journal.c revision c18536a72ddd7fe30d63e6c1500b5c930ac14594
1/* 2 * bcache journalling code, for btree insertions 3 * 4 * Copyright 2012 Google, Inc. 5 */ 6 7#include "bcache.h" 8#include "btree.h" 9#include "debug.h" 10 11#include <trace/events/bcache.h> 12 13/* 14 * Journal replay/recovery: 15 * 16 * This code is all driven from run_cache_set(); we first read the journal 17 * entries, do some other stuff, then we mark all the keys in the journal 18 * entries (same as garbage collection would), then we replay them - reinserting 19 * them into the cache in precisely the same order as they appear in the 20 * journal. 21 * 22 * We only journal keys that go in leaf nodes, which simplifies things quite a 23 * bit. 24 */ 25 26static void journal_read_endio(struct bio *bio, int error) 27{ 28 struct closure *cl = bio->bi_private; 29 closure_put(cl); 30} 31 32static int journal_read_bucket(struct cache *ca, struct list_head *list, 33 unsigned bucket_index) 34{ 35 struct journal_device *ja = &ca->journal; 36 struct bio *bio = &ja->bio; 37 38 struct journal_replay *i; 39 struct jset *j, *data = ca->set->journal.w[0].data; 40 struct closure cl; 41 unsigned len, left, offset = 0; 42 int ret = 0; 43 sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]); 44 45 closure_init_stack(&cl); 46 47 pr_debug("reading %llu", (uint64_t) bucket); 48 49 while (offset < ca->sb.bucket_size) { 50reread: left = ca->sb.bucket_size - offset; 51 len = min_t(unsigned, left, PAGE_SECTORS * 8); 52 53 bio_reset(bio); 54 bio->bi_sector = bucket + offset; 55 bio->bi_bdev = ca->bdev; 56 bio->bi_rw = READ; 57 bio->bi_size = len << 9; 58 59 bio->bi_end_io = journal_read_endio; 60 bio->bi_private = &cl; 61 bch_bio_map(bio, data); 62 63 closure_bio_submit(bio, &cl, ca); 64 closure_sync(&cl); 65 66 /* This function could be simpler now since we no longer write 67 * journal entries that overlap bucket boundaries; this means 68 * the start of a bucket will always have a valid journal entry 69 * if it has any journal entries at all. 70 */ 71 72 j = data; 73 while (len) { 74 struct list_head *where; 75 size_t blocks, bytes = set_bytes(j); 76 77 if (j->magic != jset_magic(ca->set)) 78 return ret; 79 80 if (bytes > left << 9) 81 return ret; 82 83 if (bytes > len << 9) 84 goto reread; 85 86 if (j->csum != csum_set(j)) 87 return ret; 88 89 blocks = set_blocks(j, ca->set); 90 91 while (!list_empty(list)) { 92 i = list_first_entry(list, 93 struct journal_replay, list); 94 if (i->j.seq >= j->last_seq) 95 break; 96 list_del(&i->list); 97 kfree(i); 98 } 99 100 list_for_each_entry_reverse(i, list, list) { 101 if (j->seq == i->j.seq) 102 goto next_set; 103 104 if (j->seq < i->j.last_seq) 105 goto next_set; 106 107 if (j->seq > i->j.seq) { 108 where = &i->list; 109 goto add; 110 } 111 } 112 113 where = list; 114add: 115 i = kmalloc(offsetof(struct journal_replay, j) + 116 bytes, GFP_KERNEL); 117 if (!i) 118 return -ENOMEM; 119 memcpy(&i->j, j, bytes); 120 list_add(&i->list, where); 121 ret = 1; 122 123 ja->seq[bucket_index] = j->seq; 124next_set: 125 offset += blocks * ca->sb.block_size; 126 len -= blocks * ca->sb.block_size; 127 j = ((void *) j) + blocks * block_bytes(ca); 128 } 129 } 130 131 return ret; 132} 133 134int bch_journal_read(struct cache_set *c, struct list_head *list) 135{ 136#define