Pileus Git - ~andy/linux/blob - drivers/md/bcache/request.c

   1 /*
   2  * Main bcache entry point - handle a read or a write request and decide what to
   3  * do with it; the make_request functions are called by the block layer.
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcache.h"
  10 #include "btree.h"
  11 #include "debug.h"
  12 #include "request.h"
  13 #include "writeback.h"
  14
  15 #include <linux/cgroup.h>
  16 #include <linux/module.h>
  17 #include <linux/hash.h>
  18 #include <linux/random.h>
  19 #include "blk-cgroup.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 *);
  29
  30 /* Cgroup interface */
  31
  32 #ifdef CONFIG_CGROUP_BCACHE
  33 static struct bch_cgroup bcache_default_cgroup = { .cache_mode = -1 };
  34
  35 static struct bch_cgroup *cgroup_to_bcache(struct cgroup *cgroup)
  36 {
  37         struct cgroup_subsys_state *css;
  38         return cgroup &&
  39                 (css = cgroup_subsys_state(cgroup, bcache_subsys_id))
  40                 ? container_of(css, struct bch_cgroup, css)
  41                 : &bcache_default_cgroup;
  42 }
  43
  44 struct bch_cgroup *bch_bio_to_cgroup(struct bio *bio)
  45 {
  46         struct cgroup_subsys_state *css = bio->bi_css
  47                 ? cgroup_subsys_state(bio->bi_css->cgroup, bcache_subsys_id)
  48                 : task_subsys_state(current, bcache_subsys_id);
  49
  50         return css
  51                 ? container_of(css, struct bch_cgroup, css)
  52                 : &bcache_default_cgroup;
  53 }
  54
  55 static ssize_t cache_mode_read(struct cgroup *cgrp, struct cftype *cft,
  56                         struct file *file,
  57                         char __user *buf, size_t nbytes, loff_t *ppos)
  58 {
  59         char tmp[1024];
  60         int len = bch_snprint_string_list(tmp, PAGE_SIZE, bch_cache_modes,
  61                                           cgroup_to_bcache(cgrp)->cache_mode + 1);
  62
  63         if (len < 0)
  64                 return len;
  65
  66         return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
  67 }
  68
  69 static int cache_mode_write(struct cgroup *cgrp, struct cftype *cft,
  70                             const char *buf)
  71 {
  72         int v = bch_read_string_list(buf, bch_cache_modes);
  73         if (v < 0)
  74                 return v;
  75
  76         cgroup_to_bcache(cgrp)->cache_mode = v - 1;
  77         return 0;
  78 }
  79
  80 static u64 bch_verify_read(struct cgroup *cgrp, struct cftype *cft)
  81 {
  82         return cgroup_to_bcache(cgrp)->verify;
  83 }
  84
  85 static int bch_verify_write(struct cgroup *cgrp, struct cftype *cft, u64 val)
  86 {
  87         cgroup_to_bcache(cgrp)->verify = val;
  88         return 0;
  89 }
  90
  91 static u64 bch_cache_hits_read(struct cgroup *cgrp, struct cftype *cft)
  92 {
  93         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
  94         return atomic_read(&bcachecg->stats.cache_hits);
  95 }
  96
  97 static u64 bch_cache_misses_read(struct cgroup *cgrp, struct cftype *cft)
  98 {
  99         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 100         return atomic_read(&bcachecg->stats.cache_misses);
 101 }
 102
 103 static u64 bch_cache_bypass_hits_read(struct cgroup *cgrp,
 104                                          struct cftype *cft)
 105 {
 106         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 107         return atomic_read(&bcachecg->stats.cache_bypass_hits);
 108 }
 109
 110 static u64 bch_cache_bypass_misses_read(struct cgroup *cgrp,
 111                                            struct cftype *cft)
 112 {
 113         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 114         return atomic_read(&bcachecg->stats.cache_bypass_misses);
 115 }
 116
 117 static struct cftype bch_files[] = {
 118         {
 119                 .name           = "cache_mode",
 120                 .read           = cache_mode_read,
 121                 .write_string   = cache_mode_write,
 122         },
 123         {
 124                 .name           = "verify",
 125                 .read_u64       = bch_verify_read,
 126                 .write_u64      = bch_verify_write,
 127         },
 128         {
 129                 .name           = "cache_hits",
 130                 .read_u64       = bch_cache_hits_read,
 131         },
 132         {
 133                 .name           = "cache_misses",
 134                 .read_u64       = bch_cache_misses_read,
 135         },
 136         {
 137                 .name           = "cache_bypass_hits",
 138                 .read_u64       = bch_cache_bypass_hits_read,
 139         },
 140         {
 141                 .name           = "cache_bypass_misses",
 142                 .read_u64       = bch_cache_bypass_misses_read,
 143         },
 144         { }     /* terminate */
 145 };
 146
 147 static void init_bch_cgroup(struct bch_cgroup *cg)
 148 {
 149         cg->cache_mode = -1;
 150 }
 151
 152 static struct cgroup_subsys_state *bcachecg_create(struct cgroup *cgroup)
 153 {
 154         struct bch_cgroup *cg;
 155
 156         cg = kzalloc(sizeof(*cg), GFP_KERNEL);
 157         if (!cg)
 158                 return ERR_PTR(-ENOMEM);
 159         init_bch_cgroup(cg);
 160         return &cg->css;
 161 }
 162
 163 static void bcachecg_destroy(struct cgroup *cgroup)
