drivers/md/bcache/btree.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  *
   4  * Uses a block device as cache for other block devices; optimized for SSDs.
   5  * All allocation is done in buckets, which should match the erase block size
   6  * of the device.
   7  *
   8  * Buckets containing cached data are kept on a heap sorted by priority;
   9  * bucket priority is increased on cache hit, and periodically all the buckets
  10  * on the heap have their priority scaled down. This currently is just used as
  11  * an LRU but in the future should allow for more intelligent heuristics.
  12  *
  13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  14  * counter. Garbage collection is used to remove stale pointers.
  15  *
  16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  17  * as keys are inserted we only sort the pages that have not yet been written.
  18  * When garbage collection is run, we resort the entire node.
  19  *
  20  * All configuration is done via sysfs; see Documentation/bcache.txt.
  21  */
  22
  23 #include "bcache.h"
  24 #include "btree.h"
  25 #include "debug.h"
  26 #include "extents.h"
  27 #include "writeback.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/freezer.h>
  32 #include <linux/hash.h>
  33 #include <linux/kthread.h>
  34 #include <linux/prefetch.h>
  35 #include <linux/random.h>
  36 #include <linux/rcupdate.h>
  37 #include <trace/events/bcache.h>
  38
  39 /*
  40  * Todo:
  41  * register_bcache: Return errors out to userspace correctly
  42  *
  43  * Writeback: don't undirty key until after a cache flush
  44  *
  45  * Create an iterator for key pointers
  46  *
  47  * On btree write error, mark bucket such that it won't be freed from the cache
  48  *
  49  * Journalling:
  50  *   Check for bad keys in replay
  51  *   Propagate barriers
  52  *   Refcount journal entries in journal_replay
  53  *
  54  * Garbage collection:
  55  *   Finish incremental gc
  56  *   Gc should free old UUIDs, data for invalid UUIDs
  57  *
  58  * Provide a way to list backing device UUIDs we have data cached for, and
  59  * probably how long it's been since we've seen them, and a way to invalidate
  60  * dirty data for devices that will never be attached again
  61  *
  62  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  63  * that based on that and how much dirty data we have we can keep writeback
  64  * from being starved
  65  *
  66  * Add a tracepoint or somesuch to watch for writeback starvation
  67  *
  68  * When btree depth > 1 and splitting an interior node, we have to make sure
  69  * alloc_bucket() cannot fail. This should be true but is not completely
  70  * obvious.
  71  *
  72  * Make sure all allocations get charged to the root cgroup
  73  *
  74  * Plugging?
  75  *
  76  * If data write is less than hard sector size of ssd, round up offset in open
  77  * bucket to the next whole sector
  78  *
  79  * Also lookup by cgroup in get_open_bucket()
  80  *
  81  * Superblock needs to be fleshed out for multiple cache devices
  82  *
  83  * Add a sysfs tunable for the number of writeback IOs in flight
  84  *
  85  * Add a sysfs tunable for the number of open data buckets
  86  *
  87  * IO tracking: Can we track when one process is doing io on behalf of another?
  88  * IO tracking: Don't use just an average, weigh more recent stuff higher
  89  *
  90  * Test module load/unload
  91  */
  92
  93 enum {
  94         BTREE_INSERT_STATUS_INSERT,
  95         BTREE_INSERT_STATUS_BACK_MERGE,
  96         BTREE_INSERT_STATUS_OVERWROTE,
  97         BTREE_INSERT_STATUS_FRONT_MERGE,
  98 };
  99
 100 #define MAX_NEED_GC             64
 101 #define MAX_SAVE_PRIO           72
 102
 103 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
 104
 105 #define PTR_HASH(c, k)                                                  \
 106         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 107
 108 static struct workqueue_struct *btree_io_wq;
 109
 110 static inline bool should_split(struct btree *b)
 111 {
 112         struct bset *i = write_block(b);
 113         return b->written >= btree_blocks(b) ||
 114                 (b->written + __set_blocks(i, i->keys + 15, block_bytes(b->c))
 115                  > btree_blocks(b));
 116 }
 117
 118 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 119
 120 /*
 121  * These macros are for recursing down the btree - they handle the details of
 122  * locking and looking up nodes in the cache for you. They're best treated as
 123  * mere syntax when reading code that uses them.
 124  *
 125  * op->lock determines whether we take a read or a write lock at a given depth.
 126  * If you've got a read lock and find that you need a write lock (i.e. you're
 127  * going to have to split), set op->lock and return -EINTR; btree_root() will
 128  * call you again and you'll have the correct lock.
 129  */
 130
 131 /**
 132  * btree - recurse down the btree on a specified key
 133  * @fn:         function to call, which will be passed the child node
 134  * @key:        key to recurse on
 135  * @b:          parent btree node
 136  * @op:         pointer to struct btree_op
 137  */
 138 #define btree(fn, key, b, op, ...)                                      \
 139 ({                                                                      \
 140         int _r, l = (b)->level - 1;                                     \
 141         bool _w = l <= (op)->lock;                                      \
 142         struct btree *_child = bch_btree_node_get((b)->c, key, l, _w);  \
 143         if (!IS_ERR(_child)) {                                          \
 144                 _child->parent = (b);                                   \
 145                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 146                 rw_unlock(_w, _child);                                  \
 147         } else                                                          \
 148                 _r = PTR_ERR(_child);                                   \
 149         _r;                                                             \
 150 })
 151
 152 /**
 153  * btree_root - call a function on the root of the btree
 154  * @fn:         function to call, which will be passed the child node
 155  * @c:          cache set
 156  * @op:         pointer to struct btree_op
 157  */
 158 #define btree_root(fn, c, op, ...)                                      \
 159 ({                                                                      \
 160         int _r = -EINTR;                                                \
 161         do {                                                            \
 162                 struct btree *_b = (c)->root;                           \
 163                 bool _w = insert_lock(op, _b);                          \
 164                 rw_lock(_w, _b, _b->level);                             \
 165                 if (_b == (c)->root &&                                  \
 166                     _w == insert_lock(op, _b)) {                        \
 167                         _b->parent = NULL;                              \
 168                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 169                 }                                                       \
 170                 rw_unlock(_w, _b);                                      \
 171                 if (_r == -EINTR)                                       \
 172                         schedule();                                     \
 173                 bch_cannibalize_unlock(c);                              \
 174                 if (_r == -ENOSPC) {                                    \
 175                         wait_event((c)->try_wait,                       \
 176                                    !(c)->try_harder);                   \
 177                         _r = -EINTR;                                    \
 178                 }                                                       \
 179         } while (_r == -EINTR);                                         \
 180                                                                         \
 181         finish_wait(&(c)->bucket_wait, &(op)->wait);                    \
 182         _r;                                                             \
 183 })
 184
 185 /* Btree key manipulation */
 186
 187 void bkey_put(struct cache_set *c, struct bkey *k)
 188 {
 189         unsigned i;
 190
 191         for (i = 0; i < KEY_PTRS(k); i++)
 192                 if (ptr_available(c, k, i))
 193                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 194 }
 195
 196 /* Btree IO */
 197
 198 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 199 {
 200         uint64_t crc = b->key.ptr[0];
 201         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 202
 203         crc = bch_crc64_update(crc, data, end - data);
 204         return crc ^ 0xffffffffffffffffULL;
 205 }
 206
 207 void bch_btree_node_read_done(struct btree *b)
 208 {
 209         const char *err = "bad btree header";
 210         struct bset *i = btree_bset_first(b);
 211         struct btree_iter *iter;
 212
 213         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 214         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 215         iter->used = 0;
 216
 217 #ifdef CONFIG_BCACHE_DEBUG
 218         iter->b = b;
 219 #endif
 220
 221         if (!i->seq)
 222                 goto err;
 223
 224         for (;
 225              b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
 226              i = write_block(b)) {
 227                 err = "unsupported bset version";
 228                 if (i->version > BCACHE_BSET_VERSION)
 229                         goto err;
 230
 231                 err = "bad btree header";
 232                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 233                     btree_blocks(b))
 234                         goto err;
 235
 236                 err = "bad magic";
 237                 if (i->magic != bset_magic(&b->c->sb))
 238                         goto err;
 239
 240                 err = "bad checksum";
 241                 switch (i->version) {
 242                 case 0:
 243                         if (i->csum != csum_set(i))
 244                                 goto err;
 245                         break;
 246                 case BCACHE_BSET_VERSION:
 247                         if (i->csum != btree_csum_set(b, i))
 248                                 goto err;
 249                         break;
 250                 }
 251
 252                 err = "empty set";
 253                 if (i != b->sets[0].data && !i->keys)
 254                         goto err;
 255
 256                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 257
 258                 b->written += set_blocks(i, block_bytes(b->c));
 259         }
 260
 261         err = "corrupted btree";
 262         for (i = write_block(b);
 263              bset_sector_offset(b, i) < KEY_SIZE(&b->key);
 264              i = ((void *) i) + block_bytes(b->c))
 265                 if (i->seq == b->sets[0].data->seq)
 266                         goto err;
 267
 268         bch_btree_sort_and_fix_extents(b, iter, &b->c->sort);
 269
 270         i = b->sets[0].data;
 271         err = "short btree key";
 272         if (b->sets[0].size &&
 273             bkey_cmp(&b->key, &b->sets[0].end) < 0)
 274                 goto err;
 275
 276         if (b->written < btree_blocks(b))
