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#include "accessors.h"
17
18/*
19 * Lockdep class keys for extent_buffer->lock's in this root. For a given
20 * eb, the lockdep key is determined by the btrfs_root it belongs to and
21 * the level the eb occupies in the tree.
22 *
23 * Different roots are used for different purposes and may nest inside each
24 * other and they require separate keysets. As lockdep keys should be
25 * static, assign keysets according to the purpose of the root as indicated
26 * by btrfs_root->root_key.objectid. This ensures that all special purpose
27 * roots have separate keysets.
28 *
29 * Lock-nesting across peer nodes is always done with the immediate parent
30 * node locked thus preventing deadlock. As lockdep doesn't know this, use
31 * subclass to avoid triggering lockdep warning in such cases.
32 *
33 * The key is set by the readpage_end_io_hook after the buffer has passed
34 * csum validation but before the pages are unlocked. It is also set by
35 * btrfs_init_new_buffer on freshly allocated blocks.
36 *
37 * We also add a check to make sure the highest level of the tree is the
38 * same as our lockdep setup here. If BTRFS_MAX_LEVEL changes, this code
39 * needs update as well.
40 */
41#ifdef CONFIG_DEBUG_LOCK_ALLOC
42#if BTRFS_MAX_LEVEL != 8
43#error
44#endif
45
46#define DEFINE_LEVEL(stem, level) \
47 .names[level] = "btrfs-" stem "-0" #level,
48
49#define DEFINE_NAME(stem) \
50 DEFINE_LEVEL(stem, 0) \
51 DEFINE_LEVEL(stem, 1) \
52 DEFINE_LEVEL(stem, 2) \
53 DEFINE_LEVEL(stem, 3) \
54 DEFINE_LEVEL(stem, 4) \
55 DEFINE_LEVEL(stem, 5) \
56 DEFINE_LEVEL(stem, 6) \
57 DEFINE_LEVEL(stem, 7)
58
59static struct btrfs_lockdep_keyset {
60 u64 id; /* root objectid */
61 /* Longest entry: btrfs-block-group-00 */
62 char names[BTRFS_MAX_LEVEL][24];
63 struct lock_class_key keys[BTRFS_MAX_LEVEL];
64} btrfs_lockdep_keysets[] = {
65 { .id = BTRFS_ROOT_TREE_OBJECTID, DEFINE_NAME("root") },
66 { .id = BTRFS_EXTENT_TREE_OBJECTID, DEFINE_NAME("extent") },
67 { .id = BTRFS_CHUNK_TREE_OBJECTID, DEFINE_NAME("chunk") },
68 { .id = BTRFS_DEV_TREE_OBJECTID, DEFINE_NAME("dev") },
69 { .id = BTRFS_CSUM_TREE_OBJECTID, DEFINE_NAME("csum") },
70 { .id = BTRFS_QUOTA_TREE_OBJECTID, DEFINE_NAME("quota") },
71 { .id = BTRFS_TREE_LOG_OBJECTID, DEFINE_NAME("log") },
72 { .id = BTRFS_TREE_RELOC_OBJECTID, DEFINE_NAME("treloc") },
73 { .id = BTRFS_DATA_RELOC_TREE_OBJECTID, DEFINE_NAME("dreloc") },
74 { .id = BTRFS_UUID_TREE_OBJECTID, DEFINE_NAME("uuid") },
75 { .id = BTRFS_FREE_SPACE_TREE_OBJECTID, DEFINE_NAME("free-space") },
76 { .id = BTRFS_BLOCK_GROUP_TREE_OBJECTID, DEFINE_NAME("block-group") },
77 { .id = BTRFS_RAID_STRIPE_TREE_OBJECTID, DEFINE_NAME("raid-stripe") },
78 { .id = 0, DEFINE_NAME("tree") },
79};
80
81#undef DEFINE_LEVEL
82#undef DEFINE_NAME
83
84void btrfs_set_buffer_lockdep_class(u64 objectid, struct extent_buffer *eb, int level)
85{
86 struct btrfs_lockdep_keyset *ks;
87
88 BUG_ON(level >= ARRAY_SIZE(ks->keys));
89
90 /* Find the matching keyset, id 0 is the default entry */
91 for (ks = btrfs_lockdep_keysets; ks->id; ks++)
92 if (ks->id == objectid)
93 break;
94
95 lockdep_set_class_and_name(&eb->lock, &ks->keys[level], ks->names[level]);
96}
97
98void btrfs_maybe_reset_lockdep_class(struct btrfs_root *root, struct extent_buffer *eb)
99{
100 if (test_bit(BTRFS_ROOT_RESET_LOCKDEP_CLASS, &root->state))
101 btrfs_set_buffer_lockdep_class(root->root_key.objectid,
102 eb, btrfs_header_level(eb));
103}
104
105#endif
106
107#ifdef CONFIG_BTRFS_DEBUG
108static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner)
109{
110 eb->lock_owner = owner;
111}
112#else
113static void btrfs_set_eb_lock_owner(struct extent_buffer *eb, pid_t owner) { }
114#endif
115
116/*
117 * Extent buffer locking
118 * =====================
119 *
120 * We use a rw_semaphore for tree locking, and the semantics are exactly the
121 * same:
122 *
123 * - reader/writer exclusion
124 * - writer/writer exclusion
125 * - reader/reader sharing
126 * - try-lock semantics for readers and writers
127 *
128 * The rwsem implementation does opportunistic spinning which reduces number of
129 * times the locking task needs to sleep.
