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         /*
  46          * We don't compare the low bits of the lock sequence numbers because
  47          * @path might have taken a write lock on @b, and we don't want to skip
  48          * the linked path if the sequence numbers were equal before taking that
  49          * write lock. The lock sequence number is incremented by taking and
  50          * releasing write locks and is even when unlocked:
  51          */
  52         return path->l[level].lock_seq >> 1 == b->c.lock.state.seq >> 1;
  53 }
  54
  55 static inline struct btree *btree_node_parent(struct btree_path *path,
  56                                               struct btree *b)
  57 {
  58         return btree_path_node(path, b->c.level + 1);
  59 }
  60
  61 /* Iterate over paths within a transaction: */
  62
  63 void __bch2_btree_trans_sort_paths(struct btree_trans *);
  64
  65 static inline void btree_trans_sort_paths(struct btree_trans *trans)
  66 {
  67         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
  68             trans->paths_sorted)
  69                 return;
  70         __bch2_btree_trans_sort_paths(trans);
  71 }
  72
  73 static inline struct btree_path *
  74 __trans_next_path(struct btree_trans *trans, unsigned idx)
  75 {
  76         u64 l;
  77
  78         if (idx == BTREE_ITER_MAX)
  79                 return NULL;
  80
  81         l = trans->paths_allocated >> idx;
  82         if (!l)
  83                 return NULL;
  84
  85         idx += __ffs64(l);
  86         EBUG_ON(idx >= BTREE_ITER_MAX);
  87         EBUG_ON(trans->paths[idx].idx != idx);
  88         return &trans->paths[idx];
  89 }
  90
  91 #define trans_for_each_path_from(_trans, _path, _start)                 \
  92         for (_path = __trans_next_path((_trans), _start);               \
  93              (_path);                                                   \
  94              _path = __trans_next_path((_trans), (_path)->idx + 1))
  95
  96 #define trans_for_each_path(_trans, _path)                              \
  97         trans_for_each_path_from(_trans, _path, 0)
  98
  99 static inline struct btree_path *next_btree_path(struct btree_trans *trans, struct btree_path *path)
 100 {
 101         unsigned idx = path ? path->sorted_idx + 1 : 0;
 102
 103         EBUG_ON(idx > trans->nr_sorted);
 104
 105         return idx < trans->nr_sorted
 106                 ? trans->paths + trans->sorted[idx]
 107                 : NULL;
 108 }
 109
 110 static inline struct btree_path *prev_btree_path(struct btree_trans *trans, struct btree_path *path)
 111 {
 112         unsigned idx = path ? path->sorted_idx : trans->nr_sorted;
 113
 114         return idx
 115                 ? trans->paths + trans->sorted[idx - 1]
 116                 : NULL;
 117 }
 118
 119 #define trans_for_each_path_inorder(_trans, _path, _i)                  \
 120         for (_i = 0;                                                    \
 121              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) < (_trans)->nr_sorted;\
 122              _i++)
 123
 124 #define trans_for_each_path_inorder_reverse(_trans, _path, _i)          \
 125         for (_i = trans->nr_sorted - 1;                                 \
 126              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) >= 0;\
 127              --_i)
 128
 129 static inline bool __path_has_node(const struct btree_path *path,
 130                                    const struct btree *b)
 131 {
 132         return path->l[b->c.level].b == b &&
 133                 btree_node_lock_seq_matches(path, b, b->c.level);
 134 }
 135
 136 static inline struct btree_path *
 137 __trans_next_path_with_node(struct btree_trans *trans, struct btree *b,
 138                             unsigned idx)
 139 {
 140         struct btree_path *path = __trans_next_path(trans, idx);
 141
 142         while (path && !__path_has_node(path, b))
 143                 path = __trans_next_path(trans, path->idx + 1);
 144
 145         return path;
 146 }
 147
 148 #define trans_for_each_path_with_node(_trans, _b, _path)                \
 149         for (_path = __trans_next_path_with_node((_trans), (_b), 0);    \
 150              (_path);                                                   \
 151              _path = __trans_next_path_with_node((_trans), (_b),        \
 152                                                  (_path)->idx + 1))
 153
 154 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *, struct btree_path *,
 155                          bool, unsigned long);
 156
 157 static inline struct btree_path * __must_check
 158 bch2_btree_path_make_mut(struct btree_trans *trans,
 159                          struct btree_path *path, bool intent,
 160                          unsigned long ip)
 161 {
 162         if (path->ref > 1 || path->preserve)
 163                 path = __bch2_btree_path_make_mut(trans, path, intent, ip);
 164         path->should_be_locked = false;
 165         return path;
 166 }
 167
 168 struct btree_path * __must_check
 169 __bch2_btree_path_set_pos(struct btree_trans *, struct btree_path *,
 170                         struct bpos, bool, unsigned long, int);
 171
 172 static inline struct btree_path * __must_check
 173 bch2_btree_path_set_pos(struct btree_trans *trans,
 174                    struct btree_path *path, struct bpos new_pos,
 175                    bool intent, unsigned long ip)
 176 {
 177         int cmp = bpos_cmp(new_pos, path->pos);
 178
 179         return cmp
 180                 ? __bch2_btree_path_set_pos(trans, path, new_pos, intent, ip, cmp)
