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/module.h>
  16 #include <linux/hash.h>
  17 #include <linux/random.h>
  18 #include <linux/backing-dev.h>
  19
  20 #include <trace/events/bcache.h>
  21
  22 #define CUTOFF_CACHE_ADD        95
  23 #define CUTOFF_CACHE_READA      90
  24
  25 struct kmem_cache *bch_search_cache;
  26
  27 static void bch_data_insert_start(struct closure *);
  28
  29 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
  30 {
  31         return BDEV_CACHE_MODE(&dc->sb);
  32 }
  33
  34 static bool verify(struct cached_dev *dc, struct bio *bio)
  35 {
  36         return dc->verify;
  37 }
  38
  39 static void bio_csum(struct bio *bio, struct bkey *k)
  40 {
  41         struct bio_vec bv;
  42         struct bvec_iter iter;
  43         uint64_t csum = 0;
  44
  45         bio_for_each_segment(bv, bio, iter) {
  46                 void *d = kmap(bv.bv_page) + bv.bv_offset;
  47                 csum = bch_crc64_update(csum, d, bv.bv_len);
  48                 kunmap(bv.bv_page);
  49         }
  50
  51         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
  52 }
  53
  54 /* Insert data into cache */
  55
  56 static void bch_data_insert_keys(struct closure *cl)
  57 {
  58         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
  59         atomic_t *journal_ref = NULL;
  60         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
  61         int ret;
  62
  63         /*
  64          * If we're looping, might already be waiting on
  65          * another journal write - can't wait on more than one journal write at
  66          * a time
  67          *
  68          * XXX: this looks wrong
  69          */
  70 #if 0
  71         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
  72                 closure_sync(&s->cl);
  73 #endif
  74
  75         if (!op->replace)
  76                 journal_ref = bch_journal(op->c, &op->insert_keys,
  77                                           op->flush_journal ? cl : NULL);
  78
  79         ret = bch_btree_insert(op->c, &op->insert_keys,
  80                                journal_ref, replace_key);
  81         if (ret == -ESRCH) {
  82                 op->replace_collision = true;
  83         } else if (ret) {
  84                 op->error               = -ENOMEM;
  85                 op->insert_data_done    = true;
  86         }
  87
  88         if (journal_ref)
  89                 atomic_dec_bug(journal_ref);
  90
  91         if (!op->insert_data_done) {
  92                 continue_at(cl, bch_data_insert_start, op->wq);
  93                 return;
  94         }
  95
  96         bch_keylist_free(&op->insert_keys);
  97         closure_return(cl);
  98 }
  99
 100 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
 101                                struct cache_set *c)
 102 {
 103         size_t oldsize = bch_keylist_nkeys(l);
 104         size_t newsize = oldsize + u64s;
 105
 106         /*
 107          * The journalling code doesn't handle the case where the keys to insert
 108          * is bigger than an empty write: If we just return -ENOMEM here,
 109          * bio_insert() and bio_invalidate() will insert the keys created so far
 110          * and finish the rest when the keylist is empty.
 111          */
 112         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
 113                 return -ENOMEM;
 114
 115         return __bch_keylist_realloc(l, u64s);
 116 }
 117
 118 static void bch_data_invalidate(struct closure *cl)
 119 {
 120         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 121         struct bio *bio = op->bio;
 122
 123         pr_debug("invalidating %i sectors from %llu",
 124                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 125
 126         while (bio_sectors(bio)) {
 127                 unsigned sectors = min(bio_sectors(bio),
 128                                        1U << (KEY_SIZE_BITS - 1));
 129
 130                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
 131                         goto out;
 132
 133                 bio->bi_iter.bi_sector  += sectors;
 134                 bio->bi_iter.bi_size    -= sectors << 9;
 135
 136                 bch_keylist_add(&op->insert_keys,
 137                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
 138         }
 139
 140         op->insert_data_done = true;
 141         bio_put(bio);
 142 out:
 143         continue_at(cl, bch_data_insert_keys, op->wq);
 144 }
 145
 146 static void bch_data_insert_error(struct closure *cl)
 147 {
 148         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 149
 150         /*
 151          * Our data write just errored, which means we've got a bunch of keys to
 152          * insert that point to data that wasn't succesfully written.
 153          *
 154          * We don't have to insert those keys but we still have to invalidate
 155          * that region of the cache - so, if we just strip off all the pointers
 156          * from the keys we'll accomplish just that.
