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