xref: /openbmc/linux/fs/btrfs/locking.c (revision 8ab59da2)
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  * Lockdep class keys for extent_buffer->lock's in this root.  For a given
18  * eb, the lockdep key is determined by the btrfs_root it belongs to and
19  * the level the eb occupies in the tree.
20  *
21  * Different roots are used for different purposes and may nest inside each
22  * other and they require separate keysets.  As lockdep keys should be
23  * static, assign keysets according to the purpose of the root as indicated
24  * by btrfs_root->root_key.objectid.  This ensures that all special purpose
25  * roots have separate keysets.
26  *
27  * Lock-nesting across peer nodes is always done with the immediate parent
28  * node locked thus preventing deadlock.  As lockdep doesn't know this, use
29  * subclass to avoid triggering lockdep warning in such cases.
30  *
31  * The key is set by the readpage_end_io_hook after the buffer has passed
32  * csum validation but before the pages are unlocked.  It is also set by
33  * btrfs_init_new_buffer on freshly allocated blocks.
34  *
35  * We also add a check to make sure the highest level of the tree is the
36  * same as our lockdep setup here.  If BTRFS_MAX_LEVEL changes, this code
37  * needs update as well.
38  */
39 #ifdef CONFIG_DEBUG_LOCK_ALLOC
40 #if BTRFS_MAX_LEVEL != 8
41 #error
42 #endif
43 
44 #define DEFINE_LEVEL(stem, level)					\
45 	.names[level] = "btrfs-" stem "-0" #level,
46 
47 #define DEFINE_NAME(stem)						\
48 	DEFINE_LEVEL(stem, 0)						\
49 	DEFINE_LEVEL(stem, 1)						\
50 	DEFINE_LEVEL(stem, 2)						\
51 	DEFINE_LEVEL(stem, 3)						\
52 	DEFINE_LEVEL(stem, 4)						\
53 	DEFINE_LEVEL(stem, 5)						\
54 	DEFINE_LEVEL(stem, 6)						\
55 	DEFINE_LEVEL(stem, 7)
56 
57 static struct btrfs_lockdep_keyset {
58 	u64			id;		/* root objectid */
59 	/* Longest entry: btrfs-free-space-00 */
60 	char			names[BTRFS_MAX_LEVEL][20];
61 	struct lock_class_key	keys[BTRFS_MAX_LEVEL];
62 } btrfs_lockdep_keysets[] = {
63 	{ .id = BTRFS_ROOT_TREE_OBJECTID,	DEFINE_NAME("root")	},
64 	{ .id = BTRFS_EXTENT_TREE_OBJECTID,	DEFINE_NAME("extent")	},
65 	{ .id = BTRFS_CHUNK_TREE_OBJECTID,	DEFINE_NAME("chunk")	},
66 	{ .id = BTRFS_DEV_TREE_OBJECTID,	DEFINE_NAME("dev")	},
67 	{ .id = BTRFS_CSUM_TREE_OBJECTID,	DEFINE_NAME("csum")	},
68 	{ .id = BTRFS_QUOTA_TREE_OBJECTID,	DEFINE_NAME("quota")	},
69 	{ .id = BTRFS_TREE_LOG_OBJECTID,	DEFINE_NAME("log")	},
70 	{ .id = BTRFS_TREE_RELOC_OBJECTID,	DEFINE_NAME("treloc")	},
71 	{ .id = BTRFS_DATA_RELOC_TREE_OBJECTID,	DEFINE_NAME("dreloc")	},
72 	{ .id = BTRFS_UUID_TREE_OBJECTID,	DEFINE_NAME("uuid")	},
73 	{ .id = BTRFS_FREE_SPACE_TREE_OBJECTID,	DEFINE_NAME("free-space") },
74 	{ .id = 0,				DEFINE_NAME("tree")	},
75 };
76 
77 #undef DEFINE_LEVEL
78 #undef DEFINE_NAME
79 
80 void btrfs_set_buffer_lockdep_class(u64 objectid, struct extent_buffer *eb, int level)
81 {
82 	struct btrfs_lockdep_keyset *ks;
83 
84 	BUG_ON(level >= ARRAY_SIZE(ks->keys));
85 
86 	/* Find the matching keyset, id 0 is the default entry */
87 	for (ks = btrfs_lockdep_keysets; ks->id; ks++)
88 		if (ks->id == objectid)
89 			break;
90 
91 	lockdep_set_class_and_name(&eb->lock, &ks->keys[level], ks->names[level]);
92 }
93 
94 void btrfs_maybe_reset_lockdep_class(struct btrfs_root *root, struct extent_buffer *eb)
95 {
96 	if (test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
97 		btrfs_set_buffer_lockdep_class(root->root_key.objectid,
98 					       eb, btrfs_header_level(eb));
99 }
100 
101 #endif
102 
103 /*
104  * Extent buffer locking
105  * =====================
106  *
107  * We use a rw_semaphore for tree locking, and the semantics are exactly the
108  * same:
109  *
110  * - reader/writer exclusion
111  * - writer/writer exclusion
112  * - reader/reader sharing
113  * - try-lock semantics for readers and writers
114  *
115  * The rwsem implementation does opportunistic spinning which reduces number of
116  * times the locking task needs to sleep.
