1// SPDX-License-Identifier: GPL-2.0 2/* 3 * bcache journalling code, for btree insertions 4 * 5 * Copyright 2012 Google, Inc. 6 */ 7 8#include "bcache.h" 9#include "btree.h" 10#include "debug.h" 11#include "extents.h" 12 13#include <trace/events/bcache.h> 14 15/* 16 * Journal replay/recovery: 17 * 18 * This code is all driven from run_cache_set(); we first read the journal 19 * entries, do some other stuff, then we mark all the keys in the journal 20 * entries (same as garbage collection would), then we replay them - reinserting 21 * them into the cache in precisely the same order as they appear in the 22 * journal. 23 * 24 * We only journal keys that go in leaf nodes, which simplifies things quite a 25 * bit. 26 */ 27 28static void journal_read_endio(struct bio *bio) 29{ 30 struct closure *cl = bio->bi_private; 31 32 closure_put(cl); 33} 34 35static int journal_read_bucket(struct cache *ca, struct list_head *list, 36 unsigned int bucket_index) 37{ 38 struct journal_device *ja = &ca->journal; 39 struct bio *bio = &ja->bio; 40 41 struct journal_replay *i; 42 struct jset *j, *data = ca->set->journal.w[0].data; 43 struct closure cl; 44 unsigned int len, left, offset = 0; 45 int ret = 0; 46 sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]); 47 48 closure_init_stack(&cl); 49 50 pr_debug("reading %u\n", bucket_index); 51 52 while (offset < ca->sb.bucket_size) { 53reread: left = ca->sb.bucket_size - offset; 54 len = min_t(unsigned int, left, PAGE_SECTORS << JSET_BITS); 55 56 bio_reset(bio); 57 bio->bi_iter.bi_sector = bucket + offset; 58 bio_set_dev(bio, ca->bdev); 59 bio->bi_iter.bi_size = len << 9; 60 61 bio->bi_end_io = journal_read_endio; 62 bio->bi_private = &cl; 63 bio_set_op_attrs(bio, REQ_OP_READ, 0); 64 bch_bio_map(bio, data); 65 66 closure_bio_submit(ca->set, bio, &cl); 67 closure_sync(&cl); 68 69 /* This function could be simpler now since we no longer write 70 * journal entries that overlap bucket boundaries; this means 71 * the start of a bucket will always have a valid journal entry 72 * if it has any journal entries at all. 73 */ 74 75 j = data; 76 while (len) { 77 struct list_head *where; 78 size_t blocks, bytes = set_bytes(j); 79 80 if (j->magic != jset_magic(&ca->sb)) { 81 pr_debug("%u: bad magic\n", bucket_index); 82 return ret; 83 } 84 85 if (bytes > left << 9 || 86 bytes > PAGE_SIZE << JSET_BITS) { 87 pr_info("%u: too big, %zu bytes, offset %u\n", 88 bucket_index, bytes, offset); 89 return ret; 90 } 91 92 if (bytes > len << 9) 93 goto reread; 94 95 if (j->csum != csum_set(j)) { 96 pr_info("%u: bad csum, %zu bytes, offset %u\n", 97 bucket_index, bytes, offset); 98 return ret; 99 } 100 101 blocks = set_blocks(j, block_bytes(ca)); 102 103 /* 104 * Nodes in 'list' are in linear increasing order of 105 * i->j.seq, the node on head has the smallest (oldest) 106 * journal seq, the node on tail has the biggest 107 * (latest) journal seq. 108 */ 109 110 /* 111 * Check from the oldest jset for last_seq. If 112 * i->j.seq < j->last_seq, it means the oldest jset 113 * in list is expired and useless, remove it from 114 * this list. Otherwise, j is a condidate jset for 115 * further following checks. 