1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "bcachefs.h" 4 #include "bkey_methods.h" 5 #include "bkey_buf.h" 6 #include "btree_cache.h" 7 #include "btree_iter.h" 8 #include "btree_journal_iter.h" 9 #include "btree_key_cache.h" 10 #include "btree_locking.h" 11 #include "btree_update.h" 12 #include "debug.h" 13 #include "error.h" 14 #include "extents.h" 15 #include "journal.h" 16 #include "journal_io.h" 17 #include "replicas.h" 18 #include "snapshot.h" 19 #include "trace.h" 20 21 #include <linux/random.h> 22 #include <linux/prefetch.h> 23 24 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *); 25 static inline void btree_path_list_add(struct btree_trans *, 26 btree_path_idx_t, btree_path_idx_t); 27 28 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter) 29 { 30 #ifdef TRACK_PATH_ALLOCATED 31 return iter->ip_allocated; 32 #else 33 return 0; 34 #endif 35 } 36 37 static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t); 38 static void bch2_trans_srcu_lock(struct btree_trans *); 39 40 static inline int __btree_path_cmp(const struct btree_path *l, 41 enum btree_id r_btree_id, 42 bool r_cached, 43 struct bpos r_pos, 44 unsigned r_level) 45 { 46 /* 47 * Must match lock ordering as defined by __bch2_btree_node_lock: 48 */ 49 return cmp_int(l->btree_id, r_btree_id) ?: 50 cmp_int((int) l->cached, (int) r_cached) ?: 51 bpos_cmp(l->pos, r_pos) ?: 52 -cmp_int(l->level, r_level); 53 } 54 55 static inline int btree_path_cmp(const struct btree_path *l, 56 const struct btree_path *r) 57 { 58 return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level); 59 } 60 61 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p) 62 { 63 /* Are we iterating over keys in all snapshots? */ 64 if (iter->flags & BTREE_ITER_all_snapshots) { 65 p = bpos_successor(p); 66 } else { 67 p = bpos_nosnap_successor(p); 68 p.snapshot = iter->snapshot; 69 } 70 71 return p; 72 } 73 74 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p) 75 { 76 /* Are we iterating over keys in all snapshots? */ 77 if (iter->flags & BTREE_ITER_all_snapshots) { 78 p = bpos_predecessor(p); 79 } else { 80 p = bpos_nosnap_predecessor(p); 81 p.snapshot = iter->snapshot; 82 } 83 84 return p; 85 } 86 87 static inline struct bpos btree_iter_search_key(struct btree_iter *iter) 88 { 89 struct bpos pos = iter->pos; 90 91 if ((iter->flags & BTREE_ITER_is_extents) && 92 !bkey_eq(pos, POS_MAX)) 93 pos = bkey_successor(iter, pos); 94 return pos; 95 } 96 97 static inline bool btree_path_pos_before_node(struct btree_path *path, 98 struct btree *b) 99 { 100 return bpos_lt(path->pos, b->data->min_key); 101 } 102 103 static inline bool btree_path_pos_after_node(struct btree_path *path, 104 struct btree *b) 105 { 106 return bpos_gt(path->pos, b->key.k.p); 107 } 108 109 static inline bool btree_path_pos_in_node(struct btree_path *path, 110 struct btree *b) 111 { 112 return path->btree_id == b->c.btree_id && 113 !btree_path_pos_before_node(path, b) && 114 !btree_path_pos_after_node(path, b); 115 } 116 117 /* Btree iterator: */ 118 119 #ifdef CONFIG_BCACHEFS_DEBUG 120 121 static void bch2_btree_path_verify_cached(struct btree_trans *trans, 122 struct btree_path *path) 123 { 124 struct bkey_cached *ck; 125 bool locked = btree_node_locked(path, 0); 126 127 if (!bch2_btree_node_relock(trans, path, 0)) 128 return; 129 130 ck = (void *) path->l[0].b; 131 BUG_ON(ck->key.btree_id != path->btree_id || 132 !bkey_eq(ck->key.pos, path->pos)); 133 134 if (!locked) 135 btree_node_unlock(trans, path, 0); 136 } 137 138 static void bch2_btree_path_verify_level(struct btree_trans *trans, 139 struct btree_path *path, unsigned level) 140 { 141 struct btree_path_level *l; 142 struct btree_node_iter tmp; 143 bool locked; 144 struct bkey_packed *p, *k; 145 struct printbuf buf1 = PRINTBUF; 146 struct printbuf buf2 = PRINTBUF; 147 struct printbuf buf3 = PRINTBUF; 148 const char *msg; 149 150 if (!bch2_debug_check_iterators) 151 return; 152 153 l = &path->l[level]; 154 tmp = l->iter; 155 locked = btree_node_locked(path, level); 156 157 if (path->cached) { 158 if (!level) 159 bch2_btree_path_verify_cached(trans, path); 160 return; 161 } 162 163 if (!btree_path_node(path, level)) 164 return; 165 166 if (!bch2_btree_node_relock_notrace(trans, path, level)) 167 return; 168 169 BUG_ON(!btree_path_pos_in_node(path, l->b)); 170 171 bch2_btree_node_iter_verify(&l->iter, l->b); 172 173 /* 174 * For interior nodes, the iterator will have skipped past deleted keys: 175 */ 176 p = level 177 ? bch2_btree_node_iter_prev(&tmp, l->b) 178 : bch2_btree_node_iter_prev_all(&tmp, l->b); 179 k = bch2_btree_node_iter_peek_all(&l->iter, l->b); 180 181 if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) { 182 msg = "before"; 183 goto err; 184 } 185 186 if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) { 187 msg = "after"; 188 goto err; 189 } 190 191 if (!locked) 192 btree_node_unlock(trans, path, level); 193 return; 194 err: 195 bch2_bpos_to_text(&buf1, path->pos); 196 197 if (p) { 198 struct bkey uk = bkey_unpack_key(l->b, p); 199 200 bch2_bkey_to_text(&buf2, &uk); 201 } else { 202 prt_printf(&buf2, "(none)"); 203 } 204 205 if (k) { 206 struct bkey uk = bkey_unpack_key(l->b, k); 207 208 bch2_bkey_to_text(&buf3, &uk); 209 } else { 210 prt_printf(&buf3, "(none)"); 211 } 212 213 panic("path should be %s key at level %u:\n" 214 "path pos %s\n" 215 "prev key %s\n" 216 "cur key %s\n", 217 msg, level, buf1.buf, buf2.buf, buf3.buf); 218 } 219 220 static void bch2_btree_path_verify(struct btree_trans *trans, 221 struct btree_path *path) 222 { 223 struct bch_fs *c = trans->c; 224 225 for (unsigned i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) { 226 if (!path->l[i].b) { 227 BUG_ON(!path->cached && 228 bch2_btree_id_root(c, path->btree_id)->b->c.level > i); 229 break; 230 } 231 232 bch2_btree_path_verify_level(trans, path, i); 233 } 234 235 bch2_btree_path_verify_locks(path); 236 } 237 238 void bch2_trans_verify_paths(struct btree_trans *trans) 239 { 240 struct btree_path *path; 241 unsigned iter; 242 243 trans_for_each_path(trans, path, iter) 244 bch2_btree_path_verify(trans, path); 245 } 246 247 static void bch2_btree_iter_verify(struct btree_trans *trans, struct btree_iter *iter) 248 { 249 BUG_ON(!!(iter->flags & BTREE_ITER_cached) != btree_iter_path(trans, iter)->cached); 250 251 BUG_ON((iter->flags & BTREE_ITER_is_extents) && 252 (iter->flags & BTREE_ITER_all_snapshots)); 253 254 BUG_ON(!(iter->flags & BTREE_ITER_snapshot_field) && 255 (iter->flags & BTREE_ITER_all_snapshots) && 256 !btree_type_has_snapshot_field(iter->btree_id)); 257 258 if (iter->update_path) 259 bch2_btree_path_verify(trans, &trans->paths[iter->update_path]); 260 bch2_btree_path_verify(trans, btree_iter_path(trans, iter)); 261 } 262 263 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) 264 { 265 BUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && 266 !iter->pos.snapshot); 267 268 BUG_ON(!(iter->flags & BTREE_ITER_all_snapshots) && 269 iter->pos.snapshot != iter->snapshot); 270 271 BUG_ON(iter->flags & BTREE_ITER_all_snapshots ? !bpos_eq(iter->pos, iter->k.p) : 272 !(iter->flags & BTREE_ITER_is_extents) ? !bkey_eq(iter->pos, iter->k.p) : 273 (bkey_lt(iter->pos, bkey_start_pos(&iter->k)) || 274 bkey_gt(iter->pos, iter->k.p))); 275 } 276 277 static int bch2_btree_iter_verify_ret(struct btree_trans *trans, 278 struct btree_iter *iter, struct bkey_s_c k) 279 { 280 struct btree_iter copy; 281 struct bkey_s_c prev; 282 int ret = 0; 283 284 if (!bch2_debug_check_iterators) 285 return 0; 286 287 if (!(iter->flags & BTREE_ITER_filter_snapshots)) 288 return 0; 289 290 if (bkey_err(k) || !k.k) 291 return 0; 292 293 BUG_ON(!bch2_snapshot_is_ancestor(trans->c, 294 iter->snapshot, 295 k.k->p.snapshot)); 296 297 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos, 298 BTREE_ITER_nopreserve| 299 BTREE_ITER_all_snapshots); 300 prev = bch2_btree_iter_prev(trans, ©); 301 if (!prev.k) 302 goto out; 303 304 ret = bkey_err(prev); 305 if (ret) 306 goto out; 307 308 if (bkey_eq(prev.k->p, k.k->p) && 309 bch2_snapshot_is_ancestor(trans->c, iter->snapshot, 310 prev.k->p.snapshot) > 0) { 311 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; 312 313 bch2_bkey_to_text(&buf1, k.k); 314 bch2_bkey_to_text(&buf2, prev.k); 315 316 panic("iter snap %u\n" 317 "k %s\n" 318 "prev %s\n", 319 iter->snapshot, 320 buf1.buf, buf2.buf); 321 } 322 out: 323 bch2_trans_iter_exit(trans, ©); 324 return ret; 325 } 326 327 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id, 328 struct bpos pos) 329 { 330 bch2_trans_verify_not_unlocked_or_in_restart(trans); 331 332 struct btree_path *path; 333 struct trans_for_each_path_inorder_iter iter; 334 struct printbuf buf = PRINTBUF; 335 336 btree_trans_sort_paths(trans); 337 338 trans_for_each_path_inorder(trans, path, iter) { 339 if (path->btree_id != id || 340 !btree_node_locked(path, 0) || 341 !path->should_be_locked) 342 continue; 343 344 if (!path->cached) { 345 if (bkey_ge(pos, path->l[0].b->data->min_key) && 346 bkey_le(pos, path->l[0].b->key.k.p)) 347 return; 348 } else { 349 if (bkey_eq(pos, path->pos)) 350 return; 351 } 352 } 353 354 bch2_dump_trans_paths_updates(trans); 355 bch2_bpos_to_text(&buf, pos); 356 357 panic("not locked: %s %s\n", bch2_btree_id_str(id), buf.buf); 358 } 359 360 #else 361 362 static inline void bch2_btree_path_verify_level(struct btree_trans *trans, 363 struct btree_path *path, unsigned l) {} 364 static inline void bch2_btree_path_verify(struct btree_trans *trans, 365 struct btree_path *path) {} 366 static inline void bch2_btree_iter_verify(struct btree_trans *trans, 367 struct btree_iter *iter) {} 368 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {} 369 static inline int bch2_btree_iter_verify_ret(struct btree_trans *trans, struct btree_iter *iter, 370 struct bkey_s_c k) { return 0; } 371 372 #endif 373 374 /* Btree path: fixups after btree updates */ 375 376 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter, 377 struct btree *b, 378 struct bset_tree *t, 379 struct bkey_packed *k) 380 { 381 struct btree_node_iter_set *set; 382 383 btree_node_iter_for_each(iter, set) 384 if (set->end == t->end_offset) { 385 set->k = __btree_node_key_to_offset(b, k); 386 bch2_btree_node_iter_sort(iter, b); 387 return; 388 } 389 390 bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t)); 391 } 392 393 static void __bch2_btree_path_fix_key_modified(struct btree_path *path, 394 struct btree *b, 395 struct bkey_packed *where) 396 { 397 struct btree_path_level *l = &path->l[b->c.level]; 398 399 if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b)) 400 return; 401 402 if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0) 403 bch2_btree_node_iter_advance(&l->iter, l->b); 404 } 405 406 void bch2_btree_path_fix_key_modified(struct btree_trans *trans, 407 struct btree *b, 408 struct bkey_packed *where) 409 { 410 struct btree_path *path; 411 unsigned i; 412 413 trans_for_each_path_with_node(trans, b, path, i) { 414 __bch2_btree_path_fix_key_modified(path, b, where); 415 bch2_btree_path_verify_level(trans, path, b->c.level); 416 } 417 } 418 419 static void __bch2_btree_node_iter_fix(struct btree_path *path, 420 struct btree *b, 421 struct btree_node_iter *node_iter, 422 struct bset_tree *t, 423 struct bkey_packed *where, 424 unsigned clobber_u64s, 425 unsigned new_u64s) 426 { 427 const struct bkey_packed *end = btree_bkey_last(b, t); 428 struct btree_node_iter_set *set; 429 unsigned offset = __btree_node_key_to_offset(b, where); 430 int shift = new_u64s - clobber_u64s; 431 unsigned old_end = t->end_offset - shift; 432 unsigned orig_iter_pos = node_iter->data[0].k; 433 bool iter_current_key_modified = 434 orig_iter_pos >= offset && 435 orig_iter_pos <= offset + clobber_u64s; 436 437 btree_node_iter_for_each(node_iter, set) 438 if (set->end == old_end) 439 goto found; 440 441 /* didn't find the bset in the iterator - might have to readd it: */ 442 if (new_u64s && 443 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) { 444 bch2_btree_node_iter_push(node_iter, b, where, end); 445 goto fixup_done; 446 } else { 447 /* Iterator is after key that changed */ 448 return; 449 } 450 found: 451 set->end = t->end_offset; 452 453 /* Iterator hasn't gotten to the key that changed yet: */ 454 if (set->k < offset) 455 return; 456 457 if (new_u64s && 458 bkey_iter_pos_cmp(b, where, &path->pos) >= 0) { 459 set->k = offset; 460 } else if (set->k < offset + clobber_u64s) { 461 set->k = offset + new_u64s; 462 if (set->k == set->end) 463 bch2_btree_node_iter_set_drop(node_iter, set); 464 } else { 465 /* Iterator is after key that changed */ 466 set->k = (int) set->k + shift; 467 return; 468 } 469 470 bch2_btree_node_iter_sort(node_iter, b); 471 fixup_done: 472 if (node_iter->data[0].k != orig_iter_pos) 473 iter_current_key_modified = true; 474 475 /* 476 * When a new key is added, and the node iterator now points to that 477 * key, the iterator might have skipped past deleted keys that should 478 * come after the key the iterator now points to. We have to rewind to 479 * before those deleted keys - otherwise 480 * bch2_btree_node_iter_prev_all() breaks: 481 */ 482 if (!bch2_btree_node_iter_end(node_iter) && 483 iter_current_key_modified && 484 b->c.level) { 485 struct bkey_packed *k, *k2, *p; 486 487 k = bch2_btree_node_iter_peek_all(node_iter, b); 488 489 for_each_bset(b, t) { 