drivers/md/bcache/request.c

   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 <[email protected]>
   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
  26 struct kmem_cache *bch_search_cache;
  27
  28 static void bch_data_insert_start(struct closure *cl);
  29
  30 static unsigned int cache_mode(struct cached_dev *dc)
  31 {
  32         return BDEV_CACHE_MODE(&dc->sb);
  33 }
  34
  35 static bool verify(struct cached_dev *dc)
  36 {
  37         return dc->verify;
  38 }
  39
  40 static 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
  58 static 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
  90 static 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) - sizeof(struct jset))
 103                 return -ENOMEM;
 104
 105         return __bch_keylist_realloc(l, u64s);
 106 }
 107
 108 static 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",
 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);
 135 out:
 136         continue_at(cl, bch_data_insert_keys, op->wq);
 137 }
 138
 139 static 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
 169 static 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
 187 static 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;
 254 err:
 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  */
 308 void 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  */
 324 unsigned 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
 350 static 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
 358 static 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
 363 static 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->sb.block_size - 1) ||
 398             bio_sectors(bio) & (c->sb.block_size - 1)) {
 399                 pr_debug("skipping unaligned io");
 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;
 425 found:
 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
 453 rescale:
 454         bch_rescale_priorities(c, bio_sectors(bio));
 455         return false;
 456 skip:
 457         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 458         return true;
 459 }
 460
 461 /* Cache lookup */
 462
 463 struct 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         unsigned long           start_time;
 479
 480         struct btree_op         op;
 481         struct data_insert_op   iop;
 482 };
 483
 484 static void bch_cache_read_endio(struct bio *bio)
 485 {
 486         struct bbio *b = container_of(bio, struct bbio, bio);
 487         struct closure *cl = bio->bi_private;
 488         struct search *s = container_of(cl, struct search, cl);
 489
 490         /*
 491          * If the bucket was reused while our bio was in flight, we might have
 492          * read the wrong data. Set s->error but not error so it doesn't get
 493          * counted against the cache device, but we'll still reread the data
 494          * from the backing device.
 495          */
 496
 497         if (bio->bi_status)
 498                 s->iop.status = bio->bi_status;
 499         else if (!KEY_DIRTY(&b->key) &&
 500                  ptr_stale(s->iop.c, &b->key, 0)) {
 501                 atomic_long_inc(&s->iop.c->cache_read_races);
 502                 s->iop.status = BLK_STS_IOERR;
 503         }
 504
 505         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
 506 }
 507
 508 /*
 509  * Read from a single key, handling the initial cache miss if the key starts in
 510  * the middle of the bio
 511  */
 512 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 513 {
 514         struct search *s = container_of(op, struct search, op);
 515         struct bio *n, *bio = &s->bio.bio;
 516         struct bkey *bio_key;
 517         unsigned int ptr;
 518
 519         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 520                 return MAP_CONTINUE;
 521
 522         if (KEY_INODE(k) != s->iop.inode ||
 523             KEY_START(k) > bio->bi_iter.bi_sector) {
 524                 unsigned int bio_sectors = bio_sectors(bio);
 525                 unsigned int sectors = KEY_INODE(k) == s->iop.inode
 526                         ? min_t(uint64_t, INT_MAX,
 527                                 KEY_START(k) - bio->bi_iter.bi_sector)
 528                         : INT_MAX;
 529                 int ret = s->d->cache_miss(b, s, bio, sectors);
 530
 531                 if (ret != MAP_CONTINUE)
 532                         return ret;
 533
 534                 /* if this was a complete miss we shouldn't get here */
 535                 BUG_ON(bio_sectors <= sectors);
 536         }
 537
 538         if (!KEY_SIZE(k))
 539                 return MAP_CONTINUE;
 540
 541         /* XXX: figure out best pointer - for multiple cache devices */
 542         ptr = 0;
 543
 544         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 545
 546         if (KEY_DIRTY(k))
 547                 s->read_dirty_data = true;
 548
 549         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 550                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 551                            GFP_NOIO, &s->d->bio_split);
 552
 553         bio_key = &container_of(n, struct bbio, bio)->key;
 554         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 555
 556         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 557         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 558
 559         n->bi_end_io    = bch_cache_read_endio;
 560         n->bi_private   = &s->cl;
 561
 562         /*
 563          * The bucket we're reading from might be reused while our bio
 564          * is in flight, and we could then end up reading the wrong
 565          * data.
