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/freezer.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/random.h>
  35 #include <linux/rcupdate.h>
  36 #include <trace/events/bcache.h>
  37
  38 /*
  39  * Todo:
  40  * register_bcache: Return errors out to userspace correctly
  41  *
  42  * Writeback: don't undirty key until after a cache flush
  43  *
  44  * Create an iterator for key pointers
  45  *
  46  * On btree write error, mark bucket such that it won't be freed from the cache
  47  *
  48  * Journalling:
  49  *   Check for bad keys in replay
  50  *   Propagate barriers
  51  *   Refcount journal entries in journal_replay
  52  *
  53  * Garbage collection:
  54  *   Finish incremental gc
  55  *   Gc should free old UUIDs, data for invalid UUIDs
  56  *
  57  * Provide a way to list backing device UUIDs we have data cached for, and
  58  * probably how long it's been since we've seen them, and a way to invalidate
  59  * dirty data for devices that will never be attached again
  60  *
  61  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  62  * that based on that and how much dirty data we have we can keep writeback
  63  * from being starved
  64  *
  65  * Add a tracepoint or somesuch to watch for writeback starvation
  66  *
  67  * When btree depth > 1 and splitting an interior node, we have to make sure
  68  * alloc_bucket() cannot fail. This should be true but is not completely
  69  * obvious.
  70  *
  71  * Plugging?
  72  *
  73  * If data write is less than hard sector size of ssd, round up offset in open
  74  * bucket to the next whole sector
  75  *
  76  * Superblock needs to be fleshed out for multiple cache devices
  77  *
  78  * Add a sysfs tunable for the number of writeback IOs in flight
  79  *
  80  * Add a sysfs tunable for the number of open data buckets
  81  *
  82  * IO tracking: Can we track when one process is doing io on behalf of another?
  83  * IO tracking: Don't use just an average, weigh more recent stuff higher
  84  *
  85  * Test module load/unload
  86  */
  87
  88 #define MAX_NEED_GC             64
  89 #define MAX_SAVE_PRIO           72
  90
  91 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  92
  93 #define PTR_HASH(c, k)                                                  \
  94         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  95
  96 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
  97
  98 /*
  99  * These macros are for recursing down the btree - they handle the details of
 100  * locking and looking up nodes in the cache for you. They're best treated as
 101  * mere syntax when reading code that uses them.
 102  *
 103  * op->lock determines whether we take a read or a write lock at a given depth.
 104  * If you've got a read lock and find that you need a write lock (i.e. you're
 105  * going to have to split), set op->lock and return -EINTR; btree_root() will
 106  * call you again and you'll have the correct lock.
 107  */
 108
 109 /**
 110  * btree - recurse down the btree on a specified key
 111  * @fn:         function to call, which will be passed the child node
 112  * @key:        key to recurse on
 113  * @b:          parent btree node
 114  * @op:         pointer to struct btree_op
 115  */
 116 #define btree(fn, key, b, op, ...)                                      \
 117 ({                                                                      \
 118         int _r, l = (b)->level - 1;                                     \
 119         bool _w = l <= (op)->lock;                                      \
 120         struct btree *_child = bch_btree_node_get((b)->c, op, key, l,   \
 121                                                   _w, b);               \
 122         if (!IS_ERR(_child)) {                                          \
 123                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 124                 rw_unlock(_w, _child);                                  \
 125         } else                                                          \
 126                 _r = PTR_ERR(_child);                                   \
 127         _r;                                                             \
 128 })
 129
 130 /**
 131  * btree_root - call a function on the root of the btree
 132  * @fn:         function to call, which will be passed the child node
 133  * @c:          cache set
 134  * @op:         pointer to struct btree_op
 135  */
 136 #define btree_root(fn, c, op, ...)                                      \
 137 ({                                                                      \
 138         int _r = -EINTR;                                                \
 139         do {                                                            \
 140                 struct btree *_b = (c)->root;                           \
 141                 bool _w = insert_lock(op, _b);                          \
 142                 rw_lock(_w, _b, _b->level);                             \
 143                 if (_b == (c)->root &&                                  \
 144                     _w == insert_lock(op, _b)) {                        \
 145                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 146                 }                                                       \
 147                 rw_unlock(_w, _b);                                      \
 148                 bch_cannibalize_unlock(c);                              \
 149                 if (_r == -EINTR)                                       \
 150                         schedule();                                     \
 151         } while (_r == -EINTR);                                         \
 152                                                                         \
 153         finish_wait(&(c)->btree_cache_wait, &(op)->wait);               \
 154         _r;                                                             \
 155 })
 156
 157 static inline struct bset *write_block(struct btree *b)
 158 {
 159         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 160 }
 161
 162 static void bch_btree_init_next(struct btree *b)
 163 {
 164         /* If not a leaf node, always sort */
 165         if (b->level && b->keys.nsets)
 166                 bch_btree_sort(&b->keys, &b->c->sort);
 167         else
 168                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 169
 170         if (b->written < btree_blocks(b))
 171                 bch_bset_init_next(&b->keys, write_block(b),
 172                                    bset_magic(&b->c->sb));
 173
 174 }
 175
 176 /* Btree key manipulation */
 177
 178 void bkey_put(struct cache_set *c, struct bkey *k)
 179 {
 180         unsigned i;
 181
 182         for (i = 0; i < KEY_PTRS(k); i++)
 183                 if (ptr_available(c, k, i))
 184                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 185 }
 186
 187 /* Btree IO */
 188
 189 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 190 {
 191         uint64_t crc = b->key.ptr[0];
 192         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 193
 194         crc = bch_crc64_update(crc, data, end - data);
 195         return crc ^ 0xffffffffffffffffULL;
 196 }
 197
 198 void bch_btree_node_read_done(struct btree *b)
 199 {
 200         const char *err = "bad btree header";
 201         struct bset *i = btree_bset_first(b);
 202         struct btree_iter *iter;
 203
 204         iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
 205         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 206         iter->used = 0;
 207
 208 #ifdef CONFIG_BCACHE_DEBUG
 209         iter->b = &b->keys;
 210 #endif
 211
 212         if (!i->seq)
 213                 goto err;
 214
 215         for (;
 216              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 217              i = write_block(b)) {
 218                 err = "unsupported bset version";
 219                 if (i->version > BCACHE_BSET_VERSION)
 220                         goto err;
 221
 222                 err = "bad btree header";
 223                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 224                     btree_blocks(b))
 225                         goto err;
 226
 227                 err = "bad magic";
 228                 if (i->magic != bset_magic(&b->c->sb))
 229                         goto err;
 230
 231                 err = "bad checksum";
 232                 switch (i->version) {
 233                 case 0:
 234                         if (i->csum != csum_set(i))
 235                                 goto err;
 236                         break;
 237                 case BCACHE_BSET_VERSION:
 238                         if (i->csum != btree_csum_set(b, i))
 239                                 goto err;
 240                         break;
 241                 }
 242
 243                 err = "empty set";
 244                 if (i != b->keys.set[0].data && !i->keys)
 245                         goto err;
 246
 247                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 248
 249                 b->written += set_blocks(i, block_bytes(b->c));
 250         }
 251
 252         err = "corrupted btree";
 253         for (i = write_block(b);
 254              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 255              i = ((void *) i) + block_bytes(b->c))
 256                 if (i->seq == b->keys.set[0].data->seq)
 257                         goto err;
 258
 259         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 260
 261         i = b->keys.set[0].data;
 262         err = "short btree key";
 263         if (b->keys.set[0].size &&
 264             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 265                 goto err;
 266
 267         if (b->written < btree_blocks(b))
 268                 bch_bset_init_next(&b->keys, write_block(b),
 269                                    bset_magic(&b->c->sb));
 270 out:
 271         mempool_free(iter, b->c->fill_iter);
 272         return;
 273 err:
 274         set_btree_node_io_error(b);
 275         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 276                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 277                             bset_block_offset(b, i), i->keys);
 278         goto out;
 279 }
 280
 281 static void btree_node_read_endio(struct bio *bio)
 282 {
 283         struct closure *cl = bio->bi_private;
 284         closure_put(cl);
 285 }
 286
 287 static void bch_btree_node_read(struct btree *b)
 288 {
 289         uint64_t start_time = local_clock();
 290         struct closure cl;
 291         struct bio *bio;
 292
 293         trace_bcache_btree_read(b);
 294
 295         closure_init_stack(&cl);
 296
 297         bio = bch_bbio_alloc(b->c);
 298         bio->bi_rw      = REQ_META|READ_SYNC;
 299         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 300         bio->bi_end_io  = btree_node_read_endio;
 301         bio->bi_private = &cl;
 302
 303         bch_bio_map(bio, b->keys.set[0].data);
 304
 305         bch_submit_bbio(bio, b->c, &b->key, 0);
 306         closure_sync(&cl);
 307
 308         if (bio->bi_error)
 309                 set_btree_node_io_error(b);
 310
 311         bch_bbio_free(bio, b->c);
 312
 313         if (btree_node_io_error(b))
 314                 goto err;
 315
 316         bch_btree_node_read_done(b);
 317         bch_time_stats_update(&b->c->btree_read_time, start_time);
 318
 319         return;
 320 err:
 321         bch_cache_set_error(b->c, "io error reading bucket %zu",
 322                             PTR_BUCKET_NR(b->c, &b->key, 0));
 323 }
 324
 325 static void btree_complete_write(struct btree *b, struct btree_write *w)
 326 {
 327         if (w->prio_blocked &&
 328             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 329                 wake_up_allocators(b->c);
 330
 331         if (w->journal) {
 332                 atomic_dec_bug(w->journal);
 333                 __closure_wake_up(&b->c->journal.wait);
 334         }
 335
 336         w->prio_blocked = 0;
 337         w->journal      = NULL;
 338 }
 339
 340 static void btree_node_write_unlock(struct closure *cl)
 341 {
 342         struct btree *b = container_of(cl, struct btree, io);
 343
 344         up(&b->io_mutex);
 345 }
 346
 347 static void __btree_node_write_done(struct closure *cl)
 348 {
 349         struct btree *b = container_of(cl, struct btree, io);
 350         struct btree_write *w = btree_prev_write(b);
 351
 352         bch_bbio_free(b->bio, b->c);
 353         b->bio = NULL;
 354         btree_complete_write(b, w);
 355
 356         if (btree_node_dirty(b))
 357                 schedule_delayed_work(&b->work, 30 * HZ);
 358
 359         closure_return_with_destructor(cl, btree_node_write_unlock);
 360 }
 361
 362 static void btree_node_write_done(struct closure *cl)
 363 {
 364         struct btree *b = container_of(cl, struct btree, io);
 365         struct bio_vec *bv;
 366         int n;
 367
 368         bio_for_each_segment_all(bv, b->bio, n)
 369                 __free_page(bv->bv_page);
 370
 371         __btree_node_write_done(cl);
 372 }
 373
 374 static void btree_node_write_endio(struct bio *bio)
 375 {
 376         struct closure *cl = bio->bi_private;
 377         struct btree *b = container_of(cl, struct btree, io);
 378
 379         if (bio->bi_error)
 380                 set_btree_node_io_error(b);
 381
 382         bch_bbio_count_io_errors(b->c, bio, bio->bi_error, "writing btree");
 383         closure_put(cl);
 384 }
 385
 386 static void do_btree_node_write(struct btree *b)
 387 {
 388         struct closure *cl = &b->io;
 389         struct bset *i = btree_bset_last(b);
 390         BKEY_PADDED(key) k;
 391
 392         i->version      = BCACHE_BSET_VERSION;
 393         i->csum         = btree_csum_set(b, i);
 394
 395         BUG_ON(b->bio);
 396         b->bio = bch_bbio_alloc(b->c);
 397
 398         b->bio->bi_end_io       = btree_node_write_endio;
 399         b->bio->bi_private      = cl;
 400         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 401         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 402         bch_bio_map(b->bio, i);
 403
 404         /*
 405          * If we're appending to a leaf node, we don't technically need FUA -
 406          * this write just needs to be persisted before the next journal write,
 407          * which will be marked FLUSH|FUA.
