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  * Make sure all allocations get charged to the root cgroup
  72  *
  73  * Plugging?
  74  *
  75  * If data write is less than hard sector size of ssd, round up offset in open
  76  * bucket to the next whole sector
  77  *
  78  * Also lookup by cgroup in get_open_bucket()
  79  *
  80  * Superblock needs to be fleshed out for multiple cache devices
  81  *
  82  * Add a sysfs tunable for the number of writeback IOs in flight
  83  *
  84  * Add a sysfs tunable for the number of open data buckets
  85  *
  86  * IO tracking: Can we track when one process is doing io on behalf of another?
  87  * IO tracking: Don't use just an average, weigh more recent stuff higher
  88  *
  89  * Test module load/unload
  90  */
  91
  92 #define MAX_NEED_GC             64
  93 #define MAX_SAVE_PRIO           72
  94
  95 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  96
  97 #define PTR_HASH(c, k)                                                  \
  98         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  99
 100 static struct workqueue_struct *btree_io_wq;
 101
 102 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 103
 104 /*
 105  * These macros are for recursing down the btree - they handle the details of
 106  * locking and looking up nodes in the cache for you. They're best treated as
 107  * mere syntax when reading code that uses them.
 108  *
 109  * op->lock determines whether we take a read or a write lock at a given depth.
 110  * If you've got a read lock and find that you need a write lock (i.e. you're
 111  * going to have to split), set op->lock and return -EINTR; btree_root() will
 112  * call you again and you'll have the correct lock.
 113  */
 114
 115 /**
 116  * btree - recurse down the btree on a specified key
 117  * @fn:         function to call, which will be passed the child node
 118  * @key:        key to recurse on
 119  * @b:          parent btree node
 120  * @op:         pointer to struct btree_op
 121  */
 122 #define btree(fn, key, b, op, ...)                                      \
 123 ({                                                                      \
 124         int _r, l = (b)->level - 1;                                     \
 125         bool _w = l <= (op)->lock;                                      \
 126         struct btree *_child = bch_btree_node_get((b)->c, key, l, _w);  \
 127         if (!IS_ERR(_child)) {                                          \
 128                 _child->parent = (b);                                   \
 129                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 130                 rw_unlock(_w, _child);                                  \
 131         } else                                                          \
 132                 _r = PTR_ERR(_child);                                   \
 133         _r;                                                             \
 134 })
 135
 136 /**
 137  * btree_root - call a function on the root of the btree
 138  * @fn:         function to call, which will be passed the child node
 139  * @c:          cache set
 140  * @op:         pointer to struct btree_op
 141  */
 142 #define btree_root(fn, c, op, ...)                                      \
 143 ({                                                                      \
 144         int _r = -EINTR;                                                \
 145         do {                                                            \
 146                 struct btree *_b = (c)->root;                           \
 147                 bool _w = insert_lock(op, _b);                          \
 148                 rw_lock(_w, _b, _b->level);                             \
 149                 if (_b == (c)->root &&                                  \
 150                     _w == insert_lock(op, _b)) {                        \
 151                         _b->parent = NULL;                              \
 152                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 153                 }                                                       \
 154                 rw_unlock(_w, _b);                                      \
 155                 if (_r == -EINTR)                                       \
 156                         schedule();                                     \
 157                 bch_cannibalize_unlock(c);                              \
 158                 if (_r == -ENOSPC) {                                    \
 159                         wait_event((c)->try_wait,                       \
 160                                    !(c)->try_harder);                   \
 161                         _r = -EINTR;                                    \
 162                 }                                                       \
 163         } while (_r == -EINTR);                                         \
 164                                                                         \
 165         finish_wait(&(c)->bucket_wait, &(op)->wait);                    \
 166         _r;                                                             \
 167 })
 168
 169 static inline struct bset *write_block(struct btree *b)
 170 {
 171         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 172 }
 173
 174 /* Btree key manipulation */
 175
 176 void bkey_put(struct cache_set *c, struct bkey *k)
 177 {
 178         unsigned i;
 179
 180         for (i = 0; i < KEY_PTRS(k); i++)
 181                 if (ptr_available(c, k, i))
 182                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 183 }
 184
 185 /* Btree IO */
 186
 187 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 188 {
 189         uint64_t crc = b->key.ptr[0];
 190         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 191
 192         crc = bch_crc64_update(crc, data, end - data);
 193         return crc ^ 0xffffffffffffffffULL;
 194 }
 195
 196 void bch_btree_node_read_done(struct btree *b)
 197 {
 198         const char *err = "bad btree header";
 199         struct bset *i = btree_bset_first(b);
 200         struct btree_iter *iter;
 201
 202         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 203         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 204         iter->used = 0;
 205
 206 #ifdef CONFIG_BCACHE_DEBUG
 207         iter->b = &b->keys;
 208 #endif
 209
 210         if (!i->seq)
 211                 goto err;
 212
 213         for (;
 214              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 215              i = write_block(b)) {
 216                 err = "unsupported bset version";
 217                 if (i->version > BCACHE_BSET_VERSION)
 218                         goto err;
 219
 220                 err = "bad btree header";
 221                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 222                     btree_blocks(b))
 223                         goto err;
 224
 225                 err = "bad magic";
 226                 if (i->magic != bset_magic(&b->c->sb))
 227                         goto err;
 228
 229                 err = "bad checksum";
 230                 switch (i->version) {
 231                 case 0:
 232                         if (i->csum != csum_set(i))
 233                                 goto err;
 234                         break;
 235                 case BCACHE_BSET_VERSION:
 236                         if (i->csum != btree_csum_set(b, i))
 237                                 goto err;
 238                         break;
 239                 }
 240
 241                 err = "empty set";
 242                 if (i != b->keys.set[0].data && !i->keys)
 243                         goto err;
 244
 245                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 246
 247                 b->written += set_blocks(i, block_bytes(b->c));
 248         }
 249
 250         err = "corrupted btree";
 251         for (i = write_block(b);
 252              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 253              i = ((void *) i) + block_bytes(b->c))
 254                 if (i->seq == b->keys.set[0].data->seq)
 255                         goto err;
 256
 257         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 258
 259         i = b->keys.set[0].data;
 260         err = "short btree key";
 261         if (b->keys.set[0].size &&
 262             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 263                 goto err;
 264
 265         if (b->written < btree_blocks(b))
 266                 bch_bset_init_next(&b->keys, write_block(b),
 267                                    bset_magic(&b->c->sb));
 268 out:
 269         mempool_free(iter, b->c->fill_iter);
 270         return;
 271 err:
 272         set_btree_node_io_error(b);
 273         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 274                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 275                             bset_block_offset(b, i), i->keys);
 276         goto out;
 277 }
 278
 279 static void btree_node_read_endio(struct bio *bio, int error)
 280 {
 281         struct closure *cl = bio->bi_private;
 282         closure_put(cl);
 283 }
 284
 285 static void bch_btree_node_read(struct btree *b)
 286 {
 287         uint64_t start_time = local_clock();
 288         struct closure cl;
 289         struct bio *bio;
 290
 291         trace_bcache_btree_read(b);
 292
 293         closure_init_stack(&cl);
 294
 295         bio = bch_bbio_alloc(b->c);
 296         bio->bi_rw      = REQ_META|READ_SYNC;
 297         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 298         bio->bi_end_io  = btree_node_read_endio;
 299         bio->bi_private = &cl;
 300
 301         bch_bio_map(bio, b->keys.set[0].data);
 302
 303         bch_submit_bbio(bio, b->c, &b->key, 0);
 304         closure_sync(&cl);
 305
 306         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 307                 set_btree_node_io_error(b);
 308
 309         bch_bbio_free(bio, b->c);
 310
 311         if (btree_node_io_error(b))
 312                 goto err;
 313
 314         bch_btree_node_read_done(b);
 315         bch_time_stats_update(&b->c->btree_read_time, start_time);
 316
 317         return;
 318 err:
 319         bch_cache_set_error(b->c, "io error reading bucket %zu",
 320                             PTR_BUCKET_NR(b->c, &b->key, 0));
 321 }
 322
 323 static void btree_complete_write(struct btree *b, struct btree_write *w)
 324 {
 325         if (w->prio_blocked &&
 326             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 327                 wake_up_allocators(b->c);
 328
 329         if (w->journal) {
 330                 atomic_dec_bug(w->journal);
 331                 __closure_wake_up(&b->c->journal.wait);
 332         }
 333
 334         w->prio_blocked = 0;
 335         w->journal      = NULL;
 336 }
 337
 338 static void btree_node_write_unlock(struct closure *cl)
 339 {
 340         struct btree *b = container_of(cl, struct btree, io);
 341
 342         up(&b->io_mutex);
 343 }
 344
 345 static void __btree_node_write_done(struct closure *cl)
 346 {
 347         struct btree *b = container_of(cl, struct btree, io);
 348         struct btree_write *w = btree_prev_write(b);
 349
 350         bch_bbio_free(b->bio, b->c);
 351         b->bio = NULL;
 352         btree_complete_write(b, w);
 353
 354         if (btree_node_dirty(b))
 355                 queue_delayed_work(btree_io_wq, &b->work,
 356                                    msecs_to_jiffies(30000));
 357
 358         closure_return_with_destructor(cl, btree_node_write_unlock);
 359 }
 360
 361 static void btree_node_write_done(struct closure *cl)
 362 {
 363         struct btree *b = container_of(cl, struct btree, io);
 364         struct bio_vec *bv;
 365         int n;
 366
 367         bio_for_each_segment_all(bv, b->bio, n)
 368                 __free_page(bv->bv_page);
 369
 370         __btree_node_write_done(cl);
 371 }
 372
 373 static void btree_node_write_endio(struct bio *bio, int error)
 374 {
 375         struct closure *cl = bio->bi_private;
 376         struct btree *b = container_of(cl, struct btree, io);
 377
 378         if (error)
 379                 set_btree_node_io_error(b);
 380
 381         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 382         closure_put(cl);
 383 }
 384
 385 static void do_btree_node_write(struct btree *b)
 386 {
 387         struct closure *cl = &b->io;
 388         struct bset *i = btree_bset_last(b);
 389         BKEY_PADDED(key) k;
 390
 391         i->version      = BCACHE_BSET_VERSION;
 392         i->csum         = btree_csum_set(b, i);
 393
 394         BUG_ON(b->bio);
 395         b->bio = bch_bbio_alloc(b->c);
 396
 397         b->bio->bi_end_io       = btree_node_write_endio;
 398         b->bio->bi_private      = cl;
 399         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 400         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 401         bch_bio_map(b->bio, i);
 402
 403         /*
 404          * If we're appending to a leaf node, we don't technically need FUA -
 405          * this write just needs to be persisted before the next journal write,
 406          * which will be marked FLUSH|FUA.