 277                 bch_bset_init_next(b, write_block(b), bset_magic(&b->c->sb));
 278 out:
 279         mempool_free(iter, b->c->fill_iter);
 280         return;
 281 err:
 282         set_btree_node_io_error(b);
 283         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 284                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 285                             bset_block_offset(b, i), i->keys);
 286         goto out;
 287 }
 288
 289 static void btree_node_read_endio(struct bio *bio, int error)
 290 {
 291         struct closure *cl = bio->bi_private;
 292         closure_put(cl);
 293 }
 294
 295 static void bch_btree_node_read(struct btree *b)
 296 {
 297         uint64_t start_time = local_clock();
 298         struct closure cl;
 299         struct bio *bio;
 300
 301         trace_bcache_btree_read(b);
 302
 303         closure_init_stack(&cl);
 304
 305         bio = bch_bbio_alloc(b->c);
 306         bio->bi_rw      = REQ_META|READ_SYNC;
 307         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 308         bio->bi_end_io  = btree_node_read_endio;
 309         bio->bi_private = &cl;
 310
 311         bch_bio_map(bio, b->sets[0].data);
 312
 313         bch_submit_bbio(bio, b->c, &b->key, 0);
 314         closure_sync(&cl);
 315
 316         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 317                 set_btree_node_io_error(b);
 318
 319         bch_bbio_free(bio, b->c);
 320
 321         if (btree_node_io_error(b))
 322                 goto err;
 323
 324         bch_btree_node_read_done(b);
 325         bch_time_stats_update(&b->c->btree_read_time, start_time);
 326
 327         return;
 328 err:
 329         bch_cache_set_error(b->c, "io error reading bucket %zu",
 330                             PTR_BUCKET_NR(b->c, &b->key, 0));
 331 }
 332
 333 static void btree_complete_write(struct btree *b, struct btree_write *w)
 334 {
 335         if (w->prio_blocked &&
 336             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 337                 wake_up_allocators(b->c);
 338
 339         if (w->journal) {
 340                 atomic_dec_bug(w->journal);
 341                 __closure_wake_up(&b->c->journal.wait);
 342         }
 343
 344         w->prio_blocked = 0;
 345         w->journal      = NULL;
 346 }
 347
 348 static void btree_node_write_unlock(struct closure *cl)
 349 {
 350         struct btree *b = container_of(cl, struct btree, io);
 351
 352         up(&b->io_mutex);
 353 }
 354
 355 static void __btree_node_write_done(struct closure *cl)
 356 {
 357         struct btree *b = container_of(cl, struct btree, io);
 358         struct btree_write *w = btree_prev_write(b);
 359
 360         bch_bbio_free(b->bio, b->c);
 361         b->bio = NULL;
 362         btree_complete_write(b, w);
 363
 364         if (btree_node_dirty(b))
 365                 queue_delayed_work(btree_io_wq, &b->work,
 366                                    msecs_to_jiffies(30000));
 367
 368         closure_return_with_destructor(cl, btree_node_write_unlock);
 369 }
 370
 371 static void btree_node_write_done(struct closure *cl)
 372 {
 373         struct btree *b = container_of(cl, struct btree, io);
 374         struct bio_vec *bv;
 375         int n;
 376
 377         bio_for_each_segment_all(bv, b->bio, n)
 378                 __free_page(bv->bv_page);
 379
 380         __btree_node_write_done(cl);
 381 }
 382
 383 static void btree_node_write_endio(struct bio *bio, int error)
 384 {
 385         struct closure *cl = bio->bi_private;
 386         struct btree *b = container_of(cl, struct btree, io);
 387
 388         if (error)
 389                 set_btree_node_io_error(b);
 390
 391         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 392         closure_put(cl);
 393 }
 394
 395 static void do_btree_node_write(struct btree *b)
 396 {
 397         struct closure *cl = &b->io;
 398         struct bset *i = btree_bset_last(b);
 399         BKEY_PADDED(key) k;
 400
 401         i->version      = BCACHE_BSET_VERSION;
 402         i->csum         = btree_csum_set(b, i);
 403
 404         BUG_ON(b->bio);
 405         b->bio = bch_bbio_alloc(b->c);
 406
 407         b->bio->bi_end_io       = btree_node_write_endio;
 408         b->bio->bi_private      = cl;
 409         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 410         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 411         bch_bio_map(b->bio, i);
 412
 413         /*
 414          * If we're appending to a leaf node, we don't technically need FUA -
 415          * this write just needs to be persisted before the next journal write,
 416          * which will be marked FLUSH|FUA.
 417          *
 418          * Similarly if we're writing a new btree root - the pointer is going to
 419          * be in the next journal entry.
 420          *
 421          * But if we're writing a new btree node (that isn't a root) or
 422          * appending to a non leaf btree node, we need either FUA or a flush
 423          * when we write the parent with the new pointer. FUA is cheaper than a
 424          * flush, and writes appending to leaf nodes aren't blocking anything so
 425          * just make all btree node writes FUA to keep things sane.
 426          */
 427
 428         bkey_copy(&k.key, &b->key);
 429         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 430                        bset_sector_offset(b, i));
 431
 432         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 433                 int j;
 434                 struct bio_vec *bv;
 435                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 436
 437                 bio_for_each_segment_all(bv, b->bio, j)
 438                         memcpy(page_address(bv->bv_page),
 439                                base + j * PAGE_SIZE, PAGE_SIZE);
 440
 441                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 442
 443                 continue_at(cl, btree_node_write_done, NULL);
 444         } else {
 445                 b->bio->bi_vcnt = 0;
 446                 bch_bio_map(b->bio, i);
 447
 448                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 449
 450                 closure_sync(cl);
 451                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 452         }
 453 }
 454
 455 void bch_btree_node_write(struct btree *b, struct closure *parent)
 456 {
 457         struct bset *i = btree_bset_last(b);
 458
 459         trace_bcache_btree_write(b);
 460
 461         BUG_ON(current->bio_list);
 462         BUG_ON(b->written >= btree_blocks(b));
 463         BUG_ON(b->written && !i->keys);
 464         BUG_ON(btree_bset_first(b)->seq != i->seq);
 465         bch_check_keys(b, "writing");
 466
 467         cancel_delayed_work(&b->work);
 468
 469         /* If caller isn't waiting for write, parent refcount is cache set */
 470         down(&b->io_mutex);
 471         closure_init(&b->io, parent ?: &b->c->cl);
 472
 473         clear_bit(BTREE_NODE_dirty,      &b->flags);
 474         change_bit(BTREE_NODE_write_idx, &b->flags);
 475
 476         do_btree_node_write(b);
 477
 478         b->written += set_blocks(i, block_bytes(b->c));
 479         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 480                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 481
 482         /* If not a leaf node, always sort */
 483         if (b->level && b->nsets)
 484                 bch_btree_sort(b, &b->c->sort);
 485         else
 486                 bch_btree_sort_lazy(b, &b->c->sort);
 487
 488         /*
 489          * do verify if there was more than one set initially (i.e. we did a
 490          * sort) and we sorted down to a single set:
 491          */
 492         if (i != b->sets->data && !b->nsets)
 493                 bch_btree_verify(b);
 494
 495         if (b->written < btree_blocks(b))
 496                 bch_bset_init_next(b, write_block(b), bset_magic(&b->c->sb));
 497 }
 498
 499 static void bch_btree_node_write_sync(struct btree *b)
 500 {
 501         struct closure cl;
 502
 503         closure_init_stack(&cl);
 504         bch_btree_node_write(b, &cl);
 505         closure_sync(&cl);
 506 }
 507
 508 static void btree_node_write_work(struct work_struct *w)
 509 {
 510         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 511
 512         rw_lock(true, b, b->level);
 513
 514         if (btree_node_dirty(b))
 515                 bch_btree_node_write(b, NULL);
 516         rw_unlock(true, b);
 517 }
 518
 519 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 520 {
 521         struct bset *i = btree_bset_last(b);
 522         struct btree_write *w = btree_current_write(b);
 523
 524         BUG_ON(!b->written);
 525         BUG_ON(!i->keys);
 526
 527         if (!btree_node_dirty(b))
 528                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 529
 530         set_btree_node_dirty(b);
 531
 532         if (journal_ref) {
 533                 if (w->journal &&
 534                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 535                         atomic_dec_bug(w->journal);
 536                         w->journal = NULL;
 537                 }
 538
 539                 if (!w->journal) {
 540                         w->journal = journal_ref;
 541                         atomic_inc(w->journal);
 542                 }
 543         }
 544
 545         /* Force write if set is too big */
 546         if (set_bytes(i) > PAGE_SIZE - 48 &&
 547             !current->bio_list)
 548                 bch_btree_node_write(b, NULL);
 549 }
 550
 551 /*
 552  * Btree in memory cache - allocation/freeing
 553  * mca -> memory cache
 554  */
 555
 556 static void mca_reinit(struct btree *b)
 557 {
 558         unsigned i;
 559
 560         b->flags        = 0;
 561         b->written      = 0;
 562         b->nsets        = 0;
 563
 564         for (i = 0; i < MAX_BSETS; i++)
 565                 b->sets[i].size = 0;
 566         /*
 567          * Second loop starts at 1 because b->sets[0]->data is the memory we
 568          * allocated
 569          */
 570         for (i = 1; i < MAX_BSETS; i++)
 571                 b->sets[i].data = NULL;
 572 }
 573
 574 #define mca_reserve(c)  (((c->root && c->root->level)           \
 575                           ? c->root->level : 1) * 8 + 16)
 576 #define mca_can_free(c)                                         \
 577         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 578
 579 static void mca_data_free(struct btree *b)
 580 {
 581         BUG_ON(b->io_mutex.count != 1);
 582
 583         bch_btree_keys_free(b);
 584
 585         b->c->bucket_cache_used--;
 586         list_move(&b->list, &b->c->btree_cache_freed);
 587 }
 588
 589 static void mca_bucket_free(struct btree *b)
 590 {
 591         BUG_ON(btree_node_dirty(b));
 592
 593         b->key.ptr[0] = 0;
 594         hlist_del_init_rcu(&b->hash);
 595         list_move(&b->list, &b->c->btree_cache_freeable);
 596 }
 597
 598 static unsigned btree_order(struct bkey *k)
 599 {
 600         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 601 }
 602
 603 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 604 {
 605         if (!bch_btree_keys_alloc(b,
 606                                   max_t(unsigned,
 607                                         ilog2(b->c->btree_pages),
 608                                         btree_order(k)),
 609                                   gfp)) {
 610                 b->c->bucket_cache_used++;
 611                 list_move(&b->list, &b->c->btree_cache);