 408          *
 409          * Similarly if we're writing a new btree root - the pointer is going to
 410          * be in the next journal entry.
 411          *
 412          * But if we're writing a new btree node (that isn't a root) or
 413          * appending to a non leaf btree node, we need either FUA or a flush
 414          * when we write the parent with the new pointer. FUA is cheaper than a
 415          * flush, and writes appending to leaf nodes aren't blocking anything so
 416          * just make all btree node writes FUA to keep things sane.
 417          */
 418
 419         bkey_copy(&k.key, &b->key);
 420         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 421                        bset_sector_offset(&b->keys, i));
 422
 423         if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 424                 int j;
 425                 struct bio_vec *bv;
 426                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 427
 428                 bio_for_each_segment_all(bv, b->bio, j)
 429                         memcpy(page_address(bv->bv_page),
 430                                base + j * PAGE_SIZE, PAGE_SIZE);
 431
 432                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 433
 434                 continue_at(cl, btree_node_write_done, NULL);
 435         } else {
 436                 b->bio->bi_vcnt = 0;
 437                 bch_bio_map(b->bio, i);
 438
 439                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 440
 441                 closure_sync(cl);
 442                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 443         }
 444 }
 445
 446 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 447 {
 448         struct bset *i = btree_bset_last(b);
 449
 450         lockdep_assert_held(&b->write_lock);
 451
 452         trace_bcache_btree_write(b);
 453
 454         BUG_ON(current->bio_list);
 455         BUG_ON(b->written >= btree_blocks(b));
 456         BUG_ON(b->written && !i->keys);
 457         BUG_ON(btree_bset_first(b)->seq != i->seq);
 458         bch_check_keys(&b->keys, "writing");
 459
 460         cancel_delayed_work(&b->work);
 461
 462         /* If caller isn't waiting for write, parent refcount is cache set */
 463         down(&b->io_mutex);
 464         closure_init(&b->io, parent ?: &b->c->cl);
 465
 466         clear_bit(BTREE_NODE_dirty,      &b->flags);
 467         change_bit(BTREE_NODE_write_idx, &b->flags);
 468
 469         do_btree_node_write(b);
 470
 471         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 472                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 473
 474         b->written += set_blocks(i, block_bytes(b->c));
 475 }
 476
 477 void bch_btree_node_write(struct btree *b, struct closure *parent)
 478 {
 479         unsigned nsets = b->keys.nsets;
 480
 481         lockdep_assert_held(&b->lock);
 482
 483         __bch_btree_node_write(b, parent);
 484
 485         /*
 486          * do verify if there was more than one set initially (i.e. we did a
 487          * sort) and we sorted down to a single set:
 488          */
 489         if (nsets && !b->keys.nsets)
 490                 bch_btree_verify(b);
 491
 492         bch_btree_init_next(b);
 493 }
 494
 495 static void bch_btree_node_write_sync(struct btree *b)
 496 {
 497         struct closure cl;
 498
 499         closure_init_stack(&cl);
 500
 501         mutex_lock(&b->write_lock);
 502         bch_btree_node_write(b, &cl);
 503         mutex_unlock(&b->write_lock);
 504
 505         closure_sync(&cl);
 506 }
 507
 508 static void btree_node_write_work(struct work_struct *w)
 509 {
 510         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 511
 512         mutex_lock(&b->write_lock);
 513         if (btree_node_dirty(b))
 514                 __bch_btree_node_write(b, NULL);
 515         mutex_unlock(&b->write_lock);
 516 }
 517
 518 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 519 {
 520         struct bset *i = btree_bset_last(b);
 521         struct btree_write *w = btree_current_write(b);
 522
 523         lockdep_assert_held(&b->write_lock);
 524
 525         BUG_ON(!b->written);
 526         BUG_ON(!i->keys);
 527
 528         if (!btree_node_dirty(b))
 529                 schedule_delayed_work(&b->work, 30 * HZ);
 530
 531         set_btree_node_dirty(b);
 532
 533         if (journal_ref) {
 534                 if (w->journal &&
 535                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 536                         atomic_dec_bug(w->journal);
 537                         w->journal = NULL;
 538                 }
 539
 540                 if (!w->journal) {
 541                         w->journal = journal_ref;
 542                         atomic_inc(w->journal);
 543                 }
 544         }
 545
 546         /* Force write if set is too big */
 547         if (set_bytes(i) > PAGE_SIZE - 48 &&
 548             !current->bio_list)
 549                 bch_btree_node_write(b, NULL);
 550 }
 551
 552 /*
 553  * Btree in memory cache - allocation/freeing
 554  * mca -> memory cache
 555  */
 556
 557 #define mca_reserve(c)  (((c->root && c->root->level)           \
 558                           ? c->root->level : 1) * 8 + 16)
 559 #define mca_can_free(c)                                         \
 560         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 561
 562 static void mca_data_free(struct btree *b)
 563 {
 564         BUG_ON(b->io_mutex.count != 1);
 565
 566         bch_btree_keys_free(&b->keys);
 567
 568         b->c->btree_cache_used--;
 569         list_move(&b->list, &b->c->btree_cache_freed);
 570 }
 571
 572 static void mca_bucket_free(struct btree *b)
 573 {
 574         BUG_ON(btree_node_dirty(b));
 575
 576         b->key.ptr[0] = 0;
 577         hlist_del_init_rcu(&b->hash);
 578         list_move(&b->list, &b->c->btree_cache_freeable);
 579 }
 580
 581 static unsigned btree_order(struct bkey *k)
 582 {
 583         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 584 }
 585
 586 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 587 {
 588         if (!bch_btree_keys_alloc(&b->keys,
 589                                   max_t(unsigned,
 590                                         ilog2(b->c->btree_pages),
 591                                         btree_order(k)),
 592                                   gfp)) {
 593                 b->c->btree_cache_used++;
 594                 list_move(&b->list, &b->c->btree_cache);
 595         } else {
 596                 list_move(&b->list, &b->c->btree_cache_freed);
 597         }
 598 }
 599
 600 static struct btree *mca_bucket_alloc(struct cache_set *c,
 601                                       struct bkey *k, gfp_t gfp)
 602 {
 603         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 604         if (!b)
 605                 return NULL;
 606
 607         init_rwsem(&b->lock);
 608         lockdep_set_novalidate_class(&b->lock);
 609         mutex_init(&b->write_lock);
 610         lockdep_set_novalidate_class(&b->write_lock);
 611         INIT_LIST_HEAD(&b->list);
 612         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 613         b->c = c;
 614         sema_init(&b->io_mutex, 1);
 615
 616         mca_data_alloc(b, k, gfp);
 617         return b;
 618 }
 619
 620 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 621 {
 622         struct closure cl;
 623
 624         closure_init_stack(&cl);
 625         lockdep_assert_held(&b->c->bucket_lock);
 626
 627         if (!down_write_trylock(&b->lock))
 628                 return -ENOMEM;
 629
 630         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 631
 632         if (b->keys.page_order < min_order)
 633                 goto out_unlock;
 634
 635         if (!flush) {
 636                 if (btree_node_dirty(b))
 637                         goto out_unlock;
 638
 639                 if (down_trylock(&b->io_mutex))
 640                         goto out_unlock;
 641                 up(&b->io_mutex);
 642         }
 643
 644         mutex_lock(&b->write_lock);
 645         if (btree_node_dirty(b))
 646                 __bch_btree_node_write(b, &cl);
 647         mutex_unlock(&b->write_lock);
