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_iter *iter) 248 { 249 struct btree_trans *trans = iter->trans; 250 251 BUG_ON(!!(iter->flags & BTREE_ITER_cached) != btree_iter_path(trans, iter)->cached); 252 253 BUG_ON((iter->flags & BTREE_ITER_is_extents) && 254 (iter->flags & BTREE_ITER_all_snapshots)); 255 256 BUG_ON(!(iter->flags & BTREE_ITER_snapshot_field) && 257 (iter->flags & BTREE_ITER_all_snapshots) && 258 !btree_type_has_snapshot_field(iter->btree_id)); 259 260 if (iter->update_path) 261 bch2_btree_path_verify(trans, &trans->paths[iter->update_path]); 262 bch2_btree_path_verify(trans, btree_iter_path(trans, iter)); 263 } 264 265 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) 266 { 267 BUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && 268 !iter->pos.snapshot); 269 270 BUG_ON(!(iter->flags & BTREE_ITER_all_snapshots) && 271 iter->pos.snapshot != iter->snapshot); 272 273 BUG_ON(iter->flags & BTREE_ITER_all_snapshots ? !bpos_eq(iter->pos, iter->k.p) : 274 !(iter->flags & BTREE_ITER_is_extents) ? !bkey_eq(iter->pos, iter->k.p) : 275 (bkey_lt(iter->pos, bkey_start_pos(&iter->k)) || 276 bkey_gt(iter->pos, iter->k.p))); 277 } 278 279 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) 280 { 281 struct btree_trans *trans = iter->trans; 282 struct btree_iter copy; 283 struct bkey_s_c prev; 284 int ret = 0; 285 286 if (!bch2_debug_check_iterators) 287 return 0; 288 289 if (!(iter->flags & BTREE_ITER_filter_snapshots)) 290 return 0; 291 292 if (bkey_err(k) || !k.k) 293 return 0; 294 295 BUG_ON(!bch2_snapshot_is_ancestor(trans->c, 296 iter->snapshot, 297 k.k->p.snapshot)); 298 299 bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos, 300 BTREE_ITER_nopreserve| 301 BTREE_ITER_all_snapshots); 302 prev = bch2_btree_iter_prev(©); 303 if (!prev.k) 304 goto out; 305 306 ret = bkey_err(prev); 307 if (ret) 308 goto out; 309 310 if (bkey_eq(prev.k->p, k.k->p) && 311 bch2_snapshot_is_ancestor(trans->c, iter->snapshot, 312 prev.k->p.snapshot) > 0) { 313 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; 314 315 bch2_bkey_to_text(&buf1, k.k); 316 bch2_bkey_to_text(&buf2, prev.k); 317 318 panic("iter snap %u\n" 319 "k %s\n" 320 "prev %s\n", 321 iter->snapshot, 322 buf1.buf, buf2.buf); 323 } 324 out: 325 bch2_trans_iter_exit(trans, ©); 326 return ret; 327 } 328 329 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id, 330 struct bpos pos) 331 { 332 bch2_trans_verify_not_unlocked_or_in_restart(trans); 333 334 struct btree_path *path; 335 struct trans_for_each_path_inorder_iter iter; 336 struct printbuf buf = PRINTBUF; 337 338 btree_trans_sort_paths(trans); 339 340 trans_for_each_path_inorder(trans, path, iter) { 341 if (path->btree_id != id || 342 !btree_node_locked(path, 0) || 343 !path->should_be_locked) 344 continue; 345 346 if (!path->cached) { 347 if (bkey_ge(pos, path->l[0].b->data->min_key) && 348 bkey_le(pos, path->l[0].b->key.k.p)) 349 return; 350 } else { 351 if (bkey_eq(pos, path->pos)) 352 return; 353 } 354 } 355 356 bch2_dump_trans_paths_updates(trans); 357 bch2_bpos_to_text(&buf, pos); 358 359 panic("not locked: %s %s\n", bch2_btree_id_str(id), buf.buf); 360 } 361 362 #else 363 364 static inline void bch2_btree_path_verify_level(struct btree_trans *trans, 365 struct btree_path *path, unsigned l) {} 366 static inline void bch2_btree_path_verify(struct btree_trans *trans, 367 struct btree_path *path) {} 368 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {} 369 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {} 370 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, 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_peek(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_peek(&l->iter, l->b)); 572 573 path->pos = k.k ? k.k->p : l->b->key.k.p; 574 trans->paths_sorted = false; 575 bch2_btree_path_verify_level(trans, path, l - path->l); 576 return k; 577 } 578 579 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans, 580 struct btree_path *path, 581 struct btree_path_level *l, 582 struct bkey *u) 583 { 584 struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u, 585 bch2_btree_node_iter_prev(&l->iter, l->b)); 586 587 path->pos = k.k ? k.k->p : l->b->data->min_key; 588 trans->paths_sorted = false; 589 bch2_btree_path_verify_level(trans, path, l - path->l); 590 return k; 591 } 592 593 static inline bool btree_path_advance_to_pos(struct btree_path *path, 594 struct btree_path_level *l, 595 int max_advance) 596 { 597 struct bkey_packed *k; 598 int nr_advanced = 0; 599 600 while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) && 601 bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) { 602 if (max_advance > 0 && nr_advanced >= max_advance) 603 return false; 604 605 bch2_btree_node_iter_advance(&l->iter, l->b); 606 nr_advanced++; 607 } 608 609 return true; 610 } 611 612 static inline void __btree_path_level_init(struct btree_path *path, 613 unsigned level) 614 { 615 struct btree_path_level *l = &path->l[level]; 616 617 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos); 618 619 /* 620 * Iterators to interior nodes should always be pointed at the first non 621 * whiteout: 622 */ 623 if (level) 624 bch2_btree_node_iter_peek(&l->iter, l->b); 625 } 626 627 void bch2_btree_path_level_init(struct btree_trans *trans, 628 struct btree_path *path, 629 struct btree *b) 630 { 631 BUG_ON(path->cached); 632 633 EBUG_ON(!btree_path_pos_in_node(path, b)); 634 635 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock); 636 path->l[b->c.level].b = b; 637 __btree_path_level_init(path, b->c.level); 638 } 639 640 /* Btree path: fixups after btree node updates: */ 641 642 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b) 643 { 644 struct bch_fs *c = trans->c; 645 646 trans_for_each_update(trans, i) 647 if (!i->cached && 648 i->level == b->c.level && 649 i->btree_id == b->c.btree_id && 650 bpos_cmp(i->k->k.p, b->data->min_key) >= 0 && 651 bpos_cmp(i->k->k.p, b->data->max_key) <= 0) { 652 i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v; 653 654 if (unlikely(trans->journal_replay_not_finished)) { 655 struct bkey_i *j_k = 656 bch2_journal_keys_peek_slot(c, i->btree_id, i->level, 657 i->k->k.p); 658 659 if (j_k) { 660 i->old_k = j_k->k; 661 i->old_v = &j_k->v; 662 } 663 } 664 } 665 } 666 667 /* 668 * A btree node is being replaced - update the iterator to point to the new 669 * node: 670 */ 671 void bch2_trans_node_add(struct btree_trans *trans, 672 struct btree_path *path, 673 struct btree *b) 674 { 675 struct btree_path *prev; 676 677 BUG_ON(!btree_path_pos_in_node(path, b)); 678 679 while ((prev = prev_btree_path(trans, path)) && 680 btree_path_pos_in_node(prev, b)) 681 path = prev; 682 683 for (; 684 path && btree_path_pos_in_node(path, b); 685 path = next_btree_path(trans, path)) 686 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) { 687 enum btree_node_locked_type t = 688 btree_lock_want(path, b->c.level); 689 690 if (t != BTREE_NODE_UNLOCKED) { 691 btree_node_unlock(trans, path, b->c.level); 692 six_lock_increment(&b->c.lock, (enum six_lock_type) t); 693 mark_btree_node_locked(trans, path, b->c.level, t); 694 } 695 696 bch2_btree_path_level_init(trans, path, b); 697 } 698 699 bch2_trans_revalidate_updates_in_node(trans, b); 700 } 701 702 void bch2_trans_node_drop(struct btree_trans *trans, 703 struct btree *b) 704 { 705 struct btree_path *path; 706 unsigned i, level = b->c.level; 707 708 trans_for_each_path(trans, path, i) 709 if (path->l[level].b == b) { 710 btree_node_unlock(trans, path, level); 711 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 712 } 713 } 714 715 /* 716 * A btree node has been modified in such a way as to invalidate iterators - fix 717 * them: 718 */ 719 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b) 720 { 721 struct btree_path *path; 722 unsigned i; 723 724 trans_for_each_path_with_node(trans, b, path, i) 725 __btree_path_level_init(path, b->c.level); 726 727 bch2_trans_revalidate_updates_in_node(trans, b); 728 } 729 730 /* Btree path: traverse, set_pos: */ 731 732 static inline int btree_path_lock_root(struct btree_trans *trans, 733 struct btree_path *path, 734 unsigned depth_want, 735 unsigned long trace_ip) 736 { 737 struct bch_fs *c = trans->c; 738 struct btree_root *r = bch2_btree_id_root(c, path->btree_id); 739 enum six_lock_type lock_type; 740 unsigned i; 741 int ret; 742 743 EBUG_ON(path->nodes_locked); 744 745 while (1) { 746 struct btree *b = READ_ONCE(r->b); 747 if (unlikely(!b)) { 748 BUG_ON(!r->error); 749 return r->error; 750 } 751 752 path->level = READ_ONCE(b->c.level); 753 754 if (unlikely(path->level < depth_want)) { 755 /* 756 * the root is at a lower depth than the depth we want: 757 * got to the end of the btree, or we're walking nodes 758 * greater than some depth and there are no nodes >= 759 * that depth 760 */ 761 path->level = depth_want; 762 for (i = path->level; i < BTREE_MAX_DEPTH; i++) 763 path->l[i].b = NULL; 764 return 1; 765 } 766 767 lock_type = __btree_lock_want(path, path->level); 768 ret = btree_node_lock(trans, path, &b->c, 769 path->level, lock_type, trace_ip); 770 if (unlikely(ret)) { 771 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) 772 return ret; 773 BUG(); 774 } 775 776 if (likely(b == READ_ONCE(r->b) && 777 b->c.level == path->level && 778 !race_fault())) { 779 for (i = 0; i < path->level; i++) 780 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root); 781 path->l[path->level].b = b; 782 for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++) 783 path->l[i].b = NULL; 784 785 mark_btree_node_locked(trans, path, path->level, 786 (enum btree_node_locked_type) lock_type); 787 bch2_btree_path_level_init(trans, path, b); 788 return 0; 789 } 790 791 six_unlock_type(&b->c.lock, lock_type); 792 } 793 } 794 795 noinline 796 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path) 797 { 798 struct bch_fs *c = trans->c; 799 struct btree_path_level *l = path_l(path); 800 struct btree_node_iter node_iter = l->iter; 801 struct bkey_packed *k; 802 struct bkey_buf tmp; 803 unsigned nr = test_bit(BCH_FS_started, &c->flags) 804 ? (path->level > 1 ? 0 : 2) 805 : (path->level > 1 ? 1 : 16); 806 bool was_locked = btree_node_locked(path, path->level); 807 int ret = 0; 808 809 bch2_bkey_buf_init(&tmp); 810 811 while (nr-- && !ret) { 812 if (!bch2_btree_node_relock(trans, path, path->level)) 813 break; 814 815 bch2_btree_node_iter_advance(&node_iter, l->b); 816 k = bch2_btree_node_iter_peek(&node_iter, l->b); 817 if (!k) 818 break; 819 820 bch2_bkey_buf_unpack(&tmp, c, l->b, k); 821 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, 822 path->level - 1); 823 } 824 825 if (!was_locked) 826 btree_node_unlock(trans, path, path->level); 827 828 bch2_bkey_buf_exit(&tmp, c); 829 return ret; 830 } 831 832 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path, 833 struct btree_and_journal_iter *jiter) 834 { 835 struct bch_fs *c = trans->c; 836 struct bkey_s_c k; 837 struct bkey_buf tmp; 838 unsigned nr = test_bit(BCH_FS_started, &c->flags) 839 ? (path->level > 1 ? 0 : 2) 840 : (path->level > 1 ? 