drivers/md/bcache/btree.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  *
   4  * Uses a block device as cache for other block devices; optimized for SSDs.
   5  * All allocation is done in buckets, which should match the erase block size
   6  * of the device.
   7  *
   8  * Buckets containing cached data are kept on a heap sorted by priority;
   9  * bucket priority is increased on cache hit, and periodically all the buckets
  10  * on the heap have their priority scaled down. This currently is just used as
  11  * an LRU but in the future should allow for more intelligent heuristics.
  12  *
  13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  14  * counter. Garbage collection is used to remove stale pointers.
  15  *
  16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  17  * as keys are inserted we only sort the pages that have not yet been written.
  18  * When garbage collection is run, we resort the entire node.
  19  *
  20  * All configuration is done via sysfs; see Documentation/bcache.txt.
  21  */
  22
  23 #include "bcache.h"
  24 #include "btree.h"
  25 #include "debug.h"
  26 #include "extents.h"
  27
  28 #include <linux/slab.h>
  29 #include <linux/bitops.h>
  30 #include <linux/freezer.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/random.h>
  35 #include <linux/rcupdate.h>
  36 #include <trace/events/bcache.h>
  37
  38 /*
  39  * Todo:
  40  * register_bcache: Return errors out to userspace correctly
  41  *
  42  * Writeback: don't undirty key until after a cache flush
  43  *
  44  * Create an iterator for key pointers
  45  *
  46  * On btree write error, mark bucket such that it won't be freed from the cache
  47  *
  48  * Journalling:
  49  *   Check for bad keys in replay
  50  *   Propagate barriers
  51  *   Refcount journal entries in journal_replay
  52  *
  53  * Garbage collection:
  54  *   Finish incremental gc
  55  *   Gc should free old UUIDs, data for invalid UUIDs
  56  *
  57  * Provide a way to list backing device UUIDs we have data cached for, and
  58  * probably how long it's been since we've seen them, and a way to invalidate
  59  * dirty data for devices that will never be attached again
  60  *
  61  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  62  * that based on that and how much dirty data we have we can keep writeback
  63  * from being starved
  64  *
  65  * Add a tracepoint or somesuch to watch for writeback starvation
  66  *
  67  * When btree depth > 1 and splitting an interior node, we have to make sure
  68  * alloc_bucket() cannot fail. This should be true but is not completely
  69  * obvious.
  70  *
  71  * Plugging?
  72  *
  73  * If data write is less than hard sector size of ssd, round up offset in open
  74  * bucket to the next whole sector
  75  *
  76  * Superblock needs to be fleshed out for multiple cache devices
  77  *
  78  * Add a sysfs tunable for the number of writeback IOs in flight
  79  *
  80  * Add a sysfs tunable for the number of open data buckets
  81  *
  82  * IO tracking: Can we track when one process is doing io on behalf of another?
  83  * IO tracking: Don't use just an average, weigh more recent stuff higher
  84  *
  85  * Test module load/unload
  86  */
  87
  88 #define MAX_NEED_GC             64
  89 #define MAX_SAVE_PRIO           72
  90
  91 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  92
  93 #define PTR_HASH(c, k)                                                  \
  94         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  95
  96 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
  97
  98 /*
  99  * These macros are for recursing down the btree - they handle the details of
 100  * locking and looking up nodes in the cache for you. They're best treated as
 101  * mere syntax when reading code that uses them.
 102  *
 103  * op->lock determines whether we take a read or a write lock at a given depth.
 104  * If you've got a read lock and find that you need a write lock (i.e. you're
 105  * going to have to split), set op->lock and return -EINTR; btree_root() will
 106  * call you again and you'll have the correct lock.
