drivers/md/bcache/btree.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  *
   4  * Uses a block device as cache for other block devices; optimized for SSDs.
   5  * All allocation is done in buckets, which should match the erase block size
   6  * of the device.
   7  *
   8  * Buckets containing cached data are kept on a heap sorted by priority;
   9  * bucket priority is increased on cache hit, and periodically all the buckets
  10  * on the heap have their priority scaled down. This currently is just used as
  11  * an LRU but in the future should allow for more intelligent heuristics.
  12  *
  13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  14  * counter. Garbage collection is used to remove stale pointers.
  15  *
  16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  17  * as keys are inserted we only sort the pages that have not yet been written.
  18  * When garbage collection is run, we resort the entire node.
  19  *
  20  * All configuration is done via sysfs; see Documentation/bcache.txt.
  21  */
  22
  23 #include "bcache.h"
  24 #include "btree.h"
  25 #include "debug.h"
  26 #include "extents.h"
  27
  28 #include <linux/slab.h>
  29 #include <linux/bitops.h>
  30 #include <linux/hash.h>
  31 #include <linux/kthread.h>
  32 #include <linux/prefetch.h>
  33 #include <linux/random.h>
  34 #include <linux/rcupdate.h>
  35 #include <linux/sched/clock.h>
  36 #include <trace/events/bcache.h>
  37
  38 /*
  39  * Todo:
  40  * register_bcache: Return errors out to userspace correctly
  41  *
  42  * Writeback: don't undirty key until after a cache flush
  43  *
  44  * Create an iterator for key pointers
  45  *
  46  * On btree write error, mark bucket such that it won't be freed from the cache
  47  *
  48  * Journalling:
  49  *   Check for bad keys in replay
  50  *   Propagate barriers
  51  *   Refcount journal entries in journal_replay
  52  *
  53  * Garbage collection:
  54  *   Finish incremental gc
  55  *   Gc should free old UUIDs, data for invalid UUIDs
  56  *
  57  * Provide a way to list backing device UUIDs we have data cached for, and
  58  * probably how long it's been since we've seen them, and a way to invalidate
  59  * dirty data for devices that will never be attached again
  60  *
  61  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  62  * that based on that and how much dirty data we have we can keep writeback
  63  * from being starved
  64  *
  65  * Add a tracepoint or somesuch to watch for writeback starvation
  66  *
  67  * When btree depth > 1 and splitting an interior node, we have to make sure
  68  * alloc_bucket() cannot fail. This should be true but is not completely
  69  * obvious.
  70  *
  71  * Plugging?
  72  *
  73  * If data write is less than hard sector size of ssd, round up offset in open
  74  * bucket to the next whole sector
  75  *
  76  * Superblock needs to be fleshed out for multiple cache devices
  77  *
  78  * Add a sysfs tunable for the number of writeback IOs in flight
  79  *
  80  * Add a sysfs tunable for the number of open data buckets
  81  *
  82  * IO tracking: Can we track when one process is doing io on behalf of another?
  83  * IO tracking: Don't use just an average, weigh more recent stuff higher
  84  *
  85  * Test module load/unload
  86  */
  87
  88 #define MAX_NEED_GC             64
  89 #define MAX_SAVE_PRIO           72
  90
  91 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  92
  93 #define PTR_HASH(c, k)                                                  \
  94         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  95
  96 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
  97
  98 /*
  99  * These macros are for recursing down the btree - they handle the details of
 100  * locking and looking up nodes in the cache for you. They're best treated as
 101  * mere syntax when reading code that uses them.
 102  *
 103  * op->lock determines whether we take a read or a write lock at a given depth.
 104  * If you've got a read lock and find that you need a write lock (i.e. you're
 105  * going to have to split), set op->lock and return -EINTR; btree_root() will
 106  * call you again and you'll have the correct lock.
 107  */
 108
 109 /**
 110  * btree - recurse down the btree on a specified key
 111  * @fn:         function to call, which will be passed the child node
 112  * @key:        key to recurse on
 113  * @b:          parent btree node
 114  * @op:         pointer to struct btree_op
 115  */
 116 #define btree(fn, key, b, op, ...)                                      \
 117 ({                                                                      \
 118         int _r, l = (b)->level - 1;                                     \
 119         bool _w = l <= (op)->lock;                                      \
 120         struct btree *_child = bch_btree_node_get((b)->c, op, key, l,   \
 121                                                   _w, b);               \
 122         if (!IS_ERR(_child)) {                                          \
 123                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 124                 rw_unlock(_w, _child);                                  \
 125         } else                                                          \
 126                 _r = PTR_ERR(_child);                                   \
 127         _r;                                                             \
 128 })
 129
 130 /**
 131  * btree_root - call a function on the root of the btree
 132  * @fn:         function to call, which will be passed the child node
 133  * @c:          cache set
 134  * @op:         pointer to struct btree_op
 135  */
 136 #define btree_root(fn, c, op, ...)                                      \
 137 ({                                                                      \
 138         int _r = -EINTR;                                                \
 139         do {                                                            \
 140                 struct btree *_b = (c)->root;                           \
 141                 bool _w = insert_lock(op, _b);                          \
 142                 rw_lock(_w, _b, _b->level);                             \
 143                 if (_b == (c)->root &&                                  \
 144                     _w == insert_lock(op, _b)) {                        \
 145                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 146                 }                                                       \
 147                 rw_unlock(_w, _b);                                      \
 148                 bch_cannibalize_unlock(c);                              \
 149                 if (_r == -EINTR)                                       \
 150                         schedule();                                     \
 151         } while (_r == -EINTR);                                         \
 152                                                                         \
 153         finish_wait(&(c)->btree_cache_wait, &(op)->wait);               \
 154         _r;                                                             \
 155 })
 156
 157 static inline struct bset *write_block(struct btree *b)
 158 {
 159         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 160 }
 161
 162 static void bch_btree_init_next(struct btree *b)
 163 {
 164         /* If not a leaf node, always sort */
 165         if (b->level && b->keys.nsets)
 166                 bch_btree_sort(&b->keys, &b->c->sort);
 167         else
 168                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 169
 170         if (b->written < btree_blocks(b))
 171                 bch_bset_init_next(&b->keys, write_block(b),
 172                                    bset_magic(&b->c->sb));
 173
 174 }
 175
 176 /* Btree key manipulation */
 177
 178 void bkey_put(struct cache_set *c, struct bkey *k)
 179 {
 180         unsigned i;
 181
 182         for (i = 0; i < KEY_PTRS(k); i++)
 183                 if (ptr_available(c, k, i))
 184                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 185 }
 186
 187 /* Btree IO */
 188
 189 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 190 {
 191         uint64_t crc = b->key.ptr[0];
 192         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 193
 194         crc = bch_crc64_update(crc, data, end - data);
 195         return crc ^ 0xffffffffffffffffULL;
 196 }
 197
 198 void bch_btree_node_read_done(struct btree *b)
 199 {
 200         const char *err = "bad btree header";
 201         struct bset *i = btree_bset_first(b);
 202         struct btree_iter *iter;
 203
 204         iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
 205         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 206         iter->used = 0;
 207
 208 #ifdef CONFIG_BCACHE_DEBUG
 209         iter->b = &b->keys;
 210 #endif
 211
 212         if (!i->seq)
 213                 goto err;
 214
 215         for (;
 216              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 217              i = write_block(b)) {
 218                 err = "unsupported bset version";
 219                 if (i->version > BCACHE_BSET_VERSION)
 220                         goto err;
 221
 222                 err = "bad btree header";
 223                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 224                     btree_blocks(b))
 225                         goto err;
 226
 227                 err = "bad magic";
 228                 if (i->magic != bset_magic(&b->c->sb))
 229                         goto err;
 230
 231                 err = "bad checksum";
 232                 switch (i->version) {
 233                 case 0:
 234                         if (i->csum != csum_set(i))
 235                                 goto err;
 236                         break;
 237                 case BCACHE_BSET_VERSION:
 238                         if (i->csum != btree_csum_set(b, i))
 239                                 goto err;
 240                         break;
 241                 }
 242
 243                 err = "empty set";
 244                 if (i != b->keys.set[0].data && !i->keys)
 245                         goto err;
 246
 247                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 248
 249                 b->written += set_blocks(i, block_bytes(b->c));
 250         }
 251
 252         err = "corrupted btree";
 253         for (i = write_block(b);