 648
 649         closure_sync(&cl);
 650
 651         /* wait for any in flight btree write */
 652         down(&b->io_mutex);
 653         up(&b->io_mutex);
 654
 655         return 0;
 656 out_unlock:
 657         rw_unlock(true, b);
 658         return -ENOMEM;
 659 }
 660
 661 static unsigned long bch_mca_scan(struct shrinker *shrink,
 662                                   struct shrink_control *sc)
 663 {
 664         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 665         struct btree *b, *t;
 666         unsigned long i, nr = sc->nr_to_scan;
 667         unsigned long freed = 0;
 668
 669         if (c->shrinker_disabled)
 670                 return SHRINK_STOP;
 671
 672         if (c->btree_cache_alloc_lock)
 673                 return SHRINK_STOP;
 674
 675         /* Return -1 if we can't do anything right now */
 676         if (sc->gfp_mask & __GFP_IO)
 677                 mutex_lock(&c->bucket_lock);
 678         else if (!mutex_trylock(&c->bucket_lock))
 679                 return -1;
 680
 681         /*
 682          * It's _really_ critical that we don't free too many btree nodes - we
 683          * have to always leave ourselves a reserve. The reserve is how we
 684          * guarantee that allocating memory for a new btree node can always
 685          * succeed, so that inserting keys into the btree can always succeed and
 686          * IO can always make forward progress:
 687          */
 688         nr /= c->btree_pages;
 689         if (nr == 0)
 690                 nr = 1;
 691         nr = min_t(unsigned long, nr, mca_can_free(c));
 692
 693         i = 0;
 694         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 695                 if (freed >= nr)
 696                         break;
 697
 698                 if (++i > 3 &&
 699                     !mca_reap(b, 0, false)) {
 700                         mca_data_free(b);
 701                         rw_unlock(true, b);
 702                         freed++;
 703                 }
 704         }
 705
 706         for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
 707                 if (list_empty(&c->btree_cache))
 708                         goto out;
 709
 710                 b = list_first_entry(&c->btree_cache, struct btree, list);
 711                 list_rotate_left(&c->btree_cache);
 712
 713                 if (!b->accessed &&
 714                     !mca_reap(b, 0, false)) {
 715                         mca_bucket_free(b);
 716                         mca_data_free(b);
 717                         rw_unlock(true, b);
 718                         freed++;
 719                 } else
 720                         b->accessed = 0;
 721         }
 722 out:
 723         mutex_unlock(&c->bucket_lock);
 724         return freed;
 725 }
 726
 727 static unsigned long bch_mca_count(struct shrinker *shrink,
 728                                    struct shrink_control *sc)
 729 {
 730         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 731
 732         if (c->shrinker_disabled)
 733                 return 0;
 734
 735         if (c->btree_cache_alloc_lock)
 736                 return 0;
 737
 738         return mca_can_free(c) * c->btree_pages;
 739 }
 740
 741 void bch_btree_cache_free(struct cache_set *c)
 742 {
 743         struct btree *b;
 744         struct closure cl;
 745         closure_init_stack(&cl);
 746
 747         if (c->shrink.list.next)
 748                 unregister_shrinker(&c->shrink);
 749
 750         mutex_lock(&c->bucket_lock);
 751
 752 #ifdef CONFIG_BCACHE_DEBUG
 753         if (c->verify_data)
 754                 list_move(&c->verify_data->list, &c->btree_cache);
 755
 756         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 757 #endif
 758
 759         list_splice(&c->btree_cache_freeable,
 760                     &c->btree_cache);
 761
 762         while (!list_empty(&c->btree_cache)) {
 763                 b = list_first_entry(&c->btree_cache, struct btree, list);
 764
 765                 if (btree_node_dirty(b))
 766                         btree_complete_write(b, btree_current_write(b));
 767                 clear_bit(BTREE_NODE_dirty, &b->flags);
 768
 769                 mca_data_free(b);
 770         }
 771
 772         while (!list_empty(&c->btree_cache_freed)) {
 773                 b = list_first_entry(&c->btree_cache_freed,
 774                                      struct btree, list);
 775                 list_del(&b->list);
 776                 cancel_delayed_work_sync(&b->work);
 777                 kfree(b);
 778         }
 779
 780         mutex_unlock(&c->bucket_lock);
 781 }
 782
 783 int bch_btree_cache_alloc(struct cache_set *c)
 784 {
 785         unsigned i;
 786
 787         for (i = 0; i < mca_reserve(c); i++)
 788                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 789                         return -ENOMEM;
 790
 791         list_splice_init(&c->btree_cache,
 792                          &c->btree_cache_freeable);
 793
 794 #ifdef CONFIG_BCACHE_DEBUG
 795         mutex_init(&c->verify_lock);
 796
 797         c->verify_ondisk = (void *)
 798                 __get_free_pages(GFP_KERNEL|__GFP_COMP, ilog2(bucket_pages(c)));
 799
 800         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 801
 802         if (c->verify_data &&
 803             c->verify_data->keys.set->data)
 804                 list_del_init(&c->verify_data->list);
 805         else
 806                 c->verify_data = NULL;
 807 #endif
 808
 809         c->shrink.count_objects = bch_mca_count;
 810         c->shrink.scan_objects = bch_mca_scan;
 811         c->shrink.seeks = 4;
 812         c->shrink.batch = c->btree_pages * 2;
 813
 814         if (register_shrinker(&c->shrink))
 815                 pr_warn("bcache: %s: could not register shrinker",
 816                                 __func__);
 817
 818         return 0;
 819 }
 820
 821 /* Btree in memory cache - hash table */
 822
 823 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 824 {
 825         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 826 }
 827
 828 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 829 {
 830         struct btree *b;
 831
 832         rcu_read_lock();
 833         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 834                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 835                         goto out;
 836         b = NULL;
 837 out:
 838         rcu_read_unlock();
 839         return b;
 840 }
 841
 842 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 843 {
 844         spin_lock(&c->btree_cannibalize_lock);
 845         if (likely(c->btree_cache_alloc_lock == NULL)) {
 846                 c->btree_cache_alloc_lock = current;
 847         } else if (c->btree_cache_alloc_lock != current) {
 848                 if (op)
 849                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 850                                         TASK_UNINTERRUPTIBLE);
 851                 spin_unlock(&c->btree_cannibalize_lock);
 852                 return -EINTR;
 853         }
 854         spin_unlock(&c->btree_cannibalize_lock);
 855
 856         return 0;
 857 }
 858
 859 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 860                                      struct bkey *k)
 861 {
 862         struct btree *b;
 863
 864         trace_bcache_btree_cache_cannibalize(c);
 865
 866         if (mca_cannibalize_lock(c, op))
 867                 return ERR_PTR(-EINTR);
 868
 869         list_for_each_entry_reverse(b, &c->btree_cache, list)
 870                 if (!mca_reap(b, btree_order(k), false))
 871                         return b;
 872
 873         list_for_each_entry_reverse(b, &c->btree_cache, list)
 874                 if (!mca_reap(b, btree_order(k), true))
 875                         return b;
 876
 877         WARN(1, "btree cache cannibalize failed\n");
 878         return ERR_PTR(-ENOMEM);
 879 }
 880
 881 /*
 882  * We can only have one thread cannibalizing other cached btree nodes at a time,
 883  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 884  * cannibalize_bucket() will take. This means every time we unlock the root of
 885  * the btree, we need to release this lock if we have it held.
