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 #define cmp_int(l, r)           ((l > r) - (l < r))
 563
 564 #ifdef CONFIG_PROVE_LOCKING
 565 static int btree_lock_cmp_fn(const struct lockdep_map *_a,
 566                              const struct lockdep_map *_b)
 567 {
 568         const struct btree *a = container_of(_a, struct btree, lock.dep_map);
 569         const struct btree *b = container_of(_b, struct btree, lock.dep_map);
 570
 571         return -cmp_int(a->level, b->level) ?: bkey_cmp(&a->key, &b->key);
 572 }
 573
 574 static void btree_lock_print_fn(const struct lockdep_map *map)
 575 {
 576         const struct btree *b = container_of(map, struct btree, lock.dep_map);
 577
 578         printk(KERN_CONT " l=%u %llu:%llu", b->level,
 579                KEY_INODE(&b->key), KEY_OFFSET(&b->key));
 580 }
 581 #endif
 582
 583 static struct btree *mca_bucket_alloc(struct cache_set *c,
 584                                       struct bkey *k, gfp_t gfp)
 585 {
 586         /*
 587          * kzalloc() is necessary here for initialization,
 588          * see code comments in bch_btree_keys_init().
 589          */
 590         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 591
 592         if (!b)
 593                 return NULL;
 594
 595         init_rwsem(&b->lock);
 596         lock_set_cmp_fn(&b->lock, btree_lock_cmp_fn, btree_lock_print_fn);
 597         mutex_init(&b->write_lock);
 598         lockdep_set_novalidate_class(&b->write_lock);
 599         INIT_LIST_HEAD(&b->list);
 600         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 601         b->c = c;
 602         sema_init(&b->io_mutex, 1);
 603
 604         mca_data_alloc(b, k, gfp);
 605         return b;
 606 }
 607
 608 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
 609 {
 610         struct closure cl;
 611
 612         closure_init_stack(&cl);
 613         lockdep_assert_held(&b->c->bucket_lock);
 614
 615         if (!down_write_trylock(&b->lock))
 616                 return -ENOMEM;
 617
 618         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 619
 620         if (b->keys.page_order < min_order)
 621                 goto out_unlock;
 622
 623         if (!flush) {
 624                 if (btree_node_dirty(b))
 625                         goto out_unlock;
 626
 627                 if (down_trylock(&b->io_mutex))
 628                         goto out_unlock;
 629                 up(&b->io_mutex);
 630         }
 631
 632 retry:
 633         /*
 634          * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
 635          * __bch_btree_node_write(). To avoid an extra flush, acquire
 636          * b->write_lock before checking BTREE_NODE_dirty bit.
 637          */
 638         mutex_lock(&b->write_lock);
 639         /*
 640          * If this btree node is selected in btree_flush_write() by journal
 641          * code, delay and retry until the node is flushed by journal code
 642          * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
 643          */
 644         if (btree_node_journal_flush(b)) {
 645                 pr_debug("bnode %p is flushing by journal, retry\n", b);
 646                 mutex_unlock(&b->write_lock);
 647                 udelay(1);
 648                 goto retry;
 649         }
 650
 651         if (btree_node_dirty(b))
 652                 __bch_btree_node_write(b, &cl);
 653         mutex_unlock(&b->write_lock);
 654
 655         closure_sync(&cl);
 656
 657         /* wait for any in flight btree write */
 658         down(&b->io_mutex);
 659         up(&b->io_mutex);
 660
 661         return 0;
 662 out_unlock:
 663         rw_unlock(true, b);
 664         return -ENOMEM;
 665 }
 666
 667 static unsigned long bch_mca_scan(struct shrinker *shrink,
 668                                   struct shrink_control *sc)
 669 {
 670         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 671         struct btree *b, *t;
 672         unsigned long i, nr = sc->nr_to_scan;
 673         unsigned long freed = 0;
 674         unsigned int btree_cache_used;
 675
 676         if (c->shrinker_disabled)
 677                 return SHRINK_STOP;
 678
 679         if (c->btree_cache_alloc_lock)
 680                 return SHRINK_STOP;
 681
 682         /* Return -1 if we can't do anything right now */
 683         if (sc->gfp_mask & __GFP_IO)
 684                 mutex_lock(&c->bucket_lock);
 685         else if (!mutex_trylock(&c->bucket_lock))
 686                 return -1;
 687
 688         /*
 689          * It's _really_ critical that we don't free too many btree nodes - we
 690          * have to always leave ourselves a reserve. The reserve is how we
 691          * guarantee that allocating memory for a new btree node can always
 692          * succeed, so that inserting keys into the btree can always succeed and
 693          * IO can always make forward progress:
 694          */
 695         nr /= c->btree_pages;
 696         if (nr == 0)
 697                 nr = 1;
 698         nr = min_t(unsigned long, nr, mca_can_free(c));
 699
 700         i = 0;
 701         btree_cache_used = c->btree_cache_used;
 702         list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
 703                 if (nr <= 0)
 704                         goto out;
 705
 706                 if (!mca_reap(b, 0, false)) {
 707                         mca_data_free(b);
 708                         rw_unlock(true, b);
 709                         freed++;
 710                 }
 711                 nr--;
 712                 i++;
 713         }
 714
 715         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 716                 if (nr <= 0 || i >= btree_cache_used)
 717                         goto out;
 718
 719                 if (!mca_reap(b, 0, false)) {
 720                         mca_bucket_free(b);
 721                         mca_data_free(b);
 722                         rw_unlock(true, b);
 723                         freed++;
 724                 }
 725
 726                 nr--;
 727                 i++;
 728         }
 729 out:
 730         mutex_unlock(&c->bucket_lock);
 731         return freed * c->btree_pages;
 732 }
 733
 734 static unsigned long bch_mca_count(struct shrinker *shrink,
 735                                    struct shrink_control *sc)
 736 {
 737         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 738
 739         if (c->shrinker_disabled)
 740                 return 0;
 741
 742         if (c->btree_cache_alloc_lock)
 743                 return 0;
 744
 745         return mca_can_free(c) * c->btree_pages;
 746 }
 747
 748 void bch_btree_cache_free(struct cache_set *c)
 749 {
 750         struct btree *b;
 751         struct closure cl;
 752
 753         closure_init_stack(&cl);
 754
 755         if (c->shrink.list.next)
 756                 unregister_shrinker(&c->shrink);
 757
 758         mutex_lock(&c->bucket_lock);
 759
 760 #ifdef CONFIG_BCACHE_DEBUG
 761         if (c->verify_data)
 762                 list_move(&c->verify_data->list, &c->btree_cache);
 763
 764         free_pages((unsigned long) c->verify_ondisk, ilog2(meta_bucket_pages(&c->cache->sb)));
 765 #endif
 766
 767         list_splice(&c->btree_cache_freeable,
 768                     &c->btree_cache);
 769
 770         while (!list_empty(&c->btree_cache)) {
 771                 b = list_first_entry(&c->btree_cache, struct btree, list);
 772
 773                 /*
 774                  * This function is called by cache_set_free(), no I/O
 775                  * request on cache now, it is unnecessary to acquire
 776                  * b->write_lock before clearing BTREE_NODE_dirty anymore.
 777                  */
 778                 if (btree_node_dirty(b)) {
 779                         btree_complete_write(b, btree_current_write(b));
 780                         clear_bit(BTREE_NODE_dirty, &b->flags);
 781                 }
 782                 mca_data_free(b);
 783         }
 784
 785         while (!list_empty(&c->btree_cache_freed)) {
 786                 b = list_first_entry(&c->btree_cache_freed,
 787                                      struct btree, list);
 788                 list_del(&b->list);
 789                 cancel_delayed_work_sync(&b->work);
 790                 kfree(b);
 791         }
 792
 793         mutex_unlock(&c->bucket_lock);
 794 }
 795
 796 int bch_btree_cache_alloc(struct cache_set *c)
 797 {
 798         unsigned int i;
 799
 800         for (i = 0; i < mca_reserve(c); i++)
 801                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 802                         return -ENOMEM;
 803
 804         list_splice_init(&c->btree_cache,
 805                          &c->btree_cache_freeable);
 806
 807 #ifdef CONFIG_BCACHE_DEBUG
 808         mutex_init(&c->verify_lock);
 809
 810         c->verify_ondisk = (void *)
 811                 __get_free_pages(GFP_KERNEL|__GFP_COMP,
 812                                  ilog2(meta_bucket_pages(&c->cache->sb)));
 813         if (!c->verify_ondisk) {
 814                 /*
 815                  * Don't worry about the mca_rereserve buckets
 816                  * allocated in previous for-loop, they will be
 817                  * handled properly in bch_cache_set_unregister().
