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         idx = find_next_bit(trans->paths_allocated, BTREE_ITER_MAX, idx);
  70         if (idx == BTREE_ITER_MAX)
  71                 return NULL;
  72         EBUG_ON(idx > BTREE_ITER_MAX);
  73         EBUG_ON(trans->paths[idx].idx != idx);
  74         return &trans->paths[idx];
  75 }
  76
  77 #define trans_for_each_path_from(_trans, _path, _start)                 \
  78         for (_path = __trans_next_path((_trans), _start);               \
  79              (_path);                                                   \
  80              _path = __trans_next_path((_trans), (_path)->idx + 1))
  81
  82 #define trans_for_each_path(_trans, _path)                              \
  83         trans_for_each_path_from(_trans, _path, 0)
  84
  85 static inline struct btree_path *
  86 __trans_next_path_safe(struct btree_trans *trans, unsigned *idx)
  87 {
  88         *idx = find_next_bit(trans->paths_allocated, BTREE_ITER_MAX, *idx);
  89         if (*idx == BTREE_ITER_MAX)
  90                 return NULL;
  91
  92         EBUG_ON(*idx > BTREE_ITER_MAX);
  93         return &trans->paths[*idx];
  94 }
  95
  96 /*
  97  * This version is intended to be safe for use on a btree_trans that is owned by
  98  * another thread, for bch2_btree_trans_to_text();
  99  */
 100 #define trans_for_each_path_safe_from(_trans, _path, _idx, _start)      \
 101         for (_idx = _start;                                             \
 102              (_path = __trans_next_path_safe((_trans), &_idx));         \
 103              _idx++)
 104
 105 #define trans_for_each_path_safe(_trans, _path, _idx)                   \
 106         trans_for_each_path_safe_from(_trans, _path, _idx, 0)
 107
 108 static inline struct btree_path *next_btree_path(struct btree_trans *trans, struct btree_path *path)
 109 {
 110         unsigned idx = path ? path->sorted_idx + 1 : 0;
 111
 112         EBUG_ON(idx > trans->nr_sorted);
 113
 114         return idx < trans->nr_sorted
 115                 ? trans->paths + trans->sorted[idx]
 116                 : NULL;
 117 }
 118
 119 static inline struct btree_path *prev_btree_path(struct btree_trans *trans, struct btree_path *path)
 120 {
 121         unsigned idx = path ? path->sorted_idx : trans->nr_sorted;
 122
 123         return idx
 124                 ? trans->paths + trans->sorted[idx - 1]
 125                 : NULL;
 126 }
 127
 128 #define trans_for_each_path_inorder(_trans, _path, _i)                  \
 129         for (_i = 0;                                                    \
 130              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) < (_trans)->nr_sorted;\
 131              _i++)
 132
 133 #define trans_for_each_path_inorder_reverse(_trans, _path, _i)          \
 134         for (_i = trans->nr_sorted - 1;                                 \
 135              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) >= 0;\
 136              --_i)
 137
 138 static inline bool __path_has_node(const struct btree_path *path,
 139                                    const struct btree *b)
 140 {
 141         return path->l[b->c.level].b == b &&
 142                 btree_node_lock_seq_matches(path, b, b->c.level);
 143 }
 144
 145 static inline struct btree_path *
 146 __trans_next_path_with_node(struct btree_trans *trans, struct btree *b,
 147                             unsigned idx)
 148 {
 149         struct btree_path *path = __trans_next_path(trans, idx);
 150
 151         while (path && !__path_has_node(path, b))
 152                 path = __trans_next_path(trans, path->idx + 1);
 153
 154         return path;
 155 }
 156
 157 #define trans_for_each_path_with_node(_trans, _b, _path)                \
 158         for (_path = __trans_next_path_with_node((_trans), (_b), 0);    \
 159              (_path);                                                   \
 160              _path = __trans_next_path_with_node((_trans), (_b),        \
 161                                                  (_path)->idx + 1))
 162
 163 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *, struct btree_path *,
 164                          bool, unsigned long);
 165
 166 static inline struct btree_path * __must_check
 167 bch2_btree_path_make_mut(struct btree_trans *trans,
 168                          struct btree_path *path, bool intent,
 169                          unsigned long ip)
 170 {
 171         if (path->ref > 1 || path->preserve)
 172                 path = __bch2_btree_path_make_mut(trans, path, intent, ip);
 173         path->should_be_locked = false;
 174         return path;
 175 }
 176
 177 struct btree_path * __must_check
 178 __bch2_btree_path_set_pos(struct btree_trans *, struct btree_path *,
 179                         struct bpos, bool, unsigned long, int);
 180
 181 static inline struct btree_path * __must_check
 182 bch2_btree_path_set_pos(struct btree_trans *trans,
 183                    struct btree_path *path, struct bpos new_pos,
 184                    bool intent, unsigned long ip)
 185 {
 186         int cmp = bpos_cmp(new_pos, path->pos);
 187
 188         return cmp
