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                 bch_cannibalize_unlock(c);                              \
 171                 if (_r == -ENOSPC) {                                    \
 172                         wait_event((c)->try_wait,                       \
 173                                    !(c)->try_harder);                   \
 174                         _r = -EINTR;                                    \
 175                 }                                                       \
 176         } while (_r == -EINTR);                                         \
 177                                                                         \
 178         _r;                                                             \
 179 })
 180
 181 /* Btree key manipulation */
 182
 183 void bkey_put(struct cache_set *c, struct bkey *k)
 184 {
 185         unsigned i;
 186
 187         for (i = 0; i < KEY_PTRS(k); i++)
 188                 if (ptr_available(c, k, i))
 189                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 190 }
 191
 192 /* Btree IO */
 193
 194 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 195 {
 196         uint64_t crc = b->key.ptr[0];
 197         void *data = (void *) i + 8, *end = end(i);
 198
 199         crc = bch_crc64_update(crc, data, end - data);
 200         return crc ^ 0xffffffffffffffffULL;
 201 }
 202
 203 static void bch_btree_node_read_done(struct btree *b)
 204 {
 205         const char *err = "bad btree header";
 206         struct bset *i = b->sets[0].data;
 207         struct btree_iter *iter;
 208
 209         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 210         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 211         iter->used = 0;
 212
 213 #ifdef CONFIG_BCACHE_DEBUG
 214         iter->b = b;
 215 #endif
 216
 217         if (!i->seq)
 218                 goto err;
 219
 220         for (;
 221              b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
 222              i = write_block(b)) {
 223                 err = "unsupported bset version";
 224                 if (i->version > BCACHE_BSET_VERSION)
 225                         goto err;
 226
 227                 err = "bad btree header";
 228                 if (b->written + set_blocks(i, b->c) > btree_blocks(b))
 229                         goto err;
 230
 231                 err = "bad magic";
 232                 if (i->magic != bset_magic(&b->c->sb))
 233                         goto err;
 234
 235                 err = "bad checksum";
 236                 switch (i->version) {
 237                 case 0:
 238                         if (i->csum != csum_set(i))
 239                                 goto err;
 240                         break;
 241                 case BCACHE_BSET_VERSION:
 242                         if (i->csum != btree_csum_set(b, i))
 243                                 goto err;
 244                         break;
 245                 }
 246
 247                 err = "empty set";
 248                 if (i != b->sets[0].data && !i->keys)
 249                         goto err;
 250
 251                 bch_btree_iter_push(iter, i->start, end(i));
 252
 253                 b->written += set_blocks(i, b->c);
 254         }
 255
 256         err = "corrupted btree";
 257         for (i = write_block(b);
 258              index(i, b) < btree_blocks(b);
 259              i = ((void *) i) + block_bytes(b->c))
 260                 if (i->seq == b->sets[0].data->seq)
 261                         goto err;
 262
 263         bch_btree_sort_and_fix_extents(b, iter);
 264
 265         i = b->sets[0].data;
 266         err = "short btree key";
 267         if (b->sets[0].size &&
 268             bkey_cmp(&b->key, &b->sets[0].end) < 0)
 269                 goto err;
 270
 271         if (b->written < btree_blocks(b))
 272                 bch_bset_init_next(b);
 273 out:
 274         mempool_free(iter, b->c->fill_iter);
 275         return;
 276 err:
 277         set_btree_node_io_error(b);
 278         bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys",
 279                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 280                             index(i, b), i->keys);
 281         goto out;
 282 }
 283
 284 static void btree_node_read_endio(struct bio *bio, int error)
 285 {
 286         struct closure *cl = bio->bi_private;
 287         closure_put(cl);
 288 }
 289
 290 void bch_btree_node_read(struct btree *b)
 291 {
 292         uint64_t start_time = local_clock();
 293         struct closure cl;
 294         struct bio *bio;
 295
 296         trace_bcache_btree_read(b);
 297
 298         closure_init_stack(&cl);
 299
 300         bio = bch_bbio_alloc(b->c);
 301         bio->bi_rw      = REQ_META|READ_SYNC;
 302         bio->bi_size    = KEY_SIZE(&b->key) << 9;
 303         bio->bi_end_io  = btree_node_read_endio;
 304         bio->bi_private = &cl;
 305
 306         bch_bio_map(bio, b->sets[0].data);
 307
 308         bch_submit_bbio(bio, b->c, &b->key, 0);
 309         closure_sync(&cl);
 310
 311         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 312                 set_btree_node_io_error(b);
 313
 314         bch_bbio_free(bio, b->c);
 315
 316         if (btree_node_io_error(b))
 317                 goto err;
 318
 319         bch_btree_node_read_done(b);
 320         bch_time_stats_update(&b->c->btree_read_time, start_time);
 321
 322         return;
 323 err:
 324         bch_cache_set_error(b->c, "io error reading bucket %zu",
 325                             PTR_BUCKET_NR(b->c, &b->key, 0));
 326 }
 327
 328 static void btree_complete_write(struct btree *b, struct btree_write *w)
 329 {
 330         if (w->prio_blocked &&
 331             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 332                 wake_up_allocators(b->c);
 333
 334         if (w->journal) {
 335                 atomic_dec_bug(w->journal);
 336                 __closure_wake_up(&b->c->journal.wait);
 337         }
 338
 339         w->prio_blocked = 0;
 340         w->journal      = NULL;
 341 }
 342
 343 static void __btree_node_write_done(struct closure *cl)
 344 {
 345         struct btree *b = container_of(cl, struct btree, io.cl);
 346         struct btree_write *w = btree_prev_write(b);
 347
 348         bch_bbio_free(b->bio, b->c);
 349         b->bio = NULL;
 350         btree_complete_write(b, w);
 351
 352         if (btree_node_dirty(b))
 353                 queue_delayed_work(btree_io_wq, &b->work,
 354                                    msecs_to_jiffies(30000));
 355
 356         closure_return(cl);
 357 }
 358
 359 static void btree_node_write_done(struct closure *cl)
 360 {
 361         struct btree *b = container_of(cl, struct btree, io.cl);
 362         struct bio_vec *bv;
 363         int n;
 364
 365         __bio_for_each_segment(bv, b->bio, n, 0)
 366                 __free_page(bv->bv_page);
 367
 368         __btree_node_write_done(cl);
 369 }
 370
 371 static void btree_node_write_endio(struct bio *bio, int error)
 372 {
 373         struct closure *cl = bio->bi_private;
 374         struct btree *b = container_of(cl, struct btree, io.cl);
 375
 376         if (error)
 377                 set_btree_node_io_error(b);
 378
 379         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 380         closure_put(cl);
 381 }
 382
 383 static void do_btree_node_write(struct btree *b)
 384 {
 385         struct closure *cl = &b->io.cl;
 386         struct bset *i = b->sets[b->nsets].data;
 387         BKEY_PADDED(key) k;
 388
 389         i->version      = BCACHE_BSET_VERSION;
 390         i->csum         = btree_csum_set(b, i);
 391
 392         BUG_ON(b->bio);
 393         b->bio = bch_bbio_alloc(b->c);
 394
 395         b->bio->bi_end_io       = btree_node_write_endio;
 396         b->bio->bi_private      = cl;
 397         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 398         b->bio->bi_size         = set_blocks(i, b->c) * block_bytes(b->c);
 399         bch_bio_map(b->bio, i);
 400
 401         /*
 402          * If we're appending to a leaf node, we don't technically need FUA -
 403          * this write just needs to be persisted before the next journal write,
 404          * which will be marked FLUSH|FUA.
 405          *
 406          * Similarly if we're writing a new btree root - the pointer is going to
 407          * be in the next journal entry.
 408          *
 409          * But if we're writing a new btree node (that isn't a root) or
 410          * appending to a non leaf btree node, we need either FUA or a flush
 411          * when we write the parent with the new pointer. FUA is cheaper than a
 412          * flush, and writes appending to leaf nodes aren't blocking anything so
 413          * just make all btree node writes FUA to keep things sane.
