drivers/md/bcache/btree.c

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