1 : 16); 841 bool was_locked = btree_node_locked(path, path->level); 842 int ret = 0; 843 844 bch2_bkey_buf_init(&tmp); 845 846 jiter->fail_if_too_many_whiteouts = true; 847 848 while (nr-- && !ret) { 849 if (!bch2_btree_node_relock(trans, path, path->level)) 850 break; 851 852 bch2_btree_and_journal_iter_advance(jiter); 853 k = bch2_btree_and_journal_iter_peek(jiter); 854 if (!k.k) 855 break; 856 857 bch2_bkey_buf_reassemble(&tmp, c, k); 858 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, 859 path->level - 1); 860 } 861 862 if (!was_locked) 863 btree_node_unlock(trans, path, path->level); 864 865 bch2_bkey_buf_exit(&tmp, c); 866 return ret; 867 } 868 869 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans, 870 struct btree_path *path, 871 unsigned plevel, struct btree *b) 872 { 873 struct btree_path_level *l = &path->l[plevel]; 874 bool locked = btree_node_locked(path, plevel); 875 struct bkey_packed *k; 876 struct bch_btree_ptr_v2 *bp; 877 878 if (!bch2_btree_node_relock(trans, path, plevel)) 879 return; 880 881 k = bch2_btree_node_iter_peek_all(&l->iter, l->b); 882 BUG_ON(k->type != KEY_TYPE_btree_ptr_v2); 883 884 bp = (void *) bkeyp_val(&l->b->format, k); 885 bp->mem_ptr = (unsigned long)b; 886 887 if (!locked) 888 btree_node_unlock(trans, path, plevel); 889 } 890 891 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans, 892 struct btree_path *path, 893 unsigned flags, 894 struct bkey_buf *out) 895 { 896 struct bch_fs *c = trans->c; 897 struct btree_path_level *l = path_l(path); 898 struct btree_and_journal_iter jiter; 899 struct bkey_s_c k; 900 int ret = 0; 901 902 __bch2_btree_and_journal_iter_init_node_iter(trans, &jiter, l->b, l->iter, path->pos); 903 904 k = bch2_btree_and_journal_iter_peek(&jiter); 905 if (!k.k) { 906 struct printbuf buf = PRINTBUF; 907 908 prt_str(&buf, "node not found at pos "); 909 bch2_bpos_to_text(&buf, path->pos); 910 prt_str(&buf, " at btree "); 911 bch2_btree_pos_to_text(&buf, c, l->b); 912 913 ret = bch2_fs_topology_error(c, "%s", buf.buf); 914 printbuf_exit(&buf); 915 goto err; 916 } 917 918 bch2_bkey_buf_reassemble(out, c, k); 919 920 if ((flags & BTREE_ITER_prefetch) && 921 c->opts.btree_node_prefetch) 922 ret = btree_path_prefetch_j(trans, path, &jiter); 923 924 err: 925 bch2_btree_and_journal_iter_exit(&jiter); 926 return ret; 927 } 928 929 static __always_inline int btree_path_down(struct btree_trans *trans, 930 struct btree_path *path, 931 unsigned flags, 932 unsigned long trace_ip) 933 { 934 struct bch_fs *c = trans->c; 935 struct btree_path_level *l = path_l(path); 936 struct btree *b; 937 unsigned level = path->level - 1; 938 enum six_lock_type lock_type = __btree_lock_want(path, level); 939 struct bkey_buf tmp; 940 int ret; 941 942 EBUG_ON(!btree_node_locked(path, path->level)); 943 944 bch2_bkey_buf_init(&tmp); 945 946 if (unlikely(trans->journal_replay_not_finished)) { 947 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp); 948 if (ret) 949 goto err; 950 } else { 951 struct bkey_packed *k = bch2_btree_node_iter_peek(&l->iter, l->b); 952 if (!k) { 953 struct printbuf buf = PRINTBUF; 954 955 prt_str(&buf, "node not found at pos "); 956 bch2_bpos_to_text(&buf, path->pos); 957 prt_str(&buf, " within parent node "); 958 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&l->b->key)); 959 960 bch2_fs_fatal_error(c, "%s", buf.buf); 961 printbuf_exit(&buf); 962 ret = -BCH_ERR_btree_need_topology_repair; 963 goto err; 964 } 965 966 bch2_bkey_buf_unpack(&tmp, c, l->b, k); 967 968 if ((flags & BTREE_ITER_prefetch) && 969 c->opts.btree_node_prefetch) { 970 ret = btree_path_prefetch(trans, path); 971 if (ret) 972 goto err; 973 } 974 } 975 976 b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip); 977 ret = PTR_ERR_OR_ZERO(b); 978 if (unlikely(ret)) 979 goto err; 980 981 if (likely(!trans->journal_replay_not_finished && 982 tmp.k->k.type == KEY_TYPE_btree_ptr_v2) && 983 unlikely(b != btree_node_mem_ptr(tmp.k))) 984 btree_node_mem_ptr_set(trans, path, level + 1, b); 985 986 if (btree_node_read_locked(path, level + 1)) 987 btree_node_unlock(trans, path, level + 1); 988 989 mark_btree_node_locked(trans, path, level, 990 (enum btree_node_locked_type) lock_type); 991 path->level = level; 992 bch2_btree_path_level_init(trans, path, b); 993 994 bch2_btree_path_verify_locks(path); 995 err: 996 bch2_bkey_buf_exit(&tmp, c); 997 return ret; 998 } 999 1000 static int bch2_btree_path_traverse_all(struct btree_trans *trans) 1001 { 1002 struct bch_fs *c = trans->c; 1003 struct btree_path *path; 1004 unsigned long trace_ip = _RET_IP_; 1005 unsigned i; 1006 int ret = 0; 1007 1008 if (trans->in_traverse_all) 1009 return -BCH_ERR_transaction_restart_in_traverse_all; 1010 1011 trans->in_traverse_all = true; 1012 retry_all: 1013 trans->restarted = 0; 1014 trans->last_restarted_ip = 0; 1015 1016 trans_for_each_path(trans, path, i) 1017 path->should_be_locked = false; 1018 1019 btree_trans_sort_paths(trans); 1020 1021 bch2_trans_unlock(trans); 1022 cond_resched(); 1023 trans_set_locked(trans, false); 1024 1025 if (unlikely(trans->memory_allocation_failure)) { 1026 struct closure cl; 1027 1028 closure_init_stack(&cl); 1029 1030 do { 1031 ret = bch2_btree_cache_cannibalize_lock(trans, &cl); 1032 closure_sync(&cl); 1033 } while (ret); 1034 } 1035 1036 /* Now, redo traversals in correct order: */ 1037 i = 0; 1038 while (i < trans->nr_sorted) { 1039 btree_path_idx_t idx = trans->sorted[i]; 1040 1041 /* 1042 * Traversing a path can cause another path to be added at about 1043 * the same position: 1044 */ 1045 if (trans->paths[idx].uptodate) { 1046 __btree_path_get(trans, &trans->paths[idx], false); 1047 ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_); 1048 __btree_path_put(trans, &trans->paths[idx], false); 1049 1050 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || 1051 bch2_err_matches(ret, ENOMEM)) 1052 goto retry_all; 1053 if (ret) 1054 goto err; 1055 } else { 1056 i++; 1057 } 1058 } 1059 1060 /* 1061 * We used to assert that all paths had been traversed here 1062 * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since 1063 * path->should_be_locked is not set yet, we might have unlocked and 1064 * then failed to relock a path - that's fine. 1065 */ 1066 err: 1067 bch2_btree_cache_cannibalize_unlock(trans); 1068 1069 trans->in_traverse_all = false; 1070 1071 trace_and_count(c, trans_traverse_all, trans, trace_ip); 1072 return ret; 1073 } 1074 1075 static inline bool btree_path_check_pos_in_node(struct btree_path *path, 1076 unsigned l, int check_pos) 1077 { 1078 if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b)) 1079 return false; 1080 if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b)) 1081 return false; 1082 return true; 1083 } 1084 1085 static inline bool btree_path_good_node(struct btree_trans *trans, 1086 struct btree_path *path, 1087 unsigned l, int check_pos) 1088 { 1089 return is_btree_node(path, l) && 1090 bch2_btree_node_relock(trans, path, l) && 1091 btree_path_check_pos_in_node(path, l, check_pos); 1092 } 1093 1094 static void btree_path_set_level_down(struct btree_trans *trans, 1095 struct btree_path *path, 1096 unsigned new_level) 1097 { 1098 unsigned l; 1099 1100 path->level = new_level; 1101 1102 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++) 1103 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED) 1104 btree_node_unlock(trans, path, l); 1105 1106 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1107 bch2_btree_path_verify(trans, path); 1108 } 1109 1110 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans, 1111 struct btree_path *path, 1112 int check_pos) 1113 { 1114 unsigned i, l = path->level; 1115 again: 1116 while (btree_path_node(path, l) && 1117 !btree_path_good_node(trans, path, l, check_pos)) 1118 __btree_path_set_level_up(trans, path, l++); 1119 1120 /* If we need intent locks, take them too: */ 1121 for (i = l + 1; 1122 i < path->locks_want && btree_path_node(path, i); 1123 i++) 1124 if (!bch2_btree_node_relock(trans, path, i)) { 1125 while (l <= i) 1126 __btree_path_set_level_up(trans, path, l++); 1127 goto again; 1128 } 1129 1130 return l; 1131 } 1132 1133 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans, 1134 struct btree_path *path, 1135 int check_pos) 1136 { 1137 return likely(btree_node_locked(path, path->level) && 1138 btree_path_check_pos_in_node(path, path->level, check_pos)) 1139 ? path->level 1140 : __btree_path_up_until_good_node(trans, path, check_pos); 1141 } 1142 1143 /* 1144 * This is the main state machine for walking down the btree - walks down to a 1145 * specified depth 1146 * 1147 * Returns 0 on success, -EIO on error (error reading in a btree node). 1148 * 1149 * On error, caller (peek_node()/peek_key()) must return NULL; the error is 1150 * stashed in the iterator and returned from bch2_trans_exit(). 1151 */ 1152 int bch2_btree_path_traverse_one(struct btree_trans *trans, 1153 btree_path_idx_t path_idx, 1154 unsigned flags, 1155 unsigned long trace_ip) 1156 { 1157 struct btree_path *path = &trans->paths[path_idx]; 1158 unsigned depth_want = path->level; 1159 int ret = -((int) trans->restarted); 1160 1161 if (unlikely(ret)) 1162 goto out; 1163 1164 if (unlikely(!trans->srcu_held)) 1165 bch2_trans_srcu_lock(trans); 1166 1167 trace_btree_path_traverse_start(trans, path); 1168 1169 /* 1170 * Ensure we obey path->should_be_locked: if it's set, we can't unlock 1171 * and re-traverse the path without a transaction restart: 1172 */ 1173 if (path->should_be_locked) { 1174 ret = bch2_btree_path_relock(trans, path, trace_ip); 1175 goto out; 1176 } 1177 1178 if (path->cached) { 1179 ret = bch2_btree_path_traverse_cached(trans, path, flags); 1180 goto out; 1181 } 1182 1183 path = &trans->paths[path_idx]; 1184 1185 if (unlikely(path->level >= BTREE_MAX_DEPTH)) 1186 goto out_uptodate; 1187 1188 path->level = btree_path_up_until_good_node(trans, path, 0); 1189 unsigned max_level = path->level; 1190 1191 EBUG_ON(btree_path_node(path, path->level) && 1192 !btree_node_locked(path, path->level)); 1193 1194 /* 1195 * Note: path->nodes[path->level] may be temporarily NULL here - that 1196 * would indicate to other code that we got to the end of the btree, 1197 * here it indicates that relocking the root failed - it's critical that 1198 * btree_path_lock_root() comes next and that it can't fail 1199 */ 1200 while (path->level > depth_want) { 1201 ret = btree_path_node(path, path->level) 1202 ? btree_path_down(trans, path, flags, trace_ip) 1203 : btree_path_lock_root(trans, path, depth_want, trace_ip); 1204 if (unlikely(ret)) { 1205 if (ret == 1) { 1206 /* 1207 * No nodes at this level - got to the end of 1208 * the btree: 1209 */ 1210 ret = 0; 1211 goto out; 1212 } 1213 1214 __bch2_btree_path_unlock(trans, path); 1215 path->level = depth_want; 1216 path->l[path->level].b = ERR_PTR(ret); 1217 goto out; 1218 } 1219 } 1220 1221 if (unlikely(max_level > path->level)) { 1222 struct btree_path *linked; 1223 unsigned iter; 1224 1225 trans_for_each_path_with_node(trans, path_l(path)->b, linked, iter) 1226 for (unsigned j = path->level + 1; j < max_level; j++) 1227 linked->l[j] = path->l[j]; 1228 } 1229 1230 out_uptodate: 1231 path->uptodate = BTREE_ITER_UPTODATE; 1232 trace_btree_path_traverse_end(trans, path); 1233 out: 1234 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted) 1235 panic("ret %s (%i) trans->restarted %s (%i)\n", 1236 bch2_err_str(ret), ret, 1237 bch2_err_str(trans->restarted), trans->restarted); 1238 bch2_btree_path_verify(trans, path); 1239 return ret; 1240 } 1241 1242 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst, 1243 struct btree_path *src) 1244 { 1245 unsigned i, offset = offsetof(struct btree_path, pos); 1246 1247 memcpy((void *) dst + offset, 1248 (void *) src + offset, 1249 sizeof(struct btree_path) - offset); 1250 1251 for (i = 0; i < BTREE_MAX_DEPTH; i++) { 1252 unsigned t = btree_node_locked_type(dst, i); 1253 1254 if (t != BTREE_NODE_UNLOCKED) 1255 six_lock_increment(&dst->l[i].b->c.lock, t); 1256 } 1257 } 1258 1259 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src, 1260 bool intent, unsigned long ip) 1261 { 1262 btree_path_idx_t new = btree_path_alloc(trans, src); 1263 btree_path_copy(trans, trans->paths + new, trans->paths + src); 1264 __btree_path_get(trans, trans->paths + new, intent); 1265 #ifdef TRACK_PATH_ALLOCATED 1266 trans->paths[new].ip_allocated = ip; 1267 #endif 1268 return new; 1269 } 1270 1271 __flatten 1272 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans, 1273 btree_path_idx_t path, bool intent, unsigned long ip) 1274 { 1275 struct btree_path *old = trans->paths + path; 1276 __btree_path_put(trans, trans->paths + path, intent); 1277 path = btree_path_clone(trans, path, intent, ip); 1278 trace_btree_path_clone(trans, old, trans->paths + path); 1279 trans->paths[path].preserve = false; 1280 return path; 1281 } 1282 1283 btree_path_idx_t __must_check 1284 __bch2_btree_path_set_pos(struct btree_trans *trans, 1285 btree_path_idx_t path_idx, struct bpos new_pos, 1286 bool intent, unsigned long ip) 1287 { 1288 int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos); 1289 1290 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1291 EBUG_ON(!trans->paths[path_idx].ref); 1292 1293 trace_btree_path_set_pos(trans, trans->paths + path_idx, &new_pos); 1294 1295 path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip); 1296 1297 struct btree_path *path = trans->paths + path_idx; 1298 path->pos = new_pos; 1299 trans->paths_sorted = false; 1300 1301 if (unlikely(path->cached)) { 1302 btree_node_unlock(trans, path, 0); 1303 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up); 1304 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1305 goto out; 1306 } 1307 1308 unsigned level = btree_path_up_until_good_node(trans, path, cmp); 1309 1310 if (btree_path_node(path, level)) { 1311 struct btree_path_level *l = &path->l[level]; 1312 1313 BUG_ON(!btree_node_locked(path, level)); 1314 /* 1315 * We might have to skip over many keys, or just a few: try 1316 * advancing the node iterator, and if we have to skip over too 1317 * many keys just reinit it (or if we're rewinding, since that 1318 * is expensive). 