 414          */
 415
 416         bkey_copy(&k.key, &b->key);
 417         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
 418
 419         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 420                 int j;
 421                 struct bio_vec *bv;
 422                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 423
 424                 bio_for_each_segment(bv, b->bio, j)
 425                         memcpy(page_address(bv->bv_page),
 426                                base + j * PAGE_SIZE, PAGE_SIZE);
 427
 428                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 429
 430                 continue_at(cl, btree_node_write_done, NULL);
 431         } else {
 432                 b->bio->bi_vcnt = 0;
 433                 bch_bio_map(b->bio, i);
 434
 435                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 436
 437                 closure_sync(cl);
 438                 __btree_node_write_done(cl);
 439         }
 440 }
 441
 442 void bch_btree_node_write(struct btree *b, struct closure *parent)
 443 {
 444         struct bset *i = b->sets[b->nsets].data;
 445
 446         trace_bcache_btree_write(b);
 447
 448         BUG_ON(current->bio_list);
 449         BUG_ON(b->written >= btree_blocks(b));
 450         BUG_ON(b->written && !i->keys);
 451         BUG_ON(b->sets->data->seq != i->seq);
 452         bch_check_keys(b, "writing");
 453
 454         cancel_delayed_work(&b->work);
 455
 456         /* If caller isn't waiting for write, parent refcount is cache set */
 457         closure_lock(&b->io, parent ?: &b->c->cl);
 458
 459         clear_bit(BTREE_NODE_dirty,      &b->flags);
 460         change_bit(BTREE_NODE_write_idx, &b->flags);
 461
 462         do_btree_node_write(b);
 463
 464         b->written += set_blocks(i, b->c);
 465         atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
 466                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 467
 468         bch_btree_sort_lazy(b);
 469
 470         if (b->written < btree_blocks(b))
 471                 bch_bset_init_next(b);
 472 }
 473
 474 static void bch_btree_node_write_sync(struct btree *b)
 475 {
 476         struct closure cl;
 477
 478         closure_init_stack(&cl);
 479         bch_btree_node_write(b, &cl);
 480         closure_sync(&cl);
 481 }
 482
 483 static void btree_node_write_work(struct work_struct *w)
 484 {
 485         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 486
 487         rw_lock(true, b, b->level);
 488
 489         if (btree_node_dirty(b))
 490                 bch_btree_node_write(b, NULL);
 491         rw_unlock(true, b);
 492 }
 493
 494 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 495 {
 496         struct bset *i = b->sets[b->nsets].data;
 497         struct btree_write *w = btree_current_write(b);
 498
 499         BUG_ON(!b->written);
 500         BUG_ON(!i->keys);
 501
 502         if (!btree_node_dirty(b))
 503                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 504
 505         set_btree_node_dirty(b);
 506
 507         if (journal_ref) {
 508                 if (w->journal &&
 509                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 510                         atomic_dec_bug(w->journal);
 511                         w->journal = NULL;
 512                 }
 513
 514                 if (!w->journal) {
 515                         w->journal = journal_ref;
 516                         atomic_inc(w->journal);
 517                 }
 518         }
 519
 520         /* Force write if set is too big */
 521         if (set_bytes(i) > PAGE_SIZE - 48 &&
 522             !current->bio_list)
 523                 bch_btree_node_write(b, NULL);
 524 }
 525
 526 /*
 527  * Btree in memory cache - allocation/freeing
 528  * mca -> memory cache
 529  */
 530
 531 static void mca_reinit(struct btree *b)
 532 {
 533         unsigned i;
 534
 535         b->flags        = 0;
 536         b->written      = 0;
 537         b->nsets        = 0;
 538
 539         for (i = 0; i < MAX_BSETS; i++)
 540                 b->sets[i].size = 0;
 541         /*
 542          * Second loop starts at 1 because b->sets[0]->data is the memory we
 543          * allocated
 544          */
 545         for (i = 1; i < MAX_BSETS; i++)
 546                 b->sets[i].data = NULL;
 547 }
 548
 549 #define mca_reserve(c)  (((c->root && c->root->level)           \
 550                           ? c->root->level : 1) * 8 + 16)
 551 #define mca_can_free(c)                                         \
 552         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 553
 554 static void mca_data_free(struct btree *b)
 555 {
 556         struct bset_tree *t = b->sets;
 557         BUG_ON(!closure_is_unlocked(&b->io.cl));
 558
 559         if (bset_prev_bytes(b) < PAGE_SIZE)
 560                 kfree(t->prev);
 561         else
 562                 free_pages((unsigned long) t->prev,
 563                            get_order(bset_prev_bytes(b)));
 564
 565         if (bset_tree_bytes(b) < PAGE_SIZE)
 566                 kfree(t->tree);
 567         else
 568                 free_pages((unsigned long) t->tree,
 569                            get_order(bset_tree_bytes(b)));
 570
 571         free_pages((unsigned long) t->data, b->page_order);
 572
 573         t->prev = NULL;
 574         t->tree = NULL;
 575         t->data = NULL;
 576         list_move(&b->list, &b->c->btree_cache_freed);
 577         b->c->bucket_cache_used--;
 578 }
 579
 580 static void mca_bucket_free(struct btree *b)
 581 {
 582         BUG_ON(btree_node_dirty(b));
 583
 584         b->key.ptr[0] = 0;
 585         hlist_del_init_rcu(&b->hash);
 586         list_move(&b->list, &b->c->btree_cache_freeable);
 587 }
 588
 589 static unsigned btree_order(struct bkey *k)
 590 {
 591         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 592 }
 593
 594 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 595 {
 596         struct bset_tree *t = b->sets;
 597         BUG_ON(t->data);
 598
 599         b->page_order = max_t(unsigned,
 600                               ilog2(b->c->btree_pages),
 601                               btree_order(k));
 602
 603         t->data = (void *) __get_free_pages(gfp, b->page_order);
 604         if (!t->data)
 605                 goto err;
 606
 607         t->tree = bset_tree_bytes(b) < PAGE_SIZE
 608                 ? kmalloc(bset_tree_bytes(b), gfp)
 609                 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
 610         if (!t->tree)
 611                 goto err;
 612
 613         t->prev = bset_prev_bytes(b) < PAGE_SIZE
 614                 ? kmalloc(bset_prev_bytes(b), gfp)
 615                 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
 616         if (!t->prev)
 617                 goto err;
 618
 619         list_move(&b->list, &b->c->btree_cache);
 620         b->c->bucket_cache_used++;
 621         return;
 622 err:
 623         mca_data_free(b);
 624 }
 625
 626 static struct btree *mca_bucket_alloc(struct cache_set *c,
 627                                       struct bkey *k, gfp_t gfp)
 628 {
 629         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 630         if (!b)
 631                 return NULL;
 632
 633         init_rwsem(&b->lock);
 634         lockdep_set_novalidate_class(&b->lock);
 635         INIT_LIST_HEAD(&b->list);
 636         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 637         b->c = c;
 638         closure_init_unlocked(&b->io);
 639
 640         mca_data_alloc(b, k, gfp);
 641         return b;
 642 }
 643
 644 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 645 {
 646         struct closure cl;
 647
 648         closure_init_stack(&cl);
 649         lockdep_assert_held(&b->c->bucket_lock);
 650
 651         if (!down_write_trylock(&b->lock))
 652                 return -ENOMEM;
 653
 654         BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
 655
 656         if (b->page_order < min_order ||
 657             (!flush &&
 658              (btree_node_dirty(b) ||
 659               atomic_read(&b->io.cl.remaining) != -1))) {
 660                 rw_unlock(true, b);
 661                 return -ENOMEM;
 662         }
 663
 664         if (btree_node_dirty(b))
 665                 bch_btree_node_write_sync(b);
 666
 667         /* wait for any in flight btree write */
 668         closure_wait_event(&b->io.wait, &cl,
 669                            atomic_read(&b->io.cl.remaining) == -1);
 670
 671         return 0;
 672 }
 673
 674 static unsigned long bch_mca_scan(struct shrinker *shrink,
 675                                   struct shrink_control *sc)
 676 {
 677         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 678         struct btree *b, *t;
 679         unsigned long i, nr = sc->nr_to_scan;
 680         unsigned long freed = 0;
 681
 682         if (c->shrinker_disabled)
 683                 return SHRINK_STOP;
 684
 685         if (c->try_harder)
 686                 return SHRINK_STOP;
 687
 688         /* Return -1 if we can't do anything right now */
 689         if (sc->gfp_mask & __GFP_IO)
 690                 mutex_lock(&c->bucket_lock);
 691         else if (!mutex_trylock(&c->bucket_lock))
 692                 return -1;
 693
 694         /*
 695          * It's _really_ critical that we don't free too many btree nodes - we
 696          * have to always leave ourselves a reserve. The reserve is how we
 697          * guarantee that allocating memory for a new btree node can always
 698          * succeed, so that inserting keys into the btree can always succeed and
 699          * IO can always make forward progress:
 700          */
 701         nr /= c->btree_pages;
 702         nr = min_t(unsigned long, nr, mca_can_free(c));
 703
 704         i = 0;
 705         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 706                 if (freed >= nr)
 707                         break;
 708
 709                 if (++i > 3 &&
 710                     !mca_reap(b, 0, false)) {
 711                         mca_data_free(b);
 712                         rw_unlock(true, b);
 713                         freed++;
 714                 }
 715         }
 716
 717         /*
 718          * Can happen right when we first start up, before we've read in any
 719          * btree nodes
 720          */
 721         if (list_empty(&c->btree_cache))
 722                 goto out;
 723
 724         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 725                 b = list_first_entry(&c->btree_cache, struct btree, list);
 726                 list_rotate_left(&c->btree_cache);
 727
 728                 if (!b->accessed &&
 729                     !mca_reap(b, 0, false)) {
 730                         mca_bucket_free(b);
