drivers/md/bcache/request.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Main bcache entry point - handle a read or a write request and decide what to
   4  * do with it; the make_request functions are called by the block layer.
   5  *
   6  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   7  * Copyright 2012 Google, Inc.
   8  */
   9
  10 #include "bcache.h"
  11 #include "btree.h"
  12 #include "debug.h"
  13 #include "request.h"
  14 #include "writeback.h"
  15
  16 #include <linux/module.h>
  17 #include <linux/hash.h>
  18 #include <linux/random.h>
  19 #include <linux/backing-dev.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 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
  31 {
  32         return BDEV_CACHE_MODE(&dc->sb);
  33 }
  34
  35 static bool verify(struct cached_dev *dc, struct bio *bio)
  36 {
  37         return dc->verify;
  38 }
  39
  40 static void bio_csum(struct bio *bio, struct bkey *k)
  41 {
  42         struct bio_vec bv;
  43         struct bvec_iter iter;
  44         uint64_t csum = 0;
  45
  46         bio_for_each_segment(bv, bio, iter) {
  47                 void *d = kmap(bv.bv_page) + bv.bv_offset;
  48                 csum = bch_crc64_update(csum, d, bv.bv_len);
  49                 kunmap(bv.bv_page);
  50         }
  51
  52         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
  53 }
  54
  55 /* Insert data into cache */
  56
  57 static void bch_data_insert_keys(struct closure *cl)
  58 {
  59         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
  60         atomic_t *journal_ref = NULL;
  61         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
  62         int ret;
  63
  64         /*
  65          * If we're looping, might already be waiting on
  66          * another journal write - can't wait on more than one journal write at
  67          * a time
  68          *
  69          * XXX: this looks wrong
  70          */
  71 #if 0
  72         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
  73                 closure_sync(&s->cl);
  74 #endif
  75
  76         if (!op->replace)
  77                 journal_ref = bch_journal(op->c, &op->insert_keys,
  78                                           op->flush_journal ? cl : NULL);
  79
  80         ret = bch_btree_insert(op->c, &op->insert_keys,
  81                                journal_ref, replace_key);
  82         if (ret == -ESRCH) {
  83                 op->replace_collision = true;
  84         } else if (ret) {
  85                 op->status              = BLK_STS_RESOURCE;
  86                 op->insert_data_done    = true;
  87         }
  88
  89         if (journal_ref)
  90                 atomic_dec_bug(journal_ref);
  91
  92         if (!op->insert_data_done) {
  93                 continue_at(cl, bch_data_insert_start, op->wq);
  94                 return;
  95         }
  96
  97         bch_keylist_free(&op->insert_keys);
  98         closure_return(cl);
  99 }
 100
 101 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
 102                                struct cache_set *c)
 103 {
 104         size_t oldsize = bch_keylist_nkeys(l);
 105         size_t newsize = oldsize + u64s;
 106
 107         /*
 108          * The journalling code doesn't handle the case where the keys to insert
 109          * is bigger than an empty write: If we just return -ENOMEM here,
 110          * bio_insert() and bio_invalidate() will insert the keys created so far
 111          * and finish the rest when the keylist is empty.
 112          */
 113         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
 114                 return -ENOMEM;
 115
 116         return __bch_keylist_realloc(l, u64s);
 117 }
 118
 119 static void bch_data_invalidate(struct closure *cl)
 120 {
 121         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 122         struct bio *bio = op->bio;
 123
 124         pr_debug("invalidating %i sectors from %llu",
 125                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
 126
 127         while (bio_sectors(bio)) {
 128                 unsigned sectors = min(bio_sectors(bio),
 129                                        1U << (KEY_SIZE_BITS - 1));
 130
 131                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
 132                         goto out;
 133
 134                 bio->bi_iter.bi_sector  += sectors;
 135                 bio->bi_iter.bi_size    -= sectors << 9;
 136
 137                 bch_keylist_add(&op->insert_keys,
 138                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
 139         }
 140
 141         op->insert_data_done = true;
 142         bio_put(bio);
 143 out:
 144         continue_at(cl, bch_data_insert_keys, op->wq);
 145 }
 146
 147 static void bch_data_insert_error(struct closure *cl)
 148 {
 149         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 150
 151         /*
 152          * Our data write just errored, which means we've got a bunch of keys to
 153          * insert that point to data that wasn't succesfully written.
