Loading...
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#ifdef CONFIG_BTRFS_DEBUG
17static void btrfs_assert_spinning_writers_get(struct extent_buffer *eb)
18{
19 WARN_ON(eb->spinning_writers);
20 eb->spinning_writers++;
21}
22
23static void btrfs_assert_spinning_writers_put(struct extent_buffer *eb)
24{
25 WARN_ON(eb->spinning_writers != 1);
26 eb->spinning_writers--;
27}
28
29static void btrfs_assert_no_spinning_writers(struct extent_buffer *eb)
30{
31 WARN_ON(eb->spinning_writers);
32}
33
34static void btrfs_assert_spinning_readers_get(struct extent_buffer *eb)
35{
36 atomic_inc(&eb->spinning_readers);
37}
38
39static void btrfs_assert_spinning_readers_put(struct extent_buffer *eb)
40{
41 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
42 atomic_dec(&eb->spinning_readers);
43}
44
45static void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb)
46{
47 atomic_inc(&eb->read_locks);
48}
49
50static void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb)
51{
52 atomic_dec(&eb->read_locks);
53}
54
55static void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
56{
57 BUG_ON(!atomic_read(&eb->read_locks));
58}
59
60static void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb)
61{
62 eb->write_locks++;
63}
64
65static void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb)
66{
67 eb->write_locks--;
68}
69
70void btrfs_assert_tree_locked(struct extent_buffer *eb)
71{
72 BUG_ON(!eb->write_locks);
73}
74
75#else
76static void btrfs_assert_spinning_writers_get(struct extent_buffer *eb) { }
77static void btrfs_assert_spinning_writers_put(struct extent_buffer *eb) { }
78static void btrfs_assert_no_spinning_writers(struct extent_buffer *eb) { }
79static void btrfs_assert_spinning_readers_put(struct extent_buffer *eb) { }
80static void btrfs_assert_spinning_readers_get(struct extent_buffer *eb) { }
81static void btrfs_assert_tree_read_locked(struct extent_buffer *eb) { }
82static void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb) { }
83static void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb) { }
84void btrfs_assert_tree_locked(struct extent_buffer *eb) { }
85static void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb) { }
86static void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb) { }
87#endif
88
89void btrfs_set_lock_blocking_read(struct extent_buffer *eb)
90{
91 trace_btrfs_set_lock_blocking_read(eb);
92 /*
93 * No lock is required. The lock owner may change if we have a read
94 * lock, but it won't change to or away from us. If we have the write
95 * lock, we are the owner and it'll never change.
96 */
97 if (eb->lock_nested && current->pid == eb->lock_owner)
98 return;
99 btrfs_assert_tree_read_locked(eb);
100 atomic_inc(&eb->blocking_readers);
101 btrfs_assert_spinning_readers_put(eb);
102 read_unlock(&eb->lock);
103}
104
105void btrfs_set_lock_blocking_write(struct extent_buffer *eb)
106{
107 trace_btrfs_set_lock_blocking_write(eb);
108 /*
109 * No lock is required. The lock owner may change if we have a read
110 * lock, but it won't change to or away from us. If we have the write
111 * lock, we are the owner and it'll never change.
112 */
113 if (eb->lock_nested && current->pid == eb->lock_owner)
114 return;
115 if (eb->blocking_writers == 0) {
116 btrfs_assert_spinning_writers_put(eb);
117 btrfs_assert_tree_locked(eb);
118 eb->blocking_writers++;
119 write_unlock(&eb->lock);
120 }
121}
122
123/*
124 * take a spinning read lock. This will wait for any blocking
125 * writers
126 */
127void btrfs_tree_read_lock(struct extent_buffer *eb)
128{
129 u64 start_ns = 0;
130
131 if (trace_btrfs_tree_read_lock_enabled())
132 start_ns = ktime_get_ns();
133again:
134 read_lock(&eb->lock);
135 BUG_ON(eb->blocking_writers == 0 &&
136 current->pid == eb->lock_owner);
137 if (eb->blocking_writers && current->pid == eb->lock_owner) {
138 /*
139 * This extent is already write-locked by our thread. We allow
140 * an additional read lock to be added because it's for the same
141 * thread. btrfs_find_all_roots() depends on this as it may be
142 * called on a partly (write-)locked tree.