 407          *
 408          * Similarly if we're writing a new btree root - the pointer is going to
 409          * be in the next journal entry.
 410          *
 411          * But if we're writing a new btree node (that isn't a root) or
 412          * appending to a non leaf btree node, we need either FUA or a flush
 413          * when we write the parent with the new pointer. FUA is cheaper than a
 414          * flush, and writes appending to leaf nodes aren't blocking anything so
 415          * just make all btree node writes FUA to keep things sane.
 416          */
 417
 418         bkey_copy(&k.key, &b->key);
 419         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 420                        bset_sector_offset(&b->keys, i));
 421
 422         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 423                 int j;
 424                 struct bio_vec *bv;
 425                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 426
 427                 bio_for_each_segment_all(bv, b->bio, j)
 428                         memcpy(page_address(bv->bv_page),
 429                                base + j * PAGE_SIZE, PAGE_SIZE);
 430
 431                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 432
 433                 continue_at(cl, btree_node_write_done, NULL);
 434         } else {
 435                 b->bio->bi_vcnt = 0;
 436                 bch_bio_map(b->bio, i);
 437
 438                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 439
 440                 closure_sync(cl);
 441                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 442         }
 443 }
 444
 445 void bch_btree_node_write(struct btree *b, struct closure *parent)
 446 {
 447         struct bset *i = btree_bset_last(b);
 448
 449         trace_bcache_btree_write(b);
 450
 451         BUG_ON(current->bio_list);
 452         BUG_ON(b->written >= btree_blocks(b));
 453         BUG_ON(b->written && !i->keys);
 454         BUG_ON(btree_bset_first(b)->seq != i->seq);
 455         bch_check_keys(&b->keys, "writing");
 456
 457         cancel_delayed_work(&b->work);
 458
 459         /* If caller isn't waiting for write, parent refcount is cache set */
 460         down(&b->io_mutex);
 461         closure_init(&b->io, parent ?: &b->c->cl);
 462
 463         clear_bit(BTREE_NODE_dirty,      &b->flags);
 464         change_bit(BTREE_NODE_write_idx, &b->flags);
 465
 466         do_btree_node_write(b);
 467
 468         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 469                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 470
 471         b->written += set_blocks(i, block_bytes(b->c));
 472
 473         /* If not a leaf node, always sort */
 474         if (b->level && b->keys.nsets)
 475                 bch_btree_sort(&b->keys, &b->c->sort);
 476         else
 477                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 478
 479         /*
 480          * do verify if there was more than one set initially (i.e. we did a
 481          * sort) and we sorted down to a single set:
 482          */
 483         if (i != b->keys.set->data && !b->keys.nsets)
 484                 bch_btree_verify(b);
 485
 486         if (b->written < btree_blocks(b))
 487                 bch_bset_init_next(&b->keys, write_block(b),
 488                                    bset_magic(&b->c->sb));
 489 }
 490
 491 static void bch_btree_node_write_sync(struct btree *b)
 492 {
 493         struct closure cl;
 494
 495         closure_init_stack(&cl);
 496         bch_btree_node_write(b, &cl);
 497         closure_sync(&cl);
 498 }
 499
 500 static void btree_node_write_work(struct work_struct *w)
 501 {
 502         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 503
 504         rw_lock(true, b, b->level);
 505
 506         if (btree_node_dirty(b))
 507                 bch_btree_node_write(b, NULL);
 508         rw_unlock(true, b);
 509 }
 510
 511 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 512 {
 513         struct bset *i = btree_bset_last(b);
 514         struct btree_write *w = btree_current_write(b);
 515
 516         BUG_ON(!b->written);
 517         BUG_ON(!i->keys);
 518
 519         if (!btree_node_dirty(b))
 520                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 521
 522         set_btree_node_dirty(b);
 523
 524         if (journal_ref) {
 525                 if (w->journal &&
 526                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 527                         atomic_dec_bug(w->journal);
 528                         w->journal = NULL;
 529                 }
 530
 531                 if (!w->journal) {
 532                         w->journal = journal_ref;
 533                         atomic_inc(w->journal);
 534                 }
 535         }
 536
 537         /* Force write if set is too big */
 538         if (set_bytes(i) > PAGE_SIZE - 48 &&
 539             !current->bio_list)
 540                 bch_btree_node_write(b, NULL);
 541 }
 542
 543 /*
 544  * Btree in memory cache - allocation/freeing
 545  * mca -> memory cache
 546  */
 547
 548 #define mca_reserve(c)  (((c->root && c->root->level)           \
 549                           ? c->root->level : 1) * 8 + 16)
 550 #define mca_can_free(c)                                         \
 551         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 552
 553 static void mca_data_free(struct btree *b)
 554 {
 555         BUG_ON(b->io_mutex.count != 1);
 556
 557         bch_btree_keys_free(&b->keys);
 558
 559         b->c->bucket_cache_used--;
 560         list_move(&b->list, &b->c->btree_cache_freed);
 561 }
 562
 563 static void mca_bucket_free(struct btree *b)
 564 {
 565         BUG_ON(btree_node_dirty(b));
 566
 567         b->key.ptr[0] = 0;
 568         hlist_del_init_rcu(&b->hash);
 569         list_move(&b->list, &b->c->btree_cache_freeable);
 570 }
 571
 572 static unsigned btree_order(struct bkey *k)
 573 {
 574         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 575 }
 576
 577 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 578 {
 579         if (!bch_btree_keys_alloc(&b->keys,
 580                                   max_t(unsigned,
 581                                         ilog2(b->c->btree_pages),
 582                                         btree_order(k)),
 583                                   gfp)) {
 584                 b->c->bucket_cache_used++;
 585                 list_move(&b->list, &b->c->btree_cache);
 586         } else {
 587                 list_move(&b->list, &b->c->btree_cache_freed);
 588         }
 589 }
 590
 591 static struct btree *mca_bucket_alloc(struct cache_set *c,
 592                                       struct bkey *k, gfp_t gfp)
 593 {
 594         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 595         if (!b)
 596                 return NULL;
 597
 598         init_rwsem(&b->lock);
 599         lockdep_set_novalidate_class(&b->lock);
 600         INIT_LIST_HEAD(&b->list);
 601         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 602         b->c = c;
 603         sema_init(&b->io_mutex, 1);
 604
 605         mca_data_alloc(b, k, gfp);
 606         return b;
 607 }
 608
 609 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 610 {
 611         struct closure cl;
 612
 613         closure_init_stack(&cl);
 614         lockdep_assert_held(&b->c->bucket_lock);
 615
 616         if (!down_write_trylock(&b->lock))
 617                 return -ENOMEM;
 618
 619         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 620
 621         if (b->keys.page_order < min_order)
 622                 goto out_unlock;
 623
 624         if (!flush) {
 625                 if (btree_node_dirty(b))
 626                         goto out_unlock;
 627
 628                 if (down_trylock(&b->io_mutex))
 629                         goto out_unlock;
 630                 up(&b->io_mutex);
 631         }
 632
 633         if (btree_node_dirty(b))
 634                 bch_btree_node_write_sync(b);
 635
 636         /* wait for any in flight btree write */
 637         down(&b->io_mutex);
 638         up(&b->io_mutex);
 639
 640         return 0;
 641 out_unlock:
 642         rw_unlock(true, b);
 643         return -ENOMEM;
 644 }
 645
 646 static unsigned long bch_mca_scan(struct shrinker *shrink,
 647                                   struct shrink_control *sc)
 648 {
 649         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 650         struct btree *b, *t;
 651         unsigned long i, nr = sc->nr_to_scan;
 652         unsigned long freed = 0;
 653
 654         if (c->shrinker_disabled)
 655                 return SHRINK_STOP;
 656
 657         if (c->try_harder)
 658                 return SHRINK_STOP;
 659
 660         /* Return -1 if we can't do anything right now */
 661         if (sc->gfp_mask & __GFP_IO)
 662                 mutex_lock(&c->bucket_lock);
 663         else if (!mutex_trylock(&c->bucket_lock))
 664                 return -1;
 665
 666         /*
 667          * It's _really_ critical that we don't free too many btree nodes - we