1319 */ 1320 if (cmp < 0 || 1321 !btree_path_advance_to_pos(path, l, 8)) 1322 bch2_btree_node_iter_init(&l->iter, l->b, &path->pos); 1323 1324 /* 1325 * Iterators to interior nodes should always be pointed at the first non 1326 * whiteout: 1327 */ 1328 if (unlikely(level)) 1329 bch2_btree_node_iter_peek(&l->iter, l->b); 1330 } 1331 1332 if (unlikely(level != path->level)) { 1333 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 1334 __bch2_btree_path_unlock(trans, path); 1335 } 1336 out: 1337 bch2_btree_path_verify(trans, path); 1338 return path_idx; 1339 } 1340 1341 /* Btree path: main interface: */ 1342 1343 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path) 1344 { 1345 struct btree_path *sib; 1346 1347 sib = prev_btree_path(trans, path); 1348 if (sib && !btree_path_cmp(sib, path)) 1349 return sib; 1350 1351 sib = next_btree_path(trans, path); 1352 if (sib && !btree_path_cmp(sib, path)) 1353 return sib; 1354 1355 return NULL; 1356 } 1357 1358 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path) 1359 { 1360 struct btree_path *sib; 1361 1362 sib = prev_btree_path(trans, path); 1363 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) 1364 return sib; 1365 1366 sib = next_btree_path(trans, path); 1367 if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) 1368 return sib; 1369 1370 return NULL; 1371 } 1372 1373 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path) 1374 { 1375 __bch2_btree_path_unlock(trans, trans->paths + path); 1376 btree_path_list_remove(trans, trans->paths + path); 1377 __clear_bit(path, trans->paths_allocated); 1378 } 1379 1380 static bool bch2_btree_path_can_relock(struct btree_trans *trans, struct btree_path *path) 1381 { 1382 unsigned l = path->level; 1383 1384 do { 1385 if (!btree_path_node(path, l)) 1386 break; 1387 1388 if (!is_btree_node(path, l)) 1389 return false; 1390 1391 if (path->l[l].lock_seq != path->l[l].b->c.lock.seq) 1392 return false; 1393 1394 l++; 1395 } while (l < path->locks_want); 1396 1397 return true; 1398 } 1399 1400 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent) 1401 { 1402 struct btree_path *path = trans->paths + path_idx, *dup; 1403 1404 if (!__btree_path_put(trans, path, intent)) 1405 return; 1406 1407 dup = path->preserve 1408 ? have_path_at_pos(trans, path) 1409 : have_node_at_pos(trans, path); 1410 1411 trace_btree_path_free(trans, path_idx, dup); 1412 1413 if (!dup && !(!path->preserve && !is_btree_node(path, path->level))) 1414 return; 1415 1416 if (path->should_be_locked && !trans->restarted) { 1417 if (!dup) 1418 return; 1419 1420 if (!(trans->locked 1421 ? bch2_btree_path_relock_norestart(trans, dup) 1422 : bch2_btree_path_can_relock(trans, dup))) 1423 return; 1424 } 1425 1426 if (dup) { 1427 dup->preserve |= path->preserve; 1428 dup->should_be_locked |= path->should_be_locked; 1429 } 1430 1431 __bch2_path_free(trans, path_idx); 1432 } 1433 1434 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path, 1435 bool intent) 1436 { 1437 if (!__btree_path_put(trans, trans->paths + path, intent)) 1438 return; 1439 1440 __bch2_path_free(trans, path); 1441 } 1442 1443 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count) 1444 { 1445 panic("trans->restart_count %u, should be %u, last restarted by %pS\n", 1446 trans->restart_count, restart_count, 1447 (void *) trans->last_begin_ip); 1448 } 1449 1450 static void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans) 1451 { 1452 #ifdef CONFIG_BCACHEFS_DEBUG 1453 struct printbuf buf = PRINTBUF; 1454 bch2_prt_backtrace(&buf, &trans->last_restarted_trace); 1455 panic("in transaction restart: %s, last restarted by\n%s", 1456 bch2_err_str(trans->restarted), 1457 buf.buf); 1458 #else 1459 panic("in transaction restart: %s, last restarted by %pS\n", 1460 bch2_err_str(trans->restarted), 1461 (void *) trans->last_restarted_ip); 1462 #endif 1463 } 1464 1465 void __noreturn bch2_trans_unlocked_or_in_restart_error(struct btree_trans *trans) 1466 { 1467 if (trans->restarted) 1468 bch2_trans_in_restart_error(trans); 1469 1470 if (!trans->locked) 1471 panic("trans should be locked, unlocked by %pS\n", 1472 (void *) trans->last_unlock_ip); 1473 1474 BUG(); 1475 } 1476 1477 noinline __cold 1478 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans) 1479 { 1480 prt_printf(buf, "%u transaction updates for %s journal seq %llu\n", 1481 trans->nr_updates, trans->fn, trans->journal_res.seq); 1482 printbuf_indent_add(buf, 2); 1483 1484 trans_for_each_update(trans, i) { 1485 struct bkey_s_c old = { &i->old_k, i->old_v }; 1486 1487 prt_str(buf, "update: btree="); 1488 bch2_btree_id_to_text(buf, i->btree_id); 1489 prt_printf(buf, " cached=%u %pS\n", 1490 i->cached, 1491 (void *) i->ip_allocated); 1492 1493 prt_printf(buf, " old "); 1494 bch2_bkey_val_to_text(buf, trans->c, old); 1495 prt_newline(buf); 1496 1497 prt_printf(buf, " new "); 1498 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k)); 1499 prt_newline(buf); 1500 } 1501 1502 for (struct jset_entry *e = trans->journal_entries; 1503 e != btree_trans_journal_entries_top(trans); 1504 e = vstruct_next(e)) 1505 bch2_journal_entry_to_text(buf, trans->c, e); 1506 1507 printbuf_indent_sub(buf, 2); 1508 } 1509 1510 noinline __cold 1511 void bch2_dump_trans_updates(struct btree_trans *trans) 1512 { 1513 struct printbuf buf = PRINTBUF; 1514 1515 bch2_trans_updates_to_text(&buf, trans); 1516 bch2_print_str(trans->c, buf.buf); 1517 printbuf_exit(&buf); 1518 } 1519 1520 static void bch2_btree_path_to_text_short(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx) 1521 { 1522 struct btree_path *path = trans->paths + path_idx; 1523 1524 prt_printf(out, "path: idx %3u ref %u:%u %c %c %c ", 1525 path_idx, path->ref, path->intent_ref, 1526 path->preserve ? 'P' : ' ', 1527 path->should_be_locked ? 'S' : ' ', 1528 path->cached ? 'C' : 'B'); 1529 bch2_btree_id_level_to_text(out, path->btree_id, path->level); 1530 prt_str(out, " pos "); 1531 bch2_bpos_to_text(out, path->pos); 1532 1533 if (!path->cached && btree_node_locked(path, path->level)) { 1534 prt_char(out, ' '); 1535 struct btree *b = path_l(path)->b; 1536 bch2_bpos_to_text(out, b->data->min_key); 1537 prt_char(out, '-'); 1538 bch2_bpos_to_text(out, b->key.k.p); 1539 } 1540 1541 #ifdef TRACK_PATH_ALLOCATED 1542 prt_printf(out, " %pS", (void *) path->ip_allocated); 1543 #endif 1544 } 1545 1546 static const char *btree_node_locked_str(enum btree_node_locked_type t) 1547 { 1548 switch (t) { 1549 case BTREE_NODE_UNLOCKED: 1550 return "unlocked"; 1551 case BTREE_NODE_READ_LOCKED: 1552 return "read"; 1553 case BTREE_NODE_INTENT_LOCKED: 1554 return "intent"; 1555 case BTREE_NODE_WRITE_LOCKED: 1556 return "write"; 1557 default: 1558 return NULL; 1559 } 1560 } 1561 1562 void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx) 1563 { 1564 bch2_btree_path_to_text_short(out, trans, path_idx); 1565 1566 struct btree_path *path = trans->paths + path_idx; 1567 1568 prt_printf(out, " uptodate %u locks_want %u", path->uptodate, path->locks_want); 1569 prt_newline(out); 1570 1571 printbuf_indent_add(out, 2); 1572 for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) { 1573 prt_printf(out, "l=%u locks %s seq %u node ", l, 1574 btree_node_locked_str(btree_node_locked_type(path, l)), 1575 path->l[l].lock_seq); 1576 1577 int ret = PTR_ERR_OR_ZERO(path->l[l].b); 1578 if (ret) 1579 prt_str(out, bch2_err_str(ret)); 1580 else 1581 prt_printf(out, "%px", path->l[l].b); 1582 prt_newline(out); 1583 } 1584 printbuf_indent_sub(out, 2); 1585 } 1586 1587 static noinline __cold 1588 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans, 1589 bool nosort) 1590 { 1591 struct trans_for_each_path_inorder_iter iter; 1592 1593 if (!nosort) 1594 btree_trans_sort_paths(trans); 1595 1596 trans_for_each_path_idx_inorder(trans, iter) { 1597 bch2_btree_path_to_text_short(out, trans, iter.path_idx); 1598 prt_newline(out); 1599 } 1600 } 1601 1602 noinline __cold 1603 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans) 1604 { 1605 __bch2_trans_paths_to_text(out, trans, false); 1606 } 1607 1608 static noinline __cold 1609 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort) 1610 { 1611 struct printbuf buf = PRINTBUF; 1612 1613 __bch2_trans_paths_to_text(&buf, trans, nosort); 1614 bch2_trans_updates_to_text(&buf, trans); 1615 1616 bch2_print_str(trans->c, buf.buf); 1617 printbuf_exit(&buf); 1618 } 1619 1620 noinline __cold 1621 void bch2_dump_trans_paths_updates(struct btree_trans *trans) 1622 { 1623 __bch2_dump_trans_paths_updates(trans, false); 1624 } 1625 1626 noinline __cold 1627 static void bch2_trans_update_max_paths(struct btree_trans *trans) 1628 { 1629 struct btree_transaction_stats *s = btree_trans_stats(trans); 1630 struct printbuf buf = PRINTBUF; 1631 size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths); 1632 1633 bch2_trans_paths_to_text(&buf, trans); 1634 1635 if (!buf.allocation_failure) { 1636 mutex_lock(&s->lock); 1637 if (nr > s->nr_max_paths) { 1638 s->nr_max_paths = nr; 1639 swap(s->max_paths_text, buf.buf); 1640 } 1641 mutex_unlock(&s->lock); 1642 } 1643 1644 printbuf_exit(&buf); 1645 1646 trans->nr_paths_max = nr; 1647 } 1648 1649 noinline __cold 1650 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans) 1651 { 1652 if (trace_trans_restart_too_many_iters_enabled()) { 1653 struct printbuf buf = PRINTBUF; 1654 1655 bch2_trans_paths_to_text(&buf, trans); 1656 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf); 1657 printbuf_exit(&buf); 1658 } 1659 1660 count_event(trans->c, trans_restart_too_many_iters); 1661 1662 return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters); 1663 } 1664 1665 static noinline void btree_path_overflow(struct btree_trans *trans) 1666 { 1667 bch2_dump_trans_paths_updates(trans); 1668 bch_err(trans->c, "trans path overflow"); 1669 } 1670 1671 static noinline void btree_paths_realloc(struct btree_trans *trans) 1672 { 1673 unsigned nr = trans->nr_paths * 2; 1674 1675 void *p = kvzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) + 1676 sizeof(struct btree_trans_paths) + 1677 nr * sizeof(struct btree_path) + 1678 nr * sizeof(btree_path_idx_t) + 8 + 1679 nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL); 