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 <linux/rculist.h>
  37
  38 #include <trace/events/bcache.h>
  39
  40 /*
  41  * Todo:
  42  * register_bcache: Return errors out to userspace correctly
  43  *
  44  * Writeback: don't undirty key until after a cache flush
  45  *
  46  * Create an iterator for key pointers
  47  *
  48  * On btree write error, mark bucket such that it won't be freed from the cache
  49  *
  50  * Journalling:
  51  *   Check for bad keys in replay
  52  *   Propagate barriers
  53  *   Refcount journal entries in journal_replay
  54  *
  55  * Garbage collection:
  56  *   Finish incremental gc
  57  *   Gc should free old UUIDs, data for invalid UUIDs
  58  *
  59  * Provide a way to list backing device UUIDs we have data cached for, and
  60  * probably how long it's been since we've seen them, and a way to invalidate
  61  * dirty data for devices that will never be attached again
  62  *
  63  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  64  * that based on that and how much dirty data we have we can keep writeback
  65  * from being starved
  66  *
  67  * Add a tracepoint or somesuch to watch for writeback starvation
  68  *
  69  * When btree depth > 1 and splitting an interior node, we have to make sure
  70  * alloc_bucket() cannot fail. This should be true but is not completely
  71  * obvious.
  72  *
  73  * Plugging?
  74  *
  75  * If data write is less than hard sector size of ssd, round up offset in open
  76  * bucket to the next whole sector
  77  *
  78  * Superblock needs to be fleshed out for multiple cache devices
  79  *
  80  * Add a sysfs tunable for the number of writeback IOs in flight
  81  *
  82  * Add a sysfs tunable for the number of open data buckets
  83  *
  84  * IO tracking: Can we track when one process is doing io on behalf of another?
  85  * IO tracking: Don't use just an average, weigh more recent stuff higher
  86  *
  87  * Test module load/unload
  88  */
  89
  90 #define MAX_NEED_GC             64
  91 #define MAX_SAVE_PRIO           72
  92
  93 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  94
  95 #define PTR_HASH(c, k)                                                  \
  96         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  97
  98 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
  99
 100 /*
 101  * These macros are for recursing down the btree - they handle the details of
 102  * locking and looking up nodes in the cache for you. They're best treated as
 103  * mere syntax when reading code that uses them.
 104  *
 105  * op->lock determines whether we take a read or a write lock at a given depth.
 106  * If you've got a read lock and find that you need a write lock (i.e. you're
 107  * going to have to split), set op->lock and return -EINTR; btree_root() will
 108  * call you again and you'll have the correct lock.
 109  */
 110
 111 /**
 112  * btree - recurse down the btree on a specified key
 113  * @fn:         function to call, which will be passed the child node
 114  * @key:        key to recurse on
 115  * @b:          parent btree node
 116  * @op:         pointer to struct btree_op
 117  */
 118 #define btree(fn, key, b, op, ...)                                      \
 119 ({                                                                      \
 120         int _r, l = (b)->level - 1;                                     \
 121         bool _w = l <= (op)->lock;                                      \
 122         struct btree *_child = bch_btree_node_get((b)->c, op, key, l,   \
 123                                                   _w, b);               \
 124         if (!IS_ERR(_child)) {                                          \
 125                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 126                 rw_unlock(_w, _child);                                  \
 127         } else                                                          \
 128                 _r = PTR_ERR(_child);                                   \
 129         _r;                                                             \
 130 })
 131
 132 /**
 133  * btree_root - call a function on the root of the btree
 134  * @fn:         function to call, which will be passed the child node
 135  * @c:          cache set
 136  * @op:         pointer to struct btree_op
 137  */
 138 #define btree_root(fn, c, op, ...)                                      \
 139 ({                                                                      \
 140         int _r = -EINTR;                                                \
 141         do {                                                            \
 142                 struct btree *_b = (c)->root;                           \
 143                 bool _w = insert_lock(op, _b);                          \
 144                 rw_lock(_w, _b, _b->level);                             \
 145                 if (_b == (c)->root &&                                  \
 146                     _w == insert_lock(op, _b)) {                        \
 147                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 148                 }                                                       \
 149                 rw_unlock(_w, _b);                                      \
 150                 bch_cannibalize_unlock(c);                              \
 151                 if (_r == -EINTR)                                       \
 152                         schedule();                                     \
 153         } while (_r == -EINTR);                                         \
 154                                                                         \
 155         finish_wait(&(c)->btree_cache_wait, &(op)->wait);               \
 156         _r;                                                             \
 157 })
 158
 159 static inline struct bset *write_block(struct btree *b)
 160 {
 161         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 162 }
 163
 164 static void bch_btree_init_next(struct btree *b)
 165 {
 166         /* If not a leaf node, always sort */
 167         if (b->level && b->keys.nsets)
 168                 bch_btree_sort(&b->keys, &b->c->sort);
 169         else
 170                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 171
 172         if (b->written < btree_blocks(b))
 173                 bch_bset_init_next(&b->keys, write_block(b),
 174                                    bset_magic(&b->c->sb));
 175
 176 }
 177
 178 /* Btree key manipulation */
 179
 180 void bkey_put(struct cache_set *c, struct bkey *k)
 181 {
 182         unsigned i;
 183
 184         for (i = 0; i < KEY_PTRS(k); i++)
 185                 if (ptr_available(c, k, i))
 186                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 187 }
 188
 189 /* Btree IO */
 190
 191 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 192 {
 193         uint64_t crc = b->key.ptr[0];
 194         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 195
 196         crc = bch_crc64_update(crc, data, end - data);
 197         return crc ^ 0xffffffffffffffffULL;
 198 }
 199
 200 void bch_btree_node_read_done(struct btree *b)
 201 {
 202         const char *err = "bad btree header";
 203         struct bset *i = btree_bset_first(b);
 204         struct btree_iter *iter;
 205
 206         iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
 207         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 208         iter->used = 0;
 209
 210 #ifdef CONFIG_BCACHE_DEBUG
 211         iter->b = &b->keys;
 212 #endif
 213
 214         if (!i->seq)
 215                 goto err;
 216
 217         for (;
 218              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 219              i = write_block(b)) {
 220                 err = "unsupported bset version";
 221                 if (i->version > BCACHE_BSET_VERSION)
 222                         goto err;
 223
 224                 err = "bad btree header";
 225                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 226                     btree_blocks(b))
 227                         goto err;
 228
 229                 err = "bad magic";
 230                 if (i->magic != bset_magic(&b->c->sb))
 231                         goto err;
 232
 233                 err = "bad checksum";
 234                 switch (i->version) {
 235                 case 0:
 236                         if (i->csum != csum_set(i))
 237                                 goto err;
 238                         break;
 239                 case BCACHE_BSET_VERSION:
 240                         if (i->csum != btree_csum_set(b, i))
 241                                 goto err;
 242                         break;
 243                 }
 244
 245                 err = "empty set";
 246                 if (i != b->keys.set[0].data && !i->keys)
 247                         goto err;
 248
 249                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 250
 251                 b->written += set_blocks(i, block_bytes(b->c));
 252         }
 253
 254         err = "corrupted btree";
 255         for (i = write_block(b);
 256              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 257              i = ((void *) i) + block_bytes(b->c))
 258                 if (i->seq == b->keys.set[0].data->seq)
 259                         goto err;
 260
 261         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 262
 263         i = b->keys.set[0].data;
 264         err = "short btree key";
 265         if (b->keys.set[0].size &&
 266             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 267                 goto err;
 268
 269         if (b->written < btree_blocks(b))
 270                 bch_bset_init_next(&b->keys, write_block(b),
 271                                    bset_magic(&b->c->sb));
 272 out:
 273         mempool_free(iter, b->c->fill_iter);
 274         return;
 275 err:
 276         set_btree_node_io_error(b);
 277         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 278                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 279                             bset_block_offset(b, i), i->keys);
 280         goto out;
 281 }
 282
 283 static void btree_node_read_endio(struct bio *bio)
 284 {
 285         struct closure *cl = bio->bi_private;
 286         closure_put(cl);
 287 }
 288
 289 static void bch_btree_node_read(struct btree *b)
 290 {
 291         uint64_t start_time = local_clock();
 292         struct closure cl;
 293         struct bio *bio;
 294
 295         trace_bcache_btree_read(b);
 296
 297         closure_init_stack(&cl);
 298
 299         bio = bch_bbio_alloc(b->c);
 300         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 301         bio->bi_end_io  = btree_node_read_endio;
 302         bio->bi_private = &cl;
 303         bio->bi_opf = REQ_OP_READ | REQ_META;
 304
 305         bch_bio_map(bio, b->keys.set[0].data);
 306
 307         bch_submit_bbio(bio, b->c, &b->key, 0);
 308         closure_sync(&cl);
 309
 310         if (bio->bi_status)
 311                 set_btree_node_io_error(b);
 312
 313         bch_bbio_free(bio, b->c);
 314
 315         if (btree_node_io_error(b))
 316                 goto err;
 317
 318         bch_btree_node_read_done(b);
 319         bch_time_stats_update(&b->c->btree_read_time, start_time);
 320
 321         return;
 322 err:
 323         bch_cache_set_error(b->c, "io error reading bucket %zu",
 324                             PTR_BUCKET_NR(b->c, &b->key, 0));
 325 }
 326
 327 static void btree_complete_write(struct btree *b, struct btree_write *w)
 328 {
 329         if (w->prio_blocked &&
 330             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 331                 wake_up_allocators(b->c);
 332
 333         if (w->journal) {
 334                 atomic_dec_bug(w->journal);
 335                 __closure_wake_up(&b->c->journal.wait);
 336         }
 337
 338         w->prio_blocked = 0;
 339         w->journal      = NULL;
 340 }
 341
 342 static void btree_node_write_unlock(struct closure *cl)
 343 {
 344         struct btree *b = container_of(cl, struct btree, io);
 345
 346         up(&b->io_mutex);
 347 }
 348
 349 static void __btree_node_write_done(struct closure *cl)
 350 {
 351         struct btree *b = container_of(cl, struct btree, io);
 352         struct btree_write *w = btree_prev_write(b);
 353
 354         bch_bbio_free(b->bio, b->c);
 355         b->bio = NULL;
 356         btree_complete_write(b, w);
 357
 358         if (btree_node_dirty(b))
 359                 schedule_delayed_work(&b->work, 30 * HZ);
 360
 361         closure_return_with_destructor(cl, btree_node_write_unlock);
 362 }
 363
 364 static void btree_node_write_done(struct closure *cl)
 365 {
 366         struct btree *b = container_of(cl, struct btree, io);
 367
 368         bio_free_pages(b->bio);
 369         __btree_node_write_done(cl);
 370 }
 371
 372 static void btree_node_write_endio(struct bio *bio)
 373 {
 374         struct closure *cl = bio->bi_private;
 375         struct btree *b = container_of(cl, struct btree, io);
 376
 377         if (bio->bi_status)
 378                 set_btree_node_io_error(b);
 379
 380         bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 381         closure_put(cl);
 382 }
 383
 384 static void do_btree_node_write(struct btree *b)
 385 {
 386         struct closure *cl = &b->io;
 387         struct bset *i = btree_bset_last(b);
 388         BKEY_PADDED(key) k;
 389
 390         i->version      = BCACHE_BSET_VERSION;
 391         i->csum         = btree_csum_set(b, i);
 392
 393         BUG_ON(b->bio);
 394         b->bio = bch_bbio_alloc(b->c);
 395
 396         b->bio->bi_end_io       = btree_node_write_endio;
 397         b->bio->bi_private      = cl;
 398         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 399         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 400         bch_bio_map(b->bio, i);
 401
 402         /*
 403          * If we're appending to a leaf node, we don't technically need FUA -
 404          * this write just needs to be persisted before the next journal write,
 405          * which will be marked FLUSH|FUA.
