xref: /openbmc/linux/fs/btrfs/locking.c (revision e31ac898)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2008 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include <linux/pagemap.h>
8 #include <linux/spinlock.h>
9 #include <linux/page-flags.h>
10 #include <asm/bug.h>
11 #include "misc.h"
12 #include "ctree.h"
13 #include "extent_io.h"
14 #include "locking.h"
15 
16 /*
17  * Extent buffer locking
18  * =====================
19  *
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.
23  *
24  * Required semantics:
25  *
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
34  *
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
37  * members of eb).
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.
41  *
42  * spinning semantics - the low-level rwlock is held so all other threads that
43  *                      want to take it are spinning on it.
44  *
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
48  *
49  * Write lock always allows only one thread to access the data.
50  *
51  *
52  * Debugging
53  * ---------
54  *
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.
58  *
59  *
60  * Lock recursion
61  * --------------
62  *
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.
66  *
67  * btrfs_cow_block
68  *   ..
69  *   alloc_tree_block_no_bg_flush
70  *     btrfs_alloc_tree_block
71  *       btrfs_reserve_extent
72  *         ..
73  *         load_free_space_cache
74  *           ..
75  *           btrfs_lookup_file_extent
76  *             btrfs_search_slot
77  *
78  *
79  * Locking pattern - spinning
80  * --------------------------
81  *
82  * The simple locking scenario, the +--+ denotes the spinning section.
83  *
84  * +- btrfs_tree_lock
85  * | - extent_buffer::rwlock is held
86  * | - no heavy operations should happen, eg. IO, memory allocations, large
87  * |   structure traversals
88  * +- btrfs_tree_unock
89 *
90 *
91  * Locking pattern - blocking
92  * --------------------------
93  *
94  * The blocking write uses the following scheme.  The +--+ denotes the spinning
95  * section.
96  *
97  * +- btrfs_tree_lock
98  * |
99  * +- btrfs_set_lock_blocking_write
100  *
101  *   - allowed: IO, memory allocations, etc.
102  *
103  * -- btrfs_tree_unlock - note, no explicit unblocking necessary
104  *
105  *
106  * Blocking read is similar.
107  *
108  * +- btrfs_tree_read_lock
109  * |
110  * +- btrfs_set_lock_blocking_read
111  *
112  *  - heavy operations allowed
113  *
114  * +- btrfs_tree_read_unlock_blocking
115  * |
116  * +- btrfs_tree_read_unlock
117  *
118  */
119 
120 #ifdef CONFIG_BTRFS_DEBUG
121 static inline void btrfs_assert_spinning_writers_get(struct extent_buffer *eb)
122 {
123 	WARN_ON(eb->spinning_writers);
124 	eb->spinning_writers++;
125 }
126 
127 static inline void btrfs_assert_spinning_writers_put(struct extent_buffer *eb)
128 {
129 	WARN_ON(eb->spinning_writers != 1);
130 	eb->spinning_writers--;
131 }
132 
133 static inline void btrfs_assert_no_spinning_writers(struct extent_buffer *eb)
134 {
135 	WARN_ON(eb->spinning_writers);
136 }
137 
138 static inline void btrfs_assert_spinning_readers_get(struct extent_buffer *eb)
139 {
140 	atomic_inc(&eb->spinning_readers);
141 }
142 
143 static inline void btrfs_assert_spinning_readers_put(struct extent_buffer *eb)
144 {
145 	WARN_ON(atomic_read(&eb->spinning_readers) == 0);
146 	atomic_dec(&eb->spinning_readers);
147 }
148 
149 static inline void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb)
150 {
151 	atomic_inc(&eb->read_locks);
152 }
153 
154 static inline void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb)
155 {
156 	atomic_dec(&eb->read_locks);
157 }
158 
159 static inline void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
160 {
161 	BUG_ON(!atomic_read(&eb->read_locks));
162 }
163 
164 static inline void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb)
165 {
166 	eb->write_locks++;
167 }
168 
169 static inline void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb)
170 {
171 	eb->write_locks--;
172 }
173 
174 #else
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) { }
185 #endif
186 
187 /*
188  * Mark already held read lock as blocking. Can be nested in write lock by the
189  * same thread.
190  *
191  * Use when there are potentially long operations ahead so other thread waiting
192  * on the lock will not actively spin but sleep instead.
193  *
194  * The rwlock is released and blocking reader counter is increased.
195  */
196 void btrfs_set_lock_blocking_read(struct extent_buffer *eb)
197 {
198 	trace_btrfs_set_lock_blocking_read(eb);
199 	/*
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.
203 	 */
204 	if (eb->lock_recursed && current->pid == eb->lock_owner)
205 		return;
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);
210 }
211 
212 /*
213  * Mark already held write lock as blocking.
214  *
215  * Use when there are potentially long operations ahead so other threads
216  * waiting on the lock will not actively spin but sleep instead.
217  *
218  * The rwlock is released and blocking writers is set.
