_bh_state(bh);
2478 spin_lock(&journal->j_list_lock);
2479 __jbd2_journal_refile_buffer(jh);
2480 jbd_unlock_bh_state(bh);
2481 spin_unlock(&journal->j_list_lock);
2482 __brelse(bh);
2483}
2484
2485/*
2486 * File inode in the inode list of the handle's transaction
2487 */
2488static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2489 unsigned long flags)
2490{
2491 transaction_t *transaction = handle->h_transaction;
2492 journal_t *journal;
2493
2494 if (is_handle_aborted(handle))
2495 return -EROFS;
2496 journal = transaction->t_journal;
2497
2498 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2499 transaction->t_tid);
2500
2501 /*
2502 * First check whether inode isn't already on the transaction's
2503 * lists without taking the lock. Note that this check is safe
2504 * without the lock as we cannot race with somebody removing inode
2505 * from the transaction. The reason is that we remove inode from the
2506 * transaction only in journal_release_jbd_inode() and when we commit
2507 * the transaction. We are guarded from the first case by holding
2508 * a reference to the inode. We are safe against the second case
2509 * because if jinode->i_transaction == transaction, commit code
2510 * cannot touch the transaction because we hold reference to it,
2511 * and if jinode->i_next_transaction == transaction, commit code
2512 * will only file the inode where we want it.
2513 */
2514 if ((jinode->i_transaction == transaction ||
2515 jinode->i_next_transaction == transaction) &&
2516 (jinode->i_flags & flags) == flags)
2517 return 0;
2518
2519 spin_lock(&journal->j_list_lock);
2520 jinode->i_flags |= flags;
2521 /* Is inode already attached where we need it? */
2522 if (jinode->i_transaction == transaction ||
2523 jinode->i_next_transaction == transaction)
2524 goto done;
2525
2526 /*
2527 * We only ever set this variable to 1 so the test is safe. Since
2528 * t_need_data_flush is likely to be set, we do the test to save some
2529 * cacheline bouncing
2530 */
2531 if (!transaction->t_need_data_flush)
2532 transaction->t_need_data_flush = 1;
2533 /* On some different transaction's list - should be
2534 * the committing one */
2535 if (jinode->i_transaction) {
2536 J_ASSERT(jinode->i_next_transaction == NULL);
2537 J_ASSERT(jinode->i_transaction ==
2538 journal->j_committing_transaction);
2539 jinode->i_next_transaction = transaction;
2540 goto done;
2541 }
2542 /* Not on any transaction list... */
2543 J_ASSERT(!jinode->i_next_transaction);
2544 jinode->i_transaction = transaction;
2545 list_add(&jinode->i_list, &transaction->t_inode_list);
2546done:
2547 spin_unlock(&journal->j_list_lock);
2548
2549 return 0;
2550}
2551
2552int jbd2_journal_inode_add_write(handle_t *handle, struct jbd2_inode *jinode)
2553{
2554 return jbd2_journal_file_inode(handle, jinode,
2555 JI_WRITE_DATA | JI_WAIT_DATA);
2556}
2557
2558int jbd2_journal_inode_add_wait(handle_t *handle, struct jbd2_inode *jinode)
2559{
2560 return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA);
2561}
2562
2563/*
2564 * File truncate and transaction commit interact with each other in a
2565 * non-trivial way. If a transaction writing data block A is
2566 * committing, we cannot discard the data by truncate until we have
2567 * written them. Otherwise if we crashed after the transaction with
2568 * write has committed but before the transaction with truncate has
2569 * committed, we could see stale data in block A. This function is a
2570 * helper to solve this problem. It starts writeout of the truncated
2571 * part in case it is in the committing transaction.
2572 *
2573 * Filesystem code must call this function when inode is journaled in
2574 * ordered mode before truncation happens and after the inode has been
2575 * placed on orphan list with the new inode size. The second condition
2576 * avoids the race that someone writes new data and we start
2577 * committing the transaction after this function has been called but
2578 * before a transaction for truncate is started (and furthermore it
2579 * allows us to optimize the case where the addition to orphan list
2580 * happens in the same transaction as write --- we don't have to write
2581 * any data in such case).
2582 */
2583int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2584 struct jbd2_inode *jinode,
2585 loff_t new_size)
2586{
2587 transaction_t *inode_trans, *commit_trans;
2588 int ret = 0;
2589
2590 /* This is a quick check to avoid locking if not necessary */
2591 if (!jinode->i_transaction)
2592 goto out;
2593 /* Locks are here just to force reading of recent values, it is
2594 * enough that the transaction was not committing before we started
2595 * a transaction adding the inode to orphan list */
2596 read_lock(&journal->j_state_lock);
2597 commit_trans = journal->j_committing_transaction;
2598 read_unlock(&journal->j_state_lock);
2599 spin_lock(&journal->j_list_lock);
2600 inode_trans = jinode->i_transaction;
2601 spin_unlock(&journal->j_list_lock);
2602 if (inode_trans == commit_trans) {
2603 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2604 new_size, LLONG_MAX);
2605 if (ret)
2606 jbd2_journal_abort(journal, ret);
2607 }
2608out:
2609 return ret;
2610}
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
33static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
35
36static struct kmem_cache *transaction_cache;
37int __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
52void jbd2_journal_destroy_transaction_cache(void)
53{
54 kmem_cache_destroy(transaction_cache);
55 transaction_cache = NULL;
56}
57
58void 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 * Base amount of descriptor blocks we reserve for each transaction.
67 */
68static int jbd2_descriptor_blocks_per_trans(journal_t *journal)
69{
70 int tag_space = journal->j_blocksize - sizeof(journal_header_t);
71 int tags_per_block;
72
73 /* Subtract UUID */
74 tag_space -= 16;
75 if (jbd2_journal_has_csum_v2or3(journal))
76 tag_space -= sizeof(struct jbd2_journal_block_tail);
77 /* Commit code leaves a slack space of 16 bytes at the end of block */
78 tags_per_block = (tag_space - 16) / journal_tag_bytes(journal);
79 /*
80 * Revoke descriptors are accounted separately so we need to reserve
81 * space for commit block and normal transaction descriptor blocks.
82 */
83 return 1 + DIV_ROUND_UP(journal->j_max_transaction_buffers,
84 tags_per_block);
85}
86
87/*
88 * jbd2_get_transaction: obtain a new transaction_t object.
89 *
90 * Simply initialise a new transaction. Initialize it in
91 * RUNNING state and add it to the current journal (which should not
92 * have an existing running transaction: we only make a new transaction
93 * once we have started to commit the old one).
94 *
95 * Preconditions:
96 * The journal MUST be locked. We don't perform atomic mallocs on the
97 * new transaction and we can't block without protecting against other
98 * processes trying to touch the journal while it is in transition.
99 *
100 */
101
102static void jbd2_get_transaction(journal_t *journal,
103 transaction_t *transaction)
104{
105 transaction->t_journal = journal;
106 transaction->t_state = T_RUNNING;
107 transaction->t_start_time = ktime_get();
108 transaction->t_tid = journal->j_transaction_sequence++;
109 transaction->t_expires = jiffies + journal->j_commit_interval;
110 spin_lock_init(&transaction->t_handle_lock);
111 atomic_set(&transaction->t_updates, 0);
112 atomic_set(&transaction->t_outstanding_credits,
113 jbd2_descriptor_blocks_per_trans(journal) +
114 atomic_read(&journal->j_reserved_credits));
115 atomic_set(&transaction->t_outstanding_revokes, 0);
116 atomic_set(&transaction->t_handle_count, 0);
117 INIT_LIST_HEAD(&transaction->t_inode_list);
118 INIT_LIST_HEAD(&transaction->t_private_list);
119
120 /* Set up the commit timer for the new transaction. */
121 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
122 add_timer(&journal->j_commit_timer);
123
124 J_ASSERT(journal->j_running_transaction == NULL);
125 journal->j_running_transaction = transaction;
126 transaction->t_max_wait = 0;
127 transaction->t_start = jiffies;
128 transaction->t_requested = 0;
129}
130
131/*
132 * Handle management.
133 *
134 * A handle_t is an object which represents a single atomic update to a
135 * filesystem, and which tracks all of the modifications which form part
136 * of that one update.
137 */
138
139/*
140 * Update transaction's maximum wait time, if debugging is enabled.
141 *
142 * In order for t_max_wait to be reliable, it must be protected by a
143 * lock. But doing so will mean that start_this_handle() can not be
144 * run in parallel on SMP systems, which limits our scalability. So
145 * unless debugging is enabled, we no longer update t_max_wait, which
146 * means that maximum wait time reported by the jbd2_run_stats
147 * tracepoint will always be zero.
148 */
149static inline void update_t_max_wait(transaction_t *transaction,
150 unsigned long ts)
151{
152#ifdef CONFIG_JBD2_DEBUG
153 if (jbd2_journal_enable_debug &&
154 time_after(transaction->t_start, ts)) {
155 ts = jbd2_time_diff(ts, transaction->t_start);
156 spin_lock(&transaction->t_handle_lock);
157 if (ts > transaction->t_max_wait)
158 transaction->t_max_wait = ts;
159 spin_unlock(&transaction->t_handle_lock);
160 }
161#endif
162}
163
164/*
165 * Wait until running transaction passes to T_FLUSH state and new transaction
166 * can thus be started. Also starts the commit if needed. The function expects
167 * running transaction to exist and releases j_state_lock.
168 */
169static void wait_transaction_locked(journal_t *journal)
170 __releases(journal->j_state_lock)
171{
172 DEFINE_WAIT(wait);
173 int need_to_start;
174 tid_t tid = journal->j_running_transaction->t_tid;
175
176 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177 TASK_UNINTERRUPTIBLE);
178 need_to_start = !tid_geq(journal->j_commit_request, tid);
179 read_unlock(&journal->j_state_lock);
180 if (need_to_start)
181 jbd2_log_start_commit(journal, tid);
182 jbd2_might_wait_for_commit(journal);
183 schedule();
184 finish_wait(&journal->j_wait_transaction_locked, &wait);
185}
186
187/*
188 * Wait until running transaction transitions from T_SWITCH to T_FLUSH
189 * state and new transaction can thus be started. The function releases
190 * j_state_lock.
191 */
192static void wait_transaction_switching(journal_t *journal)
193 __releases(journal->j_state_lock)
194{
195 DEFINE_WAIT(wait);
196
197 if (WARN_ON(!journal->j_running_transaction ||
198 journal->j_running_transaction->t_state != T_SWITCH))
199 return;
200 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
201 TASK_UNINTERRUPTIBLE);
202 read_unlock(&journal->j_state_lock);
203 /*
204 * We don't call jbd2_might_wait_for_commit() here as there's no
205 * waiting for outstanding handles happening anymore in T_SWITCH state
206 * and handling of reserved handles actually relies on that for
207 * correctness.
208 */
209 schedule();
210 finish_wait(&journal->j_wait_transaction_locked, &wait);
211}
212
213static void sub_reserved_credits(journal_t *journal, int blocks)
214{
215 atomic_sub(blocks, &journal->j_reserved_credits);
216 wake_up(&journal->j_wait_reserved);
217}
218
219/*
220 * Wait until we can add credits for handle to the running transaction. Called
221 * with j_state_lock held for reading. Returns 0 if handle joined the running
222 * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
223 * caller must retry.
224 */
225static int add_transaction_credits(journal_t *journal, int blocks,
226 int rsv_blocks)
227{
228 transaction_t *t = journal->j_running_transaction;
229 int needed;
230 int total = blocks + rsv_blocks;
231
232 /*
233 * If the current transaction is locked down for commit, wait
234 * for the lock to be released.
235 */
236 if (t->t_state != T_RUNNING) {
237 WARN_ON_ONCE(t->t_state >= T_FLUSH);
238 wait_transaction_locked(journal);
239 return 1;
240 }
241
242 /*
243 * If there is not enough space left in the log to write all
244 * potential buffers requested by this operation, we need to
245 * stall pending a log checkpoint to free some more log space.
246 */
247 needed = atomic_add_return(total, &t->t_outstanding_credits);
248 if (needed > journal->j_max_transaction_buffers) {
249 /*
250 * If the current transaction is already too large,
251 * then start to commit it: we can then go back and
252 * attach this handle to a new transaction.
253 */
254 atomic_sub(total, &t->t_outstanding_credits);
255
256 /*
257 * Is the number of reserved credits in the current transaction too
258 * big to fit this handle? Wait until reserved credits are freed.
