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 check_should_skip(struct cached_dev *, struct search *);
  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         free_css_id(&bcache_subsys, &cg->css);
 167         kfree(cg);
 168 }
 169
 170 struct cgroup_subsys bcache_subsys = {
 171         .create         = bcachecg_create,
 172         .destroy        = bcachecg_destroy,
 173         .subsys_id      = bcache_subsys_id,
 174         .name           = "bcache",
 175         .module         = THIS_MODULE,
 176 };
 177 EXPORT_SYMBOL_GPL(bcache_subsys);
 178 #endif
 179
 180 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
 181 {
 182 #ifdef CONFIG_CGROUP_BCACHE
 183         int r = bch_bio_to_cgroup(bio)->cache_mode;
 184         if (r >= 0)
 185                 return r;
 186 #endif
 187         return BDEV_CACHE_MODE(&dc->sb);
 188 }
 189
 190 static bool verify(struct cached_dev *dc, struct bio *bio)
 191 {
 192 #ifdef CONFIG_CGROUP_BCACHE
 193         if (bch_bio_to_cgroup(bio)->verify)
 194                 return true;
 195 #endif
 196         return dc->verify;
 197 }
 198
 199 static void bio_csum(struct bio *bio, struct bkey *k)
 200 {
 201         struct bio_vec *bv;
 202         uint64_t csum = 0;
 203         int i;
 204
 205         bio_for_each_segment(bv, bio, i) {
 206                 void *d = kmap(bv->bv_page) + bv->bv_offset;
 207                 csum = bch_crc64_update(csum, d, bv->bv_len);
 208                 kunmap(bv->bv_page);
 209         }
 210
 211         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
 212 }
 213
 214 /* Insert data into cache */
 215
 216 static void bio_invalidate(struct closure *cl)
 217 {
 218         struct btree_op *op = container_of(cl, struct btree_op, cl);
 219         struct bio *bio = op->cache_bio;
 220
 221         pr_debug("invalidating %i sectors from %llu",
 222                  bio_sectors(bio), (uint64_t) bio->bi_sector);
 223
 224         while (bio_sectors(bio)) {
 225                 unsigned len = min(bio_sectors(bio), 1U << 14);
 226
 227                 if (bch_keylist_realloc(&op->keys, 0, op->c))
 228                         goto out;
 229
 230                 bio->bi_sector  += len;
 231                 bio->bi_size    -= len << 9;
 232
 233                 bch_keylist_add(&op->keys,
 234                                 &KEY(op->inode, bio->bi_sector, len));
 235         }
 236
 237         op->insert_data_done = true;
 238         bio_put(bio);
 239 out:
 240         continue_at(cl, bch_journal, bcache_wq);
 241 }
 242
 243 struct open_bucket {
 244         struct list_head        list;
 245         struct task_struct      *last;
 246         unsigned                sectors_free;
 247         BKEY_PADDED(key);
 248 };
 249
 250 void bch_open_buckets_free(struct cache_set *c)
 251 {
 252         struct open_bucket *b;
 253
 254         while (!list_empty(&c->data_buckets)) {
 255                 b = list_first_entry(&c->data_buckets,
 256                                      struct open_bucket, list);
 257                 list_del(&b->list);
 258                 kfree(b);
 259         }
 260 }
 261
 262 int bch_open_buckets_alloc(struct cache_set *c)
 263 {
 264         int i;
 265
 266         spin_lock_init(&c->data_bucket_lock);
 267
 268         for (i = 0; i < 6; i++) {
 269                 struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
 270                 if (!b)
 271                         return -ENOMEM;
 272
 273                 list_add(&b->list, &c->data_buckets);
 274         }
 275
 276         return 0;
 277 }
 278
 279 /*
 280  * We keep multiple buckets open for writes, and try to segregate different
 281  * write streams for better cache utilization: first we look for a bucket where
 282  * the last write to it was sequential with the current write, and failing that
 283  * we look for a bucket that was last used by the same task.
 284  *
 285  * The ideas is if you've got multiple tasks pulling data into the cache at the
 286  * same time, you'll get better cache utilization if you try to segregate their
 287  * data and preserve locality.
 288  *
 289  * For example, say you've starting Firefox at the same time you're copying a
 290  * bunch of files. Firefox will likely end up being fairly hot and stay in the
 291  * cache awhile, but the data you copied might not be; if you wrote all that
 292  * data to the same buckets it'd get invalidated at the same time.
 293  *
 294  * Both of those tasks will be doing fairly random IO so we can't rely on
 295  * detecting sequential IO to segregate their data, but going off of the task
 296  * should be a sane heuristic.
 297  */
 298 static struct open_bucket *pick_data_bucket(struct cache_set *c,
 299                                             const struct bkey *search,
 300                                             struct task_struct *task,
 301                                             struct bkey *alloc)
 302 {
 303         struct open_bucket *ret, *ret_task = NULL;
 304
 305         list_for_each_entry_reverse(ret, &c->data_buckets, list)
 306                 if (!bkey_cmp(&ret->key, search))
 307                         goto found;
 308                 else if (ret->last == task)
 309                         ret_task = ret;
 310
 311         ret = ret_task ?: list_first_entry(&c->data_buckets,
 312                                            struct open_bucket, list);
 313 found:
 314         if (!ret->sectors_free && KEY_PTRS(alloc)) {
 315                 ret->sectors_free = c->sb.bucket_size;
 316                 bkey_copy(&ret->key, alloc);
 317                 bkey_init(alloc);
 318         }
 319
 320         if (!ret->sectors_free)
 321                 ret = NULL;
 322
 323         return ret;
 324 }
 325
 326 /*
 327  * Allocates some space in the cache to write to, and k to point to the newly
 328  * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
 329  * end of the newly allocated space).