 566          *
 567          * We guard against this by checking (in cache_read_endio()) if
 568          * the pointer is stale again; if so, we treat it as an error
 569          * and reread from the backing device (but we don't pass that
 570          * error up anywhere).
 571          */
 572
 573         __bch_submit_bbio(n, b->c);
 574         return n == bio ? MAP_DONE : MAP_CONTINUE;
 575 }
 576
 577 static void cache_lookup(struct closure *cl)
 578 {
 579         struct search *s = container_of(cl, struct search, iop.cl);
 580         struct bio *bio = &s->bio.bio;
 581         struct cached_dev *dc;
 582         int ret;
 583
 584         bch_btree_op_init(&s->op, -1);
 585
 586         ret = bch_btree_map_keys(&s->op, s->iop.c,
 587                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 588                                  cache_lookup_fn, MAP_END_KEY);
 589         if (ret == -EAGAIN) {
 590                 continue_at(cl, cache_lookup, bcache_wq);
 591                 return;
 592         }
 593
 594         /*
 595          * We might meet err when searching the btree, If that happens, we will
 596          * get negative ret, in this scenario we should not recover data from
 597          * backing device (when cache device is dirty) because we don't know
 598          * whether bkeys the read request covered are all clean.
 599          *
 600          * And after that happened, s->iop.status is still its initial value
 601          * before we submit s->bio.bio
 602          */
 603         if (ret < 0) {
 604                 BUG_ON(ret == -EINTR);
 605                 if (s->d && s->d->c &&
 606                                 !UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) {
 607                         dc = container_of(s->d, struct cached_dev, disk);
 608                         if (dc && atomic_read(&dc->has_dirty))
 609                                 s->recoverable = false;
 610                 }
 611                 if (!s->iop.status)
 612                         s->iop.status = BLK_STS_IOERR;
 613         }
 614
 615         closure_return(cl);
 616 }
 617
 618 /* Common code for the make_request functions */
 619
 620 static void request_endio(struct bio *bio)
 621 {
 622         struct closure *cl = bio->bi_private;
 623
 624         if (bio->bi_status) {
 625                 struct search *s = container_of(cl, struct search, cl);
 626
 627                 s->iop.status = bio->bi_status;
 628                 /* Only cache read errors are recoverable */
 629                 s->recoverable = false;
 630         }
 631
 632         bio_put(bio);
 633         closure_put(cl);
 634 }
 635
 636 static void backing_request_endio(struct bio *bio)
 637 {
 638         struct closure *cl = bio->bi_private;
 639
 640         if (bio->bi_status) {
 641                 struct search *s = container_of(cl, struct search, cl);
 642                 struct cached_dev *dc = container_of(s->d,
 643                                                      struct cached_dev, disk);
 644                 /*
 645                  * If a bio has REQ_PREFLUSH for writeback mode, it is
 646                  * speically assembled in cached_dev_write() for a non-zero
 647                  * write request which has REQ_PREFLUSH. we don't set
 648                  * s->iop.status by this failure, the status will be decided
 649                  * by result of bch_data_insert() operation.
