1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "bcachefs.h" 4 #include "alloc_foreground.h" 5 #include "bkey_buf.h" 6 #include "bkey_methods.h" 7 #include "btree_cache.h" 8 #include "btree_gc.h" 9 #include "btree_journal_iter.h" 10 #include "btree_update.h" 11 #include "btree_update_interior.h" 12 #include "btree_io.h" 13 #include "btree_iter.h" 14 #include "btree_locking.h" 15 #include "buckets.h" 16 #include "clock.h" 17 #include "error.h" 18 #include "extents.h" 19 #include "journal.h" 20 #include "journal_reclaim.h" 21 #include "keylist.h" 22 #include "recovery_passes.h" 23 #include "replicas.h" 24 #include "super-io.h" 25 #include "trace.h" 26 27 #include <linux/random.h> 28 29 const char * const bch2_btree_update_modes[] = { 30 #define x(t) #t, 31 BCH_WATERMARKS() 32 #undef x 33 NULL 34 }; 35 36 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *, 37 btree_path_idx_t, struct btree *, struct keylist *); 38 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *); 39 40 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans, 41 enum btree_id btree_id, 42 unsigned level, 43 struct bpos pos) 44 { 45 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level, 46 BTREE_ITER_NOPRESERVE| 47 BTREE_ITER_INTENT, _RET_IP_); 48 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_); 49 50 struct btree_path *path = trans->paths + path_idx; 51 bch2_btree_path_downgrade(trans, path); 52 __bch2_btree_path_unlock(trans, path); 53 return path_idx; 54 } 55 56 /* 57 * Verify that child nodes correctly span parent node's range: 58 */ 59 int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b) 60 { 61 struct bch_fs *c = trans->c; 62 struct bpos node_min = b->key.k.type == KEY_TYPE_btree_ptr_v2 63 ? bkey_i_to_btree_ptr_v2(&b->key)->v.min_key 64 : b->data->min_key; 65 struct btree_and_journal_iter iter; 66 struct bkey_s_c k; 67 struct printbuf buf = PRINTBUF; 68 struct bkey_buf prev; 69 int ret = 0; 70 71 BUG_ON(b->key.k.type == KEY_TYPE_btree_ptr_v2 && 72 !bpos_eq(bkey_i_to_btree_ptr_v2(&b->key)->v.min_key, 73 b->data->min_key)); 74 75 if (!b->c.level) 76 return 0; 77 78 bch2_bkey_buf_init(&prev); 79 bkey_init(&prev.k->k); 80 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b); 81 82 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) { 83 if (k.k->type != KEY_TYPE_btree_ptr_v2) 84 goto out; 85 86 struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k); 87 88 struct bpos expected_min = bkey_deleted(&prev.k->k) 89 ? node_min 90 : bpos_successor(prev.k->k.p); 91 92 if (!bpos_eq(expected_min, bp.v->min_key)) { 93 bch2_topology_error(c); 94 95 printbuf_reset(&buf); 96 prt_str(&buf, "end of prev node doesn't match start of next node\n"), 97 prt_printf(&buf, " in btree %s level %u node ", 98 bch2_btree_id_str(b->c.btree_id), b->c.level); 99 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key)); 100 prt_str(&buf, "\n prev "); 101 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k)); 102 prt_str(&buf, "\n next "); 103 bch2_bkey_val_to_text(&buf, c, k); 104 105 need_fsck_err(c, btree_node_topology_bad_min_key, "%s", buf.buf); 106 goto topology_repair; 107 } 108 109 bch2_bkey_buf_reassemble(&prev, c, k); 110 bch2_btree_and_journal_iter_advance(&iter); 111 } 112 113 if (bkey_deleted(&prev.k->k)) { 114 bch2_topology_error(c); 115 116 printbuf_reset(&buf); 117 prt_str(&buf, "empty interior node\n"); 118 prt_printf(&buf, " in btree %s level %u node ", 119 bch2_btree_id_str(b->c.btree_id), b->c.level); 120 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key)); 121 122 need_fsck_err(c, btree_node_topology_empty_interior_node, "%s", buf.buf); 123 goto topology_repair; 124 } else if (!bpos_eq(prev.k->k.p, b->key.k.p)) { 125 bch2_topology_error(c); 126 127 printbuf_reset(&buf); 128 prt_str(&buf, "last child node doesn't end at end of parent node\n"); 129 prt_printf(&buf, " in btree %s level %u node ", 130 bch2_btree_id_str(b->c.btree_id), b->c.level); 131 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key)); 132 prt_str(&buf, "\n last key "); 133 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k)); 134 135 need_fsck_err(c, btree_node_topology_bad_max_key, "%s", buf.buf); 136 goto topology_repair; 137 } 138 out: 139 fsck_err: 140 bch2_btree_and_journal_iter_exit(&iter); 141 bch2_bkey_buf_exit(&prev, c); 142 printbuf_exit(&buf); 143 return ret; 144 topology_repair: 145 if ((c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology)) && 146 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology) { 147 bch2_inconsistent_error(c); 148 ret = -BCH_ERR_btree_need_topology_repair; 149 } else { 150 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology); 151 } 152 goto out; 153 } 154 155 /* Calculate ideal packed bkey format for new btree nodes: */ 156 157 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b) 158 { 159 struct bkey_packed *k; 160 struct bset_tree *t; 161 struct bkey uk; 162 163 for_each_bset(b, t) 164 bset_tree_for_each_key(b, t, k) 165 if (!bkey_deleted(k)) { 166 uk = bkey_unpack_key(b, k); 167 bch2_bkey_format_add_key(s, &uk); 168 } 169 } 170 171 static struct bkey_format bch2_btree_calc_format(struct btree *b) 172 { 173 struct bkey_format_state s; 174 175 bch2_bkey_format_init(&s); 176 bch2_bkey_format_add_pos(&s, b->data->min_key); 177 bch2_bkey_format_add_pos(&s, b->data->max_key); 178 __bch2_btree_calc_format(&s, b); 179 180 return bch2_bkey_format_done(&s); 181 } 182 183 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr, 184 struct bkey_format *old_f, 185 struct bkey_format *new_f) 186 { 187 /* stupid integer promotion rules */ 188 ssize_t delta = 189 (((int) new_f->key_u64s - old_f->key_u64s) * 190 (int) nr.packed_keys) + 191 (((int) new_f->key_u64s - BKEY_U64s) * 192 (int) nr.unpacked_keys); 193 194 BUG_ON(delta + nr.live_u64s < 0); 195 196 return nr.live_u64s + delta; 197 } 198 199 /** 200 * bch2_btree_node_format_fits - check if we could rewrite node with a new format 201 * 202 * @c: filesystem handle 203 * @b: btree node to rewrite 204 * @nr: number of keys for new node (i.e. b->nr) 205 * @new_f: bkey format to translate keys to 206 * 207 * Returns: true if all re-packed keys will be able to fit in a new node. 208 * 209 * Assumes all keys will successfully pack with the new format. 210 */ 211 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b, 212 struct btree_nr_keys nr, 213 struct bkey_format *new_f) 214 { 215 size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f); 216 217 return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b); 218 } 219 220 /* Btree node freeing/allocation: */ 221 222 static void __btree_node_free(struct btree_trans *trans, struct btree *b) 223 { 224 struct bch_fs *c = trans->c; 225 226 trace_and_count(c, btree_node_free, trans, b); 227 228 BUG_ON(btree_node_write_blocked(b)); 229 BUG_ON(btree_node_dirty(b)); 230 BUG_ON(btree_node_need_write(b)); 231 BUG_ON(b == btree_node_root(c, b)); 232 BUG_ON(b->ob.nr); 233 BUG_ON(!list_empty(&b->write_blocked)); 234 BUG_ON(b->will_make_reachable); 235 236 clear_btree_node_noevict(b); 237 238 mutex_lock(&c->btree_cache.lock); 239 list_move(&b->list, &c->btree_cache.freeable); 240 mutex_unlock(&c->btree_cache.lock); 241 } 242 243 static void bch2_btree_node_free_inmem(struct btree_trans *trans, 244 struct btree_path *path, 245 struct btree *b) 246 { 247 struct bch_fs *c = trans->c; 248 unsigned i, level = b->c.level; 249 250 bch2_btree_node_lock_write_nofail(trans, path, &b->c); 251 bch2_btree_node_hash_remove(&c->btree_cache, b); 252 __btree_node_free(trans, b); 253 six_unlock_write(&b->c.lock); 254 mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED); 255 256 trans_for_each_path(trans, path, i) 257 if (path->l[level].b == b) { 258 btree_node_unlock(trans, path, level); 259 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 260 } 261 } 262 263 static void bch2_btree_node_free_never_used(struct btree_update *as, 264 struct btree_trans *trans, 265 struct btree *b) 266 { 267 struct bch_fs *c = as->c; 268 struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL]; 269 struct btree_path *path; 270 unsigned i, level = b->c.level; 271 272 BUG_ON(!list_empty(&b->write_blocked)); 273 BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as)); 274 275 b->will_make_reachable = 0; 276 closure_put(&as->cl); 277 278 clear_btree_node_will_make_reachable(b); 279 clear_btree_node_accessed(b); 280 clear_btree_node_dirty_acct(c, b); 281 clear_btree_node_need_write(b); 282 283 mutex_lock(&c->btree_cache.lock); 284 list_del_init(&b->list); 285 bch2_btree_node_hash_remove(&c->btree_cache, b); 286 mutex_unlock(&c->btree_cache.lock); 287 288 BUG_ON(p->nr >= ARRAY_SIZE(p->b)); 289 p->b[p->nr++] = b; 290 291 six_unlock_intent(&b->c.lock); 292 293 trans_for_each_path(trans, path, i) 294 if (path->l[level].b == b) { 295 btree_node_unlock(trans, path, level); 296 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); 297 } 298 } 299 300 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans, 301 struct disk_reservation *res, 302 struct closure *cl, 303 bool interior_node, 304 unsigned flags) 305 { 306 struct bch_fs *c = trans->c; 307 struct write_point *wp; 308 struct btree *b; 309 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp; 310 struct open_buckets obs = { .nr = 0 }; 311 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 }; 312 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK; 313 unsigned nr_reserve = watermark < BCH_WATERMARK_reclaim 314 ? BTREE_NODE_RESERVE 315 : 0; 316 int ret; 317 318 mutex_lock(&c->btree_reserve_cache_lock); 319 if (c->btree_reserve_cache_nr > nr_reserve) { 320 struct btree_alloc *a = 321 &c->btree_reserve_cache[--c->btree_reserve_cache_nr]; 322 323 obs = a->ob; 324 bkey_copy(&tmp.k, &a->k); 325 mutex_unlock(&c->btree_reserve_cache_lock); 326 goto mem_alloc; 327 } 328 mutex_unlock(&c->btree_reserve_cache_lock); 329 330 retry: 331 ret = bch2_alloc_sectors_start_trans(trans, 332 c->opts.metadata_target ?: 333 c->opts.foreground_target, 