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(_trans, _path, _idx)                   \
 114         for (_idx = 0;                                                  \
 115              (_path = __trans_next_path_safe((_trans), &_idx));         \
 116              _idx++)
 117
 118 static inline struct btree_path *next_btree_path(struct btree_trans *trans, struct btree_path *path)
 119 {
 120         unsigned idx = path ? path->sorted_idx + 1 : 0;
 121
 122         EBUG_ON(idx > trans->nr_sorted);
 123
 124         return idx < trans->nr_sorted
 125                 ? trans->paths + trans->sorted[idx]
 126                 : NULL;
 127 }
 128
 129 static inline struct btree_path *prev_btree_path(struct btree_trans *trans, struct btree_path *path)
 130 {
 131         unsigned idx = path ? path->sorted_idx : trans->nr_sorted;
 132
 133         return idx
 134                 ? trans->paths + trans->sorted[idx - 1]
 135                 : NULL;
 136 }
 137
 138 #define trans_for_each_path_inorder(_trans, _path, _i)                  \
 139         for (_i = 0;                                                    \
 140              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) < (_trans)->nr_sorted;\
 141              _i++)
 142
 143 #define trans_for_each_path_inorder_reverse(_trans, _path, _i)          \
 144         for (_i = trans->nr_sorted - 1;                                 \
 145              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) >= 0;\
 146              --_i)
 147
 148 static inline bool __path_has_node(const struct btree_path *path,
 149                                    const struct btree *b)
 150 {
 151         return path->l[b->c.level].b == b &&
 152                 btree_node_lock_seq_matches(path, b, b->c.level);
 153 }
 154
 155 static inline struct btree_path *
 156 __trans_next_path_with_node(struct btree_trans *trans, struct btree *b,
 157                             unsigned idx)
 158 {
 159         struct btree_path *path = __trans_next_path(trans, idx);
 160
 161         while (path && !__path_has_node(path, b))
 162                 path = __trans_next_path(trans, path->idx + 1);
 163
 164         return path;
 165 }
 166
 167 #define trans_for_each_path_with_node(_trans, _b, _path)                \
 168         for (_path = __trans_next_path_with_node((_trans), (_b), 0);    \
 169              (_path);                                                   \
 170              _path = __trans_next_path_with_node((_trans), (_b),        \
 171                                                  (_path)->idx + 1))
 172
 173 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *, struct btree_path *,
 174                          bool, unsigned long);
 175
 176 static inline struct btree_path * __must_check
 177 bch2_btree_path_make_mut(struct btree_trans *trans,
 178                          struct btree_path *path, bool intent,
 179                          unsigned long ip)
 180 {
 181         if (path->ref > 1 || path->preserve)
 182                 path = __bch2_btree_path_make_mut(trans, path, intent, ip);
 183         path->should_be_locked = false;
 184         return path;
 185 }
 186
 187 struct btree_path * __must_check
 188 __bch2_btree_path_set_pos(struct btree_trans *, struct btree_path *,
 189                         struct bpos, bool, unsigned long, int);
 190
 191 static inline struct btree_path * __must_check
 192 bch2_btree_path_set_pos(struct btree_trans *trans,
 193                    struct btree_path *path, struct bpos new_pos,
 194                    bool intent, unsigned long ip)
 195 {
 196         int cmp = bpos_cmp(new_pos, path->pos);
 197
 198         return cmp
 199                 ? __bch2_btree_path_set_pos(trans, path, new_pos, intent, ip, cmp)
 200                 : path;
 201 }
 202
 203 int __must_check bch2_btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 204                                               unsigned, unsigned long);
 205
 206 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 207                                           struct btree_path *path, unsigned flags)
 208 {
 209         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
 210                 return 0;
 211
 212         return bch2_btree_path_traverse_one(trans, path, flags, _RET_IP_);
 213 }
 214
 215 int __must_check bch2_btree_path_traverse(struct btree_trans *,
 216                                           struct btree_path *, unsigned);
 217 struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpos,
 218                                  unsigned, unsigned, unsigned, unsigned long);
 219 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 220
 221 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 222                                         struct btree_iter *, struct bpos);
 223
 224 void bch2_btree_path_level_init(struct btree_trans *, struct btree_path *, struct btree *);
 225
 226 int __bch2_trans_mutex_lock(struct btree_trans *, struct mutex *);
 227
 228 static inline int bch2_trans_mutex_lock(struct btree_trans *trans, struct mutex *lock)
 229 {
 230         return mutex_trylock(lock)
 231                 ? 0
 232                 : __bch2_trans_mutex_lock(trans, lock);
 233 }
 234
 235 #ifdef CONFIG_BCACHEFS_DEBUG
 236 void bch2_trans_verify_paths(struct btree_trans *);
 237 void bch2_assert_pos_locked(struct btree_trans *, enum btree_id,
 238                             struct bpos, bool);
 239 #else
 240 static inline void bch2_trans_verify_paths(struct btree_trans *trans) {}
 241 static inline void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 242                                           struct bpos pos, bool key_cache) {}