116 */ 117 while (!list_empty(list)) { 118 i = list_first_entry(list, 119 struct journal_replay, list); 120 if (i->j.seq >= j->last_seq) 121 break; 122 list_del(&i->list); 123 kfree(i); 124 } 125 126 /* iterate list in reverse order (from latest jset) */ 127 list_for_each_entry_reverse(i, list, list) { 128 if (j->seq == i->j.seq) 129 goto next_set; 130 131 /* 132 * if j->seq is less than any i->j.last_seq 133 * in list, j is an expired and useless jset. 134 */ 135 if (j->seq < i->j.last_seq) 136 goto next_set; 137 138 /* 139 * 'where' points to first jset in list which 140 * is elder then j. 141 */ 142 if (j->seq > i->j.seq) { 143 where = &i->list; 144 goto add; 145 } 146 } 147 148 where = list; 149add: 150 i = kmalloc(offsetof(struct journal_replay, j) + 151 bytes, GFP_KERNEL); 152 if (!i) 153 return -ENOMEM; 154 memcpy(&i->j, j, bytes); 155 /* Add to the location after 'where' points to */ 156 list_add(&i->list, where); 157 ret = 1; 158 159 if (j->seq > ja->seq[bucket_index]) 160 ja->seq[bucket_index] = j->seq; 161next_set: 162 offset += blocks * ca->sb.block_size; 163 len -= blocks * ca->sb.block_size; 164 j = ((void *) j) + blocks * block_bytes(ca); 165 } 166 } 167 168 return ret; 169} 170 171int bch_journal_read(struct cache_set *c, struct list_head *list) 172{ 173#define read_bucket(b) \ 174 ({ \ 175 ret = journal_read_bucket(ca, list, b); \ 176 __set_bit(b, bitmap); \ 177 if (ret < 0) \ 178 return ret; \ 179 ret; \ 180 }) 181 182 struct cache *ca = c->cache; 183 int ret = 0; 184 struct journal_device *ja = &ca->journal; 185 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS); 186 unsigned int i, l, r, m; 187 uint64_t seq; 188 189 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS); 190 pr_debug("%u journal buckets\n", ca->sb.njournal_buckets); 191 192 /* 193 * Read journal buckets ordered by golden ratio hash to quickly 194 * find a sequence of buckets with valid journal entries 195 */ 196 for (i = 0; i < ca->sb.njournal_buckets; i++) { 197 /* 198 * We must try the index l with ZERO first for 199 * correctness due to the scenario that the journal 200 * bucket is circular buffer which might have wrapped 201 */ 202 l = (i * 2654435769U) % ca->sb.njournal_buckets; 203 204 if (test_bit(l, bitmap)) 205 break; 206 207 if (read_bucket(l)) 208 goto bsearch; 209 } 210 211 /* 212 * If that fails, check all the buckets we haven't checked 213 * already 214 */ 215 pr_debug("falling back to linear search\n"); 216 217 for_each_clear_bit(l, bitmap, ca->sb.njournal_buckets) 218 if (read_bucket(l)) 219 goto bsearch; 220 221 /* no journal entries on this device? */ 222 if (l == ca->sb.njournal_buckets) 223 goto out; 224bsearch: 225 BUG_ON(list_empty(list)); 226 227 /* Binary search */ 228 m = l; 229 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1); 230 pr_debug("starting binary search, l %u r %u\n", l, r); 231 232 while (l + 1 < r) { 233 seq = list_entry(list->prev, struct journal_replay, 234 list)->j.seq; 235 236 m = (l + r) >> 1; 237 read_bucket(m); 238 239 if (seq != list_entry(list->prev, struct journal_replay, 240 list)->j.seq) 241 l = m; 242 else 243 r = m; 244 } 245 246 /* 247 * Read buckets in reverse order until we stop finding more 248 * journal entries 249 */ 250 pr_debug("finishing up: m %u njournal_buckets %u\n", 251 m, ca->sb.njournal_buckets); 252 l = m; 253 254 while (1) { 255 if (!l--) 256 l = ca->sb.njournal_buckets - 1; 257 258 if (l == m) 259 break; 260 261 if (test_bit(l, bitmap)) 262 continue; 263 264 if (!read_bucket(l)) 265 break; 266 } 267 268 seq = 0; 269 270 for (i = 0; i < ca->sb.njournal_buckets; i++) 271 if (ja->seq[i] > seq) { 272 seq = ja->seq[i]; 273 /* 274 * When journal_reclaim() goes to allocate for 275 * the first time, it'll use the bucket after 276 * ja->cur_idx 277 */ 278 ja->cur_idx = i; 279 ja->last_idx = ja->discard_idx = (i + 1) % 280 ca->sb.njournal_buckets; 281 282 } 283 284out: 285 if (!list_empty(list)) 286 c->journal.seq = list_entry(list->prev, 287 struct journal_replay, 288 list)->j.seq; 289 290 return 0; 291#undef read_bucket 292} 293 294void bch_journal_mark(struct cache_set *c, struct list_head *list) 295{ 296 atomic_t p = { 0 }; 297 struct bkey *k; 298 struct journal_replay *i; 299 struct journal *j = &c->journal; 300 uint64_t last = j->seq; 301 302 /* 303 * journal.pin should never fill up - we never write a journal 304 * entry when it would fill up. But if for some reason it does, we 305 * iterate over the list in reverse order so that we can just skip that 306 * refcount instead of bugging. 307 */ 308 309 list_for_each_entry_reverse(i, list, list) { 310 BUG_ON(last < i->j.seq); 311 i->pin = NULL; 312 313 while (last-- != i->j.seq) 314 if (fifo_free(&j->pin) > 1) { 315 fifo_push_front(&j->pin, p); 316 atomic_set(&fifo_front(&j->pin), 0); 317 } 318 319 if (fifo_free(&j->pin) > 1) { 320 fifo_push_front(&j->pin, p); 321 i->pin = &fifo_front(&j->pin); 322 atomic_set(i->pin, 1); 323 } 324 325 for (k = i->j.start; 326 k < bset_bkey_last(&i->j); 327 k = bkey_next(k)) 328 if (!__bch_extent_invalid(c, k)) { 329 unsigned int j; 330 331 for (j = 0; j < KEY_PTRS(k); j++) 332 if (ptr_available(c, k, j)) 333 atomic_inc(&PTR_BUCKET(c, k, j)->pin); 334 335 bch_initial_mark_key(c, 0, k); 336 } 337 } 338} 339 340static bool is_discard_enabled(struct cache_set *s) 341{ 342 struct cache *ca = s->cache; 343 344 if (ca->discard) 345 return true; 346 347 return false; 348} 349 350int bch_journal_replay(struct cache_set *s, struct list_head *list) 351{ 352 int ret = 0, keys = 0, entries = 0; 353 struct bkey *k; 354 struct journal_replay *i = 355 list_entry(list->prev, struct journal_replay, list); 356 357 uint64_t start = i->j.last_seq, end = i->j.seq, n = start; 358 struct keylist keylist; 359 360 list_for_each_entry(i, list, list) { 361 BUG_ON(i->pin && atomic_read(i->pin) != 1); 362 363 if (n != i->j.seq) { 364 if (n == start && is_discard_enabled(s)) 365 pr_info("journal entries %llu-%llu may be discarded! (replaying %llu-%llu)\n", 366 n, i->j.seq - 1, start, end); 367 else { 368 pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n", 369 n, i->j.seq - 1, start, end); 370 ret = -EIO; 371 goto err; 372 } 373 } 374 375 for (k = i->j.start; 376 k < bset_bkey_last(&i->j); 377 k = bkey_next(k)) { 378 trace_bcache_journal_replay_key(k); 379 380 bch_keylist_init_single(&keylist, k); 381 382 ret = bch_btree_insert(s, &keylist, i->pin, NULL); 383 if (ret) 384 goto err; 385 386 BUG_ON(!bch_keylist_empty(&keylist)); 387 keys++; 388 389 cond_resched(); 390 } 391 392 if (i->pin) 393 atomic_dec(i->pin); 394 n = i->j.seq + 1; 395 entries++; 396 } 397 398 pr_info("journal replay done, %i keys in %i