 886  */
 887 static void bch_cannibalize_unlock(struct cache_set *c)
 888 {
 889         spin_lock(&c->btree_cannibalize_lock);
 890         if (c->btree_cache_alloc_lock == current) {
 891                 c->btree_cache_alloc_lock = NULL;
 892                 wake_up(&c->btree_cache_wait);
 893         }
 894         spin_unlock(&c->btree_cannibalize_lock);
 895 }
 896
 897 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 898                                struct bkey *k, int level)
 899 {
 900         struct btree *b;
 901
 902         BUG_ON(current->bio_list);
 903
 904         lockdep_assert_held(&c->bucket_lock);
 905
 906         if (mca_find(c, k))
 907                 return NULL;
 908
 909         /* btree_free() doesn't free memory; it sticks the node on the end of
 910          * the list. Check if there's any freed nodes there:
 911          */
 912         list_for_each_entry(b, &c->btree_cache_freeable, list)
 913                 if (!mca_reap(b, btree_order(k), false))
 914                         goto out;
 915
 916         /* We never free struct btree itself, just the memory that holds the on
 917          * disk node. Check the freed list before allocating a new one:
 918          */
 919         list_for_each_entry(b, &c->btree_cache_freed, list)
 920                 if (!mca_reap(b, 0, false)) {
 921                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 922                         if (!b->keys.set[0].data)
 923                                 goto err;
 924                         else
 925                                 goto out;
 926                 }
 927
 928         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 929         if (!b)
 930                 goto err;
 931
 932         BUG_ON(!down_write_trylock(&b->lock));
 933         if (!b->keys.set->data)
 934                 goto err;
 935 out:
 936         BUG_ON(b->io_mutex.count != 1);
 937
 938         bkey_copy(&b->key, k);
 939         list_move(&b->list, &c->btree_cache);
 940         hlist_del_init_rcu(&b->hash);
 941         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 942
 943         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 944         b->parent       = (void *) ~0UL;
 945         b->flags        = 0;
 946         b->written      = 0;
 947         b->level        = level;
 948
 949         if (!b->level)
 950                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 951                                     &b->c->expensive_debug_checks);
 952         else
 953                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 954                                     &b->c->expensive_debug_checks);
 955
 956         return b;
 957 err:
 958         if (b)
 959                 rw_unlock(true, b);
 960
 961         b = mca_cannibalize(c, op, k);
 962         if (!IS_ERR(b))
 963                 goto out;
 964
 965         return b;
 966 }
 967
 968 /**
 969  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 970  * in from disk if necessary.
 971  *
 972  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 973  *
 974  * The btree node will have either a read or a write lock held, depending on
 975  * level and op->lock.
 976  */
 977 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 978                                  struct bkey *k, int level, bool write,
 979                                  struct btree *parent)
 980 {
 981         int i = 0;
 982         struct btree *b;
 983
 984         BUG_ON(level < 0);
 985 retry:
 986         b = mca_find(c, k);
 987
 988         if (!b) {
 989                 if (current->bio_list)
 990                         return ERR_PTR(-EAGAIN);
 991
 992                 mutex_lock(&c->bucket_lock);
 993                 b = mca_alloc(c, op, k, level);
 994                 mutex_unlock(&c->bucket_lock);
 995
 996                 if (!b)
 997                         goto retry;
 998                 if (IS_ERR(b))
 999                         return b;
1000
1001                 bch_btree_node_read(b);
1002
1003                 if (!write)
1004                         downgrade_write(&b->lock);
1005         } else {
1006                 rw_lock(write, b, level);
1007                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1008                         rw_unlock(write, b);
1009                         goto retry;
1010                 }
1011                 BUG_ON(b->level != level);
1012         }
1013
1014         b->parent = parent;
1015         b->accessed = 1;
1016
1017         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1018                 prefetch(b->keys.set[i].tree);
1019                 prefetch(b->keys.set[i].data);
1020         }
1021
1022         for (; i <= b->keys.nsets; i++)
1023                 prefetch(b->keys.set[i].data);
1024
1025         if (btree_node_io_error(b)) {
1026                 rw_unlock(write, b);
1027                 return ERR_PTR(-EIO);
1028         }
1029
1030         BUG_ON(!b->written);
1031
1032         return b;
1033 }
1034
1035 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1036 {
1037         struct btree *b;
1038
1039         mutex_lock(&parent->c->bucket_lock);
1040         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1041         mutex_unlock(&parent->c->bucket_lock);
1042
1043         if (!IS_ERR_OR_NULL(b)) {
1044                 b->parent = parent;
1045                 bch_btree_node_read(b);
1046                 rw_unlock(true, b);
1047         }
1048 }
1049
1050 /* Btree alloc */
1051
1052 static void btree_node_free(struct btree *b)
1053 {
1054         trace_bcache_btree_node_free(b);
1055
1056         BUG_ON(b == b->c->root);
1057
1058         mutex_lock(&b->write_lock);
1059
1060         if (btree_node_dirty(b))
1061                 btree_complete_write(b, btree_current_write(b));
1062         clear_bit(BTREE_NODE_dirty, &b->flags);
1063
1064         mutex_unlock(&b->write_lock);
1065
1066         cancel_delayed_work(&b->work);
1067
1068         mutex_lock(&b->c->bucket_lock);
1069         bch_bucket_free(b->c, &b->key);
1070         mca_bucket_free(b);
1071         mutex_unlock(&b->c->bucket_lock);
1072 }
1073
1074 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1075                                      int level, bool wait,
1076                                      struct btree *parent)
1077 {
1078         BKEY_PADDED(key) k;
1079         struct btree *b = ERR_PTR(-EAGAIN);
1080
1081         mutex_lock(&c->bucket_lock);
1082 retry:
1083         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1084                 goto err;
1085
1086         bkey_put(c, &k.key);
1087         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1088
1089         b = mca_alloc(c, op, &k.key, level);
1090         if (IS_ERR(b))
1091                 goto err_free;
1092
1093         if (!b) {
1094                 cache_bug(c,
1095                         "Tried to allocate bucket that was in btree cache");
1096                 goto retry;
1097         }
1098
1099         b->accessed = 1;
1100         b->parent = parent;
1101         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1102
1103         mutex_unlock(&c->bucket_lock);
1104
1105         trace_bcache_btree_node_alloc(b);
1106         return b;
1107 err_free:
1108         bch_bucket_free(c, &k.key);
1109 err:
1110         mutex_unlock(&c->bucket_lock);
1111
1112         trace_bcache_btree_node_alloc_fail(c);
1113         return b;
1114 }
1115
1116 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1117                                           struct btree_op *op, int level,
1118                                           struct btree *parent)
1119 {
1120         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1121 }
1122
1123 static struct btree *btree_node_alloc_replacement(struct btree *b,
1124                                                   struct btree_op *op)
1125 {
1126         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1127         if (!IS_ERR_OR_NULL(n)) {
1128                 mutex_lock(&n->write_lock);
1129                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1130                 bkey_copy_key(&n->key, &b->key);
1131                 mutex_unlock(&n->write_lock);
1132         }
1133
1134         return n;
1135 }
1136
1137 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1138 {
1139         unsigned i;
1140
1141         mutex_lock(&b->c->bucket_lock);
1142
1143         atomic_inc(&b->c->prio_blocked);
1144
1145         bkey_copy(k, &b->key);
1146         bkey_copy_key(k, &ZERO_KEY);
1147
1148         for (i = 0; i < KEY_PTRS(k); i++)
1149                 SET_PTR_GEN(k, i,
1150                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1151                                         PTR_BUCKET(b->c, &b->key, i)));
1152
1153         mutex_unlock(&b->c->bucket_lock);
1154 }
1155
1156 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1157 {
1158         struct cache_set *c = b->c;
1159         struct cache *ca;
1160         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1161
1162         mutex_lock(&c->bucket_lock);
1163
1164         for_each_cache(ca, c, i)
1165                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1166                         if (op)
1167                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1168                                                 TASK_UNINTERRUPTIBLE);
1169                         mutex_unlock(&c->bucket_lock);
1170                         return -EINTR;
1171                 }
1172
1173         mutex_unlock(&c->bucket_lock);
1174
1175         return mca_cannibalize_lock(b->c, op);
1176 }
1177
1178 /* Garbage collection */
1179
1180 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1181                                     struct bkey *k)
1182 {
1183         uint8_t stale = 0;
1184         unsigned i;
1185         struct bucket *g;
1186
1187         /*
1188          * ptr_invalid() can't return true for the keys that mark btree nodes as
1189          * freed, but since ptr_bad() returns true we'll never actually use them
1190          * for anything and thus we don't want mark their pointers here
1191          */
1192         if (!bkey_cmp(k, &ZERO_KEY))
1193                 return stale;
1194
1195         for (i = 0; i < KEY_PTRS(k); i++) {
1196                 if (!ptr_available(c, k, i))
1197                         continue;
1198
1199                 g = PTR_BUCKET(c, k, i);
1200
1201                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1202                         g->last_gc = PTR_GEN(k, i);
1203
1204                 if (ptr_stale(c, k, i)) {
1205                         stale = max(stale, ptr_stale(c, k, i));
1206                         continue;
1207                 }
1208
1209                 cache_bug_on(GC_MARK(g) &&
1210                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1211                              c, "inconsistent ptrs: mark = %llu, level = %i",
1212                              GC_MARK(g), level);
1213
1214                 if (level)
1215                         SET_GC_MARK(g, GC_MARK_METADATA);
1216                 else if (KEY_DIRTY(k))
1217                         SET_GC_MARK(g, GC_MARK_DIRTY);
1218                 else if (!GC_MARK(g))
1219                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1220
1221                 /* guard against overflow */
