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 "writeback.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 enum {
  93         BTREE_INSERT_STATUS_INSERT,
  94         BTREE_INSERT_STATUS_BACK_MERGE,
  95         BTREE_INSERT_STATUS_OVERWROTE,
  96         BTREE_INSERT_STATUS_FRONT_MERGE,
  97 };
  98
  99 #define MAX_NEED_GC             64
 100 #define MAX_SAVE_PRIO           72
 101
 102 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
 103
 104 #define PTR_HASH(c, k)                                                  \
 105         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 106
 107 static struct workqueue_struct *btree_io_wq;
 108
 109 static inline bool should_split(struct btree *b)
 110 {
 111         struct bset *i = write_block(b);
 112         return b->written >= btree_blocks(b) ||
 113                 (b->written + __set_blocks(i, i->keys + 15, b->c)
 114                  > btree_blocks(b));
 115 }
 116
 117 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 118
 119 /*
 120  * These macros are for recursing down the btree - they handle the details of
 121  * locking and looking up nodes in the cache for you. They're best treated as
 122  * mere syntax when reading code that uses them.
 123  *
 124  * op->lock determines whether we take a read or a write lock at a given depth.
 125  * If you've got a read lock and find that you need a write lock (i.e. you're
 126  * going to have to split), set op->lock and return -EINTR; btree_root() will
 127  * call you again and you'll have the correct lock.
 128  */
 129
 130 /**
 131  * btree - recurse down the btree on a specified key
 132  * @fn:         function to call, which will be passed the child node
 133  * @key:        key to recurse on
 134  * @b:          parent btree node
 135  * @op:         pointer to struct btree_op
 136  */
 137 #define btree(fn, key, b, op, ...)                                      \
 138 ({                                                                      \
 139         int _r, l = (b)->level - 1;                                     \
 140         bool _w = l <= (op)->lock;                                      \
 141         struct btree *_child = bch_btree_node_get((b)->c, key, l, _w);  \
 142         if (!IS_ERR(_child)) {                                          \
 143                 _child->parent = (b);                                   \
 144                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 145                 rw_unlock(_w, _child);                                  \
 146         } else                                                          \
 147                 _r = PTR_ERR(_child);                                   \
 148         _r;                                                             \
 149 })
 150
 151 /**
 152  * btree_root - call a function on the root of the btree
 153  * @fn:         function to call, which will be passed the child node
 154  * @c:          cache set
 155  * @op:         pointer to struct btree_op
 156  */
 157 #define btree_root(fn, c, op, ...)                                      \
 158 ({                                                                      \
 159         int _r = -EINTR;                                                \
 160         do {                                                            \
 161                 struct btree *_b = (c)->root;                           \
 162                 bool _w = insert_lock(op, _b);                          \
 163                 rw_lock(_w, _b, _b->level);                             \
 164                 if (_b == (c)->root &&                                  \
 165                     _w == insert_lock(op, _b)) {                        \
 166                         _b->parent = NULL;                              \
 167                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 168                 }                                                       \
 169                 rw_unlock(_w, _b);                                      \
 170                 if (_r == -EINTR)                                       \
 171                         schedule();                                     \
 172                 bch_cannibalize_unlock(c);                              \
 173                 if (_r == -ENOSPC) {                                    \
 174                         wait_event((c)->try_wait,                       \
 175                                    !(c)->try_harder);                   \
 176                         _r = -EINTR;                                    \
 177                 }                                                       \
 178         } while (_r == -EINTR);                                         \
 179                                                                         \
 180         finish_wait(&(c)->bucket_wait, &(op)->wait);                    \
 181         _r;                                                             \
 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 = end(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 = b->sets[0].data;
 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;
 218 #endif
 219
 220         if (!i->seq)
 221                 goto err;
 222
 223         for (;
 224              b->written < btree_blocks(b) && i->seq == b->sets[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, b->c) > btree_blocks(b))
 232                         goto err;
 233
 234                 err = "bad magic";
 235                 if (i->magic != bset_magic(&b->c->sb))
 236                         goto err;
 237
 238                 err = "bad checksum";
 239                 switch (i->version) {
 240                 case 0:
 241                         if (i->csum != csum_set(i))
 242                                 goto err;
 243                         break;
 244                 case BCACHE_BSET_VERSION:
 245                         if (i->csum != btree_csum_set(b, i))
 246                                 goto err;
 247                         break;
 248                 }
 249
 250                 err = "empty set";
 251                 if (i != b->sets[0].data && !i->keys)
 252                         goto err;
 253
 254                 bch_btree_iter_push(iter, i->start, end(i));
 255
 256                 b->written += set_blocks(i, b->c);
 257         }
 258
 259         err = "corrupted btree";
 260         for (i = write_block(b);
 261              bset_sector_offset(b, i) < KEY_SIZE(&b->key);
 262              i = ((void *) i) + block_bytes(b->c))
 263                 if (i->seq == b->sets[0].data->seq)
 264                         goto err;
 265
 266         bch_btree_sort_and_fix_extents(b, iter);
 267
 268         i = b->sets[0].data;
 269         err = "short btree key";
 270         if (b->sets[0].size &&
 271             bkey_cmp(&b->key, &b->sets[0].end) < 0)
 272                 goto err;
 273
 274         if (b->written < btree_blocks(b))
 275                 bch_bset_init_next(b);
 276 out:
 277         mempool_free(iter, b->c->fill_iter);
 278         return;
 279 err:
 280         set_btree_node_io_error(b);
 281         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 282                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 283                             bset_block_offset(b, i), i->keys);
 284         goto out;
 285 }
 286
 287 static void btree_node_read_endio(struct bio *bio, int error)
 288 {
 289         struct closure *cl = bio->bi_private;
 290         closure_put(cl);
 291 }
 292
 293 static void bch_btree_node_read(struct btree *b)
 294 {
 295         uint64_t start_time = local_clock();
 296         struct closure cl;
 297         struct bio *bio;
 298
 299         trace_bcache_btree_read(b);
 300
 301         closure_init_stack(&cl);
 302
 303         bio = bch_bbio_alloc(b->c);
 304         bio->bi_rw      = REQ_META|READ_SYNC;
 305         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 306         bio->bi_end_io  = btree_node_read_endio;
 307         bio->bi_private = &cl;
 308
 309         bch_bio_map(bio, b->sets[0].data);
 310
 311         bch_submit_bbio(bio, b->c, &b->key, 0);
 312         closure_sync(&cl);
 313
 314         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 315                 set_btree_node_io_error(b);
 316
 317         bch_bbio_free(bio, b->c);
 318
 319         if (btree_node_io_error(b))
 320                 goto err;
 321
 322         bch_btree_node_read_done(b);
 323         bch_time_stats_update(&b->c->btree_read_time, start_time);
 324
 325         return;
 326 err:
 327         bch_cache_set_error(b->c, "io error reading bucket %zu",
 328                             PTR_BUCKET_NR(b->c, &b->key, 0));
 329 }
 330
 331 static void btree_complete_write(struct btree *b, struct btree_write *w)
 332 {
 333         if (w->prio_blocked &&
 334             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 335                 wake_up_allocators(b->c);
 336
 337         if (w->journal) {
 338                 atomic_dec_bug(w->journal);
 339                 __closure_wake_up(&b->c->journal.wait);
 340         }
 341
 342         w->prio_blocked = 0;
 343         w->journal      = NULL;
 344 }
 345
 346 static void btree_node_write_unlock(struct closure *cl)
 347 {
 348         struct btree *b = container_of(cl, struct btree, io);
 349
 350         up(&b->io_mutex);
 351 }
 352
 353 static void __btree_node_write_done(struct closure *cl)
 354 {
 355         struct btree *b = container_of(cl, struct btree, io);
 356         struct btree_write *w = btree_prev_write(b);
 357
 358         bch_bbio_free(b->bio, b->c);
 359         b->bio = NULL;
 360         btree_complete_write(b, w);
 361
 362         if (btree_node_dirty(b))
 363                 queue_delayed_work(btree_io_wq, &b->work,
 364                                    msecs_to_jiffies(30000));
 365
 366         closure_return_with_destructor(cl, btree_node_write_unlock);
 367 }
 368
 369 static void btree_node_write_done(struct closure *cl)
 370 {
 371         struct btree *b = container_of(cl, struct btree, io);
 372         struct bio_vec *bv;
 373         int n;
 374
 375         bio_for_each_segment_all(bv, b->bio, n)
 376                 __free_page(bv->bv_page);
 377
 378         __btree_node_write_done(cl);
 379 }
 380
 381 static void btree_node_write_endio(struct bio *bio, int error)
 382 {
 383         struct closure *cl = bio->bi_private;
 384         struct btree *b = container_of(cl, struct btree, io);
 385
 386         if (error)
 387                 set_btree_node_io_error(b);
 388
 389         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 390         closure_put(cl);
 391 }
 392
 393 static void do_btree_node_write(struct btree *b)
 394 {
 395         struct closure *cl = &b->io;
 396         struct bset *i = b->sets[b->nsets].data;
 397         BKEY_PADDED(key) k;
 398
 399         i->version      = BCACHE_BSET_VERSION;
 400         i->csum         = btree_csum_set(b, i);
 401
 402         BUG_ON(b->bio);
 403         b->bio = bch_bbio_alloc(b->c);
 404
 405         b->bio->bi_end_io       = btree_node_write_endio;
 406         b->bio->bi_private      = cl;
 407         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 408         b->bio->bi_iter.bi_size = set_blocks(i, b->c) * block_bytes(b->c);
 409         bch_bio_map(b->bio, i);
 410
 411         /*
 412          * If we're appending to a leaf node, we don't technically need FUA -
 413          * this write just needs to be persisted before the next journal write,
 414          * which will be marked FLUSH|FUA.
