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 static struct workqueue_struct *btree_io_wq;
  97
  98 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
  99
 100 /*
 101  * These macros are for recursing down the btree - they handle the details of
 102  * locking and looking up nodes in the cache for you. They're best treated as
 103  * mere syntax when reading code that uses them.
 104  *
 105  * op->lock determines whether we take a read or a write lock at a given depth.
 106  * If you've got a read lock and find that you need a write lock (i.e. you're
 107  * going to have to split), set op->lock and return -EINTR; btree_root() will
 108  * call you again and you'll have the correct lock.
 109  */
 110
 111 /**
 112  * btree - recurse down the btree on a specified key
 113  * @fn:         function to call, which will be passed the child node
 114  * @key:        key to recurse on
 115  * @b:          parent btree node
 116  * @op:         pointer to struct btree_op
 117  */
 118 #define btree(fn, key, b, op, ...)                                      \
 119 ({                                                                      \
 120         int _r, l = (b)->level - 1;                                     \
 121         bool _w = l <= (op)->lock;                                      \
 122         struct btree *_child = bch_btree_node_get((b)->c, key, l, _w);  \
 123         if (!IS_ERR(_child)) {                                          \
 124                 _child->parent = (b);                                   \
 125                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 126                 rw_unlock(_w, _child);                                  \
 127         } else                                                          \
 128                 _r = PTR_ERR(_child);                                   \
 129         _r;                                                             \
 130 })
 131
 132 /**
 133  * btree_root - call a function on the root of the btree
 134  * @fn:         function to call, which will be passed the child node
 135  * @c:          cache set
 136  * @op:         pointer to struct btree_op
 137  */
 138 #define btree_root(fn, c, op, ...)                                      \
 139 ({                                                                      \
 140         int _r = -EINTR;                                                \
 141         do {                                                            \
 142                 struct btree *_b = (c)->root;                           \
 143                 bool _w = insert_lock(op, _b);                          \
 144                 rw_lock(_w, _b, _b->level);                             \
 145                 if (_b == (c)->root &&                                  \
 146                     _w == insert_lock(op, _b)) {                        \
 147                         _b->parent = NULL;                              \
 148                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 149                 }                                                       \
 150                 rw_unlock(_w, _b);                                      \
 151                 if (_r == -EINTR)                                       \
 152                         schedule();                                     \
 153                 bch_cannibalize_unlock(c);                              \
 154                 if (_r == -ENOSPC) {                                    \
 155                         wait_event((c)->try_wait,                       \
 156                                    !(c)->try_harder);                   \
 157                         _r = -EINTR;                                    \
 158                 }                                                       \
 159         } while (_r == -EINTR);                                         \
 160                                                                         \
 161         finish_wait(&(c)->bucket_wait, &(op)->wait);                    \
 162         _r;                                                             \
 163 })
 164
 165 static inline struct bset *write_block(struct btree *b)
 166 {
 167         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 168 }
 169
 170 static void bch_btree_init_next(struct btree *b)
 171 {
 172         /* If not a leaf node, always sort */
 173         if (b->level && b->keys.nsets)
 174                 bch_btree_sort(&b->keys, &b->c->sort);
 175         else
 176                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 177
 178         if (b->written < btree_blocks(b))
 179                 bch_bset_init_next(&b->keys, write_block(b),
 180                                    bset_magic(&b->c->sb));
 181
 182 }
 183
 184 /* Btree key manipulation */
 185
 186 void bkey_put(struct cache_set *c, struct bkey *k)
 187 {
 188         unsigned i;
 189
 190         for (i = 0; i < KEY_PTRS(k); i++)
 191                 if (ptr_available(c, k, i))
 192                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 193 }
 194
 195 /* Btree IO */
 196
 197 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 198 {
 199         uint64_t crc = b->key.ptr[0];
 200         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 201
 202         crc = bch_crc64_update(crc, data, end - data);
 203         return crc ^ 0xffffffffffffffffULL;
 204 }
 205
 206 void bch_btree_node_read_done(struct btree *b)
 207 {
 208         const char *err = "bad btree header";
 209         struct bset *i = btree_bset_first(b);
 210         struct btree_iter *iter;
 211
 212         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 213         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 214         iter->used = 0;
 215
 216 #ifdef CONFIG_BCACHE_DEBUG
 217         iter->b = &b->keys;
 218 #endif
 219
 220         if (!i->seq)
 221                 goto err;
 222
 223         for (;
 224              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 225              i = write_block(b)) {
 226                 err = "unsupported bset version";
 227                 if (i->version > BCACHE_BSET_VERSION)
 228                         goto err;
 229
 230                 err = "bad btree header";
 231                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 232                     btree_blocks(b))
 233                         goto err;
 234
 235                 err = "bad magic";
 236                 if (i->magic != bset_magic(&b->c->sb))
 237                         goto err;
 238
 239                 err = "bad checksum";
 240                 switch (i->version) {
 241                 case 0:
 242                         if (i->csum != csum_set(i))
 243                                 goto err;
 244                         break;
 245                 case BCACHE_BSET_VERSION:
 246                         if (i->csum != btree_csum_set(b, i))
 247                                 goto err;
 248                         break;
 249                 }
 250
 251                 err = "empty set";
 252                 if (i != b->keys.set[0].data && !i->keys)
 253                         goto err;
 254
 255                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 256
 257                 b->written += set_blocks(i, block_bytes(b->c));
 258         }
 259
 260         err = "corrupted btree";
 261         for (i = write_block(b);
 262              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 263              i = ((void *) i) + block_bytes(b->c))
 264                 if (i->seq == b->keys.set[0].data->seq)
 265                         goto err;
 266
 267         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 268
 269         i = b->keys.set[0].data;
 270         err = "short btree key";
 271         if (b->keys.set[0].size &&
 272             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 273                 goto err;
 274
 275         if (b->written < btree_blocks(b))
 276                 bch_bset_init_next(&b->keys, write_block(b),
 277                                    bset_magic(&b->c->sb));
 278 out:
 279         mempool_free(iter, b->c->fill_iter);
 280         return;
 281 err:
 282         set_btree_node_io_error(b);
 283         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 284                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 285                             bset_block_offset(b, i), i->keys);
 286         goto out;
 287 }
 288
 289 static void btree_node_read_endio(struct bio *bio, int error)
 290 {
 291         struct closure *cl = bio->bi_private;
 292         closure_put(cl);
 293 }
 294
 295 static void bch_btree_node_read(struct btree *b)
 296 {
 297         uint64_t start_time = local_clock();
 298         struct closure cl;
 299         struct bio *bio;
 300
 301         trace_bcache_btree_read(b);
 302
 303         closure_init_stack(&cl);
 304
 305         bio = bch_bbio_alloc(b->c);
 306         bio->bi_rw      = REQ_META|READ_SYNC;
 307         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 308         bio->bi_end_io  = btree_node_read_endio;
 309         bio->bi_private = &cl;
 310
 311         bch_bio_map(bio, b->keys.set[0].data);
 312
 313         bch_submit_bbio(bio, b->c, &b->key, 0);
 314         closure_sync(&cl);
 315
 316         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 317                 set_btree_node_io_error(b);
 318
 319         bch_bbio_free(bio, b->c);
 320
 321         if (btree_node_io_error(b))
 322                 goto err;
 323
 324         bch_btree_node_read_done(b);
 325         bch_time_stats_update(&b->c->btree_read_time, start_time);
 326
 327         return;
 328 err:
 329         bch_cache_set_error(b->c, "io error reading bucket %zu",
 330                             PTR_BUCKET_NR(b->c, &b->key, 0));
 331 }
 332
 333 static void btree_complete_write(struct btree *b, struct btree_write *w)
 334 {
 335         if (w->prio_blocked &&
 336             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 337                 wake_up_allocators(b->c);
 338
 339         if (w->journal) {
 340                 atomic_dec_bug(w->journal);
 341                 __closure_wake_up(&b->c->journal.wait);
 342         }
 343
 344         w->prio_blocked = 0;
 345         w->journal      = NULL;
 346 }
 347
 348 static void btree_node_write_unlock(struct closure *cl)
 349 {
 350         struct btree *b = container_of(cl, struct btree, io);
 351
 352         up(&b->io_mutex);
 353 }
 354
 355 static void __btree_node_write_done(struct closure *cl)
 356 {
 357         struct btree *b = container_of(cl, struct btree, io);
 358         struct btree_write *w = btree_prev_write(b);
 359
 360         bch_bbio_free(b->bio, b->c);
 361         b->bio = NULL;
 362         btree_complete_write(b, w);
 363
 364         if (btree_node_dirty(b))
 365                 queue_delayed_work(btree_io_wq, &b->work,
 366                                    msecs_to_jiffies(30000));
 367
 368         closure_return_with_destructor(cl, btree_node_write_unlock);
 369 }
 370
 371 static void btree_node_write_done(struct closure *cl)
 372 {
 373         struct btree *b = container_of(cl, struct btree, io);
 374         struct bio_vec *bv;
 375         int n;
 376
 377         bio_for_each_segment_all(bv, b->bio, n)
 378                 __free_page(bv->bv_page);
 379
 380         __btree_node_write_done(cl);
 381 }
 382
 383 static void btree_node_write_endio(struct bio *bio, int error)
 384 {
 385         struct closure *cl = bio->bi_private;
 386         struct btree *b = container_of(cl, struct btree, io);
 387
 388         if (error)
 389                 set_btree_node_io_error(b);
 390
 391         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 392         closure_put(cl);
 393 }
 394
 395 static void do_btree_node_write(struct btree *b)
 396 {
 397         struct closure *cl = &b->io;
 398         struct bset *i = btree_bset_last(b);
 399         BKEY_PADDED(key) k;
 400
 401         i->version      = BCACHE_BSET_VERSION;
 402         i->csum         = btree_csum_set(b, i);
 403
 404         BUG_ON(b->bio);
 405         b->bio = bch_bbio_alloc(b->c);
 406
 407         b->bio->bi_end_io       = btree_node_write_endio;
 408         b->bio->bi_private      = cl;
 409         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 410         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 411         bch_bio_map(b->bio, i);
 412
 413         /*
 414          * If we're appending to a leaf node, we don't technically need FUA -
 415          * this write just needs to be persisted before the next journal write,
 416          * which will be marked FLUSH|FUA.