130 */
131
132/*
133 * __btrfs_tree_read_lock - lock extent buffer for read
134 * @eb: the eb to be locked
135 * @nest: the nesting level to be used for lockdep
136 *
137 * This takes the read lock on the extent buffer, using the specified nesting
138 * level for lockdep purposes.
139 */
140void __btrfs_tree_read_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
141{
142 u64 start_ns = 0;
143
144 if (trace_btrfs_tree_read_lock_enabled())
145 start_ns = ktime_get_ns();
146
147 down_read_nested(&eb->lock, nest);
148 trace_btrfs_tree_read_lock(eb, start_ns);
149}
150
151void btrfs_tree_read_lock(struct extent_buffer *eb)
152{
153 __btrfs_tree_read_lock(eb, BTRFS_NESTING_NORMAL);
154}
155
156/*
157 * Try-lock for read.
158 *
159 * Return 1 if the rwlock has been taken, 0 otherwise
160 */
161int btrfs_try_tree_read_lock(struct extent_buffer *eb)
162{
163 if (down_read_trylock(&eb->lock)) {
164 trace_btrfs_try_tree_read_lock(eb);
165 return 1;
166 }
167 return 0;
168}
169
170/*
171 * Try-lock for write.
172 *
173 * Return 1 if the rwlock has been taken, 0 otherwise
174 */
175int btrfs_try_tree_write_lock(struct extent_buffer *eb)
176{
177 if (down_write_trylock(&eb->lock)) {
178 btrfs_set_eb_lock_owner(eb, current->pid);
179 trace_btrfs_try_tree_write_lock(eb);
180 return 1;
181 }
182 return 0;
183}
184
185/*
186 * Release read lock.
187 */
188void btrfs_tree_read_unlock(struct extent_buffer *eb)
189{
190 trace_btrfs_tree_read_unlock(eb);
191 up_read(&eb->lock);
192}
193
194/*
195 * Lock eb for write.
196 *
197 * @eb: the eb to lock
198 * @nest: the nesting to use for the lock
199 *
200 * Returns with the eb->lock write locked.
201 */
202void __btrfs_tree_lock(struct extent_buffer *eb, enum btrfs_lock_nesting nest)
203 __acquires(&eb->lock)
204{
205 u64 start_ns = 0;
206
207 if (trace_btrfs_tree_lock_enabled())
208 start_ns = ktime_get_ns();
209
210 down_write_nested(&eb->lock, nest);
211 btrfs_set_eb_lock_owner(eb, current->pid);
212 trace_btrfs_tree_lock(eb, start_ns);
213}
214
215void btrfs_tree_lock(struct extent_buffer *eb)
216{
217 __btrfs_tree_lock(eb, BTRFS_NESTING_NORMAL);
218}
219
220/*
221 * Release the write lock.
222 */
223void btrfs_tree_unlock(struct extent_buffer *eb)
224{
225 trace_btrfs_tree_unlock(eb);
226 btrfs_set_eb_lock_owner(eb, 0);
227 up_write(&eb->lock);
228}
229
230/*
231 * This releases any locks held in the path starting at level and going all the
232 * way up to the root.
