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 <linux/six.h>
  14
  15 #include "btree_iter.h"
  16
  17 void bch2_btree_lock_init(struct btree_bkey_cached_common *);
  18
  19 #ifdef CONFIG_LOCKDEP
  20 void bch2_assert_btree_nodes_not_locked(void);
  21 #else
  22 static inline void bch2_assert_btree_nodes_not_locked(void) {}
  23 #endif
  24
  25 static inline bool is_btree_node(struct btree_path *path, unsigned l)
  26 {
  27         return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
  28 }
  29
  30 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
  31 {
  32         return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
  33                 ? &trans->c->btree_transaction_stats[trans->fn_idx]
  34                 : NULL;
  35 }
  36
  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,
  43 };
  44
  45 static inline int btree_node_locked_type(struct btree_path *path,
  46                                          unsigned level)
  47 {
  48         return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
  49 }
  50
  51 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
  52 {
  53         return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
  54 }
  55
  56 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
  57 {
  58         return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
  59 }
  60
  61 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
  62 {
  63         return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
  64 }
  65
  66 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
  67 {
  68         return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
  69 }
  70
  71 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
  72                                                   unsigned level,
  73                                                   enum btree_node_locked_type type)
  74 {
  75         /* relying on this to avoid a branch */
  76         BUILD_BUG_ON(SIX_LOCK_read   != 0);
  77         BUILD_BUG_ON(SIX_LOCK_intent != 1);
  78
  79         path->nodes_locked &= ~(3U << (level << 1));
  80         path->nodes_locked |= (type + 1) << (level << 1);
  81 }
  82
  83 static inline void mark_btree_node_unlocked(struct btree_path *path,
  84                                             unsigned level)
  85 {
  86         EBUG_ON(btree_node_write_locked(path, level));
  87         mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
  88 }
  89
  90 static inline void mark_btree_node_locked(struct btree_trans *trans,
  91                                           struct btree_path *path,
  92                                           unsigned level,
  93                                           enum six_lock_type type)
  94 {
  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();
  98 #endif
  99 }
 100
 101 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
 102 {
 103         return level < path->locks_want
 104                 ? SIX_LOCK_intent
 105                 : SIX_LOCK_read;
 106 }
 107
 108 static inline enum btree_node_locked_type
 109 btree_lock_want(struct btree_path *path, int level)
 110 {
 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;
 118 }
 119
 120 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 121                                               struct btree_path *path, unsigned level)
 122 {
 123 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
 124         struct btree_transaction_stats *s = btree_trans_stats(trans);
 125
 126         if (s)
 127                 __bch2_time_stats_update(&s->lock_hold_times,
 128                                          path->l[level].lock_taken_time,
 129                                          local_clock());
 130 #endif
 131 }
 132
 133 /* unlock: */
 134
 135 static inline void btree_node_unlock(struct btree_trans *trans,
 136                                      struct btree_path *path, unsigned level)
 137 {
 138         int lock_type = btree_node_locked_type(path, level);
 139
 140         EBUG_ON(level >= BTREE_MAX_DEPTH);
 141
 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);
 145         }
 146         mark_btree_node_unlocked(path, level);
 147 }
 148
 149 static inline int btree_path_lowest_level_locked(struct btree_path *path)
 150 {
 151         return __ffs(path->nodes_locked) >> 1;
 152 }
 153
 154 static inline int btree_path_highest_level_locked(struct btree_path *path)
 155 {
 156         return __fls(path->nodes_locked) >> 1;
 157 }
 158
 159 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
 160                                             struct btree_path *path)
 161 {
 162         btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
 163
 164         while (path->nodes_locked)
 165                 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
 166 }
 167
 168 /*
 169  * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
 170  * succeed:
 171  */
 172 static inline void
 173 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 174                                      struct btree *b)
 175 {
 176         struct btree_path *linked;
 177
 178         EBUG_ON(path->l[b->c.level].b != b);
 179         EBUG_ON(path->l[b->c.level].lock_seq + 1 != b->c.lock.state.seq);
 180         EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
 181
 182         mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
 183
 184         trans_for_each_path_with_node(trans, b, linked)
 185                 linked->l[b->c.level].lock_seq += 2;
 186
 187         six_unlock_write(&b->c.lock);
 188 }
 189
 190 void bch2_btree_node_unlock_write(struct btree_trans *,
 191                         struct btree_path *, struct btree *);
 192
 193 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
 194
 195 /* lock: */
 196
 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,
 201                                          unsigned long ip)
 202 {
 203         int ret;
 204
 205         trans->lock_may_not_fail = lock_may_not_fail;
 206         trans->lock_must_abort  = false;
 207         trans->locking          = b;
 208