334 0, 335 writepoint_ptr(&c->btree_write_point), 336 &devs_have, 337 res->nr_replicas, 338 min(res->nr_replicas, 339 c->opts.metadata_replicas_required), 340 watermark, 0, cl, &wp); 341 if (unlikely(ret)) 342 return ERR_PTR(ret); 343 344 if (wp->sectors_free < btree_sectors(c)) { 345 struct open_bucket *ob; 346 unsigned i; 347 348 open_bucket_for_each(c, &wp->ptrs, ob, i) 349 if (ob->sectors_free < btree_sectors(c)) 350 ob->sectors_free = 0; 351 352 bch2_alloc_sectors_done(c, wp); 353 goto retry; 354 } 355 356 bkey_btree_ptr_v2_init(&tmp.k); 357 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false); 358 359 bch2_open_bucket_get(c, wp, &obs); 360 bch2_alloc_sectors_done(c, wp); 361 mem_alloc: 362 b = bch2_btree_node_mem_alloc(trans, interior_node); 363 six_unlock_write(&b->c.lock); 364 six_unlock_intent(&b->c.lock); 365 366 /* we hold cannibalize_lock: */ 367 BUG_ON(IS_ERR(b)); 368 BUG_ON(b->ob.nr); 369 370 bkey_copy(&b->key, &tmp.k); 371 b->ob = obs; 372 373 return b; 374 } 375 376 static struct btree *bch2_btree_node_alloc(struct btree_update *as, 377 struct btree_trans *trans, 378 unsigned level) 379 { 380 struct bch_fs *c = as->c; 381 struct btree *b; 382 struct prealloc_nodes *p = &as->prealloc_nodes[!!level]; 383 int ret; 384 385 BUG_ON(level >= BTREE_MAX_DEPTH); 386 BUG_ON(!p->nr); 387 388 b = p->b[--p->nr]; 389 390 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent); 391 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write); 392 393 set_btree_node_accessed(b); 394 set_btree_node_dirty_acct(c, b); 395 set_btree_node_need_write(b); 396 397 bch2_bset_init_first(b, &b->data->keys); 398 b->c.level = level; 399 b->c.btree_id = as->btree_id; 400 b->version_ondisk = c->sb.version; 401 402 memset(&b->nr, 0, sizeof(b->nr)); 403 b->data->magic = cpu_to_le64(bset_magic(c)); 404 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr)); 405 b->data->flags = 0; 406 SET_BTREE_NODE_ID(b->data, as->btree_id); 407 SET_BTREE_NODE_LEVEL(b->data, level); 408 409 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) { 410 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key); 411 412 bp->v.mem_ptr = 0; 413 bp->v.seq = b->data->keys.seq; 414 bp->v.sectors_written = 0; 415 } 416 417 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true); 418 419 bch2_btree_build_aux_trees(b); 420 421 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id); 422 BUG_ON(ret); 423 424 trace_and_count(c, btree_node_alloc, trans, b); 425 bch2_increment_clock(c, btree_sectors(c), WRITE); 426 return b; 427 } 428 429 static void btree_set_min(struct btree *b, struct bpos pos) 430 { 431 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) 432 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos; 433 b->data->min_key = pos; 434 } 435 436 static void btree_set_max(struct btree *b, struct bpos pos) 437 { 438 b->key.k.p = pos; 439 b->data->max_key = pos; 440 } 441 442 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as, 443 struct btree_trans *trans, 444 struct btree *b) 445 { 446 struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level); 447 struct bkey_format format = bch2_btree_calc_format(b); 448 449 /* 450 * The keys might expand with the new format - if they wouldn't fit in 451 * the btree node anymore, use the old format for now: 452 */ 453 if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format)) 454 format = b->format; 455 456 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1); 457 458 btree_set_min(n, b->data->min_key); 459 btree_set_max(n, b->data->max_key); 460 461 n->data->format = format; 462 btree_node_set_format(n, format); 463 464 bch2_btree_sort_into(as->c, n, b); 465 466 btree_node_reset_sib_u64s(n); 467 return n; 468 } 469 470 static struct btree *__btree_root_alloc(struct btree_update *as, 471 struct btree_trans *trans, unsigned level) 472 { 473 struct btree *b = bch2_btree_node_alloc(as, trans, level); 474 475 btree_set_min(b, POS_MIN); 476 btree_set_max(b, SPOS_MAX); 477 b->data->format = bch2_btree_calc_format(b); 478 479 btree_node_set_format(b, b->data->format); 480 bch2_btree_build_aux_trees(b); 481 482 return b; 483 } 484 485 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans) 486 { 487 struct bch_fs *c = as->c; 488 struct prealloc_nodes *p; 489 490 for (p = as->prealloc_nodes; 491 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes); 492 p++) { 493 while (p->nr) { 494 struct btree *b = p->b[--p->nr]; 495 496 mutex_lock(&c->btree_reserve_cache_lock); 497 498 if (c->btree_reserve_cache_nr < 499 ARRAY_SIZE(c->btree_reserve_cache)) { 500 struct btree_alloc *a = 501 &c->btree_reserve_cache[c->btree_reserve_cache_nr++]; 502 503 a->ob = b->ob; 504 b->ob.nr = 0; 505 bkey_copy(&a->k, &b->key); 506 } else { 507 bch2_open_buckets_put(c, &b->ob); 508 } 509 510 mutex_unlock(&c->btree_reserve_cache_lock); 511 512 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent); 513 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write); 514 __btree_node_free(trans, b); 515 six_unlock_write(&b->c.lock); 516 six_unlock_intent(&b->c.lock); 517 } 518 } 519 } 520 521 static int bch2_btree_reserve_get(struct btree_trans *trans, 522 struct btree_update *as, 523 unsigned nr_nodes[2], 524 unsigned flags, 525 struct closure *cl) 526 { 527 struct btree *b; 528 unsigned interior; 529 int ret = 0; 530 531 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX); 532 533 /* 534 * Protects reaping from the btree node cache and using the btree node 535 * open bucket reserve: 536 */ 537 ret = bch2_btree_cache_cannibalize_lock(trans, cl); 538 if (ret) 539 return ret; 540 541 for (interior = 0; interior < 2; interior++) { 542 struct prealloc_nodes *p = as->prealloc_nodes + interior; 543 544 while (p->nr < nr_nodes[interior]) { 545 b = __bch2_btree_node_alloc(trans, &as->disk_res, cl, 546 interior, flags); 547 if (IS_ERR(b)) { 548 ret = PTR_ERR(b); 549 goto err; 550 } 551 552 p->b[p->nr++] = b; 553 } 554 } 555 err: 556 bch2_btree_cache_cannibalize_unlock(trans); 557 return ret; 558 } 559 560 /* Asynchronous interior node update machinery */ 561 562 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans) 563 { 564 struct bch_fs *c = as->c; 565 566 if (as->took_gc_lock) 567 up_read(&c->gc_lock); 568 as->took_gc_lock = false; 569 570 bch2_journal_pin_drop(&c->journal, &as->journal); 571 bch2_journal_pin_flush(&c->journal, &as->journal); 572 bch2_disk_reservation_put(c, &as->disk_res); 573 bch2_btree_reserve_put(as, trans); 574 575 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total], 576 as->start_time); 577 578 mutex_lock(&c->btree_interior_update_lock); 579 list_del(&as->unwritten_list); 580 list_del(&as->list); 581 582 closure_debug_destroy(&as->cl); 583 mempool_free(as, &c->btree_interior_update_pool); 584 585 /* 586 * Have to do the wakeup with btree_interior_update_lock still held, 587 * since being on btree_interior_update_list is our ref on @c: 588 */ 589 closure_wake_up(&c->btree_interior_update_wait); 590 591 mutex_unlock(&c->btree_interior_update_lock); 592 } 593 594 static void btree_update_add_key(struct btree_update *as, 595 struct keylist *keys, struct btree *b) 596 { 597 struct bkey_i *k = &b->key; 598 599 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s > 600 ARRAY_SIZE(as->_old_keys)); 601 602 bkey_copy(keys->top, k); 603 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1; 604 605 bch2_keylist_push(keys); 606 } 607 608 /* 609 * The transactional part of an interior btree node update, where we journal the 610 * update we did to the interior node and update alloc info: 611 */ 612 static int btree_update_nodes_written_trans(struct btree_trans *trans, 613 struct btree_update *as) 614 { 615 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s); 616 int ret = PTR_ERR_OR_ZERO(e); 617 if (ret) 618 return ret; 619 620 memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64)); 621 622 trans->journal_pin = &as->journal; 623 624 for_each_keylist_key(&as->old_keys, k) { 625 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; 626 627 ret = bch2_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 628 BTREE_TRIGGER_TRANSACTIONAL); 629 if (ret) 630 return ret; 631 } 632 633 for_each_keylist_key(&as->new_keys, k) { 634 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; 635 636 ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k), 637 BTREE_TRIGGER_TRANSACTIONAL); 638 if (ret) 639 return ret; 640 } 641 642 return 0; 643 } 644 645 static void btree_update_nodes_written(struct btree_update *as) 646 { 647 struct bch_fs *c = as->c; 648 struct btree *b; 649 struct btree_trans *trans = bch2_trans_get(c); 650 u64 journal_seq = 0; 651 unsigned i; 652 int ret; 653 654 /* 655 * If we're already in an error state, it might be because a btree node 656 * was never written, and we might be trying to free that same btree 657 * node here, but it won't have been marked as allocated and we'll see 658 * spurious disk usage inconsistencies in the transactional part below 659 * if we don't skip it: 660 */ 661 ret = bch2_journal_error(&c->journal); 662 if (ret) 663 goto err; 664 665 /* 666 * Wait for any in flight writes to finish before we free the old nodes 667 * on disk: 668 */ 669 for (i = 0; i < as->nr_old_nodes; i++) { 670 __le64 seq; 671 672 b = as->old_nodes[i]; 673 674 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read); 675 seq = b->data ? b->data->keys.seq : 0; 676 six_unlock_read(&b->c.lock); 677 678 if (seq == as->old_nodes_seq[i]) 679 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner, 680 TASK_UNINTERRUPTIBLE); 681 } 682 683 /* 684 * We did an update to a parent node where the pointers we added pointed 685 * to child nodes that weren't written yet: now, the child nodes have 686 * been written so we can write out the update to the interior node. 687 */ 688 689 /* 690 * We can't call into journal reclaim here: we'd block on the journal 691 * reclaim lock, but we may need to release the open buckets we have 692 * pinned in order for other btree updates to make forward progress, and 693 * journal reclaim does btree updates when flushing bkey_cached entries, 694 * which may require allocations as well. 695 */ 696 ret = commit_do(trans, &as->disk_res, &journal_seq, 697 BCH_WATERMARK_interior_updates| 698 BCH_TRANS_COMMIT_no_enospc| 699 BCH_TRANS_COMMIT_no_check_rw| 700 BCH_TRANS_COMMIT_journal_reclaim, 701 btree_update_nodes_written_trans(trans, as)); 702 bch2_trans_unlock(trans); 703 704 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c, 705 "%s", bch2_err_str(ret)); 706 err: 707 if (as->b) { 708 709 b = as->b; 710 btree_path_idx_t path_idx = get_unlocked_mut_path(trans, 711 as->btree_id, b->c.level, b->key.k.p); 712 struct btree_path *path = trans->paths + path_idx; 713 /* 714 * @b is the node we did the final insert into: 715 * 716 * On failure to get a journal reservation, we still have to 717 * unblock the write and allow most of the write path to happen 718 * so that shutdown works, but the i->journal_seq mechanism 719 * won't work to prevent the btree write from being visible (we 720 * didn't get a journal sequence number) - instead 721 * __bch2_btree_node_write() doesn't do the actual write if 722 * we're in journal error state: 723 */ 724 725 /* 726 * Ensure transaction is unlocked before using 727 * btree_node_lock_nopath() (the use of which is always suspect, 728 * we need to work on removing this in the future) 729 * 730 * It should be, but get_unlocked_mut_path() -> bch2_path_get() 731 * calls bch2_path_upgrade(), before we call path_make_mut(), so 732 * we may rarely end up with a locked path besides the one we 733 * have here: 734 */ 735 bch2_trans_unlock(trans); 736 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent); 737 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED); 738 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock); 739 path->l[b->c.level].b = b; 740 741 bch2_btree_node_lock_write_nofail(trans, path, &b->c); 742 743 mutex_lock(&c->btree_interior_update_lock); 744 745 list_del(&as->write_blocked_list); 746 if (list_empty(&b->write_blocked)) 747 clear_btree_node_write_blocked(b); 748 749 /* 750 * Node might have been freed, recheck under 751 * btree_interior_update_lock: 752 */ 753 if (as->b == b) { 754 BUG_ON(!b->c.level); 755 BUG_ON(!btree_node_dirty(b)); 756 757 if (!ret) { 758 struct bset *last = btree_bset_last(b); 759 760 last->journal_seq = cpu_to_le64( 761 max(journal_seq, 762 le64_to_cpu(last->journal_seq))); 763 764 bch2_btree_add_journal_pin(c, b, journal_seq); 765 } else { 766 /* 767 * If we didn't get a journal sequence number we 768 * can't write this btree node, because recovery 769 * won't know to ignore this write: 770 */ 771 set_btree_node_never_write(b); 772 } 773 } 774 775 mutex_unlock(&c->btree_interior_update_lock); 776 777 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED); 778 six_unlock_write(&b->c.lock); 779 780 btree_node_write_if_need(c, b, SIX_LOCK_intent); 781 btree_node_unlock(trans, path, b->c.level); 782 bch2_path_put(trans, path_idx, true); 783 } 784 785 bch2_journal_pin_drop(&c->journal, &as->journal); 786 787 mutex_lock(&c->btree_interior_update_lock); 788 for (i = 0; i < as->nr_new_nodes; i++) { 789 b = as->new_nodes[i]; 790 791 BUG_ON(b->will_make_reachable != (unsigned long) as); 792 b->will_make_reachable = 0; 793 clear_btree_node_will_make_reachable(b); 794 } 795 mutex_unlock(&c->btree_interior_update_lock); 796 797 for (i = 0; i < as->nr_new_nodes; i++) { 798 b = as->new_nodes[i]; 799 800 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read); 801 btree_node_write_if_need(c, b, SIX_LOCK_read); 802 six_unlock_read(&b->c.lock); 803 } 804 805 for (i = 0; i < as->nr_open_buckets; i++) 806 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]); 807 808 bch2_btree_update_free(as, trans); 809 bch2_trans_put(trans); 810 } 811 812 static void btree_interior_update_work(struct work_struct *work) 813 { 814 struct bch_fs *c = 815 container_of(work, struct bch_fs, btree_interior_update_work); 816 struct btree_update *as; 817 818 while (1) { 819 mutex_lock(&c->btree_interior_update_lock); 820 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten, 821 struct btree_update, unwritten_list); 822 if (as && !as->nodes_written) 823 as = NULL; 824 mutex_unlock(&c->btree_interior_update_lock); 825 826 if (!as) 827 break; 828 829 btree_update_nodes_written(as); 830 } 831 } 832 833 static CLOSURE_CALLBACK(btree_update_set_nodes_written) 834 { 835 closure_type(as, struct btree_update, cl); 836 struct bch_fs *c = as->c; 837 838 mutex_lock(&c->btree_interior_update_lock); 839 as->nodes_written = true; 840 mutex_unlock(&c->btree_interior_update_lock); 841 842 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work); 843 } 844 845 /* 846 * We're updating @b with pointers to nodes that haven't finished writing yet: 847 * block @b from being written until @as completes 848 */ 849 static void btree_update_updated_node(struct btree_update *as, struct btree *b) 850 { 851 struct bch_fs *c = as->c; 852 853 mutex_lock(&c->btree_interior_update_lock); 854 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten); 855 856 BUG_ON(as->mode != BTREE_UPDATE_none); 857 BUG_ON(!btree_node_dirty(b)); 858 BUG_ON(!b->c.level); 859 860 as->mode = BTREE_UPDATE_node; 861 as->b = b; 862 863 set_btree_node_write_blocked(b); 864 list_add(&as->write_blocked_list, &b->write_blocked); 865 866 mutex_unlock(&c->btree_interior_update_lock); 867 } 868 869 static int bch2_update_reparent_journal_pin_flush(struct journal *j, 870 struct journal_entry_pin *_pin, u64 seq) 871 { 872 return 0; 873 } 874 875 static void btree_update_reparent(struct btree_update *as, 876 struct btree_update *child) 877 { 878 struct bch_fs *c = as->c; 879 880 lockdep_assert_held(&c->btree_interior_update_lock); 881 882 child->b = NULL; 883 child->mode = BTREE_UPDATE_update; 884 885 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, 886 bch2_update_reparent_journal_pin_flush); 887 } 888 889 static void btree_update_updated_root(struct btree_update *as, struct btree *b) 890 { 891 struct bkey_i *insert = &b->key; 892 struct bch_fs *c = as->c; 893 894 BUG_ON(as->mode != BTREE_UPDATE_none); 895 896 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) > 897 ARRAY_SIZE(as->journal_entries)); 898 899 as->journal_u64s += 900 journal_entry_set((void *) &as->journal_entries[as->journal_u64s], 901 BCH_JSET_ENTRY_btree_root, 902 b->c.btree_id, b->c.level, 903 insert, insert->k.u64s); 904 905 mutex_lock(&c->btree_interior_update_lock); 906 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten); 907 908 as->mode = BTREE_UPDATE_root; 909 mutex_unlock(&c->btree_interior_update_lock); 910 } 911 912 /* 913 * bch2_btree_update_add_new_node: 914 * 915 * This causes @as to wait on @b to be written, before it gets to 916 * bch2_btree_update_nodes_written 917 * 918 * Additionally, it sets b->will_make_reachable to prevent any additional writes 919 * to @b from happening besides the first until @b is reachable on disk 920 * 921 * And it adds @b to the list of @as's new nodes, so that we can update sector 922 * counts in bch2_btree_update_nodes_written: 923 */ 924 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b) 925 { 926 struct bch_fs *c = as->c; 927 928 closure_get(&as->cl); 929 930 mutex_lock(&c->btree_interior_update_lock); 931 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes)); 932 BUG_ON(b->will_make_reachable); 933 934 as->new_nodes[as->nr_new_nodes++] = b; 935 b->will_make_reachable = 1UL|(unsigned long) as; 936 set_btree_node_will_make_reachable(b); 937 938 mutex_unlock(&c->btree_interior_update_lock); 939 940 btree_update_add_key(as, &as->new_keys, b); 941 942 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) { 943 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data; 944 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9; 945 946 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written = 947 cpu_to_le16(sectors); 948 } 949 } 950 951 /* 952 * returns true if @b was a new node 953 */ 954 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b) 955 { 956 struct btree_update *as; 957 unsigned long v; 958 unsigned i; 959 960 mutex_lock(&c->btree_interior_update_lock); 961 /* 962 * When b->will_make_reachable != 0, it owns a ref on as->cl that's 963 * dropped when it gets written by bch2_btree_complete_write - the 964 * xchg() is for synchronization with bch2_btree_complete_write: 965 */ 966 v = xchg(&b->will_make_reachable, 0); 967 clear_btree_node_will_make_reachable(b); 968 as = (struct btree_update *) (v & ~1UL); 969 970 if (!as) { 971 mutex_unlock(&c->btree_interior_update_lock); 972 return; 973 } 974 975 for (i = 0; i < as->nr_new_nodes; i++) 976 if (as->new_nodes[i] == b) 977 goto found; 978 979 BUG(); 980 found: 981 array_remove_item(as->new_nodes, as->nr_new_nodes, i); 982 mutex_unlock(&c->btree_interior_update_lock); 983 984 if (v & 1) 985 closure_put(&as->cl); 986 } 987 988 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b) 989 { 990 while (b->ob.nr) 991 as->open_buckets[as->nr_open_buckets++] = 992 b->ob.v[--b->ob.nr]; 993 } 994 995 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j, 996 struct journal_entry_pin *_pin, u64 seq) 997 { 998 return 0; 999 } 1000 1001 /* 1002 * @b is being split/rewritten: it may have pointers to not-yet-written btree 1003 * nodes and thus outstanding btree_updates - redirect @b's 1004 * btree_updates to point to this btree_update: 1005 */ 1006 static void bch2_btree_interior_update_will_free_node(struct btree_update *as, 1007 struct btree *b) 1008 { 1009 struct bch_fs *c = as->c; 1010 struct btree_update *p, *n; 1011 struct btree_write *w; 1012 1013 set_btree_node_dying(b); 1014 1015 if (btree_node_fake(b)) 1016 return; 1017 1018 mutex_lock(&c->btree_interior_update_lock); 1019 1020 /* 1021 * Does this node have any btree_update operations preventing 1022 * it from being written? 