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 void bch2_trans_unlock_noassert(struct btree_trans *);
19
20 static inline bool is_btree_node(struct btree_path *path, unsigned l)
21 {
22 return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
23 }
24
25 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
26 {
27 return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
28 ? &trans->c->btree_transaction_stats[trans->fn_idx]
29 : NULL;
30 }
31
32 /* matches six lock types */
33 enum btree_node_locked_type {
34 BTREE_NODE_UNLOCKED = -1,
35 BTREE_NODE_READ_LOCKED = SIX_LOCK_read,
36 BTREE_NODE_INTENT_LOCKED = SIX_LOCK_intent,
37 BTREE_NODE_WRITE_LOCKED = SIX_LOCK_write,
38 };
39
40 static inline int btree_node_locked_type(struct btree_path *path,
41 unsigned level)
42 {
43 return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
44 }
45
46 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
47 {
48 return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
49 }
50
51 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
52 {
53 return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
54 }
55
56 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
57 {
58 return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
59 }
60
61 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
62 {
63 return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
64 }
65
66 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
67 unsigned level,
68 enum btree_node_locked_type type)
69 {
70 /* relying on this to avoid a branch */
71 BUILD_BUG_ON(SIX_LOCK_read != 0);
72 BUILD_BUG_ON(SIX_LOCK_intent != 1);
73
74 path->nodes_locked &= ~(3U << (level << 1));
75 path->nodes_locked |= (type + 1) << (level << 1);
76 }
77
78 static inline void mark_btree_node_unlocked(struct btree_path *path,
79 unsigned level)
80 {
81 EBUG_ON(btree_node_write_locked(path, level));
82 mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
83 }
84
85 static inline void mark_btree_node_locked(struct btree_trans *trans,
86 struct btree_path *path,
87 unsigned level,
88 enum btree_node_locked_type type)
89 {
90 mark_btree_node_locked_noreset(path, level, (enum btree_node_locked_type) type);
91 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
92 path->l[level].lock_taken_time = local_clock();
93 #endif
94 }
95
96 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
97 {
98 return level < path->locks_want
99 ? SIX_LOCK_intent
100 : SIX_LOCK_read;
101 }
102
103 static inline enum btree_node_locked_type
104 btree_lock_want(struct btree_path *path, int level)
105 {
106 if (level < path->level)
107 return BTREE_NODE_UNLOCKED;
108 if (level < path->locks_want)
109 return BTREE_NODE_INTENT_LOCKED;
110 if (level == path->level)
111 return BTREE_NODE_READ_LOCKED;
112 return BTREE_NODE_UNLOCKED;
113 }
114
115 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
116 struct btree_path *path, unsigned level)
117 {
118 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
119 __bch2_time_stats_update(&btree_trans_stats(trans)->lock_hold_times,
120 path->l[level].lock_taken_time,
121 local_clock());
122 #endif
123 }
124
125 /* unlock: */
126
127 static inline void btree_node_unlock(struct btree_trans *trans,
128 struct btree_path *path, unsigned level)
129 {
130 int lock_type = btree_node_locked_type(path, level);
131
132 EBUG_ON(level >= BTREE_MAX_DEPTH);
133 EBUG_ON(lock_type == BTREE_NODE_WRITE_LOCKED);
134
135 if (lock_type != BTREE_NODE_UNLOCKED) {
136 six_unlock_type(&path->l[level].b->c.lock, lock_type);
137 btree_trans_lock_hold_time_update(trans, path, level);
138 }
139 mark_btree_node_unlocked(path, level);
140 }
141
142 static inline int btree_path_lowest_level_locked(struct btree_path *path)
143 {
144 return __ffs(path->nodes_locked) >> 1;
145 }
146
147 static inline int btree_path_highest_level_locked(struct btree_path *path)
148 {
149 return __fls(path->nodes_locked) >> 1;
150 }
151
152 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
153 struct btree_path *path)
154 {
155 btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
156
157 while (path->nodes_locked)
158 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
159 }
160
161 /*
