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

   1 /*
   2  * Main bcache entry point - handle a read or a write request and decide what to
   3  * do with it; the make_request functions are called by the block layer.
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcache.h"
  10 #include "btree.h"
  11 #include "debug.h"
  12 #include "request.h"
  13 #include "writeback.h"
  14
  15 #include <linux/cgroup.h>
  16 #include <linux/module.h>
  17 #include <linux/hash.h>
  18 #include <linux/random.h>
  19 #include "blk-cgroup.h"
  20
  21 #include <trace/events/bcache.h>
  22
  23 #define CUTOFF_CACHE_ADD        95
  24 #define CUTOFF_CACHE_READA      90
  25
  26 struct kmem_cache *bch_search_cache;
  27
  28 static void bch_data_insert_start(struct closure *);
  29
  30 /* Cgroup interface */
  31
  32 #ifdef CONFIG_CGROUP_BCACHE
  33 static struct bch_cgroup bcache_default_cgroup = { .cache_mode = -1 };
  34
  35 static struct bch_cgroup *cgroup_to_bcache(struct cgroup *cgroup)
  36 {
  37         struct cgroup_subsys_state *css;
  38         return cgroup &&
  39                 (css = cgroup_subsys_state(cgroup, bcache_subsys_id))
  40                 ? container_of(css, struct bch_cgroup, css)
  41                 : &bcache_default_cgroup;
  42 }
  43
  44 struct bch_cgroup *bch_bio_to_cgroup(struct bio *bio)
  45 {
  46         struct cgroup_subsys_state *css = bio->bi_css
  47                 ? cgroup_subsys_state(bio->bi_css->cgroup, bcache_subsys_id)
  48                 : task_subsys_state(current, bcache_subsys_id);
  49
  50         return css
  51                 ? container_of(css, struct bch_cgroup, css)
  52                 : &bcache_default_cgroup;
  53 }
  54
  55 static ssize_t cache_mode_read(struct cgroup *cgrp, struct cftype *cft,
  56                         struct file *file,
  57                         char __user *buf, size_t nbytes, loff_t *ppos)
  58 {
  59         char tmp[1024];
  60         int len = bch_snprint_string_list(tmp, PAGE_SIZE, bch_cache_modes,
  61                                           cgroup_to_bcache(cgrp)->cache_mode + 1);
  62
  63         if (len < 0)
  64                 return len;
  65
  66         return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
  67 }
  68
  69 static int cache_mode_write(struct cgroup *cgrp, struct cftype *cft,
  70                             const char *buf)
  71 {
  72         int v = bch_read_string_list(buf, bch_cache_modes);
  73         if (v < 0)
  74                 return v;
  75
  76         cgroup_to_bcache(cgrp)->cache_mode = v - 1;
  77         return 0;
  78 }
  79
  80 static u64 bch_verify_read(struct cgroup *cgrp, struct cftype *cft)
  81 {
  82         return cgroup_to_bcache(cgrp)->verify;
  83 }
  84
  85 static int bch_verify_write(struct cgroup *cgrp, struct cftype *cft, u64 val)
  86 {
  87         cgroup_to_bcache(cgrp)->verify = val;
  88         return 0;
  89 }
  90
  91 static u64 bch_cache_hits_read(struct cgroup *cgrp, struct cftype *cft)
  92 {
  93         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
  94         return atomic_read(&bcachecg->stats.cache_hits);
  95 }
  96
  97 static u64 bch_cache_misses_read(struct cgroup *cgrp, struct cftype *cft)
  98 {
  99         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 100         return atomic_read(&bcachecg->stats.cache_misses);
 101 }
 102
 103 static u64 bch_cache_bypass_hits_read(struct cgroup *cgrp,
 104                                          struct cftype *cft)
 105 {
 106         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 107         return atomic_read(&bcachecg->stats.cache_bypass_hits);
 108 }
 109
 110 static u64 bch_cache_bypass_misses_read(struct cgroup *cgrp,
 111                                            struct cftype *cft)
 112 {
 113         struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp);
 114         return atomic_read(&bcachecg->stats.cache_bypass_misses);
 115 }
 116
 117 static struct cftype bch_files[] = {
 118         {
 119                 .name           = "cache_mode",
 120                 .read           = cache_mode_read,
 121                 .write_string   = cache_mode_write,
 122         },
 123         {
 124                 .name           = "verify",
 125                 .read_u64       = bch_verify_read,
 126                 .write_u64      = bch_verify_write,
 127         },
 128         {
 129                 .name           = "cache_hits",
 130                 .read_u64       = bch_cache_hits_read,
 131         },
 132         {
 133                 .name           = "cache_misses",
 134                 .read_u64       = bch_cache_misses_read,
 135         },
 136         {
 137                 .name           = "cache_bypass_hits",
 138                 .read_u64       = bch_cache_bypass_hits_read,
 139         },
 140         {
 141                 .name           = "cache_bypass_misses",
 142                 .read_u64       = bch_cache_bypass_misses_read,
 143         },
 144         { }     /* terminate */
 145 };
 146