 254              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 255              i = ((void *) i) + block_bytes(b->c))
 256                 if (i->seq == b->keys.set[0].data->seq)
 257                         goto err;
 258
 259         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 260
 261         i = b->keys.set[0].data;
 262         err = "short btree key";
 263         if (b->keys.set[0].size &&
 264             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 265                 goto err;
 266
 267         if (b->written < btree_blocks(b))
 268                 bch_bset_init_next(&b->keys, write_block(b),
 269                                    bset_magic(&b->c->sb));
 270 out:
 271         mempool_free(iter, b->c->fill_iter);
 272         return;
 273 err:
 274         set_btree_node_io_error(b);
 275         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 276                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 277                             bset_block_offset(b, i), i->keys);
 278         goto out;
 279 }
 280
 281 static void btree_node_read_endio(struct bio *bio)
 282 {
 283         struct closure *cl = bio->bi_private;
 284         closure_put(cl);
 285 }
 286
 287 static void bch_btree_node_read(struct btree *b)
 288 {
 289         uint64_t start_time = local_clock();
 290         struct closure cl;
 291         struct bio *bio;
 292
 293         trace_bcache_btree_read(b);
 294
 295         closure_init_stack(&cl);
 296
 297         bio = bch_bbio_alloc(b->c);
 298         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 299         bio->bi_end_io  = btree_node_read_endio;
 300         bio->bi_private = &cl;
 301         bio->bi_opf = REQ_OP_READ | REQ_META;
 302
 303         bch_bio_map(bio, b->keys.set[0].data);
 304
 305         bch_submit_bbio(bio, b->c, &b->key, 0);
 306         closure_sync(&cl);
 307
 308         if (bio->bi_error)
 309                 set_btree_node_io_error(b);
 310
 311         bch_bbio_free(bio, b->c);
 312
 313         if (btree_node_io_error(b))
 314                 goto err;
 315
 316         bch_btree_node_read_done(b);
 317         bch_time_stats_update(&b->c->btree_read_time, start_time);
 318
 319         return;
 320 err:
 321         bch_cache_set_error(b->c, "io error reading bucket %zu",
 322                             PTR_BUCKET_NR(b->c, &b->key, 0));
 323 }
 324
 325 static void btree_complete_write(struct btree *b, struct btree_write *w)
 326 {
 327         if (w->prio_blocked &&
 328             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 329                 wake_up_allocators(b->c);
 330
 331         if (w->journal) {
 332                 atomic_dec_bug(w->journal);
 333                 __closure_wake_up(&b->c->journal.wait);
 334         }
 335
 336         w->prio_blocked = 0;
 337         w->journal      = NULL;
 338 }
 339
 340 static void btree_node_write_unlock(struct closure *cl)
 341 {
 342         struct btree *b = container_of(cl, struct btree, io);
 343
 344         up(&b->io_mutex);
 345 }
 346
 347 static void __btree_node_write_done(struct closure *cl)
 348 {
 349         struct btree *b = container_of(cl, struct btree, io);
 350         struct btree_write *w = btree_prev_write(b);
 351
 352         bch_bbio_free(b->bio, b->c);
 353         b->bio = NULL;
 354         btree_complete_write(b, w);
 355
 356         if (btree_node_dirty(b))
 357                 schedule_delayed_work(&b->work, 30 * HZ);
 358
 359         closure_return_with_destructor(cl, btree_node_write_unlock);
 360 }
 361
 362 static void btree_node_write_done(struct closure *cl)
 363 {
 364         struct btree *b = container_of(cl, struct btree, io);
 365
 366         bio_free_pages(b->bio);
 367         __btree_node_write_done(cl);
 368 }
 369
 370 static void btree_node_write_endio(struct bio *bio)
 371 {
 372         struct closure *cl = bio->bi_private;
 373         struct btree *b = container_of(cl, struct btree, io);
 374
 375         if (bio->bi_error)
 376                 set_btree_node_io_error(b);
 377
 378         bch_bbio_count_io_errors(b->c, bio, bio->bi_error, "writing btree");
 379         closure_put(cl);
 380 }
 381
 382 static void do_btree_node_write(struct btree *b)
 383 {
 384         struct closure *cl = &b->io;
 385         struct bset *i = btree_bset_last(b);
 386         BKEY_PADDED(key) k;
 387
 388         i->version      = BCACHE_BSET_VERSION;
 389         i->csum         = btree_csum_set(b, i);
 390
 391         BUG_ON(b->bio);
 392         b->bio = bch_bbio_alloc(b->c);
 393
 394         b->bio->bi_end_io       = btree_node_write_endio;
 395         b->bio->bi_private      = cl;
 396         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 397         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 398         bch_bio_map(b->bio, i);
 399
 400         /*
 401          * If we're appending to a leaf node, we don't technically need FUA -
 402          * this write just needs to be persisted before the next journal write,
 403          * which will be marked FLUSH|FUA.
 404          *
 405          * Similarly if we're writing a new btree root - the pointer is going to
 406          * be in the next journal entry.
 407          *
 408          * But if we're writing a new btree node (that isn't a root) or
 409          * appending to a non leaf btree node, we need either FUA or a flush
 410          * when we write the parent with the new pointer. FUA is cheaper than a
 411          * flush, and writes appending to leaf nodes aren't blocking anything so
 412          * just make all btree node writes FUA to keep things sane.
 413          */
 414
 415         bkey_copy(&k.key, &b->key);
 416         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 417                        bset_sector_offset(&b->keys, i));
 418
 419         if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 420                 int j;
 421                 struct bio_vec *bv;
 422                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 423
 424                 bio_for_each_segment_all(bv, b->bio, j)
 425                         memcpy(page_address(bv->bv_page),
 426                                base + j * PAGE_SIZE, PAGE_SIZE);
 427
 428                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 429
 430                 continue_at(cl, btree_node_write_done, NULL);
 431         } else {
 432                 b->bio->bi_vcnt = 0;
 433                 bch_bio_map(b->bio, i);
 434
 435                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 436
 437                 closure_sync(cl);
 438                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 439         }
 440 }
 441
 442 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 443 {
 444         struct bset *i = btree_bset_last(b);
 445
 446         lockdep_assert_held(&b->write_lock);
 447
 448         trace_bcache_btree_write(b);
 449
 450         BUG_ON(current->bio_list);
 451         BUG_ON(b->written >= btree_blocks(b));
 452         BUG_ON(b->written && !i->keys);
 453         BUG_ON(btree_bset_first(b)->seq != i->seq);
 454         bch_check_keys(&b->keys, "writing");
 455
 456         cancel_delayed_work(&b->work);
 457
 458         /* If caller isn't waiting for write, parent refcount is cache set */
 459         down(&b->io_mutex);
 460         closure_init(&b->io, parent ?: &b->c->cl);
 461
 462         clear_bit(BTREE_NODE_dirty,      &b->flags);
 463         change_bit(BTREE_NODE_write_idx, &b->flags);
 464
 465         do_btree_node_write(b);
 466
 467         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 468                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 469
 470         b->written += set_blocks(i, block_bytes(b->c));
 471 }
 472
 473 void bch_btree_node_write(struct btree *b, struct closure *parent)
 474 {
 475         unsigned nsets = b->keys.nsets;
 476
 477         lockdep_assert_held(&b->lock);
 478
 479         __bch_btree_node_write(b, parent);
 480
 481         /*
 482          * do verify if there was more than one set initially (i.e. we did a
 483          * sort) and we sorted down to a single set:
 484          */
 485         if (nsets && !b->keys.nsets)
 486                 bch_btree_verify(b);
 487
 488         bch_btree_init_next(b);
 489 }
 490
 491 static void bch_btree_node_write_sync(struct btree *b)
 492 {
 493         struct closure cl;
 494
 495         closure_init_stack(&cl);
 496
 497         mutex_lock(&b->write_lock);
 498         bch_btree_node_write(b, &cl);
 499         mutex_unlock(&b->write_lock);
 500
 501         closure_sync(&cl);
 502 }
 503
 504 static void btree_node_write_work(struct work_struct *w)
 505 {
 506         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 507
 508         mutex_lock(&b->write_lock);
 509         if (btree_node_dirty(b))
 510                 __bch_btree_node_write(b, NULL);
 511         mutex_unlock(&b->write_lock);
 512 }
 513
 514 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 515 {
 516         struct bset *i = btree_bset_last(b);
 517         struct btree_write *w = btree_current_write(b);
 518
 519         lockdep_assert_held(&b->write_lock);
 520
 521         BUG_ON(!b->written);
 522         BUG_ON(!i->keys);
 523
 524         if (!btree_node_dirty(b))
 525                 schedule_delayed_work(&b->work, 30 * HZ);
 526
 527         set_btree_node_dirty(b);
 528
 529         if (journal_ref) {
 530                 if (w->journal &&
 531                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 532                         atomic_dec_bug(w->journal);
 533                         w->journal = NULL;
 534                 }
 535
 536                 if (!w->journal) {
 537                         w->journal = journal_ref;
 538                         atomic_inc(w->journal);
 539                 }
 540         }
 541
 542         /* Force write if set is too big */
 543         if (set_bytes(i) > PAGE_SIZE - 48 &&
 544             !current->bio_list)
 545                 bch_btree_node_write(b, NULL);
 546 }
 547
 548 /*
 549  * Btree in memory cache - allocation/freeing
 550  * mca -> memory cache
 551  */
 552
 553 #define mca_reserve(c)  (((c->root && c->root->level)           \
 554                           ? c->root->level : 1) * 8 + 16)
 555 #define mca_can_free(c)                                         \
 556         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 557
 558 static void mca_data_free(struct btree *b)
 559 {
 560         BUG_ON(b->io_mutex.count != 1);
 561