 650                  */
 651                 if (unlikely(s->iop.writeback &&
 652                              bio->bi_opf & REQ_PREFLUSH)) {
 653                         pr_err("Can't flush %s: returned bi_status %i",
 654                                 dc->backing_dev_name, bio->bi_status);
 655                 } else {
 656                         /* set to orig_bio->bi_status in bio_complete() */
 657                         s->iop.status = bio->bi_status;
 658                 }
 659                 s->recoverable = false;
 660                 /* should count I/O error for backing device here */
 661                 bch_count_backing_io_errors(dc, bio);
 662         }
 663
 664         bio_put(bio);
 665         closure_put(cl);
 666 }
 667
 668 static void bio_complete(struct search *s)
 669 {
 670         if (s->orig_bio) {
 671                 generic_end_io_acct(s->d->disk->queue, bio_op(s->orig_bio),
 672                                     &s->d->disk->part0, s->start_time);
 673
 674                 trace_bcache_request_end(s->d, s->orig_bio);
 675                 s->orig_bio->bi_status = s->iop.status;
 676                 bio_endio(s->orig_bio);
 677                 s->orig_bio = NULL;
 678         }
 679 }
 680
 681 static void do_bio_hook(struct search *s,
 682                         struct bio *orig_bio,
 683                         bio_end_io_t *end_io_fn)
 684 {
 685         struct bio *bio = &s->bio.bio;
 686
 687         bio_init(bio, NULL, 0);
 688         __bio_clone_fast(bio, orig_bio);
 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
 701 static void search_free(struct closure *cl)
 702 {
 703         struct search *s = container_of(cl, 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
 715 static inline struct search *search_alloc(struct bio *bio,
 716                                           struct bcache_device *d)
 717 {
 718         struct search *s;
 719
 720         s = mempool_alloc(&d->c->search, GFP_NOIO);
 721
 722         closure_init(&s->cl, NULL);
 723         do_bio_hook(s, bio, request_endio);
 724         atomic_inc(&d->c->search_inflight);
 725
 726         s->orig_bio             = bio;
 727         s->cache_miss           = NULL;
 728         s->cache_missed         = 0;
 729         s->d                    = d;
 730         s->recoverable          = 1;
 731         s->write                = op_is_write(bio_op(bio));
 732         s->read_dirty_data      = 0;
 733         s->start_time           = jiffies;
 734
 735         s->iop.c                = d->c;
 736         s->iop.bio              = NULL;
 737         s->iop.inode            = d->id;
 738         s->iop.write_point      = hash_long((unsigned long) current, 16);
 739         s->iop.write_prio       = 0;
 740         s->iop.status           = 0;
 741         s->iop.flags            = 0;
 742         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
 743         s->iop.wq               = bcache_wq;
 744
 745         return s;
 746 }
 747
 748 /* Cached devices */
 749
 750 static void cached_dev_bio_complete(struct closure *cl)
 751 {
 752         struct search *s = container_of(cl, struct search, cl);
 753         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 754
 755         cached_dev_put(dc);
 756         search_free(cl);
 757 }
 758
 759 /* Process reads */
 760
 761 static void cached_dev_read_error_done(struct closure *cl)
 762 {
 763         struct search *s = container_of(cl, struct search, cl);
 764
 765         if (s->iop.replace_collision)
 766                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 767
 768         if (s->iop.bio)
 769                 bio_free_pages(s->iop.bio);
 770
 771         cached_dev_bio_complete(cl);
 772 }
 773
 774 static void cached_dev_read_error(struct closure *cl)
 775 {
 776         struct search *s = container_of(cl, struct search, cl);
 777         struct bio *bio = &s->bio.bio;
 778
 779         /*
 780          * If read request hit dirty data (s->read_dirty_data is true),
 781          * then recovery a failed read request from cached device may
 782          * get a stale data back. So read failure recovery is only
 783          * permitted when read request hit clean data in cache device,
 784          * or when cache read race happened.