 147 static void init_bch_cgroup(struct bch_cgroup *cg)
 148 {
 149         cg->cache_mode = -1;
 150 }
 151
 152 static struct cgroup_subsys_state *bcachecg_create(struct cgroup *cgroup)
 153 {
 154         struct bch_cgroup *cg;
 155
 156         cg = kzalloc(sizeof(*cg), GFP_KERNEL);
 157         if (!cg)
 158                 return ERR_PTR(-ENOMEM);
 159         init_bch_cgroup(cg);
 160         return &cg->css;
 161 }
 162
 163 static void bcachecg_destroy(struct cgroup *cgroup)
 164 {
 165         struct bch_cgroup *cg = cgroup_to_bcache(cgroup);
 166         kfree(cg);
 167 }
 168
 169 struct cgroup_subsys bcache_subsys = {
 170         .create         = bcachecg_create,
 171         .destroy        = bcachecg_destroy,
 172         .subsys_id      = bcache_subsys_id,
 173         .name           = "bcache",
 174         .module         = THIS_MODULE,
 175 };
 176 EXPORT_SYMBOL_GPL(bcache_subsys);
 177 #endif
 178
 179 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
 180 {
 181 #ifdef CONFIG_CGROUP_BCACHE
 182         int r = bch_bio_to_cgroup(bio)->cache_mode;
 183         if (r >= 0)
 184                 return r;
 185 #endif
 186         return BDEV_CACHE_MODE(&dc->sb);
 187 }
 188
 189 static bool verify(struct cached_dev *dc, struct bio *bio)
 190 {
 191 #ifdef CONFIG_CGROUP_BCACHE
 192         if (bch_bio_to_cgroup(bio)->verify)
 193                 return true;
 194 #endif
 195         return dc->verify;
 196 }
 197
 198 static void bio_csum(struct bio *bio, struct bkey *k)
 199 {
 200         struct bio_vec bv;
 201         struct bvec_iter iter;
 202         uint64_t csum = 0;
 203
 204         bio_for_each_segment(bv, bio, iter) {
 205                 void *d = kmap(bv.bv_page) + bv.bv_offset;
 206                 csum = bch_crc64_update(csum, d, bv.bv_len);
 207                 kunmap(bv.bv_page);
 208         }
 209
 210         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
 211 }
 212
 213 /* Insert data into cache */
 214
 215 static void bch_data_insert_keys(struct closure *cl)
 216 {
 217         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 218         atomic_t *journal_ref = NULL;
 219         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
 220         int ret;
 221
 222         /*
 223          * If we're looping, might already be waiting on
 224          * another journal write - can't wait on more than one journal write at
 225          * a time
 226          *
 227          * XXX: this looks wrong
 228          */
 229 #if 0
 230         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
 231                 closure_sync(&s->cl);
 232 #endif
 233
 234         if (!op->replace)
 235                 journal_ref = bch_journal(op->c, &op->insert_keys,
 236                                           op->flush_journal ? cl : NULL);
 237
 238         ret = bch_btree_insert(op->c, &op->insert_keys,
 239                                journal_ref, replace_key);
 240         if (ret == -ESRCH) {
 241                 op->replace_collision = true;
 242         } else if (ret) {
 243                 op->error               = -ENOMEM;
 244                 op->insert_data_done    = true;
 245         }
 246
 247         if (journal_ref)
 248                 atomic_dec_bug(journal_ref);
 249
 250         if (!op->insert_data_done)
 251                 continue_at(cl, bch_data_insert_start, bcache_wq);
 252
 253         bch_keylist_free(&op->insert_keys);
 254         closure_return(cl);
 255 }
 256
 257 static void bch_data_invalidate(struct closure *cl)
 258 {
 259         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 260         struct bio *bio = op->bio;
 261
 262         pr_debug("invalidating %i sectors from %llu",
 263                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 264
 265         while (bio_sectors(bio)) {
 266                 unsigned sectors = min(bio_sectors(bio),
 267                                        1U << (KEY_SIZE_BITS - 1));
 268
 269                 if (bch_keylist_realloc(&op->insert_keys, 0, op->c))
 270                         goto out;
 271
 272                 bio->bi_iter.bi_sector  += sectors;
 273                 bio->bi_iter.bi_size    -= sectors << 9;
 274
 275                 bch_keylist_add(&op->insert_keys,
 276                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
 277         }
 278
 279         op->insert_data_done = true;
 280         bio_put(bio);
 281 out:
 282         continue_at(cl, bch_data_insert_keys, bcache_wq);
 283 }
 284
 285 static void bch_data_insert_error(struct closure *cl)
 286 {
 287         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 288
 289         /*
 290          * Our data write just errored, which means we've got a bunch of keys to
 291          * insert that point to data that wasn't succesfully written.
 292          *
 293          * We don't have to insert those keys but we still have to invalidate
 294          * that region of the cache - so, if we just strip off all the pointers
 295          * from the keys we'll accomplish just that.