 406          *
 407          * Similarly if we're writing a new btree root - the pointer is going to
 408          * be in the next journal entry.
 409          *
 410          * But if we're writing a new btree node (that isn't a root) or
 411          * appending to a non leaf btree node, we need either FUA or a flush
 412          * when we write the parent with the new pointer. FUA is cheaper than a
 413          * flush, and writes appending to leaf nodes aren't blocking anything so
 414          * just make all btree node writes FUA to keep things sane.
 415          */
 416
 417         bkey_copy(&k.key, &b->key);
 418         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 419                        bset_sector_offset(&b->keys, i));
 420
 421         if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 422                 int j;
 423                 struct bio_vec *bv;
 424                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 425
 426                 bio_for_each_segment_all(bv, b->bio, j)
 427                         memcpy(page_address(bv->bv_page),
 428                                base + j * PAGE_SIZE, PAGE_SIZE);
 429
 430                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 431
 432                 continue_at(cl, btree_node_write_done, NULL);
 433         } else {
 434                 b->bio->bi_vcnt = 0;
 435                 bch_bio_map(b->bio, i);
 436
 437                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 438
 439                 closure_sync(cl);
 440                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 441         }
 442 }
 443
 444 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 445 {
 446         struct bset *i = btree_bset_last(b);
 447
 448         lockdep_assert_held(&b->write_lock);
 449
 450         trace_bcache_btree_write(b);
 451
 452         BUG_ON(current->bio_list);
 453         BUG_ON(b->written >= btree_blocks(b));
 454         BUG_ON(b->written && !i->keys);
 455         BUG_ON(btree_bset_first(b)->seq != i->seq);
 456         bch_check_keys(&b->keys, "writing");
 457
 458         cancel_delayed_work(&b->work);
 459
 460         /* If caller isn't waiting for write, parent refcount is cache set */
 461         down(&b->io_mutex);
 462         closure_init(&b->io, parent ?: &b->c->cl);
 463
 464         clear_bit(BTREE_NODE_dirty,      &b->flags);
 465         change_bit(BTREE_NODE_write_idx, &b->flags);
 466
 467         do_btree_node_write(b);
 468
 469         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 470                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 471
 472         b->written += set_blocks(i, block_bytes(b->c));
 473 }
 474
 475 void bch_btree_node_write(struct btree *b, struct closure *parent)
 476 {
 477         unsigned nsets = b->keys.nsets;
 478
 479         lockdep_assert_held(&b->lock);
 480
 481         __bch_btree_node_write(b, parent);
 482
 483         /*
 484          * do verify if there was more than one set initially (i.e. we did a
 485          * sort) and we sorted down to a single set:
 486          */
 487         if (nsets && !b->keys.nsets)
 488                 bch_btree_verify(b);
 489
 490         bch_btree_init_next(b);
 491 }
 492
 493 static void bch_btree_node_write_sync(struct btree *b)
 494 {
 495         struct closure cl;
 496
 497         closure_init_stack(&cl);
 498
 499         mutex_lock(&b->write_lock);
 500         bch_btree_node_write(b, &cl);
 501         mutex_unlock(&b->write_lock);
 502
 503         closure_sync(&cl);
 504 }
 505
 506 static void btree_node_write_work(struct work_struct *w)
 507 {
 508         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 509
 510         mutex_lock(&b->write_lock);
 511         if (btree_node_dirty(b))
 512                 __bch_btree_node_write(b, NULL);
 513         mutex_unlock(&b->write_lock);
 514 }
 515
 516 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 517 {
 518         struct bset *i = btree_bset_last(b);
 519         struct btree_write *w = btree_current_write(b);
 520
 521         lockdep_assert_held(&b->write_lock);
 522
 523         BUG_ON(!b->written);
 524         BUG_ON(!i->keys);
 525
 526         if (!btree_node_dirty(b))
 527                 schedule_delayed_work(&b->work, 30 * HZ);
 528
 529         set_btree_node_dirty(b);
 530
 531         if (journal_ref) {
 532                 if (w->journal &&
 533                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 534                         atomic_dec_bug(w->journal);
 535                         w->journal = NULL;
 536                 }
 537
 538                 if (!w->journal) {
 539                         w->journal = journal_ref;
 540                         atomic_inc(w->journal);
 541                 }
 542         }
 543
 544         /* Force write if set is too big */
 545         if (set_bytes(i) > PAGE_SIZE - 48 &&
 546             !current->bio_list)
 547                 bch_btree_node_write(b, NULL);
 548 }
 549
 550 /*
 551  * Btree in memory cache - allocation/freeing
 552  * mca -> memory cache
 553  */
 554
 555 #define mca_reserve(c)  (((c->root && c->root->level)           \
 556                           ? c->root->level : 1) * 8 + 16)
 557 #define mca_can_free(c)                                         \
 558         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 559
 560 static void mca_data_free(struct btree *b)
 561 {
 562         BUG_ON(b->io_mutex.count != 1);
 563
 564         bch_btree_keys_free(&b->keys);
 565
 566         b->c->btree_cache_used--;
 567         list_move(&b->list, &b->c->btree_cache_freed);
 568 }
 569
 570 static void mca_bucket_free(struct btree *b)
 571 {
 572         BUG_ON(btree_node_dirty(b));
 573
 574         b->key.ptr[0] = 0;
 575         hlist_del_init_rcu(&b->hash);
 576         list_move(&b->list, &b->c->btree_cache_freeable);
 577 }
 578
 579 static unsigned btree_order(struct bkey *k)
 580 {
 581         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 582 }
 583
 584 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 585 {
 586         if (!bch_btree_keys_alloc(&b->keys,
 587                                   max_t(unsigned,
 588                                         ilog2(b->c->btree_pages),
 589                                         btree_order(k)),
 590                                   gfp)) {
 591                 b->c->btree_cache_used++;
 592                 list_move(&b->list, &b->c->btree_cache);
 593         } else {
 594                 list_move(&b->list, &b->c->btree_cache_freed);
 595         }
 596 }
 597
 598 static struct btree *mca_bucket_alloc(struct cache_set *c,
 599                                       struct bkey *k, gfp_t gfp)
 600 {
 601         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 602         if (!b)
 603                 return NULL;
 604
 605         init_rwsem(&b->lock);
 606         lockdep_set_novalidate_class(&b->lock);
 607         mutex_init(&b->write_lock);
 608         lockdep_set_novalidate_class(&b->write_lock);
 609         INIT_LIST_HEAD(&b->list);
 610         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 611         b->c = c;
 612         sema_init(&b->io_mutex, 1);
 613
 614         mca_data_alloc(b, k, gfp);
 615         return b;
 616 }
 617
 618 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 619 {
 620         struct closure cl;
 621
 622         closure_init_stack(&cl);
 623         lockdep_assert_held(&b->c->bucket_lock);
 624
 625         if (!down_write_trylock(&b->lock))
 626                 return -ENOMEM;
 627
 628         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 629
 630         if (b->keys.page_order < min_order)
 631                 goto out_unlock;
 632
 633         if (!flush) {
 634                 if (btree_node_dirty(b))
 635                         goto out_unlock;
 636
 637                 if (down_trylock(&b->io_mutex))
 638                         goto out_unlock;
 639                 up(&b->io_mutex);
 640         }
 641
 642         mutex_lock(&b->write_lock);
 643         if (btree_node_dirty(b))
 644                 __bch_btree_node_write(b, &cl);
 645         mutex_unlock(&b->write_lock);
 646
 647         closure_sync(&cl);
 648
 649         /* wait for any in flight btree write */
 650         down(&b->io_mutex);
 651         up(&b->io_mutex);
 652
 653         return 0;
 654 out_unlock:
 655         rw_unlock(true, b);
 656         return -ENOMEM;
 657 }
 658
 659 static unsigned long bch_mca_scan(struct shrinker *shrink,
 660                                   struct shrink_control *sc)
 661 {
 662         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 663         struct btree *b, *t;
 664         unsigned long i, nr = sc->nr_to_scan;
 665         unsigned long freed = 0;
 666
 667         if (c->shrinker_disabled)
 668                 return SHRINK_STOP;
 669
 670         if (c->btree_cache_alloc_lock)
 671                 return SHRINK_STOP;
 672
 673         /* Return -1 if we can't do anything right now */
 674         if (sc->gfp_mask & __GFP_IO)
 675                 mutex_lock(&c->bucket_lock);
 676         else if (!mutex_trylock(&c->bucket_lock))
 677                 return -1;
 678
 679         /*
 680          * It's _really_ critical that we don't free too many btree nodes - we
 681          * have to always leave ourselves a reserve. The reserve is how we
 682          * guarantee that allocating memory for a new btree node can always