259 */
260 if (atomic_read(&journal->j_reserved_credits) + total >
261 journal->j_max_transaction_buffers) {
262 read_unlock(&journal->j_state_lock);
263 jbd2_might_wait_for_commit(journal);
264 wait_event(journal->j_wait_reserved,
265 atomic_read(&journal->j_reserved_credits) + total <=
266 journal->j_max_transaction_buffers);
267 return 1;
268 }
269
270 wait_transaction_locked(journal);
271 return 1;
272 }
273
274 /*
275 * The commit code assumes that it can get enough log space
276 * without forcing a checkpoint. This is *critical* for
277 * correctness: a checkpoint of a buffer which is also
278 * associated with a committing transaction creates a deadlock,
279 * so commit simply cannot force through checkpoints.
280 *
281 * We must therefore ensure the necessary space in the journal
282 * *before* starting to dirty potentially checkpointed buffers
283 * in the new transaction.
284 */
285 if (jbd2_log_space_left(journal) < journal->j_max_transaction_buffers) {
286 atomic_sub(total, &t->t_outstanding_credits);
287 read_unlock(&journal->j_state_lock);
288 jbd2_might_wait_for_commit(journal);
289 write_lock(&journal->j_state_lock);
290 if (jbd2_log_space_left(journal) <
291 journal->j_max_transaction_buffers)
292 __jbd2_log_wait_for_space(journal);
293 write_unlock(&journal->j_state_lock);
294 return 1;
295 }
296
297 /* No reservation? We are done... */
298 if (!rsv_blocks)
299 return 0;
300
301 needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
302 /* We allow at most half of a transaction to be reserved */
303 if (needed > journal->j_max_transaction_buffers / 2) {
304 sub_reserved_credits(journal, rsv_blocks);
305 atomic_sub(total, &t->t_outstanding_credits);
306 read_unlock(&journal->j_state_lock);
307 jbd2_might_wait_for_commit(journal);
308 wait_event(journal->j_wait_reserved,
309 atomic_read(&journal->j_reserved_credits) + rsv_blocks
310 <= journal->j_max_transaction_buffers / 2);
311 return 1;
312 }
313 return 0;
314}
315
316/*
317 * start_this_handle: Given a handle, deal with any locking or stalling
318 * needed to make sure that there is enough journal space for the handle
319 * to begin. Attach the handle to a transaction and set up the
320 * transaction's buffer credits.
321 */
322
323static int start_this_handle(journal_t *journal, handle_t *handle,
324 gfp_t gfp_mask)
325{
326 transaction_t *transaction, *new_transaction = NULL;
327 int blocks = handle->h_total_credits;
328 int rsv_blocks = 0;
329 unsigned long ts = jiffies;
330
331 if (handle->h_rsv_handle)
332 rsv_blocks = handle->h_rsv_handle->h_total_credits;
333
334 /*
335 * Limit the number of reserved credits to 1/2 of maximum transaction
336 * size and limit the number of total credits to not exceed maximum
337 * transaction size per operation.
338 */
339 if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
340 (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
341 printk(KERN_ERR "JBD2: %s wants too many credits "
342 "credits:%d rsv_credits:%d max:%d\n",
343 current->comm, blocks, rsv_blocks,
344 journal->j_max_transaction_buffers);
345 WARN_ON(1);
346 return -ENOSPC;
347 }
348
349alloc_transaction:
350 if (!journal->j_running_transaction) {
351 /*
352 * If __GFP_FS is not present, then we may be being called from
353 * inside the fs writeback layer, so we MUST NOT fail.
354 */
355 if ((gfp_mask & __GFP_FS) == 0)
356 gfp_mask |= __GFP_NOFAIL;
357 new_transaction = kmem_cache_zalloc(transaction_cache,
358 gfp_mask);
359 if (!new_transaction)
360 return -ENOMEM;
361 }
362
363 jbd_debug(3, "New handle %p going live.\n", handle);
364
365 /*
366 * We need to hold j_state_lock until t_updates has been incremented,
367 * for proper journal barrier handling
368 */
369repeat:
370 read_lock(&journal->j_state_lock);
371 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
372 if (is_journal_aborted(journal) ||
373 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
374 read_unlock(&journal->j_state_lock);
375 jbd2_journal_free_transaction(new_transaction);
376 return -EROFS;
377 }
378
379 /*
380 * Wait on the journal's transaction barrier if necessary. Specifically
381 * we allow reserved handles to proceed because otherwise commit could
382 * deadlock on page writeback not being able to complete.
383 */
384 if (!handle->h_reserved && journal->j_barrier_count) {
385 read_unlock(&journal->j_state_lock);
386 wait_event(journal->j_wait_transaction_locked,
387 journal->j_barrier_count == 0);
388 goto repeat;
389 }
390
391 if (!journal->j_running_transaction) {
392 read_unlock(&journal->j_state_lock);
393 if (!new_transaction)
394 goto alloc_transaction;
395 write_lock(&journal->j_state_lock);
396 if (!journal->j_running_transaction &&
397 (handle->h_reserved || !journal->j_barrier_count)) {
398 jbd2_get_transaction(journal, new_transaction);
399 new_transaction = NULL;
400 }
401 write_unlock(&journal->j_state_lock);
402 goto repeat;
403 }
404
405 transaction = journal->j_running_transaction;
406
407 if (!handle->h_reserved) {
408 /* We may have dropped j_state_lock - restart in that case */
409 if (add_transaction_credits(journal, blocks, rsv_blocks))
410 goto repeat;
411 } else {
412 /*
413 * We have handle reserved so we are allowed to join T_LOCKED
414 * transaction and we don't have to check for transaction size
415 * and journal space. But we still have to wait while running
416 * transaction is being switched to a committing one as it
417 * won't wait for any handles anymore.
418 */
419 if (transaction->t_state == T_SWITCH) {
420 wait_transaction_switching(journal);
421 goto repeat;
422 }
423 sub_reserved_credits(journal, blocks);
424 handle->h_reserved = 0;
425 }
426
427 /* OK, account for the buffers that this operation expects to
428 * use and add the handle to the running transaction.
429 */
430 update_t_max_wait(transaction, ts);
431 handle->h_transaction = transaction;
432 handle->h_requested_credits = blocks;
433 handle->h_revoke_credits_requested = handle->h_revoke_credits;
434 handle->h_start_jiffies = jiffies;
435 atomic_inc(&transaction->t_updates);
436 atomic_inc(&transaction->t_handle_count);
437 jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
438 handle, blocks,
439 atomic_read(&transaction->t_outstanding_credits),
440 jbd2_log_space_left(journal));
441 read_unlock(&journal->j_state_lock);
442 current->journal_info = handle;
443
444 rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
445 jbd2_journal_free_transaction(new_transaction);
446 /*
447 * Ensure that no allocations done while the transaction is open are
448 * going to recurse back to the fs layer.
449 */
450 handle->saved_alloc_context = memalloc_nofs_save();
451 return 0;
452}
453
454/* Allocate a new handle. This should probably be in a slab... */
455static handle_t *new_handle(int nblocks)
456{
457 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
458 if (!handle)
459 return NULL;
460 handle->h_total_credits = nblocks;
461 handle->h_ref = 1;
462
463 return handle;
464}
465
466handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
467 int revoke_records, gfp_t gfp_mask,
468 unsigned int type, unsigned int line_no)
469{
470 handle_t *handle = journal_current_handle();
471 int err;
472
473 if (!journal)
474 return ERR_PTR(-EROFS);
475
476 if (handle) {
477 J_ASSERT(handle->h_transaction->t_journal == journal);
478 handle->h_ref++;
479 return handle;
480 }
481
482 nblocks += DIV_ROUND_UP(revoke_records,
483 journal->j_revoke_records_per_block);
484 handle = new_handle(nblocks);
485 if (!handle)
486 return ERR_PTR(-ENOMEM);
487 if (rsv_blocks) {
488 handle_t *rsv_handle;
489
490 rsv_handle = new_handle(rsv_blocks);
491 if (!rsv_handle) {
492 jbd2_free_handle(handle);
493 return ERR_PTR(-ENOMEM);
494 }
495 rsv_handle->h_reserved = 1;
496 rsv_handle->h_journal = journal;
497 handle->h_rsv_handle = rsv_handle;
498 }
499 handle->h_revoke_credits = revoke_records;
500
501 err = start_this_handle(journal, handle, gfp_mask);
502 if (err < 0) {
503 if (handle->h_rsv_handle)
504 jbd2_free_handle(handle->h_rsv_handle);
505 jbd2_free_handle(handle);
506 return ERR_PTR(err);
507 }
508 handle->h_type = type;
509 handle->h_line_no = line_no;
510 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
511 handle->h_transaction->t_tid, type,
512 line_no, nblocks);
513
514 return handle;
515}
516EXPORT_SYMBOL(jbd2__journal_start);
517
518
519/**
520 * handle_t *jbd2_journal_start() - Obtain a new handle.
521 * @journal: Journal to start transaction on.
522 * @nblocks: number of block buffer we might modify
523 *
524 * We make sure that the transaction can guarantee at least nblocks of
525 * modified buffers in the log. We block until the log can guarantee
526 * that much space. Additionally, if rsv_blocks > 0, we also create another
527 * handle with rsv_blocks reserved blocks in the journal. This handle is
528 * stored in h_rsv_handle. It is not attached to any particular transaction
529 * and thus doesn't block transaction commit. If the caller uses this reserved
530 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
531 * on the parent handle will dispose the reserved one. Reserved handle has to
532 * be converted to a normal handle using jbd2_journal_start_reserved() before
533 * it can be used.
534 *
535 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
536 * on failure.
537 */
538handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
539{
540 return jbd2__journal_start(journal, nblocks, 0, 0, GFP_NOFS, 0, 0);
541}
542EXPORT_SYMBOL(jbd2_journal_start);
543
544static void __jbd2_journal_unreserve_handle(handle_t *handle, transaction_t *t)
545{
546 journal_t *journal = handle->h_journal;
547
548 WARN_ON(!handle->h_reserved);
549 sub_reserved_credits(journal, handle->h_total_credits);
550 if (t)
551 atomic_sub(handle->h_total_credits, &t->t_outstanding_credits);
552}
553
554void jbd2_journal_free_reserved(handle_t *handle)
555{
556 journal_t *journal = handle->h_journal;
557
558 /* Get j_state_lock to pin running transaction if it exists */
559 read_lock(&journal->j_state_lock);
560 __jbd2_journal_unreserve_handle(handle, journal->j_running_transaction);
561 read_unlock(&journal->j_state_lock);
562 jbd2_free_handle(handle);
563}
564EXPORT_SYMBOL(jbd2_journal_free_reserved);
565
566/**
567 * int jbd2_journal_start_reserved() - start reserved handle
568 * @handle: handle to start
569 * @type: for handle statistics
570 * @line_no: for handle statistics
571 *
572 * Start handle that has been previously reserved with jbd2_journal_reserve().
573 * This attaches @handle to the running transaction (or creates one if there's
574 * not transaction running). Unlike jbd2_journal_start() this function cannot
575 * block on journal commit, checkpointing, or similar stuff. It can block on
576 * memory allocation or frozen journal though.
577 *
578 * Return 0 on success, non-zero on error - handle is freed in that case.
579 */
580int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
581 unsigned int line_no)
582{
583 journal_t *journal = handle->h_journal;
584 int ret = -EIO;
585
586 if (WARN_ON(!handle->h_reserved)) {
587 /* Someone passed in normal handle? Just stop it. */
588 jbd2_journal_stop(handle);
589 return ret;
590 }
591 /*
592 * Usefulness of mixing of reserved and unreserved handles is
593 * questionable. So far nobody seems to need it so just error out.
594 */
595 if (WARN_ON(current->journal_info)) {
596 jbd2_journal_free_reserved(handle);
597 return ret;
598 }
599
600 handle->h_journal = NULL;
601 /*
602 * GFP_NOFS is here because callers are likely from writeback or
603 * similarly constrained call sites
604 */
605 ret = start_this_handle(journal, handle, GFP_NOFS);
606 if (ret < 0) {
607 handle->h_journal = journal;
608 jbd2_journal_free_reserved(handle);
609 return ret;
610 }
611 handle->h_type = type;
612 handle->h_line_no = line_no;
613 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
614 handle->h_transaction->t_tid, type,
615 line_no, handle->h_total_credits);
616 return 0;
617}
618EXPORT_SYMBOL(jbd2_journal_start_reserved);
619
620/**
621 * int jbd2_journal_extend() - extend buffer credits.