 296          */
 297
 298         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 299
 300         while (src != op->insert_keys.top) {
 301                 struct bkey *n = bkey_next(src);
 302
 303                 SET_KEY_PTRS(src, 0);
 304                 memmove(dst, src, bkey_bytes(src));
 305
 306                 dst = bkey_next(dst);
 307                 src = n;
 308         }
 309
 310         op->insert_keys.top = dst;
 311
 312         bch_data_insert_keys(cl);
 313 }
 314
 315 static void bch_data_insert_endio(struct bio *bio, int error)
 316 {
 317         struct closure *cl = bio->bi_private;
 318         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 319
 320         if (error) {
 321                 /* TODO: We could try to recover from this. */
 322                 if (op->writeback)
 323                         op->error = error;
 324                 else if (!op->replace)
 325                         set_closure_fn(cl, bch_data_insert_error, bcache_wq);
 326                 else
 327                         set_closure_fn(cl, NULL, NULL);
 328         }
 329
 330         bch_bbio_endio(op->c, bio, error, "writing data to cache");
 331 }
 332
 333 static void bch_data_insert_start(struct closure *cl)
 334 {
 335         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 336         struct bio *bio = op->bio, *n;
 337
 338         if (op->bypass)
 339                 return bch_data_invalidate(cl);
 340
 341         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) {
 342                 set_gc_sectors(op->c);
 343                 wake_up_gc(op->c);
 344         }
 345
 346         /*
 347          * Journal writes are marked REQ_FLUSH; if the original write was a
 348          * flush, it'll wait on the journal write.
 349          */
 350         bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA);
 351
 352         do {
 353                 unsigned i;
 354                 struct bkey *k;
 355                 struct bio_set *split = op->c->bio_split;
 356
 357                 /* 1 for the device pointer and 1 for the chksum */
 358                 if (bch_keylist_realloc(&op->insert_keys,
 359                                         1 + (op->csum ? 1 : 0),
 360                                         op->c))
 361                         continue_at(cl, bch_data_insert_keys, bcache_wq);
 362
 363                 k = op->insert_keys.top;
 364                 bkey_init(k);
 365                 SET_KEY_INODE(k, op->inode);
 366                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 367
 368                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 369                                        op->write_point, op->write_prio,
 370                                        op->writeback))
 371                         goto err;
 372
 373                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 374
 375                 n->bi_end_io    = bch_data_insert_endio;
 376                 n->bi_private   = cl;
 377
 378                 if (op->writeback) {
 379                         SET_KEY_DIRTY(k, true);
 380
 381                         for (i = 0; i < KEY_PTRS(k); i++)
 382                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 383                                             GC_MARK_DIRTY);
 384                 }
 385
 386                 SET_KEY_CSUM(k, op->csum);
 387                 if (KEY_CSUM(k))
 388                         bio_csum(n, k);
 389
 390                 trace_bcache_cache_insert(k);
 391                 bch_keylist_push(&op->insert_keys);
 392
 393                 n->bi_rw |= REQ_WRITE;
 394                 bch_submit_bbio(n, op->c, k, 0);
 395         } while (n != bio);
 396
 397         op->insert_data_done = true;
 398         continue_at(cl, bch_data_insert_keys, bcache_wq);
 399 err:
 400         /* bch_alloc_sectors() blocks if s->writeback = true */
 401         BUG_ON(op->writeback);
 402
 403         /*
 404          * But if it's not a writeback write we'd rather just bail out if
 405          * there aren't any buckets ready to write to - it might take awhile and
 406          * we might be starving btree writes for gc or something.
 407          */
 408
 409         if (!op->replace) {
 410                 /*
 411                  * Writethrough write: We can't complete the write until we've
 412                  * updated the index. But we don't want to delay the write while
 413                  * we wait for buckets to be freed up, so just invalidate the
 414                  * rest of the write.
 415                  */
 416                 op->bypass = true;
 417                 return bch_data_invalidate(cl);
 418         } else {
 419                 /*
 420                  * From a cache miss, we can just insert the keys for the data
 421                  * we have written or bail out if we didn't do anything.
 422                  */
 423                 op->insert_data_done = true;
 424                 bio_put(bio);
 425
 426                 if (!bch_keylist_empty(&op->insert_keys))
 427                         continue_at(cl, bch_data_insert_keys, bcache_wq);
 428                 else
 429                         closure_return(cl);
 430         }
 431 }
 432
 433 /**
 434  * bch_data_insert - stick some data in the cache
 435  *
 436  * This is the starting point for any data to end up in a cache device; it could
 437  * be from a normal write, or a writeback write, or a write to a flash only
 438  * volume - it's also used by the moving garbage collector to compact data in
 439  * mostly empty buckets.
 440  *
 441  * It first writes the data to the cache, creating a list of keys to be inserted
 442  * (if the data had to be fragmented there will be multiple keys); after the
 443  * data is written it calls bch_journal, and after the keys have been added to
 444  * the next journal write they're inserted into the btree.
 445  *
 446  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 447  * and op->inode is used for the key inode.
 448  *
 449  * If s->bypass is true, instead of inserting the data it invalidates the
 450  * region of the cache represented by s->cache_bio and op->inode.