 417          *
 418          * Similarly if we're writing a new btree root - the pointer is going to
 419          * be in the next journal entry.
 420          *
 421          * But if we're writing a new btree node (that isn't a root) or
 422          * appending to a non leaf btree node, we need either FUA or a flush
 423          * when we write the parent with the new pointer. FUA is cheaper than a
 424          * flush, and writes appending to leaf nodes aren't blocking anything so
 425          * just make all btree node writes FUA to keep things sane.
 426          */
 427
 428         bkey_copy(&k.key, &b->key);
 429         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 430                        bset_sector_offset(&b->keys, i));
 431
 432         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 433                 int j;
 434                 struct bio_vec *bv;
 435                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 436
 437                 bio_for_each_segment_all(bv, b->bio, j)
 438                         memcpy(page_address(bv->bv_page),
 439                                base + j * PAGE_SIZE, PAGE_SIZE);
 440
 441                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 442
 443                 continue_at(cl, btree_node_write_done, NULL);
 444         } else {
 445                 b->bio->bi_vcnt = 0;
 446                 bch_bio_map(b->bio, i);
 447
 448                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 449
 450                 closure_sync(cl);
 451                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 452         }
 453 }
 454
 455 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 456 {
 457         struct bset *i = btree_bset_last(b);
 458
 459         lockdep_assert_held(&b->write_lock);
 460
 461         trace_bcache_btree_write(b);
 462
 463         BUG_ON(current->bio_list);
 464         BUG_ON(b->written >= btree_blocks(b));
 465         BUG_ON(b->written && !i->keys);
 466         BUG_ON(btree_bset_first(b)->seq != i->seq);
 467         bch_check_keys(&b->keys, "writing");
 468
 469         cancel_delayed_work(&b->work);
 470
 471         /* If caller isn't waiting for write, parent refcount is cache set */
 472         down(&b->io_mutex);
 473         closure_init(&b->io, parent ?: &b->c->cl);
 474
 475         clear_bit(BTREE_NODE_dirty,      &b->flags);
 476         change_bit(BTREE_NODE_write_idx, &b->flags);
 477
 478         do_btree_node_write(b);
 479
 480         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 481                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 482
 483         b->written += set_blocks(i, block_bytes(b->c));
 484 }
 485
 486 void bch_btree_node_write(struct btree *b, struct closure *parent)
 487 {
 488         unsigned nsets = b->keys.nsets;
 489
 490         lockdep_assert_held(&b->lock);
 491
 492         __bch_btree_node_write(b, parent);
 493
 494         /*
 495          * do verify if there was more than one set initially (i.e. we did a
 496          * sort) and we sorted down to a single set:
 497          */
 498         if (nsets && !b->keys.nsets)
 499                 bch_btree_verify(b);
 500
 501         bch_btree_init_next(b);
 502 }
 503
 504 static void bch_btree_node_write_sync(struct btree *b)
 505 {
 506         struct closure cl;
 507
 508         closure_init_stack(&cl);
 509
 510         mutex_lock(&b->write_lock);
 511         bch_btree_node_write(b, &cl);
 512         mutex_unlock(&b->write_lock);
 513
 514         closure_sync(&cl);
 515 }
 516
 517 static void btree_node_write_work(struct work_struct *w)
 518 {
 519         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 520
 521         mutex_lock(&b->write_lock);
 522         if (btree_node_dirty(b))
 523                 __bch_btree_node_write(b, NULL);
 524         mutex_unlock(&b->write_lock);
 525 }
 526
 527 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 528 {
 529         struct bset *i = btree_bset_last(b);
 530         struct btree_write *w = btree_current_write(b);
 531
 532         lockdep_assert_held(&b->write_lock);
 533
 534         BUG_ON(!b->written);
 535         BUG_ON(!i->keys);
 536
 537         if (!btree_node_dirty(b))
 538                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 539
 540         set_btree_node_dirty(b);
 541
 542         if (journal_ref) {
 543                 if (w->journal &&
 544                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 545                         atomic_dec_bug(w->journal);
 546                         w->journal = NULL;
 547                 }
 548
 549                 if (!w->journal) {
 550                         w->journal = journal_ref;
 551                         atomic_inc(w->journal);
 552                 }
 553         }
 554
 555         /* Force write if set is too big */
 556         if (set_bytes(i) > PAGE_SIZE - 48 &&
 557             !current->bio_list)
 558                 bch_btree_node_write(b, NULL);
 559 }
 560
 561 /*
 562  * Btree in memory cache - allocation/freeing
 563  * mca -> memory cache
 564  */
 565
 566 #define mca_reserve(c)  (((c->root && c->root->level)           \
 567                           ? c->root->level : 1) * 8 + 16)
 568 #define mca_can_free(c)                                         \
 569         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 570
 571 static void mca_data_free(struct btree *b)
 572 {
 573         BUG_ON(b->io_mutex.count != 1);
 574
 575         bch_btree_keys_free(&b->keys);
 576
 577         b->c->bucket_cache_used--;
 578         list_move(&b->list, &b->c->btree_cache_freed);
 579 }
 580
 581 static void mca_bucket_free(struct btree *b)
 582 {
 583         BUG_ON(btree_node_dirty(b));
 584
 585         b->key.ptr[0] = 0;
 586         hlist_del_init_rcu(&b->hash);
 587         list_move(&b->list, &b->c->btree_cache_freeable);
 588 }
 589
 590 static unsigned btree_order(struct bkey *k)
 591 {
 592         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 593 }
 594
 595 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 596 {
 597         if (!bch_btree_keys_alloc(&b->keys,
 598                                   max_t(unsigned,
 599                                         ilog2(b->c->btree_pages),
 600                                         btree_order(k)),
 601                                   gfp)) {
 602                 b->c->bucket_cache_used++;
 603                 list_move(&b->list, &b->c->btree_cache);
 604         } else {
 605                 list_move(&b->list, &b->c->btree_cache_freed);
 606         }
 607 }
 608
 609 static struct btree *mca_bucket_alloc(struct cache_set *c,
 610                                       struct bkey *k, gfp_t gfp)
 611 {
 612         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 613         if (!b)
 614                 return NULL;
 615
 616         init_rwsem(&b->lock);
 617         lockdep_set_novalidate_class(&b->lock);
 618         mutex_init(&b->write_lock);
 619         lockdep_set_novalidate_class(&b->write_lock);
 620         INIT_LIST_HEAD(&b->list);
 621         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 622         b->c = c;
 623         sema_init(&b->io_mutex, 1);
 624
 625         mca_data_alloc(b, k, gfp);
 626         return b;
 627 }
 628
 629 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 630 {
 631         struct closure cl;
 632
 633         closure_init_stack(&cl);
 634         lockdep_assert_held(&b->c->bucket_lock);
 635
 636         if (!down_write_trylock(&b->lock))
 637                 return -ENOMEM;
 638
 639         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 640
 641         if (b->keys.page_order < min_order)
 642                 goto out_unlock;
 643
 644         if (!flush) {
 645                 if (btree_node_dirty(b))
 646                         goto out_unlock;
 647
 648                 if (down_trylock(&b->io_mutex))
 649                         goto out_unlock;
 650                 up(&b->io_mutex);
 651         }
 652
 653         mutex_lock(&b->write_lock);
 654         if (btree_node_dirty(b))
 655                 __bch_btree_node_write(b, &cl);
 656         mutex_unlock(&b->write_lock);
 657
 658         closure_sync(&cl);
 659
 660         /* wait for any in flight btree write */
 661         down(&b->io_mutex);
 662         up(&b->io_mutex);
 663
 664         return 0;
 665 out_unlock:
 666         rw_unlock(true, b);
 667         return -ENOMEM;
 668 }
 669
 670 static unsigned long bch_mca_scan(struct shrinker *shrink,
 671                                   struct shrink_control *sc)
 672 {
 673         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 674         struct btree *b, *t;
 675         unsigned long i, nr = sc->nr_to_scan;
 676         unsigned long freed = 0;
 677
 678         if (c->shrinker_disabled)
 679                 return SHRINK_STOP;
 680
 681         if (c->try_harder)
 682                 return SHRINK_STOP;
 683
 684         /* Return -1 if we can't do anything right now */
 685         if (sc->gfp_mask & __GFP_IO)
 686                 mutex_lock(&c->bucket_lock);
 687         else if (!mutex_trylock(&c->bucket_lock))
 688                 return -1;
 689
 690         /*
 691          * It's _really_ critical that we don't free too many btree nodes - we
 692          * have to always leave ourselves a reserve. The reserve is how we
 693          * guarantee that allocating memory for a new btree node can always
 694          * succeed, so that inserting keys into the btree can always succeed and
 695          * IO can always make forward progress:
 696          */
 697         nr /= c->btree_pages;
 698         nr = min_t(unsigned long, nr, mca_can_free(c));
 699
 700         i = 0;
 701         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 702                 if (freed >= nr)
 703                         break;
 704
 705                 if (++i > 3 &&
 706                     !mca_reap(b, 0, false)) {
 707                         mca_data_free(b);
 708                         rw_unlock(true, b);
 709                         freed++;
 710                 }
 711         }
 712
 713         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 714                 if (list_empty(&c->btree_cache))
 715                         goto out;
 716
 717                 b = list_first_entry(&c->btree_cache, struct btree, list);
 718                 list_rotate_left(&c->btree_cache);
 719
 720                 if (!b->accessed &&
 721                     !mca_reap(b, 0, false)) {
 722                         mca_bucket_free(b);
 723                         mca_data_free(b);
 724                         rw_unlock(true, b);
 725                         freed++;
 726                 } else
 727                         b->accessed = 0;
 728         }
 729 out:
 730         mutex_unlock(&c->bucket_lock);
 731         return freed;
 732 }
 733
 734 static unsigned long bch_mca_count(struct shrinker *shrink,
 735                                    struct shrink_control *sc)
 736 {
 737         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 738
 739         if (c->shrinker_disabled)
 740                 return 0;
 741
 742         if (c->try_harder)
 743                 return 0;
 744
 745         return mca_can_free(c) * c->btree_pages;
 746 }
 747
 748 void bch_btree_cache_free(struct cache_set *c)
 749 {
 750         struct btree *b;
 751         struct closure cl;
 752         closure_init_stack(&cl);
 753
 754         if (c->shrink.list.next)
 755                 unregister_shrinker(&c->shrink);
 756
 757         mutex_lock(&c->bucket_lock);
 758
 759 #ifdef CONFIG_BCACHE_DEBUG
 760         if (c->verify_data)
 761                 list_move(&c->verify_data->list, &c->btree_cache);
 762
 763         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 764 #endif
 765
 766         list_splice(&c->btree_cache_freeable,
 767                     &c->btree_cache);
 768
 769         while (!list_empty(&c->btree_cache)) {
 770                 b = list_first_entry(&c->btree_cache, struct btree, list);
 771
 772                 if (btree_node_dirty(b))
 773                         btree_complete_write(b, btree_current_write(b));
 774                 clear_bit(BTREE_NODE_dirty, &b->flags);
 775
 776                 mca_data_free(b);
 777         }
 778
 779         while (!list_empty(&c->btree_cache_freed)) {
 780                 b = list_first_entry(&c->btree_cache_freed,
 781                                      struct btree, list);
 782                 list_del(&b->list);
 783                 cancel_delayed_work_sync(&b->work);
 784                 kfree(b);
 785         }
 786
 787         mutex_unlock(&c->bucket_lock);
 788 }
 789
 790 int bch_btree_cache_alloc(struct cache_set *c)
 791 {
 792         unsigned i;
 793
 794         for (i = 0; i < mca_reserve(c); i++)
 795                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 796                         return -ENOMEM;
 797
 798         list_splice_init(&c->btree_cache,
 799                          &c->btree_cache_freeable);
 800
 801 #ifdef CONFIG_BCACHE_DEBUG
 802         mutex_init(&c->verify_lock);
 803
 804         c->verify_ondisk = (void *)
 805                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 806
 807         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 808
 809         if (c->verify_data &&
 810             c->verify_data->keys.set->data)
 811                 list_del_init(&c->verify_data->list);
 812         else
 813                 c->verify_data = NULL;
 814 #endif
 815
 816         c->shrink.count_objects = bch_mca_count;
 817         c->shrink.scan_objects = bch_mca_scan;
 818         c->shrink.seeks = 4;
 819         c->shrink.batch = c->btree_pages * 2;
 820         register_shrinker(&c->shrink);
 821
 822         return 0;
 823 }
 824
 825 /* Btree in memory cache - hash table */
 826
 827 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 828 {
 829         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 830 }
 831
 832 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 833 {
 834         struct btree *b;
 835
 836         rcu_read_lock();
 837         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 838                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 839                         goto out;
 840         b = NULL;
 841 out:
 842         rcu_read_unlock();
 843         return b;
 844 }
 845
 846 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 847 {
 848         struct btree *b;
 849
 850         trace_bcache_btree_cache_cannibalize(c);
 851
 852         if (!c->try_harder) {
 853                 c->try_harder = current;
 854                 c->try_harder_start = local_clock();
 855         } else if (c->try_harder != current)
 856                 return ERR_PTR(-ENOSPC);
 857
 858         list_for_each_entry_reverse(b, &c->btree_cache, list)
 859                 if (!mca_reap(b, btree_order(k), false))
 860                         return b;
 861
 862         list_for_each_entry_reverse(b, &c->btree_cache, list)
 863                 if (!mca_reap(b, btree_order(k), true))
 864                         return b;
 865
 866         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->try_harder == current) {
 878                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 879                 c->try_harder = NULL;
 880                 wake_up(&c->try_wait);
 881         }
 882 }
 883
 884 static struct btree *mca_alloc(struct cache_set *c, 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, 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 bkey *k,
 964                                  int level, bool write)
 965 {
 966         int i = 0;
 967         struct btree *b;
 968
 969         BUG_ON(level < 0);
 970 retry:
 971         b = mca_find(c, k);
 972
 973         if (!b) {
 974                 if (current->bio_list)
 975                         return ERR_PTR(-EAGAIN);
 976
 977                 mutex_lock(&c->bucket_lock);
 978                 b = mca_alloc(c, k, level);
 979                 mutex_unlock(&c->bucket_lock);
 980
 981                 if (!b)
 982                         goto retry;
 983                 if (IS_ERR(b))
 984                         return b;
 985
 986                 bch_btree_node_read(b);
 987
 988                 if (!write)
 989                         downgrade_write(&b->lock);
 990         } else {
 991                 rw_lock(write, b, level);
 992                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 993                         rw_unlock(write, b);
 994                         goto retry;
 995                 }
 996                 BUG_ON(b->level != level);
 997         }
 998
 999         b->accessed = 1;
1000
1001         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1002                 prefetch(b->keys.set[i].tree);
1003                 prefetch(b->keys.set[i].data);
1004         }
1005
1006         for (; i <= b->keys.nsets; i++)
1007                 prefetch(b->keys.set[i].data);
1008
1009         if (btree_node_io_error(b)) {
1010                 rw_unlock(write, b);
1011                 return ERR_PTR(-EIO);
1012         }
1013
1014         BUG_ON(!b->written);
1015
1016         return b;
1017 }
1018
1019 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1020 {
1021         struct btree *b;
1022
1023         mutex_lock(&c->bucket_lock);
1024         b = mca_alloc(c, k, level);
1025         mutex_unlock(&c->bucket_lock);
1026
1027         if (!IS_ERR_OR_NULL(b)) {
1028                 bch_btree_node_read(b);
1029                 rw_unlock(true, b);
1030         }
1031 }
1032
1033 /* Btree alloc */
1034
1035 static void btree_node_free(struct btree *b)
1036 {
1037         trace_bcache_btree_node_free(b);
1038
1039         BUG_ON(b == b->c->root);
1040
1041         mutex_lock(&b->write_lock);
1042
1043         if (btree_node_dirty(b))
1044                 btree_complete_write(b, btree_current_write(b));
1045         clear_bit(BTREE_NODE_dirty, &b->flags);
1046
1047         mutex_unlock(&b->write_lock);
1048
1049         cancel_delayed_work(&b->work);
1050
1051         mutex_lock(&b->c->bucket_lock);
1052         bch_bucket_free(b->c, &b->key);
1053         mca_bucket_free(b);
1054         mutex_unlock(&b->c->bucket_lock);
1055 }
1056
1057 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1058 {
1059         BKEY_PADDED(key) k;
1060         struct btree *b = ERR_PTR(-EAGAIN);
1061
1062         mutex_lock(&c->bucket_lock);
1063 retry:
1064         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1065                 goto err;
1066
1067         bkey_put(c, &k.key);
1068         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1069
1070         b = mca_alloc(c, &k.key, level);
1071         if (IS_ERR(b))
1072                 goto err_free;
1073
1074         if (!b) {
1075                 cache_bug(c,
1076                         "Tried to allocate bucket that was in btree cache");
1077                 goto retry;
1078         }
1079
1080         b->accessed = 1;
1081         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1082
1083         mutex_unlock(&c->bucket_lock);
1084
1085         trace_bcache_btree_node_alloc(b);