 189                 ? __bch2_btree_path_set_pos(trans, path, new_pos, intent, ip, cmp)
 190                 : path;
 191 }
 192
 193 int __must_check bch2_btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 194                                               unsigned, unsigned long);
 195
 196 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 197                                           struct btree_path *path, unsigned flags)
 198 {
 199         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
 200                 return 0;
 201
 202         return bch2_btree_path_traverse_one(trans, path, flags, _RET_IP_);
 203 }
 204
 205 int __must_check bch2_btree_path_traverse(struct btree_trans *,
 206                                           struct btree_path *, unsigned);
 207 struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpos,
 208                                  unsigned, unsigned, unsigned, unsigned long);
 209 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 210
 211 /*
 212  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
 213  * different snapshot:
 214  */
 215 static inline struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
 216 {
 217         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
 218
 219         if (k.k && bpos_eq(path->pos, k.k->p))
 220                 return k;
 221
 222         bkey_init(u);
 223         u->p = path->pos;
 224         return (struct bkey_s_c) { u, NULL };
 225 }
 226
 227 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 228                                         struct btree_iter *, struct bpos);
 229
 230 void bch2_btree_path_level_init(struct btree_trans *, struct btree_path *, struct btree *);
 231
 232 int __bch2_trans_mutex_lock(struct btree_trans *, struct mutex *);
 233
 234 static inline int bch2_trans_mutex_lock(struct btree_trans *trans, struct mutex *lock)
 235 {
 236         return mutex_trylock(lock)
 237                 ? 0
 238                 : __bch2_trans_mutex_lock(trans, lock);
 239 }
 240
 241 #ifdef CONFIG_BCACHEFS_DEBUG
 242 void bch2_trans_verify_paths(struct btree_trans *);
 243 void bch2_assert_pos_locked(struct btree_trans *, enum btree_id,
 244                             struct bpos, bool);
 245 #else
 246 static inline void bch2_trans_verify_paths(struct btree_trans *trans) {}
 247 static inline void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 248                                           struct bpos pos, bool key_cache) {}
 249 #endif
 250
 251 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 252                                       struct btree *, struct bkey_packed *);
 253 void bch2_btree_node_iter_fix(struct btree_trans *trans, struct btree_path *,
 254                               struct btree *, struct btree_node_iter *,
 255                               struct bkey_packed *, unsigned, unsigned);
 256
 257 int bch2_btree_path_relock_intent(struct btree_trans *, struct btree_path *);
 258
 259 void bch2_path_put(struct btree_trans *, struct btree_path *, bool);
 260
 261 int bch2_trans_relock(struct btree_trans *);
 262 int bch2_trans_relock_notrace(struct btree_trans *);
 263 void bch2_trans_unlock(struct btree_trans *);
 264 void bch2_trans_unlock_long(struct btree_trans *);
 265 bool bch2_trans_locked(struct btree_trans *);
 266
 267 static inline int trans_was_restarted(struct btree_trans *trans, u32 restart_count)
 268 {
 269         return restart_count != trans->restart_count
 270                 ? -BCH_ERR_transaction_restart_nested
 271                 : 0;
 272 }
 273
 274 void __noreturn bch2_trans_restart_error(struct btree_trans *, u32);
 275
 276 static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 277                                                    u32 restart_count)
 278 {
 279         if (trans_was_restarted(trans, restart_count))
 280                 bch2_trans_restart_error(trans, restart_count);
 281 }
 282
 283 void __noreturn bch2_trans_in_restart_error(struct btree_trans *);
 284
 285 static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
 286 {
 287         if (trans->restarted)
 288                 bch2_trans_in_restart_error(trans);
 289 }
 290
 291 __always_inline
 292 static int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
 293 {
 294         BUG_ON(err <= 0);
 295         BUG_ON(!bch2_err_matches(-err, BCH_ERR_transaction_restart));
 296
 297         trans->restarted = err;
 298         trans->last_restarted_ip = _THIS_IP_;
 299         return -err;
 300 }
 301
 302 __always_inline
 303 static int btree_trans_restart(struct btree_trans *trans, int err)
 304 {
 305         btree_trans_restart_nounlock(trans, err);
 306         return -err;
 307 }
 308
 309 bool bch2_btree_node_upgrade(struct btree_trans *,
 310                              struct btree_path *, unsigned);
 311
 312 void __bch2_btree_path_downgrade(struct btree_trans *, struct btree_path *, unsigned);
 313
 314 static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 315                                              struct btree_path *path)
 316 {
 317         unsigned new_locks_want = path->level + !!path->intent_ref;
 318
 319         if (path->locks_want > new_locks_want)
