1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/transaction.c
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
7 * Copyright 1998 Red Hat corp --- All Rights Reserved
9 * Generic filesystem transaction handling code; part of the ext2fs
12 * This file manages transactions (compound commits managed by the
13 * journaling code) and handles (individual atomic operations by the
17 #include <linux/time.h>
19 #include <linux/jbd2.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/timer.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>
31 #include <trace/events/jbd2.h>
33 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
36 static struct kmem_cache *transaction_cache;
37 int __init jbd2_journal_init_transaction_cache(void)
39 J_ASSERT(!transaction_cache);
40 transaction_cache = kmem_cache_create("jbd2_transaction_s",
41 sizeof(transaction_t),
43 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
45 if (!transaction_cache) {
46 pr_emerg("JBD2: failed to create transaction cache\n");
52 void jbd2_journal_destroy_transaction_cache(void)
54 kmem_cache_destroy(transaction_cache);
55 transaction_cache = NULL;
58 void jbd2_journal_free_transaction(transaction_t *transaction)
60 if (unlikely(ZERO_OR_NULL_PTR(transaction)))
62 kmem_cache_free(transaction_cache, transaction);
66 * Base amount of descriptor blocks we reserve for each transaction.
68 static int jbd2_descriptor_blocks_per_trans(journal_t *journal)
70 int tag_space = journal->j_blocksize - sizeof(journal_header_t);
75 if (jbd2_journal_has_csum_v2or3(journal))
76 tag_space -= sizeof(struct jbd2_journal_block_tail);
77 /* Commit code leaves a slack space of 16 bytes at the end of block */
78 tags_per_block = (tag_space - 16) / journal_tag_bytes(journal);
80 * Revoke descriptors are accounted separately so we need to reserve
81 * space for commit block and normal transaction descriptor blocks.
83 return 1 + DIV_ROUND_UP(journal->j_max_transaction_buffers,
88 * jbd2_get_transaction: obtain a new transaction_t object.
90 * Simply initialise a new transaction. Initialize it in
91 * RUNNING state and add it to the current journal (which should not
92 * have an existing running transaction: we only make a new transaction
93 * once we have started to commit the old one).
96 * The journal MUST be locked. We don't perform atomic mallocs on the
97 * new transaction and we can't block without protecting against other
98 * processes trying to touch the journal while it is in transition.
102 static void jbd2_get_transaction(journal_t *journal,
103 transaction_t *transaction)
105 transaction->t_journal = journal;
106 transaction->t_state = T_RUNNING;
107 transaction->t_start_time = ktime_get();
108 transaction->t_tid = journal->j_transaction_sequence++;
109 transaction->t_expires = jiffies + journal->j_commit_interval;
110 spin_lock_init(&transaction->t_handle_lock);
111 atomic_set(&transaction->t_updates, 0);
112 atomic_set(&transaction->t_outstanding_credits,
113 jbd2_descriptor_blocks_per_trans(journal) +
114 atomic_read(&journal->j_reserved_credits));
115 atomic_set(&transaction->t_outstanding_revokes, 0);
116 atomic_set(&transaction->t_handle_count, 0);
117 INIT_LIST_HEAD(&transaction->t_inode_list);
118 INIT_LIST_HEAD(&transaction->t_private_list);
120 /* Set up the commit timer for the new transaction. */
121 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
122 add_timer(&journal->j_commit_timer);
124 J_ASSERT(journal->j_running_transaction == NULL);
125 journal->j_running_transaction = transaction;
126 transaction->t_max_wait = 0;
127 transaction->t_start = jiffies;
128 transaction->t_requested = 0;
134 * A handle_t is an object which represents a single atomic update to a
135 * filesystem, and which tracks all of the modifications which form part
136 * of that one update.
140 * Update transaction's maximum wait time, if debugging is enabled.
142 * In order for t_max_wait to be reliable, it must be protected by a
143 * lock. But doing so will mean that start_this_handle() can not be
144 * run in parallel on SMP systems, which limits our scalability. So
145 * unless debugging is enabled, we no longer update t_max_wait, which
146 * means that maximum wait time reported by the jbd2_run_stats
147 * tracepoint will always be zero.
149 static inline void update_t_max_wait(transaction_t *transaction,
152 #ifdef CONFIG_JBD2_DEBUG
153 if (jbd2_journal_enable_debug &&
154 time_after(transaction->t_start, ts)) {
155 ts = jbd2_time_diff(ts, transaction->t_start);
156 spin_lock(&transaction->t_handle_lock);
157 if (ts > transaction->t_max_wait)
158 transaction->t_max_wait = ts;
159 spin_unlock(&transaction->t_handle_lock);
165 * Wait until running transaction passes to T_FLUSH state and new transaction
166 * can thus be started. Also starts the commit if needed. The function expects
167 * running transaction to exist and releases j_state_lock.
169 static void wait_transaction_locked(journal_t *journal)
170 __releases(journal->j_state_lock)
174 tid_t tid = journal->j_running_transaction->t_tid;
176 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177 TASK_UNINTERRUPTIBLE);
178 need_to_start = !tid_geq(journal->j_commit_request, tid);
179 read_unlock(&journal->j_state_lock);
181 jbd2_log_start_commit(journal, tid);
182 jbd2_might_wait_for_commit(journal);
184 finish_wait(&journal->j_wait_transaction_locked, &wait);
188 * Wait until running transaction transitions from T_SWITCH to T_FLUSH
189 * state and new transaction can thus be started. The function releases
192 static void wait_transaction_switching(journal_t *journal)
193 __releases(journal->j_state_lock)
197 if (WARN_ON(!journal->j_running_transaction ||
198 journal->j_running_transaction->t_state != T_SWITCH)) {
199 read_unlock(&journal->j_state_lock);
202 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
203 TASK_UNINTERRUPTIBLE);
204 read_unlock(&journal->j_state_lock);
206 * We don't call jbd2_might_wait_for_commit() here as there's no
207 * waiting for outstanding handles happening anymore in T_SWITCH state
208 * and handling of reserved handles actually relies on that for
212 finish_wait(&journal->j_wait_transaction_locked, &wait);
215 static void sub_reserved_credits(journal_t *journal, int blocks)
217 atomic_sub(blocks, &journal->j_reserved_credits);
218 wake_up(&journal->j_wait_reserved);
222 * Wait until we can add credits for handle to the running transaction. Called
223 * with j_state_lock held for reading. Returns 0 if handle joined the running
224 * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
227 * Note: because j_state_lock may be dropped depending on the return
228 * value, we need to fake out sparse so ti doesn't complain about a
229 * locking imbalance. Callers of add_transaction_credits will need to
230 * make a similar accomodation.
232 static int add_transaction_credits(journal_t *journal, int blocks,
234 __must_hold(&journal->j_state_lock)
236 transaction_t *t = journal->j_running_transaction;
238 int total = blocks + rsv_blocks;
241 * If the current transaction is locked down for commit, wait
242 * for the lock to be released.
244 if (t->t_state != T_RUNNING) {
245 WARN_ON_ONCE(t->t_state >= T_FLUSH);
246 wait_transaction_locked(journal);
247 __acquire(&journal->j_state_lock); /* fake out sparse */
252 * If there is not enough space left in the log to write all
253 * potential buffers requested by this operation, we need to
254 * stall pending a log checkpoint to free some more log space.
256 needed = atomic_add_return(total, &t->t_outstanding_credits);
257 if (needed > journal->j_max_transaction_buffers) {
259 * If the current transaction is already too large,
260 * then start to commit it: we can then go back and
261 * attach this handle to a new transaction.
263 atomic_sub(total, &t->t_outstanding_credits);
266 * Is the number of reserved credits in the current transaction too
267 * big to fit this handle? Wait until reserved credits are freed.
269 if (atomic_read(&journal->j_reserved_credits) + total >
270 journal->j_max_transaction_buffers) {
271 read_unlock(&journal->j_state_lock);
272 jbd2_might_wait_for_commit(journal);
273 wait_event(journal->j_wait_reserved,
274 atomic_read(&journal->j_reserved_credits) + total <=
275 journal->j_max_transaction_buffers);
276 __acquire(&journal->j_state_lock); /* fake out sparse */
280 wait_transaction_locked(journal);
281 __acquire(&journal->j_state_lock); /* fake out sparse */
286 * The commit code assumes that it can get enough log space
287 * without forcing a checkpoint. This is *critical* for
288 * correctness: a checkpoint of a buffer which is also
289 * associated with a committing transaction creates a deadlock,
290 * so commit simply cannot force through checkpoints.
292 * We must therefore ensure the necessary space in the journal
293 * *before* starting to dirty potentially checkpointed buffers
294 * in the new transaction.
296 if (jbd2_log_space_left(journal) < journal->j_max_transaction_buffers) {
297 atomic_sub(total, &t->t_outstanding_credits);
298 read_unlock(&journal->j_state_lock);
299 jbd2_might_wait_for_commit(journal);
300 write_lock(&journal->j_state_lock);
301 if (jbd2_log_space_left(journal) <
302 journal->j_max_transaction_buffers)
303 __jbd2_log_wait_for_space(journal);
304 write_unlock(&journal->j_state_lock);
305 __acquire(&journal->j_state_lock); /* fake out sparse */
309 /* No reservation? We are done... */
313 needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
314 /* We allow at most half of a transaction to be reserved */
315 if (needed > journal->j_max_transaction_buffers / 2) {
316 sub_reserved_credits(journal, rsv_blocks);
317 atomic_sub(total, &t->t_outstanding_credits);
318 read_unlock(&journal->j_state_lock);
319 jbd2_might_wait_for_commit(journal);
320 wait_event(journal->j_wait_reserved,
321 atomic_read(&journal->j_reserved_credits) + rsv_blocks
322 <= journal->j_max_transaction_buffers / 2);
323 __acquire(&journal->j_state_lock); /* fake out sparse */
330 * start_this_handle: Given a handle, deal with any locking or stalling
331 * needed to make sure that there is enough journal space for the handle
332 * to begin. Attach the handle to a transaction and set up the
333 * transaction's buffer credits.
