drivers/md/bcache/btree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  *
   5  * Uses a block device as cache for other block devices; optimized for SSDs.
   6  * All allocation is done in buckets, which should match the erase block size
   7  * of the device.
   8  *
   9  * Buckets containing cached data are kept on a heap sorted by priority;
  10  * bucket priority is increased on cache hit, and periodically all the buckets
  11  * on the heap have their priority scaled down. This currently is just used as
  12  * an LRU but in the future should allow for more intelligent heuristics.
  13  *
  14  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  15  * counter. Garbage collection is used to remove stale pointers.
  16  *
  17  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  18  * as keys are inserted we only sort the pages that have not yet been written.
  19  * When garbage collection is run, we resort the entire node.
  20  *
  21  * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
  22  */
  23
  24 #include "bcache.h"
  25 #include "btree.h"
  26 #include "debug.h"
  27 #include "extents.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/random.h>
  35 #include <linux/rcupdate.h>
  36 #include <linux/sched/clock.h>
  37 #include <linux/rculist.h>
  38 #include <linux/delay.h>
  39 #include <trace/events/bcache.h>
  40
  41 /*
  42  * Todo:
  43  * register_bcache: Return errors out to userspace correctly
  44  *
  45  * Writeback: don't undirty key until after a cache flush
  46  *
  47  * Create an iterator for key pointers
  48  *
  49  * On btree write error, mark bucket such that it won't be freed from the cache
  50  *
  51  * Journalling:
  52  *   Check for bad keys in replay
  53  *   Propagate barriers
  54  *   Refcount journal entries in journal_replay
  55  *
  56  * Garbage collection:
  57  *   Finish incremental gc
  58  *   Gc should free old UUIDs, data for invalid UUIDs
  59  *
  60  * Provide a way to list backing device UUIDs we have data cached for, and
  61  * probably how long it's been since we've seen them, and a way to invalidate
  62  * dirty data for devices that will never be attached again
  63  *
  64  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  65  * that based on that and how much dirty data we have we can keep writeback
  66  * from being starved
  67  *
  68  * Add a tracepoint or somesuch to watch for writeback starvation
  69  *
  70  * When btree depth > 1 and splitting an interior node, we have to make sure
  71  * alloc_bucket() cannot fail. This should be true but is not completely
  72  * obvious.
  73  *
  74  * Plugging?
  75  *
  76  * If data write is less than hard sector size of ssd, round up offset in open
  77  * bucket to the next whole sector
  78  *
  79  * Superblock needs to be fleshed out for multiple cache devices
  80  *
  81  * Add a sysfs tunable for the number of writeback IOs in flight
  82  *
  83  * Add a sysfs tunable for the number of open data buckets
  84  *
  85  * IO tracking: Can we track when one process is doing io on behalf of another?
  86  * IO tracking: Don't use just an average, weigh more recent stuff higher
  87  *
  88  * Test module load/unload
  89  */
  90
  91 #define MAX_NEED_GC             64
  92 #define MAX_SAVE_PRIO           72
  93 #define MAX_GC_TIMES            100
  94 #define MIN_GC_NODES            100
  95 #define GC_SLEEP_MS             100
  96
  97 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  98
  99 #define PTR_HASH(c, k)                                                  \
 100         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 101
 102 static struct workqueue_struct *btree_io_wq;
 103
 104 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 105
 106
 107 static inline struct bset *write_block(struct btree *b)
 108 {
 109         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c->cache);
 110 }
 111
 112 static void bch_btree_init_next(struct btree *b)
 113 {
 114         /* If not a leaf node, always sort */
 115         if (b->level && b->keys.nsets)
 116                 bch_btree_sort(&b->keys, &b->c->sort);
 117         else
 118                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 119
 120         if (b->written < btree_blocks(b))
 121                 bch_bset_init_next(&b->keys, write_block(b),
 122                                    bset_magic(&b->c->cache->sb));
 123
 124 }
 125
 126 /* Btree key manipulation */
 127
 128 void bkey_put(struct cache_set *c, struct bkey *k)
 129 {
 130         unsigned int i;
 131
 132         for (i = 0; i < KEY_PTRS(k); i++)
 133                 if (ptr_available(c, k, i))
 134                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 135 }
 136
 137 /* Btree IO */
 138
 139 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 140 {
 141         uint64_t crc = b->key.ptr[0];
 142         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 143
 144         crc = crc64_be(crc, data, end - data);
 145         return crc ^ 0xffffffffffffffffULL;
 146 }
 147
 148 void bch_btree_node_read_done(struct btree *b)
 149 {
 150         const char *err = "bad btree header";
 151         struct bset *i = btree_bset_first(b);
 152         struct btree_iter *iter;
 153
 154         /*
 155          * c->fill_iter can allocate an iterator with more memory space
 156          * than static MAX_BSETS.
 157          * See the comment arount cache_set->fill_iter.
 158          */
 159         iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
 160         iter->size = b->c->cache->sb.bucket_size / b->c->cache->sb.block_size;
 161         iter->used = 0;
 162
 163 #ifdef CONFIG_BCACHE_DEBUG
 164         iter->b = &b->keys;
 165 #endif
 166
 167         if (!i->seq)
 168                 goto err;
 169
 170         for (;
 171              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 172              i = write_block(b)) {
 173                 err = "unsupported bset version";
 174                 if (i->version > BCACHE_BSET_VERSION)
 175                         goto err;
 176
 177                 err = "bad btree header";
 178                 if (b->written + set_blocks(i, block_bytes(b->c->cache)) >
 179                     btree_blocks(b))
 180                         goto err;
 181
 182                 err = "bad magic";
 183                 if (i->magic != bset_magic(&b->c->cache->sb))
 184                         goto err;
 185
 186                 err = "bad checksum";
 187                 switch (i->version) {
 188                 case 0:
 189                         if (i->csum != csum_set(i))
 190                                 goto err;
 191                         break;
 192                 case BCACHE_BSET_VERSION:
 193                         if (i->csum != btree_csum_set(b, i))
 194                                 goto err;
 195                         break;
 196                 }
 197
 198                 err = "empty set";
 199                 if (i != b->keys.set[0].data && !i->keys)
 200                         goto err;
 201
 202                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 203
 204                 b->written += set_blocks(i, block_bytes(b->c->cache));
 205         }
 206
 207         err = "corrupted btree";
 208         for (i = write_block(b);
 209              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 210              i = ((void *) i) + block_bytes(b->c->cache))
 211                 if (i->seq == b->keys.set[0].data->seq)
 212                         goto err;
 213
 214         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 215
 216         i = b->keys.set[0].data;
 217         err = "short btree key";
 218         if (b->keys.set[0].size &&
 219             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 220                 goto err;
 221
 222         if (b->written < btree_blocks(b))
 223                 bch_bset_init_next(&b->keys, write_block(b),
 224                                    bset_magic(&b->c->cache->sb));
 225 out:
 226         mempool_free(iter, &b->c->fill_iter);
 227         return;
 228 err:
 229         set_btree_node_io_error(b);
 230         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 231                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 232                             bset_block_offset(b, i), i->keys);
 233         goto out;
 234 }
 235
 236 static void btree_node_read_endio(struct bio *bio)
 237 {
 238         struct closure *cl = bio->bi_private;
 239
 240         closure_put(cl);
 241 }
 242
 243 static void bch_btree_node_read(struct btree *b)
 244 {
 245         uint64_t start_time = local_clock();
 246         struct closure cl;
 247         struct bio *bio;
 248
 249         trace_bcache_btree_read(b);
 250
 251         closure_init_stack(&cl);
 252
 253         bio = bch_bbio_alloc(b->c);
 254         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 255         bio->bi_end_io  = btree_node_read_endio;
 256         bio->bi_private = &cl;
 257         bio->bi_opf = REQ_OP_READ | REQ_META;
 258
 259         bch_bio_map(bio, b->keys.set[0].data);
 260
 261         bch_submit_bbio(bio, b->c, &b->key, 0);
 262         closure_sync(&cl);
 263
 264         if (bio->bi_status)
 265                 set_btree_node_io_error(b);
 266
 267         bch_bbio_free(bio, b->c);
 268
 269         if (btree_node_io_error(b))
 270                 goto err;
 271
 272         bch_btree_node_read_done(b);
 273         bch_time_stats_update(&b->c->btree_read_time, start_time);
 274
 275         return;
 276 err:
 277         bch_cache_set_error(b->c, "io error reading bucket %zu",
 278                             PTR_BUCKET_NR(b->c, &b->key, 0));
 279 }
 280
 281 static void btree_complete_write(struct btree *b, struct btree_write *w)
 282 {
 283         if (w->prio_blocked &&
 284             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 285                 wake_up_allocators(b->c);
 286
 287         if (w->journal) {
 288                 atomic_dec_bug(w->journal);
 289                 __closure_wake_up(&b->c->journal.wait);
 290         }
 291
 292         w->prio_blocked = 0;
 293         w->journal      = NULL;
 294 }
 295
 296 static void btree_node_write_unlock(struct closure *cl)
 297 {
 298         struct btree *b = container_of(cl, struct btree, io);
 299
 300         up(&b->io_mutex);
 301 }
 302
 303 static void __btree_node_write_done(struct closure *cl)
 304 {
 305         struct btree *b = container_of(cl, struct btree, io);
 306         struct btree_write *w = btree_prev_write(b);
 307
 308         bch_bbio_free(b->bio, b->c);
 309         b->bio = NULL;
 310         btree_complete_write(b, w);
 311
 312         if (btree_node_dirty(b))
 313                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 314
 315         closure_return_with_destructor(cl, btree_node_write_unlock);
 316 }
 317
 318 static void btree_node_write_done(struct closure *cl)
 319 {
 320         struct btree *b = container_of(cl, struct btree, io);
 321
 322         bio_free_pages(b->bio);
 323         __btree_node_write_done(cl);
 324 }
 325
 326 static void btree_node_write_endio(struct bio *bio)
 327 {
 328         struct closure *cl = bio->bi_private;
 329         struct btree *b = container_of(cl, struct btree, io);
 330
 331         if (bio->bi_status)
 332                 set_btree_node_io_error(b);
 333
 334         bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 335         closure_put(cl);
 336 }
 337
 338 static void do_btree_node_write(struct btree *b)
 339 {
 340         struct closure *cl = &b->io;
 341         struct bset *i = btree_bset_last(b);
 342         BKEY_PADDED(key) k;
 343
 344         i->version      = BCACHE_BSET_VERSION;
 345         i->csum         = btree_csum_set(b, i);
 346
 347         BUG_ON(b->bio);
 348         b->bio = bch_bbio_alloc(b->c);
 349
 350         b->bio->bi_end_io       = btree_node_write_endio;
 351         b->bio->bi_private      = cl;
 352         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c->cache));
 353         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 354         bch_bio_map(b->bio, i);
 355
 356         /*
 357          * If we're appending to a leaf node, we don't technically need FUA -
 358          * this write just needs to be persisted before the next journal write,
 359          * which will be marked FLUSH|FUA.