1680 1681 unsigned long *paths_allocated = p; 1682 memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long)); 1683 p += BITS_TO_LONGS(nr) * sizeof(unsigned long); 1684 1685 p += sizeof(struct btree_trans_paths); 1686 struct btree_path *paths = p; 1687 *trans_paths_nr(paths) = nr; 1688 memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path)); 1689 p += nr * sizeof(struct btree_path); 1690 1691 btree_path_idx_t *sorted = p; 1692 memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t)); 1693 p += nr * sizeof(btree_path_idx_t) + 8; 1694 1695 struct btree_insert_entry *updates = p; 1696 memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry)); 1697 1698 unsigned long *old = trans->paths_allocated; 1699 1700 rcu_assign_pointer(trans->paths_allocated, paths_allocated); 1701 rcu_assign_pointer(trans->paths, paths); 1702 rcu_assign_pointer(trans->sorted, sorted); 1703 rcu_assign_pointer(trans->updates, updates); 1704 1705 trans->nr_paths = nr; 1706 1707 if (old != trans->_paths_allocated) 1708 kfree_rcu_mightsleep(old); 1709 } 1710 1711 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans, 1712 btree_path_idx_t pos) 1713 { 1714 btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths); 1715 1716 if (unlikely(idx == trans->nr_paths)) { 1717 if (trans->nr_paths == BTREE_ITER_MAX) { 1718 btree_path_overflow(trans); 1719 return 0; 1720 } 1721 1722 btree_paths_realloc(trans); 1723 } 1724 1725 /* 1726 * Do this before marking the new path as allocated, since it won't be 1727 * initialized yet: 1728 */ 1729 if (unlikely(idx > trans->nr_paths_max)) 1730 bch2_trans_update_max_paths(trans); 1731 1732 __set_bit(idx, trans->paths_allocated); 1733 1734 struct btree_path *path = &trans->paths[idx]; 1735 path->ref = 0; 1736 path->intent_ref = 0; 1737 path->nodes_locked = 0; 1738 1739 btree_path_list_add(trans, pos, idx); 1740 trans->paths_sorted = false; 1741 return idx; 1742 } 1743 1744 btree_path_idx_t bch2_path_get(struct btree_trans *trans, 1745 enum btree_id btree_id, struct bpos pos, 1746 unsigned locks_want, unsigned level, 1747 unsigned flags, unsigned long ip) 1748 { 1749 struct btree_path *path; 1750 bool cached = flags & BTREE_ITER_cached; 1751 bool intent = flags & BTREE_ITER_intent; 1752 struct trans_for_each_path_inorder_iter iter; 1753 btree_path_idx_t path_pos = 0, path_idx; 1754 1755 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1756 bch2_trans_verify_locks(trans); 1757 1758 btree_trans_sort_paths(trans); 1759 1760 trans_for_each_path_inorder(trans, path, iter) { 1761 if (__btree_path_cmp(path, 1762 btree_id, 1763 cached, 1764 pos, 1765 level) > 0) 1766 break; 1767 1768 path_pos = iter.path_idx; 1769 } 1770 1771 if (path_pos && 1772 trans->paths[path_pos].cached == cached && 1773 trans->paths[path_pos].btree_id == btree_id && 1774 trans->paths[path_pos].level == level) { 1775 trace_btree_path_get(trans, trans->paths + path_pos, &pos); 1776 1777 __btree_path_get(trans, trans->paths + path_pos, intent); 1778 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip); 1779 path = trans->paths + path_idx; 1780 } else { 1781 path_idx = btree_path_alloc(trans, path_pos); 1782 path = trans->paths + path_idx; 1783 1784 __btree_path_get(trans, path, intent); 1785 path->pos = pos; 1786 path->btree_id = btree_id; 1787 path->cached = cached; 1788 path->uptodate = BTREE_ITER_NEED_TRAVERSE; 1789 path->should_be_locked = false; 1790 path->level = level; 1791 path->locks_want = locks_want; 1792 path->nodes_locked = 0; 1793 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++) 1794 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 1795 #ifdef TRACK_PATH_ALLOCATED 1796 path->ip_allocated = ip; 1797 #endif 1798 trans->paths_sorted = false; 1799 1800 trace_btree_path_alloc(trans, path); 1801 } 1802 1803 if (!(flags & BTREE_ITER_nopreserve)) 1804 path->preserve = true; 1805 1806 if (path->intent_ref) 1807 locks_want = max(locks_want, level + 1); 1808 1809 /* 1810 * If the path has locks_want greater than requested, we don't downgrade 1811 * it here - on transaction restart because btree node split needs to 1812 * upgrade locks, we might be putting/getting the iterator again. 1813 * Downgrading iterators only happens via bch2_trans_downgrade(), after 1814 * a successful transaction commit. 1815 */ 1816 1817 locks_want = min(locks_want, BTREE_MAX_DEPTH); 1818 if (locks_want > path->locks_want) 1819 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL); 1820 1821 return path_idx; 1822 } 1823 1824 btree_path_idx_t bch2_path_get_unlocked_mut(struct btree_trans *trans, 1825 enum btree_id btree_id, 1826 unsigned level, 1827 struct bpos pos) 1828 { 1829 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level, 1830 BTREE_ITER_nopreserve| 1831 BTREE_ITER_intent, _RET_IP_); 1832 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_); 1833 1834 struct btree_path *path = trans->paths + path_idx; 1835 bch2_btree_path_downgrade(trans, path); 1836 __bch2_btree_path_unlock(trans, path); 1837 return path_idx; 1838 } 1839 1840 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u) 1841 { 1842 1843 struct btree_path_level *l = path_l(path); 1844 struct bkey_packed *_k; 1845 struct bkey_s_c k; 1846 1847 if (unlikely(!l->b)) 1848 return bkey_s_c_null; 1849 1850 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE); 1851 EBUG_ON(!btree_node_locked(path, path->level)); 1852 1853 if (!path->cached) { 1854 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b); 1855 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null; 1856 1857 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos)); 1858 1859 if (!k.k || !bpos_eq(path->pos, k.k->p)) 1860 goto hole; 1861 } else { 1862 struct bkey_cached *ck = (void *) path->l[0].b; 1863 if (!ck) 1864 return bkey_s_c_null; 1865 1866 EBUG_ON(path->btree_id != ck->key.btree_id || 1867 !bkey_eq(path->pos, ck->key.pos)); 1868 1869 *u = ck->k->k; 1870 k = (struct bkey_s_c) { u, &ck->k->v }; 1871 } 1872 1873 return k; 1874 hole: 1875 bkey_init(u); 1876 u->p = path->pos; 1877 return (struct bkey_s_c) { u, NULL }; 1878 } 1879 1880 void bch2_set_btree_iter_dontneed(struct btree_iter *iter) 1881 { 1882 struct btree_trans *trans = iter->trans; 1883 1884 if (!iter->path || trans->restarted) 1885 return; 1886 1887 struct btree_path *path = btree_iter_path(trans, iter); 1888 path->preserve = false; 1889 if (path->ref == 1) 1890 path->should_be_locked = false; 1891 } 1892 /* Btree iterators: */ 1893 1894 int __must_check 1895 __bch2_btree_iter_traverse(struct btree_iter *iter) 1896 { 1897 return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags); 1898 } 1899 1900 int __must_check 1901 bch2_btree_iter_traverse(struct btree_iter *iter) 1902 { 1903 struct btree_trans *trans = iter->trans; 1904 int ret; 1905 1906 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1907 1908 iter->path = bch2_btree_path_set_pos(trans, iter->path, 1909 btree_iter_search_key(iter), 1910 iter->flags & BTREE_ITER_intent, 1911 btree_iter_ip_allocated(iter)); 1912 1913 ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags); 1914 if (ret) 1915 return ret; 1916 1917 struct btree_path *path = btree_iter_path(trans, iter); 1918 if (btree_path_node(path, path->level)) 1919 btree_path_set_should_be_locked(trans, path); 1920 return 0; 1921 } 1922 1923 /* Iterate across nodes (leaf and interior nodes) */ 1924 1925 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter) 1926 { 1927 struct btree_trans *trans = iter->trans; 1928 struct btree *b = NULL; 1929 int ret; 1930 1931 EBUG_ON(trans->paths[iter->path].cached); 1932 bch2_btree_iter_verify(iter); 1933 1934 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 1935 if (ret) 1936 goto err; 1937 1938 struct btree_path *path = btree_iter_path(trans, iter); 1939 b = btree_path_node(path, path->level); 1940 if (!b) 1941 goto out; 1942 1943 BUG_ON(bpos_lt(b->key.k.p, iter->pos)); 1944 1945 bkey_init(&iter->k); 1946 iter->k.p = iter->pos = b->key.k.p; 1947 1948 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, 1949 iter->flags & BTREE_ITER_intent, 1950 btree_iter_ip_allocated(iter)); 1951 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 1952 out: 1953 bch2_btree_iter_verify_entry_exit(iter); 1954 bch2_btree_iter_verify(iter); 1955 1956 return b; 1957 err: 1958 b = ERR_PTR(ret); 1959 goto out; 1960 } 1961 1962 /* Only kept for -tools */ 1963 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter) 1964 { 1965 struct btree *b; 1966 1967 while (b = bch2_btree_iter_peek_node(iter), 1968 bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart)) 1969 bch2_trans_begin(iter->trans); 1970 1971 return b; 1972 } 1973 1974 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter) 1975 { 1976 struct btree_trans *trans = iter->trans; 1977 struct btree *b = NULL; 1978 int ret; 1979 1980 EBUG_ON(trans->paths[iter->path].cached); 1981 bch2_trans_verify_not_unlocked_or_in_restart(trans); 1982 bch2_btree_iter_verify(iter); 1983 1984 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 1985 if (ret) 1986 goto err; 1987 1988 1989 struct btree_path *path = btree_iter_path(trans, iter); 1990 1991 /* already at end? */ 1992 if (!btree_path_node(path, path->level)) 1993 return NULL; 1994 1995 /* got to end? */ 1996 if (!btree_path_node(path, path->level + 1)) { 1997 btree_path_set_level_up(trans, path); 1998 return NULL; 1999 } 2000 2001 if (!bch2_btree_node_relock(trans, path, path->level + 1)) { 2002 __bch2_btree_path_unlock(trans, path); 2003 path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); 2004 path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); 2005 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); 2006 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path); 2007 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock); 2008 goto err; 2009 } 2010 2011 b = btree_path_node(path, path->level + 1); 2012 2013 if (bpos_eq(iter->pos, b->key.k.p)) { 2014 __btree_path_set_level_up(trans, path, path->level++); 2015 } else { 2016 if (btree_lock_want(path, path->level + 1) == BTREE_NODE_UNLOCKED) 2017 btree_node_unlock(trans, path, path->level + 1); 2018 2019 /* 2020 * Haven't gotten to the end of the parent node: go back down to 2021 * the next child node 2022 */ 2023 iter->path = bch2_btree_path_set_pos(trans, iter->path, 2024 bpos_successor(iter->pos), 2025 iter->flags & BTREE_ITER_intent, 2026 btree_iter_ip_allocated(iter)); 2027 2028 path = btree_iter_path(trans, iter); 2029 btree_path_set_level_down(trans, path, iter->min_depth); 2030 2031 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2032 if (ret) 2033 goto err; 2034 2035 path = btree_iter_path(trans, iter); 2036 b = path->l[path->level].b; 2037 } 2038 2039 bkey_init(&iter->k); 2040 iter->k.p = iter->pos = b->key.k.p; 2041 2042 iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, 2043 iter->flags & BTREE_ITER_intent, 2044 btree_iter_ip_allocated(iter)); 2045 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2046 EBUG_ON(btree_iter_path(trans, iter)->uptodate); 2047 out: 2048 bch2_btree_iter_verify_entry_exit(iter); 2049 bch2_btree_iter_verify(iter); 2050 2051 return b; 2052 err: 2053 b = ERR_PTR(ret); 2054 goto out; 2055 } 2056 2057 /* Iterate across keys (in leaf nodes only) */ 2058 2059 inline bool bch2_btree_iter_advance(struct btree_iter *iter) 2060 { 2061 struct bpos pos = iter->k.p; 2062 bool ret = !(iter->flags & BTREE_ITER_all_snapshots 2063 ? bpos_eq(pos, SPOS_MAX) 2064 : bkey_eq(pos, SPOS_MAX)); 2065 2066 if (ret && !(iter->flags & BTREE_ITER_is_extents)) 2067 pos = bkey_successor(iter, pos); 2068 bch2_btree_iter_set_pos(iter, pos); 2069 return ret; 2070 } 2071 2072 inline bool bch2_btree_iter_rewind(struct btree_iter *iter) 2073 { 2074 struct bpos pos = bkey_start_pos(&iter->k); 2075 bool ret = !