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 void bch_data_insert_start(struct closure *cl); 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 = kmap(bv.bv_page) + bv.bv_offset; 48 49 csum = bch_crc64_update(csum, d, bv.bv_len); 50 kunmap(bv.bv_page); 51 } 52 53 k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1); 54} 55 56/* Insert data into cache */ 57 58static void bch_data_insert_keys(struct closure *cl) 59{ 60 struct data_insert_op *op = container_of(cl, 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 void bch_data_insert_error(struct closure *cl) 140{ 141 struct data_insert_op *op = container_of(cl, 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); 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 void bch_data_insert_start(struct closure *cl) 188{ 189 struct data_insert_op *op = container_of(cl, 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 bio_set_op_attrs(n, REQ_OP_WRITE, 0); 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 */ 308void bch_data_insert(struct closure *cl) 309{ 310 struct data_insert_op *op = container_of(cl, 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); 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_int() & 3) == 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 hd_struct *part; 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 void cache_lookup(struct closure *cl) 579{ 580 struct search *s = container_of(cl, 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 %s: returned bi_status %i\n", 655 dc->backing_dev_name, 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 part_end_io_acct(s->part, s->orig_bio, s->start_time); 674 675 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(bio, NULL, 0); 689 __bio_clone_fast(bio, orig_bio); 690 /* 691 * bi_end_io can be set separately somewhere else, e.g. the 692 * variants in, 693 * - cache_bio->bi_end_io from cached_dev_cache_miss() 694 * - n->bi_end_io from cache_lookup_fn() 695 */ 696 bio->bi_end_io = end_io_fn; 697 bio->bi_private = &s->cl; 698 699 bio_cnt_set(bio, 3); 700} 701 702static void search_free(struct closure *cl) 703{ 704 struct search *s = container_of(cl, struct search, cl); 705 706 atomic_dec(&s->iop.c->search_inflight); 707 708 if (s->iop.bio) 709 bio_put(s->iop.bio); 710 711 bio_complete(s); 712 closure_debug_destroy(cl); 713 mempool_free(s, &s->iop.c->search); 714} 715 716static inline struct search *search_alloc(struct bio *bio, 717 struct bcache_device *d) 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->start_time = part_start_io_acct(d->disk, &s->part, bio); 736 s->iop.c = d->c; 737 s->iop.bio = NULL; 738 s->iop.inode = d->id; 739 s->iop.write_point = hash_long((unsigned long) current, 16); 740 s->iop.write_prio = 0; 741 s->iop.status = 0; 742 s->iop.flags = 0; 743 s->iop.flush_journal = op_is_flush(bio->bi_opf); 744 s->iop.wq = bcache_wq; 745 746 return s; 747} 748 749/* Cached devices */ 750 751static void cached_dev_bio_complete(struct closure *cl) 752{ 753 struct search *s = container_of(cl, struct search, cl); 754 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 755 756 cached_dev_put(dc); 757 search_free(cl); 758} 759 760/* Process reads */ 761 762static void cached_dev_read_error_done(struct closure *cl) 763{ 764 struct search *s = container_of(cl, struct search, cl); 765 766 if (s->iop.replace_collision) 767 bch_mark_cache_miss_collision(s->iop.c, s->d); 768 769 if (s->iop.bio) 770 bio_free_pages(s->iop.bio); 771 772 cached_dev_bio_complete(cl); 773} 774 775static void cached_dev_read_error(struct closure *cl) 776{ 777 struct search *s = container_of(cl, struct search, cl); 778 struct bio *bio = &s->bio.bio; 779 780 /* 781 * If read request hit dirty data (s->read_dirty_data is true), 782 * then recovery a failed read request from cached device may 783 * get a stale data back. So read failure recovery is only 784 * permitted when read request hit clean data in cache device, 785 * or when cache read race happened. 