 164 {
 165         struct bch_cgroup *cg = cgroup_to_bcache(cgroup);
 166         kfree(cg);
 167 }
 168
 169 struct cgroup_subsys bcache_subsys = {
 170         .create         = bcachecg_create,
 171         .destroy        = bcachecg_destroy,
 172         .subsys_id      = bcache_subsys_id,
 173         .name           = "bcache",
 174         .module         = THIS_MODULE,
 175 };
 176 EXPORT_SYMBOL_GPL(bcache_subsys);
 177 #endif
 178
 179 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
 180 {
 181 #ifdef CONFIG_CGROUP_BCACHE
 182         int r = bch_bio_to_cgroup(bio)->cache_mode;
 183         if (r >= 0)
 184                 return r;
 185 #endif
 186         return BDEV_CACHE_MODE(&dc->sb);
 187 }
 188
 189 static bool verify(struct cached_dev *dc, struct bio *bio)
 190 {
 191 #ifdef CONFIG_CGROUP_BCACHE
 192         if (bch_bio_to_cgroup(bio)->verify)
 193                 return true;
 194 #endif
 195         return dc->verify;
 196 }
 197
 198 static void bio_csum(struct bio *bio, struct bkey *k)
 199 {
 200         struct bio_vec bv;
 201         struct bvec_iter iter;
 202         uint64_t csum = 0;
 203
 204         bio_for_each_segment(bv, bio, iter) {
 205                 void *d = kmap(bv.bv_page) + bv.bv_offset;
 206                 csum = bch_crc64_update(csum, d, bv.bv_len);
 207                 kunmap(bv.bv_page);
 208         }
 209
 210         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
 211 }
 212
 213 /* Insert data into cache */
 214
 215 static void bch_data_insert_keys(struct closure *cl)
 216 {
 217         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 218         atomic_t *journal_ref = NULL;
 219         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
 220         int ret;
 221
 222         /*
 223          * If we're looping, might already be waiting on
 224          * another journal write - can't wait on more than one journal write at
 225          * a time
 226          *
 227          * XXX: this looks wrong
 228          */
 229 #if 0
 230         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
 231                 closure_sync(&s->cl);
 232 #endif
 233
 234         if (!op->replace)
 235                 journal_ref = bch_journal(op->c, &op->insert_keys,
 236                                           op->flush_journal ? cl : NULL);
 237
 238         ret = bch_btree_insert(op->c, &op->insert_keys,
 239                                journal_ref, replace_key);
 240         if (ret == -ESRCH) {
 241                 op->replace_collision = true;
 242         } else if (ret) {
 243                 op->error               = -ENOMEM;
 244                 op->insert_data_done    = true;
 245         }
 246
 247         if (journal_ref)
 248                 atomic_dec_bug(journal_ref);
 249
 250         if (!op->insert_data_done)
 251                 continue_at(cl, bch_data_insert_start, bcache_wq);
 252
 253         bch_keylist_free(&op->insert_keys);
 254         closure_return(cl);
 255 }
 256
 257 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
 258                                struct cache_set *c)
 259 {
 260         size_t oldsize = bch_keylist_nkeys(l);
 261         size_t newsize = oldsize + u64s;
 262
 263         /*
 264          * The journalling code doesn't handle the case where the keys to insert
 265          * is bigger than an empty write: If we just return -ENOMEM here,
 266          * bio_insert() and bio_invalidate() will insert the keys created so far
 267          * and finish the rest when the keylist is empty.
 268          */
 269         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
 270                 return -ENOMEM;
 271
 272         return __bch_keylist_realloc(l, u64s);
 273 }
 274
 275 static void bch_data_invalidate(struct closure *cl)
 276 {
 277         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 278         struct bio *bio = op->bio;
 279
 280         pr_debug("invalidating %i sectors from %llu",
 281                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 282
 283         while (bio_sectors(bio)) {
 284                 unsigned sectors = min(bio_sectors(bio),
 285                                        1U << (KEY_SIZE_BITS - 1));
 286
 287                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
 288                         goto out;
 289
 290                 bio->bi_iter.bi_sector  += sectors;
 291                 bio->bi_iter.bi_size    -= sectors << 9;
 292
 293                 bch_keylist_add(&op->insert_keys,
 294                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
 295         }
 296
 297         op->insert_data_done = true;
 298         bio_put(bio);
 299 out:
 300         continue_at(cl, bch_data_insert_keys, bcache_wq);
 301 }
 302
 303 static void bch_data_insert_error(struct closure *cl)
 304 {
 305         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 306
 307         /*
 308          * Our data write just errored, which means we've got a bunch of keys to
 309          * insert that point to data that wasn't succesfully written.
 310          *
 311          * We don't have to insert those keys but we still have to invalidate
 312          * that region of the cache - so, if we just strip off all the pointers
 313          * from the keys we'll accomplish just that.