 683          * succeed, so that inserting keys into the btree can always succeed and
 684          * IO can always make forward progress:
 685          */
 686         nr /= c->btree_pages;
 687         nr = min_t(unsigned long, nr, mca_can_free(c));
 688
 689         i = 0;
 690         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 691                 if (freed >= nr)
 692                         break;
 693
 694                 if (++i > 3 &&
 695                     !mca_reap(b, 0, false)) {
 696                         mca_data_free(b);
 697                         rw_unlock(true, b);
 698                         freed++;
 699                 }
 700         }
 701
 702         for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
 703                 if (list_empty(&c->btree_cache))
 704                         goto out;
 705
 706                 b = list_first_entry(&c->btree_cache, struct btree, list);
 707                 list_rotate_left(&c->btree_cache);
 708
 709                 if (!b->accessed &&
 710                     !mca_reap(b, 0, false)) {
 711                         mca_bucket_free(b);
 712                         mca_data_free(b);
 713                         rw_unlock(true, b);
 714                         freed++;
 715                 } else
 716                         b->accessed = 0;
 717         }
 718 out:
 719         mutex_unlock(&c->bucket_lock);
 720         return freed;
 721 }
 722
 723 static unsigned long bch_mca_count(struct shrinker *shrink,
 724                                    struct shrink_control *sc)
 725 {
 726         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 727
 728         if (c->shrinker_disabled)
 729                 return 0;
 730
 731         if (c->btree_cache_alloc_lock)
 732                 return 0;
 733
 734         return mca_can_free(c) * c->btree_pages;
 735 }
 736
 737 void bch_btree_cache_free(struct cache_set *c)
 738 {
 739         struct btree *b;
 740         struct closure cl;
 741         closure_init_stack(&cl);
 742
 743         if (c->shrink.list.next)
 744                 unregister_shrinker(&c->shrink);
 745
 746         mutex_lock(&c->bucket_lock);
 747
 748 #ifdef CONFIG_BCACHE_DEBUG
 749         if (c->verify_data)
 750                 list_move(&c->verify_data->list, &c->btree_cache);
 751
 752         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 753 #endif
 754
 755         list_splice(&c->btree_cache_freeable,
 756                     &c->btree_cache);
 757
 758         while (!list_empty(&c->btree_cache)) {
 759                 b = list_first_entry(&c->btree_cache, struct btree, list);
 760
 761                 if (btree_node_dirty(b))
 762                         btree_complete_write(b, btree_current_write(b));
 763                 clear_bit(BTREE_NODE_dirty, &b->flags);
 764
 765                 mca_data_free(b);
 766         }
 767
 768         while (!list_empty(&c->btree_cache_freed)) {
 769                 b = list_first_entry(&c->btree_cache_freed,
 770                                      struct btree, list);
 771                 list_del(&b->list);
 772                 cancel_delayed_work_sync(&b->work);
 773                 kfree(b);
 774         }
 775
 776         mutex_unlock(&c->bucket_lock);
 777 }
 778
 779 int bch_btree_cache_alloc(struct cache_set *c)
 780 {
 781         unsigned i;
 782
 783         for (i = 0; i < mca_reserve(c); i++)
 784                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 785                         return -ENOMEM;
 786
 787         list_splice_init(&c->btree_cache,
 788                          &c->btree_cache_freeable);
 789
 790 #ifdef CONFIG_BCACHE_DEBUG
 791         mutex_init(&c->verify_lock);
 792
 793         c->verify_ondisk = (void *)
 794                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 795
 796         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 797
 798         if (c->verify_data &&
 799             c->verify_data->keys.set->data)
 800                 list_del_init(&c->verify_data->list);
 801         else
 802                 c->verify_data = NULL;
 803 #endif
 804
 805         c->shrink.count_objects = bch_mca_count;
 806         c->shrink.scan_objects = bch_mca_scan;
 807         c->shrink.seeks = 4;
 808         c->shrink.batch = c->btree_pages * 2;
 809         register_shrinker(&c->shrink);
 810
 811         return 0;
 812 }
 813
 814 /* Btree in memory cache - hash table */
 815
 816 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 817 {
 818         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 819 }
 820
 821 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 822 {
 823         struct btree *b;
 824
 825         rcu_read_lock();
 826         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 827                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 828                         goto out;
 829         b = NULL;
 830 out:
 831         rcu_read_unlock();
 832         return b;
 833 }
 834
 835 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 836 {
 837         struct task_struct *old;
 838
 839         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 840         if (old && old != current) {
 841                 if (op)
 842                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 843                                         TASK_UNINTERRUPTIBLE);
 844                 return -EINTR;
 845         }
 846
 847         return 0;
 848 }
 849
 850 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 851                                      struct bkey *k)
 852 {
 853         struct btree *b;
 854
 855         trace_bcache_btree_cache_cannibalize(c);
 856
 857         if (mca_cannibalize_lock(c, op))
 858                 return ERR_PTR(-EINTR);
 859
 860         list_for_each_entry_reverse(b, &c->btree_cache, list)
 861                 if (!mca_reap(b, btree_order(k), false))
 862                         return b;
 863
 864         list_for_each_entry_reverse(b, &c->btree_cache, list)
 865                 if (!mca_reap(b, btree_order(k), true))
 866                         return b;
 867
 868         WARN(1, "btree cache cannibalize failed\n");
 869         return ERR_PTR(-ENOMEM);
 870 }
 871
 872 /*
 873  * We can only have one thread cannibalizing other cached btree nodes at a time,
 874  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 875  * cannibalize_bucket() will take. This means every time we unlock the root of
 876  * the btree, we need to release this lock if we have it held.
 877  */
 878 static void bch_cannibalize_unlock(struct cache_set *c)
 879 {
 880         if (c->btree_cache_alloc_lock == current) {
 881                 c->btree_cache_alloc_lock = NULL;
 882                 wake_up(&c->btree_cache_wait);
 883         }
 884 }
 885
 886 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 887                                struct bkey *k, int level)
 888 {
 889         struct btree *b;
 890
 891         BUG_ON(current->bio_list);
 892
 893         lockdep_assert_held(&c->bucket_lock);
 894
 895         if (mca_find(c, k))
 896                 return NULL;
 897
 898         /* btree_free() doesn't free memory; it sticks the node on the end of
 899          * the list. Check if there's any freed nodes there:
 900          */
 901         list_for_each_entry(b, &c->btree_cache_freeable, list)
 902                 if (!mca_reap(b, btree_order(k), false))
 903                         goto out;
 904
 905         /* We never free struct btree itself, just the memory that holds the on
 906          * disk node. Check the freed list before allocating a new one:
 907          */
 908         list_for_each_entry(b, &c->btree_cache_freed, list)
 909                 if (!mca_reap(b, 0, false)) {
 910                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 911                         if (!b->keys.set[0].data)
 912                                 goto err;
 913                         else
 914                                 goto out;
 915                 }
 916
 917         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 918         if (!b)
 919                 goto err;
 920
 921         BUG_ON(!down_write_trylock(&b->lock));
 922         if (!b->keys.set->data)
 923                 goto err;
 924 out:
 925         BUG_ON(b->io_mutex.count != 1);
 926
 927         bkey_copy(&b->key, k);
 928         list_move(&b->list, &c->btree_cache);
 929         hlist_del_init_rcu(&b->hash);
 930         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 931
 932         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 933         b->parent       = (void *) ~0UL;
 934         b->flags        = 0;
 935         b->written      = 0;
 936         b->level        = level;
 937
 938         if (!b->level)
 939                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 940                                     &b->c->expensive_debug_checks);
 941         else
 942                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 943                                     &b->c->expensive_debug_checks);
 944
 945         return b;
 946 err:
 947         if (b)
 948                 rw_unlock(true, b);
 949
 950         b = mca_cannibalize(c, op, k);
 951         if (!IS_ERR(b))
 952                 goto out;
 953
 954         return b;
 955 }
 956
 957 /**
 958  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 959  * in from disk if necessary.