 330  *
 331  * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
 332  * sectors were actually allocated.
 333  *
 334  * If s->writeback is true, will not fail.
 335  */
 336 static bool bch_alloc_sectors(struct bkey *k, unsigned sectors,
 337                               struct search *s)
 338 {
 339         struct cache_set *c = s->op.c;
 340         struct open_bucket *b;
 341         BKEY_PADDED(key) alloc;
 342         struct closure cl, *w = NULL;
 343         unsigned i;
 344
 345         if (s->writeback) {
 346                 closure_init_stack(&cl);
 347                 w = &cl;
 348         }
 349
 350         /*
 351          * We might have to allocate a new bucket, which we can't do with a
 352          * spinlock held. So if we have to allocate, we drop the lock, allocate
 353          * and then retry. KEY_PTRS() indicates whether alloc points to
 354          * allocated bucket(s).
 355          */
 356
 357         bkey_init(&alloc.key);
 358         spin_lock(&c->data_bucket_lock);
 359
 360         while (!(b = pick_data_bucket(c, k, s->task, &alloc.key))) {
 361                 unsigned watermark = s->op.write_prio
 362                         ? WATERMARK_MOVINGGC
 363                         : WATERMARK_NONE;
 364
 365                 spin_unlock(&c->data_bucket_lock);
 366
 367                 if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, w))
 368                         return false;
 369
 370                 spin_lock(&c->data_bucket_lock);
 371         }
 372
 373         /*
 374          * If we had to allocate, we might race and not need to allocate the
 375          * second time we call find_data_bucket(). If we allocated a bucket but
 376          * didn't use it, drop the refcount bch_bucket_alloc_set() took:
 377          */
 378         if (KEY_PTRS(&alloc.key))
 379                 __bkey_put(c, &alloc.key);
 380
 381         for (i = 0; i < KEY_PTRS(&b->key); i++)
 382                 EBUG_ON(ptr_stale(c, &b->key, i));
 383
 384         /* Set up the pointer to the space we're allocating: */
 385
 386         for (i = 0; i < KEY_PTRS(&b->key); i++)
 387                 k->ptr[i] = b->key.ptr[i];
 388
 389         sectors = min(sectors, b->sectors_free);
 390
 391         SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
 392         SET_KEY_SIZE(k, sectors);
 393         SET_KEY_PTRS(k, KEY_PTRS(&b->key));
 394
 395         /*
 396          * Move b to the end of the lru, and keep track of what this bucket was
 397          * last used for:
 398          */
 399         list_move_tail(&b->list, &c->data_buckets);
 400         bkey_copy_key(&b->key, k);
 401         b->last = s->task;
 402
 403         b->sectors_free -= sectors;
 404
 405         for (i = 0; i < KEY_PTRS(&b->key); i++) {
 406                 SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
 407
 408                 atomic_long_add(sectors,
 409                                 &PTR_CACHE(c, &b->key, i)->sectors_written);
 410         }
 411
 412         if (b->sectors_free < c->sb.block_size)
 413                 b->sectors_free = 0;
 414
 415         /*
 416          * k takes refcounts on the buckets it points to until it's inserted
 417          * into the btree, but if we're done with this bucket we just transfer
 418          * get_data_bucket()'s refcount.
 419          */
 420         if (b->sectors_free)
 421                 for (i = 0; i < KEY_PTRS(&b->key); i++)
 422                         atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
 423
 424         spin_unlock(&c->data_bucket_lock);
 425         return true;
 426 }
 427
 428 static void bch_insert_data_error(struct closure *cl)
 429 {
 430         struct btree_op *op = container_of(cl, struct btree_op, cl);
 431
 432         /*
 433          * Our data write just errored, which means we've got a bunch of keys to
 434          * insert that point to data that wasn't succesfully written.
 435          *
 436          * We don't have to insert those keys but we still have to invalidate
 437          * that region of the cache - so, if we just strip off all the pointers
 438          * from the keys we'll accomplish just that.
 439          */
 440
 441         struct bkey *src = op->keys.bottom, *dst = op->keys.bottom;
 442
 443         while (src != op->keys.top) {
 444                 struct bkey *n = bkey_next(src);
 445
 446                 SET_KEY_PTRS(src, 0);
 447                 bkey_copy(dst, src);
 448
 449                 dst = bkey_next(dst);
 450                 src = n;
 451         }
 452
 453         op->keys.top = dst;
 454
 455         bch_journal(cl);
 456 }
 457
 458 static void bch_insert_data_endio(struct bio *bio, int error)
 459 {
 460         struct closure *cl = bio->bi_private;
 461         struct btree_op *op = container_of(cl, struct btree_op, cl);
 462         struct search *s = container_of(op, struct search, op);
 463
 464         if (error) {
 465                 /* TODO: We could try to recover from this. */
 466                 if (s->writeback)
 467                         s->error = error;
 468                 else if (s->write)
 469                         set_closure_fn(cl, bch_insert_data_error, bcache_wq);
 470                 else
 471                         set_closure_fn(cl, NULL, NULL);
 472         }
 473
 474         bch_bbio_endio(op->c, bio, error, "writing data to cache");
 475 }
 476
 477 static void bch_insert_data_loop(struct closure *cl)
 478 {
 479         struct btree_op *op = container_of(cl, struct btree_op, cl);
 480         struct search *s = container_of(op, struct search, op);
 481         struct bio *bio = op->cache_bio, *n;
 482
 483         if (op->skip)
 484                 return bio_invalidate(cl);
 485
 486         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) {
 487                 set_gc_sectors(op->c);
 488                 bch_queue_gc(op->c);
 489         }
 490
 491         /*
 492          * Journal writes are marked REQ_FLUSH; if the original write was a
 493          * flush, it'll wait on the journal write.
