1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | #ifndef _BCACHEFS_BTREE_LOCKING_H |
3 | #define _BCACHEFS_BTREE_LOCKING_H |
4 | |
5 | /* |
6 | * Only for internal btree use: |
7 | * |
8 | * The btree iterator tracks what locks it wants to take, and what locks it |
9 | * currently has - here we have wrappers for locking/unlocking btree nodes and |
10 | * updating the iterator state |
11 | */ |
12 | |
13 | #include "btree_iter.h" |
14 | #include "six.h" |
15 | |
16 | void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags); |
17 | |
18 | #ifdef CONFIG_LOCKDEP |
19 | void bch2_assert_btree_nodes_not_locked(void); |
20 | #else |
21 | static inline void bch2_assert_btree_nodes_not_locked(void) {} |
22 | #endif |
23 | |
24 | void bch2_trans_unlock_noassert(struct btree_trans *); |
25 | |
26 | static inline bool is_btree_node(struct btree_path *path, unsigned l) |
27 | { |
28 | return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(ptr: path->l[l].b); |
29 | } |
30 | |
31 | static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans) |
32 | { |
33 | return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats) |
34 | ? &trans->c->btree_transaction_stats[trans->fn_idx] |
35 | : NULL; |
36 | } |
37 | |
38 | /* matches six lock types */ |
39 | enum btree_node_locked_type { |
40 | BTREE_NODE_UNLOCKED = -1, |
41 | BTREE_NODE_READ_LOCKED = SIX_LOCK_read, |
42 | BTREE_NODE_INTENT_LOCKED = SIX_LOCK_intent, |
43 | BTREE_NODE_WRITE_LOCKED = SIX_LOCK_write, |
44 | }; |
45 | |
46 | static inline int btree_node_locked_type(struct btree_path *path, |
47 | unsigned level) |
48 | { |
49 | return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3); |
50 | } |
51 | |
52 | static inline bool btree_node_write_locked(struct btree_path *path, unsigned l) |
53 | { |
54 | return btree_node_locked_type(path, level: l) == BTREE_NODE_WRITE_LOCKED; |
55 | } |
56 | |
57 | static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l) |
58 | { |
59 | return btree_node_locked_type(path, level: l) == BTREE_NODE_INTENT_LOCKED; |
60 | } |
61 | |
62 | static inline bool btree_node_read_locked(struct btree_path *path, unsigned l) |
63 | { |
64 | return btree_node_locked_type(path, level: l) == BTREE_NODE_READ_LOCKED; |
65 | } |
66 | |
67 | static inline bool btree_node_locked(struct btree_path *path, unsigned level) |
68 | { |
69 | return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED; |
70 | } |
71 | |
72 | static inline void mark_btree_node_locked_noreset(struct btree_path *path, |
73 | unsigned level, |
74 | enum btree_node_locked_type type) |
75 | { |
76 | /* relying on this to avoid a branch */ |
77 | BUILD_BUG_ON(SIX_LOCK_read != 0); |
78 | BUILD_BUG_ON(SIX_LOCK_intent != 1); |
79 | |
80 | path->nodes_locked &= ~(3U << (level << 1)); |
81 | path->nodes_locked |= (type + 1) << (level << 1); |
82 | } |
83 | |
84 | static inline void mark_btree_node_unlocked(struct btree_path *path, |
85 | unsigned level) |
86 | { |
87 | EBUG_ON(btree_node_write_locked(path, level)); |
88 | mark_btree_node_locked_noreset(path, level, type: BTREE_NODE_UNLOCKED); |
89 | } |
90 | |
91 | static inline void mark_btree_node_locked(struct btree_trans *trans, |
92 | struct btree_path *path, |
93 | unsigned level, |
94 | enum btree_node_locked_type type) |
95 | { |
96 | mark_btree_node_locked_noreset(path, level, type: (enum btree_node_locked_type) type); |
97 | #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS |
98 | path->l[level].lock_taken_time = local_clock(); |
99 | #endif |
100 | } |
101 | |
102 | static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level) |
103 | { |
104 | return level < path->locks_want |
105 | ? SIX_LOCK_intent |
106 | : SIX_LOCK_read; |
107 | } |
108 | |
