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 *, enum six_lock_init_flags);
  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 void bch2_trans_unlock_noassert(struct btree_trans *);
  26
  27 static inline bool is_btree_node(struct btree_path *path, unsigned l)
  28 {
  29         return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
  30 }
  31
  32 static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
  33 {
  34         return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
  35                 ? &trans->c->btree_transaction_stats[trans->fn_idx]
  36                 : NULL;
  37 }
  38
  39 /* matches six lock types */
  40 enum btree_node_locked_type {
  41         BTREE_NODE_UNLOCKED             = -1,
  42         BTREE_NODE_READ_LOCKED          = SIX_LOCK_read,
  43         BTREE_NODE_INTENT_LOCKED        = SIX_LOCK_intent,
  44         BTREE_NODE_WRITE_LOCKED         = SIX_LOCK_write,
  45 };
  46
  47 static inline int btree_node_locked_type(struct btree_path *path,
  48                                          unsigned level)
  49 {
  50         return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
  51 }
  52
  53 static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
  54 {
  55         return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
  56 }
  57
  58 static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
  59 {
  60         return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
  61 }
  62
  63 static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
  64 {
  65         return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
  66 }
  67
  68 static inline bool btree_node_locked(struct btree_path *path, unsigned level)
  69 {
  70         return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
  71 }
  72
  73 static inline void mark_btree_node_locked_noreset(struct btree_path *path,
  74                                                   unsigned level,
  75                                                   enum btree_node_locked_type type)
  76 {
  77         /* relying on this to avoid a branch */
  78         BUILD_BUG_ON(SIX_LOCK_read   != 0);
  79         BUILD_BUG_ON(SIX_LOCK_intent != 1);
  80
  81         path->nodes_locked &= ~(3U << (level << 1));
  82         path->nodes_locked |= (type + 1) << (level << 1);
  83 }
  84
  85 static inline void mark_btree_node_unlocked(struct btree_path *path,
  86                                             unsigned level)
  87 {
  88         EBUG_ON(btree_node_write_locked(path, level));
  89         mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
  90 }
  91
  92 static inline void mark_btree_node_locked(struct btree_trans *trans,
  93                                           struct btree_path *path,
  94                                           unsigned level,
  95                                           enum six_lock_type type)
  96 {
  97         mark_btree_node_locked_noreset(path, level, (enum btree_node_locked_type) type);
  98 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
  99         path->l[level].lock_taken_time = local_clock();
 100 #endif
 101 }
 102
 103 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
 104 {
 105         return level < path->locks_want
 106                 ? SIX_LOCK_intent
 107                 : SIX_LOCK_read;
 108 }
 109
 110 static inline enum btree_node_locked_type
 111 btree_lock_want(struct btree_path *path, int level)
 112 {
 113         if (level < path->level)
 114                 return BTREE_NODE_UNLOCKED;
 115         if (level < path->locks_want)
 116                 return BTREE_NODE_INTENT_LOCKED;
 117         if (level == path->level)
 118                 return BTREE_NODE_READ_LOCKED;
 119         return BTREE_NODE_UNLOCKED;
 120 }
 121
 122 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 123                                               struct btree_path *path, unsigned level)
 124 {
 125 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
 126         struct btree_transaction_stats *s = btree_trans_stats(trans);
 127
 128         if (s)
 129                 __bch2_time_stats_update(&s->lock_hold_times,
 130                                          path->l[level].lock_taken_time,
 131                                          local_clock());
 132 #endif
 133 }
 134
 135 /* unlock: */
 136
 137 static inline void btree_node_unlock(struct btree_trans *trans,
 138                                      struct btree_path *path, unsigned level)
 139 {
 140         int lock_type = btree_node_locked_type(path, level);
 141
 142         EBUG_ON(level >= BTREE_MAX_DEPTH);
 143
 144         if (lock_type != BTREE_NODE_UNLOCKED) {
 145                 six_unlock_type(&path->l[level].b->c.lock, lock_type);
 146                 btree_trans_lock_hold_time_update(trans, path, level);
