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