109 | static inline enum btree_node_locked_type |
110 | btree_lock_want(struct btree_path *path, int level) |
111 | { |
112 | if (level < path->level) |
113 | return BTREE_NODE_UNLOCKED; |
114 | if (level < path->locks_want) |
115 | return BTREE_NODE_INTENT_LOCKED; |
116 | if (level == path->level) |
117 | return BTREE_NODE_READ_LOCKED; |
118 | return BTREE_NODE_UNLOCKED; |
119 | } |
120 | |
121 | static void btree_trans_lock_hold_time_update(struct btree_trans *trans, |
122 | struct btree_path *path, unsigned level) |
123 | { |
124 | #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS |
125 | __bch2_time_stats_update(stats: &btree_trans_stats(trans)->lock_hold_times, |
126 | path->l[level].lock_taken_time, |
127 | local_clock()); |
128 | #endif |
129 | } |
130 | |
131 | /* unlock: */ |
132 | |
133 | static inline void btree_node_unlock(struct btree_trans *trans, |
134 | struct btree_path *path, unsigned level) |
135 | { |
136 | int lock_type = btree_node_locked_type(path, level); |
137 | |
138 | EBUG_ON(level >= BTREE_MAX_DEPTH); |
139 | |
140 | if (lock_type != BTREE_NODE_UNLOCKED) { |
141 | six_unlock_type(lock: &path->l[level].b->c.lock, type: lock_type); |
142 | btree_trans_lock_hold_time_update(trans, path, level); |
143 | } |
144 | mark_btree_node_unlocked(path, level); |
145 | } |
146 | |
147 | static inline int btree_path_lowest_level_locked(struct btree_path *path) |
148 | { |
149 | return __ffs(path->nodes_locked) >> 1; |
150 | } |
151 | |
152 | static inline int btree_path_highest_level_locked(struct btree_path *path) |
153 | { |
154 | return __fls(word: path->nodes_locked) >> 1; |
155 | } |
156 | |
157 | static inline void __bch2_btree_path_unlock(struct btree_trans *trans, |
158 | struct btree_path *path) |
159 | { |
160 | btree_path_set_dirty(path, u: BTREE_ITER_NEED_RELOCK); |
161 | |
162 | while (path->nodes_locked) |
163 | btree_node_unlock(trans, path, level: btree_path_lowest_level_locked(path)); |
164 | } |
165 | |
166 | /* |
167 | * Updates the saved lock sequence number, so that bch2_btree_node_relock() will |
168 | * succeed: |
169 | */ |
170 | static inline void |
171 | bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path, |
172 | struct btree *b) |
173 | { |
174 | struct btree_path *linked; |
175 | unsigned i; |
176 | |
177 | EBUG_ON(path->l[b->c.level].b != b); |
178 | EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock)); |
179 | EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write); |
180 | |
181 | mark_btree_node_locked_noreset(path, level: b->c.level, type: BTREE_NODE_INTENT_LOCKED); |
182 | |
183 | trans_for_each_path_with_node(trans, b, linked, i) |
184 | linked->l[b->c.level].lock_seq++; |
185 | |
186 | six_unlock_write(lock: &b->c.lock); |
187 | } |
188 | |
189 | void bch2_btree_node_unlock_write(struct btree_trans *, |
190 | struct btree_path *, struct btree *); |
191 | |
192 | int bch2_six_check_for_deadlock(struct six_lock *lock, void *p); |
193 | |
194 | /* lock: */ |
195 | |
196 | static inline int __btree_node_lock_nopath(struct btree_trans *trans, |
197 | struct btree_bkey_cached_common *b, |
198 | enum six_lock_type type, |
199 | bool lock_may_not_fail, |
200 | unsigned long ip) |
201 | { |
202 | int ret; |
203 | |
204 | trans->lock_may_not_fail = lock_may_not_fail; |
205 | trans->lock_must_abort = false; |
206 | trans->locking = b; |
207 | |
208 | ret = six_lock_ip_waiter(lock: &b->lock, type, wait: &trans->locking_wait, |
209 | should_sleep_fn: bch2_six_check_for_deadlock, p: trans, ip); |
210 | WRITE_ONCE(trans->locking, NULL); |
211 | WRITE_ONCE(trans->locking_wait.start_time, 0); |
212 | return ret; |
213 | } |
214 | |
215 | static inline int __must_check |
216 | btree_node_lock_nopath(struct btree_trans *trans, |