(iter->flags & BTREE_ITER_all_snapshots 2076 ? bpos_eq(pos, POS_MIN) 2077 : bkey_eq(pos, POS_MIN)); 2078 2079 if (ret && !(iter->flags & BTREE_ITER_is_extents)) 2080 pos = bkey_predecessor(iter, pos); 2081 bch2_btree_iter_set_pos(iter, pos); 2082 return ret; 2083 } 2084 2085 static noinline 2086 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter, 2087 struct bkey_s_c *k) 2088 { 2089 struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key; 2090 2091 trans_for_each_update(trans, i) 2092 if (!i->key_cache_already_flushed && 2093 i->btree_id == iter->btree_id && 2094 bpos_le(i->k->k.p, iter->pos) && 2095 bpos_ge(i->k->k.p, k->k ? k->k->p : end)) { 2096 iter->k = i->k->k; 2097 *k = bkey_i_to_s_c(i->k); 2098 } 2099 } 2100 2101 static noinline 2102 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter, 2103 struct bkey_s_c *k) 2104 { 2105 struct btree_path *path = btree_iter_path(trans, iter); 2106 struct bpos end = path_l(path)->b->key.k.p; 2107 2108 trans_for_each_update(trans, i) 2109 if (!i->key_cache_already_flushed && 2110 i->btree_id == iter->btree_id && 2111 bpos_ge(i->k->k.p, path->pos) && 2112 bpos_le(i->k->k.p, k->k ? k->k->p : end)) { 2113 iter->k = i->k->k; 2114 *k = bkey_i_to_s_c(i->k); 2115 } 2116 } 2117 2118 static noinline 2119 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter, 2120 struct bkey_s_c *k) 2121 { 2122 trans_for_each_update(trans, i) 2123 if (!i->key_cache_already_flushed && 2124 i->btree_id == iter->btree_id && 2125 bpos_eq(i->k->k.p, iter->pos)) { 2126 iter->k = i->k->k; 2127 *k = bkey_i_to_s_c(i->k); 2128 } 2129 } 2130 2131 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans, 2132 struct btree_iter *iter, 2133 struct bpos end_pos) 2134 { 2135 struct btree_path *path = btree_iter_path(trans, iter); 2136 2137 return bch2_journal_keys_peek_max(trans->c, iter->btree_id, 2138 path->level, 2139 path->pos, 2140 end_pos, 2141 &iter->journal_idx); 2142 } 2143 2144 static noinline 2145 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans, 2146 struct btree_iter *iter) 2147 { 2148 struct btree_path *path = btree_iter_path(trans, iter); 2149 struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos); 2150 2151 if (k) { 2152 iter->k = k->k; 2153 return bkey_i_to_s_c(k); 2154 } else { 2155 return bkey_s_c_null; 2156 } 2157 } 2158 2159 static noinline 2160 void btree_trans_peek_journal(struct btree_trans *trans, 2161 struct btree_iter *iter, 2162 struct bkey_s_c *k) 2163 { 2164 struct btree_path *path = btree_iter_path(trans, iter); 2165 struct bkey_i *next_journal = 2166 bch2_btree_journal_peek(trans, iter, 2167 k->k ? k->k->p : path_l(path)->b->key.k.p); 2168 if (next_journal) { 2169 iter->k = next_journal->k; 2170 *k = bkey_i_to_s_c(next_journal); 2171 } 2172 } 2173 2174 static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans, 2175 struct btree_iter *iter, 2176 struct bpos end_pos) 2177 { 2178 struct btree_path *path = btree_iter_path(trans, iter); 2179 2180 return bch2_journal_keys_peek_prev_min(trans->c, iter->btree_id, 2181 path->level, 2182 path->pos, 2183 end_pos, 2184 &iter->journal_idx); 2185 } 2186 2187 static noinline 2188 void btree_trans_peek_prev_journal(struct btree_trans *trans, 2189 struct btree_iter *iter, 2190 struct bkey_s_c *k) 2191 { 2192 struct btree_path *path = btree_iter_path(trans, iter); 2193 struct bkey_i *next_journal = 2194 bch2_btree_journal_peek_prev(trans, iter, 2195 k->k ? k->k->p : path_l(path)->b->key.k.p); 2196 2197 if (next_journal) { 2198 iter->k = next_journal->k; 2199 *k = bkey_i_to_s_c(next_journal); 2200 } 2201 } 2202 2203 /* 2204 * Checks btree key cache for key at iter->pos and returns it if present, or 2205 * bkey_s_c_null: 2206 */ 2207 static noinline 2208 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos) 2209 { 2210 struct btree_trans *trans = iter->trans; 2211 struct bch_fs *c = trans->c; 2212 struct bkey u; 2213 struct bkey_s_c k; 2214 int ret; 2215 2216 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2217 2218 if ((iter->flags & BTREE_ITER_key_cache_fill) && 2219 bpos_eq(iter->pos, pos)) 2220 return bkey_s_c_null; 2221 2222 if (!bch2_btree_key_cache_find(c, iter->btree_id, pos)) 2223 return bkey_s_c_null; 2224 2225 if (!iter->key_cache_path) 2226 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos, 2227 iter->flags & BTREE_ITER_intent, 0, 2228 iter->flags|BTREE_ITER_cached| 2229 BTREE_ITER_cached_nofill, 2230 _THIS_IP_); 2231 2232 iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos, 2233 iter->flags & BTREE_ITER_intent, 2234 btree_iter_ip_allocated(iter)); 2235 2236 ret = bch2_btree_path_traverse(trans, iter->key_cache_path, 2237 iter->flags|BTREE_ITER_cached) ?: 2238 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_); 2239 if (unlikely(ret)) 2240 return bkey_s_c_err(ret); 2241 2242 btree_path_set_should_be_locked(trans, trans->paths + iter->key_cache_path); 2243 2244 k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u); 2245 if (!k.k) 2246 return k; 2247 2248 if ((iter->flags & BTREE_ITER_all_snapshots) && 2249 !bpos_eq(pos, k.k->p)) 2250 return bkey_s_c_null; 2251 2252 iter->k = u; 2253 k.k = &iter->k; 2254 return k; 2255 } 2256 2257 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key) 2258 { 2259 struct btree_trans *trans = iter->trans; 2260 struct bkey_s_c k, k2; 2261 int ret; 2262 2263 EBUG_ON(btree_iter_path(trans, iter)->cached); 2264 bch2_btree_iter_verify(iter); 2265 2266 while (1) { 2267 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2268 iter->flags & BTREE_ITER_intent, 2269 btree_iter_ip_allocated(iter)); 2270 2271 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2272 if (unlikely(ret)) { 2273 /* ensure that iter->k is consistent with iter->pos: */ 2274 bch2_btree_iter_set_pos(iter, iter->pos); 2275 k = bkey_s_c_err(ret); 2276 break; 2277 } 2278 2279 struct btree_path *path = btree_iter_path(trans, iter); 2280 struct btree_path_level *l = path_l(path); 2281 2282 if (unlikely(!l->b)) { 2283 /* No btree nodes at requested level: */ 2284 bch2_btree_iter_set_pos(iter, SPOS_MAX); 2285 k = bkey_s_c_null; 2286 break; 2287 } 2288 2289 btree_path_set_should_be_locked(trans, path); 2290 2291 k = btree_path_level_peek_all(trans->c, l, &iter->k); 2292 2293 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2294 k.k && 2295 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) { 2296 k = k2; 2297 if (bkey_err(k)) { 2298 bch2_btree_iter_set_pos(iter, iter->pos); 2299 break; 2300 } 2301 } 2302 2303 if (unlikely(iter->flags & BTREE_ITER_with_journal)) 2304 btree_trans_peek_journal(trans, iter, &k); 2305 2306 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2307 trans->nr_updates)) 2308 bch2_btree_trans_peek_updates(trans, iter, &k); 2309 2310 if (k.k && bkey_deleted(k.k)) { 2311 /* 2312 * If we've got a whiteout, and it's after the search 2313 * key, advance the search key to the whiteout instead 2314 * of just after the whiteout - it might be a btree 2315 * whiteout, with a real key at the same position, since 2316 * in the btree deleted keys sort before non deleted. 2317 */ 2318 search_key = !bpos_eq(search_key, k.k->p) 2319 ? k.k->p 2320 : bpos_successor(k.k->p); 2321 continue; 2322 } 2323 2324 if (likely(k.k)) { 2325 break; 2326 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) { 2327 /* Advance to next leaf node: */ 2328 search_key = bpos_successor(l->b->key.k.p); 2329 } else { 2330 /* End of btree: */ 2331 bch2_btree_iter_set_pos(iter, SPOS_MAX); 2332 k = bkey_s_c_null; 2333 break; 2334 } 2335 } 2336 2337 bch2_btree_iter_verify(iter); 2338 return k; 2339 } 2340 2341 /** 2342 * bch2_btree_iter_peek_max() - returns first key greater than or equal to 2343 * iterator's current position 2344 * @iter: iterator to peek from 2345 * @end: search limit: returns keys less than or equal to @end 2346 * 2347 * Returns: key if found, or an error extractable with bkey_err(). 2348 */ 2349 struct bkey_s_c bch2_btree_iter_peek_max(struct btree_iter *iter, struct bpos end) 2350 { 2351 struct btree_trans *trans = iter->trans; 2352 struct bpos search_key = btree_iter_search_key(iter); 2353 struct bkey_s_c k; 2354 struct bpos iter_pos = iter->pos; 2355 int ret; 2356 2357 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2358 bch2_btree_iter_verify_entry_exit(iter); 2359 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bkey_eq(end, POS_MAX)); 2360 2361 if (iter->update_path) { 2362 bch2_path_put_nokeep(trans, iter->update_path, 2363 iter->flags & BTREE_ITER_intent); 2364 iter->update_path = 0; 2365 } 2366 2367 while (1) { 2368 k = __bch2_btree_iter_peek(iter, search_key); 2369 if (unlikely(!k.k)) 2370 goto end; 2371 if (unlikely(bkey_err(k))) 2372 goto out_no_locked; 2373 2374 if (iter->flags & BTREE_ITER_filter_snapshots) { 2375 /* 2376 * We need to check against @end before FILTER_SNAPSHOTS because 2377 * if we get to a different inode that requested we might be 2378 * seeing keys for a different snapshot tree that will all be 2379 * filtered out. 2380 * 2381 * But we can't do the full check here, because bkey_start_pos() 2382 * isn't monotonically increasing before FILTER_SNAPSHOTS, and 2383 * that's what we check against in extents mode: 2384 */ 2385 if (unlikely(!(iter->flags & BTREE_ITER_is_extents) 2386 ? bkey_gt(k.k->p, end) 2387 : k.k->p.inode > end.inode)) 2388 goto end; 2389 2390 if (iter->update_path && 2391 !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) { 2392 bch2_path_put_nokeep(trans, iter->update_path, 2393 iter->flags & BTREE_ITER_intent); 2394 iter->update_path = 0; 2395 } 2396 2397 if ((iter->flags & BTREE_ITER_intent) && 2398 !(iter->flags & BTREE_ITER_is_extents) && 2399 !iter->update_path) { 2400 struct bpos pos = k.k->p; 2401 2402 if (pos.snapshot < iter->snapshot) { 2403 search_key = bpos_successor(k.k->p); 2404 continue; 2405 } 2406 2407 pos.snapshot = iter->snapshot; 2408 2409 /* 2410 * advance, same as on exit for iter->path, but only up 2411 * to snapshot 2412 */ 2413 __btree_path_get(trans, trans->paths + iter->path, iter->flags & BTREE_ITER_intent); 2414 iter->update_path = iter->path; 2415 2416 iter->update_path = bch2_btree_path_set_pos(trans, 2417 iter->update_path, pos, 2418 iter->flags & BTREE_ITER_intent, 2419 _THIS_IP_); 2420 ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags); 2421 if (unlikely(ret)) { 2422 k = bkey_s_c_err(ret); 2423 goto out_no_locked; 2424 } 2425 } 2426 2427 /* 2428 * We can never have a key in a leaf node at POS_MAX, so 2429 * we don't have to check these successor() calls: 2430 */ 2431 if (!bch2_snapshot_is_ancestor(trans->c, 2432 iter->snapshot, 2433 k.k->p.snapshot)) { 2434 search_key = bpos_successor(k.k->p); 2435 continue; 2436 } 2437 2438 if (bkey_whiteout(k.k) && 2439 !(iter->flags & BTREE_ITER_key_cache_fill)) { 2440 search_key = bkey_successor(iter, k.k->p); 2441 continue; 2442 } 2443 } 2444 2445 /* 2446 * iter->pos should be mononotically increasing, and always be 2447 * equal to the key we just returned - except extents can 2448 * straddle iter->pos: 2449 */ 2450 if (!(iter->flags & BTREE_ITER_is_extents)) 2451 iter_pos = k.k->p; 2452 else 2453 iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k)); 2454 2455 if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_gt(iter_pos, end) : 2456 iter->flags & BTREE_ITER_is_extents ? bkey_ge(iter_pos, end) : 2457 bkey_gt(iter_pos, end))) 2458 goto end; 2459 2460 break; 2461 } 2462 2463 iter->pos = iter_pos; 2464 2465 iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p, 2466 iter->flags & BTREE_ITER_intent, 2467 btree_iter_ip_allocated(iter)); 2468 2469 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2470 out_no_locked: 2471 if (iter->update_path) { 2472 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_); 2473 if (unlikely(ret)) 2474 k = bkey_s_c_err(ret); 2475 else 2476 btree_path_set_should_be_locked(trans, trans->paths + iter->update_path); 2477 } 2478 2479 if (!