1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Main bcache entry point - handle a read or a write request and decide what to 4 * do with it; the make_request functions are called by the block layer. 5 * 6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com> 7 * Copyright 2012 Google, Inc. 8 */ 9 10#include "bcache.h" 11#include "btree.h" 12#include "debug.h" 13#include "request.h" 14#include "writeback.h" 15 16#include <linux/module.h> 17#include <linux/hash.h> 18#include <linux/random.h> 19#include <linux/backing-dev.h> 20 21#include <trace/events/bcache.h> 22 23#define CUTOFF_CACHE_ADD 95 24#define CUTOFF_CACHE_READA 90 25 26struct kmem_cache *bch_search_cache; 27 28static CLOSURE_CALLBACK(bch_data_insert_start); 29 30static unsigned int cache_mode(struct cached_dev *dc) 31{ 32 return BDEV_CACHE_MODE(&dc->sb); 33} 34 35static bool verify(struct cached_dev *dc) 36{ 37 return dc->verify; 38} 39 40static void bio_csum(struct bio *bio, struct bkey *k) 41{ 42 struct bio_vec bv; 43 struct bvec_iter iter; 44 uint64_t csum = 0; 45 46 bio_for_each_segment(bv, bio, iter) { 47 void *d = bvec_kmap_local(&bv); 48 49 csum = crc64_be(csum, d, bv.bv_len); 50 kunmap_local(d); 51 } 52 53 k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1); 54} 55 56/* Insert data into cache */ 57 58static CLOSURE_CALLBACK(bch_data_insert_keys) 59{ 60 closure_type(op, struct data_insert_op, cl); 61 atomic_t *journal_ref = NULL; 62 struct bkey *replace_key = op->replace ? &op->replace_key : NULL; 63 int ret; 64 65 if (!op->replace) 66 journal_ref = bch_journal(op->c, &op->insert_keys, 67 op->flush_journal ? cl : NULL); 68 69 ret = bch_btree_insert(op->c, &op->insert_keys, 70 journal_ref, replace_key); 71 if (ret == -ESRCH) { 72 op->replace_collision = true; 73 } else if (ret) { 74 op->status = BLK_STS_RESOURCE; 75 op->insert_data_done = true; 76 } 77 78 if (journal_ref) 79 atomic_dec_bug(journal_ref); 80 81 if (!op->insert_data_done) { 82 continue_at(cl, bch_data_insert_start, op->wq); 83 return; 84 } 85 86 bch_keylist_free(&op->insert_keys); 87 closure_return(cl); 88} 89 90static int bch_keylist_realloc(struct keylist *l, unsigned int u64s, 91 struct cache_set *c) 92{ 93 size_t oldsize = bch_keylist_nkeys(l); 94 size_t newsize = oldsize + u64s; 95 96 /* 97 * The journalling code doesn't handle the case where the keys to insert 98 * is bigger than an empty write: If we just return -ENOMEM here, 99 * bch_data_insert_keys() will insert the keys created so far 100 * and finish the rest when the keylist is empty. 101 */ 102 if (newsize * sizeof(uint64_t) > block_bytes(c->cache) - sizeof(struct jset)) 103 return -ENOMEM; 104 105 return __bch_keylist_realloc(l, u64s); 106} 107 108static void bch_data_invalidate(struct closure *cl) 109{ 110 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); 111 struct bio *bio = op->bio; 112 113 pr_debug("invalidating %i sectors from %llu\n", 114 bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector); 115 116 while (bio_sectors(bio)) { 117 unsigned int sectors = min(bio_sectors(bio), 118 1U << (KEY_SIZE_BITS - 1)); 119 120 if (bch_keylist_realloc(&op->insert_keys, 2, op->c)) 121 goto out; 122 123 bio->bi_iter.bi_sector += sectors; 124 bio->bi_iter.bi_size -= sectors << 9; 125 126 bch_keylist_add(&op->insert_keys, 127 &KEY(op->inode, 128 bio->bi_iter.bi_sector, 129 sectors)); 130 } 131 132 op->insert_data_done = true; 133 /* get in bch_data_insert() */ 134 bio_put(bio); 135out: 136 continue_at(cl, bch_data_insert_keys, op->wq); 137} 138 139static CLOSURE_CALLBACK(bch_data_insert_error) 140{ 141 closure_type(op, struct data_insert_op, cl); 142 143 /* 144 * Our data write just errored, which means we've got a bunch of keys to 145 * insert that point to data that wasn't successfully written. 146 * 147 * We don't have to insert those keys but we still have to invalidate 148 * that region of the cache - so, if we just strip off all the pointers 149 * from the keys we'll accomplish just that. 150 */ 151 152 struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys; 153 154 while (src != op->insert_keys.top) { 155 struct bkey *n = bkey_next(src); 156 157 SET_KEY_PTRS(src, 0); 158 memmove(dst, src, bkey_bytes(src)); 159 160 dst = bkey_next(dst); 161 src = n; 162 } 163 164 op->insert_keys.top = dst; 165 166 bch_data_insert_keys(&cl->work); 167} 168 169static void bch_data_insert_endio(struct bio *bio) 170{ 171 struct closure *cl = bio->bi_private; 172 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); 173 174 if (bio->bi_status) { 175 /* TODO: We could try to recover from this. */ 176 if (op->writeback) 177 op->status = bio->bi_status; 178 else if (!op->replace) 179 set_closure_fn(cl, bch_data_insert_error, op->wq); 180 else 181 set_closure_fn(cl, NULL, NULL); 182 } 183 184 bch_bbio_endio(op->c, bio, bio->bi_status, "writing data to cache"); 185} 186 187static CLOSURE_CALLBACK(bch_data_insert_start) 188{ 189 closure_type(op, struct data_insert_op, cl); 190 struct bio *bio = op->bio, *n; 191 192 if (op->bypass) 193 return bch_data_invalidate(cl); 194 195 if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) 196 wake_up_gc(op->c); 197 198 /* 199 * Journal writes are marked REQ_PREFLUSH; if the original write was a 200 * flush, it'll wait on the journal write. 