233 *
234 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
235 * cases, such as COW of the block at slot zero in the node. This ignores
236 * those rules, and it should only be called when there are no more updates to
237 * be done higher up in the tree.
238 */
239void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
240{
241 int i;
242
243 if (path->keep_locks)
244 return;
245
246 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
247 if (!path->nodes[i])
248 continue;
249 if (!path->locks[i])
250 continue;
251 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
252 path->locks[i] = 0;
253 }
254}
255
256/*
257 * Loop around taking references on and locking the root node of the tree until
258 * we end up with a lock on the root node.
259 *
260 * Return: root extent buffer with write lock held
261 */
262struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
263{
264 struct extent_buffer *eb;
265
266 while (1) {
267 eb = btrfs_root_node(root);
268
269 btrfs_maybe_reset_lockdep_class(root, eb);
270 btrfs_tree_lock(eb);
271 if (eb == root->node)
272 break;
273 btrfs_tree_unlock(eb);
274 free_extent_buffer(eb);
275 }
276 return eb;
277}
278
279/*
280 * Loop around taking references on and locking the root node of the tree until
281 * we end up with a lock on the root node.
282 *
283 * Return: root extent buffer with read lock held
284 */
285struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
286{
287 struct extent_buffer *eb;
288
289 while (1) {
290 eb = btrfs_root_node(root);
291
292 btrfs_maybe_reset_lockdep_class(root, eb);
293 btrfs_tree_read_lock(eb);
294 if (eb == root->node)
295 break;
296 btrfs_tree_read_unlock(eb);
297 free_extent_buffer(eb);
298 }
299 return eb;
300}
301
302/*
303 * Loop around taking references on and locking the root node of the tree in
304 * nowait mode until we end up with a lock on the root node or returning to
305 * avoid blocking.
306 *
307 * Return: root extent buffer with read lock held or -EAGAIN.
308 */
309struct extent_buffer *btrfs_try_read_lock_root_node(struct btrfs_root *root)
310{
311 struct extent_buffer *eb;
312
313 while (1) {
314 eb = btrfs_root_node(root);
315 if (!btrfs_try_tree_read_lock(eb)) {
316 free_extent_buffer(eb);
317 return ERR_PTR(-EAGAIN);
318 }
319 if (eb == root->node)
320 break;
321 btrfs_tree_read_unlock(eb);
322 free_extent_buffer(eb);
323 }
324 return eb;
325}
326
327/*
328 * DREW locks
329 * ==========
330 *
331 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
332 * where you want to provide A-B exclusion but not AA or BB.
333 *
334 * Currently implementation gives more priority to reader. If a reader and a
335 * writer both race to acquire their respective sides of the lock the writer
336 * would yield its lock as soon as it detects a concurrent reader. Additionally
337 * if there are pending readers no new writers would be allowed to come in and
338 * acquire the lock.
339 */
340
341void btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
342{
343 atomic_set(&lock->readers, 0);
344 atomic_set(&lock->writers, 0);
345 init_waitqueue_head(&lock->pending_readers);
346 init_waitqueue_head(&lock->pending_writers);
347}
348
349/* Return true if acquisition is successful, false otherwise */
350bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
351{
352 if (atomic_read(&lock->readers))
353 return false;
354
355 atomic_inc(&lock->writers);
356
357 /* Ensure writers count is updated before we check for pending readers */
358 smp_mb__after_atomic();
359 if (atomic_read(&lock->readers)) {
360 btrfs_drew_write_unlock(lock);
361 return false;
362 }
363
364 return true;
365}
366
367void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
368{
369 while (true) {
370 if (btrfs_drew_try_write_lock(lock))
371 return;
372 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
373 }
374}
375
376void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
377{
378 atomic_dec(&lock->writers);
379 cond_wake_up(&lock->pending_readers);
380}
381
382void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
383{
384 atomic_inc(&lock->readers);
385
386 /*
387 * Ensure the pending reader count is perceieved BEFORE this reader
388 * goes to sleep in case of active writers. This guarantees new writers
389 * won't be allowed and that the current reader will be woken up when
390 * the last active writer finishes its jobs.
391 */
392 smp_mb__after_atomic();
393
394 wait_event(lock->pending_readers, atomic_read(&lock->writers) == 0);
395}
396
397void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
398{
399 /*
400 * atomic_dec_and_test implies a full barrier, so woken up writers
401 * are guaranteed to see the decrement
402 */
403 if (atomic_dec_and_test(&lock->readers))
404 wake_up(&lock->pending_writers);
405}
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 * The locks use a custom scheme that allows to do more operations than are
21 * available fromt current locking primitives. The building blocks are still
22 * rwlock and wait queues.