 154          *
 155          * We don't have to insert those keys but we still have to invalidate
 156          * that region of the cache - so, if we just strip off all the pointers
 157          * from the keys we'll accomplish just that.
 158          */
 159
 160         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
 161
 162         while (src != op->insert_keys.top) {
 163                 struct bkey *n = bkey_next(src);
 164
 165                 SET_KEY_PTRS(src, 0);
 166                 memmove(dst, src, bkey_bytes(src));
 167
 168                 dst = bkey_next(dst);
 169                 src = n;
 170         }
 171
 172         op->insert_keys.top = dst;
 173
 174         bch_data_insert_keys(cl);
 175 }
 176
 177 static void bch_data_insert_endio(struct bio *bio)
 178 {
 179         struct closure *cl = bio->bi_private;
 180         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 181
 182         if (bio->bi_status) {
 183                 /* TODO: We could try to recover from this. */
 184                 if (op->writeback)
 185                         op->status = bio->bi_status;
 186                 else if (!op->replace)
 187                         set_closure_fn(cl, bch_data_insert_error, op->wq);
 188                 else
 189                         set_closure_fn(cl, NULL, NULL);
 190         }
 191
 192         bch_bbio_endio(op->c, bio, bio->bi_status, "writing data to cache");
 193 }
 194
 195 static void bch_data_insert_start(struct closure *cl)
 196 {
 197         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 198         struct bio *bio = op->bio, *n;
 199
 200         if (op->bypass)
 201                 return bch_data_invalidate(cl);
 202
 203         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0)
 204                 wake_up_gc(op->c);
 205
 206         /*
 207          * Journal writes are marked REQ_PREFLUSH; if the original write was a
 208          * flush, it'll wait on the journal write.
 209          */
 210         bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
 211
 212         do {
 213                 unsigned i;
 214                 struct bkey *k;
 215                 struct bio_set *split = op->c->bio_split;
 216
 217                 /* 1 for the device pointer and 1 for the chksum */
 218                 if (bch_keylist_realloc(&op->insert_keys,
 219                                         3 + (op->csum ? 1 : 0),
 220                                         op->c)) {
 221                         continue_at(cl, bch_data_insert_keys, op->wq);
 222                         return;
 223                 }
 224
 225                 k = op->insert_keys.top;
 226                 bkey_init(k);
 227                 SET_KEY_INODE(k, op->inode);
 228                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
 229
 230                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
 231                                        op->write_point, op->write_prio,
 232                                        op->writeback))
 233                         goto err;
 234
 235                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
 236
 237                 n->bi_end_io    = bch_data_insert_endio;
 238                 n->bi_private   = cl;
 239
 240                 if (op->writeback) {
 241                         SET_KEY_DIRTY(k, true);
 242
 243                         for (i = 0; i < KEY_PTRS(k); i++)
 244                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
 245                                             GC_MARK_DIRTY);
 246                 }
 247
 248                 SET_KEY_CSUM(k, op->csum);
 249                 if (KEY_CSUM(k))
 250                         bio_csum(n, k);
 251
 252                 trace_bcache_cache_insert(k);
 253                 bch_keylist_push(&op->insert_keys);
 254
 255                 bio_set_op_attrs(n, REQ_OP_WRITE, 0);
 256                 bch_submit_bbio(n, op->c, k, 0);
 257         } while (n != bio);
 258
 259         op->insert_data_done = true;
 260         continue_at(cl, bch_data_insert_keys, op->wq);
 261         return;
 262 err:
 263         /* bch_alloc_sectors() blocks if s->writeback = true */
 264         BUG_ON(op->writeback);
 265
 266         /*
 267          * But if it's not a writeback write we'd rather just bail out if
 268          * there aren't any buckets ready to write to - it might take awhile and
 269          * we might be starving btree writes for gc or something.
 270          */
 271
 272         if (!op->replace) {
 273                 /*
 274                  * Writethrough write: We can't complete the write until we've
 275                  * updated the index. But we don't want to delay the write while
 276                  * we wait for buckets to be freed up, so just invalidate the
 277                  * rest of the write.
