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         /*
  66          * If we're looping, might already be waiting on
  67          * another journal write - can't wait on more than one journal write at
  68          * a time
  69          *
  70          * XXX: this looks wrong
  71          */
  72 #if 0
  73         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
  74                 closure_sync(&s->cl);
  75 #endif
  76
  77         if (!op->replace)
  78                 journal_ref = bch_journal(op->c, &op->insert_keys,
  79                                           op->flush_journal ? cl : NULL);
  80
  81         ret = bch_btree_insert(op->c, &op->insert_keys,
  82                                journal_ref, replace_key);
  83         if (ret == -ESRCH) {
  84                 op->replace_collision = true;
  85         } else if (ret) {
  86                 op->status              = BLK_STS_RESOURCE;
  87                 op->insert_data_done    = true;
  88         }
  89
  90         if (journal_ref)
  91                 atomic_dec_bug(journal_ref);
  92
  93         if (!op->insert_data_done) {
  94                 continue_at(cl, bch_data_insert_start, op->wq);
  95                 return;
  96         }
  97
  98         bch_keylist_free(&op->insert_keys);
  99         closure_return(cl);
 100 }
 101
 102 static int bch_keylist_realloc(struct keylist *l, unsigned int u64s,
 103                                struct cache_set *c)
 104 {
 105         size_t oldsize = bch_keylist_nkeys(l);
 106         size_t newsize = oldsize + u64s;
 107
 108         /*
 109          * The journalling code doesn't handle the case where the keys to insert
 110          * is bigger than an empty write: If we just return -ENOMEM here,
 111          * bch_data_insert_keys() will insert the keys created so far
 112          * and finish the rest when the keylist is empty.
 113          */
 114         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
 115                 return -ENOMEM;
 116
 117         return __bch_keylist_realloc(l, u64s);
 118 }
 119
 120 static void bch_data_invalidate(struct closure *cl)
 121 {
 122         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 123         struct bio *bio = op->bio;
 124
 125         pr_debug("invalidating %i sectors from %llu",
 126                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 127
 128         while (bio_sectors(bio)) {
 129                 unsigned int sectors = min(bio_sectors(bio),
 130                                        1U << (KEY_SIZE_BITS - 1));
 131
 132                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
 133                         goto out;
 134
 135                 bio->bi_iter.bi_sector  += sectors;
 136                 bio->bi_iter.bi_size    -= sectors << 9;
 137
 138                 bch_keylist_add(&op->insert_keys,
 139                                 &KEY(op->inode,
 140                                      bio->bi_iter.bi_sector,
 141                                      sectors));
 142         }
 143
 144         op->insert_data_done = true;
 145         /* get in bch_data_insert() */
 146         bio_put(bio);
 147 out:
 148         continue_at(cl, bch_data_insert_keys, op->wq);
 149 }
 150
 151 static void bch_data_insert_error(struct closure *cl)
 152 {
 153         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 154
 155         /*
 156          * Our data write just errored, which means we've got a bunch of keys to
 157          * insert that point to data that wasn't successfully written.
 158          *
 159          * We don't have to insert those keys but we still have to invalidate
 160          * that region of the cache - so, if we just strip off all the pointers
 161          * from the keys we'll accomplish just that.
 162          */
 163
 164         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 165
 166         while (src != op->insert_keys.top) {
 167                 struct bkey *n = bkey_next(src);
 168
 169                 SET_KEY_PTRS(src, 0);
 170                 memmove(dst, src, bkey_bytes(src));
 171
 172                 dst = bkey_next(dst);
 173                 src = n;
 174         }
 175
 176         op->insert_keys.top = dst;
 177
 178         bch_data_insert_keys(cl);
 179 }
 180
 181 static void bch_data_insert_endio(struct bio *bio)
 182 {
 183         struct closure *cl = bio->bi_private;
 184         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 185
 186         if (bio->bi_status) {
 187                 /* TODO: We could try to recover from this. */
 188                 if (op->writeback)
 189                         op->status = bio->bi_status;
 190                 else if (!op->replace)
 191                         set_closure_fn(cl, bch_data_insert_error, op->wq);
 192                 else
 193                         set_closure_fn(cl, NULL, NULL);
 194         }
 195
 196         bch_bbio_endio(op->c, bio, bio->bi_status, "writing data to cache");
 197 }
 198
 199 static void bch_data_insert_start(struct closure *cl)
 200 {
 201         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 202         struct bio *bio = op->bio, *n;
 203
 204         if (op->bypass)
 205                 return bch_data_invalidate(cl);
 206
 207         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0)
 208                 wake_up_gc(op->c);
 209
 210         /*
 211          * Journal writes are marked REQ_PREFLUSH; if the original write was a
 212          * flush, it'll wait on the journal write.
