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(s: 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(s: transaction_cache, objp: transaction);
63}
64
65/*
66 * jbd2_get_transaction: obtain a new transaction_t object.
67 *
68 * Simply initialise a new transaction. Initialize it in
69 * RUNNING state and add it to the current journal (which should not
70 * have an existing running transaction: we only make a new transaction
71 * once we have started to commit the old one).
72 *
73 * Preconditions:
74 * The journal MUST be locked. We don't perform atomic mallocs on the
75 * new transaction and we can't block without protecting against other
76 * processes trying to touch the journal while it is in transition.
77 *
78 */
79
80static void jbd2_get_transaction(journal_t *journal,
81 transaction_t *transaction)
82{
83 transaction->t_journal = journal;
84 transaction->t_state = T_RUNNING;
85 transaction->t_start_time = ktime_get();
86 transaction->t_tid = journal->j_transaction_sequence++;
87 transaction->t_expires = jiffies + journal->j_commit_interval;
88 atomic_set(v: &transaction->t_updates, i: 0);
89 atomic_set(v: &transaction->t_outstanding_credits,
90 i: journal->j_transaction_overhead_buffers +
91 atomic_read(v: &journal->j_reserved_credits));
92 atomic_set(v: &transaction->t_outstanding_revokes, i: 0);
93 atomic_set(v: &transaction->t_handle_count, i: 0);
94 INIT_LIST_HEAD(list: &transaction->t_inode_list);
95
96 /* Set up the commit timer for the new transaction. */
97 journal->j_commit_timer.expires = round_jiffies_up(j: transaction->t_expires);
98 add_timer(timer: &journal->j_commit_timer);
99
100 J_ASSERT(journal->j_running_transaction == NULL);
101 journal->j_running_transaction = transaction;
102 transaction->t_max_wait = 0;
103 transaction->t_start = jiffies;
104 transaction->t_requested = 0;
105}
106
107/*
108 * Handle management.
109 *
110 * A handle_t is an object which represents a single atomic update to a
111 * filesystem, and which tracks all of the modifications which form part
112 * of that one update.
113 */
114
115/*
116 * t_max_wait is carefully updated here with use of atomic compare exchange.
117 * Note that there could be multiplre threads trying to do this simultaneously
118 * hence using cmpxchg to avoid any use of locks in this case.
119 */
120static inline void update_t_max_wait(transaction_t *transaction,
121 unsigned long ts)
122{
123 unsigned long oldts, newts;
124
125 if (time_after(transaction->t_start, ts)) {
126 newts = jbd2_time_diff(start: ts, end: transaction->t_start);
127 oldts = READ_ONCE(transaction->t_max_wait);
128 while (oldts < newts)
129 oldts = cmpxchg(&transaction->t_max_wait, oldts, newts);
130 }
131}
132
133/*
134 * Wait until running transaction passes to T_FLUSH state and new transaction
135 * can thus be started. Also starts the commit if needed. The function expects
136 * running transaction to exist and releases j_state_lock.
137 */
138static void wait_transaction_locked(journal_t *journal)
139 __releases(journal->j_state_lock)
140{
141 DEFINE_WAIT(wait);
142 int need_to_start;
143 tid_t tid = journal->j_running_transaction->t_tid;
144
145 prepare_to_wait_exclusive(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait,
146 TASK_UNINTERRUPTIBLE);
147 need_to_start = !tid_geq(x: journal->j_commit_request, y: tid);
148 read_unlock(&journal->j_state_lock);
149 if (need_to_start)
150 jbd2_log_start_commit(journal, tid);
151 jbd2_might_wait_for_commit(journal);
152 schedule();
153 finish_wait(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait);
154}
155
156/*
157 * Wait until running transaction transitions from T_SWITCH to T_FLUSH
158 * state and new transaction can thus be started. The function releases
159 * j_state_lock.
160 */
161static void wait_transaction_switching(journal_t *journal)
162 __releases(journal->j_state_lock)
163{
164 DEFINE_WAIT(wait);
165
166 if (WARN_ON(!journal->j_running_transaction ||
167 journal->j_running_transaction->t_state != T_SWITCH)) {
168 read_unlock(&journal->j_state_lock);
169 return;
170 }
171 prepare_to_wait_exclusive(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait,
172 TASK_UNINTERRUPTIBLE);
173 read_unlock(&journal->j_state_lock);
174 /*
175 * We don't call jbd2_might_wait_for_commit() here as there's no
176 * waiting for outstanding handles happening anymore in T_SWITCH state
177 * and handling of reserved handles actually relies on that for
178 * correctness.
179 */
180 schedule();
181 finish_wait(wq_head: &journal->j_wait_transaction_locked, wq_entry: &wait);
182}
183
184static void sub_reserved_credits(journal_t *journal, int blocks)
185{
186 atomic_sub(i: blocks, v: &journal->j_reserved_credits);
187 wake_up(&journal->j_wait_reserved);
188}
189
190/* Maximum number of blocks for user transaction payload */
191static int jbd2_max_user_trans_buffers(journal_t *journal)
192{
193 return journal->j_max_transaction_buffers -
194 journal->j_transaction_overhead_buffers;
195}
196
197/*
198 * Wait until we can add credits for handle to the running transaction. Called
199 * with j_state_lock held for reading. Returns 0 if handle joined the running
200 * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
201 * caller must retry.
202 *
203 * Note: because j_state_lock may be dropped depending on the return
204 * value, we need to fake out sparse so ti doesn't complain about a
205 * locking imbalance. Callers of add_transaction_credits will need to
206 * make a similar accomodation.
207 */
208static int add_transaction_credits(journal_t *journal, int blocks,
209 int rsv_blocks)
210__must_hold(&journal->j_state_lock)
211{
212 transaction_t *t = journal->j_running_transaction;
213 int needed;
214 int total = blocks + rsv_blocks;
215
216 /*
217 * If the current transaction is locked down for commit, wait
218 * for the lock to be released.
219 */
220 if (t->t_state != T_RUNNING) {
221 WARN_ON_ONCE(t->t_state >= T_FLUSH);
222 wait_transaction_locked(journal);
223 __acquire(&journal->j_state_lock); /* fake out sparse */
224 return 1;
225 }
226
227 /*
228 * If there is not enough space left in the log to write all
229 * potential buffers requested by this operation, we need to
230 * stall pending a log checkpoint to free some more log space.
231 */
232 needed = atomic_add_return(i: total, v: &t->t_outstanding_credits);
233 if (needed > journal->j_max_transaction_buffers) {
234 /*
235 * If the current transaction is already too large,
236 * then start to commit it: we can then go back and
237 * attach this handle to a new transaction.
238 */
239 atomic_sub(i: total, v: &t->t_outstanding_credits);
240
241 /*
242 * Is the number of reserved credits in the current transaction too
243 * big to fit this handle? Wait until reserved credits are freed.
244 */
245 if (atomic_read(v: &journal->j_reserved_credits) + total >
246 jbd2_max_user_trans_buffers(journal)) {
247 read_unlock(&journal->j_state_lock);
248 jbd2_might_wait_for_commit(journal);
249 wait_event(journal->j_wait_reserved,
250 atomic_read(&journal->j_reserved_credits) + total <=
251 jbd2_max_user_trans_buffers(journal));
252 __acquire(&journal->j_state_lock); /* fake out sparse */
253 return 1;
254 }
255
256 wait_transaction_locked(journal);
257 __acquire(&journal->j_state_lock); /* fake out sparse */
258 return 1;
259 }
260
261 /*
262 * The commit code assumes that it can get enough log space
263 * without forcing a checkpoint. This is *critical* for
264 * correctness: a checkpoint of a buffer which is also
265 * associated with a committing transaction creates a deadlock,
266 * so commit simply cannot force through checkpoints.
267 *
268 * We must therefore ensure the necessary space in the journal
269 * *before* starting to dirty potentially checkpointed buffers
270 * in the new transaction.
271 */
272 if (jbd2_log_space_left(journal) < journal->j_max_transaction_buffers) {
273 atomic_sub(i: total, v: &t->t_outstanding_credits);
274 read_unlock(&journal->j_state_lock);
275 jbd2_might_wait_for_commit(journal);
276 write_lock(&journal->j_state_lock);
277 if (jbd2_log_space_left(journal) <
278 journal->j_max_transaction_buffers)
279 __jbd2_log_wait_for_space(journal);
280 write_unlock(&journal->j_state_lock);
281 __acquire(&journal->j_state_lock); /* fake out sparse */
282 return 1;
283 }
284
285 /* No reservation? We are done... */
286 if (!rsv_blocks)
287 return 0;
288
289 needed = atomic_add_return(i: rsv_blocks, v: &journal->j_reserved_credits);
290 /* We allow at most half of a transaction to be reserved */
291 if (needed > jbd2_max_user_trans_buffers(journal) / 2) {
292 sub_reserved_credits(journal, blocks: rsv_blocks);
293 atomic_sub(i: total, v: &t->t_outstanding_credits);
294 read_unlock(&journal->j_state_lock);
295 jbd2_might_wait_for_commit(journal);
296 wait_event(journal->j_wait_reserved,
297 atomic_read(&journal->j_reserved_credits) + rsv_blocks
298 <= jbd2_max_user_trans_buffers(journal) / 2);
299 __acquire(&journal->j_state_lock); /* fake out sparse */
300 return 1;
301 }
302 return 0;
303}
304
305/*
306 * start_this_handle: Given a handle, deal with any locking or stalling
307 * needed to make sure that there is enough journal space for the handle
308 * to begin. Attach the handle to a transaction and set up the
309 * transaction's buffer credits.
310 */
311
312static int start_this_handle(journal_t *journal, handle_t *handle,
313 gfp_t gfp_mask)
314{
315 transaction_t *transaction, *new_transaction = NULL;
316 int blocks = handle->h_total_credits;
317 int rsv_blocks = 0;
318 unsigned long ts = jiffies;
319
320 if (handle->h_rsv_handle)
321 rsv_blocks = handle->h_rsv_handle->h_total_credits;
322
323 /*
324 * Limit the number of reserved credits to 1/2 of maximum transaction
325 * size and limit the number of total credits to not exceed maximum
326 * transaction size per operation.
327 */
328 if (rsv_blocks > jbd2_max_user_trans_buffers(journal) / 2 ||
329 rsv_blocks + blocks > jbd2_max_user_trans_buffers(journal)) {
330 printk(KERN_ERR "JBD2: %s wants too many credits "
331 "credits:%d rsv_credits:%d max:%d\n",
332 current->comm, blocks, rsv_blocks,
333 jbd2_max_user_trans_buffers(journal));
334 WARN_ON(1);
335 return -ENOSPC;
336 }
337
338alloc_transaction:
339 /*
340 * This check is racy but it is just an optimization of allocating new
341 * transaction early if there are high chances we'll need it. If we
342 * guess wrong, we'll retry or free unused transaction.
343 */
344 if (!data_race(journal->j_running_transaction)) {
345 /*
346 * If __GFP_FS is not present, then we may be being called from
347 * inside the fs writeback layer, so we MUST NOT fail.
348 */
349 if ((gfp_mask & __GFP_FS) == 0)
350 gfp_mask |= __GFP_NOFAIL;
351 new_transaction = kmem_cache_zalloc(transaction_cache,
352 gfp_mask);
353 if (!new_transaction)
354 return -ENOMEM;
355 }
356
357 jbd2_debug(3, "New handle %p going live.\n", handle);
358
359 /*
360 * We need to hold j_state_lock until t_updates has been incremented,
361 * for proper journal barrier handling
362 */
363repeat:
364 read_lock(&journal->j_state_lock);
365 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
366 if (is_journal_aborted(journal) ||
367 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
368 read_unlock(&journal->j_state_lock);
369 jbd2_journal_free_transaction(transaction: new_transaction);
370 return -EROFS;
371 }
372
373 /*
374 * Wait on the journal's transaction barrier if necessary. Specifically
375 * we allow reserved handles to proceed because otherwise commit could
376 * deadlock on page writeback not being able to complete.
