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         unsigned                cache_missed:1;
 467
 468         unsigned long           start_time;
 469
 470         struct btree_op         op;
 471         struct data_insert_op   iop;
 472 };
 473
 474 static void bch_cache_read_endio(struct bio *bio)
 475 {
 476         struct bbio *b = container_of(bio, struct bbio, bio);
 477         struct closure *cl = bio->bi_private;
 478         struct search *s = container_of(cl, struct search, cl);
 479
 480         /*
 481          * If the bucket was reused while our bio was in flight, we might have
 482          * read the wrong data. Set s->error but not error so it doesn't get
 483          * counted against the cache device, but we'll still reread the data
 484          * from the backing device.
 485          */
 486
 487         if (bio->bi_status)
 488                 s->iop.status = bio->bi_status;
 489         else if (!KEY_DIRTY(&b->key) &&
 490                  ptr_stale(s->iop.c, &b->key, 0)) {
 491                 atomic_long_inc(&s->iop.c->cache_read_races);
 492                 s->iop.status = BLK_STS_IOERR;
 493         }
 494
 495         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
 496 }
 497
 498 /*
 499  * Read from a single key, handling the initial cache miss if the key starts in
 500  * the middle of the bio
 501  */
 502 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
 503 {
 504         struct search *s = container_of(op, struct search, op);
 505         struct bio *n, *bio = &s->bio.bio;
 506         struct bkey *bio_key;
 507         unsigned ptr;
 508
 509         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
 510                 return MAP_CONTINUE;
 511
 512         if (KEY_INODE(k) != s->iop.inode ||
 513             KEY_START(k) > bio->bi_iter.bi_sector) {
 514                 unsigned bio_sectors = bio_sectors(bio);
 515                 unsigned sectors = KEY_INODE(k) == s->iop.inode
 516                         ? min_t(uint64_t, INT_MAX,
 517                                 KEY_START(k) - bio->bi_iter.bi_sector)
 518                         : INT_MAX;
 519
 520                 int ret = s->d->cache_miss(b, s, bio, sectors);
 521                 if (ret != MAP_CONTINUE)
 522                         return ret;
 523
 524                 /* if this was a complete miss we shouldn't get here */
 525                 BUG_ON(bio_sectors <= sectors);
 526         }
 527
 528         if (!KEY_SIZE(k))
 529                 return MAP_CONTINUE;
 530
 531         /* XXX: figure out best pointer - for multiple cache devices */
 532         ptr = 0;
 533
 534         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
 535
 536         if (KEY_DIRTY(k))
 537                 s->read_dirty_data = true;
 538
 539         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
 540                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
 541                            GFP_NOIO, s->d->bio_split);
 542
 543         bio_key = &container_of(n, struct bbio, bio)->key;
 544         bch_bkey_copy_single_ptr(bio_key, k, ptr);
 545
 546         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
 547         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
 548
 549         n->bi_end_io    = bch_cache_read_endio;
 550         n->bi_private   = &s->cl;
 551
 552         /*
 553          * The bucket we're reading from might be reused while our bio
 554          * is in flight, and we could then end up reading the wrong
 555          * data.
 556          *
 557          * We guard against this by checking (in cache_read_endio()) if
 558          * the pointer is stale again; if so, we treat it as an error
 559          * and reread from the backing device (but we don't pass that
 560          * error up anywhere).
 561          */
 562
 563         __bch_submit_bbio(n, b->c);
 564         return n == bio ? MAP_DONE : MAP_CONTINUE;
 565 }
 566
 567 static void cache_lookup(struct closure *cl)
 568 {
 569         struct search *s = container_of(cl, struct search, iop.cl);
 570         struct bio *bio = &s->bio.bio;
 571         struct cached_dev *dc;
 572         int ret;
 573
 574         bch_btree_op_init(&s->op, -1);
 575
 576         ret = bch_btree_map_keys(&s->op, s->iop.c,
 577                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
 578                                  cache_lookup_fn, MAP_END_KEY);
 579         if (ret == -EAGAIN) {
 580                 continue_at(cl, cache_lookup, bcache_wq);
 581                 return;
 582         }
 583
 584         /*
 585          * We might meet err when searching the btree, If that happens, we will
 586          * get negative ret, in this scenario we should not recover data from
 587          * backing device (when cache device is dirty) because we don't know
 588          * whether bkeys the read request covered are all clean.