 818                  */
 819                 return -ENOMEM;
 820         }
 821
 822         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 823
 824         if (c->verify_data &&
 825             c->verify_data->keys.set->data)
 826                 list_del_init(&c->verify_data->list);
 827         else
 828                 c->verify_data = NULL;
 829 #endif
 830
 831         c->shrink.count_objects = bch_mca_count;
 832         c->shrink.scan_objects = bch_mca_scan;
 833         c->shrink.seeks = 4;
 834         c->shrink.batch = c->btree_pages * 2;
 835
 836         if (register_shrinker(&c->shrink, "md-bcache:%pU", c->set_uuid))
 837                 pr_warn("bcache: %s: could not register shrinker\n",
 838                                 __func__);
 839
 840         return 0;
 841 }
 842
 843 /* Btree in memory cache - hash table */
 844
 845 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 846 {
 847         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 848 }
 849
 850 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 851 {
 852         struct btree *b;
 853
 854         rcu_read_lock();
 855         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 856                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 857                         goto out;
 858         b = NULL;
 859 out:
 860         rcu_read_unlock();
 861         return b;
 862 }
 863
 864 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 865 {
 866         spin_lock(&c->btree_cannibalize_lock);
 867         if (likely(c->btree_cache_alloc_lock == NULL)) {
 868                 c->btree_cache_alloc_lock = current;
 869         } else if (c->btree_cache_alloc_lock != current) {
 870                 if (op)
 871                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 872                                         TASK_UNINTERRUPTIBLE);
 873                 spin_unlock(&c->btree_cannibalize_lock);
 874                 return -EINTR;
 875         }
 876         spin_unlock(&c->btree_cannibalize_lock);
 877
 878         return 0;
 879 }
 880
 881 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 882                                      struct bkey *k)
 883 {
 884         struct btree *b;
 885
 886         trace_bcache_btree_cache_cannibalize(c);
 887
 888         if (mca_cannibalize_lock(c, op))
 889                 return ERR_PTR(-EINTR);
 890
 891         list_for_each_entry_reverse(b, &c->btree_cache, list)
 892                 if (!mca_reap(b, btree_order(k), false))
 893                         return b;
 894
 895         list_for_each_entry_reverse(b, &c->btree_cache, list)
 896                 if (!mca_reap(b, btree_order(k), true))
 897                         return b;
 898
 899         WARN(1, "btree cache cannibalize failed\n");
 900         return ERR_PTR(-ENOMEM);
 901 }
 902
 903 /*
 904  * We can only have one thread cannibalizing other cached btree nodes at a time,
 905  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 906  * cannibalize_bucket() will take. This means every time we unlock the root of
 907  * the btree, we need to release this lock if we have it held.
 908  */
 909 void bch_cannibalize_unlock(struct cache_set *c)
 910 {
 911         spin_lock(&c->btree_cannibalize_lock);
 912         if (c->btree_cache_alloc_lock == current) {
 913                 c->btree_cache_alloc_lock = NULL;
 914                 wake_up(&c->btree_cache_wait);
 915         }
 916         spin_unlock(&c->btree_cannibalize_lock);
 917 }
 918
 919 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 920                                struct bkey *k, int level)
 921 {
 922         struct btree *b;
 923
 924         BUG_ON(current->bio_list);
 925
 926         lockdep_assert_held(&c->bucket_lock);
 927
 928         if (mca_find(c, k))
 929                 return NULL;
 930
 931         /* btree_free() doesn't free memory; it sticks the node on the end of
 932          * the list. Check if there's any freed nodes there:
 933          */
 934         list_for_each_entry(b, &c->btree_cache_freeable, list)
 935                 if (!mca_reap(b, btree_order(k), false))
 936                         goto out;
 937
 938         /* We never free struct btree itself, just the memory that holds the on
 939          * disk node. Check the freed list before allocating a new one:
 940          */
 941         list_for_each_entry(b, &c->btree_cache_freed, list)
 942                 if (!mca_reap(b, 0, false)) {
 943                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 944                         if (!b->keys.set[0].data)
 945                                 goto err;
 946                         else
 947                                 goto out;
 948                 }
 949
 950         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 951         if (!b)
 952                 goto err;
 953
 954         BUG_ON(!down_write_trylock(&b->lock));
 955         if (!b->keys.set->data)
 956                 goto err;
 957 out:
 958         BUG_ON(b->io_mutex.count != 1);
 959
 960         bkey_copy(&b->key, k);
 961         list_move(&b->list, &c->btree_cache);
 962         hlist_del_init_rcu(&b->hash);
 963         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 964
 965         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 966         b->parent       = (void *) ~0UL;
 967         b->flags        = 0;
 968         b->written      = 0;
 969         b->level        = level;
 970
 971         if (!b->level)
 972                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 973                                     &b->c->expensive_debug_checks);
 974         else
 975                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 976                                     &b->c->expensive_debug_checks);
 977
 978         return b;
 979 err:
 980         if (b)
 981                 rw_unlock(true, b);
 982
 983         b = mca_cannibalize(c, op, k);
 984         if (!IS_ERR(b))
 985                 goto out;
 986
 987         return b;
 988 }
 989
 990 /*
 991  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 992  * in from disk if necessary.
 993  *
 994  * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
 995  *
 996  * The btree node will have either a read or a write lock held, depending on
 997  * level and op->lock.
 998  */
 999 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1000                                  struct bkey *k, int level, bool write,
1001                                  struct btree *parent)
1002 {
1003         int i = 0;
1004         struct btree *b;
1005
1006         BUG_ON(level < 0);
1007 retry:
1008         b = mca_find(c, k);
1009
1010         if (!b) {
1011                 if (current->bio_list)
1012                         return ERR_PTR(-EAGAIN);
1013
1014                 mutex_lock(&c->bucket_lock);
1015                 b = mca_alloc(c, op, k, level);
1016                 mutex_unlock(&c->bucket_lock);
1017
1018                 if (!b)
1019                         goto retry;
1020                 if (IS_ERR(b))
1021                         return b;
1022
1023                 bch_btree_node_read(b);
1024
1025                 if (!write)
1026                         downgrade_write(&b->lock);
1027         } else {
1028                 rw_lock(write, b, level);
1029                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1030                         rw_unlock(write, b);
1031                         goto retry;
1032                 }
1033                 BUG_ON(b->level != level);
1034         }
1035
1036         if (btree_node_io_error(b)) {
1037                 rw_unlock(write, b);
1038                 return ERR_PTR(-EIO);
1039         }
1040
1041         BUG_ON(!b->written);
1042
1043         b->parent = parent;
1044
1045         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1046                 prefetch(b->keys.set[i].tree);
1047                 prefetch(b->keys.set[i].data);
1048         }
1049
1050         for (; i <= b->keys.nsets; i++)
1051                 prefetch(b->keys.set[i].data);
1052
1053         return b;
1054 }
1055
1056 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1057 {
1058         struct btree *b;
1059
1060         mutex_lock(&parent->c->bucket_lock);
1061         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1062         mutex_unlock(&parent->c->bucket_lock);
1063
1064         if (!IS_ERR_OR_NULL(b)) {
1065                 b->parent = parent;
1066                 bch_btree_node_read(b);
1067                 rw_unlock(true, b);
1068         }
1069 }
1070
1071 /* Btree alloc */
1072
1073 static void btree_node_free(struct btree *b)
1074 {
1075         trace_bcache_btree_node_free(b);
1076
1077         BUG_ON(b == b->c->root);
1078
1079 retry:
1080         mutex_lock(&b->write_lock);
1081         /*
1082          * If the btree node is selected and flushing in btree_flush_write(),
1083          * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1084          * then it is safe to free the btree node here. Otherwise this btree
1085          * node will be in race condition.