1086         return b;
1087 err_free:
1088         bch_bucket_free(c, &k.key);
1089 err:
1090         mutex_unlock(&c->bucket_lock);
1091
1092         trace_bcache_btree_node_alloc_fail(b);
1093         return b;
1094 }
1095
1096 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1097 {
1098         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1099         if (!IS_ERR_OR_NULL(n)) {
1100                 mutex_lock(&n->write_lock);
1101                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1102                 bkey_copy_key(&n->key, &b->key);
1103                 mutex_unlock(&n->write_lock);
1104         }
1105
1106         return n;
1107 }
1108
1109 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1110 {
1111         unsigned i;
1112
1113         mutex_lock(&b->c->bucket_lock);
1114
1115         atomic_inc(&b->c->prio_blocked);
1116
1117         bkey_copy(k, &b->key);
1118         bkey_copy_key(k, &ZERO_KEY);
1119
1120         for (i = 0; i < KEY_PTRS(k); i++)
1121                 SET_PTR_GEN(k, i,
1122                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1123                                         PTR_BUCKET(b->c, &b->key, i)));
1124
1125         mutex_unlock(&b->c->bucket_lock);
1126 }
1127
1128 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1129 {
1130         struct cache_set *c = b->c;
1131         struct cache *ca;
1132         unsigned i, reserve = c->root->level * 2 + 1;
1133         int ret = 0;
1134
1135         mutex_lock(&c->bucket_lock);
1136
1137         for_each_cache(ca, c, i)
1138                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1139                         if (op)
1140                                 prepare_to_wait(&c->bucket_wait, &op->wait,
1141                                                 TASK_UNINTERRUPTIBLE);
1142                         ret = -EINTR;
1143                         break;
1144                 }
1145
1146         mutex_unlock(&c->bucket_lock);
1147         return ret;
1148 }
1149
1150 /* Garbage collection */
1151
1152 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1153                                     struct bkey *k)
1154 {
1155         uint8_t stale = 0;
1156         unsigned i;
1157         struct bucket *g;
1158
1159         /*
1160          * ptr_invalid() can't return true for the keys that mark btree nodes as
1161          * freed, but since ptr_bad() returns true we'll never actually use them
1162          * for anything and thus we don't want mark their pointers here
1163          */
1164         if (!bkey_cmp(k, &ZERO_KEY))
1165                 return stale;
1166
1167         for (i = 0; i < KEY_PTRS(k); i++) {
1168                 if (!ptr_available(c, k, i))
1169                         continue;
1170
1171                 g = PTR_BUCKET(c, k, i);
1172
1173                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1174                         g->gc_gen = PTR_GEN(k, i);
1175
1176                 if (ptr_stale(c, k, i)) {
1177                         stale = max(stale, ptr_stale(c, k, i));
1178                         continue;
1179                 }
1180
1181                 cache_bug_on(GC_MARK(g) &&
1182                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1183                              c, "inconsistent ptrs: mark = %llu, level = %i",
1184                              GC_MARK(g), level);
1185
1186                 if (level)
1187                         SET_GC_MARK(g, GC_MARK_METADATA);
1188                 else if (KEY_DIRTY(k))
1189                         SET_GC_MARK(g, GC_MARK_DIRTY);
1190                 else if (!GC_MARK(g))
1191                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1192
1193                 /* guard against overflow */
1194                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1195                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1196                                              MAX_GC_SECTORS_USED));
1197
1198                 BUG_ON(!GC_SECTORS_USED(g));
1199         }
1200
1201         return stale;
1202 }
1203
1204 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1205
1206 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1207 {
1208         unsigned i;
1209
1210         for (i = 0; i < KEY_PTRS(k); i++)
1211                 if (ptr_available(c, k, i) &&
1212                     !ptr_stale(c, k, i)) {
1213                         struct bucket *b = PTR_BUCKET(c, k, i);
1214
1215                         b->gen = PTR_GEN(k, i);
1216
1217                         if (level && bkey_cmp(k, &ZERO_KEY))
1218                                 b->prio = BTREE_PRIO;
1219                         else if (!level && b->prio == BTREE_PRIO)
1220                                 b->prio = INITIAL_PRIO;
1221                 }
1222
1223         __bch_btree_mark_key(c, level, k);
1224 }
1225
1226 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1227 {
1228         uint8_t stale = 0;
1229         unsigned keys = 0, good_keys = 0;
1230         struct bkey *k;
1231         struct btree_iter iter;
1232         struct bset_tree *t;
1233
1234         gc->nodes++;
1235
1236         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1237                 stale = max(stale, btree_mark_key(b, k));
1238                 keys++;
1239
1240                 if (bch_ptr_bad(&b->keys, k))
1241                         continue;
1242
1243                 gc->key_bytes += bkey_u64s(k);
1244                 gc->nkeys++;
1245                 good_keys++;
1246
1247                 gc->data += KEY_SIZE(k);
1248         }
1249
1250         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1251                 btree_bug_on(t->size &&
1252                              bset_written(&b->keys, t) &&
1253                              bkey_cmp(&b->key, &t->end) < 0,
1254                              b, "found short btree key in gc");
1255
1256         if (b->c->gc_always_rewrite)
1257                 return true;
1258
1259         if (stale > 10)
1260                 return true;
1261
1262         if ((keys - good_keys) * 2 > keys)
1263                 return true;
1264
1265         return false;
1266 }
1267
1268 #define GC_MERGE_NODES  4U
1269
1270 struct gc_merge_info {
1271         struct btree    *b;
1272         unsigned        keys;
1273 };
1274
1275 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1276                                  struct keylist *, atomic_t *, struct bkey *);
1277
1278 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1279                              struct keylist *keylist, struct gc_stat *gc,
1280                              struct gc_merge_info *r)
1281 {
1282         unsigned i, nodes = 0, keys = 0, blocks;
1283         struct btree *new_nodes[GC_MERGE_NODES];
1284         struct closure cl;
1285         struct bkey *k;
1286
1287         memset(new_nodes, 0, sizeof(new_nodes));
1288         closure_init_stack(&cl);
1289
1290         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1291                 keys += r[nodes++].keys;
1292
1293         blocks = btree_default_blocks(b->c) * 2 / 3;
1294
1295         if (nodes < 2 ||
1296             __set_blocks(b->keys.set[0].data, keys,
1297                          block_bytes(b->c)) > blocks * (nodes - 1))
1298                 return 0;
1299
1300         for (i = 0; i < nodes; i++) {
1301                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1302                 if (IS_ERR_OR_NULL(new_nodes[i]))
1303                         goto out_nocoalesce;
1304         }
1305
1306         for (i = 0; i < nodes; i++)
1307                 mutex_lock(&new_nodes[i]->write_lock);
1308
1309         for (i = nodes - 1; i > 0; --i) {
1310                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1311                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1312                 struct bkey *k, *last = NULL;
1313
1314                 keys = 0;
1315
1316                 if (i > 1) {
1317                         for (k = n2->start;
1318                              k < bset_bkey_last(n2);
1319                              k = bkey_next(k)) {
1320                                 if (__set_blocks(n1, n1->keys + keys +
1321                                                  bkey_u64s(k),
1322                                                  block_bytes(b->c)) > blocks)
1323                                         break;
1324
1325                                 last = k;
1326                                 keys += bkey_u64s(k);
1327                         }
1328                 } else {
1329                         /*
1330                          * Last node we're not getting rid of - we're getting
1331                          * rid of the node at r[0]. Have to try and fit all of
1332                          * the remaining keys into this node; we can't ensure
1333                          * they will always fit due to rounding and variable
1334                          * length keys (shouldn't be possible in practice,
1335                          * though)
1336                          */
1337                         if (__set_blocks(n1, n1->keys + n2->keys,
1338                                          block_bytes(b->c)) >
1339                             btree_blocks(new_nodes[i]))
1340                                 goto out_nocoalesce;
1341
1342                         keys = n2->keys;
1343                         /* Take the key of the node we're getting rid of */
1344                         last = &r->b->key;
1345                 }
1346
1347                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1348                        btree_blocks(new_nodes[i]));