201 */ 202 bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA); 203 204 do { 205 unsigned int i; 206 struct bkey *k; 207 struct bio_set *split = &op->c->bio_split; 208 209 /* 1 for the device pointer and 1 for the chksum */ 210 if (bch_keylist_realloc(&op->insert_keys, 211 3 + (op->csum ? 1 : 0), 212 op->c)) { 213 continue_at(cl, bch_data_insert_keys, op->wq); 214 return; 215 } 216 217 k = op->insert_keys.top; 218 bkey_init(k); 219 SET_KEY_INODE(k, op->inode); 220 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector); 221 222 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio), 223 op->write_point, op->write_prio, 224 op->writeback)) 225 goto err; 226 227 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split); 228 229 n->bi_end_io = bch_data_insert_endio; 230 n->bi_private = cl; 231 232 if (op->writeback) { 233 SET_KEY_DIRTY(k, true); 234 235 for (i = 0; i < KEY_PTRS(k); i++) 236 SET_GC_MARK(PTR_BUCKET(op->c, k, i), 237 GC_MARK_DIRTY); 238 } 239 240 SET_KEY_CSUM(k, op->csum); 241 if (KEY_CSUM(k)) 242 bio_csum(n, k); 243 244 trace_bcache_cache_insert(k); 245 bch_keylist_push(&op->insert_keys); 246 247 n->bi_opf = REQ_OP_WRITE; 248 bch_submit_bbio(n, op->c, k, 0); 249 } while (n != bio); 250 251 op->insert_data_done = true; 252 continue_at(cl, bch_data_insert_keys, op->wq); 253 return; 254err: 255 /* bch_alloc_sectors() blocks if s->writeback = true */ 256 BUG_ON(op->writeback); 257 258 /* 259 * But if it's not a writeback write we'd rather just bail out if 260 * there aren't any buckets ready to write to - it might take awhile and 261 * we might be starving btree writes for gc or something. 262 */ 263 264 if (!op->replace) { 265 /* 266 * Writethrough write: We can't complete the write until we've 267 * updated the index. But we don't want to delay the write while 268 * we wait for buckets to be freed up, so just invalidate the 269 * rest of the write. 270 */ 271 op->bypass = true; 272 return bch_data_invalidate(cl); 273 } else { 274 /* 275 * From a cache miss, we can just insert the keys for the data 276 * we have written or bail out if we didn't do anything. 277 */ 278 op->insert_data_done = true; 279 bio_put(bio); 280 281 if (!bch_keylist_empty(&op->insert_keys)) 282 continue_at(cl, bch_data_insert_keys, op->wq); 283 else 284 closure_return(cl); 285 } 286} 287 288/** 289 * bch_data_insert - stick some data in the cache 290 * @cl: closure pointer. 291 * 292 * This is the starting point for any data to end up in a cache device; it could 293 * be from a normal write, or a writeback write, or a write to a flash only 294 * volume - it's also used by the moving garbage collector to compact data in 295 * mostly empty buckets. 296 * 297 * It first writes the data to the cache, creating a list of keys to be inserted 298 * (if the data had to be fragmented there will be multiple keys); after the 299 * data is written it calls bch_journal, and after the keys have been added to 300 * the next journal write they're inserted into the btree. 301 * 302 * It inserts the data in op->bio; bi_sector is used for the key offset, 303 * and op->inode is used for the key inode. 304 * 305 * If op->bypass is true, instead of inserting the data it invalidates the 306 * region of the cache represented by op->bio and op->inode. 307 */ 308CLOSURE_CALLBACK(bch_data_insert) 309{ 310 closure_type(op, struct data_insert_op, cl); 311 312 trace_bcache_write(op->c, op->inode, op->bio, 313 op->writeback, op->bypass); 314 315 bch_keylist_init(&op->insert_keys); 316 bio_get(op->bio); 317 bch_data_insert_start(&cl->work); 318} 319 320/* 321 * Congested? Return 0 (not congested) or the limit (in sectors) 322 * beyond which we should bypass the cache due to congestion. 