 360          *
 361          * Similarly if we're writing a new btree root - the pointer is going to
 362          * be in the next journal entry.
 363          *
 364          * But if we're writing a new btree node (that isn't a root) or
 365          * appending to a non leaf btree node, we need either FUA or a flush
 366          * when we write the parent with the new pointer. FUA is cheaper than a
 367          * flush, and writes appending to leaf nodes aren't blocking anything so
 368          * just make all btree node writes FUA to keep things sane.
 369          */
 370
 371         bkey_copy(&k.key, &b->key);
 372         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 373                        bset_sector_offset(&b->keys, i));
 374
 375         if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 376                 struct bio_vec *bv;
 377                 void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 378                 struct bvec_iter_all iter_all;
 379
 380                 bio_for_each_segment_all(bv, b->bio, iter_all) {
 381                         memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
 382                         addr += PAGE_SIZE;
 383                 }
 384
 385                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 386
 387                 continue_at(cl, btree_node_write_done, NULL);
 388         } else {
 389                 /*
 390                  * No problem for multipage bvec since the bio is
 391                  * just allocated
 392                  */
 393                 b->bio->bi_vcnt = 0;
 394                 bch_bio_map(b->bio, i);
 395
 396                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 397
 398                 closure_sync(cl);
 399                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 400         }
 401 }
 402
 403 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 404 {
 405         struct bset *i = btree_bset_last(b);
 406
 407         lockdep_assert_held(&b->write_lock);
 408
 409         trace_bcache_btree_write(b);
 410
 411         BUG_ON(current->bio_list);
 412         BUG_ON(b->written >= btree_blocks(b));
 413         BUG_ON(b->written && !i->keys);
 414         BUG_ON(btree_bset_first(b)->seq != i->seq);
 415         bch_check_keys(&b->keys, "writing");
 416
 417         cancel_delayed_work(&b->work);
 418
 419         /* If caller isn't waiting for write, parent refcount is cache set */
 420         down(&b->io_mutex);
 421         closure_init(&b->io, parent ?: &b->c->cl);
 422
 423         clear_bit(BTREE_NODE_dirty,      &b->flags);
 424         change_bit(BTREE_NODE_write_idx, &b->flags);
 425
 426         do_btree_node_write(b);
 427
 428         atomic_long_add(set_blocks(i, block_bytes(b->c->cache)) * b->c->cache->sb.block_size,
 429                         &b->c->cache->btree_sectors_written);
 430
 431         b->written += set_blocks(i, block_bytes(b->c->cache));
 432 }
 433
 434 void bch_btree_node_write(struct btree *b, struct closure *parent)
 435 {
 436         unsigned int nsets = b->keys.nsets;
 437
 438         lockdep_assert_held(&b->lock);
 439
 440         __bch_btree_node_write(b, parent);
 441
 442         /*
 443          * do verify if there was more than one set initially (i.e. we did a
 444          * sort) and we sorted down to a single set:
 445          */
 446         if (nsets && !b->keys.nsets)
 447                 bch_btree_verify(b);
 448
 449         bch_btree_init_next(b);
 450 }
 451
 452 static void bch_btree_node_write_sync(struct btree *b)
 453 {
 454         struct closure cl;
 455
 456         closure_init_stack(&cl);
 457
 458         mutex_lock(&b->write_lock);
 459         bch_btree_node_write(b, &cl);
 460         mutex_unlock(&b->write_lock);
 461
 462         closure_sync(&cl);
 463 }
 464
 465 static void btree_node_write_work(struct work_struct *w)
 466 {
 467         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 468
 469         mutex_lock(&b->write_lock);
 470         if (btree_node_dirty(b))
 471                 __bch_btree_node_write(b, NULL);
 472         mutex_unlock(&b->write_lock);
 473 }
 474
 475 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 476 {
 477         struct bset *i = btree_bset_last(b);
 478         struct btree_write *w = btree_current_write(b);
 479
 480         lockdep_assert_held(&b->write_lock);
 481
 482         BUG_ON(!b->written);
 483         BUG_ON(!i->keys);
 484
 485         if (!btree_node_dirty(b))
 486                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 487
 488         set_btree_node_dirty(b);
 489
 490         /*
 491          * w->journal is always the oldest journal pin of all bkeys
 492          * in the leaf node, to make sure the oldest jset seq won't
 493          * be increased before this btree node is flushed.
 494          */
 495         if (journal_ref) {
 496                 if (w->journal &&
 497                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 498                         atomic_dec_bug(w->journal);
 499                         w->journal = NULL;
 500                 }
 501
 502                 if (!w->journal) {
 503                         w->journal = journal_ref;
 504                         atomic_inc(w->journal);
 505                 }
 506         }
 507
 508         /* Force write if set is too big */
 509         if (set_bytes(i) > PAGE_SIZE - 48 &&
 510             !current->bio_list)
 511                 bch_btree_node_write(b, NULL);
 512 }
 513
 514 /*
 515  * Btree in memory cache - allocation/freeing
 516  * mca -> memory cache
 517  */
 518
 519 #define mca_reserve(c)  (((!IS_ERR_OR_NULL(c->root) && c->root->level) \
 520                           ? c->root->level : 1) * 8 + 16)
 521 #define mca_can_free(c)                                         \
 522         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 523
 524 static void mca_data_free(struct btree *b)
 525 {
 526         BUG_ON(b->io_mutex.count != 1);
 527
 528         bch_btree_keys_free(&b->keys);
 529
 530         b->c->btree_cache_used--;
 531         list_move(&b->list, &b->c->btree_cache_freed);
 532 }
 533
 534 static void mca_bucket_free(struct btree *b)
 535 {
 536         BUG_ON(btree_node_dirty(b));
 537
 538         b->key.ptr[0] = 0;
 539         hlist_del_init_rcu(&b->hash);
 540         list_move(&b->list, &b->c->btree_cache_freeable);
 541 }
 542
 543 static unsigned int btree_order(struct bkey *k)
 544 {
 545         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 546 }
 547
 548 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 549 {
 550         if (!bch_btree_keys_alloc(&b->keys,
 551                                   max_t(unsigned int,
 552                                         ilog2(b->c->btree_pages),
 553                                         btree_order(k)),
 554                                   gfp)) {
 555                 b->c->btree_cache_used++;
 556                 list_move(&b->list, &b->c->btree_cache);
 557         } else {
 558                 list_move(&b->list, &b->c->btree_cache_freed);
 559         }
 560 }
 561
 562 static struct btree *mca_bucket_alloc(struct cache_set *c,
 563                                       struct bkey *k, gfp_t gfp)
 564 {
 565         /*
 566          * kzalloc() is necessary here for initialization,
 567          * see code comments in bch_btree_keys_init().
 568          */
 569         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 570
 571         if (!b)
 572                 return NULL;
 573
 574         init_rwsem(&b->lock);
 575         lockdep_set_novalidate_class(&b->lock);
 576         mutex_init(&b->write_lock);
 577         lockdep_set_novalidate_class(&b->write_lock);
 578         INIT_LIST_HEAD(&b->list);
 579         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 580         b->c = c;
 581         sema_init(&b->io_mutex, 1);
 582
 583         mca_data_alloc(b, k, gfp);
 584         return b;
 585 }
 586
 587 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
 588 {
 589         struct closure cl;
 590
 591         closure_init_stack(&cl);
 592         lockdep_assert_held(&b->c->bucket_lock);
 593
 594         if (!down_write_trylock(&b->lock))
 595                 return -ENOMEM;
 596
 597         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 598
 599         if (b->keys.page_order < min_order)
 600                 goto out_unlock;
 601
 602         if (!flush) {
 603                 if (btree_node_dirty(b))
 604                         goto out_unlock;
 605
 606                 if (down_trylock(&b->io_mutex))
 607                         goto out_unlock;
 608                 up(&b->io_mutex);
 609         }
 610
 611 retry:
 612         /*
 613          * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
 614          * __bch_btree_node_write(). To avoid an extra flush, acquire
 615          * b->write_lock before checking BTREE_NODE_dirty bit.