336 static int start_this_handle(journal_t *journal, handle_t *handle,
339 transaction_t *transaction, *new_transaction = NULL;
340 int blocks = handle->h_total_credits;
342 unsigned long ts = jiffies;
344 if (handle->h_rsv_handle)
345 rsv_blocks = handle->h_rsv_handle->h_total_credits;
348 * Limit the number of reserved credits to 1/2 of maximum transaction
349 * size and limit the number of total credits to not exceed maximum
350 * transaction size per operation.
352 if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
353 (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
354 printk(KERN_ERR "JBD2: %s wants too many credits "
355 "credits:%d rsv_credits:%d max:%d\n",
356 current->comm, blocks, rsv_blocks,
357 journal->j_max_transaction_buffers);
364 * This check is racy but it is just an optimization of allocating new
365 * transaction early if there are high chances we'll need it. If we
366 * guess wrong, we'll retry or free unused transaction.
368 if (!data_race(journal->j_running_transaction)) {
370 * If __GFP_FS is not present, then we may be being called from
371 * inside the fs writeback layer, so we MUST NOT fail.
373 if ((gfp_mask & __GFP_FS) == 0)
374 gfp_mask |= __GFP_NOFAIL;
375 new_transaction = kmem_cache_zalloc(transaction_cache,
377 if (!new_transaction)
381 jbd_debug(3, "New handle %p going live.\n", handle);
384 * We need to hold j_state_lock until t_updates has been incremented,
385 * for proper journal barrier handling
388 read_lock(&journal->j_state_lock);
389 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
390 if (is_journal_aborted(journal) ||
391 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
392 read_unlock(&journal->j_state_lock);
393 jbd2_journal_free_transaction(new_transaction);
398 * Wait on the journal's transaction barrier if necessary. Specifically
399 * we allow reserved handles to proceed because otherwise commit could
400 * deadlock on page writeback not being able to complete.
402 if (!handle->h_reserved && journal->j_barrier_count) {
403 read_unlock(&journal->j_state_lock);
404 wait_event(journal->j_wait_transaction_locked,
405 journal->j_barrier_count == 0);
409 if (!journal->j_running_transaction) {
410 read_unlock(&journal->j_state_lock);
411 if (!new_transaction)
412 goto alloc_transaction;
413 write_lock(&journal->j_state_lock);
414 if (!journal->j_running_transaction &&
415 (handle->h_reserved || !journal->j_barrier_count)) {
416 jbd2_get_transaction(journal, new_transaction);
417 new_transaction = NULL;
419 write_unlock(&journal->j_state_lock);
423 transaction = journal->j_running_transaction;
425 if (!handle->h_reserved) {
426 /* We may have dropped j_state_lock - restart in that case */
427 if (add_transaction_credits(journal, blocks, rsv_blocks)) {
429 * add_transaction_credits releases
430 * j_state_lock on a non-zero return
432 __release(&journal->j_state_lock);
437 * We have handle reserved so we are allowed to join T_LOCKED
438 * transaction and we don't have to check for transaction size
439 * and journal space. But we still have to wait while running
440 * transaction is being switched to a committing one as it
441 * won't wait for any handles anymore.
443 if (transaction->t_state == T_SWITCH) {
444 wait_transaction_switching(journal);
447 sub_reserved_credits(journal, blocks);
448 handle->h_reserved = 0;
451 /* OK, account for the buffers that this operation expects to
452 * use and add the handle to the running transaction.
454 update_t_max_wait(transaction, ts);
455 handle->h_transaction = transaction;
456 handle->h_requested_credits = blocks;
457 handle->h_revoke_credits_requested = handle->h_revoke_credits;
458 handle->h_start_jiffies = jiffies;
459 atomic_inc(&transaction->t_updates);
460 atomic_inc(&transaction->t_handle_count);
461 jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
463 atomic_read(&transaction->t_outstanding_credits),
464 jbd2_log_space_left(journal));
465 read_unlock(&journal->j_state_lock);
466 current->journal_info = handle;
468 rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
469 jbd2_journal_free_transaction(new_transaction);
471 * Ensure that no allocations done while the transaction is open are
472 * going to recurse back to the fs layer.
474 handle->saved_alloc_context = memalloc_nofs_save();
478 /* Allocate a new handle. This should probably be in a slab... */
479 static handle_t *new_handle(int nblocks)
481 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
484 handle->h_total_credits = nblocks;
490 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
491 int revoke_records, gfp_t gfp_mask,
492 unsigned int type, unsigned int line_no)
494 handle_t *handle = journal_current_handle();
498 return ERR_PTR(-EROFS);
501 J_ASSERT(handle->h_transaction->t_journal == journal);
506 nblocks += DIV_ROUND_UP(revoke_records,
507 journal->j_revoke_records_per_block);
508 handle = new_handle(nblocks);
510 return ERR_PTR(-ENOMEM);
512 handle_t *rsv_handle;
514 rsv_handle = new_handle(rsv_blocks);
516 jbd2_free_handle(handle);
517 return ERR_PTR(-ENOMEM);
519 rsv_handle->h_reserved = 1;
520 rsv_handle->h_journal = journal;
521 handle->h_rsv_handle = rsv_handle;
523 handle->h_revoke_credits = revoke_records;
525 err = start_this_handle(journal, handle, gfp_mask);
527 if (handle->h_rsv_handle)
528 jbd2_free_handle(handle->h_rsv_handle);
529 jbd2_free_handle(handle);
532 handle->h_type = type;
533 handle->h_line_no = line_no;
534 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
535 handle->h_transaction->t_tid, type,
540 EXPORT_SYMBOL(jbd2__journal_start);
544 * jbd2_journal_start() - Obtain a new handle.
545 * @journal: Journal to start transaction on.
546 * @nblocks: number of block buffer we might modify
548 * We make sure that the transaction can guarantee at least nblocks of
549 * modified buffers in the log. We block until the log can guarantee
550 * that much space. Additionally, if rsv_blocks > 0, we also create another
551 * handle with rsv_blocks reserved blocks in the journal. This handle is
552 * stored in h_rsv_handle. It is not attached to any particular transaction
553 * and thus doesn't block transaction commit. If the caller uses this reserved
554 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
555 * on the parent handle will dispose the reserved one. Reserved handle has to
556 * be converted to a normal handle using jbd2_journal_start_reserved() before
559 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
562 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
564 return jbd2__journal_start(journal, nblocks, 0, 0, GFP_NOFS, 0, 0);
566 EXPORT_SYMBOL(jbd2_journal_start);
568 static void __jbd2_journal_unreserve_handle(handle_t *handle, transaction_t *t)
570 journal_t *journal = handle->h_journal;
572 WARN_ON(!handle->h_reserved);
573 sub_reserved_credits(journal, handle->h_total_credits);
575 atomic_sub(handle->h_total_credits, &t->t_outstanding_credits);
578 void jbd2_journal_free_reserved(handle_t *handle)
580 journal_t *journal = handle->h_journal;
582 /* Get j_state_lock to pin running transaction if it exists */
583 read_lock(&journal->j_state_lock);
584 __jbd2_journal_unreserve_handle(handle, journal->j_running_transaction);
585 read_unlock(&journal->j_state_lock);
586 jbd2_free_handle(handle);
588 EXPORT_SYMBOL(jbd2_journal_free_reserved);
591 * jbd2_journal_start_reserved() - start reserved handle
592 * @handle: handle to start
593 * @type: for handle statistics
594 * @line_no: for handle statistics
596 * Start handle that has been previously reserved with jbd2_journal_reserve().
597 * This attaches @handle to the running transaction (or creates one if there's
598 * not transaction running). Unlike jbd2_journal_start() this function cannot
599 * block on journal commit, checkpointing, or similar stuff. It can block on
600 * memory allocation or frozen journal though.
602 * Return 0 on success, non-zero on error - handle is freed in that case.
604 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
605 unsigned int line_no)
607 journal_t *journal = handle->h_journal;
610 if (WARN_ON(!handle->h_reserved)) {
611 /* Someone passed in normal handle? Just stop it. */
612 jbd2_journal_stop(handle);
616 * Usefulness of mixing of reserved and unreserved handles is
617 * questionable. So far nobody seems to need it so just error out.
619 if (WARN_ON(current->journal_info)) {
620 jbd2_journal_free_reserved(handle);
624 handle->h_journal = NULL;
626 * GFP_NOFS is here because callers are likely from writeback or
627 * similarly constrained call sites
629 ret = start_this_handle(journal, handle, GFP_NOFS);
631 handle->h_journal = journal;
632 jbd2_journal_free_reserved(handle);
635 handle->h_type = type;
636 handle->h_line_no = line_no;
637 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
638 handle->h_transaction->t_tid, type,
639 line_no, handle->h_total_credits);
642 EXPORT_SYMBOL(jbd2_journal_start_reserved);
645 * jbd2_journal_extend() - extend buffer credits.
646 * @handle: handle to 'extend'
647 * @nblocks: nr blocks to try to extend by.
648 * @revoke_records: number of revoke records to try to extend by.
650 * Some transactions, such as large extends and truncates, can be done
651 * atomically all at once or in several stages. The operation requests
652 * a credit for a number of buffer modifications in advance, but can
653 * extend its credit if it needs more.
655 * jbd2_journal_extend tries to give the running handle more buffer credits.
656 * It does not guarantee that allocation - this is a best-effort only.
657 * The calling process MUST be able to deal cleanly with a failure to
660 * Return 0 on success, non-zero on failure.
662 * return code < 0 implies an error
663 * return code > 0 implies normal transaction-full status.