 320                 __bch2_btree_path_downgrade(trans, path, new_locks_want);
 321 }
 322
 323 void bch2_trans_downgrade(struct btree_trans *);
 324
 325 void bch2_trans_node_add(struct btree_trans *trans, struct btree *);
 326 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 327
 328 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
 329 int __must_check bch2_btree_iter_traverse(struct btree_iter *);
 330
 331 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 332 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 333 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 334
 335 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
 336 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 337
 338 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 339 {
 340         return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
 341 }
 342
 343 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 344 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 345
 346 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 347 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 348 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *);
 349
 350 bool bch2_btree_iter_advance(struct btree_iter *);
 351 bool bch2_btree_iter_rewind(struct btree_iter *);
 352
 353 static inline void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 354 {
 355         iter->k.type = KEY_TYPE_deleted;
 356         iter->k.p.inode         = iter->pos.inode       = new_pos.inode;
 357         iter->k.p.offset        = iter->pos.offset      = new_pos.offset;
 358         iter->k.p.snapshot      = iter->pos.snapshot    = new_pos.snapshot;
 359         iter->k.size = 0;
 360 }
 361
 362 static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 363 {
 364         if (unlikely(iter->update_path))
 365                 bch2_path_put(iter->trans, iter->update_path,
 366                               iter->flags & BTREE_ITER_INTENT);
 367         iter->update_path = NULL;
 368
 369         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
 370                 new_pos.snapshot = iter->snapshot;
 371
 372         __bch2_btree_iter_set_pos(iter, new_pos);
 373 }
 374
 375 static inline void bch2_btree_iter_set_pos_to_extent_start(struct btree_iter *iter)
 376 {
 377         BUG_ON(!(iter->flags & BTREE_ITER_IS_EXTENTS));
 378         iter->pos = bkey_start_pos(&iter->k);
 379 }
 380
 381 static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 snapshot)
 382 {
 383         struct bpos pos = iter->pos;
 384
 385         iter->snapshot = snapshot;
 386         pos.snapshot = snapshot;
 387         bch2_btree_iter_set_pos(iter, pos);
 388 }
 389
 390 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 391
 392 static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 393                                                unsigned btree_id,
 394                                                unsigned flags)
 395 {
 396         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
 397             btree_id_is_extents(btree_id))
 398                 flags |= BTREE_ITER_IS_EXTENTS;
 399
 400         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 401             !btree_type_has_snapshot_field(btree_id))
 402                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 403
 404         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 405             btree_type_has_snapshots(btree_id))
 406                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 407
 408         if (trans->journal_replay_not_finished)
 409                 flags |= BTREE_ITER_WITH_JOURNAL;
 410
 411         return flags;
 412 }
 413
 414 static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
 415                                              unsigned btree_id,
 416                                              unsigned flags)
 417 {
 418         if (!btree_id_cached(trans->c, btree_id)) {
 419                 flags &= ~BTREE_ITER_CACHED;
 420                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 421         } else if (!(flags & BTREE_ITER_CACHED))
 422                 flags |= BTREE_ITER_WITH_KEY_CACHE;
 423
 424         return __bch2_btree_iter_flags(trans, btree_id, flags);
 425 }
 426
 427 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 428                                           struct btree_iter *iter,
 429                                           unsigned btree_id, struct bpos pos,
 430                                           unsigned locks_want,
 431                                           unsigned depth,
 432                                           unsigned flags,
 433                                           unsigned long ip)
 434 {
 435         iter->trans             = trans;
 436         iter->update_path       = NULL;
 437         iter->key_cache_path    = NULL;
 438         iter->btree_id          = btree_id;
 439         iter->min_depth         = 0;
 440         iter->flags             = flags;
 441         iter->snapshot          = pos.snapshot;
 442         iter->pos               = pos;
 443         iter->k                 = POS_KEY(pos);
 444         iter->journal_idx       = 0;
 445 #ifdef CONFIG_BCACHEFS_DEBUG