 213          */
 214         bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
 215
 216         do {
 217                 unsigned int i;
 218                 struct bkey *k;
 219                 struct bio_set *split = &op->c->bio_split;
 220
 221                 /* 1 for the device pointer and 1 for the chksum */
 222                 if (bch_keylist_realloc(&op->insert_keys,
 223                                         3 + (op->csum ? 1 : 0),
 224                                         op->c)) {
 225                         continue_at(cl, bch_data_insert_keys, op->wq);
 226                         return;
 227                 }
 228
 229                 k = op->insert_keys.top;
 230                 bkey_init(k);
 231                 SET_KEY_INODE(k, op->inode);
 232                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 233
 234                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 235                                        op->write_point, op->write_prio,
 236                                        op->writeback))
 237                         goto err;
 238
 239                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 240
 241                 n->bi_end_io    = bch_data_insert_endio;
 242                 n->bi_private   = cl;
 243
 244                 if (op->writeback) {
 245                         SET_KEY_DIRTY(k, true);
 246
 247                         for (i = 0; i < KEY_PTRS(k); i++)
 248                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 249                                             GC_MARK_DIRTY);
 250                 }
 251
 252                 SET_KEY_CSUM(k, op->csum);
 253                 if (KEY_CSUM(k))
 254                         bio_csum(n, k);
 255
 256                 trace_bcache_cache_insert(k);
 257                 bch_keylist_push(&op->insert_keys);
 258
 259                 bio_set_op_attrs(n, REQ_OP_WRITE, 0);
 260                 bch_submit_bbio(n, op->c, k, 0);
 261         } while (n != bio);
 262
 263         op->insert_data_done = true;
 264         continue_at(cl, bch_data_insert_keys, op->wq);
 265         return;
 266 err:
 267         /* bch_alloc_sectors() blocks if s->writeback = true */
 268         BUG_ON(op->writeback);
 269
 270         /*
 271          * But if it's not a writeback write we'd rather just bail out if
 272          * there aren't any buckets ready to write to - it might take awhile and
 273          * we might be starving btree writes for gc or something.
 274          */
 275
 276         if (!op->replace) {
 277                 /*
 278                  * Writethrough write: We can't complete the write until we've
 279                  * updated the index. But we don't want to delay the write while
 280                  * we wait for buckets to be freed up, so just invalidate the
 281                  * rest of the write.
 282                  */
 283                 op->bypass = true;
 284                 return bch_data_invalidate(cl);
 285         } else {
 286                 /*
 287                  * From a cache miss, we can just insert the keys for the data
 288                  * we have written or bail out if we didn't do anything.
 289                  */
 290                 op->insert_data_done = true;
 291                 bio_put(bio);
 292
 293                 if (!bch_keylist_empty(&op->insert_keys))
 294                         continue_at(cl, bch_data_insert_keys, op->wq);
 295                 else
 296                         closure_return(cl);
 297         }
 298 }
 299
 300 /**
 301  * bch_data_insert - stick some data in the cache
 302  * @cl: closure pointer.
 303  *
 304  * This is the starting point for any data to end up in a cache device; it could
 305  * be from a normal write, or a writeback write, or a write to a flash only
 306  * volume - it's also used by the moving garbage collector to compact data in
 307  * mostly empty buckets.
 308  *
 309  * It first writes the data to the cache, creating a list of keys to be inserted
 310  * (if the data had to be fragmented there will be multiple keys); after the
 311  * data is written it calls bch_journal, and after the keys have been added to
 312  * the next journal write they're inserted into the btree.
 313  *
 314  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 315  * and op->inode is used for the key inode.
 316  *
 317  * If s->bypass is true, instead of inserting the data it invalidates the
 318  * region of the cache represented by s->cache_bio and op->inode.
 319  */
 320 void bch_data_insert(struct closure *cl)
 321 {
 322         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 323
 324         trace_bcache_write(op->c, op->inode, op->bio,
 325                            op->writeback, op->bypass);
 326
 327         bch_keylist_init(&op->insert_keys);
 328         bio_get(op->bio);
 329         bch_data_insert_start(cl);
 330 }
 331
 332 /* Congested? */
 333
 334 unsigned int bch_get_congested(struct cache_set *c)
 335 {
 336         int i;
 337         long rand;
 338
 339         if (!c->congested_read_threshold_us &&
 340             !c->congested_write_threshold_us)
 341                 return 0;
 342
 343         i = (local_clock_us() - c->congested_last_us) / 1024;
 344         if (i < 0)
 345                 return 0;
 346
 347         i += atomic_read(&c->congested);
 348         if (i >= 0)
 349                 return 0;
 350
 351         i += CONGESTED_MAX;
 352
 353         if (i > 0)
 354                 i = fract_exp_two(i, 6);
 355
 356         rand = get_random_int();
 357         i -= bitmap_weight(&rand, BITS_PER_LONG);
 358
 359         return i > 0 ? i : 1;
 360 }
 361
 362 static void add_sequential(struct task_struct *t)
 363 {
 364         ewma_add(t->sequential_io_avg,
 365                  t->sequential_io, 8, 0);
 366
 367         t->sequential_io = 0;
 368 }
 369
 370 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 371 {
 372         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 373 }
 374
 375 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 376 {
 377         struct cache_set *c = dc->disk.c;
 378         unsigned int mode = cache_mode(dc);
 379         unsigned int sectors, congested = bch_get_congested(c);
 380         struct task_struct *task = current;
 381         struct io *i;
 382
 383         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 384             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 385             (bio_op(bio) == REQ_OP_DISCARD))
 386                 goto skip;
 387
 388         if (mode == CACHE_MODE_NONE ||
 389             (mode == CACHE_MODE_WRITEAROUND &&
 390              op_is_write(bio_op(bio))))
 391                 goto skip;
 392
 393         /*
 394          * Flag for bypass if the IO is for read-ahead or background,
 395          * unless the read-ahead request is for metadata (eg, for gfs2).