 314          */
 315
 316         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 317
 318         while (src != op->insert_keys.top) {
 319                 struct bkey *n = bkey_next(src);
 320
 321                 SET_KEY_PTRS(src, 0);
 322                 memmove(dst, src, bkey_bytes(src));
 323
 324                 dst = bkey_next(dst);
 325                 src = n;
 326         }
 327
 328         op->insert_keys.top = dst;
 329
 330         bch_data_insert_keys(cl);
 331 }
 332
 333 static void bch_data_insert_endio(struct bio *bio, int error)
 334 {
 335         struct closure *cl = bio->bi_private;
 336         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 337
 338         if (error) {
 339                 /* TODO: We could try to recover from this. */
 340                 if (op->writeback)
 341                         op->error = error;
 342                 else if (!op->replace)
 343                         set_closure_fn(cl, bch_data_insert_error, bcache_wq);
 344                 else
 345                         set_closure_fn(cl, NULL, NULL);
 346         }
 347
 348         bch_bbio_endio(op->c, bio, error, "writing data to cache");
 349 }
 350
 351 static void bch_data_insert_start(struct closure *cl)
 352 {
 353         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 354         struct bio *bio = op->bio, *n;
 355
 356         if (op->bypass)
 357                 return bch_data_invalidate(cl);
 358
 359         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) {
 360                 set_gc_sectors(op->c);
 361                 wake_up_gc(op->c);
 362         }
 363
 364         /*
 365          * Journal writes are marked REQ_FLUSH; if the original write was a
 366          * flush, it'll wait on the journal write.
 367          */
 368         bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA);
 369
 370         do {
 371                 unsigned i;
 372                 struct bkey *k;
 373                 struct bio_set *split = op->c->bio_split;
 374
 375                 /* 1 for the device pointer and 1 for the chksum */
 376                 if (bch_keylist_realloc(&op->insert_keys,
 377                                         3 + (op->csum ? 1 : 0),
 378                                         op->c))
 379                         continue_at(cl, bch_data_insert_keys, bcache_wq);
 380
 381                 k = op->insert_keys.top;
 382                 bkey_init(k);
 383                 SET_KEY_INODE(k, op->inode);
 384                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 385
 386                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 387                                        op->write_point, op->write_prio,
 388                                        op->writeback))
 389                         goto err;
 390
 391                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 392
 393                 n->bi_end_io    = bch_data_insert_endio;
 394                 n->bi_private   = cl;
 395
 396                 if (op->writeback) {
 397                         SET_KEY_DIRTY(k, true);
 398
 399                         for (i = 0; i < KEY_PTRS(k); i++)
 400                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 401                                             GC_MARK_DIRTY);
 402                 }
 403
 404                 SET_KEY_CSUM(k, op->csum);
 405                 if (KEY_CSUM(k))
 406                         bio_csum(n, k);
 407
 408                 trace_bcache_cache_insert(k);
 409                 bch_keylist_push(&op->insert_keys);
 410
 411                 n->bi_rw |= REQ_WRITE;
 412                 bch_submit_bbio(n, op->c, k, 0);
 413         } while (n != bio);
 414
 415         op->insert_data_done = true;
 416         continue_at(cl, bch_data_insert_keys, bcache_wq);
 417 err:
 418         /* bch_alloc_sectors() blocks if s->writeback = true */
 419         BUG_ON(op->writeback);
 420
 421         /*
 422          * But if it's not a writeback write we'd rather just bail out if
 423          * there aren't any buckets ready to write to - it might take awhile and
 424          * we might be starving btree writes for gc or something.
 425          */
 426
 427         if (!op->replace) {
 428                 /*
 429                  * Writethrough write: We can't complete the write until we've
 430                  * updated the index. But we don't want to delay the write while
 431                  * we wait for buckets to be freed up, so just invalidate the
 432                  * rest of the write.
 433                  */
 434                 op->bypass = true;
 435                 return bch_data_invalidate(cl);
 436         } else {
 437                 /*
 438                  * From a cache miss, we can just insert the keys for the data
 439                  * we have written or bail out if we didn't do anything.
 440                  */
 441                 op->insert_data_done = true;
 442                 bio_put(bio);
 443
 444                 if (!bch_keylist_empty(&op->insert_keys))
 445                         continue_at(cl, bch_data_insert_keys, bcache_wq);
 446                 else
 447                         closure_return(cl);
 448         }
 449 }
 450
 451 /**
 452  * bch_data_insert - stick some data in the cache
 453  *
 454  * This is the starting point for any data to end up in a cache device; it could
 455  * be from a normal write, or a writeback write, or a write to a flash only
 456  * volume - it's also used by the moving garbage collector to compact data in
 457  * mostly empty buckets.
 458  *
 459  * It first writes the data to the cache, creating a list of keys to be inserted
 460  * (if the data had to be fragmented there will be multiple keys); after the
 461  * data is written it calls bch_journal, and after the keys have been added to
 462  * the next journal write they're inserted into the btree.
 463  *
 464  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 465  * and op->inode is used for the key inode.
 466  *
 467  * If s->bypass is true, instead of inserting the data it invalidates the
 468  * region of the cache represented by s->cache_bio and op->inode.
