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

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _BCACHEFS_BTREE_ITER_H
   3 #define _BCACHEFS_BTREE_ITER_H
   4
   5 #include "bset.h"
   6 #include "btree_types.h"
   7 #include "trace.h"
   8
   9 static inline int __bkey_err(const struct bkey *k)
  10 {
  11         return PTR_ERR_OR_ZERO(k);
  12 }
  13
  14 #define bkey_err(_k)    __bkey_err((_k).k)
  15
  16 static inline void __btree_path_get(struct btree_path *path, bool intent)
  17 {
  18         path->ref++;
  19         path->intent_ref += intent;
  20 }
  21
  22 static inline bool __btree_path_put(struct btree_path *path, bool intent)
  23 {
  24         EBUG_ON(!path->ref);
  25         EBUG_ON(!path->intent_ref && intent);
  26         path->intent_ref -= intent;
  27         return --path->ref == 0;
  28 }
  29
  30 static inline void btree_path_set_dirty(struct btree_path *path,
  31                                         enum btree_path_uptodate u)
  32 {
  33         path->uptodate = max_t(unsigned, path->uptodate, u);
  34 }
  35
  36 static inline struct btree *btree_path_node(struct btree_path *path,
  37                                             unsigned level)
  38 {
  39         return level < BTREE_MAX_DEPTH ? path->l[level].b : NULL;
  40 }
  41
  42 static inline bool btree_node_lock_seq_matches(const struct btree_path *path,
  43                                         const struct btree *b, unsigned level)
  44 {
  45         return path->l[level].lock_seq == six_lock_seq(&b->c.lock);
  46 }
  47
  48 static inline struct btree *btree_node_parent(struct btree_path *path,
  49                                               struct btree *b)
  50 {
  51         return btree_path_node(path, b->c.level + 1);
  52 }
  53
  54 /* Iterate over paths within a transaction: */
  55
  56 void __bch2_btree_trans_sort_paths(struct btree_trans *);
  57
  58 static inline void btree_trans_sort_paths(struct btree_trans *trans)
  59 {
  60         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
  61             trans->paths_sorted)
  62                 return;
  63         __bch2_btree_trans_sort_paths(trans);
  64 }
  65
  66 static inline struct btree_path *
  67 __trans_next_path(struct btree_trans *trans, unsigned idx)
  68 {
  69         u64 l;
  70
  71         if (idx == BTREE_ITER_MAX)
  72                 return NULL;
  73
  74         l = trans->paths_allocated >> idx;
  75         if (!l)
  76                 return NULL;
  77
  78         idx += __ffs64(l);
  79         EBUG_ON(idx >= BTREE_ITER_MAX);
  80         EBUG_ON(trans->paths[idx].idx != idx);
  81         return &trans->paths[idx];
  82 }
  83
  84 #define trans_for_each_path_from(_trans, _path, _start)                 \
  85         for (_path = __trans_next_path((_trans), _start);               \
  86              (_path);                                                   \
  87              _path = __trans_next_path((_trans), (_path)->idx + 1))
  88
  89 #define trans_for_each_path(_trans, _path)                              \
  90         trans_for_each_path_from(_trans, _path, 0)
  91
  92 static inline struct btree_path *
  93 __trans_next_path_safe(struct btree_trans *trans, unsigned *idx)
  94 {
  95         u64 l;
  96
  97         if (*idx == BTREE_ITER_MAX)
  98                 return NULL;
  99
 100         l = trans->paths_allocated >> *idx;
 101         if (!l)
 102                 return NULL;
 103
 104         *idx += __ffs64(l);
 105         EBUG_ON(*idx >= BTREE_ITER_MAX);
 106         return &trans->paths[*idx];
 107 }
 108
 109 /*
 110  * This version is intended to be safe for use on a btree_trans that is owned by
 111  * another thread, for bch2_btree_trans_to_text();
 112  */
 113 #define trans_for_each_path_safe_from(_trans, _path, _idx, _start)      \
 114         for (_idx = _start;                                             \
 115              (_path = __trans_next_path_safe((_trans), &_idx));         \
 116              _idx++)
 117
 118 #define trans_for_each_path_safe(_trans, _path, _idx)                   \
 119         trans_for_each_path_safe_from(_trans, _path, _idx, 0)
 120
 121 static inline struct btree_path *next_btree_path(struct btree_trans *trans, struct btree_path *path)
 122 {
 123         unsigned idx = path ? path->sorted_idx + 1 : 0;
 124
 125         EBUG_ON(idx > trans->nr_sorted);
 126
 127         return idx < trans->nr_sorted
 128                 ? trans->paths + trans->sorted[idx]
 129                 : NULL;
 130 }
 131
 132 static inline struct btree_path *prev_btree_path(struct btree_trans *trans, struct btree_path *path)
 133 {
 134         unsigned idx = path ? path->sorted_idx : trans->nr_sorted;
 135
 136         return idx
 137                 ? trans->paths + trans->sorted[idx - 1]
 138                 : NULL;
 139 }
 140
 141 #define trans_for_each_path_inorder(_trans, _path, _i)                  \
 142         for (_i = 0;                                                    \
 143              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) < (_trans)->nr_sorted;\
 144              _i++)
 145
 146 #define trans_for_each_path_inorder_reverse(_trans, _path, _i)          \
 147         for (_i = trans->nr_sorted - 1;                                 \
 148              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) >= 0;\
 149              --_i)
 150
 151 static inline bool __path_has_node(const struct btree_path *path,
 152                                    const struct btree *b)
