447 lines · c
1/* SPDX-License-Identifier: GPL-2.0 */2#ifndef _BCACHEFS_BTREE_LOCKING_H3#define _BCACHEFS_BTREE_LOCKING_H4 5/*6 * Only for internal btree use:7 *8 * The btree iterator tracks what locks it wants to take, and what locks it9 * currently has - here we have wrappers for locking/unlocking btree nodes and10 * updating the iterator state11 */12 13#include "btree_iter.h"14#include "six.h"15 16void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);17 18void bch2_trans_unlock_noassert(struct btree_trans *);19 20static 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 25static 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 */33enum 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 40static 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 46static 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 51static 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 56static 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 61static 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 66static 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 78static 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 85static 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_STATS92 path->l[level].lock_taken_time = local_clock();93#endif94}95 96static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)97{98 return level < path->locks_want99 ? SIX_LOCK_intent100 : SIX_LOCK_read;101}102 103static inline enum btree_node_locked_type104btree_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 115static 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_STATS119 __bch2_time_stats_update(&btree_trans_stats(trans)->lock_hold_times,120 path->l[level].lock_taken_time,121 local_clock());122#endif123}124 125/* unlock: */126 127static 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 142static inline int btree_path_lowest_level_locked(struct btree_path *path)143{144 return __ffs(path->nodes_locked) >> 1;145}146 147static inline int btree_path_highest_level_locked(struct btree_path *path)148{149 return __fls(path->nodes_locked) >> 1;150}151 152static 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() will163 * succeed:164 */165static inline void166bch2_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 184void bch2_btree_node_unlock_write(struct btree_trans *,185 struct btree_path *, struct btree *);186 187int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);188 189/* lock: */190 191static 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 203static 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 215static 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 trans->lock_may_not_fail = lock_may_not_fail;222 trans->lock_must_abort = false;223 trans->locking = b;224 225 int ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,226 bch2_six_check_for_deadlock, trans, ip);227 WRITE_ONCE(trans->locking, NULL);228 WRITE_ONCE(trans->locking_wait.start_time, 0);229 230 if (!ret)231 trace_btree_path_lock(trans, _THIS_IP_, b);232 return ret;233}234 235static inline int __must_check236btree_node_lock_nopath(struct btree_trans *trans,237 struct btree_bkey_cached_common *b,238 enum six_lock_type type,239 unsigned long ip)240{241 return __btree_node_lock_nopath(trans, b, type, false, ip);242}243 244static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,245 struct btree_bkey_cached_common *b,246 enum six_lock_type type)247{248 int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);249 250 BUG_ON(ret);251}252 253/*254 * Lock a btree node if we already have it locked on one of our linked255 * iterators:256 */257static inline bool btree_node_lock_increment(struct btree_trans *trans,258 struct btree_bkey_cached_common *b,259 unsigned level,260 enum btree_node_locked_type want)261{262 struct btree_path *path;263 unsigned i;264 265 trans_for_each_path(trans, path, i)266 if (&path->l[level].b->c == b &&267 btree_node_locked_type(path, level) >= want) {268 six_lock_increment(&b->lock, (enum six_lock_type) want);269 return true;270 }271 272 return false;273}274 275static inline int btree_node_lock(struct btree_trans *trans,276 struct btree_path *path,277 struct btree_bkey_cached_common *b,278 unsigned level,279 enum six_lock_type type,280 unsigned long ip)281{282 int ret = 0;283 284 EBUG_ON(level >= BTREE_MAX_DEPTH);285 bch2_trans_verify_not_unlocked(trans);286 287 if (likely(six_trylock_type(&b->lock, type)) ||288 btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||289 !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {290#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS291 path->l[b->level].lock_taken_time = local_clock();292#endif293 }294 295 return ret;296}297 298int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,299 struct btree_bkey_cached_common *b, bool);300 301static inline int __btree_node_lock_write(struct btree_trans *trans,302 struct btree_path *path,303 struct btree_bkey_cached_common *b,304 bool lock_may_not_fail)305{306 EBUG_ON(&path->l[b->level].b->c != b);307 EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));308 EBUG_ON(!btree_node_intent_locked(path, b->level));309 310 /*311 * six locks are unfair, and read locks block while a thread wants a312 * write lock: thus, we need to tell the cycle detector we have a write313 * lock _before_ taking the lock:314 */315 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED);316 317 return likely(six_trylock_write(&b->lock))318 ? 0319 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);320}321 322static inline int __must_check323bch2_btree_node_lock_write(struct btree_trans *trans,324 struct btree_path *path,325 struct btree_bkey_cached_common *b)326{327 return __btree_node_lock_write(trans, path, b, false);328}329 330void bch2_btree_node_lock_write_nofail(struct btree_trans *,331 struct btree_path *,332 struct btree_bkey_cached_common *);333 334/* relock: */335 336bool bch2_btree_path_relock_norestart(struct btree_trans *, struct btree_path *);337int __bch2_btree_path_relock(struct btree_trans *,338 struct btree_path *, unsigned long);339 340static inline int bch2_btree_path_relock(struct btree_trans *trans,341 struct btree_path *path, unsigned long trace_ip)342{343 return btree_node_locked(path, path->level)344 ? 0345 : __bch2_btree_path_relock(trans, path, trace_ip);346}347 348bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);349 350static inline bool bch2_btree_node_relock(struct btree_trans *trans,351 struct btree_path *path, unsigned level)352{353 EBUG_ON(btree_node_locked(path, level) &&354 !btree_node_write_locked(path, level) &&355 btree_node_locked_type(path, level) != __btree_lock_want(path, level));356 357 return likely(btree_node_locked(path, level)) ||358 (!IS_ERR_OR_NULL(path->l[level].b) &&359 __bch2_btree_node_relock(trans, path, level, true));360}361 362static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,363 struct btree_path *path, unsigned level)364{365 EBUG_ON(btree_node_locked(path, level) &&366 !btree_node_write_locked(path, level) &&367 btree_node_locked_type(path, level) != __btree_lock_want(path, level));368 369 return likely(btree_node_locked(path, level)) ||370 (!IS_ERR_OR_NULL(path->l[level].b) &&371 __bch2_btree_node_relock(trans, path, level, false));372}373 374/* upgrade */375 376bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,377 struct btree_path *, unsigned,378 struct get_locks_fail *);379 380bool __bch2_btree_path_upgrade(struct btree_trans *,381 struct btree_path *, unsigned,382 struct get_locks_fail *);383 384static inline int bch2_btree_path_upgrade(struct btree_trans *trans,385 struct btree_path *path,386 unsigned new_locks_want)387{388 struct get_locks_fail f = {};389 unsigned old_locks_want = path->locks_want;390 391 new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);392 393 if (path->locks_want < new_locks_want394 ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f)395 : path->nodes_locked)396 return 0;397 398 trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,399 old_locks_want, new_locks_want, &f);400 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);401}402 403/* misc: */404 405static inline void btree_path_set_should_be_locked(struct btree_trans *trans, struct btree_path *path)406{407 EBUG_ON(!btree_node_locked(path, path->level));408 EBUG_ON(path->uptodate);409 410 path->should_be_locked = true;411 trace_btree_path_should_be_locked(trans, path);412}413 414static 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 422static 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 431struct 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 436int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);437 438#ifdef CONFIG_BCACHEFS_DEBUG439void bch2_btree_path_verify_locks(struct btree_path *);440void bch2_trans_verify_locks(struct btree_trans *);441#else442static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}443static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}444#endif445 446#endif /* _BCACHEFS_BTREE_LOCKING_H */447