 494          */
 495         bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA);
 496
 497         do {
 498                 unsigned i;
 499                 struct bkey *k;
 500                 struct bio_set *split = s->d
 501                         ? s->d->bio_split : op->c->bio_split;
 502
 503                 /* 1 for the device pointer and 1 for the chksum */
 504                 if (bch_keylist_realloc(&op->keys,
 505                                         1 + (op->csum ? 1 : 0),
 506                                         op->c))
 507                         continue_at(cl, bch_journal, bcache_wq);
 508
 509                 k = op->keys.top;
 510                 bkey_init(k);
 511                 SET_KEY_INODE(k, op->inode);
 512                 SET_KEY_OFFSET(k, bio->bi_sector);
 513
 514                 if (!bch_alloc_sectors(k, bio_sectors(bio), s))
 515                         goto err;
 516
 517                 n = bch_bio_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 518
 519                 n->bi_end_io    = bch_insert_data_endio;
 520                 n->bi_private   = cl;
 521
 522                 if (s->writeback) {
 523                         SET_KEY_DIRTY(k, true);
 524
 525                         for (i = 0; i < KEY_PTRS(k); i++)
 526                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 527                                             GC_MARK_DIRTY);
 528                 }
 529
 530                 SET_KEY_CSUM(k, op->csum);
 531                 if (KEY_CSUM(k))
 532                         bio_csum(n, k);
 533
 534                 trace_bcache_cache_insert(k);
 535                 bch_keylist_push(&op->keys);
 536
 537                 n->bi_rw |= REQ_WRITE;
 538                 bch_submit_bbio(n, op->c, k, 0);
 539         } while (n != bio);
 540
 541         op->insert_data_done = true;
 542         continue_at(cl, bch_journal, bcache_wq);
 543 err:
 544         /* bch_alloc_sectors() blocks if s->writeback = true */
 545         BUG_ON(s->writeback);
 546
 547         /*
 548          * But if it's not a writeback write we'd rather just bail out if
 549          * there aren't any buckets ready to write to - it might take awhile and
 550          * we might be starving btree writes for gc or something.
 551          */
 552
 553         if (s->write) {
 554                 /*
 555                  * Writethrough write: We can't complete the write until we've
 556                  * updated the index. But we don't want to delay the write while
 557                  * we wait for buckets to be freed up, so just invalidate the
 558                  * rest of the write.
 559                  */
 560                 op->skip = true;
 561                 return bio_invalidate(cl);
 562         } else {
 563                 /*
 564                  * From a cache miss, we can just insert the keys for the data
 565                  * we have written or bail out if we didn't do anything.
 566                  */
 567                 op->insert_data_done = true;
 568                 bio_put(bio);
 569
 570                 if (!bch_keylist_empty(&op->keys))
 571                         continue_at(cl, bch_journal, bcache_wq);
 572                 else
 573                         closure_return(cl);
 574         }
 575 }
 576
 577 /**
 578  * bch_insert_data - stick some data in the cache
 579  *
 580  * This is the starting point for any data to end up in a cache device; it could
 581  * be from a normal write, or a writeback write, or a write to a flash only
 582  * volume - it's also used by the moving garbage collector to compact data in
 583  * mostly empty buckets.
 584  *
 585  * It first writes the data to the cache, creating a list of keys to be inserted
 586  * (if the data had to be fragmented there will be multiple keys); after the
 587  * data is written it calls bch_journal, and after the keys have been added to
 588  * the next journal write they're inserted into the btree.
 589  *
 590  * It inserts the data in op->cache_bio; bi_sector is used for the key offset,
 591  * and op->inode is used for the key inode.
 592  *
 593  * If op->skip is true, instead of inserting the data it invalidates the region
 594  * of the cache represented by op->cache_bio and op->inode.