 446         iter->ip_allocated = ip;
 447 #endif
 448         iter->path = bch2_path_get(trans, btree_id, iter->pos,
 449                                    locks_want, depth, flags, ip);
 450 }
 451
 452 void bch2_trans_iter_init_outlined(struct btree_trans *, struct btree_iter *,
 453                           enum btree_id, struct bpos, unsigned);
 454
 455 static inline void bch2_trans_iter_init(struct btree_trans *trans,
 456                           struct btree_iter *iter,
 457                           unsigned btree_id, struct bpos pos,
 458                           unsigned flags)
 459 {
 460         if (__builtin_constant_p(btree_id) &&
 461             __builtin_constant_p(flags))
 462                 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
 463                                 bch2_btree_iter_flags(trans, btree_id, flags),
 464                                 _THIS_IP_);
 465         else
 466                 bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
 467 }
 468
 469 void bch2_trans_node_iter_init(struct btree_trans *, struct btree_iter *,
 470                                enum btree_id, struct bpos,
 471                                unsigned, unsigned, unsigned);
 472 void bch2_trans_copy_iter(struct btree_iter *, struct btree_iter *);
 473
 474 static inline void set_btree_iter_dontneed(struct btree_iter *iter)
 475 {
 476         if (!iter->trans->restarted)
 477                 iter->path->preserve = false;
 478 }
 479
 480 void *__bch2_trans_kmalloc(struct btree_trans *, size_t);
 481
 482 static inline void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
 483 {
 484         size = roundup(size, 8);
 485
 486         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 487                 void *p = trans->mem + trans->mem_top;
 488
 489                 trans->mem_top += size;
 490                 memset(p, 0, size);
 491                 return p;
 492         } else {
 493                 return __bch2_trans_kmalloc(trans, size);
 494         }
 495 }
 496
 497 static inline void *bch2_trans_kmalloc_nomemzero(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                 return p;
 506         } else {
 507                 return __bch2_trans_kmalloc(trans, size);
 508         }
 509 }
 510
 511 static inline struct bkey_s_c __bch2_bkey_get_iter(struct btree_trans *trans,
 512                                 struct btree_iter *iter,
 513                                 unsigned btree_id, struct bpos pos,
 514                                 unsigned flags, unsigned type)
 515 {
 516         struct bkey_s_c k;
 517
 518         bch2_trans_iter_init(trans, iter, btree_id, pos, flags);
 519         k = bch2_btree_iter_peek_slot(iter);
 520
 521         if (!bkey_err(k) && type && k.k->type != type)
 522                 k = bkey_s_c_err(-BCH_ERR_ENOENT_bkey_type_mismatch);
 523         if (unlikely(bkey_err(k)))
 524                 bch2_trans_iter_exit(trans, iter);
 525         return k;
 526 }
 527
 528 static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 529                                 struct btree_iter *iter,
 530                                 unsigned btree_id, struct bpos pos,
 531                                 unsigned flags)
 532 {
 533         return __bch2_bkey_get_iter(trans, iter, btree_id, pos, flags, 0);
 534 }
 535
 536 #define bch2_bkey_get_iter_typed(_trans, _iter, _btree_id, _pos, _flags, _type)\
 537         bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,                 \
 538                                        _btree_id, _pos, _flags, KEY_TYPE_##_type))
 539
 540 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 541                                 unsigned btree_id, struct bpos pos,
 542                                 unsigned flags, unsigned type,
 543                                 unsigned val_size, void *val)
 544 {
 545         struct btree_iter iter;
 546         struct bkey_s_c k;
 547         int ret;
 548
 549         k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
 550         ret = bkey_err(k);
 551         if (!ret) {
 552                 unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
 553
 554                 memcpy(val, k.v, b);
 555                 if (unlikely(b < sizeof(*val)))
 556                         memset((void *) val + b, 0, sizeof(*val) - b);
 557                 bch2_trans_iter_exit(trans, &iter);
 558         }
 559
 560         return ret;
 561 }
 562
 563 #define bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags, _type, _val)\
 564         __bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags,      \
 565                                   KEY_TYPE_##_type, sizeof(*_val), _val)
 566
 567 void bch2_trans_srcu_unlock(struct btree_trans *);
 568 void bch2_trans_srcu_lock(struct btree_trans *);
 569
 570 u32 bch2_trans_begin(struct btree_trans *);
 571
 572 /*
 573  * XXX
 574  * this does not handle transaction restarts from bch2_btree_iter_next_node()
 575  * correctly
 576  */
 577 #define __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 578                               _locks_want, _depth, _flags, _b, _ret)    \