 785          */
 786         if (s->recoverable && !s->read_dirty_data) {
 787                 /* Retry from the backing device: */
 788                 trace_bcache_read_retry(s->orig_bio);
 789
 790                 s->iop.status = 0;
 791                 do_bio_hook(s, s->orig_bio, backing_request_endio);
 792
 793                 /* XXX: invalidate cache */
 794
 795                 /* I/O request sent to backing device */
 796                 closure_bio_submit(s->iop.c, bio, cl);
 797         }
 798
 799         continue_at(cl, cached_dev_read_error_done, NULL);
 800 }
 801
 802 static void cached_dev_cache_miss_done(struct closure *cl)
 803 {
 804         struct search *s = container_of(cl, struct search, cl);
 805         struct bcache_device *d = s->d;
 806
 807         if (s->iop.replace_collision)
 808                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 809
 810         if (s->iop.bio)
 811                 bio_free_pages(s->iop.bio);
 812
 813         cached_dev_bio_complete(cl);
 814         closure_put(&d->cl);
 815 }
 816
 817 static void cached_dev_read_done(struct closure *cl)
 818 {
 819         struct search *s = container_of(cl, struct search, cl);
 820         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 821
 822         /*
 823          * We had a cache miss; cache_bio now contains data ready to be inserted
 824          * into the cache.
 825          *
 826          * First, we copy the data we just read from cache_bio's bounce buffers
 827          * to the buffers the original bio pointed to:
 828          */
 829
 830         if (s->iop.bio) {
 831                 bio_reset(s->iop.bio);
 832                 s->iop.bio->bi_iter.bi_sector =
 833                         s->cache_miss->bi_iter.bi_sector;
 834                 bio_copy_dev(s->iop.bio, s->cache_miss);
 835                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 836                 bch_bio_map(s->iop.bio, NULL);
 837
 838                 bio_copy_data(s->cache_miss, s->iop.bio);
 839
 840                 bio_put(s->cache_miss);
 841                 s->cache_miss = NULL;
 842         }
 843
 844         if (verify(dc) && s->recoverable && !s->read_dirty_data)
 845                 bch_data_verify(dc, s->orig_bio);
 846
 847         closure_get(&dc->disk.cl);
 848         bio_complete(s);
 849
 850         if (s->iop.bio &&
 851             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 852                 BUG_ON(!s->iop.replace);
 853                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 854         }
 855
 856         continue_at(cl, cached_dev_cache_miss_done, NULL);
 857 }
 858
 859 static void cached_dev_read_done_bh(struct closure *cl)
 860 {
 861         struct search *s = container_of(cl, struct search, cl);
 862         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 863
 864         bch_mark_cache_accounting(s->iop.c, s->d,
 865                                   !s->cache_missed, s->iop.bypass);
 866         trace_bcache_read(s->orig_bio, !s->cache_missed, s->iop.bypass);
 867
 868         if (s->iop.status)
 869                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 870         else if (s->iop.bio || verify(dc))
 871                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 872         else
 873                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 874 }
 875
 876 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 877                                  struct bio *bio, unsigned int sectors)
 878 {
 879         int ret = MAP_CONTINUE;
 880         unsigned int reada = 0;
 881         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 882         struct bio *miss, *cache_bio;
 883
 884         s->cache_missed = 1;
 885
 886         if (s->cache_miss || s->iop.bypass) {
 887                 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split);
 888                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 889                 goto out_submit;
 890         }
 891
 892         if (!(bio->bi_opf & REQ_RAHEAD) &&
 893             !(bio->bi_opf & (REQ_META|REQ_PRIO)) &&
 894             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 895                 reada = min_t(sector_t, dc->readahead >> 9,
 896                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
 897
 898         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 899
 900         s->iop.replace_key = KEY(s->iop.inode,