 181                 : path;
 182 }
 183
 184 int __must_check bch2_btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 185                                               unsigned, unsigned long);
 186
 187 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 188                                           struct btree_path *path, unsigned flags)
 189 {
 190         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
 191                 return 0;
 192
 193         return bch2_btree_path_traverse_one(trans, path, flags, _RET_IP_);
 194 }
 195
 196 int __must_check bch2_btree_path_traverse(struct btree_trans *,
 197                                           struct btree_path *, unsigned);
 198 struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpos,
 199                                  unsigned, unsigned, unsigned, unsigned long);
 200 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 201
 202 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 203                                         struct btree_iter *, struct bpos);
 204
 205 void bch2_btree_path_level_init(struct btree_trans *, struct btree_path *, struct btree *);
 206
 207 int __bch2_trans_mutex_lock(struct btree_trans *, struct mutex *);
 208
 209 static inline int bch2_trans_mutex_lock(struct btree_trans *trans, struct mutex *lock)
 210 {
 211         return mutex_trylock(lock)
 212                 ? 0
 213                 : __bch2_trans_mutex_lock(trans, lock);
 214 }
 215
 216 #ifdef CONFIG_BCACHEFS_DEBUG
 217 void bch2_trans_verify_paths(struct btree_trans *);
 218 void bch2_assert_pos_locked(struct btree_trans *, enum btree_id,
 219                             struct bpos, bool);
 220 #else
 221 static inline void bch2_trans_verify_paths(struct btree_trans *trans) {}
 222 static inline void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 223                                           struct bpos pos, bool key_cache) {}
 224 #endif
 225
 226 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 227                                       struct btree *, struct bkey_packed *);
 228 void bch2_btree_node_iter_fix(struct btree_trans *trans, struct btree_path *,
 229                               struct btree *, struct btree_node_iter *,
 230                               struct bkey_packed *, unsigned, unsigned);
 231
 232 int bch2_btree_path_relock_intent(struct btree_trans *, struct btree_path *);
 233
 234 void bch2_path_put(struct btree_trans *, struct btree_path *, bool);
 235
 236 int bch2_trans_relock(struct btree_trans *);
 237 int bch2_trans_relock_notrace(struct btree_trans *);
 238 void bch2_trans_unlock(struct btree_trans *);
 239 bool bch2_trans_locked(struct btree_trans *);
 240
 241 static inline bool trans_was_restarted(struct btree_trans *trans, u32 restart_count)
 242 {
 243         return restart_count != trans->restart_count;
 244 }
 245
 246 void bch2_trans_restart_error(struct btree_trans *, u32);
 247
 248 static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 249                                                    u32 restart_count)
 250 {
 251         if (trans_was_restarted(trans, restart_count))
 252                 bch2_trans_restart_error(trans, restart_count);
 253 }
 254
 255 void bch2_trans_in_restart_error(struct btree_trans *);
 256
 257 static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
 258 {
 259         if (trans->restarted)
 260                 bch2_trans_in_restart_error(trans);
 261 }
 262
 263 __always_inline
 264 static inline int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
 265 {
 266         BUG_ON(err <= 0);
 267         BUG_ON(!bch2_err_matches(err, BCH_ERR_transaction_restart));
 268
 269         trans->restarted = err;
 270         trans->last_restarted_ip = _THIS_IP_;
 271         return -err;
 272 }
 273
 274 __always_inline
 275 static inline int btree_trans_restart(struct btree_trans *trans, int err)
 276 {
 277         btree_trans_restart_nounlock(trans, err);
 278         return -err;
 279 }
 280
 281 bool bch2_btree_node_upgrade(struct btree_trans *,
 282                              struct btree_path *, unsigned);
 283
 284 void __bch2_btree_path_downgrade(struct btree_trans *, struct btree_path *, unsigned);
 285
 286 static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 287                                              struct btree_path *path)
 288 {
 289         unsigned new_locks_want = path->level + !!path->intent_ref;
 290
 291         if (path->locks_want > new_locks_want)
 292                 __bch2_btree_path_downgrade(trans, path, new_locks_want);
 293 }
 294
 295 void bch2_trans_downgrade(struct btree_trans *);
 296
 297 void bch2_trans_node_add(struct btree_trans *trans, struct btree *);
 298 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 299
 300 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
 301 int __must_check bch2_btree_iter_traverse(struct btree_iter *);
 302
 303 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 304 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 305 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 306
 307 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
 308 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 309
 310 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *);
 311
 312 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 313 {
 314         return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
 315 }
 316
 317 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 318 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 319