 157          */
 158
 159         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 160
 161         while (src != op->insert_keys.top) {
 162                 struct bkey *n = bkey_next(src);
 163
 164                 SET_KEY_PTRS(src, 0);
 165                 memmove(dst, src, bkey_bytes(src));
 166
 167                 dst = bkey_next(dst);
 168                 src = n;
 169         }
 170
 171         op->insert_keys.top = dst;
 172
 173         bch_data_insert_keys(cl);
 174 }
 175
 176 static void bch_data_insert_endio(struct bio *bio)
 177 {
 178         struct closure *cl = bio->bi_private;
 179         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 180
 181         if (bio->bi_error) {
 182                 /* TODO: We could try to recover from this. */
 183                 if (op->writeback)
 184                         op->error = bio->bi_error;
 185                 else if (!op->replace)
 186                         set_closure_fn(cl, bch_data_insert_error, op->wq);
 187                 else
 188                         set_closure_fn(cl, NULL, NULL);
 189         }
 190
 191         bch_bbio_endio(op->c, bio, bio->bi_error, "writing data to cache");
 192 }
 193
 194 static void bch_data_insert_start(struct closure *cl)
 195 {
 196         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 197         struct bio *bio = op->bio, *n;
 198
 199         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0)
 200                 wake_up_gc(op->c);
 201
 202         if (op->bypass)
 203                 return bch_data_invalidate(cl);
 204
 205         /*
 206          * Journal writes are marked REQ_FLUSH; if the original write was a
 207          * flush, it'll wait on the journal write.
 208          */
 209         bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA);
 210
 211         do {
 212                 unsigned i;
 213                 struct bkey *k;
 214                 struct bio_set *split = op->c->bio_split;
 215
 216                 /* 1 for the device pointer and 1 for the chksum */
 217                 if (bch_keylist_realloc(&op->insert_keys,
 218                                         3 + (op->csum ? 1 : 0),
 219                                         op->c)) {
 220                         continue_at(cl, bch_data_insert_keys, op->wq);
 221                         return;
 222                 }
 223
 224                 k = op->insert_keys.top;
 225                 bkey_init(k);
 226                 SET_KEY_INODE(k, op->inode);
 227                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 228
 229                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 230                                        op->write_point, op->write_prio,
 231                                        op->writeback))
 232                         goto err;
 233
 234                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 235
 236                 n->bi_end_io    = bch_data_insert_endio;
 237                 n->bi_private   = cl;
 238
 239                 if (op->writeback) {
 240                         SET_KEY_DIRTY(k, true);
 241
 242                         for (i = 0; i < KEY_PTRS(k); i++)
 243                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 244                                             GC_MARK_DIRTY);
 245                 }
 246
 247                 SET_KEY_CSUM(k, op->csum);
 248                 if (KEY_CSUM(k))
 249                         bio_csum(n, k);
 250
 251                 trace_bcache_cache_insert(k);
 252                 bch_keylist_push(&op->insert_keys);
 253
 254                 n->bi_rw |= REQ_WRITE;
 255                 bch_submit_bbio(n, op->c, k, 0);
 256         } while (n != bio);
 257
 258         op->insert_data_done = true;
 259         continue_at(cl, bch_data_insert_keys, op->wq);
 260         return;
 261 err:
 262         /* bch_alloc_sectors() blocks if s->writeback = true */
 263         BUG_ON(op->writeback);
 264
 265         /*
 266          * But if it's not a writeback write we'd rather just bail out if
 267          * there aren't any buckets ready to write to - it might take awhile and
 268          * we might be starving btree writes for gc or something.
 269          */
 270
 271         if (!op->replace) {
 272                 /*
 273                  * Writethrough write: We can't complete the write until we've
 274                  * updated the index. But we don't want to delay the write while
 275                  * we wait for buckets to be freed up, so just invalidate the
 276                  * rest of the write.
 277                  */
 278                 op->bypass = true;
 279                 return bch_data_invalidate(cl);
 280         } else {
 281                 /*
 282                  * From a cache miss, we can just insert the keys for the data
 283                  * we have written or bail out if we didn't do anything.
 284                  */
 285                 op->insert_data_done = true;
 286                 bio_put(bio);
 287
 288                 if (!bch_keylist_empty(&op->insert_keys))
 289                         continue_at(cl, bch_data_insert_keys, op->wq);
 290                 else
 291                         closure_return(cl);
 292         }
 293 }
 294
 295 /**
 296  * bch_data_insert - stick some data in the cache
 297  *
 298  * This is the starting point for any data to end up in a cache device; it could
 299  * be from a normal write, or a writeback write, or a write to a flash only
 300  * volume - it's also used by the moving garbage collector to compact data in
 301  * mostly empty buckets.