 243 #endif
 244
 245 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 246                                       struct btree *, struct bkey_packed *);
 247 void bch2_btree_node_iter_fix(struct btree_trans *trans, struct btree_path *,
 248                               struct btree *, struct btree_node_iter *,
 249                               struct bkey_packed *, unsigned, unsigned);
 250
 251 int bch2_btree_path_relock_intent(struct btree_trans *, struct btree_path *);
 252
 253 void bch2_path_put(struct btree_trans *, struct btree_path *, bool);
 254
 255 int bch2_trans_relock(struct btree_trans *);
 256 int bch2_trans_relock_notrace(struct btree_trans *);
 257 void bch2_trans_unlock(struct btree_trans *);
 258 bool bch2_trans_locked(struct btree_trans *);
 259
 260 static inline bool trans_was_restarted(struct btree_trans *trans, u32 restart_count)
 261 {
 262         return restart_count != trans->restart_count;
 263 }
 264
 265 void bch2_trans_restart_error(struct btree_trans *, u32);
 266
 267 static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 268                                                    u32 restart_count)
 269 {
 270         if (trans_was_restarted(trans, restart_count))
 271                 bch2_trans_restart_error(trans, restart_count);
 272 }
 273
 274 void bch2_trans_in_restart_error(struct btree_trans *);
 275
 276 static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
 277 {
 278         if (trans->restarted)
 279                 bch2_trans_in_restart_error(trans);
 280 }
 281
 282 __always_inline
 283 static inline int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
 284 {
 285         BUG_ON(err <= 0);
 286         BUG_ON(!bch2_err_matches(err, BCH_ERR_transaction_restart));
 287
 288         trans->restarted = err;
 289         trans->last_restarted_ip = _THIS_IP_;
 290         return -err;
 291 }
 292
 293 __always_inline
 294 static inline int btree_trans_restart(struct btree_trans *trans, int err)
 295 {
 296         btree_trans_restart_nounlock(trans, err);
 297         return -err;
 298 }
 299
 300 bool bch2_btree_node_upgrade(struct btree_trans *,
 301                              struct btree_path *, unsigned);
 302
 303 void __bch2_btree_path_downgrade(struct btree_trans *, struct btree_path *, unsigned);
 304
 305 static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 306                                              struct btree_path *path)
 307 {
 308         unsigned new_locks_want = path->level + !!path->intent_ref;
 309
 310         if (path->locks_want > new_locks_want)
 311                 __bch2_btree_path_downgrade(trans, path, new_locks_want);
 312 }
 313
 314 void bch2_trans_downgrade(struct btree_trans *);
 315
 316 void bch2_trans_node_add(struct btree_trans *trans, struct btree *);
 317 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 318
 319 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
 320 int __must_check bch2_btree_iter_traverse(struct btree_iter *);
 321
 322 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 323 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 324 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 325
 326 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
 327 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 328
 329 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *);
 330
 331 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 332 {
 333         return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
 334 }
 335
 336 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 337 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 338
 339 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 340 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 341 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *);
 342
 343 bool bch2_btree_iter_advance(struct btree_iter *);
 344 bool bch2_btree_iter_rewind(struct btree_iter *);
 345
 346 static inline void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 347 {
 348         iter->k.type = KEY_TYPE_deleted;
 349         iter->k.p.inode         = iter->pos.inode       = new_pos.inode;
 350         iter->k.p.offset        = iter->pos.offset      = new_pos.offset;
 351         iter->k.p.snapshot      = iter->pos.snapshot    = new_pos.snapshot;
 352         iter->k.size = 0;
 353 }
 354
 355 static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 356 {
 357         if (unlikely(iter->update_path))
 358                 bch2_path_put(iter->trans, iter->update_path,
 359                               iter->flags & BTREE_ITER_INTENT);
 360         iter->update_path = NULL;
 361
 362         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
 363                 new_pos.snapshot = iter->snapshot;
 364
 365         __bch2_btree_iter_set_pos(iter, new_pos);
 366 }
 367
 368 static inline void bch2_btree_iter_set_pos_to_extent_start(struct btree_iter *iter)
 369 {
 370         BUG_ON(!(iter->flags & BTREE_ITER_IS_EXTENTS));
 371         iter->pos = bkey_start_pos(&iter->k);
 372 }
 373
 374 static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 snapshot)
 375 {
 376         struct bpos pos = iter->pos;
 377
 378         iter->snapshot = snapshot;
 379         pos.snapshot = snapshot;
 380         bch2_btree_iter_set_pos(iter, pos);
 381 }
 382
 383 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 384
 385 static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 386                                                unsigned btree_id,
 387                                                unsigned flags)
 388 {