162 * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
163 * succeed:
164 */
165 static inline void
166 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
167 struct btree *b)
168 {
169 struct btree_path *linked;
170 unsigned i;
171
172 EBUG_ON(path->l[b->c.level].b != b);
173 EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
174 EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
175
176 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
177
178 trans_for_each_path_with_node(trans, b, linked, i)
179 linked->l[b->c.level].lock_seq++;
180
181 six_unlock_write(&b->c.lock);
182 }
183
184 void bch2_btree_node_unlock_write(struct btree_trans *,
185 struct btree_path *, struct btree *);
186
187 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
188
189 /* lock: */
190
191 static inline void trans_set_locked(struct btree_trans *trans)
192 {
193 if (!trans->locked) {
194 lock_acquire_exclusive(&trans->dep_map, 0, 0, NULL, _THIS_IP_);
195 trans->locked = true;
196 trans->last_unlock_ip = 0;
197
198 trans->pf_memalloc_nofs = (current->flags & PF_MEMALLOC_NOFS) != 0;
199 current->flags |= PF_MEMALLOC_NOFS;
200 }
201 }
202
203 static inline void trans_set_unlocked(struct btree_trans *trans)
204 {
205 if (trans->locked) {
206 lock_release(&trans->dep_map, _THIS_IP_);
207 trans->locked = false;
208 trans->last_unlock_ip = _RET_IP_;
209
210 if (!trans->pf_memalloc_nofs)
211 current->flags &= ~PF_MEMALLOC_NOFS;
212 }
213 }
214
215 static inline int __btree_node_lock_nopath(struct btree_trans *trans,
216 struct btree_bkey_cached_common *b,
217 enum six_lock_type type,
218 bool lock_may_not_fail,
219 unsigned long ip)
220 {
221 int ret;
222
223 trans->lock_may_not_fail = lock_may_not_fail;
224 trans->lock_must_abort = false;
225 trans->locking = b;
226
227 ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,
228 bch2_six_check_for_deadlock, trans, ip);
229 WRITE_ONCE(trans->locking, NULL);
230 WRITE_ONCE(trans->locking_wait.start_time, 0);
231 return ret;
232 }
233
234 static inline int __must_check
235 btree_node_lock_nopath(struct btree_trans *trans,
236 struct btree_bkey_cached_common *b,
237 enum six_lock_type type,
238 unsigned long ip)
239 {
240 return __btree_node_lock_nopath(trans, b, type, false, ip);
241 }
242
243 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
244 struct btree_bkey_cached_common *b,
245 enum six_lock_type type)
246 {
247 int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
248
249 BUG_ON(ret);
250 }
251
252 /*
253 * Lock a btree node if we already have it locked on one of our linked
254 * iterators:
255 */
256 static inline bool btree_node_lock_increment(struct btree_trans *trans,
257 struct btree_bkey_cached_common *b,
258 unsigned level,
259 enum btree_node_locked_type want)
260 {
261 struct btree_path *path;
262 unsigned i;
263
264 trans_for_each_path(trans, path, i)
265 if (&path->l[level].b->c == b &&
266 btree_node_locked_type(path, level) >= want) {
267 six_lock_increment(&b->lock, (enum six_lock_type) want);
268 return true;
269 }
270
271 return false;
272 }
273
274 static inline int btree_node_lock(struct btree_trans *trans,
275 struct btree_path *path,
276 struct btree_bkey_cached_common *b,
277 unsigned level,
278 enum six_lock_type type,
279 unsigned long ip)
280 {
281 int ret = 0;
282
283 EBUG_ON(level >= BTREE_MAX_DEPTH);
284
285 if (likely(six_trylock_type(&b->lock, type)) ||
286 btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||
287 !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {
288 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
289 path->l[b->level].lock_taken_time = local_clock();
290 #endif
291 }
292
293 return ret;
294 }
295
296 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
297 struct btree_bkey_cached_common *b, bool);
298
299 static inline int __btree_node_lock_write(struct btree_trans *trans,
300 struct btree_path *path,
301 struct btree_bkey_cached_common *b,
302 bool lock_may_not_fail)
303 {
304 EBUG_ON(&path->l[b->level].b->c != b);
305 EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));
306 EBUG_ON(!btree_node_intent_locked(path, b->level));
307
308 /*
309 * six locks are unfair, and read locks block while a thread wants a