 302  *
 303  * It first writes the data to the cache, creating a list of keys to be inserted
 304  * (if the data had to be fragmented there will be multiple keys); after the
 305  * data is written it calls bch_journal, and after the keys have been added to
 306  * the next journal write they're inserted into the btree.
 307  *
 308  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 309  * and op->inode is used for the key inode.
 310  *
 311  * If s->bypass is true, instead of inserting the data it invalidates the
 312  * region of the cache represented by s->cache_bio and op->inode.
 313  */
 314 void bch_data_insert(struct closure *cl)
 315 {
 316         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 317
 318         trace_bcache_write(op->c, op->inode, op->bio,
 319                            op->writeback, op->bypass);
 320
 321         bch_keylist_init(&op->insert_keys);
 322         bio_get(op->bio);
 323         bch_data_insert_start(cl);
 324 }
 325
 326 /* Congested? */
 327
 328 unsigned bch_get_congested(struct cache_set *c)
 329 {
 330         int i;
 331         long rand;
 332
 333         if (!c->congested_read_threshold_us &&
 334             !c->congested_write_threshold_us)
 335                 return 0;
 336
 337         i = (local_clock_us() - c->congested_last_us) / 1024;
 338         if (i < 0)
 339                 return 0;
 340
 341         i += atomic_read(&c->congested);
 342         if (i >= 0)
 343                 return 0;
 344
 345         i += CONGESTED_MAX;
 346
 347         if (i > 0)
 348                 i = fract_exp_two(i, 6);
 349
 350         rand = get_random_int();
 351         i -= bitmap_weight(&rand, BITS_PER_LONG);
 352
 353         return i > 0 ? i : 1;
 354 }
 355
 356 static void add_sequential(struct task_struct *t)
 357 {
 358         ewma_add(t->sequential_io_avg,
 359                  t->sequential_io, 8, 0);
 360
 361         t->sequential_io = 0;
 362 }
 363
 364 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 365 {
 366         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 367 }
 368
 369 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 370 {
 371         struct cache_set *c = dc->disk.c;
 372         unsigned mode = cache_mode(dc, bio);
 373         unsigned sectors, congested = bch_get_congested(c);
 374         struct task_struct *task = current;
 375         struct io *i;
 376
 377         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 378             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 379             (bio->bi_rw & REQ_DISCARD))
 380                 goto skip;
 381
 382         if (mode == CACHE_MODE_NONE ||
 383             (mode == CACHE_MODE_WRITEAROUND &&
 384              (bio->bi_rw & REQ_WRITE)))
 385                 goto skip;
 386
 387         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 388             bio_sectors(bio) & (c->sb.block_size - 1)) {
 389                 pr_debug("skipping unaligned io");
 390                 goto skip;
 391         }
 392
 393         if (bypass_torture_test(dc)) {
 394                 if ((get_random_int() & 3) == 3)
 395                         goto skip;
 396                 else
 397                         goto rescale;
 398         }
 399
 400         if (!congested && !dc->sequential_cutoff)
 401                 goto rescale;
 402
 403         if (!congested &&
 404             mode == CACHE_MODE_WRITEBACK &&
 405             (bio->bi_rw & REQ_WRITE) &&
 406             (bio->bi_rw & REQ_SYNC))
 407                 goto rescale;
 408
 409         spin_lock(&dc->io_lock);
 410
 411         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
 412                 if (i->last == bio->bi_iter.bi_sector &&
 413                     time_before(jiffies, i->jiffies))
 414                         goto found;
 415
 416         i = list_first_entry(&dc->io_lru, struct io, lru);
 417
 418         add_sequential(task);
 419         i->sequential = 0;
 420 found:
 421         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
 422                 i->sequential   += bio->bi_iter.bi_size;
 423
 424         i->last                  = bio_end_sector(bio);
 425         i->jiffies               = jiffies + msecs_to_jiffies(5000);
 426         task->sequential_io      = i->sequential;
 427
 428         hlist_del(&i->hash);
 429         hlist_add_head(&i->hash, iohash(dc, i->last));
 430         list_move_tail(&i->lru, &dc->io_lru);
 431
 432         spin_unlock(&dc->io_lock);
 433
 434         sectors = max(task->sequential_io,
 435                       task->sequential_io_avg) >> 9;
 436
 437         if (dc->sequential_cutoff &&
 438             sectors >= dc->sequential_cutoff >> 9) {