 668          * have to always leave ourselves a reserve. The reserve is how we
 669          * guarantee that allocating memory for a new btree node can always
 670          * succeed, so that inserting keys into the btree can always succeed and
 671          * IO can always make forward progress:
 672          */
 673         nr /= c->btree_pages;
 674         nr = min_t(unsigned long, nr, mca_can_free(c));
 675
 676         i = 0;
 677         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 678                 if (freed >= nr)
 679                         break;
 680
 681                 if (++i > 3 &&
 682                     !mca_reap(b, 0, false)) {
 683                         mca_data_free(b);
 684                         rw_unlock(true, b);
 685                         freed++;
 686                 }
 687         }
 688
 689         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 690                 if (list_empty(&c->btree_cache))
 691                         goto out;
 692
 693                 b = list_first_entry(&c->btree_cache, struct btree, list);
 694                 list_rotate_left(&c->btree_cache);
 695
 696                 if (!b->accessed &&
 697                     !mca_reap(b, 0, false)) {
 698                         mca_bucket_free(b);
 699                         mca_data_free(b);
 700                         rw_unlock(true, b);
 701                         freed++;
 702                 } else
 703                         b->accessed = 0;
 704         }
 705 out:
 706         mutex_unlock(&c->bucket_lock);
 707         return freed;
 708 }
 709
 710 static unsigned long bch_mca_count(struct shrinker *shrink,
 711                                    struct shrink_control *sc)
 712 {
 713         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 714
 715         if (c->shrinker_disabled)
 716                 return 0;
 717
 718         if (c->try_harder)
 719                 return 0;
 720
 721         return mca_can_free(c) * c->btree_pages;
 722 }
 723
 724 void bch_btree_cache_free(struct cache_set *c)
 725 {
 726         struct btree *b;
 727         struct closure cl;
 728         closure_init_stack(&cl);
 729
 730         if (c->shrink.list.next)
 731                 unregister_shrinker(&c->shrink);
 732
 733         mutex_lock(&c->bucket_lock);
 734
 735 #ifdef CONFIG_BCACHE_DEBUG
 736         if (c->verify_data)
 737                 list_move(&c->verify_data->list, &c->btree_cache);
 738
 739         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 740 #endif
 741
 742         list_splice(&c->btree_cache_freeable,
 743                     &c->btree_cache);
 744
 745         while (!list_empty(&c->btree_cache)) {
 746                 b = list_first_entry(&c->btree_cache, struct btree, list);
 747
 748                 if (btree_node_dirty(b))
 749                         btree_complete_write(b, btree_current_write(b));
 750                 clear_bit(BTREE_NODE_dirty, &b->flags);
 751
 752                 mca_data_free(b);
 753         }
 754
 755         while (!list_empty(&c->btree_cache_freed)) {
 756                 b = list_first_entry(&c->btree_cache_freed,
 757                                      struct btree, list);
 758                 list_del(&b->list);
 759                 cancel_delayed_work_sync(&b->work);
 760                 kfree(b);
 761         }
 762
 763         mutex_unlock(&c->bucket_lock);
 764 }
 765
 766 int bch_btree_cache_alloc(struct cache_set *c)
 767 {
 768         unsigned i;
 769
 770         for (i = 0; i < mca_reserve(c); i++)
 771                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 772                         return -ENOMEM;
 773
 774         list_splice_init(&c->btree_cache,
 775                          &c->btree_cache_freeable);
 776
 777 #ifdef CONFIG_BCACHE_DEBUG
 778         mutex_init(&c->verify_lock);
 779
 780         c->verify_ondisk = (void *)
 781                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 782
 783         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 784
 785         if (c->verify_data &&
 786             c->verify_data->keys.set->data)
 787                 list_del_init(&c->verify_data->list);
 788         else
 789                 c->verify_data = NULL;
 790 #endif
 791
 792         c->shrink.count_objects = bch_mca_count;
 793         c->shrink.scan_objects = bch_mca_scan;
 794         c->shrink.seeks = 4;
 795         c->shrink.batch = c->btree_pages * 2;
 796         register_shrinker(&c->shrink);
 797
 798         return 0;
 799 }
 800
 801 /* Btree in memory cache - hash table */
 802
 803 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 804 {
 805         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 806 }
 807
 808 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 809 {
 810         struct btree *b;
 811
 812         rcu_read_lock();
 813         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 814                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 815                         goto out;
 816         b = NULL;
 817 out:
 818         rcu_read_unlock();
 819         return b;
 820 }
 821
 822 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 823 {
 824         struct btree *b;
 825
 826         trace_bcache_btree_cache_cannibalize(c);
 827
 828         if (!c->try_harder) {
 829                 c->try_harder = current;
 830                 c->try_harder_start = local_clock();
 831         } else if (c->try_harder != current)
 832                 return ERR_PTR(-ENOSPC);
 833
 834         list_for_each_entry_reverse(b, &c->btree_cache, list)
 835                 if (!mca_reap(b, btree_order(k), false))
 836                         return b;
 837
 838         list_for_each_entry_reverse(b, &c->btree_cache, list)
 839                 if (!mca_reap(b, btree_order(k), true))
 840                         return b;
 841
 842         return ERR_PTR(-ENOMEM);
 843 }
 844
 845 /*
 846  * We can only have one thread cannibalizing other cached btree nodes at a time,
 847  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 848  * cannibalize_bucket() will take. This means every time we unlock the root of
 849  * the btree, we need to release this lock if we have it held.
 850  */
 851 static void bch_cannibalize_unlock(struct cache_set *c)
 852 {
 853         if (c->try_harder == current) {
 854                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 855                 c->try_harder = NULL;
 856                 wake_up(&c->try_wait);
 857         }
 858 }
 859
 860 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level)
 861 {
 862         struct btree *b;
 863
 864         BUG_ON(current->bio_list);
 865
 866         lockdep_assert_held(&c->bucket_lock);
 867
 868         if (mca_find(c, k))
 869                 return NULL;
 870
 871         /* btree_free() doesn't free memory; it sticks the node on the end of
 872          * the list. Check if there's any freed nodes there:
 873          */
 874         list_for_each_entry(b, &c->btree_cache_freeable, list)
 875                 if (!mca_reap(b, btree_order(k), false))
 876                         goto out;
 877
 878         /* We never free struct btree itself, just the memory that holds the on
 879          * disk node. Check the freed list before allocating a new one:
 880          */
 881         list_for_each_entry(b, &c->btree_cache_freed, list)
 882                 if (!mca_reap(b, 0, false)) {
 883                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 884                         if (!b->keys.set[0].data)
 885                                 goto err;
 886                         else
 887                                 goto out;
 888                 }
 889
 890         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 891         if (!b)
 892                 goto err;
 893
 894         BUG_ON(!down_write_trylock(&b->lock));
 895         if (!b->keys.set->data)
 896                 goto err;
 897 out:
 898         BUG_ON(b->io_mutex.count != 1);
 899
 900         bkey_copy(&b->key, k);
 901         list_move(&b->list, &c->btree_cache);
 902         hlist_del_init_rcu(&b->hash);
 903         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 904
 905         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 906         b->parent       = (void *) ~0UL;
 907         b->flags        = 0;
 908         b->written      = 0;
 909         b->level        = level;
 910
 911         if (!b->level)
 912                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 913                                     &b->c->expensive_debug_checks);
 914         else
 915                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 916                                     &b->c->expensive_debug_checks);
 917
 918         return b;
 919 err:
 920         if (b)
 921                 rw_unlock(true, b);
 922
 923         b = mca_cannibalize(c, k);
 924         if (!IS_ERR(b))
 925                 goto out;
 926
 927         return b;
 928 }
 929
 930 /**
 931  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 932  * in from disk if necessary.