23 *
24 * Required semantics:
25 *
26 * - reader/writer exclusion
27 * - writer/writer exclusion
28 * - reader/reader sharing
29 * - spinning lock semantics
30 * - blocking lock semantics
31 * - try-lock semantics for readers and writers
32 * - one level nesting, allowing read lock to be taken by the same thread that
33 * already has write lock
34 *
35 * The extent buffer locks (also called tree locks) manage access to eb data
36 * related to the storage in the b-tree (keys, items, but not the individual
37 * members of eb).
38 * We want concurrency of many readers and safe updates. The underlying locking
39 * is done by read-write spinlock and the blocking part is implemented using
40 * counters and wait queues.
41 *
42 * spinning semantics - the low-level rwlock is held so all other threads that
43 * want to take it are spinning on it.
44 *
45 * blocking semantics - the low-level rwlock is not held but the counter
46 * denotes how many times the blocking lock was held;
47 * sleeping is possible
48 *
49 * Write lock always allows only one thread to access the data.
50 *
51 *
52 * Debugging
53 * ---------
54 *
55 * There are additional state counters that are asserted in various contexts,
56 * removed from non-debug build to reduce extent_buffer size and for
57 * performance reasons.
58 *
59 *
60 * Lock nesting
61 * ------------
62 *
63 * A write operation on a tree might indirectly start a look up on the same
64 * tree. This can happen when btrfs_cow_block locks the tree and needs to
65 * lookup free extents.
66 *
67 * btrfs_cow_block
68 * ..
69 * alloc_tree_block_no_bg_flush
70 * btrfs_alloc_tree_block
71 * btrfs_reserve_extent
72 * ..
73 * load_free_space_cache
74 * ..
75 * btrfs_lookup_file_extent
76 * btrfs_search_slot
77 *
78 *
79 * Locking pattern - spinning
80 * --------------------------
81 *
82 * The simple locking scenario, the +--+ denotes the spinning section.
83 *
84 * +- btrfs_tree_lock
85 * | - extent_buffer::rwlock is held
86 * | - no heavy operations should happen, eg. IO, memory allocations, large
87 * | structure traversals
88 * +- btrfs_tree_unock
89*
90*
91 * Locking pattern - blocking
92 * --------------------------
93 *
94 * The blocking write uses the following scheme. The +--+ denotes the spinning
95 * section.
96 *
97 * +- btrfs_tree_lock
98 * |
99 * +- btrfs_set_lock_blocking_write
100 *
101 * - allowed: IO, memory allocations, etc.
102 *
103 * -- btrfs_tree_unlock - note, no explicit unblocking necessary
104 *
105 *
106 * Blocking read is similar.
107 *
108 * +- btrfs_tree_read_lock
109 * |
110 * +- btrfs_set_lock_blocking_read
111 *
112 * - heavy operations allowed
113 *
114 * +- btrfs_tree_read_unlock_blocking
115 * |
116 * +- btrfs_tree_read_unlock
117 *
118 */
119
120#ifdef CONFIG_BTRFS_DEBUG
121static inline void btrfs_assert_spinning_writers_get(struct extent_buffer *eb)
122{
123 WARN_ON(eb->spinning_writers);
124 eb->spinning_writers++;
125}
126
127static inline void btrfs_assert_spinning_writers_put(struct extent_buffer *eb)
128{
129 WARN_ON(eb->spinning_writers != 1);
130 eb->spinning_writers--;
131}
132
133static inline void btrfs_assert_no_spinning_writers(struct extent_buffer *eb)
134{
135 WARN_ON(eb->spinning_writers);
136}
137
138static inline void btrfs_assert_spinning_readers_get(struct extent_buffer *eb)
139{
140 atomic_inc(&eb->spinning_readers);
141}
142
143static inline void btrfs_assert_spinning_readers_put(struct extent_buffer *eb)
144{
145 WARN_ON(atomic_read(&eb->spinning_readers) == 0);
146 atomic_dec(&eb->spinning_readers);
147}
148
149static inline void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb)
150{
151 atomic_inc(&eb->read_locks);
152}
153
154static inline void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb)
155{
156 atomic_dec(&eb->read_locks);
157}
158
159static inline void btrfs_assert_tree_read_locked(struct extent_buffer *eb)