 415          *
 416          * Similarly if we're writing a new btree root - the pointer is going to
 417          * be in the next journal entry.
 418          *
 419          * But if we're writing a new btree node (that isn't a root) or
 420          * appending to a non leaf btree node, we need either FUA or a flush
 421          * when we write the parent with the new pointer. FUA is cheaper than a
 422          * flush, and writes appending to leaf nodes aren't blocking anything so
 423          * just make all btree node writes FUA to keep things sane.
 424          */
 425
 426         bkey_copy(&k.key, &b->key);
 427         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
 428
 429         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 430                 int j;
 431                 struct bio_vec *bv;
 432                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 433
 434                 bio_for_each_segment_all(bv, b->bio, j)
 435                         memcpy(page_address(bv->bv_page),
 436                                base + j * PAGE_SIZE, PAGE_SIZE);
 437
 438                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 439
 440                 continue_at(cl, btree_node_write_done, NULL);
 441         } else {
 442                 b->bio->bi_vcnt = 0;
 443                 bch_bio_map(b->bio, i);
 444
 445                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 446
 447                 closure_sync(cl);
 448                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 449         }
 450 }
 451
 452 void bch_btree_node_write(struct btree *b, struct closure *parent)
 453 {
 454         struct bset *i = b->sets[b->nsets].data;
 455
 456         trace_bcache_btree_write(b);
 457
 458         BUG_ON(current->bio_list);
 459         BUG_ON(b->written >= btree_blocks(b));
 460         BUG_ON(b->written && !i->keys);
 461         BUG_ON(b->sets->data->seq != i->seq);
 462         bch_check_keys(b, "writing");
 463
 464         cancel_delayed_work(&b->work);
 465
 466         /* If caller isn't waiting for write, parent refcount is cache set */
 467         down(&b->io_mutex);
 468         closure_init(&b->io, parent ?: &b->c->cl);
 469
 470         clear_bit(BTREE_NODE_dirty,      &b->flags);
 471         change_bit(BTREE_NODE_write_idx, &b->flags);
 472
 473         do_btree_node_write(b);
 474
 475         b->written += set_blocks(i, b->c);
 476         atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
 477                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 478
 479         bch_btree_sort_lazy(b);
 480
 481         /*
 482          * do verify if there was more than one set initially (i.e. we did a
 483          * sort) and we sorted down to a single set:
 484          */
 485         if (i != b->sets->data && !b->nsets)
 486                 bch_btree_verify(b);
 487
 488         if (b->written < btree_blocks(b))
 489                 bch_bset_init_next(b);
 490 }
 491
 492 static void bch_btree_node_write_sync(struct btree *b)
 493 {
 494         struct closure cl;
 495
 496         closure_init_stack(&cl);
 497         bch_btree_node_write(b, &cl);
 498         closure_sync(&cl);
 499 }
 500
 501 static void btree_node_write_work(struct work_struct *w)
 502 {
 503         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 504
 505         rw_lock(true, b, b->level);
 506
 507         if (btree_node_dirty(b))
 508                 bch_btree_node_write(b, NULL);
 509         rw_unlock(true, b);
 510 }
 511
 512 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 513 {
 514         struct bset *i = b->sets[b->nsets].data;
 515         struct btree_write *w = btree_current_write(b);
 516
 517         BUG_ON(!b->written);
 518         BUG_ON(!i->keys);
 519
 520         if (!btree_node_dirty(b))
 521                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 522
 523         set_btree_node_dirty(b);
 524
 525         if (journal_ref) {
 526                 if (w->journal &&
 527                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 528                         atomic_dec_bug(w->journal);
 529                         w->journal = NULL;
 530                 }
 531
 532                 if (!w->journal) {
 533                         w->journal = journal_ref;
 534                         atomic_inc(w->journal);
 535                 }
 536         }
 537
 538         /* Force write if set is too big */
 539         if (set_bytes(i) > PAGE_SIZE - 48 &&
 540             !current->bio_list)
 541                 bch_btree_node_write(b, NULL);
 542 }
 543
 544 /*
 545  * Btree in memory cache - allocation/freeing
 546  * mca -> memory cache
 547  */
 548
 549 static void mca_reinit(struct btree *b)
 550 {
 551         unsigned i;
 552
 553         b->flags        = 0;
 554         b->written      = 0;
 555         b->nsets        = 0;
 556
 557         for (i = 0; i < MAX_BSETS; i++)
 558                 b->sets[i].size = 0;
 559         /*
 560          * Second loop starts at 1 because b->sets[0]->data is the memory we
 561          * allocated
 562          */
 563         for (i = 1; i < MAX_BSETS; i++)
 564                 b->sets[i].data = NULL;
 565 }
 566
 567 #define mca_reserve(c)  (((c->root && c->root->level)           \
 568                           ? c->root->level : 1) * 8 + 16)
 569 #define mca_can_free(c)                                         \
 570         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 571
 572 static void mca_data_free(struct btree *b)
 573 {
 574         struct bset_tree *t = b->sets;
 575
 576         BUG_ON(b->io_mutex.count != 1);
 577
 578         if (bset_prev_bytes(b) < PAGE_SIZE)
 579                 kfree(t->prev);
 580         else
 581                 free_pages((unsigned long) t->prev,
 582                            get_order(bset_prev_bytes(b)));
 583
 584         if (bset_tree_bytes(b) < PAGE_SIZE)
 585                 kfree(t->tree);
 586         else
 587                 free_pages((unsigned long) t->tree,
 588                            get_order(bset_tree_bytes(b)));
 589
 590         free_pages((unsigned long) t->data, b->page_order);
 591
 592         t->prev = NULL;
 593         t->tree = NULL;
 594         t->data = NULL;
 595         list_move(&b->list, &b->c->btree_cache_freed);
 596         b->c->bucket_cache_used--;
 597 }
 598
 599 static void mca_bucket_free(struct btree *b)
 600 {
 601         BUG_ON(btree_node_dirty(b));
 602
 603         b->key.ptr[0] = 0;
 604         hlist_del_init_rcu(&b->hash);
 605         list_move(&b->list, &b->c->btree_cache_freeable);
 606 }
 607
 608 static unsigned btree_order(struct bkey *k)
 609 {
 610         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 611 }
 612
 613 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 614 {
 615         struct bset_tree *t = b->sets;
 616         BUG_ON(t->data);
 617
 618         b->page_order = max_t(unsigned,
 619                               ilog2(b->c->btree_pages),
 620                               btree_order(k));
 621
 622         t->data = (void *) __get_free_pages(gfp, b->page_order);
 623         if (!t->data)
 624                 goto err;
 625
 626         t->tree = bset_tree_bytes(b) < PAGE_SIZE
 627                 ? kmalloc(bset_tree_bytes(b), gfp)
 628                 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
 629         if (!t->tree)
 630                 goto err;
 631
 632         t->prev = bset_prev_bytes(b) < PAGE_SIZE
 633                 ? kmalloc(bset_prev_bytes(b), gfp)
 634                 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
 635         if (!t->prev)
 636                 goto err;
 637
 638         list_move(&b->list, &b->c->btree_cache);
 639         b->c->bucket_cache_used++;
 640         return;
 641 err:
 642         mca_data_free(b);
 643 }
 644
 645 static struct btree *mca_bucket_alloc(struct cache_set *c,
 646                                       struct bkey *k, gfp_t gfp)
 647 {
 648         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 649         if (!b)
 650                 return NULL;
 651
 652         init_rwsem(&b->lock);
 653         lockdep_set_novalidate_class(&b->lock);
 654         INIT_LIST_HEAD(&b->list);
 655         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 656         b->c = c;
 657         sema_init(&b->io_mutex, 1);
 658
 659         mca_data_alloc(b, k, gfp);
 660         return b;
 661 }
 662
 663 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 664 {
 665         struct closure cl;
 666
 667         closure_init_stack(&cl);
 668         lockdep_assert_held(&b->c->bucket_lock);
 669
 670         if (!down_write_trylock(&b->lock))
 671                 return -ENOMEM;
 672
 673         BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
 674
 675         if (b->page_order < min_order)
 676                 goto out_unlock;
 677
 678         if (!flush) {
 679                 if (btree_node_dirty(b))
 680                         goto out_unlock;
 681
 682                 if (down_trylock(&b->io_mutex))
 683                         goto out_unlock;
 684                 up(&b->io_mutex);
 685         }
 686
 687         if (btree_node_dirty(b))
 688                 bch_btree_node_write_sync(b);
 689
 690         /* wait for any in flight btree write */
 691         down(&b->io_mutex);
 692         up(&b->io_mutex);
 693
 694         return 0;
 695 out_unlock:
 696         rw_unlock(true, b);
 697         return -ENOMEM;
 698 }
 699
 700 static unsigned long bch_mca_scan(struct shrinker *shrink,
 701                                   struct shrink_control *sc)