 451  */
 452 void bch_data_insert(struct closure *cl)
 453 {
 454         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 455
 456         trace_bcache_write(op->bio, op->writeback, op->bypass);
 457
 458         bch_keylist_init(&op->insert_keys);
 459         bio_get(op->bio);
 460         bch_data_insert_start(cl);
 461 }
 462
 463 /* Congested? */
 464
 465 unsigned bch_get_congested(struct cache_set *c)
 466 {
 467         int i;
 468         long rand;
 469
 470         if (!c->congested_read_threshold_us &&
 471             !c->congested_write_threshold_us)
 472                 return 0;
 473
 474         i = (local_clock_us() - c->congested_last_us) / 1024;
 475         if (i < 0)
 476                 return 0;
 477
 478         i += atomic_read(&c->congested);
 479         if (i >= 0)
 480                 return 0;
 481
 482         i += CONGESTED_MAX;
 483
 484         if (i > 0)
 485                 i = fract_exp_two(i, 6);
 486
 487         rand = get_random_int();
 488         i -= bitmap_weight(&rand, BITS_PER_LONG);
 489
 490         return i > 0 ? i : 1;
 491 }
 492
 493 static void add_sequential(struct task_struct *t)
 494 {
 495         ewma_add(t->sequential_io_avg,
 496                  t->sequential_io, 8, 0);
 497
 498         t->sequential_io = 0;
 499 }
 500
 501 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 502 {
 503         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 504 }
 505
 506 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 507 {
 508         struct cache_set *c = dc->disk.c;
 509         unsigned mode = cache_mode(dc, bio);
 510         unsigned sectors, congested = bch_get_congested(c);
 511         struct task_struct *task = current;
 512         struct io *i;
 513
 514         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 515             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 516             (bio->bi_rw & REQ_DISCARD))
 517                 goto skip;
 518
 519         if (mode == CACHE_MODE_NONE ||
 520             (mode == CACHE_MODE_WRITEAROUND &&
 521              (bio->bi_rw & REQ_WRITE)))
 522                 goto skip;
 523
 524         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 525             bio_sectors(bio) & (c->sb.block_size - 1)) {
 526                 pr_debug("skipping unaligned io");
 527                 goto skip;
 528         }
 529
 530         if (bypass_torture_test(dc)) {
 531                 if ((get_random_int() & 3) == 3)
 532                         goto skip;
 533                 else
 534                         goto rescale;
 535         }
 536
 537         if (!congested && !dc->sequential_cutoff)
 538                 goto rescale;
 539
 540         if (!congested &&
 541             mode == CACHE_MODE_WRITEBACK &&
 542             (bio->bi_rw & REQ_WRITE) &&
 543             (bio->bi_rw & REQ_SYNC))
 544                 goto rescale;
 545
 546         spin_lock(&dc->io_lock);
 547
 548         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
 549                 if (i->last == bio->bi_iter.bi_sector &&
 550                     time_before(jiffies, i->jiffies))
 551                         goto found;
 552
 553         i = list_first_entry(&dc->io_lru, struct io, lru);
 554
 555         add_sequential(task);
 556         i->sequential = 0;
 557 found:
 558         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
 559                 i->sequential   += bio->bi_iter.bi_size;
 560
 561         i->last                  = bio_end_sector(bio);
 562         i->jiffies               = jiffies + msecs_to_jiffies(5000);
 563         task->sequential_io      = i->sequential;
 564
 565         hlist_del(&i->hash);
 566         hlist_add_head(&i->hash, iohash(dc, i->last));
 567         list_move_tail(&i->lru, &dc->io_lru);
 568
 569         spin_unlock(&dc->io_lock);
 570
 571         sectors = max(task->sequential_io,
 572                       task->sequential_io_avg) >> 9;
 573
 574         if (dc->sequential_cutoff &&
 575             sectors >= dc->sequential_cutoff >> 9) {
 576                 trace_bcache_bypass_sequential(bio);
 577                 goto skip;
 578         }
 579
 580         if (congested && sectors >= congested) {
 581                 trace_bcache_bypass_congested(bio);
 582                 goto skip;
 583         }
 584
 585 rescale:
 586         bch_rescale_priorities(c, bio_sectors(bio));
 587         return false;
 588 skip:
 589         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 590         return true;
 591 }
 592
 593 /* Cache lookup */
 594
 595 struct search {
 596         /* Stack frame for bio_complete */
 597         struct closure          cl;
 598
 599         struct bcache_device    *d;
 600
 601         struct bbio             bio;
 602         struct bio              *orig_bio;
 603         struct bio              *cache_miss;
 604
 605         unsigned                insert_bio_sectors;
 606
 607         unsigned                recoverable:1;
 608         unsigned                write:1;
 609         unsigned                read_dirty_data:1;
 610
 611         unsigned long           start_time;
 612
 613         struct btree_op         op;
 614         struct data_insert_op   iop;
 615 };
 616
 617 static void bch_cache_read_endio(struct bio *bio, int error)
 618 {
 619         struct bbio *b = container_of(bio, struct bbio, bio);
 620         struct closure *cl = bio->bi_private;
 621         struct search *s = container_of(cl, struct search, cl);
 622
 623         /*
 624          * If the bucket was reused while our bio was in flight, we might have
 625          * read the wrong data. Set s->error but not error so it doesn't get
 626          * counted against the cache device, but we'll still reread the data
 627          * from the backing device.