665 int jbd2_journal_extend(handle_t *handle, int nblocks, int revoke_records)
667 transaction_t *transaction = handle->h_transaction;
672 if (is_handle_aborted(handle))
674 journal = transaction->t_journal;
678 read_lock(&journal->j_state_lock);
680 /* Don't extend a locked-down transaction! */
681 if (transaction->t_state != T_RUNNING) {
682 jbd_debug(3, "denied handle %p %d blocks: "
683 "transaction not running\n", handle, nblocks);
687 nblocks += DIV_ROUND_UP(
688 handle->h_revoke_credits_requested + revoke_records,
689 journal->j_revoke_records_per_block) -
691 handle->h_revoke_credits_requested,
692 journal->j_revoke_records_per_block);
693 spin_lock(&transaction->t_handle_lock);
694 wanted = atomic_add_return(nblocks,
695 &transaction->t_outstanding_credits);
697 if (wanted > journal->j_max_transaction_buffers) {
698 jbd_debug(3, "denied handle %p %d blocks: "
699 "transaction too large\n", handle, nblocks);
700 atomic_sub(nblocks, &transaction->t_outstanding_credits);
704 trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
706 handle->h_type, handle->h_line_no,
707 handle->h_total_credits,
710 handle->h_total_credits += nblocks;
711 handle->h_requested_credits += nblocks;
712 handle->h_revoke_credits += revoke_records;
713 handle->h_revoke_credits_requested += revoke_records;
716 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
718 spin_unlock(&transaction->t_handle_lock);
720 read_unlock(&journal->j_state_lock);
724 static void stop_this_handle(handle_t *handle)
726 transaction_t *transaction = handle->h_transaction;
727 journal_t *journal = transaction->t_journal;
730 J_ASSERT(journal_current_handle() == handle);
731 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
732 current->journal_info = NULL;
734 * Subtract necessary revoke descriptor blocks from handle credits. We
735 * take care to account only for revoke descriptor blocks the
736 * transaction will really need as large sequences of transactions with
737 * small numbers of revokes are relatively common.
739 revokes = handle->h_revoke_credits_requested - handle->h_revoke_credits;
741 int t_revokes, revoke_descriptors;
742 int rr_per_blk = journal->j_revoke_records_per_block;
744 WARN_ON_ONCE(DIV_ROUND_UP(revokes, rr_per_blk)
745 > handle->h_total_credits);
746 t_revokes = atomic_add_return(revokes,
747 &transaction->t_outstanding_revokes);
749 DIV_ROUND_UP(t_revokes, rr_per_blk) -
750 DIV_ROUND_UP(t_revokes - revokes, rr_per_blk);
751 handle->h_total_credits -= revoke_descriptors;
753 atomic_sub(handle->h_total_credits,
754 &transaction->t_outstanding_credits);
755 if (handle->h_rsv_handle)
756 __jbd2_journal_unreserve_handle(handle->h_rsv_handle,
758 if (atomic_dec_and_test(&transaction->t_updates))
759 wake_up(&journal->j_wait_updates);
761 rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
763 * Scope of the GFP_NOFS context is over here and so we can restore the
764 * original alloc context.
766 memalloc_nofs_restore(handle->saved_alloc_context);
770 * jbd2__journal_restart() - restart a handle .
771 * @handle: handle to restart
772 * @nblocks: nr credits requested
773 * @revoke_records: number of revoke record credits requested
774 * @gfp_mask: memory allocation flags (for start_this_handle)
776 * Restart a handle for a multi-transaction filesystem
779 * If the jbd2_journal_extend() call above fails to grant new buffer credits
780 * to a running handle, a call to jbd2_journal_restart will commit the
781 * handle's transaction so far and reattach the handle to a new
782 * transaction capable of guaranteeing the requested number of
783 * credits. We preserve reserved handle if there's any attached to the
786 int jbd2__journal_restart(handle_t *handle, int nblocks, int revoke_records,
789 transaction_t *transaction = handle->h_transaction;
795 /* If we've had an abort of any type, don't even think about
796 * actually doing the restart! */
797 if (is_handle_aborted(handle))
799 journal = transaction->t_journal;
800 tid = transaction->t_tid;
803 * First unlink the handle from its current transaction, and start the
806 jbd_debug(2, "restarting handle %p\n", handle);
807 stop_this_handle(handle);
808 handle->h_transaction = NULL;
811 * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can
812 * get rid of pointless j_state_lock traffic like this.
814 read_lock(&journal->j_state_lock);
815 need_to_start = !tid_geq(journal->j_commit_request, tid);
816 read_unlock(&journal->j_state_lock);
818 jbd2_log_start_commit(journal, tid);
819 handle->h_total_credits = nblocks +
820 DIV_ROUND_UP(revoke_records,
821 journal->j_revoke_records_per_block);
822 handle->h_revoke_credits = revoke_records;
823 ret = start_this_handle(journal, handle, gfp_mask);
824 trace_jbd2_handle_restart(journal->j_fs_dev->bd_dev,
825 ret ? 0 : handle->h_transaction->t_tid,
826 handle->h_type, handle->h_line_no,
827 handle->h_total_credits);
830 EXPORT_SYMBOL(jbd2__journal_restart);
833 int jbd2_journal_restart(handle_t *handle, int nblocks)
835 return jbd2__journal_restart(handle, nblocks, 0, GFP_NOFS);
837 EXPORT_SYMBOL(jbd2_journal_restart);
840 * Waits for any outstanding t_updates to finish.
841 * This is called with write j_state_lock held.
843 void jbd2_journal_wait_updates(journal_t *journal)
845 transaction_t *commit_transaction = journal->j_running_transaction;
847 if (!commit_transaction)
850 spin_lock(&commit_transaction->t_handle_lock);
851 while (atomic_read(&commit_transaction->t_updates)) {
854 prepare_to_wait(&journal->j_wait_updates, &wait,
855 TASK_UNINTERRUPTIBLE);
856 if (atomic_read(&commit_transaction->t_updates)) {
857 spin_unlock(&commit_transaction->t_handle_lock);
858 write_unlock(&journal->j_state_lock);
860 write_lock(&journal->j_state_lock);
861 spin_lock(&commit_transaction->t_handle_lock);
863 finish_wait(&journal->j_wait_updates, &wait);
865 spin_unlock(&commit_transaction->t_handle_lock);
869 * jbd2_journal_lock_updates () - establish a transaction barrier.
870 * @journal: Journal to establish a barrier on.
872 * This locks out any further updates from being started, and blocks
873 * until all existing updates have completed, returning only once the
874 * journal is in a quiescent state with no updates running.
876 * The journal lock should not be held on entry.
878 void jbd2_journal_lock_updates(journal_t *journal)
882 jbd2_might_wait_for_commit(journal);
884 write_lock(&journal->j_state_lock);
885 ++journal->j_barrier_count;
887 /* Wait until there are no reserved handles */
888 if (atomic_read(&journal->j_reserved_credits)) {
889 write_unlock(&journal->j_state_lock);
890 wait_event(journal->j_wait_reserved,
891 atomic_read(&journal->j_reserved_credits) == 0);
892 write_lock(&journal->j_state_lock);
895 /* Wait until there are no running t_updates */
896 jbd2_journal_wait_updates(journal);
898 write_unlock(&journal->j_state_lock);
901 * We have now established a barrier against other normal updates, but
902 * we also need to barrier against other jbd2_journal_lock_updates() calls
903 * to make sure that we serialise special journal-locked operations
906 mutex_lock(&journal->j_barrier);
910 * jbd2_journal_unlock_updates () - release barrier
911 * @journal: Journal to release the barrier on.
913 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
915 * Should be called without the journal lock held.
917 void jbd2_journal_unlock_updates (journal_t *journal)
919 J_ASSERT(journal->j_barrier_count != 0);
921 mutex_unlock(&journal->j_barrier);
922 write_lock(&journal->j_state_lock);
923 --journal->j_barrier_count;
924 write_unlock(&journal->j_state_lock);
925 wake_up(&journal->j_wait_transaction_locked);
928 static void warn_dirty_buffer(struct buffer_head *bh)
931 "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
932 "There's a risk of filesystem corruption in case of system "
934 bh->b_bdev, (unsigned long long)bh->b_blocknr);
937 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
938 static void jbd2_freeze_jh_data(struct journal_head *jh)
943 struct buffer_head *bh = jh2bh(jh);
945 J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
947 offset = offset_in_page(bh->b_data);
948 source = kmap_atomic(page);
949 /* Fire data frozen trigger just before we copy the data */
950 jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
951 memcpy(jh->b_frozen_data, source + offset, bh->b_size);
952 kunmap_atomic(source);
955 * Now that the frozen data is saved off, we need to store any matching
958 jh->b_frozen_triggers = jh->b_triggers;
962 * If the buffer is already part of the current transaction, then there
963 * is nothing we need to do. If it is already part of a prior
964 * transaction which we are still committing to disk, then we need to
965 * make sure that we do not overwrite the old copy: we do copy-out to
966 * preserve the copy going to disk. We also account the buffer against
967 * the handle's metadata buffer credits (unless the buffer is already
968 * part of the transaction, that is).
972 do_get_write_access(handle_t *handle, struct journal_head *jh,
975 struct buffer_head *bh;
976 transaction_t *transaction = handle->h_transaction;
979 char *frozen_buffer = NULL;
980 unsigned long start_lock, time_lock;
982 journal = transaction->t_journal;
984 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
986 JBUFFER_TRACE(jh, "entry");
990 /* @@@ Need to check for errors here at some point. */
992 start_lock = jiffies;
994 spin_lock(&jh->b_state_lock);
996 /* If it takes too long to lock the buffer, trace it */
997 time_lock = jbd2_time_diff(start_lock, jiffies);
998 if (time_lock > HZ/10)
999 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
1000 jiffies_to_msecs(time_lock));
1002 /* We now hold the buffer lock so it is safe to query the buffer
1003 * state. Is the buffer dirty?
1005 * If so, there are two possibilities. The buffer may be
1006 * non-journaled, and undergoing a quite legitimate writeback.
1007 * Otherwise, it is journaled, and we don't expect dirty buffers
1008 * in that state (the buffers should be marked JBD_Dirty
1009 * instead.) So either the IO is being done under our own
1010 * control and this is a bug, or it's a third party IO such as
1011 * dump(8) (which may leave the buffer scheduled for read ---
1012 * ie. locked but not dirty) or tune2fs (which may actually have
1013 * the buffer dirtied, ugh.) */
1015 if (buffer_dirty(bh)) {
1017 * First question: is this buffer already part of the current
1018 * transaction or the existing committing transaction?