143 */
144 BUG_ON(eb->lock_nested);
145 eb->lock_nested = true;
146 read_unlock(&eb->lock);
147 trace_btrfs_tree_read_lock(eb, start_ns);
148 return;
149 }
150 if (eb->blocking_writers) {
151 read_unlock(&eb->lock);
152 wait_event(eb->write_lock_wq,
153 eb->blocking_writers == 0);
154 goto again;
155 }
156 btrfs_assert_tree_read_locks_get(eb);
157 btrfs_assert_spinning_readers_get(eb);
158 trace_btrfs_tree_read_lock(eb, start_ns);
159}
160
161/*
162 * take a spinning read lock.
163 * returns 1 if we get the read lock and 0 if we don't
164 * this won't wait for blocking writers
165 */
166int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
167{
168 if (eb->blocking_writers)
169 return 0;
170
171 read_lock(&eb->lock);
172 if (eb->blocking_writers) {
173 read_unlock(&eb->lock);
174 return 0;
175 }
176 btrfs_assert_tree_read_locks_get(eb);
177 btrfs_assert_spinning_readers_get(eb);
178 trace_btrfs_tree_read_lock_atomic(eb);
179 return 1;
180}
181
182/*
183 * returns 1 if we get the read lock and 0 if we don't
184 * this won't wait for blocking writers
185 */
186int btrfs_try_tree_read_lock(struct extent_buffer *eb)
187{
188 if (eb->blocking_writers)
189 return 0;
190
191 if (!read_trylock(&eb->lock))
192 return 0;
193
194 if (eb->blocking_writers) {
195 read_unlock(&eb->lock);
196 return 0;
197 }
198 btrfs_assert_tree_read_locks_get(eb);
199 btrfs_assert_spinning_readers_get(eb);
200 trace_btrfs_try_tree_read_lock(eb);
201 return 1;
202}
203
204/*
205 * returns 1 if we get the read lock and 0 if we don't
206 * this won't wait for blocking writers or readers
207 */
208int btrfs_try_tree_write_lock(struct extent_buffer *eb)
209{
210 if (eb->blocking_writers || atomic_read(&eb->blocking_readers))
211 return 0;
212
213 write_lock(&eb->lock);
214 if (eb->blocking_writers || atomic_read(&eb->blocking_readers)) {
215 write_unlock(&eb->lock);
216 return 0;
217 }
218 btrfs_assert_tree_write_locks_get(eb);
219 btrfs_assert_spinning_writers_get(eb);
220 eb->lock_owner = current->pid;
221 trace_btrfs_try_tree_write_lock(eb);
222 return 1;
223}
224
225/*
226 * drop a spinning read lock
227 */
228void btrfs_tree_read_unlock(struct extent_buffer *eb)
229{
230 trace_btrfs_tree_read_unlock(eb);
231 /*
232 * if we're nested, we have the write lock. No new locking
233 * is needed as long as we are the lock owner.
234 * The write unlock will do a barrier for us, and the lock_nested
235 * field only matters to the lock owner.
236 */
237 if (eb->lock_nested && current->pid == eb->lock_owner) {
238 eb->lock_nested = false;
239 return;
240 }
241 btrfs_assert_tree_read_locked(eb);
242 btrfs_assert_spinning_readers_put(eb);
243 btrfs_assert_tree_read_locks_put(eb);
244 read_unlock(&eb->lock);
245}
246
247/*
248 * drop a blocking read lock
249 */
250void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
251{
252 trace_btrfs_tree_read_unlock_blocking(eb);
253 /*
254 * if we're nested, we have the write lock. No new locking
255 * is needed as long as we are the lock owner.
256 * The write unlock will do a barrier for us, and the lock_nested
257 * field only matters to the lock owner.