490 bool set_pos = false; 491 492 if (node_iter->data[0].end == t->end_offset) 493 continue; 494 495 k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t); 496 497 while ((p = bch2_bkey_prev_all(b, t, k2)) && 498 bkey_iter_cmp(b, k, p) < 0) { 499 k2 = p; 500 set_pos = true; 501 } 502 503 if (set_pos) 504 btree_node_iter_set_set_pos(node_iter, 505 b, t, k2); 506 } 507 } 508 } 509 510 void bch2_btree_node_iter_fix(struct btree_trans *trans, 511 struct btree_path *path, 512 struct btree *b, 513 struct btree_node_iter *node_iter, 514 struct bkey_packed *where, 515 unsigned clobber_u64s, 516 unsigned new_u64s) 517 { 518 struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where); 519 struct btree_path *linked; 520 unsigned i; 521 522 if (node_iter != &path->l[b->c.level].iter) { 523 __bch2_btree_node_iter_fix(path, b, node_iter, t, 524 where, clobber_u64s, new_u64s); 525 526 if (bch2_debug_check_iterators) 527 bch2_btree_node_iter_verify(node_iter, b); 528 } 529 530 trans_for_each_path_with_node(trans, b, linked, i) { 531 __bch2_btree_node_iter_fix(linked, b, 532 &linked->l[b->c.level].iter, t, 533 where, clobber_u64s, new_u64s); 534 bch2_btree_path_verify_level(trans, linked, b->c.level); 535 } 536 } 537 538 /* Btree path level: pointer to a particular btree node and node iter */ 539 540 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c, 541 struct btree_path_level *l, 542 struct bkey *u, 543 struct bkey_packed *k) 544 { 545 if (unlikely(!k)) { 546 /* 547 * signal to bch2_btree_iter_peek_slot() that we're currently at 548 * a hole 549 */ 550 u->type = KEY_TYPE_deleted; 551 return bkey_s_c_null; 552 } 553 554 return bkey_disassemble(l->b, k, u); 555 } 556 557 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c, 558 struct btree_path_level *l, 559 struct bkey *u) 560 { 561 return __btree_iter_unpack(c, l, u, 562 bch2_btree_node_iter_peek_all(&l->iter, l->b)); 563 } 564 565 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans, 566 struct btree_path *path, 567 struct btree_path_level *l, 568 struct bkey *u) 569 { 570 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u, 571 bch2_btree_node_iter_prev(&l->iter, l->b)); 572 573 path->pos = k.k ? k.k->p : l->b->data->min_key; 574 trans->paths_sorted = false; 575 bch2_btree_path_verify_level(trans, path, l - path->l); 576 return k; 577 } 578 579 static inline bool btree_path_advance_to_pos(struct btree_path *path, 580 struct btree_path_level *l, 581 int max_advance) 582 { 583 struct bkey_packed *k; 584 int nr_advanced = 0; 585 586 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) && 587 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) { 588 if (max_advance > 0 && nr_advanced >= max_advance) 589 return false; 590 591 bch2_btree_node_iter_advance(&l->iter, l->b); 592 nr_advanced++; 593 } 594 595 return true; 596 } 597 598 static inline void __btree_path_level_init(struct btree_path *path, 599 unsigned level) 600 { 601 struct btree_path_level *l = &path->l[level]; 602 603 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos); 604 605 /* 606 * Iterators to interior nodes should always be pointed at the first non 607 * whiteout: 608 */ 609 if (level) 610 bch2_btree_node_iter_peek(&l->iter, l->b); 611 } 612 613 void bch2_btree_path_level_init(struct btree_trans *trans, 614 struct btree_path *path, 615 struct btree *b) 616 { 617 BUG_ON(path->cached); 618 619 EBUG_ON(!btree_path_pos_in_node(path, b)); 620 621 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock); 622 path->l[b->c.level].b = b; 623 __btree_path_level_init(path, b->c.level); 624 } 625 626 /* Btree path: fixups after btree node updates: */ 627 628 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b) 629 { 630 struct bch_fs *c = trans->c; 631 632 trans_for_each_update(trans, i) 633 if (!i->cached && 634 i->level == b->c.level && 635 i->btree_id == b->c.btree_id && 636 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 && 637 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) { 638 i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v; 639 640 if (unlikely(trans->journal_replay_not_finished)) { 641 struct bkey_i *j_k = 642 bch2_journal_keys_peek_slot(c, i->btree_id, i->level, 643 i->k->k.p); 644 645 if (j_k) { 646 i->old_k = j_k->k; 647 i->old_v = &j_k->v; 648 } 649 } 650 } 651 } 652 653 /* 654 * A btree node is being replaced - update the iterator to point to the new 655 * node: 656 */ 657 void bch2_trans_node_add(struct btree_trans *trans, 658 struct btree_path *path, 659 struct btree *b) 660 { 661 struct btree_path *prev; 662 663 BUG_ON(!btree_path_pos_in_node(path, b)); 664 665 while ((prev = prev_btree_path(trans, path)) && 666 btree_path_pos_in_node(prev, b)) 667 path = prev; 668 669 for (; 670 path && btree_path_pos_in_node(path, b); 671 path = next_btree_path(trans, path)) 672 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) { 673 enum btree_node_locked_type t = 674 btree_lock_want(path, b->c.level); 675 676 if (t != BTREE_NODE_UNLOCKED) { 677 btree_node_unlock(trans, path, b->c.level); 678 six_lock_increment(&b->c.lock, (enum six_lock_type) t); 679 mark_btree_node_locked(trans, path, b->c.level, t); 680 } 681 682 bch2_btree_path_level_init(trans, path, b); 683 } 684 685 bch2_trans_revalidate_updates_in_node(trans, b); 686 } 687 688 void bch2_trans_node_drop(struct btree_trans *trans, 689 struct btree *b) 690 { 691 struct btree_path *path; 692 unsigned i, level = b->c.level; 693 694 trans_for_each_path(trans, path, i) 695 if (path->l[level].b == b) { 696 btree_node_unlock(trans, path, level); 697 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 698 } 699 } 700 701 /* 702 * A btree node has been modified in such a way as to invalidate iterators - fix 703 * them: 704 */ 705 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b) 706 { 707 struct btree_path *path; 708 unsigned i; 709 710 trans_for_each_path_with_node(trans, b, path, i) 711 __btree_path_level_init(path, b->c.level); 712 713 bch2_trans_revalidate_updates_in_node(trans, b); 714 } 715 716 /* Btree path: traverse, set_pos: */ 717 718 static inline int btree_path_lock_root(struct btree_trans *trans, 719 struct btree_path *path, 720 unsigned depth_want, 721 unsigned long trace_ip) 722 { 723 struct bch_fs *c = trans->c; 724 struct btree_root *r = bch2_btree_id_root(c, path->btree_id); 725 enum six_lock_type lock_type; 726 unsigned i; 727 int ret; 728 729 EBUG_ON(path->nodes_locked); 730 731 while (1) { 732 struct btree *b = READ_ONCE(r->b); 733 if (unlikely(!b)) { 734 BUG_ON(!r->error); 735 return r->error; 736 } 737 738 path->level = READ_ONCE(b->c.level); 739 740 if (unlikely(path->level < depth_want)) { 741 /* 742 * the root is at a lower depth than the depth we want: 743 * got to the end of the btree, or we're walking nodes 744 * greater than some depth and there are no nodes >= 745 * that depth 746 */ 747 path->level = depth_want; 748 for (i = path->level; i < BTREE_MAX_DEPTH; i++) 749 path->l[i].b = NULL; 750 return 1; 751 } 752 753 lock_type = __btree_lock_want(path, path->level); 754 ret = btree_node_lock(trans, path, &b->c, 755 path->level, lock_type, trace_ip); 756 if (unlikely(ret)) { 757 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 758 return ret; 759 BUG(); 760 } 761 762 if (likely(b == READ_ONCE(r->b) && 763 b->c.level == path->level && 764 !race_fault())) { 765 for (i = 0; i < path->level; i++) 766 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root); 767 path->l[path->level].b = b; 768 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++) 769 path->l[i].b = NULL; 770 771 mark_btree_node_locked(trans, path, path->level, 772 (enum btree_node_locked_type) lock_type); 773 bch2_btree_path_level_init(trans, path, b); 774 return 0; 775 } 776 777 six_unlock_type(&b->c.lock, lock_type); 778 } 779 } 780 781 noinline 782 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path) 783 { 784 struct bch_fs *c = trans->c; 785 struct btree_path_level *l = path_l(path); 786 struct btree_node_iter node_iter = l->iter; 787 struct bkey_packed *k; 788 struct bkey_buf tmp; 789 unsigned nr = test_bit(BCH_FS_started, &c->flags) 790 ? (path->level > 1 ? 0 : 2) 791 : (path->level > 1 ? 1 : 16); 792 bool was_locked = btree_node_locked(path, path->level); 793 int ret = 0; 794 795 bch2_bkey_buf_init(&tmp); 796 797 while (nr-- && !ret) { 798 if (!bch2_btree_node_relock(trans, path, path->level)) 799 break; 800 801 bch2_btree_node_iter_advance(&node_iter, l->b); 802 k = bch2_btree_node_iter_peek(&node_iter, l->b); 803 if (!k) 804 break; 805 806 bch2_bkey_buf_unpack(&tmp, c, l->b, k); 807 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, 808 path->level - 1); 809 } 810 811 if (!was_locked) 812 btree_node_unlock(trans, path, path->level); 813 814 bch2_bkey_buf_exit(&tmp, c); 815 return ret; 816 } 817 818 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path, 819 struct btree_and_journal_iter *jiter) 820 { 821 struct bch_fs *c = trans->c; 822 struct bkey_s_c k; 823 struct bkey_buf tmp; 824 unsigned nr = test_bit(BCH_FS_started, &c->flags) 825 ? (path->level > 1 ? 0 : 2) 826 : (path->level > 1 ? 1 : 16); 827 bool was_locked = btree_node_locked(path, path->level); 828 int ret = 0; 829 830 bch2_bkey_buf_init(&tmp); 831 832 jiter->fail_if_too_many_whiteouts = true; 833 834 while (nr-- && !ret) { 835 if (!bch2_btree_node_relock(trans, path, path->level)) 836 break; 837 838 bch2_btree_and_journal_iter_advance(jiter); 839 k = bch2_btree_and_journal_iter_peek(jiter); 840 if (!k.k) 841 break; 842 843 bch2_bkey_buf_reassemble(&tmp, c, k); 844 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, 845 path->level - 1); 846 } 847 848 if (!was_locked) 849 btree_node_unlock(trans, path, path->level); 850 851 bch2_bkey_buf_exit(&tmp, c); 852 return ret; 853 } 854 855 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans, 856 struct btree_path *path, 857 unsigned plevel, struct btree *b) 858 { 859 struct btree_path_level *l = &path->l[plevel]; 860 bool locked = btree_node_locked(path, plevel); 861 struct bkey_packed *k; 862 struct bch_btree_ptr_v2 *bp; 863 864 if (!bch2_btree_node_relock(trans, path, plevel)) 865 return; 866 867 k = bch2_btree_node_iter_peek_all(&l->iter, l->b); 868 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2); 869 870 bp = (void *) bkeyp_val(&l->b->format, k); 871 bp->mem_ptr = (unsigned long)b; 872 873 if (!locked) 874 btree_node_unlock(trans, path, plevel); 875 } 876 877 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans, 878 struct btree_path *path, 879 unsigned flags, 880 struct bkey_buf *out) 881 { 882 struct bch_fs *c = trans->c; 883 struct btree_path_level *l = path_l(path); 884 struct btree_and_journal_iter jiter; 885 struct bkey_s_c k; 886 int ret = 0; 887 888 __bch2_btree_and_journal_iter_init_node_iter(trans, &jiter, l->b, l->iter, path->pos); 889 890 k = bch2_btree_and_journal_iter_peek(&jiter); 891 if (!k.k) { 892 struct printbuf buf = PRINTBUF; 893 894 prt_str(&buf, "node not found at pos "); 895 bch2_bpos_to_text(&buf, path->pos); 896 prt_str(&buf, " at btree "); 897 bch2_btree_pos_to_text(&buf, c, l->b); 898 899 ret = bch2_fs_topology_error(c, "%s", buf.buf); 900 printbuf_exit(&buf); 901 goto err; 902 } 903 904 bch2_bkey_buf_reassemble(out, c, k); 905 906 if ((flags & BTREE_ITER_prefetch) && 907 c->opts.btree_node_prefetch) 908 ret = btree_path_prefetch_j(trans, path, &jiter); 909 910 err: 911 bch2_btree_and_journal_iter_exit(&jiter); 912 return ret; 913 } 914 915 static __always_inline int btree_path_down(struct btree_trans *trans, 916 struct btree_path *path, 917 unsigned flags, 918 unsigned long trace_ip) 919 { 920 struct bch_fs *c = trans->c; 921 struct btree_path_level *l = path_l(path); 922 struct btree *b; 923 unsigned level = path->level - 1; 924 enum six_lock_type lock_type = __btree_lock_want(path, level); 925 struct bkey_buf tmp; 926 int ret; 927 928 EBUG_ON(!btree_node_locked(path, path->level)); 929 930 bch2_bkey_buf_init(&tmp); 931 932 if (unlikely(trans->journal_replay_not_finished)) { 933 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp); 934 if (ret) 935 goto err; 936 } else { 937 struct bkey_packed *k = bch2_btree_node_iter_peek(&l->iter, l->b); 938 if (!k) { 939 struct printbuf buf = PRINTBUF; 940 941 prt_str(&buf, "node not found at pos "); 942 bch2_bpos_to_text(&buf, path->pos); 943 prt_str(&buf, " within parent node "); 944 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&l->b->key)); 945 946 bch2_fs_fatal_error(c, "%s", buf.buf); 947 printbuf_exit(&buf); 948 ret = -BCH_ERR_btree_need_topology_repair; 949 goto err; 950 } 951 952 bch2_bkey_buf_unpack(&tmp, c, l->b, k); 953 954 if ((flags & BTREE_ITER_prefetch) && 955 c->opts.btree_node_prefetch) { 956 ret = btree_path_prefetch(trans, path); 957 if (ret) 958 goto err; 959 } 960 } 961 962 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip); 963 ret = PTR_ERR_OR_ZERO(b); 964 if (unlikely(ret)) 965 goto err; 966 967 if (likely(!trans->journal_replay_not_finished && 968 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) && 969 unlikely(b != btree_node_mem_ptr(tmp.k))) 970 btree_node_mem_ptr_set(trans, path, level + 1, b); 971 972 if (btree_node_read_locked(path, level + 1)) 973 btree_node_unlock(trans, path, level + 1); 974 975 mark_btree_node_locked(trans, path, level, 976 (enum btree_node_locked_type) lock_type); 977 path->level = level; 978 bch2_btree_path_level_init(trans, path, b); 979 980 bch2_btree_path_verify_locks(path); 981 err: 982 bch2_bkey_buf_exit(&tmp, c); 983 return ret; 984 } 985 986 static int bch2_btree_path_traverse_all(struct btree_trans *trans) 987 { 988 struct bch_fs *c = trans->c; 989 struct btree_path *path; 990 unsigned long trace_ip = _RET_IP_; 991 unsigned i; 992 int ret = 0; 993 994 if (trans->in_traverse_all) 995 return -BCH_ERR_transaction_restart_in_traverse_all; 996 997 trans->in_traverse_all = true; 998 retry_all: 999 trans->restarted = 0; 1000 trans->last_restarted_ip = 0; 1001 1002 trans_for_each_path(trans, path, i) 1003 path->should_be_locked = false; 1004 1005 btree_trans_sort_paths(trans); 1006 1007 bch2_trans_unlock(trans); 1008 cond_resched(); 1009 trans_set_locked(trans, false); 1010 1011 if (unlikely(trans->memory_allocation_failure)) { 1012 struct closure cl; 1013 1014 closure_init_stack(&cl); 1015 1016 do { 1017 ret = bch2_btree_cache_cannibalize_lock(trans, &cl); 1018 closure_sync(&cl); 1019 } while (ret); 1020 } 1021 1022 /* Now, redo traversals in correct order: */ 1023 i = 0; 1024 while (i < trans->nr_sorted) { 1025 btree_path_idx_t idx = trans->sorted[i]; 1026 1027 /* 1028 * Traversing a path can cause another path to be added at about 1029 * the same position: 1030 */ 1031 if (trans->paths[idx].uptodate) { 1032 __btree_path_get(trans, &trans->paths[idx], false); 1033 ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_); 1034 __btree_path_put(trans, &trans->paths[idx], false); 1035 1036 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || 1037 bch2_err_matches(ret, ENOMEM)) 1038 goto retry_all; 1039 if (ret) 1040 goto err; 1041 } else { 1042 i++; 1043 } 1044 } 1045 1046 /* 1047 * We used to assert that all paths had been traversed here 1048 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since 1049 * path->should_be_locked is not set yet, we might have unlocked and 1050 * then failed to relock a path - that's fine. 1051 */ 1052 err: 1053 bch2_btree_cache_cannibalize_unlock(trans); 1054 1055 trans->in_traverse_all = false; 1056 1057 trace_and_count(c, trans_traverse_all, trans, trace_ip); 1058 return ret; 1059 } 1060 1061 static inline bool btree_path_check_pos_in_node(struct btree_path *path, 1062 unsigned l, int check_pos) 1063 { 1064 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b)) 1065 return false; 1066 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b)) 1067 return false; 1068 return true; 1069 } 1070 1071 static inline bool btree_path_good_node(struct btree_trans *trans, 1072 struct btree_path *path, 1073 unsigned l, int check_pos) 1074 { 1075 return is_btree_node(path, l) && 1076 bch2_btree_node_relock(trans, path, l) && 1077 btree_path_check_pos_in_node(path, l, check_pos); 1078 } 1079 1080 static void btree_path_set_level_down(struct btree_trans *trans, 1081 struct btree_path *path, 1082 unsigned new_level) 1083 { 1084 unsigned l; 1085 1086 path->level = new_level; 1087 1088 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++) 1089 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED) 1090 btree_node_unlock(trans, path, l); 1091 1092 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1093 bch2_btree_path_verify(trans, path); 1094 } 1095 1096 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans, 1097 struct btree_path *path, 1098 int check_pos) 1099 { 1100 unsigned i, l = path->level; 1101 again: 1102 while (btree_path_node(path, l) && 1103 !btree_path_good_node(trans, path, l, check_pos)) 1104 __btree_path_set_level_up(trans, path, l++); 1105 1106 /* If we need intent locks, take them too: */ 1107 for (i = l + 1; 1108 i < path->locks_want && btree_path_node(path, i); 1109 i++) 1110 if (!bch2_btree_node_relock(trans, path, i)) { 1111 while (l <= i) 1112 __btree_path_set_level_up(trans, path, l++); 1113 goto again; 1114 } 1115 1116 return l; 1117 } 1118 1119 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans, 1120 struct btree_path *path, 1121 int check_pos) 1122 { 1123 return likely(btree_node_locked(path, path->level) && 1124 btree_path_check_pos_in_node(path, path->level, check_pos)) 1125 ? path->level 1126 : __btree_path_up_until_good_node(trans, path, check_pos); 1127 } 1128 1129 /* 1130 * This is the main state machine for walking down the btree - walks down to a 1131 * specified depth 1132 * 1133 * Returns 0 on success, -EIO on error (error reading in a btree node). 1134 * 1135 * On error, caller (peek_node()/peek_key()) must return NULL; the error is 1136 * stashed in the iterator and returned from bch2_trans_exit(). 1137 */ 1138 int bch2_btree_path_traverse_one(struct btree_trans *trans, 1139 btree_path_idx_t path_idx, 1140 unsigned flags, 1141 unsigned long trace_ip) 1142 { 1143 struct btree_path *path = &trans->paths[path_idx]; 1144 unsigned depth_want = path->level; 1145 int ret = -((int) trans->restarted); 1146 1147 if (unlikely(ret)) 1148 goto out; 1149 1150 if (unlikely(!trans->srcu_held)) 1151 bch2_trans_srcu_lock(trans); 1152 1153 trace_btree_path_traverse_start(trans, path); 1154 1155 /* 1156 * Ensure we obey path->should_be_locked: if it's set, we can't unlock 1157 * and re-traverse the path without a transaction restart: 1158 */ 1159 if (path->should_be_locked) { 1160 ret = bch2_btree_path_relock(trans, path, trace_ip); 1161 goto out; 1162 } 1163 1164 if (path->cached) { 1165 ret = bch2_btree_path_traverse_cached(trans, path, flags); 1166 goto out; 1167 } 1168 1169 path = &trans->paths[path_idx]; 1170 1171 if (unlikely(path->level >= BTREE_MAX_DEPTH)) 1172 goto out_uptodate; 1173 1174 path->level = btree_path_up_until_good_node(trans, path, 0); 1175 unsigned max_level = path->level; 1176 1177 EBUG_ON(btree_path_node(path, path->level) && 1178 !btree_node_locked(path, path->level)); 1179 1180 /* 1181 * Note: path->nodes[path->level] may be temporarily NULL here - that 1182 * would indicate to other code that we got to the end of the btree, 1183 * here it indicates that relocking the root failed - it's critical that 1184 * btree_path_lock_root() comes next and that it can't fail 1185 */ 1186 while (path->level > depth_want) { 1187 ret = btree_path_node(path, path->level) 1188 ? btree_path_down(trans, path, flags, trace_ip) 1189 : btree_path_lock_root(trans, path, depth_want, trace_ip); 1190 if (unlikely(ret)) { 1191 if (ret == 1) { 1192 /* 1193 * No nodes at this level - got to the end of 1194 * the btree: 1195 */ 1196 ret = 0; 1197 goto out; 1198 } 1199 1200 __bch2_btree_path_unlock(trans, path); 1201 path->level = depth_want; 1202 path->l[path->level].b = ERR_PTR(ret); 1203 goto out; 1204 } 1205 } 1206 1207 if (unlikely(max_level > path->level)) { 1208 struct btree_path *linked; 1209 unsigned iter; 1210 1211 trans_for_each_path_with_node(trans, path_l(path)->b, linked, iter) 1212 for (unsigned j = path->level + 1; j < max_level; j++) 1213 linked->l[j] = path->l[j]; 1214 } 1215 1216 out_uptodate: 1217 path->uptodate = BTREE_ITER_UPTODATE; 1218 trace_btree_path_traverse_end(trans, path); 1219 out: 1220 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted) 1221 panic("ret %s (%i) trans->restarted %s (%i)\n", 1222 bch2_err_str(ret), ret, 1223 bch2_err_str(trans->restarted), trans->restarted); 1224 bch2_btree_path_verify(trans, path); 1225 return ret; 1226 } 1227 1228 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst, 1229 struct btree_path *src) 1230 { 1231 unsigned i, offset = offsetof(struct btree_path, pos); 1232 1233 memcpy((void *) dst + offset, 1234 (void *) src + offset, 1235 sizeof(struct btree_path) - offset); 1236 1237 for (i = 0; i < BTREE_MAX_DEPTH; i++) { 1238 unsigned t = btree_node_locked_type(dst, i); 1239 1240 if (t != BTREE_NODE_UNLOCKED) 1241 six_lock_increment(&dst->l[i].b->c.lock, t); 1242 } 1243 } 1244 1245 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src, 1246 bool intent, unsigned long ip) 1247 { 1248 btree_path_idx_t new = btree_path_alloc(trans, src); 1249 btree_path_copy(trans, trans->paths + new, trans->paths + src); 1250 __btree_path_get(trans, trans->paths + new, intent); 1251 #ifdef TRACK_PATH_ALLOCATED 1252 trans->paths[new].ip_allocated = ip; 1253 #endif 1254 return new; 1255 } 1256 1257 __flatten 1258 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans, 1259 btree_path_idx_t path, bool intent, unsigned long ip) 1260 { 1261 struct btree_path *old = trans->paths + path; 1262 __btree_path_put(trans, trans->paths + path, intent); 1263 path = btree_path_clone(trans, path, intent, ip); 1264 trace_btree_path_clone(trans, old, trans->paths + path); 1265 trans->paths[path].preserve = false; 1266 return path; 1267 } 1268 1269 btree_path_idx_t __must_check 1270 __bch2_btree_path_set_pos(struct btree_trans *trans, 1271 btree_path_idx_t path_idx, struct bpos new_pos, 1272 bool intent, unsigned long ip) 1273 { 1274 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos); 1275 1276 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1277 EBUG_ON(!trans->paths[path_idx].ref); 1278 1279 trace_btree_path_set_pos(trans, trans->paths + path_idx, &new_pos); 1280 1281 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip); 1282 1283 struct btree_path *path = trans->paths + path_idx; 1284 path->pos = new_pos; 1285 trans->paths_sorted = false; 1286 1287 if (unlikely(path->cached)) { 1288 btree_node_unlock(trans, path, 0); 1289 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up); 1290 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1291 goto out; 1292 } 1293 1294 unsigned level = btree_path_up_until_good_node(trans, path, cmp); 1295 1296 if (btree_path_node(path, level)) { 1297 struct btree_path_level *l = &path->l[level]; 1298 1299 BUG_ON(!btree_node_locked(path, level)); 1300 /* 1301 * We might have to skip over many keys, or just a few: try 1302 * advancing the node iterator, and if we have to skip over too 1303 * many keys just reinit it (or if we're rewinding, since that 1304 * is expensive). 1305 */ 1306 if (cmp < 0 || 1307 !btree_path_advance_to_pos(path, l, 8)) 1308 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos); 1309 1310 /* 1311 * Iterators to interior nodes should always be pointed at the first non 1312 * whiteout: 1313 */ 1314 if (unlikely(level)) 1315 bch2_btree_node_iter_peek(&l->iter, l->b); 1316 } 1317 1318 if (unlikely(level != path->level)) { 1319 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1320 __bch2_btree_path_unlock(trans, path); 1321 } 1322 out: 1323 bch2_btree_path_verify(trans, path); 1324 return path_idx; 1325 } 1326 1327 /* Btree path: main interface: */ 1328 1329 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path) 1330 { 1331 struct btree_path *sib; 1332 1333 sib = prev_btree_path(trans, path); 1334 if (sib && !btree_path_cmp(sib, path)) 1335 return sib; 1336 1337 sib = next_btree_path(trans, path); 1338 if (sib && !btree_path_cmp(sib, path)) 1339 return sib; 1340 1341 return NULL; 1342 } 1343 1344 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path) 1345 { 1346 struct btree_path *sib; 1347 1348 sib = prev_btree_path(trans, path); 1349 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) 1350 return sib; 1351 1352 sib = next_btree_path(trans, path); 1353 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) 1354 return sib; 1355 1356 return NULL; 1357 } 1358 1359 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path) 1360 { 1361 __bch2_btree_path_unlock(trans, trans->paths + path); 1362 btree_path_list_remove(trans, trans->paths + path); 1363 __clear_bit(path, trans->paths_allocated); 1364 } 1365 1366 static bool bch2_btree_path_can_relock(struct btree_trans *trans, struct btree_path *path) 1367 { 1368 unsigned l = path->level; 1369 1370 do { 1371 if (!btree_path_node(path, l)) 1372 break; 1373 1374 if (!is_btree_node(path, l)) 1375 return false; 1376 1377 if (path->l[l].lock_seq != path->l[l].b->c.lock.seq) 1378 return false; 1379 1380 l++; 1381 } while (l < path->locks_want); 1382 1383 return true; 1384 } 1385 1386 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent) 1387 { 1388 struct btree_path *path = trans->paths + path_idx, *dup; 1389 1390 if (!__btree_path_put(trans, path, intent)) 1391 return; 1392 1393 dup = path->preserve 1394 ? have_path_at_pos(trans, path) 1395 : have_node_at_pos(trans, path); 1396 1397 trace_btree_path_free(trans, path_idx, dup); 1398 1399 if (!dup && !(!path->preserve && !is_btree_node(path, path->level))) 1400 return; 1401 1402 if (path->should_be_locked && !trans->restarted) { 1403 if (!dup) 1404 return; 1405 1406 if (!(trans->locked 1407 ? bch2_btree_path_relock_norestart(trans, dup) 1408 : bch2_btree_path_can_relock(trans, dup))) 1409 return; 1410 } 1411 1412 if (dup) { 1413 dup->preserve |= path->preserve; 1414 dup->should_be_locked |= path->should_be_locked; 1415 } 1416 1417 __bch2_path_free(trans, path_idx); 1418 } 1419 1420 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path, 1421 bool intent) 1422 { 1423 if (!