 469  */
 470 void bch_data_insert(struct closure *cl)
 471 {
 472         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 473
 474         trace_bcache_write(op->bio, op->writeback, op->bypass);
 475
 476         bch_keylist_init(&op->insert_keys);
 477         bio_get(op->bio);
 478         bch_data_insert_start(cl);
 479 }
 480
 481 /* Congested? */
 482
 483 unsigned bch_get_congested(struct cache_set *c)
 484 {
 485         int i;
 486         long rand;
 487
 488         if (!c->congested_read_threshold_us &&
 489             !c->congested_write_threshold_us)
 490                 return 0;
 491
 492         i = (local_clock_us() - c->congested_last_us) / 1024;
 493         if (i < 0)
 494                 return 0;
 495
 496         i += atomic_read(&c->congested);
 497         if (i >= 0)
 498                 return 0;
 499
 500         i += CONGESTED_MAX;
 501
 502         if (i > 0)
 503                 i = fract_exp_two(i, 6);
 504
 505         rand = get_random_int();
 506         i -= bitmap_weight(&rand, BITS_PER_LONG);
 507
 508         return i > 0 ? i : 1;
 509 }
 510
 511 static void add_sequential(struct task_struct *t)
 512 {
 513         ewma_add(t->sequential_io_avg,
 514                  t->sequential_io, 8, 0);
 515
 516         t->sequential_io = 0;
 517 }
 518
 519 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 520 {
 521         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 522 }
 523
 524 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 525 {
 526         struct cache_set *c = dc->disk.c;
 527         unsigned mode = cache_mode(dc, bio);
 528         unsigned sectors, congested = bch_get_congested(c);
 529         struct task_struct *task = current;
 530         struct io *i;
 531
 532         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 533             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 534             (bio->bi_rw & REQ_DISCARD))
 535                 goto skip;
 536
 537         if (mode == CACHE_MODE_NONE ||
 538             (mode == CACHE_MODE_WRITEAROUND &&
 539              (bio->bi_rw & REQ_WRITE)))
 540                 goto skip;
 541
 542         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 543             bio_sectors(bio) & (c->sb.block_size - 1)) {
 544                 pr_debug("skipping unaligned io");
 545                 goto skip;
 546         }
 547
 548         if (bypass_torture_test(dc)) {
 549                 if ((get_random_int() & 3) == 3)
 550                         goto skip;
 551                 else
 552                         goto rescale;
 553         }
 554
 555         if (!congested && !dc->sequential_cutoff)
 556                 goto rescale;
 557
 558         if (!congested &&
 559             mode == CACHE_MODE_WRITEBACK &&
 560             (bio->bi_rw & REQ_WRITE) &&
 561             (bio->bi_rw & REQ_SYNC))
 562                 goto rescale;
 563
 564         spin_lock(&dc->io_lock);
 565
 566         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
 567                 if (i->last == bio->bi_iter.bi_sector &&
 568                     time_before(jiffies, i->jiffies))
 569                         goto found;
 570
 571         i = list_first_entry(&dc->io_lru, struct io, lru);
 572
 573         add_sequential(task);
 574         i->sequential = 0;
 575 found:
 576         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
 577                 i->sequential   += bio->bi_iter.bi_size;
 578
 579         i->last                  = bio_end_sector(bio);
 580         i->jiffies               = jiffies + msecs_to_jiffies(5000);
 581         task->sequential_io      = i->sequential;
 582
 583         hlist_del(&i->hash);
 584         hlist_add_head(&i->hash, iohash(dc, i->last));
 585         list_move_tail(&i->lru, &dc->io_lru);
 586
 587         spin_unlock(&dc->io_lock);
 588
 589         sectors = max(task->sequential_io,
 590                       task->sequential_io_avg) >> 9;
 591
 592         if (dc->sequential_cutoff &&
 593             sectors >= dc->sequential_cutoff >> 9) {
 594                 trace_bcache_bypass_sequential(bio);
 595                 goto skip;
 596         }
 597
 598         if (congested && sectors >= congested) {
 599                 trace_bcache_bypass_congested(bio);
 600                 goto skip;
 601         }
 602
 603 rescale:
 604         bch_rescale_priorities(c, bio_sectors(bio));
 605         return false;
 606 skip:
 607         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 608         return true;
 609 }
 610
 611 /* Cache lookup */
 612
 613 struct search {
 614         /* Stack frame for bio_complete */
 615         struct closure          cl;
 616
 617         struct bbio             bio;
 618         struct bio              *orig_bio;
 619         struct bio              *cache_miss;
 620         struct bcache_device    *d;
 621
 622         unsigned                insert_bio_sectors;
 623         unsigned                recoverable:1;
 624         unsigned                write:1;
 625         unsigned                read_dirty_data:1;
 626
 627         unsigned long           start_time;
 628
 629         struct btree_op         op;
 630         struct data_insert_op   iop;
 631 };
 632
 633 static void bch_cache_read_endio(struct bio *bio, int error)
 634 {
 635         struct bbio *b = container_of(bio, struct bbio, bio);
 636         struct closure *cl = bio->bi_private;
 637         struct search *s = container_of(cl, struct search, cl);
 638
 639         /*
 640          * If the bucket was reused while our bio was in flight, we might have
 641          * read the wrong data. Set s->error but not error so it doesn't get
 642          * counted against the cache device, but we'll still reread the data
 643          * from the backing device.