 731                         mca_data_free(b);
 732                         rw_unlock(true, b);
 733                         freed++;
 734                 } else
 735                         b->accessed = 0;
 736         }
 737 out:
 738         mutex_unlock(&c->bucket_lock);
 739         return freed;
 740 }
 741
 742 static unsigned long bch_mca_count(struct shrinker *shrink,
 743                                    struct shrink_control *sc)
 744 {
 745         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 746
 747         if (c->shrinker_disabled)
 748                 return 0;
 749
 750         if (c->try_harder)
 751                 return 0;
 752
 753         return mca_can_free(c) * c->btree_pages;
 754 }
 755
 756 void bch_btree_cache_free(struct cache_set *c)
 757 {
 758         struct btree *b;
 759         struct closure cl;
 760         closure_init_stack(&cl);
 761
 762         if (c->shrink.list.next)
 763                 unregister_shrinker(&c->shrink);
 764
 765         mutex_lock(&c->bucket_lock);
 766
 767 #ifdef CONFIG_BCACHE_DEBUG
 768         if (c->verify_data)
 769                 list_move(&c->verify_data->list, &c->btree_cache);
 770 #endif
 771
 772         list_splice(&c->btree_cache_freeable,
 773                     &c->btree_cache);
 774
 775         while (!list_empty(&c->btree_cache)) {
 776                 b = list_first_entry(&c->btree_cache, struct btree, list);
 777
 778                 if (btree_node_dirty(b))
 779                         btree_complete_write(b, btree_current_write(b));
 780                 clear_bit(BTREE_NODE_dirty, &b->flags);
 781
 782                 mca_data_free(b);
 783         }
 784
 785         while (!list_empty(&c->btree_cache_freed)) {
 786                 b = list_first_entry(&c->btree_cache_freed,
 787                                      struct btree, list);
 788                 list_del(&b->list);
 789                 cancel_delayed_work_sync(&b->work);
 790                 kfree(b);
 791         }
 792
 793         mutex_unlock(&c->bucket_lock);
 794 }
 795
 796 int bch_btree_cache_alloc(struct cache_set *c)
 797 {
 798         unsigned i;
 799
 800         for (i = 0; i < mca_reserve(c); i++)
 801                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 802                         return -ENOMEM;
 803
 804         list_splice_init(&c->btree_cache,
 805                          &c->btree_cache_freeable);
 806
 807 #ifdef CONFIG_BCACHE_DEBUG
 808         mutex_init(&c->verify_lock);
 809
 810         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 811
 812         if (c->verify_data &&
 813             c->verify_data->sets[0].data)
 814                 list_del_init(&c->verify_data->list);
 815         else
 816                 c->verify_data = NULL;
 817 #endif
 818
 819         c->shrink.count_objects = bch_mca_count;
 820         c->shrink.scan_objects = bch_mca_scan;
 821         c->shrink.seeks = 4;
 822         c->shrink.batch = c->btree_pages * 2;
 823         register_shrinker(&c->shrink);
 824
 825         return 0;
 826 }
 827
 828 /* Btree in memory cache - hash table */
 829
 830 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 831 {
 832         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 833 }
 834
 835 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 836 {
 837         struct btree *b;
 838
 839         rcu_read_lock();
 840         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 841                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 842                         goto out;
 843         b = NULL;
 844 out:
 845         rcu_read_unlock();
 846         return b;
 847 }
 848
 849 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 850 {
 851         struct btree *b;
 852
 853         trace_bcache_btree_cache_cannibalize(c);
 854
 855         if (!c->try_harder) {
 856                 c->try_harder = current;
 857                 c->try_harder_start = local_clock();
 858         } else if (c->try_harder != current)
 859                 return ERR_PTR(-ENOSPC);
 860
 861         list_for_each_entry_reverse(b, &c->btree_cache, list)
 862                 if (!mca_reap(b, btree_order(k), false))
 863                         return b;
 864
 865         list_for_each_entry_reverse(b, &c->btree_cache, list)
 866                 if (!mca_reap(b, btree_order(k), true))
 867                         return b;
 868
 869         return ERR_PTR(-ENOMEM);
 870 }
 871
 872 /*
 873  * We can only have one thread cannibalizing other cached btree nodes at a time,
 874  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 875  * cannibalize_bucket() will take. This means every time we unlock the root of
 876  * the btree, we need to release this lock if we have it held.
 877  */
 878 static void bch_cannibalize_unlock(struct cache_set *c)
 879 {
 880         if (c->try_harder == current) {
 881                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 882                 c->try_harder = NULL;
 883                 wake_up(&c->try_wait);
 884         }
 885 }
 886
 887 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level)
 888 {
 889         struct btree *b;
 890
 891         BUG_ON(current->bio_list);
 892
 893         lockdep_assert_held(&c->bucket_lock);
 894
 895         if (mca_find(c, k))
 896                 return NULL;
 897
 898         /* btree_free() doesn't free memory; it sticks the node on the end of
 899          * the list. Check if there's any freed nodes there:
 900          */
 901         list_for_each_entry(b, &c->btree_cache_freeable, list)
 902                 if (!mca_reap(b, btree_order(k), false))
 903                         goto out;
 904
 905         /* We never free struct btree itself, just the memory that holds the on
 906          * disk node. Check the freed list before allocating a new one:
 907          */
 908         list_for_each_entry(b, &c->btree_cache_freed, list)
 909                 if (!mca_reap(b, 0, false)) {
 910                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 911                         if (!b->sets[0].data)
 912                                 goto err;
 913                         else
 914                                 goto out;
 915                 }
 916
 917         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 918         if (!b)
 919                 goto err;
 920
 921         BUG_ON(!down_write_trylock(&b->lock));
 922         if (!b->sets->data)
 923                 goto err;
 924 out:
 925         BUG_ON(!closure_is_unlocked(&b->io.cl));
 926
 927         bkey_copy(&b->key, k);
 928         list_move(&b->list, &c->btree_cache);
 929         hlist_del_init_rcu(&b->hash);
 930         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 931
 932         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 933         b->level        = level;
 934         b->parent       = (void *) ~0UL;
 935
 936         mca_reinit(b);
 937
 938         return b;
 939 err:
 940         if (b)
 941                 rw_unlock(true, b);
 942
 943         b = mca_cannibalize(c, k);
 944         if (!IS_ERR(b))
 945                 goto out;
 946
 947         return b;
 948 }
 949
 950 /**
 951  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 952  * in from disk if necessary.
 953  *
 954  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 955  *
 956  * The btree node will have either a read or a write lock held, depending on
 957  * level and op->lock.
 958  */
 959 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 960                                  int level, bool write)
 961 {
 962         int i = 0;
 963         struct btree *b;
 964
 965         BUG_ON(level < 0);
 966 retry:
 967         b = mca_find(c, k);
 968
 969         if (!b) {
 970                 if (current->bio_list)
 971                         return ERR_PTR(-EAGAIN);
 972
 973                 mutex_lock(&c->bucket_lock);
 974                 b = mca_alloc(c, k, level);
 975                 mutex_unlock(&c->bucket_lock);
 976
 977                 if (!b)
 978                         goto retry;
 979                 if (IS_ERR(b))
 980                         return b;
 981
 982                 bch_btree_node_read(b);
 983
 984                 if (!write)
 985                         downgrade_write(&b->lock);
 986         } else {
 987                 rw_lock(write, b, level);
 988                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 989                         rw_unlock(write, b);
 990                         goto retry;
 991                 }
 992                 BUG_ON(b->level != level);
 993         }
 994
 995         b->accessed = 1;
 996
 997         for (; i <= b->nsets && b->sets[i].size; i++) {
 998                 prefetch(b->sets[i].tree);
 999                 prefetch(b->sets[i].data);
1000         }
1001
1002         for (; i <= b->nsets; i++)
1003                 prefetch(b->sets[i].data);
1004
1005         if (btree_node_io_error(b)) {
1006                 rw_unlock(write, b);
1007                 return ERR_PTR(-EIO);
1008         }
1009
1010         BUG_ON(!b->written);
1011
1012         return b;
1013 }
1014
1015 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1016 {
1017         struct btree *b;
1018
1019         mutex_lock(&c->bucket_lock);
1020         b = mca_alloc(c, k, level);
1021         mutex_unlock(&c->bucket_lock);
1022
1023         if (!IS_ERR_OR_NULL(b)) {
1024                 bch_btree_node_read(b);
1025                 rw_unlock(true, b);
1026         }
1027 }
1028
1029 /* Btree alloc */
1030
1031 static void btree_node_free(struct btree *b)
1032 {
1033         unsigned i;
1034
1035         trace_bcache_btree_node_free(b);
1036
1037         BUG_ON(b == b->c->root);
1038
1039         if (btree_node_dirty(b))
1040                 btree_complete_write(b, btree_current_write(b));
1041         clear_bit(BTREE_NODE_dirty, &b->flags);
1042
1043         cancel_delayed_work(&b->work);
1044
1045         mutex_lock(&b->c->bucket_lock);
1046
1047         for (i = 0; i < KEY_PTRS(&b->key); i++) {
1048                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
1049
1050                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1051                             PTR_BUCKET(b->c, &b->key, i));
1052         }
1053
1054         bch_bucket_free(b->c, &b->key);
1055         mca_bucket_free(b);
1056         mutex_unlock(&b->c->bucket_lock);
1057 }
1058
1059 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1060 {
1061         BKEY_PADDED(key) k;
1062         struct btree *b = ERR_PTR(-EAGAIN);
1063
1064         mutex_lock(&c->bucket_lock);
1065 retry:
1066         if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, wait))