1023 * 1024 * If so, redirect them to point to this btree_update: we can 1025 * write out our new nodes, but we won't make them visible until those 1026 * operations complete 1027 */ 1028 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) { 1029 list_del_init(&p->write_blocked_list); 1030 btree_update_reparent(as, p); 1031 1032 /* 1033 * for flush_held_btree_writes() waiting on updates to flush or 1034 * nodes to be writeable: 1035 */ 1036 closure_wake_up(&c->btree_interior_update_wait); 1037 } 1038 1039 clear_btree_node_dirty_acct(c, b); 1040 clear_btree_node_need_write(b); 1041 clear_btree_node_write_blocked(b); 1042 1043 /* 1044 * Does this node have unwritten data that has a pin on the journal? 1045 * 1046 * If so, transfer that pin to the btree_update operation - 1047 * note that if we're freeing multiple nodes, we only need to keep the 1048 * oldest pin of any of the nodes we're freeing. We'll release the pin 1049 * when the new nodes are persistent and reachable on disk: 1050 */ 1051 w = btree_current_write(b); 1052 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, 1053 bch2_btree_update_will_free_node_journal_pin_flush); 1054 bch2_journal_pin_drop(&c->journal, &w->journal); 1055 1056 w = btree_prev_write(b); 1057 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, 1058 bch2_btree_update_will_free_node_journal_pin_flush); 1059 bch2_journal_pin_drop(&c->journal, &w->journal); 1060 1061 mutex_unlock(&c->btree_interior_update_lock); 1062 1063 /* 1064 * Is this a node that isn't reachable on disk yet? 1065 * 1066 * Nodes that aren't reachable yet have writes blocked until they're 1067 * reachable - now that we've cancelled any pending writes and moved 1068 * things waiting on that write to wait on this update, we can drop this 1069 * node from the list of nodes that the other update is making 1070 * reachable, prior to freeing it: 1071 */ 1072 btree_update_drop_new_node(c, b); 1073 1074 btree_update_add_key(as, &as->old_keys, b); 1075 1076 as->old_nodes[as->nr_old_nodes] = b; 1077 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq; 1078 as->nr_old_nodes++; 1079 } 1080 1081 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans) 1082 { 1083 struct bch_fs *c = as->c; 1084 u64 start_time = as->start_time; 1085 1086 BUG_ON(as->mode == BTREE_UPDATE_none); 1087 1088 if (as->took_gc_lock) 1089 up_read(&as->c->gc_lock); 1090 as->took_gc_lock = false; 1091 1092 bch2_btree_reserve_put(as, trans); 1093 1094 continue_at(&as->cl, btree_update_set_nodes_written, 1095 as->c->btree_interior_update_worker); 1096 1097 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground], 1098 start_time); 1099 } 1100 1101 static struct btree_update * 1102 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path, 1103 unsigned level, bool split, unsigned flags) 1104 { 1105 struct bch_fs *c = trans->c; 1106 struct btree_update *as; 1107 u64 start_time = local_clock(); 1108 int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc) 1109 ? BCH_DISK_RESERVATION_NOFAIL : 0; 1110 unsigned nr_nodes[2] = { 0, 0 }; 1111 unsigned update_level = level; 1112 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK; 1113 int ret = 0; 1114 u32 restart_count = trans->restart_count; 1115 1116 BUG_ON(!path->should_be_locked); 1117 1118 if (watermark == BCH_WATERMARK_copygc) 1119 watermark = BCH_WATERMARK_btree_copygc; 1120 if (watermark < BCH_WATERMARK_btree) 1121 watermark = BCH_WATERMARK_btree; 1122 1123 flags &= ~BCH_WATERMARK_MASK; 1124 flags |= watermark; 1125 1126 if (watermark < c->journal.watermark) { 1127 struct journal_res res = { 0 }; 1128 unsigned journal_flags = watermark|JOURNAL_RES_GET_CHECK; 1129 1130 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) && 1131 watermark < BCH_WATERMARK_reclaim) 1132 journal_flags |= JOURNAL_RES_GET_NONBLOCK; 1133 1134 ret = drop_locks_do(trans, 1135 bch2_journal_res_get(&c->journal, &res, 1, journal_flags)); 1136 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) 1137 ret = -BCH_ERR_journal_reclaim_would_deadlock; 1138 if (ret) 1139 return ERR_PTR(ret); 1140 } 1141 1142 while (1) { 1143 nr_nodes[!!update_level] += 1 + split; 1144 update_level++; 1145 1146 ret = bch2_btree_path_upgrade(trans, path, update_level + 1); 1147 if (ret) 1148 return ERR_PTR(ret); 1149 1150 if (!btree_path_node(path, update_level)) { 1151 /* Allocating new root? */ 1152 nr_nodes[1] += split; 1153 update_level = BTREE_MAX_DEPTH; 1154 break; 1155 } 1156 1157 /* 1158 * Always check for space for two keys, even if we won't have to 1159 * split at prior level - it might have been a merge instead: 1160 */ 1161 if (bch2_btree_node_insert_fits(path->l[update_level].b, 1162 BKEY_BTREE_PTR_U64s_MAX * 2)) 1163 break; 1164 1165 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c); 1166 } 1167 1168 if (!down_read_trylock(&c->gc_lock)) { 1169 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0)); 1170 if (ret) { 1171 up_read(&c->gc_lock); 1172 return ERR_PTR(ret); 1173 } 1174 } 1175 1176 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS); 1177 memset(as, 0, sizeof(*as)); 1178 closure_init(&as->cl, NULL); 1179 as->c = c; 1180 as->start_time = start_time; 1181 as->ip_started = _RET_IP_; 1182 as->mode = BTREE_UPDATE_none; 1183 as->watermark = watermark; 1184 as->took_gc_lock = true; 1185 as->btree_id = path->btree_id; 1186 as->update_level = update_level; 1187 INIT_LIST_HEAD(&as->list); 1188 INIT_LIST_HEAD(&as->unwritten_list); 1189 INIT_LIST_HEAD(&as->write_blocked_list); 1190 bch2_keylist_init(&as->old_keys, as->_old_keys); 1191 bch2_keylist_init(&as->new_keys, as->_new_keys); 1192 bch2_keylist_init(&as->parent_keys, as->inline_keys); 1193 1194 mutex_lock(&c->btree_interior_update_lock); 1195 list_add_tail(&as->list, &c->btree_interior_update_list); 1196 mutex_unlock(&c->btree_interior_update_lock); 1197 1198 /* 1199 * We don't want to allocate if we're in an error state, that can cause 1200 * deadlock on emergency shutdown due to open buckets getting stuck in 1201 * the btree_reserve_cache after allocator shutdown has cleared it out. 1202 * This check needs to come after adding us to the btree_interior_update 1203 * list but before calling bch2_btree_reserve_get, to synchronize with 1204 * __bch2_fs_read_only(). 1205 */ 1206 ret = bch2_journal_error(&c->journal); 1207 if (ret) 1208 goto err; 1209 1210 ret = bch2_disk_reservation_get(c, &as->disk_res, 1211 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c), 1212 c->opts.metadata_replicas, 1213 disk_res_flags); 1214 if (ret) 1215 goto err; 1216 1217 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL); 1218 if (bch2_err_matches(ret, ENOSPC) || 1219 bch2_err_matches(ret, ENOMEM)) { 1220 struct closure cl; 1221 1222 /* 1223 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK 1224 * flag 1225 */ 1226 if (bch2_err_matches(ret, ENOSPC) && 1227 (flags & BCH_TRANS_COMMIT_journal_reclaim) && 1228 watermark < BCH_WATERMARK_reclaim) { 1229 ret = -BCH_ERR_journal_reclaim_would_deadlock; 1230 goto err; 1231 } 1232 1233 closure_init_stack(&cl); 1234 1235 do { 1236 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl); 1237 1238 bch2_trans_unlock(trans); 1239 closure_sync(&cl); 1240 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked)); 1241 } 1242 1243 if (ret) { 1244 trace_and_count(c, btree_reserve_get_fail, trans->fn, 1245 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret); 1246 goto err; 1247 } 1248 1249 ret = bch2_trans_relock(trans); 1250 if (ret) 1251 goto err; 1252 1253 bch2_trans_verify_not_restarted(trans, restart_count); 1254 return as; 1255 err: 1256 bch2_btree_update_free(as, trans); 1257 if (!bch2_err_matches(ret, ENOSPC) && 1258 !bch2_err_matches(ret, EROFS) && 1259 ret != -BCH_ERR_journal_reclaim_would_deadlock) 1260 bch_err_fn_ratelimited(c, ret); 1261 return ERR_PTR(ret); 1262 } 1263 1264 /* Btree root updates: */ 1265 1266 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) 1267 { 1268 /* Root nodes cannot be reaped */ 1269 mutex_lock(&c->btree_cache.lock); 1270 list_del_init(&b->list); 1271 mutex_unlock(&c->btree_cache.lock); 1272 1273 mutex_lock(&c->btree_root_lock); 1274 bch2_btree_id_root(c, b->c.btree_id)->b = b; 1275 mutex_unlock(&c->btree_root_lock); 1276 1277 bch2_recalc_btree_reserve(c); 1278 } 1279 1280 static void bch2_btree_set_root(struct btree_update *as, 1281 struct btree_trans *trans, 1282 struct btree_path *path, 1283 struct btree *b) 1284 { 1285 struct bch_fs *c = as->c; 1286 struct btree *old; 1287 1288 trace_and_count(c, btree_node_set_root, trans, b); 1289 1290 old = btree_node_root(c, b); 1291 1292 /* 1293 * Ensure no one is using the old root while we switch to the 1294 * new root: 1295 */ 1296 bch2_btree_node_lock_write_nofail(trans, path, &old->c); 1297 1298 bch2_btree_set_root_inmem(c, b); 1299 1300 btree_update_updated_root(as, b); 1301 1302 /* 1303 * Unlock old root after new root is visible: 1304 * 1305 * The new root isn't persistent, but that's ok: we still have 1306 * an intent lock on the new root, and any updates that would 1307 * depend on the new root would have to update the new root. 1308 */ 1309 bch2_btree_node_unlock_write(trans, path, old); 1310 } 1311 1312 /* Interior node updates: */ 1313 1314 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, 1315 struct btree_trans *trans, 1316 struct btree_path *path, 1317 struct btree *b, 1318 struct btree_node_iter *node_iter, 1319 struct bkey_i *insert) 1320 { 1321 struct bch_fs *c = as->c; 1322 struct bkey_packed *k; 1323 struct printbuf buf = PRINTBUF; 1324 unsigned long old, new, v; 1325 1326 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 && 1327 !btree_ptr_sectors_written(insert)); 1328 1329 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) 1330 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p); 1331 1332 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert), 1333 btree_node_type(b), WRITE, &buf) ?: 1334 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) { 1335 printbuf_reset(&buf); 