 960  *
 961  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 962  *
 963  * The btree node will have either a read or a write lock held, depending on
 964  * level and op->lock.
 965  */
 966 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 967                                  struct bkey *k, int level, bool write,
 968                                  struct btree *parent)
 969 {
 970         int i = 0;
 971         struct btree *b;
 972
 973         BUG_ON(level < 0);
 974 retry:
 975         b = mca_find(c, k);
 976
 977         if (!b) {
 978                 if (current->bio_list)
 979                         return ERR_PTR(-EAGAIN);
 980
 981                 mutex_lock(&c->bucket_lock);
 982                 b = mca_alloc(c, op, k, level);
 983                 mutex_unlock(&c->bucket_lock);
 984
 985                 if (!b)
 986                         goto retry;
 987                 if (IS_ERR(b))
 988                         return b;
 989
 990                 bch_btree_node_read(b);
 991
 992                 if (!write)
 993                         downgrade_write(&b->lock);
 994         } else {
 995                 rw_lock(write, b, level);
 996                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 997                         rw_unlock(write, b);
 998                         goto retry;
 999                 }
1000                 BUG_ON(b->level != level);
1001         }
1002
1003         b->parent = parent;
1004         b->accessed = 1;
1005
1006         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1007                 prefetch(b->keys.set[i].tree);
1008                 prefetch(b->keys.set[i].data);
1009         }
1010
1011         for (; i <= b->keys.nsets; i++)
1012                 prefetch(b->keys.set[i].data);
1013
1014         if (btree_node_io_error(b)) {
1015                 rw_unlock(write, b);
1016                 return ERR_PTR(-EIO);
1017         }
1018
1019         BUG_ON(!b->written);
1020
1021         return b;
1022 }
1023
1024 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1025 {
1026         struct btree *b;
1027
1028         mutex_lock(&parent->c->bucket_lock);
1029         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1030         mutex_unlock(&parent->c->bucket_lock);
1031
1032         if (!IS_ERR_OR_NULL(b)) {
1033                 b->parent = parent;
1034                 bch_btree_node_read(b);
1035                 rw_unlock(true, b);
1036         }
1037 }
1038
1039 /* Btree alloc */
1040
1041 static void btree_node_free(struct btree *b)
1042 {
1043         trace_bcache_btree_node_free(b);
1044
1045         BUG_ON(b == b->c->root);
1046
1047         mutex_lock(&b->write_lock);
1048
1049         if (btree_node_dirty(b))
1050                 btree_complete_write(b, btree_current_write(b));
1051         clear_bit(BTREE_NODE_dirty, &b->flags);
1052
1053         mutex_unlock(&b->write_lock);
1054
1055         cancel_delayed_work(&b->work);
1056
1057         mutex_lock(&b->c->bucket_lock);
1058         bch_bucket_free(b->c, &b->key);
1059         mca_bucket_free(b);
1060         mutex_unlock(&b->c->bucket_lock);
1061 }
1062
1063 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1064                                      int level, bool wait,
1065                                      struct btree *parent)
1066 {
1067         BKEY_PADDED(key) k;
1068         struct btree *b = ERR_PTR(-EAGAIN);
1069
1070         mutex_lock(&c->bucket_lock);
1071 retry:
1072         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1073                 goto err;
1074
1075         bkey_put(c, &k.key);
1076         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1077
1078         b = mca_alloc(c, op, &k.key, level);
1079         if (IS_ERR(b))
1080                 goto err_free;
1081
1082         if (!b) {
1083                 cache_bug(c,
1084                         "Tried to allocate bucket that was in btree cache");
1085                 goto retry;
1086         }
1087
1088         b->accessed = 1;
1089         b->parent = parent;
1090         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1091
1092         mutex_unlock(&c->bucket_lock);
1093
1094         trace_bcache_btree_node_alloc(b);
1095         return b;
1096 err_free:
1097         bch_bucket_free(c, &k.key);
1098 err:
1099         mutex_unlock(&c->bucket_lock);
1100
1101         trace_bcache_btree_node_alloc_fail(c);
1102         return b;
1103 }
1104
1105 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1106                                           struct btree_op *op, int level,
1107                                           struct btree *parent)
1108 {
1109         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1110 }
1111
1112 static struct btree *btree_node_alloc_replacement(struct btree *b,
1113                                                   struct btree_op *op)
1114 {
1115         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1116         if (!IS_ERR_OR_NULL(n)) {
1117                 mutex_lock(&n->write_lock);
1118                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1119                 bkey_copy_key(&n->key, &b->key);
1120                 mutex_unlock(&n->write_lock);
1121         }
1122
1123         return n;
1124 }
1125
1126 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1127 {
1128         unsigned i;
1129
1130         mutex_lock(&b->c->bucket_lock);
1131
1132         atomic_inc(&b->c->prio_blocked);
1133
1134         bkey_copy(k, &b->key);
1135         bkey_copy_key(k, &ZERO_KEY);
1136
1137         for (i = 0; i < KEY_PTRS(k); i++)
1138                 SET_PTR_GEN(k, i,
1139                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1140                                         PTR_BUCKET(b->c, &b->key, i)));
1141
1142         mutex_unlock(&b->c->bucket_lock);
1143 }
1144
1145 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1146 {
1147         struct cache_set *c = b->c;
1148         struct cache *ca;
1149         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1150
1151         mutex_lock(&c->bucket_lock);
1152
1153         for_each_cache(ca, c, i)
1154                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1155                         if (op)
1156                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1157                                                 TASK_UNINTERRUPTIBLE);
1158                         mutex_unlock(&c->bucket_lock);
1159                         return -EINTR;
1160                 }
1161
1162         mutex_unlock(&c->bucket_lock);
1163
1164         return mca_cannibalize_lock(b->c, op);
1165 }
1166
1167 /* Garbage collection */
1168
1169 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1170                                     struct bkey *k)
1171 {
1172         uint8_t stale = 0;
1173         unsigned i;
1174         struct bucket *g;
1175
1176         /*
1177          * ptr_invalid() can't return true for the keys that mark btree nodes as
1178          * freed, but since ptr_bad() returns true we'll never actually use them
1179          * for anything and thus we don't want mark their pointers here
1180          */
1181         if (!bkey_cmp(k, &ZERO_KEY))
1182                 return stale;
1183
1184         for (i = 0; i < KEY_PTRS(k); i++) {
1185                 if (!ptr_available(c, k, i))
1186                         continue;
1187
1188                 g = PTR_BUCKET(c, k, i);
1189
1190                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1191                         g->last_gc = PTR_GEN(k, i);
1192
1193                 if (ptr_stale(c, k, i)) {
1194                         stale = max(stale, ptr_stale(c, k, i));
1195                         continue;
1196                 }
1197
1198                 cache_bug_on(GC_MARK(g) &&
1199                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1200                              c, "inconsistent ptrs: mark = %llu, level = %i",
1201                              GC_MARK(g), level);
1202
1203                 if (level)
1204                         SET_GC_MARK(g, GC_MARK_METADATA);
1205                 else if (KEY_DIRTY(k))
1206                         SET_GC_MARK(g, GC_MARK_DIRTY);
1207                 else if (!GC_MARK(g))
1208                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1209
1210                 /* guard against overflow */
1211                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1212                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1213                                              MAX_GC_SECTORS_USED));
1214
1215                 BUG_ON(!GC_SECTORS_USED(g));
1216         }
1217
1218         return stale;
1219 }
1220
1221 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1222
1223 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1224 {
1225         unsigned i;
1226
1227         for (i = 0; i < KEY_PTRS(k); i++)
1228                 if (ptr_available(c, k, i) &&
1229                     !ptr_stale(c, k, i)) {
1230                         struct bucket *b = PTR_BUCKET(c, k, i);
1231
1232                         b->gen = PTR_GEN(k, i);
1233
1234                         if (level && bkey_cmp(k, &ZERO_KEY))
1235                                 b->prio = BTREE_PRIO;
1236                         else if (!level && b->prio == BTREE_PRIO)
1237                                 b->prio = INITIAL_PRIO;
1238                 }
1239
1240         __bch_btree_mark_key(c, level, k);
1241 }
1242
1243 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1244 {
1245         uint8_t stale = 0;
1246         unsigned keys = 0, good_keys = 0;
1247         struct bkey *k;
1248         struct btree_iter iter;
1249         struct bset_tree *t;
1250
1251         gc->nodes++;
1252
1253         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1254                 stale = max(stale, btree_mark_key(b, k));
1255                 keys++;
1256
1257                 if (bch_ptr_bad(&b->keys, k))
1258                         continue;
1259
1260                 gc->key_bytes += bkey_u64s(k);
1261                 gc->nkeys++;
1262                 good_keys++;
1263
1264                 gc->data += KEY_SIZE(k);
1265         }
1266
1267         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1268                 btree_bug_on(t->size &&
1269                              bset_written(&b->keys, t) &&
1270                              bkey_cmp(&b->key, &t->end) < 0,