 439                 trace_bcache_bypass_sequential(bio);
 440                 goto skip;
 441         }
 442
 443         if (congested && sectors >= congested) {
 444                 trace_bcache_bypass_congested(bio);
 445                 goto skip;
 446         }
 447
 448 rescale:
 449         bch_rescale_priorities(c, bio_sectors(bio));
 450         return false;
 451 skip:
 452         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
 453         return true;
 454 }
 455
 456 /* Cache lookup */
 457
 458 struct search {
 459         /* Stack frame for bio_complete */
 460         struct closure          cl;
 461
 462         struct bbio             bio;
 463         struct bio              *orig_bio;
 464         struct bio              *cache_miss;
 465         struct bcache_device    *d;
 466
 467         unsigned                insert_bio_sectors;
 468         unsigned                recoverable:1;
 469         unsigned                write:1;
 470         unsigned                read_dirty_data:1;
 471
 472         unsigned long           start_time;
 473
 474         struct btree_op         op;
 475         struct data_insert_op   iop;
 476 };
 477
 478 static void bch_cache_read_endio(struct bio *bio)
 479 {
 480         struct bbio *b = container_of(bio, struct bbio, bio);
 481         struct closure *cl = bio->bi_private;
 482         struct search *s = container_of(cl, struct search, cl);
 483
 484         /*
 485          * If the bucket was reused while our bio was in flight, we might have
 486          * read the wrong data. Set s->error but not error so it doesn't get
 487          * counted against the cache device, but we'll still reread the data
 488          * from the backing device.
 489          */
 490
 491         if (bio->bi_error)
 492                 s->iop.error = bio->bi_error;
 493         else if (!KEY_DIRTY(&b->key) &&
 494                  ptr_stale(s->iop.c, &b->key, 0)) {
 495                 atomic_long_inc(&s->iop.c->cache_read_races);
 496                 s->iop.error = -EINTR;
 497         }
 498
 499         bch_bbio_endio(s->iop.c, bio, bio->bi_error, "reading from cache");
 500 }
 501
 502 /*
 503  * Read from a single key, handling the initial cache miss if the key starts in
 504  * the middle of the bio
 505  */
 506 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 507 {
 508         struct search *s = container_of(op, struct search, op);
 509         struct bio *n, *bio = &s->bio.bio;
 510         struct bkey *bio_key;
 511         unsigned ptr;
 512
 513         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 514                 return MAP_CONTINUE;
 515
 516         if (KEY_INODE(k) != s->iop.inode ||
 517             KEY_START(k) > bio->bi_iter.bi_sector) {
 518                 unsigned bio_sectors = bio_sectors(bio);
 519                 unsigned sectors = KEY_INODE(k) == s->iop.inode
 520                         ? min_t(uint64_t, INT_MAX,
 521                                 KEY_START(k) - bio->bi_iter.bi_sector)
 522                         : INT_MAX;
 523
 524                 int ret = s->d->cache_miss(b, s, bio, sectors);
 525                 if (ret != MAP_CONTINUE)
 526                         return ret;
 527
 528                 /* if this was a complete miss we shouldn't get here */
 529                 BUG_ON(bio_sectors <= sectors);
 530         }
 531
 532         if (!KEY_SIZE(k))
 533                 return MAP_CONTINUE;
 534
 535         /* XXX: figure out best pointer - for multiple cache devices */
 536         ptr = 0;
 537
 538         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 539
 540         if (KEY_DIRTY(k))
 541                 s->read_dirty_data = true;
 542
 543         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 544                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 545                            GFP_NOIO, s->d->bio_split);
 546
 547         bio_key = &container_of(n, struct bbio, bio)->key;
 548         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 549
 550         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 551         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 552
 553         n->bi_end_io    = bch_cache_read_endio;
 554         n->bi_private   = &s->cl;
 555
 556         /*
 557          * The bucket we're reading from might be reused while our bio
 558          * is in flight, and we could then end up reading the wrong
 559          * data.
 560          *
 561          * We guard against this by checking (in cache_read_endio()) if
 562          * the pointer is stale again; if so, we treat it as an error
 563          * and reread from the backing device (but we don't pass that
 564          * error up anywhere).