 562         bch_btree_keys_free(&b->keys);
 563
 564         b->c->btree_cache_used--;
 565         list_move(&b->list, &b->c->btree_cache_freed);
 566 }
 567
 568 static void mca_bucket_free(struct btree *b)
 569 {
 570         BUG_ON(btree_node_dirty(b));
 571
 572         b->key.ptr[0] = 0;
 573         hlist_del_init_rcu(&b->hash);
 574         list_move(&b->list, &b->c->btree_cache_freeable);
 575 }
 576
 577 static unsigned btree_order(struct bkey *k)
 578 {
 579         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 580 }
 581
 582 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 583 {
 584         if (!bch_btree_keys_alloc(&b->keys,
 585                                   max_t(unsigned,
 586                                         ilog2(b->c->btree_pages),
 587                                         btree_order(k)),
 588                                   gfp)) {
 589                 b->c->btree_cache_used++;
 590                 list_move(&b->list, &b->c->btree_cache);
 591         } else {
 592                 list_move(&b->list, &b->c->btree_cache_freed);
 593         }
 594 }
 595
 596 static struct btree *mca_bucket_alloc(struct cache_set *c,
 597                                       struct bkey *k, gfp_t gfp)
 598 {
 599         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 600         if (!b)
 601                 return NULL;
 602
 603         init_rwsem(&b->lock);
 604         lockdep_set_novalidate_class(&b->lock);
 605         mutex_init(&b->write_lock);
 606         lockdep_set_novalidate_class(&b->write_lock);
 607         INIT_LIST_HEAD(&b->list);
 608         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 609         b->c = c;
 610         sema_init(&b->io_mutex, 1);
 611
 612         mca_data_alloc(b, k, gfp);
 613         return b;
 614 }
 615
 616 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 617 {
 618         struct closure cl;
 619
 620         closure_init_stack(&cl);
 621         lockdep_assert_held(&b->c->bucket_lock);
 622
 623         if (!down_write_trylock(&b->lock))
 624                 return -ENOMEM;
 625
 626         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 627
 628         if (b->keys.page_order < min_order)
 629                 goto out_unlock;
 630
 631         if (!flush) {
 632                 if (btree_node_dirty(b))
 633                         goto out_unlock;
 634
 635                 if (down_trylock(&b->io_mutex))
 636                         goto out_unlock;
 637                 up(&b->io_mutex);
 638         }
 639
 640         mutex_lock(&b->write_lock);
 641         if (btree_node_dirty(b))
 642                 __bch_btree_node_write(b, &cl);
 643         mutex_unlock(&b->write_lock);
 644
 645         closure_sync(&cl);
 646
 647         /* wait for any in flight btree write */
 648         down(&b->io_mutex);
 649         up(&b->io_mutex);
 650
 651         return 0;
 652 out_unlock:
 653         rw_unlock(true, b);
 654         return -ENOMEM;
 655 }
 656
 657 static unsigned long bch_mca_scan(struct shrinker *shrink,
 658                                   struct shrink_control *sc)
 659 {
 660         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 661         struct btree *b, *t;
 662         unsigned long i, nr = sc->nr_to_scan;
 663         unsigned long freed = 0;
 664
 665         if (c->shrinker_disabled)
 666                 return SHRINK_STOP;
 667
 668         if (c->btree_cache_alloc_lock)
 669                 return SHRINK_STOP;
 670
 671         /* Return -1 if we can't do anything right now */
 672         if (sc->gfp_mask & __GFP_IO)
 673                 mutex_lock(&c->bucket_lock);
 674         else if (!mutex_trylock(&c->bucket_lock))
 675                 return -1;
 676
 677         /*
 678          * It's _really_ critical that we don't free too many btree nodes - we
 679          * have to always leave ourselves a reserve. The reserve is how we
 680          * guarantee that allocating memory for a new btree node can always
 681          * succeed, so that inserting keys into the btree can always succeed and
 682          * IO can always make forward progress:
 683          */
 684         nr /= c->btree_pages;
 685         nr = min_t(unsigned long, nr, mca_can_free(c));
 686
 687         i = 0;
 688         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 689                 if (freed >= nr)
 690                         break;
 691
 692                 if (++i > 3 &&
 693                     !mca_reap(b, 0, false)) {
 694                         mca_data_free(b);
 695                         rw_unlock(true, b);
 696                         freed++;
 697                 }
 698         }
 699
 700         for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
 701                 if (list_empty(&c->btree_cache))
 702                         goto out;
 703
 704                 b = list_first_entry(&c->btree_cache, struct btree, list);
 705                 list_rotate_left(&c->btree_cache);
 706
 707                 if (!b->accessed &&
 708                     !mca_reap(b, 0, false)) {
 709                         mca_bucket_free(b);
 710                         mca_data_free(b);
 711                         rw_unlock(true, b);
 712                         freed++;
 713                 } else
 714                         b->accessed = 0;
 715         }
 716 out:
 717         mutex_unlock(&c->bucket_lock);
 718         return freed;
 719 }
 720
 721 static unsigned long bch_mca_count(struct shrinker *shrink,
 722                                    struct shrink_control *sc)
 723 {
 724         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 725
 726         if (c->shrinker_disabled)
 727                 return 0;
 728
 729         if (c->btree_cache_alloc_lock)
 730                 return 0;
 731
 732         return mca_can_free(c) * c->btree_pages;
 733 }
 734
 735 void bch_btree_cache_free(struct cache_set *c)
 736 {
 737         struct btree *b;
 738         struct closure cl;
 739         closure_init_stack(&cl);
 740
 741         if (c->shrink.list.next)
 742                 unregister_shrinker(&c->shrink);
 743
 744         mutex_lock(&c->bucket_lock);
 745
 746 #ifdef CONFIG_BCACHE_DEBUG
 747         if (c->verify_data)
 748                 list_move(&c->verify_data->list, &c->btree_cache);
 749
 750         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 751 #endif
 752
 753         list_splice(&c->btree_cache_freeable,
 754                     &c->btree_cache);
 755
 756         while (!list_empty(&c->btree_cache)) {
 757                 b = list_first_entry(&c->btree_cache, struct btree, list);
 758
 759                 if (btree_node_dirty(b))
 760                         btree_complete_write(b, btree_current_write(b));
 761                 clear_bit(BTREE_NODE_dirty, &b->flags);
 762
 763                 mca_data_free(b);
 764         }
 765
 766         while (!list_empty(&c->btree_cache_freed)) {
 767                 b = list_first_entry(&c->btree_cache_freed,
 768                                      struct btree, list);
 769                 list_del(&b->list);
 770                 cancel_delayed_work_sync(&b->work);
 771                 kfree(b);
 772         }
 773
 774         mutex_unlock(&c->bucket_lock);
 775 }
 776
 777 int bch_btree_cache_alloc(struct cache_set *c)
 778 {
 779         unsigned i;
 780
 781         for (i = 0; i < mca_reserve(c); i++)
 782                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 783                         return -ENOMEM;
 784
 785         list_splice_init(&c->btree_cache,
 786                          &c->btree_cache_freeable);
 787
 788 #ifdef CONFIG_BCACHE_DEBUG
 789         mutex_init(&c->verify_lock);
 790
 791         c->verify_ondisk = (void *)
 792                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 793
 794         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 795
 796         if (c->verify_data &&
 797             c->verify_data->keys.set->data)
 798                 list_del_init(&c->verify_data->list);
 799         else
 800                 c->verify_data = NULL;
 801 #endif
 802
 803         c->shrink.count_objects = bch_mca_count;
 804         c->shrink.scan_objects = bch_mca_scan;
 805         c->shrink.seeks = 4;
 806         c->shrink.batch = c->btree_pages * 2;
 807         register_shrinker(&c->shrink);
 808
 809         return 0;
 810 }
 811
 812 /* Btree in memory cache - hash table */
 813
 814 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 815 {
 816         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 817 }
 818
 819 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 820 {
 821         struct btree *b;
 822
 823         rcu_read_lock();
 824         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 825                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 826                         goto out;
 827         b = NULL;
 828 out:
 829         rcu_read_unlock();
 830         return b;
 831 }
 832
 833 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 834 {
 835         struct task_struct *old;
 836
 837         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 838         if (old && old != current) {
 839                 if (op)
 840                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 841                                         TASK_UNINTERRUPTIBLE);
 842                 return -EINTR;
 843         }
 844
 845         return 0;
 846 }
 847
 848 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 849                                      struct bkey *k)
 850 {
 851         struct btree *b;
 852
 853         trace_bcache_btree_cache_cannibalize(c);
 854
 855         if (mca_cannibalize_lock(c, op))
 856                 return ERR_PTR(-EINTR);
 857
 858         list_for_each_entry_reverse(b, &c->btree_cache, list)
 859                 if (!mca_reap(b, btree_order(k), false))
 860                         return b;
 861
 862         list_for_each_entry_reverse(b, &c->btree_cache, list)
 863                 if (!mca_reap(b, btree_order(k), true))
 864                         return b;
 865
 866         WARN(1, "btree cache cannibalize failed\n");
 867         return ERR_PTR(-ENOMEM);
 868 }
 869
 870 /*
 871  * We can only have one thread cannibalizing other cached btree nodes at a time,
 872  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 873  * cannibalize_bucket() will take. This means every time we unlock the root of
 874  * the btree, we need to release this lock if we have it held.