160{
161 BUG_ON(!atomic_read(&eb->read_locks));
162}
163
164static inline void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb)
165{
166 eb->write_locks++;
167}
168
169static inline void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb)
170{
171 eb->write_locks--;
172}
173
174#else
175static void btrfs_assert_spinning_writers_get(struct extent_buffer *eb) { }
176static void btrfs_assert_spinning_writers_put(struct extent_buffer *eb) { }
177static void btrfs_assert_no_spinning_writers(struct extent_buffer *eb) { }
178static void btrfs_assert_spinning_readers_put(struct extent_buffer *eb) { }
179static void btrfs_assert_spinning_readers_get(struct extent_buffer *eb) { }
180static void btrfs_assert_tree_read_locked(struct extent_buffer *eb) { }
181static void btrfs_assert_tree_read_locks_get(struct extent_buffer *eb) { }
182static void btrfs_assert_tree_read_locks_put(struct extent_buffer *eb) { }
183static void btrfs_assert_tree_write_locks_get(struct extent_buffer *eb) { }
184static void btrfs_assert_tree_write_locks_put(struct extent_buffer *eb) { }
185#endif
186
187/*
188 * Mark already held read lock as blocking. Can be nested in write lock by the
189 * same thread.
190 *
191 * Use when there are potentially long operations ahead so other thread waiting
192 * on the lock will not actively spin but sleep instead.
193 *
194 * The rwlock is released and blocking reader counter is increased.
195 */
196void btrfs_set_lock_blocking_read(struct extent_buffer *eb)
197{
198 trace_btrfs_set_lock_blocking_read(eb);
199 /*
200 * No lock is required. The lock owner may change if we have a read
201 * lock, but it won't change to or away from us. If we have the write
202 * lock, we are the owner and it'll never change.
203 */
204 if (eb->lock_nested && current->pid == eb->lock_owner)
205 return;
206 btrfs_assert_tree_read_locked(eb);
207 atomic_inc(&eb->blocking_readers);
208 btrfs_assert_spinning_readers_put(eb);
209 read_unlock(&eb->lock);
210}
211
212/*
213 * Mark already held write lock as blocking.
214 *
215 * Use when there are potentially long operations ahead so other threads
216 * waiting on the lock will not actively spin but sleep instead.
217 *
218 * The rwlock is released and blocking writers is set.
219 */
220void btrfs_set_lock_blocking_write(struct extent_buffer *eb)
221{
222 trace_btrfs_set_lock_blocking_write(eb);
223 /*
224 * No lock is required. The lock owner may change if we have a read
225 * lock, but it won't change to or away from us. If we have the write
226 * lock, we are the owner and it'll never change.
227 */
228 if (eb->lock_nested && current->pid == eb->lock_owner)
229 return;
230 if (eb->blocking_writers == 0) {
231 btrfs_assert_spinning_writers_put(eb);
232 btrfs_assert_tree_locked(eb);
233 WRITE_ONCE(eb->blocking_writers, 1);
234 write_unlock(&eb->lock);
235 }
236}
237
238/*
239 * Lock the extent buffer for read. Wait for any writers (spinning or blocking).
240 * Can be nested in write lock by the same thread.
241 *
242 * Use when the locked section does only lightweight actions and busy waiting
243 * would be cheaper than making other threads do the wait/wake loop.
244 *
245 * The rwlock is held upon exit.
246 */
247void btrfs_tree_read_lock(struct extent_buffer *eb)
248{
249 u64 start_ns = 0;
250
251 if (trace_btrfs_tree_read_lock_enabled())
252 start_ns = ktime_get_ns();
253again:
254 read_lock(&eb->lock);
255 BUG_ON(eb->blocking_writers == 0 &&
256 current->pid == eb->lock_owner);
257 if (eb->blocking_writers) {
258 if (current->pid == eb->lock_owner) {
259 /*
260 * This extent is already write-locked by our thread.
261 * We allow an additional read lock to be added because
262 * it's for the same thread. btrfs_find_all_roots()
263 * depends on this as it may be called on a partly
264 * (write-)locked tree.