 320 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 321 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 322 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *);
 323
 324 bool bch2_btree_iter_advance(struct btree_iter *);
 325 bool bch2_btree_iter_rewind(struct btree_iter *);
 326
 327 static inline void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 328 {
 329         iter->k.type = KEY_TYPE_deleted;
 330         iter->k.p.inode         = iter->pos.inode       = new_pos.inode;
 331         iter->k.p.offset        = iter->pos.offset      = new_pos.offset;
 332         iter->k.p.snapshot      = iter->pos.snapshot    = new_pos.snapshot;
 333         iter->k.size = 0;
 334 }
 335
 336 static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 337 {
 338         if (unlikely(iter->update_path))
 339                 bch2_path_put(iter->trans, iter->update_path,
 340                               iter->flags & BTREE_ITER_INTENT);
 341         iter->update_path = NULL;
 342
 343         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
 344                 new_pos.snapshot = iter->snapshot;
 345
 346         __bch2_btree_iter_set_pos(iter, new_pos);
 347 }
 348
 349 static inline void bch2_btree_iter_set_pos_to_extent_start(struct btree_iter *iter)
 350 {
 351         BUG_ON(!(iter->flags & BTREE_ITER_IS_EXTENTS));
 352         iter->pos = bkey_start_pos(&iter->k);
 353 }
 354
 355 static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 snapshot)
 356 {
 357         struct bpos pos = iter->pos;
 358
 359         iter->snapshot = snapshot;
 360         pos.snapshot = snapshot;
 361         bch2_btree_iter_set_pos(iter, pos);
 362 }
 363
 364 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 365
 366 static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 367                                                unsigned btree_id,
 368                                                unsigned flags)
 369 {
 370         if (flags & BTREE_ITER_ALL_LEVELS)
 371                 flags |= BTREE_ITER_ALL_SNAPSHOTS|__BTREE_ITER_ALL_SNAPSHOTS;
 372
 373         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
 374             btree_node_type_is_extents(btree_id))
 375                 flags |= BTREE_ITER_IS_EXTENTS;
 376
 377         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 378             !btree_type_has_snapshots(btree_id))
 379                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 380
 381         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 382             btree_type_has_snapshots(btree_id))
 383                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 384
 385         if (trans->journal_replay_not_finished)
 386                 flags |= BTREE_ITER_WITH_JOURNAL;
 387
 388         return flags;
 389 }
 390
 391 static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
 392                                              unsigned btree_id,
 393                                              unsigned flags)
 394 {
 395         if (!btree_id_cached(trans->c, btree_id)) {
 396                 flags &= ~BTREE_ITER_CACHED;
 397                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 398         } else if (!(flags & BTREE_ITER_CACHED))
 399                 flags |= BTREE_ITER_WITH_KEY_CACHE;
 400
 401         return __bch2_btree_iter_flags(trans, btree_id, flags);
 402 }
 403
 404 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 405                                           struct btree_iter *iter,
 406                                           unsigned btree_id, struct bpos pos,
 407                                           unsigned locks_want,
 408                                           unsigned depth,
 409                                           unsigned flags,
 410                                           unsigned long ip)
 411 {
 412         memset(iter, 0, sizeof(*iter));
 413         iter->trans     = trans;
 414         iter->btree_id  = btree_id;
 415         iter->flags     = flags;
 416         iter->snapshot  = pos.snapshot;
 417         iter->pos       = pos;
 418         iter->k.p       = pos;
 419
 420 #ifdef CONFIG_BCACHEFS_DEBUG
 421         iter->ip_allocated = ip;
 422 #endif
 423         iter->path = bch2_path_get(trans, btree_id, iter->pos,
 424                                    locks_want, depth, flags, ip);
 425 }
 426
 427 void bch2_trans_iter_init_outlined(struct btree_trans *, struct btree_iter *,
 428                           unsigned, struct bpos, unsigned);
 429
 430 static inline void bch2_trans_iter_init(struct btree_trans *trans,
 431                           struct btree_iter *iter,
 432                           unsigned btree_id, struct bpos pos,
 433                           unsigned flags)
 434 {
 435         if (__builtin_constant_p(btree_id) &&
 436             __builtin_constant_p(flags))
 437                 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
 438                                 bch2_btree_iter_flags(trans, btree_id, flags),
 439                                 _THIS_IP_);
 440         else
 441                 bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
 442 }
 443
 444 void bch2_trans_node_iter_init(struct btree_trans *, struct btree_iter *,
 445                                enum btree_id, struct bpos,
 446                                unsigned, unsigned, unsigned);
 447 void bch2_trans_copy_iter(struct btree_iter *, struct btree_iter *);
 448
 449 static inline void set_btree_iter_dontneed(struct btree_iter *iter)
 450 {
 451         if (!iter->trans->restarted)
 452                 iter->path->preserve = false;
 453 }
 454
