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, gfp_t gfp);
17
18 void bch2_trans_unlock_noassert(struct btree_trans *);
19 void bch2_trans_unlock_write(struct btree_trans *);
20
is_btree_node(struct btree_path * path,unsigned l)21 static inline bool is_btree_node(struct btree_path *path, unsigned l)
22 {
23 return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
24 }
25
btree_trans_stats(struct btree_trans * trans)26 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
27 {
28 return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
29 ? &trans->c->btree_transaction_stats[trans->fn_idx]
30 : NULL;
31 }
32
33 /* matches six lock types */
34 enum btree_node_locked_type {
35 BTREE_NODE_UNLOCKED = -1,
36 BTREE_NODE_READ_LOCKED = SIX_LOCK_read,
37 BTREE_NODE_INTENT_LOCKED = SIX_LOCK_intent,
38 BTREE_NODE_WRITE_LOCKED = SIX_LOCK_write,
39 };
40
btree_node_locked_type(struct btree_path * path,unsigned level)41 static inline int btree_node_locked_type(struct btree_path *path,
42 unsigned level)
43 {
44 return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
45 }
46
btree_node_write_locked(struct btree_path * path,unsigned l)47 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
48 {
49 return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
50 }
51
btree_node_intent_locked(struct btree_path * path,unsigned l)52 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
53 {
54 return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
55 }
56
btree_node_read_locked(struct btree_path * path,unsigned l)57 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
58 {
59 return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
60 }
61
btree_node_locked(struct btree_path * path,unsigned level)62 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
63 {
64 return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
65 }
66
mark_btree_node_locked_noreset(struct btree_path * path,unsigned level,enum btree_node_locked_type type)67 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
68 unsigned level,
69 enum btree_node_locked_type type)
70 {
71 /* relying on this to avoid a branch */
72 BUILD_BUG_ON(SIX_LOCK_read != 0);
73 BUILD_BUG_ON(SIX_LOCK_intent != 1);
74
75 path->nodes_locked &= ~(3U << (level << 1));
76 path->nodes_locked |= (type + 1) << (level << 1);
77 }
78
mark_btree_node_locked(struct btree_trans * trans,struct btree_path * path,unsigned level,enum btree_node_locked_type type)79 static inline void mark_btree_node_locked(struct btree_trans *trans,
80 struct btree_path *path,
81 unsigned level,
82 enum btree_node_locked_type type)
83 {
84 mark_btree_node_locked_noreset(path, level, (enum btree_node_locked_type) type);
85 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
86 path->l[level].lock_taken_time = local_clock();
87 #endif
88 }
89
__btree_lock_want(struct btree_path * path,int level)90 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
91 {
92 return level < path->locks_want
93 ? SIX_LOCK_intent
94 : SIX_LOCK_read;
95 }
96
97 static inline enum btree_node_locked_type
btree_lock_want(struct btree_path * path,int level)98 btree_lock_want(struct btree_path *path, int level)
99 {
100 if (level < path->level)
101 return BTREE_NODE_UNLOCKED;
102 if (level < path->locks_want)
103 return BTREE_NODE_INTENT_LOCKED;
104 if (level == path->level)
105 return BTREE_NODE_READ_LOCKED;
106 return BTREE_NODE_UNLOCKED;
107 }
108
btree_trans_lock_hold_time_update(struct btree_trans * trans,struct btree_path * path,unsigned level)109 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
110 struct btree_path *path, unsigned level)
111 {
112 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
113 __bch2_time_stats_update(&btree_trans_stats(trans)->lock_hold_times,
114 path->l[level].lock_taken_time,
115 local_clock());
116 #endif
117 }
118
119 /* unlock: */
120
121 void bch2_btree_node_unlock_write(struct btree_trans *,
122 struct btree_path *, struct btree *);