entries, seq %llu\n", 399 keys, entries, end); 400err: 401 while (!list_empty(list)) { 402 i = list_first_entry(list, struct journal_replay, list); 403 list_del(&i->list); 404 kfree(i); 405 } 406 407 return ret; 408} 409 410void bch_journal_space_reserve(struct journal *j) 411{ 412 j->do_reserve = true; 413} 414 415/* Journalling */ 416 417static void btree_flush_write(struct cache_set *c) 418{ 419 struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR]; 420 unsigned int i, nr; 421 int ref_nr; 422 atomic_t *fifo_front_p, *now_fifo_front_p; 423 size_t mask; 424 425 if (c->journal.btree_flushing) 426 return; 427 428 spin_lock(&c->journal.flush_write_lock); 429 if (c->journal.btree_flushing) { 430 spin_unlock(&c->journal.flush_write_lock); 431 return; 432 } 433 c->journal.btree_flushing = true; 434 spin_unlock(&c->journal.flush_write_lock); 435 436 /* get the oldest journal entry and check its refcount */ 437 spin_lock(&c->journal.lock); 438 fifo_front_p = &fifo_front(&c->journal.pin); 439 ref_nr = atomic_read(fifo_front_p); 440 if (ref_nr <= 0) { 441 /* 442 * do nothing if no btree node references 443 * the oldest journal entry 444 */ 445 spin_unlock(&c->journal.lock); 446 goto out; 447 } 448 spin_unlock(&c->journal.lock); 449 450 mask = c->journal.pin.mask; 451 nr = 0; 452 atomic_long_inc(&c->flush_write); 453 memset(btree_nodes, 0, sizeof(btree_nodes)); 454 455 mutex_lock(&c->bucket_lock); 456 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) { 457 /* 458 * It is safe to get now_fifo_front_p without holding 459 * c->journal.lock here, because we don't need to know 460 * the exactly accurate value, just check whether the 461 * front pointer of c->journal.pin is changed. 462 */ 463 now_fifo_front_p = &fifo_front(&c->journal.pin); 464 /* 465 * If the oldest journal entry is reclaimed and front 466 * pointer of c->journal.pin changes, it is unnecessary 467 * to scan c->btree_cache anymore, just quit the loop and 468 * flush out what we have already. 469 */ 470 if (now_fifo_front_p != fifo_front_p) 471 break; 472 /* 473 * quit this loop if all matching btree nodes are 474 * scanned and record in btree_nodes[] already. 475 */ 476 ref_nr = atomic_read(fifo_front_p); 477 if (nr >= ref_nr) 478 break; 479 480 if (btree_node_journal_flush(b)) 481 pr_err("BUG: flush_write bit should not be set here!\n"); 482 483 mutex_lock(&b->write_lock); 484 485 if (!btree_node_dirty(b)) { 486 mutex_unlock(&b->write_lock); 487 continue; 488 } 489 490 if (!btree_current_write(b)->journal) { 491 mutex_unlock(&b->write_lock); 492 continue; 493 } 494 495 /* 496 * Only select the btree node which exactly references 497 * the oldest journal entry. 498 * 499 * If the journal entry pointed by fifo_front_p is 500 * reclaimed in parallel, don't worry: 501 * - the list_for_each_xxx loop will quit when checking 502 * next now_fifo_front_p. 503 * - If there are matched nodes recorded in btree_nodes[], 504 * they are clean now (this is why and how the oldest 505 * journal entry can be reclaimed). These selected nodes 506 * will be ignored and skipped in the folowing for-loop. 507 */ 508 if (((btree_current_write(b)->journal - fifo_front_p) & 509 mask) != 0) { 510 mutex_unlock(&b->write_lock); 511 continue; 512 } 513 514 set_btree_node_journal_flush(b); 515 516 mutex_unlock(&b->write_lock); 517 518 btree_nodes[nr++] = b; 519 /* 520 * To avoid holding c->bucket_lock too long time, 521 * only scan for BTREE_FLUSH_NR matched btree nodes 522 * at most. If there are more btree nodes reference 523 * the oldest journal entry, try to flush them next 524 * time when btree_flush_write() is called. 