__btree_path_put(trans, trans->paths + path, intent)) 1424 return; 1425 1426 __bch2_path_free(trans, path); 1427 } 1428 1429 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count) 1430 { 1431 panic("trans->restart_count %u, should be %u, last restarted by %pS\n", 1432 trans->restart_count, restart_count, 1433 (void *) trans->last_begin_ip); 1434 } 1435 1436 static void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans) 1437 { 1438 #ifdef CONFIG_BCACHEFS_DEBUG 1439 struct printbuf buf = PRINTBUF; 1440 bch2_prt_backtrace(&buf, &trans->last_restarted_trace); 1441 panic("in transaction restart: %s, last restarted by\n%s", 1442 bch2_err_str(trans->restarted), 1443 buf.buf); 1444 #else 1445 panic("in transaction restart: %s, last restarted by %pS\n", 1446 bch2_err_str(trans->restarted), 1447 (void *) trans->last_restarted_ip); 1448 #endif 1449 } 1450 1451 void __noreturn bch2_trans_unlocked_or_in_restart_error(struct btree_trans *trans) 1452 { 1453 if (trans->restarted) 1454 bch2_trans_in_restart_error(trans); 1455 1456 if (!trans->locked) 1457 panic("trans should be locked, unlocked by %pS\n", 1458 (void *) trans->last_unlock_ip); 1459 1460 BUG(); 1461 } 1462 1463 noinline __cold 1464 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans) 1465 { 1466 prt_printf(buf, "%u transaction updates for %s journal seq %llu\n", 1467 trans->nr_updates, trans->fn, trans->journal_res.seq); 1468 printbuf_indent_add(buf, 2); 1469 1470 trans_for_each_update(trans, i) { 1471 struct bkey_s_c old = { &i->old_k, i->old_v }; 1472 1473 prt_str(buf, "update: btree="); 1474 bch2_btree_id_to_text(buf, i->btree_id); 1475 prt_printf(buf, " cached=%u %pS\n", 1476 i->cached, 1477 (void *) i->ip_allocated); 1478 1479 prt_printf(buf, " old "); 1480 bch2_bkey_val_to_text(buf, trans->c, old); 1481 prt_newline(buf); 1482 1483 prt_printf(buf, " new "); 1484 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k)); 1485 prt_newline(buf); 1486 } 1487 1488 for (struct jset_entry *e = trans->journal_entries; 1489 e != btree_trans_journal_entries_top(trans); 1490 e = vstruct_next(e)) { 1491 bch2_journal_entry_to_text(buf, trans->c, e); 1492 prt_newline(buf); 1493 } 1494 1495 printbuf_indent_sub(buf, 2); 1496 } 1497 1498 static void bch2_btree_path_to_text_short(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx) 1499 { 1500 struct btree_path *path = trans->paths + path_idx; 1501 1502 prt_printf(out, "path: idx %3u ref %u:%u %c %c %c ", 1503 path_idx, path->ref, path->intent_ref, 1504 path->preserve ? 'P' : ' ', 1505 path->should_be_locked ? 'S' : ' ', 1506 path->cached ? 'C' : 'B'); 1507 bch2_btree_id_level_to_text(out, path->btree_id, path->level); 1508 prt_str(out, " pos "); 1509 bch2_bpos_to_text(out, path->pos); 1510 1511 if (!path->cached && btree_node_locked(path, path->level)) { 1512 prt_char(out, ' '); 1513 struct btree *b = path_l(path)->b; 1514 bch2_bpos_to_text(out, b->data->min_key); 1515 prt_char(out, '-'); 1516 bch2_bpos_to_text(out, b->key.k.p); 1517 } 1518 1519 #ifdef TRACK_PATH_ALLOCATED 1520 prt_printf(out, " %pS", (void *) path->ip_allocated); 1521 #endif 1522 } 1523 1524 static const char *btree_node_locked_str(enum btree_node_locked_type t) 1525 { 1526 switch (t) { 1527 case BTREE_NODE_UNLOCKED: 1528 return "unlocked"; 1529 case BTREE_NODE_READ_LOCKED: 1530 return "read"; 1531 case BTREE_NODE_INTENT_LOCKED: 1532 return "intent"; 1533 case BTREE_NODE_WRITE_LOCKED: 1534 return "write"; 1535 default: 1536 return NULL; 1537 } 1538 } 1539 1540 void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx) 1541 { 1542 bch2_btree_path_to_text_short(out, trans, path_idx); 1543 1544 struct btree_path *path = trans->paths + path_idx; 1545 1546 prt_printf(out, " uptodate %u locks_want %u", path->uptodate, path->locks_want); 1547 prt_newline(out); 1548 1549 printbuf_indent_add(out, 2); 1550 for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) { 1551 prt_printf(out, "l=%u locks %s seq %u node ", l, 1552 btree_node_locked_str(btree_node_locked_type(path, l)), 1553 path->l[l].lock_seq); 1554 1555 int ret = PTR_ERR_OR_ZERO(path->l[l].b); 1556 if (ret) 1557 prt_str(out, bch2_err_str(ret)); 1558 else 1559 prt_printf(out, "%px", path->l[l].b); 1560 prt_newline(out); 1561 } 1562 printbuf_indent_sub(out, 2); 1563 } 1564 1565 static noinline __cold 1566 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans, 1567 bool nosort) 1568 { 1569 struct trans_for_each_path_inorder_iter iter; 1570 1571 if (!nosort) 1572 btree_trans_sort_paths(trans); 1573 1574 trans_for_each_path_idx_inorder(trans, iter) { 1575 bch2_btree_path_to_text_short(out, trans, iter.path_idx); 1576 prt_newline(out); 1577 } 1578 } 1579 1580 noinline __cold 1581 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans) 1582 { 1583 __bch2_trans_paths_to_text(out, trans, false); 1584 } 1585 1586 static noinline __cold 1587 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort) 1588 { 1589 struct printbuf buf = PRINTBUF; 1590 1591 __bch2_trans_paths_to_text(&buf, trans, nosort); 1592 bch2_trans_updates_to_text(&buf, trans); 1593 1594 bch2_print_str(trans->c, buf.buf); 1595 printbuf_exit(&buf); 1596 } 1597 1598 noinline __cold 1599 void bch2_dump_trans_paths_updates(struct btree_trans *trans) 1600 { 1601 __bch2_dump_trans_paths_updates(trans, false); 1602 } 1603 1604 noinline __cold 1605 static void bch2_trans_update_max_paths(struct btree_trans *trans) 1606 { 1607 struct btree_transaction_stats *s = btree_trans_stats(trans); 1608 struct printbuf buf = PRINTBUF; 1609 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths); 1610 1611 bch2_trans_paths_to_text(&buf, trans); 1612 1613 if (!buf.allocation_failure) { 1614 mutex_lock(&s->lock); 1615 if (nr > s->nr_max_paths) { 1616 s->nr_max_paths = nr; 1617 swap(s->max_paths_text, buf.buf); 1618 } 1619 mutex_unlock(&s->lock); 1620 } 1621 1622 printbuf_exit(&buf); 1623 1624 trans->nr_paths_max = nr; 1625 } 1626 1627 noinline __cold 1628 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans) 1629 { 1630 if (trace_trans_restart_too_many_iters_enabled()) { 1631 struct printbuf buf = PRINTBUF; 1632 1633 bch2_trans_paths_to_text(&buf, trans); 1634 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf); 1635 printbuf_exit(&buf); 1636 } 1637 1638 count_event(trans->c, trans_restart_too_many_iters); 1639 1640 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters); 1641 } 1642 1643 static noinline void btree_path_overflow(struct btree_trans *trans) 1644 { 1645 bch2_dump_trans_paths_updates(trans); 1646 bch_err(trans->c, "trans path overflow"); 1647 } 1648 1649 static noinline void btree_paths_realloc(struct btree_trans *trans) 1650 { 1651 unsigned nr = trans->nr_paths * 2; 1652 1653 void *p = kvzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) + 1654 sizeof(struct btree_trans_paths) + 1655 nr * sizeof(struct btree_path) + 1656 nr * sizeof(btree_path_idx_t) + 8 + 1657 nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL); 1658 1659 unsigned long *paths_allocated = p; 1660 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long)); 1661 p += BITS_TO_LONGS(nr) * sizeof(unsigned long); 1662 1663 p += sizeof(struct btree_trans_paths); 1664 struct btree_path *paths = p; 1665 *trans_paths_nr(paths) = nr; 1666 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path)); 1667 p += nr * sizeof(struct btree_path); 1668 1669 btree_path_idx_t *sorted = p; 1670 memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t)); 1671 p += nr * sizeof(btree_path_idx_t) + 8; 1672 1673 struct btree_insert_entry *updates = p; 1674 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry)); 1675 1676 unsigned long *old = trans->paths_allocated; 1677 1678 rcu_assign_pointer(trans->paths_allocated, paths_allocated); 1679 rcu_assign_pointer(trans->paths, paths); 1680 rcu_assign_pointer(trans->sorted, sorted); 1681 rcu_assign_pointer(trans->updates, updates); 1682 1683 trans->nr_paths = nr; 1684 1685 if (old != trans->_paths_allocated) 1686 kfree_rcu_mightsleep(old); 1687 } 1688 1689 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans, 1690 btree_path_idx_t pos) 1691 { 1692 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths); 1693 1694 if (unlikely(idx == trans->nr_paths)) { 1695 if (trans->nr_paths == BTREE_ITER_MAX) { 1696 btree_path_overflow(trans); 1697 return 0; 1698 } 1699 1700 btree_paths_realloc(trans); 1701 } 1702 1703 /* 1704 * Do this before marking the new path as allocated, since it won't be 1705 * initialized yet: 1706 */ 1707 if (unlikely(idx > trans->nr_paths_max)) 1708 bch2_trans_update_max_paths(trans); 1709 1710 __set_bit(idx, trans->paths_allocated); 1711 1712 struct btree_path *path = &trans->paths[idx]; 1713 path->ref = 0; 1714 path->intent_ref = 0; 1715 path->nodes_locked = 0; 1716 1717 btree_path_list_add(trans, pos, idx); 1718 trans->paths_sorted = false; 1719 return idx; 1720 } 1721 1722 btree_path_idx_t bch2_path_get(struct btree_trans *trans, 1723 enum btree_id btree_id, struct bpos pos, 1724 unsigned locks_want, unsigned level, 1725 unsigned flags, unsigned long ip) 1726 { 1727 struct btree_path *path; 1728 bool cached = flags & BTREE_ITER_cached; 1729 bool intent = flags & BTREE_ITER_intent; 1730 struct trans_for_each_path_inorder_iter iter; 1731 btree_path_idx_t path_pos = 0, path_idx; 1732 1733 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1734 bch2_trans_verify_locks(trans); 1735 1736 btree_trans_sort_paths(trans); 1737 1738 trans_for_each_path_inorder(trans, path, iter) { 1739 if (__btree_path_cmp(path, 1740 btree_id, 1741 cached, 1742 pos, 1743 level) > 0) 1744 break; 1745 1746 path_pos = iter.path_idx; 1747 } 1748 1749 if (path_pos && 1750 trans->paths[path_pos].cached == cached && 1751 trans->paths[path_pos].btree_id == btree_id && 1752 trans->paths[path_pos].level == level) { 1753 trace_btree_path_get(trans, trans->paths + path_pos, &pos); 1754 1755 __btree_path_get(trans, trans->paths + path_pos, intent); 1756 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip); 1757 path = trans->paths + path_idx; 1758 } else { 1759 path_idx = btree_path_alloc(trans, path_pos); 1760 path = trans->paths + path_idx; 1761 1762 __btree_path_get(trans, path, intent); 1763 path->pos = pos; 1764 path->btree_id = btree_id; 1765 path->cached = cached; 1766 path->uptodate = BTREE_ITER_NEED_TRAVERSE; 1767 path->should_be_locked = false; 1768 path->level = level; 1769 path->locks_want = locks_want; 1770 path->nodes_locked = 0; 1771 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++) 1772 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 1773 #ifdef TRACK_PATH_ALLOCATED 1774 path->ip_allocated = ip; 1775 #endif 1776 trans->paths_sorted = false; 1777 1778 trace_btree_path_alloc(trans, path); 1779 } 1780 1781 if (!(flags & BTREE_ITER_nopreserve)) 1782 path->preserve = true; 1783 1784 if (path->intent_ref) 1785 locks_want = max(locks_want, level + 1); 1786 1787 /* 1788 * If the path has locks_want greater than requested, we don't downgrade 1789 * it here - on transaction restart because btree node split needs to 1790 * upgrade locks, we might be putting/getting the iterator again. 1791 * Downgrading iterators only happens via bch2_trans_downgrade(), after 1792 * a successful transaction commit. 1793 */ 1794 1795 locks_want = min(locks_want, BTREE_MAX_DEPTH); 1796 if (locks_want > path->locks_want) 1797 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL); 1798 1799 return path_idx; 1800 } 1801 1802 btree_path_idx_t bch2_path_get_unlocked_mut(struct btree_trans *trans, 1803 enum btree_id btree_id, 1804 unsigned level, 1805 struct bpos pos) 1806 { 1807 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level, 1808 BTREE_ITER_nopreserve| 1809 BTREE_ITER_intent, _RET_IP_); 1810 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_); 1811 1812 struct btree_path *path = trans->paths + path_idx; 1813 bch2_btree_path_downgrade(trans, path); 1814 __bch2_btree_path_unlock(trans, path); 1815 return path_idx; 1816 } 1817 1818 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u) 1819 { 1820 1821 struct btree_path_level *l = path_l(path); 1822 struct bkey_packed *_k; 1823 struct bkey_s_c k; 1824 1825 if (unlikely(!l->b)) 1826 return bkey_s_c_null; 1827 1828 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE); 1829 EBUG_ON(!btree_node_locked(path, path->level)); 1830 1831 if (!path->cached) { 1832 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b); 1833 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null; 1834 1835 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos)); 1836 1837 if (!k.k || !bpos_eq(path->pos, k.k->p)) 1838 goto hole; 1839 } else { 1840 struct bkey_cached *ck = (void *) path->l[0].b; 1841 if (!ck) 1842 return bkey_s_c_null; 1843 1844 EBUG_ON(path->btree_id != ck->key.btree_id || 1845 !bkey_eq(path->pos, ck->key.pos)); 1846 1847 *u = ck->k->k; 1848 k = (struct bkey_s_c) { u, &ck->k->v }; 1849 } 1850 1851 return k; 1852 hole: 1853 bkey_init(u); 1854 u->p = path->pos; 1855 return (struct bkey_s_c) { u, NULL }; 1856 } 1857 1858 void bch2_set_btree_iter_dontneed(struct btree_trans *trans, struct btree_iter *iter) 1859 { 1860 if (!iter->path || trans->restarted) 1861 return; 1862 1863 struct btree_path *path = btree_iter_path(trans, iter); 1864 path->preserve = false; 1865 if (path->ref == 1) 1866 path->should_be_locked = false; 1867 } 1868 /* Btree iterators: */ 1869 1870 int __must_check 1871 __bch2_btree_iter_traverse(struct btree_trans *trans, struct btree_iter *iter) 1872 { 1873 return bch2_btree_path_traverse(trans, iter->path, iter->flags); 1874 } 1875 1876 int __must_check 1877 bch2_btree_iter_traverse(struct btree_trans *trans, struct btree_iter *iter) 1878 { 1879 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1880 1881 iter->path = bch2_btree_path_set_pos(trans, iter->path, 