 616          */
 617         mutex_lock(&b->write_lock);
 618         /*
 619          * If this btree node is selected in btree_flush_write() by journal
 620          * code, delay and retry until the node is flushed by journal code
 621          * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
 622          */
 623         if (btree_node_journal_flush(b)) {
 624                 pr_debug("bnode %p is flushing by journal, retry\n", b);
 625                 mutex_unlock(&b->write_lock);
 626                 udelay(1);
 627                 goto retry;
 628         }
 629
 630         if (btree_node_dirty(b))
 631                 __bch_btree_node_write(b, &cl);
 632         mutex_unlock(&b->write_lock);
 633
 634         closure_sync(&cl);
 635
 636         /* wait for any in flight btree write */
 637         down(&b->io_mutex);
 638         up(&b->io_mutex);
 639
 640         return 0;
 641 out_unlock:
 642         rw_unlock(true, b);
 643         return -ENOMEM;
 644 }
 645
 646 static unsigned long bch_mca_scan(struct shrinker *shrink,
 647                                   struct shrink_control *sc)
 648 {
 649         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 650         struct btree *b, *t;
 651         unsigned long i, nr = sc->nr_to_scan;
 652         unsigned long freed = 0;
 653         unsigned int btree_cache_used;
 654
 655         if (c->shrinker_disabled)
 656                 return SHRINK_STOP;
 657
 658         if (c->btree_cache_alloc_lock)
 659                 return SHRINK_STOP;
 660
 661         /* Return -1 if we can't do anything right now */
 662         if (sc->gfp_mask & __GFP_IO)
 663                 mutex_lock(&c->bucket_lock);
 664         else if (!mutex_trylock(&c->bucket_lock))
 665                 return -1;
 666
 667         /*
 668          * It's _really_ critical that we don't free too many btree nodes - we
 669          * have to always leave ourselves a reserve. The reserve is how we
 670          * guarantee that allocating memory for a new btree node can always
 671          * succeed, so that inserting keys into the btree can always succeed and
 672          * IO can always make forward progress:
 673          */
 674         nr /= c->btree_pages;
 675         if (nr == 0)
 676                 nr = 1;
 677         nr = min_t(unsigned long, nr, mca_can_free(c));
 678
 679         i = 0;
 680         btree_cache_used = c->btree_cache_used;
 681         list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
 682                 if (nr <= 0)
 683                         goto out;
 684
 685                 if (!mca_reap(b, 0, false)) {
 686                         mca_data_free(b);
 687                         rw_unlock(true, b);
 688                         freed++;
 689                 }
 690                 nr--;
 691                 i++;
 692         }
 693
 694         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 695                 if (nr <= 0 || i >= btree_cache_used)
 696                         goto out;
 697
 698                 if (!mca_reap(b, 0, false)) {
 699                         mca_bucket_free(b);
 700                         mca_data_free(b);
 701                         rw_unlock(true, b);
 702                         freed++;
 703                 }
 704
 705                 nr--;
 706                 i++;
 707         }
 708 out:
 709         mutex_unlock(&c->bucket_lock);
 710         return freed * c->btree_pages;
 711 }
 712
 713 static unsigned long bch_mca_count(struct shrinker *shrink,
 714                                    struct shrink_control *sc)
 715 {
 716         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 717
 718         if (c->shrinker_disabled)
 719                 return 0;
 720
 721         if (c->btree_cache_alloc_lock)
 722                 return 0;
 723
 724         return mca_can_free(c) * c->btree_pages;
 725 }
 726
 727 void bch_btree_cache_free(struct cache_set *c)
 728 {
 729         struct btree *b;
 730         struct closure cl;
 731
 732         closure_init_stack(&cl);
 733
 734         if (c->shrink.list.next)
 735                 unregister_shrinker(&c->shrink);
 736
 737         mutex_lock(&c->bucket_lock);
 738
 739 #ifdef CONFIG_BCACHE_DEBUG
 740         if (c->verify_data)
 741                 list_move(&c->verify_data->list, &c->btree_cache);
 742
 743         free_pages((unsigned long) c->verify_ondisk, ilog2(meta_bucket_pages(&c->cache->sb)));
 744 #endif
 745
 746         list_splice(&c->btree_cache_freeable,
 747                     &c->btree_cache);
 748
 749         while (!list_empty(&c->btree_cache)) {
 750                 b = list_first_entry(&c->btree_cache, struct btree, list);
 751
 752                 /*
 753                  * This function is called by cache_set_free(), no I/O
 754                  * request on cache now, it is unnecessary to acquire
 755                  * b->write_lock before clearing BTREE_NODE_dirty anymore.
 756                  */
 757                 if (btree_node_dirty(b)) {
 758                         btree_complete_write(b, btree_current_write(b));
 759                         clear_bit(BTREE_NODE_dirty, &b->flags);
 760                 }
 761                 mca_data_free(b);
 762         }
 763
 764         while (!list_empty(&c->btree_cache_freed)) {
 765                 b = list_first_entry(&c->btree_cache_freed,
 766                                      struct btree, list);
 767                 list_del(&b->list);
 768                 cancel_delayed_work_sync(&b->work);
 769                 kfree(b);
 770         }
 771
 772         mutex_unlock(&c->bucket_lock);
 773 }
 774
 775 int bch_btree_cache_alloc(struct cache_set *c)
 776 {
 777         unsigned int i;
 778
 779         for (i = 0; i < mca_reserve(c); i++)
 780                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 781                         return -ENOMEM;
 782
 783         list_splice_init(&c->btree_cache,
 784                          &c->btree_cache_freeable);
 785
 786 #ifdef CONFIG_BCACHE_DEBUG
 787         mutex_init(&c->verify_lock);
 788
 789         c->verify_ondisk = (void *)
 790                 __get_free_pages(GFP_KERNEL|__GFP_COMP,
 791                                  ilog2(meta_bucket_pages(&c->cache->sb)));
 792         if (!c->verify_ondisk) {
 793                 /*
 794                  * Don't worry about the mca_rereserve buckets
 795                  * allocated in previous for-loop, they will be
 796                  * handled properly in bch_cache_set_unregister().
 797                  */
 798                 return -ENOMEM;
 799         }
 800
 801         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 802
 803         if (c->verify_data &&
 804             c->verify_data->keys.set->data)
 805                 list_del_init(&c->verify_data->list);
 806         else
 807                 c->verify_data = NULL;
 808 #endif
 809
 810         c->shrink.count_objects = bch_mca_count;
 811         c->shrink.scan_objects = bch_mca_scan;
 812         c->shrink.seeks = 4;
 813         c->shrink.batch = c->btree_pages * 2;
 814
 815         if (register_shrinker(&c->shrink, "md-bcache:%pU", c->set_uuid))
 816                 pr_warn("bcache: %s: could not register shrinker\n",
 817                                 __func__);
 818
 819         return 0;
 820 }
 821
 822 /* Btree in memory cache - hash table */
 823
 824 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 825 {
 826         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 827 }
 828
 829 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 830 {
 831         struct btree *b;
 832
 833         rcu_read_lock();
 834         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 835                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 836                         goto out;
 837         b = NULL;
 838 out:
 839         rcu_read_unlock();
 840         return b;
 841 }
 842
 843 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 844 {
 845         spin_lock(&c->btree_cannibalize_lock);
 846         if (likely(c->btree_cache_alloc_lock == NULL)) {
 847                 c->btree_cache_alloc_lock = current;
 848         } else if (c->btree_cache_alloc_lock != current) {
 849                 if (op)
 850                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 851                                         TASK_UNINTERRUPTIBLE);
 852                 spin_unlock(&c->btree_cannibalize_lock);
 853                 return -EINTR;
 854         }
 855         spin_unlock(&c->btree_cannibalize_lock);
 856
 857         return 0;
 858 }
 859
 860 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 861                                      struct bkey *k)
 862 {
 863         struct btree *b;
 864
 865         trace_bcache_btree_cache_cannibalize(c);
 866
 867         if (mca_cannibalize_lock(c, op))
 868                 return ERR_PTR(-EINTR);
 869
 870         list_for_each_entry_reverse(b, &c->btree_cache, list)
 871                 if (!mca_reap(b, btree_order(k), false))
 872                         return b;
 873
 874         list_for_each_entry_reverse(b, &c->btree_cache, list)
 875                 if (!mca_reap(b, btree_order(k), true))
 876                         return b;
 877
 878         WARN(1, "btree cache cannibalize failed\n");
 879         return ERR_PTR(-ENOMEM);
 880 }
 881
 882 /*
 883  * We can only have one thread cannibalizing other cached btree nodes at a time,
 884  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 885  * cannibalize_bucket() will take. This means every time we unlock the root of
 886  * the btree, we need to release this lock if we have it held.
 887  */
 888 void bch_cannibalize_unlock(struct cache_set *c)
 889 {
 890         spin_lock(&c->btree_cannibalize_lock);
 891         if (c->btree_cache_alloc_lock == current) {
 892                 c->btree_cache_alloc_lock = NULL;
 893                 wake_up(&c->btree_cache_wait);
 894         }
 895         spin_unlock(&c->btree_cannibalize_lock);
 896 }
 897
 898 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 899                                struct bkey *k, int level)
 900 {
 901         struct btree *b;
 902
 903         BUG_ON(current->bio_list);
 904
 905         lockdep_assert_held(&c->bucket_lock);
 906
 907         if (mca_find(c, k))
 908                 return NULL;
 909
 910         /* btree_free() doesn't free memory; it sticks the node on the end of
 911          * the list. Check if there's any freed nodes there:
 912          */
 913         list_for_each_entry(b, &c->btree_cache_freeable, list)
 914                 if (!mca_reap(b, btree_order(k), false))
 915                         goto out;
 916
 917         /* We never free struct btree itself, just the memory that holds the on
 918          * disk node. Check the freed list before allocating a new one:
 919          */
 920         list_for_each_entry(b, &c->btree_cache_freed, list)
 921                 if (!mca_reap(b, 0, false)) {
 922                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 923                         if (!b->keys.set[0].data)
 924                                 goto err;
 925                         else
 926                                 goto out;
 927                 }
 928
 929         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 930         if (!b)
 931                 goto err;
 932
 933         BUG_ON(!down_write_trylock(&b->lock));
 934         if (!b->keys.set->data)
 935                 goto err;
 936 out:
 937         BUG_ON(b->io_mutex.count != 1);
 938
 939         bkey_copy(&b->key, k);
 940         list_move(&b->list, &c->btree_cache);
 941         hlist_del_init_rcu(&b->hash);
 942         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 943
 944         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 945         b->parent       = (void *) ~0UL;
 946         b->flags        = 0;
 947         b->written      = 0;
 948         b->level        = level;
 949
 950         if (!b->level)
 951                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 952                                     &b->c->expensive_debug_checks);
 953         else
 954                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 955                                     &b->c->expensive_debug_checks);
 956
 957         return b;
 958 err:
 959         if (b)
 960                 rw_unlock(true, b);
 961
 962         b = mca_cannibalize(c, op, k);
 963         if (!IS_ERR(b))
 964                 goto out;
 965
 966         return b;
 967 }
 968
 969 /*
 970  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 971  * in from disk if necessary.
 972  *
 973  * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
 974  *
 975  * The btree node will have either a read or a write lock held, depending on
 976  * level and op->lock.
 977  */
 978 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 979                                  struct bkey *k, int level, bool write,
 980                                  struct btree *parent)
 981 {
 982         int i = 0;
 983         struct btree *b;
 984
 985         BUG_ON(level < 0);
 986 retry:
 987         b = mca_find(c, k);
 988
 989         if (!b) {
 990                 if (current->bio_list)
 991                         return ERR_PTR(-EAGAIN);
 992
 993                 mutex_lock(&c->bucket_lock);
 994                 b = mca_alloc(c, op, k, level);
 995                 mutex_unlock(&c->bucket_lock);
 996
 997                 if (!b)
 998                         goto retry;
 999                 if (IS_ERR(b))
1000                         return b;
1001
1002                 bch_btree_node_read(b);
1003
1004                 if (!write)
1005                         downgrade_write(&b->lock);
1006         } else {
1007                 rw_lock(write, b, level);
1008                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1009                         rw_unlock(write, b);
1010                         goto retry;
1011                 }
1012                 BUG_ON(b->level != level);
1013         }
1014
1015         if (btree_node_io_error(b)) {
1016                 rw_unlock(write, b);
1017                 return ERR_PTR(-EIO);
1018         }
1019
1020         BUG_ON(!b->written);
1021
1022         b->parent = parent;
1023
1024         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1025                 prefetch(b->keys.set[i].tree);
1026                 prefetch(b->keys.set[i].data);
1027         }
1028
1029         for (; i <= b->keys.nsets; i++)
1030                 prefetch(b->keys.set[i].data);
1031
1032         return b;
1033 }
1034
1035 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1036 {
1037         struct btree *b;
1038
1039         mutex_lock(&parent->c->bucket_lock);
1040         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1041         mutex_unlock(&parent->c->bucket_lock);
1042
1043         if (!IS_ERR_OR_NULL(b)) {
1044                 b->parent = parent;
1045                 bch_btree_node_read(b);
1046                 rw_unlock(true, b);
1047         }
1048 }
1049
1050 /* Btree alloc */
1051
1052 static void btree_node_free(struct btree *b)
1053 {
1054         trace_bcache_btree_node_free(b);
1055
1056         BUG_ON(b == b->c->root);
1057
1058 retry:
1059         mutex_lock(&b->write_lock);
1060         /*
1061          * If the btree node is selected and flushing in btree_flush_write(),
1062          * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1063          * then it is safe to free the btree node here. Otherwise this btree
1064          * node will be in race condition.