525 */ 526 if (nr == BTREE_FLUSH_NR) 527 break; 528 } 529 mutex_unlock(&c->bucket_lock); 530 531 for (i = 0; i < nr; i++) { 532 b = btree_nodes[i]; 533 if (!b) { 534 pr_err("BUG: btree_nodes[%d] is NULL\n", i); 535 continue; 536 } 537 538 /* safe to check without holding b->write_lock */ 539 if (!btree_node_journal_flush(b)) { 540 pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b); 541 continue; 542 } 543 544 mutex_lock(&b->write_lock); 545 if (!btree_current_write(b)->journal) { 546 clear_bit(BTREE_NODE_journal_flush, &b->flags); 547 mutex_unlock(&b->write_lock); 548 pr_debug("bnode %p: written by others\n", b); 549 continue; 550 } 551 552 if (!btree_node_dirty(b)) { 553 clear_bit(BTREE_NODE_journal_flush, &b->flags); 554 mutex_unlock(&b->write_lock); 555 pr_debug("bnode %p: dirty bit cleaned by others\n", b); 556 continue; 557 } 558 559 __bch_btree_node_write(b, NULL); 560 clear_bit(BTREE_NODE_journal_flush, &b->flags); 561 mutex_unlock(&b->write_lock); 562 } 563 564out: 565 spin_lock(&c->journal.flush_write_lock); 566 c->journal.btree_flushing = false; 567 spin_unlock(&c->journal.flush_write_lock); 568} 569 570#define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1) 571 572static void journal_discard_endio(struct bio *bio) 573{ 574 struct journal_device *ja = 575 container_of(bio, struct journal_device, discard_bio); 576 struct cache *ca = container_of(ja, struct cache, journal); 577 578 atomic_set(&ja->discard_in_flight, DISCARD_DONE); 579 580 closure_wake_up(&ca->set->journal.wait); 581 closure_put(&ca->set->cl); 582} 583 584static void journal_discard_work(struct work_struct *work) 585{ 586 struct journal_device *ja = 587 container_of(work, struct journal_device, discard_work); 588 589 submit_bio(&ja->discard_bio); 590} 591 592static void do_journal_discard(struct cache *ca) 593{ 594 struct journal_device *ja = &ca->journal; 595 struct bio *bio = &ja->discard_bio; 596 597 if (!ca->discard) { 598 ja->discard_idx = ja->last_idx; 599 return; 600 } 601 602 switch (atomic_read(&ja->discard_in_flight)) { 603 case DISCARD_IN_FLIGHT: 604 return; 605 606 case DISCARD_DONE: 607 ja->discard_idx = (ja->discard_idx + 1) % 608 ca->sb.njournal_buckets; 609 610 atomic_set(&ja->discard_in_flight, DISCARD_READY); 611 fallthrough; 612 613 case DISCARD_READY: 614 if (ja->discard_idx == ja->last_idx) 615 return; 616 617 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT); 618 619 bio_init(bio, bio->bi_inline_vecs, 1); 620 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0); 621 bio->bi_iter.bi_sector = bucket_to_sector(ca->set, 622 ca->sb.d[ja->discard_idx]); 623 bio_set_dev(bio, ca->bdev); 624 bio->bi_iter.bi_size = bucket_bytes(ca); 625 bio->bi_end_io = journal_discard_endio; 626 627 closure_get(&ca->set->cl); 628 INIT_WORK(&ja->discard_work, journal_discard_work); 629 queue_work(bch_journal_wq, &ja->discard_work); 630 } 631} 632 633static unsigned int