 565          */
 566
 567         __bch_submit_bbio(n, b->c);
 568         return n == bio ? MAP_DONE : MAP_CONTINUE;
 569 }
 570
 571 static void cache_lookup(struct closure *cl)
 572 {
 573         struct search *s = container_of(cl, struct search, iop.cl);
 574         struct bio *bio = &s->bio.bio;
 575         int ret;
 576
 577         bch_btree_op_init(&s->op, -1);
 578
 579         ret = bch_btree_map_keys(&s->op, s->iop.c,
 580                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 581                                  cache_lookup_fn, MAP_END_KEY);
 582         if (ret == -EAGAIN) {
 583                 continue_at(cl, cache_lookup, bcache_wq);
 584                 return;
 585         }
 586
 587         closure_return(cl);
 588 }
 589
 590 /* Common code for the make_request functions */
 591
 592 static void request_endio(struct bio *bio)
 593 {
 594         struct closure *cl = bio->bi_private;
 595
 596         if (bio->bi_error) {
 597                 struct search *s = container_of(cl, struct search, cl);
 598                 s->iop.error = bio->bi_error;
 599                 /* Only cache read errors are recoverable */
 600                 s->recoverable = false;
 601         }
 602
 603         bio_put(bio);
 604         closure_put(cl);
 605 }
 606
 607 static void bio_complete(struct search *s)
 608 {
 609         if (s->orig_bio) {
 610                 generic_end_io_acct(bio_data_dir(s->orig_bio),
 611                                     &s->d->disk->part0, s->start_time);
 612
 613                 trace_bcache_request_end(s->d, s->orig_bio);
 614                 s->orig_bio->bi_error = s->iop.error;
 615                 bio_endio(s->orig_bio);
 616                 s->orig_bio = NULL;
 617         }
 618 }
 619
 620 static void do_bio_hook(struct search *s, struct bio *orig_bio)
 621 {
 622         struct bio *bio = &s->bio.bio;
 623
 624         bio_init(bio);
 625         __bio_clone_fast(bio, orig_bio);
 626         bio->bi_end_io          = request_endio;
 627         bio->bi_private         = &s->cl;
 628
 629         bio_cnt_set(bio, 3);
 630 }
 631
 632 static void search_free(struct closure *cl)
 633 {
 634         struct search *s = container_of(cl, struct search, cl);
 635         bio_complete(s);
 636
 637         if (s->iop.bio)
 638                 bio_put(s->iop.bio);
 639
 640         closure_debug_destroy(cl);
 641         mempool_free(s, s->d->c->search);
 642 }
 643
 644 static inline struct search *search_alloc(struct bio *bio,
 645                                           struct bcache_device *d)
 646 {
 647         struct search *s;
 648
 649         s = mempool_alloc(d->c->search, GFP_NOIO);
 650
 651         closure_init(&s->cl, NULL);
 652         do_bio_hook(s, bio);
 653
 654         s->orig_bio             = bio;
 655         s->cache_miss           = NULL;
 656         s->d                    = d;
 657         s->recoverable          = 1;
 658         s->write                = (bio->bi_rw & REQ_WRITE) != 0;
 659         s->read_dirty_data      = 0;
 660         s->start_time           = jiffies;
 661
 662         s->iop.c                = d->c;
 663         s->iop.bio              = NULL;
 664         s->iop.inode            = d->id;
 665         s->iop.write_point      = hash_long((unsigned long) current, 16);
 666         s->iop.write_prio       = 0;
 667         s->iop.error            = 0;
 668         s->iop.flags            = 0;
 669         s->iop.flush_journal    = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0;
 670         s->iop.wq               = bcache_wq;
 671
 672         return s;
 673 }
 674
 675 /* Cached devices */
 676
 677 static void cached_dev_bio_complete(struct closure *cl)
 678 {
 679         struct search *s = container_of(cl, struct search, cl);
 680         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 681
 682         search_free(cl);
 683         cached_dev_put(dc);
 684 }
 685
 686 /* Process reads */
 687
 688 static void cached_dev_cache_miss_done(struct closure *cl)
 689 {
 690         struct search *s = container_of(cl, struct search, cl);
 691
 692         if (s->iop.replace_collision)
 693                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 694
 695         if (s->iop.bio) {
 696                 int i;
 697                 struct bio_vec *bv;
 698
 699                 bio_for_each_segment_all(bv, s->iop.bio, i)
 700                         __free_page(bv->bv_page);
 701         }
 702
 703         cached_dev_bio_complete(cl);
 704 }
 705
 706 static void cached_dev_read_error(struct closure *cl)
 707 {
 708         struct search *s = container_of(cl, struct search, cl);
 709         struct bio *bio = &s->bio.bio;
 710
 711         if (s->recoverable) {
 712                 /* Retry from the backing device: */
 713                 trace_bcache_read_retry(s->orig_bio);
 714
 715                 s->iop.error = 0;
 716                 do_bio_hook(s, s->orig_bio);
 717
 718                 /* XXX: invalidate cache */
 719
 720                 closure_bio_submit(bio, cl);
 721         }
 722
 723         continue_at(cl, cached_dev_cache_miss_done, NULL);
 724 }
 725
 726 static void cached_dev_read_done(struct closure *cl)
 727 {
 728         struct search *s = container_of(cl, struct search, cl);
 729         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 730
 731         /*
 732          * We had a cache miss; cache_bio now contains data ready to be inserted
 733          * into the cache.