 933  *
 934  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 935  *
 936  * The btree node will have either a read or a write lock held, depending on
 937  * level and op->lock.
 938  */
 939 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 940                                  int level, bool write)
 941 {
 942         int i = 0;
 943         struct btree *b;
 944
 945         BUG_ON(level < 0);
 946 retry:
 947         b = mca_find(c, k);
 948
 949         if (!b) {
 950                 if (current->bio_list)
 951                         return ERR_PTR(-EAGAIN);
 952
 953                 mutex_lock(&c->bucket_lock);
 954                 b = mca_alloc(c, k, level);
 955                 mutex_unlock(&c->bucket_lock);
 956
 957                 if (!b)
 958                         goto retry;
 959                 if (IS_ERR(b))
 960                         return b;
 961
 962                 bch_btree_node_read(b);
 963
 964                 if (!write)
 965                         downgrade_write(&b->lock);
 966         } else {
 967                 rw_lock(write, b, level);
 968                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 969                         rw_unlock(write, b);
 970                         goto retry;
 971                 }
 972                 BUG_ON(b->level != level);
 973         }
 974
 975         b->accessed = 1;
 976
 977         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
 978                 prefetch(b->keys.set[i].tree);
 979                 prefetch(b->keys.set[i].data);
 980         }
 981
 982         for (; i <= b->keys.nsets; i++)
 983                 prefetch(b->keys.set[i].data);
 984
 985         if (btree_node_io_error(b)) {
 986                 rw_unlock(write, b);
 987                 return ERR_PTR(-EIO);
 988         }
 989
 990         BUG_ON(!b->written);
 991
 992         return b;
 993 }
 994
 995 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
 996 {
 997         struct btree *b;
 998
 999         mutex_lock(&c->bucket_lock);
1000         b = mca_alloc(c, k, level);
1001         mutex_unlock(&c->bucket_lock);
1002
1003         if (!IS_ERR_OR_NULL(b)) {
1004                 bch_btree_node_read(b);
1005                 rw_unlock(true, b);
1006         }
1007 }
1008
1009 /* Btree alloc */
1010
1011 static void btree_node_free(struct btree *b)
1012 {
1013         unsigned i;
1014
1015         trace_bcache_btree_node_free(b);
1016
1017         BUG_ON(b == b->c->root);
1018
1019         if (btree_node_dirty(b))
1020                 btree_complete_write(b, btree_current_write(b));
1021         clear_bit(BTREE_NODE_dirty, &b->flags);
1022
1023         cancel_delayed_work(&b->work);
1024
1025         mutex_lock(&b->c->bucket_lock);
1026
1027         for (i = 0; i < KEY_PTRS(&b->key); i++) {
1028                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
1029
1030                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1031                             PTR_BUCKET(b->c, &b->key, i));
1032         }
1033
1034         bch_bucket_free(b->c, &b->key);
1035         mca_bucket_free(b);
1036         mutex_unlock(&b->c->bucket_lock);
1037 }
1038
1039 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1040 {
1041         BKEY_PADDED(key) k;
1042         struct btree *b = ERR_PTR(-EAGAIN);
1043
1044         mutex_lock(&c->bucket_lock);
1045 retry:
1046         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1047                 goto err;
1048
1049         bkey_put(c, &k.key);
1050         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1051
1052         b = mca_alloc(c, &k.key, level);
1053         if (IS_ERR(b))
1054                 goto err_free;
1055
1056         if (!b) {
1057                 cache_bug(c,
1058                         "Tried to allocate bucket that was in btree cache");
1059                 goto retry;
1060         }
1061
1062         b->accessed = 1;
1063         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1064
1065         mutex_unlock(&c->bucket_lock);
1066
1067         trace_bcache_btree_node_alloc(b);
1068         return b;
1069 err_free:
1070         bch_bucket_free(c, &k.key);
1071 err:
1072         mutex_unlock(&c->bucket_lock);
1073
1074         trace_bcache_btree_node_alloc_fail(b);
1075         return b;
1076 }
1077
1078 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1079 {
1080         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1081         if (!IS_ERR_OR_NULL(n)) {
1082                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1083                 bkey_copy_key(&n->key, &b->key);
1084         }
1085
1086         return n;
1087 }
1088
1089 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1090 {
1091         unsigned i;
1092
1093         bkey_copy(k, &b->key);
1094         bkey_copy_key(k, &ZERO_KEY);
1095
1096         for (i = 0; i < KEY_PTRS(k); i++) {
1097                 uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1;
1098
1099                 SET_PTR_GEN(k, i, g);
1100         }
1101
1102         atomic_inc(&b->c->prio_blocked);
1103 }
1104
1105 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1106 {
1107         struct cache_set *c = b->c;
1108         struct cache *ca;
1109         unsigned i, reserve = c->root->level * 2 + 1;
1110         int ret = 0;
1111
1112         mutex_lock(&c->bucket_lock);
1113
1114         for_each_cache(ca, c, i)
1115                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1116                         if (op)
1117                                 prepare_to_wait(&c->bucket_wait, &op->wait,
1118                                                 TASK_UNINTERRUPTIBLE);
1119                         ret = -EINTR;
1120                         break;
1121                 }
1122
1123         mutex_unlock(&c->bucket_lock);
1124         return ret;
1125 }
1126
1127 /* Garbage collection */
1128
1129 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1130                                     struct bkey *k)
1131 {
1132         uint8_t stale = 0;
1133         unsigned i;
1134         struct bucket *g;
1135
1136         /*
1137          * ptr_invalid() can't return true for the keys that mark btree nodes as
1138          * freed, but since ptr_bad() returns true we'll never actually use them
1139          * for anything and thus we don't want mark their pointers here
1140          */
1141         if (!bkey_cmp(k, &ZERO_KEY))
1142                 return stale;
1143
1144         for (i = 0; i < KEY_PTRS(k); i++) {
1145                 if (!ptr_available(c, k, i))
1146                         continue;
1147
1148                 g = PTR_BUCKET(c, k, i);
1149
1150                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1151                         g->gc_gen = PTR_GEN(k, i);
1152
1153                 if (ptr_stale(c, k, i)) {
1154                         stale = max(stale, ptr_stale(c, k, i));
1155                         continue;
1156                 }
1157
1158                 cache_bug_on(GC_MARK(g) &&
1159                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1160                              c, "inconsistent ptrs: mark = %llu, level = %i",
1161                              GC_MARK(g), level);
1162
1163                 if (level)
1164                         SET_GC_MARK(g, GC_MARK_METADATA);
1165                 else if (KEY_DIRTY(k))
1166                         SET_GC_MARK(g, GC_MARK_DIRTY);
1167
1168                 /* guard against overflow */
1169                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1170                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1171                                              MAX_GC_SECTORS_USED));
1172
1173                 BUG_ON(!GC_SECTORS_USED(g));
1174         }
1175
1176         return stale;
1177 }
1178
1179 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1180
1181 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1182 {
1183         unsigned i;
1184
1185         for (i = 0; i < KEY_PTRS(k); i++)
1186                 if (ptr_available(c, k, i) &&
1187                     !ptr_stale(c, k, i)) {
1188                         struct bucket *b = PTR_BUCKET(c, k, i);
1189
1190                         b->gen = PTR_GEN(k, i);
1191
1192                         if (level && bkey_cmp(k, &ZERO_KEY))
1193                                 b->prio = BTREE_PRIO;
1194                         else if (!level && b->prio == BTREE_PRIO)
1195                                 b->prio = INITIAL_PRIO;
1196                 }
1197
1198         __bch_btree_mark_key(c, level, k);
1199 }
1200
1201 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1202 {
1203         uint8_t stale = 0;
1204         unsigned keys = 0, good_keys = 0;
1205         struct bkey *k;
1206         struct btree_iter iter;
1207         struct bset_tree *t;
1208
1209         gc->nodes++;
1210
1211         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1212                 stale = max(stale, btree_mark_key(b, k));
1213                 keys++;
1214
1215                 if (bch_ptr_bad(&b->keys, k))
1216                         continue;
1217
1218                 gc->key_bytes += bkey_u64s(k);
1219                 gc->nkeys++;
1220                 good_keys++;
1221
1222                 gc->data += KEY_SIZE(k);
1223         }
1224