123
btree_node_unlock(struct btree_trans * trans,struct btree_path * path,unsigned level)124 static inline void btree_node_unlock(struct btree_trans *trans,
125 struct btree_path *path, unsigned level)
126 {
127 int lock_type = btree_node_locked_type(path, level);
128
129 EBUG_ON(level >= BTREE_MAX_DEPTH);
130
131 if (lock_type != BTREE_NODE_UNLOCKED) {
132 if (unlikely(lock_type == BTREE_NODE_WRITE_LOCKED)) {
133 bch2_btree_node_unlock_write(trans, path, path->l[level].b);
134 lock_type = BTREE_NODE_INTENT_LOCKED;
135 }
136 six_unlock_type(&path->l[level].b->c.lock, lock_type);
137 btree_trans_lock_hold_time_update(trans, path, level);
138 mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
139 }
140 }
141
btree_path_lowest_level_locked(struct btree_path * path)142 static inline int btree_path_lowest_level_locked(struct btree_path *path)
143 {
144 return __ffs(path->nodes_locked) >> 1;
145 }
146
btree_path_highest_level_locked(struct btree_path * path)147 static inline int btree_path_highest_level_locked(struct btree_path *path)
148 {
149 return __fls(path->nodes_locked) >> 1;
150 }
151
__bch2_btree_path_unlock(struct btree_trans * trans,struct btree_path * path)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
__bch2_btree_node_unlock_write(struct btree_trans * trans,struct btree * b)166 __bch2_btree_node_unlock_write(struct btree_trans *trans, struct btree *b)
167 {
168 if (!b->c.lock.write_lock_recurse) {
169 struct btree_path *linked;
170 unsigned i;
171
172 trans_for_each_path_with_node(trans, b, linked, i)
173 linked->l[b->c.level].lock_seq++;
174 }
175
176 six_unlock_write(&b->c.lock);
177 }
178
179 static inline void
bch2_btree_node_unlock_write_inlined(struct btree_trans * trans,struct btree_path * path,struct btree * b)180 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
181 struct btree *b)
182 {
183 EBUG_ON(path->l[b->c.level].b != b);
184 EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
185 EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
186
187 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
188 __bch2_btree_node_unlock_write(trans, b);
189 }
190
191 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
192
193 /* lock: */
194
trans_set_locked(struct btree_trans * trans,bool try)195 static inline void trans_set_locked(struct btree_trans *trans, bool try)
196 {
197 if (!trans->locked) {
198 lock_acquire_exclusive(&trans->dep_map, 0, try, NULL, _THIS_IP_);
199 trans->locked = true;
200 trans->last_unlock_ip = 0;
201
202 trans->pf_memalloc_nofs = (current->flags & PF_MEMALLOC_NOFS) != 0;
203 current->flags |= PF_MEMALLOC_NOFS;
204 }
205 }
206
trans_set_unlocked(struct btree_trans * trans)207 static inline void trans_set_unlocked(struct btree_trans *trans)
208 {
209 if (trans->locked) {
210 lock_release(&trans->dep_map, _THIS_IP_);
211 trans->locked = false;
212 trans->last_unlock_ip = _RET_IP_;
213
214 if (!trans->pf_memalloc_nofs)
215 current->flags &= ~PF_MEMALLOC_NOFS;
216 }
217 }
218
__btree_node_lock_nopath(struct btree_trans * trans,struct btree_bkey_cached_common * b,enum six_lock_type type,bool lock_may_not_fail,unsigned long ip)219 static inline int __btree_node_lock_nopath(struct btree_trans *trans,
220 struct btree_bkey_cached_common *b,
221 enum six_lock_type type,
222 bool lock_may_not_fail,
223 unsigned long ip)
224 {
225 trans->lock_may_not_fail = lock_may_not_fail;
226 trans->lock_must_abort = false;
227 trans->locking = b;
228
229 int ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,
230 bch2_six_check_for_deadlock, trans, ip);
231 WRITE_ONCE(trans->locking, NULL);
232 WRITE_ONCE(trans->locking_wait.start_time, 0);
233
234 if (!ret)
235 trace_btree_path_lock(trans, _THIS_IP_, b);
236 return ret;
237 }
238
239 static inline int __must_check
btree_node_lock_nopath(struct btree_trans * trans,struct btree_bkey_cached_common * b,enum six_lock_type type,unsigned long ip)240 btree_node_lock_nopath(struct btree_trans *trans,
