1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Oracle. All rights reserved.
6 #include <linux/sched.h>
7 #include <linux/pagemap.h>
8 #include <linux/spinlock.h>
9 #include <linux/page-flags.h>
13 #include "extent_io.h"
17 * Extent buffer locking
18 * =====================
20 * The locks use a custom scheme that allows to do more operations than are
21 * available fromt current locking primitives. The building blocks are still
22 * rwlock and wait queues.
26 * - reader/writer exclusion
27 * - writer/writer exclusion
28 * - reader/reader sharing
29 * - spinning lock semantics
30 * - blocking lock semantics
31 * - try-lock semantics for readers and writers
32 * - one level nesting, allowing read lock to be taken by the same thread that
33 * already has write lock
35 * The extent buffer locks (also called tree locks) manage access to eb data
36 * related to the storage in the b-tree (keys, items, but not the individual
38 * We want concurrency of many readers and safe updates. The underlying locking
39 * is done by read-write spinlock and the blocking part is implemented using
40 * counters and wait queues.
42 * spinning semantics - the low-level rwlock is held so all other threads that
43 * want to take it are spinning on it.
45 * blocking semantics - the low-level rwlock is not held but the counter
46 * denotes how many times the blocking lock was held;
47 * sleeping is possible
49 * Write lock always allows only one thread to access the data.
55 * There are additional state counters that are asserted in various contexts,
56 * removed from non-debug build to reduce extent_buffer size and for
57 * performance reasons.
63 * A write operation on a tree might indirectly start a look up on the same
64 * tree. This can happen when btrfs_cow_block locks the tree and needs to
65 * lookup free extents.
69 * alloc_tree_block_no_bg_flush
70 * btrfs_alloc_tree_block
71 * btrfs_reserve_extent
73 * load_free_space_cache
75 * btrfs_lookup_file_extent
79 * Locking pattern - spinning
80 * --------------------------
82 * The simple locking scenario, the +--+ denotes the spinning section.
85 * | - extent_buffer::rwlock is held
86 * | - no heavy operations should happen, eg. IO, memory allocations, large
87 * | structure traversals
91 * Locking pattern - blocking
92 * --------------------------
94 * The blocking write uses the following scheme. The +--+ denotes the spinning
99 * +- btrfs_set_lock_blocking_write
101 * - allowed: IO, memory allocations, etc.
103 * -- btrfs_tree_unlock - note, no explicit unblocking necessary
106 * Blocking read is similar.
108 * +- btrfs_tree_read_lock
110 * +- btrfs_set_lock_blocking_read
112 * - heavy operations allowed
114 * +- btrfs_tree_read_unlock_blocking
116 * +- btrfs_tree_read_unlock
120 #ifdef CONFIG_BTRFS_DEBUG
121 static inline void btrfs_assert_spinning_writers_get(struct extent_buffer *eb)
123 WARN_ON(eb->spinning_writers);
124 eb->spinning_writers++;
127 static inline void btrfs_assert_spinning_writers_put(struct extent_buffer *eb)
129 WARN_ON(eb->spinning_writers != 1);
130 eb->spinning_writers--;
133 static inline void btrfs_assert_no_spinning_writers(struct extent_buffer *eb)
135 WARN_ON(eb->spinning_writers);
138 static inline void btrfs_assert_spinning_readers_get(struct extent_buffer *eb)
140 atomic_inc(&eb->spinning_readers);
143 static inline void btrfs_assert_spinning_readers_put(struct extent_buffer *eb)
145 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
146 atomic_dec(&eb->spinning_readers);
149 static inline void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb)
151 atomic_inc(&eb->read_locks);
154 static inline void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb)
156 atomic_dec(&eb->read_locks);
159 static inline void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
161 BUG_ON(!atomic_read(&eb->read_locks));
164 static inline void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb)
169 static inline void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb)
175 static void btrfs_assert_spinning_writers_get(struct extent_buffer *eb) { }
176 static void btrfs_assert_spinning_writers_put(struct extent_buffer *eb) { }
177 static void btrfs_assert_no_spinning_writers(struct extent_buffer *eb) { }
178 static void btrfs_assert_spinning_readers_put(struct extent_buffer *eb) { }
179 static void btrfs_assert_spinning_readers_get(struct extent_buffer *eb) { }
180 static void btrfs_assert_tree_read_locked(struct extent_buffer *eb) { }
181 static void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb) { }
182 static void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb) { }
183 static void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb) { }
184 static void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb) { }
188 * Mark already held read lock as blocking. Can be nested in write lock by the
191 * Use when there are potentially long operations ahead so other thread waiting
192 * on the lock will not actively spin but sleep instead.