 875  */
 876 static void bch_cannibalize_unlock(struct cache_set *c)
 877 {
 878         if (c->btree_cache_alloc_lock == current) {
 879                 c->btree_cache_alloc_lock = NULL;
 880                 wake_up(&c->btree_cache_wait);
 881         }
 882 }
 883
 884 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 885                                struct bkey *k, int level)
 886 {
 887         struct btree *b;
 888
 889         BUG_ON(current->bio_list);
 890
 891         lockdep_assert_held(&c->bucket_lock);
 892
 893         if (mca_find(c, k))
 894                 return NULL;
 895
 896         /* btree_free() doesn't free memory; it sticks the node on the end of
 897          * the list. Check if there's any freed nodes there:
 898          */
 899         list_for_each_entry(b, &c->btree_cache_freeable, list)
 900                 if (!mca_reap(b, btree_order(k), false))
 901                         goto out;
 902
 903         /* We never free struct btree itself, just the memory that holds the on
 904          * disk node. Check the freed list before allocating a new one:
 905          */
 906         list_for_each_entry(b, &c->btree_cache_freed, list)
 907                 if (!mca_reap(b, 0, false)) {
 908                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 909                         if (!b->keys.set[0].data)
 910                                 goto err;
 911                         else
 912                                 goto out;
 913                 }
 914
 915         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 916         if (!b)
 917                 goto err;
 918
 919         BUG_ON(!down_write_trylock(&b->lock));
 920         if (!b->keys.set->data)
 921                 goto err;
 922 out:
 923         BUG_ON(b->io_mutex.count != 1);
 924
 925         bkey_copy(&b->key, k);
 926         list_move(&b->list, &c->btree_cache);
 927         hlist_del_init_rcu(&b->hash);
 928         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 929
 930         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 931         b->parent       = (void *) ~0UL;
 932         b->flags        = 0;
 933         b->written      = 0;
 934         b->level        = level;
 935
 936         if (!b->level)
 937                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 938                                     &b->c->expensive_debug_checks);
 939         else
 940                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 941                                     &b->c->expensive_debug_checks);
 942
 943         return b;
 944 err:
 945         if (b)
 946                 rw_unlock(true, b);
 947
 948         b = mca_cannibalize(c, op, k);
 949         if (!IS_ERR(b))
 950                 goto out;
 951
 952         return b;
 953 }
 954
 955 /**
 956  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 957  * in from disk if necessary.
 958  *
 959  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 960  *
 961  * The btree node will have either a read or a write lock held, depending on
 962  * level and op->lock.
 963  */
 964 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 965                                  struct bkey *k, int level, bool write,
 966                                  struct btree *parent)
 967 {
 968         int i = 0;
 969         struct btree *b;
 970
 971         BUG_ON(level < 0);
 972 retry:
 973         b = mca_find(c, k);
 974
 975         if (!b) {
 976                 if (current->bio_list)
 977                         return ERR_PTR(-EAGAIN);
 978
 979                 mutex_lock(&c->bucket_lock);
 980                 b = mca_alloc(c, op, k, level);
 981                 mutex_unlock(&c->bucket_lock);
 982
 983                 if (!b)
 984                         goto retry;
 985                 if (IS_ERR(b))
 986                         return b;
 987
 988                 bch_btree_node_read(b);
 989
 990                 if (!write)
 991                         downgrade_write(&b->lock);
 992         } else {
 993                 rw_lock(write, b, level);
 994                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 995                         rw_unlock(write, b);
 996                         goto retry;
 997                 }
 998                 BUG_ON(b->level != level);
 999         }
1000
1001         b->parent = parent;
1002         b->accessed = 1;
1003
1004         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1005                 prefetch(b->keys.set[i].tree);
1006                 prefetch(b->keys.set[i].data);
1007         }
1008
1009         for (; i <= b->keys.nsets; i++)
1010                 prefetch(b->keys.set[i].data);
1011
1012         if (btree_node_io_error(b)) {
1013                 rw_unlock(write, b);
1014                 return ERR_PTR(-EIO);
1015         }
1016
1017         BUG_ON(!b->written);
1018
1019         return b;
1020 }
1021
1022 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1023 {
1024         struct btree *b;
1025
1026         mutex_lock(&parent->c->bucket_lock);
1027         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1028         mutex_unlock(&parent->c->bucket_lock);
1029
1030         if (!IS_ERR_OR_NULL(b)) {
1031                 b->parent = parent;
1032                 bch_btree_node_read(b);
1033                 rw_unlock(true, b);
1034         }
1035 }
1036
1037 /* Btree alloc */
1038
1039 static void btree_node_free(struct btree *b)
1040 {
1041         trace_bcache_btree_node_free(b);
1042
1043         BUG_ON(b == b->c->root);
1044
1045         mutex_lock(&b->write_lock);
1046
1047         if (btree_node_dirty(b))
1048                 btree_complete_write(b, btree_current_write(b));
1049         clear_bit(BTREE_NODE_dirty, &b->flags);
1050
1051         mutex_unlock(&b->write_lock);
1052
1053         cancel_delayed_work(&b->work);
1054
1055         mutex_lock(&b->c->bucket_lock);
1056         bch_bucket_free(b->c, &b->key);
1057         mca_bucket_free(b);
1058         mutex_unlock(&b->c->bucket_lock);
1059 }
1060
1061 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1062                                      int level, bool wait,
1063                                      struct btree *parent)
1064 {
1065         BKEY_PADDED(key) k;
1066         struct btree *b = ERR_PTR(-EAGAIN);
1067
1068         mutex_lock(&c->bucket_lock);
1069 retry:
1070         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1071                 goto err;
1072
1073         bkey_put(c, &k.key);
1074         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1075
1076         b = mca_alloc(c, op, &k.key, level);
1077         if (IS_ERR(b))
1078                 goto err_free;
1079
1080         if (!b) {
1081                 cache_bug(c,
1082                         "Tried to allocate bucket that was in btree cache");
1083                 goto retry;
1084         }
1085
1086         b->accessed = 1;
1087         b->parent = parent;
1088         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1089
1090         mutex_unlock(&c->bucket_lock);
1091
1092         trace_bcache_btree_node_alloc(b);
1093         return b;
1094 err_free:
1095         bch_bucket_free(c, &k.key);
1096 err:
1097         mutex_unlock(&c->bucket_lock);
1098
1099         trace_bcache_btree_node_alloc_fail(c);
1100         return b;
1101 }
1102
1103 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1104                                           struct btree_op *op, int level,
1105                                           struct btree *parent)
1106 {
1107         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1108 }
1109
1110 static struct btree *btree_node_alloc_replacement(struct btree *b,
1111                                                   struct btree_op *op)
1112 {
1113         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1114         if (!IS_ERR_OR_NULL(n)) {
1115                 mutex_lock(&n->write_lock);
1116                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1117                 bkey_copy_key(&n->key, &b->key);
1118                 mutex_unlock(&n->write_lock);
1119         }
1120
1121         return n;
1122 }
1123
1124 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1125 {
1126         unsigned i;
1127
1128         mutex_lock(&b->c->bucket_lock);
1129
1130         atomic_inc(&b->c->prio_blocked);
1131
1132         bkey_copy(k, &b->key);
1133         bkey_copy_key(k, &ZERO_KEY);
1134
1135         for (i = 0; i < KEY_PTRS(k); i++)
1136                 SET_PTR_GEN(k, i,
1137                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1138                                         PTR_BUCKET(b->c, &b->key, i)));
1139
1140         mutex_unlock(&b->c->bucket_lock);
1141 }
1142
1143 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1144 {
1145         struct cache_set *c = b->c;
1146         struct cache *ca;
1147         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1148
1149         mutex_lock(&c->bucket_lock);
1150
1151         for_each_cache(ca, c, i)
1152                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1153                         if (op)
1154                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1155                                                 TASK_UNINTERRUPTIBLE);
1156                         mutex_unlock(&c->bucket_lock);
1157                         return -EINTR;
1158                 }
1159
1160         mutex_unlock(&c->bucket_lock);
1161
1162         return mca_cannibalize_lock(b->c, op);
1163 }
1164
1165 /* Garbage collection */
1166
1167 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1168                                     struct bkey *k)
1169 {
1170         uint8_t stale = 0;
1171         unsigned i;
1172         struct bucket *g;
1173
1174         /*
1175          * ptr_invalid() can't return true for the keys that mark btree nodes as
1176          * freed, but since ptr_bad() returns true we'll never actually use them
1177          * for anything and thus we don't want mark their pointers here
1178          */
1179         if (!bkey_cmp(k, &ZERO_KEY))
1180                 return stale;
1181
1182         for (i = 0; i < KEY_PTRS(k); i++) {
1183                 if (!ptr_available(c, k, i))
1184                         continue;
1185
1186                 g = PTR_BUCKET(c, k, i);
1187
1188                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1189                         g->last_gc = PTR_GEN(k, i);
1190
1191                 if (ptr_stale(c, k, i)) {
1192                         stale = max(stale, ptr_stale(c, k, i));
1193                         continue;
1194                 }
1195
1196                 cache_bug_on(GC_MARK(g) &&
1197                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1198                              c, "inconsistent ptrs: mark = %llu, level = %i",
1199                              GC_MARK(g), level);
1200
1201                 if (level)
1202                         SET_GC_MARK(g, GC_MARK_METADATA);
1203                 else if (KEY_DIRTY(k))
1204                         SET_GC_MARK(g, GC_MARK_DIRTY);
1205                 else if (!GC_MARK(g))
1206                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1207
1208                 /* guard against overflow */