258 */
259 if (eb->lock_nested && current->pid == eb->lock_owner) {
260 eb->lock_nested = false;
261 return;
262 }
263 btrfs_assert_tree_read_locked(eb);
264 WARN_ON(atomic_read(&eb->blocking_readers) == 0);
265 /* atomic_dec_and_test implies a barrier */
266 if (atomic_dec_and_test(&eb->blocking_readers))
267 cond_wake_up_nomb(&eb->read_lock_wq);
268 btrfs_assert_tree_read_locks_put(eb);
269}
270
271/*
272 * take a spinning write lock. This will wait for both
273 * blocking readers or writers
274 */
275void btrfs_tree_lock(struct extent_buffer *eb)
276{
277 u64 start_ns = 0;
278
279 if (trace_btrfs_tree_lock_enabled())
280 start_ns = ktime_get_ns();
281
282 WARN_ON(eb->lock_owner == current->pid);
283again:
284 wait_event(eb->read_lock_wq, atomic_read(&eb->blocking_readers) == 0);
285 wait_event(eb->write_lock_wq, eb->blocking_writers == 0);
286 write_lock(&eb->lock);
287 if (atomic_read(&eb->blocking_readers) || eb->blocking_writers) {
288 write_unlock(&eb->lock);
289 goto again;
290 }
291 btrfs_assert_spinning_writers_get(eb);
292 btrfs_assert_tree_write_locks_get(eb);
293 eb->lock_owner = current->pid;
294 trace_btrfs_tree_lock(eb, start_ns);
295}
296
297/*
298 * drop a spinning or a blocking write lock.
299 */
300void btrfs_tree_unlock(struct extent_buffer *eb)
301{
302 int blockers = eb->blocking_writers;
303
304 BUG_ON(blockers > 1);
305
306 btrfs_assert_tree_locked(eb);
307 trace_btrfs_tree_unlock(eb);
308 eb->lock_owner = 0;
309 btrfs_assert_tree_write_locks_put(eb);
310
311 if (blockers) {
312 btrfs_assert_no_spinning_writers(eb);
313 eb->blocking_writers--;
314 /*
315 * We need to order modifying blocking_writers above with
316 * actually waking up the sleepers to ensure they see the
317 * updated value of blocking_writers
318 */
319 cond_wake_up(&eb->write_lock_wq);
320 } else {
321 btrfs_assert_spinning_writers_put(eb);
322 write_unlock(&eb->lock);
323 }
324}
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 * We use a rw_semaphore for tree locking, and the semantics are exactly the
21 * same:
22 *
23 * - reader/writer exclusion
24 * - writer/writer exclusion
25 * - reader/reader sharing
26 * - try-lock semantics for readers and writers
27 *
28 * The rwsem implementation does opportunistic spinning which reduces number of
29 * times the locking task needs to sleep.
30 */
31
32/*
33 * __btrfs_tree_read_lock - lock extent buffer for read
34 * @eb: the eb to be locked
35 * @nest: the nesting level to be used for lockdep
36 *
37 * This takes the read lock on the extent buffer, using the specified nesting
38 * level for lockdep purposes.
39 */
40void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
41{
42 u64 start_ns = 0;
43
44 if (trace_btrfs_tree_read_lock_enabled())
45 start_ns = ktime_get_ns();
46
47 down_read_nested(&eb->lock, nest);
48 eb->lock_owner = current->pid;
49 trace_btrfs_tree_read_lock(eb, start_ns);
50}
51
52void btrfs_tree_read_lock(struct extent_buffer *eb)
53{
54 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL);
55}
56
57/*
58 * Try-lock for read.
59 *
60 * Return 1 if the rwlock has been taken, 0 otherwise
61 */
62int btrfs_try_tree_read_lock(struct extent_buffer *eb)
63{
64 if (down_read_trylock(&eb->lock)) {
65 eb->lock_owner = current->pid;
66 trace_btrfs_try_tree_read_lock(eb);
67 return 1;
68 }
69 return 0;
70}
71
72/*
73 * Try-lock for write.
74 *
75 * Return 1 if the rwlock has been taken, 0 otherwise
76 */
77int btrfs_try_tree_write_lock(struct extent_buffer *eb)
78{
79 if (down_write_trylock(&eb->lock)) {
80 eb->lock_owner = current->pid;
81 trace_btrfs_try_tree_write_lock(eb);
82 return 1;
83 }
84 return 0;
85}
86
87/*
88 * Release read lock.
89 */
90void btrfs_tree_read_unlock(struct extent_buffer *eb)
91{
92 trace_btrfs_tree_read_unlock(eb);
93 eb->lock_owner = 0;
94 up_read(&eb->lock);
95}
96
97/*
98 * __btrfs_tree_lock - lock eb for write
99 * @eb: the eb to lock
100 * @nest: the nesting to use for the lock
101 *
102 * Returns with the eb->lock write locked.
