xref: /linux/fs/btrfs/locking.c (revision f898c16a0624e7f2dcb0b1cda6916c9be6489197)
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 <trace/events/btrfs.h>
12 #include "misc.h"
13 #include "ctree.h"
14 #include "extent_io.h"
15 #include "locking.h"
16 
17 /*
18  * Lockdep class keys for extent_buffer->lock's in this root.  For a given
19  * eb, the lockdep key is determined by the btrfs_root it belongs to and
20  * the level the eb occupies in the tree.
21  *
22  * Different roots are used for different purposes and may nest inside each
23  * other and they require separate keysets.  As lockdep keys should be
24  * static, assign keysets according to the purpose of the root as indicated
25  * by btrfs_root->root_key.objectid.  This ensures that all special purpose
26  * roots have separate keysets.
27  *
28  * Lock-nesting across peer nodes is always done with the immediate parent
29  * node locked thus preventing deadlock.  As lockdep doesn't know this, use
30  * subclass to avoid triggering lockdep warning in such cases.
31  *
32  * The key is set by the readpage_end_io_hook after the buffer has passed
33  * csum validation but before the pages are unlocked.  It is also set by
34  * btrfs_init_new_buffer on freshly allocated blocks.
35  *
36  * We also add a check to make sure the highest level of the tree is the
37  * same as our lockdep setup here.  If BTRFS_MAX_LEVEL changes, this code
38  * needs update as well.
39  */
40 #ifdef CONFIG_DEBUG_LOCK_ALLOC
41 #if BTRFS_MAX_LEVEL != 8
42 #error
43 #endif
44 
45 #define DEFINE_LEVEL(stem, level)					\
46 	.names[level] = "btrfs-" stem "-0" #level,
47 
48 #define DEFINE_NAME(stem)						\
49 	DEFINE_LEVEL(stem, 0)						\
50 	DEFINE_LEVEL(stem, 1)						\
51 	DEFINE_LEVEL(stem, 2)						\
52 	DEFINE_LEVEL(stem, 3)						\
53 	DEFINE_LEVEL(stem, 4)						\
54 	DEFINE_LEVEL(stem, 5)						\
55 	DEFINE_LEVEL(stem, 6)						\
56 	DEFINE_LEVEL(stem, 7)
57 
58 static struct btrfs_lockdep_keyset {
59 	u64			id;		/* root objectid */
60 	/* Longest entry: btrfs-block-group-00 */
61 	char			names[BTRFS_MAX_LEVEL][24];
62 	struct lock_class_key	keys[BTRFS_MAX_LEVEL];
63 } btrfs_lockdep_keysets[] = {
64 	{ .id = BTRFS_ROOT_TREE_OBJECTID,	DEFINE_NAME("root")	},
65 	{ .id = BTRFS_EXTENT_TREE_OBJECTID,	DEFINE_NAME("extent")	},
66 	{ .id = BTRFS_CHUNK_TREE_OBJECTID,	DEFINE_NAME("chunk")	},
67 	{ .id = BTRFS_DEV_TREE_OBJECTID,	DEFINE_NAME("dev")	},
68 	{ .id = BTRFS_CSUM_TREE_OBJECTID,	DEFINE_NAME("csum")	},
69 	{ .id = BTRFS_QUOTA_TREE_OBJECTID,	DEFINE_NAME("quota")	},
70 	{ .id = BTRFS_TREE_LOG_OBJECTID,	DEFINE_NAME("log")	},
71 	{ .id = BTRFS_TREE_RELOC_OBJECTID,	DEFINE_NAME("treloc")	},
72 	{ .id = BTRFS_DATA_RELOC_TREE_OBJECTID,	DEFINE_NAME("dreloc")	},
73 	{ .id = BTRFS_UUID_TREE_OBJECTID,	DEFINE_NAME("uuid")	},
74 	{ .id = BTRFS_FREE_SPACE_TREE_OBJECTID,	DEFINE_NAME("free-space") },
75 	{ .id = BTRFS_BLOCK_GROUP_TREE_OBJECTID, DEFINE_NAME("block-group") },
76 	{ .id = BTRFS_RAID_STRIPE_TREE_OBJECTID, DEFINE_NAME("raid-stripe") },
77 	{ .id = 0,				DEFINE_NAME("tree")	},
78 };
79 
80 #undef DEFINE_LEVEL
81 #undef DEFINE_NAME
82 
83 void btrfs_set_buffer_lockdep_class(u64 objectid, struct extent_buffer *eb, int level)
84 {
85 	struct btrfs_lockdep_keyset *ks;
86 
87 	ASSERT(level < ARRAY_SIZE(ks->keys));
88 
89 	/* Find the matching keyset, id 0 is the default entry */
90 	for (ks = btrfs_lockdep_keysets; ks->id; ks++)
91 		if (ks->id == objectid)
92 			break;
93 
94 	lockdep_set_class_and_name(&eb->lock, &ks->keys[level], ks->names[level]);
95 }
96 
97 void btrfs_maybe_reset_lockdep_class(struct btrfs_root *root, struct extent_buffer *eb)
98 {
99 	if (test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
100 		btrfs_set_buffer_lockdep_class(btrfs_root_id(root),
101 					       eb, btrfs_header_level(eb));
102 }
103 
104 #endif
105 
106 #ifdef CONFIG_BTRFS_DEBUG
107 static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner)
108 {
109 	eb->lock_owner = owner;
110 }
111 #else
112 static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner) { }
113 #endif
114 
115 /*
116  * Extent buffer locking
117  * =====================
118  *
119  * We use a rw_semaphore for tree locking, and the semantics are exactly the
120  * same:
121  *
122  * - reader/writer exclusion
123  * - writer/writer exclusion
124  * - reader/reader sharing
125  * - try-lock semantics for readers and writers
126  *
127  * The rwsem implementation does opportunistic spinning which reduces number of
128  * times the locking task needs to sleep.