1271                              b, "found short btree key in gc");
1272
1273         if (b->c->gc_always_rewrite)
1274                 return true;
1275
1276         if (stale > 10)
1277                 return true;
1278
1279         if ((keys - good_keys) * 2 > keys)
1280                 return true;
1281
1282         return false;
1283 }
1284
1285 #define GC_MERGE_NODES  4U
1286
1287 struct gc_merge_info {
1288         struct btree    *b;
1289         unsigned        keys;
1290 };
1291
1292 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1293                                  struct keylist *, atomic_t *, struct bkey *);
1294
1295 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1296                              struct gc_stat *gc, struct gc_merge_info *r)
1297 {
1298         unsigned i, nodes = 0, keys = 0, blocks;
1299         struct btree *new_nodes[GC_MERGE_NODES];
1300         struct keylist keylist;
1301         struct closure cl;
1302         struct bkey *k;
1303
1304         bch_keylist_init(&keylist);
1305
1306         if (btree_check_reserve(b, NULL))
1307                 return 0;
1308
1309         memset(new_nodes, 0, sizeof(new_nodes));
1310         closure_init_stack(&cl);
1311
1312         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1313                 keys += r[nodes++].keys;
1314
1315         blocks = btree_default_blocks(b->c) * 2 / 3;
1316
1317         if (nodes < 2 ||
1318             __set_blocks(b->keys.set[0].data, keys,
1319                          block_bytes(b->c)) > blocks * (nodes - 1))
1320                 return 0;
1321
1322         for (i = 0; i < nodes; i++) {
1323                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1324                 if (IS_ERR_OR_NULL(new_nodes[i]))
1325                         goto out_nocoalesce;
1326         }
1327
1328         /*
1329          * We have to check the reserve here, after we've allocated our new
1330          * nodes, to make sure the insert below will succeed - we also check
1331          * before as an optimization to potentially avoid a bunch of expensive
1332          * allocs/sorts
1333          */
1334         if (btree_check_reserve(b, NULL))
1335                 goto out_nocoalesce;
1336
1337         for (i = 0; i < nodes; i++)
1338                 mutex_lock(&new_nodes[i]->write_lock);
1339
1340         for (i = nodes - 1; i > 0; --i) {
1341                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1342                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1343                 struct bkey *k, *last = NULL;
1344
1345                 keys = 0;
1346
1347                 if (i > 1) {
1348                         for (k = n2->start;
1349                              k < bset_bkey_last(n2);
1350                              k = bkey_next(k)) {
1351                                 if (__set_blocks(n1, n1->keys + keys +
1352                                                  bkey_u64s(k),
1353                                                  block_bytes(b->c)) > blocks)
1354                                         break;
1355
1356                                 last = k;
1357                                 keys += bkey_u64s(k);
1358                         }
1359                 } else {
1360                         /*
1361                          * Last node we're not getting rid of - we're getting
1362                          * rid of the node at r[0]. Have to try and fit all of
1363                          * the remaining keys into this node; we can't ensure
1364                          * they will always fit due to rounding and variable
1365                          * length keys (shouldn't be possible in practice,
1366                          * though)
1367                          */
1368                         if (__set_blocks(n1, n1->keys + n2->keys,
1369                                          block_bytes(b->c)) >
1370                             btree_blocks(new_nodes[i]))
1371                                 goto out_nocoalesce;
1372
1373                         keys = n2->keys;
1374                         /* Take the key of the node we're getting rid of */
1375                         last = &r->b->key;
1376                 }
1377
1378                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1379                        btree_blocks(new_nodes[i]));
1380
1381                 if (last)
1382                         bkey_copy_key(&new_nodes[i]->key, last);
1383
1384                 memcpy(bset_bkey_last(n1),
1385                        n2->start,
1386                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1387
1388                 n1->keys += keys;
1389                 r[i].keys = n1->keys;
1390
1391                 memmove(n2->start,
1392                         bset_bkey_idx(n2, keys),
1393                         (void *) bset_bkey_last(n2) -
1394                         (void *) bset_bkey_idx(n2, keys));
1395
1396                 n2->keys -= keys;
1397
1398                 if (__bch_keylist_realloc(&keylist,
1399                                           bkey_u64s(&new_nodes[i]->key)))
1400                         goto out_nocoalesce;
1401
1402                 bch_btree_node_write(new_nodes[i], &cl);
1403                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1404         }
1405
1406         for (i = 0; i < nodes; i++)
1407                 mutex_unlock(&new_nodes[i]->write_lock);
1408
1409         closure_sync(&cl);
1410
1411         /* We emptied out this node */
1412         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1413         btree_node_free(new_nodes[0]);
1414         rw_unlock(true, new_nodes[0]);
1415         new_nodes[0] = NULL;
1416
1417         for (i = 0; i < nodes; i++) {
1418                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1419                         goto out_nocoalesce;
1420
1421                 make_btree_freeing_key(r[i].b, keylist.top);
1422                 bch_keylist_push(&keylist);
1423         }
1424
1425         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1426         BUG_ON(!bch_keylist_empty(&keylist));
1427
1428         for (i = 0; i < nodes; i++) {
1429                 btree_node_free(r[i].b);
1430                 rw_unlock(true, r[i].b);
1431
1432                 r[i].b = new_nodes[i];
1433         }
1434
1435         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1436         r[nodes - 1].b = ERR_PTR(-EINTR);
1437
1438         trace_bcache_btree_gc_coalesce(nodes);
1439         gc->nodes--;
1440
1441         bch_keylist_free(&keylist);
1442
1443         /* Invalidated our iterator */
1444         return -EINTR;
1445
1446 out_nocoalesce:
1447         closure_sync(&cl);
1448         bch_keylist_free(&keylist);
1449
1450         while ((k = bch_keylist_pop(&keylist)))
1451                 if (!bkey_cmp(k, &ZERO_KEY))
1452                         atomic_dec(&b->c->prio_blocked);
1453
1454         for (i = 0; i < nodes; i++)
1455                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1456                         btree_node_free(new_nodes[i]);
1457                         rw_unlock(true, new_nodes[i]);
1458                 }
1459         return 0;
1460 }
1461
1462 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1463                                  struct btree *replace)
1464 {
1465         struct keylist keys;
1466         struct btree *n;
1467
1468         if (btree_check_reserve(b, NULL))
1469                 return 0;
1470
1471         n = btree_node_alloc_replacement(replace, NULL);
1472
1473         /* recheck reserve after allocating replacement node */
1474         if (btree_check_reserve(b, NULL)) {
1475                 btree_node_free(n);
1476                 rw_unlock(true, n);
1477                 return 0;
1478         }
1479
1480         bch_btree_node_write_sync(n);
1481
1482         bch_keylist_init(&keys);
1483         bch_keylist_add(&keys, &n->key);
1484
1485         make_btree_freeing_key(replace, keys.top);
1486         bch_keylist_push(&keys);
1487
1488         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1489         BUG_ON(!bch_keylist_empty(&keys));
1490
1491         btree_node_free(replace);
1492         rw_unlock(true, n);
1493
1494         /* Invalidated our iterator */
1495         return -EINTR;
1496 }
1497
1498 static unsigned btree_gc_count_keys(struct btree *b)
1499 {
1500         struct bkey *k;
1501         struct btree_iter iter;
1502         unsigned ret = 0;
1503
1504         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1505                 ret += bkey_u64s(k);
1506
1507         return ret;
1508 }
1509
1510 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1511                             struct closure *writes, struct gc_stat *gc)
1512 {
1513         int ret = 0;
1514         bool should_rewrite;
1515         struct bkey *k;
1516         struct btree_iter iter;
1517         struct gc_merge_info r[GC_MERGE_NODES];
1518         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1519
1520         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1521
1522         for (i = r; i < r + ARRAY_SIZE(r); i++)
1523                 i->b = ERR_PTR(-EINTR);
1524
1525         while (1) {
1526                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1527                 if (k) {
1528                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1529                                                   true, b);
1530                         if (IS_ERR(r->b)) {
1531                                 ret = PTR_ERR(r->b);
1532                                 break;
1533                         }
1534
1535                         r->keys = btree_gc_count_keys(r->b);
1536
1537                         ret = btree_gc_coalesce(b, op, gc, r);
1538                         if (ret)
1539                                 break;
1540                 }
1541
1542                 if (!last->b)
1543                         break;
1544
1545                 if (!IS_ERR(last->b)) {
1546                         should_rewrite = btree_gc_mark_node(last->b, gc);
1547                         if (should_rewrite) {
1548                                 ret = btree_gc_rewrite_node(b, op, last->b);
1549                                 if (ret)
1550                                         break;
1551                         }
1552
1553                         if (last->b->level) {