377 */
378 if (!handle->h_reserved && journal->j_barrier_count) {
379 read_unlock(&journal->j_state_lock);
380 wait_event(journal->j_wait_transaction_locked,
381 journal->j_barrier_count == 0);
382 goto repeat;
383 }
384
385 if (!journal->j_running_transaction) {
386 read_unlock(&journal->j_state_lock);
387 if (!new_transaction)
388 goto alloc_transaction;
389 write_lock(&journal->j_state_lock);
390 if (!journal->j_running_transaction &&
391 (handle->h_reserved || !journal->j_barrier_count)) {
392 jbd2_get_transaction(journal, transaction: new_transaction);
393 new_transaction = NULL;
394 }
395 write_unlock(&journal->j_state_lock);
396 goto repeat;
397 }
398
399 transaction = journal->j_running_transaction;
400
401 if (!handle->h_reserved) {
402 /* We may have dropped j_state_lock - restart in that case */
403 if (add_transaction_credits(journal, blocks, rsv_blocks)) {
404 /*
405 * add_transaction_credits releases
406 * j_state_lock on a non-zero return
407 */
408 __release(&journal->j_state_lock);
409 goto repeat;
410 }
411 } else {
412 /*
413 * We have handle reserved so we are allowed to join T_LOCKED
414 * transaction and we don't have to check for transaction size
415 * and journal space. But we still have to wait while running
416 * transaction is being switched to a committing one as it
417 * won't wait for any handles anymore.
418 */
419 if (transaction->t_state == T_SWITCH) {
420 wait_transaction_switching(journal);
421 goto repeat;
422 }
423 sub_reserved_credits(journal, blocks);
424 handle->h_reserved = 0;
425 }
426
427 /* OK, account for the buffers that this operation expects to
428 * use and add the handle to the running transaction.
429 */
430 update_t_max_wait(transaction, ts);
431 handle->h_transaction = transaction;
432 handle->h_requested_credits = blocks;
433 handle->h_revoke_credits_requested = handle->h_revoke_credits;
434 handle->h_start_jiffies = jiffies;
435 atomic_inc(v: &transaction->t_updates);
436 atomic_inc(v: &transaction->t_handle_count);
437 jbd2_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
438 handle, blocks,
439 atomic_read(&transaction->t_outstanding_credits),
440 jbd2_log_space_left(journal));
441 read_unlock(&journal->j_state_lock);
442 current->journal_info = handle;
443
444 rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
445 jbd2_journal_free_transaction(transaction: new_transaction);
446 /*
447 * Ensure that no allocations done while the transaction is open are
448 * going to recurse back to the fs layer.
449 */
450 handle->saved_alloc_context = memalloc_nofs_save();
451 return 0;
452}
453
454/* Allocate a new handle. This should probably be in a slab... */
455static handle_t *new_handle(int nblocks)
456{
457 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
458 if (!handle)
459 return NULL;
460 handle->h_total_credits = nblocks;
461 handle->h_ref = 1;
462
463 return handle;
464}
465
466handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
467 int revoke_records, gfp_t gfp_mask,
468 unsigned int type, unsigned int line_no)
469{
470 handle_t *handle = journal_current_handle();
471 int err;
472
473 if (!journal)
474 return ERR_PTR(error: -EROFS);
475
476 if (handle) {
477 J_ASSERT(handle->h_transaction->t_journal == journal);
478 handle->h_ref++;
479 return handle;
480 }
481
482 nblocks += DIV_ROUND_UP(revoke_records,
483 journal->j_revoke_records_per_block);
484 handle = new_handle(nblocks);
485 if (!handle)
486 return ERR_PTR(error: -ENOMEM);
487 if (rsv_blocks) {
488 handle_t *rsv_handle;
489
490 rsv_handle = new_handle(nblocks: rsv_blocks);
491 if (!rsv_handle) {
492 jbd2_free_handle(handle);
493 return ERR_PTR(error: -ENOMEM);
494 }
495 rsv_handle->h_reserved = 1;
496 rsv_handle->h_journal = journal;
497 handle->h_rsv_handle = rsv_handle;
498 }
499 handle->h_revoke_credits = revoke_records;
500
501 err = start_this_handle(journal, handle, gfp_mask);
502 if (err < 0) {
503 if (handle->h_rsv_handle)
504 jbd2_free_handle(handle: handle->h_rsv_handle);
505 jbd2_free_handle(handle);
506 return ERR_PTR(error: err);
507 }
508 handle->h_type = type;
509 handle->h_line_no = line_no;
510 trace_jbd2_handle_start(dev: journal->j_fs_dev->bd_dev,
511 tid: handle->h_transaction->t_tid, type,
512 line_no, requested_blocks: nblocks);
513
514 return handle;
515}
516EXPORT_SYMBOL(jbd2__journal_start);
517
518
519/**
520 * jbd2_journal_start() - Obtain a new handle.
521 * @journal: Journal to start transaction on.
522 * @nblocks: number of block buffer we might modify
523 *
524 * We make sure that the transaction can guarantee at least nblocks of
525 * modified buffers in the log. We block until the log can guarantee
526 * that much space. Additionally, if rsv_blocks > 0, we also create another
527 * handle with rsv_blocks reserved blocks in the journal. This handle is
528 * stored in h_rsv_handle. It is not attached to any particular transaction
529 * and thus doesn't block transaction commit. If the caller uses this reserved
530 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
531 * on the parent handle will dispose the reserved one. Reserved handle has to
532 * be converted to a normal handle using jbd2_journal_start_reserved() before
533 * it can be used.
534 *
535 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
536 * on failure.
537 */
538handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
539{
540 return jbd2__journal_start(journal, nblocks, 0, 0, GFP_NOFS, 0, 0);
541}
542EXPORT_SYMBOL(jbd2_journal_start);
543
544static void __jbd2_journal_unreserve_handle(handle_t *handle, transaction_t *t)
545{
546 journal_t *journal = handle->h_journal;
547
548 WARN_ON(!handle->h_reserved);
549 sub_reserved_credits(journal, blocks: handle->h_total_credits);
550 if (t)
551 atomic_sub(i: handle->h_total_credits, v: &t->t_outstanding_credits);
552}
553
554void jbd2_journal_free_reserved(handle_t *handle)
555{
556 journal_t *journal = handle->h_journal;
557
558 /* Get j_state_lock to pin running transaction if it exists */
559 read_lock(&journal->j_state_lock);
560 __jbd2_journal_unreserve_handle(handle, t: journal->j_running_transaction);
561 read_unlock(&journal->j_state_lock);
562 jbd2_free_handle(handle);
563}
564EXPORT_SYMBOL(jbd2_journal_free_reserved);
565
566/**
567 * jbd2_journal_start_reserved() - start reserved handle
568 * @handle: handle to start
569 * @type: for handle statistics
570 * @line_no: for handle statistics
571 *
572 * Start handle that has been previously reserved with jbd2_journal_reserve().
573 * This attaches @handle to the running transaction (or creates one if there's
574 * not transaction running). Unlike jbd2_journal_start() this function cannot
575 * block on journal commit, checkpointing, or similar stuff. It can block on
576 * memory allocation or frozen journal though.
577 *
578 * Return 0 on success, non-zero on error - handle is freed in that case.
579 */
580int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
581 unsigned int line_no)
582{
583 journal_t *journal = handle->h_journal;
584 int ret = -EIO;
585
586 if (WARN_ON(!handle->h_reserved)) {
587 /* Someone passed in normal handle? Just stop it. */
588 jbd2_journal_stop(handle);
589 return ret;
590 }
591 /*
592 * Usefulness of mixing of reserved and unreserved handles is
593 * questionable. So far nobody seems to need it so just error out.
594 */
595 if (WARN_ON(current->journal_info)) {
596 jbd2_journal_free_reserved(handle);
597 return ret;
598 }
599
600 handle->h_journal = NULL;
601 /*
602 * GFP_NOFS is here because callers are likely from writeback or
603 * similarly constrained call sites
604 */
605 ret = start_this_handle(journal, handle, GFP_NOFS);
606 if (ret < 0) {
607 handle->h_journal = journal;
608 jbd2_journal_free_reserved(handle);
609 return ret;
610 }
611 handle->h_type = type;
612 handle->h_line_no = line_no;
613 trace_jbd2_handle_start(dev: journal->j_fs_dev->bd_dev,
614 tid: handle->h_transaction->t_tid, type,
615 line_no, requested_blocks: handle->h_total_credits);
616 return 0;
617}
618EXPORT_SYMBOL(jbd2_journal_start_reserved);
619
620/**
621 * jbd2_journal_extend() - extend buffer credits.
622 * @handle: handle to 'extend'
623 * @nblocks: nr blocks to try to extend by.
624 * @revoke_records: number of revoke records to try to extend by.
625 *
626 * Some transactions, such as large extends and truncates, can be done
627 * atomically all at once or in several stages. The operation requests
628 * a credit for a number of buffer modifications in advance, but can
629 * extend its credit if it needs more.
630 *
631 * jbd2_journal_extend tries to give the running handle more buffer credits.
632 * It does not guarantee that allocation - this is a best-effort only.
633 * The calling process MUST be able to deal cleanly with a failure to
634 * extend here.
635 *
636 * Return 0 on success, non-zero on failure.
637 *
638 * return code < 0 implies an error
639 * return code > 0 implies normal transaction-full status.
640 */
641int jbd2_journal_extend(handle_t *handle, int nblocks, int revoke_records)
642{
643 transaction_t *transaction = handle->h_transaction;
644 journal_t *journal;
645 int result;
646 int wanted;
647
648 if (is_handle_aborted(handle))
649 return -EROFS;
650 journal = transaction->t_journal;
651
652 result = 1;
653
654 read_lock(&journal->j_state_lock);
655
656 /* Don't extend a locked-down transaction! */
657 if (transaction->t_state != T_RUNNING) {
658 jbd2_debug(3, "denied handle %p %d blocks: "
659 "transaction not running\n", handle, nblocks);
660 goto error_out;
661 }
662
663 nblocks += DIV_ROUND_UP(
664 handle->h_revoke_credits_requested + revoke_records,
665 journal->j_revoke_records_per_block) -
666 DIV_ROUND_UP(
667 handle->h_revoke_credits_requested,
668 journal->j_revoke_records_per_block);
669 wanted = atomic_add_return(i: nblocks,
670 v: &transaction->t_outstanding_credits);
671
672 if (wanted > journal->j_max_transaction_buffers) {
673 jbd2_debug(3, "denied handle %p %d blocks: "
674 "transaction too large\n", handle, nblocks);
675 atomic_sub(i: nblocks, v: &transaction->t_outstanding_credits);
676 goto error_out;
677 }
678
679 trace_jbd2_handle_extend(dev: journal->j_fs_dev->bd_dev,
680 tid: transaction->t_tid,
681 type: handle->h_type, line_no: handle->h_line_no,
682 buffer_credits: handle->h_total_credits,
683 requested_blocks: nblocks);
684
685 handle->h_total_credits += nblocks;
686 handle->h_requested_credits += nblocks;
687 handle->h_revoke_credits += revoke_records;
688 handle->h_revoke_credits_requested += revoke_records;
689 result = 0;
690
691 jbd2_debug(3, "extended handle %p by %d\n", handle, nblocks);
692error_out:
693 read_unlock(&journal->j_state_lock);
694 return result;
695}
696
697static void stop_this_handle(handle_t *handle)
698{
699 transaction_t *transaction = handle->h_transaction;
700 journal_t *journal = transaction->t_journal;
701 int revokes;
702
703 J_ASSERT(journal_current_handle() == handle);
704 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
705 current->journal_info = NULL;
706 /*
707 * Subtract necessary revoke descriptor blocks from handle credits. We
708 * take care to account only for revoke descriptor blocks the
709 * transaction will really need as large sequences of transactions with
710 * small numbers of revokes are relatively common.