217 | struct btree_bkey_cached_common *b, |
218 | enum six_lock_type type, |
219 | unsigned long ip) |
220 | { |
221 | return __btree_node_lock_nopath(trans, b, type, lock_may_not_fail: false, ip); |
222 | } |
223 | |
224 | static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans, |
225 | struct btree_bkey_cached_common *b, |
226 | enum six_lock_type type) |
227 | { |
228 | int ret = __btree_node_lock_nopath(trans, b, type, lock_may_not_fail: true, _THIS_IP_); |
229 | |
230 | BUG_ON(ret); |
231 | } |
232 | |
233 | /* |
234 | * Lock a btree node if we already have it locked on one of our linked |
235 | * iterators: |
236 | */ |
237 | static inline bool btree_node_lock_increment(struct btree_trans *trans, |
238 | struct btree_bkey_cached_common *b, |
239 | unsigned level, |
240 | enum btree_node_locked_type want) |
241 | { |
242 | struct btree_path *path; |
243 | unsigned i; |
244 | |
245 | trans_for_each_path(trans, path, i) |
246 | if (&path->l[level].b->c == b && |
247 | btree_node_locked_type(path, level) >= want) { |
248 | six_lock_increment(&b->lock, (enum six_lock_type) want); |
249 | return true; |
250 | } |
251 | |
252 | return false; |
253 | } |
254 | |
255 | static inline int btree_node_lock(struct btree_trans *trans, |
256 | struct btree_path *path, |
257 | struct btree_bkey_cached_common *b, |
258 | unsigned level, |
259 | enum six_lock_type type, |
260 | unsigned long ip) |
261 | { |
262 | int ret = 0; |
263 | |
264 | EBUG_ON(level >= BTREE_MAX_DEPTH); |
265 | |
266 | if (likely(six_trylock_type(&b->lock, type)) || |
267 | btree_node_lock_increment(trans, b, level, want: (enum btree_node_locked_type) type) || |
268 | !(ret = btree_node_lock_nopath(trans, b, type, ip: btree_path_ip_allocated(path)))) { |
269 | #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS |
270 | path->l[b->level].lock_taken_time = local_clock(); |
271 | #endif |
272 | } |
273 | |
274 | return ret; |
275 | } |
276 | |
277 | int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *, |
278 | struct btree_bkey_cached_common *b, bool); |
279 | |
280 | static inline int __btree_node_lock_write(struct btree_trans *trans, |
281 | struct btree_path *path, |
282 | struct btree_bkey_cached_common *b, |
283 | bool lock_may_not_fail) |
284 | { |
285 | EBUG_ON(&path->l[b->level].b->c != b); |
286 | EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock)); |
287 | EBUG_ON(!btree_node_intent_locked(path, b->level)); |
288 | |
289 | /* |
290 | * six locks are unfair, and read locks block while a thread wants a |
291 | * write lock: thus, we need to tell the cycle detector we have a write |
292 | * lock _before_ taking the lock: |
293 | */ |
294 | mark_btree_node_locked_noreset(path, level: b->level, type: BTREE_NODE_WRITE_LOCKED); |
295 | |
296 | return likely(six_trylock_write(&b->lock)) |
297 | ? 0 |
298 | : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail); |
299 | } |
300 | |
301 | static inline int __must_check |
302 | bch2_btree_node_lock_write(struct btree_trans *trans, |
303 | struct btree_path *path, |
304 | struct btree_bkey_cached_common *b) |
305 | { |
306 | return __btree_node_lock_write(trans, path, b, lock_may_not_fail: false); |
307 | } |
308 | |
309 | void bch2_btree_node_lock_write_nofail(struct btree_trans *, |
310 | struct btree_path *, |
311 | struct btree_bkey_cached_common *); |
312 | |
313 | /* relock: */ |
314 | |
315 | bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *); |
316 | int __bch2_btree_path_relock(struct btree_trans *, |
317 | struct btree_path *, unsigned long); |
318 | |
319 | static inline int bch2_btree_path_relock(struct btree_trans *trans, |
320 | struct btree_path *path, unsigned long trace_ip) |
321 | { |
322 | return btree_node_locked(path, level: path->level) |