786 */ 787 if (s->recoverable && !s->read_dirty_data) { 788 /* Retry from the backing device: */ 789 trace_bcache_read_retry(s->orig_bio); 790 791 s->iop.status = 0; 792 do_bio_hook(s, s->orig_bio, backing_request_endio); 793 794 /* XXX: invalidate cache */ 795 796 /* I/O request sent to backing device */ 797 closure_bio_submit(s->iop.c, bio, cl); 798 } 799 800 continue_at(cl, cached_dev_read_error_done, NULL); 801} 802 803static void cached_dev_cache_miss_done(struct closure *cl) 804{ 805 struct search *s = container_of(cl, struct search, cl); 806 struct bcache_device *d = s->d; 807 808 if (s->iop.replace_collision) 809 bch_mark_cache_miss_collision(s->iop.c, s->d); 810 811 if (s->iop.bio) 812 bio_free_pages(s->iop.bio); 813 814 cached_dev_bio_complete(cl); 815 closure_put(&d->cl); 816} 817 818static void cached_dev_read_done(struct closure *cl) 819{ 820 struct search *s = container_of(cl, struct search, cl); 821 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 822 823 /* 824 * We had a cache miss; cache_bio now contains data ready to be inserted 825 * into the cache. 826 * 827 * First, we copy the data we just read from cache_bio's bounce buffers 828 * to the buffers the original bio pointed to: 829 */ 830 831 if (s->iop.bio) { 832 bio_reset(s->iop.bio); 833 s->iop.bio->bi_iter.bi_sector = 834 s->cache_miss->bi_iter.bi_sector; 835 bio_copy_dev(s->iop.bio, s->cache_miss); 836 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9; 837 bch_bio_map(s->iop.bio, NULL); 838 839 bio_copy_data(s->cache_miss, s->iop.bio); 840 841 bio_put(s->cache_miss); 842 s->cache_miss = NULL; 843 } 844 845 if (verify(dc) && s->recoverable && !s->read_dirty_data) 846 bch_data_verify(dc, s->orig_bio); 847 848 closure_get(&dc->disk.cl); 849 bio_complete(s); 850 851 if (s->iop.bio && 852 !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) { 853 BUG_ON(!s->iop.replace); 854 closure_call(&s->iop.cl, bch_data_insert, NULL, cl); 855 } 856 857 continue_at(cl, cached_dev_cache_miss_done, NULL); 858} 859 860static void cached_dev_read_done_bh(struct closure *cl) 861{ 862 struct search *s = container_of(cl, struct search, cl); 863 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 864 865 bch_mark_cache_accounting(s->iop.c, s->d, 866 !s->cache_missed, s->iop.bypass); 867 trace_bcache_read(s->orig_bio, !s->cache_missed, s->iop.bypass); 868 869 if (s->iop.status) 870 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq); 871 else if (s->iop.bio || verify(dc)) 872 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq); 873 else 874 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL); 875} 876 877static int cached_dev_cache_miss(struct btree *b, struct search *s, 878 struct bio *bio, unsigned int sectors) 879{ 880 int ret = MAP_CONTINUE; 881 unsigned int reada = 0; 882 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 883 struct bio *miss, *cache_bio; 884 885 s->cache_missed = 1; 886 887 if (s->cache_miss || s->iop.bypass) { 888 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split); 889 ret = miss == bio ? MAP_DONE : MAP_CONTINUE; 890 goto out_submit; 891 } 892 893 if (!(bio->bi_opf & REQ_RAHEAD) && 894 !