1086          */
1087         if (btree_node_journal_flush(b)) {
1088                 mutex_unlock(&b->write_lock);
1089                 pr_debug("bnode %p journal_flush set, retry\n", b);
1090                 udelay(1);
1091                 goto retry;
1092         }
1093
1094         if (btree_node_dirty(b)) {
1095                 btree_complete_write(b, btree_current_write(b));
1096                 clear_bit(BTREE_NODE_dirty, &b->flags);
1097         }
1098
1099         mutex_unlock(&b->write_lock);
1100
1101         cancel_delayed_work(&b->work);
1102
1103         mutex_lock(&b->c->bucket_lock);
1104         bch_bucket_free(b->c, &b->key);
1105         mca_bucket_free(b);
1106         mutex_unlock(&b->c->bucket_lock);
1107 }
1108
1109 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1110                                      int level, bool wait,
1111                                      struct btree *parent)
1112 {
1113         BKEY_PADDED(key) k;
1114         struct btree *b;
1115
1116         mutex_lock(&c->bucket_lock);
1117 retry:
1118         /* return ERR_PTR(-EAGAIN) when it fails */
1119         b = ERR_PTR(-EAGAIN);
1120         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1121                 goto err;
1122
1123         bkey_put(c, &k.key);
1124         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1125
1126         b = mca_alloc(c, op, &k.key, level);
1127         if (IS_ERR(b))
1128                 goto err_free;
1129
1130         if (!b) {
1131                 cache_bug(c,
1132                         "Tried to allocate bucket that was in btree cache");
1133                 goto retry;
1134         }
1135
1136         b->parent = parent;
1137         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->cache->sb));
1138
1139         mutex_unlock(&c->bucket_lock);
1140
1141         trace_bcache_btree_node_alloc(b);
1142         return b;
1143 err_free:
1144         bch_bucket_free(c, &k.key);
1145 err:
1146         mutex_unlock(&c->bucket_lock);
1147
1148         trace_bcache_btree_node_alloc_fail(c);
1149         return b;
1150 }
1151
1152 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1153                                           struct btree_op *op, int level,
1154                                           struct btree *parent)
1155 {
1156         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1157 }
1158
1159 static struct btree *btree_node_alloc_replacement(struct btree *b,
1160                                                   struct btree_op *op)
1161 {
1162         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1163
1164         if (!IS_ERR(n)) {
1165                 mutex_lock(&n->write_lock);
1166                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1167                 bkey_copy_key(&n->key, &b->key);
1168                 mutex_unlock(&n->write_lock);
1169         }
1170
1171         return n;
1172 }
1173
1174 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1175 {
1176         unsigned int i;
1177
1178         mutex_lock(&b->c->bucket_lock);
1179
1180         atomic_inc(&b->c->prio_blocked);
1181
1182         bkey_copy(k, &b->key);
1183         bkey_copy_key(k, &ZERO_KEY);
1184
1185         for (i = 0; i < KEY_PTRS(k); i++)
1186                 SET_PTR_GEN(k, i,
1187                             bch_inc_gen(b->c->cache,
1188                                         PTR_BUCKET(b->c, &b->key, i)));
1189
1190         mutex_unlock(&b->c->bucket_lock);
1191 }
1192
1193 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1194 {
1195         struct cache_set *c = b->c;
1196         struct cache *ca = c->cache;
1197         unsigned int reserve = (c->root->level - b->level) * 2 + 1;
1198
1199         mutex_lock(&c->bucket_lock);
1200
1201         if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1202                 if (op)
1203                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
1204                                         TASK_UNINTERRUPTIBLE);
1205                 mutex_unlock(&c->bucket_lock);
1206                 return -EINTR;
1207         }
1208
1209         mutex_unlock(&c->bucket_lock);
1210
1211         return mca_cannibalize_lock(b->c, op);
1212 }
1213
1214 /* Garbage collection */
1215
1216 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1217                                     struct bkey *k)
1218 {
1219         uint8_t stale = 0;
1220         unsigned int i;
1221         struct bucket *g;
1222
1223         /*
1224          * ptr_invalid() can't return true for the keys that mark btree nodes as
1225          * freed, but since ptr_bad() returns true we'll never actually use them
1226          * for anything and thus we don't want mark their pointers here
1227          */
1228         if (!bkey_cmp(k, &ZERO_KEY))
1229                 return stale;
1230
1231         for (i = 0; i < KEY_PTRS(k); i++) {
1232                 if (!ptr_available(c, k, i))
1233                         continue;
1234
1235                 g = PTR_BUCKET(c, k, i);
1236
1237                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1238                         g->last_gc = PTR_GEN(k, i);
1239
1240                 if (ptr_stale(c, k, i)) {
1241                         stale = max(stale, ptr_stale(c, k, i));
1242                         continue;
1243                 }
1244
1245                 cache_bug_on(GC_MARK(g) &&
1246                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1247                              c, "inconsistent ptrs: mark = %llu, level = %i",
1248                              GC_MARK(g), level);
1249
1250                 if (level)
1251                         SET_GC_MARK(g, GC_MARK_METADATA);
1252                 else if (KEY_DIRTY(k))
1253                         SET_GC_MARK(g, GC_MARK_DIRTY);
1254                 else if (!GC_MARK(g))
1255                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1256
1257                 /* guard against overflow */
1258                 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1259                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1260                                              MAX_GC_SECTORS_USED));
1261
1262                 BUG_ON(!GC_SECTORS_USED(g));
1263         }
1264
1265         return stale;
1266 }
1267
1268 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1269
1270 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1271 {
1272         unsigned int i;
1273
1274         for (i = 0; i < KEY_PTRS(k); i++)
1275                 if (ptr_available(c, k, i) &&
1276                     !ptr_stale(c, k, i)) {
1277                         struct bucket *b = PTR_BUCKET(c, k, i);
1278
1279                         b->gen = PTR_GEN(k, i);
1280
1281                         if (level && bkey_cmp(k, &ZERO_KEY))
1282                                 b->prio = BTREE_PRIO;
1283                         else if (!level && b->prio == BTREE_PRIO)
1284                                 b->prio = INITIAL_PRIO;
1285                 }
1286
1287         __bch_btree_mark_key(c, level, k);
1288 }
1289
1290 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1291 {
1292         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1293 }
1294
1295 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1296 {
1297         uint8_t stale = 0;
1298         unsigned int keys = 0, good_keys = 0;
1299         struct bkey *k;
1300         struct btree_iter iter;
1301         struct bset_tree *t;
1302
1303         gc->nodes++;
1304
1305         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1306                 stale = max(stale, btree_mark_key(b, k));
1307                 keys++;
1308
1309                 if (bch_ptr_bad(&b->keys, k))
1310                         continue;
1311
1312                 gc->key_bytes += bkey_u64s(k);
1313                 gc->nkeys++;
1314                 good_keys++;
1315
1316                 gc->data += KEY_SIZE(k);
1317         }
1318
1319         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1320                 btree_bug_on(t->size &&
1321                              bset_written(&b->keys, t) &&
1322                              bkey_cmp(&b->key, &t->end) < 0,
1323                              b, "found short btree key in gc");
1324
1325         if (b->c->gc_always_rewrite)
1326                 return true;
1327
1328         if (stale > 10)
1329                 return true;
1330
1331         if ((keys - good_keys) * 2 > keys)
1332                 return true;
1333
1334         return false;
1335 }
1336
1337 #define GC_MERGE_NODES  4U
1338
1339 struct gc_merge_info {
1340         struct btree    *b;
1341         unsigned int    keys;
1342 };
1343
1344 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1345                                  struct keylist *insert_keys,