129  */
130 
131 /*
132  * btrfs_tree_read_lock_nested - lock extent buffer for read
133  * @eb:		the eb to be locked
134  * @nest:	the nesting level to be used for lockdep
135  *
136  * This takes the read lock on the extent buffer, using the specified nesting
137  * level for lockdep purposes.
138  */
139 void btrfs_tree_read_lock_nested(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
140 {
141 	u64 start_ns = 0;
142 
143 	if (trace_btrfs_tree_read_lock_enabled())
144 		start_ns = ktime_get_ns();
145 
146 	down_read_nested(&eb->lock, nest);
147 	trace_btrfs_tree_read_lock(eb, start_ns);
148 }
149 
150 /*
151  * Try-lock for read.
152  *
153  * Return 1 if the rwlock has been taken, 0 otherwise
154  */
155 int btrfs_try_tree_read_lock(struct extent_buffer *eb)
156 {
157 	if (down_read_trylock(&eb->lock)) {
158 		trace_btrfs_try_tree_read_lock(eb);
159 		return 1;
160 	}
161 	return 0;
162 }
163 
164 /*
165  * Try-lock for write.
166  *
167  * Return 1 if the rwlock has been taken, 0 otherwise
168  */
169 int btrfs_try_tree_write_lock(struct extent_buffer *eb)
170 {
171 	if (down_write_trylock(&eb->lock)) {
172 		btrfs_set_eb_lock_owner(eb, current->pid);
173 		trace_btrfs_try_tree_write_lock(eb);
174 		return 1;
175 	}
176 	return 0;
177 }
178 
179 /*
180  * Release read lock.
181  */
182 void btrfs_tree_read_unlock(struct extent_buffer *eb)
183 {
184 	trace_btrfs_tree_read_unlock(eb);
185 	up_read(&eb->lock);
186 }
187 
188 /*
189  * Lock eb for write.
190  *
191  * @eb:		the eb to lock
192  * @nest:	the nesting to use for the lock
193  *
194  * Returns with the eb->lock write locked.
195  */
196 void btrfs_tree_lock_nested(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
197 	__acquires(&eb->lock)
198 {
199 	u64 start_ns = 0;
200 
201 	if (trace_btrfs_tree_lock_enabled())
202 		start_ns = ktime_get_ns();
203 
204 	down_write_nested(&eb->lock, nest);
205 	btrfs_set_eb_lock_owner(eb, current->pid);
206 	trace_btrfs_tree_lock(eb, start_ns);
207 }
208 
209 /*
210  * Release the write lock.
211  */
212 void btrfs_tree_unlock(struct extent_buffer *eb)
213 {
214 	trace_btrfs_tree_unlock(eb);
215 	btrfs_set_eb_lock_owner(eb, 0);
216 	up_write(&eb->lock);
217 }
218 
219 /*
220  * This releases any locks held in the path starting at level and going all the
221  * way up to the root.
222  *
223  * btrfs_search_slot will keep the lock held on higher nodes in a few corner
224  * cases, such as COW of the block at slot zero in the node.  This ignores
225  * those rules, and it should only be called when there are no more updates to
226  * be done higher up in the tree.