 702 {
 703         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 704         struct btree *b, *t;
 705         unsigned long i, nr = sc->nr_to_scan;
 706         unsigned long freed = 0;
 707
 708         if (c->shrinker_disabled)
 709                 return SHRINK_STOP;
 710
 711         if (c->try_harder)
 712                 return SHRINK_STOP;
 713
 714         /* Return -1 if we can't do anything right now */
 715         if (sc->gfp_mask & __GFP_IO)
 716                 mutex_lock(&c->bucket_lock);
 717         else if (!mutex_trylock(&c->bucket_lock))
 718                 return -1;
 719
 720         /*
 721          * It's _really_ critical that we don't free too many btree nodes - we
 722          * have to always leave ourselves a reserve. The reserve is how we
 723          * guarantee that allocating memory for a new btree node can always
 724          * succeed, so that inserting keys into the btree can always succeed and
 725          * IO can always make forward progress:
 726          */
 727         nr /= c->btree_pages;
 728         nr = min_t(unsigned long, nr, mca_can_free(c));
 729
 730         i = 0;
 731         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 732                 if (freed >= nr)
 733                         break;
 734
 735                 if (++i > 3 &&
 736                     !mca_reap(b, 0, false)) {
 737                         mca_data_free(b);
 738                         rw_unlock(true, b);
 739                         freed++;
 740                 }
 741         }
 742
 743         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 744                 if (list_empty(&c->btree_cache))
 745                         goto out;
 746
 747                 b = list_first_entry(&c->btree_cache, struct btree, list);
 748                 list_rotate_left(&c->btree_cache);
 749
 750                 if (!b->accessed &&
 751                     !mca_reap(b, 0, false)) {
 752                         mca_bucket_free(b);
 753                         mca_data_free(b);
 754                         rw_unlock(true, b);
 755                         freed++;
 756                 } else
 757                         b->accessed = 0;
 758         }
 759 out:
 760         mutex_unlock(&c->bucket_lock);
 761         return freed;
 762 }
 763
 764 static unsigned long bch_mca_count(struct shrinker *shrink,
 765                                    struct shrink_control *sc)
 766 {
 767         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 768
 769         if (c->shrinker_disabled)
 770                 return 0;
 771
 772         if (c->try_harder)
 773                 return 0;
 774
 775         return mca_can_free(c) * c->btree_pages;
 776 }
 777
 778 void bch_btree_cache_free(struct cache_set *c)
 779 {
 780         struct btree *b;
 781         struct closure cl;
 782         closure_init_stack(&cl);
 783
 784         if (c->shrink.list.next)
 785                 unregister_shrinker(&c->shrink);
 786
 787         mutex_lock(&c->bucket_lock);
 788
 789 #ifdef CONFIG_BCACHE_DEBUG
 790         if (c->verify_data)
 791                 list_move(&c->verify_data->list, &c->btree_cache);
 792
 793         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 794 #endif
 795
 796         list_splice(&c->btree_cache_freeable,
 797                     &c->btree_cache);
 798
 799         while (!list_empty(&c->btree_cache)) {
 800                 b = list_first_entry(&c->btree_cache, struct btree, list);
 801
 802                 if (btree_node_dirty(b))
 803                         btree_complete_write(b, btree_current_write(b));
 804                 clear_bit(BTREE_NODE_dirty, &b->flags);
 805
 806                 mca_data_free(b);
 807         }
 808
 809         while (!list_empty(&c->btree_cache_freed)) {
 810                 b = list_first_entry(&c->btree_cache_freed,
 811                                      struct btree, list);
 812                 list_del(&b->list);
 813                 cancel_delayed_work_sync(&b->work);
 814                 kfree(b);
 815         }
 816
 817         mutex_unlock(&c->bucket_lock);
 818 }
 819
 820 int bch_btree_cache_alloc(struct cache_set *c)
 821 {
 822         unsigned i;
 823
 824         for (i = 0; i < mca_reserve(c); i++)
 825                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 826                         return -ENOMEM;
 827
 828         list_splice_init(&c->btree_cache,
 829                          &c->btree_cache_freeable);
 830
 831 #ifdef CONFIG_BCACHE_DEBUG
 832         mutex_init(&c->verify_lock);
 833
 834         c->verify_ondisk = (void *)
 835                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 836
 837         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 838
 839         if (c->verify_data &&
 840             c->verify_data->sets[0].data)
 841                 list_del_init(&c->verify_data->list);
 842         else
 843                 c->verify_data = NULL;
 844 #endif
 845
 846         c->shrink.count_objects = bch_mca_count;
 847         c->shrink.scan_objects = bch_mca_scan;
 848         c->shrink.seeks = 4;
 849         c->shrink.batch = c->btree_pages * 2;
 850         register_shrinker(&c->shrink);
 851
 852         return 0;
 853 }
 854
 855 /* Btree in memory cache - hash table */
 856
 857 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 858 {
 859         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 860 }
 861
 862 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 863 {
 864         struct btree *b;
 865
 866         rcu_read_lock();
 867         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 868                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 869                         goto out;
 870         b = NULL;
 871 out:
 872         rcu_read_unlock();
 873         return b;
 874 }
 875
 876 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 877 {
 878         struct btree *b;
 879
 880         trace_bcache_btree_cache_cannibalize(c);
 881
 882         if (!c->try_harder) {
 883                 c->try_harder = current;
 884                 c->try_harder_start = local_clock();
 885         } else if (c->try_harder != current)
 886                 return ERR_PTR(-ENOSPC);
 887
 888         list_for_each_entry_reverse(b, &c->btree_cache, list)
 889                 if (!mca_reap(b, btree_order(k), false))
 890                         return b;
 891
 892         list_for_each_entry_reverse(b, &c->btree_cache, list)
 893                 if (!mca_reap(b, btree_order(k), true))
 894                         return b;
 895
 896         return ERR_PTR(-ENOMEM);
 897 }
 898
 899 /*
 900  * We can only have one thread cannibalizing other cached btree nodes at a time,
 901  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 902  * cannibalize_bucket() will take. This means every time we unlock the root of
 903  * the btree, we need to release this lock if we have it held.
 904  */
 905 static void bch_cannibalize_unlock(struct cache_set *c)
 906 {
 907         if (c->try_harder == current) {
 908                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 909                 c->try_harder = NULL;
 910                 wake_up(&c->try_wait);
 911         }
 912 }
 913
 914 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level)
 915 {
 916         struct btree *b;
 917
 918         BUG_ON(current->bio_list);
 919
 920         lockdep_assert_held(&c->bucket_lock);
 921
 922         if (mca_find(c, k))
 923                 return NULL;
 924
 925         /* btree_free() doesn't free memory; it sticks the node on the end of
 926          * the list. Check if there's any freed nodes there:
 927          */
 928         list_for_each_entry(b, &c->btree_cache_freeable, list)
 929                 if (!mca_reap(b, btree_order(k), false))
 930                         goto out;
 931
 932         /* We never free struct btree itself, just the memory that holds the on
 933          * disk node. Check the freed list before allocating a new one:
 934          */
 935         list_for_each_entry(b, &c->btree_cache_freed, list)
 936                 if (!mca_reap(b, 0, false)) {
 937                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 938                         if (!b->sets[0].data)
 939                                 goto err;
 940                         else
 941                                 goto out;
 942                 }
 943
 944         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 945         if (!b)
 946                 goto err;
 947
 948         BUG_ON(!down_write_trylock(&b->lock));
 949         if (!b->sets->data)
 950                 goto err;
 951 out:
 952         BUG_ON(b->io_mutex.count != 1);
 953
 954         bkey_copy(&b->key, k);
 955         list_move(&b->list, &c->btree_cache);
 956         hlist_del_init_rcu(&b->hash);
 957         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 958
 959         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 960         b->level        = level;
 961         b->parent       = (void *) ~0UL;
 962
 963         mca_reinit(b);
 964
 965         return b;
 966 err:
 967         if (b)
 968                 rw_unlock(true, b);
 969
 970         b = mca_cannibalize(c, k);
 971         if (!IS_ERR(b))
 972                 goto out;
 973
 974         return b;
 975 }
 976
 977 /**
 978  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 979  * in from disk if necessary.
 980  *
 981  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 982  *
 983  * The btree node will have either a read or a write lock held, depending on
 984  * level and op->lock.