265 */
266 BUG_ON(eb->lock_nested);
267 eb->lock_nested = true;
268 read_unlock(&eb->lock);
269 trace_btrfs_tree_read_lock(eb, start_ns);
270 return;
271 }
272 read_unlock(&eb->lock);
273 wait_event(eb->write_lock_wq,
274 READ_ONCE(eb->blocking_writers) == 0);
275 goto again;
276 }
277 btrfs_assert_tree_read_locks_get(eb);
278 btrfs_assert_spinning_readers_get(eb);
279 trace_btrfs_tree_read_lock(eb, start_ns);
280}
281
282/*
283 * Lock extent buffer for read, optimistically expecting that there are no
284 * contending blocking writers. If there are, don't wait.
285 *
286 * Return 1 if the rwlock has been taken, 0 otherwise
287 */
288int btrfs_tree_read_lock_atomic(struct extent_buffer *eb)
289{
290 if (READ_ONCE(eb->blocking_writers))
291 return 0;
292
293 read_lock(&eb->lock);
294 /* Refetch value after lock */
295 if (READ_ONCE(eb->blocking_writers)) {
296 read_unlock(&eb->lock);
297 return 0;
298 }
299 btrfs_assert_tree_read_locks_get(eb);
300 btrfs_assert_spinning_readers_get(eb);
301 trace_btrfs_tree_read_lock_atomic(eb);
302 return 1;
303}
304
305/*
306 * Try-lock for read. Don't block or wait for contending writers.
307 *
308 * Retrun 1 if the rwlock has been taken, 0 otherwise
309 */
310int btrfs_try_tree_read_lock(struct extent_buffer *eb)
311{
312 if (READ_ONCE(eb->blocking_writers))
313 return 0;
314
315 if (!read_trylock(&eb->lock))
316 return 0;
317
318 /* Refetch value after lock */
319 if (READ_ONCE(eb->blocking_writers)) {
320 read_unlock(&eb->lock);
321 return 0;
322 }
323 btrfs_assert_tree_read_locks_get(eb);
324 btrfs_assert_spinning_readers_get(eb);
325 trace_btrfs_try_tree_read_lock(eb);
326 return 1;
327}
328
329/*
330 * Try-lock for write. May block until the lock is uncontended, but does not
331 * wait until it is free.
332 *
333 * Retrun 1 if the rwlock has been taken, 0 otherwise
334 */
335int btrfs_try_tree_write_lock(struct extent_buffer *eb)
336{
337 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers))
338 return 0;
339
340 write_lock(&eb->lock);
341 /* Refetch value after lock */
342 if (READ_ONCE(eb->blocking_writers) || atomic_read(&eb->blocking_readers)) {
343 write_unlock(&eb->lock);
344 return 0;
345 }
346 btrfs_assert_tree_write_locks_get(eb);
347 btrfs_assert_spinning_writers_get(eb);
348 eb->lock_owner = current->pid;
349 trace_btrfs_try_tree_write_lock(eb);
350 return 1;
351}
352
353/*
354 * Release read lock. Must be used only if the lock is in spinning mode. If
355 * the read lock is nested, must pair with read lock before the write unlock.
356 *
357 * The rwlock is not held upon exit.
358 */
359void btrfs_tree_read_unlock(struct extent_buffer *eb)
360{
361 trace_btrfs_tree_read_unlock(eb);
362 /*
363 * if we're nested, we have the write lock. No new locking
364 * is needed as long as we are the lock owner.
365 * The write unlock will do a barrier for us, and the lock_nested
366 * field only matters to the lock owner.
367 */
368 if (eb->lock_nested && current->pid == eb->lock_owner) {
369 eb->lock_nested = false;
370 return;
371 }
372 btrfs_assert_tree_read_locked(eb);
373 btrfs_assert_spinning_readers_put(eb);
374 btrfs_assert_tree_read_locks_put(eb);
375 read_unlock(&eb->lock);
376}
377
378/*
379 * Release read lock, previously set to blocking by a pairing call to
380 * btrfs_set_lock_blocking_read(). Can be nested in write lock by the same
381 * thread.
382 *
383 * State of rwlock is unchanged, last reader wakes waiting threads.
384 */
385void btrfs_tree_read_unlock_blocking(struct extent_buffer *eb)
386{
387 trace_btrfs_tree_read_unlock_blocking(eb);
388 /*
389 * if we're nested, we have the write lock. No new locking
390 * is needed as long as we are the lock owner.