1020 if (jh->b_transaction) {
1022 jh->b_transaction == transaction ||
1023 jh->b_transaction ==
1024 journal->j_committing_transaction);
1025 if (jh->b_next_transaction)
1026 J_ASSERT_JH(jh, jh->b_next_transaction ==
1028 warn_dirty_buffer(bh);
1031 * In any case we need to clean the dirty flag and we must
1032 * do it under the buffer lock to be sure we don't race
1033 * with running write-out.
1035 JBUFFER_TRACE(jh, "Journalling dirty buffer");
1036 clear_buffer_dirty(bh);
1037 set_buffer_jbddirty(bh);
1043 if (is_handle_aborted(handle)) {
1044 spin_unlock(&jh->b_state_lock);
1050 * The buffer is already part of this transaction if b_transaction or
1051 * b_next_transaction points to it
1053 if (jh->b_transaction == transaction ||
1054 jh->b_next_transaction == transaction)
1058 * this is the first time this transaction is touching this buffer,
1059 * reset the modified flag
1064 * If the buffer is not journaled right now, we need to make sure it
1065 * doesn't get written to disk before the caller actually commits the
1068 if (!jh->b_transaction) {
1069 JBUFFER_TRACE(jh, "no transaction");
1070 J_ASSERT_JH(jh, !jh->b_next_transaction);
1071 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1073 * Make sure all stores to jh (b_modified, b_frozen_data) are
1074 * visible before attaching it to the running transaction.
1075 * Paired with barrier in jbd2_write_access_granted()
1078 spin_lock(&journal->j_list_lock);
1079 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1080 spin_unlock(&journal->j_list_lock);
1084 * If there is already a copy-out version of this buffer, then we don't
1085 * need to make another one
1087 if (jh->b_frozen_data) {
1088 JBUFFER_TRACE(jh, "has frozen data");
1089 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1093 JBUFFER_TRACE(jh, "owned by older transaction");
1094 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1095 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
1098 * There is one case we have to be very careful about. If the
1099 * committing transaction is currently writing this buffer out to disk
1100 * and has NOT made a copy-out, then we cannot modify the buffer
1101 * contents at all right now. The essence of copy-out is that it is
1102 * the extra copy, not the primary copy, which gets journaled. If the
1103 * primary copy is already going to disk then we cannot do copy-out
1106 if (buffer_shadow(bh)) {
1107 JBUFFER_TRACE(jh, "on shadow: sleep");
1108 spin_unlock(&jh->b_state_lock);
1109 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
1114 * Only do the copy if the currently-owning transaction still needs it.
1115 * If buffer isn't on BJ_Metadata list, the committing transaction is
1116 * past that stage (here we use the fact that BH_Shadow is set under
1117 * bh_state lock together with refiling to BJ_Shadow list and at this
1118 * point we know the buffer doesn't have BH_Shadow set).
1120 * Subtle point, though: if this is a get_undo_access, then we will be
1121 * relying on the frozen_data to contain the new value of the
1122 * committed_data record after the transaction, so we HAVE to force the
1123 * frozen_data copy in that case.
1125 if (jh->b_jlist == BJ_Metadata || force_copy) {
1126 JBUFFER_TRACE(jh, "generate frozen data");
1127 if (!frozen_buffer) {
1128 JBUFFER_TRACE(jh, "allocate memory for buffer");
1129 spin_unlock(&jh->b_state_lock);
1130 frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
1131 GFP_NOFS | __GFP_NOFAIL);
1134 jh->b_frozen_data = frozen_buffer;
1135 frozen_buffer = NULL;
1136 jbd2_freeze_jh_data(jh);
1140 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1141 * before attaching it to the running transaction. Paired with barrier
1142 * in jbd2_write_access_granted()
1145 jh->b_next_transaction = transaction;
1148 spin_unlock(&jh->b_state_lock);
1151 * If we are about to journal a buffer, then any revoke pending on it is
1154 jbd2_journal_cancel_revoke(handle, jh);
1157 if (unlikely(frozen_buffer)) /* It's usually NULL */
1158 jbd2_free(frozen_buffer, bh->b_size);
1160 JBUFFER_TRACE(jh, "exit");
1164 /* Fast check whether buffer is already attached to the required transaction */
1165 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1168 struct journal_head *jh;
1171 /* Dirty buffers require special handling... */
1172 if (buffer_dirty(bh))
1176 * RCU protects us from dereferencing freed pages. So the checks we do
1177 * are guaranteed not to oops. However the jh slab object can get freed
1178 * & reallocated while we work with it. So we have to be careful. When
1179 * we see jh attached to the running transaction, we know it must stay
1180 * so until the transaction is committed. Thus jh won't be freed and
1181 * will be attached to the same bh while we run. However it can
1182 * happen jh gets freed, reallocated, and attached to the transaction
1183 * just after we get pointer to it from bh. So we have to be careful
1184 * and recheck jh still belongs to our bh before we return success.
1187 if (!buffer_jbd(bh))
1189 /* This should be bh2jh() but that doesn't work with inline functions */
1190 jh = READ_ONCE(bh->b_private);
1193 /* For undo access buffer must have data copied */
1194 if (undo && !jh->b_committed_data)
1196 if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1197 READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1200 * There are two reasons for the barrier here:
1201 * 1) Make sure to fetch b_bh after we did previous checks so that we
1202 * detect when jh went through free, realloc, attach to transaction
1203 * while we were checking. Paired with implicit barrier in that path.
1204 * 2) So that access to bh done after jbd2_write_access_granted()
1205 * doesn't get reordered and see inconsistent state of concurrent
1206 * do_get_write_access().
1209 if (unlikely(jh->b_bh != bh))
1218 * jbd2_journal_get_write_access() - notify intent to modify a buffer
1219 * for metadata (not data) update.
1220 * @handle: transaction to add buffer modifications to
1221 * @bh: bh to be used for metadata writes
1223 * Returns: error code or 0 on success.
1225 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1226 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1229 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1231 struct journal_head *jh;
1234 if (is_handle_aborted(handle))
1237 if (jbd2_write_access_granted(handle, bh, false))
1240 jh = jbd2_journal_add_journal_head(bh);
1241 /* We do not want to get caught playing with fields which the
1242 * log thread also manipulates. Make sure that the buffer
1243 * completes any outstanding IO before proceeding. */
1244 rc = do_get_write_access(handle, jh, 0);
1245 jbd2_journal_put_journal_head(jh);
1251 * When the user wants to journal a newly created buffer_head
1252 * (ie. getblk() returned a new buffer and we are going to populate it
1253 * manually rather than reading off disk), then we need to keep the
1254 * buffer_head locked until it has been completely filled with new
1255 * data. In this case, we should be able to make the assertion that
1256 * the bh is not already part of an existing transaction.
1258 * The buffer should already be locked by the caller by this point.
1259 * There is no lock ranking violation: it was a newly created,
1260 * unlocked buffer beforehand. */
1263 * jbd2_journal_get_create_access () - notify intent to use newly created bh
1264 * @handle: transaction to new buffer to
1267 * Call this if you create a new bh.
1269 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1271 transaction_t *transaction = handle->h_transaction;
1273 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1276 jbd_debug(5, "journal_head %p\n", jh);
1278 if (is_handle_aborted(handle))
1280 journal = transaction->t_journal;
1283 JBUFFER_TRACE(jh, "entry");
1285 * The buffer may already belong to this transaction due to pre-zeroing
1286 * in the filesystem's new_block code. It may also be on the previous,
1287 * committing transaction's lists, but it HAS to be in Forget state in
1288 * that case: the transaction must have deleted the buffer for it to be
1291 spin_lock(&jh->b_state_lock);
1292 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1293 jh->b_transaction == NULL ||
1294 (jh->b_transaction == journal->j_committing_transaction &&
1295 jh->b_jlist == BJ_Forget)));
1297 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1298 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1300 if (jh->b_transaction == NULL) {
1302 * Previous jbd2_journal_forget() could have left the buffer
1303 * with jbddirty bit set because it was being committed. When
1304 * the commit finished, we've filed the buffer for
1305 * checkpointing and marked it dirty. Now we are reallocating
1306 * the buffer so the transaction freeing it must have
1307 * committed and so it's safe to clear the dirty bit.
1309 clear_buffer_dirty(jh2bh(jh));
1310 /* first access by this transaction */
1313 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1314 spin_lock(&journal->j_list_lock);
1315 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1316 spin_unlock(&journal->j_list_lock);
1317 } else if (jh->b_transaction == journal->j_committing_transaction) {
1318 /* first access by this transaction */
1321 JBUFFER_TRACE(jh, "set next transaction");
1322 spin_lock(&journal->j_list_lock);
1323 jh->b_next_transaction = transaction;
1324 spin_unlock(&journal->j_list_lock);
1326 spin_unlock(&jh->b_state_lock);
1329 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1330 * blocks which contain freed but then revoked metadata. We need
1331 * to cancel the revoke in case we end up freeing it yet again
1332 * and the reallocating as data - this would cause a second revoke,
1333 * which hits an assertion error.
1335 JBUFFER_TRACE(jh, "cancelling revoke");
1336 jbd2_journal_cancel_revoke(handle, jh);
1338 jbd2_journal_put_journal_head(jh);
1343 * jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1344 * non-rewindable consequences
1345 * @handle: transaction
1346 * @bh: buffer to undo
1348 * Sometimes there is a need to distinguish between metadata which has
1349 * been committed to disk and that which has not. The ext3fs code uses
1350 * this for freeing and allocating space, we have to make sure that we
1351 * do not reuse freed space until the deallocation has been committed,
1352 * since if we overwrote that space we would make the delete
1353 * un-rewindable in case of a crash.
1355 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1356 * buffer for parts of non-rewindable operations such as delete
1357 * operations on the bitmaps. The journaling code must keep a copy of
1358 * the buffer's contents prior to the undo_access call until such time
1359 * as we know that the buffer has definitely been committed to disk.
1361 * We never need to know which transaction the committed data is part
1362 * of, buffers touched here are guaranteed to be dirtied later and so
1363 * will be committed to a new transaction in due course, at which point
1364 * we can discard the old committed data pointer.