241 struct btree_bkey_cached_common *b,
242 enum six_lock_type type,
243 unsigned long ip)
244 {
245 return __btree_node_lock_nopath(trans, b, type, false, ip);
246 }
247
btree_node_lock_nopath_nofail(struct btree_trans * trans,struct btree_bkey_cached_common * b,enum six_lock_type type)248 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
249 struct btree_bkey_cached_common *b,
250 enum six_lock_type type)
251 {
252 int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
253
254 BUG_ON(ret);
255 }
256
257 /*
258 * Lock a btree node if we already have it locked on one of our linked
259 * iterators:
260 */
btree_node_lock_increment(struct btree_trans * trans,struct btree_bkey_cached_common * b,unsigned level,enum btree_node_locked_type want)261 static inline bool btree_node_lock_increment(struct btree_trans *trans,
262 struct btree_bkey_cached_common *b,
263 unsigned level,
264 enum btree_node_locked_type want)
265 {
266 struct btree_path *path;
267 unsigned i;
268
269 trans_for_each_path(trans, path, i)
270 if (&path->l[level].b->c == b &&
271 btree_node_locked_type(path, level) >= want) {
272 six_lock_increment(&b->lock, (enum six_lock_type) want);
273 return true;
274 }
275
276 return false;
277 }
278
btree_node_lock(struct btree_trans * trans,struct btree_path * path,struct btree_bkey_cached_common * b,unsigned level,enum six_lock_type type,unsigned long ip)279 static inline int btree_node_lock(struct btree_trans *trans,
280 struct btree_path *path,
281 struct btree_bkey_cached_common *b,
282 unsigned level,
283 enum six_lock_type type,
284 unsigned long ip)
285 {
286 int ret = 0;
287
288 EBUG_ON(level >= BTREE_MAX_DEPTH);
289 bch2_trans_verify_not_unlocked_or_in_restart(trans);
290
291 if (likely(six_trylock_type(&b->lock, type)) ||
292 btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||
293 !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {
294 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
295 path->l[b->level].lock_taken_time = local_clock();
296 #endif
297 }
298
299 return ret;
300 }
301
302 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
303 struct btree_bkey_cached_common *b, bool);
304
__btree_node_lock_write(struct btree_trans * trans,struct btree_path * path,struct btree_bkey_cached_common * b,bool lock_may_not_fail)305 static inline int __btree_node_lock_write(struct btree_trans *trans,
306 struct btree_path *path,
307 struct btree_bkey_cached_common *b,
308 bool lock_may_not_fail)
309 {
310 EBUG_ON(&path->l[b->level].b->c != b);
311 EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));
312 EBUG_ON(!btree_node_intent_locked(path, b->level));
313
314 /*
315 * six locks are unfair, and read locks block while a thread wants a
316 * write lock: thus, we need to tell the cycle detector we have a write
317 * lock _before_ taking the lock:
318 */
319 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED);
320
321 return likely(six_trylock_write(&b->lock))
322 ? 0
323 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
324 }
325
326 static inline int __must_check
bch2_btree_node_lock_write(struct btree_trans * trans,struct btree_path * path,struct btree_bkey_cached_common * b)327 bch2_btree_node_lock_write(struct btree_trans *trans,
328 struct btree_path *path,
329 struct btree_bkey_cached_common *b)
330 {
331 return __btree_node_lock_write(trans, path, b, false);
332 }
333
334 void bch2_btree_node_lock_write_nofail(struct btree_trans *,
335 struct btree_path *,
336 struct btree_bkey_cached_common *);
337
338 /* relock: */
339
340 bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *);
341 int __bch2_btree_path_relock(struct btree_trans *,
342 struct btree_path *, unsigned long);
343
bch2_btree_path_relock(struct btree_trans * trans,struct btree_path * path,unsigned long trace_ip)344 static inline int bch2_btree_path_relock(struct btree_trans *trans,