 734          *
 735          * First, we copy the data we just read from cache_bio's bounce buffers
 736          * to the buffers the original bio pointed to:
 737          */
 738
 739         if (s->iop.bio) {
 740                 bio_reset(s->iop.bio);
 741                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 742                 s->iop.bio->bi_bdev = s->cache_miss->bi_bdev;
 743                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 744                 bch_bio_map(s->iop.bio, NULL);
 745
 746                 bio_copy_data(s->cache_miss, s->iop.bio);
 747
 748                 bio_put(s->cache_miss);
 749                 s->cache_miss = NULL;
 750         }
 751
 752         if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
 753                 bch_data_verify(dc, s->orig_bio);
 754
 755         bio_complete(s);
 756
 757         if (s->iop.bio &&
 758             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 759                 BUG_ON(!s->iop.replace);
 760                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 761         }
 762
 763         continue_at(cl, cached_dev_cache_miss_done, NULL);
 764 }
 765
 766 static void cached_dev_read_done_bh(struct closure *cl)
 767 {
 768         struct search *s = container_of(cl, struct search, cl);
 769         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 770
 771         bch_mark_cache_accounting(s->iop.c, s->d,
 772                                   !s->cache_miss, s->iop.bypass);
 773         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 774
 775         if (s->iop.error)
 776                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 777         else if (s->iop.bio || verify(dc, &s->bio.bio))
 778                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 779         else
 780                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 781 }
 782
 783 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 784                                  struct bio *bio, unsigned sectors)
 785 {
 786         int ret = MAP_CONTINUE;
 787         unsigned reada = 0;
 788         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 789         struct bio *miss, *cache_bio;
 790
 791         if (s->cache_miss || s->iop.bypass) {
 792                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 793                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 794                 goto out_submit;
 795         }
 796
 797         if (!(bio->bi_rw & REQ_RAHEAD) &&
 798             !(bio->bi_rw & REQ_META) &&
 799             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 800                 reada = min_t(sector_t, dc->readahead >> 9,
 801                               bdev_sectors(bio->bi_bdev) - bio_end_sector(bio));
 802
 803         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 804
 805         s->iop.replace_key = KEY(s->iop.inode,
 806                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 807                                  s->insert_bio_sectors);
 808
 809         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 810         if (ret)
 811                 return ret;
 812
 813         s->iop.replace = true;
 814
 815         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 816
 817         /* btree_search_recurse()'s btree iterator is no good anymore */
 818         ret = miss == bio ? MAP_DONE : -EINTR;
 819
 820         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 821                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 822                         dc->disk.bio_split);
 823         if (!cache_bio)
 824                 goto out_submit;
 825
 826         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 827         cache_bio->bi_bdev              = miss->bi_bdev;
 828         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 829
 830         cache_bio->bi_end_io    = request_endio;
 831         cache_bio->bi_private   = &s->cl;
 832
 833         bch_bio_map(cache_bio, NULL);
 834         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 835                 goto out_put;
 836
 837         if (reada)
 838                 bch_mark_cache_readahead(s->iop.c, s->d);
 839
 840         s->cache_miss   = miss;
 841         s->iop.bio      = cache_bio;
 842         bio_get(cache_bio);
 843         closure_bio_submit(cache_bio, &s->cl);
 844
 845         return ret;
 846 out_put:
 847         bio_put(cache_bio);
 848 out_submit:
 849         miss->bi_end_io         = request_endio;
 850         miss->bi_private        = &s->cl;
 851         closure_bio_submit(miss, &s->cl);
 852         return ret;
 853 }
 854
 855 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 856 {
 857         struct closure *cl = &s->cl;
 858
 859         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 860         continue_at(cl, cached_dev_read_done_bh, NULL);
 861 }
 862
 863 /* Process writes */
 864
 865 static void cached_dev_write_complete(struct closure *cl)
 866 {
 867         struct search *s = container_of(cl, struct search, cl);
 868         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 869
 870         up_read_non_owner(&dc->writeback_lock);
 871         cached_dev_bio_complete(cl);
 872 }
 873
 874 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 875 {
 876         struct closure *cl = &s->cl;
 877         struct bio *bio = &s->bio.bio;
 878         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 879         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 880
 881         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 882
 883         down_read_non_owner(&dc->writeback_lock);
 884         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 885                 /*
 886                  * We overlap with some dirty data undergoing background
 887                  * writeback, force this write to writeback
 888                  */
 889                 s->iop.bypass = false;
 890                 s->iop.writeback = true;
 891         }
 892
 893         /*
 894          * Discards aren't _required_ to do anything, so skipping if
 895          * check_overlapping returned true is ok
 896          *
 897          * But check_overlapping drops dirty keys for which io hasn't started,
 898          * so we still want to call it.