 278                  */
 279                 op->bypass = true;
 280                 return bch_data_invalidate(cl);
 281         } else {
 282                 /*
 283                  * From a cache miss, we can just insert the keys for the data
 284                  * we have written or bail out if we didn't do anything.
 285                  */
 286                 op->insert_data_done = true;
 287                 bio_put(bio);
 288
 289                 if (!bch_keylist_empty(&op->insert_keys))
 290                         continue_at(cl, bch_data_insert_keys, op->wq);
 291                 else
 292                         closure_return(cl);
 293         }
 294 }
 295
 296 /**
 297  * bch_data_insert - stick some data in the cache
 298  *
 299  * This is the starting point for any data to end up in a cache device; it could
 300  * be from a normal write, or a writeback write, or a write to a flash only
 301  * volume - it's also used by the moving garbage collector to compact data in
 302  * mostly empty buckets.
 303  *
 304  * It first writes the data to the cache, creating a list of keys to be inserted
 305  * (if the data had to be fragmented there will be multiple keys); after the
 306  * data is written it calls bch_journal, and after the keys have been added to
 307  * the next journal write they're inserted into the btree.
 308  *
 309  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
 310  * and op->inode is used for the key inode.
 311  *
 312  * If s->bypass is true, instead of inserting the data it invalidates the
 313  * region of the cache represented by s->cache_bio and op->inode.
 314  */
 315 void bch_data_insert(struct closure *cl)
 316 {
 317         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 318
 319         trace_bcache_write(op->c, op->inode, op->bio,
 320                            op->writeback, op->bypass);
 321
 322         bch_keylist_init(&op->insert_keys);
 323         bio_get(op->bio);
 324         bch_data_insert_start(cl);
 325 }
 326
 327 /* Congested? */
 328
 329 unsigned bch_get_congested(struct cache_set *c)
 330 {
 331         int i;
 332         long rand;
 333
 334         if (!c->congested_read_threshold_us &&
 335             !c->congested_write_threshold_us)
 336                 return 0;
 337
 338         i = (local_clock_us() - c->congested_last_us) / 1024;
 339         if (i < 0)
 340                 return 0;
 341
 342         i += atomic_read(&c->congested);
 343         if (i >= 0)
 344                 return 0;
 345
 346         i += CONGESTED_MAX;
 347
 348         if (i > 0)
 349                 i = fract_exp_two(i, 6);
 350
 351         rand = get_random_int();
 352         i -= bitmap_weight(&rand, BITS_PER_LONG);
 353
 354         return i > 0 ? i : 1;
 355 }
 356
 357 static void add_sequential(struct task_struct *t)
 358 {
 359         ewma_add(t->sequential_io_avg,
 360                  t->sequential_io, 8, 0);
 361
 362         t->sequential_io = 0;
 363 }
 364
 365 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
 366 {
 367         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
 368 }
 369
 370 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 371 {
 372         struct cache_set *c = dc->disk.c;
 373         unsigned mode = cache_mode(dc, bio);
 374         unsigned sectors, congested = bch_get_congested(c);
 375         struct task_struct *task = current;
 376         struct io *i;
 377
 378         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 379             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
 380             (bio_op(bio) == REQ_OP_DISCARD))
 381                 goto skip;
 382
 383         if (mode == CACHE_MODE_NONE ||
 384             (mode == CACHE_MODE_WRITEAROUND &&
 385              op_is_write(bio_op(bio))))
 386                 goto skip;
 387
 388         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 389             bio_sectors(bio) & (c->sb.block_size - 1)) {
 390                 pr_debug("skipping unaligned io");
 391                 goto skip;
 392         }
 393
 394         if (bypass_torture_test(dc)) {
 395                 if ((get_random_int() & 3) == 3)
 396                         goto skip;
 397                 else
 398                         goto rescale;
 399         }
 400
 401         if (!congested && !dc->sequential_cutoff)
 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)
 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 (bio->bi_status)
 487                 s->iop.status = bio->bi_status;
 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.status = BLK_STS_IOERR;
 492         }
 493
 494         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "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                 return;