1065          */
1066         if (btree_node_journal_flush(b)) {
1067                 mutex_unlock(&b->write_lock);
1068                 pr_debug("bnode %p journal_flush set, retry\n", b);
1069                 udelay(1);
1070                 goto retry;
1071         }
1072
1073         if (btree_node_dirty(b)) {
1074                 btree_complete_write(b, btree_current_write(b));
1075                 clear_bit(BTREE_NODE_dirty, &b->flags);
1076         }
1077
1078         mutex_unlock(&b->write_lock);
1079
1080         cancel_delayed_work(&b->work);
1081
1082         mutex_lock(&b->c->bucket_lock);
1083         bch_bucket_free(b->c, &b->key);
1084         mca_bucket_free(b);
1085         mutex_unlock(&b->c->bucket_lock);
1086 }
1087
1088 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1089                                      int level, bool wait,
1090                                      struct btree *parent)
1091 {
1092         BKEY_PADDED(key) k;
1093         struct btree *b;
1094
1095         mutex_lock(&c->bucket_lock);
1096 retry:
1097         /* return ERR_PTR(-EAGAIN) when it fails */
1098         b = ERR_PTR(-EAGAIN);
1099         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1100                 goto err;
1101
1102         bkey_put(c, &k.key);
1103         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1104
1105         b = mca_alloc(c, op, &k.key, level);
1106         if (IS_ERR(b))
1107                 goto err_free;
1108
1109         if (!b) {
1110                 cache_bug(c,
1111                         "Tried to allocate bucket that was in btree cache");
1112                 goto retry;
1113         }
1114
1115         b->parent = parent;
1116         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->cache->sb));
1117
1118         mutex_unlock(&c->bucket_lock);
1119
1120         trace_bcache_btree_node_alloc(b);
1121         return b;
1122 err_free:
1123         bch_bucket_free(c, &k.key);
1124 err:
1125         mutex_unlock(&c->bucket_lock);
1126
1127         trace_bcache_btree_node_alloc_fail(c);
1128         return b;
1129 }
1130
1131 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1132                                           struct btree_op *op, int level,
1133                                           struct btree *parent)
1134 {
1135         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1136 }
1137
1138 static struct btree *btree_node_alloc_replacement(struct btree *b,
1139                                                   struct btree_op *op)
1140 {
1141         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1142
1143         if (!IS_ERR(n)) {
1144                 mutex_lock(&n->write_lock);
1145                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1146                 bkey_copy_key(&n->key, &b->key);
1147                 mutex_unlock(&n->write_lock);
1148         }
1149
1150         return n;
1151 }
1152
1153 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1154 {
1155         unsigned int i;
1156
1157         mutex_lock(&b->c->bucket_lock);
1158
1159         atomic_inc(&b->c->prio_blocked);
1160
1161         bkey_copy(k, &b->key);
1162         bkey_copy_key(k, &ZERO_KEY);
1163
1164         for (i = 0; i < KEY_PTRS(k); i++)
1165                 SET_PTR_GEN(k, i,
1166                             bch_inc_gen(b->c->cache,
1167                                         PTR_BUCKET(b->c, &b->key, i)));
1168
1169         mutex_unlock(&b->c->bucket_lock);
1170 }
1171
1172 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1173 {
1174         struct cache_set *c = b->c;
1175         struct cache *ca = c->cache;
1176         unsigned int reserve = (c->root->level - b->level) * 2 + 1;
1177
1178         mutex_lock(&c->bucket_lock);
1179
1180         if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1181                 if (op)
1182                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
1183                                         TASK_UNINTERRUPTIBLE);
1184                 mutex_unlock(&c->bucket_lock);
1185                 return -EINTR;
1186         }
1187
1188         mutex_unlock(&c->bucket_lock);
1189
1190         return mca_cannibalize_lock(b->c, op);
1191 }
1192
1193 /* Garbage collection */
1194
1195 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1196                                     struct bkey *k)
1197 {
1198         uint8_t stale = 0;
1199         unsigned int i;
1200         struct bucket *g;
1201
1202         /*
1203          * ptr_invalid() can't return true for the keys that mark btree nodes as
1204          * freed, but since ptr_bad() returns true we'll never actually use them
1205          * for anything and thus we don't want mark their pointers here
1206          */
1207         if (!bkey_cmp(k, &ZERO_KEY))
1208                 return stale;
1209
1210         for (i = 0; i < KEY_PTRS(k); i++) {
1211                 if (!ptr_available(c, k, i))
1212                         continue;
1213
1214                 g = PTR_BUCKET(c, k, i);
1215
1216                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1217                         g->last_gc = PTR_GEN(k, i);
1218
1219                 if (ptr_stale(c, k, i)) {
1220                         stale = max(stale, ptr_stale(c, k, i));
1221                         continue;
1222                 }
1223
1224                 cache_bug_on(GC_MARK(g) &&
1225                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1226                              c, "inconsistent ptrs: mark = %llu, level = %i",
1227                              GC_MARK(g), level);
1228
1229                 if (level)
1230                         SET_GC_MARK(g, GC_MARK_METADATA);
1231                 else if (KEY_DIRTY(k))
1232                         SET_GC_MARK(g, GC_MARK_DIRTY);
1233                 else if (!GC_MARK(g))
1234                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1235
1236                 /* guard against overflow */
1237                 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1238                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1239                                              MAX_GC_SECTORS_USED));
1240
1241                 BUG_ON(!GC_SECTORS_USED(g));
1242         }
1243
1244         return stale;
1245 }
1246
1247 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1248
1249 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1250 {
1251         unsigned int i;
1252
1253         for (i = 0; i < KEY_PTRS(k); i++)
1254                 if (ptr_available(c, k, i) &&
1255                     !ptr_stale(c, k, i)) {
1256                         struct bucket *b = PTR_BUCKET(c, k, i);
1257
1258                         b->gen = PTR_GEN(k, i);
1259
1260                         if (level && bkey_cmp(k, &ZERO_KEY))
1261                                 b->prio = BTREE_PRIO;
1262                         else if (!level && b->prio == BTREE_PRIO)
1263                                 b->prio = INITIAL_PRIO;
1264                 }
1265
1266         __bch_btree_mark_key(c, level, k);
1267 }
1268
1269 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1270 {
1271         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1272 }
1273
1274 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1275 {
1276         uint8_t stale = 0;
1277         unsigned int keys = 0, good_keys = 0;
1278         struct bkey *k;
1279         struct btree_iter iter;
1280         struct bset_tree *t;
1281
1282         gc->nodes++;
1283
1284         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1285                 stale = max(stale, btree_mark_key(b, k));
1286                 keys++;
1287
1288                 if (bch_ptr_bad(&b->keys, k))
1289                         continue;
1290
1291                 gc->key_bytes += bkey_u64s(k);
1292                 gc->nkeys++;
1293                 good_keys++;
1294
1295                 gc->data += KEY_SIZE(k);
1296         }
1297
1298         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1299                 btree_bug_on(t->size &&
1300                              bset_written(&b->keys, t) &&
1301                              bkey_cmp(&b->key, &t->end) < 0,
1302                              b, "found short btree key in gc");
1303
1304         if (b->c->gc_always_rewrite)
1305                 return true;
1306
1307         if (stale > 10)
1308                 return true;
1309
1310         if ((keys - good_keys) * 2 > keys)
1311                 return true;
1312
1313         return false;
1314 }
1315
1316 #define GC_MERGE_NODES  4U
1317
1318 struct gc_merge_info {
1319         struct btree    *b;
1320         unsigned int    keys;
1321 };
1322
1323 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1324                                  struct keylist *insert_keys,
1325                                  atomic_t *journal_ref,
1326                                  struct bkey *replace_key);
1327
1328 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1329                              struct gc_stat *gc, struct gc_merge_info *r)
1330 {
1331         unsigned int i, nodes = 0, keys = 0, blocks;
1332         struct btree *new_nodes[GC_MERGE_NODES];
1333         struct keylist keylist;
1334         struct closure cl;
1335         struct bkey *k;
1336
1337         bch_keylist_init(&keylist);
1338
1339         if (btree_check_reserve(b, NULL))
1340                 return 0;
1341
1342         memset(new_nodes, 0, sizeof(new_nodes));
1343         closure_init_stack(&cl);
1344
1345         while (nodes < GC_MERGE_NODES && !IS_ERR(r[nodes].b))
1346                 keys += r[nodes++].keys;
1347
1348         blocks = btree_default_blocks(b->c) * 2 / 3;
1349
1350         if (nodes < 2 ||
1351             __set_blocks(b->keys.set[0].data, keys,
1352                          block_bytes(b->c->cache)) > blocks * (nodes - 1))
1353                 return 0;
1354
1355         for (i = 0; i < nodes; i++) {
1356                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1357                 if (IS_ERR(new_nodes[i]))
1358                         goto out_nocoalesce;
1359         }
1360
1361         /*
1362          * We have to check the reserve here, after we've allocated our new
1363          * nodes, to make sure the insert below will succeed - we also check
1364          * before as an optimization to potentially avoid a bunch of expensive
1365          * allocs/sorts
1366          */
1367         if (btree_check_reserve(b, NULL))
1368                 goto out_nocoalesce;
1369
1370         for (i = 0; i < nodes; i++)
1371                 mutex_lock(&new_nodes[i]->write_lock);
1372
1373         for (i = nodes - 1; i > 0; --i) {
1374                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1375                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1376                 struct bkey *k, *last = NULL;
1377
1378                 keys = 0;
1379
1380                 if (i > 1) {
1381                         for (k = n2->start;
1382                              k < bset_bkey_last(n2);
1383                              k = bkey_next(k)) {
1384                                 if (__set_blocks(n1, n1->keys + keys +
1385                                                  bkey_u64s(k),
1386                                                  block_bytes(b->c->cache)) > blocks)
1387                                         break;
1388
1389                                 last = k;
1390                                 keys += bkey_u64s(k);
1391                         }
1392                 } else {
1393                         /*
1394                          * Last node we're not getting rid of - we're getting
1395                          * rid of the node at r[0]. Have to try and fit all of
1396                          * the remaining keys into this node; we can't ensure
1397                          * they will always fit due to rounding and variable
1398                          * length keys (shouldn't be possible in practice,
1399                          * though)
1400                          */
1401                         if (__set_blocks(n1, n1->keys + n2->keys,
1402                                          block_bytes(b->c->cache)) >