 209         ret = six_lock_type_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);
 213         return ret;
 214 }
 215
 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,
 220                        unsigned long ip)
 221 {
 222         return __btree_node_lock_nopath(trans, b, type, false, ip);
 223 }
 224
 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)
 228 {
 229         int ret = __btree_node_lock_nopath(trans, b, type, true, _THIS_IP_);
 230
 231         BUG_ON(ret);
 232 }
 233
 234 /*
 235  * Lock a btree node if we already have it locked on one of our linked
 236  * iterators:
 237  */
 238 static inline bool btree_node_lock_increment(struct btree_trans *trans,
 239                                              struct btree_bkey_cached_common *b,
 240                                              unsigned level,
 241                                              enum btree_node_locked_type want)
 242 {
 243         struct btree_path *path;
 244
 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);
 249                         return true;
 250                 }
 251
 252         return false;
 253 }
 254
 255 static inline int btree_node_lock(struct btree_trans *trans,
 256                         struct btree_path *path,
 257                         struct btree_bkey_cached_common *b,
 258                         unsigned level,
 259                         enum six_lock_type type,
 260                         unsigned long ip)
 261 {
 262         int ret = 0;
 263
 264         EBUG_ON(level >= BTREE_MAX_DEPTH);
 265         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
 266
 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();
 272 #endif
 273         }
 274
 275         return ret;
 276 }
 277
 278 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
 279                                  struct btree_bkey_cached_common *b, bool);
 280
 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)
 285 {
 286         EBUG_ON(&path->l[b->level].b->c != b);
 287         EBUG_ON(path->l[b->level].lock_seq != b->lock.state.seq);
 288         EBUG_ON(!btree_node_intent_locked(path, b->level));
 289
 290         /*
 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:
 294          */
 295         mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_write);
 296
 297         return likely(six_trylock_write(&b->lock))
 298                 ? 0
 299                 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
 300 }
 301
 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)
 306 {
 307         return __btree_node_lock_write(trans, path, b, false);
 308 }
 309
 310 void bch2_btree_node_lock_write_nofail(struct btree_trans *,
 311                                        struct btree_path *,
 312                                        struct btree_bkey_cached_common *);
 313
 314 /* relock: */
 315
 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);
 320
 321 static inline int bch2_btree_path_relock(struct btree_trans *trans,
 322                                 struct btree_path *path, unsigned long trace_ip)
 323 {
 324         return btree_node_locked(path, path->level)
 325                 ? 0
 326                 : __bch2_btree_path_relock(trans, path, trace_ip);
 327 }
 328
 329 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
 330
 331 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
 332                                           struct btree_path *path, unsigned level)
 333 {
 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));
 337
 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));
 341 }
 342
 343 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
 344                                                   struct btree_path *path, unsigned level)
 345 {
 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));
 349
 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));
 353 }
 354
 355 /* upgrade */
 356
 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);
 361
 362 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
 363                                           struct btree_path *path,
 364                                           unsigned new_locks_want)
 365 {
 366         unsigned old_locks_want = path->locks_want;
 367
 368         new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
 369
 370         if (path->locks_want < new_locks_want
 371             ? __bch2_btree_path_upgrade(trans, path, new_locks_want)
 372             : path->uptodate == BTREE_ITER_UPTODATE)
 373                 return 0;
 374
 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);
 378 }
 379
 380 /* misc: */
 381
 382 static inline void btree_path_set_should_be_locked(struct btree_path *path)
 383 {
 384         EBUG_ON(!btree_node_locked(path, path->level));
 385         EBUG_ON(path->uptodate);
 386
 387         path->should_be_locked = true;
 388 }
 389
 390 static inline void __btree_path_set_level_up(struct btree_trans *trans,
 391                                       struct btree_path *path,
 392                                       unsigned l)
 393 {
 394         btree_node_unlock(trans, path, l);
 395         path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
 396 }
 397
 398 static inline void btree_path_set_level_up(struct btree_trans *trans,
 399                                     struct btree_path *path)
 400 {
 401         __btree_path_set_level_up(trans, path, path->level++);
 402         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 403 }
 404
 405 /* debug */
 406
 407 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
 408                                 struct btree_path *,
 409                                 struct btree_bkey_cached_common *b,
 410                                 unsigned);
 411
 412 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
 413
 414 #ifdef CONFIG_BCACHEFS_DEBUG
 415 void bch2_btree_path_verify_locks(struct btree_path *);
 416 void bch2_trans_verify_locks(struct btree_trans *);
 417 #else
 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) {}
 420 #endif
 421
 422 #endif /* _BCACHEFS_BTREE_LOCKING_H */