1222                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1223                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1224                                              MAX_GC_SECTORS_USED));
1225
1226                 BUG_ON(!GC_SECTORS_USED(g));
1227         }
1228
1229         return stale;
1230 }
1231
1232 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1233
1234 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1235 {
1236         unsigned i;
1237
1238         for (i = 0; i < KEY_PTRS(k); i++)
1239                 if (ptr_available(c, k, i) &&
1240                     !ptr_stale(c, k, i)) {
1241                         struct bucket *b = PTR_BUCKET(c, k, i);
1242
1243                         b->gen = PTR_GEN(k, i);
1244
1245                         if (level && bkey_cmp(k, &ZERO_KEY))
1246                                 b->prio = BTREE_PRIO;
1247                         else if (!level && b->prio == BTREE_PRIO)
1248                                 b->prio = INITIAL_PRIO;
1249                 }
1250
1251         __bch_btree_mark_key(c, level, k);
1252 }
1253
1254 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1255 {
1256         uint8_t stale = 0;
1257         unsigned keys = 0, good_keys = 0;
1258         struct bkey *k;
1259         struct btree_iter iter;
1260         struct bset_tree *t;
1261
1262         gc->nodes++;
1263
1264         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1265                 stale = max(stale, btree_mark_key(b, k));
1266                 keys++;
1267
1268                 if (bch_ptr_bad(&b->keys, k))
1269                         continue;
1270
1271                 gc->key_bytes += bkey_u64s(k);
1272                 gc->nkeys++;
1273                 good_keys++;
1274
1275                 gc->data += KEY_SIZE(k);
1276         }
1277
1278         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1279                 btree_bug_on(t->size &&
1280                              bset_written(&b->keys, t) &&
1281                              bkey_cmp(&b->key, &t->end) < 0,
1282                              b, "found short btree key in gc");
1283
1284         if (b->c->gc_always_rewrite)
1285                 return true;
1286
1287         if (stale > 10)
1288                 return true;
1289
1290         if ((keys - good_keys) * 2 > keys)
1291                 return true;
1292
1293         return false;
1294 }
1295
1296 #define GC_MERGE_NODES  4U
1297
1298 struct gc_merge_info {
1299         struct btree    *b;
1300         unsigned        keys;
1301 };
1302
1303 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1304                                  struct keylist *, atomic_t *, struct bkey *);
1305
1306 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1307                              struct gc_stat *gc, struct gc_merge_info *r)
1308 {
1309         unsigned i, nodes = 0, keys = 0, blocks;
1310         struct btree *new_nodes[GC_MERGE_NODES];
1311         struct keylist keylist;
1312         struct closure cl;
1313         struct bkey *k;
1314
1315         bch_keylist_init(&keylist);
1316
1317         if (btree_check_reserve(b, NULL))
1318                 return 0;
1319
1320         memset(new_nodes, 0, sizeof(new_nodes));
1321         closure_init_stack(&cl);
1322
1323         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1324                 keys += r[nodes++].keys;
1325
1326         blocks = btree_default_blocks(b->c) * 2 / 3;
1327
1328         if (nodes < 2 ||
1329             __set_blocks(b->keys.set[0].data, keys,
1330                          block_bytes(b->c)) > blocks * (nodes - 1))
1331                 return 0;
1332
1333         for (i = 0; i < nodes; i++) {
1334                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1335                 if (IS_ERR_OR_NULL(new_nodes[i]))
1336                         goto out_nocoalesce;
1337         }
1338
1339         /*
1340          * We have to check the reserve here, after we've allocated our new
1341          * nodes, to make sure the insert below will succeed - we also check
1342          * before as an optimization to potentially avoid a bunch of expensive
1343          * allocs/sorts
1344          */
1345         if (btree_check_reserve(b, NULL))
1346                 goto out_nocoalesce;
1347
1348         for (i = 0; i < nodes; i++)
1349                 mutex_lock(&new_nodes[i]->write_lock);
1350
1351         for (i = nodes - 1; i > 0; --i) {
1352                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1353                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1354                 struct bkey *k, *last = NULL;
1355
1356                 keys = 0;
1357
1358                 if (i > 1) {
1359                         for (k = n2->start;
1360                              k < bset_bkey_last(n2);
1361                              k = bkey_next(k)) {
1362                                 if (__set_blocks(n1, n1->keys + keys +
1363                                                  bkey_u64s(k),
1364                                                  block_bytes(b->c)) > blocks)
1365                                         break;
1366
1367                                 last = k;
1368                                 keys += bkey_u64s(k);
1369                         }
1370                 } else {
1371                         /*
1372                          * Last node we're not getting rid of - we're getting
1373                          * rid of the node at r[0]. Have to try and fit all of
1374                          * the remaining keys into this node; we can't ensure
1375                          * they will always fit due to rounding and variable
1376                          * length keys (shouldn't be possible in practice,
1377                          * though)
1378                          */
1379                         if (__set_blocks(n1, n1->keys + n2->keys,
1380                                          block_bytes(b->c)) >
1381                             btree_blocks(new_nodes[i]))
1382                                 goto out_unlock_nocoalesce;
1383
1384                         keys = n2->keys;
1385                         /* Take the key of the node we're getting rid of */
1386                         last = &r->b->key;
1387                 }
1388
1389                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1390                        btree_blocks(new_nodes[i]));
1391
1392                 if (last)
1393                         bkey_copy_key(&new_nodes[i]->key, last);
1394
1395                 memcpy(bset_bkey_last(n1),
1396                        n2->start,
1397                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1398
1399                 n1->keys += keys;
1400                 r[i].keys = n1->keys;
1401
1402                 memmove(n2->start,
1403                         bset_bkey_idx(n2, keys),
1404                         (void *) bset_bkey_last(n2) -
1405                         (void *) bset_bkey_idx(n2, keys));
1406
1407                 n2->keys -= keys;
1408
1409                 if (__bch_keylist_realloc(&keylist,
1410                                           bkey_u64s(&new_nodes[i]->key)))
1411                         goto out_unlock_nocoalesce;
1412
1413                 bch_btree_node_write(new_nodes[i], &cl);
1414                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1415         }
1416
1417         for (i = 0; i < nodes; i++)
1418                 mutex_unlock(&new_nodes[i]->write_lock);
1419
1420         closure_sync(&cl);
1421
1422         /* We emptied out this node */
1423         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1424         btree_node_free(new_nodes[0]);
1425         rw_unlock(true, new_nodes[0]);
1426         new_nodes[0] = NULL;
1427
1428         for (i = 0; i < nodes; i++) {
1429                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1430                         goto out_nocoalesce;
1431
1432                 make_btree_freeing_key(r[i].b, keylist.top);
1433                 bch_keylist_push(&keylist);
1434         }
1435
1436         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1437         BUG_ON(!bch_keylist_empty(&keylist));
1438
1439         for (i = 0; i < nodes; i++) {
1440                 btree_node_free(r[i].b);
1441                 rw_unlock(true, r[i].b);
1442
1443                 r[i].b = new_nodes[i];
1444         }
1445
1446         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1447         r[nodes - 1].b = ERR_PTR(-EINTR);
1448
1449         trace_bcache_btree_gc_coalesce(nodes);
1450         gc->nodes--;
1451
1452         bch_keylist_free(&keylist);
1453
1454         /* Invalidated our iterator */
1455         return -EINTR;
1456
1457 out_unlock_nocoalesce:
1458         for (i = 0; i < nodes; i++)
1459                 mutex_unlock(&new_nodes[i]->write_lock);
1460
1461 out_nocoalesce:
1462         closure_sync(&cl);
1463         bch_keylist_free(&keylist);
1464
1465         while ((k = bch_keylist_pop(&keylist)))
1466                 if (!bkey_cmp(k, &ZERO_KEY))
1467                         atomic_dec(&b->c->prio_blocked);
1468
1469         for (i = 0; i < nodes; i++)
1470                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1471                         btree_node_free(new_nodes[i]);
1472                         rw_unlock(true, new_nodes[i]);
1473                 }
1474         return 0;
1475 }
1476
1477 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1478                                  struct btree *replace)
1479 {
1480         struct keylist keys;
1481         struct btree *n;
1482
1483         if (btree_check_reserve(b, NULL))
1484                 return 0;
1485
1486         n = btree_node_alloc_replacement(replace, NULL);
1487
1488         /* recheck reserve after allocating replacement node */
1489         if (btree_check_reserve(b, NULL)) {
1490                 btree_node_free(n);
1491                 rw_unlock(true, n);
1492                 return 0;
1493         }
1494
1495         bch_btree_node_write_sync(n);
1496
1497         bch_keylist_init(&keys);
1498         bch_keylist_add(&keys, &n->key);
1499
1500         make_btree_freeing_key(replace, keys.top);
1501         bch_keylist_push(&keys);
1502
1503         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1504         BUG_ON(!bch_keylist_empty(&keys));
1505
1506         btree_node_free(replace);
1507         rw_unlock(true, n);
1508
1509         /* Invalidated our iterator */
1510         return -EINTR;
1511 }
1512
1513 static unsigned btree_gc_count_keys(struct btree *b)
1514 {
1515         struct bkey *k;
1516         struct btree_iter iter;
1517         unsigned ret = 0;
1518
1519         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1520                 ret += bkey_u64s(k);
1521
1522         return ret;
1523 }
1524
1525 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1526                             struct closure *writes, struct gc_stat *gc)
1527 {
1528         int ret = 0;
1529         bool should_rewrite;
1530         struct bkey *k;
1531         struct btree_iter iter;
1532         struct gc_merge_info r[GC_MERGE_NODES];
1533         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1534
1535         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1536
1537         for (i = r; i < r + ARRAY_SIZE(r); i++)
1538                 i->b = ERR_PTR(-EINTR);
1539
1540         while (1) {
1541                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1542                 if (k) {