622 * @handle: handle to 'extend'
623 * @nblocks: nr blocks to try to extend by.
624 * @revoke_records: number of revoke records to try to extend by.
625 *
626 * Some transactions, such as large extends and truncates, can be done
627 * atomically all at once or in several stages. The operation requests
628 * a credit for a number of buffer modifications in advance, but can
629 * extend its credit if it needs more.
630 *
631 * jbd2_journal_extend tries to give the running handle more buffer credits.
632 * It does not guarantee that allocation - this is a best-effort only.
633 * The calling process MUST be able to deal cleanly with a failure to
634 * extend here.
635 *
636 * Return 0 on success, non-zero on failure.
637 *
638 * return code < 0 implies an error
639 * return code > 0 implies normal transaction-full status.
640 */
641int jbd2_journal_extend(handle_t *handle, int nblocks, int revoke_records)
642{
643 transaction_t *transaction = handle->h_transaction;
644 journal_t *journal;
645 int result;
646 int wanted;
647
648 if (is_handle_aborted(handle))
649 return -EROFS;
650 journal = transaction->t_journal;
651
652 result = 1;
653
654 read_lock(&journal->j_state_lock);
655
656 /* Don't extend a locked-down transaction! */
657 if (transaction->t_state != T_RUNNING) {
658 jbd_debug(3, "denied handle %p %d blocks: "
659 "transaction not running\n", handle, nblocks);
660 goto error_out;
661 }
662
663 nblocks += DIV_ROUND_UP(
664 handle->h_revoke_credits_requested + revoke_records,
665 journal->j_revoke_records_per_block) -
666 DIV_ROUND_UP(
667 handle->h_revoke_credits_requested,
668 journal->j_revoke_records_per_block);
669 spin_lock(&transaction->t_handle_lock);
670 wanted = atomic_add_return(nblocks,
671 &transaction->t_outstanding_credits);
672
673 if (wanted > journal->j_max_transaction_buffers) {
674 jbd_debug(3, "denied handle %p %d blocks: "
675 "transaction too large\n", handle, nblocks);
676 atomic_sub(nblocks, &transaction->t_outstanding_credits);
677 goto unlock;
678 }
679
680 trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
681 transaction->t_tid,
682 handle->h_type, handle->h_line_no,
683 handle->h_total_credits,
684 nblocks);
685
686 handle->h_total_credits += nblocks;
687 handle->h_requested_credits += nblocks;
688 handle->h_revoke_credits += revoke_records;
689 handle->h_revoke_credits_requested += revoke_records;
690 result = 0;
691
692 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
693unlock:
694 spin_unlock(&transaction->t_handle_lock);
695error_out:
696 read_unlock(&journal->j_state_lock);
697 return result;
698}
699
700static void stop_this_handle(handle_t *handle)
701{
702 transaction_t *transaction = handle->h_transaction;
703 journal_t *journal = transaction->t_journal;
704 int revokes;
705
706 J_ASSERT(journal_current_handle() == handle);
707 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
708 current->journal_info = NULL;
709 /*
710 * Subtract necessary revoke descriptor blocks from handle credits. We
711 * take care to account only for revoke descriptor blocks the
712 * transaction will really need as large sequences of transactions with
713 * small numbers of revokes are relatively common.
714 */
715 revokes = handle->h_revoke_credits_requested - handle->h_revoke_credits;
716 if (revokes) {
717 int t_revokes, revoke_descriptors;
718 int rr_per_blk = journal->j_revoke_records_per_block;
719
720 WARN_ON_ONCE(DIV_ROUND_UP(revokes, rr_per_blk)
721 > handle->h_total_credits);
722 t_revokes = atomic_add_return(revokes,
723 &transaction->t_outstanding_revokes);
724 revoke_descriptors =
725 DIV_ROUND_UP(t_revokes, rr_per_blk) -
726 DIV_ROUND_UP(t_revokes - revokes, rr_per_blk);
727 handle->h_total_credits -= revoke_descriptors;
728 }
729 atomic_sub(handle->h_total_credits,
730 &transaction->t_outstanding_credits);
731 if (handle->h_rsv_handle)
732 __jbd2_journal_unreserve_handle(handle->h_rsv_handle,
733 transaction);
734 if (atomic_dec_and_test(&transaction->t_updates))
735 wake_up(&journal->j_wait_updates);
736
737 rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
738 /*
739 * Scope of the GFP_NOFS context is over here and so we can restore the
740 * original alloc context.
741 */
742 memalloc_nofs_restore(handle->saved_alloc_context);
743}
744
745/**
746 * int jbd2_journal_restart() - restart a handle .
747 * @handle: handle to restart
748 * @nblocks: nr credits requested
749 * @revoke_records: number of revoke record credits requested
750 * @gfp_mask: memory allocation flags (for start_this_handle)
751 *
752 * Restart a handle for a multi-transaction filesystem
753 * operation.
754 *
755 * If the jbd2_journal_extend() call above fails to grant new buffer credits
756 * to a running handle, a call to jbd2_journal_restart will commit the
757 * handle's transaction so far and reattach the handle to a new
758 * transaction capable of guaranteeing the requested number of
759 * credits. We preserve reserved handle if there's any attached to the
760 * passed in handle.
761 */
762int jbd2__journal_restart(handle_t *handle, int nblocks, int revoke_records,
763 gfp_t gfp_mask)
764{
765 transaction_t *transaction = handle->h_transaction;
766 journal_t *journal;
767 tid_t tid;
768 int need_to_start;
769 int ret;
770
771 /* If we've had an abort of any type, don't even think about
772 * actually doing the restart! */
773 if (is_handle_aborted(handle))
774 return 0;
775 journal = transaction->t_journal;
776 tid = transaction->t_tid;
777
778 /*
779 * First unlink the handle from its current transaction, and start the
780 * commit on that.
781 */
782 jbd_debug(2, "restarting handle %p\n", handle);
783 stop_this_handle(handle);
784 handle->h_transaction = NULL;
785
786 /*
787 * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can
788 * get rid of pointless j_state_lock traffic like this.
789 */
790 read_lock(&journal->j_state_lock);
791 need_to_start = !tid_geq(journal->j_commit_request, tid);
792 read_unlock(&journal->j_state_lock);
793 if (need_to_start)
794 jbd2_log_start_commit(journal, tid);
795 handle->h_total_credits = nblocks +
796 DIV_ROUND_UP(revoke_records,
797 journal->j_revoke_records_per_block);
798 handle->h_revoke_credits = revoke_records;
799 ret = start_this_handle(journal, handle, gfp_mask);
800 trace_jbd2_handle_restart(journal->j_fs_dev->bd_dev,
801 ret ? 0 : handle->h_transaction->t_tid,
802 handle->h_type, handle->h_line_no,
803 handle->h_total_credits);
804 return ret;
805}
806EXPORT_SYMBOL(jbd2__journal_restart);
807
808
809int jbd2_journal_restart(handle_t *handle, int nblocks)
810{
811 return jbd2__journal_restart(handle, nblocks, 0, GFP_NOFS);
812}
813EXPORT_SYMBOL(jbd2_journal_restart);
814
815/**
816 * void jbd2_journal_lock_updates () - establish a transaction barrier.
817 * @journal: Journal to establish a barrier on.
818 *
819 * This locks out any further updates from being started, and blocks
820 * until all existing updates have completed, returning only once the
821 * journal is in a quiescent state with no updates running.
822 *
823 * The journal lock should not be held on entry.
824 */
825void jbd2_journal_lock_updates(journal_t *journal)
826{
827 DEFINE_WAIT(wait);
828
829 jbd2_might_wait_for_commit(journal);
830
831 write_lock(&journal->j_state_lock);
832 ++journal->j_barrier_count;
833
834 /* Wait until there are no reserved handles */
835 if (atomic_read(&journal->j_reserved_credits)) {
836 write_unlock(&journal->j_state_lock);
837 wait_event(journal->j_wait_reserved,
838 atomic_read(&journal->j_reserved_credits) == 0);
839 write_lock(&journal->j_state_lock);
840 }
841
842 /* Wait until there are no running updates */
843 while (1) {
844 transaction_t *transaction = journal->j_running_transaction;
845
846 if (!transaction)
847 break;
848
849 spin_lock(&transaction->t_handle_lock);
850 prepare_to_wait(&journal->j_wait_updates, &wait,
851 TASK_UNINTERRUPTIBLE);
852 if (!atomic_read(&transaction->t_updates)) {
853 spin_unlock(&transaction->t_handle_lock);
854 finish_wait(&journal->j_wait_updates, &wait);
855 break;
856 }
857 spin_unlock(&transaction->t_handle_lock);
858 write_unlock(&journal->j_state_lock);
859 schedule();
860 finish_wait(&journal->j_wait_updates, &wait);
861 write_lock(&journal->j_state_lock);
862 }
863 write_unlock(&journal->j_state_lock);
864
865 /*
866 * We have now established a barrier against other normal updates, but
867 * we also need to barrier against other jbd2_journal_lock_updates() calls
868 * to make sure that we serialise special journal-locked operations
869 * too.
870 */
871 mutex_lock(&journal->j_barrier);
872}
873
874/**
875 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
876 * @journal: Journal to release the barrier on.
877 *
878 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
879 *
880 * Should be called without the journal lock held.
881 */
882void jbd2_journal_unlock_updates (journal_t *journal)
883{
884 J_ASSERT(journal->j_barrier_count != 0);
885
886 mutex_unlock(&journal->j_barrier);
887 write_lock(&journal->j_state_lock);
888 --journal->j_barrier_count;
889 write_unlock(&journal->j_state_lock);
890 wake_up(&journal->j_wait_transaction_locked);
891}
892
893static void warn_dirty_buffer(struct buffer_head *bh)
894{
895 printk(KERN_WARNING
896 "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
897 "There's a risk of filesystem corruption in case of system "
898 "crash.\n",
899 bh->b_bdev, (unsigned long long)bh->b_blocknr);
900}
901
902/* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
903static void jbd2_freeze_jh_data(struct journal_head *jh)
904{
905 struct page *page;
906 int offset;
907 char *source;
908 struct buffer_head *bh = jh2bh(jh);
909
910 J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
911 page = bh->b_page;
912 offset = offset_in_page(bh->b_data);
913 source = kmap_atomic(page);
914 /* Fire data frozen trigger just before we copy the data */
915 jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
916 memcpy(jh->b_frozen_data, source + offset, bh->b_size);
917 kunmap_atomic(source);
918
919 /*
920 * Now that the frozen data is saved off, we need to store any matching
921 * triggers.
922 */
923 jh->b_frozen_triggers = jh->b_triggers;
924}
925
926/*
927 * If the buffer is already part of the current transaction, then there
928 * is nothing we need to do. If it is already part of a prior
929 * transaction which we are still committing to disk, then we need to
930 * make sure that we do not overwrite the old copy: we do copy-out to
931 * preserve the copy going to disk. We also account the buffer against
932 * the handle's metadata buffer credits (unless the buffer is already
933 * part of the transaction, that is).
934 *
935 */
936static int
937do_get_write_access(handle_t *handle, struct journal_head *jh,
938 int force_copy)
939{
940 struct buffer_head *bh;
941 transaction_t *transaction = handle->h_transaction;
942 journal_t *journal;
943 int error;
944 char *frozen_buffer = NULL;
945 unsigned long start_lock, time_lock;
946
947 journal = transaction->t_journal;
948
949 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
950
951 JBUFFER_TRACE(jh, "entry");
952repeat:
953 bh = jh2bh(jh);
954
955 /* @@@ Need to check for errors here at some point. */
956
957 start_lock = jiffies;
958 lock_buffer(bh);
959 spin_lock(&jh->b_state_lock);
960
961 /* If it takes too long to lock the buffer, trace it */
962 time_lock = jbd2_time_diff(start_lock, jiffies);
963 if (time_lock > HZ/10)
964 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
965 jiffies_to_msecs(time_lock));
966
967 /* We now hold the buffer lock so it is safe to query the buffer
968 * state. Is the buffer dirty?
969 *
970 * If so, there are two possibilities. The buffer may be
971 * non-journaled, and undergoing a quite legitimate writeback.
