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(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, 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, 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, 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, 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, (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(&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(&path->l[level].b->c.lock, 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(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, BTREE_ITER_NEED_RELOCK); 161 162 while (path->nodes_locked) 163 btree_node_unlock(trans, path, 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, b->c.level, 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(&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 void trans_set_locked(struct btree_trans *trans) 197 { 198 if (!trans->locked) { 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 } 206 207 static inline void trans_set_unlocked(struct btree_trans *trans) 208 { 209 if (trans->locked) { 210 trans->locked = false; 211 trans->last_unlock_ip = _RET_IP_; 212 213 if (!trans->pf_memalloc_nofs) 214 current->flags &= ~PF_MEMALLOC_NOFS; 215 } 216 } 217 218 static inline int __btree_node_lock_nopath(struct btree_trans *trans, 219 struct btree_bkey_cached_common *b, 220 enum six_lock_type type, 221 bool lock_may_not_fail, 222 unsigned long ip) 223 { 224 int ret; 225 226 trans->lock_may_not_fail = lock_may_not_fail; 227 trans->lock_must_abort = false; 228 trans->locking = b; 229 230 ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait, 231 bch2_six_check_for_deadlock, trans, ip); 232 WRITE_ONCE(trans->locking, NULL); 233 WRITE_ONCE(trans->locking_wait.start_time, 0); 234 return ret; 235 } 236 237 static inline int __must_check 238 btree_node_lock_nopath(struct btree_trans *trans, 239 struct btree_bkey_cached_common *b, 240 enum six_lock_type type, 241 unsigned long ip) 242 { 243 return __btree_node_lock_nopath(trans, b, type, false, ip); 244 } 245 246 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans, 247 struct btree_bkey_cached_common *b, 248 enum six_lock_type type) 249 { 250 int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_); 251 252 BUG_ON(ret); 253 } 254 255 /* 256 * Lock a btree node if we already have it locked on one of our linked 257 * iterators: 258 */ 259 static inline bool btree_node_lock_increment(struct btree_trans *trans, 260 struct btree_bkey_cached_common *b, 261 unsigned level, 262 enum btree_node_locked_type want) 263 { 264 struct btree_path *path; 265 unsigned i; 266 267 trans_for_each_path(trans, path, i) 268 if (&path->l[level].b->c == b && 269 btree_node_locked_type(path, level) >= want) { 270 six_lock_increment(&b->lock, (enum six_lock_type) want); 271 return true; 272 } 273 274 return false; 275 } 276 277 static inline int btree_node_lock(struct btree_trans *trans, 278 struct btree_path *path, 279 struct btree_bkey_cached_common *b, 280 unsigned level, 281 enum six_lock_type type, 282 unsigned long ip) 283 { 284 int ret = 0; 285 286 EBUG_ON(level >= BTREE_MAX_DEPTH); 287 288 if (likely(six_trylock_type(&b->lock, type)) || 289 btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) || 290 !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) { 291 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS 292 path->l[b->level].lock_taken_time = local_clock(); 293 #endif 294 } 295 296 return ret; 297 } 298 299 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *, 300 struct btree_bkey_cached_common *b, bool); 301 302 static inline int __btree_node_lock_write(struct btree_trans *trans, 303 struct btree_path *path, 304 struct btree_bkey_cached_common *b, 305 bool lock_may_not_fail) 306 { 307 EBUG_ON(&path->l[b->level].b->c != b); 308 EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock)); 309 EBUG_ON(!btree_node_intent_locked(path, b->level)); 310 311 /* 312 * six locks are unfair, and read locks block while a thread wants a 313 * write lock: thus, we need to tell the cycle detector we have a write 314 * lock _before_ taking the lock: 315 */ 316 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED); 317 318 return likely(six_trylock_write(&b->lock)) 319 ? 0 320 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail); 321 } 322 323 static inline int __must_check 324 bch2_btree_node_lock_write(struct btree_trans *trans, 325 struct btree_path *path, 326 struct btree_bkey_cached_common *b) 327 { 328 return __btree_node_lock_write(trans, path, b, false); 329 } 330 331 void bch2_btree_node_lock_write_nofail(struct btree_trans *, 332 struct btree_path *, 333 struct btree_bkey_cached_common *); 334 335 /* relock: */ 336 337 bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *); 338 int __bch2_btree_path_relock(struct btree_trans *, 339 struct btree_path *, unsigned long); 340 341 static inline int bch2_btree_path_relock(struct btree_trans *trans, 342 struct btree_path *path, unsigned long trace_ip) 343 { 344 return btree_node_locked(path, path->level) 345 ? 0 346 : __bch2_btree_path_relock(trans, path, trace_ip); 347 } 348 349 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace); 350 351 static inline bool bch2_btree_node_relock(struct btree_trans *trans, 352 struct btree_path *path, unsigned level) 353 { 354 EBUG_ON(btree_node_locked(path, level) && 355 !btree_node_write_locked(path, level) && 356 btree_node_locked_type(path, level) != __btree_lock_want(path, level)); 357 358 return likely(btree_node_locked(path, level)) || 359 (!IS_ERR_OR_NULL(path->l[level].b) && 360 __bch2_btree_node_relock(trans, path, level, true)); 361 } 362 363 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans, 364 struct btree_path *path, unsigned level) 365 { 366 EBUG_ON(btree_node_locked(path, level) && 367 !btree_node_write_locked(path, level) && 368 btree_node_locked_type(path, level) != __btree_lock_want(path, level)); 369 370 return likely(btree_node_locked(path, level)) || 371 (!IS_ERR_OR_NULL(path->l[level].b) && 372 __bch2_btree_node_relock(trans, path, level, false)); 373 } 374 375 /* upgrade */ 376 377 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *, 378 struct btree_path *, unsigned, 379 struct get_locks_fail *); 380 381 bool __bch2_btree_path_upgrade(struct btree_trans *, 382 struct btree_path *, unsigned, 383 struct get_locks_fail *); 384 385 static inline int bch2_btree_path_upgrade(struct btree_trans *trans, 386 struct btree_path *path, 387 unsigned new_locks_want) 388 { 389 struct get_locks_fail f = {}; 390 unsigned old_locks_want = path->locks_want; 391 392 new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH); 393 394 if (path->locks_want < new_locks_want 395 ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f) 396 : path->nodes_locked) 397 return 0; 398 399 trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path, 400 old_locks_want, new_locks_want, &f); 401 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade); 402 } 403 404 /* misc: */ 405 406 static inline void btree_path_set_should_be_locked(struct btree_path *path) 407 { 408 EBUG_ON(!btree_node_locked(path, path->level)); 409 EBUG_ON(path->uptodate); 410 411 path->should_be_locked = true; 412 } 413 414 static inline void __btree_path_set_level_up(struct btree_trans *trans, 415 struct btree_path *path, 416 unsigned l) 417 { 418 btree_node_unlock(trans, path, l); 419 path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up); 420 } 421 422 static inline void btree_path_set_level_up(struct btree_trans *trans, 423 struct btree_path *path) 424 { 425 __btree_path_set_level_up(trans, path, path->level++); 426 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 427 } 428 429 /* debug */ 430 431 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *, 432 struct btree_path *, 433 struct btree_bkey_cached_common *b, 434 unsigned); 435 436 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *); 437 438 #ifdef CONFIG_BCACHEFS_DEBUG 439 void bch2_btree_path_verify_locks(struct btree_path *); 440 void bch2_trans_verify_locks(struct btree_trans *); 441 #else 442 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {} 443 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {} 444 #endif 445 446 #endif /* _BCACHEFS_BTREE_LOCKING_H */ 447