1346                                  atomic_t *journal_ref,
1347                                  struct bkey *replace_key);
1348
1349 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1350                              struct gc_stat *gc, struct gc_merge_info *r)
1351 {
1352         unsigned int i, nodes = 0, keys = 0, blocks;
1353         struct btree *new_nodes[GC_MERGE_NODES];
1354         struct keylist keylist;
1355         struct closure cl;
1356         struct bkey *k;
1357
1358         bch_keylist_init(&keylist);
1359
1360         if (btree_check_reserve(b, NULL))
1361                 return 0;
1362
1363         memset(new_nodes, 0, sizeof(new_nodes));
1364         closure_init_stack(&cl);
1365
1366         while (nodes < GC_MERGE_NODES && !IS_ERR(r[nodes].b))
1367                 keys += r[nodes++].keys;
1368
1369         blocks = btree_default_blocks(b->c) * 2 / 3;
1370
1371         if (nodes < 2 ||
1372             __set_blocks(b->keys.set[0].data, keys,
1373                          block_bytes(b->c->cache)) > blocks * (nodes - 1))
1374                 return 0;
1375
1376         for (i = 0; i < nodes; i++) {
1377                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1378                 if (IS_ERR(new_nodes[i]))
1379                         goto out_nocoalesce;
1380         }
1381
1382         /*
1383          * We have to check the reserve here, after we've allocated our new
1384          * nodes, to make sure the insert below will succeed - we also check
1385          * before as an optimization to potentially avoid a bunch of expensive
1386          * allocs/sorts
1387          */
1388         if (btree_check_reserve(b, NULL))
1389                 goto out_nocoalesce;
1390
1391         for (i = 0; i < nodes; i++)
1392                 mutex_lock(&new_nodes[i]->write_lock);
1393
1394         for (i = nodes - 1; i > 0; --i) {
1395                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1396                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1397                 struct bkey *k, *last = NULL;
1398
1399                 keys = 0;
1400
1401                 if (i > 1) {
1402                         for (k = n2->start;
1403                              k < bset_bkey_last(n2);
1404                              k = bkey_next(k)) {
1405                                 if (__set_blocks(n1, n1->keys + keys +
1406                                                  bkey_u64s(k),
1407                                                  block_bytes(b->c->cache)) > blocks)
1408                                         break;
1409
1410                                 last = k;
1411                                 keys += bkey_u64s(k);
1412                         }
1413                 } else {
1414                         /*
1415                          * Last node we're not getting rid of - we're getting
1416                          * rid of the node at r[0]. Have to try and fit all of
1417                          * the remaining keys into this node; we can't ensure
1418                          * they will always fit due to rounding and variable
1419                          * length keys (shouldn't be possible in practice,
1420                          * though)
1421                          */
1422                         if (__set_blocks(n1, n1->keys + n2->keys,
1423                                          block_bytes(b->c->cache)) >
1424                             btree_blocks(new_nodes[i]))
1425                                 goto out_unlock_nocoalesce;
1426
1427                         keys = n2->keys;
1428                         /* Take the key of the node we're getting rid of */
1429                         last = &r->b->key;
1430                 }
1431
1432                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c->cache)) >
1433                        btree_blocks(new_nodes[i]));
1434
1435                 if (last)
1436                         bkey_copy_key(&new_nodes[i]->key, last);
1437
1438                 memcpy(bset_bkey_last(n1),
1439                        n2->start,
1440                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1441
1442                 n1->keys += keys;
1443                 r[i].keys = n1->keys;
1444
1445                 memmove(n2->start,
1446                         bset_bkey_idx(n2, keys),
1447                         (void *) bset_bkey_last(n2) -
1448                         (void *) bset_bkey_idx(n2, keys));
1449
1450                 n2->keys -= keys;
1451
1452                 if (__bch_keylist_realloc(&keylist,
1453                                           bkey_u64s(&new_nodes[i]->key)))
1454                         goto out_unlock_nocoalesce;
1455
1456                 bch_btree_node_write(new_nodes[i], &cl);
1457                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1458         }
1459
1460         for (i = 0; i < nodes; i++)
1461                 mutex_unlock(&new_nodes[i]->write_lock);
1462
1463         closure_sync(&cl);
1464
1465         /* We emptied out this node */
1466         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1467         btree_node_free(new_nodes[0]);
1468         rw_unlock(true, new_nodes[0]);
1469         new_nodes[0] = NULL;
1470
1471         for (i = 0; i < nodes; i++) {
1472                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1473                         goto out_nocoalesce;
1474
1475                 make_btree_freeing_key(r[i].b, keylist.top);
1476                 bch_keylist_push(&keylist);
1477         }
1478
1479         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1480         BUG_ON(!bch_keylist_empty(&keylist));
1481
1482         for (i = 0; i < nodes; i++) {
1483                 btree_node_free(r[i].b);
1484                 rw_unlock(true, r[i].b);
1485
1486                 r[i].b = new_nodes[i];
1487         }
1488
1489         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1490         r[nodes - 1].b = ERR_PTR(-EINTR);
1491
1492         trace_bcache_btree_gc_coalesce(nodes);
1493         gc->nodes--;
1494
1495         bch_keylist_free(&keylist);
1496
1497         /* Invalidated our iterator */
1498         return -EINTR;
1499
1500 out_unlock_nocoalesce:
1501         for (i = 0; i < nodes; i++)
1502                 mutex_unlock(&new_nodes[i]->write_lock);
1503
1504 out_nocoalesce:
1505         closure_sync(&cl);
1506
1507         while ((k = bch_keylist_pop(&keylist)))
1508                 if (!bkey_cmp(k, &ZERO_KEY))
1509                         atomic_dec(&b->c->prio_blocked);
1510         bch_keylist_free(&keylist);
1511
1512         for (i = 0; i < nodes; i++)
1513                 if (!IS_ERR(new_nodes[i])) {
1514                         btree_node_free(new_nodes[i]);
1515                         rw_unlock(true, new_nodes[i]);
1516                 }
1517         return 0;
1518 }
1519
1520 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1521                                  struct btree *replace)
1522 {
1523         struct keylist keys;
1524         struct btree *n;
1525
1526         if (btree_check_reserve(b, NULL))
1527                 return 0;
1528
1529         n = btree_node_alloc_replacement(replace, NULL);
1530
1531         /* recheck reserve after allocating replacement node */
1532         if (btree_check_reserve(b, NULL)) {
1533                 btree_node_free(n);
1534                 rw_unlock(true, n);
1535                 return 0;
1536         }
1537
1538         bch_btree_node_write_sync(n);
1539
1540         bch_keylist_init(&keys);
1541         bch_keylist_add(&keys, &n->key);
1542
1543         make_btree_freeing_key(replace, keys.top);
1544         bch_keylist_push(&keys);
1545
1546         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1547         BUG_ON(!bch_keylist_empty(&keys));
1548
1549         btree_node_free(replace);
1550         rw_unlock(true, n);
1551
1552         /* Invalidated our iterator */
1553         return -EINTR;
1554 }
1555
1556 static unsigned int btree_gc_count_keys(struct btree *b)
1557 {
1558         struct bkey *k;
1559         struct btree_iter iter;
1560         unsigned int ret = 0;
1561
1562         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1563                 ret += bkey_u64s(k);
1564
1565         return ret;
1566 }
1567
1568 static size_t btree_gc_min_nodes(struct cache_set *c)
1569 {
1570         size_t min_nodes;
1571
1572         /*
1573          * Since incremental GC would stop 100ms when front
1574          * side I/O comes, so when there are many btree nodes,
1575          * if GC only processes constant (100) nodes each time,
1576          * GC would last a long time, and the front side I/Os
1577          * would run out of the buckets (since no new bucket
1578          * can be allocated during GC), and be blocked again.