1366 * Returns error number or 0 on success.
1368 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1371 struct journal_head *jh;
1372 char *committed_data = NULL;
1374 if (is_handle_aborted(handle))
1377 if (jbd2_write_access_granted(handle, bh, true))
1380 jh = jbd2_journal_add_journal_head(bh);
1381 JBUFFER_TRACE(jh, "entry");
1384 * Do this first --- it can drop the journal lock, so we want to
1385 * make sure that obtaining the committed_data is done
1386 * atomically wrt. completion of any outstanding commits.
1388 err = do_get_write_access(handle, jh, 1);
1393 if (!jh->b_committed_data)
1394 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1395 GFP_NOFS|__GFP_NOFAIL);
1397 spin_lock(&jh->b_state_lock);
1398 if (!jh->b_committed_data) {
1399 /* Copy out the current buffer contents into the
1400 * preserved, committed copy. */
1401 JBUFFER_TRACE(jh, "generate b_committed data");
1402 if (!committed_data) {
1403 spin_unlock(&jh->b_state_lock);
1407 jh->b_committed_data = committed_data;
1408 committed_data = NULL;
1409 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1411 spin_unlock(&jh->b_state_lock);
1413 jbd2_journal_put_journal_head(jh);
1414 if (unlikely(committed_data))
1415 jbd2_free(committed_data, bh->b_size);
1420 * jbd2_journal_set_triggers() - Add triggers for commit writeout
1421 * @bh: buffer to trigger on
1422 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1424 * Set any triggers on this journal_head. This is always safe, because
1425 * triggers for a committing buffer will be saved off, and triggers for
1426 * a running transaction will match the buffer in that transaction.
1428 * Call with NULL to clear the triggers.
1430 void jbd2_journal_set_triggers(struct buffer_head *bh,
1431 struct jbd2_buffer_trigger_type *type)
1433 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1435 if (WARN_ON_ONCE(!jh))
1437 jh->b_triggers = type;
1438 jbd2_journal_put_journal_head(jh);
1441 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1442 struct jbd2_buffer_trigger_type *triggers)
1444 struct buffer_head *bh = jh2bh(jh);
1446 if (!triggers || !triggers->t_frozen)
1449 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1452 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1453 struct jbd2_buffer_trigger_type *triggers)
1455 if (!triggers || !triggers->t_abort)
1458 triggers->t_abort(triggers, jh2bh(jh));
1462 * jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1463 * @handle: transaction to add buffer to.
1464 * @bh: buffer to mark
1466 * mark dirty metadata which needs to be journaled as part of the current
1469 * The buffer must have previously had jbd2_journal_get_write_access()
1470 * called so that it has a valid journal_head attached to the buffer
1473 * The buffer is placed on the transaction's metadata list and is marked
1474 * as belonging to the transaction.
1476 * Returns error number or 0 on success.
1478 * Special care needs to be taken if the buffer already belongs to the
1479 * current committing transaction (in which case we should have frozen
1480 * data present for that commit). In that case, we don't relink the
1481 * buffer: that only gets done when the old transaction finally
1482 * completes its commit.
1484 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1486 transaction_t *transaction = handle->h_transaction;
1488 struct journal_head *jh;
1491 if (is_handle_aborted(handle))
1493 if (!buffer_jbd(bh))
1497 * We don't grab jh reference here since the buffer must be part
1498 * of the running transaction.
1501 jbd_debug(5, "journal_head %p\n", jh);
1502 JBUFFER_TRACE(jh, "entry");
1505 * This and the following assertions are unreliable since we may see jh
1506 * in inconsistent state unless we grab bh_state lock. But this is
1507 * crucial to catch bugs so let's do a reliable check until the
1508 * lockless handling is fully proven.
1510 if (data_race(jh->b_transaction != transaction &&
1511 jh->b_next_transaction != transaction)) {
1512 spin_lock(&jh->b_state_lock);
1513 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1514 jh->b_next_transaction == transaction);
1515 spin_unlock(&jh->b_state_lock);
1517 if (jh->b_modified == 1) {
1518 /* If it's in our transaction it must be in BJ_Metadata list. */
1519 if (data_race(jh->b_transaction == transaction &&
1520 jh->b_jlist != BJ_Metadata)) {
1521 spin_lock(&jh->b_state_lock);
1522 if (jh->b_transaction == transaction &&
1523 jh->b_jlist != BJ_Metadata)
1524 pr_err("JBD2: assertion failure: h_type=%u "
1525 "h_line_no=%u block_no=%llu jlist=%u\n",
1526 handle->h_type, handle->h_line_no,
1527 (unsigned long long) bh->b_blocknr,
1529 J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1530 jh->b_jlist == BJ_Metadata);
1531 spin_unlock(&jh->b_state_lock);
1536 journal = transaction->t_journal;
1537 spin_lock(&jh->b_state_lock);
1539 if (jh->b_modified == 0) {
1541 * This buffer's got modified and becoming part
1542 * of the transaction. This needs to be done
1543 * once a transaction -bzzz
1545 if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) {
1550 handle->h_total_credits--;
1554 * fastpath, to avoid expensive locking. If this buffer is already
1555 * on the running transaction's metadata list there is nothing to do.
1556 * Nobody can take it off again because there is a handle open.
1557 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1558 * result in this test being false, so we go in and take the locks.
1560 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1561 JBUFFER_TRACE(jh, "fastpath");
1562 if (unlikely(jh->b_transaction !=
1563 journal->j_running_transaction)) {
1564 printk(KERN_ERR "JBD2: %s: "
1565 "jh->b_transaction (%llu, %p, %u) != "
1566 "journal->j_running_transaction (%p, %u)\n",
1568 (unsigned long long) bh->b_blocknr,
1570 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1571 journal->j_running_transaction,
1572 journal->j_running_transaction ?
1573 journal->j_running_transaction->t_tid : 0);
1579 set_buffer_jbddirty(bh);
1582 * Metadata already on the current transaction list doesn't
1583 * need to be filed. Metadata on another transaction's list must
1584 * be committing, and will be refiled once the commit completes:
1585 * leave it alone for now.
1587 if (jh->b_transaction != transaction) {
1588 JBUFFER_TRACE(jh, "already on other transaction");
1589 if (unlikely(((jh->b_transaction !=
1590 journal->j_committing_transaction)) ||
1591 (jh->b_next_transaction != transaction))) {
1592 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1593 "bad jh for block %llu: "
1594 "transaction (%p, %u), "
1595 "jh->b_transaction (%p, %u), "
1596 "jh->b_next_transaction (%p, %u), jlist %u\n",
1598 (unsigned long long) bh->b_blocknr,
1599 transaction, transaction->t_tid,
1602 jh->b_transaction->t_tid : 0,
1603 jh->b_next_transaction,
1604 jh->b_next_transaction ?
1605 jh->b_next_transaction->t_tid : 0,
1610 /* And this case is illegal: we can't reuse another
1611 * transaction's data buffer, ever. */
1615 /* That test should have eliminated the following case: */
1616 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1618 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1619 spin_lock(&journal->j_list_lock);
1620 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1621 spin_unlock(&journal->j_list_lock);
1623 spin_unlock(&jh->b_state_lock);
1625 JBUFFER_TRACE(jh, "exit");
1630 * jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1631 * @handle: transaction handle
1632 * @bh: bh to 'forget'
1634 * We can only do the bforget if there are no commits pending against the
1635 * buffer. If the buffer is dirty in the current running transaction we
1636 * can safely unlink it.
1638 * bh may not be a journalled buffer at all - it may be a non-JBD
1639 * buffer which came off the hashtable. Check for this.
1641 * Decrements bh->b_count by one.
1643 * Allow this call even if the handle has aborted --- it may be part of
1644 * the caller's cleanup after an abort.
1646 int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh)
1648 transaction_t *transaction = handle->h_transaction;
1650 struct journal_head *jh;
1651 int drop_reserve = 0;
1653 int was_modified = 0;
1655 if (is_handle_aborted(handle))
1657 journal = transaction->t_journal;
1659 BUFFER_TRACE(bh, "entry");
1661 jh = jbd2_journal_grab_journal_head(bh);
1667 spin_lock(&jh->b_state_lock);
1669 /* Critical error: attempting to delete a bitmap buffer, maybe?
1670 * Don't do any jbd operations, and return an error. */
1671 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1672 "inconsistent data on disk")) {
1677 /* keep track of whether or not this transaction modified us */
1678 was_modified = jh->b_modified;
1681 * The buffer's going from the transaction, we must drop
1682 * all references -bzzz
1686 if (jh->b_transaction == transaction) {
1687 J_ASSERT_JH(jh, !jh->b_frozen_data);
1689 /* If we are forgetting a buffer which is already part
1690 * of this transaction, then we can just drop it from
1691 * the transaction immediately. */
1692 clear_buffer_dirty(bh);
1693 clear_buffer_jbddirty(bh);
1695 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1698 * we only want to drop a reference if this transaction
1699 * modified the buffer
1705 * We are no longer going to journal this buffer.
1706 * However, the commit of this transaction is still
1707 * important to the buffer: the delete that we are now
1708 * processing might obsolete an old log entry, so by
1709 * committing, we can satisfy the buffer's checkpoint.
1711 * So, if we have a checkpoint on the buffer, we should
1712 * now refile the buffer on our BJ_Forget list so that
1713 * we know to remove the checkpoint after we commit.