 644          */
 645
 646         if (error)
 647                 s->iop.error = error;
 648         else if (!KEY_DIRTY(&b->key) &&
 649                  ptr_stale(s->iop.c, &b->key, 0)) {
 650                 atomic_long_inc(&s->iop.c->cache_read_races);
 651                 s->iop.error = -EINTR;
 652         }
 653
 654         bch_bbio_endio(s->iop.c, bio, error, "reading from cache");
 655 }
 656
 657 /*
 658  * Read from a single key, handling the initial cache miss if the key starts in
 659  * the middle of the bio
 660  */
 661 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 662 {
 663         struct search *s = container_of(op, struct search, op);
 664         struct bio *n, *bio = &s->bio.bio;
 665         struct bkey *bio_key;
 666         unsigned ptr;
 667
 668         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 669                 return MAP_CONTINUE;
 670
 671         if (KEY_INODE(k) != s->iop.inode ||
 672             KEY_START(k) > bio->bi_iter.bi_sector) {
 673                 unsigned bio_sectors = bio_sectors(bio);
 674                 unsigned sectors = KEY_INODE(k) == s->iop.inode
 675                         ? min_t(uint64_t, INT_MAX,
 676                                 KEY_START(k) - bio->bi_iter.bi_sector)
 677                         : INT_MAX;
 678
 679                 int ret = s->d->cache_miss(b, s, bio, sectors);
 680                 if (ret != MAP_CONTINUE)
 681                         return ret;
 682
 683                 /* if this was a complete miss we shouldn't get here */
 684                 BUG_ON(bio_sectors <= sectors);
 685         }
 686
 687         if (!KEY_SIZE(k))
 688                 return MAP_CONTINUE;
 689
 690         /* XXX: figure out best pointer - for multiple cache devices */
 691         ptr = 0;
 692
 693         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 694
 695         if (KEY_DIRTY(k))
 696                 s->read_dirty_data = true;
 697
 698         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 699                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 700                            GFP_NOIO, s->d->bio_split);
 701
 702         bio_key = &container_of(n, struct bbio, bio)->key;
 703         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 704
 705         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 706         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 707
 708         n->bi_end_io    = bch_cache_read_endio;
 709         n->bi_private   = &s->cl;
 710
 711         /*
 712          * The bucket we're reading from might be reused while our bio
 713          * is in flight, and we could then end up reading the wrong
 714          * data.
 715          *
 716          * We guard against this by checking (in cache_read_endio()) if
 717          * the pointer is stale again; if so, we treat it as an error
 718          * and reread from the backing device (but we don't pass that
 719          * error up anywhere).
 720          */
 721
 722         __bch_submit_bbio(n, b->c);
 723         return n == bio ? MAP_DONE : MAP_CONTINUE;
 724 }
 725
 726 static void cache_lookup(struct closure *cl)
 727 {
 728         struct search *s = container_of(cl, struct search, iop.cl);
 729         struct bio *bio = &s->bio.bio;
 730         int ret;
 731
 732         bch_btree_op_init(&s->op, -1);
 733
 734         ret = bch_btree_map_keys(&s->op, s->iop.c,
 735                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 736                                  cache_lookup_fn, MAP_END_KEY);
 737         if (ret == -EAGAIN)
 738                 continue_at(cl, cache_lookup, bcache_wq);
 739
 740         closure_return(cl);
 741 }
 742
 743 /* Common code for the make_request functions */
 744
 745 static void request_endio(struct bio *bio, int error)
 746 {
 747         struct closure *cl = bio->bi_private;
 748
 749         if (error) {
 750                 struct search *s = container_of(cl, struct search, cl);
 751                 s->iop.error = error;
 752                 /* Only cache read errors are recoverable */
 753                 s->recoverable = false;
 754         }
 755
 756         bio_put(bio);
 757         closure_put(cl);
 758 }
 759
 760 static void bio_complete(struct search *s)
 761 {
 762         if (s->orig_bio) {
 763                 int cpu, rw = bio_data_dir(s->orig_bio);
 764                 unsigned long duration = jiffies - s->start_time;
 765
 766                 cpu = part_stat_lock();
 767                 part_round_stats(cpu, &s->d->disk->part0);
 768                 part_stat_add(cpu, &s->d->disk->part0, ticks[rw], duration);
 769                 part_stat_unlock();
 770
 771                 trace_bcache_request_end(s->d, s->orig_bio);
 772                 bio_endio(s->orig_bio, s->iop.error);
 773                 s->orig_bio = NULL;
 774         }
 775 }
 776
 777 static void do_bio_hook(struct search *s, struct bio *orig_bio)
 778 {
 779         struct bio *bio = &s->bio.bio;
 780
 781         bio_init(bio);
 782         __bio_clone_fast(bio, orig_bio);
 783         bio->bi_end_io          = request_endio;
 784         bio->bi_private         = &s->cl;
 785
 786         atomic_set(&bio->bi_cnt, 3);
 787 }
 788
 789 static void search_free(struct closure *cl)
 790 {
 791         struct search *s = container_of(cl, struct search, cl);
 792         bio_complete(s);
 793
 794         if (s->iop.bio)
 795                 bio_put(s->iop.bio);
 796
 797         closure_debug_destroy(cl);
 798         mempool_free(s, s->d->c->search);
 799 }
 800
 801 static inline struct search *search_alloc(struct bio *bio,
 802                                           struct bcache_device *d)
 803 {
 804         struct search *s;
 805
 806         s = mempool_alloc(d->c->search, GFP_NOIO);
 807
 808         closure_init(&s->cl, NULL);
 809         do_bio_hook(s, bio);
 810
 811         s->orig_bio             = bio;
 812         s->cache_miss           = NULL;
 813         s->d                    = d;
 814         s->recoverable          = 1;
 815         s->write                = (bio->bi_rw & REQ_WRITE) != 0;
 816         s->read_dirty_data      = 0;
 817         s->start_time           = jiffies;
 818
 819         s->iop.c                = d->c;
