GNU Linux-libre 5.19-rc6-gnu
[releases.git] / fs / btrfs / locking.c
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  */
40 void __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         trace_btrfs_tree_read_lock(eb, start_ns);
49 }
50
51 void btrfs_tree_read_lock(struct extent_buffer *eb)
52 {
53         __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL);
54 }
55
56 /*
57  * Try-lock for read.
58  *
59  * Return 1 if the rwlock has been taken, 0 otherwise
60  */
61 int btrfs_try_tree_read_lock(struct extent_buffer *eb)
62 {
63         if (down_read_trylock(&eb->lock)) {
64                 trace_btrfs_try_tree_read_lock(eb);
65                 return 1;
66         }
67         return 0;
68 }
69
70 /*
71  * Try-lock for write.
72  *
73  * Return 1 if the rwlock has been taken, 0 otherwise
74  */
75 int btrfs_try_tree_write_lock(struct extent_buffer *eb)
76 {
77         if (down_write_trylock(&eb->lock)) {
78                 eb->lock_owner = current->pid;
79                 trace_btrfs_try_tree_write_lock(eb);
80                 return 1;
81         }
82         return 0;
83 }
84
85 /*
86  * Release read lock.
87  */
88 void btrfs_tree_read_unlock(struct extent_buffer *eb)
89 {
90         trace_btrfs_tree_read_unlock(eb);
91         up_read(&eb->lock);
92 }
93
94 /*
95  * __btrfs_tree_lock - lock eb for write
96  * @eb:         the eb to lock
97  * @nest:       the nesting to use for the lock
98  *
99  * Returns with the eb->lock write locked.
100  */
101 void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
102         __acquires(&eb->lock)
103 {
104         u64 start_ns = 0;
105
106         if (trace_btrfs_tree_lock_enabled())
107                 start_ns = ktime_get_ns();
108
109         down_write_nested(&eb->lock, nest);
110         eb->lock_owner = current->pid;
111         trace_btrfs_tree_lock(eb, start_ns);
112 }
113
114 void btrfs_tree_lock(struct extent_buffer *eb)
115 {
116         __btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
117 }
118
119 /*
120  * Release the write lock.
121  */
122 void btrfs_tree_unlock(struct extent_buffer *eb)
123 {
124         trace_btrfs_tree_unlock(eb);
125         eb->lock_owner = 0;
126         up_write(&eb->lock);
127 }
128
129 /*
130  * This releases any locks held in the path starting at level and going all the
131  * way up to the root.
132  *
133  * btrfs_search_slot will keep the lock held on higher nodes in a few corner
134  * cases, such as COW of the block at slot zero in the node.  This ignores
135  * those rules, and it should only be called when there are no more updates to
136  * be done higher up in the tree.
137  */
138 void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
139 {
140         int i;
141
142         if (path->keep_locks)
143                 return;
144
145         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
146                 if (!path->nodes[i])
147                         continue;
148                 if (!path->locks[i])
149                         continue;
150                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
151                 path->locks[i] = 0;
152         }
153 }
154
155 /*
156  * Loop around taking references on and locking the root node of the tree until
157  * we end up with a lock on the root node.
158  *
159  * Return: root extent buffer with write lock held
160  */
161 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
162 {
163         struct extent_buffer *eb;
164
165         while (1) {
166                 eb = btrfs_root_node(root);
167                 btrfs_tree_lock(eb);
168                 if (eb == root->node)
169                         break;
170                 btrfs_tree_unlock(eb);
171                 free_extent_buffer(eb);
172         }
173         return eb;
174 }
175
176 /*
177  * Loop around taking references on and locking the root node of the tree until
178  * we end up with a lock on the root node.
179  *
180  * Return: root extent buffer with read lock held
181  */
182 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
183 {
184         struct extent_buffer *eb;
185
186         while (1) {
187                 eb = btrfs_root_node(root);
188                 btrfs_tree_read_lock(eb);
189                 if (eb == root->node)
190                         break;
191                 btrfs_tree_read_unlock(eb);
192                 free_extent_buffer(eb);
193         }
194         return eb;
195 }
196
197 /*
198  * DREW locks
199  * ==========
200  *
201  * DREW stands for double-reader-writer-exclusion lock. It's used in situation
202  * where you want to provide A-B exclusion but not AA or BB.
203  *
204  * Currently implementation gives more priority to reader. If a reader and a
205  * writer both race to acquire their respective sides of the lock the writer
206  * would yield its lock as soon as it detects a concurrent reader. Additionally
207  * if there are pending readers no new writers would be allowed to come in and
208  * acquire the lock.
209  */
210
211 int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
212 {
213         int ret;
214
215         ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
216         if (ret)
217                 return ret;
218
219         atomic_set(&lock->readers, 0);
220         init_waitqueue_head(&lock->pending_readers);
221         init_waitqueue_head(&lock->pending_writers);
222
223         return 0;
224 }
225
226 void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
227 {
228         percpu_counter_destroy(&lock->writers);
229 }
230
231 /* Return true if acquisition is successful, false otherwise */
232 bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
233 {
234         if (atomic_read(&lock->readers))
235                 return false;
236
237         percpu_counter_inc(&lock->writers);
238
239         /* Ensure writers count is updated before we check for pending readers */
240         smp_mb();
241         if (atomic_read(&lock->readers)) {
242                 btrfs_drew_write_unlock(lock);
243                 return false;
244         }
245
246         return true;
247 }
248
249 void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
250 {
251         while (true) {
252                 if (btrfs_drew_try_write_lock(lock))
253                         return;
254                 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
255         }
256 }
257
258 void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
259 {
260         percpu_counter_dec(&lock->writers);
261         cond_wake_up(&lock->pending_readers);
262 }
263
264 void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
265 {
266         atomic_inc(&lock->readers);
267
268         /*
269          * Ensure the pending reader count is perceieved BEFORE this reader
270          * goes to sleep in case of active writers. This guarantees new writers
271          * won't be allowed and that the current reader will be woken up when
272          * the last active writer finishes its jobs.
273          */
274         smp_mb__after_atomic();
275
276         wait_event(lock->pending_readers,
277                    percpu_counter_sum(&lock->writers) == 0);
278 }
279
280 void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
281 {
282         /*
283          * atomic_dec_and_test implies a full barrier, so woken up writers
284          * are guaranteed to see the decrement
285          */
286         if (atomic_dec_and_test(&lock->readers))
287                 wake_up(&lock->pending_writers);
288 }