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 <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
58static 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
83void 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
97void 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
107static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner)
108{
109 eb->lock_owner = owner;
110}
111#else
112static 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 */
139void 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 */
155int 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 */
167void 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 */
181void 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 */
197void 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 */
213void 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 */
236struct 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 */
259struct 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 */
283struct 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
315void 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 */
324bool 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
341void 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
350void 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
360void 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
375void 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}
1/*
2 * Copyright (C) 2008 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18#include <linux/sched.h>
19#include <linux/pagemap.h>
20#include <linux/spinlock.h>
21#include <linux/page-flags.h>
22#include <asm/bug.h>
23#include "ctree.h"
24#include "extent_io.h"
25#include "locking.h"
26
27static void btrfs_assert_tree_read_locked(struct extent_buffer *eb);
28
29/*
30 * if we currently have a spinning reader or writer lock
31 * (indicated by the rw flag) this will bump the count
32 * of blocking holders and drop the spinlock.
33 */
34void btrfs_set_lock_blocking_rw(struct extent_buffer *eb, int rw)
35{
36 /*
37 * no lock is required. The lock owner may change if
38 * we have a read lock, but it won't change to or away
39 * from us. If we have the write lock, we are the owner
40 * and it'll never change.
41 */
42 if (eb->lock_nested && current->pid == eb->lock_owner)
43 return;
44 if (rw == BTRFS_WRITE_LOCK) {
45 if (atomic_read(&eb->blocking_writers) == 0) {
46 WARN_ON(atomic_read(&eb->spinning_writers) != 1);
47 atomic_dec(&eb->spinning_writers);
48 btrfs_assert_tree_locked(eb);
49 atomic_inc(&eb->blocking_writers);
50 write_unlock(&eb->lock);
51 }
52 } else if (rw == BTRFS_READ_LOCK) {
53 btrfs_assert_tree_read_locked(eb);
54 atomic_inc(&eb->blocking_readers);
55 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
56 atomic_dec(&eb->spinning_readers);
57 read_unlock(&eb->lock);
58 }
59}
60
61/*
62 * if we currently have a blocking lock, take the spinlock
63 * and drop our blocking count
64 */
65void btrfs_clear_lock_blocking_rw(struct extent_buffer *eb, int rw)
66{
67 /*
68 * no lock is required. The lock owner may change if
69 * we have a read lock, but it won't change to or away
70 * from us. If we have the write lock, we are the owner
71 * and it'll never change.
72 */
73 if (eb->lock_nested && current->pid == eb->lock_owner)
74 return;
75
76 if (rw == BTRFS_WRITE_LOCK_BLOCKING) {
77 BUG_ON(atomic_read(&eb->blocking_writers) != 1);
78 write_lock(&eb->lock);
79 WARN_ON(atomic_read(&eb->spinning_writers));
80 atomic_inc(&eb->spinning_writers);
81 /*
82 * atomic_dec_and_test implies a barrier for waitqueue_active
83 */
84 if (atomic_dec_and_test(&eb->blocking_writers) &&
85 waitqueue_active(&eb->write_lock_wq))
86 wake_up(&eb->write_lock_wq);
87 } else if (rw == BTRFS_READ_LOCK_BLOCKING) {
88 BUG_ON(atomic_read(&eb->blocking_readers) == 0);
89 read_lock(&eb->lock);
90 atomic_inc(&eb->spinning_readers);
91 /*
92 * atomic_dec_and_test implies a barrier for waitqueue_active
93 */
94 if (atomic_dec_and_test(&eb->blocking_readers) &&
95 waitqueue_active(&eb->read_lock_wq))
96 wake_up(&eb->read_lock_wq);
97 }
98}
99
100/*
101 * take a spinning read lock. This will wait for any blocking
102 * writers
103 */
104void btrfs_tree_read_lock(struct extent_buffer *eb)
105{
106again:
107 BUG_ON(!atomic_read(&eb->blocking_writers) &&
108 current->pid == eb->lock_owner);
109
110 read_lock(&eb->lock);
111 if (atomic_read(&eb->blocking_writers) &&
112 current->pid == eb->lock_owner) {
113 /*
114 * This extent is already write-locked by our thread. We allow
115 * an additional read lock to be added because it's for the same
116 * thread. btrfs_find_all_roots() depends on this as it may be
117 * called on a partly (write-)locked tree.
118 */
119 BUG_ON(eb->lock_nested);
120 eb->lock_nested = 1;
121 read_unlock(&eb->lock);
122 return;
123 }
124 if (atomic_read(&eb->blocking_writers)) {
125 read_unlock(&eb->lock);
126 wait_event(eb->write_lock_wq,
127 atomic_read(&eb->blocking_writers) == 0);
128 goto again;
129 }
130 atomic_inc(&eb->read_locks);
131 atomic_inc(&eb->spinning_readers);
132}
133
134/*
135 * take a spinning read lock.