 153 {
 154         return path->l[b->c.level].b == b &&
 155                 btree_node_lock_seq_matches(path, b, b->c.level);
 156 }
 157
 158 static inline struct btree_path *
 159 __trans_next_path_with_node(struct btree_trans *trans, struct btree *b,
 160                             unsigned idx)
 161 {
 162         struct btree_path *path = __trans_next_path(trans, idx);
 163
 164         while (path && !__path_has_node(path, b))
 165                 path = __trans_next_path(trans, path->idx + 1);
 166
 167         return path;
 168 }
 169
 170 #define trans_for_each_path_with_node(_trans, _b, _path)                \
 171         for (_path = __trans_next_path_with_node((_trans), (_b), 0);    \
 172              (_path);                                                   \
 173              _path = __trans_next_path_with_node((_trans), (_b),        \
 174                                                  (_path)->idx + 1))
 175
 176 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *, struct btree_path *,
 177                          bool, unsigned long);
 178
 179 static inline struct btree_path * __must_check
 180 bch2_btree_path_make_mut(struct btree_trans *trans,
 181                          struct btree_path *path, bool intent,
 182                          unsigned long ip)
 183 {
 184         if (path->ref > 1 || path->preserve)
 185                 path = __bch2_btree_path_make_mut(trans, path, intent, ip);
 186         path->should_be_locked = false;
 187         return path;
 188 }
 189
 190 struct btree_path * __must_check
 191 __bch2_btree_path_set_pos(struct btree_trans *, struct btree_path *,
 192                         struct bpos, bool, unsigned long, int);
 193
 194 static inline struct btree_path * __must_check
 195 bch2_btree_path_set_pos(struct btree_trans *trans,
 196                    struct btree_path *path, struct bpos new_pos,
 197                    bool intent, unsigned long ip)
 198 {
 199         int cmp = bpos_cmp(new_pos, path->pos);
 200
 201         return cmp
 202                 ? __bch2_btree_path_set_pos(trans, path, new_pos, intent, ip, cmp)
 203                 : path;
 204 }
 205
 206 int __must_check bch2_btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 207                                               unsigned, unsigned long);
 208
 209 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 210                                           struct btree_path *path, unsigned flags)
 211 {
 212         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
 213                 return 0;
 214
 215         return bch2_btree_path_traverse_one(trans, path, flags, _RET_IP_);
 216 }
 217
 218 int __must_check bch2_btree_path_traverse(struct btree_trans *,
 219                                           struct btree_path *, unsigned);
 220 struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpos,
 221                                  unsigned, unsigned, unsigned, unsigned long);
 222 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 223
 224 /*
 225  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
 226  * different snapshot:
 227  */
 228 static inline struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
 229 {
 230         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
 231
 232         if (k.k && bpos_eq(path->pos, k.k->p))
 233                 return k;
 234
 235         bkey_init(u);
 236         u->p = path->pos;
 237         return (struct bkey_s_c) { u, NULL };
 238 }
 239
 240 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 241                                         struct btree_iter *, struct bpos);
 242
 243 void bch2_btree_path_level_init(struct btree_trans *, struct btree_path *, struct btree *);
 244
 245 int __bch2_trans_mutex_lock(struct btree_trans *, struct mutex *);
 246
 247 static inline int bch2_trans_mutex_lock(struct btree_trans *trans, struct mutex *lock)
 248 {
 249         return mutex_trylock(lock)
 250                 ? 0
 251                 : __bch2_trans_mutex_lock(trans, lock);
 252 }
 253
 254 #ifdef CONFIG_BCACHEFS_DEBUG
 255 void bch2_trans_verify_paths(struct btree_trans *);
 256 void bch2_assert_pos_locked(struct btree_trans *, enum btree_id,
 257                             struct bpos, bool);
 258 #else
 259 static inline void bch2_trans_verify_paths(struct btree_trans *trans) {}
 260 static inline void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 261                                           struct bpos pos, bool key_cache) {}
 262 #endif
 263
 264 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 265                                       struct btree *, struct bkey_packed *);
 266 void bch2_btree_node_iter_fix(struct btree_trans *trans, struct btree_path *,
 267                               struct btree *, struct btree_node_iter *,
 268                               struct bkey_packed *, unsigned, unsigned);
 269
 270 int bch2_btree_path_relock_intent(struct btree_trans *, struct btree_path *);
 271
 272 void bch2_path_put(struct btree_trans *, struct btree_path *, bool);
 273
 274 int bch2_trans_relock(struct btree_trans *);
 275 int bch2_trans_relock_notrace(struct btree_trans *);
 276 void bch2_trans_unlock(struct btree_trans *);
 277 bool bch2_trans_locked(struct btree_trans *);
 278
 279 static inline int trans_was_restarted(struct btree_trans *trans, u32 restart_count)
 280 {
 281         return restart_count != trans->restart_count
 282                 ? -BCH_ERR_transaction_restart_nested