 899          */
 900         if (bio->bi_rw & REQ_DISCARD)
 901                 s->iop.bypass = true;
 902
 903         if (should_writeback(dc, s->orig_bio,
 904                              cache_mode(dc, bio),
 905                              s->iop.bypass)) {
 906                 s->iop.bypass = false;
 907                 s->iop.writeback = true;
 908         }
 909
 910         if (s->iop.bypass) {
 911                 s->iop.bio = s->orig_bio;
 912                 bio_get(s->iop.bio);
 913
 914                 if (!(bio->bi_rw & REQ_DISCARD) ||
 915                     blk_queue_discard(bdev_get_queue(dc->bdev)))
 916                         closure_bio_submit(bio, cl);
 917         } else if (s->iop.writeback) {
 918                 bch_writeback_add(dc);
 919                 s->iop.bio = bio;
 920
 921                 if (bio->bi_rw & REQ_FLUSH) {
 922                         /* Also need to send a flush to the backing device */
 923                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
 924                                                              dc->disk.bio_split);
 925
 926                         flush->bi_rw    = WRITE_FLUSH;
 927                         flush->bi_bdev  = bio->bi_bdev;
 928                         flush->bi_end_io = request_endio;
 929                         flush->bi_private = cl;
 930
 931                         closure_bio_submit(flush, cl);
 932                 }
 933         } else {
 934                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
 935
 936                 closure_bio_submit(bio, cl);
 937         }
 938
 939         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 940         continue_at(cl, cached_dev_write_complete, NULL);
 941 }
 942
 943 static void cached_dev_nodata(struct closure *cl)
 944 {
 945         struct search *s = container_of(cl, struct search, cl);
 946         struct bio *bio = &s->bio.bio;
 947
 948         if (s->iop.flush_journal)
 949                 bch_journal_meta(s->iop.c, cl);
 950
 951         /* If it's a flush, we send the flush to the backing device too */
 952         closure_bio_submit(bio, cl);
 953
 954         continue_at(cl, cached_dev_bio_complete, NULL);
 955 }
 956
 957 /* Cached devices - read & write stuff */
 958
 959 static blk_qc_t cached_dev_make_request(struct request_queue *q,
 960                                         struct bio *bio)
 961 {
 962         struct search *s;
 963         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
 964         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
 965         int rw = bio_data_dir(bio);
 966
 967         generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
 968
 969         bio->bi_bdev = dc->bdev;
 970         bio->bi_iter.bi_sector += dc->sb.data_offset;
 971
 972         if (cached_dev_get(dc)) {
 973                 s = search_alloc(bio, d);
 974                 trace_bcache_request_start(s->d, bio);
 975
 976                 if (!bio->bi_iter.bi_size) {
 977                         /*
 978                          * can't call bch_journal_meta from under
 979                          * generic_make_request
 980                          */
 981                         continue_at_nobarrier(&s->cl,
 982                                               cached_dev_nodata,
 983                                               bcache_wq);
 984                 } else {
 985                         s->iop.bypass = check_should_bypass(dc, bio);
 986
 987                         if (rw)
 988                                 cached_dev_write(dc, s);
 989                         else
 990                                 cached_dev_read(dc, s);
 991                 }
 992         } else {
 993                 if ((bio->bi_rw & REQ_DISCARD) &&
 994                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
 995                         bio_endio(bio);
 996                 else
 997                         generic_make_request(bio);
 998         }
 999
1000         return BLK_QC_T_NONE;
1001 }
1002
1003 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1004                             unsigned int cmd, unsigned long arg)
1005 {
1006         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1007         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1008 }
1009
1010 static int cached_dev_congested(void *data, int bits)
1011 {
1012         struct bcache_device *d = data;
1013         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1014         struct request_queue *q = bdev_get_queue(dc->bdev);
1015         int ret = 0;
1016
1017         if (bdi_congested(&q->backing_dev_info, bits))
1018                 return 1;
1019
1020         if (cached_dev_get(dc)) {