(iter->flags & BTREE_ITER_all_snapshots)) 2480 iter->pos.snapshot = iter->snapshot; 2481 2482 ret = bch2_btree_iter_verify_ret(iter, k); 2483 if (unlikely(ret)) { 2484 bch2_btree_iter_set_pos(iter, iter->pos); 2485 k = bkey_s_c_err(ret); 2486 } 2487 2488 bch2_btree_iter_verify_entry_exit(iter); 2489 2490 return k; 2491 end: 2492 bch2_btree_iter_set_pos(iter, end); 2493 k = bkey_s_c_null; 2494 goto out_no_locked; 2495 } 2496 2497 /** 2498 * bch2_btree_iter_next() - returns first key greater than iterator's current 2499 * position 2500 * @iter: iterator to peek from 2501 * 2502 * Returns: key if found, or an error extractable with bkey_err(). 2503 */ 2504 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter) 2505 { 2506 if (!bch2_btree_iter_advance(iter)) 2507 return bkey_s_c_null; 2508 2509 return bch2_btree_iter_peek(iter); 2510 } 2511 2512 static struct bkey_s_c __bch2_btree_iter_peek_prev(struct btree_iter *iter, struct bpos search_key) 2513 { 2514 struct btree_trans *trans = iter->trans; 2515 struct bkey_s_c k, k2; 2516 2517 bch2_btree_iter_verify(iter); 2518 2519 while (1) { 2520 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2521 iter->flags & BTREE_ITER_intent, 2522 btree_iter_ip_allocated(iter)); 2523 2524 int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2525 if (unlikely(ret)) { 2526 /* ensure that iter->k is consistent with iter->pos: */ 2527 bch2_btree_iter_set_pos(iter, iter->pos); 2528 k = bkey_s_c_err(ret); 2529 break; 2530 } 2531 2532 struct btree_path *path = btree_iter_path(trans, iter); 2533 struct btree_path_level *l = path_l(path); 2534 2535 if (unlikely(!l->b)) { 2536 /* No btree nodes at requested level: */ 2537 bch2_btree_iter_set_pos(iter, SPOS_MAX); 2538 k = bkey_s_c_null; 2539 break; 2540 } 2541 2542 btree_path_set_should_be_locked(trans, path); 2543 2544 k = btree_path_level_peek_all(trans->c, l, &iter->k); 2545 if (!k.k || bpos_gt(k.k->p, search_key)) { 2546 k = btree_path_level_prev(trans, path, l, &iter->k); 2547 2548 BUG_ON(k.k && bpos_gt(k.k->p, search_key)); 2549 } 2550 2551 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2552 k.k && 2553 (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) { 2554 k = k2; 2555 if (bkey_err(k2)) { 2556 bch2_btree_iter_set_pos(iter, iter->pos); 2557 break; 2558 } 2559 } 2560 2561 if (unlikely(iter->flags & BTREE_ITER_with_journal)) 2562 btree_trans_peek_prev_journal(trans, iter, &k); 2563 2564 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2565 trans->nr_updates)) 2566 bch2_btree_trans_peek_prev_updates(trans, iter, &k); 2567 2568 if (likely(k.k && !bkey_deleted(k.k))) { 2569 break; 2570 } else if (k.k) { 2571 search_key = bpos_predecessor(k.k->p); 2572 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) { 2573 /* Advance to previous leaf node: */ 2574 search_key = bpos_predecessor(path->l[0].b->data->min_key); 2575 } else { 2576 /* Start of btree: */ 2577 bch2_btree_iter_set_pos(iter, POS_MIN); 2578 k = bkey_s_c_null; 2579 break; 2580 } 2581 } 2582 2583 bch2_btree_iter_verify(iter); 2584 return k; 2585 } 2586 2587 /** 2588 * bch2_btree_iter_peek_prev_min() - returns first key less than or equal to 2589 * iterator's current position 2590 * @iter: iterator to peek from 2591 * @end: search limit: returns keys greater than or equal to @end 2592 * 2593 * Returns: key if found, or an error extractable with bkey_err(). 2594 */ 2595 struct bkey_s_c bch2_btree_iter_peek_prev_min(struct btree_iter *iter, struct bpos end) 2596 { 2597 if ((iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots)) && 2598 !bkey_eq(iter->pos, POS_MAX)) { 2599 /* 2600 * bkey_start_pos(), for extents, is not monotonically 2601 * increasing until after filtering for snapshots: 2602 * 2603 * Thus, for extents we need to search forward until we find a 2604 * real visible extents - easiest to just use peek_slot() (which 2605 * internally uses peek() for extents) 2606 */ 2607 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter); 2608 if (bkey_err(k)) 2609 return k; 2610 2611 if (!bkey_deleted(k.k) && 2612 (!(iter->flags & BTREE_ITER_is_extents) || 2613 bkey_lt(bkey_start_pos(k.k), iter->pos))) 2614 return k; 2615 } 2616 2617 struct btree_trans *trans = iter->trans; 2618 struct bpos search_key = iter->pos; 2619 struct bkey_s_c k; 2620 btree_path_idx_t saved_path = 0; 2621 2622 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2623 bch2_btree_iter_verify_entry_exit(iter); 2624 EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bpos_eq(end, POS_MIN)); 2625 2626 while (1) { 2627 k = __bch2_btree_iter_peek_prev(iter, search_key); 2628 if (unlikely(!k.k)) 2629 goto end; 2630 if (unlikely(bkey_err(k))) 2631 goto out_no_locked; 2632 2633 if (iter->flags & BTREE_ITER_filter_snapshots) { 2634 struct btree_path *s = saved_path ? trans->paths + saved_path : NULL; 2635 if (s && bpos_lt(k.k->p, SPOS(s->pos.inode, s->pos.offset, iter->snapshot))) { 2636 /* 2637 * If we have a saved candidate, and we're past 2638 * the last possible snapshot overwrite, return 2639 * it: 2640 */ 2641 bch2_path_put_nokeep(trans, iter->path, 2642 iter->flags & BTREE_ITER_intent); 2643 iter->path = saved_path; 2644 saved_path = 0; 2645 k = bch2_btree_path_peek_slot(btree_iter_path(trans, iter), &iter->k); 2646 break; 2647 } 2648 2649 /* 2650 * We need to check against @end before FILTER_SNAPSHOTS because 2651 * if we get to a different inode that requested we might be 2652 * seeing keys for a different snapshot tree that will all be 2653 * filtered out. 2654 */ 2655 if (unlikely(bkey_lt(k.k->p, end))) 2656 goto end; 2657 2658 if (!bch2_snapshot_is_ancestor(trans->c, iter->snapshot, k.k->p.snapshot)) { 2659 search_key = bpos_predecessor(k.k->p); 2660 continue; 2661 } 2662 2663 if (k.k->p.snapshot != iter->snapshot) { 2664 /* 2665 * Have a key visible in iter->snapshot, but 2666 * might have overwrites: - save it and keep 2667 * searching. Unless it's a whiteout - then drop 2668 * our previous saved candidate: 2669 */ 2670 if (saved_path) { 2671 bch2_path_put_nokeep(trans, saved_path, 2672 iter->flags & BTREE_ITER_intent); 2673 saved_path = 0; 2674 } 2675 2676 if (!bkey_whiteout(k.k)) { 2677 saved_path = btree_path_clone(trans, iter->path, 2678 iter->flags & BTREE_ITER_intent, 2679 _THIS_IP_); 2680 trace_btree_path_save_pos(trans, 2681 trans->paths + iter->path, 2682 trans->paths + saved_path); 2683 } 2684 2685 search_key = bpos_predecessor(k.k->p); 2686 continue; 2687 } 2688 2689 if (bkey_whiteout(k.k)) { 2690 search_key = bkey_predecessor(iter, k.k->p); 2691 search_key.snapshot = U32_MAX; 2692 continue; 2693 } 2694 } 2695 2696 EBUG_ON(iter->flags & BTREE_ITER_all_snapshots ? bpos_gt(k.k->p, iter->pos) : 2697 iter->flags & BTREE_ITER_is_extents ? bkey_ge(bkey_start_pos(k.k), iter->pos) : 2698 bkey_gt(k.k->p, iter->pos)); 2699 2700 if (unlikely(iter->flags & BTREE_ITER_all_snapshots ? bpos_lt(k.k->p, end) : 2701 iter->flags & BTREE_ITER_is_extents ? bkey_le(k.k->p, end) : 2702 bkey_lt(k.k->p, end))) 2703 goto end; 2704 2705 break; 2706 } 2707 2708 /* Extents can straddle iter->pos: */ 2709 iter->pos = bpos_min(iter->pos, k.k->p);; 2710 2711 if (iter->flags & BTREE_ITER_filter_snapshots) 2712 iter->pos.snapshot = iter->snapshot; 2713 out_no_locked: 2714 if (saved_path) 2715 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_intent); 2716 2717 bch2_btree_iter_verify_entry_exit(iter); 2718 bch2_btree_iter_verify(iter); 2719 return k; 2720 end: 2721 bch2_btree_iter_set_pos(iter, end); 2722 k = bkey_s_c_null; 2723 goto out_no_locked; 2724 } 2725 2726 /** 2727 * bch2_btree_iter_prev() - returns first key less than iterator's current 2728 * position 2729 * @iter: iterator to peek from 2730 * 2731 * Returns: key if found, or an error extractable with bkey_err(). 2732 */ 2733 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter) 2734 { 2735 if (!bch2_btree_iter_rewind(iter)) 2736 return bkey_s_c_null; 2737 2738 return bch2_btree_iter_peek_prev(iter); 2739 } 2740 2741 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) 2742 { 2743 struct btree_trans *trans = iter->trans; 2744 struct bpos search_key; 2745 struct bkey_s_c k; 2746 int ret; 2747 2748 bch2_trans_verify_not_unlocked_or_in_restart(trans); 2749 bch2_btree_iter_verify(iter); 2750 bch2_btree_iter_verify_entry_exit(iter); 2751 EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_with_key_cache)); 2752 2753 /* extents can't span inode numbers: */ 2754 if ((iter->flags & BTREE_ITER_is_extents) && 2755 unlikely(iter->pos.offset == KEY_OFFSET_MAX)) { 2756 if (iter->pos.inode == KEY_INODE_MAX) 2757 return bkey_s_c_null; 2758 2759 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos)); 2760 } 2761 2762 search_key = btree_iter_search_key(iter); 2763 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, 2764 iter->flags & BTREE_ITER_intent, 2765 btree_iter_ip_allocated(iter)); 2766 2767 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); 2768 if (unlikely(ret)) { 2769 k = bkey_s_c_err(ret); 2770 goto out_no_locked; 2771 } 2772 2773 struct btree_path *path = btree_iter_path(trans, iter); 2774 if (unlikely(!btree_path_node(path, path->level))) 2775 return bkey_s_c_null; 2776 2777 if ((iter->flags & BTREE_ITER_cached) || 2778 !(iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots))) { 2779 k = bkey_s_c_null; 2780 2781 if (unlikely((iter->flags & BTREE_ITER_with_updates) && 2782 trans->nr_updates)) { 2783 bch2_btree_trans_peek_slot_updates(trans, iter, &k); 2784 if (k.k) 2785 goto out; 2786 } 2787 2788 if (unlikely(iter->flags & BTREE_ITER_with_journal) && 2789 (k = btree_trans_peek_slot_journal(trans, iter)).k) 2790 goto out; 2791 2792 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) && 2793 (k = btree_trans_peek_key_cache(iter, iter->pos)).k) { 2794 if (!bkey_err(k)) 2795 iter->k = *k.k; 2796 /* We're not returning a key from iter->path: */ 2797 goto out_no_locked; 2798 } 2799 2800 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k); 2801 if (unlikely(!k.k)) 2802 goto out_no_locked; 2803 2804 if (unlikely(k.k->type == KEY_TYPE_whiteout && 2805 (iter->flags & BTREE_ITER_filter_snapshots) && 2806 !