194 * The rwlock is released and blocking reader counter is increased.
196 void btrfs_set_lock_blocking_read(struct extent_buffer *eb)
198 trace_btrfs_set_lock_blocking_read(eb);
200 * No lock is required. The lock owner may change if we have a read
201 * lock, but it won't change to or away from us. If we have the write
202 * lock, we are the owner and it'll never change.
204 if (eb->lock_recursed && current->pid == eb->lock_owner)
206 btrfs_assert_tree_read_locked(eb);
207 atomic_inc(&eb->blocking_readers);
208 btrfs_assert_spinning_readers_put(eb);
209 read_unlock(&eb->lock);
213 * Mark already held write lock as blocking.
215 * Use when there are potentially long operations ahead so other threads
216 * waiting on the lock will not actively spin but sleep instead.
218 * The rwlock is released and blocking writers is set.
220 void btrfs_set_lock_blocking_write(struct extent_buffer *eb)
222 trace_btrfs_set_lock_blocking_write(eb);
224 * No lock is required. The lock owner may change if we have a read
225 * lock, but it won't change to or away from us. If we have the write
226 * lock, we are the owner and it'll never change.
228 if (eb->lock_recursed && current->pid == eb->lock_owner)
230 if (eb->blocking_writers == 0) {
231 btrfs_assert_spinning_writers_put(eb);
232 btrfs_assert_tree_locked(eb);
233 WRITE_ONCE(eb->blocking_writers, 1);
234 write_unlock(&eb->lock);
239 * Lock the extent buffer for read. Wait for any writers (spinning or blocking).
240 * Can be nested in write lock by the same thread.
242 * Use when the locked section does only lightweight actions and busy waiting
243 * would be cheaper than making other threads do the wait/wake loop.
245 * The rwlock is held upon exit.
247 void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest,
252 if (trace_btrfs_tree_read_lock_enabled())
253 start_ns = ktime_get_ns();
255 read_lock(&eb->lock);
256 BUG_ON(eb->blocking_writers == 0 &&
257 current->pid == eb->lock_owner);
258 if (eb->blocking_writers) {
259 if (current->pid == eb->lock_owner) {
261 * This extent is already write-locked by our thread.
262 * We allow an additional read lock to be added because
263 * it's for the same thread. btrfs_find_all_roots()
264 * depends on this as it may be called on a partly
265 * (write-)locked tree.
268 BUG_ON(eb->lock_recursed);
269 eb->lock_recursed = true;
270 read_unlock(&eb->lock);
271 trace_btrfs_tree_read_lock(eb, start_ns);
274 read_unlock(&eb->lock);
275 wait_event(eb->write_lock_wq,
276 READ_ONCE(eb->blocking_writers) == 0);
279 btrfs_assert_tree_read_locks_get(eb);
280 btrfs_assert_spinning_readers_get(eb);
281 trace_btrfs_tree_read_lock(eb, start_ns);
284 void btrfs_tree_read_lock(struct extent_buffer *eb)
286 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, false);
290 * Lock extent buffer for read, optimistically expecting that there are no
291 * contending blocking writers. If there are, don't wait.
293 * Return 1 if the rwlock has been taken, 0 otherwise
295 int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
297 if (READ_ONCE(eb->blocking_writers))
300 read_lock(&eb->lock);
301 /* Refetch value after lock */
302 if (READ_ONCE(eb->blocking_writers)) {
303 read_unlock(&eb->lock);
306 btrfs_assert_tree_read_locks_get(eb);
307 btrfs_assert_spinning_readers_get(eb);
308 trace_btrfs_tree_read_lock_atomic(eb);
313 * Try-lock for read. Don't block or wait for contending writers.