227  */
228 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
229 {
230 	int i;
231 
232 	if (path->keep_locks)
233 		return;
234 
235 	for (i = level; i < BTRFS_MAX_LEVEL; i++) {
236 		if (!path->nodes[i])
237 			continue;
238 		if (!path->locks[i])
239 			continue;
240 		btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
241 		path->locks[i] = 0;
242 	}
243 }
244 
245 /*
246  * Loop around taking references on and locking the root node of the tree until
247  * we end up with a lock on the root node.
248  *
249  * Return: root extent buffer with write lock held
250  */
251 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
252 {
253 	struct extent_buffer *eb;
254 
255 	while (1) {
256 		eb = btrfs_root_node(root);
257 
258 		btrfs_maybe_reset_lockdep_class(root, eb);
259 		btrfs_tree_lock(eb);
260 		if (eb == root->node)
261 			break;
262 		btrfs_tree_unlock(eb);
263 		free_extent_buffer(eb);
264 	}
265 	return eb;
266 }
267 
268 /*
269  * Loop around taking references on and locking the root node of the tree until
270  * we end up with a lock on the root node.
271  *
272  * Return: root extent buffer with read lock held
273  */
274 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
275 {
276 	struct extent_buffer *eb;
277 
278 	while (1) {
279 		eb = btrfs_root_node(root);
280 
281 		btrfs_maybe_reset_lockdep_class(root, eb);
282 		btrfs_tree_read_lock(eb);
283 		if (eb == root->node)
284 			break;
285 		btrfs_tree_read_unlock(eb);
286 		free_extent_buffer(eb);
287 	}
288 	return eb;
289 }
290 
291 /*
292  * Loop around taking references on and locking the root node of the tree in
293  * nowait mode until we end up with a lock on the root node or returning to
294  * avoid blocking.
295  *
296  * Return: root extent buffer with read lock held or -EAGAIN.
297  */
298 struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
299 {
300 	struct extent_buffer *eb;
301 
302 	while (1) {
303 		eb = btrfs_root_node(root);
304 		if (!btrfs_try_tree_read_lock(eb)) {
305 			free_extent_buffer(eb);
306 			return ERR_PTR(-EAGAIN);
307 		}
308 		if (eb == root->node)
309 			break;
310 		btrfs_tree_read_unlock(eb);
311 		free_extent_buffer(eb);
312 	}
313 	return eb;
314 }
315 
316 /*
317  * DREW locks
318  * ==========
319  *
320  * DREW stands for double-reader-writer-exclusion lock. It's used in situation
321  * where you want to provide A-B exclusion but not AA or BB.
322  *
323  * Currently implementation gives more priority to reader. If a reader and a
324  * writer both race to acquire their respective sides of the lock the writer
325  * would yield its lock as soon as it detects a concurrent reader. Additionally
326  * if there are pending readers no new writers would be allowed to come in and
327  * acquire the lock.
328  */
329 
330 void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
331 {
332 	atomic_set(&lock->readers, 0);
333 	atomic_set(&lock->writers, 0);
334 	init_waitqueue_head(&lock->pending_readers);
335 	init_waitqueue_head(&lock->pending_writers);
336 }
337 
338 /* Return true if acquisition is successful, false otherwise */
339 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
340 {
341 	if (atomic_read(&lock->readers))
342 		return false;
343 
344 	atomic_inc(&lock->writers);
345 
346 	/* Ensure writers count is updated before we check for pending readers */
347 	smp_mb__after_atomic();
348 	if (atomic_read(&lock->readers)) {
349 		btrfs_drew_write_unlock(lock);
350 		return false;
351 	}
352 
353 	return true;
354 }
355 
356 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
357 {
358 	while (true) {
359 		if (btrfs_drew_try_write_lock(lock))
360 			return;
361 		wait_event(lock->pending_writers, !atomic_read(&lock->readers));
362 	}
363 }
364 
365 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
366 {
367 	/*
368 	 * atomic_dec_and_test() implies a full barrier, so woken up readers are
369 	 * guaranteed to see the decrement.
370 	 */
371 	if (atomic_dec_and_test(&lock->writers))
372 		wake_up(&lock->pending_readers);
373 }
374 
375 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
376 {
377 	atomic_inc(&lock->readers);
378 
379 	/*
380 	 * Ensure the pending reader count is perceieved BEFORE this reader
381 	 * goes to sleep in case of active writers. This guarantees new writers
382 	 * won't be allowed and that the current reader will be woken up when
383 	 * the last active writer finishes its jobs.
384 	 */
385 	smp_mb__after_atomic();
386 
387 	wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
388 }
389 
390 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
391 {
392 	/*
393 	 * atomic_dec_and_test implies a full barrier, so woken up writers
394 	 * are guaranteed to see the decrement
395 	 */
396 	if (atomic_dec_and_test(&lock->readers))
397 		wake_up(&lock->pending_writers);
398 }
399