136 * returns 1 if we get the read lock and 0 if we don't
137 * this won't wait for blocking writers
138 */
139int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
140{
141 if (atomic_read(&eb->blocking_writers))
142 return 0;
143
144 read_lock(&eb->lock);
145 if (atomic_read(&eb->blocking_writers)) {
146 read_unlock(&eb->lock);
147 return 0;
148 }
149 atomic_inc(&eb->read_locks);
150 atomic_inc(&eb->spinning_readers);
151 return 1;
152}
153
154/*
155 * returns 1 if we get the read lock and 0 if we don't
156 * this won't wait for blocking writers
157 */
158int btrfs_try_tree_read_lock(struct extent_buffer *eb)
159{
160 if (atomic_read(&eb->blocking_writers))
161 return 0;
162
163 if (!read_trylock(&eb->lock))
164 return 0;
165
166 if (atomic_read(&eb->blocking_writers)) {
167 read_unlock(&eb->lock);
168 return 0;
169 }
170 atomic_inc(&eb->read_locks);
171 atomic_inc(&eb->spinning_readers);
172 return 1;
173}
174
175/*
176 * returns 1 if we get the read lock and 0 if we don't
177 * this won't wait for blocking writers or readers
178 */
179int btrfs_try_tree_write_lock(struct extent_buffer *eb)
180{
181 if (atomic_read(&eb->blocking_writers) ||
182 atomic_read(&eb->blocking_readers))
183 return 0;
184
185 write_lock(&eb->lock);
186 if (atomic_read(&eb->blocking_writers) ||
187 atomic_read(&eb->blocking_readers)) {
188 write_unlock(&eb->lock);
189 return 0;
190 }
191 atomic_inc(&eb->write_locks);
192 atomic_inc(&eb->spinning_writers);
193 eb->lock_owner = current->pid;
194 return 1;
195}
196
197/*
198 * drop a spinning read lock
199 */
200void btrfs_tree_read_unlock(struct extent_buffer *eb)
201{
202 /*
203 * if we're nested, we have the write lock. No new locking
204 * is needed as long as we are the lock owner.
205 * The write unlock will do a barrier for us, and the lock_nested
206 * field only matters to the lock owner.
207 */
208 if (eb->lock_nested && current->pid == eb->lock_owner) {
209 eb->lock_nested = 0;
210 return;
211 }
212 btrfs_assert_tree_read_locked(eb);
213 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
214 atomic_dec(&eb->spinning_readers);
215 atomic_dec(&eb->read_locks);
216 read_unlock(&eb->lock);
217}
218
219/*
220 * drop a blocking read lock
221 */
222void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
223{
224 /*
225 * if we're nested, we have the write lock. No new locking
226 * is needed as long as we are the lock owner.
227 * The write unlock will do a barrier for us, and the lock_nested
228 * field only matters to the lock owner.
229 */
230 if (eb->lock_nested && current->pid == eb->lock_owner) {
231 eb->lock_nested = 0;
232 return;
233 }
234 btrfs_assert_tree_read_locked(eb);
235 WARN_ON(atomic_read(&eb->blocking_readers) == 0);
236 /*
237 * atomic_dec_and_test implies a barrier for waitqueue_active
238 */
239 if (atomic_dec_and_test(&eb->blocking_readers) &&
240 waitqueue_active(&eb->read_lock_wq))
241 wake_up(&eb->read_lock_wq);
242 atomic_dec(&eb->read_locks);
243}
244
245/*
246 * take a spinning write lock. This will wait for both
247 * blocking readers or writers
248 */
249void btrfs_tree_lock(struct extent_buffer *eb)
250{
251 WARN_ON(eb->lock_owner == current->pid);
252again:
253 wait_event(eb->read_lock_wq, atomic_read(&eb->blocking_readers) == 0);
254 wait_event(eb->write_lock_wq, atomic_read(&eb->blocking_writers) == 0);
255 write_lock(&eb->lock);
256 if (atomic_read(&eb->blocking_readers)) {
257 write_unlock(&eb->lock);
258 wait_event(eb->read_lock_wq,
259 atomic_read(&eb->blocking_readers) == 0);
260 goto again;
261 }
262 if (atomic_read(&eb->blocking_writers)) {
263 write_unlock(&eb->lock);
264 wait_event(eb->write_lock_wq,
265 atomic_read(&eb->blocking_writers) == 0);
266 goto again;
267 }
268 WARN_ON(atomic_read(&eb->spinning_writers));
269 atomic_inc(&eb->spinning_writers);
270 atomic_inc(&eb->write_locks);
271 eb->lock_owner = current->pid;
272}
273
274/*
275 * drop a spinning or a blocking write lock.
276 */
277void btrfs_tree_unlock(struct extent_buffer *eb)
278{
279 int blockers = atomic_read(&eb->blocking_writers);
280
281 BUG_ON(blockers > 1);
282
283 btrfs_assert_tree_locked(eb);
284 eb->lock_owner = 0;
285 atomic_dec(&eb->write_locks);
286
287 if (blockers) {
288 WARN_ON(atomic_read(&eb->spinning_writers));
289 atomic_dec(&eb->blocking_writers);
290 /*
291 * Make sure counter is updated before we wake up waiters.
292 */
293 smp_mb();
294 if (waitqueue_active(&eb->write_lock_wq))
295 wake_up(&eb->write_lock_wq);
296 } else {
297 WARN_ON(atomic_read(&eb->spinning_writers) != 1);
298 atomic_dec(&eb->spinning_writers);
299 write_unlock(&eb->lock);
300 }
301}
302
303void btrfs_assert_tree_locked(struct extent_buffer *eb)
304{
305 BUG_ON(!atomic_read(&eb->write_locks));
306}
307
308static void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
309{
310 BUG_ON(!atomic_read(&eb->read_locks));
311}