219  */
220 void btrfs_set_lock_blocking_write(struct extent_buffer *eb)
221 {
222 	trace_btrfs_set_lock_blocking_write(eb);
223 	/*
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.
227 	 */
228 	if (eb->lock_recursed && current->pid == eb->lock_owner)
229 		return;
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);
235 	}
236 }
237 
238 /*
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.
241  *
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.
244  *
245  * The rwlock is held upon exit.
246  */
247 void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest,
248 			    bool recurse)
249 {
250 	u64 start_ns = 0;
251 
252 	if (trace_btrfs_tree_read_lock_enabled())
253 		start_ns = ktime_get_ns();
254 again:
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) {
260 			/*
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.
266 			 */
267 			WARN_ON(!recurse);
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);
272 			return;
273 		}
274 		read_unlock(&eb->lock);
275 		wait_event(eb->write_lock_wq,
276 			   READ_ONCE(eb->blocking_writers) == 0);
277 		goto again;
278 	}
279 	btrfs_assert_tree_read_locks_get(eb);
280 	btrfs_assert_spinning_readers_get(eb);
281 	trace_btrfs_tree_read_lock(eb, start_ns);
282 }
283 
284 void btrfs_tree_read_lock(struct extent_buffer *eb)
285 {
286 	__btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, false);
287 }
288 
289 /*
290  * Lock extent buffer for read, optimistically expecting that there are no
291  * contending blocking writers. If there are, don't wait.
292  *
293  * Return 1 if the rwlock has been taken, 0 otherwise
294  */
295 int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
296 {
297 	if (READ_ONCE(eb->blocking_writers))
298 		return 0;
299 
300 	read_lock(&eb->lock);
301 	/* Refetch value after lock */
302 	if (READ_ONCE(eb->blocking_writers)) {
303 		read_unlock(&eb->lock);
304 		return 0;
305 	}
306 	btrfs_assert_tree_read_locks_get(eb);
307 	btrfs_assert_spinning_readers_get(eb);
308 	trace_btrfs_tree_read_lock_atomic(eb);
309 	return 1;
310 }
311 
312 /*
313  * Try-lock for read. Don't block or wait for contending writers.
314  *
315  * Retrun 1 if the rwlock has been taken, 0 otherwise
316  */
317 int btrfs_try_tree_read_lock(struct extent_buffer *eb)
318 {
319 	if (READ_ONCE(eb->blocking_writers))
320 		return 0;
321 
322 	if (!read_trylock(&eb->lock))
323 		return 0;
324 
325 	/* Refetch value after lock */
326 	if (READ_ONCE(eb->blocking_writers)) {
327 		read_unlock(&eb->lock);
328 		return 0;
329 	}
330 	btrfs_assert_tree_read_locks_get(eb);
331 	btrfs_assert_spinning_readers_get(eb);
332 	trace_btrfs_try_tree_read_lock(eb);
333 	return 1;
334 }
335 
336 /*
337  * Try-lock for write. May block until the lock is uncontended, but does not
338  * wait until it is free.
339  *
340  * Retrun 1 if the rwlock has been taken, 0 otherwise
341  */
342 int btrfs_try_tree_write_lock(struct extent_buffer *eb)
343 {
344 	if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers))
345 		return 0;
346 
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);
351 		return 0;
352 	}
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);
357 	return 1;
358 }
359 
360 /*
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.
363  *
364  * The rwlock is not held upon exit.
365  */
366 void btrfs_tree_read_unlock(struct extent_buffer *eb)
367 {
368 	trace_btrfs_tree_read_unlock(eb);
369 	/*
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.
374 	 */
375 	if (eb->lock_recursed && current->pid == eb->lock_owner) {
376 		eb->lock_recursed = false;
377 		return;
378 	}
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);
383 }
384 
385 /*
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
388  * thread.
389  *
390  * State of rwlock is unchanged, last reader wakes waiting threads.
391  */
392 void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
393 {
394 	trace_btrfs_tree_read_unlock_blocking(eb);
395 	/*
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.
400 	 */
401 	if (eb->lock_recursed && current->pid == eb->lock_owner) {
402 		eb->lock_recursed = false;
403 		return;
404 	}
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);
411 }
412 
413 /*
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.
416  *
417  * The rwlock is held for write upon exit.
418  */
419 void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
420 	__acquires(&eb->lock)
421 {
422 	u64 start_ns = 0;
423 
424 	if (trace_btrfs_tree_lock_enabled())
425 		start_ns = ktime_get_ns();
426 
427 	WARN_ON(eb->lock_owner == current->pid);
428 again:
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);
436 		goto again;
437 	}
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);
442 }
443 
444 void btrfs_tree_lock(struct extent_buffer *eb)
445 {
446 	__btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
447 }
448 
449 /*
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
453  * released already.
454  *
455  * Tasks blocked and waiting are woken, rwlock is not held upon exit.