1543                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1544                                                   true, b);
1545                         if (IS_ERR(r->b)) {
1546                                 ret = PTR_ERR(r->b);
1547                                 break;
1548                         }
1549
1550                         r->keys = btree_gc_count_keys(r->b);
1551
1552                         ret = btree_gc_coalesce(b, op, gc, r);
1553                         if (ret)
1554                                 break;
1555                 }
1556
1557                 if (!last->b)
1558                         break;
1559
1560                 if (!IS_ERR(last->b)) {
1561                         should_rewrite = btree_gc_mark_node(last->b, gc);
1562                         if (should_rewrite) {
1563                                 ret = btree_gc_rewrite_node(b, op, last->b);
1564                                 if (ret)
1565                                         break;
1566                         }
1567
1568                         if (last->b->level) {
1569                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1570                                 if (ret)
1571                                         break;
1572                         }
1573
1574                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1575
1576                         /*
1577                          * Must flush leaf nodes before gc ends, since replace
1578                          * operations aren't journalled
1579                          */
1580                         mutex_lock(&last->b->write_lock);
1581                         if (btree_node_dirty(last->b))
1582                                 bch_btree_node_write(last->b, writes);
1583                         mutex_unlock(&last->b->write_lock);
1584                         rw_unlock(true, last->b);
1585                 }
1586
1587                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1588                 r->b = NULL;
1589
1590                 if (need_resched()) {
1591                         ret = -EAGAIN;
1592                         break;
1593                 }
1594         }
1595
1596         for (i = r; i < r + ARRAY_SIZE(r); i++)
1597                 if (!IS_ERR_OR_NULL(i->b)) {
1598                         mutex_lock(&i->b->write_lock);
1599                         if (btree_node_dirty(i->b))
1600                                 bch_btree_node_write(i->b, writes);
1601                         mutex_unlock(&i->b->write_lock);
1602                         rw_unlock(true, i->b);
1603                 }
1604
1605         return ret;
1606 }
1607
1608 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1609                              struct closure *writes, struct gc_stat *gc)
1610 {
1611         struct btree *n = NULL;
1612         int ret = 0;
1613         bool should_rewrite;
1614
1615         should_rewrite = btree_gc_mark_node(b, gc);
1616         if (should_rewrite) {
1617                 n = btree_node_alloc_replacement(b, NULL);
1618
1619                 if (!IS_ERR_OR_NULL(n)) {
1620                         bch_btree_node_write_sync(n);
1621
1622                         bch_btree_set_root(n);
1623                         btree_node_free(b);
1624                         rw_unlock(true, n);
1625
1626                         return -EINTR;
1627                 }
1628         }
1629
1630         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1631
1632         if (b->level) {
1633                 ret = btree_gc_recurse(b, op, writes, gc);
1634                 if (ret)
1635                         return ret;
1636         }
1637
1638         bkey_copy_key(&b->c->gc_done, &b->key);
1639
1640         return ret;
1641 }
1642
1643 static void btree_gc_start(struct cache_set *c)
1644 {
1645         struct cache *ca;
1646         struct bucket *b;
1647         unsigned i;
1648
1649         if (!c->gc_mark_valid)
1650                 return;
1651
1652         mutex_lock(&c->bucket_lock);
1653
1654         c->gc_mark_valid = 0;
1655         c->gc_done = ZERO_KEY;
1656
1657         for_each_cache(ca, c, i)
1658                 for_each_bucket(b, ca) {
1659                         b->last_gc = b->gen;
1660                         if (!atomic_read(&b->pin)) {
1661                                 SET_GC_MARK(b, 0);
1662                                 SET_GC_SECTORS_USED(b, 0);
1663                         }
1664                 }
1665
1666         mutex_unlock(&c->bucket_lock);
1667 }
1668
1669 static size_t bch_btree_gc_finish(struct cache_set *c)
1670 {
1671         size_t available = 0;
1672         struct bucket *b;
1673         struct cache *ca;
1674         unsigned i;
1675
1676         mutex_lock(&c->bucket_lock);
1677
1678         set_gc_sectors(c);
1679         c->gc_mark_valid = 1;
1680         c->need_gc      = 0;
1681
1682         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1683                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1684                             GC_MARK_METADATA);
1685
1686         /* don't reclaim buckets to which writeback keys point */
1687         rcu_read_lock();
1688         for (i = 0; i < c->nr_uuids; i++) {
1689                 struct bcache_device *d = c->devices[i];
1690                 struct cached_dev *dc;
1691                 struct keybuf_key *w, *n;
1692                 unsigned j;
1693
1694                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1695                         continue;
1696                 dc = container_of(d, struct cached_dev, disk);
1697
1698                 spin_lock(&dc->writeback_keys.lock);
1699                 rbtree_postorder_for_each_entry_safe(w, n,
1700                                         &dc->writeback_keys.keys, node)
1701                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1702                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1703                                             GC_MARK_DIRTY);
1704                 spin_unlock(&dc->writeback_keys.lock);
1705         }
1706         rcu_read_unlock();
1707
1708         for_each_cache(ca, c, i) {
1709                 uint64_t *i;
1710
1711                 ca->invalidate_needs_gc = 0;
1712
1713                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1714                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1715
1716                 for (i = ca->prio_buckets;
1717                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1718                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1719
1720                 for_each_bucket(b, ca) {
1721                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1722
1723                         if (atomic_read(&b->pin))
1724                                 continue;
1725
1726                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1727
1728                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1729                                 available++;
1730                 }
1731         }
1732
1733         mutex_unlock(&c->bucket_lock);
1734         return available;
1735 }
1736
1737 static void bch_btree_gc(struct cache_set *c)
1738 {
1739         int ret;
1740         unsigned long available;
1741         struct gc_stat stats;
1742         struct closure writes;
1743         struct btree_op op;
1744         uint64_t start_time = local_clock();
1745
1746         trace_bcache_gc_start(c);
1747
1748         memset(&stats, 0, sizeof(struct gc_stat));
1749         closure_init_stack(&writes);
1750         bch_btree_op_init(&op, SHRT_MAX);
1751
1752         btree_gc_start(c);
1753
1754         do {
1755                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1756                 closure_sync(&writes);
1757                 cond_resched();
1758
1759                 if (ret && ret != -EAGAIN)
1760                         pr_warn("gc failed!");
1761         } while (ret);
1762
1763         available = bch_btree_gc_finish(c);
1764         wake_up_allocators(c);
1765
1766         bch_time_stats_update(&c->btree_gc_time, start_time);
1767
1768         stats.key_bytes *= sizeof(uint64_t);
1769         stats.data      <<= 9;
1770         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1771         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1772
1773         trace_bcache_gc_end(c);
1774
1775         bch_moving_gc(c);
1776 }
1777
1778 static bool gc_should_run(struct cache_set *c)
1779 {
1780         struct cache *ca;
1781         unsigned i;
1782
1783         for_each_cache(ca, c, i)
1784                 if (ca->invalidate_needs_gc)
1785                         return true;
1786
1787         if (atomic_read(&c->sectors_to_gc) < 0)
1788                 return true;
1789
1790         return false;
1791 }
1792
1793 static int bch_gc_thread(void *arg)
1794 {
1795         struct cache_set *c = arg;
1796
1797         while (1) {
1798                 wait_event_interruptible(c->gc_wait,
1799                            kthread_should_stop() || gc_should_run(c));
1800
1801                 if (kthread_should_stop())
1802                         break;
1803
1804                 set_gc_sectors(c);
1805                 bch_btree_gc(c);
1806         }
1807
1808         return 0;
1809 }
1810
1811 int bch_gc_thread_start(struct cache_set *c)
1812 {
1813         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1814         if (IS_ERR(c->gc_thread))
1815                 return PTR_ERR(c->gc_thread);
1816
1817         return 0;
1818 }
1819
1820 /* Initial partial gc */
1821
1822 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1823 {
1824         int ret = 0;
1825         struct bkey *k, *p = NULL;
1826         struct btree_iter iter;
1827
1828         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1829                 bch_initial_mark_key(b->c, b->level, k);
1830
1831         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1832
1833         if (b->level) {
1834                 bch_btree_iter_init(&b->keys, &iter, NULL);
1835
1836                 do {
1837                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1838                                                        bch_ptr_bad);
1839                         if (k)
1840                                 btree_node_prefetch(b, k);
1841
1842                         if (p)
1843                                 ret = btree(check_recurse, p, b, op);
1844
1845                         p = k;
1846                 } while (p && !ret);
1847         }
1848
1849         return ret;
1850 }
1851
1852 int bch_btree_check(struct cache_set *c)
1853 {
1854         struct btree_op op;
1855
1856         bch_btree_op_init(&op, SHRT_MAX);
1857
1858         return btree_root(check_recurse, c, &op);
1859 }
1860
1861 void bch_initial_gc_finish(struct cache_set *c)
1862 {
1863         struct cache *ca;
1864         struct bucket *b;
1865         unsigned i;
1866
1867         bch_btree_gc_finish(c);
1868
1869         mutex_lock(&c->bucket_lock);
1870
1871         /*
1872          * We need to put some unused buckets directly on the prio freelist in
1873          * order to get the allocator thread started - it needs freed buckets in
1874          * order to rewrite the prios and gens, and it needs to rewrite prios
1875          * and gens in order to free buckets.
1876          *
1877          * This is only safe for buckets that have no live data in them, which
1878          * there should always be some of.