 283                 : 0;
 284 }
 285
 286 void __noreturn bch2_trans_restart_error(struct btree_trans *, u32);
 287
 288 static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 289                                                    u32 restart_count)
 290 {
 291         if (trans_was_restarted(trans, restart_count))
 292                 bch2_trans_restart_error(trans, restart_count);
 293 }
 294
 295 void __noreturn bch2_trans_in_restart_error(struct btree_trans *);
 296
 297 static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
 298 {
 299         if (trans->restarted)
 300                 bch2_trans_in_restart_error(trans);
 301 }
 302
 303 __always_inline
 304 static int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
 305 {
 306         BUG_ON(err <= 0);
 307         BUG_ON(!bch2_err_matches(-err, BCH_ERR_transaction_restart));
 308
 309         trans->restarted = err;
 310         trans->last_restarted_ip = _THIS_IP_;
 311         return -err;
 312 }
 313
 314 __always_inline
 315 static int btree_trans_restart(struct btree_trans *trans, int err)
 316 {
 317         btree_trans_restart_nounlock(trans, err);
 318         return -err;
 319 }
 320
 321 bool bch2_btree_node_upgrade(struct btree_trans *,
 322                              struct btree_path *, unsigned);
 323
 324 void __bch2_btree_path_downgrade(struct btree_trans *, struct btree_path *, unsigned);
 325
 326 static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 327                                              struct btree_path *path)
 328 {
 329         unsigned new_locks_want = path->level + !!path->intent_ref;
 330
 331         if (path->locks_want > new_locks_want)
 332                 __bch2_btree_path_downgrade(trans, path, new_locks_want);
 333 }
 334
 335 void bch2_trans_downgrade(struct btree_trans *);
 336
 337 void bch2_trans_node_add(struct btree_trans *trans, struct btree *);
 338 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 339
 340 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
 341 int __must_check bch2_btree_iter_traverse(struct btree_iter *);
 342
 343 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 344 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 345 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 346
 347 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
 348 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 349
 350 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *);
 351
 352 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 353 {
 354         return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
 355 }
 356
 357 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 358 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 359
 360 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 361 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 362 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *);
 363
 364 bool bch2_btree_iter_advance(struct btree_iter *);
 365 bool bch2_btree_iter_rewind(struct btree_iter *);
 366
 367 static inline void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 368 {
 369         iter->k.type = KEY_TYPE_deleted;
 370         iter->k.p.inode         = iter->pos.inode       = new_pos.inode;
 371         iter->k.p.offset        = iter->pos.offset      = new_pos.offset;
 372         iter->k.p.snapshot      = iter->pos.snapshot    = new_pos.snapshot;
 373         iter->k.size = 0;
 374 }
 375
 376 static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 377 {
 378         if (unlikely(iter->update_path))
 379                 bch2_path_put(iter->trans, iter->update_path,
 380                               iter->flags & BTREE_ITER_INTENT);
 381         iter->update_path = NULL;
 382
 383         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
 384                 new_pos.snapshot = iter->snapshot;
 385
 386         __bch2_btree_iter_set_pos(iter, new_pos);
 387 }
 388
 389 static inline void bch2_btree_iter_set_pos_to_extent_start(struct btree_iter *iter)
 390 {
 391         BUG_ON(!(iter->flags & BTREE_ITER_IS_EXTENTS));
 392         iter->pos = bkey_start_pos(&iter->k);
 393 }
 394
 395 static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 snapshot)
 396 {
 397         struct bpos pos = iter->pos;
 398
 399         iter->snapshot = snapshot;
 400         pos.snapshot = snapshot;
 401         bch2_btree_iter_set_pos(iter, pos);
 402 }
 403
 404 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 405
 406 static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 407                                                unsigned btree_id,
 408                                                unsigned flags)
 409 {
 410         if (flags & BTREE_ITER_ALL_LEVELS)
 411                 flags |= BTREE_ITER_ALL_SNAPSHOTS|__BTREE_ITER_ALL_SNAPSHOTS;
 412
 413         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
 414             btree_node_type_is_extents(btree_id))
 415                 flags |= BTREE_ITER_IS_EXTENTS;
 416
 417         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 418             !btree_type_has_snapshots(btree_id))
 419                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 420
 421         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 422             btree_type_has_snapshots(btree_id))
 423                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 424
 425         if (trans->journal_replay_not_finished)
 426                 flags |= BTREE_ITER_WITH_JOURNAL;
 427
 428         return flags;