 389         if (flags & BTREE_ITER_ALL_LEVELS)
 390                 flags |= BTREE_ITER_ALL_SNAPSHOTS|__BTREE_ITER_ALL_SNAPSHOTS;
 391
 392         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
 393             btree_node_type_is_extents(btree_id))
 394                 flags |= BTREE_ITER_IS_EXTENTS;
 395
 396         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 397             !btree_type_has_snapshots(btree_id))
 398                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 399
 400         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 401             btree_type_has_snapshots(btree_id))
 402                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 403
 404         if (trans->journal_replay_not_finished)
 405                 flags |= BTREE_ITER_WITH_JOURNAL;
 406
 407         return flags;
 408 }
 409
 410 static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
 411                                              unsigned btree_id,
 412                                              unsigned flags)
 413 {
 414         if (!btree_id_cached(trans->c, btree_id)) {
 415                 flags &= ~BTREE_ITER_CACHED;
 416                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 417         } else if (!(flags & BTREE_ITER_CACHED))
 418                 flags |= BTREE_ITER_WITH_KEY_CACHE;
 419
 420         return __bch2_btree_iter_flags(trans, btree_id, flags);
 421 }
 422
 423 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 424                                           struct btree_iter *iter,
 425                                           unsigned btree_id, struct bpos pos,
 426                                           unsigned locks_want,
 427                                           unsigned depth,
 428                                           unsigned flags,
 429                                           unsigned long ip)
 430 {
 431         memset(iter, 0, sizeof(*iter));
 432         iter->trans     = trans;
 433         iter->btree_id  = btree_id;
 434         iter->flags     = flags;
 435         iter->snapshot  = pos.snapshot;
 436         iter->pos       = pos;
 437         iter->k.p       = pos;
 438
 439 #ifdef CONFIG_BCACHEFS_DEBUG
 440         iter->ip_allocated = ip;
 441 #endif
 442         iter->path = bch2_path_get(trans, btree_id, iter->pos,
 443                                    locks_want, depth, flags, ip);
 444 }
 445
 446 void bch2_trans_iter_init_outlined(struct btree_trans *, struct btree_iter *,
 447                           unsigned, struct bpos, unsigned);
 448
 449 static inline void bch2_trans_iter_init(struct btree_trans *trans,
 450                           struct btree_iter *iter,
 451                           unsigned btree_id, struct bpos pos,
 452                           unsigned flags)
 453 {
 454         if (__builtin_constant_p(btree_id) &&
 455             __builtin_constant_p(flags))
 456                 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
 457                                 bch2_btree_iter_flags(trans, btree_id, flags),
 458                                 _THIS_IP_);
 459         else
 460                 bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
 461 }
 462
 463 void bch2_trans_node_iter_init(struct btree_trans *, struct btree_iter *,
 464                                enum btree_id, struct bpos,
 465                                unsigned, unsigned, unsigned);
 466 void bch2_trans_copy_iter(struct btree_iter *, struct btree_iter *);
 467
 468 static inline void set_btree_iter_dontneed(struct btree_iter *iter)
 469 {
 470         if (!iter->trans->restarted)
 471                 iter->path->preserve = false;
 472 }
 473
 474 void *__bch2_trans_kmalloc(struct btree_trans *, size_t);
 475
 476 static inline void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
 477 {
 478         size = roundup(size, 8);
 479
 480         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 481                 void *p = trans->mem + trans->mem_top;
 482
 483                 trans->mem_top += size;
 484                 memset(p, 0, size);
 485                 return p;
 486         } else {
 487                 return __bch2_trans_kmalloc(trans, size);
 488         }
 489 }
 490
 491 static inline void *bch2_trans_kmalloc_nomemzero(struct btree_trans *trans, size_t size)
 492 {
 493         size = roundup(size, 8);
 494
 495         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 496                 void *p = trans->mem + trans->mem_top;
 497
 498                 trans->mem_top += size;
 499                 return p;
 500         } else {
 501                 return __bch2_trans_kmalloc(trans, size);
 502         }
 503 }
 504
 505 static inline struct bkey_s_c __bch2_bkey_get_iter(struct btree_trans *trans,
 506                                 struct btree_iter *iter,
 507                                 unsigned btree_id, struct bpos pos,
 508                                 unsigned flags, unsigned type)
 509 {
 510         struct bkey_s_c k;
 511
 512         bch2_trans_iter_init(trans, iter, btree_id, pos, flags);
 513         k = bch2_btree_iter_peek_slot(iter);
 514
 515         if (!bkey_err(k) && type && k.k->type != type)
 516                 k = bkey_s_c_err(-BCH_ERR_ENOENT_bkey_type_mismatch);
 517         if (unlikely(bkey_err(k)))
 518                 bch2_trans_iter_exit(trans, iter);
 519         return k;
 520 }
 521
 522 static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 523                                 struct btree_iter *iter,
 524                                 unsigned btree_id, struct bpos pos,