1067                 goto err;
1068
1069         bkey_put(c, &k.key);
1070         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1071
1072         b = mca_alloc(c, &k.key, level);
1073         if (IS_ERR(b))
1074                 goto err_free;
1075
1076         if (!b) {
1077                 cache_bug(c,
1078                         "Tried to allocate bucket that was in btree cache");
1079                 goto retry;
1080         }
1081
1082         b->accessed = 1;
1083         bch_bset_init_next(b);
1084
1085         mutex_unlock(&c->bucket_lock);
1086
1087         trace_bcache_btree_node_alloc(b);
1088         return b;
1089 err_free:
1090         bch_bucket_free(c, &k.key);
1091 err:
1092         mutex_unlock(&c->bucket_lock);
1093
1094         trace_bcache_btree_node_alloc_fail(b);
1095         return b;
1096 }
1097
1098 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1099 {
1100         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1101         if (!IS_ERR_OR_NULL(n))
1102                 bch_btree_sort_into(b, n);
1103
1104         return n;
1105 }
1106
1107 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1108 {
1109         unsigned i;
1110
1111         bkey_copy(k, &b->key);
1112         bkey_copy_key(k, &ZERO_KEY);
1113
1114         for (i = 0; i < KEY_PTRS(k); i++) {
1115                 uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1;
1116
1117                 SET_PTR_GEN(k, i, g);
1118         }
1119
1120         atomic_inc(&b->c->prio_blocked);
1121 }
1122
1123 /* Garbage collection */
1124
1125 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1126 {
1127         uint8_t stale = 0;
1128         unsigned i;
1129         struct bucket *g;
1130
1131         /*
1132          * ptr_invalid() can't return true for the keys that mark btree nodes as
1133          * freed, but since ptr_bad() returns true we'll never actually use them
1134          * for anything and thus we don't want mark their pointers here
1135          */
1136         if (!bkey_cmp(k, &ZERO_KEY))
1137                 return stale;
1138
1139         for (i = 0; i < KEY_PTRS(k); i++) {
1140                 if (!ptr_available(c, k, i))
1141                         continue;
1142
1143                 g = PTR_BUCKET(c, k, i);
1144
1145                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1146                         g->gc_gen = PTR_GEN(k, i);
1147
1148                 if (ptr_stale(c, k, i)) {
1149                         stale = max(stale, ptr_stale(c, k, i));
1150                         continue;
1151                 }
1152
1153                 cache_bug_on(GC_MARK(g) &&
1154                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1155                              c, "inconsistent ptrs: mark = %llu, level = %i",
1156                              GC_MARK(g), level);
1157
1158                 if (level)
1159                         SET_GC_MARK(g, GC_MARK_METADATA);
1160                 else if (KEY_DIRTY(k))
1161                         SET_GC_MARK(g, GC_MARK_DIRTY);
1162
1163                 /* guard against overflow */
1164                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1165                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1166                                              (1 << 14) - 1));
1167
1168                 BUG_ON(!GC_SECTORS_USED(g));
1169         }
1170
1171         return stale;
1172 }
1173
1174 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1175
1176 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1177 {
1178         uint8_t stale = 0;
1179         unsigned keys = 0, good_keys = 0;
1180         struct bkey *k;
1181         struct btree_iter iter;
1182         struct bset_tree *t;
1183
1184         gc->nodes++;
1185
1186         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1187                 stale = max(stale, btree_mark_key(b, k));
1188                 keys++;
1189
1190                 if (bch_ptr_bad(b, k))
1191                         continue;
1192
1193                 gc->key_bytes += bkey_u64s(k);
1194                 gc->nkeys++;
1195                 good_keys++;
1196
1197                 gc->data += KEY_SIZE(k);
1198         }
1199
1200         for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1201                 btree_bug_on(t->size &&
1202                              bset_written(b, t) &&
1203                              bkey_cmp(&b->key, &t->end) < 0,
1204                              b, "found short btree key in gc");
1205
1206         if (b->c->gc_always_rewrite)
1207                 return true;
1208
1209         if (stale > 10)
1210                 return true;
1211
1212         if ((keys - good_keys) * 2 > keys)
1213                 return true;
1214
1215         return false;
1216 }
1217
1218 #define GC_MERGE_NODES  4U
1219
1220 struct gc_merge_info {
1221         struct btree    *b;
1222         unsigned        keys;
1223 };
1224
1225 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1226                                  struct keylist *, atomic_t *, struct bkey *);
1227
1228 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1229                              struct keylist *keylist, struct gc_stat *gc,
1230                              struct gc_merge_info *r)
1231 {
1232         unsigned i, nodes = 0, keys = 0, blocks;
1233         struct btree *new_nodes[GC_MERGE_NODES];
1234         struct closure cl;
1235         struct bkey *k;
1236
1237         memset(new_nodes, 0, sizeof(new_nodes));
1238         closure_init_stack(&cl);
1239
1240         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1241                 keys += r[nodes++].keys;
1242
1243         blocks = btree_default_blocks(b->c) * 2 / 3;
1244
1245         if (nodes < 2 ||
1246             __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
1247                 return 0;
1248
1249         for (i = 0; i < nodes; i++) {
1250                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1251                 if (IS_ERR_OR_NULL(new_nodes[i]))
1252                         goto out_nocoalesce;
1253         }
1254
1255         for (i = nodes - 1; i > 0; --i) {
1256                 struct bset *n1 = new_nodes[i]->sets->data;
1257                 struct bset *n2 = new_nodes[i - 1]->sets->data;
1258                 struct bkey *k, *last = NULL;
1259
1260                 keys = 0;
1261
1262                 if (i > 1) {
1263                         for (k = n2->start;
1264                              k < end(n2);
1265                              k = bkey_next(k)) {
1266                                 if (__set_blocks(n1, n1->keys + keys +
1267                                                  bkey_u64s(k), b->c) > blocks)
1268                                         break;
1269
1270                                 last = k;
1271                                 keys += bkey_u64s(k);
1272                         }
1273                 } else {
1274                         /*
1275                          * Last node we're not getting rid of - we're getting
1276                          * rid of the node at r[0]. Have to try and fit all of
1277                          * the remaining keys into this node; we can't ensure
1278                          * they will always fit due to rounding and variable
1279                          * length keys (shouldn't be possible in practice,
1280                          * though)
1281                          */
1282                         if (__set_blocks(n1, n1->keys + n2->keys,
1283                                          b->c) > btree_blocks(new_nodes[i]))
1284                                 goto out_nocoalesce;
1285
1286                         keys = n2->keys;
1287                         /* Take the key of the node we're getting rid of */
1288                         last = &r->b->key;
1289                 }
1290
1291                 BUG_ON(__set_blocks(n1, n1->keys + keys,
1292                                     b->c) > btree_blocks(new_nodes[i]));
1293
1294                 if (last)
1295                         bkey_copy_key(&new_nodes[i]->key, last);
1296
1297                 memcpy(end(n1),
1298                        n2->start,
1299                        (void *) node(n2, keys) - (void *) n2->start);
1300
1301                 n1->keys += keys;
1302                 r[i].keys = n1->keys;
1303
1304                 memmove(n2->start,
1305                         node(n2, keys),
1306                         (void *) end(n2) - (void *) node(n2, keys));
1307
1308                 n2->keys -= keys;
1309
1310                 if (bch_keylist_realloc(keylist,
1311                                         KEY_PTRS(&new_nodes[i]->key), b->c))
1312                         goto out_nocoalesce;
1313
1314                 bch_btree_node_write(new_nodes[i], &cl);
1315                 bch_keylist_add(keylist, &new_nodes[i]->key);
1316         }
1317
1318         for (i = 0; i < nodes; i++) {
1319                 if (bch_keylist_realloc(keylist, KEY_PTRS(&r[i].b->key), b->c))
1320                         goto out_nocoalesce;
1321
1322                 make_btree_freeing_key(r[i].b, keylist->top);
1323                 bch_keylist_push(keylist);
1324         }
1325
1326         /* We emptied out this node */
1327         BUG_ON(new_nodes[0]->sets->data->keys);
1328         btree_node_free(new_nodes[0]);
1329         rw_unlock(true, new_nodes[0]);
1330
1331         closure_sync(&cl);
1332
1333         for (i = 0; i < nodes; i++) {
1334                 btree_node_free(r[i].b);
1335                 rw_unlock(true, r[i].b);
1336
1337                 r[i].b = new_nodes[i];
1338         }
1339
1340         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1341         BUG_ON(!bch_keylist_empty(keylist));
1342
1343         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1344         r[nodes - 1].b = ERR_PTR(-EINTR);
1345
1346         trace_bcache_btree_gc_coalesce(nodes);
1347         gc->nodes--;
1348
1349         /* Invalidated our iterator */
1350         return -EINTR;
1351
1352 out_nocoalesce:
1353         closure_sync(&cl);
1354
1355         while ((k = bch_keylist_pop(keylist)))
1356                 if (!bkey_cmp(k, &ZERO_KEY))
1357                         atomic_dec(&b->c->prio_blocked);
1358
1359         for (i = 0; i < nodes; i++)
1360                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1361                         btree_node_free(new_nodes[i]);
1362                         rw_unlock(true, new_nodes[i]);
1363                 }
1364         return 0;
1365 }
1366
1367 static unsigned btree_gc_count_keys(struct btree *b)
1368 {
1369         struct bkey *k;
1370         struct btree_iter iter;
1371         unsigned ret = 0;
1372
1373         for_each_key_filter(b, k, &iter, bch_ptr_bad)