 396          */
 397         if (bio->bi_opf & (REQ_RAHEAD|REQ_BACKGROUND) &&
 398             !(bio->bi_opf & REQ_META))
 399                 goto skip;
 400
 401         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 402             bio_sectors(bio) & (c->sb.block_size - 1)) {
 403                 pr_debug("skipping unaligned io");
 404                 goto skip;
 405         }
 406
 407         if (bypass_torture_test(dc)) {
 408                 if ((get_random_int() & 3) == 3)
 409                         goto skip;
 410                 else
 411                         goto rescale;
 412         }
 413
 414         if (!congested && !dc->sequential_cutoff)
 415                 goto rescale;
 416
 417         spin_lock(&dc->io_lock);
 418
 419         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
 420                 if (i->last == bio->bi_iter.bi_sector &&
 421                     time_before(jiffies, i->jiffies))
 422                         goto found;
 423
 424         i = list_first_entry(&dc->io_lru, struct io, lru);
 425
 426         add_sequential(task);
 427         i->sequential = 0;
 428 found:
 429         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
 430                 i->sequential   += bio->bi_iter.bi_size;
 431
 432         i->last                  = bio_end_sector(bio);
 433         i->jiffies               = jiffies + msecs_to_jiffies(5000);
 434         task->sequential_io      = i->sequential;
 435
 436         hlist_del(&i->hash);
 437         hlist_add_head(&i->hash, iohash(dc, i->last));
 438         list_move_tail(&i->lru, &dc->io_lru);
 439
 440         spin_unlock(&dc->io_lock);
 441
 442         sectors = max(task->sequential_io,
 443                       task->sequential_io_avg) >> 9;
 444
 445         if (dc->sequential_cutoff &&
 446             sectors >= dc->sequential_cutoff >> 9) {
 447                 trace_bcache_bypass_sequential(bio);
 448                 goto skip;
 449         }
 450
 451         if (congested && sectors >= congested) {
 452                 trace_bcache_bypass_congested(bio);
 453                 goto skip;
 454         }
 455
 456 rescale:
 457         bch_rescale_priorities(c, bio_sectors(bio));
 458         return false;
 459 skip:
 460         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 461         return true;
 462 }
 463
 464 /* Cache lookup */
 465
 466 struct search {
 467         /* Stack frame for bio_complete */
 468         struct closure          cl;
 469
 470         struct bbio             bio;
 471         struct bio              *orig_bio;
 472         struct bio              *cache_miss;
 473         struct bcache_device    *d;
 474
 475         unsigned int            insert_bio_sectors;
 476         unsigned int            recoverable:1;
 477         unsigned int            write:1;
 478         unsigned int            read_dirty_data:1;
 479         unsigned int            cache_missed:1;
 480
 481         unsigned long           start_time;
 482
 483         struct btree_op         op;
 484         struct data_insert_op   iop;
 485 };
 486
 487 static void bch_cache_read_endio(struct bio *bio)
 488 {
 489         struct bbio *b = container_of(bio, struct bbio, bio);
 490         struct closure *cl = bio->bi_private;
 491         struct search *s = container_of(cl, struct search, cl);
 492
 493         /*
 494          * If the bucket was reused while our bio was in flight, we might have
 495          * read the wrong data. Set s->error but not error so it doesn't get
 496          * counted against the cache device, but we'll still reread the data
 497          * from the backing device.
 498          */
 499
 500         if (bio->bi_status)
 501                 s->iop.status = bio->bi_status;
 502         else if (!KEY_DIRTY(&b->key) &&
 503                  ptr_stale(s->iop.c, &b->key, 0)) {
 504                 atomic_long_inc(&s->iop.c->cache_read_races);
 505                 s->iop.status = BLK_STS_IOERR;
 506         }
 507
 508         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
 509 }
 510
 511 /*
 512  * Read from a single key, handling the initial cache miss if the key starts in
 513  * the middle of the bio
 514  */
 515 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 516 {
 517         struct search *s = container_of(op, struct search, op);
 518         struct bio *n, *bio = &s->bio.bio;
 519         struct bkey *bio_key;
 520         unsigned int ptr;
 521
 522         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 523                 return MAP_CONTINUE;
 524
 525         if (KEY_INODE(k) != s->iop.inode ||
 526             KEY_START(k) > bio->bi_iter.bi_sector) {
 527                 unsigned int bio_sectors = bio_sectors(bio);
 528                 unsigned int sectors = KEY_INODE(k) == s->iop.inode
 529                         ? min_t(uint64_t, INT_MAX,
 530                                 KEY_START(k) - bio->bi_iter.bi_sector)
 531                         : INT_MAX;
 532                 int ret = s->d->cache_miss(b, s, bio, sectors);
 533
 534                 if (ret != MAP_CONTINUE)
 535                         return ret;
 536
 537                 /* if this was a complete miss we shouldn't get here */
 538                 BUG_ON(bio_sectors <= sectors);
 539         }
 540
 541         if (!KEY_SIZE(k))
 542                 return MAP_CONTINUE;
 543
 544         /* XXX: figure out best pointer - for multiple cache devices */
 545         ptr = 0;
 546
 547         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 548
 549         if (KEY_DIRTY(k))
 550                 s->read_dirty_data = true;
 551
 552         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 553                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 554                            GFP_NOIO, &s->d->bio_split);
 555
 556         bio_key = &container_of(n, struct bbio, bio)->key;
 557         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 558
 559         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 560         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 561
 562         n->bi_end_io    = bch_cache_read_endio;
 563         n->bi_private   = &s->cl;
 564
 565         /*
 566          * The bucket we're reading from might be reused while our bio
 567          * is in flight, and we could then end up reading the wrong
 568          * data.