 429 }
 430
 431 static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
 432                                              unsigned btree_id,
 433                                              unsigned flags)
 434 {
 435         if (!btree_id_cached(trans->c, btree_id)) {
 436                 flags &= ~BTREE_ITER_CACHED;
 437                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 438         } else if (!(flags & BTREE_ITER_CACHED))
 439                 flags |= BTREE_ITER_WITH_KEY_CACHE;
 440
 441         return __bch2_btree_iter_flags(trans, btree_id, flags);
 442 }
 443
 444 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 445                                           struct btree_iter *iter,
 446                                           unsigned btree_id, struct bpos pos,
 447                                           unsigned locks_want,
 448                                           unsigned depth,
 449                                           unsigned flags,
 450                                           unsigned long ip)
 451 {
 452         memset(iter, 0, sizeof(*iter));
 453         iter->trans     = trans;
 454         iter->btree_id  = btree_id;
 455         iter->flags     = flags;
 456         iter->snapshot  = pos.snapshot;
 457         iter->pos       = pos;
 458         iter->k.p       = pos;
 459
 460 #ifdef CONFIG_BCACHEFS_DEBUG
 461         iter->ip_allocated = ip;
 462 #endif
 463         iter->path = bch2_path_get(trans, btree_id, iter->pos,
 464                                    locks_want, depth, flags, ip);
 465 }
 466
 467 void bch2_trans_iter_init_outlined(struct btree_trans *, struct btree_iter *,
 468                           enum btree_id, struct bpos, unsigned);
 469
 470 static inline void bch2_trans_iter_init(struct btree_trans *trans,
 471                           struct btree_iter *iter,
 472                           unsigned btree_id, struct bpos pos,
 473                           unsigned flags)
 474 {
 475         if (__builtin_constant_p(btree_id) &&
 476             __builtin_constant_p(flags))
 477                 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
 478                                 bch2_btree_iter_flags(trans, btree_id, flags),
 479                                 _THIS_IP_);
 480         else
 481                 bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
 482 }
 483
 484 void bch2_trans_node_iter_init(struct btree_trans *, struct btree_iter *,
 485                                enum btree_id, struct bpos,
 486                                unsigned, unsigned, unsigned);
 487 void bch2_trans_copy_iter(struct btree_iter *, struct btree_iter *);
 488
 489 static inline void set_btree_iter_dontneed(struct btree_iter *iter)
 490 {
 491         if (!iter->trans->restarted)
 492                 iter->path->preserve = false;
 493 }
 494
 495 void *__bch2_trans_kmalloc(struct btree_trans *, size_t);
 496
 497 static inline void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
 498 {
 499         size = roundup(size, 8);
 500
 501         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 502                 void *p = trans->mem + trans->mem_top;
 503
 504                 trans->mem_top += size;
 505                 memset(p, 0, size);
 506                 return p;
 507         } else {
 508                 return __bch2_trans_kmalloc(trans, size);
 509         }
 510 }
 511
 512 static inline void *bch2_trans_kmalloc_nomemzero(struct btree_trans *trans, size_t size)
 513 {
 514         size = roundup(size, 8);
 515
 516         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 517                 void *p = trans->mem + trans->mem_top;
 518
 519                 trans->mem_top += size;
 520                 return p;
 521         } else {
 522                 return __bch2_trans_kmalloc(trans, size);
 523         }
 524 }
 525
 526 static inline struct bkey_s_c __bch2_bkey_get_iter(struct btree_trans *trans,
 527                                 struct btree_iter *iter,
 528                                 unsigned btree_id, struct bpos pos,
 529                                 unsigned flags, unsigned type)
 530 {
 531         struct bkey_s_c k;
 532
 533         bch2_trans_iter_init(trans, iter, btree_id, pos, flags);
 534         k = bch2_btree_iter_peek_slot(iter);
 535
 536         if (!bkey_err(k) && type && k.k->type != type)
 537                 k = bkey_s_c_err(-BCH_ERR_ENOENT_bkey_type_mismatch);
 538         if (unlikely(bkey_err(k)))
 539                 bch2_trans_iter_exit(trans, iter);
 540         return k;
 541 }
 542
 543 static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 544                                 struct btree_iter *iter,
 545                                 unsigned btree_id, struct bpos pos,
 546                                 unsigned flags)
 547 {
 548         return __bch2_bkey_get_iter(trans, iter, btree_id, pos, flags, 0);
 549 }
 550
 551 #define bch2_bkey_get_iter_typed(_trans, _iter, _btree_id, _pos, _flags, _type)\
 552         bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,                 \
 553                                        _btree_id, _pos, _flags, KEY_TYPE_##_type))
 554
 555 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 556                                 unsigned btree_id, struct bpos pos,
 557                                 unsigned flags, unsigned type,
 558                                 unsigned val_size, void *val)
 559 {
 560         struct btree_iter iter;