 455 void *__bch2_trans_kmalloc(struct btree_trans *, size_t);
 456
 457 static inline void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
 458 {
 459         size = roundup(size, 8);
 460
 461         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 462                 void *p = trans->mem + trans->mem_top;
 463
 464                 trans->mem_top += size;
 465                 memset(p, 0, size);
 466                 return p;
 467         } else {
 468                 return __bch2_trans_kmalloc(trans, size);
 469         }
 470 }
 471
 472 static inline void *bch2_trans_kmalloc_nomemzero(struct btree_trans *trans, size_t size)
 473 {
 474         size = roundup(size, 8);
 475
 476         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 477                 void *p = trans->mem + trans->mem_top;
 478
 479                 trans->mem_top += size;
 480                 return p;
 481         } else {
 482                 return __bch2_trans_kmalloc(trans, size);
 483         }
 484 }
 485
 486 static inline struct bkey_s_c __bch2_bkey_get_iter(struct btree_trans *trans,
 487                                 struct btree_iter *iter,
 488                                 unsigned btree_id, struct bpos pos,
 489                                 unsigned flags, unsigned type)
 490 {
 491         struct bkey_s_c k;
 492
 493         bch2_trans_iter_init(trans, iter, btree_id, pos, flags);
 494         k = bch2_btree_iter_peek_slot(iter);
 495
 496         if (!bkey_err(k) && type && k.k->type != type)
 497                 k = bkey_s_c_err(-ENOENT);
 498         if (unlikely(bkey_err(k)))
 499                 bch2_trans_iter_exit(trans, iter);
 500         return k;
 501 }
 502
 503 static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 504                                 struct btree_iter *iter,
 505                                 unsigned btree_id, struct bpos pos,
 506                                 unsigned flags)
 507 {
 508         return __bch2_bkey_get_iter(trans, iter, btree_id, pos, flags, 0);
 509 }
 510
 511 #define bch2_bkey_get_iter_typed(_trans, _iter, _btree_id, _pos, _flags, _type)\
 512         bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,                 \
 513                                        _btree_id, _pos, _flags, KEY_TYPE_##_type))
 514
 515 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 516                                 unsigned btree_id, struct bpos pos,
 517                                 unsigned flags, unsigned type,
 518                                 unsigned val_size, void *val)
 519 {
 520         struct btree_iter iter;
 521         struct bkey_s_c k;
 522         int ret;
 523
 524         k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
 525         ret = bkey_err(k);
 526         if (!ret) {
 527                 unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
 528
 529                 memcpy(val, k.v, b);
 530                 if (unlikely(b < sizeof(*val)))
 531                         memset((void *) val + b, 0, sizeof(*val) - b);
 532                 bch2_trans_iter_exit(trans, &iter);
 533         }
 534
 535         return ret;
 536 }
 537
 538 #define bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags, _type, _val)\
 539         __bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags,      \
 540                                   KEY_TYPE_##_type, sizeof(*_val), _val)
 541
 542 u32 bch2_trans_begin(struct btree_trans *);
 543
 544 /*
 545  * XXX
 546  * this does not handle transaction restarts from bch2_btree_iter_next_node()
 547  * correctly
 548  */
 549 #define __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 550                               _locks_want, _depth, _flags, _b, _ret)    \
 551         for (bch2_trans_node_iter_init((_trans), &(_iter), (_btree_id), \
 552                                 _start, _locks_want, _depth, _flags);   \
 553              (_b) = bch2_btree_iter_peek_node_and_restart(&(_iter)),    \
 554              !((_ret) = PTR_ERR_OR_ZERO(_b)) && (_b);                   \
 555              (_b) = bch2_btree_iter_next_node(&(_iter)))
 556
 557 #define for_each_btree_node(_trans, _iter, _btree_id, _start,           \
 558                             _flags, _b, _ret)                           \
 559         __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 560                               0, 0, _flags, _b, _ret)
 561
 562 static inline struct bkey_s_c bch2_btree_iter_peek_prev_type(struct btree_iter *iter,
 563                                                              unsigned flags)
 564 {
 565         BUG_ON(flags & BTREE_ITER_ALL_LEVELS);
 566
 567         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 568                                                 bch2_btree_iter_peek_prev(iter);
 569 }
 570
 571 static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 572                                                         unsigned flags)
 573 {
 574         return  flags & BTREE_ITER_ALL_LEVELS ? bch2_btree_iter_peek_all_levels(iter) :
 575                 flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 576                                                 bch2_btree_iter_peek(iter);
 577 }
 578
 579 static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
 580                                                              struct bpos end,
 581                                                              unsigned flags)
 582 {
 583         if (!(flags & BTREE_ITER_SLOTS))
 584                 return bch2_btree_iter_peek_upto(iter, end);
 585
 586         if (bkey_gt(iter->pos, end))