 147         }
 148         mark_btree_node_unlocked(path, level);
 149 }
 150
 151 static inline int btree_path_lowest_level_locked(struct btree_path *path)
 152 {
 153         return __ffs(path->nodes_locked) >> 1;
 154 }
 155
 156 static inline int btree_path_highest_level_locked(struct btree_path *path)
 157 {
 158         return __fls(path->nodes_locked) >> 1;
 159 }
 160
 161 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
 162                                             struct btree_path *path)
 163 {
 164         btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
 165
 166         while (path->nodes_locked)
 167                 btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
 168 }
 169
 170 /*
 171  * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
 172  * succeed:
 173  */
 174 static inline void
 175 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 176                                      struct btree *b)
 177 {
 178         struct btree_path *linked;
 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, SIX_LOCK_intent);
 185
 186         trans_for_each_path_with_node(trans, b, linked)
 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
 247         trans_for_each_path(trans, path)
 248                 if (&path->l[level].b->c == b &&
 249                     btree_node_locked_type(path, level) >= want) {
 250                         six_lock_increment(&b->lock, (enum six_lock_type) want);
 251                         return true;
 252                 }
 253
 254         return false;
 255 }
 256
 257 static inline int btree_node_lock(struct btree_trans *trans,
 258                         struct btree_path *path,
 259                         struct btree_bkey_cached_common *b,
 260                         unsigned level,
 261                         enum six_lock_type type,
 262                         unsigned long ip)
 263 {
 264         int ret = 0;
 265
 266         EBUG_ON(level >= BTREE_MAX_DEPTH);
 267         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
 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, SIX_LOCK_write);
 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 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
 360                                struct btree_path *, unsigned);
 361 bool __bch2_btree_path_upgrade(struct btree_trans *,
 362                                struct btree_path *, unsigned);
 363
 364 static inline int bch2_btree_path_upgrade(struct btree_trans *trans,
 365                                           struct btree_path *path,
 366                                           unsigned new_locks_want)
 367 {
 368         unsigned old_locks_want = path->locks_want;
 369
 370         new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
 371
 372         if (path->locks_want < new_locks_want
 373             ? __bch2_btree_path_upgrade(trans, path, new_locks_want)
 374             : path->uptodate == BTREE_ITER_UPTODATE)
 375                 return 0;
 376
 377         trace_and_count(trans->c, trans_restart_upgrade, trans, _THIS_IP_, path,
 378                         old_locks_want, new_locks_want);
 379         return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
 380 }
 381
 382 /* misc: */
 383
 384 static inline void btree_path_set_should_be_locked(struct btree_path *path)
 385 {
 386         EBUG_ON(!btree_node_locked(path, path->level));
 387         EBUG_ON(path->uptodate);
 388
 389         path->should_be_locked = true;
 390 }
 391
 392 static inline void __btree_path_set_level_up(struct btree_trans *trans,
 393                                       struct btree_path *path,
 394                                       unsigned l)
 395 {
 396         btree_node_unlock(trans, path, l);
 397         path->l[l].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
 398 }
 399
 400 static inline void btree_path_set_level_up(struct btree_trans *trans,
 401                                     struct btree_path *path)
 402 {
 403         __btree_path_set_level_up(trans, path, path->level++);
 404         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 405 }
 406
 407 /* debug */
 408
 409 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
 410                                 struct btree_path *,
 411                                 struct btree_bkey_cached_common *b,
 412                                 unsigned);
 413
 414 int bch2_check_for_deadlock(struct btree_trans *, struct printbuf *);
 415
 416 #ifdef CONFIG_BCACHEFS_DEBUG
 417 void bch2_btree_path_verify_locks(struct btree_path *);
 418 void bch2_trans_verify_locks(struct btree_trans *);
 419 #else
 420 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 421 static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
 422 #endif
 423
 424 #endif /* _BCACHEFS_BTREE_LOCKING_H */