 569          *
 570          * We guard against this by checking (in cache_read_endio()) if
 571          * the pointer is stale again; if so, we treat it as an error
 572          * and reread from the backing device (but we don't pass that
 573          * error up anywhere).
 574          */
 575
 576         __bch_submit_bbio(n, b->c);
 577         return n == bio ? MAP_DONE : MAP_CONTINUE;
 578 }
 579
 580 static void cache_lookup(struct closure *cl)
 581 {
 582         struct search *s = container_of(cl, struct search, iop.cl);
 583         struct bio *bio = &s->bio.bio;
 584         struct cached_dev *dc;
 585         int ret;
 586
 587         bch_btree_op_init(&s->op, -1);
 588
 589         ret = bch_btree_map_keys(&s->op, s->iop.c,
 590                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 591                                  cache_lookup_fn, MAP_END_KEY);
 592         if (ret == -EAGAIN) {
 593                 continue_at(cl, cache_lookup, bcache_wq);
 594                 return;
 595         }
 596
 597         /*
 598          * We might meet err when searching the btree, If that happens, we will
 599          * get negative ret, in this scenario we should not recover data from
 600          * backing device (when cache device is dirty) because we don't know
 601          * whether bkeys the read request covered are all clean.
 602          *
 603          * And after that happened, s->iop.status is still its initial value
 604          * before we submit s->bio.bio
 605          */
 606         if (ret < 0) {
 607                 BUG_ON(ret == -EINTR);
 608                 if (s->d && s->d->c &&
 609                                 !UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) {
 610                         dc = container_of(s->d, struct cached_dev, disk);
 611                         if (dc && atomic_read(&dc->has_dirty))
 612                                 s->recoverable = false;
 613                 }
 614                 if (!s->iop.status)
 615                         s->iop.status = BLK_STS_IOERR;
 616         }
 617
 618         closure_return(cl);
 619 }
 620
 621 /* Common code for the make_request functions */
 622
 623 static void request_endio(struct bio *bio)
 624 {
 625         struct closure *cl = bio->bi_private;
 626
 627         if (bio->bi_status) {
 628                 struct search *s = container_of(cl, struct search, cl);
 629
 630                 s->iop.status = bio->bi_status;
 631                 /* Only cache read errors are recoverable */
 632                 s->recoverable = false;
 633         }
 634
 635         bio_put(bio);
 636         closure_put(cl);
 637 }
 638
 639 static void backing_request_endio(struct bio *bio)
 640 {
 641         struct closure *cl = bio->bi_private;
 642
 643         if (bio->bi_status) {
 644                 struct search *s = container_of(cl, struct search, cl);
 645                 struct cached_dev *dc = container_of(s->d,
 646                                                      struct cached_dev, disk);
 647                 /*
 648                  * If a bio has REQ_PREFLUSH for writeback mode, it is
 649                  * speically assembled in cached_dev_write() for a non-zero
 650                  * write request which has REQ_PREFLUSH. we don't set
 651                  * s->iop.status by this failure, the status will be decided
 652                  * by result of bch_data_insert() operation.
 653                  */
 654                 if (unlikely(s->iop.writeback &&
 655                              bio->bi_opf & REQ_PREFLUSH)) {
 656                         pr_err("Can't flush %s: returned bi_status %i",
 657                                 dc->backing_dev_name, bio->bi_status);
 658                 } else {
 659                         /* set to orig_bio->bi_status in bio_complete() */
 660                         s->iop.status = bio->bi_status;
 661                 }
 662                 s->recoverable = false;
 663                 /* should count I/O error for backing device here */
 664                 bch_count_backing_io_errors(dc, bio);
 665         }
 666
 667         bio_put(bio);
 668         closure_put(cl);
 669 }
 670
 671 static void bio_complete(struct search *s)
 672 {
 673         if (s->orig_bio) {
 674                 generic_end_io_acct(s->d->disk->queue, bio_op(s->orig_bio),
 675                                     &s->d->disk->part0, s->start_time);
 676
 677                 trace_bcache_request_end(s->d, s->orig_bio);
 678                 s->orig_bio->bi_status = s->iop.status;
 679                 bio_endio(s->orig_bio);
 680                 s->orig_bio = NULL;
 681         }
 682 }
 683
 684 static void do_bio_hook(struct search *s,
 685                         struct bio *orig_bio,
 686                         bio_end_io_t *end_io_fn)
 687 {
 688         struct bio *bio = &s->bio.bio;
 689
 690         bio_init(bio, NULL, 0);
 691         __bio_clone_fast(bio, orig_bio);
 692         /*
 693          * bi_end_io can be set separately somewhere else, e.g. the
 694          * variants in,
 695          * - cache_bio->bi_end_io from cached_dev_cache_miss()
 696          * - n->bi_end_io from cache_lookup_fn()
 697          */
 698         bio->bi_end_io          = end_io_fn;
 699         bio->bi_private         = &s->cl;
 700
 701         bio_cnt_set(bio, 3);
 702 }
 703
 704 static void search_free(struct closure *cl)
 705 {
 706         struct search *s = container_of(cl, struct search, cl);
 707
 708         atomic_dec(&s->d->c->search_inflight);
 709
 710         if (s->iop.bio)
 711                 bio_put(s->iop.bio);
 712
 713         bio_complete(s);
 714         closure_debug_destroy(cl);
 715         mempool_free(s, &s->d->c->search);
 716 }
 717
 718 static inline struct search *search_alloc(struct bio *bio,
 719                                           struct bcache_device *d)
 720 {
 721         struct search *s;
 722