117  */
118 
119 /*
120  * __btrfs_tree_read_lock - lock extent buffer for read
121  * @eb:		the eb to be locked
122  * @nest:	the nesting level to be used for lockdep
123  *
124  * This takes the read lock on the extent buffer, using the specified nesting
125  * level for lockdep purposes.
126  */
127 void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
128 {
129 	u64 start_ns = 0;
130 
131 	if (trace_btrfs_tree_read_lock_enabled())
132 		start_ns = ktime_get_ns();
133 
134 	down_read_nested(&eb->lock, nest);
135 	trace_btrfs_tree_read_lock(eb, start_ns);
136 }
137 
138 void btrfs_tree_read_lock(struct extent_buffer *eb)
139 {
140 	__btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL);
141 }
142 
143 /*
144  * Try-lock for read.
145  *
146  * Return 1 if the rwlock has been taken, 0 otherwise
147  */
148 int btrfs_try_tree_read_lock(struct extent_buffer *eb)
149 {
150 	if (down_read_trylock(&eb->lock)) {
151 		trace_btrfs_try_tree_read_lock(eb);
152 		return 1;
153 	}
154 	return 0;
155 }
156 
157 /*
158  * Try-lock for write.
159  *
160  * Return 1 if the rwlock has been taken, 0 otherwise
161  */
162 int btrfs_try_tree_write_lock(struct extent_buffer *eb)
163 {
164 	if (down_write_trylock(&eb->lock)) {
165 		eb->lock_owner = current->pid;
166 		trace_btrfs_try_tree_write_lock(eb);
167 		return 1;
168 	}
169 	return 0;
170 }
171 
172 /*
173  * Release read lock.
174  */
175 void btrfs_tree_read_unlock(struct extent_buffer *eb)
176 {
177 	trace_btrfs_tree_read_unlock(eb);
178 	up_read(&eb->lock);
179 }
180 
181 /*
182  * __btrfs_tree_lock - lock eb for write
183  * @eb:		the eb to lock
184  * @nest:	the nesting to use for the lock
185  *
186  * Returns with the eb->lock write locked.
187  */
188 void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
189 	__acquires(&eb->lock)
190 {
191 	u64 start_ns = 0;
192 
193 	if (trace_btrfs_tree_lock_enabled())
194 		start_ns = ktime_get_ns();
195 
196 	down_write_nested(&eb->lock, nest);
197 	eb->lock_owner = current->pid;
198 	trace_btrfs_tree_lock(eb, start_ns);
199 }
200 
201 void btrfs_tree_lock(struct extent_buffer *eb)
202 {
203 	__btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
204 }
205 
206 /*
207  * Release the write lock.
208  */
209 void btrfs_tree_unlock(struct extent_buffer *eb)
210 {
211 	trace_btrfs_tree_unlock(eb);
212 	eb->lock_owner = 0;
213 	up_write(&eb->lock);
214 }
215 
216 /*
217  * This releases any locks held in the path starting at level and going all the
218  * way up to the root.
219  *
220  * btrfs_search_slot will keep the lock held on higher nodes in a few corner
221  * cases, such as COW of the block at slot zero in the node.  This ignores
222  * those rules, and it should only be called when there are no more updates to
223  * be done higher up in the tree.
224  */
225 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
226 {
227 	int i;
228 
229 	if (path->keep_locks)
230 		return;
231 
232 	for (i = level; i < BTRFS_MAX_LEVEL; i++) {
233 		if (!path->nodes[i])
234 			continue;
235 		if (!path->locks[i])
236 			continue;
237 		btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
238 		path->locks[i] = 0;
239 	}
240 }
241 
242 /*
243  * Loop around taking references on and locking the root node of the tree until
244  * we end up with a lock on the root node.