 525                                 unsigned flags)
 526 {
 527         return __bch2_bkey_get_iter(trans, iter, btree_id, pos, flags, 0);
 528 }
 529
 530 #define bch2_bkey_get_iter_typed(_trans, _iter, _btree_id, _pos, _flags, _type)\
 531         bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,                 \
 532                                        _btree_id, _pos, _flags, KEY_TYPE_##_type))
 533
 534 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 535                                 unsigned btree_id, struct bpos pos,
 536                                 unsigned flags, unsigned type,
 537                                 unsigned val_size, void *val)
 538 {
 539         struct btree_iter iter;
 540         struct bkey_s_c k;
 541         int ret;
 542
 543         k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
 544         ret = bkey_err(k);
 545         if (!ret) {
 546                 unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
 547
 548                 memcpy(val, k.v, b);
 549                 if (unlikely(b < sizeof(*val)))
 550                         memset((void *) val + b, 0, sizeof(*val) - b);
 551                 bch2_trans_iter_exit(trans, &iter);
 552         }
 553
 554         return ret;
 555 }
 556
 557 #define bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags, _type, _val)\
 558         __bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags,      \
 559                                   KEY_TYPE_##_type, sizeof(*_val), _val)
 560
 561 u32 bch2_trans_begin(struct btree_trans *);
 562
 563 /*
 564  * XXX
 565  * this does not handle transaction restarts from bch2_btree_iter_next_node()
 566  * correctly
 567  */
 568 #define __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 569                               _locks_want, _depth, _flags, _b, _ret)    \
 570         for (bch2_trans_node_iter_init((_trans), &(_iter), (_btree_id), \
 571                                 _start, _locks_want, _depth, _flags);   \
 572              (_b) = bch2_btree_iter_peek_node_and_restart(&(_iter)),    \
 573              !((_ret) = PTR_ERR_OR_ZERO(_b)) && (_b);                   \
 574              (_b) = bch2_btree_iter_next_node(&(_iter)))
 575
 576 #define for_each_btree_node(_trans, _iter, _btree_id, _start,           \
 577                             _flags, _b, _ret)                           \
 578         __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 579                               0, 0, _flags, _b, _ret)
 580
 581 static inline struct bkey_s_c bch2_btree_iter_peek_prev_type(struct btree_iter *iter,
 582                                                              unsigned flags)
 583 {
 584         BUG_ON(flags & BTREE_ITER_ALL_LEVELS);
 585
 586         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 587                                                 bch2_btree_iter_peek_prev(iter);
 588 }
 589
 590 static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 591                                                         unsigned flags)
 592 {
 593         return  flags & BTREE_ITER_ALL_LEVELS ? bch2_btree_iter_peek_all_levels(iter) :
 594                 flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 595                                                 bch2_btree_iter_peek(iter);
 596 }
 597
 598 static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
 599                                                              struct bpos end,
 600                                                              unsigned flags)
 601 {
 602         if (!(flags & BTREE_ITER_SLOTS))
 603                 return bch2_btree_iter_peek_upto(iter, end);
 604
 605         if (bkey_gt(iter->pos, end))
 606                 return bkey_s_c_null;
 607
 608         return bch2_btree_iter_peek_slot(iter);
 609 }
 610
 611 static inline int btree_trans_too_many_iters(struct btree_trans *trans)
 612 {
 613         if (hweight64(trans->paths_allocated) > BTREE_ITER_MAX - 8) {
 614                 trace_and_count(trans->c, trans_restart_too_many_iters, trans, _THIS_IP_);
 615                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
 616         }
 617
 618         return 0;
 619 }
 620
 621 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 622
 623 static inline struct bkey_s_c
 624 __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 625                                    struct btree_iter *iter, unsigned flags)
 626 {
 627         struct bkey_s_c k;
 628
 629         while (btree_trans_too_many_iters(trans) ||
 630                (k = bch2_btree_iter_peek_type(iter, flags),
 631                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 632                 bch2_trans_begin(trans);
 633
 634         return k;
 635 }
 636
 637 static inline struct bkey_s_c
 638 __bch2_btree_iter_peek_upto_and_restart(struct btree_trans *trans,
 639                                         struct btree_iter *iter,
 640                                         struct bpos end,
 641                                         unsigned flags)
 642 {
 643         struct bkey_s_c k;
 644
 645         while (btree_trans_too_many_iters(trans) ||
 646                (k = bch2_btree_iter_peek_upto_type(iter, end, flags),
 647                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 648                 bch2_trans_begin(trans);
 649
 650         return k;
 651 }
 652
 653 #define lockrestart_do(_trans, _do)                                     \
 654 ({                                                                      \
 655         u32 _restart_count;                                             \