 589          *
 590          * And after that happened, s->iop.status is still its initial value
 591          * before we submit s->bio.bio
 592          */
 593         if (ret < 0) {
 594                 BUG_ON(ret == -EINTR);
 595                 if (s->d && s->d->c &&
 596                                 !UUID_FLASH_ONLY(&s->d->c->uuids[s->d->id])) {
 597                         dc = container_of(s->d, struct cached_dev, disk);
 598                         if (dc && atomic_read(&dc->has_dirty))
 599                                 s->recoverable = false;
 600                 }
 601                 if (!s->iop.status)
 602                         s->iop.status = BLK_STS_IOERR;
 603         }
 604
 605         closure_return(cl);
 606 }
 607
 608 /* Common code for the make_request functions */
 609
 610 static void request_endio(struct bio *bio)
 611 {
 612         struct closure *cl = bio->bi_private;
 613
 614         if (bio->bi_status) {
 615                 struct search *s = container_of(cl, struct search, cl);
 616                 s->iop.status = bio->bi_status;
 617                 /* Only cache read errors are recoverable */
 618                 s->recoverable = false;
 619         }
 620
 621         bio_put(bio);
 622         closure_put(cl);
 623 }
 624
 625 static void bio_complete(struct search *s)
 626 {
 627         if (s->orig_bio) {
 628                 struct request_queue *q = s->orig_bio->bi_disk->queue;
 629                 generic_end_io_acct(q, bio_data_dir(s->orig_bio),
 630                                     &s->d->disk->part0, s->start_time);
 631
 632                 trace_bcache_request_end(s->d, s->orig_bio);
 633                 s->orig_bio->bi_status = s->iop.status;
 634                 bio_endio(s->orig_bio);
 635                 s->orig_bio = NULL;
 636         }
 637 }
 638
 639 static void do_bio_hook(struct search *s, struct bio *orig_bio)
 640 {
 641         struct bio *bio = &s->bio.bio;
 642
 643         bio_init(bio, NULL, 0);
 644         __bio_clone_fast(bio, orig_bio);
 645         bio->bi_end_io          = request_endio;
 646         bio->bi_private         = &s->cl;
 647
 648         bio_cnt_set(bio, 3);
 649 }
 650
 651 static void search_free(struct closure *cl)
 652 {
 653         struct search *s = container_of(cl, struct search, cl);
 654
 655         if (s->iop.bio)
 656                 bio_put(s->iop.bio);
 657
 658         bio_complete(s);
 659         closure_debug_destroy(cl);
 660         mempool_free(s, s->d->c->search);
 661 }
 662
 663 static inline struct search *search_alloc(struct bio *bio,
 664                                           struct bcache_device *d)
 665 {
 666         struct search *s;
 667
 668         s = mempool_alloc(d->c->search, GFP_NOIO);
 669
 670         closure_init(&s->cl, NULL);
 671         do_bio_hook(s, bio);
 672
 673         s->orig_bio             = bio;
 674         s->cache_miss           = NULL;
 675         s->cache_missed         = 0;
 676         s->d                    = d;
 677         s->recoverable          = 1;
 678         s->write                = op_is_write(bio_op(bio));
 679         s->read_dirty_data      = 0;
 680         s->start_time           = jiffies;
 681
 682         s->iop.c                = d->c;
 683         s->iop.bio              = NULL;
 684         s->iop.inode            = d->id;
 685         s->iop.write_point      = hash_long((unsigned long) current, 16);
 686         s->iop.write_prio       = 0;
 687         s->iop.status           = 0;
 688         s->iop.flags            = 0;
 689         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
 690         s->iop.wq               = bcache_wq;
 691
 692         return s;
 693 }
 694
 695 /* Cached devices */
 696
 697 static void cached_dev_bio_complete(struct closure *cl)
 698 {
 699         struct search *s = container_of(cl, struct search, cl);
 700         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 701
 702         search_free(cl);
 703         cached_dev_put(dc);
 704 }
 705
 706 /* Process reads */
 707