free_journal_buckets(struct cache_set *c) 634{ 635 struct journal *j = &c->journal; 636 struct cache *ca = c->cache; 637 struct journal_device *ja = &c->cache->journal; 638 unsigned int n; 639 640 /* In case njournal_buckets is not power of 2 */ 641 if (ja->cur_idx >= ja->discard_idx) 642 n = ca->sb.njournal_buckets + ja->discard_idx - ja->cur_idx; 643 else 644 n = ja->discard_idx - ja->cur_idx; 645 646 if (n > (1 + j->do_reserve)) 647 return n - (1 + j->do_reserve); 648 649 return 0; 650} 651 652static void journal_reclaim(struct cache_set *c) 653{ 654 struct bkey *k = &c->journal.key; 655 struct cache *ca = c->cache; 656 uint64_t last_seq; 657 struct journal_device *ja = &ca->journal; 658 atomic_t p __maybe_unused; 659 660 atomic_long_inc(&c->reclaim); 661 662 while (!atomic_read(&fifo_front(&c->journal.pin))) 663 fifo_pop(&c->journal.pin, p); 664 665 last_seq = last_seq(&c->journal); 666 667 /* Update last_idx */ 668 669 while (ja->last_idx != ja->cur_idx && 670 ja->seq[ja->last_idx] < last_seq) 671 ja->last_idx = (ja->last_idx + 1) % 672 ca->sb.njournal_buckets; 673 674 do_journal_discard(ca); 675 676 if (c->journal.blocks_free) 677 goto out; 678 679 if (!free_journal_buckets(c)) 680 goto out; 681 682 ja->cur_idx = (ja->cur_idx + 1) % ca->sb.njournal_buckets; 683 k->ptr[0] = MAKE_PTR(0, 684 bucket_to_sector(c, ca->sb.d[ja->cur_idx]), 685 ca->sb.nr_this_dev); 686 atomic_long_inc(&c->reclaimed_journal_buckets); 687 688 bkey_init(k); 689 SET_KEY_PTRS(k, 1); 690 c->journal.blocks_free = ca->sb.bucket_size >> c->block_bits; 691 692out: 693 if (!journal_full(&c->journal)) 694 __closure_wake_up(&c->journal.wait); 695} 696 697void bch_journal_next(struct journal *j) 698{ 699 atomic_t p = { 1 }; 700 701 j->cur = (j->cur == j->w) 702 ? &j->w[1] 703 : &j->w[0]; 704 705 /* 706 * The fifo_push() needs to happen at the same time as j->seq is 707 * incremented for last_seq() to be calculated correctly 708 */ 709 BUG_ON(!fifo_push(&j->pin, p)); 710 atomic_set(&fifo_back(&j->pin), 1); 711 712 j->cur->data->seq = ++j->seq; 713 j->cur->dirty = false; 714 j->cur->need_write = false; 715 j->cur->data->keys = 0; 716 717 if (fifo_full(&j->pin)) 718 pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin)); 719} 720 721static void journal_write_endio(struct bio *bio) 722{ 723 struct journal_write *w = bio->bi_private; 724 725 cache_set_err_on(bio->bi_status, w->c, "journal io error"); 726 closure_put(&w->c->journal.io); 727} 728 729static void journal_write(struct closure *cl); 730 731static void journal_write_done(struct closure *cl) 732{ 733 struct journal *j = container_of(cl, struct journal, io); 734 struct journal_write *w = (j->cur == j->w) 735 ? &j->w[1] 736 : &j->w[0]; 737 738 __closure_wake_up(&w->wait); 739 continue_at_nobarrier(cl, journal_write, bch_journal_wq); 740} 741 742static void journal_write_unlock(struct closure *cl) 743 __releases(&c->journal.lock) 744{ 745 struct cache_set *c = container_of(cl, struct cache_set, journal.io); 746 747 c->journal.io_in_flight = 0; 748 spin_unlock(&c->journal.lock); 749} 750 751static void journal_write_unlocked(struct closure *cl) 752 __releases(c->journal.lock) 753{ 754 struct cache_set *c = container_of(cl, struct cache_set, journal.io); 755 struct cache *ca = c->cache; 756 struct