1021                 unsigned i;
1022                 struct cache *ca;
1023
1024                 for_each_cache(ca, d->c, i) {
1025                         q = bdev_get_queue(ca->bdev);
1026                         ret |= bdi_congested(&q->backing_dev_info, bits);
1027                 }
1028
1029                 cached_dev_put(dc);
1030         }
1031
1032         return ret;
1033 }
1034
1035 void bch_cached_dev_request_init(struct cached_dev *dc)
1036 {
1037         struct gendisk *g = dc->disk.disk;
1038
1039         g->queue->make_request_fn               = cached_dev_make_request;
1040         g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1041         dc->disk.cache_miss                     = cached_dev_cache_miss;
1042         dc->disk.ioctl                          = cached_dev_ioctl;
1043 }
1044
1045 /* Flash backed devices */
1046
1047 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1048                                 struct bio *bio, unsigned sectors)
1049 {
1050         unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1051
1052         swap(bio->bi_iter.bi_size, bytes);
1053         zero_fill_bio(bio);
1054         swap(bio->bi_iter.bi_size, bytes);
1055
1056         bio_advance(bio, bytes);
1057
1058         if (!bio->bi_iter.bi_size)
1059                 return MAP_DONE;
1060
1061         return MAP_CONTINUE;
1062 }
1063
1064 static void flash_dev_nodata(struct closure *cl)
1065 {
1066         struct search *s = container_of(cl, struct search, cl);
1067
1068         if (s->iop.flush_journal)
1069                 bch_journal_meta(s->iop.c, cl);
1070
1071         continue_at(cl, search_free, NULL);
1072 }
1073
1074 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1075                                              struct bio *bio)
1076 {
1077         struct search *s;
1078         struct closure *cl;
1079         struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1080         int rw = bio_data_dir(bio);
1081
1082         generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
1083
1084         s = search_alloc(bio, d);
1085         cl = &s->cl;
1086         bio = &s->bio.bio;
1087
1088         trace_bcache_request_start(s->d, bio);
1089
1090         if (!bio->bi_iter.bi_size) {
1091                 /*
1092                  * can't call bch_journal_meta from under
1093                  * generic_make_request
1094                  */
1095                 continue_at_nobarrier(&s->cl,
1096                                       flash_dev_nodata,
1097                                       bcache_wq);
1098                 return BLK_QC_T_NONE;
1099         } else if (rw) {
1100                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1101                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1102                                         &KEY(d->id, bio_end_sector(bio), 0));
1103
1104                 s->iop.bypass           = (bio->bi_rw & REQ_DISCARD) != 0;
1105                 s->iop.writeback        = true;
1106                 s->iop.bio              = bio;
1107
1108                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1109         } else {
1110                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1111         }
1112
1113         continue_at(cl, search_free, NULL);
1114         return BLK_QC_T_NONE;
1115 }
1116
1117 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1118                            unsigned int cmd, unsigned long arg)
1119 {
1120         return -ENOTTY;
1121 }
1122
1123 static int flash_dev_congested(void *data, int bits)
1124 {
1125         struct bcache_device *d = data;
1126         struct request_queue *q;
1127         struct cache *ca;
1128         unsigned i;
1129         int ret = 0;
1130
1131         for_each_cache(ca, d->c, i) {
1132                 q = bdev_get_queue(ca->bdev);
1133                 ret |= bdi_congested(&q->backing_dev_info, bits);
1134         }
1135
1136         return ret;
1137 }
1138
1139 void bch_flash_dev_request_init(struct bcache_device *d)
1140 {
1141         struct gendisk *g = d->disk;
1142
1143         g->queue->make_request_fn               = flash_dev_make_request;
1144         g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1145         d->cache_miss                           = flash_dev_cache_miss;
1146         d->ioctl                                = flash_dev_ioctl;
1147 }
1148
1149 void bch_request_exit(void)
1150 {
1151         if (bch_search_cache)
1152                 kmem_cache_destroy(bch_search_cache);
1153 }
1154
1155 int __init bch_request_init(void)
1156 {
1157         bch_search_cache = KMEM_CACHE(search, 0);
1158         if (!bch_search_cache)
1159                 return -ENOMEM;
1160
1161         return 0;
1162 }