972 * Otherwise, it is journaled, and we don't expect dirty buffers
973 * in that state (the buffers should be marked JBD_Dirty
974 * instead.) So either the IO is being done under our own
975 * control and this is a bug, or it's a third party IO such as
976 * dump(8) (which may leave the buffer scheduled for read ---
977 * ie. locked but not dirty) or tune2fs (which may actually have
978 * the buffer dirtied, ugh.) */
979
980 if (buffer_dirty(bh)) {
981 /*
982 * First question: is this buffer already part of the current
983 * transaction or the existing committing transaction?
984 */
985 if (jh->b_transaction) {
986 J_ASSERT_JH(jh,
987 jh->b_transaction == transaction ||
988 jh->b_transaction ==
989 journal->j_committing_transaction);
990 if (jh->b_next_transaction)
991 J_ASSERT_JH(jh, jh->b_next_transaction ==
992 transaction);
993 warn_dirty_buffer(bh);
994 }
995 /*
996 * In any case we need to clean the dirty flag and we must
997 * do it under the buffer lock to be sure we don't race
998 * with running write-out.
999 */
1000 JBUFFER_TRACE(jh, "Journalling dirty buffer");
1001 clear_buffer_dirty(bh);
1002 set_buffer_jbddirty(bh);
1003 }
1004
1005 unlock_buffer(bh);
1006
1007 error = -EROFS;
1008 if (is_handle_aborted(handle)) {
1009 spin_unlock(&jh->b_state_lock);
1010 goto out;
1011 }
1012 error = 0;
1013
1014 /*
1015 * The buffer is already part of this transaction if b_transaction or
1016 * b_next_transaction points to it
1017 */
1018 if (jh->b_transaction == transaction ||
1019 jh->b_next_transaction == transaction)
1020 goto done;
1021
1022 /*
1023 * this is the first time this transaction is touching this buffer,
1024 * reset the modified flag
1025 */
1026 jh->b_modified = 0;
1027
1028 /*
1029 * If the buffer is not journaled right now, we need to make sure it
1030 * doesn't get written to disk before the caller actually commits the
1031 * new data
1032 */
1033 if (!jh->b_transaction) {
1034 JBUFFER_TRACE(jh, "no transaction");
1035 J_ASSERT_JH(jh, !jh->b_next_transaction);
1036 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1037 /*
1038 * Make sure all stores to jh (b_modified, b_frozen_data) are
1039 * visible before attaching it to the running transaction.
1040 * Paired with barrier in jbd2_write_access_granted()
1041 */
1042 smp_wmb();
1043 spin_lock(&journal->j_list_lock);
1044 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1045 spin_unlock(&journal->j_list_lock);
1046 goto done;
1047 }
1048 /*
1049 * If there is already a copy-out version of this buffer, then we don't
1050 * need to make another one
1051 */
1052 if (jh->b_frozen_data) {
1053 JBUFFER_TRACE(jh, "has frozen data");
1054 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1055 goto attach_next;
1056 }
1057
1058 JBUFFER_TRACE(jh, "owned by older transaction");
1059 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1060 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
1061
1062 /*
1063 * There is one case we have to be very careful about. If the
1064 * committing transaction is currently writing this buffer out to disk
1065 * and has NOT made a copy-out, then we cannot modify the buffer
1066 * contents at all right now. The essence of copy-out is that it is
1067 * the extra copy, not the primary copy, which gets journaled. If the
1068 * primary copy is already going to disk then we cannot do copy-out
1069 * here.
1070 */
1071 if (buffer_shadow(bh)) {
1072 JBUFFER_TRACE(jh, "on shadow: sleep");
1073 spin_unlock(&jh->b_state_lock);
1074 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
1075 goto repeat;
1076 }
1077
1078 /*
1079 * Only do the copy if the currently-owning transaction still needs it.
1080 * If buffer isn't on BJ_Metadata list, the committing transaction is
1081 * past that stage (here we use the fact that BH_Shadow is set under
1082 * bh_state lock together with refiling to BJ_Shadow list and at this
1083 * point we know the buffer doesn't have BH_Shadow set).
1084 *
1085 * Subtle point, though: if this is a get_undo_access, then we will be
1086 * relying on the frozen_data to contain the new value of the
1087 * committed_data record after the transaction, so we HAVE to force the
1088 * frozen_data copy in that case.
1089 */
1090 if (jh->b_jlist == BJ_Metadata || force_copy) {
1091 JBUFFER_TRACE(jh, "generate frozen data");
1092 if (!frozen_buffer) {
1093 JBUFFER_TRACE(jh, "allocate memory for buffer");
1094 spin_unlock(&jh->b_state_lock);
1095 frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
1096 GFP_NOFS | __GFP_NOFAIL);
1097 goto repeat;
1098 }
1099 jh->b_frozen_data = frozen_buffer;
1100 frozen_buffer = NULL;
1101 jbd2_freeze_jh_data(jh);
1102 }
1103attach_next:
1104 /*
1105 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1106 * before attaching it to the running transaction. Paired with barrier
1107 * in jbd2_write_access_granted()
1108 */
1109 smp_wmb();
1110 jh->b_next_transaction = transaction;
1111
1112done:
1113 spin_unlock(&jh->b_state_lock);
1114
1115 /*
1116 * If we are about to journal a buffer, then any revoke pending on it is
1117 * no longer valid
1118 */
1119 jbd2_journal_cancel_revoke(handle, jh);
1120
1121out:
1122 if (unlikely(frozen_buffer)) /* It's usually NULL */
1123 jbd2_free(frozen_buffer, bh->b_size);
1124
1125 JBUFFER_TRACE(jh, "exit");
1126 return error;
1127}
1128
1129/* Fast check whether buffer is already attached to the required transaction */
1130static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1131 bool undo)
1132{
1133 struct journal_head *jh;
1134 bool ret = false;
1135
1136 /* Dirty buffers require special handling... */
1137 if (buffer_dirty(bh))
1138 return false;
1139
1140 /*
1141 * RCU protects us from dereferencing freed pages. So the checks we do
1142 * are guaranteed not to oops. However the jh slab object can get freed
1143 * & reallocated while we work with it. So we have to be careful. When
1144 * we see jh attached to the running transaction, we know it must stay
1145 * so until the transaction is committed. Thus jh won't be freed and
1146 * will be attached to the same bh while we run. However it can
1147 * happen jh gets freed, reallocated, and attached to the transaction
1148 * just after we get pointer to it from bh. So we have to be careful
1149 * and recheck jh still belongs to our bh before we return success.
1150 */
1151 rcu_read_lock();
1152 if (!buffer_jbd(bh))
1153 goto out;
1154 /* This should be bh2jh() but that doesn't work with inline functions */
1155 jh = READ_ONCE(bh->b_private);
1156 if (!jh)
1157 goto out;
1158 /* For undo access buffer must have data copied */
1159 if (undo && !jh->b_committed_data)
1160 goto out;
1161 if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1162 READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1163 goto out;
1164 /*
1165 * There are two reasons for the barrier here:
1166 * 1) Make sure to fetch b_bh after we did previous checks so that we
1167 * detect when jh went through free, realloc, attach to transaction
1168 * while we were checking. Paired with implicit barrier in that path.
1169 * 2) So that access to bh done after jbd2_write_access_granted()
1170 * doesn't get reordered and see inconsistent state of concurrent
1171 * do_get_write_access().
1172 */
1173 smp_mb();
1174 if (unlikely(jh->b_bh != bh))
1175 goto out;
1176 ret = true;
1177out:
1178 rcu_read_unlock();
1179 return ret;
1180}
1181
1182/**
1183 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1184 * @handle: transaction to add buffer modifications to
1185 * @bh: bh to be used for metadata writes
1186 *
1187 * Returns: error code or 0 on success.
1188 *
1189 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1190 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1191 */
1192
1193int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1194{
1195 struct journal_head *jh;
1196 int rc;
1197
1198 if (is_handle_aborted(handle))
1199 return -EROFS;
1200
1201 if (jbd2_write_access_granted(handle, bh, false))
1202 return 0;
1203
1204 jh = jbd2_journal_add_journal_head(bh);
1205 /* We do not want to get caught playing with fields which the
1206 * log thread also manipulates. Make sure that the buffer
1207 * completes any outstanding IO before proceeding. */
1208 rc = do_get_write_access(handle, jh, 0);
1209 jbd2_journal_put_journal_head(jh);
1210 return rc;
1211}
1212
1213
1214/*
1215 * When the user wants to journal a newly created buffer_head
1216 * (ie. getblk() returned a new buffer and we are going to populate it
1217 * manually rather than reading off disk), then we need to keep the
1218 * buffer_head locked until it has been completely filled with new
1219 * data. In this case, we should be able to make the assertion that
1220 * the bh is not already part of an existing transaction.
1221 *
1222 * The buffer should already be locked by the caller by this point.
1223 * There is no lock ranking violation: it was a newly created,
1224 * unlocked buffer beforehand. */
1225
1226/**
1227 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1228 * @handle: transaction to new buffer to
1229 * @bh: new buffer.
1230 *
1231 * Call this if you create a new bh.
1232 */
1233int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1234{
1235 transaction_t *transaction = handle->h_transaction;
1236 journal_t *journal;
1237 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1238 int err;
1239
1240 jbd_debug(5, "journal_head %p\n", jh);
1241 err = -EROFS;
1242 if (is_handle_aborted(handle))
1243 goto out;
1244 journal = transaction->t_journal;
1245 err = 0;
1246
1247 JBUFFER_TRACE(jh, "entry");
1248 /*
1249 * The buffer may already belong to this transaction due to pre-zeroing
1250 * in the filesystem's new_block code. It may also be on the previous,
1251 * committing transaction's lists, but it HAS to be in Forget state in
1252 * that case: the transaction must have deleted the buffer for it to be
1253 * reused here.
1254 */
1255 spin_lock(&jh->b_state_lock);
1256 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1257 jh->b_transaction == NULL ||
1258 (jh->b_transaction == journal->j_committing_transaction &&
1259 jh->b_jlist == BJ_Forget)));
1260
1261 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1262 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1263
1264 if (jh->b_transaction == NULL) {
1265 /*
1266 * Previous jbd2_journal_forget() could have left the buffer
1267 * with jbddirty bit set because it was being committed. When
1268 * the commit finished, we've filed the buffer for
1269 * checkpointing and marked it dirty. Now we are reallocating
1270 * the buffer so the transaction freeing it must have
1271 * committed and so it's safe to clear the dirty bit.
1272 */
1273 clear_buffer_dirty(jh2bh(jh));
1274 /* first access by this transaction */
1275 jh->b_modified = 0;
1276
1277 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1278 spin_lock(&journal->j_list_lock);
1279 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1280 spin_unlock(&journal->j_list_lock);
1281 } else if (jh->b_transaction == journal->j_committing_transaction) {
1282 /* first access by this transaction */
1283 jh->b_modified = 0;
1284
1285 JBUFFER_TRACE(jh, "set next transaction");
1286 spin_lock(&journal->j_list_lock);
1287 jh->b_next_transaction = transaction;
1288 spin_unlock(&journal->j_list_lock);
1289 }
1290 spin_unlock(&jh->b_state_lock);
1291
1292 /*
1293 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1294 * blocks which contain freed but then revoked metadata. We need
1295 * to cancel the revoke in case we end up freeing it yet again
1296 * and the reallocating as data - this would cause a second revoke,
1297 * which hits an assertion error.
1298 */
1299 JBUFFER_TRACE(jh, "cancelling revoke");
1300 jbd2_journal_cancel_revoke(handle, jh);
1301out:
1302 jbd2_journal_put_journal_head(jh);
1303 return err;
1304}
1305
1306/**
1307 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1308 * non-rewindable consequences
1309 * @handle: transaction
1310 * @bh: buffer to undo
1311 *
1312 * Sometimes there is a need to distinguish between metadata which has
1313 * been committed to disk and that which has not. The ext3fs code uses
1314 * this for freeing and allocating space, we have to make sure that we
1315 * do not reuse freed space until the deallocation has been committed,
1316 * since if we overwrote that space we would make the delete
1317 * un-rewindable in case of a crash.
1318 *
1319 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1320 * buffer for parts of non-rewindable operations such as delete
1321 * operations on the bitmaps. The journaling code must keep a copy of
1322 * the buffer's contents prior to the undo_access call until such time
1323 * as we know that the buffer has definitely been committed to disk.