1403                             btree_blocks(new_nodes[i]))
1404                                 goto out_unlock_nocoalesce;
1405
1406                         keys = n2->keys;
1407                         /* Take the key of the node we're getting rid of */
1408                         last = &r->b->key;
1409                 }
1410
1411                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c->cache)) >
1412                        btree_blocks(new_nodes[i]));
1413
1414                 if (last)
1415                         bkey_copy_key(&new_nodes[i]->key, last);
1416
1417                 memcpy(bset_bkey_last(n1),
1418                        n2->start,
1419                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1420
1421                 n1->keys += keys;
1422                 r[i].keys = n1->keys;
1423
1424                 memmove(n2->start,
1425                         bset_bkey_idx(n2, keys),
1426                         (void *) bset_bkey_last(n2) -
1427                         (void *) bset_bkey_idx(n2, keys));
1428
1429                 n2->keys -= keys;
1430
1431                 if (__bch_keylist_realloc(&keylist,
1432                                           bkey_u64s(&new_nodes[i]->key)))
1433                         goto out_unlock_nocoalesce;
1434
1435                 bch_btree_node_write(new_nodes[i], &cl);
1436                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1437         }
1438
1439         for (i = 0; i < nodes; i++)
1440                 mutex_unlock(&new_nodes[i]->write_lock);
1441
1442         closure_sync(&cl);
1443
1444         /* We emptied out this node */
1445         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1446         btree_node_free(new_nodes[0]);
1447         rw_unlock(true, new_nodes[0]);
1448         new_nodes[0] = NULL;
1449
1450         for (i = 0; i < nodes; i++) {
1451                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1452                         goto out_nocoalesce;
1453
1454                 make_btree_freeing_key(r[i].b, keylist.top);
1455                 bch_keylist_push(&keylist);
1456         }
1457
1458         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1459         BUG_ON(!bch_keylist_empty(&keylist));
1460
1461         for (i = 0; i < nodes; i++) {
1462                 btree_node_free(r[i].b);
1463                 rw_unlock(true, r[i].b);
1464
1465                 r[i].b = new_nodes[i];
1466         }
1467
1468         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1469         r[nodes - 1].b = ERR_PTR(-EINTR);
1470
1471         trace_bcache_btree_gc_coalesce(nodes);
1472         gc->nodes--;
1473
1474         bch_keylist_free(&keylist);
1475
1476         /* Invalidated our iterator */
1477         return -EINTR;
1478
1479 out_unlock_nocoalesce:
1480         for (i = 0; i < nodes; i++)
1481                 mutex_unlock(&new_nodes[i]->write_lock);
1482
1483 out_nocoalesce:
1484         closure_sync(&cl);
1485
1486         while ((k = bch_keylist_pop(&keylist)))
1487                 if (!bkey_cmp(k, &ZERO_KEY))
1488                         atomic_dec(&b->c->prio_blocked);
1489         bch_keylist_free(&keylist);
1490
1491         for (i = 0; i < nodes; i++)
1492                 if (!IS_ERR(new_nodes[i])) {
1493                         btree_node_free(new_nodes[i]);
1494                         rw_unlock(true, new_nodes[i]);
1495                 }
1496         return 0;
1497 }
1498
1499 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1500                                  struct btree *replace)
1501 {
1502         struct keylist keys;
1503         struct btree *n;
1504
1505         if (btree_check_reserve(b, NULL))
1506                 return 0;
1507
1508         n = btree_node_alloc_replacement(replace, NULL);
1509
1510         /* recheck reserve after allocating replacement node */
1511         if (btree_check_reserve(b, NULL)) {
1512                 btree_node_free(n);
1513                 rw_unlock(true, n);
1514                 return 0;
1515         }
1516
1517         bch_btree_node_write_sync(n);
1518
1519         bch_keylist_init(&keys);
1520         bch_keylist_add(&keys, &n->key);
1521
1522         make_btree_freeing_key(replace, keys.top);
1523         bch_keylist_push(&keys);
1524
1525         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1526         BUG_ON(!bch_keylist_empty(&keys));
1527
1528         btree_node_free(replace);
1529         rw_unlock(true, n);
1530
1531         /* Invalidated our iterator */
1532         return -EINTR;
1533 }
1534
1535 static unsigned int btree_gc_count_keys(struct btree *b)
1536 {
1537         struct bkey *k;
1538         struct btree_iter iter;
1539         unsigned int ret = 0;
1540
1541         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1542                 ret += bkey_u64s(k);
1543
1544         return ret;
1545 }
1546
1547 static size_t btree_gc_min_nodes(struct cache_set *c)
1548 {
1549         size_t min_nodes;
1550
1551         /*
1552          * Since incremental GC would stop 100ms when front
1553          * side I/O comes, so when there are many btree nodes,
1554          * if GC only processes constant (100) nodes each time,
1555          * GC would last a long time, and the front side I/Os
1556          * would run out of the buckets (since no new bucket
1557          * can be allocated during GC), and be blocked again.
1558          * So GC should not process constant nodes, but varied
1559          * nodes according to the number of btree nodes, which
1560          * realized by dividing GC into constant(100) times,
1561          * so when there are many btree nodes, GC can process
1562          * more nodes each time, otherwise, GC will process less
1563          * nodes each time (but no less than MIN_GC_NODES)
1564          */
1565         min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1566         if (min_nodes < MIN_GC_NODES)
1567                 min_nodes = MIN_GC_NODES;
1568
1569         return min_nodes;
1570 }
1571
1572
1573 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1574                             struct closure *writes, struct gc_stat *gc)
1575 {
1576         int ret = 0;
1577         bool should_rewrite;
1578         struct bkey *k;
1579         struct btree_iter iter;
1580         struct gc_merge_info r[GC_MERGE_NODES];
1581         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1582
1583         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1584
1585         for (i = r; i < r + ARRAY_SIZE(r); i++)
1586                 i->b = ERR_PTR(-EINTR);
1587
1588         while (1) {
1589                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1590                 if (k) {
1591                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1592                                                   true, b);
1593                         if (IS_ERR(r->b)) {
1594                                 ret = PTR_ERR(r->b);
1595                                 break;
1596                         }
1597
1598                         r->keys = btree_gc_count_keys(r->b);
1599
1600                         ret = btree_gc_coalesce(b, op, gc, r);
1601                         if (ret)
1602                                 break;
1603                 }
1604
1605                 if (!last->b)
1606                         break;
1607
1608                 if (!IS_ERR(last->b)) {
1609                         should_rewrite = btree_gc_mark_node(last->b, gc);
1610                         if (should_rewrite) {
1611                                 ret = btree_gc_rewrite_node(b, op, last->b);
1612                                 if (ret)
1613                                         break;
1614                         }
1615
1616                         if (last->b->level) {
1617                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1618                                 if (ret)
1619                                         break;
1620                         }
1621
1622                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1623
1624                         /*
1625                          * Must flush leaf nodes before gc ends, since replace
1626                          * operations aren't journalled
1627                          */
1628                         mutex_lock(&last->b->write_lock);
1629                         if (btree_node_dirty(last->b))
1630                                 bch_btree_node_write(last->b, writes);
1631                         mutex_unlock(&last->b->write_lock);
1632                         rw_unlock(true, last->b);
1633                 }
1634
1635                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1636                 r->b = NULL;
1637
1638                 if (atomic_read(&b->c->search_inflight) &&
1639                     gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1640                         gc->nodes_pre =  gc->nodes;
1641                         ret = -EAGAIN;
1642                         break;
1643                 }
1644
1645                 if (need_resched()) {
1646                         ret = -EAGAIN;
1647                         break;
1648                 }
1649         }
1650
1651         for (i = r; i < r + ARRAY_SIZE(r); i++)
1652                 if (!IS_ERR_OR_NULL(i->b)) {
1653                         mutex_lock(&i->b->write_lock);
1654                         if (btree_node_dirty(i->b))
1655                                 bch_btree_node_write(i->b, writes);
1656                         mutex_unlock(&i->b->write_lock);
1657                         rw_unlock(true, i->b);
1658                 }
1659
1660         return ret;
1661 }
1662
1663 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1664                              struct closure *writes, struct gc_stat *gc)
1665 {
1666         struct btree *n = NULL;
1667         int ret = 0;
1668         bool should_rewrite;
1669
1670         should_rewrite = btree_gc_mark_node(b, gc);
1671         if (should_rewrite) {
1672                 n = btree_node_alloc_replacement(b, NULL);
1673
1674                 if (!IS_ERR(n)) {
1675                         bch_btree_node_write_sync(n);
1676
1677                         bch_btree_set_root(n);
1678                         btree_node_free(b);
1679                         rw_unlock(true, n);
1680
1681                         return -EINTR;
1682                 }
1683         }
1684
1685         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1686
1687         if (b->level) {
1688                 ret = btree_gc_recurse(b, op, writes, gc);
1689                 if (ret)
1690                         return ret;
1691         }
1692
1693         bkey_copy_key(&b->c->gc_done, &b->key);
1694
1695         return ret;
1696 }
1697
1698 static void btree_gc_start(struct cache_set *c)
1699 {
1700         struct cache *ca;
1701         struct bucket *b;
1702
1703         if (!c->gc_mark_valid)
1704                 return;
1705
1706         mutex_lock(&c->bucket_lock);
1707
1708         c->gc_mark_valid = 0;
1709         c->gc_done = ZERO_KEY;
1710
1711         ca = c->cache;
1712         for_each_bucket(b, ca) {
1713                 b->last_gc = b->gen;
1714                 if (!atomic_read(&b->pin)) {
1715                         SET_GC_MARK(b, 0);
1716                         SET_GC_SECTORS_USED(b, 0);
1717                 }
1718         }
1719
1720         mutex_unlock(&c->bucket_lock);
1721 }
1722
1723 static void bch_btree_gc_finish(struct cache_set *c)
1724 {
1725         struct bucket *b;
1726         struct cache *ca;
1727         unsigned int i, j;
1728         uint64_t *k;
1729
1730         mutex_lock(&c->bucket_lock);
1731
1732         set_gc_sectors(c);
1733         c->gc_mark_valid = 1;
1734         c->need_gc      = 0;
1735
1736         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1737                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1738                             GC_MARK_METADATA);
1739
1740         /* don't reclaim buckets to which writeback keys point */
1741         rcu_read_lock();
1742         for (i = 0; i < c->devices_max_used; i++) {
1743                 struct bcache_device *d = c->devices[i];
1744                 struct cached_dev *dc;
