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 extern struct lock_class_key bch2_btree_node_lock_key;
  18
  19 static inline bool is_btree_node(struct btree_path *path, unsigned l)
  20 {
  21         return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
  22 }
  23
  24 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
  25 {
  26         return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
  27                 ? &trans->c->btree_transaction_stats[trans->fn_idx]
  28                 : NULL;
  29 }
  30
  31 /* matches six lock types */
  32 enum btree_node_locked_type {
  33         BTREE_NODE_UNLOCKED             = -1,
  34         BTREE_NODE_READ_LOCKED          = SIX_LOCK_read,
  35         BTREE_NODE_INTENT_LOCKED        = SIX_LOCK_intent,
  36         BTREE_NODE_WRITE_LOCKED         = SIX_LOCK_write,
  37 };
  38
  39 static inline int btree_node_locked_type(struct btree_path *path,
  40                                          unsigned level)
  41 {
  42         return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
  43 }
  44
  45 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
  46 {
  47         return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
  48 }
  49
  50 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
  51 {
  52         return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
  53 }
  54
  55 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
  56 {
  57         return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
  58 }
  59
  60 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
  61 {
  62         return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
  63 }
  64
  65 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
  66                                                   unsigned level,
  67                                                   enum btree_node_locked_type type)
  68 {
  69         /* relying on this to avoid a branch */
  70         BUILD_BUG_ON(SIX_LOCK_read   != 0);
  71         BUILD_BUG_ON(SIX_LOCK_intent != 1);
  72
  73         path->nodes_locked &= ~(3U << (level << 1));
  74         path->nodes_locked |= (type + 1) << (level << 1);
  75 }
  76
  77 static inline void mark_btree_node_unlocked(struct btree_path *path,
  78                                             unsigned level)
  79 {
  80         EBUG_ON(btree_node_write_locked(path, level));
  81         mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
  82 }
  83
  84 static inline void mark_btree_node_locked(struct btree_trans *trans,
  85                                           struct btree_path *path,
  86                                           unsigned level,
  87                                           enum six_lock_type type)
  88 {
  89         mark_btree_node_locked_noreset(path, level, type);
  90 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
  91         path->l[level].lock_taken_time = local_clock();
  92 #endif
  93 }
  94
  95 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
  96 {
  97         return level < path->locks_want
  98                 ? SIX_LOCK_intent
  99                 : SIX_LOCK_read;
 100 }
 101
 102 static inline enum btree_node_locked_type
 103 btree_lock_want(struct btree_path *path, int level)
 104 {
 105         if (level < path->level)
 106                 return BTREE_NODE_UNLOCKED;
 107         if (level < path->locks_want)
 108                 return BTREE_NODE_INTENT_LOCKED;
 109         if (level == path->level)
 110                 return BTREE_NODE_READ_LOCKED;
 111         return BTREE_NODE_UNLOCKED;
 112 }
 113
 114 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 115                                               struct btree_path *path, unsigned level)
 116 {
 117 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
 118         struct btree_transaction_stats *s = btree_trans_stats(trans);
 119
 120         if (s)
 121                 __bch2_time_stats_update(&s->lock_hold_times,
 122                                          path->l[level].lock_taken_time,
 123                                          local_clock());
 124 #endif
 125 }
 126
 127 /* unlock: */
 128
 129 static inline void btree_node_unlock(struct btree_trans *trans,
 130                                      struct btree_path *path, unsigned level)
 131 {
 132         int lock_type = btree_node_locked_type(path, level);
 133
 134         EBUG_ON(level >= BTREE_MAX_DEPTH);
 135
 136         if (lock_type != BTREE_NODE_UNLOCKED) {
 137                 six_unlock_type(&path->l[level].b->c.lock, lock_type);
 138                 btree_trans_lock_hold_time_update(trans, path, level);
 139         }
 140         mark_btree_node_unlocked(path, level);
 141 }
 142
 143 static inline int btree_path_lowest_level_locked(struct btree_path *path)
 144 {
 145         return __ffs(path->nodes_locked) >> 1;
 146 }
 147
 148 static inline int btree_path_highest_level_locked(struct btree_path *path)
 149 {
 150         return __fls(path->nodes_locked) >> 1;
 151 }
 152
 153 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
 154                                             struct btree_path *path)