1579          * So GC should not process constant nodes, but varied
1580          * nodes according to the number of btree nodes, which
1581          * realized by dividing GC into constant(100) times,
1582          * so when there are many btree nodes, GC can process
1583          * more nodes each time, otherwise, GC will process less
1584          * nodes each time (but no less than MIN_GC_NODES)
1585          */
1586         min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1587         if (min_nodes < MIN_GC_NODES)
1588                 min_nodes = MIN_GC_NODES;
1589
1590         return min_nodes;
1591 }
1592
1593
1594 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1595                             struct closure *writes, struct gc_stat *gc)
1596 {
1597         int ret = 0;
1598         bool should_rewrite;
1599         struct bkey *k;
1600         struct btree_iter iter;
1601         struct gc_merge_info r[GC_MERGE_NODES];
1602         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1603
1604         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1605
1606         for (i = r; i < r + ARRAY_SIZE(r); i++)
1607                 i->b = ERR_PTR(-EINTR);
1608
1609         while (1) {
1610                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1611                 if (k) {
1612                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1613                                                   true, b);
1614                         if (IS_ERR(r->b)) {
1615                                 ret = PTR_ERR(r->b);
1616                                 break;
1617                         }
1618
1619                         r->keys = btree_gc_count_keys(r->b);
1620
1621                         ret = btree_gc_coalesce(b, op, gc, r);
1622                         if (ret)
1623                                 break;
1624                 }
1625
1626                 if (!last->b)
1627                         break;
1628
1629                 if (!IS_ERR(last->b)) {
1630                         should_rewrite = btree_gc_mark_node(last->b, gc);
1631                         if (should_rewrite) {
1632                                 ret = btree_gc_rewrite_node(b, op, last->b);
1633                                 if (ret)
1634                                         break;
1635                         }
1636
1637                         if (last->b->level) {
1638                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1639                                 if (ret)
1640                                         break;
1641                         }
1642
1643                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1644
1645                         /*
1646                          * Must flush leaf nodes before gc ends, since replace
1647                          * operations aren't journalled
1648                          */
1649                         mutex_lock(&last->b->write_lock);
1650                         if (btree_node_dirty(last->b))
1651                                 bch_btree_node_write(last->b, writes);
1652                         mutex_unlock(&last->b->write_lock);
1653                         rw_unlock(true, last->b);
1654                 }
1655
1656                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1657                 r->b = NULL;
1658
1659                 if (atomic_read(&b->c->search_inflight) &&
1660                     gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1661                         gc->nodes_pre =  gc->nodes;
1662                         ret = -EAGAIN;
1663                         break;
1664                 }
1665
1666                 if (need_resched()) {
1667                         ret = -EAGAIN;
1668                         break;
1669                 }
1670         }
1671
1672         for (i = r; i < r + ARRAY_SIZE(r); i++)
1673                 if (!IS_ERR_OR_NULL(i->b)) {
1674                         mutex_lock(&i->b->write_lock);
1675                         if (btree_node_dirty(i->b))
1676                                 bch_btree_node_write(i->b, writes);
1677                         mutex_unlock(&i->b->write_lock);
1678                         rw_unlock(true, i->b);
1679                 }
1680
1681         return ret;
1682 }
1683
1684 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1685                              struct closure *writes, struct gc_stat *gc)
1686 {
1687         struct btree *n = NULL;
1688         int ret = 0;
1689         bool should_rewrite;
1690
1691         should_rewrite = btree_gc_mark_node(b, gc);
1692         if (should_rewrite) {
1693                 n = btree_node_alloc_replacement(b, NULL);
1694
1695                 if (!IS_ERR(n)) {
1696                         bch_btree_node_write_sync(n);
1697
1698                         bch_btree_set_root(n);
1699                         btree_node_free(b);
1700                         rw_unlock(true, n);
1701
1702                         return -EINTR;
1703                 }
1704         }
1705
1706         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1707
1708         if (b->level) {
1709                 ret = btree_gc_recurse(b, op, writes, gc);
1710                 if (ret)
1711                         return ret;
1712         }
1713
1714         bkey_copy_key(&b->c->gc_done, &b->key);
1715
1716         return ret;
1717 }
1718
1719 static void btree_gc_start(struct cache_set *c)
1720 {
1721         struct cache *ca;
1722         struct bucket *b;
1723
1724         if (!c->gc_mark_valid)
1725                 return;
1726
1727         mutex_lock(&c->bucket_lock);
1728
1729         c->gc_mark_valid = 0;
1730         c->gc_done = ZERO_KEY;
1731
1732         ca = c->cache;
1733         for_each_bucket(b, ca) {
1734                 b->last_gc = b->gen;
1735                 if (!atomic_read(&b->pin)) {
1736                         SET_GC_MARK(b, 0);
1737                         SET_GC_SECTORS_USED(b, 0);
1738                 }
1739         }
1740
1741         mutex_unlock(&c->bucket_lock);
1742 }
1743
1744 static void bch_btree_gc_finish(struct cache_set *c)
1745 {
1746         struct bucket *b;
1747         struct cache *ca;
1748         unsigned int i, j;
1749         uint64_t *k;
1750
1751         mutex_lock(&c->bucket_lock);
1752
1753         set_gc_sectors(c);
1754         c->gc_mark_valid = 1;
1755         c->need_gc      = 0;
1756
1757         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1758                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1759                             GC_MARK_METADATA);
1760
1761         /* don't reclaim buckets to which writeback keys point */
1762         rcu_read_lock();
1763         for (i = 0; i < c->devices_max_used; i++) {
1764                 struct bcache_device *d = c->devices[i];
1765                 struct cached_dev *dc;
1766                 struct keybuf_key *w, *n;
1767
1768                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1769                         continue;
1770                 dc = container_of(d, struct cached_dev, disk);
1771
1772                 spin_lock(&dc->writeback_keys.lock);
1773                 rbtree_postorder_for_each_entry_safe(w, n,
1774                                         &dc->writeback_keys.keys, node)
1775                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1776                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1777                                             GC_MARK_DIRTY);
1778                 spin_unlock(&dc->writeback_keys.lock);
1779         }
1780         rcu_read_unlock();
1781
1782         c->avail_nbuckets = 0;
1783
1784         ca = c->cache;
1785         ca->invalidate_needs_gc = 0;
1786
1787         for (k = ca->sb.d; k < ca->sb.d + ca->sb.keys; k++)
1788                 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1789
1790         for (k = ca->prio_buckets;
1791              k < ca->prio_buckets + prio_buckets(ca) * 2; k++)
1792                 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1793
1794         for_each_bucket(b, ca) {
1795                 c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1796
1797                 if (atomic_read(&b->pin))
1798                         continue;
1799
1800                 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1801
1802                 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1803                         c->avail_nbuckets++;
1804         }
1805
1806         mutex_unlock(&c->bucket_lock);
1807 }
1808
1809 static void bch_btree_gc(struct cache_set *c)
1810 {
1811         int ret;
1812         struct gc_stat stats;
1813         struct closure writes;
1814         struct btree_op op;
1815         uint64_t start_time = local_clock();
1816
1817         trace_bcache_gc_start(c);
1818
1819         memset(&stats, 0, sizeof(struct gc_stat));
1820         closure_init_stack(&writes);
1821         bch_btree_op_init(&op, SHRT_MAX);
1822
1823         btree_gc_start(c);
1824
1825         /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1826         do {
1827                 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
1828                 closure_sync(&writes);
1829                 cond_resched();
1830
1831                 if (ret == -EAGAIN)
1832                         schedule_timeout_interruptible(msecs_to_jiffies
1833                                                        (GC_SLEEP_MS));
1834                 else if (ret)
1835                         pr_warn("gc failed!\n");
1836         } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1837
1838         bch_btree_gc_finish(c);
1839         wake_up_allocators(c);
1840
1841         bch_time_stats_update(&c->btree_gc_time, start_time);
1842
1843         stats.key_bytes *= sizeof(uint64_t);
1844         stats.data      <<= 9;
1845         bch_update_bucket_in_use(c, &stats);
1846         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1847
1848         trace_bcache_gc_end(c);
1849
1850         bch_moving_gc(c);
1851 }
1852
1853 static bool gc_should_run(struct cache_set *c)
1854 {
1855         struct cache *ca = c->cache;
1856
1857         if (ca->invalidate_needs_gc)
1858                 return true;
1859
1860         if (atomic_read(&c->sectors_to_gc) < 0)
1861                 return true;
1862
1863         return false;
1864 }
1865
1866 static int bch_gc_thread(void *arg)
1867 {
1868         struct cache_set *c = arg;
1869
1870         while (1) {
1871                 wait_event_interruptible(c->gc_wait,
1872                            kthread_should_stop() ||
1873                            test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1874                            gc_should_run(c));
1875
1876                 if (kthread_should_stop() ||
1877                     test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1878                         break;
1879
1880                 set_gc_sectors(c);
1881                 bch_btree_gc(c);
1882         }
1883
1884         wait_for_kthread_stop();
1885         return 0;
1886 }
1887
1888 int bch_gc_thread_start(struct cache_set *c)
1889 {
1890         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1891         return PTR_ERR_OR_ZERO(c->gc_thread);
1892 }
1893
1894 /* Initial partial gc */
1895
1896 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1897 {
1898         int ret = 0;
1899         struct bkey *k, *p = NULL;
1900         struct btree_iter iter;
1901
1902         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1903                 bch_initial_mark_key(b->c, b->level, k);
1904
1905         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1906
1907         if (b->level) {
1908                 bch_btree_iter_init(&b->keys, &iter, NULL);
1909
1910                 do {
1911                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1912                                                        bch_ptr_bad);
1913                         if (k) {
1914                                 btree_node_prefetch(b, k);
1915                                 /*
1916                                  * initiallize c->gc_stats.nodes
1917                                  * for incremental GC
1918                                  */
1919                                 b->c->gc_stats.nodes++;
1920                         }
1921
1922                         if (p)
1923                                 ret = bcache_btree(check_recurse, p, b, op);
1924
1925                         p = k;
1926                 } while (p && !ret);
1927         }
1928
1929         return ret;
1930 }
1931
1932
1933 static int bch_btree_check_thread(void *arg)
1934 {
1935         int ret;
1936         struct btree_check_info *info = arg;
1937         struct btree_check_state *check_state = info->state;
1938         struct cache_set *c = check_state->c;
1939         struct btree_iter iter;
1940         struct bkey *k, *p;
1941         int cur_idx, prev_idx, skip_nr;
1942
1943         k = p = NULL;
1944         cur_idx = prev_idx = 0;
1945         ret = 0;
1946
1947         /* root node keys are checked before thread created */
1948         bch_btree_iter_init(&c->root->keys, &iter, NULL);
1949         k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
1950         BUG_ON(!k);
1951
1952         p = k;
1953         while (k) {
1954                 /*
1955                  * Fetch a root node key index, skip the keys which
1956                  * should be fetched by other threads, then check the
1957                  * sub-tree indexed by the fetched key.