711 */
712 revokes = handle->h_revoke_credits_requested - handle->h_revoke_credits;
713 if (revokes) {
714 int t_revokes, revoke_descriptors;
715 int rr_per_blk = journal->j_revoke_records_per_block;
716
717 WARN_ON_ONCE(DIV_ROUND_UP(revokes, rr_per_blk)
718 > handle->h_total_credits);
719 t_revokes = atomic_add_return(i: revokes,
720 v: &transaction->t_outstanding_revokes);
721 revoke_descriptors =
722 DIV_ROUND_UP(t_revokes, rr_per_blk) -
723 DIV_ROUND_UP(t_revokes - revokes, rr_per_blk);
724 handle->h_total_credits -= revoke_descriptors;
725 }
726 atomic_sub(i: handle->h_total_credits,
727 v: &transaction->t_outstanding_credits);
728 if (handle->h_rsv_handle)
729 __jbd2_journal_unreserve_handle(handle: handle->h_rsv_handle,
730 t: transaction);
731 if (atomic_dec_and_test(v: &transaction->t_updates))
732 wake_up(&journal->j_wait_updates);
733
734 rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
735 /*
736 * Scope of the GFP_NOFS context is over here and so we can restore the
737 * original alloc context.
738 */
739 memalloc_nofs_restore(flags: handle->saved_alloc_context);
740}
741
742/**
743 * jbd2__journal_restart() - restart a handle .
744 * @handle: handle to restart
745 * @nblocks: nr credits requested
746 * @revoke_records: number of revoke record credits requested
747 * @gfp_mask: memory allocation flags (for start_this_handle)
748 *
749 * Restart a handle for a multi-transaction filesystem
750 * operation.
751 *
752 * If the jbd2_journal_extend() call above fails to grant new buffer credits
753 * to a running handle, a call to jbd2_journal_restart will commit the
754 * handle's transaction so far and reattach the handle to a new
755 * transaction capable of guaranteeing the requested number of
756 * credits. We preserve reserved handle if there's any attached to the
757 * passed in handle.
758 */
759int jbd2__journal_restart(handle_t *handle, int nblocks, int revoke_records,
760 gfp_t gfp_mask)
761{
762 transaction_t *transaction = handle->h_transaction;
763 journal_t *journal;
764 tid_t tid;
765 int need_to_start;
766 int ret;
767
768 /* If we've had an abort of any type, don't even think about
769 * actually doing the restart! */
770 if (is_handle_aborted(handle))
771 return 0;
772 journal = transaction->t_journal;
773 tid = transaction->t_tid;
774
775 /*
776 * First unlink the handle from its current transaction, and start the
777 * commit on that.
778 */
779 jbd2_debug(2, "restarting handle %p\n", handle);
780 stop_this_handle(handle);
781 handle->h_transaction = NULL;
782
783 /*
784 * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can
785 * get rid of pointless j_state_lock traffic like this.
786 */
787 read_lock(&journal->j_state_lock);
788 need_to_start = !tid_geq(x: journal->j_commit_request, y: tid);
789 read_unlock(&journal->j_state_lock);
790 if (need_to_start)
791 jbd2_log_start_commit(journal, tid);
792 handle->h_total_credits = nblocks +
793 DIV_ROUND_UP(revoke_records,
794 journal->j_revoke_records_per_block);
795 handle->h_revoke_credits = revoke_records;
796 ret = start_this_handle(journal, handle, gfp_mask);
797 trace_jbd2_handle_restart(dev: journal->j_fs_dev->bd_dev,
798 tid: ret ? 0 : handle->h_transaction->t_tid,
799 type: handle->h_type, line_no: handle->h_line_no,
800 requested_blocks: handle->h_total_credits);
801 return ret;
802}
803EXPORT_SYMBOL(jbd2__journal_restart);
804
805
806int jbd2_journal_restart(handle_t *handle, int nblocks)
807{
808 return jbd2__journal_restart(handle, nblocks, 0, GFP_NOFS);
809}
810EXPORT_SYMBOL(jbd2_journal_restart);
811
812/*
813 * Waits for any outstanding t_updates to finish.
814 * This is called with write j_state_lock held.
815 */
816void jbd2_journal_wait_updates(journal_t *journal)
817{
818 DEFINE_WAIT(wait);
819
820 while (1) {
821 /*
822 * Note that the running transaction can get freed under us if
823 * this transaction is getting committed in
824 * jbd2_journal_commit_transaction() ->
825 * jbd2_journal_free_transaction(). This can only happen when we
826 * release j_state_lock -> schedule() -> acquire j_state_lock.
827 * Hence we should everytime retrieve new j_running_transaction
828 * value (after j_state_lock release acquire cycle), else it may
829 * lead to use-after-free of old freed transaction.
830 */
831 transaction_t *transaction = journal->j_running_transaction;
832
833 if (!transaction)
834 break;
835
836 prepare_to_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait,
837 TASK_UNINTERRUPTIBLE);
838 if (!atomic_read(v: &transaction->t_updates)) {
839 finish_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait);
840 break;
841 }
842 write_unlock(&journal->j_state_lock);
843 schedule();
844 finish_wait(wq_head: &journal->j_wait_updates, wq_entry: &wait);
845 write_lock(&journal->j_state_lock);
846 }
847}
848
849/**
850 * jbd2_journal_lock_updates () - establish a transaction barrier.
851 * @journal: Journal to establish a barrier on.
852 *
853 * This locks out any further updates from being started, and blocks
854 * until all existing updates have completed, returning only once the
855 * journal is in a quiescent state with no updates running.
856 *
857 * The journal lock should not be held on entry.
858 */
859void jbd2_journal_lock_updates(journal_t *journal)
860{
861 jbd2_might_wait_for_commit(journal);
862
863 write_lock(&journal->j_state_lock);
864 ++journal->j_barrier_count;
865
866 /* Wait until there are no reserved handles */
867 if (atomic_read(v: &journal->j_reserved_credits)) {
868 write_unlock(&journal->j_state_lock);
869 wait_event(journal->j_wait_reserved,
870 atomic_read(&journal->j_reserved_credits) == 0);
871 write_lock(&journal->j_state_lock);
872 }
873
874 /* Wait until there are no running t_updates */
875 jbd2_journal_wait_updates(journal);
876
877 write_unlock(&journal->j_state_lock);
878
879 /*
880 * We have now established a barrier against other normal updates, but
881 * we also need to barrier against other jbd2_journal_lock_updates() calls
882 * to make sure that we serialise special journal-locked operations
883 * too.
884 */
885 mutex_lock(lock: &journal->j_barrier);
886}
887
888/**
889 * jbd2_journal_unlock_updates () - release barrier
890 * @journal: Journal to release the barrier on.
891 *
892 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
893 *
894 * Should be called without the journal lock held.
895 */
896void jbd2_journal_unlock_updates (journal_t *journal)
897{
898 J_ASSERT(journal->j_barrier_count != 0);
899
900 mutex_unlock(lock: &journal->j_barrier);
901 write_lock(&journal->j_state_lock);
902 --journal->j_barrier_count;
903 write_unlock(&journal->j_state_lock);
904 wake_up_all(&journal->j_wait_transaction_locked);
905}
906
907static void warn_dirty_buffer(struct buffer_head *bh)
908{
909 printk(KERN_WARNING
910 "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
911 "There's a risk of filesystem corruption in case of system "
912 "crash.\n",
913 bh->b_bdev, (unsigned long long)bh->b_blocknr);
914}
915
916/* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
917static void jbd2_freeze_jh_data(struct journal_head *jh)
918{
919 char *source;
920 struct buffer_head *bh = jh2bh(jh);
921
922 J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
923 source = kmap_local_folio(folio: bh->b_folio, offset: bh_offset(bh));
924 /* Fire data frozen trigger just before we copy the data */
925 jbd2_buffer_frozen_trigger(jh, mapped_data: source, triggers: jh->b_triggers);
926 memcpy(to: jh->b_frozen_data, from: source, len: bh->b_size);
927 kunmap_local(source);
928
929 /*
930 * Now that the frozen data is saved off, we need to store any matching
931 * triggers.
932 */
933 jh->b_frozen_triggers = jh->b_triggers;
934}
935
936/*
937 * If the buffer is already part of the current transaction, then there
938 * is nothing we need to do. If it is already part of a prior
939 * transaction which we are still committing to disk, then we need to
940 * make sure that we do not overwrite the old copy: we do copy-out to
941 * preserve the copy going to disk. We also account the buffer against
942 * the handle's metadata buffer credits (unless the buffer is already
943 * part of the transaction, that is).
944 *
945 */
946static int
947do_get_write_access(handle_t *handle, struct journal_head *jh,
948 int force_copy)
949{
950 struct buffer_head *bh;
951 transaction_t *transaction = handle->h_transaction;
952 journal_t *journal;
953 int error;
954 char *frozen_buffer = NULL;
955 unsigned long start_lock, time_lock;
956
957 journal = transaction->t_journal;
958
959 jbd2_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
960
961 JBUFFER_TRACE(jh, "entry");
962repeat:
963 bh = jh2bh(jh);
964
965 /* @@@ Need to check for errors here at some point. */
966
967 start_lock = jiffies;
968 lock_buffer(bh);
969 spin_lock(lock: &jh->b_state_lock);
970
971 /* If it takes too long to lock the buffer, trace it */
972 time_lock = jbd2_time_diff(start: start_lock, end: jiffies);
973 if (time_lock > HZ/10)
974 trace_jbd2_lock_buffer_stall(dev: bh->b_bdev->bd_dev,
975 stall_ms: jiffies_to_msecs(j: time_lock));
976
977 /* We now hold the buffer lock so it is safe to query the buffer
978 * state. Is the buffer dirty?
979 *
980 * If so, there are two possibilities. The buffer may be
981 * non-journaled, and undergoing a quite legitimate writeback.
982 * Otherwise, it is journaled, and we don't expect dirty buffers
983 * in that state (the buffers should be marked JBD_Dirty
984 * instead.) So either the IO is being done under our own
985 * control and this is a bug, or it's a third party IO such as
986 * dump(8) (which may leave the buffer scheduled for read ---
987 * ie. locked but not dirty) or tune2fs (which may actually have
988 * the buffer dirtied, ugh.) */
989
990 if (buffer_dirty(bh) && jh->b_transaction) {
991 warn_dirty_buffer(bh);
992 /*
993 * We need to clean the dirty flag and we must do it under the
994 * buffer lock to be sure we don't race with running write-out.
995 */
996 JBUFFER_TRACE(jh, "Journalling dirty buffer");
997 clear_buffer_dirty(bh);
998 /*
999 * The buffer is going to be added to BJ_Reserved list now and
1000 * nothing guarantees jbd2_journal_dirty_metadata() will be
1001 * ever called for it. So we need to set jbddirty bit here to
1002 * make sure the buffer is dirtied and written out when the
1003 * journaling machinery is done with it.
1004 */
1005 set_buffer_jbddirty(bh);
1006 }
1007
1008 error = -EROFS;
1009 if (is_handle_aborted(handle)) {
1010 spin_unlock(lock: &jh->b_state_lock);
1011 unlock_buffer(bh);
1012 goto out;
1013 }
1014 error = 0;
1015
1016 /*
1017 * The buffer is already part of this transaction if b_transaction or
1018 * b_next_transaction points to it
1019 */
1020 if (jh->b_transaction == transaction ||
1021 jh->b_next_transaction == transaction) {
1022 unlock_buffer(bh);
1023 goto done;
1024 }
1025
1026 /*
1027 * this is the first time this transaction is touching this buffer,
1028 * reset the modified flag
1029 */
1030 jh->b_modified = 0;
1031
1032 /*
1033 * If the buffer is not journaled right now, we need to make sure it
1034 * doesn't get written to disk before the caller actually commits the
1035 * new data
1036 */
1037 if (!jh->b_transaction) {
1038 JBUFFER_TRACE(jh, "no transaction");
1039 J_ASSERT_JH(jh, !jh->b_next_transaction);
1040 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1041 /*
1042 * Make sure all stores to jh (b_modified, b_frozen_data) are
1043 * visible before attaching it to the running transaction.