315 * Retrun 1 if the rwlock has been taken, 0 otherwise
317 int btrfs_try_tree_read_lock(struct extent_buffer *eb)
319 if (READ_ONCE(eb->blocking_writers))
322 if (!read_trylock(&eb->lock))
325 /* Refetch value after lock */
326 if (READ_ONCE(eb->blocking_writers)) {
327 read_unlock(&eb->lock);
330 btrfs_assert_tree_read_locks_get(eb);
331 btrfs_assert_spinning_readers_get(eb);
332 trace_btrfs_try_tree_read_lock(eb);
337 * Try-lock for write. May block until the lock is uncontended, but does not
338 * wait until it is free.
340 * Retrun 1 if the rwlock has been taken, 0 otherwise
342 int btrfs_try_tree_write_lock(struct extent_buffer *eb)
344 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers))
347 write_lock(&eb->lock);
348 /* Refetch value after lock */
349 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers)) {
350 write_unlock(&eb->lock);
353 btrfs_assert_tree_write_locks_get(eb);
354 btrfs_assert_spinning_writers_get(eb);
355 eb->lock_owner = current->pid;
356 trace_btrfs_try_tree_write_lock(eb);
361 * Release read lock. Must be used only if the lock is in spinning mode. If
362 * the read lock is nested, must pair with read lock before the write unlock.
364 * The rwlock is not held upon exit.
366 void btrfs_tree_read_unlock(struct extent_buffer *eb)
368 trace_btrfs_tree_read_unlock(eb);
370 * if we're nested, we have the write lock. No new locking
371 * is needed as long as we are the lock owner.
372 * The write unlock will do a barrier for us, and the lock_recursed
373 * field only matters to the lock owner.
375 if (eb->lock_recursed && current->pid == eb->lock_owner) {
376 eb->lock_recursed = false;
379 btrfs_assert_tree_read_locked(eb);
380 btrfs_assert_spinning_readers_put(eb);
381 btrfs_assert_tree_read_locks_put(eb);
382 read_unlock(&eb->lock);
386 * Release read lock, previously set to blocking by a pairing call to
387 * btrfs_set_lock_blocking_read(). Can be nested in write lock by the same
390 * State of rwlock is unchanged, last reader wakes waiting threads.
392 void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
394 trace_btrfs_tree_read_unlock_blocking(eb);
396 * if we're nested, we have the write lock. No new locking
397 * is needed as long as we are the lock owner.
398 * The write unlock will do a barrier for us, and the lock_recursed
399 * field only matters to the lock owner.
401 if (eb->lock_recursed && current->pid == eb->lock_owner) {
402 eb->lock_recursed = false;
405 btrfs_assert_tree_read_locked(eb);
406 WARN_ON(atomic_read(&eb->blocking_readers) == 0);
407 /* atomic_dec_and_test implies a barrier */
408 if (atomic_dec_and_test(&eb->blocking_readers))
409 cond_wake_up_nomb(&eb->read_lock_wq);
410 btrfs_assert_tree_read_locks_put(eb);
414 * Lock for write. Wait for all blocking and spinning readers and writers. This
415 * starts context where reader lock could be nested by the same thread.
417 * The rwlock is held for write upon exit.
419 void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
420 __acquires(&eb->lock)
424 if (trace_btrfs_tree_lock_enabled())
425 start_ns = ktime_get_ns();
427 WARN_ON(eb->lock_owner == current->pid);
429 wait_event(eb->read_lock_wq, atomic_read(&eb->blocking_readers) == 0);
430 wait_event(eb->write_lock_wq, READ_ONCE(eb->blocking_writers) == 0);
431 write_lock(&eb->lock);
432 /* Refetch value after lock */
433 if (atomic_read(&eb->blocking_readers) ||
434 READ_ONCE(eb->blocking_writers)) {
435 write_unlock(&eb->lock);
438 btrfs_assert_spinning_writers_get(eb);
439 btrfs_assert_tree_write_locks_get(eb);
440 eb->lock_owner = current->pid;
441 trace_btrfs_tree_lock(eb, start_ns);
444 void btrfs_tree_lock(struct extent_buffer *eb)
446 __btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
450 * Release the write lock, either blocking or spinning (ie. there's no need
451 * for an explicit blocking unlock, like btrfs_tree_read_unlock_blocking).
452 * This also ends the context for nesting, the read lock must have been
455 * Tasks blocked and waiting are woken, rwlock is not held upon exit.