1374                 ret += bkey_u64s(k);
1375
1376         return ret;
1377 }
1378
1379 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1380                             struct closure *writes, struct gc_stat *gc)
1381 {
1382         unsigned i;
1383         int ret = 0;
1384         bool should_rewrite;
1385         struct btree *n;
1386         struct bkey *k;
1387         struct keylist keys;
1388         struct btree_iter iter;
1389         struct gc_merge_info r[GC_MERGE_NODES];
1390         struct gc_merge_info *last = r + GC_MERGE_NODES - 1;
1391
1392         bch_keylist_init(&keys);
1393         bch_btree_iter_init(b, &iter, &b->c->gc_done);
1394
1395         for (i = 0; i < GC_MERGE_NODES; i++)
1396                 r[i].b = ERR_PTR(-EINTR);
1397
1398         while (1) {
1399                 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1400                 if (k) {
1401                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1402                         if (IS_ERR(r->b)) {
1403                                 ret = PTR_ERR(r->b);
1404                                 break;
1405                         }
1406
1407                         r->keys = btree_gc_count_keys(r->b);
1408
1409                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1410                         if (ret)
1411                                 break;
1412                 }
1413
1414                 if (!last->b)
1415                         break;
1416
1417                 if (!IS_ERR(last->b)) {
1418                         should_rewrite = btree_gc_mark_node(last->b, gc);
1419                         if (should_rewrite) {
1420                                 n = btree_node_alloc_replacement(last->b,
1421                                                                  false);
1422
1423                                 if (!IS_ERR_OR_NULL(n)) {
1424                                         bch_btree_node_write_sync(n);
1425                                         bch_keylist_add(&keys, &n->key);
1426
1427                                         make_btree_freeing_key(last->b,
1428                                                                keys.top);
1429                                         bch_keylist_push(&keys);
1430
1431                                         btree_node_free(last->b);
1432
1433                                         bch_btree_insert_node(b, op, &keys,
1434                                                               NULL, NULL);
1435                                         BUG_ON(!bch_keylist_empty(&keys));
1436
1437                                         rw_unlock(true, last->b);
1438                                         last->b = n;
1439
1440                                         /* Invalidated our iterator */
1441                                         ret = -EINTR;
1442                                         break;
1443                                 }
1444                         }
1445
1446                         if (last->b->level) {
1447                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1448                                 if (ret)
1449                                         break;
1450                         }
1451
1452                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1453
1454                         /*
1455                          * Must flush leaf nodes before gc ends, since replace
1456                          * operations aren't journalled
1457                          */
1458                         if (btree_node_dirty(last->b))
1459                                 bch_btree_node_write(last->b, writes);
1460                         rw_unlock(true, last->b);
1461                 }
1462
1463                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1464                 r->b = NULL;
1465
1466                 if (need_resched()) {
1467                         ret = -EAGAIN;
1468                         break;
1469                 }
1470         }
1471
1472         for (i = 0; i < GC_MERGE_NODES; i++)
1473                 if (!IS_ERR_OR_NULL(r[i].b)) {
1474                         if (btree_node_dirty(r[i].b))
1475                                 bch_btree_node_write(r[i].b, writes);
1476                         rw_unlock(true, r[i].b);
1477                 }
1478
1479         bch_keylist_free(&keys);
1480
1481         return ret;
1482 }
1483
1484 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1485                              struct closure *writes, struct gc_stat *gc)
1486 {
1487         struct btree *n = NULL;
1488         int ret = 0;
1489         bool should_rewrite;
1490
1491         should_rewrite = btree_gc_mark_node(b, gc);
1492         if (should_rewrite) {
1493                 n = btree_node_alloc_replacement(b, false);
1494
1495                 if (!IS_ERR_OR_NULL(n)) {
1496                         bch_btree_node_write_sync(n);
1497                         bch_btree_set_root(n);
1498                         btree_node_free(b);
1499                         rw_unlock(true, n);
1500
1501                         return -EINTR;
1502                 }
1503         }
1504
1505         if (b->level) {
1506                 ret = btree_gc_recurse(b, op, writes, gc);
1507                 if (ret)
1508                         return ret;
1509         }
1510
1511         bkey_copy_key(&b->c->gc_done, &b->key);
1512
1513         return ret;
1514 }
1515
1516 static void btree_gc_start(struct cache_set *c)
1517 {
1518         struct cache *ca;
1519         struct bucket *b;
1520         unsigned i;
1521
1522         if (!c->gc_mark_valid)
1523                 return;
1524
1525         mutex_lock(&c->bucket_lock);
1526
1527         c->gc_mark_valid = 0;
1528         c->gc_done = ZERO_KEY;
1529
1530         for_each_cache(ca, c, i)
1531                 for_each_bucket(b, ca) {
1532                         b->gc_gen = b->gen;
1533                         if (!atomic_read(&b->pin)) {
1534                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1535                                 SET_GC_SECTORS_USED(b, 0);
1536                         }
1537                 }
1538
1539         mutex_unlock(&c->bucket_lock);
1540 }
1541
1542 size_t bch_btree_gc_finish(struct cache_set *c)
1543 {
1544         size_t available = 0;
1545         struct bucket *b;
1546         struct cache *ca;
1547         unsigned i;
1548
1549         mutex_lock(&c->bucket_lock);
1550
1551         set_gc_sectors(c);
1552         c->gc_mark_valid = 1;
1553         c->need_gc      = 0;
1554
1555         if (c->root)
1556                 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1557                         SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1558                                     GC_MARK_METADATA);
1559
1560         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1561                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1562                             GC_MARK_METADATA);
1563
1564         for_each_cache(ca, c, i) {
1565                 uint64_t *i;
1566
1567                 ca->invalidate_needs_gc = 0;
1568
1569                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1570                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1571
1572                 for (i = ca->prio_buckets;
1573                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1574                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1575
1576                 for_each_bucket(b, ca) {
1577                         b->last_gc      = b->gc_gen;
1578                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1579
1580                         if (!atomic_read(&b->pin) &&
1581                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1582                                 available++;
1583                                 if (!GC_SECTORS_USED(b))
1584                                         bch_bucket_add_unused(ca, b);
1585                         }
1586                 }
1587         }
1588
1589         mutex_unlock(&c->bucket_lock);
1590         return available;
1591 }
1592
1593 static void bch_btree_gc(struct cache_set *c)
1594 {
1595         int ret;
1596         unsigned long available;
1597         struct gc_stat stats;
1598         struct closure writes;
1599         struct btree_op op;
1600         uint64_t start_time = local_clock();
1601
1602         trace_bcache_gc_start(c);
1603
1604         memset(&stats, 0, sizeof(struct gc_stat));
1605         closure_init_stack(&writes);
1606         bch_btree_op_init(&op, SHRT_MAX);
1607
1608         btree_gc_start(c);
1609
1610         do {
1611                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1612                 closure_sync(&writes);
1613
1614                 if (ret && ret != -EAGAIN)
1615                         pr_warn("gc failed!");
1616         } while (ret);
1617
1618         available = bch_btree_gc_finish(c);
1619         wake_up_allocators(c);
1620
1621         bch_time_stats_update(&c->btree_gc_time, start_time);
1622
1623         stats.key_bytes *= sizeof(uint64_t);
1624         stats.data      <<= 9;
1625         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1626         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1627
1628         trace_bcache_gc_end(c);
1629
1630         bch_moving_gc(c);
1631 }
1632
1633 static int bch_gc_thread(void *arg)
1634 {
1635         struct cache_set *c = arg;
1636         struct cache *ca;
1637         unsigned i;
1638
1639         while (1) {
1640 again:
1641                 bch_btree_gc(c);
1642
1643                 set_current_state(TASK_INTERRUPTIBLE);
1644                 if (kthread_should_stop())
1645                         break;
1646
1647                 mutex_lock(&c->bucket_lock);
1648
1649                 for_each_cache(ca, c, i)
1650                         if (ca->invalidate_needs_gc) {
1651                                 mutex_unlock(&c->bucket_lock);
1652                                 set_current_state(TASK_RUNNING);
1653                                 goto again;
1654                         }
1655
1656                 mutex_unlock(&c->bucket_lock);
1657
1658                 try_to_freeze();
1659                 schedule();
1660         }
1661
1662         return 0;
1663 }
1664
1665 int bch_gc_thread_start(struct cache_set *c)
1666 {
1667         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1668         if (IS_ERR(c->gc_thread))
1669                 return PTR_ERR(c->gc_thread);
1670
1671         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1672         return 0;
1673 }
1674
1675 /* Initial partial gc */
1676
