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