 985  */
 986 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 987                                  int level, bool write)
 988 {
 989         int i = 0;
 990         struct btree *b;
 991
 992         BUG_ON(level < 0);
 993 retry:
 994         b = mca_find(c, k);
 995
 996         if (!b) {
 997                 if (current->bio_list)
 998                         return ERR_PTR(-EAGAIN);
 999
1000                 mutex_lock(&c->bucket_lock);
1001                 b = mca_alloc(c, k, level);
1002                 mutex_unlock(&c->bucket_lock);
1003
1004                 if (!b)
1005                         goto retry;
1006                 if (IS_ERR(b))
1007                         return b;
1008
1009                 bch_btree_node_read(b);
1010
1011                 if (!write)
1012                         downgrade_write(&b->lock);
1013         } else {
1014                 rw_lock(write, b, level);
1015                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1016                         rw_unlock(write, b);
1017                         goto retry;
1018                 }
1019                 BUG_ON(b->level != level);
1020         }
1021
1022         b->accessed = 1;
1023
1024         for (; i <= b->nsets && b->sets[i].size; i++) {
1025                 prefetch(b->sets[i].tree);
1026                 prefetch(b->sets[i].data);
1027         }
1028
1029         for (; i <= b->nsets; i++)
1030                 prefetch(b->sets[i].data);
1031
1032         if (btree_node_io_error(b)) {
1033                 rw_unlock(write, b);
1034                 return ERR_PTR(-EIO);
1035         }
1036
1037         BUG_ON(!b->written);
1038
1039         return b;
1040 }
1041
1042 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1043 {
1044         struct btree *b;
1045
1046         mutex_lock(&c->bucket_lock);
1047         b = mca_alloc(c, k, level);
1048         mutex_unlock(&c->bucket_lock);
1049
1050         if (!IS_ERR_OR_NULL(b)) {
1051                 bch_btree_node_read(b);
1052                 rw_unlock(true, b);
1053         }
1054 }
1055
1056 /* Btree alloc */
1057
1058 static void btree_node_free(struct btree *b)
1059 {
1060         unsigned i;
1061
1062         trace_bcache_btree_node_free(b);
1063
1064         BUG_ON(b == b->c->root);
1065
1066         if (btree_node_dirty(b))
1067                 btree_complete_write(b, btree_current_write(b));
1068         clear_bit(BTREE_NODE_dirty, &b->flags);
1069
1070         cancel_delayed_work(&b->work);
1071
1072         mutex_lock(&b->c->bucket_lock);
1073
1074         for (i = 0; i < KEY_PTRS(&b->key); i++) {
1075                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
1076
1077                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1078                             PTR_BUCKET(b->c, &b->key, i));
1079         }
1080
1081         bch_bucket_free(b->c, &b->key);
1082         mca_bucket_free(b);
1083         mutex_unlock(&b->c->bucket_lock);
1084 }
1085
1086 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1087 {
1088         BKEY_PADDED(key) k;
1089         struct btree *b = ERR_PTR(-EAGAIN);
1090
1091         mutex_lock(&c->bucket_lock);
1092 retry:
1093         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1094                 goto err;
1095
1096         bkey_put(c, &k.key);
1097         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1098
1099         b = mca_alloc(c, &k.key, level);
1100         if (IS_ERR(b))
1101                 goto err_free;
1102
1103         if (!b) {
1104                 cache_bug(c,
1105                         "Tried to allocate bucket that was in btree cache");
1106                 goto retry;
1107         }
1108
1109         b->accessed = 1;
1110         bch_bset_init_next(b);
1111
1112         mutex_unlock(&c->bucket_lock);
1113
1114         trace_bcache_btree_node_alloc(b);
1115         return b;
1116 err_free:
1117         bch_bucket_free(c, &k.key);
1118 err:
1119         mutex_unlock(&c->bucket_lock);
1120
1121         trace_bcache_btree_node_alloc_fail(b);
1122         return b;
1123 }
1124
1125 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1126 {
1127         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1128         if (!IS_ERR_OR_NULL(n))
1129                 bch_btree_sort_into(b, n);
1130
1131         return n;
1132 }
1133
1134 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1135 {
1136         unsigned i;
1137
1138         bkey_copy(k, &b->key);
1139         bkey_copy_key(k, &ZERO_KEY);
1140
1141         for (i = 0; i < KEY_PTRS(k); i++) {
1142                 uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1;
1143
1144                 SET_PTR_GEN(k, i, g);
1145         }
1146
1147         atomic_inc(&b->c->prio_blocked);
1148 }
1149
1150 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1151 {
1152         struct cache_set *c = b->c;
1153         struct cache *ca;
1154         unsigned i, reserve = c->root->level * 2 + 1;
1155         int ret = 0;
1156
1157         mutex_lock(&c->bucket_lock);
1158
1159         for_each_cache(ca, c, i)
1160                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1161                         if (op)
1162                                 prepare_to_wait(&c->bucket_wait, &op->wait,
1163                                                 TASK_UNINTERRUPTIBLE);
1164                         ret = -EINTR;
1165                         break;
1166                 }
1167
1168         mutex_unlock(&c->bucket_lock);
1169         return ret;
1170 }
1171
1172 /* Garbage collection */
1173
1174 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1175 {
1176         uint8_t stale = 0;
1177         unsigned i;
1178         struct bucket *g;
1179
1180         /*
1181          * ptr_invalid() can't return true for the keys that mark btree nodes as
1182          * freed, but since ptr_bad() returns true we'll never actually use them
1183          * for anything and thus we don't want mark their pointers here
1184          */
1185         if (!bkey_cmp(k, &ZERO_KEY))
1186                 return stale;
1187
1188         for (i = 0; i < KEY_PTRS(k); i++) {
1189                 if (!ptr_available(c, k, i))
1190                         continue;
1191
1192                 g = PTR_BUCKET(c, k, i);
1193
1194                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1195                         g->gc_gen = PTR_GEN(k, i);
1196
1197                 if (ptr_stale(c, k, i)) {
1198                         stale = max(stale, ptr_stale(c, k, i));
1199                         continue;
1200                 }
1201
1202                 cache_bug_on(GC_MARK(g) &&
1203                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1204                              c, "inconsistent ptrs: mark = %llu, level = %i",
1205                              GC_MARK(g), level);
1206
1207                 if (level)
1208                         SET_GC_MARK(g, GC_MARK_METADATA);
1209                 else if (KEY_DIRTY(k))
1210                         SET_GC_MARK(g, GC_MARK_DIRTY);
1211
1212                 /* guard against overflow */
1213                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1214                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1215                                              (1 << 14) - 1));
1216
1217                 BUG_ON(!GC_SECTORS_USED(g));
1218         }
1219
1220         return stale;
1221 }
1222
1223 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1224
1225 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1226 {
1227         uint8_t stale = 0;
1228         unsigned keys = 0, good_keys = 0;
1229         struct bkey *k;
1230         struct btree_iter iter;
1231         struct bset_tree *t;
1232
1233         gc->nodes++;
1234
1235         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1236                 stale = max(stale, btree_mark_key(b, k));
1237                 keys++;
1238
1239                 if (bch_ptr_bad(b, k))
1240                         continue;
1241
1242                 gc->key_bytes += bkey_u64s(k);
1243                 gc->nkeys++;
1244                 good_keys++;
1245
1246                 gc->data += KEY_SIZE(k);
1247         }
1248
1249         for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1250                 btree_bug_on(t->size &&
1251                              bset_written(b, t) &&
1252                              bkey_cmp(&b->key, &t->end) < 0,
1253                              b, "found short btree key in gc");
1254
1255         if (b->c->gc_always_rewrite)
1256                 return true;
1257
1258         if (stale > 10)
1259                 return true;
1260
1261         if ((keys - good_keys) * 2 > keys)
1262                 return true;
1263
1264         return false;
1265 }
1266
1267 #define GC_MERGE_NODES  4U
1268
1269 struct gc_merge_info {
1270         struct btree    *b;
1271         unsigned        keys;
1272 };
1273
1274 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1275                                  struct keylist *, atomic_t *, struct bkey *);
1276
1277 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1278                              struct keylist *keylist, struct gc_stat *gc,
1279                              struct gc_merge_info *r)
1280 {
1281         unsigned i, nodes = 0, keys = 0, blocks;
1282         struct btree *new_nodes[GC_MERGE_NODES];
1283         struct closure cl;
1284         struct bkey *k;
1285
1286         memset(new_nodes, 0, sizeof(new_nodes));
1287         closure_init_stack(&cl);
1288
1289         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1290                 keys += r[nodes++].keys;
1291
1292         blocks = btree_default_blocks(b->c) * 2 / 3;
1293
1294         if (nodes < 2 ||
1295             __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
1296                 return 0;
1297
1298         for (i = 0; i < nodes; i++) {
1299                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1300                 if (IS_ERR_OR_NULL(new_nodes[i]))
1301                         goto out_nocoalesce;
1302         }
1303
1304         for (i = nodes - 1; i > 0; --i) {
1305                 struct bset *n1 = new_nodes[i]->sets->data;
1306                 struct bset *n2 = new_nodes[i - 1]->sets->data;
1307                 struct bkey *k, *last = NULL;
1308
1309                 keys = 0;
1310
1311                 if (i > 1) {
1312                         for (k = n2->start;
1313                              k < end(n2);
1314                              k = bkey_next(k)) {
1315                                 if (__set_blocks(n1, n1->keys + keys +
1316                                                  bkey_u64s(k), b->c) > blocks)
1317                                         break;
1318
1319                                 last = k;
1320                                 keys += bkey_u64s(k);
1321                         }
1322                 } else {
1323                         /*
1324                          * Last node we're not getting rid of - we're getting
1325                          * rid of the node at r[0]. Have to try and fit all of
1326                          * the remaining keys into this node; we can't ensure
1327                          * they will always fit due to rounding and variable
1328                          * length keys (shouldn't be possible in practice,
1329                          * though)
1330                          */
1331                         if (__set_blocks(n1, n1->keys + n2->keys,
1332                                          b->c) > btree_blocks(new_nodes[i]))
1333                                 goto out_nocoalesce;
1334
1335                         keys = n2->keys;
1336                         /* Take the key of the node we're getting rid of */
1337                         last = &r->b->key;
1338                 }
1339
1340                 BUG_ON(__set_blocks(n1, n1->keys + keys,
1341                                     b->c) > btree_blocks(new_nodes[i]));
1342
1343                 if (last)
1344                         bkey_copy_key(&new_nodes[i]->key, last);
1345
1346                 memcpy(end(n1),
1347                        n2->start,
1348                        (void *) node(n2, keys) - (void *) n2->start);
1349
1350                 n1->keys += keys;
1351                 r[i].keys = n1->keys;
1352
1353                 memmove(n2->start,
1354                         node(n2, keys),
1355                         (void *) end(n2) - (void *) node(n2, keys));