read_bucket(b) \ 137 ({ \ 138 int ret = journal_read_bucket(ca, list, b); \ 139 __set_bit(b, bitmap); \ 140 if (ret < 0) \ 141 return ret; \ 142 ret; \ 143 }) 144 145 struct cache *ca; 146 unsigned iter; 147 148 for_each_cache(ca, c, iter) { 149 struct journal_device *ja = &ca->journal; 150 unsigned long bitmap[SB_JOURNAL_BUCKETS / BITS_PER_LONG]; 151 unsigned i, l, r, m; 152 uint64_t seq; 153 154 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS); 155 pr_debug("%u journal buckets", ca->sb.njournal_buckets); 156 157 /* 158 * Read journal buckets ordered by golden ratio hash to quickly 159 * find a sequence of buckets with valid journal entries 160 */ 161 for (i = 0; i < ca->sb.njournal_buckets; i++) { 162 l = (i * 2654435769U) % ca->sb.njournal_buckets; 163 164 if (test_bit(l, bitmap)) 165 break; 166 167 if (read_bucket(l)) 168 goto bsearch; 169 } 170 171 /* 172 * If that fails, check all the buckets we haven't checked 173 * already 174 */ 175 pr_debug("falling back to linear search"); 176 177 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets); 178 l < ca->sb.njournal_buckets; 179 l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1)) 180 if (read_bucket(l)) 181 goto bsearch; 182 183 if (list_empty(list)) 184 continue; 185bsearch: 186 /* Binary search */ 187 m = r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1); 188 pr_debug("starting binary search, l %u r %u", l, r); 189 190 while (l + 1 < r) { 191 seq = list_entry(list->prev, struct journal_replay, 192 list)->j.seq; 193 194 m = (l + r) >> 1; 195 read_bucket(m); 196 197 if (seq != list_entry(list->prev, struct journal_replay, 198 list)->j.seq) 199 l = m; 200 else 201 r = m; 202 } 203 204 /* 205 * Read buckets in reverse order until we stop finding more 206 * journal entries 207 */ 208 pr_debug("finishing up: m %u njournal_buckets %u", 209 m, ca->sb.njournal_buckets); 210 l = m; 211 212 while (1) { 213 if (!l--) 214 l = ca->sb.njournal_buckets - 1; 215 216 if (l == m) 217 break; 218 219 if (test_bit(l, bitmap)) 220 continue; 221 222 if (!read_bucket(l)) 223 break; 224 } 225 226 seq = 0; 227 228 for (i = 0; i < ca->sb.njournal_buckets; i++) 229 if (ja->seq[i] > seq) { 230 seq = ja->seq[i]; 231 ja->cur_idx = ja->discard_idx = 232 ja->last_idx = i; 233 234 } 235 } 236 237 if (!list_empty(list)) 238 c->journal.seq = list_entry(list->prev, 239 struct journal_replay, 240 list)->j.seq; 241 242 return 0; 243#undef read_bucket 244} 245 246void bch_journal_mark(struct cache_set *c, struct list_head *list) 247{ 248 atomic_t p = { 0 }; 249 struct bkey *k; 250 struct journal_replay *i; 251 struct journal *j = &c->journal; 252 uint64_t last = j->seq; 253 254 /* 255 * journal.pin should never fill up - we never write a journal 256 * entry when it would fill up. But if for some reason it does, we 257 * iterate over the list in reverse order so that we can just skip that 258 * refcount instead of bugging. 