 708 static void cached_dev_cache_miss_done(struct closure *cl)
 709 {
 710         struct search *s = container_of(cl, struct search, cl);
 711
 712         if (s->iop.replace_collision)
 713                 bch_mark_cache_miss_collision(s->iop.c, s->d);
 714
 715         if (s->iop.bio)
 716                 bio_free_pages(s->iop.bio);
 717
 718         cached_dev_bio_complete(cl);
 719 }
 720
 721 static void cached_dev_read_error(struct closure *cl)
 722 {
 723         struct search *s = container_of(cl, struct search, cl);
 724         struct bio *bio = &s->bio.bio;
 725
 726         /*
 727          * If read request hit dirty data (s->read_dirty_data is true),
 728          * then recovery a failed read request from cached device may
 729          * get a stale data back. So read failure recovery is only
 730          * permitted when read request hit clean data in cache device,
 731          * or when cache read race happened.
 732          */
 733         if (s->recoverable && !s->read_dirty_data) {
 734                 /* Retry from the backing device: */
 735                 trace_bcache_read_retry(s->orig_bio);
 736
 737                 s->iop.status = 0;
 738                 do_bio_hook(s, s->orig_bio);
 739
 740                 /* XXX: invalidate cache */
 741
 742                 closure_bio_submit(bio, cl);
 743         }
 744
 745         continue_at(cl, cached_dev_cache_miss_done, NULL);
 746 }
 747
 748 static void cached_dev_read_done(struct closure *cl)
 749 {
 750         struct search *s = container_of(cl, struct search, cl);
 751         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 752
 753         /*
 754          * We had a cache miss; cache_bio now contains data ready to be inserted
 755          * into the cache.
 756          *
 757          * First, we copy the data we just read from cache_bio's bounce buffers
 758          * to the buffers the original bio pointed to:
 759          */
 760
 761         if (s->iop.bio) {
 762                 bio_reset(s->iop.bio);
 763                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
 764                 bio_copy_dev(s->iop.bio, s->cache_miss);
 765                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
 766                 bch_bio_map(s->iop.bio, NULL);
 767
 768                 bio_copy_data(s->cache_miss, s->iop.bio);
 769
 770                 bio_put(s->cache_miss);
 771                 s->cache_miss = NULL;
 772         }
 773
 774         if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
 775                 bch_data_verify(dc, s->orig_bio);
 776
 777         bio_complete(s);
 778
 779         if (s->iop.bio &&
 780             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
 781                 BUG_ON(!s->iop.replace);
 782                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 783         }
 784
 785         continue_at(cl, cached_dev_cache_miss_done, NULL);
 786 }
 787
 788 static void cached_dev_read_done_bh(struct closure *cl)
 789 {
 790         struct search *s = container_of(cl, struct search, cl);
 791         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 792
 793         bch_mark_cache_accounting(s->iop.c, s->d,
 794                                   !s->cache_missed, s->iop.bypass);
 795         trace_bcache_read(s->orig_bio, !s->cache_missed, s->iop.bypass);
 796
 797         if (s->iop.status)
 798                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
 799         else if (s->iop.bio || verify(dc, &s->bio.bio))
 800                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
 801         else
 802                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
 803 }
 804
 805 static int cached_dev_cache_miss(struct btree *b, struct search *s,
 806                                  struct bio *bio, unsigned sectors)
 807 {
 808         int ret = MAP_CONTINUE;
 809         unsigned reada = 0;
 810         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 811         struct bio *miss, *cache_bio;
 812
 813         s->cache_missed = 1;
 814
 815         if (s->cache_miss || s->iop.bypass) {