 612         } else {
 613                 list_move(&b->list, &b->c->btree_cache_freed);
 614         }
 615 }
 616
 617 static struct btree *mca_bucket_alloc(struct cache_set *c,
 618                                       struct bkey *k, gfp_t gfp)
 619 {
 620         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 621         if (!b)
 622                 return NULL;
 623
 624         init_rwsem(&b->lock);
 625         lockdep_set_novalidate_class(&b->lock);
 626         INIT_LIST_HEAD(&b->list);
 627         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 628         b->c = c;
 629         sema_init(&b->io_mutex, 1);
 630
 631         mca_data_alloc(b, k, gfp);
 632         return b;
 633 }
 634
 635 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 636 {
 637         struct closure cl;
 638
 639         closure_init_stack(&cl);
 640         lockdep_assert_held(&b->c->bucket_lock);
 641
 642         if (!down_write_trylock(&b->lock))
 643                 return -ENOMEM;
 644
 645         BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
 646
 647         if (b->page_order < min_order)
 648                 goto out_unlock;
 649
 650         if (!flush) {
 651                 if (btree_node_dirty(b))
 652                         goto out_unlock;
 653
 654                 if (down_trylock(&b->io_mutex))
 655                         goto out_unlock;
 656                 up(&b->io_mutex);
 657         }
 658
 659         if (btree_node_dirty(b))
 660                 bch_btree_node_write_sync(b);
 661
 662         /* wait for any in flight btree write */
 663         down(&b->io_mutex);
 664         up(&b->io_mutex);
 665
 666         return 0;
 667 out_unlock:
 668         rw_unlock(true, b);
 669         return -ENOMEM;
 670 }
 671
 672 static unsigned long bch_mca_scan(struct shrinker *shrink,
 673                                   struct shrink_control *sc)
 674 {
 675         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 676         struct btree *b, *t;
 677         unsigned long i, nr = sc->nr_to_scan;
 678         unsigned long freed = 0;
 679
 680         if (c->shrinker_disabled)
 681                 return SHRINK_STOP;
 682
 683         if (c->try_harder)
 684                 return SHRINK_STOP;
 685
 686         /* Return -1 if we can't do anything right now */
 687         if (sc->gfp_mask & __GFP_IO)
 688                 mutex_lock(&c->bucket_lock);
 689         else if (!mutex_trylock(&c->bucket_lock))
 690                 return -1;
 691
 692         /*
 693          * It's _really_ critical that we don't free too many btree nodes - we
 694          * have to always leave ourselves a reserve. The reserve is how we
 695          * guarantee that allocating memory for a new btree node can always
 696          * succeed, so that inserting keys into the btree can always succeed and
 697          * IO can always make forward progress:
 698          */
 699         nr /= c->btree_pages;
 700         nr = min_t(unsigned long, nr, mca_can_free(c));
 701
 702         i = 0;
 703         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 704                 if (freed >= nr)
 705                         break;
 706
 707                 if (++i > 3 &&
 708                     !mca_reap(b, 0, false)) {
 709                         mca_data_free(b);
 710                         rw_unlock(true, b);
 711                         freed++;
 712                 }
 713         }
 714
 715         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 716                 if (list_empty(&c->btree_cache))
 717                         goto out;
 718
 719                 b = list_first_entry(&c->btree_cache, struct btree, list);
 720                 list_rotate_left(&c->btree_cache);
 721
 722                 if (!b->accessed &&
 723                     !mca_reap(b, 0, false)) {
 724                         mca_bucket_free(b);
 725                         mca_data_free(b);
 726                         rw_unlock(true, b);
 727                         freed++;
 728                 } else
 729                         b->accessed = 0;
 730         }
 731 out:
 732         mutex_unlock(&c->bucket_lock);
 733         return freed;
 734 }
 735
 736 static unsigned long bch_mca_count(struct shrinker *shrink,
 737                                    struct shrink_control *sc)
 738 {
 739         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 740
 741         if (c->shrinker_disabled)
 742                 return 0;
 743
 744         if (c->try_harder)
 745                 return 0;
 746
 747         return mca_can_free(c) * c->btree_pages;
 748 }
 749
 750 void bch_btree_cache_free(struct cache_set *c)
 751 {
 752         struct btree *b;
 753         struct closure cl;
 754         closure_init_stack(&cl);
 755
 756         if (c->shrink.list.next)
 757                 unregister_shrinker(&c->shrink);
 758
 759         mutex_lock(&c->bucket_lock);
 760
 761 #ifdef CONFIG_BCACHE_DEBUG
 762         if (c->verify_data)
 763                 list_move(&c->verify_data->list, &c->btree_cache);
 764
 765         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 766 #endif
 767
 768         list_splice(&c->btree_cache_freeable,
 769                     &c->btree_cache);
 770
 771         while (!list_empty(&c->btree_cache)) {
 772                 b = list_first_entry(&c->btree_cache, struct btree, list);
 773
 774                 if (btree_node_dirty(b))
 775                         btree_complete_write(b, btree_current_write(b));
 776                 clear_bit(BTREE_NODE_dirty, &b->flags);
 777
 778                 mca_data_free(b);
 779         }
 780
 781         while (!list_empty(&c->btree_cache_freed)) {
 782                 b = list_first_entry(&c->btree_cache_freed,
 783                                      struct btree, list);
 784                 list_del(&b->list);
 785                 cancel_delayed_work_sync(&b->work);
 786                 kfree(b);
 787         }
 788
 789         mutex_unlock(&c->bucket_lock);
 790 }
 791
 792 int bch_btree_cache_alloc(struct cache_set *c)
 793 {
 794         unsigned i;
 795
 796         for (i = 0; i < mca_reserve(c); i++)
 797                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 798                         return -ENOMEM;
 799
 800         list_splice_init(&c->btree_cache,
 801                          &c->btree_cache_freeable);
 802
 803 #ifdef CONFIG_BCACHE_DEBUG
 804         mutex_init(&c->verify_lock);
 805
 806         c->verify_ondisk = (void *)
 807                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 808
 809         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 810
 811         if (c->verify_data &&
 812             c->verify_data->sets[0].data)
 813                 list_del_init(&c->verify_data->list);
 814         else
 815                 c->verify_data = NULL;
 816 #endif
 817
 818         c->shrink.count_objects = bch_mca_count;
 819         c->shrink.scan_objects = bch_mca_scan;
 820         c->shrink.seeks = 4;
 821         c->shrink.batch = c->btree_pages * 2;
 822         register_shrinker(&c->shrink);
 823
 824         return 0;
 825 }
 826
 827 /* Btree in memory cache - hash table */
 828
 829 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 830 {
 831         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 832 }
 833
 834 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 835 {
 836         struct btree *b;
 837
 838         rcu_read_lock();
 839         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 840                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 841                         goto out;
 842         b = NULL;
 843 out:
 844         rcu_read_unlock();
 845         return b;
 846 }
 847
 848 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 849 {
 850         struct btree *b;
 851
 852         trace_bcache_btree_cache_cannibalize(c);
 853
 854         if (!c->try_harder) {
 855                 c->try_harder = current;
 856                 c->try_harder_start = local_clock();
 857         } else if (c->try_harder != current)
 858                 return ERR_PTR(-ENOSPC);
 859
 860         list_for_each_entry_reverse(b, &c->btree_cache, list)
 861                 if (!mca_reap(b, btree_order(k), false))
 862                         return b;
 863
 864         list_for_each_entry_reverse(b, &c->btree_cache, list)
 865                 if (!mca_reap(b, btree_order(k), true))
 866                         return b;
 867
 868         return ERR_PTR(-ENOMEM);
 869 }
 870
 871 /*
 872  * We can only have one thread cannibalizing other cached btree nodes at a time,
 873  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 874  * cannibalize_bucket() will take. This means every time we unlock the root of
 875  * the btree, we need to release this lock if we have it held.
 876  */
 877 static void bch_cannibalize_unlock(struct cache_set *c)
 878 {
 879         if (c->try_harder == current) {
 880                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 881                 c->try_harder = NULL;
 882                 wake_up(&c->try_wait);
 883         }
 884 }
 885
 886 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level)
 887 {
 888         struct btree *b;
 889
 890         BUG_ON(current->bio_list);
 891
 892         lockdep_assert_held(&c->bucket_lock);
 893
 894         if (mca_find(c, k))
 895                 return NULL;
 896
 897         /* btree_free() doesn't free memory; it sticks the node on the end of
 898          * the list. Check if there's any freed nodes there:
 899          */
 900         list_for_each_entry(b, &c->btree_cache_freeable, list)
 901                 if (!mca_reap(b, btree_order(k), false))
 902                         goto out;
 903
 904         /* We never free struct btree itself, just the memory that holds the on
 905          * disk node. Check the freed list before allocating a new one:
 906          */
 907         list_for_each_entry(b, &c->btree_cache_freed, list)
 908                 if (!mca_reap(b, 0, false)) {
 909                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 910                         if (!b->sets[0].data)
 911                                 goto err;
 912                         else
 913                                 goto out;
 914                 }
 915
 916         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 917         if (!b)
 918                 goto err;
 919
 920         BUG_ON(!down_write_trylock(&b->lock));
 921         if (!b->sets->data)
 922                 goto err;
 923 out:
 924         BUG_ON(b->io_mutex.count != 1);
 925
 926         bkey_copy(&b->key, k);
 927         list_move(&b->list, &c->btree_cache);
 928         hlist_del_init_rcu(&b->hash);
 929         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 930
 931         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 932         b->level        = level;
 933         b->parent       = (void *) ~0UL;
 934
 935         if (!b->level)
 936                 b->ops  = &bch_extent_keys_ops;
 937         else
 938                 b->ops  = &bch_btree_keys_ops;
 939
 940         mca_reinit(b);
 941
 942         return b;
 943 err:
 944         if (b)
 945                 rw_unlock(true, b);
 946
 947         b = mca_cannibalize(c, k);
 948         if (!IS_ERR(b))
 949                 goto out;
 950
 951         return b;
 952 }
 953
 954 /**
 955  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 956  * in from disk if necessary.