 587                 return bkey_s_c_null;
 588
 589         return bch2_btree_iter_peek_slot(iter);
 590 }
 591
 592 static inline int btree_trans_too_many_iters(struct btree_trans *trans)
 593 {
 594         if (hweight64(trans->paths_allocated) > BTREE_ITER_MAX - 8) {
 595                 trace_and_count(trans->c, trans_restart_too_many_iters, trans, _THIS_IP_);
 596                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
 597         }
 598
 599         return 0;
 600 }
 601
 602 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 603
 604 static inline struct bkey_s_c
 605 __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 606                                    struct btree_iter *iter, unsigned flags)
 607 {
 608         struct bkey_s_c k;
 609
 610         while (btree_trans_too_many_iters(trans) ||
 611                (k = bch2_btree_iter_peek_type(iter, flags),
 612                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 613                 bch2_trans_begin(trans);
 614
 615         return k;
 616 }
 617
 618 static inline struct bkey_s_c
 619 __bch2_btree_iter_peek_upto_and_restart(struct btree_trans *trans,
 620                                         struct btree_iter *iter,
 621                                         struct bpos end,
 622                                         unsigned flags)
 623 {
 624         struct bkey_s_c k;
 625
 626         while (btree_trans_too_many_iters(trans) ||
 627                (k = bch2_btree_iter_peek_upto_type(iter, end, flags),
 628                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 629                 bch2_trans_begin(trans);
 630
 631         return k;
 632 }
 633
 634 #define lockrestart_do(_trans, _do)                                     \
 635 ({                                                                      \
 636         u32 _restart_count;                                             \
 637         int _ret;                                                       \
 638                                                                         \
 639         do {                                                            \
 640                 _restart_count = bch2_trans_begin(_trans);              \
 641                 _ret = (_do);                                           \
 642         } while (bch2_err_matches(_ret, BCH_ERR_transaction_restart));  \
 643                                                                         \
 644         if (!_ret)                                                      \
 645                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 646                                                                         \
 647         _ret;                                                           \
 648 })
 649
 650 /*
 651  * nested_lockrestart_do(), nested_commit_do():
 652  *
 653  * These are like lockrestart_do() and commit_do(), with two differences:
 654  *
 655  *  - We don't call bch2_trans_begin() unless we had a transaction restart
 656  *  - We return -BCH_ERR_transaction_restart_nested if we succeeded after a
 657  *  transaction restart
 658  */
 659 #define nested_lockrestart_do(_trans, _do)                              \
 660 ({                                                                      \
 661         u32 _restart_count, _orig_restart_count;                        \
 662         int _ret;                                                       \
 663                                                                         \
 664         _restart_count = _orig_restart_count = (_trans)->restart_count; \
 665                                                                         \
 666         while (bch2_err_matches(_ret = (_do), BCH_ERR_transaction_restart))\
 667                 _restart_count = bch2_trans_begin(_trans);              \
 668                                                                         \
 669         if (!_ret)                                                      \
 670                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 671                                                                         \
 672         if (!_ret && trans_was_restarted(_trans, _orig_restart_count))  \
 673                 _ret = -BCH_ERR_transaction_restart_nested;             \
 674                                                                         \
 675         _ret;                                                           \
 676 })
 677
 678 #define for_each_btree_key2(_trans, _iter, _btree_id,                   \
 679                             _start, _flags, _k, _do)                    \
 680 ({                                                                      \
 681         int _ret = 0;                                                   \
 682                                                                         \
 683         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 684                              (_start), (_flags));                       \
 685                                                                         \
 686         while (1) {                                                     \
 687                 u32 _restart_count = bch2_trans_begin(_trans);          \
 688                                                                         \
 689                 _ret = 0;                                               \
 690                 (_k) = bch2_btree_iter_peek_type(&(_iter), (_flags));   \
 691                 if (!(_k).k)                                            \