 561         struct bkey_s_c k;
 562         int ret;
 563
 564         k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
 565         ret = bkey_err(k);
 566         if (!ret) {
 567                 unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
 568
 569                 memcpy(val, k.v, b);
 570                 if (unlikely(b < sizeof(*val)))
 571                         memset((void *) val + b, 0, sizeof(*val) - b);
 572                 bch2_trans_iter_exit(trans, &iter);
 573         }
 574
 575         return ret;
 576 }
 577
 578 #define bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags, _type, _val)\
 579         __bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags,      \
 580                                   KEY_TYPE_##_type, sizeof(*_val), _val)
 581
 582 u32 bch2_trans_begin(struct btree_trans *);
 583
 584 /*
 585  * XXX
 586  * this does not handle transaction restarts from bch2_btree_iter_next_node()
 587  * correctly
 588  */
 589 #define __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 590                               _locks_want, _depth, _flags, _b, _ret)    \
 591         for (bch2_trans_node_iter_init((_trans), &(_iter), (_btree_id), \
 592                                 _start, _locks_want, _depth, _flags);   \
 593              (_b) = bch2_btree_iter_peek_node_and_restart(&(_iter)),    \
 594              !((_ret) = PTR_ERR_OR_ZERO(_b)) && (_b);                   \
 595              (_b) = bch2_btree_iter_next_node(&(_iter)))
 596
 597 #define for_each_btree_node(_trans, _iter, _btree_id, _start,           \
 598                             _flags, _b, _ret)                           \
 599         __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 600                               0, 0, _flags, _b, _ret)
 601
 602 static inline struct bkey_s_c bch2_btree_iter_peek_prev_type(struct btree_iter *iter,
 603                                                              unsigned flags)
 604 {
 605         BUG_ON(flags & BTREE_ITER_ALL_LEVELS);
 606
 607         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 608                                                 bch2_btree_iter_peek_prev(iter);
 609 }
 610
 611 static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 612                                                         unsigned flags)
 613 {
 614         return  flags & BTREE_ITER_ALL_LEVELS ? bch2_btree_iter_peek_all_levels(iter) :
 615                 flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 616                                                 bch2_btree_iter_peek(iter);
 617 }
 618
 619 static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
 620                                                              struct bpos end,
 621                                                              unsigned flags)
 622 {
 623         if (!(flags & BTREE_ITER_SLOTS))
 624                 return bch2_btree_iter_peek_upto(iter, end);
 625
 626         if (bkey_gt(iter->pos, end))
 627                 return bkey_s_c_null;
 628
 629         return bch2_btree_iter_peek_slot(iter);
 630 }
 631
 632 static inline int btree_trans_too_many_iters(struct btree_trans *trans)
 633 {
 634         if (hweight64(trans->paths_allocated) > BTREE_ITER_MAX - 8) {
 635                 trace_and_count(trans->c, trans_restart_too_many_iters, trans, _THIS_IP_);
 636                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
 637         }
 638
 639         return 0;
 640 }
 641
 642 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 643
 644 static inline struct bkey_s_c
 645 __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 646                                    struct btree_iter *iter, unsigned flags)
 647 {
 648         struct bkey_s_c k;
 649
 650         while (btree_trans_too_many_iters(trans) ||
 651                (k = bch2_btree_iter_peek_type(iter, flags),
 652                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 653                 bch2_trans_begin(trans);
 654
 655         return k;
 656 }
 657
 658 static inline struct bkey_s_c
 659 __bch2_btree_iter_peek_upto_and_restart(struct btree_trans *trans,
 660                                         struct btree_iter *iter,
 661                                         struct bpos end,
 662                                         unsigned flags)
 663 {
 664         struct bkey_s_c k;
 665
 666         while (btree_trans_too_many_iters(trans) ||
 667                (k = bch2_btree_iter_peek_upto_type(iter, end, flags),
 668                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 669                 bch2_trans_begin(trans);
 670
 671         return k;
 672 }
 673
 674 #define lockrestart_do(_trans, _do)                                     \
 675 ({                                                                      \
 676         u32 _restart_count;                                             \
 677         int _ret2;                                                      \
 678                                                                         \
 679         do {                                                            \
 680                 _restart_count = bch2_trans_begin(_trans);              \
 681                 _ret2 = (_do);                                          \
 682         } while (bch2_err_matches(_ret2, BCH_ERR_transaction_restart)); \
 683                                                                         \