1745                 struct keybuf_key *w, *n;
1746
1747                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1748                         continue;
1749                 dc = container_of(d, struct cached_dev, disk);
1750
1751                 spin_lock(&dc->writeback_keys.lock);
1752                 rbtree_postorder_for_each_entry_safe(w, n,
1753                                         &dc->writeback_keys.keys, node)
1754                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1755                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1756                                             GC_MARK_DIRTY);
1757                 spin_unlock(&dc->writeback_keys.lock);
1758         }
1759         rcu_read_unlock();
1760
1761         c->avail_nbuckets = 0;
1762
1763         ca = c->cache;
1764         ca->invalidate_needs_gc = 0;
1765
1766         for (k = ca->sb.d; k < ca->sb.d + ca->sb.keys; k++)
1767                 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1768
1769         for (k = ca->prio_buckets;
1770              k < ca->prio_buckets + prio_buckets(ca) * 2; k++)
1771                 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1772
1773         for_each_bucket(b, ca) {
1774                 c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1775
1776                 if (atomic_read(&b->pin))
1777                         continue;
1778
1779                 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1780
1781                 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1782                         c->avail_nbuckets++;
1783         }
1784
1785         mutex_unlock(&c->bucket_lock);
1786 }
1787
1788 static void bch_btree_gc(struct cache_set *c)
1789 {
1790         int ret;
1791         struct gc_stat stats;
1792         struct closure writes;
1793         struct btree_op op;
1794         uint64_t start_time = local_clock();
1795
1796         trace_bcache_gc_start(c);
1797
1798         memset(&stats, 0, sizeof(struct gc_stat));
1799         closure_init_stack(&writes);
1800         bch_btree_op_init(&op, SHRT_MAX);
1801
1802         btree_gc_start(c);
1803
1804         /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1805         do {
1806                 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
1807                 closure_sync(&writes);
1808                 cond_resched();
1809
1810                 if (ret == -EAGAIN)
1811                         schedule_timeout_interruptible(msecs_to_jiffies
1812                                                        (GC_SLEEP_MS));
1813                 else if (ret)
1814                         pr_warn("gc failed!\n");
1815         } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1816
1817         bch_btree_gc_finish(c);
1818         wake_up_allocators(c);
1819
1820         bch_time_stats_update(&c->btree_gc_time, start_time);
1821
1822         stats.key_bytes *= sizeof(uint64_t);
1823         stats.data      <<= 9;
1824         bch_update_bucket_in_use(c, &stats);
1825         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1826
1827         trace_bcache_gc_end(c);
1828
1829         bch_moving_gc(c);
1830 }
1831
1832 static bool gc_should_run(struct cache_set *c)
1833 {
1834         struct cache *ca = c->cache;
1835
1836         if (ca->invalidate_needs_gc)
1837                 return true;
1838
1839         if (atomic_read(&c->sectors_to_gc) < 0)
1840                 return true;
1841
1842         return false;
1843 }
1844
1845 static int bch_gc_thread(void *arg)
1846 {
1847         struct cache_set *c = arg;
1848
1849         while (1) {
1850                 wait_event_interruptible(c->gc_wait,
1851                            kthread_should_stop() ||
1852                            test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1853                            gc_should_run(c));
1854
1855                 if (kthread_should_stop() ||
1856                     test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1857                         break;
1858
1859                 set_gc_sectors(c);
1860                 bch_btree_gc(c);
1861         }
1862
1863         wait_for_kthread_stop();
1864         return 0;
1865 }
1866
1867 int bch_gc_thread_start(struct cache_set *c)
1868 {
1869         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1870         return PTR_ERR_OR_ZERO(c->gc_thread);
1871 }
1872
1873 /* Initial partial gc */
1874
1875 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1876 {
1877         int ret = 0;
1878         struct bkey *k, *p = NULL;
1879         struct btree_iter iter;
1880
1881         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1882                 bch_initial_mark_key(b->c, b->level, k);
1883
1884         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1885
1886         if (b->level) {
1887                 bch_btree_iter_init(&b->keys, &iter, NULL);
1888
1889                 do {
1890                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1891                                                        bch_ptr_bad);
1892                         if (k) {
1893                                 btree_node_prefetch(b, k);
1894                                 /*
1895                                  * initiallize c->gc_stats.nodes
1896                                  * for incremental GC
1897                                  */
1898                                 b->c->gc_stats.nodes++;
1899                         }
1900
1901                         if (p)
1902                                 ret = bcache_btree(check_recurse, p, b, op);
1903
1904                         p = k;
1905                 } while (p && !ret);
1906         }
1907
1908         return ret;
1909 }
1910
1911
1912 static int bch_btree_check_thread(void *arg)
1913 {
1914         int ret;
1915         struct btree_check_info *info = arg;
1916         struct btree_check_state *check_state = info->state;
1917         struct cache_set *c = check_state->c;
1918         struct btree_iter iter;
1919         struct bkey *k, *p;
1920         int cur_idx, prev_idx, skip_nr;
1921
1922         k = p = NULL;
1923         cur_idx = prev_idx = 0;
1924         ret = 0;
1925
1926         /* root node keys are checked before thread created */
1927         bch_btree_iter_init(&c->root->keys, &iter, NULL);
1928         k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
1929         BUG_ON(!k);
1930
1931         p = k;
1932         while (k) {
1933                 /*
1934                  * Fetch a root node key index, skip the keys which
1935                  * should be fetched by other threads, then check the
1936                  * sub-tree indexed by the fetched key.
1937                  */
1938                 spin_lock(&check_state->idx_lock);
1939                 cur_idx = check_state->key_idx;
1940                 check_state->key_idx++;
1941                 spin_unlock(&check_state->idx_lock);
1942
1943                 skip_nr = cur_idx - prev_idx;
1944
1945                 while (skip_nr) {
1946                         k = bch_btree_iter_next_filter(&iter,
1947                                                        &c->root->keys,
1948                                                        bch_ptr_bad);
1949                         if (k)
1950                                 p = k;
1951                         else {
1952                                 /*
1953                                  * No more keys to check in root node,
1954                                  * current checking threads are enough,
1955                                  * stop creating more.
1956                                  */
1957                                 atomic_set(&check_state->enough, 1);
1958                                 /* Update check_state->enough earlier */
1959                                 smp_mb__after_atomic();
1960                                 goto out;
1961                         }
1962                         skip_nr--;
1963                         cond_resched();
1964                 }
1965
1966                 if (p) {
1967                         struct btree_op op;
1968
1969                         btree_node_prefetch(c->root, p);
1970                         c->gc_stats.nodes++;
1971                         bch_btree_op_init(&op, 0);
1972                         ret = bcache_btree(check_recurse, p, c->root, &op);
1973                         /*
1974                          * The op may be added to cache_set's btree_cache_wait
1975                          * in mca_cannibalize(), must ensure it is removed from
1976                          * the list and release btree_cache_alloc_lock before
1977                          * free op memory.
1978                          * Otherwise, the btree_cache_wait will be damaged.
1979                          */
1980                         bch_cannibalize_unlock(c);
1981                         finish_wait(&c->btree_cache_wait, &(&op)->wait);
1982                         if (ret)
1983                                 goto out;
1984                 }
1985                 p = NULL;
1986                 prev_idx = cur_idx;
1987                 cond_resched();
1988         }
1989
1990 out:
1991         info->result = ret;
1992         /* update check_state->started among all CPUs */
1993         smp_mb__before_atomic();
1994         if (atomic_dec_and_test(&check_state->started))
1995                 wake_up(&check_state->wait);
1996
1997         return ret;
1998 }
1999
2000
2001
2002 static int bch_btree_chkthread_nr(void)
2003 {
2004         int n = num_online_cpus()/2;
2005
2006         if (n == 0)
2007                 n = 1;
2008         else if (n > BCH_BTR_CHKTHREAD_MAX)
2009                 n = BCH_BTR_CHKTHREAD_MAX;
2010
2011         return n;
2012 }
2013
2014 int bch_btree_check(struct cache_set *c)
2015 {
2016         int ret = 0;
2017         int i;
2018         struct bkey *k = NULL;
2019         struct btree_iter iter;
2020         struct btree_check_state check_state;
2021
2022         /* check and mark root node keys */
2023         for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2024                 bch_initial_mark_key(c, c->root->level, k);
2025
2026         bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2027
2028         if (c->root->level == 0)
2029                 return 0;
2030
2031         memset(&check_state, 0, sizeof(struct btree_check_state));
2032         check_state.c = c;
2033         check_state.total_threads = bch_btree_chkthread_nr();
2034         check_state.key_idx = 0;
2035         spin_lock_init(&check_state.idx_lock);
2036         atomic_set(&check_state.started, 0);
2037         atomic_set(&check_state.enough, 0);
2038         init_waitqueue_head(&check_state.wait);
2039
2040         rw_lock(0, c->root, c->root->level);
2041         /*
2042          * Run multiple threads to check btree nodes in parallel,
2043          * if check_state.enough is non-zero, it means current
2044          * running check threads are enough, unncessary to create
2045          * more.
2046          */
2047         for (i = 0; i < check_state.total_threads; i++) {
2048                 /* fetch latest check_state.enough earlier */
2049                 smp_mb__before_atomic();
2050                 if (atomic_read(&check_state.enough))
2051                         break;
2052
2053                 check_state.infos[i].result = 0;
2054                 check_state.infos[i].state = &check_state;
2055
2056                 check_state.infos[i].thread =
2057                         kthread_run(bch_btree_check_thread,
2058                                     &check_state.infos[i],
2059                                     "bch_btrchk[%d]", i);
2060                 if (IS_ERR(check_state.infos[i].thread)) {
2061                         pr_err("fails to run thread bch_btrchk[%d]\n", i);
2062                         for (--i; i >= 0; i--)
2063                                 kthread_stop(check_state.infos[i].thread);
2064                         ret = -ENOMEM;
2065                         goto out;
2066                 }
2067                 atomic_inc(&check_state.started);
2068         }
2069
2070         /*
2071          * Must wait for all threads to stop.