1324 *
1325 * We never need to know which transaction the committed data is part
1326 * of, buffers touched here are guaranteed to be dirtied later and so
1327 * will be committed to a new transaction in due course, at which point
1328 * we can discard the old committed data pointer.
1329 *
1330 * Returns error number or 0 on success.
1331 */
1332int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1333{
1334 int err;
1335 struct journal_head *jh;
1336 char *committed_data = NULL;
1337
1338 if (is_handle_aborted(handle))
1339 return -EROFS;
1340
1341 if (jbd2_write_access_granted(handle, bh, true))
1342 return 0;
1343
1344 jh = jbd2_journal_add_journal_head(bh);
1345 JBUFFER_TRACE(jh, "entry");
1346
1347 /*
1348 * Do this first --- it can drop the journal lock, so we want to
1349 * make sure that obtaining the committed_data is done
1350 * atomically wrt. completion of any outstanding commits.
1351 */
1352 err = do_get_write_access(handle, jh, 1);
1353 if (err)
1354 goto out;
1355
1356repeat:
1357 if (!jh->b_committed_data)
1358 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1359 GFP_NOFS|__GFP_NOFAIL);
1360
1361 spin_lock(&jh->b_state_lock);
1362 if (!jh->b_committed_data) {
1363 /* Copy out the current buffer contents into the
1364 * preserved, committed copy. */
1365 JBUFFER_TRACE(jh, "generate b_committed data");
1366 if (!committed_data) {
1367 spin_unlock(&jh->b_state_lock);
1368 goto repeat;
1369 }
1370
1371 jh->b_committed_data = committed_data;
1372 committed_data = NULL;
1373 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1374 }
1375 spin_unlock(&jh->b_state_lock);
1376out:
1377 jbd2_journal_put_journal_head(jh);
1378 if (unlikely(committed_data))
1379 jbd2_free(committed_data, bh->b_size);
1380 return err;
1381}
1382
1383/**
1384 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1385 * @bh: buffer to trigger on
1386 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1387 *
1388 * Set any triggers on this journal_head. This is always safe, because
1389 * triggers for a committing buffer will be saved off, and triggers for
1390 * a running transaction will match the buffer in that transaction.
1391 *
1392 * Call with NULL to clear the triggers.
1393 */
1394void jbd2_journal_set_triggers(struct buffer_head *bh,
1395 struct jbd2_buffer_trigger_type *type)
1396{
1397 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1398
1399 if (WARN_ON(!jh))
1400 return;
1401 jh->b_triggers = type;
1402 jbd2_journal_put_journal_head(jh);
1403}
1404
1405void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1406 struct jbd2_buffer_trigger_type *triggers)
1407{
1408 struct buffer_head *bh = jh2bh(jh);
1409
1410 if (!triggers || !triggers->t_frozen)
1411 return;
1412
1413 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1414}
1415
1416void jbd2_buffer_abort_trigger(struct journal_head *jh,
1417 struct jbd2_buffer_trigger_type *triggers)
1418{
1419 if (!triggers || !triggers->t_abort)
1420 return;
1421
1422 triggers->t_abort(triggers, jh2bh(jh));
1423}
1424
1425/**
1426 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1427 * @handle: transaction to add buffer to.
1428 * @bh: buffer to mark
1429 *
1430 * mark dirty metadata which needs to be journaled as part of the current
1431 * transaction.
1432 *
1433 * The buffer must have previously had jbd2_journal_get_write_access()
1434 * called so that it has a valid journal_head attached to the buffer
1435 * head.
1436 *
1437 * The buffer is placed on the transaction's metadata list and is marked
1438 * as belonging to the transaction.
1439 *
1440 * Returns error number or 0 on success.
1441 *
1442 * Special care needs to be taken if the buffer already belongs to the
1443 * current committing transaction (in which case we should have frozen
1444 * data present for that commit). In that case, we don't relink the
1445 * buffer: that only gets done when the old transaction finally
1446 * completes its commit.
1447 */
1448int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1449{
1450 transaction_t *transaction = handle->h_transaction;
1451 journal_t *journal;
1452 struct journal_head *jh;
1453 int ret = 0;
1454
1455 if (is_handle_aborted(handle))
1456 return -EROFS;
1457 if (!buffer_jbd(bh))
1458 return -EUCLEAN;
1459
1460 /*
1461 * We don't grab jh reference here since the buffer must be part
1462 * of the running transaction.
1463 */
1464 jh = bh2jh(bh);
1465 jbd_debug(5, "journal_head %p\n", jh);
1466 JBUFFER_TRACE(jh, "entry");
1467
1468 /*
1469 * This and the following assertions are unreliable since we may see jh
1470 * in inconsistent state unless we grab bh_state lock. But this is
1471 * crucial to catch bugs so let's do a reliable check until the
1472 * lockless handling is fully proven.
1473 */
1474 if (jh->b_transaction != transaction &&
1475 jh->b_next_transaction != transaction) {
1476 spin_lock(&jh->b_state_lock);
1477 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1478 jh->b_next_transaction == transaction);
1479 spin_unlock(&jh->b_state_lock);
1480 }
1481 if (jh->b_modified == 1) {
1482 /* If it's in our transaction it must be in BJ_Metadata list. */
1483 if (jh->b_transaction == transaction &&
1484 jh->b_jlist != BJ_Metadata) {
1485 spin_lock(&jh->b_state_lock);
1486 if (jh->b_transaction == transaction &&
1487 jh->b_jlist != BJ_Metadata)
1488 pr_err("JBD2: assertion failure: h_type=%u "
1489 "h_line_no=%u block_no=%llu jlist=%u\n",
1490 handle->h_type, handle->h_line_no,
1491 (unsigned long long) bh->b_blocknr,
1492 jh->b_jlist);
1493 J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1494 jh->b_jlist == BJ_Metadata);
1495 spin_unlock(&jh->b_state_lock);
1496 }
1497 goto out;
1498 }
1499
1500 journal = transaction->t_journal;
1501 spin_lock(&jh->b_state_lock);
1502
1503 if (jh->b_modified == 0) {
1504 /*
1505 * This buffer's got modified and becoming part
1506 * of the transaction. This needs to be done
1507 * once a transaction -bzzz
1508 */
1509 if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) {
1510 ret = -ENOSPC;
1511 goto out_unlock_bh;
1512 }
1513 jh->b_modified = 1;
1514 handle->h_total_credits--;
1515 }
1516
1517 /*
1518 * fastpath, to avoid expensive locking. If this buffer is already
1519 * on the running transaction's metadata list there is nothing to do.
1520 * Nobody can take it off again because there is a handle open.
1521 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1522 * result in this test being false, so we go in and take the locks.
1523 */
1524 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1525 JBUFFER_TRACE(jh, "fastpath");
1526 if (unlikely(jh->b_transaction !=
1527 journal->j_running_transaction)) {
1528 printk(KERN_ERR "JBD2: %s: "
1529 "jh->b_transaction (%llu, %p, %u) != "
1530 "journal->j_running_transaction (%p, %u)\n",
1531 journal->j_devname,
1532 (unsigned long long) bh->b_blocknr,
1533 jh->b_transaction,
1534 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1535 journal->j_running_transaction,
1536 journal->j_running_transaction ?
1537 journal->j_running_transaction->t_tid : 0);
1538 ret = -EINVAL;
1539 }
1540 goto out_unlock_bh;
1541 }
1542
1543 set_buffer_jbddirty(bh);
1544
1545 /*
1546 * Metadata already on the current transaction list doesn't
1547 * need to be filed. Metadata on another transaction's list must
1548 * be committing, and will be refiled once the commit completes:
1549 * leave it alone for now.
1550 */
1551 if (jh->b_transaction != transaction) {
1552 JBUFFER_TRACE(jh, "already on other transaction");
1553 if (unlikely(((jh->b_transaction !=
1554 journal->j_committing_transaction)) ||
1555 (jh->b_next_transaction != transaction))) {
1556 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1557 "bad jh for block %llu: "
1558 "transaction (%p, %u), "
1559 "jh->b_transaction (%p, %u), "
1560 "jh->b_next_transaction (%p, %u), jlist %u\n",
1561 journal->j_devname,
1562 (unsigned long long) bh->b_blocknr,
1563 transaction, transaction->t_tid,
1564 jh->b_transaction,
1565 jh->b_transaction ?
1566 jh->b_transaction->t_tid : 0,
1567 jh->b_next_transaction,
1568 jh->b_next_transaction ?
1569 jh->b_next_transaction->t_tid : 0,
1570 jh->b_jlist);
1571 WARN_ON(1);
1572 ret = -EINVAL;
1573 }
1574 /* And this case is illegal: we can't reuse another
1575 * transaction's data buffer, ever. */
1576 goto out_unlock_bh;
1577 }
1578
1579 /* That test should have eliminated the following case: */
1580 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1581
1582 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1583 spin_lock(&journal->j_list_lock);
1584 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1585 spin_unlock(&journal->j_list_lock);
1586out_unlock_bh:
1587 spin_unlock(&jh->b_state_lock);
1588out:
1589 JBUFFER_TRACE(jh, "exit");
1590 return ret;
1591}
1592
1593/**
1594 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1595 * @handle: transaction handle
1596 * @bh: bh to 'forget'
1597 *
1598 * We can only do the bforget if there are no commits pending against the
1599 * buffer. If the buffer is dirty in the current running transaction we
1600 * can safely unlink it.
1601 *
1602 * bh may not be a journalled buffer at all - it may be a non-JBD
1603 * buffer which came off the hashtable. Check for this.
1604 *
1605 * Decrements bh->b_count by one.
1606 *
1607 * Allow this call even if the handle has aborted --- it may be part of
1608 * the caller's cleanup after an abort.
1609 */
1610int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh)
1611{
1612 transaction_t *transaction = handle->h_transaction;
1613 journal_t *journal;
1614 struct journal_head *jh;
1615 int drop_reserve = 0;
1616 int err = 0;
1617 int was_modified = 0;
1618
1619 if (is_handle_aborted(handle))
1620 return -EROFS;
1621 journal = transaction->t_journal;
1622
1623 BUFFER_TRACE(bh, "entry");
1624
1625 jh = jbd2_journal_grab_journal_head(bh);
1626 if (!jh) {
1627 __bforget(bh);
1628 return 0;
1629 }
1630
1631 spin_lock(&jh->b_state_lock);
1632
1633 /* Critical error: attempting to delete a bitmap buffer, maybe?
1634 * Don't do any jbd operations, and return an error. */
1635 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1636 "inconsistent data on disk")) {
1637 err = -EIO;
1638 goto drop;
1639 }
1640
1641 /* keep track of whether or not this transaction modified us */
1642 was_modified = jh->b_modified;
1643
1644 /*
1645 * The buffer's going from the transaction, we must drop
1646 * all references -bzzz
1647 */
1648 jh->b_modified = 0;
1649
1650 if (jh->b_transaction == transaction) {
1651 J_ASSERT_JH(jh, !jh->b_frozen_data);
1652
1653 /* If we are forgetting a buffer which is already part
1654 * of this transaction, then we can just drop it from
1655 * the transaction immediately. */
1656 clear_buffer_dirty(bh);
1657 clear_buffer_jbddirty(bh);
1658
1659 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1660
1661 /*
1662 * we only want to drop a reference if this transaction
1663 * modified the buffer
1664 */
1665 if (was_modified)
1666 drop_reserve = 1;
1667
1668 /*
1669 * We are no longer going to journal this buffer.
1670 * However, the commit of this transaction is still
1671 * important to the buffer: the delete that we are now
1672 * processing might obsolete an old log entry, so by
1673 * committing, we can satisfy the buffer's checkpoint.
1674 *
1675 * So, if we have a checkpoint on the buffer, we should
1676 * now refile the buffer on our BJ_Forget list so that
1677 * we know to remove the checkpoint after we commit.
