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