1044 * Paired with barrier in jbd2_write_access_granted()
1045 */
1046 smp_wmb();
1047 spin_lock(lock: &journal->j_list_lock);
1048 if (test_clear_buffer_dirty(bh)) {
1049 /*
1050 * Execute buffer dirty clearing and jh->b_transaction
1051 * assignment under journal->j_list_lock locked to
1052 * prevent bh being removed from checkpoint list if
1053 * the buffer is in an intermediate state (not dirty
1054 * and jh->b_transaction is NULL).
1055 */
1056 JBUFFER_TRACE(jh, "Journalling dirty buffer");
1057 set_buffer_jbddirty(bh);
1058 }
1059 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1060 spin_unlock(lock: &journal->j_list_lock);
1061 unlock_buffer(bh);
1062 goto done;
1063 }
1064 unlock_buffer(bh);
1065
1066 /*
1067 * If there is already a copy-out version of this buffer, then we don't
1068 * need to make another one
1069 */
1070 if (jh->b_frozen_data) {
1071 JBUFFER_TRACE(jh, "has frozen data");
1072 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1073 goto attach_next;
1074 }
1075
1076 JBUFFER_TRACE(jh, "owned by older transaction");
1077 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1078 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
1079
1080 /*
1081 * There is one case we have to be very careful about. If the
1082 * committing transaction is currently writing this buffer out to disk
1083 * and has NOT made a copy-out, then we cannot modify the buffer
1084 * contents at all right now. The essence of copy-out is that it is
1085 * the extra copy, not the primary copy, which gets journaled. If the
1086 * primary copy is already going to disk then we cannot do copy-out
1087 * here.
1088 */
1089 if (buffer_shadow(bh)) {
1090 JBUFFER_TRACE(jh, "on shadow: sleep");
1091 spin_unlock(lock: &jh->b_state_lock);
1092 wait_on_bit_io(word: &bh->b_state, bit: BH_Shadow, TASK_UNINTERRUPTIBLE);
1093 goto repeat;
1094 }
1095
1096 /*
1097 * Only do the copy if the currently-owning transaction still needs it.
1098 * If buffer isn't on BJ_Metadata list, the committing transaction is
1099 * past that stage (here we use the fact that BH_Shadow is set under
1100 * bh_state lock together with refiling to BJ_Shadow list and at this
1101 * point we know the buffer doesn't have BH_Shadow set).
1102 *
1103 * Subtle point, though: if this is a get_undo_access, then we will be
1104 * relying on the frozen_data to contain the new value of the
1105 * committed_data record after the transaction, so we HAVE to force the
1106 * frozen_data copy in that case.
1107 */
1108 if (jh->b_jlist == BJ_Metadata || force_copy) {
1109 JBUFFER_TRACE(jh, "generate frozen data");
1110 if (!frozen_buffer) {
1111 JBUFFER_TRACE(jh, "allocate memory for buffer");
1112 spin_unlock(lock: &jh->b_state_lock);
1113 frozen_buffer = jbd2_alloc(size: jh2bh(jh)->b_size,
1114 GFP_NOFS | __GFP_NOFAIL);
1115 goto repeat;
1116 }
1117 jh->b_frozen_data = frozen_buffer;
1118 frozen_buffer = NULL;
1119 jbd2_freeze_jh_data(jh);
1120 }
1121attach_next:
1122 /*
1123 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1124 * before attaching it to the running transaction. Paired with barrier
1125 * in jbd2_write_access_granted()
1126 */
1127 smp_wmb();
1128 jh->b_next_transaction = transaction;
1129
1130done:
1131 spin_unlock(lock: &jh->b_state_lock);
1132
1133 /*
1134 * If we are about to journal a buffer, then any revoke pending on it is
1135 * no longer valid
1136 */
1137 jbd2_journal_cancel_revoke(handle, jh);
1138
1139out:
1140 if (unlikely(frozen_buffer)) /* It's usually NULL */
1141 jbd2_free(ptr: frozen_buffer, size: bh->b_size);
1142
1143 JBUFFER_TRACE(jh, "exit");
1144 return error;
1145}
1146
1147/* Fast check whether buffer is already attached to the required transaction */
1148static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1149 bool undo)
1150{
1151 struct journal_head *jh;
1152 bool ret = false;
1153
1154 /* Dirty buffers require special handling... */
1155 if (buffer_dirty(bh))
1156 return false;
1157
1158 /*
1159 * RCU protects us from dereferencing freed pages. So the checks we do
1160 * are guaranteed not to oops. However the jh slab object can get freed
1161 * & reallocated while we work with it. So we have to be careful. When
1162 * we see jh attached to the running transaction, we know it must stay
1163 * so until the transaction is committed. Thus jh won't be freed and
1164 * will be attached to the same bh while we run. However it can
1165 * happen jh gets freed, reallocated, and attached to the transaction
1166 * just after we get pointer to it from bh. So we have to be careful
1167 * and recheck jh still belongs to our bh before we return success.
1168 */
1169 rcu_read_lock();
1170 if (!buffer_jbd(bh))
1171 goto out;
1172 /* This should be bh2jh() but that doesn't work with inline functions */
1173 jh = READ_ONCE(bh->b_private);
1174 if (!jh)
1175 goto out;
1176 /* For undo access buffer must have data copied */
1177 if (undo && !jh->b_committed_data)
1178 goto out;
1179 if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1180 READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1181 goto out;
1182 /*
1183 * There are two reasons for the barrier here:
1184 * 1) Make sure to fetch b_bh after we did previous checks so that we
1185 * detect when jh went through free, realloc, attach to transaction
1186 * while we were checking. Paired with implicit barrier in that path.
1187 * 2) So that access to bh done after jbd2_write_access_granted()
1188 * doesn't get reordered and see inconsistent state of concurrent
1189 * do_get_write_access().
1190 */
1191 smp_mb();
1192 if (unlikely(jh->b_bh != bh))
1193 goto out;
1194 ret = true;
1195out:
1196 rcu_read_unlock();
1197 return ret;
1198}
1199
1200/**
1201 * jbd2_journal_get_write_access() - notify intent to modify a buffer
1202 * for metadata (not data) update.
1203 * @handle: transaction to add buffer modifications to
1204 * @bh: bh to be used for metadata writes
1205 *
1206 * Returns: error code or 0 on success.
1207 *
1208 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1209 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1210 */
1211
1212int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1213{
1214 struct journal_head *jh;
1215 journal_t *journal;
1216 int rc;
1217
1218 if (is_handle_aborted(handle))
1219 return -EROFS;
1220
1221 journal = handle->h_transaction->t_journal;
1222 if (jbd2_check_fs_dev_write_error(journal)) {
1223 /*
1224 * If the fs dev has writeback errors, it may have failed
1225 * to async write out metadata buffers in the background.
1226 * In this case, we could read old data from disk and write
1227 * it out again, which may lead to on-disk filesystem
1228 * inconsistency. Aborting journal can avoid it happen.
1229 */
1230 jbd2_journal_abort(journal, -EIO);
1231 return -EIO;
1232 }
1233
1234 if (jbd2_write_access_granted(handle, bh, undo: false))
1235 return 0;
1236
1237 jh = jbd2_journal_add_journal_head(bh);
1238 /* We do not want to get caught playing with fields which the
1239 * log thread also manipulates. Make sure that the buffer
1240 * completes any outstanding IO before proceeding. */
1241 rc = do_get_write_access(handle, jh, force_copy: 0);
1242 jbd2_journal_put_journal_head(jh);
1243 return rc;
1244}
1245
1246
1247/*
1248 * When the user wants to journal a newly created buffer_head
1249 * (ie. getblk() returned a new buffer and we are going to populate it
1250 * manually rather than reading off disk), then we need to keep the
1251 * buffer_head locked until it has been completely filled with new
1252 * data. In this case, we should be able to make the assertion that
1253 * the bh is not already part of an existing transaction.
1254 *
1255 * The buffer should already be locked by the caller by this point.
1256 * There is no lock ranking violation: it was a newly created,
1257 * unlocked buffer beforehand. */
1258
1259/**
1260 * jbd2_journal_get_create_access () - notify intent to use newly created bh
1261 * @handle: transaction to new buffer to
1262 * @bh: new buffer.
1263 *
1264 * Call this if you create a new bh.
1265 */
1266int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1267{
1268 transaction_t *transaction = handle->h_transaction;
1269 journal_t *journal;
1270 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1271 int err;
1272
1273 jbd2_debug(5, "journal_head %p\n", jh);
1274 err = -EROFS;
1275 if (is_handle_aborted(handle))
1276 goto out;
1277 journal = transaction->t_journal;
1278 err = 0;
1279
1280 JBUFFER_TRACE(jh, "entry");
1281 /*
1282 * The buffer may already belong to this transaction due to pre-zeroing
1283 * in the filesystem's new_block code. It may also be on the previous,
1284 * committing transaction's lists, but it HAS to be in Forget state in
1285 * that case: the transaction must have deleted the buffer for it to be
1286 * reused here.
1287 */
1288 spin_lock(lock: &jh->b_state_lock);
1289 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1290 jh->b_transaction == NULL ||
1291 (jh->b_transaction == journal->j_committing_transaction &&
1292 jh->b_jlist == BJ_Forget)));
1293
1294 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1295 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1296
1297 if (jh->b_transaction == NULL) {
1298 /*
1299 * Previous jbd2_journal_forget() could have left the buffer
1300 * with jbddirty bit set because it was being committed. When
1301 * the commit finished, we've filed the buffer for
1302 * checkpointing and marked it dirty. Now we are reallocating
1303 * the buffer so the transaction freeing it must have
1304 * committed and so it's safe to clear the dirty bit.
1305 */
1306 clear_buffer_dirty(bh: jh2bh(jh));
1307 /* first access by this transaction */
1308 jh->b_modified = 0;
1309
1310 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1311 spin_lock(lock: &journal->j_list_lock);
1312 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1313 spin_unlock(lock: &journal->j_list_lock);
1314 } else if (jh->b_transaction == journal->j_committing_transaction) {
1315 /* first access by this transaction */
1316 jh->b_modified = 0;
1317
1318 JBUFFER_TRACE(jh, "set next transaction");
1319 spin_lock(lock: &journal->j_list_lock);
1320 jh->b_next_transaction = transaction;
1321 spin_unlock(lock: &journal->j_list_lock);
1322 }
1323 spin_unlock(lock: &jh->b_state_lock);
1324
1325 /*
1326 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1327 * blocks which contain freed but then revoked metadata. We need
1328 * to cancel the revoke in case we end up freeing it yet again
1329 * and the reallocating as data - this would cause a second revoke,
1330 * which hits an assertion error.
1331 */
1332 JBUFFER_TRACE(jh, "cancelling revoke");
1333 jbd2_journal_cancel_revoke(handle, jh);
1334out:
1335 jbd2_journal_put_journal_head(jh);
1336 return err;
1337}
1338
1339/**
1340 * jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1341 * non-rewindable consequences
1342 * @handle: transaction
1343 * @bh: buffer to undo
1344 *
1345 * Sometimes there is a need to distinguish between metadata which has
1346 * been committed to disk and that which has not. The ext3fs code uses
1347 * this for freeing and allocating space, we have to make sure that we
1348 * do not reuse freed space until the deallocation has been committed,
1349 * since if we overwrote that space we would make the delete
1350 * un-rewindable in case of a crash.