1349
1350                 if (last)
1351                         bkey_copy_key(&new_nodes[i]->key, last);
1352
1353                 memcpy(bset_bkey_last(n1),
1354                        n2->start,
1355                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1356
1357                 n1->keys += keys;
1358                 r[i].keys = n1->keys;
1359
1360                 memmove(n2->start,
1361                         bset_bkey_idx(n2, keys),
1362                         (void *) bset_bkey_last(n2) -
1363                         (void *) bset_bkey_idx(n2, keys));
1364
1365                 n2->keys -= keys;
1366
1367                 if (__bch_keylist_realloc(keylist,
1368                                           bkey_u64s(&new_nodes[i]->key)))
1369                         goto out_nocoalesce;
1370
1371                 bch_btree_node_write(new_nodes[i], &cl);
1372                 bch_keylist_add(keylist, &new_nodes[i]->key);
1373         }
1374
1375         for (i = 0; i < nodes; i++)
1376                 mutex_unlock(&new_nodes[i]->write_lock);
1377
1378         closure_sync(&cl);
1379
1380         /* We emptied out this node */
1381         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1382         btree_node_free(new_nodes[0]);
1383         rw_unlock(true, new_nodes[0]);
1384
1385         for (i = 0; i < nodes; i++) {
1386                 if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key)))
1387                         goto out_nocoalesce;
1388
1389                 make_btree_freeing_key(r[i].b, keylist->top);
1390                 bch_keylist_push(keylist);
1391         }
1392
1393         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1394         BUG_ON(!bch_keylist_empty(keylist));
1395
1396         for (i = 0; i < nodes; i++) {
1397                 btree_node_free(r[i].b);
1398                 rw_unlock(true, r[i].b);
1399
1400                 r[i].b = new_nodes[i];
1401         }
1402
1403         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1404         r[nodes - 1].b = ERR_PTR(-EINTR);
1405
1406         trace_bcache_btree_gc_coalesce(nodes);
1407         gc->nodes--;
1408
1409         /* Invalidated our iterator */
1410         return -EINTR;
1411
1412 out_nocoalesce:
1413         closure_sync(&cl);
1414
1415         while ((k = bch_keylist_pop(keylist)))
1416                 if (!bkey_cmp(k, &ZERO_KEY))
1417                         atomic_dec(&b->c->prio_blocked);
1418
1419         for (i = 0; i < nodes; i++)
1420                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1421                         btree_node_free(new_nodes[i]);
1422                         rw_unlock(true, new_nodes[i]);
1423                 }
1424         return 0;
1425 }
1426
1427 static unsigned btree_gc_count_keys(struct btree *b)
1428 {
1429         struct bkey *k;
1430         struct btree_iter iter;
1431         unsigned ret = 0;
1432
1433         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1434                 ret += bkey_u64s(k);
1435
1436         return ret;
1437 }
1438
1439 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1440                             struct closure *writes, struct gc_stat *gc)
1441 {
1442         int ret = 0;
1443         bool should_rewrite;
1444         struct btree *n;
1445         struct bkey *k;
1446         struct keylist keys;
1447         struct btree_iter iter;
1448         struct gc_merge_info r[GC_MERGE_NODES];
1449         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1450
1451         bch_keylist_init(&keys);
1452         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1453
1454         for (i = r; i < r + ARRAY_SIZE(r); i++)
1455                 i->b = ERR_PTR(-EINTR);
1456
1457         while (1) {
1458                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1459                 if (k) {
1460                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1461                         if (IS_ERR(r->b)) {
1462                                 ret = PTR_ERR(r->b);
1463                                 break;
1464                         }
1465
1466                         r->keys = btree_gc_count_keys(r->b);
1467
1468                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1469                         if (ret)
1470                                 break;
1471                 }
1472
1473                 if (!last->b)
1474                         break;
1475
1476                 if (!IS_ERR(last->b)) {
1477                         should_rewrite = btree_gc_mark_node(last->b, gc);
1478                         if (should_rewrite &&
1479                             !btree_check_reserve(b, NULL)) {
1480                                 n = btree_node_alloc_replacement(last->b,
1481                                                                  false);
1482
1483                                 if (!IS_ERR_OR_NULL(n)) {
1484                                         bch_btree_node_write_sync(n);
1485
1486                                         bch_keylist_add(&keys, &n->key);
1487
1488                                         make_btree_freeing_key(last->b,
1489                                                                keys.top);
1490                                         bch_keylist_push(&keys);
1491
1492                                         bch_btree_insert_node(b, op, &keys,
1493                                                               NULL, NULL);
1494                                         BUG_ON(!bch_keylist_empty(&keys));
1495
1496                                         btree_node_free(last->b);
1497                                         rw_unlock(true, last->b);
1498                                         last->b = n;
1499
1500                                         /* Invalidated our iterator */
1501                                         ret = -EINTR;
1502                                         break;
1503                                 }
1504                         }
1505
1506                         if (last->b->level) {
1507                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1508                                 if (ret)
1509                                         break;
1510                         }
1511
1512                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1513
1514                         /*
1515                          * Must flush leaf nodes before gc ends, since replace
1516                          * operations aren't journalled
1517                          */
1518                         mutex_lock(&last->b->write_lock);
1519                         if (btree_node_dirty(last->b))
1520                                 bch_btree_node_write(last->b, writes);
1521                         mutex_unlock(&last->b->write_lock);
1522                         rw_unlock(true, last->b);
1523                 }
1524
1525                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1526                 r->b = NULL;
1527
1528                 if (need_resched()) {
1529                         ret = -EAGAIN;
1530                         break;
1531                 }
1532         }
1533
1534         for (i = r; i < r + ARRAY_SIZE(r); i++)
1535                 if (!IS_ERR_OR_NULL(i->b)) {
1536                         mutex_lock(&i->b->write_lock);
1537                         if (btree_node_dirty(i->b))
1538                                 bch_btree_node_write(i->b, writes);
1539                         mutex_unlock(&i->b->write_lock);
1540                         rw_unlock(true, i->b);
1541                 }
1542
1543         bch_keylist_free(&keys);
1544
1545         return ret;
1546 }
1547
1548 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1549                              struct closure *writes, struct gc_stat *gc)
1550 {
1551         struct btree *n = NULL;
1552         int ret = 0;
1553         bool should_rewrite;
1554
1555         should_rewrite = btree_gc_mark_node(b, gc);
1556         if (should_rewrite) {
1557                 n = btree_node_alloc_replacement(b, false);
1558
1559                 if (!IS_ERR_OR_NULL(n)) {
1560                         bch_btree_node_write_sync(n);
1561
1562                         bch_btree_set_root(n);
1563                         btree_node_free(b);
1564                         rw_unlock(true, n);
1565
1566                         return -EINTR;
1567                 }
1568         }
1569
1570         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1571
1572         if (b->level) {
1573                 ret = btree_gc_recurse(b, op, writes, gc);
1574                 if (ret)
1575                         return ret;
1576         }
1577
1578         bkey_copy_key(&b->c->gc_done, &b->key);
1579
1580         return ret;
1581 }
1582
1583 static void btree_gc_start(struct cache_set *c)
1584 {
1585         struct cache *ca;
1586         struct bucket *b;
1587         unsigned i;
1588
1589         if (!c->gc_mark_valid)
1590                 return;
1591
1592         mutex_lock(&c->bucket_lock);
1593
1594         c->gc_mark_valid = 0;
1595         c->gc_done = ZERO_KEY;
1596
1597         for_each_cache(ca, c, i)
1598                 for_each_bucket(b, ca) {
1599                         b->gc_gen = b->gen;
1600                         if (!atomic_read(&b->pin)) {
1601                                 SET_GC_MARK(b, 0);
1602                                 SET_GC_SECTORS_USED(b, 0);
1603                         }
1604                 }
1605
1606         mutex_unlock(&c->bucket_lock);
1607 }
1608
1609 size_t bch_btree_gc_finish(struct cache_set *c)