 692                         break;                                          \
 693                                                                         \
 694                 _ret = bkey_err(_k) ?: (_do);                           \
 695                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 696                         continue;                                       \
 697                 if (_ret)                                               \
 698                         break;                                          \
 699                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 700                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 701                         break;                                          \
 702         }                                                               \
 703                                                                         \
 704         bch2_trans_iter_exit((_trans), &(_iter));                       \
 705         _ret;                                                           \
 706 })
 707
 708 #define for_each_btree_key2_upto(_trans, _iter, _btree_id,              \
 709                             _start, _end, _flags, _k, _do)              \
 710 ({                                                                      \
 711         int _ret = 0;                                                   \
 712                                                                         \
 713         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 714                              (_start), (_flags));                       \
 715                                                                         \
 716         while (1) {                                                     \
 717                 u32 _restart_count = bch2_trans_begin(_trans);          \
 718                                                                         \
 719                 _ret = 0;                                               \
 720                 (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, (_flags));\
 721                 if (!(_k).k)                                            \
 722                         break;                                          \
 723                                                                         \
 724                 _ret = bkey_err(_k) ?: (_do);                           \
 725                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 726                         continue;                                       \
 727                 if (_ret)                                               \
 728                         break;                                          \
 729                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 730                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 731                         break;                                          \
 732         }                                                               \
 733                                                                         \
 734         bch2_trans_iter_exit((_trans), &(_iter));                       \
 735         _ret;                                                           \
 736 })
 737
 738 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,            \
 739                                    _start, _flags, _k, _do)             \
 740 ({                                                                      \
 741         int _ret = 0;                                                   \
 742                                                                         \
 743         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 744                              (_start), (_flags));                       \
 745                                                                         \
 746         while (1) {                                                     \
 747                 u32 _restart_count = bch2_trans_begin(_trans);          \
 748                 (_k) = bch2_btree_iter_peek_prev_type(&(_iter), (_flags));\
 749                 if (!(_k).k) {                                          \
 750                         _ret = 0;                                       \
 751                         break;                                          \
 752                 }                                                       \
 753                                                                         \
 754                 _ret = bkey_err(_k) ?: (_do);                           \
 755                 if (bch2_err_matches(_ret, BCH_ERR_transaction_restart))\
 756                         continue;                                       \
 757                 if (_ret)                                               \
 758                         break;                                          \
 759                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 760                 if (!bch2_btree_iter_rewind(&(_iter)))                  \
 761                         break;                                          \
 762         }                                                               \
 763                                                                         \
 764         bch2_trans_iter_exit((_trans), &(_iter));                       \
 765         _ret;                                                           \
 766 })
 767
 768 #define for_each_btree_key_commit(_trans, _iter, _btree_id,             \
 769                                   _start, _iter_flags, _k,              \