2072          */
2073         wait_event(check_state.wait, atomic_read(&check_state.started) == 0);
2074
2075         for (i = 0; i < check_state.total_threads; i++) {
2076                 if (check_state.infos[i].result) {
2077                         ret = check_state.infos[i].result;
2078                         goto out;
2079                 }
2080         }
2081
2082 out:
2083         rw_unlock(0, c->root);
2084         return ret;
2085 }
2086
2087 void bch_initial_gc_finish(struct cache_set *c)
2088 {
2089         struct cache *ca = c->cache;
2090         struct bucket *b;
2091
2092         bch_btree_gc_finish(c);
2093
2094         mutex_lock(&c->bucket_lock);
2095
2096         /*
2097          * We need to put some unused buckets directly on the prio freelist in
2098          * order to get the allocator thread started - it needs freed buckets in
2099          * order to rewrite the prios and gens, and it needs to rewrite prios
2100          * and gens in order to free buckets.
2101          *
2102          * This is only safe for buckets that have no live data in them, which
2103          * there should always be some of.
2104          */
2105         for_each_bucket(b, ca) {
2106                 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2107                     fifo_full(&ca->free[RESERVE_BTREE]))
2108                         break;
2109
2110                 if (bch_can_invalidate_bucket(ca, b) &&
2111                     !GC_MARK(b)) {
2112                         __bch_invalidate_one_bucket(ca, b);
2113                         if (!fifo_push(&ca->free[RESERVE_PRIO],
2114                            b - ca->buckets))
2115                                 fifo_push(&ca->free[RESERVE_BTREE],
2116                                           b - ca->buckets);
2117                 }
2118         }
2119
2120         mutex_unlock(&c->bucket_lock);
2121 }
2122
2123 /* Btree insertion */
2124
2125 static bool btree_insert_key(struct btree *b, struct bkey *k,
2126                              struct bkey *replace_key)
2127 {
2128         unsigned int status;
2129
2130         BUG_ON(bkey_cmp(k, &b->key) > 0);
2131
2132         status = bch_btree_insert_key(&b->keys, k, replace_key);
2133         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2134                 bch_check_keys(&b->keys, "%u for %s", status,
2135                                replace_key ? "replace" : "insert");
2136
2137                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2138                                               status);
2139                 return true;
2140         } else
2141                 return false;
2142 }
2143
2144 static size_t insert_u64s_remaining(struct btree *b)
2145 {
2146         long ret = bch_btree_keys_u64s_remaining(&b->keys);
2147
2148         /*
2149          * Might land in the middle of an existing extent and have to split it
2150          */
2151         if (b->keys.ops->is_extents)
2152                 ret -= KEY_MAX_U64S;
2153
2154         return max(ret, 0L);
2155 }
2156
2157 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2158                                   struct keylist *insert_keys,
2159                                   struct bkey *replace_key)
2160 {
2161         bool ret = false;
2162         int oldsize = bch_count_data(&b->keys);
2163
2164         while (!bch_keylist_empty(insert_keys)) {
2165                 struct bkey *k = insert_keys->keys;
2166
2167                 if (bkey_u64s(k) > insert_u64s_remaining(b))
2168                         break;
2169
2170                 if (bkey_cmp(k, &b->key) <= 0) {
2171                         if (!b->level)
2172                                 bkey_put(b->c, k);
2173
2174                         ret |= btree_insert_key(b, k, replace_key);
2175                         bch_keylist_pop_front(insert_keys);
2176                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2177                         BKEY_PADDED(key) temp;
2178                         bkey_copy(&temp.key, insert_keys->keys);
2179
2180                         bch_cut_back(&b->key, &temp.key);
2181                         bch_cut_front(&b->key, insert_keys->keys);
2182
2183                         ret |= btree_insert_key(b, &temp.key, replace_key);
2184                         break;
2185                 } else {
2186                         break;
2187                 }
2188         }
2189
2190         if (!ret)
2191                 op->insert_collision = true;
2192
2193         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2194
2195         BUG_ON(bch_count_data(&b->keys) < oldsize);
2196         return ret;
2197 }
2198
2199 static int btree_split(struct btree *b, struct btree_op *op,
2200                        struct keylist *insert_keys,
2201                        struct bkey *replace_key)
2202 {
2203         bool split;
2204         struct btree *n1, *n2 = NULL, *n3 = NULL;
2205         uint64_t start_time = local_clock();
2206         struct closure cl;
2207         struct keylist parent_keys;
2208
2209         closure_init_stack(&cl);
2210         bch_keylist_init(&parent_keys);
2211
2212         if (btree_check_reserve(b, op)) {
2213                 if (!b->level)
2214                         return -EINTR;
2215                 else
2216                         WARN(1, "insufficient reserve for split\n");
2217         }
2218
2219         n1 = btree_node_alloc_replacement(b, op);
2220         if (IS_ERR(n1))
2221                 goto err;
2222
2223         split = set_blocks(btree_bset_first(n1),
2224                            block_bytes(n1->c->cache)) > (btree_blocks(b) * 4) / 5;
2225
2226         if (split) {
2227                 unsigned int keys = 0;
2228
2229                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2230
2231                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2232                 if (IS_ERR(n2))
2233                         goto err_free1;
2234
2235                 if (!b->parent) {
2236                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2237                         if (IS_ERR(n3))
2238                                 goto err_free2;
2239                 }
2240
2241                 mutex_lock(&n1->write_lock);
2242                 mutex_lock(&n2->write_lock);
2243
2244                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2245
2246                 /*
2247                  * Has to be a linear search because we don't have an auxiliary
2248                  * search tree yet
2249                  */
2250
2251                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2252                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2253                                                         keys));
2254
2255                 bkey_copy_key(&n1->key,
2256                               bset_bkey_idx(btree_bset_first(n1), keys));
2257                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2258
2259                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2260                 btree_bset_first(n1)->keys = keys;
2261
2262                 memcpy(btree_bset_first(n2)->start,
2263                        bset_bkey_last(btree_bset_first(n1)),
2264                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2265
2266                 bkey_copy_key(&n2->key, &b->key);
2267
2268                 bch_keylist_add(&parent_keys, &n2->key);
2269                 bch_btree_node_write(n2, &cl);
2270                 mutex_unlock(&n2->write_lock);
2271                 rw_unlock(true, n2);
2272         } else {
2273                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2274
2275                 mutex_lock(&n1->write_lock);
2276                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2277         }
2278
2279         bch_keylist_add(&parent_keys, &n1->key);
2280         bch_btree_node_write(n1, &cl);
2281         mutex_unlock(&n1->write_lock);
2282
2283         if (n3) {
2284                 /* Depth increases, make a new root */
2285                 mutex_lock(&n3->write_lock);
2286                 bkey_copy_key(&n3->key, &MAX_KEY);
2287                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2288                 bch_btree_node_write(n3, &cl);
2289                 mutex_unlock(&n3->write_lock);
2290
2291                 closure_sync(&cl);
2292                 bch_btree_set_root(n3);
2293                 rw_unlock(true, n3);
2294         } else if (!b->parent) {
2295                 /* Root filled up but didn't need to be split */
2296                 closure_sync(&cl);
2297                 bch_btree_set_root(n1);
2298         } else {
2299                 /* Split a non root node */
2300                 closure_sync(&cl);
2301                 make_btree_freeing_key(b, parent_keys.top);
2302                 bch_keylist_push(&parent_keys);
2303
2304                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2305                 BUG_ON(!bch_keylist_empty(&parent_keys));
2306         }
2307
2308         btree_node_free(b);
2309         rw_unlock(true, n1);
2310
2311         bch_time_stats_update(&b->c->btree_split_time, start_time);
2312
2313         return 0;
2314 err_free2:
2315         bkey_put(b->c, &n2->key);
2316         btree_node_free(n2);
2317         rw_unlock(true, n2);
2318 err_free1:
2319         bkey_put(b->c, &n1->key);
2320         btree_node_free(n1);
2321         rw_unlock(true, n1);
2322 err:
2323         WARN(1, "bcache: btree split failed (level %u)", b->level);
2324
2325         if (n3 == ERR_PTR(-EAGAIN) ||
2326             n2 == ERR_PTR(-EAGAIN) ||
2327             n1 == ERR_PTR(-EAGAIN))
2328                 return -EAGAIN;
2329
2330         return -ENOMEM;
2331 }
2332
2333 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2334                                  struct keylist *insert_keys,
2335                                  atomic_t *journal_ref,
2336                                  struct bkey *replace_key)
2337 {
2338         struct closure cl;
2339
2340         BUG_ON(b->level && replace_key);
2341
2342         closure_init_stack(&cl);
2343
2344         mutex_lock(&b->write_lock);
2345
2346         if (write_block(b) != btree_bset_last(b) &&
2347             b->keys.last_set_unwritten)
2348                 bch_btree_init_next(b); /* just wrote a set */
2349
2350         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2351                 mutex_unlock(&b->write_lock);
2352                 goto split;
2353         }
2354
2355         BUG_ON(write_block(b) != btree_bset_last(b));
2356
2357         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2358                 if (!b->level)
2359                         bch_btree_leaf_dirty(b, journal_ref);
2360                 else
2361                         bch_btree_node_write(b, &cl);
2362         }
2363
2364         mutex_unlock(&b->write_lock);
2365
2366         /* wait for btree node write if necessary, after unlock */
2367         closure_sync(&cl);
2368
2369         return 0;
2370 split:
2371         if (current->bio_list) {
2372                 op->lock = b->c->root->level + 1;
2373                 return -EAGAIN;
2374         } else if (op->lock <= b->c->root->level) {
2375                 op->lock = b->c->root->level + 1;
2376                 return -EINTR;
2377         } else {
2378                 /* Invalidated all iterators */
2379                 int ret = btree_split(b, op, insert_keys, replace_key);
2380
2381                 if (bch_keylist_empty(insert_keys))
2382                         return 0;
2383                 else if (!ret)
2384                         return -EINTR;
2385                 return ret;
2386         }
2387 }
2388
2389 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2390                                struct bkey *check_key)
2391 {
2392         int ret = -EINTR;
2393         uint64_t btree_ptr = b->key.ptr[0];
2394         unsigned long seq = b->seq;
2395         struct keylist insert;
2396         bool upgrade = op->lock == -1;
2397
2398         bch_keylist_init(&insert);
2399
2400         if (upgrade) {
2401                 rw_unlock(false, b);
2402                 rw_lock(true, b, b->level);
2403
2404                 if (b->key.ptr[0] != btree_ptr ||
2405                     b->seq != seq + 1) {