 595  */
 596 void bch_insert_data(struct closure *cl)
 597 {
 598         struct btree_op *op = container_of(cl, struct btree_op, cl);
 599
 600         bch_keylist_init(&op->keys);
 601         bio_get(op->cache_bio);
 602         bch_insert_data_loop(cl);
 603 }
 604
 605 void bch_btree_insert_async(struct closure *cl)
 606 {
 607         struct btree_op *op = container_of(cl, struct btree_op, cl);
 608         struct search *s = container_of(op, struct search, op);
 609
 610         if (bch_btree_insert(op, op->c)) {
 611                 s->error                = -ENOMEM;
 612                 op->insert_data_done    = true;
 613         }
 614
 615         if (op->insert_data_done) {
 616                 bch_keylist_free(&op->keys);
 617                 closure_return(cl);
 618         } else
 619                 continue_at(cl, bch_insert_data_loop, bcache_wq);
 620 }
 621
 622 /* Common code for the make_request functions */
 623
 624 static void request_endio(struct bio *bio, int error)
 625 {
 626         struct closure *cl = bio->bi_private;
 627
 628         if (error) {
 629                 struct search *s = container_of(cl, struct search, cl);
 630                 s->error = error;
 631                 /* Only cache read errors are recoverable */
 632                 s->recoverable = false;
 633         }
 634
 635         bio_put(bio);
 636         closure_put(cl);
 637 }
 638
 639 void bch_cache_read_endio(struct bio *bio, int error)
 640 {
 641         struct bbio *b = container_of(bio, struct bbio, bio);
 642         struct closure *cl = bio->bi_private;
 643         struct search *s = container_of(cl, struct search, cl);
 644
 645         /*
 646          * If the bucket was reused while our bio was in flight, we might have
 647          * read the wrong data. Set s->error but not error so it doesn't get
 648          * counted against the cache device, but we'll still reread the data
 649          * from the backing device.
 650          */
 651
 652         if (error)
 653                 s->error = error;
 654         else if (ptr_stale(s->op.c, &b->key, 0)) {
 655                 atomic_long_inc(&s->op.c->cache_read_races);
 656                 s->error = -EINTR;
 657         }
 658
 659         bch_bbio_endio(s->op.c, bio, error, "reading from cache");
 660 }
 661
 662 static void bio_complete(struct search *s)
 663 {
 664         if (s->orig_bio) {
 665                 int cpu, rw = bio_data_dir(s->orig_bio);
 666                 unsigned long duration = jiffies - s->start_time;
 667
 668                 cpu = part_stat_lock();
 669                 part_round_stats(cpu, &s->d->disk->part0);
 670                 part_stat_add(cpu, &s->d->disk->part0, ticks[rw], duration);
 671                 part_stat_unlock();
 672
 673                 trace_bcache_request_end(s, s->orig_bio);
 674                 bio_endio(s->orig_bio, s->error);
 675                 s->orig_bio = NULL;
 676         }
 677 }
 678
 679 static void do_bio_hook(struct search *s)
 680 {
 681         struct bio *bio = &s->bio.bio;
 682         memcpy(bio, s->orig_bio, sizeof(struct bio));
 683
 684         bio->bi_end_io          = request_endio;
 685         bio->bi_private         = &s->cl;
 686         atomic_set(&bio->bi_cnt, 3);
 687 }
 688
 689 static void search_free(struct closure *cl)
 690 {
 691         struct search *s = container_of(cl, struct search, cl);
 692         bio_complete(s);
 693
 694         if (s->op.cache_bio)
 695                 bio_put(s->op.cache_bio);
 696
 697         if (s->unaligned_bvec)
 698                 mempool_free(s->bio.bio.bi_io_vec, s->d->unaligned_bvec);
 699
 700         closure_debug_destroy(cl);
 701         mempool_free(s, s->d->c->search);
 702 }
 703
 704 static struct search *search_alloc(struct bio *bio, struct bcache_device *d)
 705 {
 706         struct bio_vec *bv;
 707         struct search *s = mempool_alloc(d->c->search, GFP_NOIO);
 708         memset(s, 0, offsetof(struct search, op.keys));
 709
 710         __closure_init(&s->cl, NULL);
 711
 712         s->op.inode             = d->id;
 713         s->op.c                 = d->c;
 714         s->d                    = d;
 715         s->op.lock              = -1;
 716         s->task                 = current;
 717         s->orig_bio             = bio;
 718         s->write                = (bio->bi_rw & REQ_WRITE) != 0;
 719         s->op.flush_journal     = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0;
 720         s->op.skip              = (bio->bi_rw & REQ_DISCARD) != 0;
 721         s->recoverable          = 1;
 722         s->start_time           = jiffies;
 723         do_bio_hook(s);
 724
 725         if (bio->bi_size != bio_segments(bio) * PAGE_SIZE) {
 726                 bv = mempool_alloc(d->unaligned_bvec, GFP_NOIO);
 727                 memcpy(bv, bio_iovec(bio),
 728                        sizeof(struct bio_vec) * bio_segments(bio));
 729
 730                 s->bio.bio.bi_io_vec    = bv;
 731                 s->unaligned_bvec       = 1;
 732         }
 733
 734         return s;
 735 }
 736
 737 static void btree_read_async(struct closure *cl)
 738 {
 739         struct btree_op *op = container_of(cl, struct btree_op, cl);
 740
 741         int ret = btree_root(search_recurse, op->c, op);
 742
 743         if (ret == -EAGAIN)
 744                 continue_at(cl, btree_read_async, bcache_wq);
 745
 746         closure_return(cl);
 747 }
 748
 749 /* Cached devices */
 750