1356
1357                 n2->keys -= keys;
1358
1359                 if (__bch_keylist_realloc(keylist,
1360                                           bkey_u64s(&new_nodes[i]->key)))
1361                         goto out_nocoalesce;
1362
1363                 bch_btree_node_write(new_nodes[i], &cl);
1364                 bch_keylist_add(keylist, &new_nodes[i]->key);
1365         }
1366
1367         for (i = 0; i < nodes; i++) {
1368                 if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key)))
1369                         goto out_nocoalesce;
1370
1371                 make_btree_freeing_key(r[i].b, keylist->top);
1372                 bch_keylist_push(keylist);
1373         }
1374
1375         /* We emptied out this node */
1376         BUG_ON(new_nodes[0]->sets->data->keys);
1377         btree_node_free(new_nodes[0]);
1378         rw_unlock(true, new_nodes[0]);
1379
1380         closure_sync(&cl);
1381
1382         for (i = 0; i < nodes; i++) {
1383                 btree_node_free(r[i].b);
1384                 rw_unlock(true, r[i].b);
1385
1386                 r[i].b = new_nodes[i];
1387         }
1388
1389         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1390         BUG_ON(!bch_keylist_empty(keylist));
1391
1392         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1393         r[nodes - 1].b = ERR_PTR(-EINTR);
1394
1395         trace_bcache_btree_gc_coalesce(nodes);
1396         gc->nodes--;
1397
1398         /* Invalidated our iterator */
1399         return -EINTR;
1400
1401 out_nocoalesce:
1402         closure_sync(&cl);
1403
1404         while ((k = bch_keylist_pop(keylist)))
1405                 if (!bkey_cmp(k, &ZERO_KEY))
1406                         atomic_dec(&b->c->prio_blocked);
1407
1408         for (i = 0; i < nodes; i++)
1409                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1410                         btree_node_free(new_nodes[i]);
1411                         rw_unlock(true, new_nodes[i]);
1412                 }
1413         return 0;
1414 }
1415
1416 static unsigned btree_gc_count_keys(struct btree *b)
1417 {
1418         struct bkey *k;
1419         struct btree_iter iter;
1420         unsigned ret = 0;
1421
1422         for_each_key_filter(b, k, &iter, bch_ptr_bad)
1423                 ret += bkey_u64s(k);
1424
1425         return ret;
1426 }
1427
1428 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1429                             struct closure *writes, struct gc_stat *gc)
1430 {
1431         unsigned i;
1432         int ret = 0;
1433         bool should_rewrite;
1434         struct btree *n;
1435         struct bkey *k;
1436         struct keylist keys;
1437         struct btree_iter iter;
1438         struct gc_merge_info r[GC_MERGE_NODES];
1439         struct gc_merge_info *last = r + GC_MERGE_NODES - 1;
1440
1441         bch_keylist_init(&keys);
1442         bch_btree_iter_init(b, &iter, &b->c->gc_done);
1443
1444         for (i = 0; i < GC_MERGE_NODES; i++)
1445                 r[i].b = ERR_PTR(-EINTR);
1446
1447         while (1) {
1448                 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1449                 if (k) {
1450                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1451                         if (IS_ERR(r->b)) {
1452                                 ret = PTR_ERR(r->b);
1453                                 break;
1454                         }
1455
1456                         r->keys = btree_gc_count_keys(r->b);
1457
1458                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1459                         if (ret)
1460                                 break;
1461                 }
1462
1463                 if (!last->b)
1464                         break;
1465
1466                 if (!IS_ERR(last->b)) {
1467                         should_rewrite = btree_gc_mark_node(last->b, gc);
1468                         if (should_rewrite &&
1469                             !btree_check_reserve(b, NULL)) {
1470                                 n = btree_node_alloc_replacement(last->b,
1471                                                                  false);
1472
1473                                 if (!IS_ERR_OR_NULL(n)) {
1474                                         bch_btree_node_write_sync(n);
1475                                         bch_keylist_add(&keys, &n->key);
1476
1477                                         make_btree_freeing_key(last->b,
1478                                                                keys.top);
1479                                         bch_keylist_push(&keys);
1480
1481                                         btree_node_free(last->b);
1482
1483                                         bch_btree_insert_node(b, op, &keys,
1484                                                               NULL, NULL);
1485                                         BUG_ON(!bch_keylist_empty(&keys));
1486
1487                                         rw_unlock(true, last->b);
1488                                         last->b = n;
1489
1490                                         /* Invalidated our iterator */
1491                                         ret = -EINTR;
1492                                         break;
1493                                 }
1494                         }
1495
1496                         if (last->b->level) {
1497                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1498                                 if (ret)
1499                                         break;
1500                         }
1501
1502                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1503
1504                         /*
1505                          * Must flush leaf nodes before gc ends, since replace
1506                          * operations aren't journalled
1507                          */
1508                         if (btree_node_dirty(last->b))
1509                                 bch_btree_node_write(last->b, writes);
1510                         rw_unlock(true, last->b);
1511                 }
1512
1513                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1514                 r->b = NULL;
1515
1516                 if (need_resched()) {
1517                         ret = -EAGAIN;
1518                         break;
1519                 }
1520         }
1521
1522         for (i = 0; i < GC_MERGE_NODES; i++)
1523                 if (!IS_ERR_OR_NULL(r[i].b)) {
1524                         if (btree_node_dirty(r[i].b))
1525                                 bch_btree_node_write(r[i].b, writes);
1526                         rw_unlock(true, r[i].b);
1527                 }
1528
1529         bch_keylist_free(&keys);
1530
1531         return ret;
1532 }
1533
1534 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1535                              struct closure *writes, struct gc_stat *gc)
1536 {
1537         struct btree *n = NULL;
1538         int ret = 0;
1539         bool should_rewrite;
1540
1541         should_rewrite = btree_gc_mark_node(b, gc);
1542         if (should_rewrite) {
1543                 n = btree_node_alloc_replacement(b, false);
1544
1545                 if (!IS_ERR_OR_NULL(n)) {
1546                         bch_btree_node_write_sync(n);
1547                         bch_btree_set_root(n);
1548                         btree_node_free(b);
1549                         rw_unlock(true, n);
1550
1551                         return -EINTR;
1552                 }
1553         }
1554
1555         if (b->level) {
1556                 ret = btree_gc_recurse(b, op, writes, gc);
1557                 if (ret)
1558                         return ret;
1559         }
1560
1561         bkey_copy_key(&b->c->gc_done, &b->key);
1562
1563         return ret;
1564 }
1565
1566 static void btree_gc_start(struct cache_set *c)
1567 {
1568         struct cache *ca;
1569         struct bucket *b;
1570         unsigned i;
1571
1572         if (!c->gc_mark_valid)
1573                 return;
1574
1575         mutex_lock(&c->bucket_lock);
1576
1577         c->gc_mark_valid = 0;
1578         c->gc_done = ZERO_KEY;
1579
1580         for_each_cache(ca, c, i)
1581                 for_each_bucket(b, ca) {
1582                         b->gc_gen = b->gen;
1583                         if (!atomic_read(&b->pin)) {
1584                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1585                                 SET_GC_SECTORS_USED(b, 0);
1586                         }
1587                 }
1588
1589         mutex_unlock(&c->bucket_lock);
1590 }
1591
1592 size_t bch_btree_gc_finish(struct cache_set *c)
1593 {
1594         size_t available = 0;
1595         struct bucket *b;
1596         struct cache *ca;
1597         unsigned i;
1598
1599         mutex_lock(&c->bucket_lock);
1600
1601         set_gc_sectors(c);
1602         c->gc_mark_valid = 1;
1603         c->need_gc      = 0;
1604
1605         if (c->root)
1606                 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1607                         SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1608                                     GC_MARK_METADATA);
1609
1610         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1611                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1612                             GC_MARK_METADATA);
1613
1614         /* don't reclaim buckets to which writeback keys point */
1615         rcu_read_lock();
1616         for (i = 0; i < c->nr_uuids; i++) {
1617                 struct bcache_device *d = c->devices[i];
1618                 struct cached_dev *dc;
1619                 struct keybuf_key *w, *n;
1620                 unsigned j;
1621
1622                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1623                         continue;
1624                 dc = container_of(d, struct cached_dev, disk);
1625
1626                 spin_lock(&dc->writeback_keys.lock);
1627                 rbtree_postorder_for_each_entry_safe(w, n,
1628                                         &dc->writeback_keys.keys, node)
1629                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1630                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1631                                             GC_MARK_DIRTY);
1632                 spin_unlock(&dc->writeback_keys.lock);
1633         }
1634         rcu_read_unlock();
1635
1636         for_each_cache(ca, c, i) {
1637                 uint64_t *i;
1638
1639                 ca->invalidate_needs_gc = 0;
1640
1641                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1642                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1643
1644                 for (i = ca->prio_buckets;
1645                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1646                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1647
1648                 for_each_bucket(b, ca) {
1649                         b->last_gc      = b->gc_gen;
1650                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1651
1652                         if (!atomic_read(&b->pin) &&
1653                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1654                                 available++;
1655                                 if (!GC_SECTORS_USED(b))
1656                                         bch_bucket_add_unused(ca, b);
1657                         }
1658                 }
1659         }
1660
1661         mutex_unlock(&c->bucket_lock);
1662         return available;
1663 }
1664
1665 static void bch_btree_gc(struct cache_set *c)
1666 {
1667         int ret;
1668         unsigned long available;
1669         struct gc_stat stats;
1670         struct closure writes;
1671         struct btree_op op;
1672         uint64_t start_time = local_clock();
1673
1674         trace_bcache_gc_start(c);
1675
1676         memset(&stats, 0, sizeof(struct gc_stat));
1677         closure_init_stack(&writes);
1678         bch_btree_op_init(&op, SHRT_MAX);
1679
1680         btree_gc_start(c);
1681
1682         do {
1683                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1684                 closure_sync(&writes);
1685
1686                 if (ret && ret != -EAGAIN)
1687                         pr_warn("gc failed!");
1688         } while (ret);
1689
1690         available = bch_btree_gc_finish(c);
1691         wake_up_allocators(c);
1692
1693         bch_time_stats_update(&c->btree_gc_time, start_time);