1610 {
1611         size_t available = 0;
1612         struct bucket *b;
1613         struct cache *ca;
1614         unsigned i;
1615
1616         mutex_lock(&c->bucket_lock);
1617
1618         set_gc_sectors(c);
1619         c->gc_mark_valid = 1;
1620         c->need_gc      = 0;
1621
1622         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1623                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1624                             GC_MARK_METADATA);
1625
1626         /* don't reclaim buckets to which writeback keys point */
1627         rcu_read_lock();
1628         for (i = 0; i < c->nr_uuids; i++) {
1629                 struct bcache_device *d = c->devices[i];
1630                 struct cached_dev *dc;
1631                 struct keybuf_key *w, *n;
1632                 unsigned j;
1633
1634                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1635                         continue;
1636                 dc = container_of(d, struct cached_dev, disk);
1637
1638                 spin_lock(&dc->writeback_keys.lock);
1639                 rbtree_postorder_for_each_entry_safe(w, n,
1640                                         &dc->writeback_keys.keys, node)
1641                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1642                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1643                                             GC_MARK_DIRTY);
1644                 spin_unlock(&dc->writeback_keys.lock);
1645         }
1646         rcu_read_unlock();
1647
1648         for_each_cache(ca, c, i) {
1649                 uint64_t *i;
1650
1651                 ca->invalidate_needs_gc = 0;
1652
1653                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1654                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1655
1656                 for (i = ca->prio_buckets;
1657                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1658                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1659
1660                 for_each_bucket(b, ca) {
1661                         b->last_gc      = b->gc_gen;
1662                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1663
1664                         if (atomic_read(&b->pin))
1665                                 continue;
1666
1667                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1668
1669                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1670                                 available++;
1671
1672                         if (!GC_MARK(b))
1673                                 bch_bucket_add_unused(ca, b);
1674                 }
1675         }
1676
1677         mutex_unlock(&c->bucket_lock);
1678         return available;
1679 }
1680
1681 static void bch_btree_gc(struct cache_set *c)
1682 {
1683         int ret;
1684         unsigned long available;
1685         struct gc_stat stats;
1686         struct closure writes;
1687         struct btree_op op;
1688         uint64_t start_time = local_clock();
1689
1690         trace_bcache_gc_start(c);
1691
1692         memset(&stats, 0, sizeof(struct gc_stat));
1693         closure_init_stack(&writes);
1694         bch_btree_op_init(&op, SHRT_MAX);
1695
1696         btree_gc_start(c);
1697
1698         do {
1699                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1700                 closure_sync(&writes);
1701
1702                 if (ret && ret != -EAGAIN)
1703                         pr_warn("gc failed!");
1704         } while (ret);
1705
1706         available = bch_btree_gc_finish(c);
1707         wake_up_allocators(c);
1708
1709         bch_time_stats_update(&c->btree_gc_time, start_time);
1710
1711         stats.key_bytes *= sizeof(uint64_t);
1712         stats.data      <<= 9;
1713         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1714         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1715
1716         trace_bcache_gc_end(c);
1717
1718         bch_moving_gc(c);
1719 }
1720
1721 static int bch_gc_thread(void *arg)
1722 {
1723         struct cache_set *c = arg;
1724         struct cache *ca;
1725         unsigned i;
1726
1727         while (1) {
1728 again:
1729                 bch_btree_gc(c);
1730
1731                 set_current_state(TASK_INTERRUPTIBLE);
1732                 if (kthread_should_stop())
1733                         break;
1734
1735                 mutex_lock(&c->bucket_lock);
1736
1737                 for_each_cache(ca, c, i)
1738                         if (ca->invalidate_needs_gc) {
1739                                 mutex_unlock(&c->bucket_lock);
1740                                 set_current_state(TASK_RUNNING);
1741                                 goto again;
1742                         }
1743
1744                 mutex_unlock(&c->bucket_lock);
1745
1746                 try_to_freeze();
1747                 schedule();
1748         }
1749
1750         return 0;
1751 }
1752
1753 int bch_gc_thread_start(struct cache_set *c)
1754 {
1755         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1756         if (IS_ERR(c->gc_thread))
1757                 return PTR_ERR(c->gc_thread);
1758
1759         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1760         return 0;
1761 }
1762
1763 /* Initial partial gc */
1764
1765 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1766 {
1767         int ret = 0;
1768         struct bkey *k, *p = NULL;
1769         struct btree_iter iter;
1770
1771         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1772                 bch_initial_mark_key(b->c, b->level, k);
1773
1774         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1775
1776         if (b->level) {
1777                 bch_btree_iter_init(&b->keys, &iter, NULL);
1778
1779                 do {
1780                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1781                                                        bch_ptr_bad);
1782                         if (k)
1783                                 btree_node_prefetch(b->c, k, b->level - 1);
1784
1785                         if (p)
1786                                 ret = btree(check_recurse, p, b, op);
1787
1788                         p = k;
1789                 } while (p && !ret);
1790         }
1791
1792         return ret;
1793 }
1794
1795 int bch_btree_check(struct cache_set *c)
1796 {
1797         struct btree_op op;
1798
1799         bch_btree_op_init(&op, SHRT_MAX);
1800
1801         return btree_root(check_recurse, c, &op);
1802 }
1803
1804 /* Btree insertion */
1805
1806 static bool btree_insert_key(struct btree *b, struct bkey *k,
1807                              struct bkey *replace_key)
1808 {
1809         unsigned status;
1810
1811         BUG_ON(bkey_cmp(k, &b->key) > 0);
1812
1813         status = bch_btree_insert_key(&b->keys, k, replace_key);
1814         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1815                 bch_check_keys(&b->keys, "%u for %s", status,
1816                                replace_key ? "replace" : "insert");
1817
1818                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1819                                               status);
1820                 return true;
1821         } else
1822                 return false;
1823 }
1824
1825 static size_t insert_u64s_remaining(struct btree *b)
1826 {
1827         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1828
1829         /*
1830          * Might land in the middle of an existing extent and have to split it
1831          */
1832         if (b->keys.ops->is_extents)
1833                 ret -= KEY_MAX_U64S;
1834
1835         return max(ret, 0L);
1836 }
1837
1838 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1839                                   struct keylist *insert_keys,
1840                                   struct bkey *replace_key)
1841 {
1842         bool ret = false;
1843         int oldsize = bch_count_data(&b->keys);
1844
1845         while (!bch_keylist_empty(insert_keys)) {
1846                 struct bkey *k = insert_keys->keys;
1847
1848                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1849                         break;
1850
1851                 if (bkey_cmp(k, &b->key) <= 0) {
1852                         if (!b->level)
1853                                 bkey_put(b->c, k);
1854
1855                         ret |= btree_insert_key(b, k, replace_key);
1856                         bch_keylist_pop_front(insert_keys);
1857                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1858                         BKEY_PADDED(key) temp;
1859                         bkey_copy(&temp.key, insert_keys->keys);
1860
1861                         bch_cut_back(&b->key, &temp.key);
1862                         bch_cut_front(&b->key, insert_keys->keys);
1863
1864                         ret |= btree_insert_key(b, &temp.key, replace_key);
1865                         break;
1866                 } else {
1867                         break;
1868                 }
1869         }
1870
1871         if (!ret)
1872                 op->insert_collision = true;
1873
1874         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1875
1876         BUG_ON(bch_count_data(&b->keys) < oldsize);
1877         return ret;
1878 }
1879
1880 static int btree_split(struct btree *b, struct btree_op *op,
1881                        struct keylist *insert_keys,
1882                        struct bkey *replace_key)
1883 {
1884         bool split;
1885         struct btree *n1, *n2 = NULL, *n3 = NULL;
1886         uint64_t start_time = local_clock();
1887         struct closure cl;