1209                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1210                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1211                                              MAX_GC_SECTORS_USED));
1212
1213                 BUG_ON(!GC_SECTORS_USED(g));
1214         }
1215
1216         return stale;
1217 }
1218
1219 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1220
1221 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1222 {
1223         unsigned i;
1224
1225         for (i = 0; i < KEY_PTRS(k); i++)
1226                 if (ptr_available(c, k, i) &&
1227                     !ptr_stale(c, k, i)) {
1228                         struct bucket *b = PTR_BUCKET(c, k, i);
1229
1230                         b->gen = PTR_GEN(k, i);
1231
1232                         if (level && bkey_cmp(k, &ZERO_KEY))
1233                                 b->prio = BTREE_PRIO;
1234                         else if (!level && b->prio == BTREE_PRIO)
1235                                 b->prio = INITIAL_PRIO;
1236                 }
1237
1238         __bch_btree_mark_key(c, level, k);
1239 }
1240
1241 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1242 {
1243         uint8_t stale = 0;
1244         unsigned keys = 0, good_keys = 0;
1245         struct bkey *k;
1246         struct btree_iter iter;
1247         struct bset_tree *t;
1248
1249         gc->nodes++;
1250
1251         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1252                 stale = max(stale, btree_mark_key(b, k));
1253                 keys++;
1254
1255                 if (bch_ptr_bad(&b->keys, k))
1256                         continue;
1257
1258                 gc->key_bytes += bkey_u64s(k);
1259                 gc->nkeys++;
1260                 good_keys++;
1261
1262                 gc->data += KEY_SIZE(k);
1263         }
1264
1265         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1266                 btree_bug_on(t->size &&
1267                              bset_written(&b->keys, t) &&
1268                              bkey_cmp(&b->key, &t->end) < 0,
1269                              b, "found short btree key in gc");
1270
1271         if (b->c->gc_always_rewrite)
1272                 return true;
1273
1274         if (stale > 10)
1275                 return true;
1276
1277         if ((keys - good_keys) * 2 > keys)
1278                 return true;
1279
1280         return false;
1281 }
1282
1283 #define GC_MERGE_NODES  4U
1284
1285 struct gc_merge_info {
1286         struct btree    *b;
1287         unsigned        keys;
1288 };
1289
1290 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1291                                  struct keylist *, atomic_t *, struct bkey *);
1292
1293 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1294                              struct gc_stat *gc, struct gc_merge_info *r)
1295 {
1296         unsigned i, nodes = 0, keys = 0, blocks;
1297         struct btree *new_nodes[GC_MERGE_NODES];
1298         struct keylist keylist;
1299         struct closure cl;
1300         struct bkey *k;
1301
1302         bch_keylist_init(&keylist);
1303
1304         if (btree_check_reserve(b, NULL))
1305                 return 0;
1306
1307         memset(new_nodes, 0, sizeof(new_nodes));
1308         closure_init_stack(&cl);
1309
1310         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1311                 keys += r[nodes++].keys;
1312
1313         blocks = btree_default_blocks(b->c) * 2 / 3;
1314
1315         if (nodes < 2 ||
1316             __set_blocks(b->keys.set[0].data, keys,
1317                          block_bytes(b->c)) > blocks * (nodes - 1))
1318                 return 0;
1319
1320         for (i = 0; i < nodes; i++) {
1321                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1322                 if (IS_ERR_OR_NULL(new_nodes[i]))
1323                         goto out_nocoalesce;
1324         }
1325
1326         /*
1327          * We have to check the reserve here, after we've allocated our new
1328          * nodes, to make sure the insert below will succeed - we also check
1329          * before as an optimization to potentially avoid a bunch of expensive
1330          * allocs/sorts
1331          */
1332         if (btree_check_reserve(b, NULL))
1333                 goto out_nocoalesce;
1334
1335         for (i = 0; i < nodes; i++)
1336                 mutex_lock(&new_nodes[i]->write_lock);
1337
1338         for (i = nodes - 1; i > 0; --i) {
1339                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1340                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1341                 struct bkey *k, *last = NULL;
1342
1343                 keys = 0;
1344
1345                 if (i > 1) {
1346                         for (k = n2->start;
1347                              k < bset_bkey_last(n2);
1348                              k = bkey_next(k)) {
1349                                 if (__set_blocks(n1, n1->keys + keys +
1350                                                  bkey_u64s(k),
1351                                                  block_bytes(b->c)) > blocks)
1352                                         break;
1353
1354                                 last = k;
1355                                 keys += bkey_u64s(k);
1356                         }
1357                 } else {
1358                         /*
1359                          * Last node we're not getting rid of - we're getting
1360                          * rid of the node at r[0]. Have to try and fit all of
1361                          * the remaining keys into this node; we can't ensure
1362                          * they will always fit due to rounding and variable
1363                          * length keys (shouldn't be possible in practice,
1364                          * though)
1365                          */
1366                         if (__set_blocks(n1, n1->keys + n2->keys,
1367                                          block_bytes(b->c)) >
1368                             btree_blocks(new_nodes[i]))
1369                                 goto out_nocoalesce;
1370
1371                         keys = n2->keys;
1372                         /* Take the key of the node we're getting rid of */
1373                         last = &r->b->key;
1374                 }
1375
1376                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1377                        btree_blocks(new_nodes[i]));
1378
1379                 if (last)
1380                         bkey_copy_key(&new_nodes[i]->key, last);
1381
1382                 memcpy(bset_bkey_last(n1),
1383                        n2->start,
1384                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1385
1386                 n1->keys += keys;
1387                 r[i].keys = n1->keys;
1388
1389                 memmove(n2->start,
1390                         bset_bkey_idx(n2, keys),
1391                         (void *) bset_bkey_last(n2) -
1392                         (void *) bset_bkey_idx(n2, keys));
1393
1394                 n2->keys -= keys;
1395
1396                 if (__bch_keylist_realloc(&keylist,
1397                                           bkey_u64s(&new_nodes[i]->key)))
1398                         goto out_nocoalesce;
1399
1400                 bch_btree_node_write(new_nodes[i], &cl);
1401                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1402         }
1403
1404         for (i = 0; i < nodes; i++)
1405                 mutex_unlock(&new_nodes[i]->write_lock);
1406
1407         closure_sync(&cl);
1408
1409         /* We emptied out this node */
1410         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1411         btree_node_free(new_nodes[0]);
1412         rw_unlock(true, new_nodes[0]);
1413         new_nodes[0] = NULL;
1414
1415         for (i = 0; i < nodes; i++) {
1416                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1417                         goto out_nocoalesce;
1418
1419                 make_btree_freeing_key(r[i].b, keylist.top);
1420                 bch_keylist_push(&keylist);
1421         }
1422
1423         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1424         BUG_ON(!bch_keylist_empty(&keylist));
1425
1426         for (i = 0; i < nodes; i++) {
1427                 btree_node_free(r[i].b);
1428                 rw_unlock(true, r[i].b);
1429
1430                 r[i].b = new_nodes[i];
1431         }
1432
1433         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1434         r[nodes - 1].b = ERR_PTR(-EINTR);
1435
1436         trace_bcache_btree_gc_coalesce(nodes);
1437         gc->nodes--;
1438
1439         bch_keylist_free(&keylist);
1440
1441         /* Invalidated our iterator */
1442         return -EINTR;
1443
1444 out_nocoalesce:
1445         closure_sync(&cl);
1446         bch_keylist_free(&keylist);
1447
1448         while ((k = bch_keylist_pop(&keylist)))
1449                 if (!bkey_cmp(k, &ZERO_KEY))
1450                         atomic_dec(&b->c->prio_blocked);
1451
1452         for (i = 0; i < nodes; i++)
1453                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1454                         btree_node_free(new_nodes[i]);
1455                         rw_unlock(true, new_nodes[i]);
1456                 }
1457         return 0;
1458 }
1459
1460 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1461                                  struct btree *replace)
1462 {
1463         struct keylist keys;
1464         struct btree *n;
1465
1466         if (btree_check_reserve(b, NULL))
1467                 return 0;
1468
1469         n = btree_node_alloc_replacement(replace, NULL);
1470
1471         /* recheck reserve after allocating replacement node */
1472         if (btree_check_reserve(b, NULL)) {
1473                 btree_node_free(n);
1474                 rw_unlock(true, n);
1475                 return 0;
1476         }
1477
1478         bch_btree_node_write_sync(n);
1479
1480         bch_keylist_init(&keys);
1481         bch_keylist_add(&keys, &n->key);
1482
1483         make_btree_freeing_key(replace, keys.top);
1484         bch_keylist_push(&keys);
1485
1486         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1487         BUG_ON(!bch_keylist_empty(&keys));
1488
1489         btree_node_free(replace);
1490         rw_unlock(true, n);
1491
1492         /* Invalidated our iterator */
1493         return -EINTR;
1494 }
1495
1496 static unsigned btree_gc_count_keys(struct btree *b)
1497 {
1498         struct bkey *k;
1499         struct btree_iter iter;
1500         unsigned ret = 0;
1501
1502         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1503                 ret += bkey_u64s(k);
1504
1505         return ret;
1506 }
1507
1508 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1509                             struct closure *writes, struct gc_stat *gc)
1510 {
1511         int ret = 0;
1512         bool should_rewrite;
1513         struct bkey *k;
1514         struct btree_iter iter;
1515         struct gc_merge_info r[GC_MERGE_NODES];
1516         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1517
1518         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1519
1520         for (i = r; i < r + ARRAY_SIZE(r); i++)
1521                 i->b = ERR_PTR(-EINTR);
1522
1523         while (1) {
1524                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1525                 if (k) {
1526                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1527                                                   true, b);