 901                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 902                                  s->insert_bio_sectors);
 903
 904         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 905         if (ret)
 906                 return ret;
 907
 908         s->iop.replace = true;
 909
 910         miss = bio_next_split(bio, sectors, GFP_NOIO, &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(GFP_NOWAIT,
 916                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 917                         &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         bio_copy_dev(cache_bio, miss);
 923         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 924
 925         cache_bio->bi_end_io    = backing_request_endio;
 926         cache_bio->bi_private   = &s->cl;
 927
 928         bch_bio_map(cache_bio, NULL);
 929         if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 930                 goto out_put;
 931
 932         if (reada)
 933                 bch_mark_cache_readahead(s->iop.c, s->d);
 934
 935         s->cache_miss   = miss;
 936         s->iop.bio      = cache_bio;
 937         bio_get(cache_bio);
 938         /* I/O request sent to backing device */
 939         closure_bio_submit(s->iop.c, cache_bio, &s->cl);
 940
 941         return ret;
 942 out_put:
 943         bio_put(cache_bio);
 944 out_submit:
 945         miss->bi_end_io         = backing_request_endio;
 946         miss->bi_private        = &s->cl;
 947         /* I/O request sent to backing device */
 948         closure_bio_submit(s->iop.c, miss, &s->cl);
 949         return ret;
 950 }
 951
 952 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 953 {
 954         struct closure *cl = &s->cl;
 955
 956         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 957         continue_at(cl, cached_dev_read_done_bh, NULL);
 958 }
 959
 960 /* Process writes */
 961
 962 static void cached_dev_write_complete(struct closure *cl)
 963 {
 964         struct search *s = container_of(cl, struct search, cl);
 965         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 966
 967         up_read_non_owner(&dc->writeback_lock);
 968         cached_dev_bio_complete(cl);
 969 }
 970
 971 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 972 {
 973         struct closure *cl = &s->cl;
 974         struct bio *bio = &s->bio.bio;
 975         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 976         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 977
 978         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 979
 980         down_read_non_owner(&dc->writeback_lock);
 981         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 982                 /*
 983                  * We overlap with some dirty data undergoing background
 984                  * writeback, force this write to writeback
 985                  */
 986                 s->iop.bypass = false;
 987                 s->iop.writeback = true;
 988         }
 989
 990         /*
 991          * Discards aren't _required_ to do anything, so skipping if
 992          * check_overlapping returned true is ok
 993          *
 994          * But check_overlapping drops dirty keys for which io hasn't started,
 995          * so we still want to call it.
 996          */
 997         if (bio_op(bio) == REQ_OP_DISCARD)
 998                 s->iop.bypass = true;
 999
1000         if (should_writeback(dc, s->orig_bio,
1001                              cache_mode(dc),
1002                              s->iop.bypass)) {
1003                 s->iop.bypass = false;
1004                 s->iop.writeback = true;
1005         }
1006
1007         if (s->iop.bypass) {
1008                 s->iop.bio = s->orig_bio;
1009                 bio_get(s->iop.bio);
1010
1011                 if (bio_op(bio) == REQ_OP_DISCARD &&
1012                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1013                         goto insert_data;
1014
1015                 /* I/O request sent to backing device */
1016                 bio->bi_end_io = backing_request_endio;
1017                 closure_bio_submit(s->iop.c, bio, cl);
1018
1019         } else if (s->iop.writeback) {
1020                 bch_writeback_add(dc);
1021                 s->iop.bio = bio;
1022
1023                 if (bio->bi_opf & REQ_PREFLUSH) {
1024                         /*
1025                          * Also need to send a flush to the backing
1026                          * device.