323 */ 324unsigned int bch_get_congested(const struct cache_set *c) 325{ 326 int i; 327 328 if (!c->congested_read_threshold_us && 329 !c->congested_write_threshold_us) 330 return 0; 331 332 i = (local_clock_us() - c->congested_last_us) / 1024; 333 if (i < 0) 334 return 0; 335 336 i += atomic_read(&c->congested); 337 if (i >= 0) 338 return 0; 339 340 i += CONGESTED_MAX; 341 342 if (i > 0) 343 i = fract_exp_two(i, 6); 344 345 i -= hweight32(get_random_u32()); 346 347 return i > 0 ? i : 1; 348} 349 350static void add_sequential(struct task_struct *t) 351{ 352 ewma_add(t->sequential_io_avg, 353 t->sequential_io, 8, 0); 354 355 t->sequential_io = 0; 356} 357 358static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k) 359{ 360 return &dc->io_hash[hash_64(k, RECENT_IO_BITS)]; 361} 362 363static bool check_should_bypass(struct cached_dev *dc, struct bio *bio) 364{ 365 struct cache_set *c = dc->disk.c; 366 unsigned int mode = cache_mode(dc); 367 unsigned int sectors, congested; 368 struct task_struct *task = current; 369 struct io *i; 370 371 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) || 372 c->gc_stats.in_use > CUTOFF_CACHE_ADD || 373 (bio_op(bio) == REQ_OP_DISCARD)) 374 goto skip; 375 376 if (mode == CACHE_MODE_NONE || 377 (mode == CACHE_MODE_WRITEAROUND && 378 op_is_write(bio_op(bio)))) 379 goto skip; 380 381 /* 382 * If the bio is for read-ahead or background IO, bypass it or 383 * not depends on the following situations, 384 * - If the IO is for meta data, always cache it and no bypass 385 * - If the IO is not meta data, check dc->cache_reada_policy, 386 * BCH_CACHE_READA_ALL: cache it and not bypass 387 * BCH_CACHE_READA_META_ONLY: not cache it and bypass 388 * That is, read-ahead request for metadata always get cached 389 * (eg, for gfs2 or xfs). 390 */ 391 if ((bio->bi_opf & (REQ_RAHEAD|REQ_BACKGROUND))) { 392 if (!(bio->bi_opf & (REQ_META|REQ_PRIO)) && 393 (dc->cache_readahead_policy != BCH_CACHE_READA_ALL)) 394 goto skip; 395 } 396 397 if (bio->bi_iter.bi_sector & (c->cache->sb.block_size - 1) || 398 bio_sectors(bio) & (c->cache->sb.block_size - 1)) { 399 pr_debug("skipping unaligned io\n"); 400 goto skip; 401 } 402 403 if (bypass_torture_test(dc)) { 404 if (get_random_u32_below(4) == 3) 405 goto skip; 406 else 407 goto rescale; 408 } 409 410 congested = bch_get_congested(c); 411 if (!congested && !dc->sequential_cutoff) 412 goto rescale; 413 414 spin_lock(&dc->io_lock); 415 416 hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash) 417 if (i->last == bio->bi_iter.bi_sector && 418 time_before(jiffies, i->jiffies)) 419 goto found; 420 421 i = list_first_entry(&dc->io_lru, struct io, lru); 422 423 add_sequential(task); 424 i->sequential = 0; 425found: 426 if (i->sequential + bio->bi_iter.bi_size > i->sequential) 427 i->sequential += bio->bi_iter.bi_size; 428 429 i->last = bio_end_sector(bio); 430 i->jiffies = jiffies + msecs_to_jiffies(5000); 431 task->sequential_io = i->sequential; 432 433 hlist_del(&i->hash); 434 hlist_add_head(&i->hash, iohash(dc, i->last)); 435 list_move_tail(&i->lru, &dc->io_lru); 436 437 spin_unlock(&dc->io_lock); 438 439 sectors = max(task->sequential_io, 440 task->sequential_io_avg) >> 9; 441 442 if (dc->sequential_cutoff && 443 sectors >= dc->sequential_cutoff >> 9) { 444 trace_bcache_bypass_sequential(bio); 445 goto skip; 446 } 447 448 if (congested && sectors >= congested) { 449 trace_bcache_bypass_congested(bio); 450 goto skip; 451 } 452 453rescale: 454 bch_rescale_priorities(c, bio_sectors(bio)); 455 return false; 456skip: 457 bch_mark_sectors_bypassed(c, dc, bio_sectors(bio)); 458 return true; 459} 460 461/* Cache lookup */ 462 463struct search { 464 /* Stack frame for bio_complete */ 465 struct closure cl; 466 467 struct bbio bio; 468 struct bio *orig_bio; 469 struct bio *cache_miss; 470 struct bcache_device *d; 471 472 unsigned int insert_bio_sectors; 473 unsigned int recoverable:1; 474 unsigned int write:1; 475 unsigned int read_dirty_data:1; 476 unsigned int cache_missed:1; 477 478 struct block_device *orig_bdev; 479 unsigned long start_time; 480 481 struct btree_op op; 482 struct data_insert_op iop; 483}; 484 485static void bch_cache_read_endio(struct bio *bio) 486{ 487 struct bbio *b = container_of(bio, struct bbio, bio); 488 struct closure *cl = bio->bi_private; 489 struct search *s = container_of(cl, struct search, cl); 490 491 /* 492 * If the bucket was reused while our bio was in flight, we might have 493 * read the wrong data. Set s->error but not error so it doesn't get 494 * counted against the cache device, but we'll still reread the data 495 * from the backing device. 