1677 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1678                                    unsigned long **seen)
1679 {
1680         int ret = 0;
1681         unsigned i;
1682         struct bkey *k, *p = NULL;
1683         struct bucket *g;
1684         struct btree_iter iter;
1685
1686         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1687                 for (i = 0; i < KEY_PTRS(k); i++) {
1688                         if (!ptr_available(b->c, k, i))
1689                                 continue;
1690
1691                         g = PTR_BUCKET(b->c, k, i);
1692
1693                         if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1694                                                 seen[PTR_DEV(k, i)]) ||
1695                             !ptr_stale(b->c, k, i)) {
1696                                 g->gen = PTR_GEN(k, i);
1697
1698                                 if (b->level)
1699                                         g->prio = BTREE_PRIO;
1700                                 else if (g->prio == BTREE_PRIO)
1701                                         g->prio = INITIAL_PRIO;
1702                         }
1703                 }
1704
1705                 btree_mark_key(b, k);
1706         }
1707
1708         if (b->level) {
1709                 bch_btree_iter_init(b, &iter, NULL);
1710
1711                 do {
1712                         k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1713                         if (k)
1714                                 btree_node_prefetch(b->c, k, b->level - 1);
1715
1716                         if (p)
1717                                 ret = btree(check_recurse, p, b, op, seen);
1718
1719                         p = k;
1720                 } while (p && !ret);
1721         }
1722
1723         return 0;
1724 }
1725
1726 int bch_btree_check(struct cache_set *c)
1727 {
1728         int ret = -ENOMEM;
1729         unsigned i;
1730         unsigned long *seen[MAX_CACHES_PER_SET];
1731         struct btree_op op;
1732
1733         memset(seen, 0, sizeof(seen));
1734         bch_btree_op_init(&op, SHRT_MAX);
1735
1736         for (i = 0; c->cache[i]; i++) {
1737                 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1738                 seen[i] = kmalloc(n, GFP_KERNEL);
1739                 if (!seen[i])
1740                         goto err;
1741
1742                 /* Disables the seen array until prio_read() uses it too */
1743                 memset(seen[i], 0xFF, n);
1744         }
1745
1746         ret = btree_root(check_recurse, c, &op, seen);
1747 err:
1748         for (i = 0; i < MAX_CACHES_PER_SET; i++)
1749                 kfree(seen[i]);
1750         return ret;
1751 }
1752
1753 /* Btree insertion */
1754
1755 static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
1756 {
1757         struct bset *i = b->sets[b->nsets].data;
1758
1759         memmove((uint64_t *) where + bkey_u64s(insert),
1760                 where,
1761                 (void *) end(i) - (void *) where);
1762
1763         i->keys += bkey_u64s(insert);
1764         bkey_copy(where, insert);
1765         bch_bset_fix_lookup_table(b, where);
1766 }
1767
1768 static bool fix_overlapping_extents(struct btree *b, struct bkey *insert,
1769                                     struct btree_iter *iter,
1770                                     struct bkey *replace_key)
1771 {
1772         void subtract_dirty(struct bkey *k, uint64_t offset, int sectors)
1773         {
1774                 if (KEY_DIRTY(k))
1775                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1776                                                      offset, -sectors);
1777         }
1778
1779         uint64_t old_offset;
1780         unsigned old_size, sectors_found = 0;
1781
1782         while (1) {
1783                 struct bkey *k = bch_btree_iter_next(iter);
1784                 if (!k ||
1785                     bkey_cmp(&START_KEY(k), insert) >= 0)
1786                         break;
1787
1788                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1789                         continue;
1790
1791                 old_offset = KEY_START(k);
1792                 old_size = KEY_SIZE(k);
1793
1794                 /*
1795                  * We might overlap with 0 size extents; we can't skip these
1796                  * because if they're in the set we're inserting to we have to
1797                  * adjust them so they don't overlap with the key we're
1798                  * inserting. But we don't want to check them for replace
1799                  * operations.
1800                  */
1801
1802                 if (replace_key && KEY_SIZE(k)) {
1803                         /*
1804                          * k might have been split since we inserted/found the
1805                          * key we're replacing
1806                          */
1807                         unsigned i;
1808                         uint64_t offset = KEY_START(k) -
1809                                 KEY_START(replace_key);
1810
1811                         /* But it must be a subset of the replace key */
1812                         if (KEY_START(k) < KEY_START(replace_key) ||
1813                             KEY_OFFSET(k) > KEY_OFFSET(replace_key))
1814                                 goto check_failed;
1815
1816                         /* We didn't find a key that we were supposed to */
1817                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
1818                                 goto check_failed;
1819
1820                         if (KEY_PTRS(replace_key) != KEY_PTRS(k))
1821                                 goto check_failed;
1822
1823                         /* skip past gen */
1824                         offset <<= 8;
1825
1826                         BUG_ON(!KEY_PTRS(replace_key));
1827
1828                         for (i = 0; i < KEY_PTRS(replace_key); i++)
1829                                 if (k->ptr[i] != replace_key->ptr[i] + offset)
1830                                         goto check_failed;
1831
1832                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1833                 }
1834
1835                 if (bkey_cmp(insert, k) < 0 &&
1836                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1837                         /*
1838                          * We overlapped in the middle of an existing key: that
1839                          * means we have to split the old key. But we have to do
1840                          * slightly different things depending on whether the
1841                          * old key has been written out yet.
1842                          */
1843
1844                         struct bkey *top;
1845
1846                         subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert));
1847
1848                         if (bkey_written(b, k)) {
1849                                 /*
1850                                  * We insert a new key to cover the top of the
1851                                  * old key, and the old key is modified in place
1852                                  * to represent the bottom split.
1853                                  *
1854                                  * It's completely arbitrary whether the new key
1855                                  * is the top or the bottom, but it has to match
1856                                  * up with what btree_sort_fixup() does - it
1857                                  * doesn't check for this kind of overlap, it
1858                                  * depends on us inserting a new key for the top
1859                                  * here.
1860                                  */
1861                                 top = bch_bset_search(b, &b->sets[b->nsets],
1862                                                       insert);
1863                                 shift_keys(b, top, k);
1864                         } else {
1865                                 BKEY_PADDED(key) temp;
1866                                 bkey_copy(&temp.key, k);
1867                                 shift_keys(b, k, &temp.key);
1868                                 top = bkey_next(k);
1869                         }
1870
1871                         bch_cut_front(insert, top);
1872                         bch_cut_back(&START_KEY(insert), k);
1873                         bch_bset_fix_invalidated_key(b, k);
1874                         return false;
1875                 }
1876
1877                 if (bkey_cmp(insert, k) < 0) {
1878                         bch_cut_front(insert, k);
1879                 } else {
1880                         if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0)
1881                                 old_offset = KEY_START(insert);
1882
1883                         if (bkey_written(b, k) &&
1884                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1885                                 /*
1886                                  * Completely overwrote, so we don't have to
1887                                  * invalidate the binary search tree
1888                                  */
1889                                 bch_cut_front(k, k);
1890                         } else {
1891                                 __bch_cut_back(&START_KEY(insert), k);
1892                                 bch_bset_fix_invalidated_key(b, k);
1893                         }
1894                 }
1895
1896                 subtract_dirty(k, old_offset, old_size - KEY_SIZE(k));
1897         }
1898
1899 check_failed:
1900         if (replace_key) {
1901                 if (!sectors_found) {
1902                         return true;
1903                 } else if (sectors_found < KEY_SIZE(insert)) {
1904                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1905                                        (KEY_SIZE(insert) - sectors_found));
1906                         SET_KEY_SIZE(insert, sectors_found);
1907                 }
1908         }
1909
1910         return false;
1911 }
1912
1913 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1914                              struct bkey *k, struct bkey *replace_key)
1915 {
1916         struct bset *i = b->sets[b->nsets].data;
1917         struct bkey *m, *prev;
1918         unsigned status = BTREE_INSERT_STATUS_INSERT;
1919
1920         BUG_ON(bkey_cmp(k, &b->key) > 0);
1921         BUG_ON(b->level && !KEY_PTRS(k));
1922         BUG_ON(!b->level && !KEY_OFFSET(k));
1923
1924         if (!b->level) {
1925                 struct btree_iter iter;
1926
1927                 /*
1928                  * bset_search() returns the first key that is strictly greater
1929                  * than the search key - but for back merging, we want to find
1930                  * the previous key.