1958                  */
1959                 spin_lock(&check_state->idx_lock);
1960                 cur_idx = check_state->key_idx;
1961                 check_state->key_idx++;
1962                 spin_unlock(&check_state->idx_lock);
1963
1964                 skip_nr = cur_idx - prev_idx;
1965
1966                 while (skip_nr) {
1967                         k = bch_btree_iter_next_filter(&iter,
1968                                                        &c->root->keys,
1969                                                        bch_ptr_bad);
1970                         if (k)
1971                                 p = k;
1972                         else {
1973                                 /*
1974                                  * No more keys to check in root node,
1975                                  * current checking threads are enough,
1976                                  * stop creating more.
1977                                  */
1978                                 atomic_set(&check_state->enough, 1);
1979                                 /* Update check_state->enough earlier */
1980                                 smp_mb__after_atomic();
1981                                 goto out;
1982                         }
1983                         skip_nr--;
1984                         cond_resched();
1985                 }
1986
1987                 if (p) {
1988                         struct btree_op op;
1989
1990                         btree_node_prefetch(c->root, p);
1991                         c->gc_stats.nodes++;
1992                         bch_btree_op_init(&op, 0);
1993                         ret = bcache_btree(check_recurse, p, c->root, &op);
1994                         /*
1995                          * The op may be added to cache_set's btree_cache_wait
1996                          * in mca_cannibalize(), must ensure it is removed from
1997                          * the list and release btree_cache_alloc_lock before
1998                          * free op memory.
1999                          * Otherwise, the btree_cache_wait will be damaged.
2000                          */
2001                         bch_cannibalize_unlock(c);
2002                         finish_wait(&c->btree_cache_wait, &(&op)->wait);
2003                         if (ret)
2004                                 goto out;
2005                 }
2006                 p = NULL;
2007                 prev_idx = cur_idx;
2008                 cond_resched();
2009         }
2010
2011 out:
2012         info->result = ret;
2013         /* update check_state->started among all CPUs */
2014         smp_mb__before_atomic();
2015         if (atomic_dec_and_test(&check_state->started))
2016                 wake_up(&check_state->wait);
2017
2018         return ret;
2019 }
2020
2021
2022
2023 static int bch_btree_chkthread_nr(void)
2024 {
2025         int n = num_online_cpus()/2;
2026
2027         if (n == 0)
2028                 n = 1;
2029         else if (n > BCH_BTR_CHKTHREAD_MAX)
2030                 n = BCH_BTR_CHKTHREAD_MAX;
2031
2032         return n;
2033 }
2034
2035 int bch_btree_check(struct cache_set *c)
2036 {
2037         int ret = 0;
2038         int i;
2039         struct bkey *k = NULL;
2040         struct btree_iter iter;
2041         struct btree_check_state check_state;
2042
2043         /* check and mark root node keys */
2044         for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2045                 bch_initial_mark_key(c, c->root->level, k);
2046
2047         bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2048
2049         if (c->root->level == 0)
2050                 return 0;
2051
2052         memset(&check_state, 0, sizeof(struct btree_check_state));
2053         check_state.c = c;
2054         check_state.total_threads = bch_btree_chkthread_nr();
2055         check_state.key_idx = 0;
2056         spin_lock_init(&check_state.idx_lock);
2057         atomic_set(&check_state.started, 0);
2058         atomic_set(&check_state.enough, 0);
2059         init_waitqueue_head(&check_state.wait);
2060
2061         rw_lock(0, c->root, c->root->level);
2062         /*
2063          * Run multiple threads to check btree nodes in parallel,
2064          * if check_state.enough is non-zero, it means current
2065          * running check threads are enough, unncessary to create
2066          * more.
2067          */
2068         for (i = 0; i < check_state.total_threads; i++) {
2069                 /* fetch latest check_state.enough earlier */
2070                 smp_mb__before_atomic();
2071                 if (atomic_read(&check_state.enough))
2072                         break;
2073
2074                 check_state.infos[i].result = 0;
2075                 check_state.infos[i].state = &check_state;
2076
2077                 check_state.infos[i].thread =
2078                         kthread_run(bch_btree_check_thread,
2079                                     &check_state.infos[i],
2080                                     "bch_btrchk[%d]", i);
2081                 if (IS_ERR(check_state.infos[i].thread)) {
2082                         pr_err("fails to run thread bch_btrchk[%d]\n", i);
2083                         for (--i; i >= 0; i--)
2084                                 kthread_stop(check_state.infos[i].thread);
2085                         ret = -ENOMEM;
2086                         goto out;
2087                 }
2088                 atomic_inc(&check_state.started);
2089         }
2090
2091         /*
2092          * Must wait for all threads to stop.
2093          */
2094         wait_event(check_state.wait, atomic_read(&check_state.started) == 0);
2095
2096         for (i = 0; i < check_state.total_threads; i++) {
2097                 if (check_state.infos[i].result) {
2098                         ret = check_state.infos[i].result;
2099                         goto out;
2100                 }
2101         }
2102
2103 out:
2104         rw_unlock(0, c->root);
2105         return ret;
2106 }
2107
2108 void bch_initial_gc_finish(struct cache_set *c)
2109 {
2110         struct cache *ca = c->cache;
2111         struct bucket *b;
2112
2113         bch_btree_gc_finish(c);
2114
2115         mutex_lock(&c->bucket_lock);
2116
2117         /*
2118          * We need to put some unused buckets directly on the prio freelist in
2119          * order to get the allocator thread started - it needs freed buckets in
2120          * order to rewrite the prios and gens, and it needs to rewrite prios
2121          * and gens in order to free buckets.
2122          *
2123          * This is only safe for buckets that have no live data in them, which
2124          * there should always be some of.
2125          */
2126         for_each_bucket(b, ca) {
2127                 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2128                     fifo_full(&ca->free[RESERVE_BTREE]))
2129                         break;
2130
2131                 if (bch_can_invalidate_bucket(ca, b) &&
2132                     !GC_MARK(b)) {
2133                         __bch_invalidate_one_bucket(ca, b);
2134                         if (!fifo_push(&ca->free[RESERVE_PRIO],
2135                            b - ca->buckets))
2136                                 fifo_push(&ca->free[RESERVE_BTREE],
2137                                           b - ca->buckets);
2138                 }
2139         }
2140
2141         mutex_unlock(&c->bucket_lock);
2142 }
2143
2144 /* Btree insertion */
2145
2146 static bool btree_insert_key(struct btree *b, struct bkey *k,
2147                              struct bkey *replace_key)
2148 {
2149         unsigned int status;
2150
2151         BUG_ON(bkey_cmp(k, &b->key) > 0);
2152
2153         status = bch_btree_insert_key(&b->keys, k, replace_key);
2154         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2155                 bch_check_keys(&b->keys, "%u for %s", status,
2156                                replace_key ? "replace" : "insert");
2157
2158                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2159                                               status);
2160                 return true;
2161         } else
2162                 return false;
2163 }
2164
2165 static size_t insert_u64s_remaining(struct btree *b)
2166 {
2167         long ret = bch_btree_keys_u64s_remaining(&b->keys);
2168
2169         /*
2170          * Might land in the middle of an existing extent and have to split it
2171          */
2172         if (b->keys.ops->is_extents)
2173                 ret -= KEY_MAX_U64S;
2174
2175         return max(ret, 0L);
2176 }
2177
2178 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2179                                   struct keylist *insert_keys,
2180                                   struct bkey *replace_key)
2181 {
2182         bool ret = false;
2183         int oldsize = bch_count_data(&b->keys);
2184
2185         while (!bch_keylist_empty(insert_keys)) {
2186                 struct bkey *k = insert_keys->keys;
2187
2188                 if (bkey_u64s(k) > insert_u64s_remaining(b))
2189                         break;
2190
2191                 if (bkey_cmp(k, &b->key) <= 0) {
2192                         if (!b->level)
2193                                 bkey_put(b->c, k);
2194
2195                         ret |= btree_insert_key(b, k, replace_key);
2196                         bch_keylist_pop_front(insert_keys);
2197                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2198                         BKEY_PADDED(key) temp;
2199                         bkey_copy(&temp.key, insert_keys->keys);
2200
2201                         bch_cut_back(&b->key, &temp.key);
2202                         bch_cut_front(&b->key, insert_keys->keys);
2203
2204                         ret |= btree_insert_key(b, &temp.key, replace_key);
2205                         break;
2206                 } else {
2207                         break;
2208                 }
2209         }
2210
2211         if (!ret)
2212                 op->insert_collision = true;
2213