496 */ 497 498 if (bio->bi_status) 499 s->iop.status = bio->bi_status; 500 else if (!KEY_DIRTY(&b->key) && 501 ptr_stale(s->iop.c, &b->key, 0)) { 502 atomic_long_inc(&s->iop.c->cache_read_races); 503 s->iop.status = BLK_STS_IOERR; 504 } 505 506 bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache"); 507} 508 509/* 510 * Read from a single key, handling the initial cache miss if the key starts in 511 * the middle of the bio 512 */ 513static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k) 514{ 515 struct search *s = container_of(op, struct search, op); 516 struct bio *n, *bio = &s->bio.bio; 517 struct bkey *bio_key; 518 unsigned int ptr; 519 520 if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0) 521 return MAP_CONTINUE; 522 523 if (KEY_INODE(k) != s->iop.inode || 524 KEY_START(k) > bio->bi_iter.bi_sector) { 525 unsigned int bio_sectors = bio_sectors(bio); 526 unsigned int sectors = KEY_INODE(k) == s->iop.inode 527 ? min_t(uint64_t, INT_MAX, 528 KEY_START(k) - bio->bi_iter.bi_sector) 529 : INT_MAX; 530 int ret = s->d->cache_miss(b, s, bio, sectors); 531 532 if (ret != MAP_CONTINUE) 533 return ret; 534 535 /* if this was a complete miss we shouldn't get here */ 536 BUG_ON(bio_sectors <= sectors); 537 } 538 539 if (!KEY_SIZE(k)) 540 return MAP_CONTINUE; 541 542 /* XXX: figure out best pointer - for multiple cache devices */ 543 ptr = 0; 544 545 PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO; 546 547 if (KEY_DIRTY(k)) 548 s->read_dirty_data = true; 549 550 n = bio_next_split(bio, min_t(uint64_t, INT_MAX, 551 KEY_OFFSET(k) - bio->bi_iter.bi_sector), 552 GFP_NOIO, &s->d->bio_split); 553 554 bio_key = &container_of(n, struct bbio, bio)->key; 555 bch_bkey_copy_single_ptr(bio_key, k, ptr); 556 557 bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key); 558 bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key); 559 560 n->bi_end_io = bch_cache_read_endio; 561 n->bi_private = &s->cl; 562 563 /* 564 * The bucket we're reading from might be reused while our bio 565 * is in flight, and we could then end up reading the wrong 566 * data. 567 * 568 * We guard against this by checking (in cache_read_endio()) if 569 * the pointer is stale again; if so, we treat it as an error 570 * and reread from the backing device (but we don't pass that 571 * error up anywhere). 572 */ 573 574 __bch_submit_bbio(n, b->c); 575 return n == bio ? MAP_DONE : MAP_CONTINUE; 576} 577 578static CLOSURE_CALLBACK(cache_lookup) 579{ 580 closure_type(s, struct search, iop.cl); 581 struct bio *bio = &s->bio.bio; 582 struct cached_dev *dc; 583 int ret; 584 585 bch_btree_op_init(&s->op, -1); 586 587 ret = bch_btree_map_keys(&s->op, s->iop.c, 588 &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0), 589 cache_lookup_fn, MAP_END_KEY); 590 if (ret == -EAGAIN) { 591 continue_at(cl, cache_lookup, bcache_wq); 592 return; 593 } 594 595 /* 596 * We might meet err when searching the btree, If that happens, we will 597 * get negative ret, in this scenario we should not recover data from 598 * backing device (when cache device is dirty) because we don't know 599 * whether bkeys the read request covered are all clean. 600 * 601 * And after that happened, s->iop.status is still its initial value 602 * before we submit s->bio.bio 603 */ 604 if (ret < 0) { 605 BUG_ON(ret == -EINTR); 606 if (s->d && s->d->c && 607 !UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) { 608 dc = container_of(s->d, struct cached_dev, disk); 609 if (dc && atomic_read(&dc->has_dirty)) 610 s->recoverable = false; 611 } 612 if (!s->iop.status) 613 s->iop.status = BLK_STS_IOERR; 614 } 615 616 closure_return(cl); 617} 618 619/* Common code for the make_request functions */ 620 621static void request_endio(struct bio *bio) 622{ 623 struct closure *cl = bio->bi_private; 624 625 if (bio->bi_status) { 626 struct search *s = container_of(cl, struct search, cl); 627 628 s->iop.status = bio->bi_status; 629 /* Only cache read errors are recoverable */ 630 s->recoverable = false; 631 } 632 633 bio_put(bio); 634 closure_put(cl); 635} 636 637static void backing_request_endio(struct bio *bio) 638{ 639 struct closure *cl = bio->bi_private; 640 641 if (bio->bi_status) { 642 struct search *s = container_of(cl, struct search, cl); 643 struct cached_dev *dc = container_of(s->d, 644 struct cached_dev, disk); 645 /* 646 * If a bio has REQ_PREFLUSH for writeback mode, it is 647 * speically assembled in cached_dev_write() for a non-zero 648 * write request which has REQ_PREFLUSH. we don't set 649 * s->iop.status by this failure, the status will be decided 650 * by result of bch_data_insert() operation. 651 */ 652 if (unlikely(s->iop.writeback && 653 bio->bi_opf & REQ_PREFLUSH)) { 654 pr_err("Can't flush %pg: returned bi_status %i\n", 655 dc->bdev, bio->bi_status); 656 } else { 657 /* set to orig_bio->bi_status in bio_complete() */ 658 s->iop.status = bio->bi_status; 659 } 660 s->recoverable = false; 661 /* should count I/O error for backing device here */ 662 bch_count_backing_io_errors(dc, bio); 663 } 664 665 bio_put(bio); 666 closure_put(cl); 667} 668 669static void bio_complete(struct search *s) 670{ 671 if (s->orig_bio) { 672 /* Count on bcache device */ 673 bio_end_io_acct_remapped(s->orig_bio, s->start_time, 674 s->orig_bdev); 675 trace_bcache_request_end(s->d, s->orig_bio); 676 s->orig_bio->bi_status = s->iop.status; 