 579         for (bch2_trans_node_iter_init((_trans), &(_iter), (_btree_id), \
 580                                 _start, _locks_want, _depth, _flags);   \
 581              (_b) = bch2_btree_iter_peek_node_and_restart(&(_iter)),    \
 582              !((_ret) = PTR_ERR_OR_ZERO(_b)) && (_b);                   \
 583              (_b) = bch2_btree_iter_next_node(&(_iter)))
 584
 585 #define for_each_btree_node(_trans, _iter, _btree_id, _start,           \
 586                             _flags, _b, _ret)                           \
 587         __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 588                               0, 0, _flags, _b, _ret)
 589
 590 static inline struct bkey_s_c bch2_btree_iter_peek_prev_type(struct btree_iter *iter,
 591                                                              unsigned flags)
 592 {
 593         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 594                                                 bch2_btree_iter_peek_prev(iter);
 595 }
 596
 597 static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 598                                                         unsigned flags)
 599 {
 600         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 601                                                 bch2_btree_iter_peek(iter);
 602 }
 603
 604 static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
 605                                                              struct bpos end,
 606                                                              unsigned flags)
 607 {
 608         if (!(flags & BTREE_ITER_SLOTS))
 609                 return bch2_btree_iter_peek_upto(iter, end);
 610
 611         if (bkey_gt(iter->pos, end))
 612                 return bkey_s_c_null;
 613
 614         return bch2_btree_iter_peek_slot(iter);
 615 }
 616
 617 int __bch2_btree_trans_too_many_iters(struct btree_trans *);
 618
 619 static inline int btree_trans_too_many_iters(struct btree_trans *trans)
 620 {
 621         if (bitmap_weight(trans->paths_allocated, BTREE_ITER_MAX) > BTREE_ITER_MAX - 8)
 622                 return __bch2_btree_trans_too_many_iters(trans);
 623
 624         return 0;
 625 }
 626
 627 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 628
 629 static inline struct bkey_s_c
 630 __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 631                                    struct btree_iter *iter, unsigned flags)
 632 {
 633         struct bkey_s_c k;
 634
 635         while (btree_trans_too_many_iters(trans) ||
 636                (k = bch2_btree_iter_peek_type(iter, flags),
 637                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 638                 bch2_trans_begin(trans);
 639
 640         return k;
 641 }
 642
 643 static inline struct bkey_s_c
 644 __bch2_btree_iter_peek_upto_and_restart(struct btree_trans *trans,
 645                                         struct btree_iter *iter,
 646                                         struct bpos end,
 647                                         unsigned flags)
 648 {
 649         struct bkey_s_c k;
 650
 651         while (btree_trans_too_many_iters(trans) ||
 652                (k = bch2_btree_iter_peek_upto_type(iter, end, flags),
 653                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 654                 bch2_trans_begin(trans);
 655
 656         return k;
 657 }
 658
 659 #define lockrestart_do(_trans, _do)                                     \
 660 ({                                                                      \
 661         u32 _restart_count;                                             \
 662         int _ret2;                                                      \
 663                                                                         \
 664         do {                                                            \
 665                 _restart_count = bch2_trans_begin(_trans);              \
 666                 _ret2 = (_do);                                          \
 667         } while (bch2_err_matches(_ret2, BCH_ERR_transaction_restart)); \
 668                                                                         \
 669         if (!_ret2)                                                     \
 670                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 671                                                                         \
 672         _ret2;                                                          \
 673 })
 674
 675 /*
 676  * nested_lockrestart_do(), nested_commit_do():
 677  *
 678  * These are like lockrestart_do() and commit_do(), with two differences:
 679  *
 680  *  - We don't call bch2_trans_begin() unless we had a transaction restart
 681  *  - We return -BCH_ERR_transaction_restart_nested if we succeeded after a
 682  *  transaction restart
 683  */
 684 #define nested_lockrestart_do(_trans, _do)                              \
 685 ({                                                                      \
 686         u32 _restart_count, _orig_restart_count;                        \
 687         int _ret2;                                                      \
 688                                                                         \