1351 *
1352 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1353 * buffer for parts of non-rewindable operations such as delete
1354 * operations on the bitmaps. The journaling code must keep a copy of
1355 * the buffer's contents prior to the undo_access call until such time
1356 * as we know that the buffer has definitely been committed to disk.
1357 *
1358 * We never need to know which transaction the committed data is part
1359 * of, buffers touched here are guaranteed to be dirtied later and so
1360 * will be committed to a new transaction in due course, at which point
1361 * we can discard the old committed data pointer.
1362 *
1363 * Returns error number or 0 on success.
1364 */
1365int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1366{
1367 int err;
1368 struct journal_head *jh;
1369 char *committed_data = NULL;
1370
1371 if (is_handle_aborted(handle))
1372 return -EROFS;
1373
1374 if (jbd2_write_access_granted(handle, bh, undo: true))
1375 return 0;
1376
1377 jh = jbd2_journal_add_journal_head(bh);
1378 JBUFFER_TRACE(jh, "entry");
1379
1380 /*
1381 * Do this first --- it can drop the journal lock, so we want to
1382 * make sure that obtaining the committed_data is done
1383 * atomically wrt. completion of any outstanding commits.
1384 */
1385 err = do_get_write_access(handle, jh, force_copy: 1);
1386 if (err)
1387 goto out;
1388
1389repeat:
1390 if (!jh->b_committed_data)
1391 committed_data = jbd2_alloc(size: jh2bh(jh)->b_size,
1392 GFP_NOFS|__GFP_NOFAIL);
1393
1394 spin_lock(lock: &jh->b_state_lock);
1395 if (!jh->b_committed_data) {
1396 /* Copy out the current buffer contents into the
1397 * preserved, committed copy. */
1398 JBUFFER_TRACE(jh, "generate b_committed data");
1399 if (!committed_data) {
1400 spin_unlock(lock: &jh->b_state_lock);
1401 goto repeat;
1402 }
1403
1404 jh->b_committed_data = committed_data;
1405 committed_data = NULL;
1406 memcpy(to: jh->b_committed_data, from: bh->b_data, len: bh->b_size);
1407 }
1408 spin_unlock(lock: &jh->b_state_lock);
1409out:
1410 jbd2_journal_put_journal_head(jh);
1411 if (unlikely(committed_data))
1412 jbd2_free(ptr: committed_data, size: bh->b_size);
1413 return err;
1414}
1415
1416/**
1417 * jbd2_journal_set_triggers() - Add triggers for commit writeout
1418 * @bh: buffer to trigger on
1419 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1420 *
1421 * Set any triggers on this journal_head. This is always safe, because
1422 * triggers for a committing buffer will be saved off, and triggers for
1423 * a running transaction will match the buffer in that transaction.
1424 *
1425 * Call with NULL to clear the triggers.
1426 */
1427void jbd2_journal_set_triggers(struct buffer_head *bh,
1428 struct jbd2_buffer_trigger_type *type)
1429{
1430 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1431
1432 if (WARN_ON_ONCE(!jh))
1433 return;
1434 jh->b_triggers = type;
1435 jbd2_journal_put_journal_head(jh);
1436}
1437
1438void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1439 struct jbd2_buffer_trigger_type *triggers)
1440{
1441 struct buffer_head *bh = jh2bh(jh);
1442
1443 if (!triggers || !triggers->t_frozen)
1444 return;
1445
1446 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1447}
1448
1449void jbd2_buffer_abort_trigger(struct journal_head *jh,
1450 struct jbd2_buffer_trigger_type *triggers)
1451{
1452 if (!triggers || !triggers->t_abort)
1453 return;
1454
1455 triggers->t_abort(triggers, jh2bh(jh));
1456}
1457
1458/**
1459 * jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1460 * @handle: transaction to add buffer to.
1461 * @bh: buffer to mark
1462 *
1463 * mark dirty metadata which needs to be journaled as part of the current
1464 * transaction.
1465 *
1466 * The buffer must have previously had jbd2_journal_get_write_access()
1467 * called so that it has a valid journal_head attached to the buffer
1468 * head.
1469 *
1470 * The buffer is placed on the transaction's metadata list and is marked
1471 * as belonging to the transaction.
1472 *
1473 * Returns error number or 0 on success.
1474 *
1475 * Special care needs to be taken if the buffer already belongs to the
1476 * current committing transaction (in which case we should have frozen
1477 * data present for that commit). In that case, we don't relink the
1478 * buffer: that only gets done when the old transaction finally
1479 * completes its commit.
1480 */
1481int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1482{
1483 transaction_t *transaction = handle->h_transaction;
1484 journal_t *journal;
1485 struct journal_head *jh;
1486 int ret = 0;
1487
1488 if (!buffer_jbd(bh))
1489 return -EUCLEAN;
1490
1491 /*
1492 * We don't grab jh reference here since the buffer must be part
1493 * of the running transaction.
1494 */
1495 jh = bh2jh(bh);
1496 jbd2_debug(5, "journal_head %p\n", jh);
1497 JBUFFER_TRACE(jh, "entry");
1498
1499 /*
1500 * This and the following assertions are unreliable since we may see jh
1501 * in inconsistent state unless we grab bh_state lock. But this is
1502 * crucial to catch bugs so let's do a reliable check until the
1503 * lockless handling is fully proven.
1504 */
1505 if (data_race(jh->b_transaction != transaction &&
1506 jh->b_next_transaction != transaction)) {
1507 spin_lock(lock: &jh->b_state_lock);
1508 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1509 jh->b_next_transaction == transaction);
1510 spin_unlock(lock: &jh->b_state_lock);
1511 }
1512 if (data_race(jh->b_modified == 1)) {
1513 /* If it's in our transaction it must be in BJ_Metadata list. */
1514 if (data_race(jh->b_transaction == transaction &&
1515 jh->b_jlist != BJ_Metadata)) {
1516 spin_lock(lock: &jh->b_state_lock);
1517 if (jh->b_transaction == transaction &&
1518 jh->b_jlist != BJ_Metadata)
1519 pr_err("JBD2: assertion failure: h_type=%u "
1520 "h_line_no=%u block_no=%llu jlist=%u\n",
1521 handle->h_type, handle->h_line_no,
1522 (unsigned long long) bh->b_blocknr,
1523 jh->b_jlist);
1524 J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1525 jh->b_jlist == BJ_Metadata);
1526 spin_unlock(lock: &jh->b_state_lock);
1527 }
1528 goto out;
1529 }
1530
1531 spin_lock(lock: &jh->b_state_lock);
1532
1533 if (is_handle_aborted(handle)) {
1534 /*
1535 * Check journal aborting with @jh->b_state_lock locked,
1536 * since 'jh->b_transaction' could be replaced with
1537 * 'jh->b_next_transaction' during old transaction
1538 * committing if journal aborted, which may fail
1539 * assertion on 'jh->b_frozen_data == NULL'.
1540 */
1541 ret = -EROFS;
1542 goto out_unlock_bh;
1543 }
1544
1545 journal = transaction->t_journal;
1546
1547 if (jh->b_modified == 0) {
1548 /*
1549 * This buffer's got modified and becoming part
1550 * of the transaction. This needs to be done
1551 * once a transaction -bzzz
1552 */
1553 if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) {
1554 ret = -ENOSPC;
1555 goto out_unlock_bh;
1556 }
1557 jh->b_modified = 1;
1558 handle->h_total_credits--;
1559 }
1560
1561 /*
1562 * fastpath, to avoid expensive locking. If this buffer is already
1563 * on the running transaction's metadata list there is nothing to do.
1564 * Nobody can take it off again because there is a handle open.
1565 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1566 * result in this test being false, so we go in and take the locks.
1567 */
1568 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1569 JBUFFER_TRACE(jh, "fastpath");
1570 if (unlikely(jh->b_transaction !=
1571 journal->j_running_transaction)) {
1572 printk(KERN_ERR "JBD2: %s: "
1573 "jh->b_transaction (%llu, %p, %u) != "
1574 "journal->j_running_transaction (%p, %u)\n",
1575 journal->j_devname,
1576 (unsigned long long) bh->b_blocknr,
1577 jh->b_transaction,
1578 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1579 journal->j_running_transaction,
1580 journal->j_running_transaction ?
1581 journal->j_running_transaction->t_tid : 0);
1582 ret = -EINVAL;
1583 }
1584 goto out_unlock_bh;
1585 }
1586
1587 set_buffer_jbddirty(bh);
1588
1589 /*
1590 * Metadata already on the current transaction list doesn't
1591 * need to be filed. Metadata on another transaction's list must
1592 * be committing, and will be refiled once the commit completes:
1593 * leave it alone for now.
1594 */
1595 if (jh->b_transaction != transaction) {
1596 JBUFFER_TRACE(jh, "already on other transaction");
1597 if (unlikely(((jh->b_transaction !=
1598 journal->j_committing_transaction)) ||
1599 (jh->b_next_transaction != transaction))) {
1600 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1601 "bad jh for block %llu: "
1602 "transaction (%p, %u), "
1603 "jh->b_transaction (%p, %u), "
1604 "jh->b_next_transaction (%p, %u), jlist %u\n",
1605 journal->j_devname,
1606 (unsigned long long) bh->b_blocknr,
1607 transaction, transaction->t_tid,
1608 jh->b_transaction,
1609 jh->b_transaction ?
1610 jh->b_transaction->t_tid : 0,
1611 jh->b_next_transaction,
1612 jh->b_next_transaction ?
1613 jh->b_next_transaction->t_tid : 0,
1614 jh->b_jlist);
1615 WARN_ON(1);
1616 ret = -EINVAL;
1617 }
1618 /* And this case is illegal: we can't reuse another
1619 * transaction's data buffer, ever. */
1620 goto out_unlock_bh;
1621 }
1622
1623 /* That test should have eliminated the following case: */
1624 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1625
1626 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1627 spin_lock(lock: &journal->j_list_lock);
1628 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1629 spin_unlock(lock: &journal->j_list_lock);
1630out_unlock_bh:
1631 spin_unlock(lock: &jh->b_state_lock);
1632out:
1633 JBUFFER_TRACE(jh, "exit");
1634 return ret;
1635}
1636
1637/**
1638 * jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1639 * @handle: transaction handle
1640 * @bh: bh to 'forget'
1641 *
1642 * We can only do the bforget if there are no commits pending against the
1643 * buffer. If the buffer is dirty in the current running transaction we
1644 * can safely unlink it.
1645 *
1646 * bh may not be a journalled buffer at all - it may be a non-JBD
1647 * buffer which came off the hashtable. Check for this.
1648 *
1649 * Decrements bh->b_count by one.
1650 *
1651 * Allow this call even if the handle has aborted --- it may be part of
1652 * the caller's cleanup after an abort.
1653 */
1654int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh)
1655{
1656 transaction_t *transaction = handle->h_transaction;
1657 journal_t *journal;
1658 struct journal_head *jh;
1659 int drop_reserve = 0;
1660 int err = 0;
1661 int was_modified = 0;
1662
1663 if (is_handle_aborted(handle))
1664 return -EROFS;
1665 journal = transaction->t_journal;
1666
1667 BUFFER_TRACE(bh, "entry");
1668
1669 jh = jbd2_journal_grab_journal_head(bh);
1670 if (!jh) {
1671 __bforget(bh);
1672 return 0;
1673 }
1674
1675 spin_lock(lock: &jh->b_state_lock);
1676
1677 /* Critical error: attempting to delete a bitmap buffer, maybe?