1027                          */
1028                         struct bio *flush;
1029
1030                         flush = bio_alloc_bioset(GFP_NOIO, 0,
1031                                                  &dc->disk.bio_split);
1032                         if (!flush) {
1033                                 s->iop.status = BLK_STS_RESOURCE;
1034                                 goto insert_data;
1035                         }
1036                         bio_copy_dev(flush, bio);
1037                         flush->bi_end_io = backing_request_endio;
1038                         flush->bi_private = cl;
1039                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1040                         /* I/O request sent to backing device */
1041                         closure_bio_submit(s->iop.c, flush, cl);
1042                 }
1043         } else {
1044                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, &dc->disk.bio_split);
1045                 /* I/O request sent to backing device */
1046                 bio->bi_end_io = backing_request_endio;
1047                 closure_bio_submit(s->iop.c, bio, cl);
1048         }
1049
1050 insert_data:
1051         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1052         continue_at(cl, cached_dev_write_complete, NULL);
1053 }
1054
1055 static void cached_dev_nodata(struct closure *cl)
1056 {
1057         struct search *s = container_of(cl, struct search, cl);
1058         struct bio *bio = &s->bio.bio;
1059
1060         if (s->iop.flush_journal)
1061                 bch_journal_meta(s->iop.c, cl);
1062
1063         /* If it's a flush, we send the flush to the backing device too */
1064         bio->bi_end_io = backing_request_endio;
1065         closure_bio_submit(s->iop.c, bio, cl);
1066
1067         continue_at(cl, cached_dev_bio_complete, NULL);
1068 }
1069
1070 struct detached_dev_io_private {
1071         struct bcache_device    *d;
1072         unsigned long           start_time;
1073         bio_end_io_t            *bi_end_io;
1074         void                    *bi_private;
1075 };
1076
1077 static void detached_dev_end_io(struct bio *bio)
1078 {
1079         struct detached_dev_io_private *ddip;
1080
1081         ddip = bio->bi_private;
1082         bio->bi_end_io = ddip->bi_end_io;
1083         bio->bi_private = ddip->bi_private;
1084
1085         generic_end_io_acct(ddip->d->disk->queue, bio_op(bio),
1086                             &ddip->d->disk->part0, ddip->start_time);
1087
1088         if (bio->bi_status) {
1089                 struct cached_dev *dc = container_of(ddip->d,
1090                                                      struct cached_dev, disk);
1091                 /* should count I/O error for backing device here */
1092                 bch_count_backing_io_errors(dc, bio);
1093         }
1094
1095         kfree(ddip);
1096         bio->bi_end_io(bio);
1097 }
1098
1099 static void detached_dev_do_request(struct bcache_device *d, struct bio *bio)
1100 {
1101         struct detached_dev_io_private *ddip;
1102         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1103
1104         /*
1105          * no need to call closure_get(&dc->disk.cl),
1106          * because upper layer had already opened bcache device,
1107          * which would call closure_get(&dc->disk.cl)
1108          */
1109         ddip = kzalloc(sizeof(struct detached_dev_io_private), GFP_NOIO);
1110         ddip->d = d;
1111         ddip->start_time = jiffies;
1112         ddip->bi_end_io = bio->bi_end_io;
1113         ddip->bi_private = bio->bi_private;
1114         bio->bi_end_io = detached_dev_end_io;
1115         bio->bi_private = ddip;
1116
1117         if ((bio_op(bio) == REQ_OP_DISCARD) &&
1118             !blk_queue_discard(bdev_get_queue(dc->bdev)))
1119                 bio->bi_end_io(bio);
1120         else
1121                 generic_make_request(bio);
1122 }
1123
1124 static void quit_max_writeback_rate(struct cache_set *c,
1125                                     struct cached_dev *this_dc)
1126 {
1127         int i;
1128         struct bcache_device *d;
1129         struct cached_dev *dc;
1130
1131         /*
1132          * mutex bch_register_lock may compete with other parallel requesters,
1133          * or attach/detach operations on other backing device. Waiting to
1134          * the mutex lock may increase I/O request latency for seconds or more.
1135          * To avoid such situation, if mutext_trylock() failed, only writeback
1136          * rate of current cached device is set to 1, and __update_write_back()
1137          * will decide writeback rate of other cached devices (remember now
1138          * c->idle_counter is 0 already).
1139          */
1140         if (mutex_trylock(&bch_register_lock)) {
1141                 for (i = 0; i < c->devices_max_used; i++) {
1142                         if (!c->devices[i])
1143                                 continue;
1144
1145                         if (UUID_FLASH_ONLY(&c->uuids[i]))
1146                                 continue;
1147
1148                         d = c->devices[i];
1149                         dc = container_of(d, struct cached_dev, disk);
1150                         /*
1151                          * set writeback rate to default minimum value,
1152                          * then let update_writeback_rate() to decide the
1153                          * upcoming rate.