456  */
457 void btrfs_tree_unlock(struct extent_buffer *eb)
458 {
459 	/*
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
462 	 */
463 	int blockers = eb->blocking_writers;
464 
465 	BUG_ON(blockers > 1);
466 
467 	btrfs_assert_tree_locked(eb);
468 	trace_btrfs_tree_unlock(eb);
469 	eb->lock_owner = 0;
470 	btrfs_assert_tree_write_locks_put(eb);
471 
472 	if (blockers) {
473 		btrfs_assert_no_spinning_writers(eb);
474 		/* Unlocked write */
475 		WRITE_ONCE(eb->blocking_writers, 0);
476 		/*
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
480 		 */
481 		cond_wake_up(&eb->write_lock_wq);
482 	} else {
483 		btrfs_assert_spinning_writers_put(eb);
484 		write_unlock(&eb->lock);
485 	}
486 }
487 
488 /*
489  * Set all locked nodes in the path to blocking locks.  This should be done
490  * before scheduling
491  */
492 void btrfs_set_path_blocking(struct btrfs_path *p)
493 {
494 	int i;
495 
496 	for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
497 		if (!p->nodes[i] || !p->locks[i])
498 			continue;
499 		/*
500 		 * If we currently have a spinning reader or writer lock this
501 		 * will bump the count of blocking holders and drop the
502 		 * spinlock.
503 		 */
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;
510 		}
511 	}
512 }
513 
514 /*
515  * This releases any locks held in the path starting at level and going all the
516  * way up to the root.
517  *
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.
522  */
523 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
524 {
525 	int i;
526 
527 	if (path->keep_locks)
528 		return;
529 
530 	for (i = level; i < BTRFS_MAX_LEVEL; i++) {
531 		if (!path->nodes[i])
532 			continue;
533 		if (!path->locks[i])
534 			continue;
535 		btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
536 		path->locks[i] = 0;
537 	}
538 }
539 
540 /*
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.
543  *
544  * Return: root extent buffer with write lock held
545  */
546 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
547 {
548 	struct extent_buffer *eb;
549 
550 	while (1) {
551 		eb = btrfs_root_node(root);
552 		btrfs_tree_lock(eb);
553 		if (eb == root->node)
554 			break;
555 		btrfs_tree_unlock(eb);
556 		free_extent_buffer(eb);
557 	}
558 	return eb;
559 }
560 
561 /*
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.
564  *
565  * Return: root extent buffer with read lock held
566  */
567 struct extent_buffer *__btrfs_read_lock_root_node(struct btrfs_root *root,
568 						  bool recurse)
569 {
570 	struct extent_buffer *eb;
571 
572 	while (1) {
573 		eb = btrfs_root_node(root);
574 		__btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL, recurse);
575 		if (eb == root->node)
576 			break;
577 		btrfs_tree_read_unlock(eb);
578 		free_extent_buffer(eb);
579 	}
580 	return eb;
581 }
582 
583 /*
584  * DREW locks
585  * ==========
586  *
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.
589  *
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
594  * acquire the lock.
595  */
596 
597 int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
598 {
599 	int ret;
600 
601 	ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
602 	if (ret)
603 		return ret;
604 
605 	atomic_set(&lock->readers, 0);
606 	init_waitqueue_head(&lock->pending_readers);
607 	init_waitqueue_head(&lock->pending_writers);
608 
609 	return 0;
610 }
611 
612 void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
613 {
614 	percpu_counter_destroy(&lock->writers);
615 }
616 
617 /* Return true if acquisition is successful, false otherwise */
618 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
619 {
620 	if (atomic_read(&lock->readers))
621 		return false;
622 
623 	percpu_counter_inc(&lock->writers);
624 
625 	/* Ensure writers count is updated before we check for pending readers */
626 	smp_mb();
627 	if (atomic_read(&lock->readers)) {
628 		btrfs_drew_write_unlock(lock);
629 		return false;
630 	}
631 
632 	return true;
633 }
634 
635 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
636 {
637 	while (true) {
638 		if (btrfs_drew_try_write_lock(lock))
639 			return;
640 		wait_event(lock->pending_writers, !atomic_read(&lock->readers));
641 	}
642 }
643 
644 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
645 {
646 	percpu_counter_dec(&lock->writers);
647 	cond_wake_up(&lock->pending_readers);
648 }
649 
650 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
651 {
652 	atomic_inc(&lock->readers);
653 
654 	/*
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.
659 	 */
660 	smp_mb__after_atomic();
661 
662 	wait_event(lock->pending_readers,
663 		   percpu_counter_sum(&lock->writers) == 0);
664 }
665 
666 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
667 {
668 	/*
669 	 * atomic_dec_and_test implies a full barrier, so woken up writers
670 	 * are guaranteed to see the decrement
671 	 */
672 	if (atomic_dec_and_test(&lock->readers))
673 		wake_up(&lock->pending_writers);
674 }
675