(bio->bi_opf & (REQ_META|REQ_PRIO)) && 895 s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA) 896 reada = min_t(sector_t, dc->readahead >> 9, 897 get_capacity(bio->bi_disk) - bio_end_sector(bio)); 898 899 s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada); 900 901 s->iop.replace_key = KEY(s->iop.inode, 902 bio->bi_iter.bi_sector + s->insert_bio_sectors, 903 s->insert_bio_sectors); 904 905 ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key); 906 if (ret) 907 return ret; 908 909 s->iop.replace = true; 910 911 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split); 912 913 /* btree_search_recurse()'s btree iterator is no good anymore */ 914 ret = miss == bio ? MAP_DONE : -EINTR; 915 916 cache_bio = bio_alloc_bioset(GFP_NOWAIT, 917 DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS), 918 &dc->disk.bio_split); 919 if (!cache_bio) 920 goto out_submit; 921 922 cache_bio->bi_iter.bi_sector = miss->bi_iter.bi_sector; 923 bio_copy_dev(cache_bio, miss); 924 cache_bio->bi_iter.bi_size = s->insert_bio_sectors << 9; 925 926 cache_bio->bi_end_io = backing_request_endio; 927 cache_bio->bi_private = &s->cl; 928 929 bch_bio_map(cache_bio, NULL); 930 if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO)) 931 goto out_put; 932 933 if (reada) 934 bch_mark_cache_readahead(s->iop.c, s->d); 935 936 s->cache_miss = miss; 937 s->iop.bio = cache_bio; 938 bio_get(cache_bio); 939 /* I/O request sent to backing device */ 940 closure_bio_submit(s->iop.c, cache_bio, &s->cl); 941 942 return ret; 943out_put: 944 bio_put(cache_bio); 945out_submit: 946 miss->bi_end_io = backing_request_endio; 947 miss->bi_private = &s->cl; 948 /* I/O request sent to backing device */ 949 closure_bio_submit(s->iop.c, miss, &s->cl); 950 return ret; 951} 952 953static void cached_dev_read(struct cached_dev *dc, struct search *s) 954{ 955 struct closure *cl = &s->cl; 956 957 closure_call(&s->iop.cl, cache_lookup, NULL, cl); 958 continue_at(cl, cached_dev_read_done_bh, NULL); 959} 960 961/* Process writes */ 962 963static void cached_dev_write_complete(struct closure *cl) 964{ 965 struct search *s = container_of(cl, struct search, cl); 966 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); 967 968 up_read_non_owner(&dc->writeback_lock); 969 cached_dev_bio_complete(cl); 970} 971 972static void cached_dev_write(struct cached_dev *dc, struct search *s) 973{ 974 struct closure *cl = &s->cl; 975 struct bio *bio = &s->bio.bio; 976 struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0); 977 struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0); 978 979 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end); 980 981 down_read_non_owner(&dc->writeback_lock); 982 if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) { 983 /* 984 * We overlap with some dirty data undergoing background 985 * writeback, force this write to writeback 986 */ 987 s->iop.bypass = false; 988 s->iop.writeback = true; 989 } 990 991 /* 992 * Discards aren't _required_ to do anything, so skipping if 993 * check_overlapping returned true is ok 994 * 995 * But check_overlapping drops dirty keys for which io hasn't started, 996 * so we still want to call it. 997 */ 998 if (bio_op(bio) == REQ_OP_DISCARD) 999 s->iop.bypass = true; 1000 1001 if (should_writeback(dc, s->orig_bio, 1002 cache_mode(dc), 1003 s->iop.bypass)) { 1004 s->iop.bypass = false; 1005 s->iop.writeback = true; 1006 } 1007 1008 if (s->iop.bypass) { 1009 s->iop.bio = s->orig_bio; 1010 bio_get(s->iop.bio); 1011 1012 if (bio_op(bio) == REQ_OP_DISCARD && 1013 !blk_queue_discard(bdev_get_queue(dc->bdev))) 1014 goto insert_data; 1015 1016 /* I/O request sent to backing device */ 1017 bio->bi_end_io = backing_request_endio; 1018 closure_bio_submit(s->iop.c, bio, cl); 1019 1020 } else if (s->iop.writeback) { 1021 bch_writeback_add(dc); 1022 s->iop.bio = bio; 1023 1024 if (bio->bi_opf & REQ_PREFLUSH) { 1025 /* 1026 * Also need to send a flush to the backing 1027 * device. 