1336 prt_printf(&buf, "inserting invalid bkey\n "); 1337 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert)); 1338 prt_printf(&buf, "\n "); 1339 bch2_bkey_invalid(c, bkey_i_to_s_c(insert), 1340 btree_node_type(b), WRITE, &buf); 1341 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf); 1342 1343 bch2_fs_inconsistent(c, "%s", buf.buf); 1344 dump_stack(); 1345 } 1346 1347 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) > 1348 ARRAY_SIZE(as->journal_entries)); 1349 1350 as->journal_u64s += 1351 journal_entry_set((void *) &as->journal_entries[as->journal_u64s], 1352 BCH_JSET_ENTRY_btree_keys, 1353 b->c.btree_id, b->c.level, 1354 insert, insert->k.u64s); 1355 1356 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) && 1357 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0) 1358 bch2_btree_node_iter_advance(node_iter, b); 1359 1360 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert); 1361 set_btree_node_dirty_acct(c, b); 1362 1363 v = READ_ONCE(b->flags); 1364 do { 1365 old = new = v; 1366 1367 new &= ~BTREE_WRITE_TYPE_MASK; 1368 new |= BTREE_WRITE_interior; 1369 new |= 1 << BTREE_NODE_need_write; 1370 } while ((v = cmpxchg(&b->flags, old, new)) != old); 1371 1372 printbuf_exit(&buf); 1373 } 1374 1375 static void 1376 __bch2_btree_insert_keys_interior(struct btree_update *as, 1377 struct btree_trans *trans, 1378 struct btree_path *path, 1379 struct btree *b, 1380 struct btree_node_iter node_iter, 1381 struct keylist *keys) 1382 { 1383 struct bkey_i *insert = bch2_keylist_front(keys); 1384 struct bkey_packed *k; 1385 1386 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree); 1387 1388 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) && 1389 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0)) 1390 ; 1391 1392 while (!bch2_keylist_empty(keys)) { 1393 insert = bch2_keylist_front(keys); 1394 1395 if (bpos_gt(insert->k.p, b->key.k.p)) 1396 break; 1397 1398 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert); 1399 bch2_keylist_pop_front(keys); 1400 } 1401 } 1402 1403 /* 1404 * Move keys from n1 (original replacement node, now lower node) to n2 (higher 1405 * node) 1406 */ 1407 static void __btree_split_node(struct btree_update *as, 1408 struct btree_trans *trans, 1409 struct btree *b, 1410 struct btree *n[2]) 1411 { 1412 struct bkey_packed *k; 1413 struct bpos n1_pos = POS_MIN; 1414 struct btree_node_iter iter; 1415 struct bset *bsets[2]; 1416 struct bkey_format_state format[2]; 1417 struct bkey_packed *out[2]; 1418 struct bkey uk; 1419 unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5; 1420 struct { unsigned nr_keys, val_u64s; } nr_keys[2]; 1421 int i; 1422 1423 memset(&nr_keys, 0, sizeof(nr_keys)); 1424 1425 for (i = 0; i < 2; i++) { 1426 BUG_ON(n[i]->nsets != 1); 1427 1428 bsets[i] = btree_bset_first(n[i]); 1429 out[i] = bsets[i]->start; 1430 1431 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1); 1432 bch2_bkey_format_init(&format[i]); 1433 } 1434 1435 u64s = 0; 1436 for_each_btree_node_key(b, k, &iter) { 1437 if (bkey_deleted(k)) 1438 continue; 1439 1440 uk = bkey_unpack_key(b, k); 1441 1442 if (b->c.level && 1443 u64s < n1_u64s && 1444 u64s + k->u64s >= n1_u64s && 1445 bch2_key_deleted_in_journal(trans, b->c.btree_id, b->c.level, uk.p)) 1446 n1_u64s += k->u64s; 1447 1448 i = u64s >= n1_u64s; 1449 u64s += k->u64s; 1450 if (!i) 1451 n1_pos = uk.p; 1452 bch2_bkey_format_add_key(&format[i], &uk); 1453 1454 nr_keys[i].nr_keys++; 1455 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k); 1456 } 1457 1458 btree_set_min(n[0], b->data->min_key); 1459 btree_set_max(n[0], n1_pos); 1460 btree_set_min(n[1], bpos_successor(n1_pos)); 1461 btree_set_max(n[1], b->data->max_key); 1462 1463 for (i = 0; i < 2; i++) { 1464 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key); 1465 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key); 1466 1467 n[i]->data->format = bch2_bkey_format_done(&format[i]); 1468 1469 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s + 1470 nr_keys[i].val_u64s; 1471 if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b)) 1472 n[i]->data->format = b->format; 1473 1474 btree_node_set_format(n[i], n[i]->data->format); 1475 } 1476 1477 u64s = 0; 1478 for_each_btree_node_key(b, k, &iter) { 1479 if (bkey_deleted(k)) 1480 continue; 1481 1482 i = u64s >= n1_u64s; 1483 u64s += k->u64s; 1484 1485 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k) 1486 ? &b->format: &bch2_bkey_format_current, k)) 1487 out[i]->format = KEY_FORMAT_LOCAL_BTREE; 1488 else 1489 bch2_bkey_unpack(b, (void *) out[i], k); 1490 1491 out[i]->needs_whiteout = false; 1492 1493 btree_keys_account_key_add(&n[i]->nr, 0, out[i]); 1494 out[i] = bkey_p_next(out[i]); 1495 } 1496 1497 for (i = 0; i < 2; i++) { 1498 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data); 1499 1500 BUG_ON(!bsets[i]->u64s); 1501 1502 set_btree_bset_end(n[i], n[i]->set); 1503 1504 btree_node_reset_sib_u64s(n[i]); 1505 1506 bch2_verify_btree_nr_keys(n[i]); 1507 1508 BUG_ON(bch2_btree_node_check_topology(trans, n[i])); 1509 } 1510 } 1511 1512 /* 1513 * For updates to interior nodes, we've got to do the insert before we split 1514 * because the stuff we're inserting has to be inserted atomically. Post split, 1515 * the keys might have to go in different nodes and the split would no longer be 1516 * atomic. 1517 * 1518 * Worse, if the insert is from btree node coalescing, if we do the insert after 1519 * we do the split (and pick the pivot) - the pivot we pick might be between 1520 * nodes that were coalesced, and thus in the middle of a child node post 1521 * coalescing: 1522 */ 1523 static void btree_split_insert_keys(struct btree_update *as, 1524 struct btree_trans *trans, 1525 btree_path_idx_t path_idx, 1526 struct btree *b, 1527 struct keylist *keys) 1528 { 1529 struct btree_path *path = trans->paths + path_idx; 1530 1531 if (!bch2_keylist_empty(keys) && 1532 bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) { 1533 struct btree_node_iter node_iter; 1534 1535 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p); 1536 1537 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys); 1538 1539 BUG_ON(bch2_btree_node_check_topology(trans, b)); 1540 } 1541 } 1542 1543 static int btree_split(struct btree_update *as, struct btree_trans *trans, 1544 btree_path_idx_t path, struct btree *b, 1545 struct keylist *keys) 1546 { 1547 struct bch_fs *c = as->c; 1548 struct btree *parent = btree_node_parent(trans->paths + path, b); 1549 struct btree *n1, *n2 = NULL, *n3 = NULL; 1550 btree_path_idx_t path1 = 0, path2 = 0; 1551 u64 start_time = local_clock(); 1552 int ret = 0; 1553 1554 bch2_verify_btree_nr_keys(b); 1555 BUG_ON(!parent && (b != btree_node_root(c, b))); 1556 BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1)); 1557 1558 ret = bch2_btree_node_check_topology(trans, b); 1559 if (ret) 1560 return ret; 1561 1562 bch2_btree_interior_update_will_free_node(as, b); 1563 1564 if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) { 1565 struct btree *n[2]; 1566 1567 trace_and_count(c, btree_node_split, trans, b); 1568 1569 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level); 1570 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level); 1571 1572 __btree_split_node(as, trans, b, n); 1573 1574 if (keys) { 1575 btree_split_insert_keys(as, trans, path, n1, keys); 1576 btree_split_insert_keys(as, trans, path, n2, keys); 1577 BUG_ON(!bch2_keylist_empty(keys)); 1578 } 1579 1580 bch2_btree_build_aux_trees(n2); 1581 bch2_btree_build_aux_trees(n1); 1582 1583 bch2_btree_update_add_new_node(as, n1); 1584 bch2_btree_update_add_new_node(as, n2); 1585 six_unlock_write(&n2->c.lock); 1586 six_unlock_write(&n1->c.lock); 1587 1588 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p); 1589 six_lock_increment(&n1->c.lock, SIX_LOCK_intent); 1590 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED); 1591 bch2_btree_path_level_init(trans, trans->paths + path1, n1); 1592 1593 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p); 1594 six_lock_increment(&n2->c.lock, SIX_LOCK_intent); 1595 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED); 1596 bch2_btree_path_level_init(trans, trans->paths + path2, n2); 1597 1598 /* 1599 * Note that on recursive parent_keys == keys, so we 1600 * can't start adding new keys to parent_keys before emptying it 1601 * out (which we did with btree_split_insert_keys() above) 1602 */ 1603 bch2_keylist_add(&as->parent_keys, &n1->key); 1604 bch2_keylist_add(&as->parent_keys, &n2->key); 1605 1606 if (!parent) { 1607 /* Depth increases, make a new root */ 1608 n3 = __btree_root_alloc(as, trans, b->c.level + 1); 1609 1610 bch2_btree_update_add_new_node(as, n3); 1611 six_unlock_write(&n3->c.lock); 1612 1613 trans->paths[path2].locks_want++; 1614 BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level)); 1615 six_lock_increment(&n3->c.lock, SIX_LOCK_intent); 1616 mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED); 1617 bch2_btree_path_level_init(trans, trans->paths + path2, n3); 1618 1619 n3->sib_u64s[0] = U16_MAX; 1620 n3->sib_u64s[1] = U16_MAX; 1621 1622 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys); 1623 } 1624 } else { 1625 trace_and_count(c, btree_node_compact, trans, b); 1626 1627 n1 = bch2_btree_node_alloc_replacement(as, trans, b); 1628 1629 if (keys) { 1630 btree_split_insert_keys(as, trans, path, n1, keys); 1631 BUG_ON(!bch2_keylist_empty(keys)); 1632 } 1633 1634 bch2_btree_build_aux_trees(n1); 1635 bch2_btree_update_add_new_node(as, n1); 1636 six_unlock_write(&n1->c.lock); 1637 1638 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p); 1639 six_lock_increment(&n1->c.lock, SIX_LOCK_intent); 1640 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED); 1641 bch2_btree_path_level_init(trans, trans->paths + path1, n1); 1642 1643 if (parent) 1644 bch2_keylist_add(&as->parent_keys, &n1->key); 1645 } 1646 1647 /* New nodes all written, now make them visible: */ 1648 1649 if (parent) { 1650 /* Split