1716 spin_lock(&journal->j_list_lock);
1717 if (jh->b_cp_transaction) {
1718 __jbd2_journal_temp_unlink_buffer(jh);
1719 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1721 __jbd2_journal_unfile_buffer(jh);
1722 jbd2_journal_put_journal_head(jh);
1724 spin_unlock(&journal->j_list_lock);
1725 } else if (jh->b_transaction) {
1726 J_ASSERT_JH(jh, (jh->b_transaction ==
1727 journal->j_committing_transaction));
1728 /* However, if the buffer is still owned by a prior
1729 * (committing) transaction, we can't drop it yet... */
1730 JBUFFER_TRACE(jh, "belongs to older transaction");
1731 /* ... but we CAN drop it from the new transaction through
1732 * marking the buffer as freed and set j_next_transaction to
1733 * the new transaction, so that not only the commit code
1734 * knows it should clear dirty bits when it is done with the
1735 * buffer, but also the buffer can be checkpointed only
1736 * after the new transaction commits. */
1738 set_buffer_freed(bh);
1740 if (!jh->b_next_transaction) {
1741 spin_lock(&journal->j_list_lock);
1742 jh->b_next_transaction = transaction;
1743 spin_unlock(&journal->j_list_lock);
1745 J_ASSERT(jh->b_next_transaction == transaction);
1748 * only drop a reference if this transaction modified
1756 * Finally, if the buffer is not belongs to any
1757 * transaction, we can just drop it now if it has no
1760 spin_lock(&journal->j_list_lock);
1761 if (!jh->b_cp_transaction) {
1762 JBUFFER_TRACE(jh, "belongs to none transaction");
1763 spin_unlock(&journal->j_list_lock);
1768 * Otherwise, if the buffer has been written to disk,
1769 * it is safe to remove the checkpoint and drop it.
1771 if (!buffer_dirty(bh)) {
1772 __jbd2_journal_remove_checkpoint(jh);
1773 spin_unlock(&journal->j_list_lock);
1778 * The buffer is still not written to disk, we should
1779 * attach this buffer to current transaction so that the
1780 * buffer can be checkpointed only after the current
1781 * transaction commits.
1783 clear_buffer_dirty(bh);
1784 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1785 spin_unlock(&journal->j_list_lock);
1789 spin_unlock(&jh->b_state_lock);
1790 jbd2_journal_put_journal_head(jh);
1792 /* no need to reserve log space for this block -bzzz */
1793 handle->h_total_credits++;
1799 * jbd2_journal_stop() - complete a transaction
1800 * @handle: transaction to complete.
1802 * All done for a particular handle.
1804 * There is not much action needed here. We just return any remaining
1805 * buffer credits to the transaction and remove the handle. The only
1806 * complication is that we need to start a commit operation if the
1807 * filesystem is marked for synchronous update.
1809 * jbd2_journal_stop itself will not usually return an error, but it may
1810 * do so in unusual circumstances. In particular, expect it to
1811 * return -EIO if a jbd2_journal_abort has been executed since the
1812 * transaction began.
1814 int jbd2_journal_stop(handle_t *handle)
1816 transaction_t *transaction = handle->h_transaction;
1818 int err = 0, wait_for_commit = 0;
1822 if (--handle->h_ref > 0) {
1823 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1825 if (is_handle_aborted(handle))
1831 * Handle is already detached from the transaction so there is
1832 * nothing to do other than free the handle.
1834 memalloc_nofs_restore(handle->saved_alloc_context);
1837 journal = transaction->t_journal;
1838 tid = transaction->t_tid;
1840 if (is_handle_aborted(handle))
1843 jbd_debug(4, "Handle %p going down\n", handle);
1844 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1845 tid, handle->h_type, handle->h_line_no,
1846 jiffies - handle->h_start_jiffies,
1847 handle->h_sync, handle->h_requested_credits,
1848 (handle->h_requested_credits -
1849 handle->h_total_credits));
1852 * Implement synchronous transaction batching. If the handle
1853 * was synchronous, don't force a commit immediately. Let's
1854 * yield and let another thread piggyback onto this
1855 * transaction. Keep doing that while new threads continue to
1856 * arrive. It doesn't cost much - we're about to run a commit
1857 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1858 * operations by 30x or more...
1860 * We try and optimize the sleep time against what the
1861 * underlying disk can do, instead of having a static sleep
1862 * time. This is useful for the case where our storage is so
1863 * fast that it is more optimal to go ahead and force a flush
1864 * and wait for the transaction to be committed than it is to
1865 * wait for an arbitrary amount of time for new writers to
1866 * join the transaction. We achieve this by measuring how
1867 * long it takes to commit a transaction, and compare it with
1868 * how long this transaction has been running, and if run time
1869 * < commit time then we sleep for the delta and commit. This
1870 * greatly helps super fast disks that would see slowdowns as
1871 * more threads started doing fsyncs.
1873 * But don't do this if this process was the most recent one
1874 * to perform a synchronous write. We do this to detect the
1875 * case where a single process is doing a stream of sync
1876 * writes. No point in waiting for joiners in that case.
1878 * Setting max_batch_time to 0 disables this completely.
1881 if (handle->h_sync && journal->j_last_sync_writer != pid &&
1882 journal->j_max_batch_time) {
1883 u64 commit_time, trans_time;
1885 journal->j_last_sync_writer = pid;
1887 read_lock(&journal->j_state_lock);
1888 commit_time = journal->j_average_commit_time;
1889 read_unlock(&journal->j_state_lock);
1891 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1892 transaction->t_start_time));
1894 commit_time = max_t(u64, commit_time,
1895 1000*journal->j_min_batch_time);
1896 commit_time = min_t(u64, commit_time,
1897 1000*journal->j_max_batch_time);
1899 if (trans_time < commit_time) {
1900 ktime_t expires = ktime_add_ns(ktime_get(),
1902 set_current_state(TASK_UNINTERRUPTIBLE);
1903 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1908 transaction->t_synchronous_commit = 1;
1911 * If the handle is marked SYNC, we need to set another commit
1912 * going! We also want to force a commit if the transaction is too
1915 if (handle->h_sync ||
1916 time_after_eq(jiffies, transaction->t_expires)) {
1917 /* Do this even for aborted journals: an abort still
1918 * completes the commit thread, it just doesn't write
1919 * anything to disk. */
1921 jbd_debug(2, "transaction too old, requesting commit for "
1922 "handle %p\n", handle);
1923 /* This is non-blocking */
1924 jbd2_log_start_commit(journal, tid);
1927 * Special case: JBD2_SYNC synchronous updates require us
1928 * to wait for the commit to complete.
1930 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1931 wait_for_commit = 1;
1935 * Once stop_this_handle() drops t_updates, the transaction could start
1936 * committing on us and eventually disappear. So we must not
1937 * dereference transaction pointer again after calling
1938 * stop_this_handle().
1940 stop_this_handle(handle);
1942 if (wait_for_commit)
1943 err = jbd2_log_wait_commit(journal, tid);
1946 if (handle->h_rsv_handle)
1947 jbd2_free_handle(handle->h_rsv_handle);
1948 jbd2_free_handle(handle);
1954 * List management code snippets: various functions for manipulating the
1955 * transaction buffer lists.
1960 * Append a buffer to a transaction list, given the transaction's list head
1963 * j_list_lock is held.
1965 * jh->b_state_lock is held.
1969 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1972 jh->b_tnext = jh->b_tprev = jh;
1975 /* Insert at the tail of the list to preserve order */
1976 struct journal_head *first = *list, *last = first->b_tprev;
1978 jh->b_tnext = first;
1979 last->b_tnext = first->b_tprev = jh;
1984 * Remove a buffer from a transaction list, given the transaction's list
1987 * Called with j_list_lock held, and the journal may not be locked.
1989 * jh->b_state_lock is held.
1993 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1996 *list = jh->b_tnext;
2000 jh->b_tprev->b_tnext = jh->b_tnext;
2001 jh->b_tnext->b_tprev = jh->b_tprev;
2005 * Remove a buffer from the appropriate transaction list.
2007 * Note that this function can *change* the value of
2008 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
2009 * t_reserved_list. If the caller is holding onto a copy of one of these
2010 * pointers, it could go bad. Generally the caller needs to re-read the
2011 * pointer from the transaction_t.
2013 * Called under j_list_lock.
2015 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
2017 struct journal_head **list = NULL;
2018 transaction_t *transaction;
2019 struct buffer_head *bh = jh2bh(jh);
2021 lockdep_assert_held(&jh->b_state_lock);
2022 transaction = jh->b_transaction;
2024 assert_spin_locked(&transaction->t_journal->j_list_lock);
2026 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2027 if (jh->b_jlist != BJ_None)
2028 J_ASSERT_JH(jh, transaction != NULL);
2030 switch (jh->b_jlist) {
2034 transaction->t_nr_buffers--;
2035 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
2036 list = &transaction->t_buffers;
2039 list = &transaction->t_forget;
2042 list = &transaction->t_shadow_list;
2045 list = &transaction->t_reserved_list;
2049 __blist_del_buffer(list, jh);
2050 jh->b_jlist = BJ_None;
2051 if (transaction && is_journal_aborted(transaction->t_journal))
2052 clear_buffer_jbddirty(bh);
2053 else if (test_clear_buffer_jbddirty(bh))
2054 mark_buffer_dirty(bh); /* Expose it to the VM */
2058 * Remove buffer from all transactions. The caller is responsible for dropping
2059 * the jh reference that belonged to the transaction.
2061 * Called with bh_state lock and j_list_lock
2063 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
2065 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2066 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2068 __jbd2_journal_temp_unlink_buffer(jh);
2069 jh->b_transaction = NULL;
2072 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
2074 struct buffer_head *bh = jh2bh(jh);
2076 /* Get reference so that buffer cannot be freed before we unlock it */
2078 spin_lock(&jh->b_state_lock);
2079 spin_lock(&journal->j_list_lock);
2080 __jbd2_journal_unfile_buffer(jh);
2081 spin_unlock(&journal->j_list_lock);
2082 spin_unlock(&jh->b_state_lock);
2083 jbd2_journal_put_journal_head(jh);
2088 * Called from jbd2_journal_try_to_free_buffers().
2090 * Called under jh->b_state_lock
2093 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
2095 struct journal_head *jh;
2099 if (buffer_locked(bh) || buffer_dirty(bh))
2102 if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
2105 spin_lock(&journal->j_list_lock);
2106 if (jh->b_cp_transaction != NULL) {
2107 /* written-back checkpointed metadata buffer */
2108 JBUFFER_TRACE(jh, "remove from checkpoint list");
2109 __jbd2_journal_remove_checkpoint(jh);
2111 spin_unlock(&journal->j_list_lock);
2117 * jbd2_journal_try_to_free_buffers() - try to free page buffers.