 957  *
 958  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 959  *
 960  * The btree node will have either a read or a write lock held, depending on
 961  * level and op->lock.
 962  */
 963 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 964                                  int level, bool write)
 965 {
 966         int i = 0;
 967         struct btree *b;
 968
 969         BUG_ON(level < 0);
 970 retry:
 971         b = mca_find(c, k);
 972
 973         if (!b) {
 974                 if (current->bio_list)
 975                         return ERR_PTR(-EAGAIN);
 976
 977                 mutex_lock(&c->bucket_lock);
 978                 b = mca_alloc(c, k, level);
 979                 mutex_unlock(&c->bucket_lock);
 980
 981                 if (!b)
 982                         goto retry;
 983                 if (IS_ERR(b))
 984                         return b;
 985
 986                 bch_btree_node_read(b);
 987
 988                 if (!write)
 989                         downgrade_write(&b->lock);
 990         } else {
 991                 rw_lock(write, b, level);
 992                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 993                         rw_unlock(write, b);
 994                         goto retry;
 995                 }
 996                 BUG_ON(b->level != level);
 997         }
 998
 999         b->accessed = 1;
1000
1001         for (; i <= b->nsets && b->sets[i].size; i++) {
1002                 prefetch(b->sets[i].tree);
1003                 prefetch(b->sets[i].data);
1004         }
1005
1006         for (; i <= b->nsets; i++)
1007                 prefetch(b->sets[i].data);
1008
1009         if (btree_node_io_error(b)) {
1010                 rw_unlock(write, b);
1011                 return ERR_PTR(-EIO);
1012         }
1013
1014         BUG_ON(!b->written);
1015
1016         return b;
1017 }
1018
1019 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1020 {
1021         struct btree *b;
1022
1023         mutex_lock(&c->bucket_lock);
1024         b = mca_alloc(c, k, level);
1025         mutex_unlock(&c->bucket_lock);
1026
1027         if (!IS_ERR_OR_NULL(b)) {
1028                 bch_btree_node_read(b);
1029                 rw_unlock(true, b);
1030         }
1031 }
1032
1033 /* Btree alloc */
1034
1035 static void btree_node_free(struct btree *b)
1036 {
1037         unsigned i;
1038
1039         trace_bcache_btree_node_free(b);
1040
1041         BUG_ON(b == b->c->root);
1042
1043         if (btree_node_dirty(b))
1044                 btree_complete_write(b, btree_current_write(b));
1045         clear_bit(BTREE_NODE_dirty, &b->flags);
1046
1047         cancel_delayed_work(&b->work);
1048
1049         mutex_lock(&b->c->bucket_lock);
1050
1051         for (i = 0; i < KEY_PTRS(&b->key); i++) {
1052                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
1053
1054                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1055                             PTR_BUCKET(b->c, &b->key, i));
1056         }
1057
1058         bch_bucket_free(b->c, &b->key);
1059         mca_bucket_free(b);
1060         mutex_unlock(&b->c->bucket_lock);
1061 }
1062
1063 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1064 {
1065         BKEY_PADDED(key) k;
1066         struct btree *b = ERR_PTR(-EAGAIN);
1067
1068         mutex_lock(&c->bucket_lock);
1069 retry:
1070         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1071                 goto err;
1072
1073         bkey_put(c, &k.key);
1074         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1075
1076         b = mca_alloc(c, &k.key, level);
1077         if (IS_ERR(b))
1078                 goto err_free;
1079
1080         if (!b) {
1081                 cache_bug(c,
1082                         "Tried to allocate bucket that was in btree cache");
1083                 goto retry;
1084         }
1085
1086         b->accessed = 1;
1087         bch_bset_init_next(b, b->sets->data, bset_magic(&b->c->sb));
1088
1089         mutex_unlock(&c->bucket_lock);
1090
1091         trace_bcache_btree_node_alloc(b);
1092         return b;
1093 err_free:
1094         bch_bucket_free(c, &k.key);
1095 err:
1096         mutex_unlock(&c->bucket_lock);
1097
1098         trace_bcache_btree_node_alloc_fail(b);
1099         return b;
1100 }
1101
1102 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1103 {
1104         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1105         if (!IS_ERR_OR_NULL(n)) {
1106                 bch_btree_sort_into(b, n, &b->c->sort);
1107                 bkey_copy_key(&n->key, &b->key);
1108         }
1109
1110         return n;
1111 }
1112
1113 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1114 {
1115         unsigned i;
1116
1117         bkey_copy(k, &b->key);
1118         bkey_copy_key(k, &ZERO_KEY);
1119
1120         for (i = 0; i < KEY_PTRS(k); i++) {
1121                 uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1;
1122
1123                 SET_PTR_GEN(k, i, g);
1124         }
1125
1126         atomic_inc(&b->c->prio_blocked);
1127 }
1128
1129 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1130 {
1131         struct cache_set *c = b->c;
1132         struct cache *ca;
1133         unsigned i, reserve = c->root->level * 2 + 1;
1134         int ret = 0;
1135
1136         mutex_lock(&c->bucket_lock);
1137
1138         for_each_cache(ca, c, i)
1139                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1140                         if (op)
1141                                 prepare_to_wait(&c->bucket_wait, &op->wait,
1142                                                 TASK_UNINTERRUPTIBLE);
1143                         ret = -EINTR;
1144                         break;
1145                 }
1146
1147         mutex_unlock(&c->bucket_lock);
1148         return ret;
1149 }
1150
1151 /* Garbage collection */
1152
1153 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1154 {
1155         uint8_t stale = 0;
1156         unsigned i;
1157         struct bucket *g;
1158
1159         /*
1160          * ptr_invalid() can't return true for the keys that mark btree nodes as
1161          * freed, but since ptr_bad() returns true we'll never actually use them
1162          * for anything and thus we don't want mark their pointers here
1163          */
1164         if (!bkey_cmp(k, &ZERO_KEY))
1165                 return stale;
1166
1167         for (i = 0; i < KEY_PTRS(k); i++) {
1168                 if (!ptr_available(c, k, i))
1169                         continue;
1170
1171                 g = PTR_BUCKET(c, k, i);
1172
1173                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1174                         g->gc_gen = PTR_GEN(k, i);
1175
1176                 if (ptr_stale(c, k, i)) {
1177                         stale = max(stale, ptr_stale(c, k, i));
1178                         continue;
1179                 }
1180
1181                 cache_bug_on(GC_MARK(g) &&
1182                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1183                              c, "inconsistent ptrs: mark = %llu, level = %i",
1184                              GC_MARK(g), level);
1185
1186                 if (level)
1187                         SET_GC_MARK(g, GC_MARK_METADATA);
1188                 else if (KEY_DIRTY(k))
1189                         SET_GC_MARK(g, GC_MARK_DIRTY);
1190
1191                 /* guard against overflow */
1192                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1193                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1194                                              (1 << 14) - 1));
1195
1196                 BUG_ON(!GC_SECTORS_USED(g));
1197         }
1198
1199         return stale;
1200 }
1201
1202 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1203
1204 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1205 {
1206         uint8_t stale = 0;
1207         unsigned keys = 0, good_keys = 0;
1208         struct bkey *k;
1209         struct btree_iter iter;
1210         struct bset_tree *t;
1211
1212         gc->nodes++;
1213
1214         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1215                 stale = max(stale, btree_mark_key(b, k));
1216                 keys++;
1217
1218                 if (bch_ptr_bad(b, k))
1219                         continue;
1220
1221                 gc->key_bytes += bkey_u64s(k);
1222                 gc->nkeys++;
1223                 good_keys++;
1224
1225                 gc->data += KEY_SIZE(k);
1226         }
1227
1228         for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1229                 btree_bug_on(t->size &&
1230                              bset_written(b, t) &&
1231                              bkey_cmp(&b->key, &t->end) < 0,
1232                              b, "found short btree key in gc");
1233
1234         if (b->c->gc_always_rewrite)
1235                 return true;
1236
1237         if (stale > 10)
1238                 return true;
1239
1240         if ((keys - good_keys) * 2 > keys)
1241                 return true;
1242
1243         return false;
1244 }
1245
1246 #define GC_MERGE_NODES  4U
1247
1248 struct gc_merge_info {
1249         struct btree    *b;
1250         unsigned        keys;
1251 };
1252
1253 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1254                                  struct keylist *, atomic_t *, struct bkey *);
1255
1256 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1257                              struct keylist *keylist, struct gc_stat *gc,
1258                              struct gc_merge_info *r)
1259 {
1260         unsigned i, nodes = 0, keys = 0, blocks;
1261         struct btree *new_nodes[GC_MERGE_NODES];
1262         struct closure cl;
1263         struct bkey *k;
1264
1265         memset(new_nodes, 0, sizeof(new_nodes));
1266         closure_init_stack(&cl);
1267
1268         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1269                 keys += r[nodes++].keys;
1270
1271         blocks = btree_default_blocks(b->c) * 2 / 3;
1272
1273         if (nodes < 2 ||
1274             __set_blocks(b->sets[0].data, keys,
1275                          block_bytes(b->c)) > blocks * (nodes - 1))
1276                 return 0;
1277
1278         for (i = 0; i < nodes; i++) {
1279                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1280                 if (IS_ERR_OR_NULL(new_nodes[i]))
1281                         goto out_nocoalesce;
1282         }
1283
1284         for (i = nodes - 1; i > 0; --i) {
1285                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1286                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1287                 struct bkey *k, *last = NULL;
1288
1289                 keys = 0;
1290
1291                 if (i > 1) {
1292                         for (k = n2->start;
1293                              k < bset_bkey_last(n2);
1294                              k = bkey_next(k)) {
1295                                 if (__set_blocks(n1, n1->keys + keys +
1296                                                  bkey_u64s(k),
1297                                                  block_bytes(b->c)) > blocks)
1298                                         break;
1299
1300                                 last = k;
1301                                 keys += bkey_u64s(k);
1302                         }
1303                 } else {
1304                         /*
1305                          * Last node we're not getting rid of - we're getting
1306                          * rid of the node at r[0]. Have to try and fit all of
1307                          * the remaining keys into this node; we can't ensure
1308                          * they will always fit due to rounding and variable
1309                          * length keys (shouldn't be possible in practice,
1310                          * though)
1311                          */
1312                         if (__set_blocks(n1, n1->keys + n2->keys,
1313                                          block_bytes(b->c)) >
1314                             btree_blocks(new_nodes[i]))
1315                                 goto out_nocoalesce;
1316
1317                         keys = n2->keys;
1318                         /* Take the key of the node we're getting rid of */
1319                         last = &r->b->key;
1320                 }
1321
1322                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1323                        btree_blocks(new_nodes[i]));
1324
1325                 if (last)
1326                         bkey_copy_key(&new_nodes[i]->key, last);