1694
1695         stats.key_bytes *= sizeof(uint64_t);
1696         stats.data      <<= 9;
1697         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1698         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1699
1700         trace_bcache_gc_end(c);
1701
1702         bch_moving_gc(c);
1703 }
1704
1705 static int bch_gc_thread(void *arg)
1706 {
1707         struct cache_set *c = arg;
1708         struct cache *ca;
1709         unsigned i;
1710
1711         while (1) {
1712 again:
1713                 bch_btree_gc(c);
1714
1715                 set_current_state(TASK_INTERRUPTIBLE);
1716                 if (kthread_should_stop())
1717                         break;
1718
1719                 mutex_lock(&c->bucket_lock);
1720
1721                 for_each_cache(ca, c, i)
1722                         if (ca->invalidate_needs_gc) {
1723                                 mutex_unlock(&c->bucket_lock);
1724                                 set_current_state(TASK_RUNNING);
1725                                 goto again;
1726                         }
1727
1728                 mutex_unlock(&c->bucket_lock);
1729
1730                 try_to_freeze();
1731                 schedule();
1732         }
1733
1734         return 0;
1735 }
1736
1737 int bch_gc_thread_start(struct cache_set *c)
1738 {
1739         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1740         if (IS_ERR(c->gc_thread))
1741                 return PTR_ERR(c->gc_thread);
1742
1743         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1744         return 0;
1745 }
1746
1747 /* Initial partial gc */
1748
1749 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1750                                    unsigned long **seen)
1751 {
1752         int ret = 0;
1753         unsigned i;
1754         struct bkey *k, *p = NULL;
1755         struct bucket *g;
1756         struct btree_iter iter;
1757
1758         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1759                 for (i = 0; i < KEY_PTRS(k); i++) {
1760                         if (!ptr_available(b->c, k, i))
1761                                 continue;
1762
1763                         g = PTR_BUCKET(b->c, k, i);
1764
1765                         if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1766                                                 seen[PTR_DEV(k, i)]) ||
1767                             !ptr_stale(b->c, k, i)) {
1768                                 g->gen = PTR_GEN(k, i);
1769
1770                                 if (b->level)
1771                                         g->prio = BTREE_PRIO;
1772                                 else if (g->prio == BTREE_PRIO)
1773                                         g->prio = INITIAL_PRIO;
1774                         }
1775                 }
1776
1777                 btree_mark_key(b, k);
1778         }
1779
1780         if (b->level) {
1781                 bch_btree_iter_init(b, &iter, NULL);
1782
1783                 do {
1784                         k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1785                         if (k)
1786                                 btree_node_prefetch(b->c, k, b->level - 1);
1787
1788                         if (p)
1789                                 ret = btree(check_recurse, p, b, op, seen);
1790
1791                         p = k;
1792                 } while (p && !ret);
1793         }
1794
1795         return 0;
1796 }
1797
1798 int bch_btree_check(struct cache_set *c)
1799 {
1800         int ret = -ENOMEM;
1801         unsigned i;
1802         unsigned long *seen[MAX_CACHES_PER_SET];
1803         struct btree_op op;
1804
1805         memset(seen, 0, sizeof(seen));
1806         bch_btree_op_init(&op, SHRT_MAX);
1807
1808         for (i = 0; c->cache[i]; i++) {
1809                 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1810                 seen[i] = kmalloc(n, GFP_KERNEL);
1811                 if (!seen[i])
1812                         goto err;
1813
1814                 /* Disables the seen array until prio_read() uses it too */
1815                 memset(seen[i], 0xFF, n);
1816         }
1817
1818         ret = btree_root(check_recurse, c, &op, seen);
1819 err:
1820         for (i = 0; i < MAX_CACHES_PER_SET; i++)
1821                 kfree(seen[i]);
1822         return ret;
1823 }
1824
1825 /* Btree insertion */
1826
1827 static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
1828 {
1829         struct bset *i = b->sets[b->nsets].data;
1830
1831         memmove((uint64_t *) where + bkey_u64s(insert),
1832                 where,
1833                 (void *) end(i) - (void *) where);
1834
1835         i->keys += bkey_u64s(insert);
1836         bkey_copy(where, insert);
1837         bch_bset_fix_lookup_table(b, where);
1838 }
1839
1840 static bool fix_overlapping_extents(struct btree *b, struct bkey *insert,
1841                                     struct btree_iter *iter,
1842                                     struct bkey *replace_key)
1843 {
1844         void subtract_dirty(struct bkey *k, uint64_t offset, int sectors)
1845         {
1846                 if (KEY_DIRTY(k))
1847                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1848                                                      offset, -sectors);
1849         }
1850
1851         uint64_t old_offset;
1852         unsigned old_size, sectors_found = 0;
1853
1854         while (1) {
1855                 struct bkey *k = bch_btree_iter_next(iter);
1856                 if (!k ||
1857                     bkey_cmp(&START_KEY(k), insert) >= 0)
1858                         break;
1859
1860                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1861                         continue;
1862
1863                 old_offset = KEY_START(k);
1864                 old_size = KEY_SIZE(k);
1865
1866                 /*
1867                  * We might overlap with 0 size extents; we can't skip these
1868                  * because if they're in the set we're inserting to we have to
1869                  * adjust them so they don't overlap with the key we're
1870                  * inserting. But we don't want to check them for replace
1871                  * operations.
1872                  */
1873
1874                 if (replace_key && KEY_SIZE(k)) {
1875                         /*
1876                          * k might have been split since we inserted/found the
1877                          * key we're replacing
1878                          */
1879                         unsigned i;
1880                         uint64_t offset = KEY_START(k) -
1881                                 KEY_START(replace_key);
1882
1883                         /* But it must be a subset of the replace key */
1884                         if (KEY_START(k) < KEY_START(replace_key) ||
1885                             KEY_OFFSET(k) > KEY_OFFSET(replace_key))
1886                                 goto check_failed;
1887
1888                         /* We didn't find a key that we were supposed to */
1889                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
1890                                 goto check_failed;
1891
1892                         if (KEY_PTRS(k) != KEY_PTRS(replace_key) ||
1893                             KEY_DIRTY(k) != KEY_DIRTY(replace_key))
1894                                 goto check_failed;
1895
1896                         /* skip past gen */
1897                         offset <<= 8;
1898
1899                         BUG_ON(!KEY_PTRS(replace_key));
1900
1901                         for (i = 0; i < KEY_PTRS(replace_key); i++)
1902                                 if (k->ptr[i] != replace_key->ptr[i] + offset)
1903                                         goto check_failed;
1904
1905                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1906                 }
1907
1908                 if (bkey_cmp(insert, k) < 0 &&
1909                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1910                         /*
1911                          * We overlapped in the middle of an existing key: that
1912                          * means we have to split the old key. But we have to do
1913                          * slightly different things depending on whether the
1914                          * old key has been written out yet.
1915                          */
1916
1917                         struct bkey *top;
1918
1919                         subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert));
1920
1921                         if (bkey_written(b, k)) {
1922                                 /*
1923                                  * We insert a new key to cover the top of the
1924                                  * old key, and the old key is modified in place
1925                                  * to represent the bottom split.
1926                                  *
1927                                  * It's completely arbitrary whether the new key
1928                                  * is the top or the bottom, but it has to match
1929                                  * up with what btree_sort_fixup() does - it
1930                                  * doesn't check for this kind of overlap, it
1931                                  * depends on us inserting a new key for the top
1932                                  * here.
1933                                  */
1934                                 top = bch_bset_search(b, &b->sets[b->nsets],
1935                                                       insert);
1936                                 shift_keys(b, top, k);
1937                         } else {
1938                                 BKEY_PADDED(key) temp;
1939                                 bkey_copy(&temp.key, k);
1940                                 shift_keys(b, k, &temp.key);
1941                                 top = bkey_next(k);
1942                         }
1943
1944                         bch_cut_front(insert, top);
1945                         bch_cut_back(&START_KEY(insert), k);
1946                         bch_bset_fix_invalidated_key(b, k);
1947                         return false;
1948                 }
1949
1950                 if (bkey_cmp(insert, k) < 0) {
1951                         bch_cut_front(insert, k);
1952                 } else {
1953                         if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0)
1954                                 old_offset = KEY_START(insert);
1955
1956                         if (bkey_written(b, k) &&
1957                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1958                                 /*
1959                                  * Completely overwrote, so we don't have to
1960                                  * invalidate the binary search tree
1961                                  */
1962                                 bch_cut_front(k, k);
1963                         } else {
1964                                 __bch_cut_back(&START_KEY(insert), k);
1965                                 bch_bset_fix_invalidated_key(b, k);
1966                         }
1967                 }
1968
1969                 subtract_dirty(k, old_offset, old_size - KEY_SIZE(k));
1970         }
1971
1972 check_failed:
1973         if (replace_key) {
1974                 if (!sectors_found) {
1975                         return true;
1976                 } else if (sectors_found < KEY_SIZE(insert)) {
1977                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1978                                        (KEY_SIZE(insert) - sectors_found));
1979                         SET_KEY_SIZE(insert, sectors_found);
1980                 }
1981         }
1982
1983         return false;
1984 }
1985
1986 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1987                              struct bkey *k, struct bkey *replace_key)
1988 {
1989         struct bset *i = b->sets[b->nsets].data;
1990         struct bkey *m, *prev;
1991         unsigned status = BTREE_INSERT_STATUS_INSERT;
1992
1993         BUG_ON(bkey_cmp(k, &b->key) > 0);
1994         BUG_ON(b->level && !KEY_PTRS(k));
1995         BUG_ON(!b->level && !KEY_OFFSET(k));
1996
1997         if (!b->level) {
1998                 struct btree_iter iter;
1999
2000                 /*
2001                  * bset_search() returns the first key that is strictly greater
2002                  * than the search key - but for back merging, we want to find
2003                  * the previous key.