2118 * @journal: journal for operation
2119 * @page: to try and free
2121 * For all the buffers on this page,
2122 * if they are fully written out ordered data, move them onto BUF_CLEAN
2123 * so try_to_free_buffers() can reap them.
2125 * This function returns non-zero if we wish try_to_free_buffers()
2126 * to be called. We do this if the page is releasable by try_to_free_buffers().
2127 * We also do it if the page has locked or dirty buffers and the caller wants
2128 * us to perform sync or async writeout.
2130 * This complicates JBD locking somewhat. We aren't protected by the
2131 * BKL here. We wish to remove the buffer from its committing or
2132 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2134 * This may *change* the value of transaction_t->t_datalist, so anyone
2135 * who looks at t_datalist needs to lock against this function.
2137 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2138 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
2139 * will come out of the lock with the buffer dirty, which makes it
2140 * ineligible for release here.
2142 * Who else is affected by this? hmm... Really the only contender
2143 * is do_get_write_access() - it could be looking at the buffer while
2144 * journal_try_to_free_buffer() is changing its state. But that
2145 * cannot happen because we never reallocate freed data as metadata
2146 * while the data is part of a transaction. Yes?
2148 * Return 0 on failure, 1 on success
2150 int jbd2_journal_try_to_free_buffers(journal_t *journal, struct page *page)
2152 struct buffer_head *head;
2153 struct buffer_head *bh;
2156 J_ASSERT(PageLocked(page));
2158 head = page_buffers(page);
2161 struct journal_head *jh;
2164 * We take our own ref against the journal_head here to avoid
2165 * having to add tons of locking around each instance of
2166 * jbd2_journal_put_journal_head().
2168 jh = jbd2_journal_grab_journal_head(bh);
2172 spin_lock(&jh->b_state_lock);
2173 __journal_try_to_free_buffer(journal, bh);
2174 spin_unlock(&jh->b_state_lock);
2175 jbd2_journal_put_journal_head(jh);
2178 } while ((bh = bh->b_this_page) != head);
2180 ret = try_to_free_buffers(page);
2186 * This buffer is no longer needed. If it is on an older transaction's
2187 * checkpoint list we need to record it on this transaction's forget list
2188 * to pin this buffer (and hence its checkpointing transaction) down until
2189 * this transaction commits. If the buffer isn't on a checkpoint list, we
2191 * Returns non-zero if JBD no longer has an interest in the buffer.
2193 * Called under j_list_lock.
2195 * Called under jh->b_state_lock.
2197 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2200 struct buffer_head *bh = jh2bh(jh);
2202 if (jh->b_cp_transaction) {
2203 JBUFFER_TRACE(jh, "on running+cp transaction");
2204 __jbd2_journal_temp_unlink_buffer(jh);
2206 * We don't want to write the buffer anymore, clear the
2207 * bit so that we don't confuse checks in
2208 * __journal_file_buffer
2210 clear_buffer_dirty(bh);
2211 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2214 JBUFFER_TRACE(jh, "on running transaction");
2215 __jbd2_journal_unfile_buffer(jh);
2216 jbd2_journal_put_journal_head(jh);
2222 * jbd2_journal_invalidatepage
2224 * This code is tricky. It has a number of cases to deal with.
2226 * There are two invariants which this code relies on:
2228 * i_size must be updated on disk before we start calling invalidatepage on the
2231 * This is done in ext3 by defining an ext3_setattr method which
2232 * updates i_size before truncate gets going. By maintaining this
2233 * invariant, we can be sure that it is safe to throw away any buffers
2234 * attached to the current transaction: once the transaction commits,
2235 * we know that the data will not be needed.
2237 * Note however that we can *not* throw away data belonging to the
2238 * previous, committing transaction!
2240 * Any disk blocks which *are* part of the previous, committing
2241 * transaction (and which therefore cannot be discarded immediately) are
2242 * not going to be reused in the new running transaction
2244 * The bitmap committed_data images guarantee this: any block which is
2245 * allocated in one transaction and removed in the next will be marked
2246 * as in-use in the committed_data bitmap, so cannot be reused until
2247 * the next transaction to delete the block commits. This means that
2248 * leaving committing buffers dirty is quite safe: the disk blocks
2249 * cannot be reallocated to a different file and so buffer aliasing is
2253 * The above applies mainly to ordered data mode. In writeback mode we
2254 * don't make guarantees about the order in which data hits disk --- in
2255 * particular we don't guarantee that new dirty data is flushed before
2256 * transaction commit --- so it is always safe just to discard data
2257 * immediately in that mode. --sct
2261 * The journal_unmap_buffer helper function returns zero if the buffer
2262 * concerned remains pinned as an anonymous buffer belonging to an older
2265 * We're outside-transaction here. Either or both of j_running_transaction
2266 * and j_committing_transaction may be NULL.
2268 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2271 transaction_t *transaction;
2272 struct journal_head *jh;
2275 BUFFER_TRACE(bh, "entry");
2278 * It is safe to proceed here without the j_list_lock because the
2279 * buffers cannot be stolen by try_to_free_buffers as long as we are
2280 * holding the page lock. --sct
2283 jh = jbd2_journal_grab_journal_head(bh);
2285 goto zap_buffer_unlocked;
2287 /* OK, we have data buffer in journaled mode */
2288 write_lock(&journal->j_state_lock);
2289 spin_lock(&jh->b_state_lock);
2290 spin_lock(&journal->j_list_lock);
2293 * We cannot remove the buffer from checkpoint lists until the
2294 * transaction adding inode to orphan list (let's call it T)
2295 * is committed. Otherwise if the transaction changing the
2296 * buffer would be cleaned from the journal before T is
2297 * committed, a crash will cause that the correct contents of
2298 * the buffer will be lost. On the other hand we have to
2299 * clear the buffer dirty bit at latest at the moment when the
2300 * transaction marking the buffer as freed in the filesystem
2301 * structures is committed because from that moment on the
2302 * block can be reallocated and used by a different page.
2303 * Since the block hasn't been freed yet but the inode has
2304 * already been added to orphan list, it is safe for us to add
2305 * the buffer to BJ_Forget list of the newest transaction.
2307 * Also we have to clear buffer_mapped flag of a truncated buffer
2308 * because the buffer_head may be attached to the page straddling
2309 * i_size (can happen only when blocksize < pagesize) and thus the
2310 * buffer_head can be reused when the file is extended again. So we end
2311 * up keeping around invalidated buffers attached to transactions'
2312 * BJ_Forget list just to stop checkpointing code from cleaning up
2313 * the transaction this buffer was modified in.
2315 transaction = jh->b_transaction;
2316 if (transaction == NULL) {
2317 /* First case: not on any transaction. If it
2318 * has no checkpoint link, then we can zap it:
2319 * it's a writeback-mode buffer so we don't care
2320 * if it hits disk safely. */
2321 if (!jh->b_cp_transaction) {
2322 JBUFFER_TRACE(jh, "not on any transaction: zap");
2326 if (!buffer_dirty(bh)) {
2327 /* bdflush has written it. We can drop it now */
2328 __jbd2_journal_remove_checkpoint(jh);
2332 /* OK, it must be in the journal but still not
2333 * written fully to disk: it's metadata or
2334 * journaled data... */
2336 if (journal->j_running_transaction) {
2337 /* ... and once the current transaction has
2338 * committed, the buffer won't be needed any
2340 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2341 may_free = __dispose_buffer(jh,
2342 journal->j_running_transaction);
2345 /* There is no currently-running transaction. So the
2346 * orphan record which we wrote for this file must have
2347 * passed into commit. We must attach this buffer to
2348 * the committing transaction, if it exists. */
2349 if (journal->j_committing_transaction) {
2350 JBUFFER_TRACE(jh, "give to committing trans");
2351 may_free = __dispose_buffer(jh,
2352 journal->j_committing_transaction);
2355 /* The orphan record's transaction has
2356 * committed. We can cleanse this buffer */
2357 clear_buffer_jbddirty(bh);
2358 __jbd2_journal_remove_checkpoint(jh);
2362 } else if (transaction == journal->j_committing_transaction) {
2363 JBUFFER_TRACE(jh, "on committing transaction");
2365 * The buffer is committing, we simply cannot touch
2366 * it. If the page is straddling i_size we have to wait
2367 * for commit and try again.
2370 spin_unlock(&journal->j_list_lock);
2371 spin_unlock(&jh->b_state_lock);
2372 write_unlock(&journal->j_state_lock);
2373 jbd2_journal_put_journal_head(jh);
2377 * OK, buffer won't be reachable after truncate. We just clear
2378 * b_modified to not confuse transaction credit accounting, and
2379 * set j_next_transaction to the running transaction (if there
2380 * is one) and mark buffer as freed so that commit code knows
2381 * it should clear dirty bits when it is done with the buffer.
2383 set_buffer_freed(bh);
2384 if (journal->j_running_transaction && buffer_jbddirty(bh))
2385 jh->b_next_transaction = journal->j_running_transaction;
2387 spin_unlock(&journal->j_list_lock);
2388 spin_unlock(&jh->b_state_lock);
2389 write_unlock(&journal->j_state_lock);
2390 jbd2_journal_put_journal_head(jh);
2393 /* Good, the buffer belongs to the running transaction.
2394 * We are writing our own transaction's data, not any
2395 * previous one's, so it is safe to throw it away
2396 * (remember that we expect the filesystem to have set
2397 * i_size already for this truncate so recovery will not
2398 * expose the disk blocks we are discarding here.) */
2399 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2400 JBUFFER_TRACE(jh, "on running transaction");
2401 may_free = __dispose_buffer(jh, transaction);
2406 * This is tricky. Although the buffer is truncated, it may be reused
2407 * if blocksize < pagesize and it is attached to the page straddling
2408 * EOF. Since the buffer might have been added to BJ_Forget list of the
2409 * running transaction, journal_get_write_access() won't clear
2410 * b_modified and credit accounting gets confused. So clear b_modified
2414 spin_unlock(&journal->j_list_lock);
2415 spin_unlock(&jh->b_state_lock);
2416 write_unlock(&journal->j_state_lock);
2417 jbd2_journal_put_journal_head(jh);
2418 zap_buffer_unlocked:
2419 clear_buffer_dirty(bh);
2420 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2421 clear_buffer_mapped(bh);
2422 clear_buffer_req(bh);
2423 clear_buffer_new(bh);
2424 clear_buffer_delay(bh);
2425 clear_buffer_unwritten(bh);
2431 * jbd2_journal_invalidatepage()
2432 * @journal: journal to use for flush...