 723         s = mempool_alloc(&d->c->search, GFP_NOIO);
 724
 725         closure_init(&s->cl, NULL);
 726         do_bio_hook(s, bio, request_endio);
 727         atomic_inc(&d->c->search_inflight);
 728
 729         s->orig_bio             = bio;
 730         s->cache_miss           = NULL;
 731         s->cache_missed         = 0;
 732         s->d                    = d;
 733         s->recoverable          = 1;
 734         s->write                = op_is_write(bio_op(bio));
 735         s->read_dirty_data      = 0;
 736         s->start_time           = jiffies;
 737
 738         s->iop.c                = d->c;
 739         s->iop.bio              = NULL;
 740         s->iop.inode            = d->id;
 741         s->iop.write_point      = hash_long((unsigned long) current, 16);
 742         s->iop.write_prio       = 0;
 743         s->iop.status           = 0;
 744         s->iop.flags            = 0;
 745         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
 746         s->iop.wq               = bcache_wq;
 747
 748         return s;
 749 }
 750
 751 /* Cached devices */
 752
 753 static void cached_dev_bio_complete(struct closure *cl)
 754 {
 755         struct search *s = container_of(cl, struct search, cl);
 756         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 757
 758         search_free(cl);
 759         cached_dev_put(dc);
 760 }
 761
 762 /* Process reads */
 763
 764 static void cached_dev_cache_miss_done(struct closure *cl)
 765 {
 766         struct search *s = container_of(cl, struct search, cl);
 767
 768         if (s->iop.replace_collision)
 769                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 770
 771         if (s->iop.bio)
 772                 bio_free_pages(s->iop.bio);
 773
 774         cached_dev_bio_complete(cl);
 775 }
 776
 777 static void cached_dev_read_error(struct closure *cl)
 778 {
 779         struct search *s = container_of(cl, struct search, cl);
 780         struct bio *bio = &s->bio.bio;
 781
 782         /*
 783          * If read request hit dirty data (s->read_dirty_data is true),
 784          * then recovery a failed read request from cached device may
 785          * get a stale data back. So read failure recovery is only
 786          * permitted when read request hit clean data in cache device,
 787          * or when cache read race happened.
 788          */
 789         if (s->recoverable && !s->read_dirty_data) {
 790                 /* Retry from the backing device: */
 791                 trace_bcache_read_retry(s->orig_bio);
 792
 793                 s->iop.status = 0;
 794                 do_bio_hook(s, s->orig_bio, backing_request_endio);
 795
 796                 /* XXX: invalidate cache */
 797
 798                 /* I/O request sent to backing device */
 799                 closure_bio_submit(s->iop.c, bio, cl);
 800         }
 801
 802         continue_at(cl, cached_dev_cache_miss_done, NULL);
 803 }
 804
 805 static void cached_dev_read_done(struct closure *cl)
 806 {
 807         struct search *s = container_of(cl, struct search, cl);
 808         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 809
 810         /*
 811          * We had a cache miss; cache_bio now contains data ready to be inserted
 812          * into the cache.
 813          *
 814          * First, we copy the data we just read from cache_bio's bounce buffers
 815          * to the buffers the original bio pointed to:
 816          */
 817
 818         if (s->iop.bio) {
 819                 bio_reset(s->iop.bio);
 820                 s->iop.bio->bi_iter.bi_sector =
 821                         s->cache_miss->bi_iter.bi_sector;
 822                 bio_copy_dev(s->iop.bio, s->cache_miss);
 823                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 824                 bch_bio_map(s->iop.bio, NULL);
 825
 826                 bio_copy_data(s->cache_miss, s->iop.bio);
 827
 828                 bio_put(s->cache_miss);
 829                 s->cache_miss = NULL;
 830         }
 831
 832         if (verify(dc) && s->recoverable && !s->read_dirty_data)
 833                 bch_data_verify(dc, s->orig_bio);
 834
 835         bio_complete(s);
 836
 837         if (s->iop.bio &&
 838             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 839                 BUG_ON(!s->iop.replace);
 840                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 841         }
 842
 843         continue_at(cl, cached_dev_cache_miss_done, NULL);
 844 }
 845
 846 static void cached_dev_read_done_bh(struct closure *cl)
 847 {
 848         struct search *s = container_of(cl, struct search, cl);
 849         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 850
 851         bch_mark_cache_accounting(s->iop.c, s->d,
 852                                   !s->cache_missed, s->iop.bypass);
 853         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 854
 855         if (s->iop.status)
 856                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 857         else if (s->iop.bio || verify(dc))
 858                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 859         else
 860                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 861 }
 862
 863 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 864                                  struct bio *bio, unsigned int sectors)
 865 {
 866         int ret = MAP_CONTINUE;
 867         unsigned int reada = 0;
 868         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 869         struct bio *miss, *cache_bio;
 870
 871         s->cache_missed = 1;
 872
 873         if (s->cache_miss || s->iop.bypass) {
 874                 miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split);
 875                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 876                 goto out_submit;
 877         }
 878
 879         if (!(bio->bi_opf & REQ_RAHEAD) &&
 880             !(bio->bi_opf & REQ_META) &&