1678 * Don't do any jbd operations, and return an error. */
1679 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1680 "inconsistent data on disk")) {
1681 err = -EIO;
1682 goto drop;
1683 }
1684
1685 /* keep track of whether or not this transaction modified us */
1686 was_modified = jh->b_modified;
1687
1688 /*
1689 * The buffer's going from the transaction, we must drop
1690 * all references -bzzz
1691 */
1692 jh->b_modified = 0;
1693
1694 if (jh->b_transaction == transaction) {
1695 J_ASSERT_JH(jh, !jh->b_frozen_data);
1696
1697 /* If we are forgetting a buffer which is already part
1698 * of this transaction, then we can just drop it from
1699 * the transaction immediately. */
1700 clear_buffer_dirty(bh);
1701 clear_buffer_jbddirty(bh);
1702
1703 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1704
1705 /*
1706 * we only want to drop a reference if this transaction
1707 * modified the buffer
1708 */
1709 if (was_modified)
1710 drop_reserve = 1;
1711
1712 /*
1713 * We are no longer going to journal this buffer.
1714 * However, the commit of this transaction is still
1715 * important to the buffer: the delete that we are now
1716 * processing might obsolete an old log entry, so by
1717 * committing, we can satisfy the buffer's checkpoint.
1718 *
1719 * So, if we have a checkpoint on the buffer, we should
1720 * now refile the buffer on our BJ_Forget list so that
1721 * we know to remove the checkpoint after we commit.
1722 */
1723
1724 spin_lock(lock: &journal->j_list_lock);
1725 if (jh->b_cp_transaction) {
1726 __jbd2_journal_temp_unlink_buffer(jh);
1727 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1728 } else {
1729 __jbd2_journal_unfile_buffer(jh);
1730 jbd2_journal_put_journal_head(jh);
1731 }
1732 spin_unlock(lock: &journal->j_list_lock);
1733 } else if (jh->b_transaction) {
1734 J_ASSERT_JH(jh, (jh->b_transaction ==
1735 journal->j_committing_transaction));
1736 /* However, if the buffer is still owned by a prior
1737 * (committing) transaction, we can't drop it yet... */
1738 JBUFFER_TRACE(jh, "belongs to older transaction");
1739 /* ... but we CAN drop it from the new transaction through
1740 * marking the buffer as freed and set j_next_transaction to
1741 * the new transaction, so that not only the commit code
1742 * knows it should clear dirty bits when it is done with the
1743 * buffer, but also the buffer can be checkpointed only
1744 * after the new transaction commits. */
1745
1746 set_buffer_freed(bh);
1747
1748 if (!jh->b_next_transaction) {
1749 spin_lock(lock: &journal->j_list_lock);
1750 jh->b_next_transaction = transaction;
1751 spin_unlock(lock: &journal->j_list_lock);
1752 } else {
1753 J_ASSERT(jh->b_next_transaction == transaction);
1754
1755 /*
1756 * only drop a reference if this transaction modified
1757 * the buffer
1758 */
1759 if (was_modified)
1760 drop_reserve = 1;
1761 }
1762 } else {
1763 /*
1764 * Finally, if the buffer is not belongs to any
1765 * transaction, we can just drop it now if it has no
1766 * checkpoint.
1767 */
1768 spin_lock(lock: &journal->j_list_lock);
1769 if (!jh->b_cp_transaction) {
1770 JBUFFER_TRACE(jh, "belongs to none transaction");
1771 spin_unlock(lock: &journal->j_list_lock);
1772 goto drop;
1773 }
1774
1775 /*
1776 * Otherwise, if the buffer has been written to disk,
1777 * it is safe to remove the checkpoint and drop it.
1778 */
1779 if (jbd2_journal_try_remove_checkpoint(jh) >= 0) {
1780 spin_unlock(lock: &journal->j_list_lock);
1781 goto drop;
1782 }
1783
1784 /*
1785 * The buffer is still not written to disk, we should
1786 * attach this buffer to current transaction so that the
1787 * buffer can be checkpointed only after the current
1788 * transaction commits.
1789 */
1790 clear_buffer_dirty(bh);
1791 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1792 spin_unlock(lock: &journal->j_list_lock);
1793 }
1794drop:
1795 __brelse(bh);
1796 spin_unlock(lock: &jh->b_state_lock);
1797 jbd2_journal_put_journal_head(jh);
1798 if (drop_reserve) {
1799 /* no need to reserve log space for this block -bzzz */
1800 handle->h_total_credits++;
1801 }
1802 return err;
1803}
1804
1805/**
1806 * jbd2_journal_stop() - complete a transaction
1807 * @handle: transaction to complete.
1808 *
1809 * All done for a particular handle.
1810 *
1811 * There is not much action needed here. We just return any remaining
1812 * buffer credits to the transaction and remove the handle. The only
1813 * complication is that we need to start a commit operation if the
1814 * filesystem is marked for synchronous update.
1815 *
1816 * jbd2_journal_stop itself will not usually return an error, but it may
1817 * do so in unusual circumstances. In particular, expect it to
1818 * return -EIO if a jbd2_journal_abort has been executed since the
1819 * transaction began.
1820 */
1821int jbd2_journal_stop(handle_t *handle)
1822{
1823 transaction_t *transaction = handle->h_transaction;
1824 journal_t *journal;
1825 int err = 0, wait_for_commit = 0;
1826 tid_t tid;
1827 pid_t pid;
1828
1829 if (--handle->h_ref > 0) {
1830 jbd2_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1831 handle->h_ref);
1832 if (is_handle_aborted(handle))
1833 return -EIO;
1834 return 0;
1835 }
1836 if (!transaction) {
1837 /*
1838 * Handle is already detached from the transaction so there is
1839 * nothing to do other than free the handle.
1840 */
1841 memalloc_nofs_restore(flags: handle->saved_alloc_context);
1842 goto free_and_exit;
1843 }
1844 journal = transaction->t_journal;
1845 tid = transaction->t_tid;
1846
1847 if (is_handle_aborted(handle))
1848 err = -EIO;
1849
1850 jbd2_debug(4, "Handle %p going down\n", handle);
1851 trace_jbd2_handle_stats(dev: journal->j_fs_dev->bd_dev,
1852 tid, type: handle->h_type, line_no: handle->h_line_no,
1853 interval: jiffies - handle->h_start_jiffies,
1854 sync: handle->h_sync, requested_blocks: handle->h_requested_credits,
1855 dirtied_blocks: (handle->h_requested_credits -
1856 handle->h_total_credits));
1857
1858 /*
1859 * Implement synchronous transaction batching. If the handle
1860 * was synchronous, don't force a commit immediately. Let's
1861 * yield and let another thread piggyback onto this
1862 * transaction. Keep doing that while new threads continue to
1863 * arrive. It doesn't cost much - we're about to run a commit
1864 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1865 * operations by 30x or more...
1866 *
1867 * We try and optimize the sleep time against what the
1868 * underlying disk can do, instead of having a static sleep
1869 * time. This is useful for the case where our storage is so
1870 * fast that it is more optimal to go ahead and force a flush
1871 * and wait for the transaction to be committed than it is to
1872 * wait for an arbitrary amount of time for new writers to
1873 * join the transaction. We achieve this by measuring how
1874 * long it takes to commit a transaction, and compare it with
1875 * how long this transaction has been running, and if run time
1876 * < commit time then we sleep for the delta and commit. This
1877 * greatly helps super fast disks that would see slowdowns as
1878 * more threads started doing fsyncs.
1879 *
1880 * But don't do this if this process was the most recent one
1881 * to perform a synchronous write. We do this to detect the
1882 * case where a single process is doing a stream of sync
1883 * writes. No point in waiting for joiners in that case.
1884 *
1885 * Setting max_batch_time to 0 disables this completely.
1886 */
1887 pid = current->pid;
1888 if (handle->h_sync && journal->j_last_sync_writer != pid &&
1889 journal->j_max_batch_time) {
1890 u64 commit_time, trans_time;
1891
1892 journal->j_last_sync_writer = pid;
1893
1894 read_lock(&journal->j_state_lock);
1895 commit_time = journal->j_average_commit_time;
1896 read_unlock(&journal->j_state_lock);
1897
1898 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1899 transaction->t_start_time));
1900
1901 commit_time = max_t(u64, commit_time,
1902 1000*journal->j_min_batch_time);
1903 commit_time = min_t(u64, commit_time,
1904 1000*journal->j_max_batch_time);
1905
1906 if (trans_time < commit_time) {
1907 ktime_t expires = ktime_add_ns(ktime_get(),
1908 commit_time);
1909 set_current_state(TASK_UNINTERRUPTIBLE);
1910 schedule_hrtimeout(expires: &expires, mode: HRTIMER_MODE_ABS);
1911 }
1912 }
1913
1914 if (handle->h_sync)
1915 transaction->t_synchronous_commit = 1;
1916
1917 /*
1918 * If the handle is marked SYNC, we need to set another commit
1919 * going! We also want to force a commit if the transaction is too
1920 * old now.
1921 */
1922 if (handle->h_sync ||
1923 time_after_eq(jiffies, transaction->t_expires)) {
1924 /* Do this even for aborted journals: an abort still
1925 * completes the commit thread, it just doesn't write
1926 * anything to disk. */
1927
1928 jbd2_debug(2, "transaction too old, requesting commit for "
1929 "handle %p\n", handle);
1930 /* This is non-blocking */
1931 jbd2_log_start_commit(journal, tid);
1932
1933 /*
1934 * Special case: JBD2_SYNC synchronous updates require us
1935 * to wait for the commit to complete.
1936 */
1937 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1938 wait_for_commit = 1;
1939 }
1940
1941 /*
1942 * Once stop_this_handle() drops t_updates, the transaction could start
1943 * committing on us and eventually disappear. So we must not
1944 * dereference transaction pointer again after calling
1945 * stop_this_handle().
1946 */
1947 stop_this_handle(handle);
1948
1949 if (wait_for_commit)
1950 err = jbd2_log_wait_commit(journal, tid);
1951
1952free_and_exit:
1953 if (handle->h_rsv_handle)
1954 jbd2_free_handle(handle: handle->h_rsv_handle);
1955 jbd2_free_handle(handle);
1956 return err;
1957}
1958
1959/*
1960 *
1961 * List management code snippets: various functions for manipulating the
1962 * transaction buffer lists.
1963 *
1964 */
1965
1966/*
1967 * Append a buffer to a transaction list, given the transaction's list head
1968 * pointer.
1969 *
1970 * j_list_lock is held.
1971 *
1972 * jh->b_state_lock is held.
1973 */
1974
1975static inline void
1976__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1977{
1978 if (!*list) {
1979 jh->b_tnext = jh->b_tprev = jh;
1980 *list = jh;
1981 } else {
1982 /* Insert at the tail of the list to preserve order */
1983 struct journal_head *first = *list, *last = first->b_tprev;
1984 jh->b_tprev = last;
1985 jh->b_tnext = first;
1986 last->b_tnext = first->b_tprev = jh;
1987 }
1988}
1989
1990/*
1991 * Remove a buffer from a transaction list, given the transaction's list
1992 * head pointer.
1993 *
1994 * Called with j_list_lock held, and the journal may not be locked.
1995 *
1996 * jh->b_state_lock is held.
1997 */
1998
1999static inline void
2000__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
2001{
2002 if (*list == jh) {
2003 *list = jh->b_tnext;
2004 if (*list == jh)
2005 *list = NULL;
2006 }
2007 jh->b_tprev->b_tnext = jh->b_tnext;
2008 jh->b_tnext->b_tprev = jh->b_tprev;
2009}
2010
2011/*
2012 * Remove a buffer from the appropriate transaction list.
2013 *
2014 * Note that this function can *change* the value of
2015 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
2016 * t_reserved_list. If the caller is holding onto a copy of one of these
2017 * pointers, it could go bad. Generally the caller needs to re-read the
2018 * pointer from the transaction_t.
2019 *
2020 * Called under j_list_lock.