1225         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1226                 btree_bug_on(t->size &&
1227                              bset_written(&b->keys, t) &&
1228                              bkey_cmp(&b->key, &t->end) < 0,
1229                              b, "found short btree key in gc");
1230
1231         if (b->c->gc_always_rewrite)
1232                 return true;
1233
1234         if (stale > 10)
1235                 return true;
1236
1237         if ((keys - good_keys) * 2 > keys)
1238                 return true;
1239
1240         return false;
1241 }
1242
1243 #define GC_MERGE_NODES  4U
1244
1245 struct gc_merge_info {
1246         struct btree    *b;
1247         unsigned        keys;
1248 };
1249
1250 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1251                                  struct keylist *, atomic_t *, struct bkey *);
1252
1253 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1254                              struct keylist *keylist, struct gc_stat *gc,
1255                              struct gc_merge_info *r)
1256 {
1257         unsigned i, nodes = 0, keys = 0, blocks;
1258         struct btree *new_nodes[GC_MERGE_NODES];
1259         struct closure cl;
1260         struct bkey *k;
1261
1262         memset(new_nodes, 0, sizeof(new_nodes));
1263         closure_init_stack(&cl);
1264
1265         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1266                 keys += r[nodes++].keys;
1267
1268         blocks = btree_default_blocks(b->c) * 2 / 3;
1269
1270         if (nodes < 2 ||
1271             __set_blocks(b->keys.set[0].data, keys,
1272                          block_bytes(b->c)) > blocks * (nodes - 1))
1273                 return 0;
1274
1275         for (i = 0; i < nodes; i++) {
1276                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1277                 if (IS_ERR_OR_NULL(new_nodes[i]))
1278                         goto out_nocoalesce;
1279         }
1280
1281         for (i = nodes - 1; i > 0; --i) {
1282                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1283                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1284                 struct bkey *k, *last = NULL;
1285
1286                 keys = 0;
1287
1288                 if (i > 1) {
1289                         for (k = n2->start;
1290                              k < bset_bkey_last(n2);
1291                              k = bkey_next(k)) {
1292                                 if (__set_blocks(n1, n1->keys + keys +
1293                                                  bkey_u64s(k),
1294                                                  block_bytes(b->c)) > blocks)
1295                                         break;
1296
1297                                 last = k;
1298                                 keys += bkey_u64s(k);
1299                         }
1300                 } else {
1301                         /*
1302                          * Last node we're not getting rid of - we're getting
1303                          * rid of the node at r[0]. Have to try and fit all of
1304                          * the remaining keys into this node; we can't ensure
1305                          * they will always fit due to rounding and variable
1306                          * length keys (shouldn't be possible in practice,
1307                          * though)
1308                          */
1309                         if (__set_blocks(n1, n1->keys + n2->keys,
1310                                          block_bytes(b->c)) >
1311                             btree_blocks(new_nodes[i]))
1312                                 goto out_nocoalesce;
1313
1314                         keys = n2->keys;
1315                         /* Take the key of the node we're getting rid of */
1316                         last = &r->b->key;
1317                 }
1318
1319                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1320                        btree_blocks(new_nodes[i]));
1321
1322                 if (last)
1323                         bkey_copy_key(&new_nodes[i]->key, last);
1324
1325                 memcpy(bset_bkey_last(n1),
1326                        n2->start,
1327                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1328
1329                 n1->keys += keys;
1330                 r[i].keys = n1->keys;
1331
1332                 memmove(n2->start,
1333                         bset_bkey_idx(n2, keys),
1334                         (void *) bset_bkey_last(n2) -
1335                         (void *) bset_bkey_idx(n2, keys));
1336
1337                 n2->keys -= keys;
1338
1339                 if (__bch_keylist_realloc(keylist,
1340                                           bkey_u64s(&new_nodes[i]->key)))
1341                         goto out_nocoalesce;
1342
1343                 bch_btree_node_write(new_nodes[i], &cl);
1344                 bch_keylist_add(keylist, &new_nodes[i]->key);
1345         }
1346
1347         for (i = 0; i < nodes; i++) {
1348                 if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key)))
1349                         goto out_nocoalesce;
1350
1351                 make_btree_freeing_key(r[i].b, keylist->top);
1352                 bch_keylist_push(keylist);
1353         }
1354
1355         /* We emptied out this node */
1356         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1357         btree_node_free(new_nodes[0]);
1358         rw_unlock(true, new_nodes[0]);
1359
1360         closure_sync(&cl);
1361
1362         for (i = 0; i < nodes; i++) {
1363                 btree_node_free(r[i].b);
1364                 rw_unlock(true, r[i].b);
1365
1366                 r[i].b = new_nodes[i];
1367         }
1368
1369         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1370         BUG_ON(!bch_keylist_empty(keylist));
1371
1372         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1373         r[nodes - 1].b = ERR_PTR(-EINTR);
1374
1375         trace_bcache_btree_gc_coalesce(nodes);
1376         gc->nodes--;
1377
1378         /* Invalidated our iterator */
1379         return -EINTR;
1380
1381 out_nocoalesce:
1382         closure_sync(&cl);
1383
1384         while ((k = bch_keylist_pop(keylist)))
1385                 if (!bkey_cmp(k, &ZERO_KEY))
1386                         atomic_dec(&b->c->prio_blocked);
1387
1388         for (i = 0; i < nodes; i++)
1389                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1390                         btree_node_free(new_nodes[i]);
1391                         rw_unlock(true, new_nodes[i]);
1392                 }
1393         return 0;
1394 }
1395
1396 static unsigned btree_gc_count_keys(struct btree *b)
1397 {
1398         struct bkey *k;
1399         struct btree_iter iter;
1400         unsigned ret = 0;
1401
1402         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1403                 ret += bkey_u64s(k);
1404
1405         return ret;
1406 }
1407
1408 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1409                             struct closure *writes, struct gc_stat *gc)
1410 {
1411         unsigned i;
1412         int ret = 0;
1413         bool should_rewrite;
1414         struct btree *n;
1415         struct bkey *k;
1416         struct keylist keys;
1417         struct btree_iter iter;
1418         struct gc_merge_info r[GC_MERGE_NODES];
1419         struct gc_merge_info *last = r + GC_MERGE_NODES - 1;
1420
1421         bch_keylist_init(&keys);
1422         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1423
1424         for (i = 0; i < GC_MERGE_NODES; i++)
1425                 r[i].b = ERR_PTR(-EINTR);
1426
1427         while (1) {
1428                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1429                 if (k) {
1430                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1431                         if (IS_ERR(r->b)) {
1432                                 ret = PTR_ERR(r->b);
1433                                 break;
1434                         }
1435
1436                         r->keys = btree_gc_count_keys(r->b);
1437
1438                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1439                         if (ret)
1440                                 break;
1441                 }
1442
1443                 if (!last->b)
1444                         break;
1445
1446                 if (!IS_ERR(last->b)) {
1447                         should_rewrite = btree_gc_mark_node(last->b, gc);
1448                         if (should_rewrite &&
1449                             !btree_check_reserve(b, NULL)) {
1450                                 n = btree_node_alloc_replacement(last->b,
1451                                                                  false);
1452
1453                                 if (!IS_ERR_OR_NULL(n)) {
1454                                         bch_btree_node_write_sync(n);
1455                                         bch_keylist_add(&keys, &n->key);
1456
1457                                         make_btree_freeing_key(last->b,
1458                                                                keys.top);
1459                                         bch_keylist_push(&keys);
1460
1461                                         btree_node_free(last->b);
1462
1463                                         bch_btree_insert_node(b, op, &keys,