 628          */
 629
 630         if (error)
 631                 s->iop.error = error;
 632         else if (ptr_stale(s->iop.c, &b->key, 0)) {
 633                 atomic_long_inc(&s->iop.c->cache_read_races);
 634                 s->iop.error = -EINTR;
 635         }
 636
 637         bch_bbio_endio(s->iop.c, bio, error, "reading from cache");
 638 }
 639
 640 /*
 641  * Read from a single key, handling the initial cache miss if the key starts in
 642  * the middle of the bio
 643  */
 644 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 645 {
 646         struct search *s = container_of(op, struct search, op);
 647         struct bio *n, *bio = &s->bio.bio;
 648         struct bkey *bio_key;
 649         unsigned ptr;
 650
 651         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 652                 return MAP_CONTINUE;
 653
 654         if (KEY_INODE(k) != s->iop.inode ||
 655             KEY_START(k) > bio->bi_iter.bi_sector) {
 656                 unsigned bio_sectors = bio_sectors(bio);
 657                 unsigned sectors = KEY_INODE(k) == s->iop.inode
 658                         ? min_t(uint64_t, INT_MAX,
 659                                 KEY_START(k) - bio->bi_iter.bi_sector)
 660                         : INT_MAX;
 661
 662                 int ret = s->d->cache_miss(b, s, bio, sectors);
 663                 if (ret != MAP_CONTINUE)
 664                         return ret;
 665
 666                 /* if this was a complete miss we shouldn't get here */
 667                 BUG_ON(bio_sectors <= sectors);
 668         }
 669
 670         if (!KEY_SIZE(k))
 671                 return MAP_CONTINUE;
 672
 673         /* XXX: figure out best pointer - for multiple cache devices */
 674         ptr = 0;
 675
 676         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 677
 678         if (KEY_DIRTY(k))
 679                 s->read_dirty_data = true;
 680
 681         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 682                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 683                            GFP_NOIO, s->d->bio_split);
 684
 685         bio_key = &container_of(n, struct bbio, bio)->key;
 686         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 687
 688         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 689         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 690
 691         n->bi_end_io    = bch_cache_read_endio;
 692         n->bi_private   = &s->cl;
 693
 694         /*
 695          * The bucket we're reading from might be reused while our bio
 696          * is in flight, and we could then end up reading the wrong
 697          * data.
 698          *
 699          * We guard against this by checking (in cache_read_endio()) if
 700          * the pointer is stale again; if so, we treat it as an error
 701          * and reread from the backing device (but we don't pass that
 702          * error up anywhere).
 703          */
 704
 705         __bch_submit_bbio(n, b->c);
 706         return n == bio ? MAP_DONE : MAP_CONTINUE;
 707 }
 708
 709 static void cache_lookup(struct closure *cl)
 710 {
 711         struct search *s = container_of(cl, struct search, iop.cl);
 712         struct bio *bio = &s->bio.bio;
 713
 714         int ret = bch_btree_map_keys(&s->op, s->iop.c,
 715                                      &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 716                                      cache_lookup_fn, MAP_END_KEY);
 717         if (ret == -EAGAIN)
 718                 continue_at(cl, cache_lookup, bcache_wq);
 719
 720         closure_return(cl);
 721 }
 722
 723 /* Common code for the make_request functions */
 724
 725 static void request_endio(struct bio *bio, int error)
 726 {
 727         struct closure *cl = bio->bi_private;
 728
 729         if (error) {
 730                 struct search *s = container_of(cl, struct search, cl);
 731                 s->iop.error = error;
 732                 /* Only cache read errors are recoverable */
 733                 s->recoverable = false;
 734         }
 735
 736         bio_put(bio);
 737         closure_put(cl);
 738 }
 739
 740 static void bio_complete(struct search *s)
 741 {
 742         if (s->orig_bio) {
 743                 int cpu, rw = bio_data_dir(s->orig_bio);
 744                 unsigned long duration = jiffies - s->start_time;
 745
 746                 cpu = part_stat_lock();
 747                 part_round_stats(cpu, &s->d->disk->part0);
 748                 part_stat_add(cpu, &s->d->disk->part0, ticks[rw], duration);
 749                 part_stat_unlock();
 750
 751                 trace_bcache_request_end(s->d, s->orig_bio);
 752                 bio_endio(s->orig_bio, s->iop.error);
 753                 s->orig_bio = NULL;
 754         }
 755 }
 756
 757 static void do_bio_hook(struct search *s)
 758 {
 759         struct bio *bio = &s->bio.bio;
 760
 761         bio_init(bio);
 762         __bio_clone_fast(bio, s->orig_bio);
 763         bio->bi_end_io          = request_endio;
 764         bio->bi_private         = &s->cl;
 765
 766         atomic_set(&bio->bi_cnt, 3);
 767 }
 768
 769 static void search_free(struct closure *cl)
 770 {
 771         struct search *s = container_of(cl, struct search, cl);
 772         bio_complete(s);
 773
 774         if (s->iop.bio)
 775                 bio_put(s->iop.bio);
 776
 777         closure_debug_destroy(cl);
 778         mempool_free(s, s->d->c->search);
 779 }
 780
 781 static struct search *search_alloc(struct bio *bio, struct bcache_device *d)
 782 {
 783         struct search *s;
 784
 785         s = mempool_alloc(d->c->search, GFP_NOIO);
 786         memset(s, 0, offsetof(struct search, iop.insert_keys));
 787
 788         __closure_init(&s->cl, NULL);
 789
 790         s->iop.inode            = d->id;
 791         s->iop.c                = d->c;
 792         s->d                    = d;
 793         s->op.lock              = -1;
 794         s->iop.write_point      = hash_long((unsigned long) current, 16);
 795         s->orig_bio             = bio;
 796         s->write                = (bio->bi_rw & REQ_WRITE) != 0;
 797         s->iop.flush_journal    = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0;
 798         s->recoverable          = 1;
 799         s->start_time           = jiffies;
 800         do_bio_hook(s);
 801
 802         return s;
 803 }
 804
 805 /* Cached devices */
 806
 807 static void cached_dev_bio_complete(struct closure *cl)
 808 {
 809         struct search *s = container_of(cl, struct search, cl);
 810         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 811
 812         search_free(cl);
 813         cached_dev_put(dc);
 814 }
 815
 816 /* Process reads */
 817
 818 static void cached_dev_cache_miss_done(struct closure *cl)
 819 {
 820         struct search *s = container_of(cl, struct search, cl);
 821
 822         if (s->iop.replace_collision)
 823                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 824
 825         if (s->iop.bio) {
 826                 int i;
 827                 struct bio_vec *bv;
 828
 829                 bio_for_each_segment_all(bv, s->iop.bio, i)
 830                         __free_page(bv->bv_page);
 831         }
 832
 833         cached_dev_bio_complete(cl);
 834 }
 835