 684         if (!_ret2)                                                     \
 685                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 686                                                                         \
 687         _ret2;                                                          \
 688 })
 689
 690 /*
 691  * nested_lockrestart_do(), nested_commit_do():
 692  *
 693  * These are like lockrestart_do() and commit_do(), with two differences:
 694  *
 695  *  - We don't call bch2_trans_begin() unless we had a transaction restart
 696  *  - We return -BCH_ERR_transaction_restart_nested if we succeeded after a
 697  *  transaction restart
 698  */
 699 #define nested_lockrestart_do(_trans, _do)                              \
 700 ({                                                                      \
 701         u32 _restart_count, _orig_restart_count;                        \
 702         int _ret2;                                                      \
 703                                                                         \
 704         _restart_count = _orig_restart_count = (_trans)->restart_count; \
 705                                                                         \
 706         while (bch2_err_matches(_ret2 = (_do), BCH_ERR_transaction_restart))\
 707                 _restart_count = bch2_trans_begin(_trans);              \
 708                                                                         \
 709         if (!_ret2)                                                     \
 710                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 711                                                                         \
 712         _ret2 ?: trans_was_restarted(_trans, _restart_count);           \
 713 })
 714
 715 #define for_each_btree_key2(_trans, _iter, _btree_id,                   \
 716                             _start, _flags, _k, _do)                    \
 717 ({                                                                      \
 718         int _ret3 = 0;                                                  \
 719                                                                         \
 720         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 721                              (_start), (_flags));                       \
 722                                                                         \
 723         while (1) {                                                     \
 724                 u32 _restart_count = bch2_trans_begin(_trans);          \
 725                                                                         \
 726                 _ret3 = 0;                                              \
 727                 (_k) = bch2_btree_iter_peek_type(&(_iter), (_flags));   \
 728                 if (!(_k).k)                                            \
 729                         break;                                          \
 730                                                                         \
 731                 _ret3 = bkey_err(_k) ?: (_do);                          \
 732                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 733                         continue;                                       \
 734                 if (_ret3)                                              \
 735                         break;                                          \
 736                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 737                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 738                         break;                                          \
 739         }                                                               \
 740                                                                         \
 741         bch2_trans_iter_exit((_trans), &(_iter));                       \
 742         _ret3;                                                          \
 743 })
 744
 745 #define for_each_btree_key2_upto(_trans, _iter, _btree_id,              \
 746                             _start, _end, _flags, _k, _do)              \
 747 ({                                                                      \
 748         int _ret3 = 0;                                                  \
 749                                                                         \
 750         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 751                              (_start), (_flags));                       \
 752                                                                         \
 753         while (1) {                                                     \
 754                 u32 _restart_count = bch2_trans_begin(_trans);          \
 755                                                                         \
 756                 _ret3 = 0;                                              \
 757                 (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, (_flags));\
 758                 if (!(_k).k)                                            \
 759                         break;                                          \
 760                                                                         \
 761                 _ret3 = bkey_err(_k) ?: (_do);                          \
 762                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 763                         continue;                                       \
 764                 if (_ret3)                                              \