1028 */ 1029 struct bio *flush; 1030 1031 flush = bio_alloc_bioset(GFP_NOIO, 0, 1032 &dc->disk.bio_split); 1033 if (!flush) { 1034 s->iop.status = BLK_STS_RESOURCE; 1035 goto insert_data; 1036 } 1037 bio_copy_dev(flush, bio); 1038 flush->bi_end_io = backing_request_endio; 1039 flush->bi_private = cl; 1040 flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH; 1041 /* I/O request sent to backing device */ 1042 closure_bio_submit(s->iop.c, flush, cl); 1043 } 1044 } else { 1045 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, &dc->disk.bio_split); 1046 /* I/O request sent to backing device */ 1047 bio->bi_end_io = backing_request_endio; 1048 closure_bio_submit(s->iop.c, bio, cl); 1049 } 1050 1051insert_data: 1052 closure_call(&s->iop.cl, bch_data_insert, NULL, cl); 1053 continue_at(cl, cached_dev_write_complete, NULL); 1054} 1055 1056static void cached_dev_nodata(struct closure *cl) 1057{ 1058 struct search *s = container_of(cl, struct search, cl); 1059 struct bio *bio = &s->bio.bio; 1060 1061 if (s->iop.flush_journal) 1062 bch_journal_meta(s->iop.c, cl); 1063 1064 /* If it's a flush, we send the flush to the backing device too */ 1065 bio->bi_end_io = backing_request_endio; 1066 closure_bio_submit(s->iop.c, bio, cl); 1067 1068 continue_at(cl, cached_dev_bio_complete, NULL); 1069} 1070 1071struct detached_dev_io_private { 1072 struct bcache_device *d; 1073 unsigned long start_time; 1074 bio_end_io_t *bi_end_io; 1075 void *bi_private; 1076 struct hd_struct *part; 1077}; 1078 1079static void detached_dev_end_io(struct bio *bio) 1080{ 1081 struct detached_dev_io_private *ddip; 1082 1083 ddip = bio->bi_private; 1084 bio->bi_end_io = ddip->bi_end_io; 1085 bio->bi_private = ddip->bi_private; 1086 1087 /* Count on the bcache device */ 1088 part_end_io_acct(ddip->part, bio, ddip->start_time); 1089 1090 if (bio->bi_status) { 1091 struct cached_dev *dc = container_of(ddip->d, 1092 struct cached_dev, disk); 1093 /* should count I/O error for backing device here */ 1094 bch_count_backing_io_errors(dc, bio); 1095 } 1096 1097 kfree(ddip); 1098 bio->bi_end_io(bio); 1099} 1100 1101static void detached_dev_do_request(struct bcache_device *d, struct bio *bio) 1102{ 1103 struct detached_dev_io_private *ddip; 1104 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1105 1106 /* 1107 * no need to call closure_get(&dc->disk.cl), 1108 * because upper layer had already opened bcache device, 1109 * which would call closure_get(&dc->disk.cl) 1110 */ 1111 ddip = kzalloc(sizeof(struct detached_dev_io_private), GFP_NOIO); 1112 if (!ddip) { 1113 bio->bi_status = BLK_STS_RESOURCE; 1114 bio->bi_end_io(bio); 1115 return; 1116 } 1117 1118 ddip->d = d; 1119 /* Count on the bcache device */ 1120 ddip->start_time = part_start_io_acct(d->disk, &ddip->part, bio); 1121 ddip->bi_end_io = bio->bi_end_io; 1122 ddip->bi_private = bio->bi_private; 1123 bio->bi_end_io = detached_dev_end_io; 1124 bio->bi_private = ddip; 1125 1126 if ((bio_op(bio) == REQ_OP_DISCARD) && 1127 !blk_queue_discard(bdev_get_queue(dc->bdev))) 1128 bio->bi_end_io(bio); 1129 else 1130 submit_bio_noacct(bio); 1131} 1132 1133static void quit_max_writeback_rate(struct cache_set *c, 1134 struct cached_dev *this_dc) 1135{ 1136 int i; 1137 struct bcache_device *d; 1138 struct cached_dev *dc; 1139 1140 /* 1141 * mutex bch_register_lock may compete with other parallel requesters, 1142 * or attach/detach operations on other backing device. Waiting to 1143 * the mutex lock may increase I/O request latency for seconds or more. 1144 * To avoid such situation, if mutext_trylock() failed, only writeback 1145 * rate of current cached device is set to 1, and __update_write_back() 1146 * will decide writeback rate of other cached devices (remember now 1147 * c->idle_counter is 0 already). 