103 */
104void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
105 __acquires(&eb->lock)
106{
107 u64 start_ns = 0;
108
109 if (trace_btrfs_tree_lock_enabled())
110 start_ns = ktime_get_ns();
111
112 down_write_nested(&eb->lock, nest);
113 eb->lock_owner = current->pid;
114 trace_btrfs_tree_lock(eb, start_ns);
115}
116
117void btrfs_tree_lock(struct extent_buffer *eb)
118{
119 __btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
120}
121
122/*
123 * Release the write lock.
124 */
125void btrfs_tree_unlock(struct extent_buffer *eb)
126{
127 trace_btrfs_tree_unlock(eb);
128 eb->lock_owner = 0;
129 up_write(&eb->lock);
130}
131
132/*
133 * This releases any locks held in the path starting at level and going all the
134 * way up to the root.
135 *
136 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
137 * cases, such as COW of the block at slot zero in the node. This ignores
138 * those rules, and it should only be called when there are no more updates to
139 * be done higher up in the tree.
140 */
141void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
142{
143 int i;
144
145 if (path->keep_locks)
146 return;
147
148 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
149 if (!path->nodes[i])
150 continue;
151 if (!path->locks[i])
152 continue;
153 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
154 path->locks[i] = 0;
155 }
156}
157
158/*
159 * Loop around taking references on and locking the root node of the tree until
160 * we end up with a lock on the root node.
161 *
162 * Return: root extent buffer with write lock held
163 */
164struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
165{
166 struct extent_buffer *eb;
167
168 while (1) {
169 eb = btrfs_root_node(root);
170 btrfs_tree_lock(eb);
171 if (eb == root->node)
172 break;
173 btrfs_tree_unlock(eb);
174 free_extent_buffer(eb);
175 }
176 return eb;
177}
178
179/*
180 * Loop around taking references on and locking the root node of the tree until
181 * we end up with a lock on the root node.
182 *
183 * Return: root extent buffer with read lock held
184 */
185struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
186{
187 struct extent_buffer *eb;
188
189 while (1) {
190 eb = btrfs_root_node(root);
191 btrfs_tree_read_lock(eb);
192 if (eb == root->node)
193 break;
194 btrfs_tree_read_unlock(eb);
195 free_extent_buffer(eb);
196 }
197 return eb;
198}
199
200/*
201 * DREW locks
202 * ==========
203 *
204 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
205 * where you want to provide A-B exclusion but not AA or BB.
206 *
207 * Currently implementation gives more priority to reader. If a reader and a
208 * writer both race to acquire their respective sides of the lock the writer
209 * would yield its lock as soon as it detects a concurrent reader. Additionally
210 * if there are pending readers no new writers would be allowed to come in and
211 * acquire the lock.
212 */
213
214int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
215{
216 int ret;
217
218 ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
219 if (ret)
220 return ret;
221
222 atomic_set(&lock->readers, 0);
223 init_waitqueue_head(&lock->pending_readers);
224 init_waitqueue_head(&lock->pending_writers);
225
226 return 0;
227}
228
229void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
230{
231 percpu_counter_destroy(&lock->writers);
232}
233
234/* Return true if acquisition is successful, false otherwise */
235bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
236{
237 if (atomic_read(&lock->readers))
238 return false;
239
240 percpu_counter_inc(&lock->writers);
241
242 /* Ensure writers count is updated before we check for pending readers */
243 smp_mb();
244 if (atomic_read(&lock->readers)) {
245 btrfs_drew_write_unlock(lock);
246 return false;
247 }
248
249 return true;
250}
251
252void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
253{
254 while (true) {
255 if (btrfs_drew_try_write_lock(lock))
256 return;
257 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
258 }
259}
260
261void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
262{
263 percpu_counter_dec(&lock->writers);
264 cond_wake_up(&lock->pending_readers);
265}
266
267void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
268{
269 atomic_inc(&lock->readers);
270
271 /*
272 * Ensure the pending reader count is perceieved BEFORE this reader
273 * goes to sleep in case of active writers. This guarantees new writers
274 * won't be allowed and that the current reader will be woken up when
275 * the last active writer finishes its jobs.
276 */
277 smp_mb__after_atomic();
278
279 wait_event(lock->pending_readers,
280 percpu_counter_sum(&lock->writers) == 0);
281}
282
283void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
284{
285 /*
286 * atomic_dec_and_test implies a full barrier, so woken up writers
287 * are guaranteed to see the decrement
288 */
289 if (atomic_dec_and_test(&lock->readers))
290 wake_up(&lock->pending_writers);
291}