1678 */
1679
1680 spin_lock(&journal->j_list_lock);
1681 if (jh->b_cp_transaction) {
1682 __jbd2_journal_temp_unlink_buffer(jh);
1683 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1684 } else {
1685 __jbd2_journal_unfile_buffer(jh);
1686 jbd2_journal_put_journal_head(jh);
1687 }
1688 spin_unlock(&journal->j_list_lock);
1689 } else if (jh->b_transaction) {
1690 J_ASSERT_JH(jh, (jh->b_transaction ==
1691 journal->j_committing_transaction));
1692 /* However, if the buffer is still owned by a prior
1693 * (committing) transaction, we can't drop it yet... */
1694 JBUFFER_TRACE(jh, "belongs to older transaction");
1695 /* ... but we CAN drop it from the new transaction through
1696 * marking the buffer as freed and set j_next_transaction to
1697 * the new transaction, so that not only the commit code
1698 * knows it should clear dirty bits when it is done with the
1699 * buffer, but also the buffer can be checkpointed only
1700 * after the new transaction commits. */
1701
1702 set_buffer_freed(bh);
1703
1704 if (!jh->b_next_transaction) {
1705 spin_lock(&journal->j_list_lock);
1706 jh->b_next_transaction = transaction;
1707 spin_unlock(&journal->j_list_lock);
1708 } else {
1709 J_ASSERT(jh->b_next_transaction == transaction);
1710
1711 /*
1712 * only drop a reference if this transaction modified
1713 * the buffer
1714 */
1715 if (was_modified)
1716 drop_reserve = 1;
1717 }
1718 } else {
1719 /*
1720 * Finally, if the buffer is not belongs to any
1721 * transaction, we can just drop it now if it has no
1722 * checkpoint.
1723 */
1724 spin_lock(&journal->j_list_lock);
1725 if (!jh->b_cp_transaction) {
1726 JBUFFER_TRACE(jh, "belongs to none transaction");
1727 spin_unlock(&journal->j_list_lock);
1728 goto drop;
1729 }
1730
1731 /*
1732 * Otherwise, if the buffer has been written to disk,
1733 * it is safe to remove the checkpoint and drop it.
1734 */
1735 if (!buffer_dirty(bh)) {
1736 __jbd2_journal_remove_checkpoint(jh);
1737 spin_unlock(&journal->j_list_lock);
1738 goto drop;
1739 }
1740
1741 /*
1742 * The buffer is still not written to disk, we should
1743 * attach this buffer to current transaction so that the
1744 * buffer can be checkpointed only after the current
1745 * transaction commits.
1746 */
1747 clear_buffer_dirty(bh);
1748 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1749 spin_unlock(&journal->j_list_lock);
1750 }
1751drop:
1752 __brelse(bh);
1753 spin_unlock(&jh->b_state_lock);
1754 jbd2_journal_put_journal_head(jh);
1755 if (drop_reserve) {
1756 /* no need to reserve log space for this block -bzzz */
1757 handle->h_total_credits++;
1758 }
1759 return err;
1760}
1761
1762/**
1763 * int jbd2_journal_stop() - complete a transaction
1764 * @handle: transaction to complete.
1765 *
1766 * All done for a particular handle.
1767 *
1768 * There is not much action needed here. We just return any remaining
1769 * buffer credits to the transaction and remove the handle. The only
1770 * complication is that we need to start a commit operation if the
1771 * filesystem is marked for synchronous update.
1772 *
1773 * jbd2_journal_stop itself will not usually return an error, but it may
1774 * do so in unusual circumstances. In particular, expect it to
1775 * return -EIO if a jbd2_journal_abort has been executed since the
1776 * transaction began.
1777 */
1778int jbd2_journal_stop(handle_t *handle)
1779{
1780 transaction_t *transaction = handle->h_transaction;
1781 journal_t *journal;
1782 int err = 0, wait_for_commit = 0;
1783 tid_t tid;
1784 pid_t pid;
1785
1786 if (--handle->h_ref > 0) {
1787 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1788 handle->h_ref);
1789 if (is_handle_aborted(handle))
1790 return -EIO;
1791 return 0;
1792 }
1793 if (!transaction) {
1794 /*
1795 * Handle is already detached from the transaction so there is
1796 * nothing to do other than free the handle.
1797 */
1798 memalloc_nofs_restore(handle->saved_alloc_context);
1799 goto free_and_exit;
1800 }
1801 journal = transaction->t_journal;
1802 tid = transaction->t_tid;
1803
1804 if (is_handle_aborted(handle))
1805 err = -EIO;
1806
1807 jbd_debug(4, "Handle %p going down\n", handle);
1808 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1809 tid, handle->h_type, handle->h_line_no,
1810 jiffies - handle->h_start_jiffies,
1811 handle->h_sync, handle->h_requested_credits,
1812 (handle->h_requested_credits -
1813 handle->h_total_credits));
1814
1815 /*
1816 * Implement synchronous transaction batching. If the handle
1817 * was synchronous, don't force a commit immediately. Let's
1818 * yield and let another thread piggyback onto this
1819 * transaction. Keep doing that while new threads continue to
1820 * arrive. It doesn't cost much - we're about to run a commit
1821 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1822 * operations by 30x or more...
1823 *
1824 * We try and optimize the sleep time against what the
1825 * underlying disk can do, instead of having a static sleep
1826 * time. This is useful for the case where our storage is so
1827 * fast that it is more optimal to go ahead and force a flush
1828 * and wait for the transaction to be committed than it is to
1829 * wait for an arbitrary amount of time for new writers to
1830 * join the transaction. We achieve this by measuring how
1831 * long it takes to commit a transaction, and compare it with
1832 * how long this transaction has been running, and if run time
1833 * < commit time then we sleep for the delta and commit. This
1834 * greatly helps super fast disks that would see slowdowns as
1835 * more threads started doing fsyncs.
1836 *
1837 * But don't do this if this process was the most recent one
1838 * to perform a synchronous write. We do this to detect the
1839 * case where a single process is doing a stream of sync
1840 * writes. No point in waiting for joiners in that case.
1841 *
1842 * Setting max_batch_time to 0 disables this completely.
1843 */
1844 pid = current->pid;
1845 if (handle->h_sync && journal->j_last_sync_writer != pid &&
1846 journal->j_max_batch_time) {
1847 u64 commit_time, trans_time;
1848
1849 journal->j_last_sync_writer = pid;
1850
1851 read_lock(&journal->j_state_lock);
1852 commit_time = journal->j_average_commit_time;
1853 read_unlock(&journal->j_state_lock);
1854
1855 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1856 transaction->t_start_time));
1857
1858 commit_time = max_t(u64, commit_time,
1859 1000*journal->j_min_batch_time);
1860 commit_time = min_t(u64, commit_time,
1861 1000*journal->j_max_batch_time);
1862
1863 if (trans_time < commit_time) {
1864 ktime_t expires = ktime_add_ns(ktime_get(),
1865 commit_time);
1866 set_current_state(TASK_UNINTERRUPTIBLE);
1867 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1868 }
1869 }
1870
1871 if (handle->h_sync)
1872 transaction->t_synchronous_commit = 1;
1873
1874 /*
1875 * If the handle is marked SYNC, we need to set another commit
1876 * going! We also want to force a commit if the transaction is too
1877 * old now.
1878 */
1879 if (handle->h_sync ||
1880 time_after_eq(jiffies, transaction->t_expires)) {
1881 /* Do this even for aborted journals: an abort still
1882 * completes the commit thread, it just doesn't write
1883 * anything to disk. */
1884
1885 jbd_debug(2, "transaction too old, requesting commit for "
1886 "handle %p\n", handle);
1887 /* This is non-blocking */
1888 jbd2_log_start_commit(journal, tid);
1889
1890 /*
1891 * Special case: JBD2_SYNC synchronous updates require us
1892 * to wait for the commit to complete.
1893 */
1894 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1895 wait_for_commit = 1;
1896 }
1897
1898 /*
1899 * Once stop_this_handle() drops t_updates, the transaction could start
1900 * committing on us and eventually disappear. So we must not
1901 * dereference transaction pointer again after calling
1902 * stop_this_handle().
1903 */
1904 stop_this_handle(handle);
1905
1906 if (wait_for_commit)
1907 err = jbd2_log_wait_commit(journal, tid);
1908
1909free_and_exit:
1910 if (handle->h_rsv_handle)
1911 jbd2_free_handle(handle->h_rsv_handle);
1912 jbd2_free_handle(handle);
1913 return err;
1914}
1915
1916/*
1917 *
1918 * List management code snippets: various functions for manipulating the
1919 * transaction buffer lists.
1920 *
1921 */
1922
1923/*
1924 * Append a buffer to a transaction list, given the transaction's list head
1925 * pointer.
1926 *
1927 * j_list_lock is held.
1928 *
1929 * jh->b_state_lock is held.
1930 */
1931
1932static inline void
1933__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1934{
1935 if (!*list) {
1936 jh->b_tnext = jh->b_tprev = jh;
1937 *list = jh;
1938 } else {
1939 /* Insert at the tail of the list to preserve order */
1940 struct journal_head *first = *list, *last = first->b_tprev;
1941 jh->b_tprev = last;
1942 jh->b_tnext = first;
1943 last->b_tnext = first->b_tprev = jh;
1944 }
1945}
1946
1947/*
1948 * Remove a buffer from a transaction list, given the transaction's list
1949 * head pointer.
1950 *
1951 * Called with j_list_lock held, and the journal may not be locked.
1952 *
1953 * jh->b_state_lock is held.
1954 */
1955
1956static inline void
1957__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1958{
1959 if (*list == jh) {
1960 *list = jh->b_tnext;
1961 if (*list == jh)
1962 *list = NULL;
1963 }
1964 jh->b_tprev->b_tnext = jh->b_tnext;
1965 jh->b_tnext->b_tprev = jh->b_tprev;
1966}
1967
1968/*
1969 * Remove a buffer from the appropriate transaction list.
1970 *
1971 * Note that this function can *change* the value of
1972 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1973 * t_reserved_list. If the caller is holding onto a copy of one of these
1974 * pointers, it could go bad. Generally the caller needs to re-read the
1975 * pointer from the transaction_t.
1976 *
1977 * Called under j_list_lock.
1978 */
1979static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1980{
1981 struct journal_head **list = NULL;
1982 transaction_t *transaction;
1983 struct buffer_head *bh = jh2bh(jh);
1984
1985 lockdep_assert_held(&jh->b_state_lock);
1986 transaction = jh->b_transaction;
1987 if (transaction)
1988 assert_spin_locked(&transaction->t_journal->j_list_lock);
1989
1990 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1991 if (jh->b_jlist != BJ_None)
1992 J_ASSERT_JH(jh, transaction != NULL);
1993
1994 switch (jh->b_jlist) {
1995 case BJ_None:
1996 return;
1997 case BJ_Metadata:
1998 transaction->t_nr_buffers--;
1999 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
2000 list = &transaction->t_buffers;
2001 break;
2002 case BJ_Forget:
2003 list = &transaction->t_forget;
2004 break;
2005 case BJ_Shadow:
2006 list = &transaction->t_shadow_list;
2007 break;
2008 case BJ_Reserved:
2009 list = &transaction->t_reserved_list;
2010 break;
2011 }
2012
2013 __blist_del_buffer(list, jh);
2014 jh->b_jlist = BJ_None;
2015 if (transaction && is_journal_aborted(transaction->t_journal))
2016 clear_buffer_jbddirty(bh);
2017 else if (test_clear_buffer_jbddirty(bh))
2018 mark_buffer_dirty(bh); /* Expose it to the VM */
2019}
2020
2021/*
2022 * Remove buffer from all transactions. The caller is responsible for dropping
2023 * the jh reference that belonged to the transaction.
2024 *
2025 * Called with bh_state lock and j_list_lock
2026 */
2027static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
2028{
2029 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2030 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2031
2032 __jbd2_journal_temp_unlink_buffer(jh);
2033 jh->b_transaction = NULL;
2034}
2035
2036void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
2037{
2038 struct buffer_head *bh = jh2bh(jh);
2039
2040 /* Get reference so that buffer cannot be freed before we unlock it */
2041 get_bh(bh);
2042 spin_lock(&jh->b_state_lock);
2043 spin_lock(&journal->j_list_lock);
2044 __jbd2_journal_unfile_buffer(jh);
2045 spin_unlock(&journal->j_list_lock);
2046 spin_unlock(&jh->b_state_lock);
2047 jbd2_journal_put_journal_head(jh);
2048 __brelse(bh);
2049}
2050
2051/*
2052 * Called from jbd2_journal_try_to_free_buffers().