 580         }
 581
 582         closure_return(cl);
 583 }
 584
 585 /* Common code for the make_request functions */
 586
 587 static void request_endio(struct bio *bio)
 588 {
 589         struct closure *cl = bio->bi_private;
 590
 591         if (bio->bi_status) {
 592                 struct search *s = container_of(cl, struct search, cl);
 593                 s->iop.status = bio->bi_status;
 594                 /* Only cache read errors are recoverable */
 595                 s->recoverable = false;
 596         }
 597
 598         bio_put(bio);
 599         closure_put(cl);
 600 }
 601
 602 static void bio_complete(struct search *s)
 603 {
 604         if (s->orig_bio) {
 605                 struct request_queue *q = s->orig_bio->bi_disk->queue;
 606                 generic_end_io_acct(q, bio_data_dir(s->orig_bio),
 607                                     &s->d->disk->part0, s->start_time);
 608
 609                 trace_bcache_request_end(s->d, s->orig_bio);
 610                 s->orig_bio->bi_status = s->iop.status;
 611                 bio_endio(s->orig_bio);
 612                 s->orig_bio = NULL;
 613         }
 614 }
 615
 616 static void do_bio_hook(struct search *s, struct bio *orig_bio)
 617 {
 618         struct bio *bio = &s->bio.bio;
 619
 620         bio_init(bio, NULL, 0);
 621         __bio_clone_fast(bio, orig_bio);
 622         bio->bi_end_io          = request_endio;
 623         bio->bi_private         = &s->cl;
 624
 625         bio_cnt_set(bio, 3);
 626 }
 627
 628 static void search_free(struct closure *cl)
 629 {
 630         struct search *s = container_of(cl, struct search, cl);
 631         bio_complete(s);
 632
 633         if (s->iop.bio)
 634                 bio_put(s->iop.bio);
 635
 636         closure_debug_destroy(cl);
 637         mempool_free(s, s->d->c->search);
 638 }
 639
 640 static inline struct search *search_alloc(struct bio *bio,
 641                                           struct bcache_device *d)
 642 {
 643         struct search *s;
 644
 645         s = mempool_alloc(d->c->search, GFP_NOIO);
 646
 647         closure_init(&s->cl, NULL);
 648         do_bio_hook(s, bio);
 649
 650         s->orig_bio             = bio;
 651         s->cache_miss           = NULL;
 652         s->d                    = d;
 653         s->recoverable          = 1;
 654         s->write                = op_is_write(bio_op(bio));
 655         s->read_dirty_data      = 0;
 656         s->start_time           = jiffies;
 657
 658         s->iop.c                = d->c;
 659         s->iop.bio              = NULL;
 660         s->iop.inode            = d->id;
 661         s->iop.write_point      = hash_long((unsigned long) current, 16);
 662         s->iop.write_prio       = 0;
 663         s->iop.status           = 0;
 664         s->iop.flags            = 0;
 665         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
 666         s->iop.wq               = bcache_wq;
 667
 668         return s;
 669 }
 670
 671 /* Cached devices */
 672
 673 static void cached_dev_bio_complete(struct closure *cl)
 674 {
 675         struct search *s = container_of(cl, struct search, cl);
 676         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 677
 678         search_free(cl);
 679         cached_dev_put(dc);
 680 }
 681
 682 /* Process reads */
 683
 684 static void cached_dev_cache_miss_done(struct closure *cl)
 685 {
 686         struct search *s = container_of(cl, struct search, cl);
 687
 688         if (s->iop.replace_collision)
 689                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 690
 691         if (s->iop.bio)
 692                 bio_free_pages(s->iop.bio);
 693
 694         cached_dev_bio_complete(cl);
 695 }
 696
 697 static void cached_dev_read_error(struct closure *cl)
 698 {
 699         struct search *s = container_of(cl, struct search, cl);
 700         struct bio *bio = &s->bio.bio;
 701
 702         if (s->recoverable) {
 703                 /* Retry from the backing device: */
 704                 trace_bcache_read_retry(s->orig_bio);
 705
 706                 s->iop.status = 0;
 707                 do_bio_hook(s, s->orig_bio);
 708
 709                 /* XXX: invalidate cache */
 710
 711                 closure_bio_submit(bio, cl);
 712         }
 713
 714         continue_at(cl, cached_dev_cache_miss_done, NULL);
 715 }
 716
 717 static void cached_dev_read_done(struct closure *cl)
 718 {
 719         struct search *s = container_of(cl, struct search, cl);
 720         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 721
 722         /*
 723          * We had a cache miss; cache_bio now contains data ready to be inserted
 724          * into the cache.