323 | ? 0 |
324 | : __bch2_btree_path_relock(trans, path, trace_ip); |
325 | } |
326 | |
327 | bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace); |
328 | |
329 | static inline bool bch2_btree_node_relock(struct btree_trans *trans, |
330 | struct btree_path *path, unsigned level) |
331 | { |
332 | EBUG_ON(btree_node_locked(path, level) && |
333 | !btree_node_write_locked(path, level) && |
334 | btree_node_locked_type(path, level) != __btree_lock_want(path, level)); |
335 | |
336 | return likely(btree_node_locked(path, level)) || |
337 | (!IS_ERR_OR_NULL(ptr: path->l[level].b) && |
338 | __bch2_btree_node_relock(trans, path, level, trace: true)); |
339 | } |
340 | |
341 | static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans, |
342 | struct btree_path *path, unsigned level) |
343 | { |
344 | EBUG_ON(btree_node_locked(path, level) && |
345 | !btree_node_write_locked(path, level) && |
346 | btree_node_locked_type(path, level) != __btree_lock_want(path, level)); |
347 | |
348 | return likely(btree_node_locked(path, level)) || |
349 | (!IS_ERR_OR_NULL(ptr: path->l[level].b) && |
350 | __bch2_btree_node_relock(trans, path, level, trace: false)); |
351 | } |
352 | |
353 | /* upgrade */ |
354 | |
355 | bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *, |
356 | struct btree_path *, unsigned, |
357 | struct get_locks_fail *); |
358 | |
359 | bool __bch2_btree_path_upgrade(struct btree_trans *, |
360 | struct btree_path *, unsigned, |
361 | struct get_locks_fail *); |
362 | |
363 | static inline int bch2_btree_path_upgrade(struct btree_trans *trans, |
364 | struct btree_path *path, |
365 | unsigned new_locks_want) |
366 | { |
367 | struct get_locks_fail f; |
368 | unsigned old_locks_want = path->locks_want; |
369 | |
370 | new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH); |
371 | |
372 | if (path->locks_want < new_locks_want |
373 | ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f) |
374 | : path->uptodate == BTREE_ITER_UPTODATE) |
375 | return 0; |
376 | |
377 | trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path, |
378 | old_locks_want, new_locks_want, &f); |
379 | return btree_trans_restart(trans, err: BCH_ERR_transaction_restart_upgrade); |
380 | } |
381 | |
382 | /* misc: */ |
383 | |
384 | static inline void btree_path_set_should_be_locked(struct btree_path *path) |
385 | { |
386 | EBUG_ON(!btree_node_locked(path, path->level)); |
387 | EBUG_ON(path->uptodate); |
388 | |
389 | path->should_be_locked = true; |
390 | } |
391 | |
392 | static inline void __btree_path_set_level_up(struct btree_trans *trans, |
393 | struct btree_path *path, |
394 | unsigned l) |
395 | { |
396 | btree_node_unlock(trans, path, level: l); |
397 | path->l[l].b = ERR_PTR(error: -BCH_ERR_no_btree_node_up); |
398 | } |
399 | |
400 | static inline void btree_path_set_level_up(struct btree_trans *trans, |
401 | struct btree_path *path) |
402 | { |
403 | __btree_path_set_level_up(trans, path, l: path->level++); |
404 | btree_path_set_dirty(path, u: BTREE_ITER_NEED_TRAVERSE); |
405 | } |
406 | |
407 | /* debug */ |
408 | |
409 | struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *, |
410 | struct btree_path *, |
411 | struct btree_bkey_cached_common *b, |
412 | unsigned); |
413 | |
414 | int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *); |
415 | |
416 | #ifdef CONFIG_BCACHEFS_DEBUG |
417 | void bch2_btree_path_verify_locks(struct btree_path *); |
418 | void bch2_trans_verify_locks(struct btree_trans *); |
419 | #else |
420 | static inline void bch2_btree_path_verify_locks(struct btree_path *path) {} |
421 | static inline void bch2_trans_verify_locks(struct btree_trans *trans) {} |
422 | #endif |
423 | |
424 | #endif /* _BCACHEFS_BTREE_LOCKING_H */ |
425 | |