245  *
246  * Return: root extent buffer with write lock held
247  */
248 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
249 {
250 	struct extent_buffer *eb;
251 
252 	while (1) {
253 		eb = btrfs_root_node(root);
254 
255 		btrfs_maybe_reset_lockdep_class(root, eb);
256 		btrfs_tree_lock(eb);
257 		if (eb == root->node)
258 			break;
259 		btrfs_tree_unlock(eb);
260 		free_extent_buffer(eb);
261 	}
262 	return eb;
263 }
264 
265 /*
266  * Loop around taking references on and locking the root node of the tree until
267  * we end up with a lock on the root node.
268  *
269  * Return: root extent buffer with read lock held
270  */
271 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
272 {
273 	struct extent_buffer *eb;
274 
275 	while (1) {
276 		eb = btrfs_root_node(root);
277 
278 		btrfs_maybe_reset_lockdep_class(root, eb);
279 		btrfs_tree_read_lock(eb);
280 		if (eb == root->node)
281 			break;
282 		btrfs_tree_read_unlock(eb);
283 		free_extent_buffer(eb);
284 	}
285 	return eb;
286 }
287 
288 /*
289  * Loop around taking references on and locking the root node of the tree in
290  * nowait mode until we end up with a lock on the root node or returning to
291  * avoid blocking.
292  *
293  * Return: root extent buffer with read lock held or -EAGAIN.
294  */
295 struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
296 {
297 	struct extent_buffer *eb;
298 
299 	while (1) {
300 		eb = btrfs_root_node(root);
301 		if (!btrfs_try_tree_read_lock(eb)) {
302 			free_extent_buffer(eb);
303 			return ERR_PTR(-EAGAIN);
304 		}
305 		if (eb == root->node)
306 			break;
307 		btrfs_tree_read_unlock(eb);
308 		free_extent_buffer(eb);
309 	}
310 	return eb;
311 }
312 
313 /*
314  * DREW locks
315  * ==========
316  *
317  * DREW stands for double-reader-writer-exclusion lock. It's used in situation
318  * where you want to provide A-B exclusion but not AA or BB.
319  *
320  * Currently implementation gives more priority to reader. If a reader and a
321  * writer both race to acquire their respective sides of the lock the writer
322  * would yield its lock as soon as it detects a concurrent reader. Additionally
323  * if there are pending readers no new writers would be allowed to come in and
324  * acquire the lock.
325  */
326 
327 int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
328 {
329 	int ret;
330 
331 	ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
332 	if (ret)
333 		return ret;
334 
335 	atomic_set(&lock->readers, 0);
336 	init_waitqueue_head(&lock->pending_readers);
337 	init_waitqueue_head(&lock->pending_writers);
338 
339 	return 0;
340 }
341 
342 void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
343 {
344 	percpu_counter_destroy(&lock->writers);
345 }
346 
347 /* Return true if acquisition is successful, false otherwise */
348 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
349 {
350 	if (atomic_read(&lock->readers))
351 		return false;
352 
353 	percpu_counter_inc(&lock->writers);
354 
355 	/* Ensure writers count is updated before we check for pending readers */
356 	smp_mb();
357 	if (atomic_read(&lock->readers)) {
358 		btrfs_drew_write_unlock(lock);
359 		return false;
360 	}
361 
362 	return true;
363 }
364 
365 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
366 {
367 	while (true) {
368 		if (btrfs_drew_try_write_lock(lock))
369 			return;
370 		wait_event(lock->pending_writers, !atomic_read(&lock->readers));
371 	}
372 }
373 
374 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
375 {
376 	percpu_counter_dec(&lock->writers);
377 	cond_wake_up(&lock->pending_readers);
378 }
379 
380 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
381 {
382 	atomic_inc(&lock->readers);
383 
384 	/*
385 	 * Ensure the pending reader count is perceieved BEFORE this reader
386 	 * goes to sleep in case of active writers. This guarantees new writers
387 	 * won't be allowed and that the current reader will be woken up when
388 	 * the last active writer finishes its jobs.
389 	 */
390 	smp_mb__after_atomic();
391 
392 	wait_event(lock->pending_readers,
393 		   percpu_counter_sum(&lock->writers) == 0);
394 }
395 
396 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
397 {
398 	/*
399 	 * atomic_dec_and_test implies a full barrier, so woken up writers
400 	 * are guaranteed to see the decrement
401 	 */
402 	if (atomic_dec_and_test(&lock->readers))
403 		wake_up(&lock->pending_writers);
404 }
405