(iter->flags & BTREE_ITER_key_cache_fill))) 2807 iter->k.type = KEY_TYPE_deleted; 2808 } else { 2809 struct bpos next; 2810 struct bpos end = iter->pos; 2811 2812 if (iter->flags & BTREE_ITER_is_extents) 2813 end.offset = U64_MAX; 2814 2815 EBUG_ON(btree_iter_path(trans, iter)->level); 2816 2817 if (iter->flags & BTREE_ITER_intent) { 2818 struct btree_iter iter2; 2819 2820 bch2_trans_copy_iter(&iter2, iter); 2821 k = bch2_btree_iter_peek_max(&iter2, end); 2822 2823 if (k.k && !bkey_err(k)) { 2824 swap(iter->key_cache_path, iter2.key_cache_path); 2825 iter->k = iter2.k; 2826 k.k = &iter->k; 2827 } 2828 bch2_trans_iter_exit(trans, &iter2); 2829 } else { 2830 struct bpos pos = iter->pos; 2831 2832 k = bch2_btree_iter_peek_max(iter, end); 2833 if (unlikely(bkey_err(k))) 2834 bch2_btree_iter_set_pos(iter, pos); 2835 else 2836 iter->pos = pos; 2837 } 2838 2839 if (unlikely(bkey_err(k))) 2840 goto out_no_locked; 2841 2842 next = k.k ? bkey_start_pos(k.k) : POS_MAX; 2843 2844 if (bkey_lt(iter->pos, next)) { 2845 bkey_init(&iter->k); 2846 iter->k.p = iter->pos; 2847 2848 if (iter->flags & BTREE_ITER_is_extents) { 2849 bch2_key_resize(&iter->k, 2850 min_t(u64, KEY_SIZE_MAX, 2851 (next.inode == iter->pos.inode 2852 ? next.offset 2853 : KEY_OFFSET_MAX) - 2854 iter->pos.offset)); 2855 EBUG_ON(!iter->k.size); 2856 } 2857 2858 k = (struct bkey_s_c) { &iter->k, NULL }; 2859 } 2860 } 2861 out: 2862 btree_path_set_should_be_locked(trans, btree_iter_path(trans, iter)); 2863 out_no_locked: 2864 bch2_btree_iter_verify_entry_exit(iter); 2865 bch2_btree_iter_verify(iter); 2866 ret = bch2_btree_iter_verify_ret(iter, k); 2867 if (unlikely(ret)) 2868 return bkey_s_c_err(ret); 2869 2870 return k; 2871 } 2872 2873 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter) 2874 { 2875 if (!bch2_btree_iter_advance(iter)) 2876 return bkey_s_c_null; 2877 2878 return bch2_btree_iter_peek_slot(iter); 2879 } 2880 2881 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter) 2882 { 2883 if (!bch2_btree_iter_rewind(iter)) 2884 return bkey_s_c_null; 2885 2886 return bch2_btree_iter_peek_slot(iter); 2887 } 2888 2889 /* Obsolete, but still used by rust wrapper in -tools */ 2890 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter) 2891 { 2892 struct bkey_s_c k; 2893 2894 while (btree_trans_too_many_iters(iter->trans) || 2895 (k = bch2_btree_iter_peek_type(iter, iter->flags), 2896 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart))) 2897 bch2_trans_begin(iter->trans); 2898 2899 return k; 2900 } 2901 2902 /* new transactional stuff: */ 2903 2904 #ifdef CONFIG_BCACHEFS_DEBUG 2905 static void btree_trans_verify_sorted_refs(struct btree_trans *trans) 2906 { 2907 struct btree_path *path; 2908 unsigned i; 2909 2910 BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1); 2911 2912 trans_for_each_path(trans, path, i) { 2913 BUG_ON(path->sorted_idx >= trans->nr_sorted); 2914 BUG_ON(trans->sorted[path->sorted_idx] != i); 2915 } 2916 2917 for (i = 0; i < trans->nr_sorted; i++) { 2918 unsigned idx = trans->sorted[i]; 2919 2920 BUG_ON(!test_bit(idx, trans->paths_allocated)); 2921 BUG_ON(trans->paths[idx].sorted_idx != i); 2922 } 2923 } 2924 2925 static void btree_trans_verify_sorted(struct btree_trans *trans) 2926 { 2927 struct btree_path *path, *prev = NULL; 2928 struct trans_for_each_path_inorder_iter iter; 2929 2930 if (!bch2_debug_check_iterators) 2931 return; 2932 2933 trans_for_each_path_inorder(trans, path, iter) { 2934 if (prev && btree_path_cmp(prev, path) > 0) { 2935 __bch2_dump_trans_paths_updates(trans, true); 2936 panic("trans paths out of order!\n"); 2937 } 2938 prev = path; 2939 } 2940 } 2941 #else 2942 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {} 2943 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {} 2944 #endif 2945 2946 void __bch2_btree_trans_sort_paths(struct btree_trans *trans) 2947 { 2948 int i, l = 0, r = trans->nr_sorted, inc = 1; 2949 bool swapped; 2950 2951 btree_trans_verify_sorted_refs(trans); 2952 2953 if (trans->paths_sorted) 2954 goto out; 2955 2956 /* 2957 * Cocktail shaker sort: this is efficient because iterators will be 2958 * mostly sorted. 2959 */ 2960 do { 2961 swapped = false; 2962 2963 for (i = inc > 0 ? l : r - 2; 2964 i + 1 < r && i >= l; 2965 i += inc) { 2966 if (btree_path_cmp(trans->paths + trans->sorted[i], 2967 trans->paths + trans->sorted[i + 1]) > 0) { 2968 swap(trans->sorted[i], trans->sorted[i + 1]); 2969 trans->paths[trans->sorted[i]].sorted_idx = i; 2970 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1; 2971 swapped = true; 2972 } 2973 } 2974 2975 if (inc > 0) 2976 --r; 2977 else 2978 l++; 2979 inc = -inc; 2980 } while (swapped); 2981 2982 trans->paths_sorted = true; 2983 out: 2984 btree_trans_verify_sorted(trans); 2985 } 2986 2987 static inline void btree_path_list_remove(struct btree_trans *trans, 2988 struct btree_path *path) 2989 { 2990 EBUG_ON(path->sorted_idx >= trans->nr_sorted); 2991 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS 2992 trans->nr_sorted--; 2993 memmove_u64s_down_small(trans->sorted + path->sorted_idx, 2994 trans->sorted + path->sorted_idx + 1, 2995 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 2996 sizeof(u64) / sizeof(btree_path_idx_t))); 2997 #else 2998 array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx); 2999 #endif 3000 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++) 3001 trans->paths[trans->sorted[i]].sorted_idx = i; 3002 } 3003 3004 static inline void btree_path_list_add(struct btree_trans *trans, 3005 btree_path_idx_t pos, 3006 btree_path_idx_t path_idx) 3007 { 3008 struct btree_path *path = trans->paths + path_idx; 3009 3010 path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted; 3011 3012 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS 3013 memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1, 3014 trans->sorted + path->sorted_idx, 3015 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 3016 sizeof(u64) / sizeof(btree_path_idx_t))); 3017 trans->nr_sorted++; 3018 trans->sorted[path->sorted_idx] = path_idx; 3019 #else 3020 array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx); 3021 #endif 3022 3023 for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++) 3024 trans->paths[trans->sorted[i]].sorted_idx = i; 3025 3026 btree_trans_verify_sorted_refs(trans); 3027 } 3028 3029 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter) 3030 { 3031 if (iter->update_path) 3032 bch2_path_put_nokeep(trans, iter->update_path, 3033 iter->flags & BTREE_ITER_intent); 3034 if (iter->path) 3035 bch2_path_put(trans, iter->path, 3036 iter->flags & BTREE_ITER_intent); 3037 if (iter->key_cache_path) 3038 bch2_path_put(trans, iter->key_cache_path, 3039 iter->flags & BTREE_ITER_intent); 3040 iter->path = 0; 3041 iter->update_path = 0; 3042 iter->key_cache_path = 0; 3043 iter->trans = NULL; 3044 } 3045 3046 void bch2_trans_iter_init_outlined(struct btree_trans *trans, 3047 struct btree_iter *iter, 3048 enum btree_id btree_id, struct bpos pos, 3049 unsigned flags) 3050 { 3051 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0, 3052 bch2_btree_iter_flags(trans, btree_id, 0, flags), 3053 _RET_IP_); 3054 } 3055 3056 void bch2_trans_node_iter_init(struct btree_trans *trans, 3057 struct btree_iter *iter, 3058 enum btree_id btree_id, 3059 struct bpos pos, 3060 unsigned locks_want, 3061 unsigned depth, 3062 unsigned flags) 3063 { 3064 flags |= BTREE_ITER_not_extents; 3065 flags |= BTREE_ITER_snapshot_field; 3066 flags |= BTREE_ITER_all_snapshots; 3067 3068 if (!depth && btree_id_cached(trans->c, btree_id)) 3069 flags |= BTREE_ITER_with_key_cache; 3070 3071 bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth, 3072 bch2_btree_iter_flags(trans, btree_id, depth, flags), 3073 _RET_IP_); 3074 3075 iter->min_depth = depth; 3076 3077 struct btree_path *path = btree_iter_path(trans, iter); 3078 BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH)); 3079 BUG_ON(path->level != depth); 3080 BUG_ON(iter->min_depth != depth); 3081 } 3082 3083 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src) 3084 { 3085 struct btree_trans *trans = src->trans; 3086 3087 *dst = *src; 3088 #ifdef TRACK_PATH_ALLOCATED 3089 dst->ip_allocated = _RET_IP_; 3090 #endif 3091 if (src->path) 3092 __btree_path_get(trans, trans->paths + src->path, src->flags & BTREE_ITER_intent); 3093 if (src->update_path) 3094 __btree_path_get(trans, trans->paths + src->update_path, src->flags & BTREE_ITER_intent); 3095 dst->key_cache_path = 0; 3096 } 3097 3098 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size) 3099 { 3100 struct bch_fs *c = trans->c; 3101 unsigned new_top = trans->mem_top + size; 3102 unsigned old_bytes = trans->mem_bytes; 3103 unsigned new_bytes = roundup_pow_of_two(new_top); 3104 int ret; 3105 void *new_mem; 3106 void *p; 3107 3108 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX); 3109 3110 struct btree_transaction_stats *s = btree_trans_stats(trans); 3111 s->max_mem = max(s->max_mem, new_bytes); 3112 3113 if (trans->used_mempool) { 3114 if (trans->mem_bytes >= new_bytes) 3115 goto out_change_top; 3116 3117 /* No more space from mempool item, need malloc new one */ 3118 new_mem = kmalloc(new_bytes, GFP_NOWAIT|__GFP_NOWARN); 3119 if (unlikely(!new_mem)) { 3120 bch2_trans_unlock(trans); 3121 3122 new_mem = kmalloc(new_bytes, GFP_KERNEL); 3123 if (!new_mem) 3124 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc); 3125 3126 ret = bch2_trans_relock(trans); 3127 if (ret) { 3128 kfree(new_mem); 3129 return ERR_PTR(ret); 3130 } 3131 } 3132 memcpy(new_mem, trans->mem, trans->mem_top); 3133 trans->used_mempool = false; 3134 mempool_free(trans->mem, &c->btree_trans_mem_pool); 3135 goto out_new_mem; 3136 } 3137 3138 new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN); 3139 if (unlikely(!new_mem)) { 3140 bch2_trans_unlock(trans); 3141 3142 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL); 3143 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) { 3144 new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL); 3145 new_bytes = BTREE_TRANS_MEM_MAX; 3146 memcpy(new_mem, trans->mem, trans->mem_top); 3147 trans->used_mempool = true; 3148 kfree(trans->mem); 3149 } 3150 3151 if (!new_mem) 3152 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc); 3153 3154 trans->mem = new_mem; 3155 trans->mem_bytes = new_bytes; 3156 3157 ret = bch2_trans_relock(trans); 3158 if (ret) 3159 return ERR_PTR(ret); 3160 } 3161 out_new_mem: 3162 trans->mem = new_mem; 3163 trans->mem_bytes = new_bytes; 3164 3165 if (old_bytes) { 3166 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes); 3167 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced)); 3168 } 3169 out_change_top: 3170 p = trans->mem + trans->mem_top; 3171 trans->mem_top += size; 3172 memset(p, 0, size); 3173 return p; 3174 } 3175 3176 static inline void check_srcu_held_too_long(struct btree_trans *trans) 3177 { 3178 WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10), 3179 "btree trans held srcu lock (delaying memory reclaim) for %lu seconds", 3180 (jiffies - trans->srcu_lock_time) / HZ); 3181 } 3182 3183 void bch2_trans_srcu_unlock(struct btree_trans *trans) 3184 { 3185 if (trans->srcu_held) { 3186 struct bch_fs *c = trans->c; 3187 struct btree_path *path; 3188 unsigned i; 3189 3190 trans_for_each_path(trans, path, i) 3191 if (path->cached && !btree_node_locked(path, 0)) 3192 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset); 3193 3194 check_srcu_held_too_long(trans); 3195 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); 3196 trans->srcu_held = false; 3197 } 3198 } 3199 3200 static void bch2_trans_srcu_lock(struct btree_trans *trans) 3201 { 3202 if (!trans->srcu_held) { 3203 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier); 3204 trans->srcu_lock_time = jiffies; 3205 trans->srcu_held = true; 3206 } 3207 } 3208 3209 /** 3210 * bch2_trans_begin() - reset a transaction after a interrupted attempt 3211 * @trans: transaction to reset 3212 * 3213 * Returns: current restart counter, to be used with trans_was_restarted() 3214 * 3215 * While iterating over nodes or updating nodes a attempt to lock a btree node 3216 * may return BCH_ERR_transaction_restart when the trylock fails. When this 3217 * occurs bch2_trans_begin() should be called and the transaction retried. 