1888         struct keylist parent_keys;
1889
1890         closure_init_stack(&cl);
1891         bch_keylist_init(&parent_keys);
1892
1893         if (!b->level &&
1894             btree_check_reserve(b, op))
1895                 return -EINTR;
1896
1897         n1 = btree_node_alloc_replacement(b, true);
1898         if (IS_ERR(n1))
1899                 goto err;
1900
1901         split = set_blocks(btree_bset_first(n1),
1902                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1903
1904         if (split) {
1905                 unsigned keys = 0;
1906
1907                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1908
1909                 n2 = bch_btree_node_alloc(b->c, b->level, true);
1910                 if (IS_ERR(n2))
1911                         goto err_free1;
1912
1913                 if (!b->parent) {
1914                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
1915                         if (IS_ERR(n3))
1916                                 goto err_free2;
1917                 }
1918
1919                 mutex_lock(&n1->write_lock);
1920                 mutex_lock(&n2->write_lock);
1921
1922                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1923
1924                 /*
1925                  * Has to be a linear search because we don't have an auxiliary
1926                  * search tree yet
1927                  */
1928
1929                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
1930                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
1931                                                         keys));
1932
1933                 bkey_copy_key(&n1->key,
1934                               bset_bkey_idx(btree_bset_first(n1), keys));
1935                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
1936
1937                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
1938                 btree_bset_first(n1)->keys = keys;
1939
1940                 memcpy(btree_bset_first(n2)->start,
1941                        bset_bkey_last(btree_bset_first(n1)),
1942                        btree_bset_first(n2)->keys * sizeof(uint64_t));
1943
1944                 bkey_copy_key(&n2->key, &b->key);
1945
1946                 bch_keylist_add(&parent_keys, &n2->key);
1947                 bch_btree_node_write(n2, &cl);
1948                 mutex_unlock(&n2->write_lock);
1949                 rw_unlock(true, n2);
1950         } else {
1951                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
1952
1953                 mutex_lock(&n1->write_lock);
1954                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1955         }
1956
1957         bch_keylist_add(&parent_keys, &n1->key);
1958         bch_btree_node_write(n1, &cl);
1959         mutex_unlock(&n1->write_lock);
1960
1961         if (n3) {
1962                 /* Depth increases, make a new root */
1963                 mutex_lock(&n3->write_lock);
1964                 bkey_copy_key(&n3->key, &MAX_KEY);
1965                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
1966                 bch_btree_node_write(n3, &cl);
1967                 mutex_unlock(&n3->write_lock);
1968
1969                 closure_sync(&cl);
1970                 bch_btree_set_root(n3);
1971                 rw_unlock(true, n3);
1972         } else if (!b->parent) {
1973                 /* Root filled up but didn't need to be split */
1974                 closure_sync(&cl);
1975                 bch_btree_set_root(n1);
1976         } else {
1977                 /* Split a non root node */
1978                 closure_sync(&cl);
1979                 make_btree_freeing_key(b, parent_keys.top);
1980                 bch_keylist_push(&parent_keys);
1981
1982                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
1983                 BUG_ON(!bch_keylist_empty(&parent_keys));
1984         }
1985
1986         btree_node_free(b);
1987         rw_unlock(true, n1);
1988
1989         bch_time_stats_update(&b->c->btree_split_time, start_time);
1990
1991         return 0;
1992 err_free2:
1993         bkey_put(b->c, &n2->key);
1994         btree_node_free(n2);
1995         rw_unlock(true, n2);
1996 err_free1:
1997         bkey_put(b->c, &n1->key);
1998         btree_node_free(n1);
1999         rw_unlock(true, n1);
2000 err:
2001         WARN(1, "bcache: btree split failed");
2002
2003         if (n3 == ERR_PTR(-EAGAIN) ||
2004             n2 == ERR_PTR(-EAGAIN) ||
2005             n1 == ERR_PTR(-EAGAIN))
2006                 return -EAGAIN;
2007
2008         return -ENOMEM;
2009 }
2010
2011 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2012                                  struct keylist *insert_keys,
2013                                  atomic_t *journal_ref,
2014                                  struct bkey *replace_key)
2015 {
2016         struct closure cl;
2017
2018         BUG_ON(b->level && replace_key);
2019
2020         closure_init_stack(&cl);
2021
2022         mutex_lock(&b->write_lock);
2023
2024         if (write_block(b) != btree_bset_last(b) &&
2025             b->keys.last_set_unwritten)
2026                 bch_btree_init_next(b); /* just wrote a set */
2027
2028         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2029                 mutex_unlock(&b->write_lock);
2030                 goto split;
2031         }
2032
2033         BUG_ON(write_block(b) != btree_bset_last(b));
2034
2035         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2036                 if (!b->level)
2037                         bch_btree_leaf_dirty(b, journal_ref);
2038                 else
2039                         bch_btree_node_write(b, &cl);
2040         }
2041
2042         mutex_unlock(&b->write_lock);
2043
2044         /* wait for btree node write if necessary, after unlock */
2045         closure_sync(&cl);
2046
2047         return 0;
2048 split:
2049         if (current->bio_list) {
2050                 op->lock = b->c->root->level + 1;
2051                 return -EAGAIN;
2052         } else if (op->lock <= b->c->root->level) {
2053                 op->lock = b->c->root->level + 1;
2054                 return -EINTR;
2055         } else {
2056                 /* Invalidated all iterators */
2057                 int ret = btree_split(b, op, insert_keys, replace_key);
2058
2059                 if (bch_keylist_empty(insert_keys))
2060                         return 0;
2061                 else if (!ret)
2062                         return -EINTR;
2063                 return ret;
2064         }
2065 }
2066
2067 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2068                                struct bkey *check_key)
2069 {
2070         int ret = -EINTR;
2071         uint64_t btree_ptr = b->key.ptr[0];
2072         unsigned long seq = b->seq;
2073         struct keylist insert;
2074         bool upgrade = op->lock == -1;
2075
2076         bch_keylist_init(&insert);
2077
2078         if (upgrade) {
2079                 rw_unlock(false, b);
2080                 rw_lock(true, b, b->level);
2081
2082                 if (b->key.ptr[0] != btree_ptr ||
2083                     b->seq != seq + 1)
2084                         goto out;
2085         }
2086
2087         SET_KEY_PTRS(check_key, 1);
2088         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2089
2090         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2091
2092         bch_keylist_add(&insert, check_key);
2093
2094         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2095
2096         BUG_ON(!ret && !bch_keylist_empty(&insert));
2097 out:
2098         if (upgrade)
2099                 downgrade_write(&b->lock);
2100         return ret;
2101 }
2102
2103 struct btree_insert_op {
2104         struct btree_op op;
2105         struct keylist  *keys;
2106         atomic_t        *journal_ref;
2107         struct bkey     *replace_key;
2108 };
2109
2110 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2111 {
2112         struct btree_insert_op *op = container_of(b_op,
2113                                         struct btree_insert_op, op);
2114
2115         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2116                                         op->journal_ref, op->replace_key);
2117         if (ret && !bch_keylist_empty(op->keys))
2118                 return ret;
2119         else
2120                 return MAP_DONE;
2121 }
2122
2123 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2124                      atomic_t *journal_ref, struct bkey *replace_key)
2125 {
2126         struct btree_insert_op op;
2127         int ret = 0;
2128
2129         BUG_ON(current->bio_list);
2130         BUG_ON(bch_keylist_empty(keys));
2131
2132         bch_btree_op_init(&op.op, 0);
2133         op.keys         = keys;
2134         op.journal_ref  = journal_ref;
2135         op.replace_key  = replace_key;
2136
2137         while (!ret && !bch_keylist_empty(keys)) {
2138                 op.op.lock = 0;
2139                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2140                                                &START_KEY(keys->keys),
2141                                                btree_insert_fn);
2142         }
2143
2144         if (ret) {
2145                 struct bkey *k;
2146
2147                 pr_err("error %i", ret);
2148
2149                 while ((k = bch_keylist_pop(keys)))
2150                         bkey_put(c, k);
2151         } else if (op.op.insert_collision)
2152                 ret = -ESRCH;
2153
2154         return ret;
2155 }
2156
2157 void bch_btree_set_root(struct btree *b)
2158 {
2159         unsigned i;
2160         struct closure cl;