 689         _restart_count = _orig_restart_count = (_trans)->restart_count; \
 690                                                                         \
 691         while (bch2_err_matches(_ret2 = (_do), BCH_ERR_transaction_restart))\
 692                 _restart_count = bch2_trans_begin(_trans);              \
 693                                                                         \
 694         if (!_ret2)                                                     \
 695                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 696                                                                         \
 697         _ret2 ?: trans_was_restarted(_trans, _restart_count);           \
 698 })
 699
 700 #define for_each_btree_key2(_trans, _iter, _btree_id,                   \
 701                             _start, _flags, _k, _do)                    \
 702 ({                                                                      \
 703         int _ret3 = 0;                                                  \
 704                                                                         \
 705         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 706                              (_start), (_flags));                       \
 707                                                                         \
 708         while (1) {                                                     \
 709                 u32 _restart_count = bch2_trans_begin(_trans);          \
 710                                                                         \
 711                 _ret3 = 0;                                              \
 712                 (_k) = bch2_btree_iter_peek_type(&(_iter), (_flags));   \
 713                 if (!(_k).k)                                            \
 714                         break;                                          \
 715                                                                         \
 716                 _ret3 = bkey_err(_k) ?: (_do);                          \
 717                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 718                         continue;                                       \
 719                 if (_ret3)                                              \
 720                         break;                                          \
 721                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 722                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 723                         break;                                          \
 724         }                                                               \
 725                                                                         \
 726         bch2_trans_iter_exit((_trans), &(_iter));                       \
 727         _ret3;                                                          \
 728 })
 729
 730 #define for_each_btree_key2_upto(_trans, _iter, _btree_id,              \
 731                             _start, _end, _flags, _k, _do)              \
 732 ({                                                                      \
 733         int _ret3 = 0;                                                  \
 734                                                                         \
 735         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 736                              (_start), (_flags));                       \
 737                                                                         \
 738         while (1) {                                                     \
 739                 u32 _restart_count = bch2_trans_begin(_trans);          \
 740                                                                         \
 741                 _ret3 = 0;                                              \
 742                 (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, (_flags));\
 743                 if (!(_k).k)                                            \
 744                         break;                                          \
 745                                                                         \
 746                 _ret3 = bkey_err(_k) ?: (_do);                          \
 747                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 748                         continue;                                       \
 749                 if (_ret3)                                              \
 750                         break;                                          \
 751                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 752                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 753                         break;                                          \
 754         }                                                               \
 755                                                                         \
 756         bch2_trans_iter_exit((_trans), &(_iter));                       \
 757         _ret3;                                                          \
 758 })
 759
 760 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,            \