259 */ 260 261 list_for_each_entry_reverse(i, list, list) { 262 BUG_ON(last < i->j.seq); 263 i->pin = NULL; 264 265 while (last-- != i->j.seq) 266 if (fifo_free(&j->pin) > 1) { 267 fifo_push_front(&j->pin, p); 268 atomic_set(&fifo_front(&j->pin), 0); 269 } 270 271 if (fifo_free(&j->pin) > 1) { 272 fifo_push_front(&j->pin, p); 273 i->pin = &fifo_front(&j->pin); 274 atomic_set(i->pin, 1); 275 } 276 277 for (k = i->j.start; 278 k < end(&i->j); 279 k = bkey_next(k)) { 280 unsigned j; 281 282 for (j = 0; j < KEY_PTRS(k); j++) { 283 struct bucket *g = PTR_BUCKET(c, k, j); 284 atomic_inc(&g->pin); 285 286 if (g->prio == BTREE_PRIO && 287 !ptr_stale(c, k, j)) 288 g->prio = INITIAL_PRIO; 289 } 290 291 __bch_btree_mark_key(c, 0, k); 292 } 293 } 294} 295 296int bch_journal_replay(struct cache_set *s, struct list_head *list) 297{ 298 int ret = 0, keys = 0, entries = 0; 299 struct bkey *k; 300 struct journal_replay *i = 301 list_entry(list->prev, struct journal_replay, list); 302 303 uint64_t start = i->j.last_seq, end = i->j.seq, n = start; 304 struct keylist keylist; 305 struct btree_op op; 306 307 bch_keylist_init(&keylist); 308 bch_btree_op_init_stack(&op); 309 op.lock = SHRT_MAX; 310 311 list_for_each_entry(i, list, list) { 312 BUG_ON(i->pin && atomic_read(i->pin) != 1); 313 314 cache_set_err_on(n != i->j.seq, s, 315"bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)", 316 n, i->j.seq - 1, start, end); 317 318 for (k = i->j.start; 319 k < end(&i->j); 320 k = bkey_next(k)) { 321 trace_bcache_journal_replay_key(k); 322 323 bkey_copy(keylist.top, k); 324 bch_keylist_push(&keylist); 325 326 ret = bch_btree_insert(&op, s, &keylist, i->pin); 327 if (ret) 328 goto err; 329 330 BUG_ON(!bch_keylist_empty(&keylist)); 331 keys++; 332 333 cond_resched(); 334 } 335 336 if (i->pin) 337 atomic_dec(i->pin); 338 n = i->j.seq + 1; 339 entries++; 340 } 341 342 pr_info("journal replay done, %i keys in %i entries, seq %llu", 343 keys, entries, end); 344 345 while (!list_empty(list)) { 346 i = list_first_entry(list, struct journal_replay, list); 347 list_del(&i->list); 348 kfree(i); 349 } 350err: 351 closure_sync(&op.cl); 352 return ret; 353} 354 355/* Journalling */ 356 357static void btree_flush_write(struct cache_set *c) 358{ 359 /* 360 * Try to find the btree node with that references the oldest journal 361 * entry, best is our current candidate and is locked if non NULL: 362 */ 363 struct btree *b, *best; 364 unsigned i; 365retry: 366 best = NULL; 367 368 for_each_cached_btree(b, c, i) 369 if (btree_current_write(b)->journal) { 370 if (!best) 371 best = b; 372 else if (journal_pin_cmp(c, 373 btree_current_write(best)->journal, 374 btree_current_write(b)->journal)) { 375 best = b; 376 } 377 } 378 379 b = best; 380 if (b) { 381 rw_lock(true, b, b->level); 382 383 if (!btree_current_write(b)->journal) { 384 rw_unlock(true, b); 385 /* We raced */ 386 goto retry; 387 } 388 389 bch_btree_node_write(b, NULL); 390 rw_unlock(true, b); 391 } 392} 393 394#define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1) 395 396static void journal_discard_endio(struct bio *bio, int error) 397{ 398 struct journal_device *ja = 399 container_of(bio, struct journal_device, discard_bio); 400 struct cache *ca = container_of(ja, struct cache, journal); 401 402 atomic_set(&ja->discard_in_flight, DISCARD_DONE); 403 404 closure_wake_up(&ca->set->journal.wait); 405 closure_put(&ca->set->cl); 406} 407 408static void journal_discard_work(struct work_struct *work) 409{ 410 struct journal_device *ja = 411 container_of(work, struct journal_device, discard_work); 412 413 submit_bio(0, &ja->discard_bio); 414} 415 416static void do_journal_discard(struct cache *ca) 417{ 418 struct