 751 static void cached_dev_bio_complete(struct closure *cl)
 752 {
 753         struct search *s = container_of(cl, struct search, cl);
 754         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 755
 756         search_free(cl);
 757         cached_dev_put(dc);
 758 }
 759
 760 /* Process reads */
 761
 762 static void cached_dev_read_complete(struct closure *cl)
 763 {
 764         struct search *s = container_of(cl, struct search, cl);
 765
 766         if (s->op.insert_collision)
 767                 bch_mark_cache_miss_collision(s);
 768
 769         if (s->op.cache_bio) {
 770                 int i;
 771                 struct bio_vec *bv;
 772
 773                 __bio_for_each_segment(bv, s->op.cache_bio, i, 0)
 774                         __free_page(bv->bv_page);
 775         }
 776
 777         cached_dev_bio_complete(cl);
 778 }
 779
 780 static void request_read_error(struct closure *cl)
 781 {
 782         struct search *s = container_of(cl, struct search, cl);
 783         struct bio_vec *bv;
 784         int i;
 785
 786         if (s->recoverable) {
 787                 /* Retry from the backing device: */
 788                 trace_bcache_read_retry(s->orig_bio);
 789
 790                 s->error = 0;
 791                 bv = s->bio.bio.bi_io_vec;
 792                 do_bio_hook(s);
 793                 s->bio.bio.bi_io_vec = bv;
 794
 795                 if (!s->unaligned_bvec)
 796                         bio_for_each_segment(bv, s->orig_bio, i)
 797                                 bv->bv_offset = 0, bv->bv_len = PAGE_SIZE;
 798                 else
 799                         memcpy(s->bio.bio.bi_io_vec,
 800                                bio_iovec(s->orig_bio),
 801                                sizeof(struct bio_vec) *
 802                                bio_segments(s->orig_bio));
 803
 804                 /* XXX: invalidate cache */
 805
 806                 closure_bio_submit(&s->bio.bio, &s->cl, s->d);
 807         }
 808
 809         continue_at(cl, cached_dev_read_complete, NULL);
 810 }
 811
 812 static void request_read_done(struct closure *cl)
 813 {
 814         struct search *s = container_of(cl, struct search, cl);
 815         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 816
 817         /*
 818          * s->cache_bio != NULL implies that we had a cache miss; cache_bio now
 819          * contains data ready to be inserted into the cache.
 820          *
 821          * First, we copy the data we just read from cache_bio's bounce buffers
 822          * to the buffers the original bio pointed to:
 823          */
 824
 825         if (s->op.cache_bio) {
 826                 bio_reset(s->op.cache_bio);
 827                 s->op.cache_bio->bi_sector      = s->cache_miss->bi_sector;
 828                 s->op.cache_bio->bi_bdev        = s->cache_miss->bi_bdev;
 829                 s->op.cache_bio->bi_size        = s->cache_bio_sectors << 9;
 830                 bch_bio_map(s->op.cache_bio, NULL);
 831
 832                 bio_copy_data(s->cache_miss, s->op.cache_bio);
 833
 834                 bio_put(s->cache_miss);
 835                 s->cache_miss = NULL;
 836         }
 837
 838         if (verify(dc, &s->bio.bio) && s->recoverable)
 839                 bch_data_verify(s);
 840
 841         bio_complete(s);
 842
 843         if (s->op.cache_bio &&
 844             !test_bit(CACHE_SET_STOPPING, &s->op.c->flags)) {
 845                 s->op.type = BTREE_REPLACE;
 846                 closure_call(&s->op.cl, bch_insert_data, NULL, cl);
 847         }
 848
 849         continue_at(cl, cached_dev_read_complete, NULL);
 850 }
 851
 852 static void request_read_done_bh(struct closure *cl)
 853 {
 854         struct search *s = container_of(cl, struct search, cl);
 855         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 856
 857         bch_mark_cache_accounting(s, !s->cache_miss, s->op.skip);
 858         trace_bcache_read(s->orig_bio, !s->cache_miss, s->op.skip);
 859
 860         if (s->error)
 861                 continue_at_nobarrier(cl, request_read_error, bcache_wq);
 862         else if (s->op.cache_bio || verify(dc, &s->bio.bio))
 863                 continue_at_nobarrier(cl, request_read_done, bcache_wq);
 864         else
 865                 continue_at_nobarrier(cl, cached_dev_read_complete, NULL);
 866 }
 867
 868 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 869                                  struct bio *bio, unsigned sectors)
 870 {
 871         int ret = 0;
 872         unsigned reada;
 873         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 874         struct bio *miss;
 875
 876         miss = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 877         if (miss == bio)
 878                 s->op.lookup_done = true;
 879
 880         miss->bi_end_io         = request_endio;
 881         miss->bi_private        = &s->cl;
 882
 883         if (s->cache_miss || s->op.skip)
 884                 goto out_submit;
 885
 886         if (miss != bio ||
 887             (bio->bi_rw & REQ_RAHEAD) ||
 888             (bio->bi_rw & REQ_META) ||
 889             s->op.c->gc_stats.in_use >= CUTOFF_CACHE_READA)
 890                 reada = 0;
 891         else {
 892                 reada = min(dc->readahead >> 9,
 893                             sectors - bio_sectors(miss));
 894
 895                 if (bio_end_sector(miss) + reada > bdev_sectors(miss->bi_bdev))
 896                         reada = bdev_sectors(miss->bi_bdev) -