 765                         break;                                          \
 766                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 767                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 768                         break;                                          \
 769         }                                                               \
 770                                                                         \
 771         bch2_trans_iter_exit((_trans), &(_iter));                       \
 772         _ret3;                                                          \
 773 })
 774
 775 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,            \
 776                                    _start, _flags, _k, _do)             \
 777 ({                                                                      \
 778         int _ret3 = 0;                                                  \
 779                                                                         \
 780         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 781                              (_start), (_flags));                       \
 782                                                                         \
 783         while (1) {                                                     \
 784                 u32 _restart_count = bch2_trans_begin(_trans);          \
 785                 (_k) = bch2_btree_iter_peek_prev_type(&(_iter), (_flags));\
 786                 if (!(_k).k) {                                          \
 787                         _ret3 = 0;                                      \
 788                         break;                                          \
 789                 }                                                       \
 790                                                                         \
 791                 _ret3 = bkey_err(_k) ?: (_do);                          \
 792                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 793                         continue;                                       \
 794                 if (_ret3)                                              \
 795                         break;                                          \
 796                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 797                 if (!bch2_btree_iter_rewind(&(_iter)))                  \
 798                         break;                                          \
 799         }                                                               \
 800                                                                         \
 801         bch2_trans_iter_exit((_trans), &(_iter));                       \
 802         _ret3;                                                          \
 803 })
 804
 805 #define for_each_btree_key_commit(_trans, _iter, _btree_id,             \
 806                                   _start, _iter_flags, _k,              \
 807                                   _disk_res, _journal_seq, _commit_flags,\
 808                                   _do)                                  \
 809         for_each_btree_key2(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 810                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 811                                         (_journal_seq), (_commit_flags)))
 812
 813 #define for_each_btree_key_reverse_commit(_trans, _iter, _btree_id,     \
 814                                   _start, _iter_flags, _k,              \
 815                                   _disk_res, _journal_seq, _commit_flags,\
 816                                   _do)                                  \
 817         for_each_btree_key_reverse(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 818                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 819                                         (_journal_seq), (_commit_flags)))
 820
 821 #define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,        \
 822                                   _start, _end, _iter_flags, _k,        \
 823                                   _disk_res, _journal_seq, _commit_flags,\
 824                                   _do)                                  \
 825         for_each_btree_key2_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 826                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 827                                         (_journal_seq), (_commit_flags)))
 828
 829 #define for_each_btree_key(_trans, _iter, _btree_id,                    \
 830                            _start, _flags, _k, _ret)                    \
 831         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 832                                   (_start), (_flags));                  \
 833              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 834              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 835              bch2_btree_iter_advance(&(_iter)))
 836
 837 #define for_each_btree_key_upto(_trans, _iter, _btree_id,               \
 838                                 _start, _end, _flags, _k, _ret)         \
 839         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 840                                   (_start), (_flags));                  \
 841              (_k) = __bch2_btree_iter_peek_upto_and_restart((_trans),   \
 842                                                 &(_iter), _end, _flags),\
 843              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 844              bch2_btree_iter_advance(&(_iter)))