journal_device *ja = &ca->journal; 419 struct bio *bio = &ja->discard_bio; 420 421 if (!ca->discard) { 422 ja->discard_idx = ja->last_idx; 423 return; 424 } 425 426 switch (atomic_read(&ja->discard_in_flight)) { 427 case DISCARD_IN_FLIGHT: 428 return; 429 430 case DISCARD_DONE: 431 ja->discard_idx = (ja->discard_idx + 1) % 432 ca->sb.njournal_buckets; 433 434 atomic_set(&ja->discard_in_flight, DISCARD_READY); 435 /* fallthrough */ 436 437 case DISCARD_READY: 438 if (ja->discard_idx == ja->last_idx) 439 return; 440 441 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT); 442 443 bio_init(bio); 444 bio->bi_sector = bucket_to_sector(ca->set, 445 ca->sb.d[ja->discard_idx]); 446 bio->bi_bdev = ca->bdev; 447 bio->bi_rw = REQ_WRITE|REQ_DISCARD; 448 bio->bi_max_vecs = 1; 449 bio->bi_io_vec = bio->bi_inline_vecs; 450 bio->bi_size = bucket_bytes(ca); 451 bio->bi_end_io = journal_discard_endio; 452 453 closure_get(&ca->set->cl); 454 INIT_WORK(&ja->discard_work, journal_discard_work); 455 schedule_work(&ja->discard_work); 456 } 457} 458 459static void journal_reclaim(struct cache_set *c) 460{ 461 struct bkey *k = &c->journal.key; 462 struct cache *ca; 463 uint64_t last_seq; 464 unsigned iter, n = 0; 465 atomic_t p; 466 467 while (!atomic_read(&fifo_front(&c->journal.pin))) 468 fifo_pop(&c->journal.pin, p); 469 470 last_seq = last_seq(&c->journal); 471 472 /* Update last_idx */ 473 474 for_each_cache(ca, c, iter) { 475 struct journal_device *ja = &ca->journal; 476 477 while (ja->last_idx != ja->cur_idx && 478 ja->seq[ja->last_idx] < last_seq) 479 ja->last_idx = (ja->last_idx + 1) % 480 ca->sb.njournal_buckets; 481 } 482 483 for_each_cache(ca, c, iter) 484 do_journal_discard(ca); 485 486 if (c->journal.blocks_free) 487 goto out; 488 489 /* 490 * Allocate: 491 * XXX: Sort by free journal space 492 */ 493 494 for_each_cache(ca, c, iter) { 495 struct journal_device *ja = &ca->journal; 496 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets; 497 498 /* No space available on this device */ 499 if (next == ja->discard_idx) 500 continue; 501 502 ja->cur_idx = next; 503 k->ptr[n++] = PTR(0, 504 bucket_to_sector(c, ca->sb.d[ja->cur_idx]), 505 ca->sb.nr_this_dev); 506 } 507 508 bkey_init(k); 509 SET_KEY_PTRS(k, n); 510 511 if (n) 512 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits; 513out: 514 if (!journal_full(&c->journal)) 515 __closure_wake_up(&c->journal.wait); 516} 517 518void bch_journal_next(struct journal *j) 519{ 520 atomic_t p = { 1 }; 521 522 j->cur = (j->cur == j->w) 523 ? &j->w[1] 524 : &j->w[0]; 525 526 /* 527 * The fifo_push() needs to happen at the same time as j->seq is 528 * incremented for last_seq() to be calculated correctly 529 */ 530 BUG_ON(!fifo_push(&j->pin, p)); 531 atomic_set(&fifo_back(&j->pin), 1); 532 533 j->cur->data->seq = ++j->seq; 534 j->cur->need_write = false; 535 j->cur->data->keys = 0; 536 537 if (fifo_full(&j->pin)) 538 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin)); 539} 540 541static void journal_write_endio(struct bio *bio, int error) 542{ 543 struct journal_write *w = bio->bi_private; 544 545 cache_set_err_on(error, w->c, "journal io error"); 546 closure_put(&w->c->journal.io); 547} 548 549static void journal_write(struct closure *); 550 551static void journal_write_done(struct closure *cl) 552{ 