2406                         op->lock = b->level;
2407                         goto out;
2408                 }
2409         }
2410
2411         SET_KEY_PTRS(check_key, 1);
2412         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2413
2414         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2415
2416         bch_keylist_add(&insert, check_key);
2417
2418         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2419
2420         BUG_ON(!ret && !bch_keylist_empty(&insert));
2421 out:
2422         if (upgrade)
2423                 downgrade_write(&b->lock);
2424         return ret;
2425 }
2426
2427 struct btree_insert_op {
2428         struct btree_op op;
2429         struct keylist  *keys;
2430         atomic_t        *journal_ref;
2431         struct bkey     *replace_key;
2432 };
2433
2434 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2435 {
2436         struct btree_insert_op *op = container_of(b_op,
2437                                         struct btree_insert_op, op);
2438
2439         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2440                                         op->journal_ref, op->replace_key);
2441         if (ret && !bch_keylist_empty(op->keys))
2442                 return ret;
2443         else
2444                 return MAP_DONE;
2445 }
2446
2447 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2448                      atomic_t *journal_ref, struct bkey *replace_key)
2449 {
2450         struct btree_insert_op op;
2451         int ret = 0;
2452
2453         BUG_ON(current->bio_list);
2454         BUG_ON(bch_keylist_empty(keys));
2455
2456         bch_btree_op_init(&op.op, 0);
2457         op.keys         = keys;
2458         op.journal_ref  = journal_ref;
2459         op.replace_key  = replace_key;
2460
2461         while (!ret && !bch_keylist_empty(keys)) {
2462                 op.op.lock = 0;
2463                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2464                                                &START_KEY(keys->keys),
2465                                                btree_insert_fn);
2466         }
2467
2468         if (ret) {
2469                 struct bkey *k;
2470
2471                 pr_err("error %i\n", ret);
2472
2473                 while ((k = bch_keylist_pop(keys)))
2474                         bkey_put(c, k);
2475         } else if (op.op.insert_collision)
2476                 ret = -ESRCH;
2477
2478         return ret;
2479 }
2480
2481 void bch_btree_set_root(struct btree *b)
2482 {
2483         unsigned int i;
2484         struct closure cl;
2485
2486         closure_init_stack(&cl);
2487
2488         trace_bcache_btree_set_root(b);
2489
2490         BUG_ON(!b->written);
2491
2492         for (i = 0; i < KEY_PTRS(&b->key); i++)
2493                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2494
2495         mutex_lock(&b->c->bucket_lock);
2496         list_del_init(&b->list);
2497         mutex_unlock(&b->c->bucket_lock);
2498
2499         b->c->root = b;
2500
2501         bch_journal_meta(b->c, &cl);
2502         closure_sync(&cl);
2503 }
2504
2505 /* Map across nodes or keys */
2506
2507 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2508                                        struct bkey *from,
2509                                        btree_map_nodes_fn *fn, int flags)
2510 {
2511         int ret = MAP_CONTINUE;
2512
2513         if (b->level) {
2514                 struct bkey *k;
2515                 struct btree_iter iter;
2516
2517                 bch_btree_iter_init(&b->keys, &iter, from);
2518
2519                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2520                                                        bch_ptr_bad))) {
2521                         ret = bcache_btree(map_nodes_recurse, k, b,
2522                                     op, from, fn, flags);
2523                         from = NULL;
2524
2525                         if (ret != MAP_CONTINUE)
2526                                 return ret;
2527                 }
2528         }
2529
2530         if (!b->level || flags == MAP_ALL_NODES)
2531                 ret = fn(op, b);
2532
2533         return ret;
2534 }
2535
2536 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2537                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2538 {
2539         return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
2540 }
2541
2542 int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2543                                       struct bkey *from, btree_map_keys_fn *fn,
2544                                       int flags)
2545 {
2546         int ret = MAP_CONTINUE;
2547         struct bkey *k;
2548         struct btree_iter iter;
2549
2550         bch_btree_iter_init(&b->keys, &iter, from);
2551
2552         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2553                 ret = !b->level
2554                         ? fn(op, b, k)
2555                         : bcache_btree(map_keys_recurse, k,
2556                                        b, op, from, fn, flags);
2557                 from = NULL;
2558
2559                 if (ret != MAP_CONTINUE)
2560                         return ret;
2561         }
2562
2563         if (!b->level && (flags & MAP_END_KEY))
2564                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2565                                      KEY_OFFSET(&b->key), 0));
2566
2567         return ret;
2568 }
2569
2570 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2571                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2572 {
2573         return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
2574 }
2575
2576 /* Keybuf code */
2577
2578 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2579 {
2580         /* Overlapping keys compare equal */
2581         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2582                 return -1;
2583         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2584                 return 1;
2585         return 0;
2586 }
2587
2588 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2589                                             struct keybuf_key *r)
2590 {
2591         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2592 }
2593
2594 struct refill {
2595         struct btree_op op;
2596         unsigned int    nr_found;
2597         struct keybuf   *buf;
2598         struct bkey     *end;
2599         keybuf_pred_fn  *pred;
2600 };
2601
2602 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2603                             struct bkey *k)
2604 {
2605         struct refill *refill = container_of(op, struct refill, op);
2606         struct keybuf *buf = refill->buf;
2607         int ret = MAP_CONTINUE;
2608
2609         if (bkey_cmp(k, refill->end) > 0) {
2610                 ret = MAP_DONE;
2611                 goto out;
2612         }
2613
2614         if (!KEY_SIZE(k)) /* end key */
2615                 goto out;
2616
2617         if (refill->pred(buf, k)) {
2618                 struct keybuf_key *w;
2619
2620                 spin_lock(&buf->lock);
2621
2622                 w = array_alloc(&buf->freelist);
2623                 if (!w) {
2624                         spin_unlock(&buf->lock);
2625                         return MAP_DONE;
2626                 }
2627
2628                 w->private = NULL;
2629                 bkey_copy(&w->key, k);
2630
2631                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2632                         array_free(&buf->freelist, w);
2633                 else
2634                         refill->nr_found++;
2635
2636                 if (array_freelist_empty(&buf->freelist))
2637                         ret = MAP_DONE;
2638
2639                 spin_unlock(&buf->lock);
2640         }
2641 out:
2642         buf->last_scanned = *k;
2643         return ret;
2644 }
2645
2646 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2647                        struct bkey *end, keybuf_pred_fn *pred)
2648 {
2649         struct bkey start = buf->last_scanned;
2650         struct refill refill;
2651
2652         cond_resched();
2653
2654         bch_btree_op_init(&refill.op, -1);
2655         refill.nr_found = 0;
2656         refill.buf      = buf;
2657         refill.end      = end;
2658         refill.pred     = pred;
2659
2660         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2661                            refill_keybuf_fn, MAP_END_KEY);
2662
2663         trace_bcache_keyscan(refill.nr_found,
2664                              KEY_INODE(&start), KEY_OFFSET(&start),
2665                              KEY_INODE(&buf->last_scanned),
2666                              KEY_OFFSET(&buf->last_scanned));
2667
2668         spin_lock(&buf->lock);
2669
2670         if (!RB_EMPTY_ROOT(&buf->keys)) {
2671                 struct keybuf_key *w;
2672
2673                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2674                 buf->start      = START_KEY(&w->key);
2675
2676                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2677                 buf->end        = w->key;
2678         } else {
2679                 buf->start      = MAX_KEY;
2680                 buf->end        = MAX_KEY;
2681         }
2682
2683         spin_unlock(&buf->lock);
2684 }
2685
2686 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2687 {
2688         rb_erase(&w->node, &buf->keys);
2689         array_free(&buf->freelist, w);
2690 }
2691
2692 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2693 {
2694         spin_lock(&buf->lock);
2695         __bch_keybuf_del(buf, w);
2696         spin_unlock(&buf->lock);
2697 }
2698
2699 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2700                                   struct bkey *end)
2701 {
2702         bool ret = false;
2703         struct keybuf_key *p, *w, s;
2704
2705         s.key = *start;
2706
2707         if (bkey_cmp(end, &buf->start) <= 0 ||
2708             bkey_cmp(start, &buf->end) >= 0)
2709                 return false;
2710
2711         spin_lock(&buf->lock);
2712         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2713
2714         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2715                 p = w;
2716                 w = RB_NEXT(w, node);
2717
2718                 if (p->private)
2719                         ret = true;
2720                 else
2721                         __bch_keybuf_del(buf, p);
2722         }
2723
2724         spin_unlock(&buf->lock);
2725         return ret;
2726 }
2727
2728 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2729 {
2730         struct keybuf_key *w;
2731
2732         spin_lock(&buf->lock);
2733
2734         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2735
2736         while (w && w->private)
2737                 w = RB_NEXT(w, node);
2738
2739         if (w)
2740                 w->private = ERR_PTR(-EINTR);
2741
2742         spin_unlock(&buf->lock);
2743         return w;
2744 }
2745
2746 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2747                                           struct keybuf *buf,
2748                                           struct bkey *end,
2749                                           keybuf_pred_fn *pred)
2750 {
2751         struct keybuf_key *ret;
2752
2753         while (1) {
2754                 ret = bch_keybuf_next(buf);
2755                 if (ret)
2756                         break;
2757
2758                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2759                         pr_debug("scan finished\n");
2760                         break;
2761                 }
2762
2763                 bch_refill_keybuf(c, buf, end, pred);
2764         }
2765
2766         return ret;
2767 }
2768
2769 void bch_keybuf_init(struct keybuf *buf)
2770 {
2771         buf->last_scanned       = MAX_KEY;
2772         buf->keys               = RB_ROOT;
2773
2774         spin_lock_init(&buf->lock);
2775         array_allocator_init(&buf->freelist);
2776 }
2777
2778 void bch_btree_exit(void)
2779 {
2780         if (btree_io_wq)
2781                 destroy_workqueue(btree_io_wq);
2782 }
2783
2784 int __init bch_btree_init(void)
2785 {
2786         btree_io_wq = alloc_workqueue("bch_btree_io", WQ_MEM_RECLAIM, 0);
2787         if (!btree_io_wq)
2788                 return -ENOMEM;
2789
2790         return 0;
2791 }