2004                  */
2005                 prev = NULL;
2006                 m = bch_btree_iter_init(b, &iter, PRECEDING_KEY(&START_KEY(k)));
2007
2008                 if (fix_overlapping_extents(b, k, &iter, replace_key)) {
2009                         op->insert_collision = true;
2010                         return false;
2011                 }
2012
2013                 if (KEY_DIRTY(k))
2014                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
2015                                                      KEY_START(k), KEY_SIZE(k));
2016
2017                 while (m != end(i) &&
2018                        bkey_cmp(k, &START_KEY(m)) > 0)
2019                         prev = m, m = bkey_next(m);
2020
2021                 if (key_merging_disabled(b->c))
2022                         goto insert;
2023
2024                 /* prev is in the tree, if we merge we're done */
2025                 status = BTREE_INSERT_STATUS_BACK_MERGE;
2026                 if (prev &&
2027                     bch_bkey_try_merge(b, prev, k))
2028                         goto merged;
2029
2030                 status = BTREE_INSERT_STATUS_OVERWROTE;
2031                 if (m != end(i) &&
2032                     KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
2033                         goto copy;
2034
2035                 status = BTREE_INSERT_STATUS_FRONT_MERGE;
2036                 if (m != end(i) &&
2037                     bch_bkey_try_merge(b, k, m))
2038                         goto copy;
2039         } else {
2040                 BUG_ON(replace_key);
2041                 m = bch_bset_search(b, &b->sets[b->nsets], k);
2042         }
2043
2044 insert: shift_keys(b, m, k);
2045 copy:   bkey_copy(m, k);
2046 merged:
2047         bch_check_keys(b, "%u for %s", status,
2048                        replace_key ? "replace" : "insert");
2049
2050         if (b->level && !KEY_OFFSET(k))
2051                 btree_current_write(b)->prio_blocked++;
2052
2053         trace_bcache_btree_insert_key(b, k, replace_key != NULL, status);
2054
2055         return true;
2056 }
2057
2058 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2059                                   struct keylist *insert_keys,
2060                                   struct bkey *replace_key)
2061 {
2062         bool ret = false;
2063         int oldsize = bch_count_data(b);
2064
2065         while (!bch_keylist_empty(insert_keys)) {
2066                 struct bset *i = write_block(b);
2067                 struct bkey *k = insert_keys->keys;
2068
2069                 if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c)
2070                     > btree_blocks(b))
2071                         break;
2072
2073                 if (bkey_cmp(k, &b->key) <= 0) {
2074                         if (!b->level)
2075                                 bkey_put(b->c, k);
2076
2077                         ret |= btree_insert_key(b, op, k, replace_key);
2078                         bch_keylist_pop_front(insert_keys);
2079                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2080                         BKEY_PADDED(key) temp;
2081                         bkey_copy(&temp.key, insert_keys->keys);
2082
2083                         bch_cut_back(&b->key, &temp.key);
2084                         bch_cut_front(&b->key, insert_keys->keys);
2085
2086                         ret |= btree_insert_key(b, op, &temp.key, replace_key);
2087                         break;
2088                 } else {
2089                         break;
2090                 }
2091         }
2092
2093         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2094
2095         BUG_ON(bch_count_data(b) < oldsize);
2096         return ret;
2097 }
2098
2099 static int btree_split(struct btree *b, struct btree_op *op,
2100                        struct keylist *insert_keys,
2101                        struct bkey *replace_key)
2102 {
2103         bool split;
2104         struct btree *n1, *n2 = NULL, *n3 = NULL;
2105         uint64_t start_time = local_clock();
2106         struct closure cl;
2107         struct keylist parent_keys;
2108
2109         closure_init_stack(&cl);
2110         bch_keylist_init(&parent_keys);
2111
2112         if (!b->level &&
2113             btree_check_reserve(b, op))
2114                 return -EINTR;
2115
2116         n1 = btree_node_alloc_replacement(b, true);
2117         if (IS_ERR(n1))
2118                 goto err;
2119
2120         split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
2121
2122         if (split) {
2123                 unsigned keys = 0;
2124
2125                 trace_bcache_btree_node_split(b, n1->sets[0].data->keys);
2126
2127                 n2 = bch_btree_node_alloc(b->c, b->level, true);
2128                 if (IS_ERR(n2))
2129                         goto err_free1;
2130
2131                 if (!b->parent) {
2132                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
2133                         if (IS_ERR(n3))
2134                                 goto err_free2;
2135                 }
2136
2137                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2138
2139                 /*
2140                  * Has to be a linear search because we don't have an auxiliary
2141                  * search tree yet
2142                  */
2143
2144                 while (keys < (n1->sets[0].data->keys * 3) / 5)
2145                         keys += bkey_u64s(node(n1->sets[0].data, keys));
2146
2147                 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
2148                 keys += bkey_u64s(node(n1->sets[0].data, keys));
2149
2150                 n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
2151                 n1->sets[0].data->keys = keys;
2152
2153                 memcpy(n2->sets[0].data->start,
2154                        end(n1->sets[0].data),
2155                        n2->sets[0].data->keys * sizeof(uint64_t));
2156
2157                 bkey_copy_key(&n2->key, &b->key);
2158
2159                 bch_keylist_add(&parent_keys, &n2->key);
2160                 bch_btree_node_write(n2, &cl);
2161                 rw_unlock(true, n2);
2162         } else {
2163                 trace_bcache_btree_node_compact(b, n1->sets[0].data->keys);
2164
2165                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2166         }
2167
2168         bch_keylist_add(&parent_keys, &n1->key);
2169         bch_btree_node_write(n1, &cl);
2170
2171         if (n3) {
2172                 /* Depth increases, make a new root */
2173                 bkey_copy_key(&n3->key, &MAX_KEY);
2174                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2175                 bch_btree_node_write(n3, &cl);
2176
2177                 closure_sync(&cl);
2178                 bch_btree_set_root(n3);
2179                 rw_unlock(true, n3);
2180
2181                 btree_node_free(b);
2182         } else if (!b->parent) {
2183                 /* Root filled up but didn't need to be split */
2184                 closure_sync(&cl);
2185                 bch_btree_set_root(n1);
2186
2187                 btree_node_free(b);
2188         } else {
2189                 /* Split a non root node */
2190                 closure_sync(&cl);
2191                 make_btree_freeing_key(b, parent_keys.top);
2192                 bch_keylist_push(&parent_keys);
2193
2194                 btree_node_free(b);
2195
2196                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2197                 BUG_ON(!bch_keylist_empty(&parent_keys));
2198         }
2199
2200         rw_unlock(true, n1);
2201
2202         bch_time_stats_update(&b->c->btree_split_time, start_time);
2203
2204         return 0;
2205 err_free2:
2206         bkey_put(b->c, &n2->key);
2207         btree_node_free(n2);
2208         rw_unlock(true, n2);
2209 err_free1:
2210         bkey_put(b->c, &n1->key);
2211         btree_node_free(n1);
2212         rw_unlock(true, n1);
2213 err:
2214         WARN(1, "bcache: btree split failed");
2215
2216         if (n3 == ERR_PTR(-EAGAIN) ||
2217             n2 == ERR_PTR(-EAGAIN) ||
2218             n1 == ERR_PTR(-EAGAIN))
2219                 return -EAGAIN;
2220
2221         return -ENOMEM;
2222 }
2223
2224 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2225                                  struct keylist *insert_keys,
2226                                  atomic_t *journal_ref,
2227                                  struct bkey *replace_key)
2228 {
2229         BUG_ON(b->level && replace_key);
2230
2231         if (should_split(b)) {
2232                 if (current->bio_list) {
2233                         op->lock = b->c->root->level + 1;
2234                         return -EAGAIN;
2235                 } else if (op->lock <= b->c->root->level) {
2236                         op->lock = b->c->root->level + 1;
2237                         return -EINTR;
2238                 } else {
2239                         /* Invalidated all iterators */
2240                         return btree_split(b, op, insert_keys, replace_key) ?:
2241                                 -EINTR;
2242                 }
2243         } else {
2244                 BUG_ON(write_block(b) != b->sets[b->nsets].data);
2245
2246                 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2247                         if (!b->level)
2248                                 bch_btree_leaf_dirty(b, journal_ref);
2249                         else
2250                                 bch_btree_node_write_sync(b);
2251                 }
2252
2253                 return 0;
2254         }
2255 }
2256
2257 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2258                                struct bkey *check_key)
2259 {
2260         int ret = -EINTR;
2261         uint64_t btree_ptr = b->key.ptr[0];
2262         unsigned long seq = b->seq;
2263         struct keylist insert;
2264         bool upgrade = op->lock == -1;
2265
2266         bch_keylist_init(&insert);
2267
2268         if (upgrade) {
2269                 rw_unlock(false, b);
2270                 rw_lock(true, b, b->level);
2271
2272                 if (b->key.ptr[0] != btree_ptr ||
2273                     b->seq != seq + 1)
2274                         goto out;
2275         }
2276
2277         SET_KEY_PTRS(check_key, 1);
2278         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2279
2280         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2281
2282         bch_keylist_add(&insert, check_key);
2283
2284         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2285
2286         BUG_ON(!ret && !bch_keylist_empty(&insert));
2287 out:
2288         if (upgrade)
2289                 downgrade_write(&b->lock);
2290         return ret;
2291 }
2292
2293 struct btree_insert_op {
2294         struct btree_op op;
2295         struct keylist  *keys;
2296         atomic_t        *journal_ref;
2297         struct bkey     *replace_key;
2298 };
2299
2300 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2301 {
2302         struct btree_insert_op *op = container_of(b_op,
2303                                         struct btree_insert_op, op);
2304