457 void btrfs_tree_unlock(struct extent_buffer *eb)
460 * This is read both locked and unlocked but always by the same thread
461 * that already owns the lock so we don't need to use READ_ONCE
463 int blockers = eb->blocking_writers;
465 BUG_ON(blockers > 1);
467 btrfs_assert_tree_locked(eb);
468 trace_btrfs_tree_unlock(eb);
470 btrfs_assert_tree_write_locks_put(eb);
473 btrfs_assert_no_spinning_writers(eb);
475 WRITE_ONCE(eb->blocking_writers, 0);
477 * We need to order modifying blocking_writers above with
478 * actually waking up the sleepers to ensure they see the
479 * updated value of blocking_writers
481 cond_wake_up(&eb->write_lock_wq);
483 btrfs_assert_spinning_writers_put(eb);
484 write_unlock(&eb->lock);
489 * Set all locked nodes in the path to blocking locks. This should be done
492 void btrfs_set_path_blocking(struct btrfs_path *p)
496 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
497 if (!p->nodes[i] || !p->locks[i])
500 * If we currently have a spinning reader or writer lock this
501 * will bump the count of blocking holders and drop the
504 if (p->locks[i] == BTRFS_READ_LOCK) {
505 btrfs_set_lock_blocking_read(p->nodes[i]);
506 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
507 } else if (p->locks[i] == BTRFS_WRITE_LOCK) {
508 btrfs_set_lock_blocking_write(p->nodes[i]);
509 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
515 * This releases any locks held in the path starting at level and going all the
516 * way up to the root.
518 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
519 * cases, such as COW of the block at slot zero in the node. This ignores
520 * those rules, and it should only be called when there are no more updates to
521 * be done higher up in the tree.
523 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
527 if (path->keep_locks)
530 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
535 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
541 * Loop around taking references on and locking the root node of the tree until
542 * we end up with a lock on the root node.
544 * Return: root extent buffer with write lock held
546 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
548 struct extent_buffer *eb;
551 eb = btrfs_root_node(root);
553 if (eb == root->node)
555 btrfs_tree_unlock(eb);
556 free_extent_buffer(eb);
562 * Loop around taking references on and locking the root node of the tree until
563 * we end up with a lock on the root node.
565 * Return: root extent buffer with read lock held
567 struct extent_buffer *__btrfs_read_lock_root_node(struct btrfs_root *root,
570 struct extent_buffer *eb;
573 eb = btrfs_root_node(root);
574 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, recurse);
575 if (eb == root->node)
577 btrfs_tree_read_unlock(eb);
578 free_extent_buffer(eb);
587 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
588 * where you want to provide A-B exclusion but not AA or BB.
590 * Currently implementation gives more priority to reader. If a reader and a
591 * writer both race to acquire their respective sides of the lock the writer
592 * would yield its lock as soon as it detects a concurrent reader. Additionally
593 * if there are pending readers no new writers would be allowed to come in and
597 int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
601 ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
605 atomic_set(&lock->readers, 0);
606 init_waitqueue_head(&lock->pending_readers);
607 init_waitqueue_head(&lock->pending_writers);
612 void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
614 percpu_counter_destroy(&lock->writers);
617 /* Return true if acquisition is successful, false otherwise */
618 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
620 if (atomic_read(&lock->readers))
623 percpu_counter_inc(&lock->writers);
625 /* Ensure writers count is updated before we check for pending readers */
627 if (atomic_read(&lock->readers)) {
628 btrfs_drew_write_unlock(lock);
635 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
638 if (btrfs_drew_try_write_lock(lock))
640 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
644 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
646 percpu_counter_dec(&lock->writers);
647 cond_wake_up(&lock->pending_readers);
650 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
652 atomic_inc(&lock->readers);
655 * Ensure the pending reader count is perceieved BEFORE this reader
656 * goes to sleep in case of active writers. This guarantees new writers
657 * won't be allowed and that the current reader will be woken up when
658 * the last active writer finishes its jobs.
660 smp_mb__after_atomic();
662 wait_event(lock->pending_readers,
663 percpu_counter_sum(&lock->writers) == 0);
666 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
669 * atomic_dec_and_test implies a full barrier, so woken up writers
670 * are guaranteed to see the decrement
672 if (atomic_dec_and_test(&lock->readers))
673 wake_up(&lock->pending_writers);