2053 *
2054 * Called under jh->b_state_lock
2055 */
2056static void
2057__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
2058{
2059 struct journal_head *jh;
2060
2061 jh = bh2jh(bh);
2062
2063 if (buffer_locked(bh) || buffer_dirty(bh))
2064 goto out;
2065
2066 if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
2067 goto out;
2068
2069 spin_lock(&journal->j_list_lock);
2070 if (jh->b_cp_transaction != NULL) {
2071 /* written-back checkpointed metadata buffer */
2072 JBUFFER_TRACE(jh, "remove from checkpoint list");
2073 __jbd2_journal_remove_checkpoint(jh);
2074 }
2075 spin_unlock(&journal->j_list_lock);
2076out:
2077 return;
2078}
2079
2080/**
2081 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
2082 * @journal: journal for operation
2083 * @page: to try and free
2084 *
2085 * For all the buffers on this page,
2086 * if they are fully written out ordered data, move them onto BUF_CLEAN
2087 * so try_to_free_buffers() can reap them.
2088 *
2089 * This function returns non-zero if we wish try_to_free_buffers()
2090 * to be called. We do this if the page is releasable by try_to_free_buffers().
2091 * We also do it if the page has locked or dirty buffers and the caller wants
2092 * us to perform sync or async writeout.
2093 *
2094 * This complicates JBD locking somewhat. We aren't protected by the
2095 * BKL here. We wish to remove the buffer from its committing or
2096 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2097 *
2098 * This may *change* the value of transaction_t->t_datalist, so anyone
2099 * who looks at t_datalist needs to lock against this function.
2100 *
2101 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2102 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
2103 * will come out of the lock with the buffer dirty, which makes it
2104 * ineligible for release here.
2105 *
2106 * Who else is affected by this? hmm... Really the only contender
2107 * is do_get_write_access() - it could be looking at the buffer while
2108 * journal_try_to_free_buffer() is changing its state. But that
2109 * cannot happen because we never reallocate freed data as metadata
2110 * while the data is part of a transaction. Yes?
2111 *
2112 * Return 0 on failure, 1 on success
2113 */
2114int jbd2_journal_try_to_free_buffers(journal_t *journal, struct page *page)
2115{
2116 struct buffer_head *head;
2117 struct buffer_head *bh;
2118 bool has_write_io_error = false;
2119 int ret = 0;
2120
2121 J_ASSERT(PageLocked(page));
2122
2123 head = page_buffers(page);
2124 bh = head;
2125 do {
2126 struct journal_head *jh;
2127
2128 /*
2129 * We take our own ref against the journal_head here to avoid
2130 * having to add tons of locking around each instance of
2131 * jbd2_journal_put_journal_head().
2132 */
2133 jh = jbd2_journal_grab_journal_head(bh);
2134 if (!jh)
2135 continue;
2136
2137 spin_lock(&jh->b_state_lock);
2138 __journal_try_to_free_buffer(journal, bh);
2139 spin_unlock(&jh->b_state_lock);
2140 jbd2_journal_put_journal_head(jh);
2141 if (buffer_jbd(bh))
2142 goto busy;
2143
2144 /*
2145 * If we free a metadata buffer which has been failed to
2146 * write out, the jbd2 checkpoint procedure will not detect
2147 * this failure and may lead to filesystem inconsistency
2148 * after cleanup journal tail.
2149 */
2150 if (buffer_write_io_error(bh)) {
2151 pr_err("JBD2: Error while async write back metadata bh %llu.",
2152 (unsigned long long)bh->b_blocknr);
2153 has_write_io_error = true;
2154 }
2155 } while ((bh = bh->b_this_page) != head);
2156
2157 ret = try_to_free_buffers(page);
2158
2159busy:
2160 if (has_write_io_error)
2161 jbd2_journal_abort(journal, -EIO);
2162
2163 return ret;
2164}
2165
2166/*
2167 * This buffer is no longer needed. If it is on an older transaction's
2168 * checkpoint list we need to record it on this transaction's forget list
2169 * to pin this buffer (and hence its checkpointing transaction) down until
2170 * this transaction commits. If the buffer isn't on a checkpoint list, we
2171 * release it.
2172 * Returns non-zero if JBD no longer has an interest in the buffer.
2173 *
2174 * Called under j_list_lock.
2175 *
2176 * Called under jh->b_state_lock.
2177 */
2178static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2179{
2180 int may_free = 1;
2181 struct buffer_head *bh = jh2bh(jh);
2182
2183 if (jh->b_cp_transaction) {
2184 JBUFFER_TRACE(jh, "on running+cp transaction");
2185 __jbd2_journal_temp_unlink_buffer(jh);
2186 /*
2187 * We don't want to write the buffer anymore, clear the
2188 * bit so that we don't confuse checks in
2189 * __journal_file_buffer
2190 */
2191 clear_buffer_dirty(bh);
2192 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2193 may_free = 0;
2194 } else {
2195 JBUFFER_TRACE(jh, "on running transaction");
2196 __jbd2_journal_unfile_buffer(jh);
2197 jbd2_journal_put_journal_head(jh);
2198 }
2199 return may_free;
2200}
2201
2202/*
2203 * jbd2_journal_invalidatepage
2204 *
2205 * This code is tricky. It has a number of cases to deal with.
2206 *
2207 * There are two invariants which this code relies on:
2208 *
2209 * i_size must be updated on disk before we start calling invalidatepage on the
2210 * data.
2211 *
2212 * This is done in ext3 by defining an ext3_setattr method which
2213 * updates i_size before truncate gets going. By maintaining this
2214 * invariant, we can be sure that it is safe to throw away any buffers
2215 * attached to the current transaction: once the transaction commits,
2216 * we know that the data will not be needed.
2217 *
2218 * Note however that we can *not* throw away data belonging to the
2219 * previous, committing transaction!
2220 *
2221 * Any disk blocks which *are* part of the previous, committing
2222 * transaction (and which therefore cannot be discarded immediately) are
2223 * not going to be reused in the new running transaction
2224 *
2225 * The bitmap committed_data images guarantee this: any block which is
2226 * allocated in one transaction and removed in the next will be marked
2227 * as in-use in the committed_data bitmap, so cannot be reused until
2228 * the next transaction to delete the block commits. This means that
2229 * leaving committing buffers dirty is quite safe: the disk blocks
2230 * cannot be reallocated to a different file and so buffer aliasing is
2231 * not possible.
2232 *
2233 *
2234 * The above applies mainly to ordered data mode. In writeback mode we
2235 * don't make guarantees about the order in which data hits disk --- in
2236 * particular we don't guarantee that new dirty data is flushed before
2237 * transaction commit --- so it is always safe just to discard data
2238 * immediately in that mode. --sct
2239 */
2240
2241/*
2242 * The journal_unmap_buffer helper function returns zero if the buffer
2243 * concerned remains pinned as an anonymous buffer belonging to an older
2244 * transaction.
2245 *
2246 * We're outside-transaction here. Either or both of j_running_transaction
2247 * and j_committing_transaction may be NULL.
2248 */
2249static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2250 int partial_page)
2251{
2252 transaction_t *transaction;
2253 struct journal_head *jh;
2254 int may_free = 1;
2255
2256 BUFFER_TRACE(bh, "entry");
2257
2258 /*
2259 * It is safe to proceed here without the j_list_lock because the
2260 * buffers cannot be stolen by try_to_free_buffers as long as we are
2261 * holding the page lock. --sct
2262 */
2263
2264 jh = jbd2_journal_grab_journal_head(bh);
2265 if (!jh)
2266 goto zap_buffer_unlocked;
2267
2268 /* OK, we have data buffer in journaled mode */
2269 write_lock(&journal->j_state_lock);
2270 spin_lock(&jh->b_state_lock);
2271 spin_lock(&journal->j_list_lock);
2272
2273 /*
2274 * We cannot remove the buffer from checkpoint lists until the
2275 * transaction adding inode to orphan list (let's call it T)
2276 * is committed. Otherwise if the transaction changing the
2277 * buffer would be cleaned from the journal before T is
2278 * committed, a crash will cause that the correct contents of
2279 * the buffer will be lost. On the other hand we have to
2280 * clear the buffer dirty bit at latest at the moment when the
2281 * transaction marking the buffer as freed in the filesystem
2282 * structures is committed because from that moment on the
2283 * block can be reallocated and used by a different page.
2284 * Since the block hasn't been freed yet but the inode has
2285 * already been added to orphan list, it is safe for us to add
2286 * the buffer to BJ_Forget list of the newest transaction.
2287 *
2288 * Also we have to clear buffer_mapped flag of a truncated buffer
2289 * because the buffer_head may be attached to the page straddling
2290 * i_size (can happen only when blocksize < pagesize) and thus the
2291 * buffer_head can be reused when the file is extended again. So we end
2292 * up keeping around invalidated buffers attached to transactions'
2293 * BJ_Forget list just to stop checkpointing code from cleaning up
2294 * the transaction this buffer was modified in.
2295 */
2296 transaction = jh->b_transaction;
2297 if (transaction == NULL) {
2298 /* First case: not on any transaction. If it
2299 * has no checkpoint link, then we can zap it:
2300 * it's a writeback-mode buffer so we don't care
2301 * if it hits disk safely. */
2302 if (!jh->b_cp_transaction) {
2303 JBUFFER_TRACE(jh, "not on any transaction: zap");
2304 goto zap_buffer;
2305 }
2306
2307 if (!buffer_dirty(bh)) {
2308 /* bdflush has written it. We can drop it now */
2309 __jbd2_journal_remove_checkpoint(jh);
2310 goto zap_buffer;
2311 }
2312
2313 /* OK, it must be in the journal but still not
2314 * written fully to disk: it's metadata or
2315 * journaled data... */
2316
2317 if (journal->j_running_transaction) {
2318 /* ... and once the current transaction has
2319 * committed, the buffer won't be needed any
2320 * longer. */
2321 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2322 may_free = __dispose_buffer(jh,
2323 journal->j_running_transaction);
2324 goto zap_buffer;
2325 } else {
2326 /* There is no currently-running transaction. So the
2327 * orphan record which we wrote for this file must have
2328 * passed into commit. We must attach this buffer to
2329 * the committing transaction, if it exists. */
2330 if (journal->j_committing_transaction) {
2331 JBUFFER_TRACE(jh, "give to committing trans");
2332 may_free = __dispose_buffer(jh,
2333 journal->j_committing_transaction);
2334 goto zap_buffer;
2335 } else {
2336 /* The orphan record's transaction has
2337 * committed. We can cleanse this buffer */
2338 clear_buffer_jbddirty(bh);
2339 __jbd2_journal_remove_checkpoint(jh);
2340 goto zap_buffer;
2341 }
2342 }
2343 } else if (transaction == journal->j_committing_transaction) {
2344 JBUFFER_TRACE(jh, "on committing transaction");
2345 /*
2346 * The buffer is committing, we simply cannot touch
2347 * it. If the page is straddling i_size we have to wait
2348 * for commit and try again.
2349 */
2350 if (partial_page) {
2351 spin_unlock(&journal->j_list_lock);
2352 spin_unlock(&jh->b_state_lock);
2353 write_unlock(&journal->j_state_lock);
2354 jbd2_journal_put_journal_head(jh);
2355 return -EBUSY;
2356 }
2357 /*
2358 * OK, buffer won't be reachable after truncate. We just clear
2359 * b_modified to not confuse transaction credit accounting, and
2360 * set j_next_transaction to the running transaction (if there
2361 * is one) and mark buffer as freed so that commit code knows
2362 * it should clear dirty bits when it is done with the buffer.
2363 */
2364 set_buffer_freed(bh);
2365 if (journal->j_running_transaction && buffer_jbddirty(bh))
2366 jh->b_next_transaction = journal->j_running_transaction;
2367 jh->b_modified = 0;
2368 spin_unlock(&journal->j_list_lock);
2369 spin_unlock(&jh->b_state_lock);
2370 write_unlock(&journal->j_state_lock);
2371 jbd2_journal_put_journal_head(jh);
2372 return 0;
2373 } else {
2374 /* Good, the buffer belongs to the running transaction.
2375 * We are writing our own transaction's data, not any
2376 * previous one's, so it is safe to throw it away
2377 * (remember that we expect the filesystem to have set
2378 * i_size already for this truncate so recovery will not
2379 * expose the disk blocks we are discarding here.) */
2380 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2381 JBUFFER_TRACE(jh, "on running transaction");
2382 may_free = __dispose_buffer(jh, transaction);
2383 }
2384
2385zap_buffer:
2386 /*
2387 * This is tricky. Although the buffer is truncated, it may be reused
2388 * if blocksize < pagesize and it is attached to the page straddling
2389 * EOF. Since the buffer might have been added to BJ_Forget list of the
2390 * running transaction, journal_get_write_access() won't clear
2391 * b_modified and credit accounting gets confused. So clear b_modified
2392 * here.