677 bio_endio(s->orig_bio); 678 s->orig_bio = NULL; 679 } 680} 681 682static void do_bio_hook(struct search *s, 683 struct bio *orig_bio, 684 bio_end_io_t *end_io_fn) 685{ 686 struct bio *bio = &s->bio.bio; 687 688 bio_init_clone(orig_bio->bi_bdev, bio, orig_bio, GFP_NOIO); 689 /* 690 * bi_end_io can be set separately somewhere else, e.g. the 691 * variants in, 692 * - cache_bio->bi_end_io from cached_dev_cache_miss() 693 * - n->bi_end_io from cache_lookup_fn() 694 */ 695 bio->bi_end_io = end_io_fn; 696 bio->bi_private = &s->cl; 697 698 bio_cnt_set(bio, 3); 699} 700 701static CLOSURE_CALLBACK(search_free) 702{ 703 closure_type(s, struct search, cl); 704 705 atomic_dec(&s->iop.c->search_inflight); 706 707 if (s->iop.bio) 708 bio_put(s->iop.bio); 709 710 bio_complete(s); 711 closure_debug_destroy(cl); 712 mempool_free(s, &s->iop.c->search); 713} 714 715static inline struct search *search_alloc(struct bio *bio, 716 struct bcache_device *d, struct block_device *orig_bdev, 717 unsigned long start_time) 718{ 719 struct search *s; 720 721 s = mempool_alloc(&d->c->search, GFP_NOIO); 722 723 closure_init(&s->cl, NULL); 724 do_bio_hook(s, bio, request_endio); 725 atomic_inc(&d->c->search_inflight); 726 727 s->orig_bio = bio; 728 s->cache_miss = NULL; 729 s->cache_missed = 0; 730 s->d = d; 731 s->recoverable = 1; 732 s->write = op_is_write(bio_op(bio)); 733 s->read_dirty_data = 0; 734 /* Count on the bcache device */ 735 s->orig_bdev = orig_bdev; 736 s->start_time = start_time; 737 s->iop.c = d->c; 738 s->iop.bio = NULL; 739 s->iop.inode = d->id; 740 s->iop.write_point = hash_long((unsigned long) current, 16); 741 s->iop.write_prio = 0; 742 s->iop.status = 0; 743 s->iop.flags = 0; 744 s->iop.flush_journal = op_is_flush(bio->bi_opf); 745 s->iop.wq = bcache_wq; 746 747 return s; 748} 749 750/* Cached devices */ 751 752static CLOSURE_CALLBACK(cached_dev_bio_complete) 753{ 754 closure_type(s, struct search, cl); 755 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 756 757 cached_dev_put(dc); 758 search_free(&cl->work); 759} 760 761/* Process reads */ 762 763static CLOSURE_CALLBACK(cached_dev_read_error_done) 764{ 765 closure_type(s, struct search, cl); 766 767 if (s->iop.replace_collision) 768 bch_mark_cache_miss_collision(s->iop.c, s->d); 769 770 if (s->iop.bio) 771 bio_free_pages(s->iop.bio); 772 773 cached_dev_bio_complete(&cl->work); 774} 775 776static CLOSURE_CALLBACK(cached_dev_read_error) 777{ 778 closure_type(s, struct search, cl); 779 struct bio *bio = &s->bio.bio; 780 781 /* 782 * If read request hit dirty data (s->read_dirty_data is true), 783 * then recovery a failed read request from cached device may 784 * get a stale data back. So read failure recovery is only 785 * permitted when read request hit clean data in cache device, 786 * or when cache read race happened. 787 */ 788 if (s->recoverable && !s->read_dirty_data) { 789 /* Retry from the backing device: */ 790 trace_bcache_read_retry(s->orig_bio); 791 792 s->iop.status = 0; 793 do_bio_hook(s, s->orig_bio, backing_request_endio); 794 795 /* XXX: invalidate cache */ 796 797 /* I/O request sent to backing device */ 798 closure_bio_submit(s->iop.c, bio, cl); 799 } 800 801 continue_at(cl, cached_dev_read_error_done, NULL); 802} 803 804static CLOSURE_CALLBACK(cached_dev_cache_miss_done) 805{ 806 closure_type(s, struct search, cl); 807 struct bcache_device *d = s->d; 808 809 if (s->iop.replace_collision) 810 bch_mark_cache_miss_collision(s->iop.c, s->d); 811 812 if (s->iop.bio) 813 bio_free_pages(s->iop.bio); 814 815 cached_dev_bio_complete(&cl->work); 816 closure_put(&d->cl); 817} 818 819static CLOSURE_CALLBACK(cached_dev_read_done) 820{ 821 closure_type(s, struct search, cl); 822 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 823 824 /* 825 * We had a cache miss; cache_bio now contains data ready to be inserted 826 * into the cache. 827 * 828 * First, we copy the data we just read from cache_bio's bounce buffers 829 * to the buffers the original bio pointed to: 830 */ 831 832 if (s->iop.bio) { 833 bio_reset(s->iop.bio, s->cache_miss->bi_bdev, REQ_OP_READ); 834 s->iop.bio->bi_iter.bi_sector = 835 s->cache_miss->bi_iter.bi_sector; 836 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9; 837 bio_clone_blkg_association(s->iop.bio, s->cache_miss); 838 bch_bio_map(s->iop.bio, NULL); 839 840 bio_copy_data(s->cache_miss, s->iop.bio); 841 842 bio_put(s->cache_miss); 843 s->cache_miss = NULL; 844 } 845 846 if (verify(dc) && s->recoverable && !s->read_dirty_data) 847 bch_data_verify(dc, s->orig_bio); 848 849 closure_get(&dc->disk.cl); 850 bio_complete(s); 851 852 if (s->iop.bio && 853 !