3218 */ 3219 u32 bch2_trans_begin(struct btree_trans *trans) 3220 { 3221 struct btree_path *path; 3222 unsigned i; 3223 u64 now; 3224 3225 bch2_trans_reset_updates(trans); 3226 3227 trans->restart_count++; 3228 trans->mem_top = 0; 3229 trans->journal_entries = NULL; 3230 3231 trans_for_each_path(trans, path, i) { 3232 path->should_be_locked = false; 3233 3234 /* 3235 * If the transaction wasn't restarted, we're presuming to be 3236 * doing something new: dont keep iterators excpt the ones that 3237 * are in use - except for the subvolumes btree: 3238 */ 3239 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes) 3240 path->preserve = false; 3241 3242 /* 3243 * XXX: we probably shouldn't be doing this if the transaction 3244 * was restarted, but currently we still overflow transaction 3245 * iterators if we do that 3246 */ 3247 if (!path->ref && !path->preserve) 3248 __bch2_path_free(trans, i); 3249 else 3250 path->preserve = false; 3251 } 3252 3253 now = local_clock(); 3254 3255 if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) && 3256 time_after64(now, trans->last_begin_time + 10)) 3257 __bch2_time_stats_update(&btree_trans_stats(trans)->duration, 3258 trans->last_begin_time, now); 3259 3260 if (!trans->restarted && 3261 (need_resched() || 3262 time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) { 3263 bch2_trans_unlock(trans); 3264 cond_resched(); 3265 now = local_clock(); 3266 } 3267 trans->last_begin_time = now; 3268 3269 if (unlikely(trans->srcu_held && 3270 time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10)))) 3271 bch2_trans_srcu_unlock(trans); 3272 3273 trans->last_begin_ip = _RET_IP_; 3274 3275 trans_set_locked(trans, false); 3276 3277 if (trans->restarted) { 3278 bch2_btree_path_traverse_all(trans); 3279 trans->notrace_relock_fail = false; 3280 } 3281 3282 bch2_trans_verify_not_unlocked_or_in_restart(trans); 3283 return trans->restart_count; 3284 } 3285 3286 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" }; 3287 3288 unsigned bch2_trans_get_fn_idx(const char *fn) 3289 { 3290 for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++) 3291 if (!bch2_btree_transaction_fns[i] || 3292 bch2_btree_transaction_fns[i] == fn) { 3293 bch2_btree_transaction_fns[i] = fn; 3294 return i; 3295 } 3296 3297 pr_warn_once("BCH_TRANSACTIONS_NR not big enough!"); 3298 return 0; 3299 } 3300 3301 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx) 3302 __acquires(&c->btree_trans_barrier) 3303 { 3304 struct btree_trans *trans; 3305 3306 if (IS_ENABLED(__KERNEL__)) { 3307 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL); 3308 if (trans) { 3309 memset(trans, 0, offsetof(struct btree_trans, list)); 3310 goto got_trans; 3311 } 3312 } 3313 3314 trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS); 3315 memset(trans, 0, sizeof(*trans)); 3316 3317 seqmutex_lock(&c->btree_trans_lock); 3318 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) { 3319 struct btree_trans *pos; 3320 pid_t pid = current->pid; 3321 3322 trans->locking_wait.task = current; 3323 3324 list_for_each_entry(pos, &c->btree_trans_list, list) { 3325 struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task); 3326 /* 3327 * We'd much prefer to be stricter here and completely 3328 * disallow multiple btree_trans in the same thread - 3329 * but the data move path calls bch2_write when we 3330 * already have a btree_trans initialized. 3331 */ 3332 BUG_ON(pos_task && 3333 pid == pos_task->pid && 3334 pos->locked); 3335 } 3336 } 3337 3338 list_add(&trans->list, &c->btree_trans_list); 3339 seqmutex_unlock(&c->btree_trans_lock); 3340 got_trans: 3341 trans->c = c; 3342 trans->last_begin_time = local_clock(); 3343 trans->fn_idx = fn_idx; 3344 trans->locking_wait.task = current; 3345 trans->journal_replay_not_finished = 3346 unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)) && 3347 atomic_inc_not_zero(&c->journal_keys.ref); 3348 trans->nr_paths = ARRAY_SIZE(trans->_paths); 3349 trans->paths_allocated = trans->_paths_allocated; 3350 trans->sorted = trans->_sorted; 3351 trans->paths = trans->_paths; 3352 trans->updates = trans->_updates; 3353 3354 *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL; 3355 3356 trans->paths_allocated[0] = 1; 3357 3358 static struct lock_class_key lockdep_key; 3359 lockdep_init_map(&trans->dep_map, "bcachefs_btree", &lockdep_key, 0); 3360 3361 if (fn_idx < BCH_TRANSACTIONS_NR) { 3362 trans->fn = bch2_btree_transaction_fns[fn_idx]; 3363 3364 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx]; 3365 3366 if (s->max_mem) { 3367 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem); 3368 3369 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL); 3370 if (likely(trans->mem)) 3371 trans->mem_bytes = expected_mem_bytes; 3372 } 3373 3374 trans->nr_paths_max = s->nr_max_paths; 3375 trans->journal_entries_size = s->journal_entries_size; 3376 } 3377 3378 trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier); 3379 trans->srcu_lock_time = jiffies; 3380 trans->srcu_held = true; 3381 trans_set_locked(trans, false); 3382 3383 closure_init_stack_release(&trans->ref); 3384 return trans; 3385 } 3386 3387 static void check_btree_paths_leaked(struct btree_trans *trans) 3388 { 3389 #ifdef CONFIG_BCACHEFS_DEBUG 3390 struct bch_fs *c = trans->c; 3391 struct btree_path *path; 3392 unsigned i; 3393 3394 trans_for_each_path(trans, path, i) 3395 if (path->ref) 3396 goto leaked; 3397 return; 3398 leaked: 3399 bch_err(c, "btree paths leaked from %s!", trans->fn); 3400 trans_for_each_path(trans, path, i) 3401 if (path->ref) 3402 printk(KERN_ERR " btree %s %pS\n", 3403 bch2_btree_id_str(path->btree_id), 3404 (void *) path->ip_allocated); 3405 /* Be noisy about this: */ 3406 bch2_fatal_error(c); 3407 #endif 3408 } 3409 3410 void bch2_trans_put(struct btree_trans *trans) 3411 __releases(&c->btree_trans_barrier) 3412 { 3413 struct bch_fs *c = trans->c; 3414 3415 if (trans->restarted) 3416 bch2_trans_in_restart_error(trans); 3417 3418 bch2_trans_unlock(trans); 3419 3420 trans_for_each_update(trans, i) 3421 __btree_path_put(trans, trans->paths + i->path, true); 3422 trans->nr_updates = 0; 3423 3424 check_btree_paths_leaked(trans); 3425 3426 if (trans->srcu_held) { 3427 check_srcu_held_too_long(trans); 3428 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); 3429 } 3430 3431 if (unlikely(trans->journal_replay_not_finished)) 3432 bch2_journal_keys_put(c); 3433 3434 /* 3435 * trans->ref protects trans->locking_wait.task, btree_paths array; used 3436 * by cycle detector 3437 */ 3438 closure_return_sync(&trans->ref); 3439 trans->locking_wait.task = NULL; 3440 3441 #ifdef CONFIG_BCACHEFS_DEBUG 3442 darray_exit(&trans->last_restarted_trace); 3443 #endif 3444 3445 unsigned long *paths_allocated = trans->paths_allocated; 3446 trans->paths_allocated = NULL; 3447 trans->paths = NULL; 3448 3449 if (paths_allocated != trans->_paths_allocated) 3450 kvfree_rcu_mightsleep(paths_allocated); 3451 3452 if (trans->used_mempool) 3453 mempool_free(trans->mem, &c->btree_trans_mem_pool); 3454 else 3455 kfree(trans->mem); 3456 3457 /* Userspace doesn't have a real percpu implementation: */ 3458 if (IS_ENABLED(__KERNEL__)) 3459 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans); 3460 3461 if (trans) { 3462 seqmutex_lock(&c->btree_trans_lock); 3463 list_del(&trans->list); 3464 seqmutex_unlock(&c->btree_trans_lock); 3465 3466 mempool_free(trans, &c->btree_trans_pool); 3467 } 3468 } 3469 3470 bool bch2_current_has_btree_trans(struct bch_fs *c) 3471 { 3472 seqmutex_lock(&c->btree_trans_lock); 3473 struct btree_trans *trans; 3474 bool ret = false; 3475 list_for_each_entry(trans, &c->btree_trans_list, list) 3476 if (trans->locking_wait.task == current && 3477 trans->locked) { 3478 ret = true; 3479 break; 3480 } 3481 seqmutex_unlock(&c->btree_trans_lock); 3482 return ret; 3483 } 3484 3485 static void __maybe_unused 3486 bch2_btree_bkey_cached_common_to_text(struct printbuf *out, 3487 struct btree_bkey_cached_common *b) 3488 { 3489 struct six_lock_count c = six_lock_counts(&b->lock); 3490 struct task_struct *owner; 3491 pid_t pid; 3492 3493 rcu_read_lock(); 3494 owner = READ_ONCE(b->lock.owner); 3495 pid = owner ? owner->pid : 0; 3496 rcu_read_unlock(); 3497 3498 prt_printf(out, "\t%px %c ", b, b->cached ? 'c' : 'b'); 3499 bch2_btree_id_to_text(out, b->btree_id); 3500 prt_printf(out, " l=%u:", b->level); 3501 bch2_bpos_to_text(out, btree_node_pos(b)); 3502 3503 prt_printf(out, "\t locks %u:%u:%u held by pid %u", 3504 c.n[0], c.n[1], c.n[2], pid); 3505 } 3506 3507 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans) 3508 { 3509 struct btree_bkey_cached_common *b; 3510 static char lock_types[] = { 'r', 'i', 'w' }; 3511 struct task_struct *task = READ_ONCE(trans->locking_wait.task); 3512 unsigned l, idx; 3513 3514 /* before rcu_read_lock(): */ 3515 bch2_printbuf_make_room(out, 4096); 3516 3517 if (!out->nr_tabstops) { 3518 printbuf_tabstop_push(out, 16); 3519 printbuf_tabstop_push(out, 32); 3520 } 3521 3522 prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn); 3523 3524 /* trans->paths is rcu protected vs. freeing */ 3525 rcu_read_lock(); 3526 out->atomic++; 3527 3528 struct btree_path *paths = rcu_dereference(trans->paths); 3529 if (!paths) 3530 goto out; 3531 3532 unsigned long *paths_allocated = trans_paths_allocated(paths); 3533 3534 trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) { 3535 struct btree_path *path = paths + idx; 3536 if (!path->nodes_locked) 3537 continue; 3538 3539 prt_printf(out, " path %u %c ", 3540 idx, 3541 path->cached ? 'c' : 'b'); 3542 bch2_btree_id_to_text(out, path->btree_id); 3543 prt_printf(out, " l=%u:", path->level); 3544 bch2_bpos_to_text(out, path->pos); 3545 prt_newline(out); 3546 3547 for (l = 0; l < BTREE_MAX_DEPTH; l++) { 3548 if (btree_node_locked(path, l) && 3549 !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) { 3550 prt_printf(out, " %c l=%u ", 3551 lock_types[btree_node_locked_type(path, l)], l); 3552 bch2_btree_bkey_cached_common_to_text(out, b); 3553 prt_newline(out); 3554 } 3555 } 3556 } 3557 3558 b = READ_ONCE(trans->locking); 3559 if (b) { 3560 prt_printf(out, " blocked for %lluus on\n", 3561 div_u64(local_clock() - trans->locking_wait.start_time, 1000)); 3562 prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]); 3563 bch2_btree_bkey_cached_common_to_text(out, b); 3564 prt_newline(out); 3565 } 3566 out: 3567 --out->atomic; 3568 rcu_read_unlock(); 3569 } 3570 3571 void bch2_fs_btree_iter_exit(struct bch_fs *c) 3572 { 3573 struct btree_transaction_stats *s; 3574 struct btree_trans *trans; 3575 int cpu; 3576 3577 if (c->btree_trans_bufs) 3578 for_each_possible_cpu(cpu) { 3579 struct btree_trans *trans = 3580 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans; 3581 3582 if (trans) { 3583 seqmutex_lock(&c->btree_trans_lock); 3584 list_del(&trans->list); 3585 seqmutex_unlock(&c->btree_trans_lock); 3586 } 3587 kfree(trans); 3588 } 3589 free_percpu(c->btree_trans_bufs); 3590 3591 trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list); 3592 if (trans) 3593 panic("%s leaked btree_trans\n", trans->fn); 3594 3595 for (s = c->btree_transaction_stats; 3596 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); 3597 s++) { 3598 kfree(s->max_paths_text); 3599 bch2_time_stats_exit(&s->lock_hold_times); 3600 } 3601 3602 if (c->btree_trans_barrier_initialized) { 3603 synchronize_srcu_expedited(&c->btree_trans_barrier); 3604 cleanup_srcu_struct(&c->btree_trans_barrier); 3605 } 3606 mempool_exit(&c->btree_trans_mem_pool); 3607 mempool_exit(&c->btree_trans_pool); 3608 } 3609 3610 void bch2_fs_btree_iter_init_early(struct bch_fs *c) 3611 { 3612 struct btree_transaction_stats *s; 3613 3614 for (s = c->btree_transaction_stats; 3615 s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); 3616 s++) { 3617 bch2_time_stats_init(&s->duration); 3618 bch2_time_stats_init(&s->lock_hold_times); 3619 mutex_init(&s->lock); 3620 } 3621 3622 INIT_LIST_HEAD(&c->btree_trans_list); 3623 seqmutex_init(&c->btree_trans_lock); 3624 } 3625 3626 int bch2_fs_btree_iter_init(struct bch_fs *c) 3627 { 3628 int ret; 3629 3630 c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf); 3631 if (!c->btree_trans_bufs) 3632 return -ENOMEM; 3633 3634 ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1, 3635 sizeof(struct btree_trans)) ?: 3636 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1, 3637 BTREE_TRANS_MEM_MAX) ?: 3638 init_srcu_struct(&c->btree_trans_barrier); 3639 if (ret) 3640 return ret; 3641 3642 /* 3643 * static annotation (hackily done) for lock ordering of reclaim vs. 3644 * btree node locks: 3645 */ 3646 #ifdef CONFIG_LOCKDEP 3647 fs_reclaim_acquire(GFP_KERNEL); 3648 struct btree_trans *trans = bch2_trans_get(c); 3649 trans_set_locked(trans, false); 3650 bch2_trans_put(trans); 3651 fs_reclaim_release(GFP_KERNEL); 3652 #endif 3653 3654 c->btree_trans_barrier_initialized = true; 3655 return 0; 3656 3657 } 3658