 897                                 bio_end_sector(miss);
 898         }
 899
 900         s->cache_bio_sectors = bio_sectors(miss) + reada;
 901         s->op.cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 902                         DIV_ROUND_UP(s->cache_bio_sectors, PAGE_SECTORS),
 903                         dc->disk.bio_split);
 904
 905         if (!s->op.cache_bio)
 906                 goto out_submit;
 907
 908         s->op.cache_bio->bi_sector      = miss->bi_sector;
 909         s->op.cache_bio->bi_bdev        = miss->bi_bdev;
 910         s->op.cache_bio->bi_size        = s->cache_bio_sectors << 9;
 911
 912         s->op.cache_bio->bi_end_io      = request_endio;
 913         s->op.cache_bio->bi_private     = &s->cl;
 914
 915         /* btree_search_recurse()'s btree iterator is no good anymore */
 916         ret = -EINTR;
 917         if (!bch_btree_insert_check_key(b, &s->op, s->op.cache_bio))
 918                 goto out_put;
 919
 920         bch_bio_map(s->op.cache_bio, NULL);
 921         if (bio_alloc_pages(s->op.cache_bio, __GFP_NOWARN|GFP_NOIO))
 922                 goto out_put;
 923
 924         s->cache_miss = miss;
 925         bio_get(s->op.cache_bio);
 926
 927         closure_bio_submit(s->op.cache_bio, &s->cl, s->d);
 928
 929         return ret;
 930 out_put:
 931         bio_put(s->op.cache_bio);
 932         s->op.cache_bio = NULL;
 933 out_submit:
 934         closure_bio_submit(miss, &s->cl, s->d);
 935         return ret;
 936 }
 937
 938 static void request_read(struct cached_dev *dc, struct search *s)
 939 {
 940         struct closure *cl = &s->cl;
 941
 942         check_should_skip(dc, s);
 943         closure_call(&s->op.cl, btree_read_async, NULL, cl);
 944
 945         continue_at(cl, request_read_done_bh, NULL);
 946 }
 947
 948 /* Process writes */
 949
 950 static void cached_dev_write_complete(struct closure *cl)
 951 {
 952         struct search *s = container_of(cl, struct search, cl);
 953         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 954
 955         up_read_non_owner(&dc->writeback_lock);
 956         cached_dev_bio_complete(cl);
 957 }
 958
 959 static void request_write(struct cached_dev *dc, struct search *s)
 960 {
 961         struct closure *cl = &s->cl;
 962         struct bio *bio = &s->bio.bio;
 963         struct bkey start, end;
 964         start = KEY(dc->disk.id, bio->bi_sector, 0);
 965         end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 966
 967         bch_keybuf_check_overlapping(&s->op.c->moving_gc_keys, &start, &end);
 968
 969         check_should_skip(dc, s);
 970         down_read_non_owner(&dc->writeback_lock);
 971
 972         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 973                 s->op.skip      = false;
 974                 s->writeback    = true;
 975         }
 976
 977         if (bio->bi_rw & REQ_DISCARD)
 978                 goto skip;
 979
 980         if (should_writeback(dc, s->orig_bio,
 981                              cache_mode(dc, bio),
 982                              s->op.skip)) {
 983                 s->op.skip = false;
 984                 s->writeback = true;
 985         }
 986
 987         if (s->op.skip)
 988                 goto skip;
 989
 990         trace_bcache_write(s->orig_bio, s->writeback, s->op.skip);
 991
 992         if (!s->writeback) {
 993                 s->op.cache_bio = bio_clone_bioset(bio, GFP_NOIO,
 994                                                    dc->disk.bio_split);
 995
 996                 closure_bio_submit(bio, cl, s->d);
 997         } else {
 998                 bch_writeback_add(dc);
 999                 s->op.cache_bio = bio;
1000
1001                 if (bio->bi_rw & REQ_FLUSH) {
1002                         /* Also need to send a flush to the backing device */
1003                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
1004                                                              dc->disk.bio_split);
1005
1006                         flush->bi_rw    = WRITE_FLUSH;
1007                         flush->bi_bdev  = bio->bi_bdev;
1008                         flush->bi_end_io = request_endio;
1009                         flush->bi_private = cl;
1010
1011                         closure_bio_submit(flush, cl, s->d);
1012                 }
1013         }
1014 out:
1015         closure_call(&s->op.cl, bch_insert_data, NULL, cl);
1016         continue_at(cl, cached_dev_write_complete, NULL);
1017 skip:
1018         s->op.skip = true;
1019         s->op.cache_bio = s->orig_bio;
1020         bio_get(s->op.cache_bio);
1021
1022         if ((bio->bi_rw & REQ_DISCARD) &&
1023             !blk_queue_discard(bdev_get_queue(dc->bdev)))
1024                 goto out;
1025
1026         closure_bio_submit(bio, cl, s->d);
1027         goto out;
1028 }
1029
1030 static void request_nodata(struct cached_dev *dc, struct search *s)
1031 {
1032         struct closure *cl = &s->cl;
1033         struct bio *bio = &s->bio.bio;
1034
1035         if (bio->bi_rw & REQ_DISCARD) {
1036                 request_write(dc, s);
1037                 return;
1038         }
1039
1040         if (s->op.flush_journal)
1041                 bch_journal_meta(s->op.c, cl);
1042
1043         closure_bio_submit(bio, cl, s->d);
1044
1045         continue_at(cl, cached_dev_bio_complete, NULL);
1046 }
1047
1048 /* Cached devices - read & write stuff */
1049
1050 unsigned bch_get_congested(struct cache_set *c)
1051 {
1052         int i;
1053         long rand;
1054