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) { 854 BUG_ON(!s->iop.replace); 855 closure_call(&s->iop.cl, bch_data_insert, NULL, cl); 856 } 857 858 continue_at(cl, cached_dev_cache_miss_done, NULL); 859} 860 861static CLOSURE_CALLBACK(cached_dev_read_done_bh) 862{ 863 closure_type(s, struct search, cl); 864 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 865 866 bch_mark_cache_accounting(s->iop.c, s->d, 867 !s->cache_missed, s->iop.bypass); 868 trace_bcache_read(s->orig_bio, !s->cache_missed, s->iop.bypass); 869 870 if (s->iop.status) 871 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq); 872 else if (s->iop.bio || verify(dc)) 873 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq); 874 else 875 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL); 876} 877 878static int cached_dev_cache_miss(struct btree *b, struct search *s, 879 struct bio *bio, unsigned int sectors) 880{ 881 int ret = MAP_CONTINUE; 882 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 883 struct bio *miss, *cache_bio; 884 unsigned int size_limit; 885 886 s->cache_missed = 1; 887 888 if (s->cache_miss || s->iop.bypass) { 889 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split); 890 ret = miss == bio ? MAP_DONE : MAP_CONTINUE; 891 goto out_submit; 892 } 893 894 /* Limitation for valid replace key size and cache_bio bvecs number */ 895 size_limit = min_t(unsigned int, BIO_MAX_VECS * PAGE_SECTORS, 896 (1 << KEY_SIZE_BITS) - 1); 897 s->insert_bio_sectors = min3(size_limit, sectors, bio_sectors(bio)); 898 899 s->iop.replace_key = KEY(s->iop.inode, 900 bio->bi_iter.bi_sector + s->insert_bio_sectors, 901 s->insert_bio_sectors); 902 903 ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key); 904 if (ret) 905 return ret; 906 907 s->iop.replace = true; 908 909 miss = bio_next_split(bio, s->insert_bio_sectors, GFP_NOIO, 910 &s->d->bio_split); 911 912 /* btree_search_recurse()'s btree iterator is no good anymore */ 913 ret = miss == bio ? MAP_DONE : -EINTR; 914 915 cache_bio = bio_alloc_bioset(miss->bi_bdev, 916 DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS), 917 0, GFP_NOWAIT, &dc->disk.bio_split); 918 if (!cache_bio) 919 goto out_submit; 920 921 cache_bio->bi_iter.bi_sector = miss->bi_iter.bi_sector; 922 cache_bio->bi_iter.bi_size = s->insert_bio_sectors << 9; 923 924 cache_bio->bi_end_io = backing_request_endio; 925 cache_bio->bi_private = &s->cl; 926 927 bch_bio_map(cache_bio, NULL); 928 if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO)) 929 goto out_put; 930 931 s->cache_miss = miss; 932 s->iop.bio = cache_bio; 933 bio_get(cache_bio); 934 /* I/O request sent to backing device */ 935 closure_bio_submit(s->iop.c, cache_bio, &s->cl); 936 937 return ret; 938out_put: 939 bio_put(cache_bio); 940out_submit: 941 miss->bi_end_io = backing_request_endio; 942 miss->bi_private = &s->cl; 943 /* I/O request sent to backing device */ 944 closure_bio_submit(s->iop.c, miss, &s->cl); 945 return ret; 946} 947 948static void cached_dev_read(struct cached_dev *dc, struct search *s) 949{ 950 struct closure *cl = &s->cl; 951 952 closure_call(&s->iop.cl, cache_lookup, NULL, cl); 953 continue_at(cl, cached_dev_read_done_bh, NULL); 954} 955 956/* Process writes */ 957 958static CLOSURE_CALLBACK(cached_dev_write_complete) 959{ 960 closure_type(s, struct search, cl); 961 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 962 963 up_read_non_owner(&dc->writeback_lock); 964 cached_dev_bio_complete(&cl->work); 965} 966 967static void cached_dev_write(struct cached_dev *dc, struct search *s) 968{ 969 struct closure *cl = &s->cl; 970 struct bio *bio = &s->bio.bio; 971 struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0); 972 struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0); 973 974 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end); 975 976 down_read_non_owner(&dc->writeback_lock); 977 if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) { 978 /* 979 * We overlap with some dirty data undergoing background 980 * writeback, force this write to writeback 981 */ 982 s->iop.bypass = false; 983 s->iop.writeback = true; 984 } 985 986 /* 987 * Discards aren't _required_ to do anything, so skipping if 988 * check_overlapping returned true is ok 989 * 990 * But check_overlapping drops dirty keys for which io hasn't started, 991 * so we still want to call it. 992 */ 993 if (bio_op(bio) == REQ_OP_DISCARD) 994 s->iop.bypass = true; 995 996 if (should_writeback(dc, s->orig_bio, 997 cache_mode(dc), 998 s->iop.bypass)) { 999 s->iop.bypass = false; 1000 s->iop.writeback = true; 1001 } 1002 1003 if (s->iop.bypass) { 1004 s->iop.bio = s->orig_bio; 1005 bio_get(s->iop.bio); 1006 1007 if (bio_op(bio) == REQ_OP_DISCARD && 1008 !bdev_max_discard_sectors(dc->bdev)) 1009 goto insert_data; 1010 1011 /* I/O request sent to backing device */ 1012 bio->bi_end_io = backing_request_endio; 1013 closure_bio_submit(s->iop.c, bio, cl); 1014 1015 } else if (s->iop.writeback) { 1016 bch_writeback_add(dc); 1017 s->iop.bio = bio; 1018 1019 if (bio->bi_opf & REQ_PREFLUSH) { 1020 /* 1021 * Also need to send a flush to the backing 1022 * device. 