2021 */
2022static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
2023{
2024 struct journal_head **list = NULL;
2025 transaction_t *transaction;
2026 struct buffer_head *bh = jh2bh(jh);
2027
2028 lockdep_assert_held(&jh->b_state_lock);
2029 transaction = jh->b_transaction;
2030 if (transaction)
2031 assert_spin_locked(&transaction->t_journal->j_list_lock);
2032
2033 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2034 if (jh->b_jlist != BJ_None)
2035 J_ASSERT_JH(jh, transaction != NULL);
2036
2037 switch (jh->b_jlist) {
2038 case BJ_None:
2039 return;
2040 case BJ_Metadata:
2041 transaction->t_nr_buffers--;
2042 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
2043 list = &transaction->t_buffers;
2044 break;
2045 case BJ_Forget:
2046 list = &transaction->t_forget;
2047 break;
2048 case BJ_Shadow:
2049 list = &transaction->t_shadow_list;
2050 break;
2051 case BJ_Reserved:
2052 list = &transaction->t_reserved_list;
2053 break;
2054 }
2055
2056 __blist_del_buffer(list, jh);
2057 jh->b_jlist = BJ_None;
2058 if (transaction && is_journal_aborted(journal: transaction->t_journal))
2059 clear_buffer_jbddirty(bh);
2060 else if (test_clear_buffer_jbddirty(bh))
2061 mark_buffer_dirty(bh); /* Expose it to the VM */
2062}
2063
2064/*
2065 * Remove buffer from all transactions. The caller is responsible for dropping
2066 * the jh reference that belonged to the transaction.
2067 *
2068 * Called with bh_state lock and j_list_lock
2069 */
2070static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
2071{
2072 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2073 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2074
2075 __jbd2_journal_temp_unlink_buffer(jh);
2076 jh->b_transaction = NULL;
2077}
2078
2079/**
2080 * jbd2_journal_try_to_free_buffers() - try to free page buffers.
2081 * @journal: journal for operation
2082 * @folio: Folio to detach data from.
2083 *
2084 * For all the buffers on this page,
2085 * if they are fully written out ordered data, move them onto BUF_CLEAN
2086 * so try_to_free_buffers() can reap them.
2087 *
2088 * This function returns non-zero if we wish try_to_free_buffers()
2089 * to be called. We do this if the page is releasable by try_to_free_buffers().
2090 * We also do it if the page has locked or dirty buffers and the caller wants
2091 * us to perform sync or async writeout.
2092 *
2093 * This complicates JBD locking somewhat. We aren't protected by the
2094 * BKL here. We wish to remove the buffer from its committing or
2095 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2096 *
2097 * This may *change* the value of transaction_t->t_datalist, so anyone
2098 * who looks at t_datalist needs to lock against this function.
2099 *
2100 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2101 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
2102 * will come out of the lock with the buffer dirty, which makes it
2103 * ineligible for release here.
2104 *
2105 * Who else is affected by this? hmm... Really the only contender
2106 * is do_get_write_access() - it could be looking at the buffer while
2107 * journal_try_to_free_buffer() is changing its state. But that
2108 * cannot happen because we never reallocate freed data as metadata
2109 * while the data is part of a transaction. Yes?
2110 *
2111 * Return false on failure, true on success
2112 */
2113bool jbd2_journal_try_to_free_buffers(journal_t *journal, struct folio *folio)
2114{
2115 struct buffer_head *head;
2116 struct buffer_head *bh;
2117 bool ret = false;
2118
2119 J_ASSERT(folio_test_locked(folio));
2120
2121 head = folio_buffers(folio);
2122 bh = head;
2123 do {
2124 struct journal_head *jh;
2125
2126 /*
2127 * We take our own ref against the journal_head here to avoid
2128 * having to add tons of locking around each instance of
2129 * jbd2_journal_put_journal_head().
2130 */
2131 jh = jbd2_journal_grab_journal_head(bh);
2132 if (!jh)
2133 continue;
2134
2135 spin_lock(lock: &jh->b_state_lock);
2136 if (!jh->b_transaction && !jh->b_next_transaction) {
2137 spin_lock(lock: &journal->j_list_lock);
2138 /* Remove written-back checkpointed metadata buffer */
2139 if (jh->b_cp_transaction != NULL)
2140 jbd2_journal_try_remove_checkpoint(jh);
2141 spin_unlock(lock: &journal->j_list_lock);
2142 }
2143 spin_unlock(lock: &jh->b_state_lock);
2144 jbd2_journal_put_journal_head(jh);
2145 if (buffer_jbd(bh))
2146 goto busy;
2147 } while ((bh = bh->b_this_page) != head);
2148
2149 ret = try_to_free_buffers(folio);
2150busy:
2151 return ret;
2152}
2153
2154/*
2155 * This buffer is no longer needed. If it is on an older transaction's
2156 * checkpoint list we need to record it on this transaction's forget list
2157 * to pin this buffer (and hence its checkpointing transaction) down until
2158 * this transaction commits. If the buffer isn't on a checkpoint list, we
2159 * release it.
2160 * Returns non-zero if JBD no longer has an interest in the buffer.
2161 *
2162 * Called under j_list_lock.
2163 *
2164 * Called under jh->b_state_lock.
2165 */
2166static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2167{
2168 int may_free = 1;
2169 struct buffer_head *bh = jh2bh(jh);
2170
2171 if (jh->b_cp_transaction) {
2172 JBUFFER_TRACE(jh, "on running+cp transaction");
2173 __jbd2_journal_temp_unlink_buffer(jh);
2174 /*
2175 * We don't want to write the buffer anymore, clear the
2176 * bit so that we don't confuse checks in
2177 * __jbd2_journal_file_buffer
2178 */
2179 clear_buffer_dirty(bh);
2180 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2181 may_free = 0;
2182 } else {
2183 JBUFFER_TRACE(jh, "on running transaction");
2184 __jbd2_journal_unfile_buffer(jh);
2185 jbd2_journal_put_journal_head(jh);
2186 }
2187 return may_free;
2188}
2189
2190/*
2191 * jbd2_journal_invalidate_folio
2192 *
2193 * This code is tricky. It has a number of cases to deal with.
2194 *
2195 * There are two invariants which this code relies on:
2196 *
2197 * i_size must be updated on disk before we start calling invalidate_folio
2198 * on the data.
2199 *
2200 * This is done in ext3 by defining an ext3_setattr method which
2201 * updates i_size before truncate gets going. By maintaining this
2202 * invariant, we can be sure that it is safe to throw away any buffers
2203 * attached to the current transaction: once the transaction commits,
2204 * we know that the data will not be needed.
2205 *
2206 * Note however that we can *not* throw away data belonging to the
2207 * previous, committing transaction!
2208 *
2209 * Any disk blocks which *are* part of the previous, committing
2210 * transaction (and which therefore cannot be discarded immediately) are
2211 * not going to be reused in the new running transaction
2212 *
2213 * The bitmap committed_data images guarantee this: any block which is
2214 * allocated in one transaction and removed in the next will be marked
2215 * as in-use in the committed_data bitmap, so cannot be reused until
2216 * the next transaction to delete the block commits. This means that
2217 * leaving committing buffers dirty is quite safe: the disk blocks
2218 * cannot be reallocated to a different file and so buffer aliasing is
2219 * not possible.
2220 *
2221 *
2222 * The above applies mainly to ordered data mode. In writeback mode we
2223 * don't make guarantees about the order in which data hits disk --- in
2224 * particular we don't guarantee that new dirty data is flushed before
2225 * transaction commit --- so it is always safe just to discard data
2226 * immediately in that mode. --sct
2227 */
2228
2229/*
2230 * The journal_unmap_buffer helper function returns zero if the buffer
2231 * concerned remains pinned as an anonymous buffer belonging to an older
2232 * transaction.
2233 *
2234 * We're outside-transaction here. Either or both of j_running_transaction
2235 * and j_committing_transaction may be NULL.
2236 */
2237static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2238 int partial_page)
2239{
2240 transaction_t *transaction;
2241 struct journal_head *jh;
2242 int may_free = 1;
2243
2244 BUFFER_TRACE(bh, "entry");
2245
2246 /*
2247 * It is safe to proceed here without the j_list_lock because the
2248 * buffers cannot be stolen by try_to_free_buffers as long as we are
2249 * holding the page lock. --sct
2250 */
2251
2252 jh = jbd2_journal_grab_journal_head(bh);
2253 if (!jh)
2254 goto zap_buffer_unlocked;
2255
2256 /* OK, we have data buffer in journaled mode */
2257 write_lock(&journal->j_state_lock);
2258 spin_lock(lock: &jh->b_state_lock);
2259 spin_lock(lock: &journal->j_list_lock);
2260
2261 /*
2262 * We cannot remove the buffer from checkpoint lists until the
2263 * transaction adding inode to orphan list (let's call it T)
2264 * is committed. Otherwise if the transaction changing the
2265 * buffer would be cleaned from the journal before T is
2266 * committed, a crash will cause that the correct contents of
2267 * the buffer will be lost. On the other hand we have to
2268 * clear the buffer dirty bit at latest at the moment when the
2269 * transaction marking the buffer as freed in the filesystem
2270 * structures is committed because from that moment on the
2271 * block can be reallocated and used by a different page.
2272 * Since the block hasn't been freed yet but the inode has
2273 * already been added to orphan list, it is safe for us to add
2274 * the buffer to BJ_Forget list of the newest transaction.
2275 *
2276 * Also we have to clear buffer_mapped flag of a truncated buffer
2277 * because the buffer_head may be attached to the page straddling
2278 * i_size (can happen only when blocksize < pagesize) and thus the
2279 * buffer_head can be reused when the file is extended again. So we end
2280 * up keeping around invalidated buffers attached to transactions'
2281 * BJ_Forget list just to stop checkpointing code from cleaning up
2282 * the transaction this buffer was modified in.
2283 */
2284 transaction = jh->b_transaction;
2285 if (transaction == NULL) {
2286 /* First case: not on any transaction. If it
2287 * has no checkpoint link, then we can zap it:
2288 * it's a writeback-mode buffer so we don't care
2289 * if it hits disk safely. */
2290 if (!jh->b_cp_transaction) {
2291 JBUFFER_TRACE(jh, "not on any transaction: zap");
2292 goto zap_buffer;
2293 }
2294
2295 if (!buffer_dirty(bh)) {
2296 /* bdflush has written it. We can drop it now */
2297 __jbd2_journal_remove_checkpoint(jh);
2298 goto zap_buffer;
2299 }
2300
2301 /* OK, it must be in the journal but still not
2302 * written fully to disk: it's metadata or
2303 * journaled data... */
2304
2305 if (journal->j_running_transaction) {
2306 /* ... and once the current transaction has
2307 * committed, the buffer won't be needed any
2308 * longer. */
2309 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2310 may_free = __dispose_buffer(jh,
2311 transaction: journal->j_running_transaction);
2312 goto zap_buffer;
2313 } else {
2314 /* There is no currently-running transaction. So the
2315 * orphan record which we wrote for this file must have
2316 * passed into commit. We must attach this buffer to
2317 * the committing transaction, if it exists. */
2318 if (journal->j_committing_transaction) {
2319 JBUFFER_TRACE(jh, "give to committing trans");
2320 may_free = __dispose_buffer(jh,
2321 transaction: journal->j_committing_transaction);
2322 goto zap_buffer;
2323 } else {
2324 /* The orphan record's transaction has
2325 * committed. We can cleanse this buffer */
2326 clear_buffer_jbddirty(bh);
2327 __jbd2_journal_remove_checkpoint(jh);
2328 goto zap_buffer;
2329 }
2330 }
2331 } else if (transaction == journal->j_committing_transaction) {
2332 JBUFFER_TRACE(jh, "on committing transaction");
2333 /*
2334 * The buffer is committing, we simply cannot touch
2335 * it. If the page is straddling i_size we have to wait
2336 * for commit and try again.