1882 btree_iter_search_key(iter), 1883 iter->flags & BTREE_ITER_intent, 1884 btree_iter_ip_allocated(iter)); 1885 1886 int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 1887 if (ret) 1888 return ret; 1889 1890 struct btree_path *path = btree_iter_path(trans, iter); 1891 if (btree_path_node(path, path->level)) 1892 btree_path_set_should_be_locked(trans, path); 1893 return 0; 1894 } 1895 1896 /* Iterate across nodes (leaf and interior nodes) */ 1897 1898 struct btree *bch2_btree_iter_peek_node(struct btree_trans *trans, 1899 struct btree_iter *iter) 1900 { 1901 struct btree *b = NULL; 1902 int ret; 1903 1904 EBUG_ON(trans->paths[iter->path].cached); 1905 bch2_btree_iter_verify(trans, iter); 1906 1907 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 1908 if (ret) 1909 goto err; 1910 1911 struct btree_path *path = btree_iter_path(trans, iter); 1912 b = btree_path_node(path, path->level); 1913 if (!b) 1914 goto out; 1915 1916 BUG_ON(bpos_lt(b->key.k.p, iter->pos)); 1917 1918 bkey_init(&iter->k); 1919 iter->k.p = iter->pos = b->key.k.p; 1920 1921 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, 1922 iter->flags & BTREE_ITER_intent, 1923 btree_iter_ip_allocated(iter)); 1924 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 1925 out: 1926 bch2_btree_iter_verify_entry_exit(iter); 1927 bch2_btree_iter_verify(trans, iter); 1928 1929 return b; 1930 err: 1931 b = ERR_PTR(ret); 1932 goto out; 1933 } 1934 1935 /* Only kept for -tools */ 1936 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_trans *trans, 1937 struct btree_iter *iter) 1938 { 1939 struct btree *b; 1940 1941 while (b = bch2_btree_iter_peek_node(trans, iter), 1942 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart)) 1943 bch2_trans_begin(trans); 1944 1945 return b; 1946 } 1947 1948 struct btree *bch2_btree_iter_next_node(struct btree_trans *trans, struct btree_iter *iter) 1949 { 1950 struct btree *b = NULL; 1951 int ret; 1952 1953 EBUG_ON(trans->paths[iter->path].cached); 1954 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1955 bch2_btree_iter_verify(trans, iter); 1956 1957 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 1958 if (ret) 1959 goto err; 1960 1961 1962 struct btree_path *path = btree_iter_path(trans, iter); 1963 1964 /* already at end? */ 1965 if (!btree_path_node(path, path->level)) 1966 return NULL; 1967 1968 /* got to end? */ 1969 if (!btree_path_node(path, path->level + 1)) { 1970 btree_path_set_level_up(trans, path); 1971 return NULL; 1972 } 1973 1974 if (!bch2_btree_node_relock(trans, path, path->level + 1)) { 1975 __bch2_btree_path_unlock(trans, path); 1976 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); 1977 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); 1978 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1979 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path); 1980 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock); 1981 goto err; 1982 } 1983 1984 b = btree_path_node(path, path->level + 1); 1985 1986 if (bpos_eq(iter->pos, b->key.k.p)) { 1987 __btree_path_set_level_up(trans, path, path->level++); 1988 } else { 1989 if (btree_lock_want(path, path->level + 1) == BTREE_NODE_UNLOCKED) 1990 btree_node_unlock(trans, path, path->level + 1); 1991 1992 /* 1993 * Haven't gotten to the end of the parent node: go back down to 1994 * the next child node 1995 */ 1996 iter->path = bch2_btree_path_set_pos(trans, iter->path, 1997 bpos_successor(iter->pos), 1998 iter->flags & BTREE_ITER_intent, 1999 btree_iter_ip_allocated(iter)); 2000 2001 path = btree_iter_path(trans, iter); 2002 btree_path_set_level_down(trans, path, iter->min_depth); 2003 2004 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2005 if (ret) 2006 goto err; 2007 2008 path = btree_iter_path(trans, iter); 2009 b = path->l[path->level].b; 2010 } 2011 2012 bkey_init(&iter->k); 2013 iter->k.p = iter->pos = b->key.k.p; 2014 2015 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, 2016 iter->flags & BTREE_ITER_intent, 2017 btree_iter_ip_allocated(iter)); 2018 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2019 EBUG_ON(btree_iter_path(trans, iter)->uptodate); 2020 out: 2021 bch2_btree_iter_verify_entry_exit(iter); 2022 bch2_btree_iter_verify(trans, iter); 2023 2024 return b; 2025 err: 2026 b = ERR_PTR(ret); 2027 goto out; 2028 } 2029 2030 /* Iterate across keys (in leaf nodes only) */ 2031 2032 inline bool bch2_btree_iter_advance(struct btree_trans *trans, struct btree_iter *iter) 2033 { 2034 struct bpos pos = iter->k.p; 2035 bool ret = !(iter->flags & BTREE_ITER_all_snapshots 2036 ? bpos_eq(pos, SPOS_MAX) 2037 : bkey_eq(pos, SPOS_MAX)); 2038 2039 if (ret && !(iter->flags & BTREE_ITER_is_extents)) 2040 pos = bkey_successor(iter, pos); 2041 bch2_btree_iter_set_pos(trans, iter, pos); 2042 return ret; 2043 } 2044 2045 inline bool bch2_btree_iter_rewind(struct btree_trans *trans, struct btree_iter *iter) 2046 { 2047 struct bpos pos = bkey_start_pos(&iter->k); 2048 bool ret = !(iter->flags & BTREE_ITER_all_snapshots 2049 ? bpos_eq(pos, POS_MIN) 2050 : bkey_eq(pos, POS_MIN)); 2051 2052 if (ret && !(iter->flags & BTREE_ITER_is_extents)) 2053 pos = bkey_predecessor(iter, pos); 2054 bch2_btree_iter_set_pos(trans, iter, pos); 2055 return ret; 2056 } 2057 2058 static noinline 2059 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter, 2060 struct bkey_s_c *k) 2061 { 2062 struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key; 2063 2064 trans_for_each_update(trans, i) 2065 if (!i->key_cache_already_flushed && 2066 i->btree_id == iter->btree_id && 2067 bpos_le(i->k->k.p, iter->pos) && 2068 bpos_ge(i->k->k.p, k->k ? k->k->p : end)) { 2069 iter->k = i->k->k; 2070 *k = bkey_i_to_s_c(i->k); 2071 } 2072 } 2073 2074 static noinline 2075 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter, 2076 struct bkey_s_c *k) 2077 { 2078 struct btree_path *path = btree_iter_path(trans, iter); 2079 struct bpos end = path_l(path)->b->key.k.p; 2080 2081 trans_for_each_update(trans, i) 2082 if (!i->key_cache_already_flushed && 2083 i->btree_id == iter->btree_id && 2084 bpos_ge(i->k->k.p, path->pos) && 2085 bpos_le(i->k->k.p, k->k ? k->k->p : end)) { 2086 iter->k = i->k->k; 2087 *k = bkey_i_to_s_c(i->k); 2088 } 2089 } 2090 2091 static noinline 2092 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter, 2093 struct bkey_s_c *k) 2094 { 2095 trans_for_each_update(trans, i) 2096 if (!i->key_cache_already_flushed && 2097 i->btree_id == iter->btree_id && 2098 bpos_eq(i->k->k.p, iter->pos)) { 2099 iter->k = i->k->k; 2100 *k = bkey_i_to_s_c(i->k); 2101 } 2102 } 2103 2104 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans, 2105 struct btree_iter *iter, 2106 struct bpos end_pos) 2107 { 2108 struct btree_path *path = btree_iter_path(trans, iter); 2109 2110 return bch2_journal_keys_peek_max(trans->c, iter->btree_id, 2111 path->level, 2112 path->pos, 2113 end_pos, 2114 &iter->journal_idx); 2115 } 2116 2117 static noinline 2118 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans, 2119 struct btree_iter *iter) 2120 { 2121 struct btree_path *path = btree_iter_path(trans, iter); 2122 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos); 2123 2124 if (k) { 2125 iter->k = k->k; 2126 return bkey_i_to_s_c(k); 2127 } else { 2128 return bkey_s_c_null; 2129 } 2130 } 2131 2132 static noinline 2133 void btree_trans_peek_journal(struct btree_trans *trans, 2134 struct btree_iter *iter, 2135 struct bkey_s_c *k) 2136 { 2137 struct btree_path *path = btree_iter_path(trans, iter); 2138 struct bkey_i *next_journal = 2139 bch2_btree_journal_peek(trans, iter, 2140 k->k ? k->k->p : path_l(path)->b->key.k.p); 2141 if (next_journal) { 2142 iter->k = next_journal->k; 2143 *k = bkey_i_to_s_c(next_journal); 2144 } 2145 } 2146 2147 static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans, 2148 struct btree_iter *iter, 2149 struct bpos end_pos) 2150 { 2151 struct btree_path *path = btree_iter_path(trans, iter); 2152 2153 return bch2_journal_keys_peek_prev_min(trans->c, iter->btree_id, 2154 path->level, 2155 path->pos, 2156 end_pos, 2157 &iter->journal_idx); 2158 } 2159 2160 static noinline 2161 void btree_trans_peek_prev_journal(struct btree_trans *trans, 2162 struct btree_iter *iter, 2163 struct bkey_s_c *k) 2164 { 2165 struct btree_path *path = btree_iter_path(trans, iter); 2166 struct bkey_i *next_journal = 2167 bch2_btree_journal_peek_prev(trans, iter, 2168 k->k ? k->k->p : path_l(path)->b->key.k.p); 2169 2170 if (next_journal) { 2171 iter->k = next_journal->k; 2172 *k = bkey_i_to_s_c(next_journal); 2173 } 2174 } 2175 2176 /* 2177 * Checks btree key cache for key at iter->pos and returns it if present, or 2178 * bkey_s_c_null: 2179 */ 2180 static noinline 2181 struct bkey_s_c btree_trans_peek_key_cache(struct btree_trans *trans, struct btree_iter *iter, 2182 struct bpos pos) 2183 { 2184 struct bch_fs *c = trans->c; 2185 struct bkey u; 2186 struct bkey_s_c k; 2187 int ret; 2188 2189 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2190 2191 if ((iter->flags & BTREE_ITER_key_cache_fill) && 2192 bpos_eq(iter->pos, pos)) 2193 return bkey_s_c_null; 2194 2195 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos)) 2196 return bkey_s_c_null; 2197 2198 if (!iter->key_cache_path) 2199 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos, 2200 iter->flags & BTREE_ITER_intent, 0, 2201 iter->flags|BTREE_ITER_cached| 2202 BTREE_ITER_cached_nofill, 2203 _THIS_IP_); 2204 2205 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos, 2206 iter->flags & BTREE_ITER_intent, 2207 btree_iter_ip_allocated(iter)); 2208 2209 ret = bch2_btree_path_traverse(trans, iter->key_cache_path, 2210 iter->flags|BTREE_ITER_cached) ?: 2211 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_); 2212 if (unlikely(ret)) 2213 return bkey_s_c_err(ret); 2214 2215 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u); 2216 if (!k.k) 2217 return k; 2218 2219 if ((iter->flags & BTREE_ITER_all_snapshots) && 2220 !bpos_eq(pos, k.k->p)) 2221 return bkey_s_c_null; 2222 2223 iter->k = u; 2224 k.k = &iter->k; 2225 btree_path_set_should_be_locked(trans, trans->paths + iter->key_cache_path); 2226 return k; 2227 } 2228 2229 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_trans *trans, struct btree_iter *iter, 2230 struct bpos search_key) 2231 { 2232 struct bkey_s_c k, k2; 2233 int ret; 2234 2235 EBUG_ON(btree_iter_path(trans, iter)->cached); 2236 bch2_btree_iter_verify(trans, iter); 2237 2238 while (1) { 2239 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2240 iter->flags & BTREE_ITER_intent, 2241 btree_iter_ip_allocated(iter)); 2242 2243 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2244 if (unlikely(ret)) { 2245 /* ensure that iter->k is consistent with iter->pos: */ 2246 bch2_btree_iter_set_pos(trans, iter, iter->pos); 2247 k = bkey_s_c_err(ret); 2248 break; 2249 } 2250 2251 struct btree_path *path = btree_iter_path(trans, iter); 2252 struct btree_path_level *l = path_l(path); 2253 2254 if (unlikely(!l->b)) { 2255 /* No btree nodes at requested level: */ 2256 bch2_btree_iter_set_pos(trans, iter, SPOS_MAX); 2257 k = bkey_s_c_null; 2258 break; 2259 } 2260 2261 btree_path_set_should_be_locked(trans, path); 2262 2263 k = btree_path_level_peek_all(trans->c, l, &iter->k); 2264 2265 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2266 k.k && 2267 (k2 = btree_trans_peek_key_cache(trans, iter, k.k->p)).k) { 2268 k = k2; 2269 if (bkey_err(k)) { 2270 bch2_btree_iter_set_pos(trans, iter, iter->pos); 2271 break; 2272 } 2273 } 2274 2275 if (unlikely(iter->flags & BTREE_ITER_with_journal)) 2276 btree_trans_peek_journal(trans, iter, &k); 2277 2278 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2279 trans->nr_updates)) 2280 bch2_btree_trans_peek_updates(trans, iter, &k); 2281 2282 if (k.k && bkey_deleted(k.k)) { 2283 /* 2284 * If we've got a whiteout, and it's after the search 2285 * key, advance the search key to the whiteout instead 2286 * of just after the whiteout - it might be a btree 2287 * whiteout, with a real key at the same position, since 2288 * in the btree deleted keys sort before non deleted. 2289 */ 2290 search_key = !bpos_eq(search_key, k.k->p) 2291 ? k.k->p 2292 : bpos_successor(k.k->p); 2293 continue; 2294 } 2295 2296 if (likely(k.k)) { 2297 break; 2298 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) { 2299 /* Advance to next leaf node: */ 2300 search_key = bpos_successor(l->b->key.k.p); 2301 } else { 2302 /* End of btree: */ 2303 bch2_btree_iter_set_pos(trans, iter, SPOS_MAX); 2304 k = bkey_s_c_null; 2305 break; 2306 } 2307 } 2308 2309 bch2_btree_iter_verify(trans, iter); 2310 return k; 2311 } 2312 2313 /** 2314 * bch2_btree_iter_peek_max() - returns first key greater than or equal to 2315 * iterator's current position 2316 * @trans: btree transaction object 2317 * @iter: iterator to peek from 2318 * @end: search limit: returns keys less than or equal to @end 2319 * 2320 * Returns: key if found, or an error extractable with bkey_err(). 