1023 */ 1024 struct bio *flush; 1025 1026 flush = bio_alloc_bioset(bio->bi_bdev, 0, 1027 REQ_OP_WRITE | REQ_PREFLUSH, 1028 GFP_NOIO, &dc->disk.bio_split); 1029 if (!flush) { 1030 s->iop.status = BLK_STS_RESOURCE; 1031 goto insert_data; 1032 } 1033 flush->bi_end_io = backing_request_endio; 1034 flush->bi_private = cl; 1035 /* I/O request sent to backing device */ 1036 closure_bio_submit(s->iop.c, flush, cl); 1037 } 1038 } else { 1039 s->iop.bio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO, 1040 &dc->disk.bio_split); 1041 /* I/O request sent to backing device */ 1042 bio->bi_end_io = backing_request_endio; 1043 closure_bio_submit(s->iop.c, bio, cl); 1044 } 1045 1046insert_data: 1047 closure_call(&s->iop.cl, bch_data_insert, NULL, cl); 1048 continue_at(cl, cached_dev_write_complete, NULL); 1049} 1050 1051static CLOSURE_CALLBACK(cached_dev_nodata) 1052{ 1053 closure_type(s, struct search, cl); 1054 struct bio *bio = &s->bio.bio; 1055 1056 if (s->iop.flush_journal) 1057 bch_journal_meta(s->iop.c, cl); 1058 1059 /* If it's a flush, we send the flush to the backing device too */ 1060 bio->bi_end_io = backing_request_endio; 1061 closure_bio_submit(s->iop.c, bio, cl); 1062 1063 continue_at(cl, cached_dev_bio_complete, NULL); 1064} 1065 1066struct detached_dev_io_private { 1067 struct bcache_device *d; 1068 unsigned long start_time; 1069 bio_end_io_t *bi_end_io; 1070 void *bi_private; 1071 struct block_device *orig_bdev; 1072}; 1073 1074static void detached_dev_end_io(struct bio *bio) 1075{ 1076 struct detached_dev_io_private *ddip; 1077 1078 ddip = bio->bi_private; 1079 bio->bi_end_io = ddip->bi_end_io; 1080 bio->bi_private = ddip->bi_private; 1081 1082 /* Count on the bcache device */ 1083 bio_end_io_acct_remapped(bio, ddip->start_time, ddip->orig_bdev); 1084 1085 if (bio->bi_status) { 1086 struct cached_dev *dc = container_of(ddip->d, 1087 struct cached_dev, disk); 1088 /* should count I/O error for backing device here */ 1089 bch_count_backing_io_errors(dc, bio); 1090 } 1091 1092 kfree(ddip); 1093 bio->bi_end_io(bio); 1094} 1095 1096static void detached_dev_do_request(struct bcache_device *d, struct bio *bio, 1097 struct block_device *orig_bdev, unsigned long start_time) 1098{ 1099 struct detached_dev_io_private *ddip; 1100 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1101 1102 /* 1103 * no need to call closure_get(&dc->disk.cl), 1104 * because upper layer had already opened bcache device, 1105 * which would call closure_get(&dc->disk.cl) 1106 */ 1107 ddip = kzalloc(sizeof(struct detached_dev_io_private), GFP_NOIO); 1108 if (!ddip) { 1109 bio->bi_status = BLK_STS_RESOURCE; 1110 bio->bi_end_io(bio); 1111 return; 1112 } 1113 1114 ddip->d = d; 1115 /* Count on the bcache device */ 1116 ddip->orig_bdev = orig_bdev; 1117 ddip->start_time = start_time; 1118 ddip->bi_end_io = bio->bi_end_io; 1119 ddip->bi_private = bio->bi_private; 1120 bio->bi_end_io = detached_dev_end_io; 1121 bio->bi_private = ddip; 1122 1123 if ((bio_op(bio) == REQ_OP_DISCARD) && 1124 !bdev_max_discard_sectors(dc->bdev)) 1125 bio->bi_end_io(bio); 1126 else 1127 submit_bio_noacct(bio); 1128} 1129 1130static void quit_max_writeback_rate(struct cache_set *c, 1131 struct cached_dev *this_dc) 1132{ 1133 int i; 1134 struct bcache_device *d; 1135 struct cached_dev *dc; 1136 1137 /* 1138 * mutex bch_register_lock may compete with other parallel requesters, 1139 * or attach/detach operations on other backing device. Waiting to 1140 * the mutex lock may increase I/O request latency for seconds or more. 1141 * To avoid such situation, if mutext_trylock() failed, only writeback 1142 * rate of current cached device is set to 1, and __update_write_back() 1143 * will decide writeback rate of other cached devices (remember now 1144 * c->idle_counter is 0 already). 