1148 */ 1149 if (mutex_trylock(&bch_register_lock)) { 1150 for (i = 0; i < c->devices_max_used; i++) { 1151 if (!c->devices[i]) 1152 continue; 1153 1154 if (UUID_FLASH_ONLY(&c->uuids[i])) 1155 continue; 1156 1157 d = c->devices[i]; 1158 dc = container_of(d, struct cached_dev, disk); 1159 /* 1160 * set writeback rate to default minimum value, 1161 * then let update_writeback_rate() to decide the 1162 * upcoming rate. 1163 */ 1164 atomic_long_set(&dc->writeback_rate.rate, 1); 1165 } 1166 mutex_unlock(&bch_register_lock); 1167 } else 1168 atomic_long_set(&this_dc->writeback_rate.rate, 1); 1169} 1170 1171/* Cached devices - read & write stuff */ 1172 1173blk_qc_t cached_dev_submit_bio(struct bio *bio) 1174{ 1175 struct search *s; 1176 struct bcache_device *d = bio->bi_disk->private_data; 1177 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1178 int rw = bio_data_dir(bio); 1179 1180 if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags)) || 1181 dc->io_disable)) { 1182 bio->bi_status = BLK_STS_IOERR; 1183 bio_endio(bio); 1184 return BLK_QC_T_NONE; 1185 } 1186 1187 if (likely(d->c)) { 1188 if (atomic_read(&d->c->idle_counter)) 1189 atomic_set(&d->c->idle_counter, 0); 1190 /* 1191 * If at_max_writeback_rate of cache set is true and new I/O 1192 * comes, quit max writeback rate of all cached devices 1193 * attached to this cache set, and set at_max_writeback_rate 1194 * to false. 1195 */ 1196 if (unlikely(atomic_read(&d->c->at_max_writeback_rate) == 1)) { 1197 atomic_set(&d->c->at_max_writeback_rate, 0); 1198 quit_max_writeback_rate(d->c, dc); 1199 } 1200 } 1201 1202 bio_set_dev(bio, dc->bdev); 1203 bio->bi_iter.bi_sector += dc->sb.data_offset; 1204 1205 if (cached_dev_get(dc)) { 1206 s = search_alloc(bio, d); 1207 trace_bcache_request_start(s->d, bio); 1208 1209 if (!bio->bi_iter.bi_size) { 1210 /* 1211 * can't call bch_journal_meta from under 1212 * submit_bio_noacct 1213 */ 1214 continue_at_nobarrier(&s->cl, 1215 cached_dev_nodata, 1216 bcache_wq); 1217 } else { 1218 s->iop.bypass = check_should_bypass(dc, bio); 1219 1220 if (rw) 1221 cached_dev_write(dc, s); 1222 else 1223 cached_dev_read(dc, s); 1224 } 1225 } else 1226 /* I/O request sent to backing device */ 1227 detached_dev_do_request(d, bio); 1228 1229 return BLK_QC_T_NONE; 1230} 1231 1232static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode, 1233 unsigned int cmd, unsigned long arg) 1234{ 1235 struct cached_dev *dc = container_of(d, struct cached_dev, disk); 1236 1237 if (dc->io_disable) 1238 return -EIO; 1239 1240 return __blkdev_driver_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 void flash_dev_nodata(struct closure *cl) 1269{ 1270 struct search *s = container_of(cl, 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 1278blk_qc_t flash_dev_submit_bio(struct bio *bio) 1279{ 1280 struct search *s; 1281 struct closure *cl; 1282 struct bcache_device *d = bio->bi_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 BLK_QC_T_NONE; 1288 } 1289 1290 s = search_alloc(bio, d); 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 BLK_QC_T_NONE; 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 return BLK_QC_T_NONE; 1320} 1321 1322static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode, 1323 unsigned int cmd, unsigned long arg) 1324{ 1325 return -ENOTTY; 1326} 1327 1328void bch_flash_dev_request_init(struct bcache_device *d) 1329{ 1330 d->cache_miss = flash_dev_cache_miss; 1331 d->ioctl = flash_dev_ioctl; 1332} 1333 1334void bch_request_exit(void) 1335{ 1336 kmem_cache_destroy(bch_search_cache); 1337} 1338 1339int __init bch_request_init(void) 1340{ 1341 bch_search_cache = KMEM_CACHE(search, 0); 1342 if (!bch_search_cache) 1343 return -ENOMEM; 1344 1345 return 0; 1346} 1347