391 * The write unlock will do a barrier for us, and the lock_nested
392 * field only matters to the lock owner.
393 */
394 if (eb->lock_nested && current->pid == eb->lock_owner) {
395 eb->lock_nested = false;
396 return;
397 }
398 btrfs_assert_tree_read_locked(eb);
399 WARN_ON(atomic_read(&eb->blocking_readers) == 0);
400 /* atomic_dec_and_test implies a barrier */
401 if (atomic_dec_and_test(&eb->blocking_readers))
402 cond_wake_up_nomb(&eb->read_lock_wq);
403 btrfs_assert_tree_read_locks_put(eb);
404}
405
406/*
407 * Lock for write. Wait for all blocking and spinning readers and writers. This
408 * starts context where reader lock could be nested by the same thread.
409 *
410 * The rwlock is held for write upon exit.
411 */
412void btrfs_tree_lock(struct extent_buffer *eb)
413 __acquires(&eb->lock)
414{
415 u64 start_ns = 0;
416
417 if (trace_btrfs_tree_lock_enabled())
418 start_ns = ktime_get_ns();
419
420 WARN_ON(eb->lock_owner == current->pid);
421again:
422 wait_event(eb->read_lock_wq, atomic_read(&eb->blocking_readers) == 0);
423 wait_event(eb->write_lock_wq, READ_ONCE(eb->blocking_writers) == 0);
424 write_lock(&eb->lock);
425 /* Refetch value after lock */
426 if (atomic_read(&eb->blocking_readers) ||
427 READ_ONCE(eb->blocking_writers)) {
428 write_unlock(&eb->lock);
429 goto again;
430 }
431 btrfs_assert_spinning_writers_get(eb);
432 btrfs_assert_tree_write_locks_get(eb);
433 eb->lock_owner = current->pid;
434 trace_btrfs_tree_lock(eb, start_ns);
435}
436
437/*
438 * Release the write lock, either blocking or spinning (ie. there's no need
439 * for an explicit blocking unlock, like btrfs_tree_read_unlock_blocking).
440 * This also ends the context for nesting, the read lock must have been
441 * released already.
442 *
443 * Tasks blocked and waiting are woken, rwlock is not held upon exit.
444 */
445void btrfs_tree_unlock(struct extent_buffer *eb)
446{
447 /*
448 * This is read both locked and unlocked but always by the same thread
449 * that already owns the lock so we don't need to use READ_ONCE
450 */
451 int blockers = eb->blocking_writers;
452
453 BUG_ON(blockers > 1);
454
455 btrfs_assert_tree_locked(eb);
456 trace_btrfs_tree_unlock(eb);
457 eb->lock_owner = 0;
458 btrfs_assert_tree_write_locks_put(eb);
459
460 if (blockers) {
461 btrfs_assert_no_spinning_writers(eb);
462 /* Unlocked write */
463 WRITE_ONCE(eb->blocking_writers, 0);
464 /*
465 * We need to order modifying blocking_writers above with
466 * actually waking up the sleepers to ensure they see the
467 * updated value of blocking_writers
468 */
469 cond_wake_up(&eb->write_lock_wq);
470 } else {
471 btrfs_assert_spinning_writers_put(eb);
472 write_unlock(&eb->lock);
473 }
474}
475
476/*
477 * Set all locked nodes in the path to blocking locks. This should be done
478 * before scheduling
479 */
480void btrfs_set_path_blocking(struct btrfs_path *p)
481{
482 int i;
483
484 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
485 if (!p->nodes[i] || !p->locks[i])
486 continue;
487 /*
488 * If we currently have a spinning reader or writer lock this
489 * will bump the count of blocking holders and drop the
490 * spinlock.
491 */
492 if (p->locks[i] == BTRFS_READ_LOCK) {
493 btrfs_set_lock_blocking_read(p->nodes[i]);
494 p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
495 } else if (p->locks[i] == BTRFS_WRITE_LOCK) {
496 btrfs_set_lock_blocking_write(p->nodes[i]);
497 p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
498 }
499 }
500}
501
502/*
503 * This releases any locks held in the path starting at level and going all the
504 * way up to the root.
505 *
506 * btrfs_search_slot will keep the lock held on higher nodes in a few corner
507 * cases, such as COW of the block at slot zero in the node. This ignores
508 * those rules, and it should only be called when there are no more updates to
509 * be done higher up in the tree.