1145 */ 1146 if (mutex_trylock(&bch_register_lock)) { 1147 for (i = 0; i < c->devices_max_used; i++) { 1148 if (!c->devices[i]) 1149 continue; 1150 1151 if (UUID_FLASH_ONLY(&c->uuids[i])) 1152 continue; 1153 1154 d = c->devices[i]; 1155 dc = container_of(d, struct cached_dev, disk); 1156 /* 1157 * set writeback rate to default minimum value, 1158 * then let update_writeback_rate() to decide the 1159 * upcoming rate. 1160 */ 1161 atomic_long_set(&dc->writeback_rate.rate, 1); 1162 } 1163 mutex_unlock(&bch_register_lock); 1164 } else 1165 atomic_long_set(&this_dc->writeback_rate.rate, 1); 1166} 1167 1168/* Cached devices - read & write stuff */ 1169 1170void cached_dev_submit_bio(struct bio *bio) 1171{ 1172 struct search *s; 1173 struct block_device *orig_bdev = bio->bi_bdev; 1174 struct bcache_device *d = orig_bdev->bd_disk->private_data; 1175 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1176 unsigned long start_time; 1177 int rw = bio_data_dir(bio); 1178 1179 if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags)) || 1180 dc->io_disable)) { 1181 bio->bi_status = BLK_STS_IOERR; 1182 bio_endio(bio); 1183 return; 1184 } 1185 1186 if (likely(d->c)) { 1187 if (atomic_read(&d->c->idle_counter)) 1188 atomic_set(&d->c->idle_counter, 0); 1189 /* 1190 * If at_max_writeback_rate of cache set is true and new I/O 1191 * comes, quit max writeback rate of all cached devices 1192 * attached to this cache set, and set at_max_writeback_rate 1193 * to false. 1194 */ 1195 if (unlikely(atomic_read(&d->c->at_max_writeback_rate) == 1)) { 1196 atomic_set(&d->c->at_max_writeback_rate, 0); 1197 quit_max_writeback_rate(d->c, dc); 1198 } 1199 } 1200 1201 start_time = bio_start_io_acct(bio); 1202 1203 bio_set_dev(bio, dc->bdev); 1204 bio->bi_iter.bi_sector += dc->sb.data_offset; 1205 1206 if (cached_dev_get(dc)) { 1207 s = search_alloc(bio, d, orig_bdev, start_time); 1208 trace_bcache_request_start(s->d, bio); 1209 1210 if (!bio->bi_iter.bi_size) { 1211 /* 1212 * can't call bch_journal_meta from under 1213 * submit_bio_noacct 1214 */ 1215 continue_at_nobarrier(&s->cl, 1216 cached_dev_nodata, 1217 bcache_wq); 1218 } else { 1219 s->iop.bypass = check_should_bypass(dc, bio); 1220 1221 if (rw) 1222 cached_dev_write(dc, s); 1223 else 1224 cached_dev_read(dc, s); 1225 } 1226 } else 1227 /* I/O request sent to backing device */ 1228 detached_dev_do_request(d, bio, orig_bdev, start_time); 1229} 1230 1231static int cached_dev_ioctl(struct bcache_device *d, blk_mode_t mode, 1232 unsigned int cmd, unsigned long arg) 1233{ 1234 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1235 1236 if (dc->io_disable) 1237 return -EIO; 1238 if (!dc->bdev->bd_disk->fops->ioctl) 1239 return -ENOTTY; 1240 return dc->bdev->bd_disk->fops->ioctl(dc->bdev, mode, cmd, arg); 1241} 1242 1243void bch_cached_dev_request_init(struct cached_dev *dc) 1244{ 1245 dc->disk.cache_miss = cached_dev_cache_miss; 1246 dc->disk.ioctl = cached_dev_ioctl; 1247} 1248 1249/* Flash backed devices */ 1250 1251static int flash_dev_cache_miss(struct btree *b, struct search *s, 1252 struct bio *bio, unsigned int sectors) 1253{ 1254 unsigned int bytes = min(sectors, bio_sectors(bio)) << 9; 1255 1256 swap(bio->bi_iter.bi_size, bytes); 1257 zero_fill_bio(bio); 1258 swap(bio->bi_iter.bi_size, bytes); 1259 1260 bio_advance(bio, bytes); 1261 1262 if (!bio->bi_iter.bi_size) 1263 return MAP_DONE; 1264 1265 return MAP_CONTINUE; 1266} 1267 1268static CLOSURE_CALLBACK(flash_dev_nodata) 1269{ 1270 closure_type(s, struct search, cl); 1271 1272 if (s->iop.flush_journal) 1273 bch_journal_meta(s->iop.c, cl); 1274 1275 continue_at(cl, search_free, NULL); 1276} 1277 1278void flash_dev_submit_bio(struct bio *bio) 1279{ 1280 struct search *s; 1281 struct closure *cl; 1282 struct bcache_device *d = bio->bi_bdev->bd_disk->private_data; 1283 1284 if (unlikely(d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags))) { 1285 bio->bi_status = BLK_STS_IOERR; 1286 bio_endio(bio); 1287 return; 1288 } 1289 1290 s = search_alloc(bio, d, bio->bi_bdev, bio_start_io_acct(bio)); 1291 cl = &s->cl; 1292 bio = &s->bio.bio; 1293 1294 trace_bcache_request_start(s->d, bio); 1295 1296 if (!bio->bi_iter.bi_size) { 1297 /* 1298 * can't call bch_journal_meta from under submit_bio_noacct 1299 */ 1300 continue_at_nobarrier(&s->cl, 1301 flash_dev_nodata, 1302 bcache_wq); 1303 return; 1304 } else if (bio_data_dir(bio)) { 1305 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, 1306 &KEY(d->id, bio->bi_iter.bi_sector, 0), 1307 &KEY(d->id, bio_end_sector(bio), 0)); 1308 1309 s->iop.bypass = (bio_op(bio) == REQ_OP_DISCARD) != 0; 1310 s->iop.writeback = true; 1311 s->iop.bio = bio; 1312 1313 closure_call(&s->iop.cl, bch_data_insert, NULL, cl); 1314 } else { 1315 closure_call(&s->iop.cl, cache_lookup, NULL, cl); 1316 } 1317 1318 continue_at(cl, search_free, NULL); 1319} 1320 1321static int flash_dev_ioctl(struct bcache_device *d, blk_mode_t mode, 1322 unsigned int cmd, unsigned long arg) 1323{ 1324 return -ENOTTY; 1325} 1326 1327void bch_flash_dev_request_init(struct bcache_device *d) 1328{ 1329 d->cache_miss = flash_dev_cache_miss; 1330 d->ioctl = flash_dev_ioctl; 1331} 1332 1333void bch_request_exit(void) 1334{ 1335 kmem_cache_destroy(bch_search_cache); 1336} 1337 1338int __init bch_request_init(void) 1339{ 1340 bch_search_cache = KMEM_CACHE(search, 0); 1341 if (!bch_search_cache) 1342 return -ENOMEM; 1343 1344 return 0; 1345} 1346