2321 */ 2322 struct bkey_s_c bch2_btree_iter_peek_max(struct btree_trans *trans, struct btree_iter *iter, 2323 struct bpos end) 2324 { 2325 struct bpos search_key = btree_iter_search_key(iter); 2326 struct bkey_s_c k; 2327 struct bpos iter_pos = iter->pos; 2328 int ret; 2329 2330 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2331 bch2_btree_iter_verify_entry_exit(iter); 2332 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bkey_eq(end, POS_MAX)); 2333 2334 ret = trans_maybe_inject_restart(trans, _RET_IP_); 2335 if (unlikely(ret)) { 2336 k = bkey_s_c_err(ret); 2337 goto out_no_locked; 2338 } 2339 2340 if (iter->update_path) { 2341 bch2_path_put_nokeep(trans, iter->update_path, 2342 iter->flags & BTREE_ITER_intent); 2343 iter->update_path = 0; 2344 } 2345 2346 while (1) { 2347 k = __bch2_btree_iter_peek(trans, iter, search_key); 2348 if (unlikely(!k.k)) 2349 goto end; 2350 if (unlikely(bkey_err(k))) 2351 goto out_no_locked; 2352 2353 if (iter->flags & BTREE_ITER_filter_snapshots) { 2354 /* 2355 * We need to check against @end before FILTER_SNAPSHOTS because 2356 * if we get to a different inode that requested we might be 2357 * seeing keys for a different snapshot tree that will all be 2358 * filtered out. 2359 * 2360 * But we can't do the full check here, because bkey_start_pos() 2361 * isn't monotonically increasing before FILTER_SNAPSHOTS, and 2362 * that's what we check against in extents mode: 2363 */ 2364 if (unlikely(!(iter->flags & BTREE_ITER_is_extents) 2365 ? bkey_gt(k.k->p, end) 2366 : k.k->p.inode > end.inode)) 2367 goto end; 2368 2369 if (iter->update_path && 2370 !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) { 2371 bch2_path_put_nokeep(trans, iter->update_path, 2372 iter->flags & BTREE_ITER_intent); 2373 iter->update_path = 0; 2374 } 2375 2376 if ((iter->flags & BTREE_ITER_intent) && 2377 !(iter->flags & BTREE_ITER_is_extents) && 2378 !iter->update_path) { 2379 struct bpos pos = k.k->p; 2380 2381 if (pos.snapshot < iter->snapshot) { 2382 search_key = bpos_successor(k.k->p); 2383 continue; 2384 } 2385 2386 pos.snapshot = iter->snapshot; 2387 2388 /* 2389 * advance, same as on exit for iter->path, but only up 2390 * to snapshot 2391 */ 2392 __btree_path_get(trans, trans->paths + iter->path, iter->flags & BTREE_ITER_intent); 2393 iter->update_path = iter->path; 2394 2395 iter->update_path = bch2_btree_path_set_pos(trans, 2396 iter->update_path, pos, 2397 iter->flags & BTREE_ITER_intent, 2398 _THIS_IP_); 2399 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags); 2400 if (unlikely(ret)) { 2401 k = bkey_s_c_err(ret); 2402 goto out_no_locked; 2403 } 2404 } 2405 2406 /* 2407 * We can never have a key in a leaf node at POS_MAX, so 2408 * we don't have to check these successor() calls: 2409 */ 2410 if (!bch2_snapshot_is_ancestor(trans->c, 2411 iter->snapshot, 2412 k.k->p.snapshot)) { 2413 search_key = bpos_successor(k.k->p); 2414 continue; 2415 } 2416 2417 if (bkey_whiteout(k.k) && 2418 !(iter->flags & BTREE_ITER_key_cache_fill)) { 2419 search_key = bkey_successor(iter, k.k->p); 2420 continue; 2421 } 2422 } 2423 2424 /* 2425 * iter->pos should be mononotically increasing, and always be 2426 * equal to the key we just returned - except extents can 2427 * straddle iter->pos: 2428 */ 2429 if (!(iter->flags & BTREE_ITER_is_extents)) 2430 iter_pos = k.k->p; 2431 else 2432 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k)); 2433 2434 if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_gt(iter_pos, end) : 2435 iter->flags & BTREE_ITER_is_extents ? bkey_ge(iter_pos, end) : 2436 bkey_gt(iter_pos, end))) 2437 goto end; 2438 2439 break; 2440 } 2441 2442 iter->pos = iter_pos; 2443 2444 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p, 2445 iter->flags & BTREE_ITER_intent, 2446 btree_iter_ip_allocated(iter)); 2447 2448 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2449 out_no_locked: 2450 if (iter->update_path) { 2451 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_); 2452 if (unlikely(ret)) 2453 k = bkey_s_c_err(ret); 2454 else 2455 btree_path_set_should_be_locked(trans, trans->paths + iter->update_path); 2456 } 2457 2458 if (!(iter->flags & BTREE_ITER_all_snapshots)) 2459 iter->pos.snapshot = iter->snapshot; 2460 2461 ret = bch2_btree_iter_verify_ret(trans, iter, k); 2462 if (unlikely(ret)) { 2463 bch2_btree_iter_set_pos(trans, iter, iter->pos); 2464 k = bkey_s_c_err(ret); 2465 } 2466 2467 bch2_btree_iter_verify_entry_exit(iter); 2468 2469 return k; 2470 end: 2471 bch2_btree_iter_set_pos(trans, iter, end); 2472 k = bkey_s_c_null; 2473 goto out_no_locked; 2474 } 2475 2476 /** 2477 * bch2_btree_iter_next() - returns first key greater than iterator's current 2478 * position 2479 * @trans: btree transaction object 2480 * @iter: iterator to peek from 2481 * 2482 * Returns: key if found, or an error extractable with bkey_err(). 2483 */ 2484 struct bkey_s_c bch2_btree_iter_next(struct btree_trans *trans, struct btree_iter *iter) 2485 { 2486 if (!bch2_btree_iter_advance(trans, iter)) 2487 return bkey_s_c_null; 2488 2489 return bch2_btree_iter_peek(trans, iter); 2490 } 2491 2492 static struct bkey_s_c __bch2_btree_iter_peek_prev(struct btree_trans *trans, struct btree_iter *iter, 2493 struct bpos search_key) 2494 { 2495 struct bkey_s_c k, k2; 2496 2497 bch2_btree_iter_verify(trans, iter); 2498 2499 while (1) { 2500 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2501 iter->flags & BTREE_ITER_intent, 2502 btree_iter_ip_allocated(iter)); 2503 2504 int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2505 if (unlikely(ret)) { 2506 /* ensure that iter->k is consistent with iter->pos: */ 2507 bch2_btree_iter_set_pos(trans, iter, iter->pos); 2508 k = bkey_s_c_err(ret); 2509 break; 2510 } 2511 2512 struct btree_path *path = btree_iter_path(trans, iter); 2513 struct btree_path_level *l = path_l(path); 2514 2515 if (unlikely(!l->b)) { 2516 /* No btree nodes at requested level: */ 2517 bch2_btree_iter_set_pos(trans, iter, SPOS_MAX); 2518 k = bkey_s_c_null; 2519 break; 2520 } 2521 2522 btree_path_set_should_be_locked(trans, path); 2523 2524 k = btree_path_level_peek_all(trans->c, l, &iter->k); 2525 if (!k.k || bpos_gt(k.k->p, search_key)) { 2526 k = btree_path_level_prev(trans, path, l, &iter->k); 2527 2528 BUG_ON(k.k && bpos_gt(k.k->p, search_key)); 2529 } 2530 2531 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2532 k.k && 2533 (k2 = btree_trans_peek_key_cache(trans, iter, k.k->p)).k) { 2534 k = k2; 2535 if (bkey_err(k2)) { 2536 bch2_btree_iter_set_pos(trans, iter, iter->pos); 2537 break; 2538 } 2539 } 2540 2541 if (unlikely(iter->flags & BTREE_ITER_with_journal)) 2542 btree_trans_peek_prev_journal(trans, iter, &k); 2543 2544 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2545 trans->nr_updates)) 2546 bch2_btree_trans_peek_prev_updates(trans, iter, &k); 2547 2548 if (likely(k.k && !bkey_deleted(k.k))) { 2549 break; 2550 } else if (k.k) { 2551 search_key = bpos_predecessor(k.k->p); 2552 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) { 2553 /* Advance to previous leaf node: */ 2554 search_key = bpos_predecessor(path->l[0].b->data->min_key); 2555 } else { 2556 /* Start of btree: */ 2557 bch2_btree_iter_set_pos(trans, iter, POS_MIN); 2558 k = bkey_s_c_null; 2559 break; 2560 } 2561 } 2562 2563 bch2_btree_iter_verify(trans, iter); 2564 return k; 2565 } 2566 2567 /** 2568 * bch2_btree_iter_peek_prev_min() - returns first key less than or equal to 2569 * iterator's current position 2570 * @trans: btree transaction object 2571 * @iter: iterator to peek from 2572 * @end: search limit: returns keys greater than or equal to @end 2573 * 2574 * Returns: key if found, or an error extractable with bkey_err(). 2575 */ 2576 struct bkey_s_c bch2_btree_iter_peek_prev_min(struct btree_trans *trans, struct btree_iter *iter, 2577 struct bpos end) 2578 { 2579 if ((iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots)) && 2580 !bkey_eq(iter->pos, POS_MAX)) { 2581 /* 2582 * bkey_start_pos(), for extents, is not monotonically 2583 * increasing until after filtering for snapshots: 2584 * 2585 * Thus, for extents we need to search forward until we find a 2586 * real visible extents - easiest to just use peek_slot() (which 2587 * internally uses peek() for extents) 2588 */ 2589 struct bkey_s_c k = bch2_btree_iter_peek_slot(trans, iter); 2590 if (bkey_err(k)) 2591 return k; 2592 2593 if (!bkey_deleted(k.k) && 2594 (!(iter->flags & BTREE_ITER_is_extents) || 2595 bkey_lt(bkey_start_pos(k.k), iter->pos))) 2596 return k; 2597 } 2598 2599 struct bpos search_key = iter->pos; 2600 struct bkey_s_c k; 2601 btree_path_idx_t saved_path = 0; 2602 2603 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2604 bch2_btree_iter_verify_entry_exit(iter); 2605 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bpos_eq(end, POS_MIN)); 2606 2607 int ret = trans_maybe_inject_restart(trans, _RET_IP_); 2608 if (unlikely(ret)) { 2609 k = bkey_s_c_err(ret); 2610 goto out_no_locked; 2611 } 2612 2613 while (1) { 2614 k = __bch2_btree_iter_peek_prev(trans, iter, search_key); 2615 if (unlikely(!k.k)) 2616 goto end; 2617 if (unlikely(bkey_err(k))) 2618 goto out_no_locked; 2619 2620 if (iter->flags & BTREE_ITER_filter_snapshots) { 2621 struct btree_path *s = saved_path ? trans->paths + saved_path : NULL; 2622 if (s && bpos_lt(k.k->p, SPOS(s->pos.inode, s->pos.offset, iter->snapshot))) { 2623 /* 2624 * If we have a saved candidate, and we're past 2625 * the last possible snapshot overwrite, return 2626 * it: 2627 */ 2628 bch2_path_put_nokeep(trans, iter->path, 2629 iter->flags & BTREE_ITER_intent); 2630 iter->path = saved_path; 2631 saved_path = 0; 2632 k = bch2_btree_path_peek_slot(btree_iter_path(trans, iter), &iter->k); 2633 break; 2634 } 2635 2636 /* 2637 * We need to check against @end before FILTER_SNAPSHOTS because 2638 * if we get to a different inode that requested we might be 2639 * seeing keys for a different snapshot tree that will all be 2640 * filtered out. 2641 */ 2642 if (unlikely(bkey_lt(k.k->p, end))) 2643 goto end; 2644 2645 if (!bch2_snapshot_is_ancestor(trans->c, iter->snapshot, k.k->p.snapshot)) { 2646 search_key = bpos_predecessor(k.k->p); 2647 continue; 2648 } 2649 2650 if (k.k->p.snapshot != iter->snapshot) { 2651 /* 2652 * Have a key visible in iter->snapshot, but 2653 * might have overwrites: - save it and keep 2654 * searching. Unless it's a whiteout - then drop 2655 * our previous saved candidate: 2656 */ 2657 if (saved_path) { 2658 bch2_path_put_nokeep(trans, saved_path, 2659 iter->flags & BTREE_ITER_intent); 2660 saved_path = 0; 2661 } 2662 2663 if (!bkey_whiteout(k.k)) { 2664 saved_path = btree_path_clone(trans, iter->path, 2665 iter->flags & BTREE_ITER_intent, 2666 _THIS_IP_); 2667 trace_btree_path_save_pos(trans, 2668 trans->paths + iter->path, 2669 trans->paths + saved_path); 2670 } 2671 2672 search_key = bpos_predecessor(k.k->p); 2673 continue; 2674 } 2675 2676 if (bkey_whiteout(k.k)) { 2677 search_key = bkey_predecessor(iter, k.k->p); 2678 search_key.snapshot = U32_MAX; 2679 continue; 2680 } 2681 } 2682 2683 EBUG_ON(iter->flags & BTREE_ITER_all_snapshots ? bpos_gt(k.k->p, iter->pos) : 2684 iter->flags & BTREE_ITER_is_extents ? bkey_ge(bkey_start_pos(k.k), iter->pos) : 2685 bkey_gt(k.k->p, iter->pos)); 2686 2687 if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_lt(k.k->p, end) : 2688 iter->flags & BTREE_ITER_is_extents ? bkey_le(k.k->p, end) : 2689 bkey_lt(k.k->p, end))) 2690 goto end; 2691 2692 break; 2693 } 2694 2695 /* Extents can straddle iter->pos: */ 2696 iter->pos = bpos_min(iter->pos, k.k->p);; 2697 2698 if (iter->flags & BTREE_ITER_filter_snapshots) 2699 iter->pos.snapshot = iter->snapshot; 2700 out_no_locked: 2701 if (saved_path) 2702 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_intent); 2703 2704 bch2_btree_iter_verify_entry_exit(iter); 2705 bch2_btree_iter_verify(trans, iter); 2706 return k; 2707 end: 2708 bch2_btree_iter_set_pos(trans, iter, end); 2709 k = bkey_s_c_null; 2710 goto out_no_locked; 2711 } 2712 2713 /** 2714 * bch2_btree_iter_prev() - returns first key less than iterator's current 2715 * position 2716 * @trans: btree transaction object 2717 * @iter: iterator to peek from 2718 * 2719 * Returns: key if found, or an error extractable with bkey_err(). 2720 */ 2721 struct bkey_s_c bch2_btree_iter_prev(struct btree_trans *trans, struct btree_iter *iter) 2722 { 2723 if (!bch2_btree_iter_rewind(trans, iter)) 2724 return bkey_s_c_null; 2725 2726 return bch2_btree_iter_peek_prev(trans, iter); 2727 } 2728 2729 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_trans *trans, struct btree_iter *iter) 2730 { 2731 struct bpos search_key; 2732 struct bkey_s_c k; 2733 int ret; 2734 2735 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2736 bch2_btree_iter_verify(trans, iter); 2737 bch2_btree_iter_verify_entry_exit(iter); 2738 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_with_key_cache)); 2739 2740 ret = trans_maybe_inject_restart(trans, _RET_IP_); 2741 if (unlikely(ret)) { 2742 k = bkey_s_c_err(ret); 2743 goto out_no_locked; 2744 } 2745 2746 /* extents can't span inode numbers: */ 2747 if ((iter->flags & BTREE_ITER_is_extents) && 2748 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) { 2749 if (iter->pos.inode == KEY_INODE_MAX) 2750 return bkey_s_c_null; 2751 2752 bch2_btree_iter_set_pos(trans, iter, bpos_nosnap_successor(iter->pos)); 2753 } 2754 2755 search_key = btree_iter_search_key(iter); 2756 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2757 iter->flags & BTREE_ITER_intent, 2758 btree_iter_ip_allocated(iter)); 2759 2760 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2761 if (unlikely(ret)) { 2762 k = bkey_s_c_err(ret); 2763 goto out_no_locked; 2764 } 2765 2766 struct btree_path *path = btree_iter_path(trans, iter); 2767 if (unlikely(!btree_path_node(path, path->level))) 2768 return bkey_s_c_null; 2769 2770 if ((iter->flags & BTREE_ITER_cached) || 2771 !(iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots))) { 2772 k = bkey_s_c_null; 2773 2774 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2775 trans->nr_updates)) { 2776 bch2_btree_trans_peek_slot_updates(trans, iter, &k); 2777 if (k.k) 2778 goto out; 2779 } 2780 2781 if (unlikely(iter->flags & BTREE_ITER_with_journal) && 2782 (k = btree_trans_peek_slot_journal(trans, iter)).k) 2783 goto out; 2784 2785 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2786 (k = btree_trans_peek_key_cache(trans, iter, iter->pos)).k) { 2787 if (!bkey_err(k)) 2788 iter->k = *k.k; 2789 /* We're not returning a key from iter->path: */ 2790 goto out_no_locked; 2791 } 2792 2793 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k); 2794 if (unlikely(!k.k)) 2795 goto out_no_locked; 2796 2797 if (unlikely(k.k->type == KEY_TYPE_whiteout && 2798 (iter->flags & BTREE_ITER_filter_snapshots) && 2799 !(iter->flags & BTREE_ITER_key_cache_fill))) 2800 iter->k.type = KEY_TYPE_deleted; 2801 } else { 2802 struct bpos next; 2803 struct bpos end = iter->pos; 2804 2805 if (iter->flags & BTREE_ITER_is_extents) 2806 end.offset = U64_MAX; 2807 2808 EBUG_ON(btree_iter_path(trans, iter)->level); 2809 2810 if (iter->flags & BTREE_ITER_intent) { 2811 struct btree_iter iter2; 2812 2813 bch2_trans_copy_iter(trans, &iter2, iter); 2814 k = bch2_btree_iter_peek_max(trans, &iter2, end); 2815 2816 if (k.k && !bkey_err(k)) { 2817 swap(iter->key_cache_path, iter2.key_cache_path); 2818 iter->k = iter2.k; 2819 k.k = &iter->k; 2820 } 2821 bch2_trans_iter_exit(trans, &iter2); 2822 } else { 2823 struct bpos pos = iter->pos; 2824 2825 k = bch2_btree_iter_peek_max(trans, iter, end); 2826 if (unlikely(bkey_err(k))) 2827 bch2_btree_iter_set_pos(trans, iter, pos); 2828 else 2829 iter->pos = pos; 2830 } 2831 2832 if (unlikely(bkey_err(k))) 2833 goto out_no_locked; 2834 2835 next = k.k ? bkey_start_pos(k.k) : POS_MAX; 2836 2837 if (bkey_lt(iter->pos, next)) { 2838 bkey_init(&iter->k); 2839 iter->k.p = iter->pos; 2840 2841 if (iter->flags & BTREE_ITER_is_extents) { 2842 bch2_key_resize(&iter->k, 2843 min_t(u64, KEY_SIZE_MAX, 2844 (next.inode == iter->pos.inode 2845 ? next.offset 2846 : KEY_OFFSET_MAX) - 2847 iter->pos.offset)); 2848 EBUG_ON(!iter->k.size); 2849 } 2850 2851 k = (struct bkey_s_c) { &iter->k, NULL }; 2852 } 2853 } 2854 out: 2855 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2856 out_no_locked: 2857 bch2_btree_iter_verify_entry_exit(iter); 2858 bch2_btree_iter_verify(trans, iter); 2859 ret = bch2_btree_iter_verify_ret(trans, iter, k); 2860 if (unlikely(ret)) 2861 return bkey_s_c_err(ret); 2862 2863 return k; 2864 } 2865 2866 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_trans *trans, struct btree_iter *iter) 2867 { 2868 if (!bch2_btree_iter_advance(trans, iter)) 2869 return bkey_s_c_null; 2870 2871 return bch2_btree_iter_peek_slot(trans, iter); 2872 } 2873 2874 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_trans *trans, struct btree_iter *iter) 2875 { 2876 if (!bch2_btree_iter_rewind(trans, iter)) 2877 return bkey_s_c_null; 2878 2879 return bch2_btree_iter_peek_slot(trans, iter); 2880 } 2881 2882 /* Obsolete, but still used by rust wrapper in -tools */ 2883 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_trans *trans, struct btree_iter *iter) 2884 { 2885 struct bkey_s_c k; 2886 2887 while (btree_trans_too_many_iters(trans) || 2888 (k = bch2_btree_iter_peek_type(trans, iter, iter->flags), 2889 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart))) 2890 bch2_trans_begin(trans); 2891 2892 return k; 2893 } 2894 2895 /* new transactional stuff: */ 2896 2897 #ifdef CONFIG_BCACHEFS_DEBUG 2898 static void btree_trans_verify_sorted_refs(struct btree_trans *trans) 2899 { 2900 struct btree_path *path; 2901 unsigned i; 2902 2903 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1); 2904 2905 trans_for_each_path(trans, path, i) { 2906 BUG_ON(path->sorted_idx >= trans->nr_sorted); 2907 BUG_ON(trans->sorted[path->sorted_idx] != i); 2908 } 2909 2910 for (i = 0; i < trans->nr_sorted; i++) { 2911 unsigned idx = trans->sorted[i]; 2912 2913 BUG_ON(!test_bit(idx, trans->paths_allocated)); 2914 BUG_ON(trans->paths[idx].sorted_idx != i); 2915 } 2916 } 2917 2918 static void btree_trans_verify_sorted(struct btree_trans *trans) 2919 { 2920 struct btree_path *path, *prev = NULL; 2921 struct trans_for_each_path_inorder_iter iter; 2922 2923 if (!bch2_debug_check_iterators) 2924 return; 2925 2926 trans_for_each_path_inorder(trans, path, iter) { 2927 if (prev && btree_path_cmp(prev, path) > 0) { 2928 __bch2_dump_trans_paths_updates(trans, true); 2929 panic("trans paths out of order!\n"); 2930 } 2931 prev = path; 2932 } 2933 } 2934 #else 2935 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {} 2936 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {} 2937 #endif 2938 2939 void __bch2_btree_trans_sort_paths(struct btree_trans *trans) 2940 { 2941 int i, l = 0, r = trans->nr_sorted, inc = 1; 2942 bool swapped; 2943 2944 btree_trans_verify_sorted_refs(trans); 2945 2946 if (trans->paths_sorted) 2947 goto out; 2948 2949 /* 2950 * Cocktail shaker sort: this is efficient because iterators will be 2951 * mostly sorted. 2952 */ 2953 do { 2954 swapped = false; 2955 2956 for (i = inc > 0 ? l : r - 2; 2957 i + 1 < r && i >= l; 2958 i += inc) { 2959 if (btree_path_cmp(trans->paths + trans->sorted[i], 2960 trans->paths + trans->sorted[i + 1]) > 0) { 2961 swap(trans->sorted[i], trans->sorted[i + 1]); 2962 trans->paths[trans->sorted[i]].sorted_idx = i; 2963 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1; 2964 swapped = true; 2965 } 2966 } 2967 2968 if (inc > 0) 2969 --r; 2970 else 2971 l++; 2972 inc = -inc; 2973 } while (swapped); 2974 2975 trans->paths_sorted = true; 2976 out: 2977 btree_trans_verify_sorted(trans); 2978 } 2979 2980 static inline void btree_path_list_remove(struct btree_trans *trans, 2981 struct btree_path *path) 2982 { 2983 EBUG_ON(path->sorted_idx >= trans->nr_sorted); 2984 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS 2985 trans->nr_sorted--; 2986 memmove_u64s_down_small(trans->sorted + path->sorted_idx, 2987 trans->sorted + path->sorted_idx + 1, 2988 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 2989 sizeof(u64) / sizeof(btree_path_idx_t))); 2990 #else 2991 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx); 2992 #endif 2993 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++) 2994 trans->paths[trans->sorted[i]].sorted_idx = i; 2995 } 2996 2997 static inline void btree_path_list_add(struct btree_trans *trans, 2998 btree_path_idx_t pos, 2999 btree_path_idx_t path_idx) 3000 { 3001 struct btree_path *path = trans->paths + path_idx; 3002 3003 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted; 3004 3005 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS 3006 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1, 3007 trans->sorted + path->sorted_idx, 3008 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 3009 sizeof(u64) / sizeof(btree_path_idx_t))); 3010 trans->nr_sorted++; 3011 trans->sorted[path->sorted_idx] = path_idx; 3012 #else 3013 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx); 3014 #endif 3015 3016 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++) 3017 trans->paths[trans->sorted[i]].sorted_idx = i; 3018 3019 btree_trans_verify_sorted_refs(trans); 3020 } 3021 3022 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter) 3023 { 3024 if (iter->update_path) 3025 bch2_path_put_nokeep(trans, iter->update_path, 3026 iter->flags & BTREE_ITER_intent); 3027 if (iter->path) 3028 bch2_path_put(trans, iter->path, 3029 iter->flags & BTREE_ITER_intent); 3030 if (iter->key_cache_path) 3031 bch2_path_put(trans, iter->key_cache_path, 3032 iter->flags & BTREE_ITER_intent); 3033 iter->path = 0; 3034 iter->update_path = 0; 3035 iter->key_cache_path = 0; 3036 } 3037 3038 void bch2_trans_iter_init_outlined(struct btree_trans *trans, 3039 struct btree_iter *iter, 3040 enum btree_id btree_id, struct bpos pos, 3041 unsigned flags) 3042 { 3043 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0, 3044 bch2_btree_iter_flags(trans, btree_id, 0, flags), 3045 _RET_IP_); 3046 } 3047 3048 void bch2_trans_node_iter_init(struct btree_trans *trans, 3049 struct btree_iter *iter, 3050 enum btree_id btree_id, 3051 struct bpos pos, 3052 unsigned locks_want, 3053 unsigned depth, 3054 unsigned flags) 3055 { 3056 flags |= BTREE_ITER_not_extents; 3057 flags |= BTREE_ITER_snapshot_field; 3058 flags |= BTREE_ITER_all_snapshots; 3059 3060 if (!depth && btree_id_cached(trans->c, btree_id)) 3061 flags |= BTREE_ITER_with_key_cache; 3062 3063 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth, 3064 bch2_btree_iter_flags(trans, btree_id, depth, flags), 3065 _RET_IP_); 3066 3067 iter->min_depth = depth; 3068 3069 struct btree_path *path = btree_iter_path(trans, iter); 3070 BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH)); 3071 BUG_ON(path->level != depth); 3072 BUG_ON(iter->min_depth != depth); 3073 } 3074 3075 void bch2_trans_copy_iter(struct btree_trans *trans, 3076 struct btree_iter *dst, struct btree_iter *src) 3077 { 3078 *dst = *src; 3079 #ifdef TRACK_PATH_ALLOCATED 3080 dst->ip_allocated = _RET_IP_; 3081 #endif 3082 if (src->path) 3083 __btree_path_get(trans, trans->paths + src->path, src->flags & BTREE_ITER_intent); 3084 if (src->update_path) 3085 __btree_path_get(trans, trans->paths + src->update_path, src->flags & BTREE_ITER_intent); 3086 dst->key_cache_path = 0; 3087 } 3088 3089 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size) 3090 { 3091 struct bch_fs *c = trans->c; 3092 unsigned new_top = trans->mem_top + size; 3093 unsigned old_bytes = trans->mem_bytes; 3094 unsigned new_bytes = roundup_pow_of_two(new_top); 3095 int ret; 3096 void *new_mem; 3097 void *p; 3098 3099 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX); 3100 3101 ret = trans_maybe_inject_restart(trans, _RET_IP_); 3102 if (ret) 3103 return ERR_PTR(ret); 3104 3105 struct btree_transaction_stats *s = btree_trans_stats(trans); 3106 s->max_mem = max(s->max_mem, new_bytes); 3107 3108 if (trans->used_mempool) { 3109 if (trans->mem_bytes >= new_bytes) 3110 goto out_change_top; 3111 3112 /* No more space from mempool item, need malloc new one */ 3113 new_mem = kmalloc(new_bytes, GFP_NOWAIT|__GFP_NOWARN); 3114 if (unlikely(!new_mem)) { 3115 bch2_trans_unlock(trans); 3116 3117 new_mem = kmalloc(new_bytes, GFP_KERNEL); 3118 if (!new_mem) 3119 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc); 3120 3121 ret = bch2_trans_relock(trans); 3122 if (ret) { 3123 kfree(new_mem); 3124 return ERR_PTR(ret); 3125 } 3126 } 3127 memcpy(new_mem, trans->mem, trans->mem_top); 3128 trans->used_mempool = false; 3129 mempool_free(trans->mem, &c->btree_trans_mem_pool); 3130 goto out_new_mem; 3131 } 3132 3133 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN); 3134 if (unlikely(!new_mem)) { 3135 bch2_trans_unlock(trans); 3136 3137 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL); 3138 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) { 3139 new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL); 3140 new_bytes = BTREE_TRANS_MEM_MAX; 3141 memcpy(new_mem, trans->mem, trans->mem_top); 3142 trans->used_mempool = true; 3143 kfree(trans->mem); 3144 } 3145 3146 if (!new_mem) 3147 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc); 3148 3149 trans->mem = new_mem; 3150 trans->mem_bytes = new_bytes; 3151 3152 ret = bch2_trans_relock(trans); 3153 if (ret) 3154 return ERR_PTR(ret); 3155 } 3156 out_new_mem: 3157 trans->mem = new_mem; 3158 trans->mem_bytes = new_bytes; 3159 3160 if (old_bytes) { 3161 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes); 3162 return ERR_PTR(btree_trans_restart_ip(trans, 3163 BCH_ERR_transaction_restart_mem_realloced, _RET_IP_)); 3164 } 3165 out_change_top: 3166 p = trans->mem + trans->mem_top; 3167 trans->mem_top += size; 3168 memset(p, 0, size); 3169 return p; 3170 } 3171 3172 static inline void check_srcu_held_too_long(struct btree_trans *trans) 3173 { 3174 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10), 3175 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds", 3176 (jiffies - trans->srcu_lock_time) / HZ); 3177 } 3178 3179 void bch2_trans_srcu_unlock(struct btree_trans *trans) 3180 { 3181 if (trans->srcu_held) { 3182 struct bch_fs *c = trans->c; 3183 struct btree_path *path; 3184 unsigned i; 3185 3186 trans_for_each_path(trans, path, i) 3187 if (path->cached && !btree_node_locked(path, 0)) 3188 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset); 3189 3190 check_srcu_held_too_long(trans); 3191 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); 3192 trans->srcu_held = false; 3193 } 3194 } 3195 3196 static void bch2_trans_srcu_lock(struct btree_trans *trans) 3197 { 3198 if (!trans->srcu_held) { 3199 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier); 3200 trans->srcu_lock_time = jiffies; 3201 trans->srcu_held = true; 3202 } 3203 } 3204 3205 /** 3206 * bch2_trans_begin() - reset a transaction after a interrupted attempt 3207 * @trans: transaction to reset 3208 * 3209 * Returns: current restart counter, to be used with trans_was_restarted() 3210 * 3211 * While iterating over nodes or updating nodes a attempt to lock a btree node 3212 * may return BCH_ERR_transaction_restart when the trylock fails. When this 3213 * occurs bch2_trans_begin() should be called and the transaction retried. 