 656         int _ret;                                                       \
 657                                                                         \
 658         do {                                                            \
 659                 _restart_count = bch2_trans_begin(_trans);              \
 660                 _ret = (_do);                                           \
 661         } while (bch2_err_matches(_ret, BCH_ERR_transaction_restart));  \
 662                                                                         \
 663         if (!_ret)                                                      \
 664                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 665                                                                         \
 666         _ret;                                                           \
 667 })
 668
 669 /*
 670  * nested_lockrestart_do(), nested_commit_do():
 671  *
 672  * These are like lockrestart_do() and commit_do(), with two differences:
 673  *
 674  *  - We don't call bch2_trans_begin() unless we had a transaction restart
 675  *  - We return -BCH_ERR_transaction_restart_nested if we succeeded after a
 676  *  transaction restart
 677  */
 678 #define nested_lockrestart_do(_trans, _do)                              \
 679 ({                                                                      \
 680         u32 _restart_count, _orig_restart_count;                        \
 681         int _ret;                                                       \
 682                                                                         \
 683         _restart_count = _orig_restart_count = (_trans)->restart_count; \
 684                                                                         \
 685         while (bch2_err_matches(_ret = (_do), BCH_ERR_transaction_restart))\
 686                 _restart_count = bch2_trans_begin(_trans);              \
 687                                                                         \
 688         if (!_ret)                                                      \
 689                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 690                                                                         \
 691         if (!_ret && trans_was_restarted(_trans, _orig_restart_count))  \
 692                 _ret = -BCH_ERR_transaction_restart_nested;             \
 693                                                                         \
 694         _ret;                                                           \
 695 })
 696
 697 #define for_each_btree_key2(_trans, _iter, _btree_id,                   \
 698                             _start, _flags, _k, _do)                    \
 699 ({                                                                      \
 700         int _ret = 0;                                                   \
 701                                                                         \
 702         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 703                              (_start), (_flags));                       \
 704                                                                         \
 705         while (1) {                                                     \
 706                 u32 _restart_count = bch2_trans_begin(_trans);          \
 707                                                                         \
 708                 _ret = 0;                                               \
 709                 (_k) = bch2_btree_iter_peek_type(&(_iter), (_flags));   \
 710                 if (!(_k).k)                                            \
 711                         break;                                          \
 712                                                                         \
 713                 _ret = bkey_err(_k) ?: (_do);                           \
 714                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 715                         continue;                                       \
 716                 if (_ret)                                               \
 717                         break;                                          \
 718                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 719                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 720                         break;                                          \
 721         }                                                               \
 722                                                                         \
 723         bch2_trans_iter_exit((_trans), &(_iter));                       \
 724         _ret;                                                           \
 725 })
 726
 727 #define for_each_btree_key2_upto(_trans, _iter, _btree_id,              \
 728                             _start, _end, _flags, _k, _do)              \
 729 ({                                                                      \
 730         int _ret = 0;                                                   \
 731                                                                         \
 732         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 733                              (_start), (_flags));                       \
 734                                                                         \
 735         while (1) {                                                     \
 736                 u32 _restart_count = bch2_trans_begin(_trans);          \
 737                                                                         \