2214         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2215
2216         BUG_ON(bch_count_data(&b->keys) < oldsize);
2217         return ret;
2218 }
2219
2220 static int btree_split(struct btree *b, struct btree_op *op,
2221                        struct keylist *insert_keys,
2222                        struct bkey *replace_key)
2223 {
2224         bool split;
2225         struct btree *n1, *n2 = NULL, *n3 = NULL;
2226         uint64_t start_time = local_clock();
2227         struct closure cl;
2228         struct keylist parent_keys;
2229
2230         closure_init_stack(&cl);
2231         bch_keylist_init(&parent_keys);
2232
2233         if (btree_check_reserve(b, op)) {
2234                 if (!b->level)
2235                         return -EINTR;
2236                 else
2237                         WARN(1, "insufficient reserve for split\n");
2238         }
2239
2240         n1 = btree_node_alloc_replacement(b, op);
2241         if (IS_ERR(n1))
2242                 goto err;
2243
2244         split = set_blocks(btree_bset_first(n1),
2245                            block_bytes(n1->c->cache)) > (btree_blocks(b) * 4) / 5;
2246
2247         if (split) {
2248                 unsigned int keys = 0;
2249
2250                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2251
2252                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2253                 if (IS_ERR(n2))
2254                         goto err_free1;
2255
2256                 if (!b->parent) {
2257                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2258                         if (IS_ERR(n3))
2259                                 goto err_free2;
2260                 }
2261
2262                 mutex_lock(&n1->write_lock);
2263                 mutex_lock(&n2->write_lock);
2264
2265                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2266
2267                 /*
2268                  * Has to be a linear search because we don't have an auxiliary
2269                  * search tree yet
2270                  */
2271
2272                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2273                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2274                                                         keys));
2275
2276                 bkey_copy_key(&n1->key,
2277                               bset_bkey_idx(btree_bset_first(n1), keys));
2278                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2279
2280                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2281                 btree_bset_first(n1)->keys = keys;
2282
2283                 memcpy(btree_bset_first(n2)->start,
2284                        bset_bkey_last(btree_bset_first(n1)),
2285                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2286
2287                 bkey_copy_key(&n2->key, &b->key);
2288
2289                 bch_keylist_add(&parent_keys, &n2->key);
2290                 bch_btree_node_write(n2, &cl);
2291                 mutex_unlock(&n2->write_lock);
2292                 rw_unlock(true, n2);
2293         } else {
2294                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2295
2296                 mutex_lock(&n1->write_lock);
2297                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2298         }
2299
2300         bch_keylist_add(&parent_keys, &n1->key);
2301         bch_btree_node_write(n1, &cl);
2302         mutex_unlock(&n1->write_lock);
2303
2304         if (n3) {
2305                 /* Depth increases, make a new root */
2306                 mutex_lock(&n3->write_lock);
2307                 bkey_copy_key(&n3->key, &MAX_KEY);
2308                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2309                 bch_btree_node_write(n3, &cl);
2310                 mutex_unlock(&n3->write_lock);
2311
2312                 closure_sync(&cl);
2313                 bch_btree_set_root(n3);
2314                 rw_unlock(true, n3);
2315         } else if (!b->parent) {
2316                 /* Root filled up but didn't need to be split */
2317                 closure_sync(&cl);
2318                 bch_btree_set_root(n1);
2319         } else {
2320                 /* Split a non root node */
2321                 closure_sync(&cl);
2322                 make_btree_freeing_key(b, parent_keys.top);
2323                 bch_keylist_push(&parent_keys);
2324
2325                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2326                 BUG_ON(!bch_keylist_empty(&parent_keys));
2327         }
2328
2329         btree_node_free(b);
2330         rw_unlock(true, n1);
2331
2332         bch_time_stats_update(&b->c->btree_split_time, start_time);
2333
2334         return 0;
2335 err_free2:
2336         bkey_put(b->c, &n2->key);
2337         btree_node_free(n2);
2338         rw_unlock(true, n2);
2339 err_free1:
2340         bkey_put(b->c, &n1->key);
2341         btree_node_free(n1);
2342         rw_unlock(true, n1);
2343 err:
2344         WARN(1, "bcache: btree split failed (level %u)", b->level);
2345
2346         if (n3 == ERR_PTR(-EAGAIN) ||
2347             n2 == ERR_PTR(-EAGAIN) ||
2348             n1 == ERR_PTR(-EAGAIN))
2349                 return -EAGAIN;
2350
2351         return -ENOMEM;
2352 }
2353
2354 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2355                                  struct keylist *insert_keys,
2356                                  atomic_t *journal_ref,
2357                                  struct bkey *replace_key)
2358 {
2359         struct closure cl;
2360
2361         BUG_ON(b->level && replace_key);
2362
2363         closure_init_stack(&cl);
2364
2365         mutex_lock(&b->write_lock);
2366
2367         if (write_block(b) != btree_bset_last(b) &&
2368             b->keys.last_set_unwritten)
2369                 bch_btree_init_next(b); /* just wrote a set */
2370
2371         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2372                 mutex_unlock(&b->write_lock);
2373                 goto split;
2374         }
2375
2376         BUG_ON(write_block(b) != btree_bset_last(b));
2377
2378         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2379                 if (!b->level)
2380                         bch_btree_leaf_dirty(b, journal_ref);
2381                 else
2382                         bch_btree_node_write(b, &cl);
2383         }
2384
2385         mutex_unlock(&b->write_lock);
2386
2387         /* wait for btree node write if necessary, after unlock */
2388         closure_sync(&cl);
2389
2390         return 0;
2391 split:
2392         if (current->bio_list) {
2393                 op->lock = b->c->root->level + 1;
2394                 return -EAGAIN;
2395         } else if (op->lock <= b->c->root->level) {
2396                 op->lock = b->c->root->level + 1;
2397                 return -EINTR;
2398         } else {
2399                 /* Invalidated all iterators */
2400                 int ret = btree_split(b, op, insert_keys, replace_key);
2401
2402                 if (bch_keylist_empty(insert_keys))
2403                         return 0;
2404                 else if (!ret)
2405                         return -EINTR;
2406                 return ret;
2407         }
2408 }
2409
2410 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2411                                struct bkey *check_key)
2412 {
2413         int ret = -EINTR;
2414         uint64_t btree_ptr = b->key.ptr[0];
2415         unsigned long seq = b->seq;
2416         struct keylist insert;
2417         bool upgrade = op->lock == -1;
2418
2419         bch_keylist_init(&insert);
2420
2421         if (upgrade) {
2422                 rw_unlock(false, b);
2423                 rw_lock(true, b, b->level);
2424
2425                 if (b->key.ptr[0] != btree_ptr ||
2426                     b->seq != seq + 1) {
2427                         op->lock = b->level;
2428                         goto out;
2429                 }
2430         }
2431
2432         SET_KEY_PTRS(check_key, 1);
2433         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2434
2435         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2436
2437         bch_keylist_add(&insert, check_key);
2438
2439         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2440
2441         BUG_ON(!ret && !bch_keylist_empty(&insert));
2442 out:
2443         if (upgrade)
2444                 downgrade_write(&b->lock);
2445         return ret;
2446 }
2447
2448 struct btree_insert_op {
2449         struct btree_op op;
2450         struct keylist  *keys;
2451         atomic_t        *journal_ref;
2452         struct bkey     *replace_key;
2453 };
2454
2455 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2456 {
2457         struct btree_insert_op *op = container_of(b_op,
2458                                         struct btree_insert_op, op);
2459
2460         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2461                                         op->journal_ref, op->replace_key);
2462         if (ret && !bch_keylist_empty(op->keys))
2463                 return ret;
2464         else
2465                 return MAP_DONE;
2466 }
2467
2468 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2469                      atomic_t *journal_ref, struct bkey *replace_key)
2470 {
2471         struct btree_insert_op op;
2472         int ret = 0;
2473
2474         BUG_ON(current->bio_list);
2475         BUG_ON(bch_keylist_empty(keys));
2476
2477         bch_btree_op_init(&op.op, 0);
2478         op.keys         = keys;
2479         op.journal_ref  = journal_ref;
2480         op.replace_key  = replace_key;
2481
2482         while (!ret && !bch_keylist_empty(keys)) {
2483                 op.op.lock = 0;
2484                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2485                                                &START_KEY(keys->keys),
2486                                                btree_insert_fn);
2487         }
2488
2489         if (ret) {
2490                 struct bkey *k;
2491
2492                 pr_err("error %i\n", ret);
2493
