1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _BCACHEFS_BTREE_LOCKING_H
3 #define _BCACHEFS_BTREE_LOCKING_H
6 * Only for internal btree use:
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
13 #include <linux/six.h>
15 #include "btree_iter.h"
17 void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);
20 void bch2_assert_btree_nodes_not_locked(void);
22 static inline void bch2_assert_btree_nodes_not_locked(void) {}
25 static inline bool is_btree_node(struct btree_path *path, unsigned l)
27 return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
30 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
32 return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
33 ? &trans->c->btree_transaction_stats[trans->fn_idx]
37 /* matches six lock types */
38 enum btree_node_locked_type {
39 BTREE_NODE_UNLOCKED = -1,
40 BTREE_NODE_READ_LOCKED = SIX_LOCK_read,
41 BTREE_NODE_INTENT_LOCKED = SIX_LOCK_intent,
42 BTREE_NODE_WRITE_LOCKED = SIX_LOCK_write,
45 static inline int btree_node_locked_type(struct btree_path *path,
48 return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
51 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
53 return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
56 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
58 return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
61 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
63 return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
66 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
68 return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
71 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
73 enum btree_node_locked_type type)
75 /* relying on this to avoid a branch */
76 BUILD_BUG_ON(SIX_LOCK_read != 0);
77 BUILD_BUG_ON(SIX_LOCK_intent != 1);
79 path->nodes_locked &= ~(3U << (level << 1));
80 path->nodes_locked |= (type + 1) << (level << 1);
83 static inline void mark_btree_node_unlocked(struct btree_path *path,
86 EBUG_ON(btree_node_write_locked(path, level));
87 mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
90 static inline void mark_btree_node_locked(struct btree_trans *trans,
91 struct btree_path *path,
93 enum six_lock_type type)
95 mark_btree_node_locked_noreset(path, level, type);
96 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
97 path->l[level].lock_taken_time = local_clock();
101 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
103 return level < path->locks_want
108 static inline enum btree_node_locked_type
109 btree_lock_want(struct btree_path *path, int level)
111 if (level < path->level)
112 return BTREE_NODE_UNLOCKED;
113 if (level < path->locks_want)
114 return BTREE_NODE_INTENT_LOCKED;
115 if (level == path->level)
116 return BTREE_NODE_READ_LOCKED;
117 return BTREE_NODE_UNLOCKED;
120 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
121 struct btree_path *path, unsigned level)
123 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
124 struct btree_transaction_stats *s = btree_trans_stats(trans);
127 __bch2_time_stats_update(&s->lock_hold_times,
128 path->l[level].lock_taken_time,
135 static inline void btree_node_unlock(struct btree_trans *trans,
136 struct btree_path *path, unsigned level)
138 int lock_type = btree_node_locked_type(path, level);
140 EBUG_ON(level >= BTREE_MAX_DEPTH);
142 if (lock_type != BTREE_NODE_UNLOCKED) {
143 six_unlock_type(&path->l[level].b->c.lock, lock_type);
144 btree_trans_lock_hold_time_update(trans, path, level);
146 mark_btree_node_unlocked(path, level);
149 static inline int btree_path_lowest_level_locked(struct btree_path *path)
151 return __ffs(path->nodes_locked) >> 1;
154 static inline int btree_path_highest_level_locked(struct btree_path *path)
156 return __fls(path->nodes_locked) >> 1;
159 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
160 struct btree_path *path)
162 btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
164 while (path->nodes_locked)
165 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
169 * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
173 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
176 struct btree_path *linked;
178 EBUG_ON(path->l[b->c.level].b != b);
179 EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
180 EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
182 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
184 trans_for_each_path_with_node(trans, b, linked)
185 linked->l[b->c.level].lock_seq++;
187 six_unlock_write(&b->c.lock);
190 void bch2_btree_node_unlock_write(struct btree_trans *,
191 struct btree_path *, struct btree *);
193 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
197 static inline int __btree_node_lock_nopath(struct btree_trans *trans,
198 struct btree_bkey_cached_common *b,
199 enum six_lock_type type,
200 bool lock_may_not_fail,
205 trans->lock_may_not_fail = lock_may_not_fail;
206 trans->lock_must_abort = false;
209 ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,
210 bch2_six_check_for_deadlock, trans, ip);
211 WRITE_ONCE(trans->locking, NULL);
212 WRITE_ONCE(trans->locking_wait.start_time, 0);
216 static inline int __must_check
217 btree_node_lock_nopath(struct btree_trans *trans,
218 struct btree_bkey_cached_common *b,
219 enum six_lock_type type,
222 return __btree_node_lock_nopath(trans, b, type, false, ip);