510 */
511void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
512{
513 int i;
514
515 if (path->keep_locks)
516 return;
517
518 for (i = level; i < BTRFS_MAX_LEVEL; i++) {
519 if (!path->nodes[i])
520 continue;
521 if (!path->locks[i])
522 continue;
523 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
524 path->locks[i] = 0;
525 }
526}
527
528/*
529 * Loop around taking references on and locking the root node of the tree until
530 * we end up with a lock on the root node.
531 *
532 * Return: root extent buffer with write lock held
533 */
534struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
535{
536 struct extent_buffer *eb;
537
538 while (1) {
539 eb = btrfs_root_node(root);
540 btrfs_tree_lock(eb);
541 if (eb == root->node)
542 break;
543 btrfs_tree_unlock(eb);
544 free_extent_buffer(eb);
545 }
546 return eb;
547}
548
549/*
550 * Loop around taking references on and locking the root node of the tree until
551 * we end up with a lock on the root node.
552 *
553 * Return: root extent buffer with read lock held
554 */
555struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
556{
557 struct extent_buffer *eb;
558
559 while (1) {
560 eb = btrfs_root_node(root);
561 btrfs_tree_read_lock(eb);
562 if (eb == root->node)
563 break;
564 btrfs_tree_read_unlock(eb);
565 free_extent_buffer(eb);
566 }
567 return eb;
568}
569
570/*
571 * DREW locks
572 * ==========
573 *
574 * DREW stands for double-reader-writer-exclusion lock. It's used in situation
575 * where you want to provide A-B exclusion but not AA or BB.
576 *
577 * Currently implementation gives more priority to reader. If a reader and a
578 * writer both race to acquire their respective sides of the lock the writer
579 * would yield its lock as soon as it detects a concurrent reader. Additionally
580 * if there are pending readers no new writers would be allowed to come in and
581 * acquire the lock.
582 */
583
584int btrfs_drew_lock_init(struct btrfs_drew_lock *lock)
585{
586 int ret;
587
588 ret = percpu_counter_init(&lock->writers, 0, GFP_KERNEL);
589 if (ret)
590 return ret;
591
592 atomic_set(&lock->readers, 0);
593 init_waitqueue_head(&lock->pending_readers);
594 init_waitqueue_head(&lock->pending_writers);
595
596 return 0;
597}
598
599void btrfs_drew_lock_destroy(struct btrfs_drew_lock *lock)
600{
601 percpu_counter_destroy(&lock->writers);
602}
603
604/* Return true if acquisition is successful, false otherwise */
605bool btrfs_drew_try_write_lock(struct btrfs_drew_lock *lock)
606{
607 if (atomic_read(&lock->readers))
608 return false;
609
610 percpu_counter_inc(&lock->writers);
611
612 /* Ensure writers count is updated before we check for pending readers */
613 smp_mb();
614 if (atomic_read(&lock->readers)) {
615 btrfs_drew_write_unlock(lock);
616 return false;
617 }
618
619 return true;
620}
621
622void btrfs_drew_write_lock(struct btrfs_drew_lock *lock)
623{
624 while (true) {
625 if (btrfs_drew_try_write_lock(lock))
626 return;
627 wait_event(lock->pending_writers, !atomic_read(&lock->readers));
628 }
629}
630
631void btrfs_drew_write_unlock(struct btrfs_drew_lock *lock)
632{
633 percpu_counter_dec(&lock->writers);
634 cond_wake_up(&lock->pending_readers);
635}
636
637void btrfs_drew_read_lock(struct btrfs_drew_lock *lock)
638{
639 atomic_inc(&lock->readers);
640
641 /*
642 * Ensure the pending reader count is perceieved BEFORE this reader
643 * goes to sleep in case of active writers. This guarantees new writers
644 * won't be allowed and that the current reader will be woken up when
645 * the last active writer finishes its jobs.
646 */
647 smp_mb__after_atomic();
648
649 wait_event(lock->pending_readers,
650 percpu_counter_sum(&lock->writers) == 0);
651}
652
653void btrfs_drew_read_unlock(struct btrfs_drew_lock *lock)
654{
655 /*
656 * atomic_dec_and_test implies a full barrier, so woken up writers
657 * are guaranteed to see the decrement
658 */
659 if (atomic_dec_and_test(&lock->readers))
660 wake_up(&lock->pending_writers);
661}