 155 {
 156         btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
 157
 158         while (path->nodes_locked)
 159                 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
 160 }
 161
 162 /*
 163  * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
 164  * succeed:
 165  */
 166 static inline void
 167 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 168                                      struct btree *b)
 169 {
 170         struct btree_path *linked;
 171
 172         EBUG_ON(path->l[b->c.level].b != b);
 173         EBUG_ON(path->l[b->c.level].lock_seq + 1 != b->c.lock.state.seq);
 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, SIX_LOCK_intent);
 177
 178         trans_for_each_path_with_node(trans, b, linked)
 179                 linked->l[b->c.level].lock_seq += 2;
 180
 181         six_unlock_write(&b->c.lock);
 182 }
 183
 184 void bch2_btree_node_unlock_write(struct btree_trans *,
 185                         struct btree_path *, struct btree *);
 186
 187 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
 188
 189 /* lock: */
 190
 191 static inline int __btree_node_lock_nopath(struct btree_trans *trans,
 192                                          struct btree_bkey_cached_common *b,
 193                                          enum six_lock_type type,
 194                                          bool lock_may_not_fail)
 195 {
 196         int ret;
 197
 198         trans->lock_may_not_fail = lock_may_not_fail;
 199         trans->lock_must_abort  = false;
 200         trans->locking          = b;
 201
 202         ret = six_lock_type_waiter(&b->lock, type, &trans->locking_wait,
 203                                    bch2_six_check_for_deadlock, trans);
 204         WRITE_ONCE(trans->locking, NULL);
 205         WRITE_ONCE(trans->locking_wait.start_time, 0);
 206         return ret;
 207 }
 208
 209 static inline int __must_check
 210 btree_node_lock_nopath(struct btree_trans *trans,
 211                        struct btree_bkey_cached_common *b,
 212                        enum six_lock_type type)
 213 {
 214         return __btree_node_lock_nopath(trans, b, type, false);
 215 }
 216
 217 static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
 218                                          struct btree_bkey_cached_common *b,
 219                                          enum six_lock_type type)
 220 {
 221         int ret = __btree_node_lock_nopath(trans, b, type, true);
 222
 223         BUG_ON(ret);
 224 }
 225
 226 /*
 227  * Lock a btree node if we already have it locked on one of our linked
 228  * iterators:
 229  */
 230 static inline bool btree_node_lock_increment(struct btree_trans *trans,
 231                                              struct btree_bkey_cached_common *b,
 232                                              unsigned level,
 233                                              enum btree_node_locked_type want)
 234 {
 235         struct btree_path *path;
 236
 237         trans_for_each_path(trans, path)
 238                 if (&path->l[level].b->c == b &&
 239                     btree_node_locked_type(path, level) >= want) {
 240                         six_lock_increment(&b->lock, want);
 241                         return true;
 242                 }
 243
 244         return false;
 245 }
 246
 247 static inline int btree_node_lock(struct btree_trans *trans,
 248                         struct btree_path *path,
 249                         struct btree_bkey_cached_common *b,
 250                         unsigned level,
 251                         enum six_lock_type type,
 252                         unsigned long ip)
 253 {
 254         int ret = 0;
 255
 256         EBUG_ON(level >= BTREE_MAX_DEPTH);
 257         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
 258
 259         if (likely(six_trylock_type(&b->lock, type)) ||
 260             btree_node_lock_increment(trans, b, level, type) ||
 261             !(ret = btree_node_lock_nopath(trans, b, type))) {
 262 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
 263                 path->l[b->level].lock_taken_time = local_clock();
 264 #endif
 265         }
 266
 267         return ret;
 268 }
 269
 270 int __bch2_btree_node_lock_write(struct btree_trans *, struct btree_path *,
 271                                  struct btree_bkey_cached_common *b, bool);
 272
 273 static inline int __btree_node_lock_write(struct btree_trans *trans,
 274                                           struct btree_path *path,
 275                                           struct btree_bkey_cached_common *b,
 276                                           bool lock_may_not_fail)
 277 {
 278         EBUG_ON(&path->l[b->level].b->c != b);
 279         EBUG_ON(path->l[b->level].lock_seq != b->lock.state.seq);
 280         EBUG_ON(!btree_node_intent_locked(path, b->level));
 281
 282         /*
 283          * six locks are unfair, and read locks block while a thread wants a
 284          * write lock: thus, we need to tell the cycle detector we have a write
 285          * lock _before_ taking the lock:
 286          */
 287         mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_write);
 288
 289         return likely(six_trylock_write(&b->lock))
 290                 ? 0
 291                 : __bch2_btree_node_lock_write(trans, path, b, lock_may_not_fail);
 292 }
 293
 294 static inline void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
 295                                               struct btree_path *path,
 296                                               struct btree_bkey_cached_common *b)
 297 {
 298         int ret = __btree_node_lock_write(trans, path, b, true);
 299         BUG_ON(ret);
 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 /* relock: */
 311
 312 bool bch2_btree_path_relock_norestart(struct btree_trans *,
 313                                       struct btree_path *, unsigned long);
 314 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned, bool trace);
 315
 316 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
 317                                           struct btree_path *path, unsigned level)
 318 {
 319         EBUG_ON(btree_node_locked(path, level) &&
 320                 !btree_node_write_locked(path, level) &&
 321                 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
 322
 323         return likely(btree_node_locked(path, level)) ||
 324                 (!IS_ERR_OR_NULL(path->l[level].b) &&
 325                  __bch2_btree_node_relock(trans, path, level, true));
 326 }
 327
 328 static inline bool bch2_btree_node_relock_notrace(struct btree_trans *trans,
 329                                                   struct btree_path *path, unsigned level)
 330 {
 331         EBUG_ON(btree_node_locked(path, level) &&
 332                 !btree_node_write_locked(path, level) &&
 333                 btree_node_locked_type(path, level) != __btree_lock_want(path, level));
 334
 335         return likely(btree_node_locked(path, level)) ||
 336                 (!IS_ERR_OR_NULL(path->l[level].b) &&
 337                  __bch2_btree_node_relock(trans, path, level, false));
 338 }
 339
 340 static inline int bch2_btree_path_relock(struct btree_trans *trans,
 341                                 struct btree_path *path, unsigned long trace_ip)
 342 {
 343         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 344                 trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
 345                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 346         }
 347
 348         return 0;
 349 }
 350
 351 /* upgrade */
 352
 353 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
 354                                struct btree_path *, unsigned);
 355 bool __bch2_btree_path_upgrade(struct btree_trans *,
 356                                struct btree_path *, unsigned);
 357
 358 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
 359                                           struct btree_path *path,
 360                                           unsigned new_locks_want)
 361 {
 362         unsigned old_locks_want = path->locks_want;
 363
 364         new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
 365
 366         if (path->locks_want < new_locks_want
 367             ? __bch2_btree_path_upgrade(trans, path, new_locks_want)
 368             : path->uptodate == BTREE_ITER_UPTODATE)
 369                 return 0;
 370
 371         trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
 372                         old_locks_want, new_locks_want);
 373         return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
 374 }
 375
 376 /* misc: */
 377
 378 static inline void btree_path_set_should_be_locked(struct btree_path *path)
 379 {
 380         EBUG_ON(!btree_node_locked(path, path->level));
 381         EBUG_ON(path->uptodate);
 382
 383         path->should_be_locked = true;
 384 }
 385
 386 static inline void __btree_path_set_level_up(struct btree_trans *trans,
 387                                       struct btree_path *path,
 388                                       unsigned l)
 389 {
 390         btree_node_unlock(trans, path, l);
 391         path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
 392 }
 393
 394 static inline void btree_path_set_level_up(struct btree_trans *trans,
 395                                     struct btree_path *path)
 396 {
 397         __btree_path_set_level_up(trans, path, path->level++);
 398         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 399 }
 400
 401 /* debug */
 402
 403 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
 404                                 struct btree_path *,
 405                                 struct btree_bkey_cached_common *b,
 406                                 unsigned);
 407
 408 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
 409
 410 #ifdef CONFIG_BCACHEFS_DEBUG
 411 void bch2_btree_path_verify_locks(struct btree_path *);
 412 void bch2_trans_verify_locks(struct btree_trans *);
 413 #else
 414 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 415 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
 416 #endif
 417
 418 #endif /* _BCACHEFS_BTREE_LOCKING_H */