2494                 while ((k = bch_keylist_pop(keys)))
2495                         bkey_put(c, k);
2496         } else if (op.op.insert_collision)
2497                 ret = -ESRCH;
2498
2499         return ret;
2500 }
2501
2502 void bch_btree_set_root(struct btree *b)
2503 {
2504         unsigned int i;
2505         struct closure cl;
2506
2507         closure_init_stack(&cl);
2508
2509         trace_bcache_btree_set_root(b);
2510
2511         BUG_ON(!b->written);
2512
2513         for (i = 0; i < KEY_PTRS(&b->key); i++)
2514                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2515
2516         mutex_lock(&b->c->bucket_lock);
2517         list_del_init(&b->list);
2518         mutex_unlock(&b->c->bucket_lock);
2519
2520         b->c->root = b;
2521
2522         bch_journal_meta(b->c, &cl);
2523         closure_sync(&cl);
2524 }
2525
2526 /* Map across nodes or keys */
2527
2528 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2529                                        struct bkey *from,
2530                                        btree_map_nodes_fn *fn, int flags)
2531 {
2532         int ret = MAP_CONTINUE;
2533
2534         if (b->level) {
2535                 struct bkey *k;
2536                 struct btree_iter iter;
2537
2538                 bch_btree_iter_init(&b->keys, &iter, from);
2539
2540                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2541                                                        bch_ptr_bad))) {
2542                         ret = bcache_btree(map_nodes_recurse, k, b,
2543                                     op, from, fn, flags);
2544                         from = NULL;
2545
2546                         if (ret != MAP_CONTINUE)
2547                                 return ret;
2548                 }
2549         }
2550
2551         if (!b->level || flags == MAP_ALL_NODES)
2552                 ret = fn(op, b);
2553
2554         return ret;
2555 }
2556
2557 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2558                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2559 {
2560         return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
2561 }
2562
2563 int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2564                                       struct bkey *from, btree_map_keys_fn *fn,
2565                                       int flags)
2566 {
2567         int ret = MAP_CONTINUE;
2568         struct bkey *k;
2569         struct btree_iter iter;
2570
2571         bch_btree_iter_init(&b->keys, &iter, from);
2572
2573         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2574                 ret = !b->level
2575                         ? fn(op, b, k)
2576                         : bcache_btree(map_keys_recurse, k,
2577                                        b, op, from, fn, flags);
2578                 from = NULL;
2579
2580                 if (ret != MAP_CONTINUE)
2581                         return ret;
2582         }
2583
2584         if (!b->level && (flags & MAP_END_KEY))
2585                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2586                                      KEY_OFFSET(&b->key), 0));
2587
2588         return ret;
2589 }
2590
2591 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2592                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2593 {
2594         return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
2595 }
2596
2597 /* Keybuf code */
2598
2599 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2600 {
2601         /* Overlapping keys compare equal */
2602         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2603                 return -1;
2604         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2605                 return 1;
2606         return 0;
2607 }
2608
2609 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2610                                             struct keybuf_key *r)
2611 {
2612         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2613 }
2614
2615 struct refill {
2616         struct btree_op op;
2617         unsigned int    nr_found;
2618         struct keybuf   *buf;
2619         struct bkey     *end;
2620         keybuf_pred_fn  *pred;
2621 };
2622
2623 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2624                             struct bkey *k)
2625 {
2626         struct refill *refill = container_of(op, struct refill, op);
2627         struct keybuf *buf = refill->buf;
2628         int ret = MAP_CONTINUE;
2629
2630         if (bkey_cmp(k, refill->end) > 0) {
2631                 ret = MAP_DONE;
2632                 goto out;
2633         }
2634
2635         if (!KEY_SIZE(k)) /* end key */
2636                 goto out;
2637
2638         if (refill->pred(buf, k)) {
2639                 struct keybuf_key *w;
2640
2641                 spin_lock(&buf->lock);
2642
2643                 w = array_alloc(&buf->freelist);
2644                 if (!w) {
2645                         spin_unlock(&buf->lock);
2646                         return MAP_DONE;
2647                 }
2648
2649                 w->private = NULL;
2650                 bkey_copy(&w->key, k);
2651
2652                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2653                         array_free(&buf->freelist, w);
2654                 else
2655                         refill->nr_found++;
2656
2657                 if (array_freelist_empty(&buf->freelist))
2658                         ret = MAP_DONE;
2659
2660                 spin_unlock(&buf->lock);
2661         }
2662 out:
2663         buf->last_scanned = *k;
2664         return ret;
2665 }
2666
2667 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2668                        struct bkey *end, keybuf_pred_fn *pred)
2669 {
2670         struct bkey start = buf->last_scanned;
2671         struct refill refill;
2672
2673         cond_resched();
2674
2675         bch_btree_op_init(&refill.op, -1);
2676         refill.nr_found = 0;
2677         refill.buf      = buf;
2678         refill.end      = end;
2679         refill.pred     = pred;
2680
2681         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2682                            refill_keybuf_fn, MAP_END_KEY);
2683
2684         trace_bcache_keyscan(refill.nr_found,
2685                              KEY_INODE(&start), KEY_OFFSET(&start),
2686                              KEY_INODE(&buf->last_scanned),
2687                              KEY_OFFSET(&buf->last_scanned));
2688
2689         spin_lock(&buf->lock);
2690
2691         if (!RB_EMPTY_ROOT(&buf->keys)) {
2692                 struct keybuf_key *w;
2693
2694                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2695                 buf->start      = START_KEY(&w->key);
2696
2697                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2698                 buf->end        = w->key;
2699         } else {
2700                 buf->start      = MAX_KEY;
2701                 buf->end        = MAX_KEY;
2702         }
2703
2704         spin_unlock(&buf->lock);
2705 }
2706
2707 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2708 {
2709         rb_erase(&w->node, &buf->keys);
2710         array_free(&buf->freelist, w);
2711 }
2712
2713 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2714 {
2715         spin_lock(&buf->lock);
2716         __bch_keybuf_del(buf, w);
2717         spin_unlock(&buf->lock);
2718 }
2719
2720 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2721                                   struct bkey *end)
2722 {
2723         bool ret = false;
2724         struct keybuf_key *p, *w, s;
2725
2726         s.key = *start;
2727
2728         if (bkey_cmp(end, &buf->start) <= 0 ||
2729             bkey_cmp(start, &buf->end) >= 0)
2730                 return false;
2731
2732         spin_lock(&buf->lock);
2733         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2734
2735         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2736                 p = w;
2737                 w = RB_NEXT(w, node);
2738
2739                 if (p->private)
2740                         ret = true;
2741                 else
2742                         __bch_keybuf_del(buf, p);
2743         }
2744
2745         spin_unlock(&buf->lock);
2746         return ret;
2747 }
2748
2749 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2750 {
2751         struct keybuf_key *w;
2752
2753         spin_lock(&buf->lock);
2754
2755         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2756
2757         while (w && w->private)
2758                 w = RB_NEXT(w, node);
2759
2760         if (w)
2761                 w->private = ERR_PTR(-EINTR);
2762
2763         spin_unlock(&buf->lock);
2764         return w;
2765 }
2766
2767 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2768                                           struct keybuf *buf,
2769                                           struct bkey *end,
2770                                           keybuf_pred_fn *pred)
2771 {
2772         struct keybuf_key *ret;
2773
2774         while (1) {
2775                 ret = bch_keybuf_next(buf);
2776                 if (ret)
2777                         break;
2778
2779                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2780                         pr_debug("scan finished\n");
2781                         break;
2782                 }
2783
2784                 bch_refill_keybuf(c, buf, end, pred);
2785         }
2786
2787         return ret;
2788 }
2789
2790 void bch_keybuf_init(struct keybuf *buf)
2791 {
2792         buf->last_scanned       = MAX_KEY;
2793         buf->keys               = RB_ROOT;
2794
2795         spin_lock_init(&buf->lock);
2796         array_allocator_init(&buf->freelist);
2797 }
2798
2799 void bch_btree_exit(void)
2800 {
2801         if (btree_io_wq)
2802                 destroy_workqueue(btree_io_wq);
2803 }
2804
2805 int __init bch_btree_init(void)
2806 {
2807         btree_io_wq = alloc_workqueue("bch_btree_io", WQ_MEM_RECLAIM, 0);
2808         if (!btree_io_wq)
2809                 return -ENOMEM;
2810
2811         return 0;
2812 }