 820         s->iop.bio              = NULL;
 821         s->iop.inode            = d->id;
 822         s->iop.write_point      = hash_long((unsigned long) current, 16);
 823         s->iop.write_prio       = 0;
 824         s->iop.error            = 0;
 825         s->iop.flags            = 0;
 826         s->iop.flush_journal    = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0;
 827
 828         return s;
 829 }
 830
 831 /* Cached devices */
 832
 833 static void cached_dev_bio_complete(struct closure *cl)
 834 {
 835         struct search *s = container_of(cl, struct search, cl);
 836         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 837
 838         search_free(cl);
 839         cached_dev_put(dc);
 840 }
 841
 842 /* Process reads */
 843
 844 static void cached_dev_cache_miss_done(struct closure *cl)
 845 {
 846         struct search *s = container_of(cl, struct search, cl);
 847
 848         if (s->iop.replace_collision)
 849                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 850
 851         if (s->iop.bio) {
 852                 int i;
 853                 struct bio_vec *bv;
 854
 855                 bio_for_each_segment_all(bv, s->iop.bio, i)
 856                         __free_page(bv->bv_page);
 857         }
 858
 859         cached_dev_bio_complete(cl);
 860 }
 861
 862 static void cached_dev_read_error(struct closure *cl)
 863 {
 864         struct search *s = container_of(cl, struct search, cl);
 865         struct bio *bio = &s->bio.bio;
 866
 867         if (s->recoverable) {
 868                 /* Retry from the backing device: */
 869                 trace_bcache_read_retry(s->orig_bio);
 870
 871                 s->iop.error = 0;
 872                 do_bio_hook(s, s->orig_bio);
 873
 874                 /* XXX: invalidate cache */
 875
 876                 closure_bio_submit(bio, cl, s->d);
 877         }
 878
 879         continue_at(cl, cached_dev_cache_miss_done, NULL);
 880 }
 881
 882 static void cached_dev_read_done(struct closure *cl)
 883 {
 884         struct search *s = container_of(cl, struct search, cl);
 885         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 886
 887         /*
 888          * We had a cache miss; cache_bio now contains data ready to be inserted
 889          * into the cache.
 890          *
 891          * First, we copy the data we just read from cache_bio's bounce buffers
 892          * to the buffers the original bio pointed to:
 893          */
 894
 895         if (s->iop.bio) {
 896                 bio_reset(s->iop.bio);
 897                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 898                 s->iop.bio->bi_bdev = s->cache_miss->bi_bdev;
 899                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 900                 bch_bio_map(s->iop.bio, NULL);
 901
 902                 bio_copy_data(s->cache_miss, s->iop.bio);
 903
 904                 bio_put(s->cache_miss);
 905                 s->cache_miss = NULL;
 906         }
 907
 908         if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
 909                 bch_data_verify(dc, s->orig_bio);
 910
 911         bio_complete(s);
 912
 913         if (s->iop.bio &&
 914             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 915                 BUG_ON(!s->iop.replace);
 916                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 917         }
 918
 919         continue_at(cl, cached_dev_cache_miss_done, NULL);
 920 }
 921
 922 static void cached_dev_read_done_bh(struct closure *cl)
 923 {
 924         struct search *s = container_of(cl, struct search, cl);
 925         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 926
 927         bch_mark_cache_accounting(s->iop.c, s->d,
 928                                   !s->cache_miss, s->iop.bypass);
 929         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 930
 931         if (s->iop.error)
 932                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 933         else if (s->iop.bio || verify(dc, &s->bio.bio))
 934                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 935         else
 936                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 937 }
 938
 939 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 940                                  struct bio *bio, unsigned sectors)
 941 {
 942         int ret = MAP_CONTINUE;
 943         unsigned reada = 0;
 944         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 945         struct bio *miss, *cache_bio;
 946
 947         if (s->cache_miss || s->iop.bypass) {
 948                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 949                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 950                 goto out_submit;
 951         }
 952
 953         if (!(bio->bi_rw & REQ_RAHEAD) &&
 954             !(bio->bi_rw & REQ_META) &&
 955             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 956                 reada = min_t(sector_t, dc->readahead >> 9,
 957                               bdev_sectors(bio->bi_bdev) - bio_end_sector(bio));
 958
 959         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 960
 961         s->iop.replace_key = KEY(s->iop.inode,
 962                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 963                                  s->insert_bio_sectors);
 964
 965         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 966         if (ret)
 967                 return ret;
 968
 969         s->iop.replace = true;
 970
 971         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 972
 973         /* btree_search_recurse()'s btree iterator is no good anymore */
 974         ret = miss == bio ? MAP_DONE : -EINTR;
 975
 976         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 977                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 978                         dc->disk.bio_split);
 979         if (!cache_bio)
 980                 goto out_submit;
 981
 982         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 983         cache_bio->bi_bdev              = miss->bi_bdev;
 984         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 985
 986         cache_bio->bi_end_io    = request_endio;
 987         cache_bio->bi_private   = &s->cl;