journal_write *w = c->journal.cur; 757 struct bkey *k = &c->journal.key; 758 unsigned int i, sectors = set_blocks(w->data, block_bytes(ca)) * 759 ca->sb.block_size; 760 761 struct bio *bio; 762 struct bio_list list; 763 764 bio_list_init(&list); 765 766 if (!w->need_write) { 767 closure_return_with_destructor(cl, journal_write_unlock); 768 return; 769 } else if (journal_full(&c->journal)) { 770 journal_reclaim(c); 771 spin_unlock(&c->journal.lock); 772 773 btree_flush_write(c); 774 continue_at(cl, journal_write, bch_journal_wq); 775 return; 776 } 777 778 c->journal.blocks_free -= set_blocks(w->data, block_bytes(ca)); 779 780 w->data->btree_level = c->root->level; 781 782 bkey_copy(&w->data->btree_root, &c->root->key); 783 bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket); 784 785 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0]; 786 w->data->magic = jset_magic(&ca->sb); 787 w->data->version = BCACHE_JSET_VERSION; 788 w->data->last_seq = last_seq(&c->journal); 789 w->data->csum = csum_set(w->data); 790 791 for (i = 0; i < KEY_PTRS(k); i++) { 792 ca = PTR_CACHE(c, k, i); 793 bio = &ca->journal.bio; 794 795 atomic_long_add(sectors, &ca->meta_sectors_written); 796 797 bio_reset(bio); 798 bio->bi_iter.bi_sector = PTR_OFFSET(k, i); 799 bio_set_dev(bio, ca->bdev); 800 bio->bi_iter.bi_size = sectors << 9; 801 802 bio->bi_end_io = journal_write_endio; 803 bio->bi_private = w; 804 bio_set_op_attrs(bio, REQ_OP_WRITE, 805 REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA); 806 bch_bio_map(bio, w->data); 807 808 trace_bcache_journal_write(bio, w->data->keys); 809 bio_list_add(&list, bio); 810 811 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors); 812 813 ca->journal.seq[ca->journal.cur_idx] = w->data->seq; 814 } 815 816 /* If KEY_PTRS(k) == 0, this jset gets lost in air */ 817 BUG_ON(i == 0); 818 819 atomic_dec_bug(&fifo_back(&c->journal.pin)); 820 bch_journal_next(&c->journal); 821 journal_reclaim(c); 822 823 spin_unlock(&c->journal.lock); 824 825 while ((bio = bio_list_pop(&list))) 826 closure_bio_submit(c, bio, cl); 827 828 continue_at(cl, journal_write_done, NULL); 829} 830 831static void journal_write(struct closure *cl) 832{ 833 struct cache_set *c = container_of(cl, struct cache_set, journal.io); 834 835 spin_lock(&c->journal.lock); 836 journal_write_unlocked(cl); 837} 838 839static void journal_try_write(struct cache_set *c) 840 __releases(c->journal.lock) 841{ 842 struct closure *cl = &c->journal.io; 843 struct journal_write *w = c->journal.cur; 844 845 w->need_write = true; 846 847 if (!c->journal.io_in_flight) { 848 c->journal.io_in_flight = 1; 849 closure_call(cl, journal_write_unlocked, NULL, &c->cl); 850 } else { 851 spin_unlock(&c->journal.lock); 852 } 853} 854 855static struct journal_write *journal_wait_for_write(struct cache_set *c, 856 unsigned int nkeys) 857 __acquires(&c->journal.lock) 858{ 859 size_t sectors; 860 struct closure cl; 861 bool wait = false; 862 struct cache *ca = c->cache; 863 864 closure_init_stack(&cl); 865 866 spin_lock(&c->journal.lock); 867 868 while (1) { 869 struct journal_write *w = c->journal.cur; 870 871 sectors = __set_blocks(w->data, w->data->keys + nkeys, 872 block_bytes(ca)) * ca->sb.block_size; 873 874 if (sectors <= min_t(size_t, 875 c->journal.blocks_free * ca->sb.block_size, 876 PAGE_SECTORS << JSET_BITS)) 877 return w; 878 879 if (wait) 880 closure_wait(&c->journal.wait, &cl); 881 882 if (!journal_full(&c->journal)) { 883 if (wait) 884 trace_bcache_journal_entry_full(c); 885 886 /* 887 * XXX: If we were inserting so many keys that they 888 * won't fit in an _empty_ journal write, we'll 889 * deadlock. For now, handle this in 890 * bch_keylist_realloc() - but something to think about. 