 845
 846 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,          \
 847                            _start, _flags, _k, _ret)                    \
 848         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 849                                   (_start), (_flags));                  \
 850              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 851              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 852              bch2_btree_iter_advance(&(_iter)))
 853
 854 #define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,     \
 855                            _start, _end, _flags, _k, _ret)              \
 856         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 857                                   (_start), (_flags));                  \
 858              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
 859              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 860              bch2_btree_iter_advance(&(_iter)))
 861
 862 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _ret)    \
 863         for (;                                                          \
 864              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 865              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 866              bch2_btree_iter_advance(&(_iter)))
 867
 868 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)  \
 869         for (;                                                          \
 870              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 871              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 872              bch2_btree_iter_advance(&(_iter)))
 873
 874 #define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
 875         for (;                                                                  \
 876              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),     \
 877              !((_ret) = bkey_err(_k)) && (_k).k;                                \
 878              bch2_btree_iter_advance(&(_iter)))
 879
 880 #define drop_locks_do(_trans, _do)                                      \
 881 ({                                                                      \
 882         bch2_trans_unlock(_trans);                                      \
 883         _do ?: bch2_trans_relock(_trans);                               \
 884 })
 885
 886 #define allocate_dropping_locks_errcode(_trans, _do)                    \
 887 ({                                                                      \
 888         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 889         int _ret = _do;                                                 \
 890                                                                         \
 891         if (bch2_err_matches(_ret, ENOMEM)) {                           \
 892                 _gfp = GFP_KERNEL;                                      \
 893                 _ret = drop_locks_do(trans, _do);                       \
 894         }                                                               \
 895         _ret;                                                           \
 896 })
 897
 898 #define allocate_dropping_locks(_trans, _ret, _do)                      \
 899 ({                                                                      \
 900         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 901         typeof(_do) _p = _do;                                           \
 902                                                                         \
 903         _ret = 0;                                                       \
 904         if (unlikely(!_p)) {                                            \
 905                 _gfp = GFP_KERNEL;                                      \
 906                 _ret = drop_locks_do(trans, ((_p = _do), 0));           \
 907         }                                                               \
 908         _p;                                                             \
 909 })
 910
 911 /* new multiple iterator interface: */
 912
 913 void bch2_trans_updates_to_text(struct printbuf *, struct btree_trans *);
 914 void bch2_btree_path_to_text(struct printbuf *, struct btree_path *);
 915 void bch2_trans_paths_to_text(struct printbuf *, struct btree_trans *);
 916 void bch2_dump_trans_updates(struct btree_trans *);
 917 void bch2_dump_trans_paths_updates(struct btree_trans *);
 918
 919 struct btree_trans *__bch2_trans_get(struct bch_fs *, unsigned);
 920 void bch2_trans_put(struct btree_trans *);
 921
 922 extern const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
 923 unsigned bch2_trans_get_fn_idx(const char *);
 924
 925 #define bch2_trans_get(_c)                                              \
 926 ({                                                                      \
 927         static unsigned trans_fn_idx;                                   \
 928                                                                         \
 929         if (unlikely(!trans_fn_idx))                                    \
 930                 trans_fn_idx = bch2_trans_get_fn_idx(__func__);         \
 931         __bch2_trans_get(_c, trans_fn_idx);                             \
 932 })
 933
 934 void bch2_btree_trans_to_text(struct printbuf *, struct btree_trans *);
 935
 936 void bch2_fs_btree_iter_exit(struct bch_fs *);
 937 int bch2_fs_btree_iter_init(struct bch_fs *);
 938
 939 #endif /* _BCACHEFS_BTREE_ITER_H */