 988
 989         bch_bio_map(cache_bio, NULL);
 990         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 991                 goto out_put;
 992
 993         if (reada)
 994                 bch_mark_cache_readahead(s->iop.c, s->d);
 995
 996         s->cache_miss   = miss;
 997         s->iop.bio      = cache_bio;
 998         bio_get(cache_bio);
 999         closure_bio_submit(cache_bio, &s->cl, s->d);
1000
1001         return ret;
1002 out_put:
1003         bio_put(cache_bio);
1004 out_submit:
1005         miss->bi_end_io         = request_endio;
1006         miss->bi_private        = &s->cl;
1007         closure_bio_submit(miss, &s->cl, s->d);
1008         return ret;
1009 }
1010
1011 static void cached_dev_read(struct cached_dev *dc, struct search *s)
1012 {
1013         struct closure *cl = &s->cl;
1014
1015         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1016         continue_at(cl, cached_dev_read_done_bh, NULL);
1017 }
1018
1019 /* Process writes */
1020
1021 static void cached_dev_write_complete(struct closure *cl)
1022 {
1023         struct search *s = container_of(cl, struct search, cl);
1024         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
1025
1026         up_read_non_owner(&dc->writeback_lock);
1027         cached_dev_bio_complete(cl);
1028 }
1029
1030 static void cached_dev_write(struct cached_dev *dc, struct search *s)
1031 {
1032         struct closure *cl = &s->cl;
1033         struct bio *bio = &s->bio.bio;
1034         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
1035         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
1036
1037         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
1038
1039         down_read_non_owner(&dc->writeback_lock);
1040         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
1041                 /*
1042                  * We overlap with some dirty data undergoing background
1043                  * writeback, force this write to writeback
1044                  */
1045                 s->iop.bypass = false;
1046                 s->iop.writeback = true;
1047         }
1048
1049         /*
1050          * Discards aren't _required_ to do anything, so skipping if
1051          * check_overlapping returned true is ok
1052          *
1053          * But check_overlapping drops dirty keys for which io hasn't started,
1054          * so we still want to call it.
1055          */
1056         if (bio->bi_rw & REQ_DISCARD)
1057                 s->iop.bypass = true;
1058
1059         if (should_writeback(dc, s->orig_bio,
1060                              cache_mode(dc, bio),
1061                              s->iop.bypass)) {
1062                 s->iop.bypass = false;
1063                 s->iop.writeback = true;
1064         }
1065
1066         if (s->iop.bypass) {
1067                 s->iop.bio = s->orig_bio;
1068                 bio_get(s->iop.bio);
1069
1070                 if (!(bio->bi_rw & REQ_DISCARD) ||
1071                     blk_queue_discard(bdev_get_queue(dc->bdev)))
1072                         closure_bio_submit(bio, cl, s->d);
1073         } else if (s->iop.writeback) {
1074                 bch_writeback_add(dc);
1075                 s->iop.bio = bio;
1076
1077                 if (bio->bi_rw & REQ_FLUSH) {
1078                         /* Also need to send a flush to the backing device */
1079                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
1080                                                              dc->disk.bio_split);
1081
1082                         flush->bi_rw    = WRITE_FLUSH;
1083                         flush->bi_bdev  = bio->bi_bdev;
1084                         flush->bi_end_io = request_endio;
1085                         flush->bi_private = cl;
1086
1087                         closure_bio_submit(flush, cl, s->d);
1088                 }
1089         } else {
1090                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
1091
1092                 closure_bio_submit(bio, cl, s->d);
1093         }
1094
1095         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1096         continue_at(cl, cached_dev_write_complete, NULL);
1097 }
1098
1099 static void cached_dev_nodata(struct closure *cl)
1100 {
1101         struct search *s = container_of(cl, struct search, cl);
1102         struct bio *bio = &s->bio.bio;
1103
1104         if (s->iop.flush_journal)
1105                 bch_journal_meta(s->iop.c, cl);
1106
1107         /* If it's a flush, we send the flush to the backing device too */
1108         closure_bio_submit(bio, cl, s->d);
1109
1110         continue_at(cl, cached_dev_bio_complete, NULL);
1111 }
1112
1113 /* Cached devices - read & write stuff */
1114
1115 static void cached_dev_make_request(struct request_queue *q, struct bio *bio)
1116 {
1117         struct search *s;
1118         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1119         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1120         int cpu, rw = bio_data_dir(bio);
1121
1122         cpu = part_stat_lock();
1123         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1124         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1125         part_stat_unlock();
1126
1127         bio->bi_bdev = dc->bdev;
1128         bio->bi_iter.bi_sector += dc->sb.data_offset;
1129
1130         if (cached_dev_get(dc)) {
1131                 s = search_alloc(bio, d);
1132                 trace_bcache_request_start(s->d, bio);
1133
1134                 if (!bio->bi_iter.bi_size) {
1135                         /*
1136                          * can't call bch_journal_meta from under
1137                          * generic_make_request
1138                          */
1139                         continue_at_nobarrier(&s->cl,
1140                                               cached_dev_nodata,
1141                                               bcache_wq);
1142                 } else {
1143                         s->iop.bypass = check_should_bypass(dc, bio);
1144
1145                         if (rw)
1146                                 cached_dev_write(dc, s);
1147                         else
1148                                 cached_dev_read(dc, s);
1149                 }
1150         } else {
1151                 if ((bio->bi_rw & REQ_DISCARD) &&