345 struct btree_path *path, unsigned long trace_ip)
346 {
347 return btree_node_locked(path, path->level)
348 ? 0
349 : __bch2_btree_path_relock(trans, path, trace_ip);
350 }
351
352 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
353
bch2_btree_node_relock(struct btree_trans * trans,struct btree_path * path,unsigned level)354 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
355 struct btree_path *path, unsigned level)
356 {
357 EBUG_ON(btree_node_locked(path, level) &&
358 !btree_node_write_locked(path, level) &&
359 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
360
361 return likely(btree_node_locked(path, level)) ||
362 (!IS_ERR_OR_NULL(path->l[level].b) &&
363 __bch2_btree_node_relock(trans, path, level, true));
364 }
365
bch2_btree_node_relock_notrace(struct btree_trans * trans,struct btree_path * path,unsigned level)366 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
367 struct btree_path *path, unsigned level)
368 {
369 EBUG_ON(btree_node_locked(path, level) &&
370 !btree_node_write_locked(path, level) &&
371 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
372
373 return likely(btree_node_locked(path, level)) ||
374 (!IS_ERR_OR_NULL(path->l[level].b) &&
375 __bch2_btree_node_relock(trans, path, level, false));
376 }
377
378 /* upgrade */
379
380 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
381 struct btree_path *, unsigned,
382 struct get_locks_fail *);
383
384 bool __bch2_btree_path_upgrade(struct btree_trans *,
385 struct btree_path *, unsigned,
386 struct get_locks_fail *);
387
bch2_btree_path_upgrade(struct btree_trans * trans,struct btree_path * path,unsigned new_locks_want)388 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
389 struct btree_path *path,
390 unsigned new_locks_want)
391 {
392 struct get_locks_fail f = {};
393 unsigned old_locks_want = path->locks_want;
394
395 new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
396
397 if (path->locks_want < new_locks_want
398 ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f)
399 : path->nodes_locked)
400 return 0;
401
402 trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
403 old_locks_want, new_locks_want, &f);
404 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
405 }
406
407 /* misc: */
408
btree_path_set_should_be_locked(struct btree_trans * trans,struct btree_path * path)409 static inline void btree_path_set_should_be_locked(struct btree_trans *trans, struct btree_path *path)
410 {
411 EBUG_ON(!btree_node_locked(path, path->level));
412 EBUG_ON(path->uptodate);
413
414 path->should_be_locked = true;
415 trace_btree_path_should_be_locked(trans, path);
416 }
417
__btree_path_set_level_up(struct btree_trans * trans,struct btree_path * path,unsigned l)418 static inline void __btree_path_set_level_up(struct btree_trans *trans,
419 struct btree_path *path,
420 unsigned l)
421 {
422 btree_node_unlock(trans, path, l);
423 path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
424 }
425
btree_path_set_level_up(struct btree_trans * trans,struct btree_path * path)426 static inline void btree_path_set_level_up(struct btree_trans *trans,
427 struct btree_path *path)
428 {
429 __btree_path_set_level_up(trans, path, path->level++);
430 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
431 }
432
433 /* debug */
434
435 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
436 struct btree_path *,
437 struct btree_bkey_cached_common *b,
438 unsigned);
439
440 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
441
442 #ifdef CONFIG_BCACHEFS_DEBUG
443 void bch2_btree_path_verify_locks(struct btree_path *);
444 void bch2_trans_verify_locks(struct btree_trans *);
445 #else
bch2_btree_path_verify_locks(struct btree_path * path)446 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
bch2_trans_verify_locks(struct btree_trans * trans)447 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
448 #endif
449
450 #endif /* _BCACHEFS_BTREE_LOCKING_H */
451