553 struct journal *j = container_of(cl, struct journal, io); 554 struct journal_write *w = (j->cur == j->w) 555 ? &j->w[1] 556 : &j->w[0]; 557 558 __closure_wake_up(&w->wait); 559 continue_at_nobarrier(cl, journal_write, system_wq); 560} 561 562static void journal_write_unlocked(struct closure *cl) 563 __releases(c->journal.lock) 564{ 565 struct cache_set *c = container_of(cl, struct cache_set, journal.io); 566 struct cache *ca; 567 struct journal_write *w = c->journal.cur; 568 struct bkey *k = &c->journal.key; 569 unsigned i, sectors = set_blocks(w->data, c) * c->sb.block_size; 570 571 struct bio *bio; 572 struct bio_list list; 573 bio_list_init(&list); 574 575 if (!w->need_write) { 576 /* 577 * XXX: have to unlock closure before we unlock journal lock, 578 * else we race with bch_journal(). But this way we race 579 * against cache set unregister. Doh. 580 */ 581 set_closure_fn(cl, NULL, NULL); 582 closure_sub(cl, CLOSURE_RUNNING + 1); 583 spin_unlock(&c->journal.lock); 584 return; 585 } else if (journal_full(&c->journal)) { 586 journal_reclaim(c); 587 spin_unlock(&c->journal.lock); 588 589 btree_flush_write(c); 590 continue_at(cl, journal_write, system_wq); 591 } 592 593 c->journal.blocks_free -= set_blocks(w->data, c); 594 595 w->data->btree_level = c->root->level; 596 597 bkey_copy(&w->data->btree_root, &c->root->key); 598 bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket); 599 600 for_each_cache(ca, c, i) 601 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0]; 602 603 w->data->magic = jset_magic(c); 604 w->data->version = BCACHE_JSET_VERSION; 605 w->data->last_seq = last_seq(&c->journal); 606 w->data->csum = csum_set(w->data); 607 608 for (i = 0; i < KEY_PTRS(k); i++) { 609 ca = PTR_CACHE(c, k, i); 610 bio = &ca->journal.bio; 611 612 atomic_long_add(sectors, &ca->meta_sectors_written); 613 614 bio_reset(bio); 615 bio->bi_sector = PTR_OFFSET(k, i); 616 bio->bi_bdev = ca->bdev; 617 bio->bi_rw = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA; 618 bio->bi_size = sectors << 9; 619 620 bio->bi_end_io = journal_write_endio; 621 bio->bi_private = w; 622 bch_bio_map(bio, w->data); 623 624 trace_bcache_journal_write(bio); 625 bio_list_add(&list, bio); 626 627 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors); 628 629 ca->journal.seq[ca->journal.cur_idx] = w->data->seq; 630 } 631 632 atomic_dec_bug(&fifo_back(&c->journal.pin)); 633 bch_journal_next(&c->journal); 634 journal_reclaim(c); 635 636 spin_unlock(&c->journal.lock); 637 638 while ((bio = bio_list_pop(&list))) 639 closure_bio_submit(bio, cl, c->cache[0]); 640 641 continue_at(cl, journal_write_done, NULL); 642} 643 644static void journal_write(struct closure *cl) 645{ 646 struct cache_set *c = container_of(cl, struct cache_set, journal.io); 647 648 spin_lock(&c->journal.lock); 649 journal_write_unlocked(cl); 650} 651 652static void journal_try_write(struct cache_set *c) 653 __releases(c->journal.lock) 654{ 655 struct closure *cl = &c->journal.io; 656 struct journal_write *w = c->journal.cur; 657 658 w->need_write = true; 659 660 if (closure_trylock(cl, &c->cl)) 661 journal_write_unlocked(cl); 662 else 663 spin_unlock(&c->journal.lock); 664} 665 666static struct journal_write *journal_wait_for_write(struct cache_set *c, 667 unsigned nkeys) 668{ 669 size_t sectors; 