xref: /linux/fs/btrfs/locking.c (revision 364eeb79a213fcf9164208b53764223ad522d6b3)
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  * Release read lock.
166  */
167 void btrfs_tree_read_unlock(struct extent_buffer *eb)
168 {
169 	trace_btrfs_tree_read_unlock(eb);
170 	up_read(&eb->lock);
171 }
172 
173 /*
174  * Lock eb for write.
175  *
176  * @eb:		the eb to lock
177  * @nest:	the nesting to use for the lock
178  *
179  * Returns with the eb->lock write locked.
180  */
181 void btrfs_tree_lock_nested(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
182 	__acquires(&eb->lock)
183 {
184 	u64 start_ns = 0;
185 
186 	if (trace_btrfs_tree_lock_enabled())
187 		start_ns = ktime_get_ns();
188 
189 	down_write_nested(&eb->lock, nest);
190 	btrfs_set_eb_lock_owner(eb, current->pid);
191 	trace_btrfs_tree_lock(eb, start_ns);
192 }
193 
194 /*
195  * Release the write lock.
196  */
197 void btrfs_tree_unlock(struct extent_buffer *eb)
198 {
199 	trace_btrfs_tree_unlock(eb);
200 	btrfs_set_eb_lock_owner(eb, 0);
201 	up_write(&eb->lock);
202 }
203 
204 /*
205  * This releases any locks held in the path starting at level and going all the
206  * way up to the root.
207  *
208  * btrfs_search_slot will keep the lock held on higher nodes in a few corner
209  * cases, such as COW of the block at slot zero in the node.  This ignores
210  * those rules, and it should only be called when there are no more updates to
211  * be done higher up in the tree.
212  */
213 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
214 {
215 	int i;
216 
217 	if (path->keep_locks)
218 		return;
219 
220 	for (i = level; i < BTRFS_MAX_LEVEL; i++) {
221 		if (!path->nodes[i])
222 			continue;
223 		if (!path->locks[i])
224 			continue;
225 		btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
226 		path->locks[i] = 0;
227 	}
228 }
229 
230 /*
231  * Loop around taking references on and locking the root node of the tree until
232  * we end up with a lock on the root node.
233  *
234  * Return: root extent buffer with write lock held
235  */
236 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
237 {
238 	struct extent_buffer *eb;
239 
240 	while (1) {
241 		eb = btrfs_root_node(root);
242 
243 		btrfs_maybe_reset_lockdep_class(root, eb);
244 		btrfs_tree_lock(eb);
245 		if (eb == root->node)
246 			break;
247 		btrfs_tree_unlock(eb);
248 		free_extent_buffer(eb);
249 	}
250 	return eb;
251 }
252 
253 /*
254  * Loop around taking references on and locking the root node of the tree until
255  * we end up with a lock on the root node.
256  *
257  * Return: root extent buffer with read lock held
258  */
259 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
260 {
261 	struct extent_buffer *eb;
262 
263 	while (1) {
264 		eb = btrfs_root_node(root);
265 
266 		btrfs_maybe_reset_lockdep_class(root, eb);
267 		btrfs_tree_read_lock(eb);
268 		if (eb == root->node)
269 			break;
270 		btrfs_tree_read_unlock(eb);
271 		free_extent_buffer(eb);
272 	}
273 	return eb;
274 }
275 
276 /*
277  * Loop around taking references on and locking the root node of the tree in
278  * nowait mode until we end up with a lock on the root node or returning to
279  * avoid blocking.
280  *
281  * Return: root extent buffer with read lock held or -EAGAIN.
282  */
283 struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
284 {
285 	struct extent_buffer *eb;
286 
287 	while (1) {
288 		eb = btrfs_root_node(root);
289 		if (!btrfs_try_tree_read_lock(eb)) {
290 			free_extent_buffer(eb);
291 			return ERR_PTR(-EAGAIN);
292 		}
293 		if (eb == root->node)
294 			break;
295 		btrfs_tree_read_unlock(eb);
296 		free_extent_buffer(eb);
297 	}
298 	return eb;
299 }
300 
301 /*
302  * DREW locks
303  * ==========
304  *
305  * DREW stands for double-reader-writer-exclusion lock. It's used in situation
306  * where you want to provide A-B exclusion but not AA or BB.
307  *
308  * Currently implementation gives more priority to reader. If a reader and a
309  * writer both race to acquire their respective sides of the lock the writer
310  * would yield its lock as soon as it detects a concurrent reader. Additionally
311  * if there are pending readers no new writers would be allowed to come in and
312  * acquire the lock.
313  */
314 
315 void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
316 {
317 	atomic_set(&lock->readers, 0);
318 	atomic_set(&lock->writers, 0);
319 	init_waitqueue_head(&lock->pending_readers);
320 	init_waitqueue_head(&lock->pending_writers);
321 }
322 
323 /* Return true if acquisition is successful, false otherwise */
324 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
325 {
326 	if (atomic_read(&lock->readers))
327 		return false;
328 
329 	atomic_inc(&lock->writers);
330 
331 	/* Ensure writers count is updated before we check for pending readers */
332 	smp_mb__after_atomic();
333 	if (atomic_read(&lock->readers)) {
334 		btrfs_drew_write_unlock(lock);
335 		return false;
336 	}
337 
338 	return true;
339 }
340 
341 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
342 {
343 	while (true) {
344 		if (btrfs_drew_try_write_lock(lock))
345 			return;
346 		wait_event(lock->pending_writers, !atomic_read(&lock->readers));
347 	}
348 }
349 
350 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
351 {
352 	/*
353 	 * atomic_dec_and_test() implies a full barrier, so woken up readers are
354 	 * guaranteed to see the decrement.
355 	 */
356 	if (atomic_dec_and_test(&lock->writers))
357 		wake_up(&lock->pending_readers);
358 }
359 
360 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
361 {
362 	atomic_inc(&lock->readers);
363 
364 	/*
365 	 * Ensure the pending reader count is perceieved BEFORE this reader
366 	 * goes to sleep in case of active writers. This guarantees new writers
367 	 * won't be allowed and that the current reader will be woken up when
368 	 * the last active writer finishes its jobs.
369 	 */
370 	smp_mb__after_atomic();
371 
372 	wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
373 }
374 
375 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
376 {
377 	/*
378 	 * atomic_dec_and_test implies a full barrier, so woken up writers
379 	 * are guaranteed to see the decrement
380 	 */
381 	if (atomic_dec_and_test(&lock->readers))
382 		wake_up(&lock->pending_writers);
383 }
384