git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_locking.h

   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         struct btree_transaction_stats *s = btree_trans_stats(trans);
 126
 127         if (s)
 128                 __bch2_time_stats_update(&s->lock_hold_times,
 129                                          path->l[level].lock_taken_time,
 130                                          local_clock());
 131 #endif
 132 }
 133
 134 /* unlock: */
 135
 136 static inline void btree_node_unlock(struct btree_trans *trans,
 137                                      struct btree_path *path, unsigned level)
 138 {
 139         int lock_type = btree_node_locked_type(path, level);
 140
 141         EBUG_ON(level >= BTREE_MAX_DEPTH);
 142
 143         if (lock_type != BTREE_NODE_UNLOCKED) {
 144                 six_unlock_type(&path->l[level].b->c.lock, lock_type);
 145                 btree_trans_lock_hold_time_update(trans, path, level);
 146         }
 147         mark_btree_node_unlocked(path, level);
 148 }
 149
 150 static inline int btree_path_lowest_level_locked(struct btree_path *path)
 151 {
 152         return __ffs(path->nodes_locked) >> 1;
 153 }
 154
 155 static inline int btree_path_highest_level_locked(struct btree_path *path)
 156 {
 157         return __fls(path->nodes_locked) >> 1;
 158 }
 159
 160 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
 161                                             struct btree_path *path)
 162 {
 163         btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
 164
 165         while (path->nodes_locked)
 166                 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
 167 }
 168
 169 /*
 170  * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
 171  * succeed:
 172  */
 173 static inline void
 174 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 175                                      struct btree *b)
 176 {
 177         struct btree_path *linked;
 178         unsigned i;
 179
 180         EBUG_ON(path->l[b->c.level].b != b);
 181         EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
 182         EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
 183
 184         mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 185
 186         trans_for_each_path_with_node(trans, b, linked, i)
 187                 linked->l[b->c.level].lock_seq++;
 188
 189         six_unlock_write(&b->c.lock);
 190 }
 191
 192 void bch2_btree_node_unlock_write(struct btree_trans *,
 193                         struct btree_path *, struct btree *);
 194
 195 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
 196
 197 /* lock: */
 198
 199 static inline int __btree_node_lock_nopath(struct btree_trans *trans,
 200                                          struct btree_bkey_cached_common *b,
 201                                          enum six_lock_type type,
 202                                          bool lock_may_not_fail,
 203                                          unsigned long ip)
 204 {
 205         int ret;
 206
 207         trans->lock_may_not_fail = lock_may_not_fail;
 208         trans->lock_must_abort  = false;
 209         trans->locking          = b;
 210
 211         ret = six_lock_ip_waiter(&b->lock, type, &trans->locking_wait,
 212                                  bch2_six_check_for_deadlock, trans, ip);
 213         WRITE_ONCE(trans->locking, NULL);
 214         WRITE_ONCE(trans->locking_wait.start_time, 0);
 215         return ret;
 216 }
 217
 218 static inline int __must_check
 219 btree_node_lock_nopath(struct btree_trans *trans,
 220                        struct btree_bkey_cached_common *b,
 221                        enum six_lock_type type,
 222                        unsigned long ip)
 223 {
 224         return __btree_node_lock_nopath(trans, b, type, false, ip);
 225 }
 226
 227 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
 228                                          struct btree_bkey_cached_common *b,
 229                                          enum six_lock_type type)
 230 {
 231         int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
 232
 233         BUG_ON(ret);
 234 }
 235
 236 /*
 237  * Lock a btree node if we already have it locked on one of our linked
 238  * iterators:
 239  */
 240 static inline bool btree_node_lock_increment(struct btree_trans *trans,
 241                                              struct btree_bkey_cached_common *b,
 242                                              unsigned level,
 243                                              enum btree_node_locked_type want)
 244 {
 245         struct btree_path *path;
 246         unsigned i;
 247
 248         trans_for_each_path(trans, path, i)
 249                 if (&path->l[level].b->c == b &&
 250                     btree_node_locked_type(path, level) >= want) {
 251                         six_lock_increment(&b->lock, (enum six_lock_type) want);
 252                         return true;
 253                 }
 254
 255         return false;
 256 }
 257
 258 static inline int btree_node_lock(struct btree_trans *trans,
 259                         struct btree_path *path,
 260                         struct btree_bkey_cached_common *b,
 261                         unsigned level,
 262                         enum six_lock_type type,
 263                         unsigned long ip)
 264 {
 265         int ret = 0;
 266
 267         EBUG_ON(level >= BTREE_MAX_DEPTH);
 268
 269         if (likely(six_trylock_type(&b->lock, type)) ||
 270             btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||
 271             !(ret = btree_node_lock_nopath(trans, b, type, btree_path_ip_allocated(path)))) {
 272 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
 273                 path->l[b->level].lock_taken_time = local_clock();
 274 #endif
 275         }
 276
 277         return ret;
 278 }
 279
 280 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
 281                                  struct btree_bkey_cached_common *b, bool);
 282
 283 static inline int __btree_node_lock_write(struct btree_trans *trans,
