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(objectid: root->root_key.objectid,
101 eb, level: 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 - 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(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(sem: &eb->lock, subclass: nest);
147 trace_btrfs_tree_read_lock(eb, start_ns);
148}
149
150void btrfs_tree_read_lock(struct extent_buffer *eb)
151{
152 __btrfs_tree_read_lock(eb, nest: BTRFS_NESTING_NORMAL);
153}
154
155/*
156 * Try-lock for read.
157 *
158 * Return 1 if the rwlock has been taken, 0 otherwise
159 */
160int btrfs_try_tree_read_lock(struct extent_buffer *eb)
161{
162 if (down_read_trylock(sem: &eb->lock)) {
163 trace_btrfs_try_tree_read_lock(eb);
164 return 1;
165 }
166 return 0;
167}
168
169/*
170 * Try-lock for write.
171 *
172 * Return 1 if the rwlock has been taken, 0 otherwise
173 */
174int btrfs_try_tree_write_lock(struct extent_buffer *eb)
175{
176 if (down_write_trylock(sem: &eb->lock)) {
177 btrfs_set_eb_lock_owner(eb, current->pid);
178 trace_btrfs_try_tree_write_lock(eb);
179 return 1;
180 }
181 return 0;
182}
183
184/*
185 * Release read lock.
186 */
187void btrfs_tree_read_unlock(struct extent_buffer *eb)
188{
189 trace_btrfs_tree_read_unlock(eb);
190 up_read(sem: &eb->lock);
191}
192
193/*
194 * Lock eb for write.
195 *
196 * @eb: the eb to lock
197 * @nest: the nesting to use for the lock
198 *
199 * Returns with the eb->lock write locked.
200 */
201void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
202 __acquires(&eb->lock)
203{
204 u64 start_ns = 0;
205
206 if (trace_btrfs_tree_lock_enabled())
207 start_ns = ktime_get_ns();
208
209 down_write_nested(sem: &eb->lock, subclass: nest);
210 btrfs_set_eb_lock_owner(eb, current->pid);
211 trace_btrfs_tree_lock(eb, start_ns);
212}
213
214void btrfs_tree_lock(struct extent_buffer *eb)
215{
216 __btrfs_tree_lock(eb, nest: BTRFS_NESTING_NORMAL);
217}
218
219/*
220 * Release the write lock.
221 */
222void btrfs_tree_unlock(struct extent_buffer *eb)
223{
224 trace_btrfs_tree_unlock(eb);
225 btrfs_set_eb_lock_owner(eb, owner: 0);
226 up_write(sem: &eb->lock);
227}
228
229/*
230 * This releases any locks held in the path starting at level and going all the
231 * way up to the root.
232 *
233 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
234 * cases, such as COW of the block at slot zero in the node. This ignores
235 * those rules, and it should only be called when there are no more updates to
236 * be done higher up in the tree.
237 */
238void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
239{
240 int i;
241
242 if (path->keep_locks)
243 return;
244
245 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
246 if (!path->nodes[i])
247 continue;
248 if (!path->locks[i])
249 continue;
250 btrfs_tree_unlock_rw(eb: path->nodes[i], rw: path->locks[i]);
251 path->locks[i] = 0;
252 }
253}
254
255/*
256 * Loop around taking references on and locking the root node of the tree until
257 * we end up with a lock on the root node.
258 *
259 * Return: root extent buffer with write lock held
260 */
261struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
262{
263 struct extent_buffer *eb;
264
265 while (1) {
266 eb = btrfs_root_node(root);
267
268 btrfs_maybe_reset_lockdep_class(root, eb);
269 btrfs_tree_lock(eb);
270 if (eb == root->node)
271 break;
272 btrfs_tree_unlock(eb);
273 free_extent_buffer(eb);
274 }
275 return eb;
276}
277
278/*
279 * Loop around taking references on and locking the root node of the tree until
280 * we end up with a lock on the root node.
281 *
282 * Return: root extent buffer with read lock held
283 */
284struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
285{
286 struct extent_buffer *eb;
287
288 while (1) {
289 eb = btrfs_root_node(root);
290
291 btrfs_maybe_reset_lockdep_class(root, eb);
292 btrfs_tree_read_lock(eb);
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 * Loop around taking references on and locking the root node of the tree in
303 * nowait mode until we end up with a lock on the root node or returning to
304 * avoid blocking.
305 *
306 * Return: root extent buffer with read lock held or -EAGAIN.
307 */
308struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
309{
310 struct extent_buffer *eb;
311
312 while (1) {
313 eb = btrfs_root_node(root);
314 if (!btrfs_try_tree_read_lock(eb)) {
315 free_extent_buffer(eb);
316 return ERR_PTR(error: -EAGAIN);
317 }
318 if (eb == root->node)
319 break;
320 btrfs_tree_read_unlock(eb);
321 free_extent_buffer(eb);
322 }
323 return eb;
324}
325
326/*
327 * DREW locks
328 * ==========
329 *
330 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
331 * where you want to provide A-B exclusion but not AA or BB.
332 *
333 * Currently implementation gives more priority to reader. If a reader and a
334 * writer both race to acquire their respective sides of the lock the writer
335 * would yield its lock as soon as it detects a concurrent reader. Additionally
336 * if there are pending readers no new writers would be allowed to come in and
337 * acquire the lock.
338 */
339
340void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
341{
342 atomic_set(v: &lock->readers, i: 0);
343 atomic_set(v: &lock->writers, i: 0);
344 init_waitqueue_head(&lock->pending_readers);
345 init_waitqueue_head(&lock->pending_writers);
346}
347
348/* Return true if acquisition is successful, false otherwise */
349bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
350{
351 if (atomic_read(v: &lock->readers))
352 return false;
353
354 atomic_inc(v: &lock->writers);
355
356 /* Ensure writers count is updated before we check for pending readers */
357 smp_mb__after_atomic();
358 if (atomic_read(v: &lock->readers)) {
359 btrfs_drew_write_unlock(lock);
360 return false;
361 }
362
363 return true;
364}
365
366void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
367{
368 while (true) {
369 if (btrfs_drew_try_write_lock(lock))
370 return;
371 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
372 }
373}
374
375void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
376{
377 atomic_dec(v: &lock->writers);
378 cond_wake_up(wq: &lock->pending_readers);
379}
380
381void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
382{
383 atomic_inc(v: &lock->readers);
384
385 /*
386 * Ensure the pending reader count is perceieved BEFORE this reader
387 * goes to sleep in case of active writers. This guarantees new writers
388 * won't be allowed and that the current reader will be woken up when
389 * the last active writer finishes its jobs.
390 */
391 smp_mb__after_atomic();
392
393 wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
394}
395
396void 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(v: &lock->readers))
403 wake_up(&lock->pending_writers);
404}
405

source code of linux/fs/btrfs/locking.c