a non root node */ 1651 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys); 1652 if (ret) 1653 goto err; 1654 } else if (n3) { 1655 bch2_btree_set_root(as, trans, trans->paths + path, n3); 1656 } else { 1657 /* Root filled up but didn't need to be split */ 1658 bch2_btree_set_root(as, trans, trans->paths + path, n1); 1659 } 1660 1661 if (n3) { 1662 bch2_btree_update_get_open_buckets(as, n3); 1663 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0); 1664 } 1665 if (n2) { 1666 bch2_btree_update_get_open_buckets(as, n2); 1667 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0); 1668 } 1669 bch2_btree_update_get_open_buckets(as, n1); 1670 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0); 1671 1672 /* 1673 * The old node must be freed (in memory) _before_ unlocking the new 1674 * nodes - else another thread could re-acquire a read lock on the old 1675 * node after another thread has locked and updated the new node, thus 1676 * seeing stale data: 1677 */ 1678 bch2_btree_node_free_inmem(trans, trans->paths + path, b); 1679 1680 if (n3) 1681 bch2_trans_node_add(trans, trans->paths + path, n3); 1682 if (n2) 1683 bch2_trans_node_add(trans, trans->paths + path2, n2); 1684 bch2_trans_node_add(trans, trans->paths + path1, n1); 1685 1686 if (n3) 1687 six_unlock_intent(&n3->c.lock); 1688 if (n2) 1689 six_unlock_intent(&n2->c.lock); 1690 six_unlock_intent(&n1->c.lock); 1691 out: 1692 if (path2) { 1693 __bch2_btree_path_unlock(trans, trans->paths + path2); 1694 bch2_path_put(trans, path2, true); 1695 } 1696 if (path1) { 1697 __bch2_btree_path_unlock(trans, trans->paths + path1); 1698 bch2_path_put(trans, path1, true); 1699 } 1700 1701 bch2_trans_verify_locks(trans); 1702 1703 bch2_time_stats_update(&c->times[n2 1704 ? BCH_TIME_btree_node_split 1705 : BCH_TIME_btree_node_compact], 1706 start_time); 1707 return ret; 1708 err: 1709 if (n3) 1710 bch2_btree_node_free_never_used(as, trans, n3); 1711 if (n2) 1712 bch2_btree_node_free_never_used(as, trans, n2); 1713 bch2_btree_node_free_never_used(as, trans, n1); 1714 goto out; 1715 } 1716 1717 static void 1718 bch2_btree_insert_keys_interior(struct btree_update *as, 1719 struct btree_trans *trans, 1720 struct btree_path *path, 1721 struct btree *b, 1722 struct keylist *keys) 1723 { 1724 struct btree_path *linked; 1725 unsigned i; 1726 1727 __bch2_btree_insert_keys_interior(as, trans, path, b, 1728 path->l[b->c.level].iter, keys); 1729 1730 btree_update_updated_node(as, b); 1731 1732 trans_for_each_path_with_node(trans, b, linked, i) 1733 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b); 1734 1735 bch2_trans_verify_paths(trans); 1736 } 1737 1738 /** 1739 * bch2_btree_insert_node - insert bkeys into a given btree node 1740 * 1741 * @as: btree_update object 1742 * @trans: btree_trans object 1743 * @path_idx: path that points to current node 1744 * @b: node to insert keys into 1745 * @keys: list of keys to insert 1746 * 1747 * Returns: 0 on success, typically transaction restart error on failure 1748 * 1749 * Inserts as many keys as it can into a given btree node, splitting it if full. 1750 * If a split occurred, this function will return early. This can only happen 1751 * for leaf nodes -- inserts into interior nodes have to be atomic. 1752 */ 1753 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans, 1754 btree_path_idx_t path_idx, struct btree *b, 1755 struct keylist *keys) 1756 { 1757 struct bch_fs *c = as->c; 1758 struct btree_path *path = trans->paths + path_idx; 1759 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s); 1760 int old_live_u64s = b->nr.live_u64s; 1761 int live_u64s_added, u64s_added; 1762 int ret; 1763 1764 lockdep_assert_held(&c->gc_lock); 1765 BUG_ON(!btree_node_intent_locked(path, b->c.level)); 1766 BUG_ON(!b->c.level); 1767 BUG_ON(!as || as->b); 1768 bch2_verify_keylist_sorted(keys); 1769 1770 ret = bch2_btree_node_lock_write(trans, path, &b->c); 1771 if (ret) 1772 return ret; 1773 1774 bch2_btree_node_prep_for_write(trans, path, b); 1775 1776 if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) { 1777 bch2_btree_node_unlock_write(trans, path, b); 1778 goto split; 1779 } 1780 1781 ret = bch2_btree_node_check_topology(trans, b); 1782 if (ret) { 1783 bch2_btree_node_unlock_write(trans, path, b); 1784 return ret; 1785 } 1786 1787 bch2_btree_insert_keys_interior(as, trans, path, b, keys); 1788 1789 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s; 1790 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s; 1791 1792 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0) 1793 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added); 1794 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0) 1795 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added); 1796 1797 if (u64s_added > live_u64s_added && 1798 bch2_maybe_compact_whiteouts(c, b)) 1799 bch2_trans_node_reinit_iter(trans, b); 1800 1801 bch2_btree_node_unlock_write(trans, path, b); 1802 1803 BUG_ON(bch2_btree_node_check_topology(trans, b)); 1804 return 0; 1805 split: 1806 /* 1807 * We could attempt to avoid the transaction restart, by calling 1808 * bch2_btree_path_upgrade() and allocating more nodes: 1809 */ 1810 if (b->c.level >= as->update_level) { 1811 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b); 1812 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race); 1813 } 1814 1815 return btree_split(as, trans, path_idx, b, keys); 1816 } 1817 1818 int bch2_btree_split_leaf(struct btree_trans *trans, 1819 btree_path_idx_t path, 1820 unsigned flags) 1821 { 1822 /* btree_split & merge may both cause paths array to be reallocated */ 1823 struct btree *b = path_l(trans->paths + path)->b; 1824 struct btree_update *as; 1825 unsigned l; 1826 int ret = 0; 1827 1828 as = bch2_btree_update_start(trans, trans->paths + path, 1829 trans->paths[path].level, 1830 true, flags); 1831 if (IS_ERR(as)) 1832 return PTR_ERR(as); 1833 1834 ret = btree_split(as, trans, path, b, NULL); 1835 if (ret) { 1836 bch2_btree_update_free(as, trans); 1837 return ret; 1838 } 1839 1840 bch2_btree_update_done(as, trans); 1841 1842 for (l = trans->paths[path].level + 1; 1843 btree_node_intent_locked(&trans->paths[path], l) && !ret; 1844 l++) 1845 ret = bch2_foreground_maybe_merge(trans, path, l, flags); 1846 1847 return ret; 1848 } 1849 1850 static void __btree_increase_depth(struct btree_update *as, struct btree_trans *trans, 1851 btree_path_idx_t path_idx) 1852 { 1853 struct bch_fs *c = as->c; 1854 struct btree_path *path = trans->paths + path_idx; 1855 struct btree *n, *b = bch2_btree_id_root(c, path->btree_id)->b; 1856 1857 BUG_ON(!btree_node_locked(path, b->c.level)); 1858 1859 n = __btree_root_alloc(as, trans, b->c.level + 1); 1860 1861 bch2_btree_update_add_new_node(as, n); 1862 six_unlock_write(&n->c.lock); 1863 1864 path->locks_want++; 1865 BUG_ON(btree_node_locked(path, n->c.level)); 1866 six_lock_increment(&n->c.lock, SIX_LOCK_intent); 1867 mark_btree_node_locked(trans, path, n->c.level, BTREE_NODE_INTENT_LOCKED); 1868 bch2_btree_path_level_init(trans, path, n); 1869 1870 n->sib_u64s[0] = U16_MAX; 1871 n->sib_u64s[1] = U16_MAX; 1872 1873 bch2_keylist_add(&as->parent_keys, &b->key); 1874 btree_split_insert_keys(as, trans, path_idx, n, &as->parent_keys); 1875 1876 bch2_btree_set_root(as, trans, path, n); 1877 bch2_btree_update_get_open_buckets(as, n); 1878 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); 1879 bch2_trans_node_add(trans, path, n); 1880 six_unlock_intent(&n->c.lock); 1881 1882 mutex_lock(&c->btree_cache.lock); 1883 list_add_tail(&b->list, &c->btree_cache.live); 1884 mutex_unlock(&c->btree_cache.lock); 1885 1886 bch2_trans_verify_locks(trans); 1887 } 1888 1889 int bch2_btree_increase_depth(struct btree_trans *trans, btree_path_idx_t path, unsigned flags) 1890 { 1891 struct bch_fs *c = trans->c; 1892 struct btree *b = bch2_btree_id_root(c, trans->paths[path].btree_id)->b; 1893 1894 if (btree_node_fake(b)) 1895 return bch2_btree_split_leaf(trans, path, flags); 1896 1897 struct btree_update *as = 1898 bch2_btree_update_start(trans, trans->paths + path, b->c.level, true, flags); 1899 if (IS_ERR(as)) 1900 return PTR_ERR(as); 1901 1902 __btree_increase_depth(as, trans, path); 1903 bch2_btree_update_done(as, trans); 1904 return 0; 1905 } 1906 1907 int __bch2_foreground_maybe_merge(struct btree_trans *trans, 1908 btree_path_idx_t path, 1909 unsigned level, 1910 unsigned flags, 1911 enum btree_node_sibling sib) 1912 { 1913 struct bch_fs *c = trans->c; 1914 struct btree_update *as; 1915 struct bkey_format_state new_s; 1916 struct bkey_format new_f; 1917 struct bkey_i delete; 1918 struct btree *b, *m, *n, *prev, *next, *parent; 1919 struct bpos sib_pos; 1920 size_t sib_u64s; 1921 enum btree_id btree = trans->paths[path].btree_id; 1922 btree_path_idx_t sib_path = 0, new_path = 0; 1923 u64 start_time = local_clock(); 1924 int ret = 0; 1925 1926 BUG_ON(!trans->paths[path].should_be_locked); 1927 BUG_ON(!btree_node_locked(&trans->paths[path], level)); 1928 1929 b = trans->paths[path].l[level].b; 1930 1931 if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) || 1932 (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) { 1933 b->sib_u64s[sib] = U16_MAX; 1934 return 0; 1935 } 1936 1937 sib_pos = sib == btree_prev_sib 1938 ? bpos_predecessor(b->data->min_key) 1939 : bpos_successor(b->data->max_key); 1940 1941 sib_path = bch2_path_get(trans, btree, sib_pos, 1942 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_); 1943 ret = bch2_btree_path_traverse(trans, sib_path, false); 1944 if (ret) 1945 goto err; 1946 1947 btree_path_set_should_be_locked(trans->paths + sib_path); 1948 1949 m = trans->paths[sib_path].l[level].b; 1950 1951 if (btree_node_parent(trans->paths + path, b) != 1952 btree_node_parent(trans->paths + sib_path, m)) { 1953 b->sib_u64s[sib] = U16_MAX; 1954 goto out; 1955 } 1956 1957 if (sib == btree_prev_sib) { 1958 prev = m; 1959 next = b; 1960 } else { 1961 prev = b; 1962 next = m; 1963 } 1964 1965 