 770                                   _disk_res, _journal_seq, _commit_flags,\
 771                                   _do)                                  \
 772         for_each_btree_key2(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 773                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 774                                         (_journal_seq), (_commit_flags)))
 775
 776 #define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,        \
 777                                   _start, _end, _iter_flags, _k,        \
 778                                   _disk_res, _journal_seq, _commit_flags,\
 779                                   _do)                                  \
 780         for_each_btree_key2_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 781                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 782                                         (_journal_seq), (_commit_flags)))
 783
 784 #define for_each_btree_key(_trans, _iter, _btree_id,                    \
 785                            _start, _flags, _k, _ret)                    \
 786         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 787                                   (_start), (_flags));                  \
 788              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 789              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 790              bch2_btree_iter_advance(&(_iter)))
 791
 792 #define for_each_btree_key_upto(_trans, _iter, _btree_id,               \
 793                                 _start, _end, _flags, _k, _ret)         \
 794         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 795                                   (_start), (_flags));                  \
 796              (_k) = __bch2_btree_iter_peek_upto_and_restart((_trans),   \
 797                                                 &(_iter), _end, _flags),\
 798              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 799              bch2_btree_iter_advance(&(_iter)))
 800
 801 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,          \
 802                            _start, _flags, _k, _ret)                    \
 803         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 804                                   (_start), (_flags));                  \
 805              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 806              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 807              bch2_btree_iter_advance(&(_iter)))
 808
 809 #define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,     \
 810                            _start, _end, _flags, _k, _ret)              \
 811         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 812                                   (_start), (_flags));                  \
 813              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
 814              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 815              bch2_btree_iter_advance(&(_iter)))
 816
 817 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _ret)    \
 818         for (;                                                          \
 819              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 820              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 821              bch2_btree_iter_advance(&(_iter)))
 822
 823 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)  \
 824         for (;                                                          \
 825              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 826              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 827              bch2_btree_iter_advance(&(_iter)))
 828
 829 #define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
 830         for (;                                                                  \
 831              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),     \
 832              !((_ret) = bkey_err(_k)) && (_k).k;                                \
 833              bch2_btree_iter_advance(&(_iter)))
 834
 835 /* new multiple iterator interface: */
 836
 837 void bch2_trans_updates_to_text(struct printbuf *, struct btree_trans *);
 838 void bch2_btree_path_to_text(struct printbuf *, struct btree_path *);
 839 void bch2_trans_paths_to_text(struct printbuf *, struct btree_trans *);
 840 void bch2_dump_trans_updates(struct btree_trans *);
 841 void bch2_dump_trans_paths_updates(struct btree_trans *);
 842 void __bch2_trans_init(struct btree_trans *, struct bch_fs *, unsigned);
 843 void bch2_trans_exit(struct btree_trans *);
 844
 845 extern const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
 846 unsigned bch2_trans_get_fn_idx(const char *);
 847
 848 #define bch2_trans_init(_trans, _c, _nr_iters, _mem)                    \
 849 do {                                                                    \
 850         static unsigned trans_fn_idx;                                   \
 851                                                                         \
 852         if (unlikely(!trans_fn_idx))                                    \
 853                 trans_fn_idx = bch2_trans_get_fn_idx(__func__);         \
 854                                                                         \
 855         __bch2_trans_init(_trans, _c, trans_fn_idx);                    \
 856 } while (0)
 857
 858 void bch2_btree_trans_to_text(struct printbuf *, struct btree_trans *);
 859
 860 void bch2_fs_btree_iter_exit(struct bch_fs *);
 861 int bch2_fs_btree_iter_init(struct bch_fs *);
 862
 863 #endif /* _BCACHEFS_BTREE_ITER_H */