3214 */ 3215 u32 bch2_trans_begin(struct btree_trans *trans) 3216 { 3217 struct btree_path *path; 3218 unsigned i; 3219 u64 now; 3220 3221 bch2_trans_reset_updates(trans); 3222 3223 trans->restart_count++; 3224 trans->mem_top = 0; 3225 trans->journal_entries = NULL; 3226 3227 trans_for_each_path(trans, path, i) { 3228 path->should_be_locked = false; 3229 3230 /* 3231 * If the transaction wasn't restarted, we're presuming to be 3232 * doing something new: dont keep iterators excpt the ones that 3233 * are in use - except for the subvolumes btree: 3234 */ 3235 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes) 3236 path->preserve = false; 3237 3238 /* 3239 * XXX: we probably shouldn't be doing this if the transaction 3240 * was restarted, but currently we still overflow transaction 3241 * iterators if we do that 3242 */ 3243 if (!path->ref && !path->preserve) 3244 __bch2_path_free(trans, i); 3245 else 3246 path->preserve = false; 3247 } 3248 3249 now = local_clock(); 3250 3251 if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) && 3252 time_after64(now, trans->last_begin_time + 10)) 3253 __bch2_time_stats_update(&btree_trans_stats(trans)->duration, 3254 trans->last_begin_time, now); 3255 3256 if (!trans->restarted && 3257 (need_resched() || 3258 time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) { 3259 bch2_trans_unlock(trans); 3260 cond_resched(); 3261 now = local_clock(); 3262 } 3263 trans->last_begin_time = now; 3264 3265 if (unlikely(trans->srcu_held && 3266 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10)))) 3267 bch2_trans_srcu_unlock(trans); 3268 3269 trans->last_begin_ip = _RET_IP_; 3270 3271 #ifdef CONFIG_BCACHEFS_INJECT_TRANSACTION_RESTARTS 3272 if (trans->restarted) { 3273 trans->restart_count_this_trans++; 3274 } else { 3275 trans->restart_count_this_trans = 0; 3276 } 3277 #endif 3278 3279 trans_set_locked(trans, false); 3280 3281 if (trans->restarted) { 3282 bch2_btree_path_traverse_all(trans); 3283 trans->notrace_relock_fail = false; 3284 } 3285 3286 bch2_trans_verify_not_unlocked_or_in_restart(trans); 3287 return trans->restart_count; 3288 } 3289 3290 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" }; 3291 3292 unsigned bch2_trans_get_fn_idx(const char *fn) 3293 { 3294 for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++) 3295 if (!bch2_btree_transaction_fns[i] || 3296 bch2_btree_transaction_fns[i] == fn) { 3297 bch2_btree_transaction_fns[i] = fn; 3298 return i; 3299 } 3300 3301 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!"); 3302 return 0; 3303 } 3304 3305 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx) 3306 __acquires(&c->btree_trans_barrier) 3307 { 3308 struct btree_trans *trans; 3309 3310 if (IS_ENABLED(__KERNEL__)) { 3311 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL); 3312 if (trans) { 3313 memset(trans, 0, offsetof(struct btree_trans, list)); 3314 goto got_trans; 3315 } 3316 } 3317 3318 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS); 3319 memset(trans, 0, sizeof(*trans)); 3320 3321 seqmutex_lock(&c->btree_trans_lock); 3322 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) { 3323 struct btree_trans *pos; 3324 pid_t pid = current->pid; 3325 3326 trans->locking_wait.task = current; 3327 3328 list_for_each_entry(pos, &c->btree_trans_list, list) { 3329 struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task); 3330 /* 3331 * We'd much prefer to be stricter here and completely 3332 * disallow multiple btree_trans in the same thread - 3333 * but the data move path calls bch2_write when we 3334 * already have a btree_trans initialized. 3335 */ 3336 BUG_ON(pos_task && 3337 pid == pos_task->pid && 3338 pos->locked); 3339 } 3340 } 3341 3342 list_add(&trans->list, &c->btree_trans_list); 3343 seqmutex_unlock(&c->btree_trans_lock); 3344 got_trans: 3345 trans->c = c; 3346 trans->last_begin_time = local_clock(); 3347 trans->fn_idx = fn_idx; 3348 trans->locking_wait.task = current; 3349 trans->journal_replay_not_finished = 3350 unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)) && 3351 atomic_inc_not_zero(&c->journal_keys.ref); 3352 trans->nr_paths = ARRAY_SIZE(trans->_paths); 3353 trans->paths_allocated = trans->_paths_allocated; 3354 trans->sorted = trans->_sorted; 3355 trans->paths = trans->_paths; 3356 trans->updates = trans->_updates; 3357 3358 *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL; 3359 3360 trans->paths_allocated[0] = 1; 3361 3362 static struct lock_class_key lockdep_key; 3363 lockdep_init_map(&trans->dep_map, "bcachefs_btree", &lockdep_key, 0); 3364 3365 if (fn_idx < BCH_TRANSACTIONS_NR) { 3366 trans->fn = bch2_btree_transaction_fns[fn_idx]; 3367 3368 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx]; 3369 3370 if (s->max_mem) { 3371 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem); 3372 3373 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL); 3374 if (likely(trans->mem)) 3375 trans->mem_bytes = expected_mem_bytes; 3376 } 3377 3378 trans->nr_paths_max = s->nr_max_paths; 3379 trans->journal_entries_size = s->journal_entries_size; 3380 } 3381 3382 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier); 3383 trans->srcu_lock_time = jiffies; 3384 trans->srcu_held = true; 3385 trans_set_locked(trans, false); 3386 3387 closure_init_stack_release(&trans->ref); 3388 return trans; 3389 } 3390 3391 static void check_btree_paths_leaked(struct btree_trans *trans) 3392 { 3393 #ifdef CONFIG_BCACHEFS_DEBUG 3394 struct bch_fs *c = trans->c; 3395 struct btree_path *path; 3396 unsigned i; 3397 3398 trans_for_each_path(trans, path, i) 3399 if (path->ref) 3400 goto leaked; 3401 return; 3402 leaked: 3403 bch_err(c, "btree paths leaked from %s!", trans->fn); 3404 trans_for_each_path(trans, path, i) 3405 if (path->ref) 3406 printk(KERN_ERR " btree %s %pS\n", 3407 bch2_btree_id_str(path->btree_id), 3408 (void *) path->ip_allocated); 3409 /* Be noisy about this: */ 3410 bch2_fatal_error(c); 3411 #endif 3412 } 3413 3414 void bch2_trans_put(struct btree_trans *trans) 3415 __releases(&c->btree_trans_barrier) 3416 { 3417 struct bch_fs *c = trans->c; 3418 3419 if (trans->restarted) 3420 bch2_trans_in_restart_error(trans); 3421 3422 bch2_trans_unlock(trans); 3423 3424 trans_for_each_update(trans, i) 3425 __btree_path_put(trans, trans->paths + i->path, true); 3426 trans->nr_updates = 0; 3427 3428 check_btree_paths_leaked(trans); 3429 3430 if (trans->srcu_held) { 3431 check_srcu_held_too_long(trans); 3432 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); 3433 } 3434 3435 if (unlikely(trans->journal_replay_not_finished)) 3436 bch2_journal_keys_put(c); 3437 3438 /* 3439 * trans->ref protects trans->locking_wait.task, btree_paths array; used 3440 * by cycle detector 3441 */ 3442 closure_return_sync(&trans->ref); 3443 trans->locking_wait.task = NULL; 3444 3445 #ifdef CONFIG_BCACHEFS_DEBUG 3446 darray_exit(&trans->last_restarted_trace); 3447 #endif 3448 3449 unsigned long *paths_allocated = trans->paths_allocated; 3450 trans->paths_allocated = NULL; 3451 trans->paths = NULL; 3452 3453 if (paths_allocated != trans->_paths_allocated) 3454 kvfree_rcu_mightsleep(paths_allocated); 3455 3456 if (trans->used_mempool) 3457 mempool_free(trans->mem, &c->btree_trans_mem_pool); 3458 else 3459 kfree(trans->mem); 3460 3461 /* Userspace doesn't have a real percpu implementation: */ 3462 if (IS_ENABLED(__KERNEL__)) 3463 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans); 3464 3465 if (trans) { 3466 seqmutex_lock(&c->btree_trans_lock); 3467 list_del(&trans->list); 3468 seqmutex_unlock(&c->btree_trans_lock); 3469 3470 mempool_free(trans, &c->btree_trans_pool); 3471 } 3472 } 3473 3474 bool bch2_current_has_btree_trans(struct bch_fs *c) 3475 { 3476 seqmutex_lock(&c->btree_trans_lock); 3477 struct btree_trans *trans; 3478 bool ret = false; 3479 list_for_each_entry(trans, &c->btree_trans_list, list) 3480 if (trans->locking_wait.task == current && 3481 trans->locked) { 3482 ret = true; 3483 break; 3484 } 3485 seqmutex_unlock(&c->btree_trans_lock); 3486 return ret; 3487 } 3488 3489 static void __maybe_unused 3490 bch2_btree_bkey_cached_common_to_text(struct printbuf *out, 3491 struct btree_bkey_cached_common *b) 3492 { 3493 struct six_lock_count c = six_lock_counts(&b->lock); 3494 struct task_struct *owner; 3495 pid_t pid; 3496 3497 rcu_read_lock(); 3498 owner = READ_ONCE(b->lock.owner); 3499 pid = owner ? owner->pid : 0; 3500 rcu_read_unlock(); 3501 3502 prt_printf(out, "\t%px %c ", b, b->cached ? 'c' : 'b'); 3503 bch2_btree_id_to_text(out, b->btree_id); 3504 prt_printf(out, " l=%u:", b->level); 3505 bch2_bpos_to_text(out, btree_node_pos(b)); 3506 3507 prt_printf(out, "\t locks %u:%u:%u held by pid %u", 3508 c.n[0], c.n[1], c.n[2], pid); 3509 } 3510 3511 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans) 3512 { 3513 struct btree_bkey_cached_common *b; 3514 static char lock_types[] = { 'r', 'i', 'w' }; 3515 struct task_struct *task = READ_ONCE(trans->locking_wait.task); 3516 unsigned l, idx; 3517 3518 /* before rcu_read_lock(): */ 3519 bch2_printbuf_make_room(out, 4096); 3520 3521 if (!out->nr_tabstops) { 3522 printbuf_tabstop_push(out, 16); 3523 printbuf_tabstop_push(out, 32); 3524 } 3525 3526 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn); 3527 3528 /* trans->paths is rcu protected vs. freeing */ 3529 rcu_read_lock(); 3530 out->atomic++; 3531 3532 struct btree_path *paths = rcu_dereference(trans->paths); 3533 if (!paths) 3534 goto out; 3535 3536 unsigned long *paths_allocated = trans_paths_allocated(paths); 3537 3538 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) { 3539 struct btree_path *path = paths + idx; 3540 if (!path->nodes_locked) 3541 continue; 3542 3543 prt_printf(out, " path %u %c ", 3544 idx, 3545 path->cached ? 'c' : 'b'); 3546 bch2_btree_id_to_text(out, path->btree_id); 3547 prt_printf(out, " l=%u:", path->level); 3548 bch2_bpos_to_text(out, path->pos); 3549 prt_newline(out); 3550 3551 for (l = 0; l < BTREE_MAX_DEPTH; l++) { 3552 if (btree_node_locked(path, l) && 3553 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) { 3554 prt_printf(out, " %c l=%u ", 3555 lock_types[btree_node_locked_type(path, l)], l); 3556 bch2_btree_bkey_cached_common_to_text(out, b); 3557 prt_newline(out); 3558 } 3559 } 3560 } 3561 3562 b = READ_ONCE(trans->locking); 3563 if (b) { 3564 prt_printf(out, " blocked for %lluus on\n", 3565 div_u64(local_clock() - trans->locking_wait.start_time, 1000)); 3566 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]); 3567 bch2_btree_bkey_cached_common_to_text(out, b); 3568 prt_newline(out); 3569 } 3570 out: 3571 --out->atomic; 3572 rcu_read_unlock(); 3573 } 3574 3575 void bch2_fs_btree_iter_exit(struct bch_fs *c) 3576 { 3577 struct btree_transaction_stats *s; 3578 struct btree_trans *trans; 3579 int cpu; 3580 3581 if (c->btree_trans_bufs) 3582 for_each_possible_cpu(cpu) { 3583 struct btree_trans *trans = 3584 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans; 3585 3586 if (trans) { 3587 seqmutex_lock(&c->btree_trans_lock); 3588 list_del(&trans->list); 3589 seqmutex_unlock(&c->btree_trans_lock); 3590 } 3591 kfree(trans); 3592 } 3593 free_percpu(c->btree_trans_bufs); 3594 3595 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list); 3596 if (trans) 3597 panic("%s leaked btree_trans\n", trans->fn); 3598 3599 for (s = c->btree_transaction_stats; 3600 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); 3601 s++) { 3602 kfree(s->max_paths_text); 3603 bch2_time_stats_exit(&s->lock_hold_times); 3604 } 3605 3606 if (c->btree_trans_barrier_initialized) { 3607 synchronize_srcu_expedited(&c->btree_trans_barrier); 3608 cleanup_srcu_struct(&c->btree_trans_barrier); 3609 } 3610 mempool_exit(&c->btree_trans_mem_pool); 3611 mempool_exit(&c->btree_trans_pool); 3612 } 3613 3614 void bch2_fs_btree_iter_init_early(struct bch_fs *c) 3615 { 3616 struct btree_transaction_stats *s; 3617 3618 for (s = c->btree_transaction_stats; 3619 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); 3620 s++) { 3621 bch2_time_stats_init(&s->duration); 3622 bch2_time_stats_init(&s->lock_hold_times); 3623 mutex_init(&s->lock); 3624 } 3625 3626 INIT_LIST_HEAD(&c->btree_trans_list); 3627 seqmutex_init(&c->btree_trans_lock); 3628 } 3629 3630 int bch2_fs_btree_iter_init(struct bch_fs *c) 3631 { 3632 int ret; 3633 3634 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf); 3635 if (!c->btree_trans_bufs) 3636 return -ENOMEM; 3637 3638 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1, 3639 sizeof(struct btree_trans)) ?: 3640 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1, 3641 BTREE_TRANS_MEM_MAX) ?: 3642 init_srcu_struct(&c->btree_trans_barrier); 3643 if (ret) 3644 return ret; 3645 3646 /* 3647 * static annotation (hackily done) for lock ordering of reclaim vs. 3648 * btree node locks: 3649 */ 3650 #ifdef CONFIG_LOCKDEP 3651 fs_reclaim_acquire(GFP_KERNEL); 3652 struct btree_trans *trans = bch2_trans_get(c); 3653 trans_set_locked(trans, false); 3654 bch2_trans_put(trans); 3655 fs_reclaim_release(GFP_KERNEL); 3656 #endif 3657 3658 c->btree_trans_barrier_initialized = true; 3659 return 0; 3660 3661 } 3662