if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) { 1966 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; 1967 1968 bch2_bpos_to_text(&buf1, prev->data->max_key); 1969 bch2_bpos_to_text(&buf2, next->data->min_key); 1970 bch_err(c, 1971 "%s(): btree topology error:\n" 1972 " prev ends at %s\n" 1973 " next starts at %s", 1974 __func__, buf1.buf, buf2.buf); 1975 printbuf_exit(&buf1); 1976 printbuf_exit(&buf2); 1977 ret = bch2_topology_error(c); 1978 goto err; 1979 } 1980 1981 bch2_bkey_format_init(&new_s); 1982 bch2_bkey_format_add_pos(&new_s, prev->data->min_key); 1983 __bch2_btree_calc_format(&new_s, prev); 1984 __bch2_btree_calc_format(&new_s, next); 1985 bch2_bkey_format_add_pos(&new_s, next->data->max_key); 1986 new_f = bch2_bkey_format_done(&new_s); 1987 1988 sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) + 1989 btree_node_u64s_with_format(m->nr, &m->format, &new_f); 1990 1991 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) { 1992 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c); 1993 sib_u64s /= 2; 1994 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c); 1995 } 1996 1997 sib_u64s = min(sib_u64s, btree_max_u64s(c)); 1998 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1); 1999 b->sib_u64s[sib] = sib_u64s; 2000 2001 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold) 2002 goto out; 2003 2004 parent = btree_node_parent(trans->paths + path, b); 2005 as = bch2_btree_update_start(trans, trans->paths + path, level, false, 2006 BCH_TRANS_COMMIT_no_enospc|flags); 2007 ret = PTR_ERR_OR_ZERO(as); 2008 if (ret) 2009 goto err; 2010 2011 trace_and_count(c, btree_node_merge, trans, b); 2012 2013 bch2_btree_interior_update_will_free_node(as, b); 2014 bch2_btree_interior_update_will_free_node(as, m); 2015 2016 n = bch2_btree_node_alloc(as, trans, b->c.level); 2017 2018 SET_BTREE_NODE_SEQ(n->data, 2019 max(BTREE_NODE_SEQ(b->data), 2020 BTREE_NODE_SEQ(m->data)) + 1); 2021 2022 btree_set_min(n, prev->data->min_key); 2023 btree_set_max(n, next->data->max_key); 2024 2025 n->data->format = new_f; 2026 btree_node_set_format(n, new_f); 2027 2028 bch2_btree_sort_into(c, n, prev); 2029 bch2_btree_sort_into(c, n, next); 2030 2031 bch2_btree_build_aux_trees(n); 2032 bch2_btree_update_add_new_node(as, n); 2033 six_unlock_write(&n->c.lock); 2034 2035 new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p); 2036 six_lock_increment(&n->c.lock, SIX_LOCK_intent); 2037 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED); 2038 bch2_btree_path_level_init(trans, trans->paths + new_path, n); 2039 2040 bkey_init(&delete.k); 2041 delete.k.p = prev->key.k.p; 2042 bch2_keylist_add(&as->parent_keys, &delete); 2043 bch2_keylist_add(&as->parent_keys, &n->key); 2044 2045 bch2_trans_verify_paths(trans); 2046 2047 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys); 2048 if (ret) 2049 goto err_free_update; 2050 2051 bch2_trans_verify_paths(trans); 2052 2053 bch2_btree_update_get_open_buckets(as, n); 2054 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); 2055 2056 bch2_btree_node_free_inmem(trans, trans->paths + path, b); 2057 bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m); 2058 2059 bch2_trans_node_add(trans, trans->paths + path, n); 2060 2061 bch2_trans_verify_paths(trans); 2062 2063 six_unlock_intent(&n->c.lock); 2064 2065 bch2_btree_update_done(as, trans); 2066 2067 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time); 2068 out: 2069 err: 2070 if (new_path) 2071 bch2_path_put(trans, new_path, true); 2072 bch2_path_put(trans, sib_path, true); 2073 bch2_trans_verify_locks(trans); 2074 return ret; 2075 err_free_update: 2076 bch2_btree_node_free_never_used(as, trans, n); 2077 bch2_btree_update_free(as, trans); 2078 goto out; 2079 } 2080 2081 int bch2_btree_node_rewrite(struct btree_trans *trans, 2082 struct btree_iter *iter, 2083 struct btree *b, 2084 unsigned flags) 2085 { 2086 struct bch_fs *c = trans->c; 2087 struct btree *n, *parent; 2088 struct btree_update *as; 2089 btree_path_idx_t new_path = 0; 2090 int ret; 2091 2092 flags |= BCH_TRANS_COMMIT_no_enospc; 2093 2094 struct btree_path *path = btree_iter_path(trans, iter); 2095 parent = btree_node_parent(path, b); 2096 as = bch2_btree_update_start(trans, path, b->c.level, false, flags); 2097 ret = PTR_ERR_OR_ZERO(as); 2098 if (ret) 2099 goto out; 2100 2101 bch2_btree_interior_update_will_free_node(as, b); 2102 2103 n = bch2_btree_node_alloc_replacement(as, trans, b); 2104 2105 bch2_btree_build_aux_trees(n); 2106 bch2_btree_update_add_new_node(as, n); 2107 six_unlock_write(&n->c.lock); 2108 2109 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p); 2110 six_lock_increment(&n->c.lock, SIX_LOCK_intent); 2111 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED); 2112 bch2_btree_path_level_init(trans, trans->paths + new_path, n); 2113 2114 trace_and_count(c, btree_node_rewrite, trans, b); 2115 2116 if (parent) { 2117 bch2_keylist_add(&as->parent_keys, &n->key); 2118 ret = bch2_btree_insert_node(as, trans, iter->path, parent, &as->parent_keys); 2119 if (ret) 2120 goto err; 2121 } else { 2122 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n); 2123 } 2124 2125 bch2_btree_update_get_open_buckets(as, n); 2126 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); 2127 2128 bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b); 2129 2130 bch2_trans_node_add(trans, trans->paths + iter->path, n); 2131 six_unlock_intent(&n->c.lock); 2132 2133 bch2_btree_update_done(as, trans); 2134 out: 2135 if (new_path) 2136 bch2_path_put(trans, new_path, true); 2137 bch2_trans_downgrade(trans); 2138 return ret; 2139 err: 2140 bch2_btree_node_free_never_used(as, trans, n); 2141 bch2_btree_update_free(as, trans); 2142 goto out; 2143 } 2144 2145 struct async_btree_rewrite { 2146 struct bch_fs *c; 2147 struct work_struct work; 2148 struct list_head list; 2149 enum btree_id btree_id; 2150 unsigned level; 2151 struct bpos pos; 2152 __le64 seq; 2153 }; 2154 2155 static int async_btree_node_rewrite_trans(struct btree_trans *trans, 2156 struct async_btree_rewrite *a) 2157 { 2158 struct bch_fs *c = trans->c; 2159 struct btree_iter iter; 2160 struct btree *b; 2161 int ret; 2162 2163 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos, 2164 BTREE_MAX_DEPTH, a->level, 0); 2165 b = bch2_btree_iter_peek_node(&iter); 2166 ret = PTR_ERR_OR_ZERO(b); 2167 if (ret) 2168 goto out; 2169 2170 if (!b || b->data->keys.seq != a->seq) { 2171 struct printbuf buf = PRINTBUF; 2172 2173 if (b) 2174 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key)); 2175 else 2176 prt_str(&buf, "(null"); 2177 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s", 2178 __func__, a->seq, buf.buf); 2179 printbuf_exit(&buf); 2180 goto out; 2181 } 2182 2183 ret = bch2_btree_node_rewrite(trans, &iter, b, 0); 2184 out: 2185 bch2_trans_iter_exit(trans, &iter); 2186 2187 return ret; 2188 } 2189 2190 static void async_btree_node_rewrite_work(struct work_struct *work) 2191 { 2192 struct async_btree_rewrite *a = 2193 container_of(work, struct async_btree_rewrite, work); 2194 struct bch_fs *c = a->c; 2195 int ret; 2196 2197 ret = bch2_trans_do(c, NULL, NULL, 0, 2198 async_btree_node_rewrite_trans(trans, a)); 2199 bch_err_fn_ratelimited(c, ret); 2200 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite); 2201 kfree(a); 2202 } 2203 2204 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b) 2205 { 2206 struct async_btree_rewrite *a; 2207 int ret; 2208 2209 a = kmalloc(sizeof(*a), GFP_NOFS); 2210 if (!a) { 2211 bch_err(c, "%s: error allocating memory", __func__); 2212 return; 2213 } 2214 2215 a->c = c; 2216 a->btree_id = b->c.btree_id; 2217 a->level = b->c.level; 2218 a->pos = b->key.k.p; 2219 a->seq = b->data->keys.seq; 2220 INIT_WORK(&a->work, async_btree_node_rewrite_work); 2221 2222 if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) { 2223 mutex_lock(&c->pending_node_rewrites_lock); 2224 list_add(&a->list, &c->pending_node_rewrites); 2225 mutex_unlock(&c->pending_node_rewrites_lock); 2226 return; 2227 } 2228 2229 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) { 2230 if (test_bit(BCH_FS_started, &c->flags)) { 2231 bch_err(c, "%s: error getting c->writes ref", __func__); 2232 kfree(a); 2233 return; 2234 } 2235 2236 ret = bch2_fs_read_write_early(c); 2237 bch_err_msg(c, ret, "going read-write"); 2238 if (ret) { 2239 kfree(a); 2240 return; 2241 } 2242 2243 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite); 2244 } 2245 2246 queue_work(c->btree_node_rewrite_worker, &a->work); 2247 } 2248 2249 void bch2_do_pending_node_rewrites(struct bch_fs *c) 2250 { 2251 struct async_btree_rewrite *a, *n; 2252 2253 mutex_lock(&c->pending_node_rewrites_lock); 2254 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) { 2255 list_del(&a->list); 2256 2257 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite); 2258 queue_work(c->btree_node_rewrite_worker, &a->work); 2259 } 2260 mutex_unlock(&c->pending_node_rewrites_lock); 2261 } 2262 2263 void bch2_free_pending_node_rewrites(struct bch_fs *c) 2264 { 2265 struct async_btree_rewrite *a, *n; 2266 2267 mutex_lock(&c->pending_node_rewrites_lock); 2268 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) { 2269 list_del(&a->list); 2270 2271 kfree(a); 2272 } 2273 mutex_unlock(&c->pending_node_rewrites_lock); 2274 } 2275 2276 static int __bch2_btree_node_update_key(struct btree_trans *trans, 2277 struct btree_iter *iter, 2278 struct btree *b, struct btree *new_hash, 2279 struct bkey_i *new_key, 2280 unsigned commit_flags, 2281 bool skip_triggers) 2282 { 2283 struct bch_fs *c = trans->c; 2284 struct btree_iter iter2 = { NULL }; 2285 struct btree *parent; 2286 int ret; 2287 2288 if (!skip_triggers) { 2289 ret = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1, 2290 bkey_i_to_s_c(&b->key), 2291 BTREE_TRIGGER_TRANSACTIONAL) ?: 2292 bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1, 2293 bkey_i_to_s(new_key), 2294 BTREE_TRIGGER_TRANSACTIONAL); 2295 if (ret) 2296 return ret; 2297 } 2298 