1879          */
1880         for_each_cache(ca, c, i) {
1881                 for_each_bucket(b, ca) {
1882                         if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1883                             fifo_full(&ca->free[RESERVE_BTREE]))
1884                                 break;
1885
1886                         if (bch_can_invalidate_bucket(ca, b) &&
1887                             !GC_MARK(b)) {
1888                                 __bch_invalidate_one_bucket(ca, b);
1889                                 if (!fifo_push(&ca->free[RESERVE_PRIO],
1890                                    b - ca->buckets))
1891                                         fifo_push(&ca->free[RESERVE_BTREE],
1892                                                   b - ca->buckets);
1893                         }
1894                 }
1895         }
1896
1897         mutex_unlock(&c->bucket_lock);
1898 }
1899
1900 /* Btree insertion */
1901
1902 static bool btree_insert_key(struct btree *b, struct bkey *k,
1903                              struct bkey *replace_key)
1904 {
1905         unsigned status;
1906
1907         BUG_ON(bkey_cmp(k, &b->key) > 0);
1908
1909         status = bch_btree_insert_key(&b->keys, k, replace_key);
1910         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1911                 bch_check_keys(&b->keys, "%u for %s", status,
1912                                replace_key ? "replace" : "insert");
1913
1914                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1915                                               status);
1916                 return true;
1917         } else
1918                 return false;
1919 }
1920
1921 static size_t insert_u64s_remaining(struct btree *b)
1922 {
1923         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1924
1925         /*
1926          * Might land in the middle of an existing extent and have to split it
1927          */
1928         if (b->keys.ops->is_extents)
1929                 ret -= KEY_MAX_U64S;
1930
1931         return max(ret, 0L);
1932 }
1933
1934 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1935                                   struct keylist *insert_keys,
1936                                   struct bkey *replace_key)
1937 {
1938         bool ret = false;
1939         int oldsize = bch_count_data(&b->keys);
1940
1941         while (!bch_keylist_empty(insert_keys)) {
1942                 struct bkey *k = insert_keys->keys;
1943
1944                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1945                         break;
1946
1947                 if (bkey_cmp(k, &b->key) <= 0) {
1948                         if (!b->level)
1949                                 bkey_put(b->c, k);
1950
1951                         ret |= btree_insert_key(b, k, replace_key);
1952                         bch_keylist_pop_front(insert_keys);
1953                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1954                         BKEY_PADDED(key) temp;
1955                         bkey_copy(&temp.key, insert_keys->keys);
1956
1957                         bch_cut_back(&b->key, &temp.key);
1958                         bch_cut_front(&b->key, insert_keys->keys);
1959
1960                         ret |= btree_insert_key(b, &temp.key, replace_key);
1961                         break;
1962                 } else {
1963                         break;
1964                 }
1965         }
1966
1967         if (!ret)
1968                 op->insert_collision = true;
1969
1970         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1971
1972         BUG_ON(bch_count_data(&b->keys) < oldsize);
1973         return ret;
1974 }
1975
1976 static int btree_split(struct btree *b, struct btree_op *op,
1977                        struct keylist *insert_keys,
1978                        struct bkey *replace_key)
1979 {
1980         bool split;
1981         struct btree *n1, *n2 = NULL, *n3 = NULL;
1982         uint64_t start_time = local_clock();
1983         struct closure cl;
1984         struct keylist parent_keys;
1985
1986         closure_init_stack(&cl);
1987         bch_keylist_init(&parent_keys);
1988
1989         if (btree_check_reserve(b, op)) {
1990                 if (!b->level)
1991                         return -EINTR;
1992                 else
1993                         WARN(1, "insufficient reserve for split\n");
1994         }
1995
1996         n1 = btree_node_alloc_replacement(b, op);
1997         if (IS_ERR(n1))
1998                 goto err;
1999
2000         split = set_blocks(btree_bset_first(n1),
2001                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2002
2003         if (split) {
2004                 unsigned keys = 0;
2005
2006                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2007
2008                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2009                 if (IS_ERR(n2))
2010                         goto err_free1;
2011
2012                 if (!b->parent) {
2013                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2014                         if (IS_ERR(n3))
2015                                 goto err_free2;
2016                 }
2017
2018                 mutex_lock(&n1->write_lock);
2019                 mutex_lock(&n2->write_lock);
2020
2021                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2022
2023                 /*
2024                  * Has to be a linear search because we don't have an auxiliary
2025                  * search tree yet
2026                  */
2027
2028                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2029                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2030                                                         keys));
2031
2032                 bkey_copy_key(&n1->key,
2033                               bset_bkey_idx(btree_bset_first(n1), keys));
2034                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2035
2036                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2037                 btree_bset_first(n1)->keys = keys;
2038
2039                 memcpy(btree_bset_first(n2)->start,
2040                        bset_bkey_last(btree_bset_first(n1)),
2041                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2042
2043                 bkey_copy_key(&n2->key, &b->key);
2044
2045                 bch_keylist_add(&parent_keys, &n2->key);
2046                 bch_btree_node_write(n2, &cl);
2047                 mutex_unlock(&n2->write_lock);
2048                 rw_unlock(true, n2);
2049         } else {
2050                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2051
2052                 mutex_lock(&n1->write_lock);
2053                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2054         }
2055
2056         bch_keylist_add(&parent_keys, &n1->key);
2057         bch_btree_node_write(n1, &cl);
2058         mutex_unlock(&n1->write_lock);
2059
2060         if (n3) {
2061                 /* Depth increases, make a new root */
2062                 mutex_lock(&n3->write_lock);
2063                 bkey_copy_key(&n3->key, &MAX_KEY);
2064                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2065                 bch_btree_node_write(n3, &cl);
2066                 mutex_unlock(&n3->write_lock);
2067
2068                 closure_sync(&cl);
2069                 bch_btree_set_root(n3);
2070                 rw_unlock(true, n3);
2071         } else if (!b->parent) {
2072                 /* Root filled up but didn't need to be split */
2073                 closure_sync(&cl);
2074                 bch_btree_set_root(n1);
2075         } else {
2076                 /* Split a non root node */
2077                 closure_sync(&cl);
2078                 make_btree_freeing_key(b, parent_keys.top);
2079                 bch_keylist_push(&parent_keys);
2080
2081                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2082                 BUG_ON(!bch_keylist_empty(&parent_keys));
2083         }
2084
2085         btree_node_free(b);
2086         rw_unlock(true, n1);
2087
2088         bch_time_stats_update(&b->c->btree_split_time, start_time);
2089
2090         return 0;
2091 err_free2:
2092         bkey_put(b->c, &n2->key);
2093         btree_node_free(n2);
2094         rw_unlock(true, n2);
2095 err_free1:
2096         bkey_put(b->c, &n1->key);
2097         btree_node_free(n1);
2098         rw_unlock(true, n1);
2099 err:
2100         WARN(1, "bcache: btree split failed (level %u)", b->level);
2101
2102         if (n3 == ERR_PTR(-EAGAIN) ||
2103             n2 == ERR_PTR(-EAGAIN) ||
2104             n1 == ERR_PTR(-EAGAIN))
2105                 return -EAGAIN;
2106
2107         return -ENOMEM;
2108 }
2109
2110 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2111                                  struct keylist *insert_keys,
2112                                  atomic_t *journal_ref,
2113                                  struct bkey *replace_key)
2114 {
2115         struct closure cl;
2116
2117         BUG_ON(b->level && replace_key);
2118
2119         closure_init_stack(&cl);
2120
2121         mutex_lock(&b->write_lock);
2122
2123         if (write_block(b) != btree_bset_last(b) &&
2124             b->keys.last_set_unwritten)
2125                 bch_btree_init_next(b); /* just wrote a set */
2126
2127         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2128                 mutex_unlock(&b->write_lock);
2129                 goto split;
2130         }
2131
2132         BUG_ON(write_block(b) != btree_bset_last(b));
2133
2134         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2135                 if (!b->level)
2136                         bch_btree_leaf_dirty(b, journal_ref);
2137                 else
2138                         bch_btree_node_write(b, &cl);
2139         }
2140
2141         mutex_unlock(&b->write_lock);
2142
2143         /* wait for btree node write if necessary, after unlock */
2144         closure_sync(&cl);
2145
2146         return 0;
2147 split:
2148         if (current->bio_list) {
2149                 op->lock = b->c->root->level + 1;
2150                 return -EAGAIN;
2151         } else if (op->lock <= b->c->root->level) {
2152                 op->lock = b->c->root->level + 1;
2153                 return -EINTR;
2154         } else {
2155                 /* Invalidated all iterators */
2156                 int ret = btree_split(b, op, insert_keys, replace_key);
2157
2158                 if (bch_keylist_empty(insert_keys))
2159                         return 0;
2160                 else if (!ret)
2161                         return -EINTR;
2162                 return ret;
2163         }
2164 }
2165
2166 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2167                                struct bkey *check_key)
2168 {
2169         int ret = -EINTR;
2170         uint64_t btree_ptr = b->key.ptr[0];
2171         unsigned long seq = b->seq;
2172         struct keylist insert;
2173         bool upgrade = op->lock == -1;
2174
2175         bch_keylist_init(&insert);
2176
2177         if (upgrade) {
2178                 rw_unlock(false, b);
2179                 rw_lock(true, b, b->level);
2180
2181                 if (b->key.ptr[0] != btree_ptr ||
2182                    b->seq != seq + 1) {
2183                        op->lock = b->level;
2184                         goto out;
2185                }
2186         }
2187
2188         SET_KEY_PTRS(check_key, 1);
2189         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2190