 761                                    _start, _flags, _k, _do)             \
 762 ({                                                                      \
 763         int _ret3 = 0;                                                  \
 764                                                                         \
 765         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 766                              (_start), (_flags));                       \
 767                                                                         \
 768         while (1) {                                                     \
 769                 u32 _restart_count = bch2_trans_begin(_trans);          \
 770                 (_k) = bch2_btree_iter_peek_prev_type(&(_iter), (_flags));\
 771                 if (!(_k).k) {                                          \
 772                         _ret3 = 0;                                      \
 773                         break;                                          \
 774                 }                                                       \
 775                                                                         \
 776                 _ret3 = bkey_err(_k) ?: (_do);                          \
 777                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 778                         continue;                                       \
 779                 if (_ret3)                                              \
 780                         break;                                          \
 781                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 782                 if (!bch2_btree_iter_rewind(&(_iter)))                  \
 783                         break;                                          \
 784         }                                                               \
 785                                                                         \
 786         bch2_trans_iter_exit((_trans), &(_iter));                       \
 787         _ret3;                                                          \
 788 })
 789
 790 #define for_each_btree_key_commit(_trans, _iter, _btree_id,             \
 791                                   _start, _iter_flags, _k,              \
 792                                   _disk_res, _journal_seq, _commit_flags,\
 793                                   _do)                                  \
 794         for_each_btree_key2(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 795                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 796                                         (_journal_seq), (_commit_flags)))
 797
 798 #define for_each_btree_key_reverse_commit(_trans, _iter, _btree_id,     \
 799                                   _start, _iter_flags, _k,              \
 800                                   _disk_res, _journal_seq, _commit_flags,\
 801                                   _do)                                  \
 802         for_each_btree_key_reverse(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 803                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 804                                         (_journal_seq), (_commit_flags)))
 805
 806 #define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,        \
 807                                   _start, _end, _iter_flags, _k,        \
 808                                   _disk_res, _journal_seq, _commit_flags,\
 809                                   _do)                                  \
 810         for_each_btree_key2_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 811                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 812                                         (_journal_seq), (_commit_flags)))
 813
 814 #define for_each_btree_key(_trans, _iter, _btree_id,                    \
 815                            _start, _flags, _k, _ret)                    \
 816         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 817                                   (_start), (_flags));                  \
 818              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 819              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 820              bch2_btree_iter_advance(&(_iter)))
 821
 822 #define for_each_btree_key_upto(_trans, _iter, _btree_id,               \
 823                                 _start, _end, _flags, _k, _ret)         \
 824         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 825                                   (_start), (_flags));                  \
 826              (_k) = __bch2_btree_iter_peek_upto_and_restart((_trans),   \
 827                                                 &(_iter), _end, _flags),\
 828              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 829              bch2_btree_iter_advance(&(_iter)))
 830
 831 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,          \
 832                            _start, _flags, _k, _ret)                    \
 833         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 834                                   (_start), (_flags));                  \