891 */ 892 BUG_ON(!w->data->keys); 893 894 journal_try_write(c); /* unlocks */ 895 } else { 896 if (wait) 897 trace_bcache_journal_full(c); 898 899 journal_reclaim(c); 900 spin_unlock(&c->journal.lock); 901 902 btree_flush_write(c); 903 } 904 905 closure_sync(&cl); 906 spin_lock(&c->journal.lock); 907 wait = true; 908 } 909} 910 911static void journal_write_work(struct work_struct *work) 912{ 913 struct cache_set *c = container_of(to_delayed_work(work), 914 struct cache_set, 915 journal.work); 916 spin_lock(&c->journal.lock); 917 if (c->journal.cur->dirty) 918 journal_try_write(c); 919 else 920 spin_unlock(&c->journal.lock); 921} 922 923/* 924 * Entry point to the journalling code - bio_insert() and btree_invalidate() 925 * pass bch_journal() a list of keys to be journalled, and then 926 * bch_journal() hands those same keys off to btree_insert_async() 927 */ 928 929atomic_t *bch_journal(struct cache_set *c, 930 struct keylist *keys, 931 struct closure *parent) 932{ 933 struct journal_write *w; 934 atomic_t *ret; 935 936 /* No journaling if CACHE_SET_IO_DISABLE set already */ 937 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags))) 938 return NULL; 939 940 if (!CACHE_SYNC(&c->cache->sb)) 941 return NULL; 942 943 w = journal_wait_for_write(c, bch_keylist_nkeys(keys)); 944 945 memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys)); 946 w->data->keys += bch_keylist_nkeys(keys); 947 948 ret = &fifo_back(&c->journal.pin); 949 atomic_inc(ret); 950 951 if (parent) { 952 closure_wait(&w->wait, parent); 953 journal_try_write(c); 954 } else if (!w->dirty) { 955 w->dirty = true; 956 queue_delayed_work(bch_flush_wq, &c->journal.work, 957 msecs_to_jiffies(c->journal_delay_ms)); 958 spin_unlock(&c->journal.lock); 959 } else { 960 spin_unlock(&c->journal.lock); 961 } 962 963 964 return ret; 965} 966 967void bch_journal_meta(struct cache_set *c, struct closure *cl) 968{ 969 struct keylist keys; 970 atomic_t *ref; 971 972 bch_keylist_init(&keys); 973 974 ref = bch_journal(c, &keys, cl); 975 if (ref) 976 atomic_dec_bug(ref); 977} 978 979void bch_journal_free(struct cache_set *c) 980{ 981 free_pages((unsigned long) c->journal.w[1].data, JSET_BITS); 982 free_pages((unsigned long) c->journal.w[0].data, JSET_BITS); 983 free_fifo(&c->journal.pin); 984} 985 986int bch_journal_alloc(struct cache_set *c) 987{ 988 struct journal *j = &c->journal; 989 990 spin_lock_init(&j->lock); 991 spin_lock_init(&j->flush_write_lock); 992 INIT_DELAYED_WORK(&j->work, journal_write_work); 993 994 c->journal_delay_ms = 100; 995 996 j->w[0].c = c; 997 j->w[1].c = c; 998 999 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) || 1000 !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) || 1001 !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS))) 1002 return -ENOMEM; 1003 1004 return 0; 1005} 1006