2191         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2192
2193         bch_keylist_add(&insert, check_key);
2194
2195         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2196
2197         BUG_ON(!ret && !bch_keylist_empty(&insert));
2198 out:
2199         if (upgrade)
2200                 downgrade_write(&b->lock);
2201         return ret;
2202 }
2203
2204 struct btree_insert_op {
2205         struct btree_op op;
2206         struct keylist  *keys;
2207         atomic_t        *journal_ref;
2208         struct bkey     *replace_key;
2209 };
2210
2211 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2212 {
2213         struct btree_insert_op *op = container_of(b_op,
2214                                         struct btree_insert_op, op);
2215
2216         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2217                                         op->journal_ref, op->replace_key);
2218         if (ret && !bch_keylist_empty(op->keys))
2219                 return ret;
2220         else
2221                 return MAP_DONE;
2222 }
2223
2224 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2225                      atomic_t *journal_ref, struct bkey *replace_key)
2226 {
2227         struct btree_insert_op op;
2228         int ret = 0;
2229
2230         BUG_ON(current->bio_list);
2231         BUG_ON(bch_keylist_empty(keys));
2232
2233         bch_btree_op_init(&op.op, 0);
2234         op.keys         = keys;
2235         op.journal_ref  = journal_ref;
2236         op.replace_key  = replace_key;
2237
2238         while (!ret && !bch_keylist_empty(keys)) {
2239                 op.op.lock = 0;
2240                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2241                                                &START_KEY(keys->keys),
2242                                                btree_insert_fn);
2243         }
2244
2245         if (ret) {
2246                 struct bkey *k;
2247
2248                 pr_err("error %i", ret);
2249
2250                 while ((k = bch_keylist_pop(keys)))
2251                         bkey_put(c, k);
2252         } else if (op.op.insert_collision)
2253                 ret = -ESRCH;
2254
2255         return ret;
2256 }
2257
2258 void bch_btree_set_root(struct btree *b)
2259 {
2260         unsigned i;
2261         struct closure cl;
2262
2263         closure_init_stack(&cl);
2264
2265         trace_bcache_btree_set_root(b);
2266
2267         BUG_ON(!b->written);
2268
2269         for (i = 0; i < KEY_PTRS(&b->key); i++)
2270                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2271
2272         mutex_lock(&b->c->bucket_lock);
2273         list_del_init(&b->list);
2274         mutex_unlock(&b->c->bucket_lock);
2275
2276         b->c->root = b;
2277
2278         bch_journal_meta(b->c, &cl);
2279         closure_sync(&cl);
2280 }
2281
2282 /* Map across nodes or keys */
2283
2284 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2285                                        struct bkey *from,
2286                                        btree_map_nodes_fn *fn, int flags)
2287 {
2288         int ret = MAP_CONTINUE;
2289
2290         if (b->level) {
2291                 struct bkey *k;
2292                 struct btree_iter iter;
2293
2294                 bch_btree_iter_init(&b->keys, &iter, from);
2295
2296                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2297                                                        bch_ptr_bad))) {
2298                         ret = btree(map_nodes_recurse, k, b,
2299                                     op, from, fn, flags);
2300                         from = NULL;
2301
2302                         if (ret != MAP_CONTINUE)
2303                                 return ret;
2304                 }
2305         }
2306
2307         if (!b->level || flags == MAP_ALL_NODES)
2308                 ret = fn(op, b);
2309
2310         return ret;
2311 }
2312
2313 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2314                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2315 {
2316         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2317 }
2318
2319 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2320                                       struct bkey *from, btree_map_keys_fn *fn,
2321                                       int flags)
2322 {
2323         int ret = MAP_CONTINUE;
2324         struct bkey *k;
2325         struct btree_iter iter;
2326
2327         bch_btree_iter_init(&b->keys, &iter, from);
2328
2329         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2330                 ret = !b->level
2331                         ? fn(op, b, k)
2332                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2333                 from = NULL;
2334
2335                 if (ret != MAP_CONTINUE)
2336                         return ret;
2337         }
2338
2339         if (!b->level && (flags & MAP_END_KEY))
2340                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2341                                      KEY_OFFSET(&b->key), 0));
2342
2343         return ret;
2344 }
2345
2346 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2347                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2348 {
2349         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2350 }
2351
2352 /* Keybuf code */
2353
2354 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2355 {
2356         /* Overlapping keys compare equal */
2357         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2358                 return -1;
2359         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2360                 return 1;
2361         return 0;
2362 }
2363
2364 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2365                                             struct keybuf_key *r)
2366 {
2367         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2368 }
2369
2370 struct refill {
2371         struct btree_op op;
2372         unsigned        nr_found;
2373         struct keybuf   *buf;
2374         struct bkey     *end;
2375         keybuf_pred_fn  *pred;
2376 };
2377
2378 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2379                             struct bkey *k)
2380 {
2381         struct refill *refill = container_of(op, struct refill, op);
2382         struct keybuf *buf = refill->buf;
2383         int ret = MAP_CONTINUE;
2384
2385         if (bkey_cmp(k, refill->end) > 0) {
2386                 ret = MAP_DONE;
2387                 goto out;
2388         }
2389
2390         if (!KEY_SIZE(k)) /* end key */
2391                 goto out;
2392
2393         if (refill->pred(buf, k)) {
2394                 struct keybuf_key *w;
2395
2396                 spin_lock(&buf->lock);
2397
2398                 w = array_alloc(&buf->freelist);
2399                 if (!w) {
2400                         spin_unlock(&buf->lock);
2401                         return MAP_DONE;
2402                 }
2403
2404                 w->private = NULL;
2405                 bkey_copy(&w->key, k);
2406
2407                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2408                         array_free(&buf->freelist, w);
2409                 else
2410                         refill->nr_found++;
2411
2412                 if (array_freelist_empty(&buf->freelist))
2413                         ret = MAP_DONE;
2414
2415                 spin_unlock(&buf->lock);
2416         }
2417 out:
2418         buf->last_scanned = *k;
2419         return ret;
2420 }
2421
2422 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2423                        struct bkey *end, keybuf_pred_fn *pred)
2424 {
2425         struct bkey start = buf->last_scanned;
2426         struct refill refill;
2427
2428         cond_resched();
2429
2430         bch_btree_op_init(&refill.op, -1);
2431         refill.nr_found = 0;
2432         refill.buf      = buf;
2433         refill.end      = end;
2434         refill.pred     = pred;
2435
2436         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2437                            refill_keybuf_fn, MAP_END_KEY);
2438
2439         trace_bcache_keyscan(refill.nr_found,
2440                              KEY_INODE(&start), KEY_OFFSET(&start),
2441                              KEY_INODE(&buf->last_scanned),
2442                              KEY_OFFSET(&buf->last_scanned));
2443
2444         spin_lock(&buf->lock);
2445
2446         if (!RB_EMPTY_ROOT(&buf->keys)) {
2447                 struct keybuf_key *w;
2448                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2449                 buf->start      = START_KEY(&w->key);
2450
2451                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2452                 buf->end        = w->key;
2453         } else {
2454                 buf->start      = MAX_KEY;
2455                 buf->end        = MAX_KEY;
2456         }
2457
2458         spin_unlock(&buf->lock);
2459 }
2460
2461 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2462 {
2463         rb_erase(&w->node, &buf->keys);
2464         array_free(&buf->freelist, w);
2465 }
2466
2467 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2468 {
2469         spin_lock(&buf->lock);
2470         __bch_keybuf_del(buf, w);
2471         spin_unlock(&buf->lock);
2472 }
2473
2474 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2475                                   struct bkey *end)
2476 {
2477         bool ret = false;
2478         struct keybuf_key *p, *w, s;
2479         s.key = *start;
2480
2481         if (bkey_cmp(end, &buf->start) <= 0 ||
2482             bkey_cmp(start, &buf->end) >= 0)
2483                 return false;
2484
2485         spin_lock(&buf->lock);
2486         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2487
2488         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2489                 p = w;
2490                 w = RB_NEXT(w, node);
2491
2492                 if (p->private)
2493                         ret = true;
2494                 else
2495                         __bch_keybuf_del(buf, p);
2496         }
2497
2498         spin_unlock(&buf->lock);
2499         return ret;
2500 }
2501
2502 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2503 {
2504         struct keybuf_key *w;
2505         spin_lock(&buf->lock);
2506
2507         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2508
2509         while (w && w->private)
2510                 w = RB_NEXT(w, node);
2511
2512         if (w)
2513                 w->private = ERR_PTR(-EINTR);
2514
2515         spin_unlock(&buf->lock);
2516         return w;
2517 }
2518
2519 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2520                                           struct keybuf *buf,
2521                                           struct bkey *end,
2522                                           keybuf_pred_fn *pred)
2523 {
2524         struct keybuf_key *ret;
2525
2526         while (1) {
2527                 ret = bch_keybuf_next(buf);
2528                 if (ret)
2529                         break;
2530
2531                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2532                         pr_debug("scan finished");
2533                         break;
2534                 }
2535
2536                 bch_refill_keybuf(c, buf, end, pred);
2537         }
2538
2539         return ret;
2540 }
2541
2542 void bch_keybuf_init(struct keybuf *buf)
2543 {
2544         buf->last_scanned       = MAX_KEY;
2545         buf->keys               = RB_ROOT;
2546
2547         spin_lock_init(&buf->lock);
2548         array_allocator_init(&buf->freelist);
2549 }