 836 static void cached_dev_read_error(struct closure *cl)
 837 {
 838         struct search *s = container_of(cl, struct search, cl);
 839         struct bio *bio = &s->bio.bio;
 840
 841         if (s->recoverable) {
 842                 /* Retry from the backing device: */
 843                 trace_bcache_read_retry(s->orig_bio);
 844
 845                 s->iop.error = 0;
 846                 do_bio_hook(s);
 847
 848                 /* XXX: invalidate cache */
 849
 850                 closure_bio_submit(bio, cl, s->d);
 851         }
 852
 853         continue_at(cl, cached_dev_cache_miss_done, NULL);
 854 }
 855
 856 static void cached_dev_read_done(struct closure *cl)
 857 {
 858         struct search *s = container_of(cl, struct search, cl);
 859         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 860
 861         /*
 862          * We had a cache miss; cache_bio now contains data ready to be inserted
 863          * into the cache.
 864          *
 865          * First, we copy the data we just read from cache_bio's bounce buffers
 866          * to the buffers the original bio pointed to:
 867          */
 868
 869         if (s->iop.bio) {
 870                 bio_reset(s->iop.bio);
 871                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 872                 s->iop.bio->bi_bdev = s->cache_miss->bi_bdev;
 873                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 874                 bch_bio_map(s->iop.bio, NULL);
 875
 876                 bio_copy_data(s->cache_miss, s->iop.bio);
 877
 878                 bio_put(s->cache_miss);
 879                 s->cache_miss = NULL;
 880         }
 881
 882         if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
 883                 bch_data_verify(dc, s->orig_bio);
 884
 885         bio_complete(s);
 886
 887         if (s->iop.bio &&
 888             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 889                 BUG_ON(!s->iop.replace);
 890                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 891         }
 892
 893         continue_at(cl, cached_dev_cache_miss_done, NULL);
 894 }
 895
 896 static void cached_dev_read_done_bh(struct closure *cl)
 897 {
 898         struct search *s = container_of(cl, struct search, cl);
 899         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 900
 901         bch_mark_cache_accounting(s->iop.c, s->d,
 902                                   !s->cache_miss, s->iop.bypass);
 903         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 904
 905         if (s->iop.error)
 906                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 907         else if (s->iop.bio || verify(dc, &s->bio.bio))
 908                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 909         else
 910                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 911 }
 912
 913 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 914                                  struct bio *bio, unsigned sectors)
 915 {
 916         int ret = MAP_CONTINUE;
 917         unsigned reada = 0;
 918         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 919         struct bio *miss, *cache_bio;
 920
 921         if (s->cache_miss || s->iop.bypass) {
 922                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 923                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 924                 goto out_submit;
 925         }
 926
 927         if (!(bio->bi_rw & REQ_RAHEAD) &&
 928             !(bio->bi_rw & REQ_META) &&
 929             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 930                 reada = min_t(sector_t, dc->readahead >> 9,
 931                               bdev_sectors(bio->bi_bdev) - bio_end_sector(bio));
 932
 933         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 934
 935         s->iop.replace_key = KEY(s->iop.inode,
 936                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 937                                  s->insert_bio_sectors);
 938
 939         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 940         if (ret)
 941                 return ret;
 942
 943         s->iop.replace = true;
 944
 945         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 946
 947         /* btree_search_recurse()'s btree iterator is no good anymore */
 948         ret = miss == bio ? MAP_DONE : -EINTR;
 949
 950         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 951                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 952                         dc->disk.bio_split);
 953         if (!cache_bio)
 954                 goto out_submit;
 955
 956         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 957         cache_bio->bi_bdev              = miss->bi_bdev;
 958         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 959
 960         cache_bio->bi_end_io    = request_endio;
 961         cache_bio->bi_private   = &s->cl;
 962
 963         bch_bio_map(cache_bio, NULL);
 964         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 965                 goto out_put;
 966
 967         if (reada)
 968                 bch_mark_cache_readahead(s->iop.c, s->d);
 969
 970         s->cache_miss   = miss;
 971         s->iop.bio      = cache_bio;
 972         bio_get(cache_bio);
 973         closure_bio_submit(cache_bio, &s->cl, s->d);
 974
 975         return ret;
 976 out_put:
 977         bio_put(cache_bio);
 978 out_submit:
 979         miss->bi_end_io         = request_endio;
 980         miss->bi_private        = &s->cl;
 981         closure_bio_submit(miss, &s->cl, s->d);
 982         return ret;
 983 }
 984
 985 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 986 {
 987         struct closure *cl = &s->cl;
 988
 989         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 990         continue_at(cl, cached_dev_read_done_bh, NULL);
 991 }
 992
 993 /* Process writes */
 994
 995 static void cached_dev_write_complete(struct closure *cl)
 996 {
 997         struct search *s = container_of(cl, struct search, cl);
 998         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 999
1000         up_read_non_owner(&dc->writeback_lock);
1001         cached_dev_bio_complete(cl);
1002 }
1003
1004 static void cached_dev_write(struct cached_dev *dc, struct search *s)
1005 {
1006         struct closure *cl = &s->cl;
1007         struct bio *bio = &s->bio.bio;
1008         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
1009         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
1010
1011         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
1012
1013         down_read_non_owner(&dc->writeback_lock);
1014         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
1015                 /*
1016                  * We overlap with some dirty data undergoing background
1017                  * writeback, force this write to writeback
1018                  */
1019                 s->iop.bypass = false;
1020                 s->iop.writeback = true;
1021         }
1022
1023         /*
1024          * Discards aren't _required_ to do anything, so skipping if
1025          * check_overlapping returned true is ok
1026          *
1027          * But check_overlapping drops dirty keys for which io hasn't started,
1028          * so we still want to call it.