1931                  */
1932                 prev = NULL;
1933                 m = bch_btree_iter_init(b, &iter, PRECEDING_KEY(&START_KEY(k)));
1934
1935                 if (fix_overlapping_extents(b, k, &iter, replace_key)) {
1936                         op->insert_collision = true;
1937                         return false;
1938                 }
1939
1940                 if (KEY_DIRTY(k))
1941                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1942                                                      KEY_START(k), KEY_SIZE(k));
1943
1944                 while (m != end(i) &&
1945                        bkey_cmp(k, &START_KEY(m)) > 0)
1946                         prev = m, m = bkey_next(m);
1947
1948                 if (key_merging_disabled(b->c))
1949                         goto insert;
1950
1951                 /* prev is in the tree, if we merge we're done */
1952                 status = BTREE_INSERT_STATUS_BACK_MERGE;
1953                 if (prev &&
1954                     bch_bkey_try_merge(b, prev, k))
1955                         goto merged;
1956
1957                 status = BTREE_INSERT_STATUS_OVERWROTE;
1958                 if (m != end(i) &&
1959                     KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
1960                         goto copy;
1961
1962                 status = BTREE_INSERT_STATUS_FRONT_MERGE;
1963                 if (m != end(i) &&
1964                     bch_bkey_try_merge(b, k, m))
1965                         goto copy;
1966         } else {
1967                 BUG_ON(replace_key);
1968                 m = bch_bset_search(b, &b->sets[b->nsets], k);
1969         }
1970
1971 insert: shift_keys(b, m, k);
1972 copy:   bkey_copy(m, k);
1973 merged:
1974         bch_check_keys(b, "%u for %s", status,
1975                        replace_key ? "replace" : "insert");
1976
1977         if (b->level && !KEY_OFFSET(k))
1978                 btree_current_write(b)->prio_blocked++;
1979
1980         trace_bcache_btree_insert_key(b, k, replace_key != NULL, status);
1981
1982         return true;
1983 }
1984
1985 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1986                                   struct keylist *insert_keys,
1987                                   struct bkey *replace_key)
1988 {
1989         bool ret = false;
1990         int oldsize = bch_count_data(b);
1991
1992         while (!bch_keylist_empty(insert_keys)) {
1993                 struct bset *i = write_block(b);
1994                 struct bkey *k = insert_keys->keys;
1995
1996                 if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c)
1997                     > btree_blocks(b))
1998                         break;
1999
2000                 if (bkey_cmp(k, &b->key) <= 0) {
2001                         if (!b->level)
2002                                 bkey_put(b->c, k);
2003
2004                         ret |= btree_insert_key(b, op, k, replace_key);
2005                         bch_keylist_pop_front(insert_keys);
2006                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2007                         BKEY_PADDED(key) temp;
2008                         bkey_copy(&temp.key, insert_keys->keys);
2009
2010                         bch_cut_back(&b->key, &temp.key);
2011                         bch_cut_front(&b->key, insert_keys->keys);
2012
2013                         ret |= btree_insert_key(b, op, &temp.key, replace_key);
2014                         break;
2015                 } else {
2016                         break;
2017                 }
2018         }
2019
2020         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2021
2022         BUG_ON(bch_count_data(b) < oldsize);
2023         return ret;
2024 }
2025
2026 static int btree_split(struct btree *b, struct btree_op *op,
2027                        struct keylist *insert_keys,
2028                        struct bkey *replace_key)
2029 {
2030         bool split;
2031         struct btree *n1, *n2 = NULL, *n3 = NULL;
2032         uint64_t start_time = local_clock();
2033         struct closure cl;
2034         struct keylist parent_keys;
2035
2036         closure_init_stack(&cl);
2037         bch_keylist_init(&parent_keys);
2038
2039         n1 = btree_node_alloc_replacement(b, true);
2040         if (IS_ERR(n1))
2041                 goto err;
2042
2043         split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
2044
2045         if (split) {
2046                 unsigned keys = 0;
2047
2048                 trace_bcache_btree_node_split(b, n1->sets[0].data->keys);
2049
2050                 n2 = bch_btree_node_alloc(b->c, b->level, true);
2051                 if (IS_ERR(n2))
2052                         goto err_free1;
2053
2054                 if (!b->parent) {
2055                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
2056                         if (IS_ERR(n3))
2057                                 goto err_free2;
2058                 }
2059
2060                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2061
2062                 /*
2063                  * Has to be a linear search because we don't have an auxiliary
2064                  * search tree yet
2065                  */
2066
2067                 while (keys < (n1->sets[0].data->keys * 3) / 5)
2068                         keys += bkey_u64s(node(n1->sets[0].data, keys));
2069
2070                 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
2071                 keys += bkey_u64s(node(n1->sets[0].data, keys));
2072
2073                 n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
2074                 n1->sets[0].data->keys = keys;
2075
2076                 memcpy(n2->sets[0].data->start,
2077                        end(n1->sets[0].data),
2078                        n2->sets[0].data->keys * sizeof(uint64_t));
2079
2080                 bkey_copy_key(&n2->key, &b->key);
2081
2082                 bch_keylist_add(&parent_keys, &n2->key);
2083                 bch_btree_node_write(n2, &cl);
2084                 rw_unlock(true, n2);
2085         } else {
2086                 trace_bcache_btree_node_compact(b, n1->sets[0].data->keys);
2087
2088                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2089         }
2090
2091         bch_keylist_add(&parent_keys, &n1->key);
2092         bch_btree_node_write(n1, &cl);
2093
2094         if (n3) {
2095                 /* Depth increases, make a new root */
2096                 bkey_copy_key(&n3->key, &MAX_KEY);
2097                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2098                 bch_btree_node_write(n3, &cl);
2099
2100                 closure_sync(&cl);
2101                 bch_btree_set_root(n3);
2102                 rw_unlock(true, n3);
2103
2104                 btree_node_free(b);
2105         } else if (!b->parent) {
2106                 /* Root filled up but didn't need to be split */
2107                 closure_sync(&cl);
2108                 bch_btree_set_root(n1);
2109
2110                 btree_node_free(b);
2111         } else {
2112                 /* Split a non root node */
2113                 closure_sync(&cl);
2114                 make_btree_freeing_key(b, parent_keys.top);
2115                 bch_keylist_push(&parent_keys);
2116
2117                 btree_node_free(b);
2118
2119                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2120                 BUG_ON(!bch_keylist_empty(&parent_keys));
2121         }
2122
2123         rw_unlock(true, n1);
2124
2125         bch_time_stats_update(&b->c->btree_split_time, start_time);
2126
2127         return 0;
2128 err_free2:
2129         btree_node_free(n2);
2130         rw_unlock(true, n2);
2131 err_free1:
2132         btree_node_free(n1);
2133         rw_unlock(true, n1);
2134 err:
2135         if (n3 == ERR_PTR(-EAGAIN) ||
2136             n2 == ERR_PTR(-EAGAIN) ||
2137             n1 == ERR_PTR(-EAGAIN))
2138                 return -EAGAIN;
2139
2140         pr_warn("couldn't split");
2141         return -ENOMEM;
2142 }
2143
2144 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2145                                  struct keylist *insert_keys,
2146                                  atomic_t *journal_ref,
2147                                  struct bkey *replace_key)
2148 {
2149         BUG_ON(b->level && replace_key);
2150
2151         if (should_split(b)) {
2152                 if (current->bio_list) {
2153                         op->lock = b->c->root->level + 1;
2154                         return -EAGAIN;
2155                 } else if (op->lock <= b->c->root->level) {
2156                         op->lock = b->c->root->level + 1;
2157                         return -EINTR;
2158                 } else {
2159                         /* Invalidated all iterators */
2160                         return btree_split(b, op, insert_keys, replace_key) ?:
2161                                 -EINTR;
2162                 }
2163         } else {
2164                 BUG_ON(write_block(b) != b->sets[b->nsets].data);
2165
2166                 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2167                         if (!b->level)
2168                                 bch_btree_leaf_dirty(b, journal_ref);
2169                         else
2170                                 bch_btree_node_write_sync(b);
2171                 }
2172
2173                 return 0;
2174         }
2175 }
2176
2177 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2178                                struct bkey *check_key)
2179 {
2180         int ret = -EINTR;
2181         uint64_t btree_ptr = b->key.ptr[0];
2182         unsigned long seq = b->seq;
2183         struct keylist insert;
2184         bool upgrade = op->lock == -1;
2185
2186         bch_keylist_init(&insert);
2187
2188         if (upgrade) {
2189                 rw_unlock(false, b);
2190                 rw_lock(true, b, b->level);
2191
2192                 if (b->key.ptr[0] != btree_ptr ||
2193                     b->seq != seq + 1)
2194                         goto out;
2195         }
2196
2197         SET_KEY_PTRS(check_key, 1);
2198         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2199
2200         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2201
2202         bch_keylist_add(&insert, check_key);
2203
2204         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2205
2206         BUG_ON(!ret && !bch_keylist_empty(&insert));
2207 out:
2208         if (upgrade)
2209                 downgrade_write(&b->lock);
2210         return ret;
2211 }
2212
2213 struct btree_insert_op {
2214         struct btree_op op;
2215         struct keylist  *keys;
2216         atomic_t        *journal_ref;
2217         struct bkey     *replace_key;
2218 };
2219
2220 int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2221 {
2222         struct btree_insert_op *op = container_of(b_op,
2223                                         struct btree_insert_op, op);
2224
2225         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2226                                         op->journal_ref, op->replace_key);