1554                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1555                                 if (ret)
1556                                         break;
1557                         }
1558
1559                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1560
1561                         /*
1562                          * Must flush leaf nodes before gc ends, since replace
1563                          * operations aren't journalled
1564                          */
1565                         mutex_lock(&last->b->write_lock);
1566                         if (btree_node_dirty(last->b))
1567                                 bch_btree_node_write(last->b, writes);
1568                         mutex_unlock(&last->b->write_lock);
1569                         rw_unlock(true, last->b);
1570                 }
1571
1572                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1573                 r->b = NULL;
1574
1575                 if (need_resched()) {
1576                         ret = -EAGAIN;
1577                         break;
1578                 }
1579         }
1580
1581         for (i = r; i < r + ARRAY_SIZE(r); i++)
1582                 if (!IS_ERR_OR_NULL(i->b)) {
1583                         mutex_lock(&i->b->write_lock);
1584                         if (btree_node_dirty(i->b))
1585                                 bch_btree_node_write(i->b, writes);
1586                         mutex_unlock(&i->b->write_lock);
1587                         rw_unlock(true, i->b);
1588                 }
1589
1590         return ret;
1591 }
1592
1593 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1594                              struct closure *writes, struct gc_stat *gc)
1595 {
1596         struct btree *n = NULL;
1597         int ret = 0;
1598         bool should_rewrite;
1599
1600         should_rewrite = btree_gc_mark_node(b, gc);
1601         if (should_rewrite) {
1602                 n = btree_node_alloc_replacement(b, NULL);
1603
1604                 if (!IS_ERR_OR_NULL(n)) {
1605                         bch_btree_node_write_sync(n);
1606
1607                         bch_btree_set_root(n);
1608                         btree_node_free(b);
1609                         rw_unlock(true, n);
1610
1611                         return -EINTR;
1612                 }
1613         }
1614
1615         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1616
1617         if (b->level) {
1618                 ret = btree_gc_recurse(b, op, writes, gc);
1619                 if (ret)
1620                         return ret;
1621         }
1622
1623         bkey_copy_key(&b->c->gc_done, &b->key);
1624
1625         return ret;
1626 }
1627
1628 static void btree_gc_start(struct cache_set *c)
1629 {
1630         struct cache *ca;
1631         struct bucket *b;
1632         unsigned i;
1633
1634         if (!c->gc_mark_valid)
1635                 return;
1636
1637         mutex_lock(&c->bucket_lock);
1638
1639         c->gc_mark_valid = 0;
1640         c->gc_done = ZERO_KEY;
1641
1642         for_each_cache(ca, c, i)
1643                 for_each_bucket(b, ca) {
1644                         b->last_gc = b->gen;
1645                         if (!atomic_read(&b->pin)) {
1646                                 SET_GC_MARK(b, 0);
1647                                 SET_GC_SECTORS_USED(b, 0);
1648                         }
1649                 }
1650
1651         mutex_unlock(&c->bucket_lock);
1652 }
1653
1654 static size_t bch_btree_gc_finish(struct cache_set *c)
1655 {
1656         size_t available = 0;
1657         struct bucket *b;
1658         struct cache *ca;
1659         unsigned i;
1660
1661         mutex_lock(&c->bucket_lock);
1662
1663         set_gc_sectors(c);
1664         c->gc_mark_valid = 1;
1665         c->need_gc      = 0;
1666
1667         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1668                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1669                             GC_MARK_METADATA);
1670
1671         /* don't reclaim buckets to which writeback keys point */
1672         rcu_read_lock();
1673         for (i = 0; i < c->nr_uuids; i++) {
1674                 struct bcache_device *d = c->devices[i];
1675                 struct cached_dev *dc;
1676                 struct keybuf_key *w, *n;
1677                 unsigned j;
1678
1679                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1680                         continue;
1681                 dc = container_of(d, struct cached_dev, disk);
1682
1683                 spin_lock(&dc->writeback_keys.lock);
1684                 rbtree_postorder_for_each_entry_safe(w, n,
1685                                         &dc->writeback_keys.keys, node)
1686                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1687                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1688                                             GC_MARK_DIRTY);
1689                 spin_unlock(&dc->writeback_keys.lock);
1690         }
1691         rcu_read_unlock();
1692
1693         for_each_cache(ca, c, i) {
1694                 uint64_t *i;
1695
1696                 ca->invalidate_needs_gc = 0;
1697
1698                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1699                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1700
1701                 for (i = ca->prio_buckets;
1702                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1703                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1704
1705                 for_each_bucket(b, ca) {
1706                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1707
1708                         if (atomic_read(&b->pin))
1709                                 continue;
1710
1711                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1712
1713                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1714                                 available++;
1715                 }
1716         }
1717
1718         mutex_unlock(&c->bucket_lock);
1719         return available;
1720 }
1721
1722 static void bch_btree_gc(struct cache_set *c)
1723 {
1724         int ret;
1725         unsigned long available;
1726         struct gc_stat stats;
1727         struct closure writes;
1728         struct btree_op op;
1729         uint64_t start_time = local_clock();
1730
1731         trace_bcache_gc_start(c);
1732
1733         memset(&stats, 0, sizeof(struct gc_stat));
1734         closure_init_stack(&writes);
1735         bch_btree_op_init(&op, SHRT_MAX);
1736
1737         btree_gc_start(c);
1738
1739         do {
1740                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1741                 closure_sync(&writes);
1742                 cond_resched();
1743
1744                 if (ret && ret != -EAGAIN)
1745                         pr_warn("gc failed!");
1746         } while (ret);
1747
1748         available = bch_btree_gc_finish(c);
1749         wake_up_allocators(c);
1750
1751         bch_time_stats_update(&c->btree_gc_time, start_time);
1752
1753         stats.key_bytes *= sizeof(uint64_t);
1754         stats.data      <<= 9;
1755         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1756         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1757
1758         trace_bcache_gc_end(c);
1759
1760         bch_moving_gc(c);
1761 }
1762
1763 static bool gc_should_run(struct cache_set *c)
1764 {
1765         struct cache *ca;
1766         unsigned i;
1767
1768         for_each_cache(ca, c, i)
1769                 if (ca->invalidate_needs_gc)
1770                         return true;
1771
1772         if (atomic_read(&c->sectors_to_gc) < 0)
1773                 return true;
1774
1775         return false;
1776 }
1777
1778 static int bch_gc_thread(void *arg)
1779 {
1780         struct cache_set *c = arg;
1781
1782         while (1) {
1783                 wait_event_interruptible(c->gc_wait,
1784                            kthread_should_stop() || gc_should_run(c));
1785
1786                 if (kthread_should_stop())
1787                         break;
1788
1789                 set_gc_sectors(c);
1790                 bch_btree_gc(c);
1791         }
1792
1793         return 0;
1794 }
1795
1796 int bch_gc_thread_start(struct cache_set *c)
1797 {
1798         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1799         if (IS_ERR(c->gc_thread))
1800                 return PTR_ERR(c->gc_thread);
1801
1802         return 0;
1803 }
1804
1805 /* Initial partial gc */
1806
1807 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1808 {
1809         int ret = 0;
1810         struct bkey *k, *p = NULL;
1811         struct btree_iter iter;
1812
1813         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1814                 bch_initial_mark_key(b->c, b->level, k);
1815
1816         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1817
1818         if (b->level) {
1819                 bch_btree_iter_init(&b->keys, &iter, NULL);
1820
1821                 do {
1822                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1823                                                        bch_ptr_bad);
1824                         if (k)
1825                                 btree_node_prefetch(b, k);
1826
1827                         if (p)
1828                                 ret = btree(check_recurse, p, b, op);
1829
1830                         p = k;
1831                 } while (p && !ret);
1832         }
1833
1834         return ret;
1835 }
1836
1837 int bch_btree_check(struct cache_set *c)
1838 {
1839         struct btree_op op;
1840
1841         bch_btree_op_init(&op, SHRT_MAX);
1842
1843         return btree_root(check_recurse, c, &op);
1844 }
1845
1846 void bch_initial_gc_finish(struct cache_set *c)
1847 {
1848         struct cache *ca;
1849         struct bucket *b;
1850         unsigned i;
1851
1852         bch_btree_gc_finish(c);
1853
1854         mutex_lock(&c->bucket_lock);
1855
1856         /*
1857          * We need to put some unused buckets directly on the prio freelist in
1858          * order to get the allocator thread started - it needs freed buckets in
1859          * order to rewrite the prios and gens, and it needs to rewrite prios
1860          * and gens in order to free buckets.
1861          *
1862          * This is only safe for buckets that have no live data in them, which
1863          * there should always be some of.