 835              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 836              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 837              bch2_btree_iter_advance(&(_iter)))
 838
 839 #define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,     \
 840                            _start, _end, _flags, _k, _ret)              \
 841         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 842                                   (_start), (_flags));                  \
 843              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
 844              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 845              bch2_btree_iter_advance(&(_iter)))
 846
 847 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _ret)    \
 848         for (;                                                          \
 849              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 850              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 851              bch2_btree_iter_advance(&(_iter)))
 852
 853 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)  \
 854         for (;                                                          \
 855              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 856              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 857              bch2_btree_iter_advance(&(_iter)))
 858
 859 #define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
 860         for (;                                                                  \
 861              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),     \
 862              !((_ret) = bkey_err(_k)) && (_k).k;                                \
 863              bch2_btree_iter_advance(&(_iter)))
 864
 865 #define drop_locks_do(_trans, _do)                                      \
 866 ({                                                                      \
 867         bch2_trans_unlock(_trans);                                      \
 868         _do ?: bch2_trans_relock(_trans);                               \
 869 })
 870
 871 #define allocate_dropping_locks_errcode(_trans, _do)                    \
 872 ({                                                                      \
 873         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 874         int _ret = _do;                                                 \
 875                                                                         \
 876         if (bch2_err_matches(_ret, ENOMEM)) {                           \
 877                 _gfp = GFP_KERNEL;                                      \
 878                 _ret = drop_locks_do(trans, _do);                       \
 879         }                                                               \
 880         _ret;                                                           \
 881 })
 882
 883 #define allocate_dropping_locks(_trans, _ret, _do)                      \
 884 ({                                                                      \
 885         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 886         typeof(_do) _p = _do;                                           \
 887                                                                         \
 888         _ret = 0;                                                       \
 889         if (unlikely(!_p)) {                                            \
 890                 _gfp = GFP_KERNEL;                                      \
 891                 _ret = drop_locks_do(trans, ((_p = _do), 0));           \
 892         }                                                               \
 893         _p;                                                             \
 894 })
 895
 896 /* new multiple iterator interface: */
 897
 898 void bch2_trans_updates_to_text(struct printbuf *, struct btree_trans *);
 899 void bch2_btree_path_to_text(struct printbuf *, struct btree_path *);
 900 void bch2_trans_paths_to_text(struct printbuf *, struct btree_trans *);
 901 void bch2_dump_trans_updates(struct btree_trans *);
 902 void bch2_dump_trans_paths_updates(struct btree_trans *);
 903
 904 struct btree_trans *__bch2_trans_get(struct bch_fs *, unsigned);
 905 void bch2_trans_put(struct btree_trans *);
 906
 907 extern const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
 908 unsigned bch2_trans_get_fn_idx(const char *);
 909
 910 #define bch2_trans_get(_c)                                              \
 911 ({                                                                      \
 912         static unsigned trans_fn_idx;                                   \
 913                                                                         \
 914         if (unlikely(!trans_fn_idx))                                    \
 915                 trans_fn_idx = bch2_trans_get_fn_idx(__func__);         \
 916         __bch2_trans_get(_c, trans_fn_idx);                             \
 917 })
 918
 919 void bch2_btree_trans_to_text(struct printbuf *, struct btree_trans *);
 920
 921 void bch2_fs_btree_iter_exit(struct bch_fs *);
 922 int bch2_fs_btree_iter_init(struct bch_fs *);
 923
 924 #endif /* _BCACHEFS_BTREE_ITER_H */