225 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
226 struct btree_bkey_cached_common *b,
227 enum six_lock_type type)
229 int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
235 * Lock a btree node if we already have it locked on one of our linked
238 static inline bool btree_node_lock_increment(struct btree_trans *trans,
239 struct btree_bkey_cached_common *b,
241 enum btree_node_locked_type want)
243 struct btree_path *path;
245 trans_for_each_path(trans, path)
246 if (&path->l[level].b->c == b &&
247 btree_node_locked_type(path, level) >= want) {
248 six_lock_increment(&b->lock, want);
255 static inline int btree_node_lock(struct btree_trans *trans,
256 struct btree_path *path,
257 struct btree_bkey_cached_common *b,
259 enum six_lock_type type,
264 EBUG_ON(level >= BTREE_MAX_DEPTH);
265 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
267 if (likely(six_trylock_type(&b->lock, type)) ||
268 btree_node_lock_increment(trans, b, level, type) ||
269 !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {
270 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
271 path->l[b->level].lock_taken_time = local_clock();
278 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
279 struct btree_bkey_cached_common *b, bool);
281 static inline int __btree_node_lock_write(struct btree_trans *trans,
282 struct btree_path *path,
283 struct btree_bkey_cached_common *b,
284 bool lock_may_not_fail)
286 EBUG_ON(&path->l[b->level].b->c != b);
287 EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));
288 EBUG_ON(!btree_node_intent_locked(path, b->level));
291 * six locks are unfair, and read locks block while a thread wants a
292 * write lock: thus, we need to tell the cycle detector we have a write
293 * lock _before_ taking the lock:
295 mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_write);
297 return likely(six_trylock_write(&b->lock))
299 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
302 static inline int __must_check
303 bch2_btree_node_lock_write(struct btree_trans *trans,
304 struct btree_path *path,
305 struct btree_bkey_cached_common *b)
307 return __btree_node_lock_write(trans, path, b, false);
310 void bch2_btree_node_lock_write_nofail(struct btree_trans *,
312 struct btree_bkey_cached_common *);
316 bool bch2_btree_path_relock_norestart(struct btree_trans *,
317 struct btree_path *, unsigned long);
318 int __bch2_btree_path_relock(struct btree_trans *,
319 struct btree_path *, unsigned long);
321 static inline int bch2_btree_path_relock(struct btree_trans *trans,
322 struct btree_path *path, unsigned long trace_ip)
324 return btree_node_locked(path, path->level)
326 : __bch2_btree_path_relock(trans, path, trace_ip);
329 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
331 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
332 struct btree_path *path, unsigned level)
334 EBUG_ON(btree_node_locked(path, level) &&
335 !btree_node_write_locked(path, level) &&
336 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
338 return likely(btree_node_locked(path, level)) ||
339 (!IS_ERR_OR_NULL(path->l[level].b) &&
340 __bch2_btree_node_relock(trans, path, level, true));
343 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
344 struct btree_path *path, unsigned level)
346 EBUG_ON(btree_node_locked(path, level) &&
347 !btree_node_write_locked(path, level) &&
348 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
350 return likely(btree_node_locked(path, level)) ||
351 (!IS_ERR_OR_NULL(path->l[level].b) &&
352 __bch2_btree_node_relock(trans, path, level, false));
357 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
358 struct btree_path *, unsigned);
359 bool __bch2_btree_path_upgrade(struct btree_trans *,
360 struct btree_path *, unsigned);
362 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
363 struct btree_path *path,
364 unsigned new_locks_want)
366 unsigned old_locks_want = path->locks_want;
368 new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
370 if (path->locks_want < new_locks_want
371 ? __bch2_btree_path_upgrade(trans, path, new_locks_want)
372 : path->uptodate == BTREE_ITER_UPTODATE)
375 trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
376 old_locks_want, new_locks_want);
377 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
382 static inline void btree_path_set_should_be_locked(struct btree_path *path)
384 EBUG_ON(!btree_node_locked(path, path->level));
385 EBUG_ON(path->uptodate);
387 path->should_be_locked = true;
390 static inline void __btree_path_set_level_up(struct btree_trans *trans,
391 struct btree_path *path,
394 btree_node_unlock(trans, path, l);
395 path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
398 static inline void btree_path_set_level_up(struct btree_trans *trans,
399 struct btree_path *path)
401 __btree_path_set_level_up(trans, path, path->level++);
402 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
407 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
409 struct btree_bkey_cached_common *b,
412 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
414 #ifdef CONFIG_BCACHEFS_DEBUG
415 void bch2_btree_path_verify_locks(struct btree_path *);
416 void bch2_trans_verify_locks(struct btree_trans *);
418 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
419 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
422 #endif /* _BCACHEFS_BTREE_LOCKING_H */