1029          */
1030         if (bio->bi_rw & REQ_DISCARD)
1031                 s->iop.bypass = true;
1032
1033         if (should_writeback(dc, s->orig_bio,
1034                              cache_mode(dc, bio),
1035                              s->iop.bypass)) {
1036                 s->iop.bypass = false;
1037                 s->iop.writeback = true;
1038         }
1039
1040         if (s->iop.bypass) {
1041                 s->iop.bio = s->orig_bio;
1042                 bio_get(s->iop.bio);
1043
1044                 if (!(bio->bi_rw & REQ_DISCARD) ||
1045                     blk_queue_discard(bdev_get_queue(dc->bdev)))
1046                         closure_bio_submit(bio, cl, s->d);
1047         } else if (s->iop.writeback) {
1048                 bch_writeback_add(dc);
1049                 s->iop.bio = bio;
1050
1051                 if (bio->bi_rw & REQ_FLUSH) {
1052                         /* Also need to send a flush to the backing device */
1053                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
1054                                                              dc->disk.bio_split);
1055
1056                         flush->bi_rw    = WRITE_FLUSH;
1057                         flush->bi_bdev  = bio->bi_bdev;
1058                         flush->bi_end_io = request_endio;
1059                         flush->bi_private = cl;
1060
1061                         closure_bio_submit(flush, cl, s->d);
1062                 }
1063         } else {
1064                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
1065
1066                 closure_bio_submit(bio, cl, s->d);
1067         }
1068
1069         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1070         continue_at(cl, cached_dev_write_complete, NULL);
1071 }
1072
1073 static void cached_dev_nodata(struct closure *cl)
1074 {
1075         struct search *s = container_of(cl, struct search, cl);
1076         struct bio *bio = &s->bio.bio;
1077
1078         if (s->iop.flush_journal)
1079                 bch_journal_meta(s->iop.c, cl);
1080
1081         /* If it's a flush, we send the flush to the backing device too */
1082         closure_bio_submit(bio, cl, s->d);
1083
1084         continue_at(cl, cached_dev_bio_complete, NULL);
1085 }
1086
1087 /* Cached devices - read & write stuff */
1088
1089 static void cached_dev_make_request(struct request_queue *q, struct bio *bio)
1090 {
1091         struct search *s;
1092         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1093         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1094         int cpu, rw = bio_data_dir(bio);
1095
1096         cpu = part_stat_lock();
1097         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1098         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1099         part_stat_unlock();
1100
1101         bio->bi_bdev = dc->bdev;
1102         bio->bi_iter.bi_sector += dc->sb.data_offset;
1103
1104         if (cached_dev_get(dc)) {
1105                 s = search_alloc(bio, d);
1106                 trace_bcache_request_start(s->d, bio);
1107
1108                 if (!bio->bi_iter.bi_size) {
1109                         /*
1110                          * can't call bch_journal_meta from under
1111                          * generic_make_request
1112                          */
1113                         continue_at_nobarrier(&s->cl,
1114                                               cached_dev_nodata,
1115                                               bcache_wq);
1116                 } else {
1117                         s->iop.bypass = check_should_bypass(dc, bio);
1118
1119                         if (rw)
1120                                 cached_dev_write(dc, s);
1121                         else
1122                                 cached_dev_read(dc, s);
1123                 }
1124         } else {
1125                 if ((bio->bi_rw & REQ_DISCARD) &&
1126                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1127                         bio_endio(bio, 0);
1128                 else
1129                         bch_generic_make_request(bio, &d->bio_split_hook);
1130         }
1131 }
1132
1133 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1134                             unsigned int cmd, unsigned long arg)
1135 {
1136         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1137         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1138 }
1139
1140 static int cached_dev_congested(void *data, int bits)
1141 {
1142         struct bcache_device *d = data;
1143         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1144         struct request_queue *q = bdev_get_queue(dc->bdev);
1145         int ret = 0;
1146
1147         if (bdi_congested(&q->backing_dev_info, bits))
1148                 return 1;
1149
1150         if (cached_dev_get(dc)) {
1151                 unsigned i;
1152                 struct cache *ca;
1153
1154                 for_each_cache(ca, d->c, i) {
1155                         q = bdev_get_queue(ca->bdev);
1156                         ret |= bdi_congested(&q->backing_dev_info, bits);