 725          *
 726          * First, we copy the data we just read from cache_bio's bounce buffers
 727          * to the buffers the original bio pointed to:
 728          */
 729
 730         if (s->iop.bio) {
 731                 bio_reset(s->iop.bio);
 732                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 733                 bio_copy_dev(s->iop.bio, s->cache_miss);
 734                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 735                 bch_bio_map(s->iop.bio, NULL);
 736
 737                 bio_copy_data(s->cache_miss, s->iop.bio);
 738
 739                 bio_put(s->cache_miss);
 740                 s->cache_miss = NULL;
 741         }
 742
 743         if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
 744                 bch_data_verify(dc, s->orig_bio);
 745
 746         bio_complete(s);
 747
 748         if (s->iop.bio &&
 749             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 750                 BUG_ON(!s->iop.replace);
 751                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 752         }
 753
 754         continue_at(cl, cached_dev_cache_miss_done, NULL);
 755 }
 756
 757 static void cached_dev_read_done_bh(struct closure *cl)
 758 {
 759         struct search *s = container_of(cl, struct search, cl);
 760         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 761
 762         bch_mark_cache_accounting(s->iop.c, s->d,
 763                                   !s->cache_miss, s->iop.bypass);
 764         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
 765
 766         if (s->iop.status)
 767                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 768         else if (s->iop.bio || verify(dc, &s->bio.bio))
 769                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 770         else
 771                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 772 }
 773
 774 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 775                                  struct bio *bio, unsigned sectors)
 776 {
 777         int ret = MAP_CONTINUE;
 778         unsigned reada = 0;
 779         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 780         struct bio *miss, *cache_bio;
 781
 782         if (s->cache_miss || s->iop.bypass) {
 783                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 784                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 785                 goto out_submit;
 786         }
 787
 788         if (!(bio->bi_opf & REQ_RAHEAD) &&
 789             !(bio->bi_opf & REQ_META) &&
 790             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 791                 reada = min_t(sector_t, dc->readahead >> 9,
 792                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
 793
 794         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 795
 796         s->iop.replace_key = KEY(s->iop.inode,
 797                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 798                                  s->insert_bio_sectors);
 799
 800         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 801         if (ret)
 802                 return ret;
 803
 804         s->iop.replace = true;
 805
 806         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 807
 808         /* btree_search_recurse()'s btree iterator is no good anymore */
 809         ret = miss == bio ? MAP_DONE : -EINTR;
 810
 811         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 812                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 813                         dc->disk.bio_split);
 814         if (!cache_bio)
 815                 goto out_submit;
 816
 817         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 818         bio_copy_dev(cache_bio, miss);
 819         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 820
 821         cache_bio->bi_end_io    = request_endio;
 822         cache_bio->bi_private   = &s->cl;
 823
 824         bch_bio_map(cache_bio, NULL);
 825         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 826                 goto out_put;
 827
 828         if (reada)
 829                 bch_mark_cache_readahead(s->iop.c, s->d);
 830
 831         s->cache_miss   = miss;
 832         s->iop.bio      = cache_bio;
 833         bio_get(cache_bio);
 834         closure_bio_submit(cache_bio, &s->cl);
 835
 836         return ret;
 837 out_put:
 838         bio_put(cache_bio);
 839 out_submit:
 840         miss->bi_end_io         = request_endio;
 841         miss->bi_private        = &s->cl;
 842         closure_bio_submit(miss, &s->cl);
 843         return ret;
 844 }
 845
 846 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 847 {
 848         struct closure *cl = &s->cl;
 849
 850         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 851         continue_at(cl, cached_dev_read_done_bh, NULL);
 852 }
 853
 854 /* Process writes */
 855
 856 static void cached_dev_write_complete(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         up_read_non_owner(&dc->writeback_lock);
 862         cached_dev_bio_complete(cl);
 863 }
 864
 865 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 866 {
 867         struct closure *cl = &s->cl;
 868         struct bio *bio = &s->bio.bio;
 869         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 870         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 871
 872         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 873
 874         down_read_non_owner(&dc->writeback_lock);
 875         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 876                 /*
 877                  * We overlap with some dirty data undergoing background
 878                  * writeback, force this write to writeback
 879                  */
 880                 s->iop.bypass = false;
 881                 s->iop.writeback = true;
 882         }
 883
 884         /*
 885          * Discards aren't _required_ to do anything, so skipping if
 886          * check_overlapping returned true is ok
 887          *
 888          * But check_overlapping drops dirty keys for which io hasn't started,
 889          * so we still want to call it.