2433 * @page: page to flush
2434 * @offset: start of the range to invalidate
2435 * @length: length of the range to invalidate
2437 * Reap page buffers containing data after in the specified range in page.
2438 * Can return -EBUSY if buffers are part of the committing transaction and
2439 * the page is straddling i_size. Caller then has to wait for current commit
2442 int jbd2_journal_invalidatepage(journal_t *journal,
2444 unsigned int offset,
2445 unsigned int length)
2447 struct buffer_head *head, *bh, *next;
2448 unsigned int stop = offset + length;
2449 unsigned int curr_off = 0;
2450 int partial_page = (offset || length < PAGE_SIZE);
2454 if (!PageLocked(page))
2456 if (!page_has_buffers(page))
2459 BUG_ON(stop > PAGE_SIZE || stop < length);
2461 /* We will potentially be playing with lists other than just the
2462 * data lists (especially for journaled data mode), so be
2463 * cautious in our locking. */
2465 head = bh = page_buffers(page);
2467 unsigned int next_off = curr_off + bh->b_size;
2468 next = bh->b_this_page;
2470 if (next_off > stop)
2473 if (offset <= curr_off) {
2474 /* This block is wholly outside the truncation point */
2476 ret = journal_unmap_buffer(journal, bh, partial_page);
2482 curr_off = next_off;
2485 } while (bh != head);
2487 if (!partial_page) {
2488 if (may_free && try_to_free_buffers(page))
2489 J_ASSERT(!page_has_buffers(page));
2495 * File a buffer on the given transaction list.
2497 void __jbd2_journal_file_buffer(struct journal_head *jh,
2498 transaction_t *transaction, int jlist)
2500 struct journal_head **list = NULL;
2502 struct buffer_head *bh = jh2bh(jh);
2504 lockdep_assert_held(&jh->b_state_lock);
2505 assert_spin_locked(&transaction->t_journal->j_list_lock);
2507 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2508 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2509 jh->b_transaction == NULL);
2511 if (jh->b_transaction && jh->b_jlist == jlist)
2514 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2515 jlist == BJ_Shadow || jlist == BJ_Forget) {
2517 * For metadata buffers, we track dirty bit in buffer_jbddirty
2518 * instead of buffer_dirty. We should not see a dirty bit set
2519 * here because we clear it in do_get_write_access but e.g.
2520 * tune2fs can modify the sb and set the dirty bit at any time
2521 * so we try to gracefully handle that.
2523 if (buffer_dirty(bh))
2524 warn_dirty_buffer(bh);
2525 if (test_clear_buffer_dirty(bh) ||
2526 test_clear_buffer_jbddirty(bh))
2530 if (jh->b_transaction)
2531 __jbd2_journal_temp_unlink_buffer(jh);
2533 jbd2_journal_grab_journal_head(bh);
2534 jh->b_transaction = transaction;
2538 J_ASSERT_JH(jh, !jh->b_committed_data);
2539 J_ASSERT_JH(jh, !jh->b_frozen_data);
2542 transaction->t_nr_buffers++;
2543 list = &transaction->t_buffers;
2546 list = &transaction->t_forget;
2549 list = &transaction->t_shadow_list;
2552 list = &transaction->t_reserved_list;
2556 __blist_add_buffer(list, jh);
2557 jh->b_jlist = jlist;
2560 set_buffer_jbddirty(bh);
2563 void jbd2_journal_file_buffer(struct journal_head *jh,
2564 transaction_t *transaction, int jlist)
2566 spin_lock(&jh->b_state_lock);
2567 spin_lock(&transaction->t_journal->j_list_lock);
2568 __jbd2_journal_file_buffer(jh, transaction, jlist);
2569 spin_unlock(&transaction->t_journal->j_list_lock);
2570 spin_unlock(&jh->b_state_lock);
2574 * Remove a buffer from its current buffer list in preparation for
2575 * dropping it from its current transaction entirely. If the buffer has
2576 * already started to be used by a subsequent transaction, refile the
2577 * buffer on that transaction's metadata list.
2579 * Called under j_list_lock
2580 * Called under jh->b_state_lock
2582 * When this function returns true, there's no next transaction to refile to
2583 * and the caller has to drop jh reference through
2584 * jbd2_journal_put_journal_head().
2586 bool __jbd2_journal_refile_buffer(struct journal_head *jh)
2588 int was_dirty, jlist;
2589 struct buffer_head *bh = jh2bh(jh);
2591 lockdep_assert_held(&jh->b_state_lock);
2592 if (jh->b_transaction)
2593 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2595 /* If the buffer is now unused, just drop it. */
2596 if (jh->b_next_transaction == NULL) {
2597 __jbd2_journal_unfile_buffer(jh);
2602 * It has been modified by a later transaction: add it to the new
2603 * transaction's metadata list.
2606 was_dirty = test_clear_buffer_jbddirty(bh);
2607 __jbd2_journal_temp_unlink_buffer(jh);
2610 * b_transaction must be set, otherwise the new b_transaction won't
2611 * be holding jh reference
2613 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2616 * We set b_transaction here because b_next_transaction will inherit
2617 * our jh reference and thus __jbd2_journal_file_buffer() must not
2620 WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2621 WRITE_ONCE(jh->b_next_transaction, NULL);
2622 if (buffer_freed(bh))
2624 else if (jh->b_modified)
2625 jlist = BJ_Metadata;
2627 jlist = BJ_Reserved;
2628 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2629 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2632 set_buffer_jbddirty(bh);
2637 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2638 * bh reference so that we can safely unlock bh.
2640 * The jh and bh may be freed by this call.
2642 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2646 spin_lock(&jh->b_state_lock);
2647 spin_lock(&journal->j_list_lock);
2648 drop = __jbd2_journal_refile_buffer(jh);
2649 spin_unlock(&jh->b_state_lock);
2650 spin_unlock(&journal->j_list_lock);
2652 jbd2_journal_put_journal_head(jh);
2656 * File inode in the inode list of the handle's transaction
2658 static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2659 unsigned long flags, loff_t start_byte, loff_t end_byte)
2661 transaction_t *transaction = handle->h_transaction;
2664 if (is_handle_aborted(handle))
2666 journal = transaction->t_journal;
2668 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2669 transaction->t_tid);
2671 spin_lock(&journal->j_list_lock);
2672 jinode->i_flags |= flags;
2674 if (jinode->i_dirty_end) {
2675 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2676 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2678 jinode->i_dirty_start = start_byte;
2679 jinode->i_dirty_end = end_byte;
2682 /* Is inode already attached where we need it? */
2683 if (jinode->i_transaction == transaction ||
2684 jinode->i_next_transaction == transaction)
2688 * We only ever set this variable to 1 so the test is safe. Since
2689 * t_need_data_flush is likely to be set, we do the test to save some
2690 * cacheline bouncing
2692 if (!transaction->t_need_data_flush)
2693 transaction->t_need_data_flush = 1;
2694 /* On some different transaction's list - should be
2695 * the committing one */
2696 if (jinode->i_transaction) {
2697 J_ASSERT(jinode->i_next_transaction == NULL);
2698 J_ASSERT(jinode->i_transaction ==
2699 journal->j_committing_transaction);
2700 jinode->i_next_transaction = transaction;
2703 /* Not on any transaction list... */
2704 J_ASSERT(!jinode->i_next_transaction);
2705 jinode->i_transaction = transaction;
2706 list_add(&jinode->i_list, &transaction->t_inode_list);
2708 spin_unlock(&journal->j_list_lock);
2713 int jbd2_journal_inode_ranged_write(handle_t *handle,
2714 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2716 return jbd2_journal_file_inode(handle, jinode,
2717 JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2718 start_byte + length - 1);
2721 int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2722 loff_t start_byte, loff_t length)
2724 return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2725 start_byte, start_byte + length - 1);
2729 * File truncate and transaction commit interact with each other in a
2730 * non-trivial way. If a transaction writing data block A is
2731 * committing, we cannot discard the data by truncate until we have
2732 * written them. Otherwise if we crashed after the transaction with
2733 * write has committed but before the transaction with truncate has
2734 * committed, we could see stale data in block A. This function is a
2735 * helper to solve this problem. It starts writeout of the truncated
2736 * part in case it is in the committing transaction.
2738 * Filesystem code must call this function when inode is journaled in
2739 * ordered mode before truncation happens and after the inode has been
2740 * placed on orphan list with the new inode size. The second condition
2741 * avoids the race that someone writes new data and we start
2742 * committing the transaction after this function has been called but
2743 * before a transaction for truncate is started (and furthermore it
2744 * allows us to optimize the case where the addition to orphan list
2745 * happens in the same transaction as write --- we don't have to write
2746 * any data in such case).
2748 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2749 struct jbd2_inode *jinode,
2752 transaction_t *inode_trans, *commit_trans;
2755 /* This is a quick check to avoid locking if not necessary */
2756 if (!jinode->i_transaction)
2758 /* Locks are here just to force reading of recent values, it is
2759 * enough that the transaction was not committing before we started
2760 * a transaction adding the inode to orphan list */
2761 read_lock(&journal->j_state_lock);
2762 commit_trans = journal->j_committing_transaction;
2763 read_unlock(&journal->j_state_lock);
2764 spin_lock(&journal->j_list_lock);
2765 inode_trans = jinode->i_transaction;
2766 spin_unlock(&journal->j_list_lock);
2767 if (inode_trans == commit_trans) {
2768 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2769 new_size, LLONG_MAX);
2771 jbd2_journal_abort(journal, ret);