 881             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 882                 reada = min_t(sector_t, dc->readahead >> 9,
 883                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
 884
 885         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 886
 887         s->iop.replace_key = KEY(s->iop.inode,
 888                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 889                                  s->insert_bio_sectors);
 890
 891         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 892         if (ret)
 893                 return ret;
 894
 895         s->iop.replace = true;
 896
 897         miss = bio_next_split(bio, sectors, GFP_NOIO, &s->d->bio_split);
 898
 899         /* btree_search_recurse()'s btree iterator is no good anymore */
 900         ret = miss == bio ? MAP_DONE : -EINTR;
 901
 902         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 903                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 904                         &dc->disk.bio_split);
 905         if (!cache_bio)
 906                 goto out_submit;
 907
 908         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 909         bio_copy_dev(cache_bio, miss);
 910         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 911
 912         cache_bio->bi_end_io    = backing_request_endio;
 913         cache_bio->bi_private   = &s->cl;
 914
 915         bch_bio_map(cache_bio, NULL);
 916         if (bch_bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 917                 goto out_put;
 918
 919         if (reada)
 920                 bch_mark_cache_readahead(s->iop.c, s->d);
 921
 922         s->cache_miss   = miss;
 923         s->iop.bio      = cache_bio;
 924         bio_get(cache_bio);
 925         /* I/O request sent to backing device */
 926         closure_bio_submit(s->iop.c, cache_bio, &s->cl);
 927
 928         return ret;
 929 out_put:
 930         bio_put(cache_bio);
 931 out_submit:
 932         miss->bi_end_io         = backing_request_endio;
 933         miss->bi_private        = &s->cl;
 934         /* I/O request sent to backing device */
 935         closure_bio_submit(s->iop.c, miss, &s->cl);
 936         return ret;
 937 }
 938
 939 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 940 {
 941         struct closure *cl = &s->cl;
 942
 943         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 944         continue_at(cl, cached_dev_read_done_bh, NULL);
 945 }
 946
 947 /* Process writes */
 948
 949 static void cached_dev_write_complete(struct closure *cl)
 950 {
 951         struct search *s = container_of(cl, struct search, cl);
 952         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 953
 954         up_read_non_owner(&dc->writeback_lock);
 955         cached_dev_bio_complete(cl);
 956 }
 957
 958 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 959 {
 960         struct closure *cl = &s->cl;
 961         struct bio *bio = &s->bio.bio;
 962         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 963         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 964
 965         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 966
 967         down_read_non_owner(&dc->writeback_lock);
 968         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 969                 /*
 970                  * We overlap with some dirty data undergoing background
 971                  * writeback, force this write to writeback
 972                  */
 973                 s->iop.bypass = false;
 974                 s->iop.writeback = true;
 975         }
 976
 977         /*
 978          * Discards aren't _required_ to do anything, so skipping if
 979          * check_overlapping returned true is ok
 980          *
 981          * But check_overlapping drops dirty keys for which io hasn't started,
 982          * so we still want to call it.
 983          */
 984         if (bio_op(bio) == REQ_OP_DISCARD)
 985                 s->iop.bypass = true;
 986
 987         if (should_writeback(dc, s->orig_bio,
 988                              cache_mode(dc),
 989                              s->iop.bypass)) {
 990                 s->iop.bypass = false;
 991                 s->iop.writeback = true;
 992         }
 993
 994         if (s->iop.bypass) {
 995                 s->iop.bio = s->orig_bio;
 996                 bio_get(s->iop.bio);
 997
 998                 if (bio_op(bio) == REQ_OP_DISCARD &&
 999                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1000                         goto insert_data;
1001
1002                 /* I/O request sent to backing device */
1003                 bio->bi_end_io = backing_request_endio;
1004                 closure_bio_submit(s->iop.c, bio, cl);
1005
1006         } else if (s->iop.writeback) {
1007                 bch_writeback_add(dc);
1008                 s->iop.bio = bio;
1009
1010                 if (bio->bi_opf & REQ_PREFLUSH) {
1011                         /*
1012                          * Also need to send a flush to the backing
1013                          * device.
1014                          */
1015                         struct bio *flush;
1016
1017                         flush = bio_alloc_bioset(GFP_NOIO, 0,
1018                                                  &dc->disk.bio_split);
1019                         if (!flush) {
1020                                 s->iop.status = BLK_STS_RESOURCE;
1021                                 goto insert_data;
1022                         }
1023                         bio_copy_dev(flush, bio);
1024                         flush->bi_end_io = backing_request_endio;
1025                         flush->bi_private = cl;
1026                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1027                         /* I/O request sent to backing device */
1028                         closure_bio_submit(s->iop.c, flush, cl);
1029                 }
1030         } else {