 284                                           struct btree_path *path,
 285                                           struct btree_bkey_cached_common *b,
 286                                           bool lock_may_not_fail)
 287 {
 288         EBUG_ON(&path->l[b->level].b->c != b);
 289         EBUG_ON(path->l[b->level].lock_seq != six_lock_seq(&b->lock));
 290         EBUG_ON(!btree_node_intent_locked(path, b->level));
 291
 292         /*
 293          * six locks are unfair, and read locks block while a thread wants a
 294          * write lock: thus, we need to tell the cycle detector we have a write
 295          * lock _before_ taking the lock:
 296          */
 297         mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_WRITE_LOCKED);
 298
 299         return likely(six_trylock_write(&b->lock))
 300                 ? 0
 301                 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
 302 }
 303
 304 static inline int __must_check
 305 bch2_btree_node_lock_write(struct btree_trans *trans,
 306                            struct btree_path *path,
 307                            struct btree_bkey_cached_common *b)
 308 {
 309         return __btree_node_lock_write(trans, path, b, false);
 310 }
 311
 312 void bch2_btree_node_lock_write_nofail(struct btree_trans *,
 313                                        struct btree_path *,
 314                                        struct btree_bkey_cached_common *);
 315
 316 /* relock: */
 317
 318 bool bch2_btree_path_relock_norestart(struct btree_trans *,
 319                                       struct btree_path *, unsigned long);
 320 int __bch2_btree_path_relock(struct btree_trans *,
 321                              struct btree_path *, unsigned long);
 322
 323 static inline int bch2_btree_path_relock(struct btree_trans *trans,
 324                                 struct btree_path *path, unsigned long trace_ip)
 325 {
 326         return btree_node_locked(path, path->level)
 327                 ? 0
 328                 : __bch2_btree_path_relock(trans, path, trace_ip);
 329 }
 330
 331 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
 332
 333 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
 334                                           struct btree_path *path, unsigned level)
 335 {
 336         EBUG_ON(btree_node_locked(path, level) &&
 337                 !btree_node_write_locked(path, level) &&
 338                 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
 339
 340         return likely(btree_node_locked(path, level)) ||
 341                 (!IS_ERR_OR_NULL(path->l[level].b) &&
 342                  __bch2_btree_node_relock(trans, path, level, true));
 343 }
 344
 345 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
 346                                                   struct btree_path *path, unsigned level)
 347 {
 348         EBUG_ON(btree_node_locked(path, level) &&
 349                 !btree_node_write_locked(path, level) &&
 350                 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
 351
 352         return likely(btree_node_locked(path, level)) ||
 353                 (!IS_ERR_OR_NULL(path->l[level].b) &&
 354                  __bch2_btree_node_relock(trans, path, level, false));
 355 }
 356
 357 /* upgrade */
 358
 359
 360 struct get_locks_fail {
 361         unsigned        l;
 362         struct btree    *b;
 363 };
 364
 365 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
 366                                struct btree_path *, unsigned,
 367                                struct get_locks_fail *);
 368
 369 bool __bch2_btree_path_upgrade(struct btree_trans *,
 370                                struct btree_path *, unsigned,
 371                                struct get_locks_fail *);
 372
 373 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
 374                                           struct btree_path *path,
 375                                           unsigned new_locks_want)
 376 {
 377         struct get_locks_fail f;
 378         unsigned old_locks_want = path->locks_want;
 379
 380         new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
 381
 382         if (path->locks_want < new_locks_want
 383             ? __bch2_btree_path_upgrade(trans, path, new_locks_want, &f)
 384             : path->uptodate == BTREE_ITER_UPTODATE)
 385                 return 0;
 386
 387         trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
 388                         old_locks_want, new_locks_want, &f);
 389         return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
 390 }
 391
 392 /* misc: */
 393
 394 static inline void btree_path_set_should_be_locked(struct btree_path *path)
 395 {
 396         EBUG_ON(!btree_node_locked(path, path->level));
 397         EBUG_ON(path->uptodate);
 398
 399         path->should_be_locked = true;
 400 }
 401
 402 static inline void __btree_path_set_level_up(struct btree_trans *trans,
 403                                       struct btree_path *path,
 404                                       unsigned l)
 405 {
 406         btree_node_unlock(trans, path, l);
 407         path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
 408 }
 409
 410 static inline void btree_path_set_level_up(struct btree_trans *trans,
 411                                     struct btree_path *path)
 412 {
 413         __btree_path_set_level_up(trans, path, path->level++);
 414         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 415 }
 416
 417 /* debug */
 418
 419 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
 420                                 struct btree_path *,
 421                                 struct btree_bkey_cached_common *b,
 422                                 unsigned);
 423
 424 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
 425
 426 #ifdef CONFIG_BCACHEFS_DEBUG
 427 void bch2_btree_path_verify_locks(struct btree_path *);
 428 void bch2_trans_verify_locks(struct btree_trans *);
 429 #else
 430 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 431 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
 432 #endif
 433
 434 #endif /* _BCACHEFS_BTREE_LOCKING_H */