2161
2162         closure_init_stack(&cl);
2163
2164         trace_bcache_btree_set_root(b);
2165
2166         BUG_ON(!b->written);
2167
2168         for (i = 0; i < KEY_PTRS(&b->key); i++)
2169                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2170
2171         mutex_lock(&b->c->bucket_lock);
2172         list_del_init(&b->list);
2173         mutex_unlock(&b->c->bucket_lock);
2174
2175         b->c->root = b;
2176
2177         bch_journal_meta(b->c, &cl);
2178         closure_sync(&cl);
2179 }
2180
2181 /* Map across nodes or keys */
2182
2183 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2184                                        struct bkey *from,
2185                                        btree_map_nodes_fn *fn, int flags)
2186 {
2187         int ret = MAP_CONTINUE;
2188
2189         if (b->level) {
2190                 struct bkey *k;
2191                 struct btree_iter iter;
2192
2193                 bch_btree_iter_init(&b->keys, &iter, from);
2194
2195                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2196                                                        bch_ptr_bad))) {
2197                         ret = btree(map_nodes_recurse, k, b,
2198                                     op, from, fn, flags);
2199                         from = NULL;
2200
2201                         if (ret != MAP_CONTINUE)
2202                                 return ret;
2203                 }
2204         }
2205
2206         if (!b->level || flags == MAP_ALL_NODES)
2207                 ret = fn(op, b);
2208
2209         return ret;
2210 }
2211
2212 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2213                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2214 {
2215         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2216 }
2217
2218 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2219                                       struct bkey *from, btree_map_keys_fn *fn,
2220                                       int flags)
2221 {
2222         int ret = MAP_CONTINUE;
2223         struct bkey *k;
2224         struct btree_iter iter;
2225
2226         bch_btree_iter_init(&b->keys, &iter, from);
2227
2228         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2229                 ret = !b->level
2230                         ? fn(op, b, k)
2231                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2232                 from = NULL;
2233
2234                 if (ret != MAP_CONTINUE)
2235                         return ret;
2236         }
2237
2238         if (!b->level && (flags & MAP_END_KEY))
2239                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2240                                      KEY_OFFSET(&b->key), 0));
2241
2242         return ret;
2243 }
2244
2245 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2246                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2247 {
2248         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2249 }
2250
2251 /* Keybuf code */
2252
2253 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2254 {
2255         /* Overlapping keys compare equal */
2256         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2257                 return -1;
2258         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2259                 return 1;
2260         return 0;
2261 }
2262
2263 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2264                                             struct keybuf_key *r)
2265 {
2266         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2267 }
2268
2269 struct refill {
2270         struct btree_op op;
2271         unsigned        nr_found;
2272         struct keybuf   *buf;
2273         struct bkey     *end;
2274         keybuf_pred_fn  *pred;
2275 };
2276
2277 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2278                             struct bkey *k)
2279 {
2280         struct refill *refill = container_of(op, struct refill, op);
2281         struct keybuf *buf = refill->buf;
2282         int ret = MAP_CONTINUE;
2283
2284         if (bkey_cmp(k, refill->end) >= 0) {
2285                 ret = MAP_DONE;
2286                 goto out;
2287         }
2288
2289         if (!KEY_SIZE(k)) /* end key */
2290                 goto out;
2291
2292         if (refill->pred(buf, k)) {
2293                 struct keybuf_key *w;
2294
2295                 spin_lock(&buf->lock);
2296
2297                 w = array_alloc(&buf->freelist);
2298                 if (!w) {
2299                         spin_unlock(&buf->lock);
2300                         return MAP_DONE;
2301                 }
2302
2303                 w->private = NULL;
2304                 bkey_copy(&w->key, k);
2305
2306                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2307                         array_free(&buf->freelist, w);
2308                 else
2309                         refill->nr_found++;
2310
2311                 if (array_freelist_empty(&buf->freelist))
2312                         ret = MAP_DONE;
2313
2314                 spin_unlock(&buf->lock);
2315         }
2316 out:
2317         buf->last_scanned = *k;
2318         return ret;
2319 }
2320
2321 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2322                        struct bkey *end, keybuf_pred_fn *pred)
2323 {
2324         struct bkey start = buf->last_scanned;
2325         struct refill refill;
2326
2327         cond_resched();
2328
2329         bch_btree_op_init(&refill.op, -1);
2330         refill.nr_found = 0;
2331         refill.buf      = buf;
2332         refill.end      = end;
2333         refill.pred     = pred;
2334
2335         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2336                            refill_keybuf_fn, MAP_END_KEY);
2337
2338         trace_bcache_keyscan(refill.nr_found,
2339                              KEY_INODE(&start), KEY_OFFSET(&start),
2340                              KEY_INODE(&buf->last_scanned),
2341                              KEY_OFFSET(&buf->last_scanned));
2342
2343         spin_lock(&buf->lock);
2344
2345         if (!RB_EMPTY_ROOT(&buf->keys)) {
2346                 struct keybuf_key *w;
2347                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2348                 buf->start      = START_KEY(&w->key);
2349
2350                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2351                 buf->end        = w->key;
2352         } else {
2353                 buf->start      = MAX_KEY;
2354                 buf->end        = MAX_KEY;
2355         }
2356
2357         spin_unlock(&buf->lock);
2358 }
2359
2360 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2361 {
2362         rb_erase(&w->node, &buf->keys);
2363         array_free(&buf->freelist, w);
2364 }
2365
2366 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2367 {
2368         spin_lock(&buf->lock);
2369         __bch_keybuf_del(buf, w);
2370         spin_unlock(&buf->lock);
2371 }
2372
2373 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2374                                   struct bkey *end)
2375 {
2376         bool ret = false;
2377         struct keybuf_key *p, *w, s;
2378         s.key = *start;
2379
2380         if (bkey_cmp(end, &buf->start) <= 0 ||
2381             bkey_cmp(start, &buf->end) >= 0)
2382                 return false;
2383
2384         spin_lock(&buf->lock);
2385         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2386
2387         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2388                 p = w;
2389                 w = RB_NEXT(w, node);
2390
2391                 if (p->private)
2392                         ret = true;
2393                 else
2394                         __bch_keybuf_del(buf, p);
2395         }
2396
2397         spin_unlock(&buf->lock);
2398         return ret;
2399 }
2400
2401 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2402 {
2403         struct keybuf_key *w;
2404         spin_lock(&buf->lock);
2405
2406         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2407
2408         while (w && w->private)
2409                 w = RB_NEXT(w, node);
2410
2411         if (w)
2412                 w->private = ERR_PTR(-EINTR);
2413
2414         spin_unlock(&buf->lock);
2415         return w;
2416 }
2417
2418 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2419                                           struct keybuf *buf,
2420                                           struct bkey *end,
2421                                           keybuf_pred_fn *pred)
2422 {
2423         struct keybuf_key *ret;
2424
2425         while (1) {
2426                 ret = bch_keybuf_next(buf);
2427                 if (ret)
2428                         break;
2429
2430                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2431                         pr_debug("scan finished");
2432                         break;
2433                 }
2434
2435                 bch_refill_keybuf(c, buf, end, pred);
2436         }
2437
2438         return ret;
2439 }
2440
2441 void bch_keybuf_init(struct keybuf *buf)
2442 {
2443         buf->last_scanned       = MAX_KEY;
2444         buf->keys               = RB_ROOT;
2445
2446         spin_lock_init(&buf->lock);
2447         array_allocator_init(&buf->freelist);
2448 }
2449
2450 void bch_btree_exit(void)
2451 {
2452         if (btree_io_wq)
2453                 destroy_workqueue(btree_io_wq);
2454 }
2455
2456 int __init bch_btree_init(void)
2457 {
2458         btree_io_wq = create_singlethread_workqueue("bch_btree_io");
2459         if (!btree_io_wq)
2460                 return -ENOMEM;
2461
2462         return 0;
2463 }