1152                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1153                         bio_endio(bio, 0);
1154                 else
1155                         bch_generic_make_request(bio, &d->bio_split_hook);
1156         }
1157 }
1158
1159 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1160                             unsigned int cmd, unsigned long arg)
1161 {
1162         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1163         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1164 }
1165
1166 static int cached_dev_congested(void *data, int bits)
1167 {
1168         struct bcache_device *d = data;
1169         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1170         struct request_queue *q = bdev_get_queue(dc->bdev);
1171         int ret = 0;
1172
1173         if (bdi_congested(&q->backing_dev_info, bits))
1174                 return 1;
1175
1176         if (cached_dev_get(dc)) {
1177                 unsigned i;
1178                 struct cache *ca;
1179
1180                 for_each_cache(ca, d->c, i) {
1181                         q = bdev_get_queue(ca->bdev);
1182                         ret |= bdi_congested(&q->backing_dev_info, bits);
1183                 }
1184
1185                 cached_dev_put(dc);
1186         }
1187
1188         return ret;
1189 }
1190
1191 void bch_cached_dev_request_init(struct cached_dev *dc)
1192 {
1193         struct gendisk *g = dc->disk.disk;
1194
1195         g->queue->make_request_fn               = cached_dev_make_request;
1196         g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1197         dc->disk.cache_miss                     = cached_dev_cache_miss;
1198         dc->disk.ioctl                          = cached_dev_ioctl;
1199 }
1200
1201 /* Flash backed devices */
1202
1203 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1204                                 struct bio *bio, unsigned sectors)
1205 {
1206         struct bio_vec bv;
1207         struct bvec_iter iter;
1208
1209         /* Zero fill bio */
1210
1211         bio_for_each_segment(bv, bio, iter) {
1212                 unsigned j = min(bv.bv_len >> 9, sectors);
1213
1214                 void *p = kmap(bv.bv_page);
1215                 memset(p + bv.bv_offset, 0, j << 9);
1216                 kunmap(bv.bv_page);
1217
1218                 sectors -= j;
1219         }
1220
1221         bio_advance(bio, min(sectors << 9, bio->bi_iter.bi_size));
1222
1223         if (!bio->bi_iter.bi_size)
1224                 return MAP_DONE;
1225
1226         return MAP_CONTINUE;
1227 }
1228
1229 static void flash_dev_nodata(struct closure *cl)
1230 {
1231         struct search *s = container_of(cl, struct search, cl);
1232
1233         if (s->iop.flush_journal)
1234                 bch_journal_meta(s->iop.c, cl);
1235
1236         continue_at(cl, search_free, NULL);
1237 }
1238
1239 static void flash_dev_make_request(struct request_queue *q, struct bio *bio)
1240 {
1241         struct search *s;
1242         struct closure *cl;
1243         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1244         int cpu, rw = bio_data_dir(bio);
1245
1246         cpu = part_stat_lock();
1247         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1248         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1249         part_stat_unlock();
1250
1251         s = search_alloc(bio, d);
1252         cl = &s->cl;
1253         bio = &s->bio.bio;
1254
1255         trace_bcache_request_start(s->d, bio);
1256
1257         if (!bio->bi_iter.bi_size) {
1258                 /*
1259                  * can't call bch_journal_meta from under
1260                  * generic_make_request
1261                  */
1262                 continue_at_nobarrier(&s->cl,
1263                                       flash_dev_nodata,
1264                                       bcache_wq);
1265         } else if (rw) {
1266                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1267                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1268                                         &KEY(d->id, bio_end_sector(bio), 0));
1269
1270                 s->iop.bypass           = (bio->bi_rw & REQ_DISCARD) != 0;
1271                 s->iop.writeback        = true;
1272                 s->iop.bio              = bio;
1273
1274                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1275         } else {
1276                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1277         }
1278
1279         continue_at(cl, search_free, NULL);
1280 }
1281
1282 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1283                            unsigned int cmd, unsigned long arg)
1284 {
1285         return -ENOTTY;
1286 }
1287
1288 static int flash_dev_congested(void *data, int bits)
1289 {
1290         struct bcache_device *d = data;
1291         struct request_queue *q;
1292         struct cache *ca;
1293         unsigned i;
1294         int ret = 0;
1295
1296         for_each_cache(ca, d->c, i) {
1297                 q = bdev_get_queue(ca->bdev);
1298                 ret |= bdi_congested(&q->backing_dev_info, bits);
1299         }
1300
1301         return ret;
1302 }
1303
1304 void bch_flash_dev_request_init(struct bcache_device *d)
1305 {
1306         struct gendisk *g = d->disk;
1307
1308         g->queue->make_request_fn               = flash_dev_make_request;
1309         g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1310         d->cache_miss                           = flash_dev_cache_miss;
1311         d->ioctl                                = flash_dev_ioctl;
1312 }
1313
1314 void bch_request_exit(void)
1315 {
1316 #ifdef CONFIG_CGROUP_BCACHE
1317         cgroup_unload_subsys(&bcache_subsys);
1318 #endif
1319         if (bch_search_cache)
1320                 kmem_cache_destroy(bch_search_cache);
1321 }
1322
1323 int __init bch_request_init(void)
1324 {
1325         bch_search_cache = KMEM_CACHE(search, 0);
1326         if (!bch_search_cache)
1327                 return -ENOMEM;
1328
1329 #ifdef CONFIG_CGROUP_BCACHE
1330         cgroup_load_subsys(&bcache_subsys);
1331         init_bch_cgroup(&bcache_default_cgroup);
1332
1333         cgroup_add_cftypes(&bcache_subsys, bch_files);
1334 #endif
1335         return 0;
1336 }