 816                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 817                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
 818                 goto out_submit;
 819         }
 820
 821         if (!(bio->bi_opf & REQ_RAHEAD) &&
 822             !(bio->bi_opf & REQ_META) &&
 823             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
 824                 reada = min_t(sector_t, dc->readahead >> 9,
 825                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
 826
 827         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
 828
 829         s->iop.replace_key = KEY(s->iop.inode,
 830                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
 831                                  s->insert_bio_sectors);
 832
 833         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
 834         if (ret)
 835                 return ret;
 836
 837         s->iop.replace = true;
 838
 839         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
 840
 841         /* btree_search_recurse()'s btree iterator is no good anymore */
 842         ret = miss == bio ? MAP_DONE : -EINTR;
 843
 844         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
 845                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
 846                         dc->disk.bio_split);
 847         if (!cache_bio)
 848                 goto out_submit;
 849
 850         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
 851         bio_copy_dev(cache_bio, miss);
 852         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
 853
 854         cache_bio->bi_end_io    = request_endio;
 855         cache_bio->bi_private   = &s->cl;
 856
 857         bch_bio_map(cache_bio, NULL);
 858         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
 859                 goto out_put;
 860
 861         if (reada)
 862                 bch_mark_cache_readahead(s->iop.c, s->d);
 863
 864         s->cache_miss   = miss;
 865         s->iop.bio      = cache_bio;
 866         bio_get(cache_bio);
 867         closure_bio_submit(cache_bio, &s->cl);
 868
 869         return ret;
 870 out_put:
 871         bio_put(cache_bio);
 872 out_submit:
 873         miss->bi_end_io         = request_endio;
 874         miss->bi_private        = &s->cl;
 875         closure_bio_submit(miss, &s->cl);
 876         return ret;
 877 }
 878
 879 static void cached_dev_read(struct cached_dev *dc, struct search *s)
 880 {
 881         struct closure *cl = &s->cl;
 882
 883         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
 884         continue_at(cl, cached_dev_read_done_bh, NULL);
 885 }
 886
 887 /* Process writes */
 888
 889 static void cached_dev_write_complete(struct closure *cl)
 890 {
 891         struct search *s = container_of(cl, struct search, cl);
 892         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
 893
 894         up_read_non_owner(&dc->writeback_lock);
 895         cached_dev_bio_complete(cl);
 896 }
 897
 898 static void cached_dev_write(struct cached_dev *dc, struct search *s)
 899 {
 900         struct closure *cl = &s->cl;
 901         struct bio *bio = &s->bio.bio;
 902         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
 903         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
 904
 905         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
 906
 907         down_read_non_owner(&dc->writeback_lock);
 908         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
 909                 /*
 910                  * We overlap with some dirty data undergoing background
 911                  * writeback, force this write to writeback
 912                  */
 913                 s->iop.bypass = false;
 914                 s->iop.writeback = true;
 915         }
 916
 917         /*
 918          * Discards aren't _required_ to do anything, so skipping if
 919          * check_overlapping returned true is ok
 920          *
 921          * But check_overlapping drops dirty keys for which io hasn't started,
 922          * so we still want to call it.