310 * write lock: thus, we need to tell the cycle detector we have a write
311 * lock _before_ taking the lock:
312 */
313 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED);
314
315 return likely(six_trylock_write(&b->lock))
316 ? 0
317 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
318 }
319
320 static inline int __must_check
321 bch2_btree_node_lock_write(struct btree_trans *trans,
322 struct btree_path *path,
323 struct btree_bkey_cached_common *b)
324 {
325 return __btree_node_lock_write(trans, path, b, false);
326 }
327
328 void bch2_btree_node_lock_write_nofail(struct btree_trans *,
329 struct btree_path *,
330 struct btree_bkey_cached_common *);
331
332 /* relock: */
333
334 bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *);
335 int __bch2_btree_path_relock(struct btree_trans *,
336 struct btree_path *, unsigned long);
337
338 static inline int bch2_btree_path_relock(struct btree_trans *trans,
339 struct btree_path *path, unsigned long trace_ip)
340 {
341 return btree_node_locked(path, path->level)
342 ? 0
343 : __bch2_btree_path_relock(trans, path, trace_ip);
344 }
345
346 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
347
348 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
349 struct btree_path *path, unsigned level)
350 {
351 EBUG_ON(btree_node_locked(path, level) &&
352 !btree_node_write_locked(path, level) &&
353 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
354
355 return likely(btree_node_locked(path, level)) ||
356 (!IS_ERR_OR_NULL(path->l[level].b) &&
357 __bch2_btree_node_relock(trans, path, level, true));
358 }
359
360 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
361 struct btree_path *path, unsigned level)
362 {
363 EBUG_ON(btree_node_locked(path, level) &&
364 !btree_node_write_locked(path, level) &&
365 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
366
367 return likely(btree_node_locked(path, level)) ||
368 (!IS_ERR_OR_NULL(path->l[level].b) &&
369 __bch2_btree_node_relock(trans, path, level, false));
370 }
371
372 /* upgrade */
373
374 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
375 struct btree_path *, unsigned,
376 struct get_locks_fail *);
377
378 bool __bch2_btree_path_upgrade(struct btree_trans *,
379 struct btree_path *, unsigned,
380 struct get_locks_fail *);
381
382 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
383 struct btree_path *path,
384 unsigned new_locks_want)
385 {
386 struct get_locks_fail f = {};
387 unsigned old_locks_want = path->locks_want;
388
389 new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
390
391 if (path->locks_want < new_locks_want
392 ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f)
393 : path->nodes_locked)
394 return 0;
395
396 trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
397 old_locks_want, new_locks_want, &f);
398 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
399 }
400
401 /* misc: */
402
403 static inline void btree_path_set_should_be_locked(struct btree_path *path)
404 {
405 EBUG_ON(!btree_node_locked(path, path->level));
406 EBUG_ON(path->uptodate);
407
408 path->should_be_locked = true;
409 }
410
411 static inline void __btree_path_set_level_up(struct btree_trans *trans,
412 struct btree_path *path,
413 unsigned l)
414 {
415 btree_node_unlock(trans, path, l);
416 path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
417 }
418
419 static inline void btree_path_set_level_up(struct btree_trans *trans,
420 struct btree_path *path)
421 {
422 __btree_path_set_level_up(trans, path, path->level++);
423 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
424 }
425
426 /* debug */
427
428 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
429 struct btree_path *,
430 struct btree_bkey_cached_common *b,
431 unsigned);
432
433 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
434
435 #ifdef CONFIG_BCACHEFS_DEBUG
436 void bch2_btree_path_verify_locks(struct btree_path *);
437 void bch2_trans_verify_locks(struct btree_trans *);
438 #else
439 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
440 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
441 #endif
442
443 #endif /* _BCACHEFS_BTREE_LOCKING_H */
444
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.