1055         if (!c->congested_read_threshold_us &&
1056             !c->congested_write_threshold_us)
1057                 return 0;
1058
1059         i = (local_clock_us() - c->congested_last_us) / 1024;
1060         if (i < 0)
1061                 return 0;
1062
1063         i += atomic_read(&c->congested);
1064         if (i >= 0)
1065                 return 0;
1066
1067         i += CONGESTED_MAX;
1068
1069         if (i > 0)
1070                 i = fract_exp_two(i, 6);
1071
1072         rand = get_random_int();
1073         i -= bitmap_weight(&rand, BITS_PER_LONG);
1074
1075         return i > 0 ? i : 1;
1076 }
1077
1078 static void add_sequential(struct task_struct *t)
1079 {
1080         ewma_add(t->sequential_io_avg,
1081                  t->sequential_io, 8, 0);
1082
1083         t->sequential_io = 0;
1084 }
1085
1086 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
1087 {
1088         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
1089 }
1090
1091 static void check_should_skip(struct cached_dev *dc, struct search *s)
1092 {
1093         struct cache_set *c = s->op.c;
1094         struct bio *bio = &s->bio.bio;
1095         unsigned mode = cache_mode(dc, bio);
1096         unsigned sectors, congested = bch_get_congested(c);
1097
1098         if (atomic_read(&dc->disk.detaching) ||
1099             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
1100             (bio->bi_rw & REQ_DISCARD))
1101                 goto skip;
1102
1103         if (mode == CACHE_MODE_NONE ||
1104             (mode == CACHE_MODE_WRITEAROUND &&
1105              (bio->bi_rw & REQ_WRITE)))
1106                 goto skip;
1107
1108         if (bio->bi_sector   & (c->sb.block_size - 1) ||
1109             bio_sectors(bio) & (c->sb.block_size - 1)) {
1110                 pr_debug("skipping unaligned io");
1111                 goto skip;
1112         }
1113
1114         if (!congested && !dc->sequential_cutoff)
1115                 goto rescale;
1116
1117         if (!congested &&
1118             mode == CACHE_MODE_WRITEBACK &&
1119             (bio->bi_rw & REQ_WRITE) &&
1120             (bio->bi_rw & REQ_SYNC))
1121                 goto rescale;
1122
1123         if (dc->sequential_merge) {
1124                 struct io *i;
1125
1126                 spin_lock(&dc->io_lock);
1127
1128                 hlist_for_each_entry(i, iohash(dc, bio->bi_sector), hash)
1129                         if (i->last == bio->bi_sector &&
1130                             time_before(jiffies, i->jiffies))
1131                                 goto found;
1132
1133                 i = list_first_entry(&dc->io_lru, struct io, lru);
1134
1135                 add_sequential(s->task);
1136                 i->sequential = 0;
1137 found:
1138                 if (i->sequential + bio->bi_size > i->sequential)
1139                         i->sequential   += bio->bi_size;
1140
1141                 i->last                  = bio_end_sector(bio);
1142                 i->jiffies               = jiffies + msecs_to_jiffies(5000);
1143                 s->task->sequential_io   = i->sequential;
1144
1145                 hlist_del(&i->hash);
1146                 hlist_add_head(&i->hash, iohash(dc, i->last));
1147                 list_move_tail(&i->lru, &dc->io_lru);
1148
1149                 spin_unlock(&dc->io_lock);
1150         } else {
1151                 s->task->sequential_io = bio->bi_size;
1152
1153                 add_sequential(s->task);
1154         }
1155
1156         sectors = max(s->task->sequential_io,
1157                       s->task->sequential_io_avg) >> 9;
1158
1159         if (dc->sequential_cutoff &&
1160             sectors >= dc->sequential_cutoff >> 9) {
1161                 trace_bcache_bypass_sequential(s->orig_bio);
1162                 goto skip;
1163         }
1164
1165         if (congested && sectors >= congested) {
1166                 trace_bcache_bypass_congested(s->orig_bio);
1167                 goto skip;
1168         }
1169
1170 rescale:
1171         bch_rescale_priorities(c, bio_sectors(bio));
1172         return;
1173 skip:
1174         bch_mark_sectors_bypassed(s, bio_sectors(bio));
1175         s->op.skip = true;
1176 }
1177
1178 static void cached_dev_make_request(struct request_queue *q, struct bio *bio)
1179 {
1180         struct search *s;
1181         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1182         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1183         int cpu, rw = bio_data_dir(bio);
1184
1185         cpu = part_stat_lock();
1186         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1187         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1188         part_stat_unlock();
1189
1190         bio->bi_bdev = dc->bdev;
1191         bio->bi_sector += dc->sb.data_offset;
1192
1193         if (cached_dev_get(dc)) {
1194                 s = search_alloc(bio, d);
1195                 trace_bcache_request_start(s, bio);
1196
1197                 if (!bio_has_data(bio))
1198                         request_nodata(dc, s);
1199                 else if (rw)
1200                         request_write(dc, s);
1201                 else
1202                         request_read(dc, s);
1203         } else {
1204                 if ((bio->bi_rw & REQ_DISCARD) &&
1205                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1206                         bio_endio(bio, 0);
1207                 else
1208                         bch_generic_make_request(bio, &d->bio_split_hook);
1209         }