1157                 }
1158
1159                 cached_dev_put(dc);
1160         }
1161
1162         return ret;
1163 }
1164
1165 void bch_cached_dev_request_init(struct cached_dev *dc)
1166 {
1167         struct gendisk *g = dc->disk.disk;
1168
1169         g->queue->make_request_fn               = cached_dev_make_request;
1170         g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1171         dc->disk.cache_miss                     = cached_dev_cache_miss;
1172         dc->disk.ioctl                          = cached_dev_ioctl;
1173 }
1174
1175 /* Flash backed devices */
1176
1177 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1178                                 struct bio *bio, unsigned sectors)
1179 {
1180         struct bio_vec bv;
1181         struct bvec_iter iter;
1182
1183         /* Zero fill bio */
1184
1185         bio_for_each_segment(bv, bio, iter) {
1186                 unsigned j = min(bv.bv_len >> 9, sectors);
1187
1188                 void *p = kmap(bv.bv_page);
1189                 memset(p + bv.bv_offset, 0, j << 9);
1190                 kunmap(bv.bv_page);
1191
1192                 sectors -= j;
1193         }
1194
1195         bio_advance(bio, min(sectors << 9, bio->bi_iter.bi_size));
1196
1197         if (!bio->bi_iter.bi_size)
1198                 return MAP_DONE;
1199
1200         return MAP_CONTINUE;
1201 }
1202
1203 static void flash_dev_nodata(struct closure *cl)
1204 {
1205         struct search *s = container_of(cl, struct search, cl);
1206
1207         if (s->iop.flush_journal)
1208                 bch_journal_meta(s->iop.c, cl);
1209
1210         continue_at(cl, search_free, NULL);
1211 }
1212
1213 static void flash_dev_make_request(struct request_queue *q, struct bio *bio)
1214 {
1215         struct search *s;
1216         struct closure *cl;
1217         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1218         int cpu, rw = bio_data_dir(bio);
1219
1220         cpu = part_stat_lock();
1221         part_stat_inc(cpu, &d->disk->part0, ios[rw]);
1222         part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio));
1223         part_stat_unlock();
1224
1225         s = search_alloc(bio, d);
1226         cl = &s->cl;
1227         bio = &s->bio.bio;
1228
1229         trace_bcache_request_start(s->d, bio);
1230
1231         if (!bio->bi_iter.bi_size) {
1232                 /*
1233                  * can't call bch_journal_meta from under
1234                  * generic_make_request
1235                  */
1236                 continue_at_nobarrier(&s->cl,
1237                                       flash_dev_nodata,
1238                                       bcache_wq);
1239         } else if (rw) {
1240                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1241                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1242                                         &KEY(d->id, bio_end_sector(bio), 0));
1243
1244                 s->iop.bypass           = (bio->bi_rw & REQ_DISCARD) != 0;
1245                 s->iop.writeback        = true;
1246                 s->iop.bio              = bio;
1247
1248                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1249         } else {
1250                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1251         }
1252
1253         continue_at(cl, search_free, NULL);
1254 }
1255
1256 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1257                            unsigned int cmd, unsigned long arg)
1258 {
1259         return -ENOTTY;
1260 }
1261
1262 static int flash_dev_congested(void *data, int bits)
1263 {
1264         struct bcache_device *d = data;
1265         struct request_queue *q;
1266         struct cache *ca;
1267         unsigned i;
1268         int ret = 0;
1269
1270         for_each_cache(ca, d->c, i) {
1271                 q = bdev_get_queue(ca->bdev);
1272                 ret |= bdi_congested(&q->backing_dev_info, bits);
1273         }
1274
1275         return ret;
1276 }
1277
1278 void bch_flash_dev_request_init(struct bcache_device *d)
1279 {
1280         struct gendisk *g = d->disk;
1281
1282         g->queue->make_request_fn               = flash_dev_make_request;
1283         g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1284         d->cache_miss                           = flash_dev_cache_miss;
1285         d->ioctl                                = flash_dev_ioctl;
1286 }
1287
1288 void bch_request_exit(void)
1289 {
1290 #ifdef CONFIG_CGROUP_BCACHE
1291         cgroup_unload_subsys(&bcache_subsys);
1292 #endif
1293         if (bch_search_cache)
1294                 kmem_cache_destroy(bch_search_cache);
1295 }
1296
1297 int __init bch_request_init(void)
1298 {
1299         bch_search_cache = KMEM_CACHE(search, 0);
1300         if (!bch_search_cache)
1301                 return -ENOMEM;
1302
1303 #ifdef CONFIG_CGROUP_BCACHE
1304         cgroup_load_subsys(&bcache_subsys);
1305         init_bch_cgroup(&bcache_default_cgroup);
1306
1307         cgroup_add_cftypes(&bcache_subsys, bch_files);
1308 #endif
1309         return 0;
1310 }