1528                         if (IS_ERR(r->b)) {
1529                                 ret = PTR_ERR(r->b);
1530                                 break;
1531                         }
1532
1533                         r->keys = btree_gc_count_keys(r->b);
1534
1535                         ret = btree_gc_coalesce(b, op, gc, r);
1536                         if (ret)
1537                                 break;
1538                 }
1539
1540                 if (!last->b)
1541                         break;
1542
1543                 if (!IS_ERR(last->b)) {
1544                         should_rewrite = btree_gc_mark_node(last->b, gc);
1545                         if (should_rewrite) {
1546                                 ret = btree_gc_rewrite_node(b, op, last->b);
1547                                 if (ret)
1548                                         break;
1549                         }
1550
1551                         if (last->b->level) {
1552                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1553                                 if (ret)
1554                                         break;
1555                         }
1556
1557                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1558
1559                         /*
1560                          * Must flush leaf nodes before gc ends, since replace
1561                          * operations aren't journalled
1562                          */
1563                         mutex_lock(&last->b->write_lock);
1564                         if (btree_node_dirty(last->b))
1565                                 bch_btree_node_write(last->b, writes);
1566                         mutex_unlock(&last->b->write_lock);
1567                         rw_unlock(true, last->b);
1568                 }
1569
1570                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1571                 r->b = NULL;
1572
1573                 if (need_resched()) {
1574                         ret = -EAGAIN;
1575                         break;
1576                 }
1577         }
1578
1579         for (i = r; i < r + ARRAY_SIZE(r); i++)
1580                 if (!IS_ERR_OR_NULL(i->b)) {
1581                         mutex_lock(&i->b->write_lock);
1582                         if (btree_node_dirty(i->b))
1583                                 bch_btree_node_write(i->b, writes);
1584                         mutex_unlock(&i->b->write_lock);
1585                         rw_unlock(true, i->b);
1586                 }
1587
1588         return ret;
1589 }
1590
1591 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1592                              struct closure *writes, struct gc_stat *gc)
1593 {
1594         struct btree *n = NULL;
1595         int ret = 0;
1596         bool should_rewrite;
1597
1598         should_rewrite = btree_gc_mark_node(b, gc);
1599         if (should_rewrite) {
1600                 n = btree_node_alloc_replacement(b, NULL);
1601
1602                 if (!IS_ERR_OR_NULL(n)) {
1603                         bch_btree_node_write_sync(n);
1604
1605                         bch_btree_set_root(n);
1606                         btree_node_free(b);
1607                         rw_unlock(true, n);
1608
1609                         return -EINTR;
1610                 }
1611         }
1612
1613         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1614
1615         if (b->level) {
1616                 ret = btree_gc_recurse(b, op, writes, gc);
1617                 if (ret)
1618                         return ret;
1619         }
1620
1621         bkey_copy_key(&b->c->gc_done, &b->key);
1622
1623         return ret;
1624 }
1625
1626 static void btree_gc_start(struct cache_set *c)
1627 {
1628         struct cache *ca;
1629         struct bucket *b;
1630         unsigned i;
1631
1632         if (!c->gc_mark_valid)
1633                 return;
1634
1635         mutex_lock(&c->bucket_lock);
1636
1637         c->gc_mark_valid = 0;
1638         c->gc_done = ZERO_KEY;
1639
1640         for_each_cache(ca, c, i)
1641                 for_each_bucket(b, ca) {
1642                         b->last_gc = b->gen;
1643                         if (!atomic_read(&b->pin)) {
1644                                 SET_GC_MARK(b, 0);
1645                                 SET_GC_SECTORS_USED(b, 0);
1646                         }
1647                 }
1648
1649         mutex_unlock(&c->bucket_lock);
1650 }
1651
1652 static size_t bch_btree_gc_finish(struct cache_set *c)
1653 {
1654         size_t available = 0;
1655         struct bucket *b;
1656         struct cache *ca;
1657         unsigned i;
1658
1659         mutex_lock(&c->bucket_lock);
1660
1661         set_gc_sectors(c);
1662         c->gc_mark_valid = 1;
1663         c->need_gc      = 0;
1664
1665         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1666                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1667                             GC_MARK_METADATA);
1668
1669         /* don't reclaim buckets to which writeback keys point */
1670         rcu_read_lock();
1671         for (i = 0; i < c->nr_uuids; i++) {
1672                 struct bcache_device *d = c->devices[i];
1673                 struct cached_dev *dc;
1674                 struct keybuf_key *w, *n;
1675                 unsigned j;
1676
1677                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1678                         continue;
1679                 dc = container_of(d, struct cached_dev, disk);
1680
1681                 spin_lock(&dc->writeback_keys.lock);
1682                 rbtree_postorder_for_each_entry_safe(w, n,
1683                                         &dc->writeback_keys.keys, node)
1684                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1685                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1686                                             GC_MARK_DIRTY);
1687                 spin_unlock(&dc->writeback_keys.lock);
1688         }
1689         rcu_read_unlock();
1690
1691         for_each_cache(ca, c, i) {
1692                 uint64_t *i;
1693
1694                 ca->invalidate_needs_gc = 0;
1695
1696                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1697                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1698
1699                 for (i = ca->prio_buckets;
1700                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1701                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1702
1703                 for_each_bucket(b, ca) {
1704                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1705
1706                         if (atomic_read(&b->pin))
1707                                 continue;
1708
1709                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1710
1711                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1712                                 available++;
1713                 }
1714         }
1715
1716         mutex_unlock(&c->bucket_lock);
1717         return available;
1718 }
1719
1720 static void bch_btree_gc(struct cache_set *c)
1721 {
1722         int ret;
1723         unsigned long available;
1724         struct gc_stat stats;
1725         struct closure writes;
1726         struct btree_op op;
1727         uint64_t start_time = local_clock();
1728
1729         trace_bcache_gc_start(c);
1730
1731         memset(&stats, 0, sizeof(struct gc_stat));
1732         closure_init_stack(&writes);
1733         bch_btree_op_init(&op, SHRT_MAX);
1734
1735         btree_gc_start(c);
1736
1737         do {
1738                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1739                 closure_sync(&writes);
1740                 cond_resched();
1741
1742                 if (ret && ret != -EAGAIN)
1743                         pr_warn("gc failed!");
1744         } while (ret);
1745
1746         available = bch_btree_gc_finish(c);
1747         wake_up_allocators(c);
1748
1749         bch_time_stats_update(&c->btree_gc_time, start_time);
1750
1751         stats.key_bytes *= sizeof(uint64_t);
1752         stats.data      <<= 9;
1753         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1754         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1755
1756         trace_bcache_gc_end(c);
1757
1758         bch_moving_gc(c);
1759 }
1760
1761 static bool gc_should_run(struct cache_set *c)
1762 {
1763         struct cache *ca;
1764         unsigned i;
1765
1766         for_each_cache(ca, c, i)
1767                 if (ca->invalidate_needs_gc)
1768                         return true;
1769
1770         if (atomic_read(&c->sectors_to_gc) < 0)
1771                 return true;
1772
1773         return false;
1774 }
1775
1776 static int bch_gc_thread(void *arg)
1777 {
1778         struct cache_set *c = arg;
1779
1780         while (1) {
1781                 wait_event_interruptible(c->gc_wait,
1782                            kthread_should_stop() || gc_should_run(c));
1783
1784                 if (kthread_should_stop())
1785                         break;
1786
1787                 set_gc_sectors(c);
1788                 bch_btree_gc(c);
1789         }
1790
1791         return 0;
1792 }
1793
1794 int bch_gc_thread_start(struct cache_set *c)
1795 {
1796         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1797         if (IS_ERR(c->gc_thread))
1798                 return PTR_ERR(c->gc_thread);
1799
1800         return 0;
1801 }
1802
1803 /* Initial partial gc */
1804
1805 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1806 {
1807         int ret = 0;
1808         struct bkey *k, *p = NULL;
1809         struct btree_iter iter;
1810
1811         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1812                 bch_initial_mark_key(b->c, b->level, k);
1813
1814         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1815
1816         if (b->level) {
1817                 bch_btree_iter_init(&b->keys, &iter, NULL);
1818
1819                 do {
1820                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1821                                                        bch_ptr_bad);
1822                         if (k)
1823                                 btree_node_prefetch(b, k);
1824
1825                         if (p)
1826                                 ret = btree(check_recurse, p, b, op);
1827
1828                         p = k;
1829                 } while (p && !ret);
1830         }
1831
1832         return ret;
1833 }
1834
1835 int bch_btree_check(struct cache_set *c)
1836 {
1837         struct btree_op op;
1838
1839         bch_btree_op_init(&op, SHRT_MAX);
1840
1841         return btree_root(check_recurse, c, &op);
1842 }
1843
1844 void bch_initial_gc_finish(struct cache_set *c)
1845 {
1846         struct cache *ca;
1847         struct bucket *b;
1848         unsigned i;
1849
1850         bch_btree_gc_finish(c);
1851
1852         mutex_lock(&c->bucket_lock);
1853
1854         /*
1855          * We need to put some unused buckets directly on the prio freelist in
1856          * order to get the allocator thread started - it needs freed buckets in
1857          * order to rewrite the prios and gens, and it needs to rewrite prios
1858          * and gens in order to free buckets.
1859          *
1860          * This is only safe for buckets that have no live data in them, which
1861          * there should always be some of.