 890          */
 891         if (bio_op(bio) == REQ_OP_DISCARD)
 892                 s->iop.bypass = true;
 893
 894         if (should_writeback(dc, s->orig_bio,
 895                              cache_mode(dc, bio),
 896                              s->iop.bypass)) {
 897                 s->iop.bypass = false;
 898                 s->iop.writeback = true;
 899         }
 900
 901         if (s->iop.bypass) {
 902                 s->iop.bio = s->orig_bio;
 903                 bio_get(s->iop.bio);
 904
 905                 if ((bio_op(bio) != REQ_OP_DISCARD) ||
 906                     blk_queue_discard(bdev_get_queue(dc->bdev)))
 907                         closure_bio_submit(bio, cl);
 908         } else if (s->iop.writeback) {
 909                 bch_writeback_add(dc);
 910                 s->iop.bio = bio;
 911
 912                 if (bio->bi_opf & REQ_PREFLUSH) {
 913                         /* Also need to send a flush to the backing device */
 914                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
 915                                                              dc->disk.bio_split);
 916
 917                         bio_copy_dev(flush, bio);
 918                         flush->bi_end_io = request_endio;
 919                         flush->bi_private = cl;
 920                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 921
 922                         closure_bio_submit(flush, cl);
 923                 }
 924         } else {
 925                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
 926
 927                 closure_bio_submit(bio, cl);
 928         }
 929
 930         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 931         continue_at(cl, cached_dev_write_complete, NULL);
 932 }
 933
 934 static void cached_dev_nodata(struct closure *cl)
 935 {
 936         struct search *s = container_of(cl, struct search, cl);
 937         struct bio *bio = &s->bio.bio;
 938
 939         if (s->iop.flush_journal)
 940                 bch_journal_meta(s->iop.c, cl);
 941
 942         /* If it's a flush, we send the flush to the backing device too */
 943         closure_bio_submit(bio, cl);
 944
 945         continue_at(cl, cached_dev_bio_complete, NULL);
 946 }
 947
 948 /* Cached devices - read & write stuff */
 949
 950 static blk_qc_t cached_dev_make_request(struct request_queue *q,
 951                                         struct bio *bio)
 952 {
 953         struct search *s;
 954         struct bcache_device *d = bio->bi_disk->private_data;
 955         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
 956         int rw = bio_data_dir(bio);
 957
 958         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
 959
 960         bio_set_dev(bio, dc->bdev);
 961         bio->bi_iter.bi_sector += dc->sb.data_offset;
 962
 963         if (cached_dev_get(dc)) {
 964                 s = search_alloc(bio, d);
 965                 trace_bcache_request_start(s->d, bio);
 966
 967                 if (!bio->bi_iter.bi_size) {
 968                         /*
 969                          * can't call bch_journal_meta from under
 970                          * generic_make_request
 971                          */
 972                         continue_at_nobarrier(&s->cl,
 973                                               cached_dev_nodata,
 974                                               bcache_wq);
 975                 } else {
 976                         s->iop.bypass = check_should_bypass(dc, bio);
 977
 978                         if (rw)
 979                                 cached_dev_write(dc, s);
 980                         else
 981                                 cached_dev_read(dc, s);
 982                 }
 983         } else {
 984                 if ((bio_op(bio) == REQ_OP_DISCARD) &&
 985                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
 986                         bio_endio(bio);
 987                 else
 988                         generic_make_request(bio);
 989         }
 990
 991         return BLK_QC_T_NONE;
 992 }
 993
 994 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
 995                             unsigned int cmd, unsigned long arg)
 996 {
 997         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
 998         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
 999 }
1000
1001 static int cached_dev_congested(void *data, int bits)
1002 {
1003         struct bcache_device *d = data;
1004         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1005         struct request_queue *q = bdev_get_queue(dc->bdev);
1006         int ret = 0;
1007
1008         if (bdi_congested(q->backing_dev_info, bits))
1009                 return 1;
1010
1011         if (cached_dev_get(dc)) {