1154                          */
1155                         atomic_long_set(&dc->writeback_rate.rate, 1);
1156                 }
1157                 mutex_unlock(&bch_register_lock);
1158         } else
1159                 atomic_long_set(&this_dc->writeback_rate.rate, 1);
1160 }
1161
1162 /* Cached devices - read & write stuff */
1163
1164 static blk_qc_t cached_dev_make_request(struct request_queue *q,
1165                                         struct bio *bio)
1166 {
1167         struct search *s;
1168         struct bcache_device *d = bio->bi_disk->private_data;
1169         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1170         int rw = bio_data_dir(bio);
1171
1172         if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags)) ||
1173                      dc->io_disable)) {
1174                 bio->bi_status = BLK_STS_IOERR;
1175                 bio_endio(bio);
1176                 return BLK_QC_T_NONE;
1177         }
1178
1179         if (likely(d->c)) {
1180                 if (atomic_read(&d->c->idle_counter))
1181                         atomic_set(&d->c->idle_counter, 0);
1182                 /*
1183                  * If at_max_writeback_rate of cache set is true and new I/O
1184                  * comes, quit max writeback rate of all cached devices
1185                  * attached to this cache set, and set at_max_writeback_rate
1186                  * to false.
1187                  */
1188                 if (unlikely(atomic_read(&d->c->at_max_writeback_rate) == 1)) {
1189                         atomic_set(&d->c->at_max_writeback_rate, 0);
1190                         quit_max_writeback_rate(d->c, dc);
1191                 }
1192         }
1193
1194         generic_start_io_acct(q,
1195                               bio_op(bio),
1196                               bio_sectors(bio),
1197                               &d->disk->part0);
1198
1199         bio_set_dev(bio, dc->bdev);
1200         bio->bi_iter.bi_sector += dc->sb.data_offset;
1201
1202         if (cached_dev_get(dc)) {
1203                 s = search_alloc(bio, d);
1204                 trace_bcache_request_start(s->d, bio);
1205
1206                 if (!bio->bi_iter.bi_size) {
1207                         /*
1208                          * can't call bch_journal_meta from under
1209                          * generic_make_request
1210                          */
1211                         continue_at_nobarrier(&s->cl,
1212                                               cached_dev_nodata,
1213                                               bcache_wq);
1214                 } else {
1215                         s->iop.bypass = check_should_bypass(dc, bio);
1216
1217                         if (rw)
1218                                 cached_dev_write(dc, s);
1219                         else
1220                                 cached_dev_read(dc, s);
1221                 }
1222         } else
1223                 /* I/O request sent to backing device */
1224                 detached_dev_do_request(d, bio);
1225
1226         return BLK_QC_T_NONE;
1227 }
1228
1229 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1230                             unsigned int cmd, unsigned long arg)
1231 {
1232         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1233
1234         if (dc->io_disable)
1235                 return -EIO;
1236
1237         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1238 }
1239
1240 static int cached_dev_congested(void *data, int bits)
1241 {
1242         struct bcache_device *d = data;
1243         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1244         struct request_queue *q = bdev_get_queue(dc->bdev);
1245         int ret = 0;
1246
1247         if (bdi_congested(q->backing_dev_info, bits))
1248                 return 1;
1249
1250         if (cached_dev_get(dc)) {
1251                 unsigned int i;
1252                 struct cache *ca;
1253
1254                 for_each_cache(ca, d->c, i) {
1255                         q = bdev_get_queue(ca->bdev);
1256                         ret |= bdi_congested(q->backing_dev_info, bits);
1257                 }
1258
1259                 cached_dev_put(dc);
1260         }
1261
1262         return ret;
1263 }
1264
1265 void bch_cached_dev_request_init(struct cached_dev *dc)
1266 {
1267         struct gendisk *g = dc->disk.disk;
1268