 923          */
 924         if (bio_op(bio) == REQ_OP_DISCARD)
 925                 s->iop.bypass = true;
 926
 927         if (should_writeback(dc, s->orig_bio,
 928                              cache_mode(dc, bio),
 929                              s->iop.bypass)) {
 930                 s->iop.bypass = false;
 931                 s->iop.writeback = true;
 932         }
 933
 934         if (s->iop.bypass) {
 935                 s->iop.bio = s->orig_bio;
 936                 bio_get(s->iop.bio);
 937
 938                 if ((bio_op(bio) != REQ_OP_DISCARD) ||
 939                     blk_queue_discard(bdev_get_queue(dc->bdev)))
 940                         closure_bio_submit(bio, cl);
 941         } else if (s->iop.writeback) {
 942                 bch_writeback_add(dc);
 943                 s->iop.bio = bio;
 944
 945                 if (bio->bi_opf & REQ_PREFLUSH) {
 946                         /* Also need to send a flush to the backing device */
 947                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
 948                                                              dc->disk.bio_split);
 949
 950                         bio_copy_dev(flush, bio);
 951                         flush->bi_end_io = request_endio;
 952                         flush->bi_private = cl;
 953                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 954
 955                         closure_bio_submit(flush, cl);
 956                 }
 957         } else {
 958                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
 959
 960                 closure_bio_submit(bio, cl);
 961         }
 962
 963         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
 964         continue_at(cl, cached_dev_write_complete, NULL);
 965 }
 966
 967 static void cached_dev_nodata(struct closure *cl)
 968 {
 969         struct search *s = container_of(cl, struct search, cl);
 970         struct bio *bio = &s->bio.bio;
 971
 972         if (s->iop.flush_journal)
 973                 bch_journal_meta(s->iop.c, cl);
 974
 975         /* If it's a flush, we send the flush to the backing device too */
 976         closure_bio_submit(bio, cl);
 977
 978         continue_at(cl, cached_dev_bio_complete, NULL);
 979 }
 980
 981 /* Cached devices - read & write stuff */
 982
 983 static blk_qc_t cached_dev_make_request(struct request_queue *q,
 984                                         struct bio *bio)
 985 {
 986         struct search *s;
 987         struct bcache_device *d = bio->bi_disk->private_data;
 988         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
 989         int rw = bio_data_dir(bio);
 990
 991         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
 992
 993         bio_set_dev(bio, dc->bdev);
 994         bio->bi_iter.bi_sector += dc->sb.data_offset;
 995
 996         if (cached_dev_get(dc)) {
 997                 s = search_alloc(bio, d);
 998                 trace_bcache_request_start(s->d, bio);
 999
1000                 if (!bio->bi_iter.bi_size) {
1001                         /*
1002                          * can't call bch_journal_meta from under
1003                          * generic_make_request
1004                          */
1005                         continue_at_nobarrier(&s->cl,
1006                                               cached_dev_nodata,
1007                                               bcache_wq);
1008                 } else {
1009                         s->iop.bypass = check_should_bypass(dc, bio);
1010
1011                         if (rw)
1012                                 cached_dev_write(dc, s);
1013                         else
1014                                 cached_dev_read(dc, s);
1015                 }
1016         } else {
1017                 if ((bio_op(bio) == REQ_OP_DISCARD) &&
1018                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1019                         bio_endio(bio);
1020                 else
1021                         generic_make_request(bio);
1022         }
1023
1024         return BLK_QC_T_NONE;
1025 }
1026
1027 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1028                             unsigned int cmd, unsigned long arg)
1029 {
1030         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1031         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1032 }
1033
1034 static int cached_dev_congested(void *data, int bits)
1035 {
1036         struct bcache_device *d = data;
1037         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1038         struct request_queue *q = bdev_get_queue(dc->bdev);
1039         int ret = 0;
1040
1041         if (bdi_congested(q->backing_dev_info, bits))
1042                 return 1;
1043
1044         if (cached_dev_get(dc)) {