2337 */
2338 if (partial_page) {
2339 spin_unlock(lock: &journal->j_list_lock);
2340 spin_unlock(lock: &jh->b_state_lock);
2341 write_unlock(&journal->j_state_lock);
2342 jbd2_journal_put_journal_head(jh);
2343 /* Already zapped buffer? Nothing to do... */
2344 if (!bh->b_bdev)
2345 return 0;
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(lock: &journal->j_list_lock);
2360 spin_unlock(lock: &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(lock: &journal->j_list_lock);
2387 spin_unlock(lock: &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_invalidate_folio()
2404 * @journal: journal to use for flush...
2405 * @folio: folio 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_invalidate_folio(journal_t *journal, struct folio *folio,
2415 size_t offset, size_t length)
2416{
2417 struct buffer_head *head, *bh, *next;
2418 unsigned int stop = offset + length;
2419 unsigned int curr_off = 0;
2420 int partial_page = (offset || length < folio_size(folio));
2421 int may_free = 1;
2422 int ret = 0;
2423
2424 if (!folio_test_locked(folio))
2425 BUG();
2426 head = folio_buffers(folio);
2427 if (!head)
2428 return 0;
2429
2430 BUG_ON(stop > folio_size(folio) || stop < length);
2431
2432 /* We will potentially be playing with lists other than just the
2433 * data lists (especially for journaled data mode), so be
2434 * cautious in our locking. */
2435
2436 bh = head;
2437 do {
2438 unsigned int next_off = curr_off + bh->b_size;
2439 next = bh->b_this_page;
2440
2441 if (next_off > stop)
2442 return 0;
2443
2444 if (offset <= curr_off) {
2445 /* This block is wholly outside the truncation point */
2446 lock_buffer(bh);
2447 ret = journal_unmap_buffer(journal, bh, partial_page);
2448 unlock_buffer(bh);
2449 if (ret < 0)
2450 return ret;
2451 may_free &= ret;
2452 }
2453 curr_off = next_off;
2454 bh = next;
2455
2456 } while (bh != head);
2457
2458 if (!partial_page) {
2459 if (may_free && try_to_free_buffers(folio))
2460 J_ASSERT(!folio_buffers(folio));
2461 }
2462 return 0;
2463}
2464
2465/*
2466 * File a buffer on the given transaction list.
2467 */
2468void __jbd2_journal_file_buffer(struct journal_head *jh,
2469 transaction_t *transaction, int jlist)
2470{
2471 struct journal_head **list = NULL;
2472 int was_dirty = 0;
2473 struct buffer_head *bh = jh2bh(jh);
2474
2475 lockdep_assert_held(&jh->b_state_lock);
2476 assert_spin_locked(&transaction->t_journal->j_list_lock);
2477
2478 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2479 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2480 jh->b_transaction == NULL);
2481
2482 if (jh->b_transaction && jh->b_jlist == jlist)
2483 return;
2484
2485 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2486 jlist == BJ_Shadow || jlist == BJ_Forget) {
2487 /*
2488 * For metadata buffers, we track dirty bit in buffer_jbddirty
2489 * instead of buffer_dirty. We should not see a dirty bit set
2490 * here because we clear it in do_get_write_access but e.g.
2491 * tune2fs can modify the sb and set the dirty bit at any time
2492 * so we try to gracefully handle that.
2493 */
2494 if (buffer_dirty(bh))
2495 warn_dirty_buffer(bh);
2496 if (test_clear_buffer_dirty(bh) ||
2497 test_clear_buffer_jbddirty(bh))
2498 was_dirty = 1;
2499 }
2500
2501 if (jh->b_transaction)
2502 __jbd2_journal_temp_unlink_buffer(jh);
2503 else
2504 jbd2_journal_grab_journal_head(bh);
2505 jh->b_transaction = transaction;
2506
2507 switch (jlist) {
2508 case BJ_None:
2509 J_ASSERT_JH(jh, !jh->b_committed_data);
2510 J_ASSERT_JH(jh, !jh->b_frozen_data);
2511 return;
2512 case BJ_Metadata:
2513 transaction->t_nr_buffers++;
2514 list = &transaction->t_buffers;
2515 break;
2516 case BJ_Forget:
2517 list = &transaction->t_forget;
2518 break;
2519 case BJ_Shadow:
2520 list = &transaction->t_shadow_list;
2521 break;
2522 case BJ_Reserved:
2523 list = &transaction->t_reserved_list;
2524 break;
2525 }
2526
2527 __blist_add_buffer(list, jh);
2528 jh->b_jlist = jlist;
2529
2530 if (was_dirty)
2531 set_buffer_jbddirty(bh);
2532}
2533
2534void jbd2_journal_file_buffer(struct journal_head *jh,
2535 transaction_t *transaction, int jlist)
2536{
2537 spin_lock(lock: &jh->b_state_lock);
2538 spin_lock(lock: &transaction->t_journal->j_list_lock);
2539 __jbd2_journal_file_buffer(jh, transaction, jlist);
2540 spin_unlock(lock: &transaction->t_journal->j_list_lock);
2541 spin_unlock(lock: &jh->b_state_lock);
2542}
2543
2544/*
2545 * Remove a buffer from its current buffer list in preparation for
2546 * dropping it from its current transaction entirely. If the buffer has
2547 * already started to be used by a subsequent transaction, refile the
2548 * buffer on that transaction's metadata list.
2549 *
2550 * Called under j_list_lock
2551 * Called under jh->b_state_lock
2552 *
2553 * When this function returns true, there's no next transaction to refile to
2554 * and the caller has to drop jh reference through
2555 * jbd2_journal_put_journal_head().
2556 */
2557bool __jbd2_journal_refile_buffer(struct journal_head *jh)
2558{
2559 int was_dirty, jlist;
2560 struct buffer_head *bh = jh2bh(jh);
2561
2562 lockdep_assert_held(&jh->b_state_lock);
2563 if (jh->b_transaction)
2564 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2565
2566 /* If the buffer is now unused, just drop it. */
2567 if (jh->b_next_transaction == NULL) {
2568 __jbd2_journal_unfile_buffer(jh);
2569 return true;
2570 }
2571
2572 /*
2573 * It has been modified by a later transaction: add it to the new
2574 * transaction's metadata list.
2575 */
2576
2577 was_dirty = test_clear_buffer_jbddirty(bh);
2578 __jbd2_journal_temp_unlink_buffer(jh);
2579
2580 /*
2581 * b_transaction must be set, otherwise the new b_transaction won't
2582 * be holding jh reference
2583 */
2584 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2585
2586 /*
2587 * We set b_transaction here because b_next_transaction will inherit
2588 * our jh reference and thus __jbd2_journal_file_buffer() must not
2589 * take a new one.
2590 */
2591 WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2592 WRITE_ONCE(jh->b_next_transaction, NULL);
2593 if (buffer_freed(bh))
2594 jlist = BJ_Forget;
2595 else if (jh->b_modified)
2596 jlist = BJ_Metadata;
2597 else
2598 jlist = BJ_Reserved;
2599 __jbd2_journal_file_buffer(jh, transaction: jh->b_transaction, jlist);
2600 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2601
2602 if (was_dirty)
2603 set_buffer_jbddirty(bh);
2604 return false;
2605}
2606
2607/*
2608 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2609 * bh reference so that we can safely unlock bh.
2610 *
2611 * The jh and bh may be freed by this call.
2612 */
2613void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2614{
2615 bool drop;
2616
2617 spin_lock(lock: &jh->b_state_lock);
2618 spin_lock(lock: &journal->j_list_lock);
2619 drop = __jbd2_journal_refile_buffer(jh);
2620 spin_unlock(lock: &jh->b_state_lock);
2621 spin_unlock(lock: &journal->j_list_lock);
2622 if (drop)
2623 jbd2_journal_put_journal_head(jh);
2624}
2625
2626/*
2627 * File inode in the inode list of the handle's transaction
2628 */
2629static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2630 unsigned long flags, loff_t start_byte, loff_t end_byte)
2631{
2632 transaction_t *transaction = handle->h_transaction;
2633 journal_t *journal;
2634
2635 if (is_handle_aborted(handle))
2636 return -EROFS;
2637 journal = transaction->t_journal;
2638
2639 jbd2_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2640 transaction->t_tid);
2641
2642 spin_lock(lock: &journal->j_list_lock);
2643 jinode->i_flags |= flags;
2644
2645 if (jinode->i_dirty_end) {
2646 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2647 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2648 } else {
2649 jinode->i_dirty_start = start_byte;
2650 jinode->i_dirty_end = end_byte;
2651 }
2652
2653 /* Is inode already attached where we need it? */
2654 if (jinode->i_transaction == transaction ||
2655 jinode->i_next_transaction == transaction)
2656 goto done;
2657
2658 /*
2659 * We only ever set this variable to 1 so the test is safe. Since
2660 * t_need_data_flush is likely to be set, we do the test to save some
2661 * cacheline bouncing
2662 */
2663 if (!transaction->t_need_data_flush)
2664 transaction->t_need_data_flush = 1;
2665 /* On some different transaction's list - should be
2666 * the committing one */
2667 if (jinode->i_transaction) {
2668 J_ASSERT(jinode->i_next_transaction == NULL);
2669 J_ASSERT(jinode->i_transaction ==
2670 journal->j_committing_transaction);
2671 jinode->i_next_transaction = transaction;
2672 goto done;
2673 }
2674 /* Not on any transaction list... */
2675 J_ASSERT(!jinode->i_next_transaction);
2676 jinode->i_transaction = transaction;
2677 list_add(new: &jinode->i_list, head: &transaction->t_inode_list);
2678done:
2679 spin_unlock(lock: &journal->j_list_lock);
2680
2681 return 0;
2682}
2683
2684int jbd2_journal_inode_ranged_write(handle_t *handle,
2685 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2686{
2687 return jbd2_journal_file_inode(handle, jinode,
2688 JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2689 end_byte: start_byte + length - 1);
2690}
2691
2692int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2693 loff_t start_byte, loff_t length)
2694{
2695 return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2696 start_byte, end_byte: start_byte + length - 1);
2697}
2698
2699/*
2700 * File truncate and transaction commit interact with each other in a
2701 * non-trivial way. If a transaction writing data block A is
2702 * committing, we cannot discard the data by truncate until we have
2703 * written them. Otherwise if we crashed after the transaction with
2704 * write has committed but before the transaction with truncate has
2705 * committed, we could see stale data in block A. This function is a
2706 * helper to solve this problem. It starts writeout of the truncated
2707 * part in case it is in the committing transaction.
2708 *
2709 * Filesystem code must call this function when inode is journaled in
2710 * ordered mode before truncation happens and after the inode has been
2711 * placed on orphan list with the new inode size. The second condition
2712 * avoids the race that someone writes new data and we start
2713 * committing the transaction after this function has been called but
2714 * before a transaction for truncate is started (and furthermore it
2715 * allows us to optimize the case where the addition to orphan list
2716 * happens in the same transaction as write --- we don't have to write
2717 * any data in such case).
2718 */
2719int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2720 struct jbd2_inode *jinode,
2721 loff_t new_size)
2722{
2723 transaction_t *inode_trans, *commit_trans;
2724 int ret = 0;
2725
2726 /* This is a quick check to avoid locking if not necessary */
2727 if (!jinode->i_transaction)
2728 goto out;
2729 /* Locks are here just to force reading of recent values, it is
2730 * enough that the transaction was not committing before we started
2731 * a transaction adding the inode to orphan list */
2732 read_lock(&journal->j_state_lock);
2733 commit_trans = journal->j_committing_transaction;
2734 read_unlock(&journal->j_state_lock);
2735 spin_lock(lock: &journal->j_list_lock);
2736 inode_trans = jinode->i_transaction;
2737 spin_unlock(lock: &journal->j_list_lock);
2738 if (inode_trans == commit_trans) {
2739 ret = filemap_fdatawrite_range(mapping: jinode->i_vfs_inode->i_mapping,
2740 start: new_size, LLONG_MAX);
2741 if (ret)
2742 jbd2_journal_abort(journal, ret);
2743 }
2744out:
2745 return ret;
2746}
2747