1862          */
1863         for_each_cache(ca, c, i) {
1864                 for_each_bucket(b, ca) {
1865                         if (fifo_full(&ca->free[RESERVE_PRIO]))
1866                                 break;
1867
1868                         if (bch_can_invalidate_bucket(ca, b) &&
1869                             !GC_MARK(b)) {
1870                                 __bch_invalidate_one_bucket(ca, b);
1871                                 fifo_push(&ca->free[RESERVE_PRIO],
1872                                           b - ca->buckets);
1873                         }
1874                 }
1875         }
1876
1877         mutex_unlock(&c->bucket_lock);
1878 }
1879
1880 /* Btree insertion */
1881
1882 static bool btree_insert_key(struct btree *b, struct bkey *k,
1883                              struct bkey *replace_key)
1884 {
1885         unsigned status;
1886
1887         BUG_ON(bkey_cmp(k, &b->key) > 0);
1888
1889         status = bch_btree_insert_key(&b->keys, k, replace_key);
1890         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1891                 bch_check_keys(&b->keys, "%u for %s", status,
1892                                replace_key ? "replace" : "insert");
1893
1894                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1895                                               status);
1896                 return true;
1897         } else
1898                 return false;
1899 }
1900
1901 static size_t insert_u64s_remaining(struct btree *b)
1902 {
1903         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1904
1905         /*
1906          * Might land in the middle of an existing extent and have to split it
1907          */
1908         if (b->keys.ops->is_extents)
1909                 ret -= KEY_MAX_U64S;
1910
1911         return max(ret, 0L);
1912 }
1913
1914 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1915                                   struct keylist *insert_keys,
1916                                   struct bkey *replace_key)
1917 {
1918         bool ret = false;
1919         int oldsize = bch_count_data(&b->keys);
1920
1921         while (!bch_keylist_empty(insert_keys)) {
1922                 struct bkey *k = insert_keys->keys;
1923
1924                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1925                         break;
1926
1927                 if (bkey_cmp(k, &b->key) <= 0) {
1928                         if (!b->level)
1929                                 bkey_put(b->c, k);
1930
1931                         ret |= btree_insert_key(b, k, replace_key);
1932                         bch_keylist_pop_front(insert_keys);
1933                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1934                         BKEY_PADDED(key) temp;
1935                         bkey_copy(&temp.key, insert_keys->keys);
1936
1937                         bch_cut_back(&b->key, &temp.key);
1938                         bch_cut_front(&b->key, insert_keys->keys);
1939
1940                         ret |= btree_insert_key(b, &temp.key, replace_key);
1941                         break;
1942                 } else {
1943                         break;
1944                 }
1945         }
1946
1947         if (!ret)
1948                 op->insert_collision = true;
1949
1950         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1951
1952         BUG_ON(bch_count_data(&b->keys) < oldsize);
1953         return ret;
1954 }
1955
1956 static int btree_split(struct btree *b, struct btree_op *op,
1957                        struct keylist *insert_keys,
1958                        struct bkey *replace_key)
1959 {
1960         bool split;
1961         struct btree *n1, *n2 = NULL, *n3 = NULL;
1962         uint64_t start_time = local_clock();
1963         struct closure cl;
1964         struct keylist parent_keys;
1965
1966         closure_init_stack(&cl);
1967         bch_keylist_init(&parent_keys);
1968
1969         if (btree_check_reserve(b, op)) {
1970                 if (!b->level)
1971                         return -EINTR;
1972                 else
1973                         WARN(1, "insufficient reserve for split\n");
1974         }
1975
1976         n1 = btree_node_alloc_replacement(b, op);
1977         if (IS_ERR(n1))
1978                 goto err;
1979
1980         split = set_blocks(btree_bset_first(n1),
1981                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1982
1983         if (split) {
1984                 unsigned keys = 0;
1985
1986                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1987
1988                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1989                 if (IS_ERR(n2))
1990                         goto err_free1;
1991
1992                 if (!b->parent) {
1993                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
1994                         if (IS_ERR(n3))
1995                                 goto err_free2;
1996                 }
1997
1998                 mutex_lock(&n1->write_lock);
1999                 mutex_lock(&n2->write_lock);
2000
2001                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2002
2003                 /*
2004                  * Has to be a linear search because we don't have an auxiliary
2005                  * search tree yet
2006                  */
2007
2008                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2009                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2010                                                         keys));
2011
2012                 bkey_copy_key(&n1->key,
2013                               bset_bkey_idx(btree_bset_first(n1), keys));
2014                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2015
2016                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2017                 btree_bset_first(n1)->keys = keys;
2018
2019                 memcpy(btree_bset_first(n2)->start,
2020                        bset_bkey_last(btree_bset_first(n1)),
2021                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2022
2023                 bkey_copy_key(&n2->key, &b->key);
2024
2025                 bch_keylist_add(&parent_keys, &n2->key);
2026                 bch_btree_node_write(n2, &cl);
2027                 mutex_unlock(&n2->write_lock);
2028                 rw_unlock(true, n2);
2029         } else {
2030                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2031
2032                 mutex_lock(&n1->write_lock);
2033                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2034         }
2035
2036         bch_keylist_add(&parent_keys, &n1->key);
2037         bch_btree_node_write(n1, &cl);
2038         mutex_unlock(&n1->write_lock);
2039
2040         if (n3) {
2041                 /* Depth increases, make a new root */
2042                 mutex_lock(&n3->write_lock);
2043                 bkey_copy_key(&n3->key, &MAX_KEY);
2044                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2045                 bch_btree_node_write(n3, &cl);
2046                 mutex_unlock(&n3->write_lock);
2047
2048                 closure_sync(&cl);
2049                 bch_btree_set_root(n3);
2050                 rw_unlock(true, n3);
2051         } else if (!b->parent) {
2052                 /* Root filled up but didn't need to be split */
2053                 closure_sync(&cl);
2054                 bch_btree_set_root(n1);
2055         } else {
2056                 /* Split a non root node */
2057                 closure_sync(&cl);
2058                 make_btree_freeing_key(b, parent_keys.top);
2059                 bch_keylist_push(&parent_keys);
2060
2061                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2062                 BUG_ON(!bch_keylist_empty(&parent_keys));
2063         }
2064
2065         btree_node_free(b);
2066         rw_unlock(true, n1);
2067
2068         bch_time_stats_update(&b->c->btree_split_time, start_time);
2069
2070         return 0;
2071 err_free2:
2072         bkey_put(b->c, &n2->key);
2073         btree_node_free(n2);
2074         rw_unlock(true, n2);
2075 err_free1:
2076         bkey_put(b->c, &n1->key);
2077         btree_node_free(n1);
2078         rw_unlock(true, n1);
2079 err:
2080         WARN(1, "bcache: btree split failed (level %u)", b->level);
2081
2082         if (n3 == ERR_PTR(-EAGAIN) ||
2083             n2 == ERR_PTR(-EAGAIN) ||
2084             n1 == ERR_PTR(-EAGAIN))
2085                 return -EAGAIN;
2086
2087         return -ENOMEM;
2088 }
2089
2090 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2091                                  struct keylist *insert_keys,
2092                                  atomic_t *journal_ref,
2093                                  struct bkey *replace_key)
2094 {
2095         struct closure cl;
2096
2097         BUG_ON(b->level && replace_key);
2098
2099         closure_init_stack(&cl);
2100
2101         mutex_lock(&b->write_lock);
2102
2103         if (write_block(b) != btree_bset_last(b) &&
2104             b->keys.last_set_unwritten)
2105                 bch_btree_init_next(b); /* just wrote a set */
2106
2107         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2108                 mutex_unlock(&b->write_lock);
2109                 goto split;
2110         }
2111
2112         BUG_ON(write_block(b) != btree_bset_last(b));
2113
2114         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2115                 if (!b->level)
2116                         bch_btree_leaf_dirty(b, journal_ref);
2117                 else
2118                         bch_btree_node_write(b, &cl);
2119         }
2120
2121         mutex_unlock(&b->write_lock);
2122
2123         /* wait for btree node write if necessary, after unlock */
2124         closure_sync(&cl);
2125
2126         return 0;
2127 split:
2128         if (current->bio_list) {
2129                 op->lock = b->c->root->level + 1;
2130                 return -EAGAIN;
2131         } else if (op->lock <= b->c->root->level) {
2132                 op->lock = b->c->root->level + 1;
2133                 return -EINTR;
2134         } else {
2135                 /* Invalidated all iterators */
2136                 int ret = btree_split(b, op, insert_keys, replace_key);
2137
2138                 if (bch_keylist_empty(insert_keys))
2139                         return 0;
2140                 else if (!ret)
2141                         return -EINTR;
2142                 return ret;
2143         }
2144 }
2145
2146 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2147                                struct bkey *check_key)
2148 {
2149         int ret = -EINTR;
2150         uint64_t btree_ptr = b->key.ptr[0];
2151         unsigned long seq = b->seq;
2152         struct keylist insert;
2153         bool upgrade = op->lock == -1;
2154
2155         bch_keylist_init(&insert);
2156
2157         if (upgrade) {
2158                 rw_unlock(false, b);
2159                 rw_lock(true, b, b->level);
2160
2161                 if (b->key.ptr[0] != btree_ptr ||
2162                    b->seq != seq + 1) {
2163                        op->lock = b->level;
2164                         goto out;
2165                }
2166         }
2167
2168         SET_KEY_PTRS(check_key, 1);
2169         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2170
2171         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2172
2173         bch_keylist_add(&insert, check_key);
2174