2393 */
2394 jh->b_modified = 0;
2395 spin_unlock(&journal->j_list_lock);
2396 spin_unlock(&jh->b_state_lock);
2397 write_unlock(&journal->j_state_lock);
2398 jbd2_journal_put_journal_head(jh);
2399zap_buffer_unlocked:
2400 clear_buffer_dirty(bh);
2401 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2402 clear_buffer_mapped(bh);
2403 clear_buffer_req(bh);
2404 clear_buffer_new(bh);
2405 clear_buffer_delay(bh);
2406 clear_buffer_unwritten(bh);
2407 bh->b_bdev = NULL;
2408 return may_free;
2409}
2410
2411/**
2412 * void jbd2_journal_invalidatepage()
2413 * @journal: journal to use for flush...
2414 * @page: page to flush
2415 * @offset: start of the range to invalidate
2416 * @length: length of the range to invalidate
2417 *
2418 * Reap page buffers containing data after in the specified range in page.
2419 * Can return -EBUSY if buffers are part of the committing transaction and
2420 * the page is straddling i_size. Caller then has to wait for current commit
2421 * and try again.
2422 */
2423int jbd2_journal_invalidatepage(journal_t *journal,
2424 struct page *page,
2425 unsigned int offset,
2426 unsigned int length)
2427{
2428 struct buffer_head *head, *bh, *next;
2429 unsigned int stop = offset + length;
2430 unsigned int curr_off = 0;
2431 int partial_page = (offset || length < PAGE_SIZE);
2432 int may_free = 1;
2433 int ret = 0;
2434
2435 if (!PageLocked(page))
2436 BUG();
2437 if (!page_has_buffers(page))
2438 return 0;
2439
2440 BUG_ON(stop > PAGE_SIZE || stop < length);
2441
2442 /* We will potentially be playing with lists other than just the
2443 * data lists (especially for journaled data mode), so be
2444 * cautious in our locking. */
2445
2446 head = bh = page_buffers(page);
2447 do {
2448 unsigned int next_off = curr_off + bh->b_size;
2449 next = bh->b_this_page;
2450
2451 if (next_off > stop)
2452 return 0;
2453
2454 if (offset <= curr_off) {
2455 /* This block is wholly outside the truncation point */
2456 lock_buffer(bh);
2457 ret = journal_unmap_buffer(journal, bh, partial_page);
2458 unlock_buffer(bh);
2459 if (ret < 0)
2460 return ret;
2461 may_free &= ret;
2462 }
2463 curr_off = next_off;
2464 bh = next;
2465
2466 } while (bh != head);
2467
2468 if (!partial_page) {
2469 if (may_free && try_to_free_buffers(page))
2470 J_ASSERT(!page_has_buffers(page));
2471 }
2472 return 0;
2473}
2474
2475/*
2476 * File a buffer on the given transaction list.
2477 */
2478void __jbd2_journal_file_buffer(struct journal_head *jh,
2479 transaction_t *transaction, int jlist)
2480{
2481 struct journal_head **list = NULL;
2482 int was_dirty = 0;
2483 struct buffer_head *bh = jh2bh(jh);
2484
2485 lockdep_assert_held(&jh->b_state_lock);
2486 assert_spin_locked(&transaction->t_journal->j_list_lock);
2487
2488 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2489 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2490 jh->b_transaction == NULL);
2491
2492 if (jh->b_transaction && jh->b_jlist == jlist)
2493 return;
2494
2495 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2496 jlist == BJ_Shadow || jlist == BJ_Forget) {
2497 /*
2498 * For metadata buffers, we track dirty bit in buffer_jbddirty
2499 * instead of buffer_dirty. We should not see a dirty bit set
2500 * here because we clear it in do_get_write_access but e.g.
2501 * tune2fs can modify the sb and set the dirty bit at any time
2502 * so we try to gracefully handle that.
2503 */
2504 if (buffer_dirty(bh))
2505 warn_dirty_buffer(bh);
2506 if (test_clear_buffer_dirty(bh) ||
2507 test_clear_buffer_jbddirty(bh))
2508 was_dirty = 1;
2509 }
2510
2511 if (jh->b_transaction)
2512 __jbd2_journal_temp_unlink_buffer(jh);
2513 else
2514 jbd2_journal_grab_journal_head(bh);
2515 jh->b_transaction = transaction;
2516
2517 switch (jlist) {
2518 case BJ_None:
2519 J_ASSERT_JH(jh, !jh->b_committed_data);
2520 J_ASSERT_JH(jh, !jh->b_frozen_data);
2521 return;
2522 case BJ_Metadata:
2523 transaction->t_nr_buffers++;
2524 list = &transaction->t_buffers;
2525 break;
2526 case BJ_Forget:
2527 list = &transaction->t_forget;
2528 break;
2529 case BJ_Shadow:
2530 list = &transaction->t_shadow_list;
2531 break;
2532 case BJ_Reserved:
2533 list = &transaction->t_reserved_list;
2534 break;
2535 }
2536
2537 __blist_add_buffer(list, jh);
2538 jh->b_jlist = jlist;
2539
2540 if (was_dirty)
2541 set_buffer_jbddirty(bh);
2542}
2543
2544void jbd2_journal_file_buffer(struct journal_head *jh,
2545 transaction_t *transaction, int jlist)
2546{
2547 spin_lock(&jh->b_state_lock);
2548 spin_lock(&transaction->t_journal->j_list_lock);
2549 __jbd2_journal_file_buffer(jh, transaction, jlist);
2550 spin_unlock(&transaction->t_journal->j_list_lock);
2551 spin_unlock(&jh->b_state_lock);
2552}
2553
2554/*
2555 * Remove a buffer from its current buffer list in preparation for
2556 * dropping it from its current transaction entirely. If the buffer has
2557 * already started to be used by a subsequent transaction, refile the
2558 * buffer on that transaction's metadata list.
2559 *
2560 * Called under j_list_lock
2561 * Called under jh->b_state_lock
2562 *
2563 * When this function returns true, there's no next transaction to refile to
2564 * and the caller has to drop jh reference through
2565 * jbd2_journal_put_journal_head().
2566 */
2567bool __jbd2_journal_refile_buffer(struct journal_head *jh)
2568{
2569 int was_dirty, jlist;
2570 struct buffer_head *bh = jh2bh(jh);
2571
2572 lockdep_assert_held(&jh->b_state_lock);
2573 if (jh->b_transaction)
2574 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2575
2576 /* If the buffer is now unused, just drop it. */
2577 if (jh->b_next_transaction == NULL) {
2578 __jbd2_journal_unfile_buffer(jh);
2579 return true;
2580 }
2581
2582 /*
2583 * It has been modified by a later transaction: add it to the new
2584 * transaction's metadata list.
2585 */
2586
2587 was_dirty = test_clear_buffer_jbddirty(bh);
2588 __jbd2_journal_temp_unlink_buffer(jh);
2589
2590 /*
2591 * b_transaction must be set, otherwise the new b_transaction won't
2592 * be holding jh reference
2593 */
2594 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2595
2596 /*
2597 * We set b_transaction here because b_next_transaction will inherit
2598 * our jh reference and thus __jbd2_journal_file_buffer() must not
2599 * take a new one.
2600 */
2601 WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2602 WRITE_ONCE(jh->b_next_transaction, NULL);
2603 if (buffer_freed(bh))
2604 jlist = BJ_Forget;
2605 else if (jh->b_modified)
2606 jlist = BJ_Metadata;
2607 else
2608 jlist = BJ_Reserved;
2609 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2610 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2611
2612 if (was_dirty)
2613 set_buffer_jbddirty(bh);
2614 return false;
2615}
2616
2617/*
2618 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2619 * bh reference so that we can safely unlock bh.
2620 *
2621 * The jh and bh may be freed by this call.
2622 */
2623void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2624{
2625 bool drop;
2626
2627 spin_lock(&jh->b_state_lock);
2628 spin_lock(&journal->j_list_lock);
2629 drop = __jbd2_journal_refile_buffer(jh);
2630 spin_unlock(&jh->b_state_lock);
2631 spin_unlock(&journal->j_list_lock);
2632 if (drop)
2633 jbd2_journal_put_journal_head(jh);
2634}
2635
2636/*
2637 * File inode in the inode list of the handle's transaction
2638 */
2639static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2640 unsigned long flags, loff_t start_byte, loff_t end_byte)
2641{
2642 transaction_t *transaction = handle->h_transaction;
2643 journal_t *journal;
2644
2645 if (is_handle_aborted(handle))
2646 return -EROFS;
2647 journal = transaction->t_journal;
2648
2649 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2650 transaction->t_tid);
2651
2652 spin_lock(&journal->j_list_lock);
2653 jinode->i_flags |= flags;
2654
2655 if (jinode->i_dirty_end) {
2656 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2657 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2658 } else {
2659 jinode->i_dirty_start = start_byte;
2660 jinode->i_dirty_end = end_byte;
2661 }
2662
2663 /* Is inode already attached where we need it? */
2664 if (jinode->i_transaction == transaction ||
2665 jinode->i_next_transaction == transaction)
2666 goto done;
2667
2668 /*
2669 * We only ever set this variable to 1 so the test is safe. Since
2670 * t_need_data_flush is likely to be set, we do the test to save some
2671 * cacheline bouncing
2672 */
2673 if (!transaction->t_need_data_flush)
2674 transaction->t_need_data_flush = 1;
2675 /* On some different transaction's list - should be
2676 * the committing one */
2677 if (jinode->i_transaction) {
2678 J_ASSERT(jinode->i_next_transaction == NULL);
2679 J_ASSERT(jinode->i_transaction ==
2680 journal->j_committing_transaction);
2681 jinode->i_next_transaction = transaction;
2682 goto done;
2683 }
2684 /* Not on any transaction list... */
2685 J_ASSERT(!jinode->i_next_transaction);
2686 jinode->i_transaction = transaction;
2687 list_add(&jinode->i_list, &transaction->t_inode_list);
2688done:
2689 spin_unlock(&journal->j_list_lock);
2690
2691 return 0;
2692}
2693
2694int jbd2_journal_inode_ranged_write(handle_t *handle,
2695 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2696{
2697 return jbd2_journal_file_inode(handle, jinode,
2698 JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2699 start_byte + length - 1);
2700}
2701
2702int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2703 loff_t start_byte, loff_t length)
2704{
2705 return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2706 start_byte, start_byte + length - 1);
2707}
2708
2709/*
2710 * File truncate and transaction commit interact with each other in a
2711 * non-trivial way. If a transaction writing data block A is
2712 * committing, we cannot discard the data by truncate until we have
2713 * written them. Otherwise if we crashed after the transaction with
2714 * write has committed but before the transaction with truncate has
2715 * committed, we could see stale data in block A. This function is a
2716 * helper to solve this problem. It starts writeout of the truncated
2717 * part in case it is in the committing transaction.
2718 *
2719 * Filesystem code must call this function when inode is journaled in
2720 * ordered mode before truncation happens and after the inode has been
2721 * placed on orphan list with the new inode size. The second condition
2722 * avoids the race that someone writes new data and we start
2723 * committing the transaction after this function has been called but
2724 * before a transaction for truncate is started (and furthermore it
2725 * allows us to optimize the case where the addition to orphan list
2726 * happens in the same transaction as write --- we don't have to write
2727 * any data in such case).
2728 */
2729int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2730 struct jbd2_inode *jinode,
2731 loff_t new_size)
2732{
2733 transaction_t *inode_trans, *commit_trans;
2734 int ret = 0;
2735
2736 /* This is a quick check to avoid locking if not necessary */
2737 if (!jinode->i_transaction)
2738 goto out;
2739 /* Locks are here just to force reading of recent values, it is
2740 * enough that the transaction was not committing before we started
2741 * a transaction adding the inode to orphan list */
2742 read_lock(&journal->j_state_lock);
2743 commit_trans = journal->j_committing_transaction;
2744 read_unlock(&journal->j_state_lock);
2745 spin_lock(&journal->j_list_lock);
2746 inode_trans = jinode->i_transaction;
2747 spin_unlock(&journal->j_list_lock);
2748 if (inode_trans == commit_trans) {
2749 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2750 new_size, LLONG_MAX);
2751 if (ret)
2752 jbd2_journal_abort(journal, ret);
2753 }
2754out:
2755 return ret;
2756}