1464                                                               NULL, NULL);
1465                                         BUG_ON(!bch_keylist_empty(&keys));
1466
1467                                         rw_unlock(true, last->b);
1468                                         last->b = n;
1469
1470                                         /* Invalidated our iterator */
1471                                         ret = -EINTR;
1472                                         break;
1473                                 }
1474                         }
1475
1476                         if (last->b->level) {
1477                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1478                                 if (ret)
1479                                         break;
1480                         }
1481
1482                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1483
1484                         /*
1485                          * Must flush leaf nodes before gc ends, since replace
1486                          * operations aren't journalled
1487                          */
1488                         if (btree_node_dirty(last->b))
1489                                 bch_btree_node_write(last->b, writes);
1490                         rw_unlock(true, last->b);
1491                 }
1492
1493                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1494                 r->b = NULL;
1495
1496                 if (need_resched()) {
1497                         ret = -EAGAIN;
1498                         break;
1499                 }
1500         }
1501
1502         for (i = 0; i < GC_MERGE_NODES; i++)
1503                 if (!IS_ERR_OR_NULL(r[i].b)) {
1504                         if (btree_node_dirty(r[i].b))
1505                                 bch_btree_node_write(r[i].b, writes);
1506                         rw_unlock(true, r[i].b);
1507                 }
1508
1509         bch_keylist_free(&keys);
1510
1511         return ret;
1512 }
1513
1514 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1515                              struct closure *writes, struct gc_stat *gc)
1516 {
1517         struct btree *n = NULL;
1518         int ret = 0;
1519         bool should_rewrite;
1520
1521         should_rewrite = btree_gc_mark_node(b, gc);
1522         if (should_rewrite) {
1523                 n = btree_node_alloc_replacement(b, false);
1524
1525                 if (!IS_ERR_OR_NULL(n)) {
1526                         bch_btree_node_write_sync(n);
1527                         bch_btree_set_root(n);
1528                         btree_node_free(b);
1529                         rw_unlock(true, n);
1530
1531                         return -EINTR;
1532                 }
1533         }
1534
1535         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1536
1537         if (b->level) {
1538                 ret = btree_gc_recurse(b, op, writes, gc);
1539                 if (ret)
1540                         return ret;
1541         }
1542
1543         bkey_copy_key(&b->c->gc_done, &b->key);
1544
1545         return ret;
1546 }
1547
1548 static void btree_gc_start(struct cache_set *c)
1549 {
1550         struct cache *ca;
1551         struct bucket *b;
1552         unsigned i;
1553
1554         if (!c->gc_mark_valid)
1555                 return;
1556
1557         mutex_lock(&c->bucket_lock);
1558
1559         c->gc_mark_valid = 0;
1560         c->gc_done = ZERO_KEY;
1561
1562         for_each_cache(ca, c, i)
1563                 for_each_bucket(b, ca) {
1564                         b->gc_gen = b->gen;
1565                         if (!atomic_read(&b->pin)) {
1566                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1567                                 SET_GC_SECTORS_USED(b, 0);
1568                         }
1569                 }
1570
1571         mutex_unlock(&c->bucket_lock);
1572 }
1573
1574 size_t bch_btree_gc_finish(struct cache_set *c)
1575 {
1576         size_t available = 0;
1577         struct bucket *b;
1578         struct cache *ca;
1579         unsigned i;
1580
1581         mutex_lock(&c->bucket_lock);
1582
1583         set_gc_sectors(c);
1584         c->gc_mark_valid = 1;
1585         c->need_gc      = 0;
1586
1587         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1588                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1589                             GC_MARK_METADATA);
1590
1591         /* don't reclaim buckets to which writeback keys point */
1592         rcu_read_lock();
1593         for (i = 0; i < c->nr_uuids; i++) {
1594                 struct bcache_device *d = c->devices[i];
1595                 struct cached_dev *dc;
1596                 struct keybuf_key *w, *n;
1597                 unsigned j;
1598
1599                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1600                         continue;
1601                 dc = container_of(d, struct cached_dev, disk);
1602
1603                 spin_lock(&dc->writeback_keys.lock);
1604                 rbtree_postorder_for_each_entry_safe(w, n,
1605                                         &dc->writeback_keys.keys, node)
1606                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1607                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1608                                             GC_MARK_DIRTY);
1609                 spin_unlock(&dc->writeback_keys.lock);
1610         }
1611         rcu_read_unlock();
1612
1613         for_each_cache(ca, c, i) {
1614                 uint64_t *i;
1615
1616                 ca->invalidate_needs_gc = 0;
1617
1618                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1619                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1620
1621                 for (i = ca->prio_buckets;
1622                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1623                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1624
1625                 for_each_bucket(b, ca) {
1626                         b->last_gc      = b->gc_gen;
1627                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1628
1629                         if (!atomic_read(&b->pin) &&
1630                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1631                                 available++;
1632                                 if (!GC_SECTORS_USED(b))
1633                                         bch_bucket_add_unused(ca, b);
1634                         }
1635                 }
1636         }
1637
1638         mutex_unlock(&c->bucket_lock);
1639         return available;
1640 }
1641
1642 static void bch_btree_gc(struct cache_set *c)
1643 {
1644         int ret;
1645         unsigned long available;
1646         struct gc_stat stats;
1647         struct closure writes;
1648         struct btree_op op;
1649         uint64_t start_time = local_clock();
1650
1651         trace_bcache_gc_start(c);
1652
1653         memset(&stats, 0, sizeof(struct gc_stat));
1654         closure_init_stack(&writes);
1655         bch_btree_op_init(&op, SHRT_MAX);
1656
1657         btree_gc_start(c);
1658
1659         do {
1660                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1661                 closure_sync(&writes);
1662
1663                 if (ret && ret != -EAGAIN)
1664                         pr_warn("gc failed!");
1665         } while (ret);
1666
1667         available = bch_btree_gc_finish(c);
1668         wake_up_allocators(c);
1669
1670         bch_time_stats_update(&c->btree_gc_time, start_time);
1671
1672         stats.key_bytes *= sizeof(uint64_t);
1673         stats.data      <<= 9;
1674         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1675         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1676
1677         trace_bcache_gc_end(c);
1678
1679         bch_moving_gc(c);
1680 }
1681
1682 static int bch_gc_thread(void *arg)
1683 {
1684         struct cache_set *c = arg;
1685         struct cache *ca;
1686         unsigned i;
1687
1688         while (1) {
1689 again:
1690                 bch_btree_gc(c);
1691
1692                 set_current_state(TASK_INTERRUPTIBLE);
1693                 if (kthread_should_stop())
1694                         break;
1695
1696                 mutex_lock(&c->bucket_lock);
1697
1698                 for_each_cache(ca, c, i)
1699                         if (ca->invalidate_needs_gc) {
1700                                 mutex_unlock(&c->bucket_lock);
1701                                 set_current_state(TASK_RUNNING);
1702                                 goto again;
1703                         }
1704
1705                 mutex_unlock(&c->bucket_lock);
1706
1707                 try_to_freeze();
1708                 schedule();
1709         }
1710
1711         return 0;
1712 }
1713
1714 int bch_gc_thread_start(struct cache_set *c)
1715 {
1716         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1717         if (IS_ERR(c->gc_thread))
1718                 return PTR_ERR(c->gc_thread);
1719
1720         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1721         return 0;
1722 }
1723
1724 /* Initial partial gc */
1725
1726 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1727 {
1728         int ret = 0;
1729         struct bkey *k, *p = NULL;
1730         struct btree_iter iter;
1731
1732         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1733                 bch_initial_mark_key(b->c, b->level, k);