 738                 _ret = 0;                                               \
 739                 (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, (_flags));\
 740                 if (!(_k).k)                                            \
 741                         break;                                          \
 742                                                                         \
 743                 _ret = bkey_err(_k) ?: (_do);                           \
 744                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 745                         continue;                                       \
 746                 if (_ret)                                               \
 747                         break;                                          \
 748                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 749                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 750                         break;                                          \
 751         }                                                               \
 752                                                                         \
 753         bch2_trans_iter_exit((_trans), &(_iter));                       \
 754         _ret;                                                           \
 755 })
 756
 757 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,            \
 758                                    _start, _flags, _k, _do)             \
 759 ({                                                                      \
 760         int _ret = 0;                                                   \
 761                                                                         \
 762         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 763                              (_start), (_flags));                       \
 764                                                                         \
 765         while (1) {                                                     \
 766                 u32 _restart_count = bch2_trans_begin(_trans);          \
 767                 (_k) = bch2_btree_iter_peek_prev_type(&(_iter), (_flags));\
 768                 if (!(_k).k) {                                          \
 769                         _ret = 0;                                       \
 770                         break;                                          \
 771                 }                                                       \
 772                                                                         \
 773                 _ret = bkey_err(_k) ?: (_do);                           \
 774                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 775                         continue;                                       \
 776                 if (_ret)                                               \
 777                         break;                                          \
 778                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 779                 if (!bch2_btree_iter_rewind(&(_iter)))                  \
 780                         break;                                          \
 781         }                                                               \
 782                                                                         \
 783         bch2_trans_iter_exit((_trans), &(_iter));                       \
 784         _ret;                                                           \
 785 })
 786
 787 #define for_each_btree_key_commit(_trans, _iter, _btree_id,             \
 788                                   _start, _iter_flags, _k,              \
 789                                   _disk_res, _journal_seq, _commit_flags,\
 790                                   _do)                                  \
 791         for_each_btree_key2(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 792                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 793                                         (_journal_seq), (_commit_flags)))
 794
 795 #define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,        \
 796                                   _start, _end, _iter_flags, _k,        \
 797                                   _disk_res, _journal_seq, _commit_flags,\
 798                                   _do)                                  \
 799         for_each_btree_key2_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 800                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 801                                         (_journal_seq), (_commit_flags)))
 802
 803 #define for_each_btree_key(_trans, _iter, _btree_id,                    \
 804                            _start, _flags, _k, _ret)                    \
 805         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 806                                   (_start), (_flags));                  \
 807              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 808              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 809              bch2_btree_iter_advance(&(_iter)))
 810
 811 #define for_each_btree_key_upto(_trans, _iter, _btree_id,               \
 812                                 _start, _end, _flags, _k, _ret)         \
 813         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 814                                   (_start), (_flags));                  \
 815              (_k) = __bch2_btree_iter_peek_upto_and_restart((_trans),   \
 816                                                 &(_iter), _end, _flags),\