1269         g->queue->make_request_fn               = cached_dev_make_request;
1270         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1271         dc->disk.cache_miss                     = cached_dev_cache_miss;
1272         dc->disk.ioctl                          = cached_dev_ioctl;
1273 }
1274
1275 /* Flash backed devices */
1276
1277 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1278                                 struct bio *bio, unsigned int sectors)
1279 {
1280         unsigned int bytes = min(sectors, bio_sectors(bio)) << 9;
1281
1282         swap(bio->bi_iter.bi_size, bytes);
1283         zero_fill_bio(bio);
1284         swap(bio->bi_iter.bi_size, bytes);
1285
1286         bio_advance(bio, bytes);
1287
1288         if (!bio->bi_iter.bi_size)
1289                 return MAP_DONE;
1290
1291         return MAP_CONTINUE;
1292 }
1293
1294 static void flash_dev_nodata(struct closure *cl)
1295 {
1296         struct search *s = container_of(cl, struct search, cl);
1297
1298         if (s->iop.flush_journal)
1299                 bch_journal_meta(s->iop.c, cl);
1300
1301         continue_at(cl, search_free, NULL);
1302 }
1303
1304 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1305                                              struct bio *bio)
1306 {
1307         struct search *s;
1308         struct closure *cl;
1309         struct bcache_device *d = bio->bi_disk->private_data;
1310
1311         if (unlikely(d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags))) {
1312                 bio->bi_status = BLK_STS_IOERR;
1313                 bio_endio(bio);
1314                 return BLK_QC_T_NONE;
1315         }
1316
1317         generic_start_io_acct(q, bio_op(bio), bio_sectors(bio), &d->disk->part0);
1318
1319         s = search_alloc(bio, d);
1320         cl = &s->cl;
1321         bio = &s->bio.bio;
1322
1323         trace_bcache_request_start(s->d, bio);
1324
1325         if (!bio->bi_iter.bi_size) {
1326                 /*
1327                  * can't call bch_journal_meta from under
1328                  * generic_make_request
1329                  */
1330                 continue_at_nobarrier(&s->cl,
1331                                       flash_dev_nodata,
1332                                       bcache_wq);
1333                 return BLK_QC_T_NONE;
1334         } else if (bio_data_dir(bio)) {
1335                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1336                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1337                                         &KEY(d->id, bio_end_sector(bio), 0));
1338
1339                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1340                 s->iop.writeback        = true;
1341                 s->iop.bio              = bio;
1342
1343                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1344         } else {
1345                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1346         }
1347
1348         continue_at(cl, search_free, NULL);
1349         return BLK_QC_T_NONE;
1350 }
1351
1352 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1353                            unsigned int cmd, unsigned long arg)
1354 {
1355         return -ENOTTY;
1356 }
1357
1358 static int flash_dev_congested(void *data, int bits)
1359 {
1360         struct bcache_device *d = data;
1361         struct request_queue *q;
1362         struct cache *ca;
1363         unsigned int i;
1364         int ret = 0;
1365
1366         for_each_cache(ca, d->c, i) {
1367                 q = bdev_get_queue(ca->bdev);
1368                 ret |= bdi_congested(q->backing_dev_info, bits);
1369         }
1370
1371         return ret;
1372 }
1373
1374 void bch_flash_dev_request_init(struct bcache_device *d)
1375 {
1376         struct gendisk *g = d->disk;
1377
1378         g->queue->make_request_fn               = flash_dev_make_request;
1379         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1380         d->cache_miss                           = flash_dev_cache_miss;
1381         d->ioctl                                = flash_dev_ioctl;
1382 }
1383
1384 void bch_request_exit(void)
1385 {
1386         kmem_cache_destroy(bch_search_cache);
1387 }
1388
1389 int __init bch_request_init(void)
1390 {
1391         bch_search_cache = KMEM_CACHE(search, 0);
1392         if (!bch_search_cache)
1393                 return -ENOMEM;
1394
1395         return 0;
1396 }