1734
1735         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1736
1737         if (b->level) {
1738                 bch_btree_iter_init(&b->keys, &iter, NULL);
1739
1740                 do {
1741                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1742                                                        bch_ptr_bad);
1743                         if (k)
1744                                 btree_node_prefetch(b->c, k, b->level - 1);
1745
1746                         if (p)
1747                                 ret = btree(check_recurse, p, b, op);
1748
1749                         p = k;
1750                 } while (p && !ret);
1751         }
1752
1753         return ret;
1754 }
1755
1756 int bch_btree_check(struct cache_set *c)
1757 {
1758         struct btree_op op;
1759
1760         bch_btree_op_init(&op, SHRT_MAX);
1761
1762         return btree_root(check_recurse, c, &op);
1763 }
1764
1765 /* Btree insertion */
1766
1767 static bool btree_insert_key(struct btree *b, struct bkey *k,
1768                              struct bkey *replace_key)
1769 {
1770         unsigned status;
1771
1772         BUG_ON(bkey_cmp(k, &b->key) > 0);
1773
1774         status = bch_btree_insert_key(&b->keys, k, replace_key);
1775         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1776                 bch_check_keys(&b->keys, "%u for %s", status,
1777                                replace_key ? "replace" : "insert");
1778
1779                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1780                                               status);
1781                 return true;
1782         } else
1783                 return false;
1784 }
1785
1786 static size_t insert_u64s_remaining(struct btree *b)
1787 {
1788         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1789
1790         /*
1791          * Might land in the middle of an existing extent and have to split it
1792          */
1793         if (b->keys.ops->is_extents)
1794                 ret -= KEY_MAX_U64S;
1795
1796         return max(ret, 0L);
1797 }
1798
1799 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1800                                   struct keylist *insert_keys,
1801                                   struct bkey *replace_key)
1802 {
1803         bool ret = false;
1804         int oldsize = bch_count_data(&b->keys);
1805
1806         while (!bch_keylist_empty(insert_keys)) {
1807                 struct bkey *k = insert_keys->keys;
1808
1809                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1810                         break;
1811
1812                 if (bkey_cmp(k, &b->key) <= 0) {
1813                         if (!b->level)
1814                                 bkey_put(b->c, k);
1815
1816                         ret |= btree_insert_key(b, k, replace_key);
1817                         bch_keylist_pop_front(insert_keys);
1818                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1819                         BKEY_PADDED(key) temp;
1820                         bkey_copy(&temp.key, insert_keys->keys);
1821
1822                         bch_cut_back(&b->key, &temp.key);
1823                         bch_cut_front(&b->key, insert_keys->keys);
1824
1825                         ret |= btree_insert_key(b, &temp.key, replace_key);
1826                         break;
1827                 } else {
1828                         break;
1829                 }
1830         }
1831
1832         if (!ret)
1833                 op->insert_collision = true;
1834
1835         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1836
1837         BUG_ON(bch_count_data(&b->keys) < oldsize);
1838         return ret;
1839 }
1840
1841 static int btree_split(struct btree *b, struct btree_op *op,
1842                        struct keylist *insert_keys,
1843                        struct bkey *replace_key)
1844 {
1845         bool split;
1846         struct btree *n1, *n2 = NULL, *n3 = NULL;
1847         uint64_t start_time = local_clock();
1848         struct closure cl;
1849         struct keylist parent_keys;
1850
1851         closure_init_stack(&cl);
1852         bch_keylist_init(&parent_keys);
1853
1854         if (!b->level &&
1855             btree_check_reserve(b, op))
1856                 return -EINTR;
1857
1858         n1 = btree_node_alloc_replacement(b, true);
1859         if (IS_ERR(n1))
1860                 goto err;
1861
1862         split = set_blocks(btree_bset_first(n1),
1863                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1864
1865         if (split) {
1866                 unsigned keys = 0;
1867
1868                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1869
1870                 n2 = bch_btree_node_alloc(b->c, b->level, true);
1871                 if (IS_ERR(n2))
1872                         goto err_free1;
1873
1874                 if (!b->parent) {
1875                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
1876                         if (IS_ERR(n3))
1877                                 goto err_free2;
1878                 }
1879
1880                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1881
1882                 /*
1883                  * Has to be a linear search because we don't have an auxiliary
1884                  * search tree yet
1885                  */
1886
1887                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
1888                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
1889                                                         keys));
1890
1891                 bkey_copy_key(&n1->key,
1892                               bset_bkey_idx(btree_bset_first(n1), keys));
1893                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
1894
1895                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
1896                 btree_bset_first(n1)->keys = keys;
1897
1898                 memcpy(btree_bset_first(n2)->start,
1899                        bset_bkey_last(btree_bset_first(n1)),
1900                        btree_bset_first(n2)->keys * sizeof(uint64_t));
1901
1902                 bkey_copy_key(&n2->key, &b->key);
1903
1904                 bch_keylist_add(&parent_keys, &n2->key);
1905                 bch_btree_node_write(n2, &cl);
1906                 rw_unlock(true, n2);
1907         } else {
1908                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
1909
1910                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1911         }
1912
1913         bch_keylist_add(&parent_keys, &n1->key);
1914         bch_btree_node_write(n1, &cl);
1915
1916         if (n3) {
1917                 /* Depth increases, make a new root */
1918                 bkey_copy_key(&n3->key, &MAX_KEY);
1919                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
1920                 bch_btree_node_write(n3, &cl);
1921
1922                 closure_sync(&cl);
1923                 bch_btree_set_root(n3);
1924                 rw_unlock(true, n3);
1925
1926                 btree_node_free(b);
1927         } else if (!b->parent) {
1928                 /* Root filled up but didn't need to be split */
1929                 closure_sync(&cl);
1930                 bch_btree_set_root(n1);
1931
1932                 btree_node_free(b);
1933         } else {
1934                 /* Split a non root node */
1935                 closure_sync(&cl);
1936                 make_btree_freeing_key(b, parent_keys.top);
1937                 bch_keylist_push(&parent_keys);
1938
1939                 btree_node_free(b);
1940
1941                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
1942                 BUG_ON(!bch_keylist_empty(&parent_keys));
1943         }
1944
1945         rw_unlock(true, n1);
1946
1947         bch_time_stats_update(&b->c->btree_split_time, start_time);
1948
1949         return 0;
1950 err_free2:
1951         bkey_put(b->c, &n2->key);
1952         btree_node_free(n2);
1953         rw_unlock(true, n2);
1954 err_free1:
1955         bkey_put(b->c, &n1->key);
1956         btree_node_free(n1);
1957         rw_unlock(true, n1);
1958 err:
1959         WARN(1, "bcache: btree split failed");
1960
1961         if (n3 == ERR_PTR(-EAGAIN) ||
1962             n2 == ERR_PTR(-EAGAIN) ||
1963             n1 == ERR_PTR(-EAGAIN))
1964                 return -EAGAIN;
1965
1966         return -ENOMEM;
1967 }
1968
1969 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1970                                  struct keylist *insert_keys,
1971                                  atomic_t *journal_ref,
1972                                  struct bkey *replace_key)
1973 {
1974         BUG_ON(b->level && replace_key);
1975
1976         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
1977                 if (current->bio_list) {
1978                         op->lock = b->c->root->level + 1;
1979                         return -EAGAIN;
1980                 } else if (op->lock <= b->c->root->level) {
1981                         op->lock = b->c->root->level + 1;
1982                         return -EINTR;
1983                 } else {
1984                         /* Invalidated all iterators */
1985                         int ret = btree_split(b, op, insert_keys, replace_key);
1986
1987                         return bch_keylist_empty(insert_keys) ?