2305         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2306                                         op->journal_ref, op->replace_key);
2307         if (ret && !bch_keylist_empty(op->keys))
2308                 return ret;
2309         else
2310                 return MAP_DONE;
2311 }
2312
2313 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2314                      atomic_t *journal_ref, struct bkey *replace_key)
2315 {
2316         struct btree_insert_op op;
2317         int ret = 0;
2318
2319         BUG_ON(current->bio_list);
2320         BUG_ON(bch_keylist_empty(keys));
2321
2322         bch_btree_op_init(&op.op, 0);
2323         op.keys         = keys;
2324         op.journal_ref  = journal_ref;
2325         op.replace_key  = replace_key;
2326
2327         while (!ret && !bch_keylist_empty(keys)) {
2328                 op.op.lock = 0;
2329                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2330                                                &START_KEY(keys->keys),
2331                                                btree_insert_fn);
2332         }
2333
2334         if (ret) {
2335                 struct bkey *k;
2336
2337                 pr_err("error %i", ret);
2338
2339                 while ((k = bch_keylist_pop(keys)))
2340                         bkey_put(c, k);
2341         } else if (op.op.insert_collision)
2342                 ret = -ESRCH;
2343
2344         return ret;
2345 }
2346
2347 void bch_btree_set_root(struct btree *b)
2348 {
2349         unsigned i;
2350         struct closure cl;
2351
2352         closure_init_stack(&cl);
2353
2354         trace_bcache_btree_set_root(b);
2355
2356         BUG_ON(!b->written);
2357
2358         for (i = 0; i < KEY_PTRS(&b->key); i++)
2359                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2360
2361         mutex_lock(&b->c->bucket_lock);
2362         list_del_init(&b->list);
2363         mutex_unlock(&b->c->bucket_lock);
2364
2365         b->c->root = b;
2366
2367         bch_journal_meta(b->c, &cl);
2368         closure_sync(&cl);
2369 }
2370
2371 /* Map across nodes or keys */
2372
2373 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2374                                        struct bkey *from,
2375                                        btree_map_nodes_fn *fn, int flags)
2376 {
2377         int ret = MAP_CONTINUE;
2378
2379         if (b->level) {
2380                 struct bkey *k;
2381                 struct btree_iter iter;
2382
2383                 bch_btree_iter_init(b, &iter, from);
2384
2385                 while ((k = bch_btree_iter_next_filter(&iter, b,
2386                                                        bch_ptr_bad))) {
2387                         ret = btree(map_nodes_recurse, k, b,
2388                                     op, from, fn, flags);
2389                         from = NULL;
2390
2391                         if (ret != MAP_CONTINUE)
2392                                 return ret;
2393                 }
2394         }
2395
2396         if (!b->level || flags == MAP_ALL_NODES)
2397                 ret = fn(op, b);
2398
2399         return ret;
2400 }
2401
2402 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2403                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2404 {
2405         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2406 }
2407
2408 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2409                                       struct bkey *from, btree_map_keys_fn *fn,
2410                                       int flags)
2411 {
2412         int ret = MAP_CONTINUE;
2413         struct bkey *k;
2414         struct btree_iter iter;
2415
2416         bch_btree_iter_init(b, &iter, from);
2417
2418         while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) {
2419                 ret = !b->level
2420                         ? fn(op, b, k)
2421                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2422                 from = NULL;
2423
2424                 if (ret != MAP_CONTINUE)
2425                         return ret;
2426         }
2427
2428         if (!b->level && (flags & MAP_END_KEY))
2429                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2430                                      KEY_OFFSET(&b->key), 0));
2431
2432         return ret;
2433 }
2434
2435 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2436                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2437 {
2438         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2439 }
2440
2441 /* Keybuf code */
2442
2443 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2444 {
2445         /* Overlapping keys compare equal */
2446         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2447                 return -1;
2448         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2449                 return 1;
2450         return 0;
2451 }
2452
2453 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2454                                             struct keybuf_key *r)
2455 {
2456         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2457 }
2458
2459 struct refill {
2460         struct btree_op op;
2461         unsigned        nr_found;
2462         struct keybuf   *buf;
2463         struct bkey     *end;
2464         keybuf_pred_fn  *pred;
2465 };
2466
2467 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2468                             struct bkey *k)
2469 {
2470         struct refill *refill = container_of(op, struct refill, op);
2471         struct keybuf *buf = refill->buf;
2472         int ret = MAP_CONTINUE;
2473
2474         if (bkey_cmp(k, refill->end) >= 0) {
2475                 ret = MAP_DONE;
2476                 goto out;
2477         }
2478
2479         if (!KEY_SIZE(k)) /* end key */
2480                 goto out;
2481
2482         if (refill->pred(buf, k)) {
2483                 struct keybuf_key *w;
2484
2485                 spin_lock(&buf->lock);
2486
2487                 w = array_alloc(&buf->freelist);
2488                 if (!w) {
2489                         spin_unlock(&buf->lock);
2490                         return MAP_DONE;
2491                 }
2492
2493                 w->private = NULL;
2494                 bkey_copy(&w->key, k);
2495
2496                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2497                         array_free(&buf->freelist, w);
2498                 else
2499                         refill->nr_found++;
2500
2501                 if (array_freelist_empty(&buf->freelist))
2502                         ret = MAP_DONE;
2503
2504                 spin_unlock(&buf->lock);
2505         }
2506 out:
2507         buf->last_scanned = *k;
2508         return ret;
2509 }
2510
2511 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2512                        struct bkey *end, keybuf_pred_fn *pred)
2513 {
2514         struct bkey start = buf->last_scanned;
2515         struct refill refill;
2516
2517         cond_resched();
2518
2519         bch_btree_op_init(&refill.op, -1);
2520         refill.nr_found = 0;
2521         refill.buf      = buf;
2522         refill.end      = end;
2523         refill.pred     = pred;
2524
2525         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2526                            refill_keybuf_fn, MAP_END_KEY);
2527
2528         trace_bcache_keyscan(refill.nr_found,
2529                              KEY_INODE(&start), KEY_OFFSET(&start),
2530                              KEY_INODE(&buf->last_scanned),
2531                              KEY_OFFSET(&buf->last_scanned));
2532
2533         spin_lock(&buf->lock);
2534
2535         if (!RB_EMPTY_ROOT(&buf->keys)) {
2536                 struct keybuf_key *w;
2537                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2538                 buf->start      = START_KEY(&w->key);
2539
2540                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2541                 buf->end        = w->key;
2542         } else {
2543                 buf->start      = MAX_KEY;
2544                 buf->end        = MAX_KEY;
2545         }
2546
2547         spin_unlock(&buf->lock);
2548 }
2549
2550 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2551 {
2552         rb_erase(&w->node, &buf->keys);
2553         array_free(&buf->freelist, w);
2554 }
2555
2556 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2557 {
2558         spin_lock(&buf->lock);
2559         __bch_keybuf_del(buf, w);
2560         spin_unlock(&buf->lock);
2561 }
2562
2563 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2564                                   struct bkey *end)
2565 {
2566         bool ret = false;
2567         struct keybuf_key *p, *w, s;
2568         s.key = *start;
2569
2570         if (bkey_cmp(end, &buf->start) <= 0 ||
2571             bkey_cmp(start, &buf->end) >= 0)
2572                 return false;
2573
2574         spin_lock(&buf->lock);
2575         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2576
2577         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2578                 p = w;
2579                 w = RB_NEXT(w, node);
2580
2581                 if (p->private)
2582                         ret = true;
2583                 else
2584                         __bch_keybuf_del(buf, p);
2585         }
2586
2587         spin_unlock(&buf->lock);
2588         return ret;
2589 }
2590
2591 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2592 {
2593         struct keybuf_key *w;
2594         spin_lock(&buf->lock);
2595
2596         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2597
2598         while (w && w->private)
2599                 w = RB_NEXT(w, node);
2600
2601         if (w)
2602                 w->private = ERR_PTR(-EINTR);
2603
2604         spin_unlock(&buf->lock);
2605         return w;
2606 }
2607
2608 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2609                                           struct keybuf *buf,
2610                                           struct bkey *end,
2611                                           keybuf_pred_fn *pred)
2612 {
2613         struct keybuf_key *ret;
2614
2615         while (1) {
2616                 ret = bch_keybuf_next(buf);
2617                 if (ret)
2618                         break;
2619
2620                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2621                         pr_debug("scan finished");
2622                         break;
2623                 }
2624
2625                 bch_refill_keybuf(c, buf, end, pred);
2626         }
2627
2628         return ret;
2629 }
2630
2631 void bch_keybuf_init(struct keybuf *buf)
2632 {
2633         buf->last_scanned       = MAX_KEY;
2634         buf->keys               = RB_ROOT;
2635
2636         spin_lock_init(&buf->lock);
2637         array_allocator_init(&buf->freelist);
2638 }
2639
2640 void bch_btree_exit(void)
2641 {
2642         if (btree_io_wq)
2643                 destroy_workqueue(btree_io_wq);
2644 }
2645
2646 int __init bch_btree_init(void)
2647 {
2648         btree_io_wq = create_singlethread_workqueue("bch_btree_io");
2649         if (!btree_io_wq)
2650                 return -ENOMEM;
2651
2652         return 0;
2653 }