1864          */
1865         for_each_cache(ca, c, i) {
1866                 for_each_bucket(b, ca) {
1867                         if (fifo_full(&ca->free[RESERVE_PRIO]))
1868                                 break;
1869
1870                         if (bch_can_invalidate_bucket(ca, b) &&
1871                             !GC_MARK(b)) {
1872                                 __bch_invalidate_one_bucket(ca, b);
1873                                 fifo_push(&ca->free[RESERVE_PRIO],
1874                                           b - ca->buckets);
1875                         }
1876                 }
1877         }
1878
1879         mutex_unlock(&c->bucket_lock);
1880 }
1881
1882 /* Btree insertion */
1883
1884 static bool btree_insert_key(struct btree *b, struct bkey *k,
1885                              struct bkey *replace_key)
1886 {
1887         unsigned status;
1888
1889         BUG_ON(bkey_cmp(k, &b->key) > 0);
1890
1891         status = bch_btree_insert_key(&b->keys, k, replace_key);
1892         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1893                 bch_check_keys(&b->keys, "%u for %s", status,
1894                                replace_key ? "replace" : "insert");
1895
1896                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1897                                               status);
1898                 return true;
1899         } else
1900                 return false;
1901 }
1902
1903 static size_t insert_u64s_remaining(struct btree *b)
1904 {
1905         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1906
1907         /*
1908          * Might land in the middle of an existing extent and have to split it
1909          */
1910         if (b->keys.ops->is_extents)
1911                 ret -= KEY_MAX_U64S;
1912
1913         return max(ret, 0L);
1914 }
1915
1916 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1917                                   struct keylist *insert_keys,
1918                                   struct bkey *replace_key)
1919 {
1920         bool ret = false;
1921         int oldsize = bch_count_data(&b->keys);
1922
1923         while (!bch_keylist_empty(insert_keys)) {
1924                 struct bkey *k = insert_keys->keys;
1925
1926                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1927                         break;
1928
1929                 if (bkey_cmp(k, &b->key) <= 0) {
1930                         if (!b->level)
1931                                 bkey_put(b->c, k);
1932
1933                         ret |= btree_insert_key(b, k, replace_key);
1934                         bch_keylist_pop_front(insert_keys);
1935                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1936                         BKEY_PADDED(key) temp;
1937                         bkey_copy(&temp.key, insert_keys->keys);
1938
1939                         bch_cut_back(&b->key, &temp.key);
1940                         bch_cut_front(&b->key, insert_keys->keys);
1941
1942                         ret |= btree_insert_key(b, &temp.key, replace_key);
1943                         break;
1944                 } else {
1945                         break;
1946                 }
1947         }
1948
1949         if (!ret)
1950                 op->insert_collision = true;
1951
1952         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1953
1954         BUG_ON(bch_count_data(&b->keys) < oldsize);
1955         return ret;
1956 }
1957
1958 static int btree_split(struct btree *b, struct btree_op *op,
1959                        struct keylist *insert_keys,
1960                        struct bkey *replace_key)
1961 {
1962         bool split;
1963         struct btree *n1, *n2 = NULL, *n3 = NULL;
1964         uint64_t start_time = local_clock();
1965         struct closure cl;
1966         struct keylist parent_keys;
1967
1968         closure_init_stack(&cl);
1969         bch_keylist_init(&parent_keys);
1970
1971         if (btree_check_reserve(b, op)) {
1972                 if (!b->level)
1973                         return -EINTR;
1974                 else
1975                         WARN(1, "insufficient reserve for split\n");
1976         }
1977
1978         n1 = btree_node_alloc_replacement(b, op);
1979         if (IS_ERR(n1))
1980                 goto err;
1981
1982         split = set_blocks(btree_bset_first(n1),
1983                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1984
1985         if (split) {
1986                 unsigned keys = 0;
1987
1988                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1989
1990                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1991                 if (IS_ERR(n2))
1992                         goto err_free1;
1993
1994                 if (!b->parent) {
1995                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
1996                         if (IS_ERR(n3))
1997                                 goto err_free2;
1998                 }
1999
2000                 mutex_lock(&n1->write_lock);
2001                 mutex_lock(&n2->write_lock);
2002
2003                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2004
2005                 /*
2006                  * Has to be a linear search because we don't have an auxiliary
2007                  * search tree yet
2008                  */
2009
2010                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2011                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2012                                                         keys));
2013
2014                 bkey_copy_key(&n1->key,
2015                               bset_bkey_idx(btree_bset_first(n1), keys));
2016                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2017
2018                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2019                 btree_bset_first(n1)->keys = keys;
2020
2021                 memcpy(btree_bset_first(n2)->start,
2022                        bset_bkey_last(btree_bset_first(n1)),
2023                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2024
2025                 bkey_copy_key(&n2->key, &b->key);
2026
2027                 bch_keylist_add(&parent_keys, &n2->key);
2028                 bch_btree_node_write(n2, &cl);
2029                 mutex_unlock(&n2->write_lock);
2030                 rw_unlock(true, n2);
2031         } else {
2032                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2033
2034                 mutex_lock(&n1->write_lock);
2035                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2036         }
2037
2038         bch_keylist_add(&parent_keys, &n1->key);
2039         bch_btree_node_write(n1, &cl);
2040         mutex_unlock(&n1->write_lock);
2041
2042         if (n3) {
2043                 /* Depth increases, make a new root */
2044                 mutex_lock(&n3->write_lock);
2045                 bkey_copy_key(&n3->key, &MAX_KEY);
2046                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2047                 bch_btree_node_write(n3, &cl);
2048                 mutex_unlock(&n3->write_lock);
2049
2050                 closure_sync(&cl);
2051                 bch_btree_set_root(n3);
2052                 rw_unlock(true, n3);
2053         } else if (!b->parent) {
2054                 /* Root filled up but didn't need to be split */
2055                 closure_sync(&cl);
2056                 bch_btree_set_root(n1);
2057         } else {
2058                 /* Split a non root node */
2059                 closure_sync(&cl);
2060                 make_btree_freeing_key(b, parent_keys.top);
2061                 bch_keylist_push(&parent_keys);
2062
2063                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2064                 BUG_ON(!bch_keylist_empty(&parent_keys));
2065         }
2066
2067         btree_node_free(b);
2068         rw_unlock(true, n1);
2069
2070         bch_time_stats_update(&b->c->btree_split_time, start_time);
2071
2072         return 0;
2073 err_free2:
2074         bkey_put(b->c, &n2->key);
2075         btree_node_free(n2);
2076         rw_unlock(true, n2);
2077 err_free1:
2078         bkey_put(b->c, &n1->key);
2079         btree_node_free(n1);
2080         rw_unlock(true, n1);
2081 err:
2082         WARN(1, "bcache: btree split failed (level %u)", b->level);
2083
2084         if (n3 == ERR_PTR(-EAGAIN) ||
2085             n2 == ERR_PTR(-EAGAIN) ||
2086             n1 == ERR_PTR(-EAGAIN))
2087                 return -EAGAIN;
2088
2089         return -ENOMEM;
2090 }
2091
2092 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2093                                  struct keylist *insert_keys,
2094                                  atomic_t *journal_ref,
2095                                  struct bkey *replace_key)
2096 {
2097         struct closure cl;
2098
2099         BUG_ON(b->level && replace_key);
2100
2101         closure_init_stack(&cl);
2102
2103         mutex_lock(&b->write_lock);
2104
2105         if (write_block(b) != btree_bset_last(b) &&
2106             b->keys.last_set_unwritten)
2107                 bch_btree_init_next(b); /* just wrote a set */
2108
2109         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2110                 mutex_unlock(&b->write_lock);
2111                 goto split;
2112         }
2113
2114         BUG_ON(write_block(b) != btree_bset_last(b));
2115
2116         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2117                 if (!b->level)
2118                         bch_btree_leaf_dirty(b, journal_ref);
2119                 else
2120                         bch_btree_node_write(b, &cl);
2121         }
2122
2123         mutex_unlock(&b->write_lock);
2124
2125         /* wait for btree node write if necessary, after unlock */
2126         closure_sync(&cl);
2127
2128         return 0;
2129 split:
2130         if (current->bio_list) {
2131                 op->lock = b->c->root->level + 1;
2132                 return -EAGAIN;
2133         } else if (op->lock <= b->c->root->level) {
2134                 op->lock = b->c->root->level + 1;
2135                 return -EINTR;
2136         } else {
2137                 /* Invalidated all iterators */
2138                 int ret = btree_split(b, op, insert_keys, replace_key);
2139
2140                 if (bch_keylist_empty(insert_keys))
2141                         return 0;
2142                 else if (!ret)
2143                         return -EINTR;
2144                 return ret;
2145         }
2146 }
2147
2148 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2149                                struct bkey *check_key)
2150 {
2151         int ret = -EINTR;
2152         uint64_t btree_ptr = b->key.ptr[0];
2153         unsigned long seq = b->seq;
2154         struct keylist insert;
2155         bool upgrade = op->lock == -1;
2156
2157         bch_keylist_init(&insert);
2158
2159         if (upgrade) {
2160                 rw_unlock(false, b);
2161                 rw_lock(true, b, b->level);
2162
2163                 if (b->key.ptr[0] != btree_ptr ||
2164                    b->seq != seq + 1) {
2165                        op->lock = b->level;
2166                         goto out;
2167                }
2168         }
2169
2170         SET_KEY_PTRS(check_key, 1);
2171         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2172
2173         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2174
2175         bch_keylist_add(&insert, check_key);
2176