1988                                 0 : ret ?: -EINTR;
1989                 }
1990         } else {
1991                 BUG_ON(write_block(b) != btree_bset_last(b));
1992
1993                 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
1994                         if (!b->level)
1995                                 bch_btree_leaf_dirty(b, journal_ref);
1996                         else
1997                                 bch_btree_node_write_sync(b);
1998                 }
1999
2000                 return 0;
2001         }
2002 }
2003
2004 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2005                                struct bkey *check_key)
2006 {
2007         int ret = -EINTR;
2008         uint64_t btree_ptr = b->key.ptr[0];
2009         unsigned long seq = b->seq;
2010         struct keylist insert;
2011         bool upgrade = op->lock == -1;
2012
2013         bch_keylist_init(&insert);
2014
2015         if (upgrade) {
2016                 rw_unlock(false, b);
2017                 rw_lock(true, b, b->level);
2018
2019                 if (b->key.ptr[0] != btree_ptr ||
2020                     b->seq != seq + 1)
2021                         goto out;
2022         }
2023
2024         SET_KEY_PTRS(check_key, 1);
2025         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2026
2027         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2028
2029         bch_keylist_add(&insert, check_key);
2030
2031         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2032
2033         BUG_ON(!ret && !bch_keylist_empty(&insert));
2034 out:
2035         if (upgrade)
2036                 downgrade_write(&b->lock);
2037         return ret;
2038 }
2039
2040 struct btree_insert_op {
2041         struct btree_op op;
2042         struct keylist  *keys;
2043         atomic_t        *journal_ref;
2044         struct bkey     *replace_key;
2045 };
2046
2047 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2048 {
2049         struct btree_insert_op *op = container_of(b_op,
2050                                         struct btree_insert_op, op);
2051
2052         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2053                                         op->journal_ref, op->replace_key);
2054         if (ret && !bch_keylist_empty(op->keys))
2055                 return ret;
2056         else
2057                 return MAP_DONE;
2058 }
2059
2060 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2061                      atomic_t *journal_ref, struct bkey *replace_key)
2062 {
2063         struct btree_insert_op op;
2064         int ret = 0;
2065
2066         BUG_ON(current->bio_list);
2067         BUG_ON(bch_keylist_empty(keys));
2068
2069         bch_btree_op_init(&op.op, 0);
2070         op.keys         = keys;
2071         op.journal_ref  = journal_ref;
2072         op.replace_key  = replace_key;
2073
2074         while (!ret && !bch_keylist_empty(keys)) {
2075                 op.op.lock = 0;
2076                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2077                                                &START_KEY(keys->keys),
2078                                                btree_insert_fn);
2079         }
2080
2081         if (ret) {
2082                 struct bkey *k;
2083
2084                 pr_err("error %i", ret);
2085
2086                 while ((k = bch_keylist_pop(keys)))
2087                         bkey_put(c, k);
2088         } else if (op.op.insert_collision)
2089                 ret = -ESRCH;
2090
2091         return ret;
2092 }
2093
2094 void bch_btree_set_root(struct btree *b)
2095 {
2096         unsigned i;
2097         struct closure cl;
2098
2099         closure_init_stack(&cl);
2100
2101         trace_bcache_btree_set_root(b);
2102
2103         BUG_ON(!b->written);
2104
2105         for (i = 0; i < KEY_PTRS(&b->key); i++)
2106                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2107
2108         mutex_lock(&b->c->bucket_lock);
2109         list_del_init(&b->list);
2110         mutex_unlock(&b->c->bucket_lock);
2111
2112         b->c->root = b;
2113
2114         bch_journal_meta(b->c, &cl);
2115         closure_sync(&cl);
2116 }
2117
2118 /* Map across nodes or keys */
2119
2120 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2121                                        struct bkey *from,
2122                                        btree_map_nodes_fn *fn, int flags)
2123 {
2124         int ret = MAP_CONTINUE;
2125
2126         if (b->level) {
2127                 struct bkey *k;
2128                 struct btree_iter iter;
2129
2130                 bch_btree_iter_init(&b->keys, &iter, from);
2131
2132                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2133                                                        bch_ptr_bad))) {
2134                         ret = btree(map_nodes_recurse, k, b,
2135                                     op, from, fn, flags);
2136                         from = NULL;
2137
2138                         if (ret != MAP_CONTINUE)
2139                                 return ret;
2140                 }
2141         }
2142
2143         if (!b->level || flags == MAP_ALL_NODES)
2144                 ret = fn(op, b);
2145
2146         return ret;
2147 }
2148
2149 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2150                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2151 {
2152         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2153 }
2154
2155 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2156                                       struct bkey *from, btree_map_keys_fn *fn,
2157                                       int flags)
2158 {
2159         int ret = MAP_CONTINUE;
2160         struct bkey *k;
2161         struct btree_iter iter;
2162
2163         bch_btree_iter_init(&b->keys, &iter, from);
2164
2165         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2166                 ret = !b->level
2167                         ? fn(op, b, k)
2168                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2169                 from = NULL;
2170
2171                 if (ret != MAP_CONTINUE)
2172                         return ret;
2173         }
2174
2175         if (!b->level && (flags & MAP_END_KEY))
2176                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2177                                      KEY_OFFSET(&b->key), 0));
2178
2179         return ret;
2180 }
2181
2182 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2183                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2184 {
2185         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2186 }
2187
2188 /* Keybuf code */
2189
2190 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2191 {
2192         /* Overlapping keys compare equal */
2193         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2194                 return -1;
2195         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2196                 return 1;
2197         return 0;
2198 }
2199
2200 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2201                                             struct keybuf_key *r)
2202 {
2203         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2204 }
2205
2206 struct refill {
2207         struct btree_op op;
2208         unsigned        nr_found;
2209         struct keybuf   *buf;
2210         struct bkey     *end;
2211         keybuf_pred_fn  *pred;
2212 };
2213
2214 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2215                             struct bkey *k)
2216 {
2217         struct refill *refill = container_of(op, struct refill, op);
2218         struct keybuf *buf = refill->buf;
2219         int ret = MAP_CONTINUE;
2220
2221         if (bkey_cmp(k, refill->end) >= 0) {
2222                 ret = MAP_DONE;
2223                 goto out;
2224         }
2225
2226         if (!KEY_SIZE(k)) /* end key */
2227                 goto out;
2228
2229         if (refill->pred(buf, k)) {
2230                 struct keybuf_key *w;
2231
2232                 spin_lock(&buf->lock);
2233
2234                 w = array_alloc(&buf->freelist);
2235                 if (!w) {
2236                         spin_unlock(&buf->lock);
2237                         return MAP_DONE;
2238                 }
2239
2240                 w->private = NULL;
2241                 bkey_copy(&w->key, k);
2242
2243                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2244                         array_free(&buf->freelist, w);
2245                 else
2246                         refill->nr_found++;
2247
2248                 if (array_freelist_empty(&buf->freelist))
2249                         ret = MAP_DONE;
2250
2251                 spin_unlock(&buf->lock);
2252         }
2253 out:
2254         buf->last_scanned = *k;
2255         return ret;
2256 }
2257
2258 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2259                        struct bkey *end, keybuf_pred_fn *pred)
2260 {
2261         struct bkey start = buf->last_scanned;
2262         struct refill refill;
2263
2264         cond_resched();
2265
2266         bch_btree_op_init(&refill.op, -1);
2267         refill.nr_found = 0;
2268         refill.buf      = buf;
2269         refill.end      = end;
2270         refill.pred     = pred;
2271
2272         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2273                            refill_keybuf_fn, MAP_END_KEY);
2274
2275         trace_bcache_keyscan(refill.nr_found,
2276                              KEY_INODE(&start), KEY_OFFSET(&start),
2277                              KEY_INODE(&buf->last_scanned),
2278                              KEY_OFFSET(&buf->last_scanned));
2279
2280         spin_lock(&buf->lock);
2281
2282         if (!RB_EMPTY_ROOT(&buf->keys)) {
2283                 struct keybuf_key *w;
2284                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2285                 buf->start      = START_KEY(&w->key);
2286
2287                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2288                 buf->end        = w->key;
2289         } else {
2290                 buf->start      = MAX_KEY;
2291                 buf->end        = MAX_KEY;
2292         }
2293
2294         spin_unlock(&buf->lock);
2295 }
2296
2297 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2298 {
2299         rb_erase(&w->node, &buf->keys);
2300         array_free(&buf->freelist, w);
2301 }
2302
2303 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2304 {
2305         spin_lock(&buf->lock);
2306         __bch_keybuf_del(buf, w);
2307         spin_unlock(&buf->lock);
2308 }
2309
2310 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2311                                   struct bkey *end)
2312 {
2313         bool ret = false;
2314         struct keybuf_key *p, *w, s;
2315         s.key = *start;
2316
2317         if (bkey_cmp(end, &buf->start) <= 0 ||
2318             bkey_cmp(start, &buf->end) >= 0)
2319                 return false;
2320
2321         spin_lock(&buf->lock);
2322         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2323
2324         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2325                 p = w;
2326                 w = RB_NEXT(w, node);
2327
2328                 if (p->private)
2329                         ret = true;
2330                 else
2331                         __bch_keybuf_del(buf, p);
2332         }
2333
2334         spin_unlock(&buf->lock);
2335         return ret;
2336 }
2337
2338 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2339 {
2340         struct keybuf_key *w;
2341         spin_lock(&buf->lock);
2342
2343         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2344
2345         while (w && w->private)
2346                 w = RB_NEXT(w, node);
2347
2348         if (w)
2349                 w->private = ERR_PTR(-EINTR);
2350
2351         spin_unlock(&buf->lock);
2352         return w;
2353 }
2354
2355 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2356                                           struct keybuf *buf,
2357                                           struct bkey *end,
2358                                           keybuf_pred_fn *pred)
2359 {
2360         struct keybuf_key *ret;
2361
2362         while (1) {
2363                 ret = bch_keybuf_next(buf);
2364                 if (ret)
2365                         break;
2366
2367                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2368                         pr_debug("scan finished");
2369                         break;
2370                 }
2371
2372                 bch_refill_keybuf(c, buf, end, pred);
2373         }
2374
2375         return ret;
2376 }
2377
2378 void bch_keybuf_init(struct keybuf *buf)
2379 {
2380         buf->last_scanned       = MAX_KEY;
2381         buf->keys               = RB_ROOT;
2382
2383         spin_lock_init(&buf->lock);
2384         array_allocator_init(&buf->freelist);
2385 }
2386
2387 void bch_btree_exit(void)
2388 {
2389         if (btree_io_wq)
2390                 destroy_workqueue(btree_io_wq);
2391 }
2392
2393 int __init bch_btree_init(void)
2394 {
2395         btree_io_wq = create_singlethread_workqueue("bch_btree_io");
2396         if (!btree_io_wq)
2397                 return -ENOMEM;
2398
2399         return 0;
2400 }