2175         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2176
2177         BUG_ON(!ret && !bch_keylist_empty(&insert));
2178 out:
2179         if (upgrade)
2180                 downgrade_write(&b->lock);
2181         return ret;
2182 }
2183
2184 struct btree_insert_op {
2185         struct btree_op op;
2186         struct keylist  *keys;
2187         atomic_t        *journal_ref;
2188         struct bkey     *replace_key;
2189 };
2190
2191 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2192 {
2193         struct btree_insert_op *op = container_of(b_op,
2194                                         struct btree_insert_op, op);
2195
2196         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2197                                         op->journal_ref, op->replace_key);
2198         if (ret && !bch_keylist_empty(op->keys))
2199                 return ret;
2200         else
2201                 return MAP_DONE;
2202 }
2203
2204 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2205                      atomic_t *journal_ref, struct bkey *replace_key)
2206 {
2207         struct btree_insert_op op;
2208         int ret = 0;
2209
2210         BUG_ON(current->bio_list);
2211         BUG_ON(bch_keylist_empty(keys));
2212
2213         bch_btree_op_init(&op.op, 0);
2214         op.keys         = keys;
2215         op.journal_ref  = journal_ref;
2216         op.replace_key  = replace_key;
2217
2218         while (!ret && !bch_keylist_empty(keys)) {
2219                 op.op.lock = 0;
2220                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2221                                                &START_KEY(keys->keys),
2222                                                btree_insert_fn);
2223         }
2224
2225         if (ret) {
2226                 struct bkey *k;
2227
2228                 pr_err("error %i", ret);
2229
2230                 while ((k = bch_keylist_pop(keys)))
2231                         bkey_put(c, k);
2232         } else if (op.op.insert_collision)
2233                 ret = -ESRCH;
2234
2235         return ret;
2236 }
2237
2238 void bch_btree_set_root(struct btree *b)
2239 {
2240         unsigned i;
2241         struct closure cl;
2242
2243         closure_init_stack(&cl);
2244
2245         trace_bcache_btree_set_root(b);
2246
2247         BUG_ON(!b->written);
2248
2249         for (i = 0; i < KEY_PTRS(&b->key); i++)
2250                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2251
2252         mutex_lock(&b->c->bucket_lock);
2253         list_del_init(&b->list);
2254         mutex_unlock(&b->c->bucket_lock);
2255
2256         b->c->root = b;
2257
2258         bch_journal_meta(b->c, &cl);
2259         closure_sync(&cl);
2260 }
2261
2262 /* Map across nodes or keys */
2263
2264 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2265                                        struct bkey *from,
2266                                        btree_map_nodes_fn *fn, int flags)
2267 {
2268         int ret = MAP_CONTINUE;
2269
2270         if (b->level) {
2271                 struct bkey *k;
2272                 struct btree_iter iter;
2273
2274                 bch_btree_iter_init(&b->keys, &iter, from);
2275
2276                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2277                                                        bch_ptr_bad))) {
2278                         ret = btree(map_nodes_recurse, k, b,
2279                                     op, from, fn, flags);
2280                         from = NULL;
2281
2282                         if (ret != MAP_CONTINUE)
2283                                 return ret;
2284                 }
2285         }
2286
2287         if (!b->level || flags == MAP_ALL_NODES)
2288                 ret = fn(op, b);
2289
2290         return ret;
2291 }
2292
2293 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2294                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2295 {
2296         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2297 }
2298
2299 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2300                                       struct bkey *from, btree_map_keys_fn *fn,
2301                                       int flags)
2302 {
2303         int ret = MAP_CONTINUE;
2304         struct bkey *k;
2305         struct btree_iter iter;
2306
2307         bch_btree_iter_init(&b->keys, &iter, from);
2308
2309         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2310                 ret = !b->level
2311                         ? fn(op, b, k)
2312                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2313                 from = NULL;
2314
2315                 if (ret != MAP_CONTINUE)
2316                         return ret;
2317         }
2318
2319         if (!b->level && (flags & MAP_END_KEY))
2320                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2321                                      KEY_OFFSET(&b->key), 0));
2322
2323         return ret;
2324 }
2325
2326 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2327                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2328 {
2329         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2330 }
2331
2332 /* Keybuf code */
2333
2334 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2335 {
2336         /* Overlapping keys compare equal */
2337         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2338                 return -1;
2339         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2340                 return 1;
2341         return 0;
2342 }
2343
2344 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2345                                             struct keybuf_key *r)
2346 {
2347         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2348 }
2349
2350 struct refill {
2351         struct btree_op op;
2352         unsigned        nr_found;
2353         struct keybuf   *buf;
2354         struct bkey     *end;
2355         keybuf_pred_fn  *pred;
2356 };
2357
2358 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2359                             struct bkey *k)
2360 {
2361         struct refill *refill = container_of(op, struct refill, op);
2362         struct keybuf *buf = refill->buf;
2363         int ret = MAP_CONTINUE;
2364
2365         if (bkey_cmp(k, refill->end) >= 0) {
2366                 ret = MAP_DONE;
2367                 goto out;
2368         }
2369
2370         if (!KEY_SIZE(k)) /* end key */
2371                 goto out;
2372
2373         if (refill->pred(buf, k)) {
2374                 struct keybuf_key *w;
2375
2376                 spin_lock(&buf->lock);
2377
2378                 w = array_alloc(&buf->freelist);
2379                 if (!w) {
2380                         spin_unlock(&buf->lock);
2381                         return MAP_DONE;
2382                 }
2383
2384                 w->private = NULL;
2385                 bkey_copy(&w->key, k);
2386
2387                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2388                         array_free(&buf->freelist, w);
2389                 else
2390                         refill->nr_found++;
2391
2392                 if (array_freelist_empty(&buf->freelist))
2393                         ret = MAP_DONE;
2394
2395                 spin_unlock(&buf->lock);
2396         }
2397 out:
2398         buf->last_scanned = *k;
2399         return ret;
2400 }
2401
2402 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2403                        struct bkey *end, keybuf_pred_fn *pred)
2404 {
2405         struct bkey start = buf->last_scanned;
2406         struct refill refill;
2407
2408         cond_resched();
2409
2410         bch_btree_op_init(&refill.op, -1);
2411         refill.nr_found = 0;
2412         refill.buf      = buf;
2413         refill.end      = end;
2414         refill.pred     = pred;
2415
2416         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2417                            refill_keybuf_fn, MAP_END_KEY);
2418
2419         trace_bcache_keyscan(refill.nr_found,
2420                              KEY_INODE(&start), KEY_OFFSET(&start),
2421                              KEY_INODE(&buf->last_scanned),
2422                              KEY_OFFSET(&buf->last_scanned));
2423
2424         spin_lock(&buf->lock);
2425
2426         if (!RB_EMPTY_ROOT(&buf->keys)) {
2427                 struct keybuf_key *w;
2428                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2429                 buf->start      = START_KEY(&w->key);
2430
2431                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2432                 buf->end        = w->key;
2433         } else {
2434                 buf->start      = MAX_KEY;
2435                 buf->end        = MAX_KEY;
2436         }
2437
2438         spin_unlock(&buf->lock);
2439 }
2440
2441 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2442 {
2443         rb_erase(&w->node, &buf->keys);
2444         array_free(&buf->freelist, w);
2445 }
2446
2447 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2448 {
2449         spin_lock(&buf->lock);
2450         __bch_keybuf_del(buf, w);
2451         spin_unlock(&buf->lock);
2452 }
2453
2454 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2455                                   struct bkey *end)
2456 {
2457         bool ret = false;
2458         struct keybuf_key *p, *w, s;
2459         s.key = *start;
2460
2461         if (bkey_cmp(end, &buf->start) <= 0 ||
2462             bkey_cmp(start, &buf->end) >= 0)
2463                 return false;
2464
2465         spin_lock(&buf->lock);
2466         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2467
2468         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2469                 p = w;
2470                 w = RB_NEXT(w, node);
2471
2472                 if (p->private)
2473                         ret = true;
2474                 else
2475                         __bch_keybuf_del(buf, p);
2476         }
2477
2478         spin_unlock(&buf->lock);
2479         return ret;
2480 }
2481
2482 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2483 {
2484         struct keybuf_key *w;
2485         spin_lock(&buf->lock);
2486
2487         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2488
2489         while (w && w->private)
2490                 w = RB_NEXT(w, node);
2491
2492         if (w)
2493                 w->private = ERR_PTR(-EINTR);
2494
2495         spin_unlock(&buf->lock);
2496         return w;
2497 }
2498
2499 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2500                                           struct keybuf *buf,
2501                                           struct bkey *end,
2502                                           keybuf_pred_fn *pred)
2503 {
2504         struct keybuf_key *ret;
2505
2506         while (1) {
2507                 ret = bch_keybuf_next(buf);
2508                 if (ret)
2509                         break;
2510
2511                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2512                         pr_debug("scan finished");
2513                         break;
2514                 }
2515
2516                 bch_refill_keybuf(c, buf, end, pred);
2517         }
2518
2519         return ret;
2520 }
2521
2522 void bch_keybuf_init(struct keybuf *buf)
2523 {
2524         buf->last_scanned       = MAX_KEY;
2525         buf->keys               = RB_ROOT;
2526
2527         spin_lock_init(&buf->lock);
2528         array_allocator_init(&buf->freelist);
2529 }