 817              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 818              bch2_btree_iter_advance(&(_iter)))
 819
 820 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,          \
 821                            _start, _flags, _k, _ret)                    \
 822         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 823                                   (_start), (_flags));                  \
 824              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 825              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 826              bch2_btree_iter_advance(&(_iter)))
 827
 828 #define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,     \
 829                            _start, _end, _flags, _k, _ret)              \
 830         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 831                                   (_start), (_flags));                  \
 832              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
 833              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 834              bch2_btree_iter_advance(&(_iter)))
 835
 836 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _ret)    \
 837         for (;                                                          \
 838              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 839              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 840              bch2_btree_iter_advance(&(_iter)))
 841
 842 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)  \
 843         for (;                                                          \
 844              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 845              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 846              bch2_btree_iter_advance(&(_iter)))
 847
 848 #define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
 849         for (;                                                                  \
 850              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),     \
 851              !((_ret) = bkey_err(_k)) && (_k).k;                                \
 852              bch2_btree_iter_advance(&(_iter)))
 853
 854 #define drop_locks_do(_trans, _do)                                      \
 855 ({                                                                      \
 856         bch2_trans_unlock(_trans);                                      \
 857         _do ?: bch2_trans_relock(_trans);                               \
 858 })
 859
 860 #define allocate_dropping_locks_errcode(_trans, _do)                    \
 861 ({                                                                      \
 862         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 863         int _ret = _do;                                                 \
 864                                                                         \
 865         if (bch2_err_matches(_ret, ENOMEM)) {                           \
 866                 _gfp = GFP_KERNEL;                                      \
 867                 _ret = drop_locks_do(trans, _do);                       \
 868         }                                                               \
 869         _ret;                                                           \
 870 })
 871
 872 #define allocate_dropping_locks(_trans, _ret, _do)                      \
 873 ({                                                                      \
 874         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 875         typeof(_do) _p = _do;                                           \
 876                                                                         \
 877         _ret = 0;                                                       \
 878         if (unlikely(!_p)) {                                            \
 879                 _gfp = GFP_KERNEL;                                      \
 880                 _ret = drop_locks_do(trans, ((_p = _do), 0));           \
 881         }                                                               \
 882         _p;                                                             \
 883 })
 884
 885 /* new multiple iterator interface: */
 886
 887 void bch2_trans_updates_to_text(struct printbuf *, struct btree_trans *);
 888 void bch2_btree_path_to_text(struct printbuf *, struct btree_path *);
 889 void bch2_trans_paths_to_text(struct printbuf *, struct btree_trans *);
 890 void bch2_dump_trans_updates(struct btree_trans *);
 891 void bch2_dump_trans_paths_updates(struct btree_trans *);
 892 void __bch2_trans_init(struct btree_trans *, struct bch_fs *, unsigned);
 893 void bch2_trans_exit(struct btree_trans *);
 894
 895 extern const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
 896 unsigned bch2_trans_get_fn_idx(const char *);
 897
 898 #define bch2_trans_init(_trans, _c, _nr_iters, _mem)                    \
 899 do {                                                                    \
 900         static unsigned trans_fn_idx;                                   \
 901                                                                         \
 902         if (unlikely(!trans_fn_idx))                                    \
 903                 trans_fn_idx = bch2_trans_get_fn_idx(__func__);         \
 904                                                                         \
 905         __bch2_trans_init(_trans, _c, trans_fn_idx);                    \
 906 } while (0)
 907
 908 void bch2_btree_trans_to_text(struct printbuf *, struct btree_trans *);
 909
 910 void bch2_fs_btree_iter_exit(struct bch_fs *);
 911 int bch2_fs_btree_iter_init(struct bch_fs *);
 912
 913 #endif /* _BCACHEFS_BTREE_ITER_H */