1327
1328                 memcpy(bset_bkey_last(n1),
1329                        n2->start,
1330                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1331
1332                 n1->keys += keys;
1333                 r[i].keys = n1->keys;
1334
1335                 memmove(n2->start,
1336                         bset_bkey_idx(n2, keys),
1337                         (void *) bset_bkey_last(n2) -
1338                         (void *) bset_bkey_idx(n2, keys));
1339
1340                 n2->keys -= keys;
1341
1342                 if (__bch_keylist_realloc(keylist,
1343                                           bkey_u64s(&new_nodes[i]->key)))
1344                         goto out_nocoalesce;
1345
1346                 bch_btree_node_write(new_nodes[i], &cl);
1347                 bch_keylist_add(keylist, &new_nodes[i]->key);
1348         }
1349
1350         for (i = 0; i < nodes; i++) {
1351                 if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key)))
1352                         goto out_nocoalesce;
1353
1354                 make_btree_freeing_key(r[i].b, keylist->top);
1355                 bch_keylist_push(keylist);
1356         }
1357
1358         /* We emptied out this node */
1359         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1360         btree_node_free(new_nodes[0]);
1361         rw_unlock(true, new_nodes[0]);
1362
1363         closure_sync(&cl);
1364
1365         for (i = 0; i < nodes; i++) {
1366                 btree_node_free(r[i].b);
1367                 rw_unlock(true, r[i].b);
1368
1369                 r[i].b = new_nodes[i];
1370         }
1371
1372         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1373         BUG_ON(!bch_keylist_empty(keylist));
1374
1375         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1376         r[nodes - 1].b = ERR_PTR(-EINTR);
1377
1378         trace_bcache_btree_gc_coalesce(nodes);
1379         gc->nodes--;
1380
1381         /* Invalidated our iterator */
1382         return -EINTR;
1383
1384 out_nocoalesce:
1385         closure_sync(&cl);
1386
1387         while ((k = bch_keylist_pop(keylist)))
1388                 if (!bkey_cmp(k, &ZERO_KEY))
1389                         atomic_dec(&b->c->prio_blocked);
1390
1391         for (i = 0; i < nodes; i++)
1392                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1393                         btree_node_free(new_nodes[i]);
1394                         rw_unlock(true, new_nodes[i]);
1395                 }
1396         return 0;
1397 }
1398
1399 static unsigned btree_gc_count_keys(struct btree *b)
1400 {
1401         struct bkey *k;
1402         struct btree_iter iter;
1403         unsigned ret = 0;
1404
1405         for_each_key_filter(b, k, &iter, bch_ptr_bad)
1406                 ret += bkey_u64s(k);
1407
1408         return ret;
1409 }
1410
1411 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1412                             struct closure *writes, struct gc_stat *gc)
1413 {
1414         unsigned i;
1415         int ret = 0;
1416         bool should_rewrite;
1417         struct btree *n;
1418         struct bkey *k;
1419         struct keylist keys;
1420         struct btree_iter iter;
1421         struct gc_merge_info r[GC_MERGE_NODES];
1422         struct gc_merge_info *last = r + GC_MERGE_NODES - 1;
1423
1424         bch_keylist_init(&keys);
1425         bch_btree_iter_init(b, &iter, &b->c->gc_done);
1426
1427         for (i = 0; i < GC_MERGE_NODES; i++)
1428                 r[i].b = ERR_PTR(-EINTR);
1429
1430         while (1) {
1431                 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1432                 if (k) {
1433                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1434                         if (IS_ERR(r->b)) {
1435                                 ret = PTR_ERR(r->b);
1436                                 break;
1437                         }
1438
1439                         r->keys = btree_gc_count_keys(r->b);
1440
1441                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1442                         if (ret)
1443                                 break;
1444                 }
1445
1446                 if (!last->b)
1447                         break;
1448
1449                 if (!IS_ERR(last->b)) {
1450                         should_rewrite = btree_gc_mark_node(last->b, gc);
1451                         if (should_rewrite &&
1452                             !btree_check_reserve(b, NULL)) {
1453                                 n = btree_node_alloc_replacement(last->b,
1454                                                                  false);
1455
1456                                 if (!IS_ERR_OR_NULL(n)) {
1457                                         bch_btree_node_write_sync(n);
1458                                         bch_keylist_add(&keys, &n->key);
1459
1460                                         make_btree_freeing_key(last->b,
1461                                                                keys.top);
1462                                         bch_keylist_push(&keys);
1463
1464                                         btree_node_free(last->b);
1465
1466                                         bch_btree_insert_node(b, op, &keys,
1467                                                               NULL, NULL);
1468                                         BUG_ON(!bch_keylist_empty(&keys));
1469
1470                                         rw_unlock(true, last->b);
1471                                         last->b = n;
1472
1473                                         /* Invalidated our iterator */
1474                                         ret = -EINTR;
1475                                         break;
1476                                 }
1477                         }
1478
1479                         if (last->b->level) {
1480                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1481                                 if (ret)
1482                                         break;
1483                         }
1484
1485                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1486
1487                         /*
1488                          * Must flush leaf nodes before gc ends, since replace
1489                          * operations aren't journalled
1490                          */
1491                         if (btree_node_dirty(last->b))
1492                                 bch_btree_node_write(last->b, writes);
1493                         rw_unlock(true, last->b);
1494                 }
1495
1496                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1497                 r->b = NULL;
1498
1499                 if (need_resched()) {
1500                         ret = -EAGAIN;
1501                         break;
1502                 }
1503         }
1504
1505         for (i = 0; i < GC_MERGE_NODES; i++)
1506                 if (!IS_ERR_OR_NULL(r[i].b)) {
1507                         if (btree_node_dirty(r[i].b))
1508                                 bch_btree_node_write(r[i].b, writes);
1509                         rw_unlock(true, r[i].b);
1510                 }
1511
1512         bch_keylist_free(&keys);
1513
1514         return ret;
1515 }
1516
1517 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1518                              struct closure *writes, struct gc_stat *gc)
1519 {
1520         struct btree *n = NULL;
1521         int ret = 0;
1522         bool should_rewrite;
1523
1524         should_rewrite = btree_gc_mark_node(b, gc);
1525         if (should_rewrite) {
1526                 n = btree_node_alloc_replacement(b, false);
1527
1528                 if (!IS_ERR_OR_NULL(n)) {
1529                         bch_btree_node_write_sync(n);
1530                         bch_btree_set_root(n);
1531                         btree_node_free(b);
1532                         rw_unlock(true, n);
1533
1534                         return -EINTR;
1535                 }
1536         }
1537
1538         if (b->level) {
1539                 ret = btree_gc_recurse(b, op, writes, gc);
1540                 if (ret)
1541                         return ret;
1542         }
1543
1544         bkey_copy_key(&b->c->gc_done, &b->key);
1545
1546         return ret;
1547 }
1548
1549 static void btree_gc_start(struct cache_set *c)
1550 {
1551         struct cache *ca;
1552         struct bucket *b;
1553         unsigned i;
1554
1555         if (!c->gc_mark_valid)
1556                 return;
1557
1558         mutex_lock(&c->bucket_lock);
1559
1560         c->gc_mark_valid = 0;
1561         c->gc_done = ZERO_KEY;
1562
1563         for_each_cache(ca, c, i)
1564                 for_each_bucket(b, ca) {
1565                         b->gc_gen = b->gen;
1566                         if (!atomic_read(&b->pin)) {
1567                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1568                                 SET_GC_SECTORS_USED(b, 0);
1569                         }
1570                 }
1571
1572         mutex_unlock(&c->bucket_lock);
1573 }
1574
1575 size_t bch_btree_gc_finish(struct cache_set *c)
1576 {
1577         size_t available = 0;
1578         struct bucket *b;
1579         struct cache *ca;
1580         unsigned i;
1581
1582         mutex_lock(&c->bucket_lock);
1583
1584         set_gc_sectors(c);
1585         c->gc_mark_valid = 1;
1586         c->need_gc      = 0;
1587
1588         if (c->root)
1589                 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1590                         SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1591                                     GC_MARK_METADATA);
1592
1593         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1594                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1595                             GC_MARK_METADATA);
1596
1597         /* don't reclaim buckets to which writeback keys point */
1598         rcu_read_lock();
1599         for (i = 0; i < c->nr_uuids; i++) {
1600                 struct bcache_device *d = c->devices[i];
1601                 struct cached_dev *dc;
1602                 struct keybuf_key *w, *n;
1603                 unsigned j;
1604
1605                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1606                         continue;
1607                 dc = container_of(d, struct cached_dev, disk);
1608
1609                 spin_lock(&dc->writeback_keys.lock);
1610                 rbtree_postorder_for_each_entry_safe(w, n,
1611                                         &dc->writeback_keys.keys, node)
1612                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1613                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1614                                             GC_MARK_DIRTY);
1615                 spin_unlock(&dc->writeback_keys.lock);
1616         }
1617         rcu_read_unlock();
1618
1619         for_each_cache(ca, c, i) {
1620                 uint64_t *i;
1621
1622                 ca->invalidate_needs_gc = 0;
1623
1624                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1625                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1626
1627                 for (i = ca->prio_buckets;
1628                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1629                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1630
1631                 for_each_bucket(b, ca) {
1632                         b->last_gc      = b->gc_gen;
1633                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1634
1635                         if (!atomic_read(&b->pin) &&
1636                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1637                                 available++;
1638                                 if (!GC_SECTORS_USED(b))
1639                                         bch_bucket_add_unused(ca, b);
1640                         }
1641                 }
1642         }
1643
1644         mutex_unlock(&c->bucket_lock);
1645         return available;
1646 }
1647
1648 static void bch_btree_gc(struct cache_set *c)
1649 {
1650         int ret;
1651         unsigned long available;
1652         struct gc_stat stats;
1653         struct closure writes;
1654         struct btree_op op;
1655         uint64_t start_time = local_clock();
1656
1657         trace_bcache_gc_start(c);
1658
1659         memset(&stats, 0, sizeof(struct gc_stat));
1660         closure_init_stack(&writes);
1661         bch_btree_op_init(&op, SHRT_MAX);
1662
1663         btree_gc_start(c);
1664
1665         do {
1666                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1667                 closure_sync(&writes);