670 struct closure cl; 671 672 closure_init_stack(&cl); 673 674 spin_lock(&c->journal.lock); 675 676 while (1) { 677 struct journal_write *w = c->journal.cur; 678 679 sectors = __set_blocks(w->data, w->data->keys + nkeys, 680 c) * c->sb.block_size; 681 682 if (sectors <= min_t(size_t, 683 c->journal.blocks_free * c->sb.block_size, 684 PAGE_SECTORS << JSET_BITS)) 685 return w; 686 687 /* XXX: tracepoint */ 688 if (!journal_full(&c->journal)) { 689 trace_bcache_journal_entry_full(c); 690 691 /* 692 * XXX: If we were inserting so many keys that they 693 * won't fit in an _empty_ journal write, we'll 694 * deadlock. For now, handle this in 695 * bch_keylist_realloc() - but something to think about. 696 */ 697 BUG_ON(!w->data->keys); 698 699 closure_wait(&w->wait, &cl); 700 journal_try_write(c); /* unlocks */ 701 } else { 702 trace_bcache_journal_full(c); 703 704 closure_wait(&c->journal.wait, &cl); 705 journal_reclaim(c); 706 spin_unlock(&c->journal.lock); 707 708 btree_flush_write(c); 709 } 710 711 closure_sync(&cl); 712 spin_lock(&c->journal.lock); 713 } 714} 715 716static void journal_write_work(struct work_struct *work) 717{ 718 struct cache_set *c = container_of(to_delayed_work(work), 719 struct cache_set, 720 journal.work); 721 spin_lock(&c->journal.lock); 722 journal_try_write(c); 723} 724 725/* 726 * Entry point to the journalling code - bio_insert() and btree_invalidate() 727 * pass bch_journal() a list of keys to be journalled, and then 728 * bch_journal() hands those same keys off to btree_insert_async() 729 */ 730 731atomic_t *bch_journal(struct cache_set *c, 732 struct keylist *keys, 733 struct closure *parent) 734{ 735 struct journal_write *w; 736 atomic_t *ret; 737 738 if (!CACHE_SYNC(&c->sb)) 739 return NULL; 740 741 w = journal_wait_for_write(c, bch_keylist_nkeys(keys)); 742 743 memcpy(end(w->data), keys->keys, bch_keylist_bytes(keys)); 744 w->data->keys += bch_keylist_nkeys(keys); 745 746 ret = &fifo_back(&c->journal.pin); 747 atomic_inc(ret); 748 749 if (parent) { 750 closure_wait(&w->wait, parent); 751 journal_try_write(c); 752 } else if (!w->need_write) { 753 schedule_delayed_work(&c->journal.work, 754 msecs_to_jiffies(c->journal_delay_ms)); 755 spin_unlock(&c->journal.lock); 756 } else { 757 spin_unlock(&c->journal.lock); 758 } 759 760 761 return ret; 762} 763 764void bch_journal_meta(struct cache_set *c, struct closure *cl) 765{ 766 struct keylist keys; 767 atomic_t *ref; 768 769 bch_keylist_init(&keys); 770 771 ref = bch_journal(c, &keys, cl); 772 if (ref) 773 atomic_dec_bug(ref); 774} 775 776void bch_journal_free(struct cache_set *c) 777{ 778 free_pages((unsigned long) c->journal.w[1].data, JSET_BITS); 779 free_pages((unsigned long) c->journal.w[0].data, JSET_BITS); 780 free_fifo(&c->journal.pin); 781} 782 783int bch_journal_alloc(struct cache_set *c) 784{ 785 struct journal *j = &c->journal; 786 787 closure_init_unlocked(&j->io); 788 spin_lock_init(&j->lock); 789 INIT_DELAYED_WORK(&j->work, journal_write_work); 790 791 c->journal_delay_ms = 100; 792 793 j->w[0].c = c; 794 j->w[1].c = c; 795 796 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) || 797 !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) || 798 !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS))) 799 return -ENOMEM; 800 801 return 0; 802} 803