1045                 unsigned i;
1046                 struct cache *ca;
1047
1048                 for_each_cache(ca, d->c, i) {
1049                         q = bdev_get_queue(ca->bdev);
1050                         ret |= bdi_congested(q->backing_dev_info, bits);
1051                 }
1052
1053                 cached_dev_put(dc);
1054         }
1055
1056         return ret;
1057 }
1058
1059 void bch_cached_dev_request_init(struct cached_dev *dc)
1060 {
1061         struct gendisk *g = dc->disk.disk;
1062
1063         g->queue->make_request_fn               = cached_dev_make_request;
1064         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1065         dc->disk.cache_miss                     = cached_dev_cache_miss;
1066         dc->disk.ioctl                          = cached_dev_ioctl;
1067 }
1068
1069 /* Flash backed devices */
1070
1071 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1072                                 struct bio *bio, unsigned sectors)
1073 {
1074         unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1075
1076         swap(bio->bi_iter.bi_size, bytes);
1077         zero_fill_bio(bio);
1078         swap(bio->bi_iter.bi_size, bytes);
1079
1080         bio_advance(bio, bytes);
1081
1082         if (!bio->bi_iter.bi_size)
1083                 return MAP_DONE;
1084
1085         return MAP_CONTINUE;
1086 }
1087
1088 static void flash_dev_nodata(struct closure *cl)
1089 {
1090         struct search *s = container_of(cl, struct search, cl);
1091
1092         if (s->iop.flush_journal)
1093                 bch_journal_meta(s->iop.c, cl);
1094
1095         continue_at(cl, search_free, NULL);
1096 }
1097
1098 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1099                                              struct bio *bio)
1100 {
1101         struct search *s;
1102         struct closure *cl;
1103         struct bcache_device *d = bio->bi_disk->private_data;
1104         int rw = bio_data_dir(bio);
1105
1106         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1107
1108         s = search_alloc(bio, d);
1109         cl = &s->cl;
1110         bio = &s->bio.bio;
1111
1112         trace_bcache_request_start(s->d, bio);
1113
1114         if (!bio->bi_iter.bi_size) {
1115                 /*
1116                  * can't call bch_journal_meta from under
1117                  * generic_make_request
1118                  */
1119                 continue_at_nobarrier(&s->cl,
1120                                       flash_dev_nodata,
1121                                       bcache_wq);
1122                 return BLK_QC_T_NONE;
1123         } else if (rw) {
1124                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1125                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1126                                         &KEY(d->id, bio_end_sector(bio), 0));
1127
1128                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1129                 s->iop.writeback        = true;
1130                 s->iop.bio              = bio;
1131
1132                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1133         } else {
1134                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1135         }
1136
1137         continue_at(cl, search_free, NULL);
1138         return BLK_QC_T_NONE;
1139 }
1140
1141 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1142                            unsigned int cmd, unsigned long arg)
1143 {
1144         return -ENOTTY;
1145 }
1146
1147 static int flash_dev_congested(void *data, int bits)
1148 {
1149         struct bcache_device *d = data;
1150         struct request_queue *q;
1151         struct cache *ca;
1152         unsigned i;
1153         int ret = 0;
1154
1155         for_each_cache(ca, d->c, i) {
1156                 q = bdev_get_queue(ca->bdev);
1157                 ret |= bdi_congested(q->backing_dev_info, bits);
1158         }
1159
1160         return ret;
1161 }
1162
1163 void bch_flash_dev_request_init(struct bcache_device *d)
1164 {
1165         struct gendisk *g = d->disk;
1166
1167         g->queue->make_request_fn               = flash_dev_make_request;
1168         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1169         d->cache_miss                           = flash_dev_cache_miss;
1170         d->ioctl                                = flash_dev_ioctl;
1171 }
1172
1173 void bch_request_exit(void)
1174 {
1175         if (bch_search_cache)
1176                 kmem_cache_destroy(bch_search_cache);
1177 }
1178
1179 int __init bch_request_init(void)
1180 {
1181         bch_search_cache = KMEM_CACHE(search, 0);
1182         if (!bch_search_cache)
1183                 return -ENOMEM;
1184
1185         return 0;
1186 }