2299 if (new_hash) { 2300 bkey_copy(&new_hash->key, new_key); 2301 ret = bch2_btree_node_hash_insert(&c->btree_cache, 2302 new_hash, b->c.level, b->c.btree_id); 2303 BUG_ON(ret); 2304 } 2305 2306 parent = btree_node_parent(btree_iter_path(trans, iter), b); 2307 if (parent) { 2308 bch2_trans_copy_iter(&iter2, iter); 2309 2310 iter2.path = bch2_btree_path_make_mut(trans, iter2.path, 2311 iter2.flags & BTREE_ITER_INTENT, 2312 _THIS_IP_); 2313 2314 struct btree_path *path2 = btree_iter_path(trans, &iter2); 2315 BUG_ON(path2->level != b->c.level); 2316 BUG_ON(!bpos_eq(path2->pos, new_key->k.p)); 2317 2318 btree_path_set_level_up(trans, path2); 2319 2320 trans->paths_sorted = false; 2321 2322 ret = bch2_btree_iter_traverse(&iter2) ?: 2323 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN); 2324 if (ret) 2325 goto err; 2326 } else { 2327 BUG_ON(btree_node_root(c, b) != b); 2328 2329 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, 2330 jset_u64s(new_key->k.u64s)); 2331 ret = PTR_ERR_OR_ZERO(e); 2332 if (ret) 2333 return ret; 2334 2335 journal_entry_set(e, 2336 BCH_JSET_ENTRY_btree_root, 2337 b->c.btree_id, b->c.level, 2338 new_key, new_key->k.u64s); 2339 } 2340 2341 ret = bch2_trans_commit(trans, NULL, NULL, commit_flags); 2342 if (ret) 2343 goto err; 2344 2345 bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c); 2346 2347 if (new_hash) { 2348 mutex_lock(&c->btree_cache.lock); 2349 bch2_btree_node_hash_remove(&c->btree_cache, new_hash); 2350 bch2_btree_node_hash_remove(&c->btree_cache, b); 2351 2352 bkey_copy(&b->key, new_key); 2353 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b); 2354 BUG_ON(ret); 2355 mutex_unlock(&c->btree_cache.lock); 2356 } else { 2357 bkey_copy(&b->key, new_key); 2358 } 2359 2360 bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b); 2361 out: 2362 bch2_trans_iter_exit(trans, &iter2); 2363 return ret; 2364 err: 2365 if (new_hash) { 2366 mutex_lock(&c->btree_cache.lock); 2367 bch2_btree_node_hash_remove(&c->btree_cache, b); 2368 mutex_unlock(&c->btree_cache.lock); 2369 } 2370 goto out; 2371 } 2372 2373 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter, 2374 struct btree *b, struct bkey_i *new_key, 2375 unsigned commit_flags, bool skip_triggers) 2376 { 2377 struct bch_fs *c = trans->c; 2378 struct btree *new_hash = NULL; 2379 struct btree_path *path = btree_iter_path(trans, iter); 2380 struct closure cl; 2381 int ret = 0; 2382 2383 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1); 2384 if (ret) 2385 return ret; 2386 2387 closure_init_stack(&cl); 2388 2389 /* 2390 * check btree_ptr_hash_val() after @b is locked by 2391 * btree_iter_traverse(): 2392 */ 2393 if (btree_ptr_hash_val(new_key) != b->hash_val) { 2394 ret = bch2_btree_cache_cannibalize_lock(trans, &cl); 2395 if (ret) { 2396 ret = drop_locks_do(trans, (closure_sync(&cl), 0)); 2397 if (ret) 2398 return ret; 2399 } 2400 2401 new_hash = bch2_btree_node_mem_alloc(trans, false); 2402 } 2403 2404 path->intent_ref++; 2405 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key, 2406 commit_flags, skip_triggers); 2407 --path->intent_ref; 2408 2409 if (new_hash) { 2410 mutex_lock(&c->btree_cache.lock); 2411 list_move(&new_hash->list, &c->btree_cache.freeable); 2412 mutex_unlock(&c->btree_cache.lock); 2413 2414 six_unlock_write(&new_hash->c.lock); 2415 six_unlock_intent(&new_hash->c.lock); 2416 } 2417 closure_sync(&cl); 2418 bch2_btree_cache_cannibalize_unlock(trans); 2419 return ret; 2420 } 2421 2422 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans, 2423 struct btree *b, struct bkey_i *new_key, 2424 unsigned commit_flags, bool skip_triggers) 2425 { 2426 struct btree_iter iter; 2427 int ret; 2428 2429 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p, 2430 BTREE_MAX_DEPTH, b->c.level, 2431 BTREE_ITER_INTENT); 2432 ret = bch2_btree_iter_traverse(&iter); 2433 if (ret) 2434 goto out; 2435 2436 /* has node been freed? */ 2437 if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) { 2438 /* node has been freed: */ 2439 BUG_ON(!btree_node_dying(b)); 2440 goto out; 2441 } 2442 2443 BUG_ON(!btree_node_hashed(b)); 2444 2445 struct bch_extent_ptr *ptr; 2446 bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr, 2447 !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev)); 2448 2449 ret = bch2_btree_node_update_key(trans, &iter, b, new_key, 2450 commit_flags, skip_triggers); 2451 out: 2452 bch2_trans_iter_exit(trans, &iter); 2453 return ret; 2454 } 2455 2456 /* Init code: */ 2457 2458 /* 2459 * Only for filesystem bringup, when first reading the btree roots or allocating 2460 * btree roots when initializing a new filesystem: 2461 */ 2462 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b) 2463 { 2464 BUG_ON(btree_node_root(c, b)); 2465 2466 bch2_btree_set_root_inmem(c, b); 2467 } 2468 2469 static int __bch2_btree_root_alloc_fake(struct btree_trans *trans, enum btree_id id, unsigned level) 2470 { 2471 struct bch_fs *c = trans->c; 2472 struct closure cl; 2473 struct btree *b; 2474 int ret; 2475 2476 closure_init_stack(&cl); 2477 2478 do { 2479 ret = bch2_btree_cache_cannibalize_lock(trans, &cl); 2480 closure_sync(&cl); 2481 } while (ret); 2482 2483 b = bch2_btree_node_mem_alloc(trans, false); 2484 bch2_btree_cache_cannibalize_unlock(trans); 2485 2486 set_btree_node_fake(b); 2487 set_btree_node_need_rewrite(b); 2488 b->c.level = level; 2489 b->c.btree_id = id; 2490 2491 bkey_btree_ptr_init(&b->key); 2492 b->key.k.p = SPOS_MAX; 2493 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id; 2494 2495 bch2_bset_init_first(b, &b->data->keys); 2496 bch2_btree_build_aux_trees(b); 2497 2498 b->data->flags = 0; 2499 btree_set_min(b, POS_MIN); 2500 btree_set_max(b, SPOS_MAX); 2501 b->data->format = bch2_btree_calc_format(b); 2502 btree_node_set_format(b, b->data->format); 2503 2504 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, 2505 b->c.level, b->c.btree_id); 2506 BUG_ON(ret); 2507 2508 bch2_btree_set_root_inmem(c, b); 2509 2510 six_unlock_write(&b->c.lock); 2511 six_unlock_intent(&b->c.lock); 2512 return 0; 2513 } 2514 2515 void bch2_btree_root_alloc_fake(struct bch_fs *c, enum btree_id id, unsigned level) 2516 { 2517 bch2_trans_run(c, __bch2_btree_root_alloc_fake(trans, id, level)); 2518 } 2519 2520 static void bch2_btree_update_to_text(struct printbuf *out, struct btree_update *as) 2521 { 2522 prt_printf(out, "%ps: btree=%s watermark=%s mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n", 2523 (void *) as->ip_started, 2524 bch2_btree_id_str(as->btree_id), 2525 bch2_watermarks[as->watermark], 2526 bch2_btree_update_modes[as->mode], 2527 as->nodes_written, 2528 closure_nr_remaining(&as->cl), 2529 as->journal.seq); 2530 } 2531 2532 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c) 2533 { 2534 struct btree_update *as; 2535 2536 mutex_lock(&c->btree_interior_update_lock); 2537 list_for_each_entry(as, &c->btree_interior_update_list, list) 2538 bch2_btree_update_to_text(out, as); 2539 mutex_unlock(&c->btree_interior_update_lock); 2540 } 2541 2542 static bool bch2_btree_interior_updates_pending(struct bch_fs *c) 2543 { 2544 bool ret; 2545 2546 mutex_lock(&c->btree_interior_update_lock); 2547 ret = !list_empty(&c->btree_interior_update_list); 2548 mutex_unlock(&c->btree_interior_update_lock); 2549 2550 return ret; 2551 } 2552 2553 bool bch2_btree_interior_updates_flush(struct bch_fs *c) 2554 { 2555 bool ret = bch2_btree_interior_updates_pending(c); 2556 2557 if (ret) 2558 closure_wait_event(&c->btree_interior_update_wait, 2559 !bch2_btree_interior_updates_pending(c)); 2560 return ret; 2561 } 2562 2563 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry) 2564 { 2565 struct btree_root *r = bch2_btree_id_root(c, entry->btree_id); 2566 2567 mutex_lock(&c->btree_root_lock); 2568 2569 r->level = entry->level; 2570 r->alive = true; 2571 bkey_copy(&r->key, (struct bkey_i *) entry->start); 2572 2573 mutex_unlock(&c->btree_root_lock); 2574 } 2575 2576 struct jset_entry * 2577 bch2_btree_roots_to_journal_entries(struct bch_fs *c, 2578 struct jset_entry *end, 2579 unsigned long skip) 2580 { 2581 unsigned i; 2582 2583 mutex_lock(&c->btree_root_lock); 2584 2585 for (i = 0; i < btree_id_nr_alive(c); i++) { 2586 struct btree_root *r = bch2_btree_id_root(c, i); 2587 2588 if (r->alive && !test_bit(i, &skip)) { 2589 journal_entry_set(end, BCH_JSET_ENTRY_btree_root, 2590 i, r->level, &r->key, r->key.k.u64s); 2591 end = vstruct_next(end); 2592 } 2593 } 2594 2595 mutex_unlock(&c->btree_root_lock); 2596 2597 return end; 2598 } 2599 2600 void bch2_fs_btree_interior_update_exit(struct bch_fs *c) 2601 { 2602 if (c->btree_node_rewrite_worker) 2603 destroy_workqueue(c->btree_node_rewrite_worker); 2604 if (c->btree_interior_update_worker) 2605 destroy_workqueue(c->btree_interior_update_worker); 2606 mempool_exit(&c->btree_interior_update_pool); 2607 } 2608 2609 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c) 2610 { 2611 mutex_init(&c->btree_reserve_cache_lock); 2612 INIT_LIST_HEAD(&c->btree_interior_update_list); 2613 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten); 2614 mutex_init(&c->btree_interior_update_lock); 2615 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work); 2616 2617 INIT_LIST_HEAD(&c->pending_node_rewrites); 2618 mutex_init(&c->pending_node_rewrites_lock); 2619 } 2620 2621 int bch2_fs_btree_interior_update_init(struct bch_fs *c) 2622 { 2623 c->btree_interior_update_worker = 2624 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 8); 2625 if (!c->btree_interior_update_worker) 2626 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init; 2627 2628 c->btree_node_rewrite_worker = 2629 alloc_ordered_workqueue("btree_node_rewrite", WQ_UNBOUND); 2630 if (!c->btree_node_rewrite_worker) 2631 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init; 2632 2633 if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1, 2634 sizeof(struct btree_update))) 2635 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init; 2636 2637 return 0; 2638 } 2639