1668
1669                 if (ret && ret != -EAGAIN)
1670                         pr_warn("gc failed!");
1671         } while (ret);
1672
1673         available = bch_btree_gc_finish(c);
1674         wake_up_allocators(c);
1675
1676         bch_time_stats_update(&c->btree_gc_time, start_time);
1677
1678         stats.key_bytes *= sizeof(uint64_t);
1679         stats.data      <<= 9;
1680         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1681         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1682
1683         trace_bcache_gc_end(c);
1684
1685         bch_moving_gc(c);
1686 }
1687
1688 static int bch_gc_thread(void *arg)
1689 {
1690         struct cache_set *c = arg;
1691         struct cache *ca;
1692         unsigned i;
1693
1694         while (1) {
1695 again:
1696                 bch_btree_gc(c);
1697
1698                 set_current_state(TASK_INTERRUPTIBLE);
1699                 if (kthread_should_stop())
1700                         break;
1701
1702                 mutex_lock(&c->bucket_lock);
1703
1704                 for_each_cache(ca, c, i)
1705                         if (ca->invalidate_needs_gc) {
1706                                 mutex_unlock(&c->bucket_lock);
1707                                 set_current_state(TASK_RUNNING);
1708                                 goto again;
1709                         }
1710
1711                 mutex_unlock(&c->bucket_lock);
1712
1713                 try_to_freeze();
1714                 schedule();
1715         }
1716
1717         return 0;
1718 }
1719
1720 int bch_gc_thread_start(struct cache_set *c)
1721 {
1722         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1723         if (IS_ERR(c->gc_thread))
1724                 return PTR_ERR(c->gc_thread);
1725
1726         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1727         return 0;
1728 }
1729
1730 /* Initial partial gc */
1731
1732 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1733                                    unsigned long **seen)
1734 {
1735         int ret = 0;
1736         unsigned i;
1737         struct bkey *k, *p = NULL;
1738         struct bucket *g;
1739         struct btree_iter iter;
1740
1741         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1742                 for (i = 0; i < KEY_PTRS(k); i++) {
1743                         if (!ptr_available(b->c, k, i))
1744                                 continue;
1745
1746                         g = PTR_BUCKET(b->c, k, i);
1747
1748                         if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1749                                                 seen[PTR_DEV(k, i)]) ||
1750                             !ptr_stale(b->c, k, i)) {
1751                                 g->gen = PTR_GEN(k, i);
1752
1753                                 if (b->level)
1754                                         g->prio = BTREE_PRIO;
1755                                 else if (g->prio == BTREE_PRIO)
1756                                         g->prio = INITIAL_PRIO;
1757                         }
1758                 }
1759
1760                 btree_mark_key(b, k);
1761         }
1762
1763         if (b->level) {
1764                 bch_btree_iter_init(b, &iter, NULL);
1765
1766                 do {
1767                         k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
1768                         if (k)
1769                                 btree_node_prefetch(b->c, k, b->level - 1);
1770
1771                         if (p)
1772                                 ret = btree(check_recurse, p, b, op, seen);
1773
1774                         p = k;
1775                 } while (p && !ret);
1776         }
1777
1778         return 0;
1779 }
1780
1781 int bch_btree_check(struct cache_set *c)
1782 {
1783         int ret = -ENOMEM;
1784         unsigned i;
1785         unsigned long *seen[MAX_CACHES_PER_SET];
1786         struct btree_op op;
1787
1788         memset(seen, 0, sizeof(seen));
1789         bch_btree_op_init(&op, SHRT_MAX);
1790
1791         for (i = 0; c->cache[i]; i++) {
1792                 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1793                 seen[i] = kmalloc(n, GFP_KERNEL);
1794                 if (!seen[i])
1795                         goto err;
1796
1797                 /* Disables the seen array until prio_read() uses it too */
1798                 memset(seen[i], 0xFF, n);
1799         }
1800
1801         ret = btree_root(check_recurse, c, &op, seen);
1802 err:
1803         for (i = 0; i < MAX_CACHES_PER_SET; i++)
1804                 kfree(seen[i]);
1805         return ret;
1806 }
1807
1808 /* Btree insertion */
1809
1810 static bool fix_overlapping_extents(struct btree *b, struct bkey *insert,
1811                                     struct btree_iter *iter,
1812                                     struct bkey *replace_key)
1813 {
1814         void subtract_dirty(struct bkey *k, uint64_t offset, int sectors)
1815         {
1816                 if (KEY_DIRTY(k))
1817                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1818                                                      offset, -sectors);
1819         }
1820
1821         uint64_t old_offset;
1822         unsigned old_size, sectors_found = 0;
1823
1824         while (1) {
1825                 struct bkey *k = bch_btree_iter_next(iter);
1826                 if (!k)
1827                         break;
1828
1829                 if (bkey_cmp(&START_KEY(k), insert) >= 0) {
1830                         if (KEY_SIZE(k))
1831                                 break;
1832                         else
1833                                 continue;
1834                 }
1835
1836                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1837                         continue;
1838
1839                 old_offset = KEY_START(k);
1840                 old_size = KEY_SIZE(k);
1841
1842                 /*
1843                  * We might overlap with 0 size extents; we can't skip these
1844                  * because if they're in the set we're inserting to we have to
1845                  * adjust them so they don't overlap with the key we're
1846                  * inserting. But we don't want to check them for replace
1847                  * operations.
1848                  */
1849
1850                 if (replace_key && KEY_SIZE(k)) {
1851                         /*
1852                          * k might have been split since we inserted/found the
1853                          * key we're replacing
1854                          */
1855                         unsigned i;
1856                         uint64_t offset = KEY_START(k) -
1857                                 KEY_START(replace_key);
1858
1859                         /* But it must be a subset of the replace key */
1860                         if (KEY_START(k) < KEY_START(replace_key) ||
1861                             KEY_OFFSET(k) > KEY_OFFSET(replace_key))
1862                                 goto check_failed;
1863
1864                         /* We didn't find a key that we were supposed to */
1865                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
1866                                 goto check_failed;
1867
1868                         if (KEY_PTRS(k) != KEY_PTRS(replace_key) ||
1869                             KEY_DIRTY(k) != KEY_DIRTY(replace_key))
1870                                 goto check_failed;
1871
1872                         /* skip past gen */
1873                         offset <<= 8;
1874
1875                         BUG_ON(!KEY_PTRS(replace_key));
1876
1877                         for (i = 0; i < KEY_PTRS(replace_key); i++)
1878                                 if (k->ptr[i] != replace_key->ptr[i] + offset)
1879                                         goto check_failed;
1880
1881                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1882                 }
1883
1884                 if (bkey_cmp(insert, k) < 0 &&
1885                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1886                         /*
1887                          * We overlapped in the middle of an existing key: that
1888                          * means we have to split the old key. But we have to do
1889                          * slightly different things depending on whether the
1890                          * old key has been written out yet.
1891                          */
1892
1893                         struct bkey *top;
1894
1895                         subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert));
1896
1897                         if (bkey_written(b, k)) {
1898                                 /*
1899                                  * We insert a new key to cover the top of the
1900                                  * old key, and the old key is modified in place
1901                                  * to represent the bottom split.
1902                                  *
1903                                  * It's completely arbitrary whether the new key
1904                                  * is the top or the bottom, but it has to match
1905                                  * up with what btree_sort_fixup() does - it
1906                                  * doesn't check for this kind of overlap, it
1907                                  * depends on us inserting a new key for the top
1908                                  * here.
1909                                  */
1910                                 top = bch_bset_search(b, bset_tree_last(b),
1911                                                       insert);
1912                                 bch_bset_insert(b, top, k);
1913                         } else {
1914                                 BKEY_PADDED(key) temp;
1915                                 bkey_copy(&temp.key, k);
1916                                 bch_bset_insert(b, k, &temp.key);
1917                                 top = bkey_next(k);
1918                         }
1919
1920                         bch_cut_front(insert, top);
1921                         bch_cut_back(&START_KEY(insert), k);
1922                         bch_bset_fix_invalidated_key(b, k);
1923                         return false;
1924                 }
1925
1926                 if (bkey_cmp(insert, k) < 0) {
1927                         bch_cut_front(insert, k);
1928                 } else {
1929                         if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0)
1930                                 old_offset = KEY_START(insert);
1931
1932                         if (bkey_written(b, k) &&
1933                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1934                                 /*
1935                                  * Completely overwrote, so we don't have to
1936                                  * invalidate the binary search tree
1937                                  */
1938                                 bch_cut_front(k, k);
1939                         } else {
1940                                 __bch_cut_back(&START_KEY(insert), k);
1941                                 bch_bset_fix_invalidated_key(b, k);
1942                         }
1943                 }
1944
1945                 subtract_dirty(k, old_offset, old_size - KEY_SIZE(k));
1946         }
1947
1948 check_failed:
1949         if (replace_key) {
1950                 if (!sectors_found) {
1951                         return true;
1952                 } else if (sectors_found < KEY_SIZE(insert)) {
1953                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1954                                        (KEY_SIZE(insert) - sectors_found));
1955                         SET_KEY_SIZE(insert, sectors_found);
1956                 }
1957         }
1958
1959         return false;
1960 }
1961
1962 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1963                              struct bkey *k, struct bkey *replace_key)
1964 {
1965         struct bset *i = btree_bset_last(b);
1966         struct bkey *m, *prev;
1967         unsigned status = BTREE_INSERT_STATUS_INSERT;
1968
1969         BUG_ON(bkey_cmp(k, &b->key) > 0);
1970         BUG_ON(b->level && !KEY_PTRS(k));
1971         BUG_ON(!b->level && !KEY_OFFSET(k));
1972
1973         if (!b->level) {
1974                 struct btree_iter iter;
1975
1976                 /*
1977                  * bset_search() returns the first key that is strictly greater
1978                  * than the search key - but for back merging, we want to find
1979                  * the previous key.