2177         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2178
2179         BUG_ON(!ret && !bch_keylist_empty(&insert));
2180 out:
2181         if (upgrade)
2182                 downgrade_write(&b->lock);
2183         return ret;
2184 }
2185
2186 struct btree_insert_op {
2187         struct btree_op op;
2188         struct keylist  *keys;
2189         atomic_t        *journal_ref;
2190         struct bkey     *replace_key;
2191 };
2192
2193 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2194 {
2195         struct btree_insert_op *op = container_of(b_op,
2196                                         struct btree_insert_op, op);
2197
2198         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2199                                         op->journal_ref, op->replace_key);
2200         if (ret && !bch_keylist_empty(op->keys))
2201                 return ret;
2202         else
2203                 return MAP_DONE;
2204 }
2205
2206 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2207                      atomic_t *journal_ref, struct bkey *replace_key)
2208 {
2209         struct btree_insert_op op;
2210         int ret = 0;
2211
2212         BUG_ON(current->bio_list);
2213         BUG_ON(bch_keylist_empty(keys));
2214
2215         bch_btree_op_init(&op.op, 0);
2216         op.keys         = keys;
2217         op.journal_ref  = journal_ref;
2218         op.replace_key  = replace_key;
2219
2220         while (!ret && !bch_keylist_empty(keys)) {
2221                 op.op.lock = 0;
2222                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2223                                                &START_KEY(keys->keys),
2224                                                btree_insert_fn);
2225         }
2226
2227         if (ret) {
2228                 struct bkey *k;
2229
2230                 pr_err("error %i", ret);
2231
2232                 while ((k = bch_keylist_pop(keys)))
2233                         bkey_put(c, k);
2234         } else if (op.op.insert_collision)
2235                 ret = -ESRCH;
2236
2237         return ret;
2238 }
2239
2240 void bch_btree_set_root(struct btree *b)
2241 {
2242         unsigned i;
2243         struct closure cl;
2244
2245         closure_init_stack(&cl);
2246
2247         trace_bcache_btree_set_root(b);
2248
2249         BUG_ON(!b->written);
2250
2251         for (i = 0; i < KEY_PTRS(&b->key); i++)
2252                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2253
2254         mutex_lock(&b->c->bucket_lock);
2255         list_del_init(&b->list);
2256         mutex_unlock(&b->c->bucket_lock);
2257
2258         b->c->root = b;
2259
2260         bch_journal_meta(b->c, &cl);
2261         closure_sync(&cl);
2262 }
2263
2264 /* Map across nodes or keys */
2265
2266 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2267                                        struct bkey *from,
2268                                        btree_map_nodes_fn *fn, int flags)
2269 {
2270         int ret = MAP_CONTINUE;
2271
2272         if (b->level) {
2273                 struct bkey *k;
2274                 struct btree_iter iter;
2275
2276                 bch_btree_iter_init(&b->keys, &iter, from);
2277
2278                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2279                                                        bch_ptr_bad))) {
2280                         ret = btree(map_nodes_recurse, k, b,
2281                                     op, from, fn, flags);
2282                         from = NULL;
2283
2284                         if (ret != MAP_CONTINUE)
2285                                 return ret;
2286                 }
2287         }
2288
2289         if (!b->level || flags == MAP_ALL_NODES)
2290                 ret = fn(op, b);
2291
2292         return ret;
2293 }
2294
2295 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2296                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2297 {
2298         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2299 }
2300
2301 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2302                                       struct bkey *from, btree_map_keys_fn *fn,
2303                                       int flags)
2304 {
2305         int ret = MAP_CONTINUE;
2306         struct bkey *k;
2307         struct btree_iter iter;
2308
2309         bch_btree_iter_init(&b->keys, &iter, from);
2310
2311         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2312                 ret = !b->level
2313                         ? fn(op, b, k)
2314                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2315                 from = NULL;
2316
2317                 if (ret != MAP_CONTINUE)
2318                         return ret;
2319         }
2320
2321         if (!b->level && (flags & MAP_END_KEY))
2322                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2323                                      KEY_OFFSET(&b->key), 0));
2324
2325         return ret;
2326 }
2327
2328 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2329                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2330 {
2331         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2332 }
2333
2334 /* Keybuf code */
2335
2336 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2337 {
2338         /* Overlapping keys compare equal */
2339         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2340                 return -1;
2341         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2342                 return 1;
2343         return 0;
2344 }
2345
2346 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2347                                             struct keybuf_key *r)
2348 {
2349         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2350 }
2351
2352 struct refill {
2353         struct btree_op op;
2354         unsigned        nr_found;
2355         struct keybuf   *buf;
2356         struct bkey     *end;
2357         keybuf_pred_fn  *pred;
2358 };
2359
2360 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2361                             struct bkey *k)
2362 {
2363         struct refill *refill = container_of(op, struct refill, op);
2364         struct keybuf *buf = refill->buf;
2365         int ret = MAP_CONTINUE;
2366
2367         if (bkey_cmp(k, refill->end) >= 0) {
2368                 ret = MAP_DONE;
2369                 goto out;
2370         }
2371
2372         if (!KEY_SIZE(k)) /* end key */
2373                 goto out;
2374
2375         if (refill->pred(buf, k)) {
2376                 struct keybuf_key *w;
2377
2378                 spin_lock(&buf->lock);
2379
2380                 w = array_alloc(&buf->freelist);
2381                 if (!w) {
2382                         spin_unlock(&buf->lock);
2383                         return MAP_DONE;
2384                 }
2385
2386                 w->private = NULL;
2387                 bkey_copy(&w->key, k);
2388
2389                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2390                         array_free(&buf->freelist, w);
2391                 else
2392                         refill->nr_found++;
2393
2394                 if (array_freelist_empty(&buf->freelist))
2395                         ret = MAP_DONE;
2396
2397                 spin_unlock(&buf->lock);
2398         }
2399 out:
2400         buf->last_scanned = *k;
2401         return ret;
2402 }
2403
2404 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2405                        struct bkey *end, keybuf_pred_fn *pred)
2406 {
2407         struct bkey start = buf->last_scanned;
2408         struct refill refill;
2409
2410         cond_resched();
2411
2412         bch_btree_op_init(&refill.op, -1);
2413         refill.nr_found = 0;
2414         refill.buf      = buf;
2415         refill.end      = end;
2416         refill.pred     = pred;
2417
2418         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2419                            refill_keybuf_fn, MAP_END_KEY);
2420
2421         trace_bcache_keyscan(refill.nr_found,
2422                              KEY_INODE(&start), KEY_OFFSET(&start),
2423                              KEY_INODE(&buf->last_scanned),
2424                              KEY_OFFSET(&buf->last_scanned));
2425
2426         spin_lock(&buf->lock);
2427
2428         if (!RB_EMPTY_ROOT(&buf->keys)) {
2429                 struct keybuf_key *w;
2430                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2431                 buf->start      = START_KEY(&w->key);
2432
2433                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2434                 buf->end        = w->key;
2435         } else {
2436                 buf->start      = MAX_KEY;
2437                 buf->end        = MAX_KEY;
2438         }
2439
2440         spin_unlock(&buf->lock);
2441 }
2442
2443 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2444 {
2445         rb_erase(&w->node, &buf->keys);
2446         array_free(&buf->freelist, w);
2447 }
2448
2449 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2450 {
2451         spin_lock(&buf->lock);
2452         __bch_keybuf_del(buf, w);
2453         spin_unlock(&buf->lock);
2454 }
2455
2456 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2457                                   struct bkey *end)
2458 {
2459         bool ret = false;
2460         struct keybuf_key *p, *w, s;
2461         s.key = *start;
2462
2463         if (bkey_cmp(end, &buf->start) <= 0 ||
2464             bkey_cmp(start, &buf->end) >= 0)
2465                 return false;
2466
2467         spin_lock(&buf->lock);
2468         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2469
2470         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2471                 p = w;
2472                 w = RB_NEXT(w, node);
2473
2474                 if (p->private)
2475                         ret = true;
2476                 else
2477                         __bch_keybuf_del(buf, p);
2478         }
2479
2480         spin_unlock(&buf->lock);
2481         return ret;
2482 }
2483
2484 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2485 {
2486         struct keybuf_key *w;
2487         spin_lock(&buf->lock);
2488
2489         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2490
2491         while (w && w->private)
2492                 w = RB_NEXT(w, node);
2493
2494         if (w)
2495                 w->private = ERR_PTR(-EINTR);
2496
2497         spin_unlock(&buf->lock);
2498         return w;
2499 }
2500
2501 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2502                                           struct keybuf *buf,
2503                                           struct bkey *end,
2504                                           keybuf_pred_fn *pred)
2505 {
2506         struct keybuf_key *ret;
2507
2508         while (1) {
2509                 ret = bch_keybuf_next(buf);
2510                 if (ret)
2511                         break;
2512
2513                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2514                         pr_debug("scan finished");
2515                         break;
2516                 }
2517
2518                 bch_refill_keybuf(c, buf, end, pred);
2519         }
2520
2521         return ret;
2522 }
2523
2524 void bch_keybuf_init(struct keybuf *buf)
2525 {
2526         buf->last_scanned       = MAX_KEY;
2527         buf->keys               = RB_ROOT;
2528
2529         spin_lock_init(&buf->lock);
2530         array_allocator_init(&buf->freelist);
2531 }