1210 }
1211
1212 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1213                             unsigned int cmd, unsigned long arg)
1214 {
1215         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1216         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1217 }
1218
1219 static int cached_dev_congested(void *data, int bits)
1220 {
1221         struct bcache_device *d = data;
1222         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1223         struct request_queue *q = bdev_get_queue(dc->bdev);
1224         int ret = 0;
1225
1226         if (bdi_congested(&q->backing_dev_info, bits))
1227                 return 1;
1228
1229         if (cached_dev_get(dc)) {
1230                 unsigned i;
1231                 struct cache *ca;
1232
1233                 for_each_cache(ca, d->c, i) {
1234                         q = bdev_get_queue(ca->bdev);
1235                         ret |= bdi_congested(&q->backing_dev_info, bits);
1236                 }
1237
1238                 cached_dev_put(dc);
1239         }
1240
1241         return ret;
1242 }
1243
1244 void bch_cached_dev_request_init(struct cached_dev *dc)
1245 {
1246         struct gendisk *g = dc->disk.disk;
1247
1248         g->queue->make_request_fn               = cached_dev_make_request;
1249         g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1250         dc->disk.cache_miss                     = cached_dev_cache_miss;
1251         dc->disk.ioctl                          = cached_dev_ioctl;
1252 }
1253
1254 /* Flash backed devices */
1255
1256 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1257                                 struct bio *bio, unsigned sectors)
1258 {
1259         struct bio_vec *bv;
1260         int i;
1261
1262         /* Zero fill bio */
1263
1264         bio_for_each_segment(bv, bio, i) {
1265                 unsigned j = min(bv->bv_len >> 9, sectors);
1266
1267                 void *p = kmap(bv->bv_page);
1268                 memset(p + bv->bv_offset, 0, j << 9);
1269                 kunmap(bv->bv_page);
1270
1271                 sectors -= j;
1272         }
1273
1274         bio_advance(bio, min(sectors << 9, bio->bi_size));
1275
1276         if (!bio->bi_size)
1277                 s->op.lookup_done = true;
1278
1279         return 0;
1280 }
1281
1282 static void flash_dev_make_request(struct request_queue *q, struct bio *bio)
1283 {
1284         struct search *s;
1285         struct closure *cl;
1286         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1287         int cpu, rw = bio_data_dir(bio);
1288
1289         cpu = part_stat_lock();
1290         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1291         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1292         part_stat_unlock();
1293
1294         s = search_alloc(bio, d);
1295         cl = &s->cl;
1296         bio = &s->bio.bio;
1297
1298         trace_bcache_request_start(s, bio);
1299
1300         if (bio_has_data(bio) && !rw) {
1301                 closure_call(&s->op.cl, btree_read_async, NULL, cl);
1302         } else if (bio_has_data(bio) || s->op.skip) {
1303                 bch_keybuf_check_overlapping(&s->op.c->moving_gc_keys,
1304                                         &KEY(d->id, bio->bi_sector, 0),
1305                                         &KEY(d->id, bio_end_sector(bio), 0));
1306
1307                 s->writeback    = true;
1308                 s->op.cache_bio = bio;
1309
1310                 closure_call(&s->op.cl, bch_insert_data, NULL, cl);
1311         } else {
1312                 /* No data - probably a cache flush */
1313                 if (s->op.flush_journal)
1314                         bch_journal_meta(s->op.c, cl);
1315         }
1316
1317         continue_at(cl, search_free, NULL);
1318 }
1319
1320 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1321                            unsigned int cmd, unsigned long arg)
1322 {
1323         return -ENOTTY;
1324 }
1325
1326 static int flash_dev_congested(void *data, int bits)
1327 {
1328         struct bcache_device *d = data;
1329         struct request_queue *q;
1330         struct cache *ca;
1331         unsigned i;
1332         int ret = 0;
1333
1334         for_each_cache(ca, d->c, i) {
1335                 q = bdev_get_queue(ca->bdev);
1336                 ret |= bdi_congested(&q->backing_dev_info, bits);
1337         }
1338
1339         return ret;
1340 }
1341
1342 void bch_flash_dev_request_init(struct bcache_device *d)
1343 {
1344         struct gendisk *g = d->disk;
1345
1346         g->queue->make_request_fn               = flash_dev_make_request;
1347         g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1348         d->cache_miss                           = flash_dev_cache_miss;
1349         d->ioctl                                = flash_dev_ioctl;
1350 }
1351
1352 void bch_request_exit(void)
1353 {
1354 #ifdef CONFIG_CGROUP_BCACHE
1355         cgroup_unload_subsys(&bcache_subsys);
1356 #endif
1357         if (bch_search_cache)
1358                 kmem_cache_destroy(bch_search_cache);
1359 }
1360
1361 int __init bch_request_init(void)
1362 {
1363         bch_search_cache = KMEM_CACHE(search, 0);
1364         if (!bch_search_cache)
1365                 return -ENOMEM;
1366
1367 #ifdef CONFIG_CGROUP_BCACHE
1368         cgroup_load_subsys(&bcache_subsys);
1369         init_bch_cgroup(&bcache_default_cgroup);
1370
1371         cgroup_add_cftypes(&bcache_subsys, bch_files);
1372 #endif
1373         return 0;
1374 }