1980                  */
1981                 prev = NULL;
1982                 m = bch_btree_iter_init(b, &iter, PRECEDING_KEY(&START_KEY(k)));
1983
1984                 if (fix_overlapping_extents(b, k, &iter, replace_key)) {
1985                         op->insert_collision = true;
1986                         return false;
1987                 }
1988
1989                 if (KEY_DIRTY(k))
1990                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1991                                                      KEY_START(k), KEY_SIZE(k));
1992
1993                 while (m != bset_bkey_last(i) &&
1994                        bkey_cmp(k, &START_KEY(m)) > 0)
1995                         prev = m, m = bkey_next(m);
1996
1997                 if (key_merging_disabled(b->c))
1998                         goto insert;
1999
2000                 /* prev is in the tree, if we merge we're done */
2001                 status = BTREE_INSERT_STATUS_BACK_MERGE;
2002                 if (prev &&
2003                     bch_bkey_try_merge(b, prev, k))
2004                         goto merged;
2005
2006                 status = BTREE_INSERT_STATUS_OVERWROTE;
2007                 if (m != bset_bkey_last(i) &&
2008                     KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
2009                         goto copy;
2010
2011                 status = BTREE_INSERT_STATUS_FRONT_MERGE;
2012                 if (m != bset_bkey_last(i) &&
2013                     bch_bkey_try_merge(b, k, m))
2014                         goto copy;
2015         } else {
2016                 BUG_ON(replace_key);
2017                 m = bch_bset_search(b, bset_tree_last(b), k);
2018         }
2019
2020 insert: bch_bset_insert(b, m, k);
2021 copy:   bkey_copy(m, k);
2022 merged:
2023         bch_check_keys(b, "%u for %s", status,
2024                        replace_key ? "replace" : "insert");
2025
2026         if (b->level && !KEY_OFFSET(k))
2027                 btree_current_write(b)->prio_blocked++;
2028
2029         trace_bcache_btree_insert_key(b, k, replace_key != NULL, status);
2030
2031         return true;
2032 }
2033
2034 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2035                                   struct keylist *insert_keys,
2036                                   struct bkey *replace_key)
2037 {
2038         bool ret = false;
2039         int oldsize = bch_count_data(b);
2040
2041         while (!bch_keylist_empty(insert_keys)) {
2042                 struct bset *i = write_block(b);
2043                 struct bkey *k = insert_keys->keys;
2044
2045                 if (b->written +
2046                     __set_blocks(i, i->keys + bkey_u64s(k),
2047                                  block_bytes(b->c)) > btree_blocks(b))
2048                         break;
2049
2050                 if (bkey_cmp(k, &b->key) <= 0) {
2051                         if (!b->level)
2052                                 bkey_put(b->c, k);
2053
2054                         ret |= btree_insert_key(b, op, k, replace_key);
2055                         bch_keylist_pop_front(insert_keys);
2056                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2057                         BKEY_PADDED(key) temp;
2058                         bkey_copy(&temp.key, insert_keys->keys);
2059
2060                         bch_cut_back(&b->key, &temp.key);
2061                         bch_cut_front(&b->key, insert_keys->keys);
2062
2063                         ret |= btree_insert_key(b, op, &temp.key, replace_key);
2064                         break;
2065                 } else {
2066                         break;
2067                 }
2068         }
2069
2070         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2071
2072         BUG_ON(bch_count_data(b) < oldsize);
2073         return ret;
2074 }
2075
2076 static int btree_split(struct btree *b, struct btree_op *op,
2077                        struct keylist *insert_keys,
2078                        struct bkey *replace_key)
2079 {
2080         bool split;
2081         struct btree *n1, *n2 = NULL, *n3 = NULL;
2082         uint64_t start_time = local_clock();
2083         struct closure cl;
2084         struct keylist parent_keys;
2085
2086         closure_init_stack(&cl);
2087         bch_keylist_init(&parent_keys);
2088
2089         if (!b->level &&
2090             btree_check_reserve(b, op))
2091                 return -EINTR;
2092
2093         n1 = btree_node_alloc_replacement(b, true);
2094         if (IS_ERR(n1))
2095                 goto err;
2096
2097         split = set_blocks(btree_bset_first(n1),
2098                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2099
2100         if (split) {
2101                 unsigned keys = 0;
2102
2103                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2104
2105                 n2 = bch_btree_node_alloc(b->c, b->level, true);
2106                 if (IS_ERR(n2))
2107                         goto err_free1;
2108
2109                 if (!b->parent) {
2110                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
2111                         if (IS_ERR(n3))
2112                                 goto err_free2;
2113                 }
2114
2115                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2116
2117                 /*
2118                  * Has to be a linear search because we don't have an auxiliary
2119                  * search tree yet
2120                  */
2121
2122                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2123                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2124                                                         keys));
2125
2126                 bkey_copy_key(&n1->key,
2127                               bset_bkey_idx(btree_bset_first(n1), keys));
2128                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2129
2130                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2131                 btree_bset_first(n1)->keys = keys;
2132
2133                 memcpy(btree_bset_first(n2)->start,
2134                        bset_bkey_last(btree_bset_first(n1)),
2135                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2136
2137                 bkey_copy_key(&n2->key, &b->key);
2138
2139                 bch_keylist_add(&parent_keys, &n2->key);
2140                 bch_btree_node_write(n2, &cl);
2141                 rw_unlock(true, n2);
2142         } else {
2143                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2144
2145                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2146         }
2147
2148         bch_keylist_add(&parent_keys, &n1->key);
2149         bch_btree_node_write(n1, &cl);
2150
2151         if (n3) {
2152                 /* Depth increases, make a new root */
2153                 bkey_copy_key(&n3->key, &MAX_KEY);
2154                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2155                 bch_btree_node_write(n3, &cl);
2156
2157                 closure_sync(&cl);
2158                 bch_btree_set_root(n3);
2159                 rw_unlock(true, n3);
2160
2161                 btree_node_free(b);
2162         } else if (!b->parent) {
2163                 /* Root filled up but didn't need to be split */
2164                 closure_sync(&cl);
2165                 bch_btree_set_root(n1);
2166
2167                 btree_node_free(b);
2168         } else {
2169                 /* Split a non root node */
2170                 closure_sync(&cl);
2171                 make_btree_freeing_key(b, parent_keys.top);
2172                 bch_keylist_push(&parent_keys);
2173
2174                 btree_node_free(b);
2175
2176                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2177                 BUG_ON(!bch_keylist_empty(&parent_keys));
2178         }
2179
2180         rw_unlock(true, n1);
2181
2182         bch_time_stats_update(&b->c->btree_split_time, start_time);
2183
2184         return 0;
2185 err_free2:
2186         bkey_put(b->c, &n2->key);
2187         btree_node_free(n2);
2188         rw_unlock(true, n2);
2189 err_free1:
2190         bkey_put(b->c, &n1->key);
2191         btree_node_free(n1);
2192         rw_unlock(true, n1);
2193 err:
2194         WARN(1, "bcache: btree split failed");
2195
2196         if (n3 == ERR_PTR(-EAGAIN) ||
2197             n2 == ERR_PTR(-EAGAIN) ||
2198             n1 == ERR_PTR(-EAGAIN))
2199                 return -EAGAIN;
2200
2201         return -ENOMEM;
2202 }
2203
2204 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2205                                  struct keylist *insert_keys,
2206                                  atomic_t *journal_ref,
2207                                  struct bkey *replace_key)
2208 {
2209         BUG_ON(b->level && replace_key);
2210
2211         if (should_split(b)) {
2212                 if (current->bio_list) {
2213                         op->lock = b->c->root->level + 1;
2214                         return -EAGAIN;
2215                 } else if (op->lock <= b->c->root->level) {
2216                         op->lock = b->c->root->level + 1;
2217                         return -EINTR;
2218                 } else {
2219                         /* Invalidated all iterators */
2220                         return btree_split(b, op, insert_keys, replace_key) ?:
2221                                 -EINTR;
2222                 }
2223         } else {
2224                 BUG_ON(write_block(b) != btree_bset_last(b));
2225
2226                 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2227                         if (!b->level)
2228                                 bch_btree_leaf_dirty(b, journal_ref);
2229                         else
2230                                 bch_btree_node_write_sync(b);
2231                 }
2232
2233                 return 0;
2234         }
2235 }
2236
2237 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2238                                struct bkey *check_key)
2239 {
2240         int ret = -EINTR;
2241         uint64_t btree_ptr = b->key.ptr[0];
2242         unsigned long seq = b->seq;
2243         struct keylist insert;
2244         bool upgrade = op->lock == -1;
2245
2246         bch_keylist_init(&insert);
2247
2248         if (upgrade) {
2249                 rw_unlock(false, b);
2250                 rw_lock(true, b, b->level);
2251
2252                 if (b->key.ptr[0] != btree_ptr ||
2253                     b->seq != seq + 1)
2254                         goto out;
2255         }
2256
2257         SET_KEY_PTRS(check_key, 1);
2258         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2259
2260         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2261
2262         bch_keylist_add(&insert, check_key);
2263
2264         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2265
2266         BUG_ON(!ret && !bch_keylist_empty(&insert));
2267 out:
2268         if (upgrade)
2269                 downgrade_write(&b->lock);
2270         return ret;
2271 }
2272
2273 struct btree_insert_op {
2274         struct btree_op op;
2275         struct keylist  *keys;
2276         atomic_t        *journal_ref;
2277         struct bkey     *replace_key;
2278 };
2279
2280 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2281 {