1012                 unsigned i;
1013                 struct cache *ca;
1014
1015                 for_each_cache(ca, d->c, i) {
1016                         q = bdev_get_queue(ca->bdev);
1017                         ret |= bdi_congested(q->backing_dev_info, bits);
1018                 }
1019
1020                 cached_dev_put(dc);
1021         }
1022
1023         return ret;
1024 }
1025
1026 void bch_cached_dev_request_init(struct cached_dev *dc)
1027 {
1028         struct gendisk *g = dc->disk.disk;
1029
1030         g->queue->make_request_fn               = cached_dev_make_request;
1031         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1032         dc->disk.cache_miss                     = cached_dev_cache_miss;
1033         dc->disk.ioctl                          = cached_dev_ioctl;
1034 }
1035
1036 /* Flash backed devices */
1037
1038 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1039                                 struct bio *bio, unsigned sectors)
1040 {
1041         unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1042
1043         swap(bio->bi_iter.bi_size, bytes);
1044         zero_fill_bio(bio);
1045         swap(bio->bi_iter.bi_size, bytes);
1046
1047         bio_advance(bio, bytes);
1048
1049         if (!bio->bi_iter.bi_size)
1050                 return MAP_DONE;
1051
1052         return MAP_CONTINUE;
1053 }
1054
1055 static void flash_dev_nodata(struct closure *cl)
1056 {
1057         struct search *s = container_of(cl, struct search, cl);
1058
1059         if (s->iop.flush_journal)
1060                 bch_journal_meta(s->iop.c, cl);
1061
1062         continue_at(cl, search_free, NULL);
1063 }
1064
1065 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1066                                              struct bio *bio)
1067 {
1068         struct search *s;
1069         struct closure *cl;
1070         struct bcache_device *d = bio->bi_disk->private_data;
1071         int rw = bio_data_dir(bio);
1072
1073         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1074
1075         s = search_alloc(bio, d);
1076         cl = &s->cl;
1077         bio = &s->bio.bio;
1078
1079         trace_bcache_request_start(s->d, bio);
1080
1081         if (!bio->bi_iter.bi_size) {
1082                 /*
1083                  * can't call bch_journal_meta from under
1084                  * generic_make_request
1085                  */
1086                 continue_at_nobarrier(&s->cl,
1087                                       flash_dev_nodata,
1088                                       bcache_wq);
1089                 return BLK_QC_T_NONE;
1090         } else if (rw) {
1091                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1092                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1093                                         &KEY(d->id, bio_end_sector(bio), 0));
1094
1095                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1096                 s->iop.writeback        = true;
1097                 s->iop.bio              = bio;
1098
1099                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1100         } else {
1101                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1102         }
1103
1104         continue_at(cl, search_free, NULL);
1105         return BLK_QC_T_NONE;
1106 }
1107
1108 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1109                            unsigned int cmd, unsigned long arg)
1110 {
1111         return -ENOTTY;
1112 }
1113
1114 static int flash_dev_congested(void *data, int bits)
1115 {
1116         struct bcache_device *d = data;
1117         struct request_queue *q;
1118         struct cache *ca;
1119         unsigned i;
1120         int ret = 0;
1121
1122         for_each_cache(ca, d->c, i) {
1123                 q = bdev_get_queue(ca->bdev);
1124                 ret |= bdi_congested(q->backing_dev_info, bits);
1125         }
1126
1127         return ret;
1128 }
1129
1130 void bch_flash_dev_request_init(struct bcache_device *d)
1131 {
1132         struct gendisk *g = d->disk;
1133
1134         g->queue->make_request_fn               = flash_dev_make_request;
1135         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1136         d->cache_miss                           = flash_dev_cache_miss;
1137         d->ioctl                                = flash_dev_ioctl;
1138 }
1139
1140 void bch_request_exit(void)
1141 {
1142         if (bch_search_cache)
1143                 kmem_cache_destroy(bch_search_cache);
1144 }
1145
1146 int __init bch_request_init(void)
1147 {
1148         bch_search_cache = KMEM_CACHE(search, 0);
1149         if (!bch_search_cache)
1150                 return -ENOMEM;
1151
1152         return 0;
1153 }