1031                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, &dc->disk.bio_split);
1032                 /* I/O request sent to backing device */
1033                 bio->bi_end_io = backing_request_endio;
1034                 closure_bio_submit(s->iop.c, bio, cl);
1035         }
1036
1037 insert_data:
1038         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1039         continue_at(cl, cached_dev_write_complete, NULL);
1040 }
1041
1042 static void cached_dev_nodata(struct closure *cl)
1043 {
1044         struct search *s = container_of(cl, struct search, cl);
1045         struct bio *bio = &s->bio.bio;
1046
1047         if (s->iop.flush_journal)
1048                 bch_journal_meta(s->iop.c, cl);
1049
1050         /* If it's a flush, we send the flush to the backing device too */
1051         bio->bi_end_io = backing_request_endio;
1052         closure_bio_submit(s->iop.c, bio, cl);
1053
1054         continue_at(cl, cached_dev_bio_complete, NULL);
1055 }
1056
1057 struct detached_dev_io_private {
1058         struct bcache_device    *d;
1059         unsigned long           start_time;
1060         bio_end_io_t            *bi_end_io;
1061         void                    *bi_private;
1062 };
1063
1064 static void detached_dev_end_io(struct bio *bio)
1065 {
1066         struct detached_dev_io_private *ddip;
1067
1068         ddip = bio->bi_private;
1069         bio->bi_end_io = ddip->bi_end_io;
1070         bio->bi_private = ddip->bi_private;
1071
1072         generic_end_io_acct(ddip->d->disk->queue, bio_op(bio),
1073                             &ddip->d->disk->part0, ddip->start_time);
1074
1075         if (bio->bi_status) {
1076                 struct cached_dev *dc = container_of(ddip->d,
1077                                                      struct cached_dev, disk);
1078                 /* should count I/O error for backing device here */
1079                 bch_count_backing_io_errors(dc, bio);
1080         }
1081
1082         kfree(ddip);
1083         bio->bi_end_io(bio);
1084 }
1085
1086 static void detached_dev_do_request(struct bcache_device *d, struct bio *bio)
1087 {
1088         struct detached_dev_io_private *ddip;
1089         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1090
1091         /*
1092          * no need to call closure_get(&dc->disk.cl),
1093          * because upper layer had already opened bcache device,
1094          * which would call closure_get(&dc->disk.cl)
1095          */
1096         ddip = kzalloc(sizeof(struct detached_dev_io_private), GFP_NOIO);
1097         ddip->d = d;
1098         ddip->start_time = jiffies;
1099         ddip->bi_end_io = bio->bi_end_io;
1100         ddip->bi_private = bio->bi_private;
1101         bio->bi_end_io = detached_dev_end_io;
1102         bio->bi_private = ddip;
1103
1104         if ((bio_op(bio) == REQ_OP_DISCARD) &&
1105             !blk_queue_discard(bdev_get_queue(dc->bdev)))
1106                 bio->bi_end_io(bio);
1107         else
1108                 generic_make_request(bio);
1109 }
1110
1111 static void quit_max_writeback_rate(struct cache_set *c,
1112                                     struct cached_dev *this_dc)
1113 {
1114         int i;
1115         struct bcache_device *d;
1116         struct cached_dev *dc;
1117
1118         /*
1119          * mutex bch_register_lock may compete with other parallel requesters,
1120          * or attach/detach operations on other backing device. Waiting to
1121          * the mutex lock may increase I/O request latency for seconds or more.
1122          * To avoid such situation, if mutext_trylock() failed, only writeback
1123          * rate of current cached device is set to 1, and __update_write_back()
1124          * will decide writeback rate of other cached devices (remember now
1125          * c->idle_counter is 0 already).
1126          */
1127         if (mutex_trylock(&bch_register_lock)) {
1128                 for (i = 0; i < c->devices_max_used; i++) {
1129                         if (!c->devices[i])
1130                                 continue;
1131
1132                         if (UUID_FLASH_ONLY(&c->uuids[i]))
1133                                 continue;
1134
1135                         d = c->devices[i];
1136                         dc = container_of(d, struct cached_dev, disk);
1137                         /*
1138                          * set writeback rate to default minimum value,
1139                          * then let update_writeback_rate() to decide the
1140                          * upcoming rate.
1141                          */
1142                         atomic_long_set(&dc->writeback_rate.rate, 1);
1143                 }
1144                 mutex_unlock(&bch_register_lock);
1145         } else
1146                 atomic_long_set(&this_dc->writeback_rate.rate, 1);
1147 }
1148
1149 /* Cached devices - read & write stuff */
1150
1151 static blk_qc_t cached_dev_make_request(struct request_queue *q,
1152                                         struct bio *bio)
1153 {
1154         struct search *s;
1155         struct bcache_device *d = bio->bi_disk->private_data;
1156         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1157         int rw = bio_data_dir(bio);
1158
1159         if (unlikely((d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags)) ||
1160                      dc->io_disable)) {
1161                 bio->bi_status = BLK_STS_IOERR;
1162                 bio_endio(bio);
1163                 return BLK_QC_T_NONE;
1164         }
1165
1166         if (likely(d->c)) {
1167                 if (atomic_read(&d->c->idle_counter))
1168                         atomic_set(&d->c->idle_counter, 0);
1169                 /*
1170                  * If at_max_writeback_rate of cache set is true and new I/O
1171                  * comes, quit max writeback rate of all cached devices
1172                  * attached to this cache set, and set at_max_writeback_rate
1173                  * to false.