2282         struct btree_insert_op *op = container_of(b_op,
2283                                         struct btree_insert_op, op);
2284
2285         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2286                                         op->journal_ref, op->replace_key);
2287         if (ret && !bch_keylist_empty(op->keys))
2288                 return ret;
2289         else
2290                 return MAP_DONE;
2291 }
2292
2293 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2294                      atomic_t *journal_ref, struct bkey *replace_key)
2295 {
2296         struct btree_insert_op op;
2297         int ret = 0;
2298
2299         BUG_ON(current->bio_list);
2300         BUG_ON(bch_keylist_empty(keys));
2301
2302         bch_btree_op_init(&op.op, 0);
2303         op.keys         = keys;
2304         op.journal_ref  = journal_ref;
2305         op.replace_key  = replace_key;
2306
2307         while (!ret && !bch_keylist_empty(keys)) {
2308                 op.op.lock = 0;
2309                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2310                                                &START_KEY(keys->keys),
2311                                                btree_insert_fn);
2312         }
2313
2314         if (ret) {
2315                 struct bkey *k;
2316
2317                 pr_err("error %i", ret);
2318
2319                 while ((k = bch_keylist_pop(keys)))
2320                         bkey_put(c, k);
2321         } else if (op.op.insert_collision)
2322                 ret = -ESRCH;
2323
2324         return ret;
2325 }
2326
2327 void bch_btree_set_root(struct btree *b)
2328 {
2329         unsigned i;
2330         struct closure cl;
2331
2332         closure_init_stack(&cl);
2333
2334         trace_bcache_btree_set_root(b);
2335
2336         BUG_ON(!b->written);
2337
2338         for (i = 0; i < KEY_PTRS(&b->key); i++)
2339                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2340
2341         mutex_lock(&b->c->bucket_lock);
2342         list_del_init(&b->list);
2343         mutex_unlock(&b->c->bucket_lock);
2344
2345         b->c->root = b;
2346
2347         bch_journal_meta(b->c, &cl);
2348         closure_sync(&cl);
2349 }
2350
2351 /* Map across nodes or keys */
2352
2353 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2354                                        struct bkey *from,
2355                                        btree_map_nodes_fn *fn, int flags)
2356 {
2357         int ret = MAP_CONTINUE;
2358
2359         if (b->level) {
2360                 struct bkey *k;
2361                 struct btree_iter iter;
2362
2363                 bch_btree_iter_init(b, &iter, from);
2364
2365                 while ((k = bch_btree_iter_next_filter(&iter, b,
2366                                                        bch_ptr_bad))) {
2367                         ret = btree(map_nodes_recurse, k, b,
2368                                     op, from, fn, flags);
2369                         from = NULL;
2370
2371                         if (ret != MAP_CONTINUE)
2372                                 return ret;
2373                 }
2374         }
2375
2376         if (!b->level || flags == MAP_ALL_NODES)
2377                 ret = fn(op, b);
2378
2379         return ret;
2380 }
2381
2382 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2383                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2384 {
2385         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2386 }
2387
2388 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2389                                       struct bkey *from, btree_map_keys_fn *fn,
2390                                       int flags)
2391 {
2392         int ret = MAP_CONTINUE;
2393         struct bkey *k;
2394         struct btree_iter iter;
2395
2396         bch_btree_iter_init(b, &iter, from);
2397
2398         while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) {
2399                 ret = !b->level
2400                         ? fn(op, b, k)
2401                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2402                 from = NULL;
2403
2404                 if (ret != MAP_CONTINUE)
2405                         return ret;
2406         }
2407
2408         if (!b->level && (flags & MAP_END_KEY))
2409                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2410                                      KEY_OFFSET(&b->key), 0));
2411
2412         return ret;
2413 }
2414
2415 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2416                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2417 {
2418         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2419 }
2420
2421 /* Keybuf code */
2422
2423 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2424 {
2425         /* Overlapping keys compare equal */
2426         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2427                 return -1;
2428         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2429                 return 1;
2430         return 0;
2431 }
2432
2433 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2434                                             struct keybuf_key *r)
2435 {
2436         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2437 }
2438
2439 struct refill {
2440         struct btree_op op;
2441         unsigned        nr_found;
2442         struct keybuf   *buf;
2443         struct bkey     *end;
2444         keybuf_pred_fn  *pred;
2445 };
2446
2447 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2448                             struct bkey *k)
2449 {
2450         struct refill *refill = container_of(op, struct refill, op);
2451         struct keybuf *buf = refill->buf;
2452         int ret = MAP_CONTINUE;
2453
2454         if (bkey_cmp(k, refill->end) >= 0) {
2455                 ret = MAP_DONE;
2456                 goto out;
2457         }
2458
2459         if (!KEY_SIZE(k)) /* end key */
2460                 goto out;
2461
2462         if (refill->pred(buf, k)) {
2463                 struct keybuf_key *w;
2464
2465                 spin_lock(&buf->lock);
2466
2467                 w = array_alloc(&buf->freelist);
2468                 if (!w) {
2469                         spin_unlock(&buf->lock);
2470                         return MAP_DONE;
2471                 }
2472
2473                 w->private = NULL;
2474                 bkey_copy(&w->key, k);
2475
2476                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2477                         array_free(&buf->freelist, w);
2478                 else
2479                         refill->nr_found++;
2480
2481                 if (array_freelist_empty(&buf->freelist))
2482                         ret = MAP_DONE;
2483
2484                 spin_unlock(&buf->lock);
2485         }
2486 out:
2487         buf->last_scanned = *k;
2488         return ret;
2489 }
2490
2491 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2492                        struct bkey *end, keybuf_pred_fn *pred)
2493 {
2494         struct bkey start = buf->last_scanned;
2495         struct refill refill;
2496
2497         cond_resched();
2498
2499         bch_btree_op_init(&refill.op, -1);
2500         refill.nr_found = 0;
2501         refill.buf      = buf;
2502         refill.end      = end;
2503         refill.pred     = pred;
2504
2505         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2506                            refill_keybuf_fn, MAP_END_KEY);
2507
2508         trace_bcache_keyscan(refill.nr_found,
2509                              KEY_INODE(&start), KEY_OFFSET(&start),
2510                              KEY_INODE(&buf->last_scanned),
2511                              KEY_OFFSET(&buf->last_scanned));
2512
2513         spin_lock(&buf->lock);
2514
2515         if (!RB_EMPTY_ROOT(&buf->keys)) {
2516                 struct keybuf_key *w;
2517                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2518                 buf->start      = START_KEY(&w->key);
2519
2520                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2521                 buf->end        = w->key;
2522         } else {
2523                 buf->start      = MAX_KEY;
2524                 buf->end        = MAX_KEY;
2525         }
2526
2527         spin_unlock(&buf->lock);
2528 }
2529
2530 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2531 {
2532         rb_erase(&w->node, &buf->keys);
2533         array_free(&buf->freelist, w);
2534 }
2535
2536 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2537 {
2538         spin_lock(&buf->lock);
2539         __bch_keybuf_del(buf, w);
2540         spin_unlock(&buf->lock);
2541 }
2542
2543 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2544                                   struct bkey *end)
2545 {
2546         bool ret = false;
2547         struct keybuf_key *p, *w, s;
2548         s.key = *start;
2549
2550         if (bkey_cmp(end, &buf->start) <= 0 ||
2551             bkey_cmp(start, &buf->end) >= 0)
2552                 return false;
2553
2554         spin_lock(&buf->lock);
2555         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2556
2557         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2558                 p = w;
2559                 w = RB_NEXT(w, node);
2560
2561                 if (p->private)
2562                         ret = true;
2563                 else
2564                         __bch_keybuf_del(buf, p);
2565         }
2566
2567         spin_unlock(&buf->lock);
2568         return ret;
2569 }
2570
2571 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2572 {
2573         struct keybuf_key *w;
2574         spin_lock(&buf->lock);
2575
2576         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2577
2578         while (w && w->private)
2579                 w = RB_NEXT(w, node);
2580
2581         if (w)
2582                 w->private = ERR_PTR(-EINTR);
2583
2584         spin_unlock(&buf->lock);
2585         return w;
2586 }
2587
2588 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2589                                           struct keybuf *buf,
2590                                           struct bkey *end,
2591                                           keybuf_pred_fn *pred)
2592 {
2593         struct keybuf_key *ret;
2594
2595         while (1) {
2596                 ret = bch_keybuf_next(buf);
2597                 if (ret)
2598                         break;
2599
2600                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2601                         pr_debug("scan finished");
2602                         break;
2603                 }
2604
2605                 bch_refill_keybuf(c, buf, end, pred);
2606         }
2607
2608         return ret;
2609 }
2610
2611 void bch_keybuf_init(struct keybuf *buf)
2612 {
2613         buf->last_scanned       = MAX_KEY;
2614         buf->keys               = RB_ROOT;
2615
2616         spin_lock_init(&buf->lock);
2617         array_allocator_init(&buf->freelist);
2618 }
2619
2620 void bch_btree_exit(void)
2621 {
2622         if (btree_io_wq)
2623                 destroy_workqueue(btree_io_wq);
2624 }
2625
2626 int __init bch_btree_init(void)
2627 {
2628         btree_io_wq = create_singlethread_workqueue("bch_btree_io");
2629         if (!btree_io_wq)
2630                 return -ENOMEM;
2631
2632         return 0;
2633 }