2227         if (ret && !bch_keylist_empty(op->keys))
2228                 return ret;
2229         else
2230                 return MAP_DONE;
2231 }
2232
2233 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2234                      atomic_t *journal_ref, struct bkey *replace_key)
2235 {
2236         struct btree_insert_op op;
2237         int ret = 0;
2238
2239         BUG_ON(current->bio_list);
2240         BUG_ON(bch_keylist_empty(keys));
2241
2242         bch_btree_op_init(&op.op, 0);
2243         op.keys         = keys;
2244         op.journal_ref  = journal_ref;
2245         op.replace_key  = replace_key;
2246
2247         while (!ret && !bch_keylist_empty(keys)) {
2248                 op.op.lock = 0;
2249                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2250                                                &START_KEY(keys->keys),
2251                                                btree_insert_fn);
2252         }
2253
2254         if (ret) {
2255                 struct bkey *k;
2256
2257                 pr_err("error %i", ret);
2258
2259                 while ((k = bch_keylist_pop(keys)))
2260                         bkey_put(c, k);
2261         } else if (op.op.insert_collision)
2262                 ret = -ESRCH;
2263
2264         return ret;
2265 }
2266
2267 void bch_btree_set_root(struct btree *b)
2268 {
2269         unsigned i;
2270         struct closure cl;
2271
2272         closure_init_stack(&cl);
2273
2274         trace_bcache_btree_set_root(b);
2275
2276         BUG_ON(!b->written);
2277
2278         for (i = 0; i < KEY_PTRS(&b->key); i++)
2279                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2280
2281         mutex_lock(&b->c->bucket_lock);
2282         list_del_init(&b->list);
2283         mutex_unlock(&b->c->bucket_lock);
2284
2285         b->c->root = b;
2286
2287         bch_journal_meta(b->c, &cl);
2288         closure_sync(&cl);
2289 }
2290
2291 /* Map across nodes or keys */
2292
2293 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2294                                        struct bkey *from,
2295                                        btree_map_nodes_fn *fn, int flags)
2296 {
2297         int ret = MAP_CONTINUE;
2298
2299         if (b->level) {
2300                 struct bkey *k;
2301                 struct btree_iter iter;
2302
2303                 bch_btree_iter_init(b, &iter, from);
2304
2305                 while ((k = bch_btree_iter_next_filter(&iter, b,
2306                                                        bch_ptr_bad))) {
2307                         ret = btree(map_nodes_recurse, k, b,
2308                                     op, from, fn, flags);
2309                         from = NULL;
2310
2311                         if (ret != MAP_CONTINUE)
2312                                 return ret;
2313                 }
2314         }
2315
2316         if (!b->level || flags == MAP_ALL_NODES)
2317                 ret = fn(op, b);
2318
2319         return ret;
2320 }
2321
2322 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2323                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2324 {
2325         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2326 }
2327
2328 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2329                                       struct bkey *from, btree_map_keys_fn *fn,
2330                                       int flags)
2331 {
2332         int ret = MAP_CONTINUE;
2333         struct bkey *k;
2334         struct btree_iter iter;
2335
2336         bch_btree_iter_init(b, &iter, from);
2337
2338         while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) {
2339                 ret = !b->level
2340                         ? fn(op, b, k)
2341                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2342                 from = NULL;
2343
2344                 if (ret != MAP_CONTINUE)
2345                         return ret;
2346         }
2347
2348         if (!b->level && (flags & MAP_END_KEY))
2349                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2350                                      KEY_OFFSET(&b->key), 0));
2351
2352         return ret;
2353 }
2354
2355 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2356                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2357 {
2358         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2359 }
2360
2361 /* Keybuf code */
2362
2363 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2364 {
2365         /* Overlapping keys compare equal */
2366         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2367                 return -1;
2368         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2369                 return 1;
2370         return 0;
2371 }
2372
2373 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2374                                             struct keybuf_key *r)
2375 {
2376         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2377 }
2378
2379 struct refill {
2380         struct btree_op op;
2381         unsigned        nr_found;
2382         struct keybuf   *buf;
2383         struct bkey     *end;
2384         keybuf_pred_fn  *pred;
2385 };
2386
2387 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2388                             struct bkey *k)
2389 {
2390         struct refill *refill = container_of(op, struct refill, op);
2391         struct keybuf *buf = refill->buf;
2392         int ret = MAP_CONTINUE;
2393
2394         if (bkey_cmp(k, refill->end) >= 0) {
2395                 ret = MAP_DONE;
2396                 goto out;
2397         }
2398
2399         if (!KEY_SIZE(k)) /* end key */
2400                 goto out;
2401
2402         if (refill->pred(buf, k)) {
2403                 struct keybuf_key *w;
2404
2405                 spin_lock(&buf->lock);
2406
2407                 w = array_alloc(&buf->freelist);
2408                 if (!w) {
2409                         spin_unlock(&buf->lock);
2410                         return MAP_DONE;
2411                 }
2412
2413                 w->private = NULL;
2414                 bkey_copy(&w->key, k);
2415
2416                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2417                         array_free(&buf->freelist, w);
2418                 else
2419                         refill->nr_found++;
2420
2421                 if (array_freelist_empty(&buf->freelist))
2422                         ret = MAP_DONE;
2423
2424                 spin_unlock(&buf->lock);
2425         }
2426 out:
2427         buf->last_scanned = *k;
2428         return ret;
2429 }
2430
2431 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2432                        struct bkey *end, keybuf_pred_fn *pred)
2433 {
2434         struct bkey start = buf->last_scanned;
2435         struct refill refill;
2436
2437         cond_resched();
2438
2439         bch_btree_op_init(&refill.op, -1);
2440         refill.nr_found = 0;
2441         refill.buf      = buf;
2442         refill.end      = end;
2443         refill.pred     = pred;
2444
2445         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2446                            refill_keybuf_fn, MAP_END_KEY);
2447
2448         trace_bcache_keyscan(refill.nr_found,
2449                              KEY_INODE(&start), KEY_OFFSET(&start),
2450                              KEY_INODE(&buf->last_scanned),
2451                              KEY_OFFSET(&buf->last_scanned));
2452
2453         spin_lock(&buf->lock);
2454
2455         if (!RB_EMPTY_ROOT(&buf->keys)) {
2456                 struct keybuf_key *w;
2457                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2458                 buf->start      = START_KEY(&w->key);
2459
2460                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2461                 buf->end        = w->key;
2462         } else {
2463                 buf->start      = MAX_KEY;
2464                 buf->end        = MAX_KEY;
2465         }
2466
2467         spin_unlock(&buf->lock);
2468 }
2469
2470 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2471 {
2472         rb_erase(&w->node, &buf->keys);
2473         array_free(&buf->freelist, w);
2474 }
2475
2476 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2477 {
2478         spin_lock(&buf->lock);
2479         __bch_keybuf_del(buf, w);
2480         spin_unlock(&buf->lock);
2481 }
2482
2483 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2484                                   struct bkey *end)
2485 {
2486         bool ret = false;
2487         struct keybuf_key *p, *w, s;
2488         s.key = *start;
2489
2490         if (bkey_cmp(end, &buf->start) <= 0 ||
2491             bkey_cmp(start, &buf->end) >= 0)
2492                 return false;
2493
2494         spin_lock(&buf->lock);
2495         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2496
2497         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2498                 p = w;
2499                 w = RB_NEXT(w, node);
2500
2501                 if (p->private)
2502                         ret = true;
2503                 else
2504                         __bch_keybuf_del(buf, p);
2505         }
2506
2507         spin_unlock(&buf->lock);
2508         return ret;
2509 }
2510
2511 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2512 {
2513         struct keybuf_key *w;
2514         spin_lock(&buf->lock);
2515
2516         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2517
2518         while (w && w->private)
2519                 w = RB_NEXT(w, node);
2520
2521         if (w)
2522                 w->private = ERR_PTR(-EINTR);
2523
2524         spin_unlock(&buf->lock);
2525         return w;
2526 }
2527
2528 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2529                                           struct keybuf *buf,
2530                                           struct bkey *end,
2531                                           keybuf_pred_fn *pred)
2532 {
2533         struct keybuf_key *ret;
2534
2535         while (1) {
2536                 ret = bch_keybuf_next(buf);
2537                 if (ret)
2538                         break;
2539
2540                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2541                         pr_debug("scan finished");
2542                         break;
2543                 }
2544
2545                 bch_refill_keybuf(c, buf, end, pred);
2546         }
2547
2548         return ret;
2549 }
2550
2551 void bch_keybuf_init(struct keybuf *buf)
2552 {
2553         buf->last_scanned       = MAX_KEY;
2554         buf->keys               = RB_ROOT;
2555
2556         spin_lock_init(&buf->lock);
2557         array_allocator_init(&buf->freelist);
2558 }
2559
2560 void bch_btree_exit(void)
2561 {
2562         if (btree_io_wq)
2563                 destroy_workqueue(btree_io_wq);
2564 }
2565
2566 int __init bch_btree_init(void)
2567 {
2568         btree_io_wq = create_singlethread_workqueue("bch_btree_io");
2569         if (!btree_io_wq)
2570                 return -ENOMEM;
2571
2572         return 0;
2573 }