1174                  */
1175                 if (unlikely(atomic_read(&d->c->at_max_writeback_rate) == 1)) {
1176                         atomic_set(&d->c->at_max_writeback_rate, 0);
1177                         quit_max_writeback_rate(d->c, dc);
1178                 }
1179         }
1180
1181         generic_start_io_acct(q,
1182                               bio_op(bio),
1183                               bio_sectors(bio),
1184                               &d->disk->part0);
1185
1186         bio_set_dev(bio, dc->bdev);
1187         bio->bi_iter.bi_sector += dc->sb.data_offset;
1188
1189         if (cached_dev_get(dc)) {
1190                 s = search_alloc(bio, d);
1191                 trace_bcache_request_start(s->d, bio);
1192
1193                 if (!bio->bi_iter.bi_size) {
1194                         /*
1195                          * can't call bch_journal_meta from under
1196                          * generic_make_request
1197                          */
1198                         continue_at_nobarrier(&s->cl,
1199                                               cached_dev_nodata,
1200                                               bcache_wq);
1201                 } else {
1202                         s->iop.bypass = check_should_bypass(dc, bio);
1203
1204                         if (rw)
1205                                 cached_dev_write(dc, s);
1206                         else
1207                                 cached_dev_read(dc, s);
1208                 }
1209         } else
1210                 /* I/O request sent to backing device */
1211                 detached_dev_do_request(d, bio);
1212
1213         return BLK_QC_T_NONE;
1214 }
1215
1216 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1217                             unsigned int cmd, unsigned long arg)
1218 {
1219         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1220
1221         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1222 }
1223
1224 static int cached_dev_congested(void *data, int bits)
1225 {
1226         struct bcache_device *d = data;
1227         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1228         struct request_queue *q = bdev_get_queue(dc->bdev);
1229         int ret = 0;
1230
1231         if (bdi_congested(q->backing_dev_info, bits))
1232                 return 1;
1233
1234         if (cached_dev_get(dc)) {
1235                 unsigned int i;
1236                 struct cache *ca;
1237
1238                 for_each_cache(ca, d->c, i) {
1239                         q = bdev_get_queue(ca->bdev);
1240                         ret |= bdi_congested(q->backing_dev_info, bits);
1241                 }
1242
1243                 cached_dev_put(dc);
1244         }
1245
1246         return ret;
1247 }
1248
1249 void bch_cached_dev_request_init(struct cached_dev *dc)
1250 {
1251         struct gendisk *g = dc->disk.disk;
1252
1253         g->queue->make_request_fn               = cached_dev_make_request;
1254         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1255         dc->disk.cache_miss                     = cached_dev_cache_miss;
1256         dc->disk.ioctl                          = cached_dev_ioctl;
1257 }
1258
1259 /* Flash backed devices */
1260
1261 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1262                                 struct bio *bio, unsigned int sectors)
1263 {
1264         unsigned int bytes = min(sectors, bio_sectors(bio)) << 9;
1265
1266         swap(bio->bi_iter.bi_size, bytes);
1267         zero_fill_bio(bio);
1268         swap(bio->bi_iter.bi_size, bytes);
1269
1270         bio_advance(bio, bytes);
1271
1272         if (!bio->bi_iter.bi_size)
1273                 return MAP_DONE;
1274
1275         return MAP_CONTINUE;
1276 }
1277
1278 static void flash_dev_nodata(struct closure *cl)
1279 {
1280         struct search *s = container_of(cl, struct search, cl);
1281
1282         if (s->iop.flush_journal)
1283                 bch_journal_meta(s->iop.c, cl);
1284
1285         continue_at(cl, search_free, NULL);
1286 }
1287
1288 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1289                                              struct bio *bio)
1290 {
1291         struct search *s;
1292         struct closure *cl;
1293         struct bcache_device *d = bio->bi_disk->private_data;
1294
1295         if (unlikely(d->c && test_bit(CACHE_SET_IO_DISABLE, &d->c->flags))) {
1296                 bio->bi_status = BLK_STS_IOERR;
1297                 bio_endio(bio);
1298                 return BLK_QC_T_NONE;
1299         }
1300
1301         generic_start_io_acct(q, bio_op(bio), bio_sectors(bio), &d->disk->part0);
1302
1303         s = search_alloc(bio, d);
1304         cl = &s->cl;
1305         bio = &s->bio.bio;
1306
1307         trace_bcache_request_start(s->d, bio);
1308
1309         if (!bio->bi_iter.bi_size) {
1310                 /*
1311                  * can't call bch_journal_meta from under
1312                  * generic_make_request
1313                  */
1314                 continue_at_nobarrier(&s->cl,
1315                                       flash_dev_nodata,
1316                                       bcache_wq);
1317                 return BLK_QC_T_NONE;
1318         } else if (bio_data_dir(bio)) {
1319                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1320                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1321                                         &KEY(d->id, bio_end_sector(bio), 0));
1322
1323                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1324                 s->iop.writeback        = true;
1325                 s->iop.bio              = bio;
1326
1327                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1328         } else {
1329                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1330         }
1331
1332         continue_at(cl, search_free, NULL);
1333         return BLK_QC_T_NONE;
1334 }
1335
1336 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1337                            unsigned int cmd, unsigned long arg)
1338 {
1339         return -ENOTTY;
1340 }
1341
1342 static int flash_dev_congested(void *data, int bits)
1343 {
1344         struct bcache_device *d = data;
1345         struct request_queue *q;
1346         struct cache *ca;
1347         unsigned int i;
1348         int ret = 0;
1349
1350         for_each_cache(ca, d->c, i) {
1351                 q = bdev_get_queue(ca->bdev);
1352                 ret |= bdi_congested(q->backing_dev_info, bits);
1353         }
1354
1355         return ret;
1356 }
1357
1358 void bch_flash_dev_request_init(struct bcache_device *d)
1359 {
1360         struct gendisk *g = d->disk;
1361
1362         g->queue->make_request_fn               = flash_dev_make_request;
1363         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1364         d->cache_miss                           = flash_dev_cache_miss;
1365         d->ioctl                                = flash_dev_ioctl;
1366 }
1367
1368 void bch_request_exit(void)
1369 {
1370         kmem_cache_destroy(bch_search_cache);
1371 }
1372
1373 int __init bch_request_init(void)
1374 {
1375         bch_search_cache = KMEM_CACHE(search, 0);
1376         if (!bch_search_cache)
1377                 return -ENOMEM;
1378
1379         return 0;
1380 }