 107  */
 108
 109 /**
 110  * btree - recurse down the btree on a specified key
 111  * @fn:         function to call, which will be passed the child node
 112  * @key:        key to recurse on
 113  * @b:          parent btree node
 114  * @op:         pointer to struct btree_op
 115  */
 116 #define btree(fn, key, b, op, ...)                                      \
 117 ({                                                                      \
 118         int _r, l = (b)->level - 1;                                     \
 119         bool _w = l <= (op)->lock;                                      \
 120         struct btree *_child = bch_btree_node_get((b)->c, key, l, _w);  \
 121         if (!IS_ERR(_child)) {                                          \
 122                 _child->parent = (b);                                   \
 123                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 124                 rw_unlock(_w, _child);                                  \
 125         } else                                                          \
 126                 _r = PTR_ERR(_child);                                   \
 127         _r;                                                             \
 128 })
 129
 130 /**
 131  * btree_root - call a function on the root of the btree
 132  * @fn:         function to call, which will be passed the child node
 133  * @c:          cache set
 134  * @op:         pointer to struct btree_op
 135  */
 136 #define btree_root(fn, c, op, ...)                                      \
 137 ({                                                                      \
 138         int _r = -EINTR;                                                \
 139         do {                                                            \
 140                 struct btree *_b = (c)->root;                           \
 141                 bool _w = insert_lock(op, _b);                          \
 142                 rw_lock(_w, _b, _b->level);                             \
 143                 if (_b == (c)->root &&                                  \
 144                     _w == insert_lock(op, _b)) {                        \
 145                         _b->parent = NULL;                              \
 146                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 147                 }                                                       \
 148                 rw_unlock(_w, _b);                                      \
 149                 if (_r == -EINTR)                                       \
 150                         schedule();                                     \
 151                 bch_cannibalize_unlock(c);                              \
 152                 if (_r == -ENOSPC) {                                    \
 153                         wait_event((c)->try_wait,                       \
 154                                    !(c)->try_harder);                   \
 155                         _r = -EINTR;                                    \
 156                 }                                                       \
 157         } while (_r == -EINTR);                                         \
 158                                                                         \
 159         finish_wait(&(c)->bucket_wait, &(op)->wait);                    \
 160         _r;                                                             \
 161 })
 162
 163 static inline struct bset *write_block(struct btree *b)
 164 {
 165         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 166 }
 167
 168 static void bch_btree_init_next(struct btree *b)
 169 {
 170         /* If not a leaf node, always sort */
 171         if (b->level && b->keys.nsets)
 172                 bch_btree_sort(&b->keys, &b->c->sort);
 173         else
 174                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 175
 176         if (b->written < btree_blocks(b))
 177                 bch_bset_init_next(&b->keys, write_block(b),
 178                                    bset_magic(&b->c->sb));
 179
 180 }
 181
 182 /* Btree key manipulation */
 183
 184 void bkey_put(struct cache_set *c, struct bkey *k)
 185 {
 186         unsigned i;
 187
 188         for (i = 0; i < KEY_PTRS(k); i++)
 189                 if (ptr_available(c, k, i))
 190                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 191 }
 192
 193 /* Btree IO */
 194
 195 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 196 {
 197         uint64_t crc = b->key.ptr[0];
 198         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 199
 200         crc = bch_crc64_update(crc, data, end - data);
 201         return crc ^ 0xffffffffffffffffULL;
 202 }
 203
 204 void bch_btree_node_read_done(struct btree *b)
 205 {
 206         const char *err = "bad btree header";
 207         struct bset *i = btree_bset_first(b);
 208         struct btree_iter *iter;
 209
 210         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 211         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 212         iter->used = 0;
 213
 214 #ifdef CONFIG_BCACHE_DEBUG
 215         iter->b = &b->keys;
 216 #endif
 217
 218         if (!i->seq)
 219                 goto err;
 220
 221         for (;
 222              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 223              i = write_block(b)) {
 224                 err = "unsupported bset version";
 225                 if (i->version > BCACHE_BSET_VERSION)
 226                         goto err;
 227
 228                 err = "bad btree header";
 229                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 230                     btree_blocks(b))
 231                         goto err;
 232
 233                 err = "bad magic";
 234                 if (i->magic != bset_magic(&b->c->sb))
 235                         goto err;
 236
 237                 err = "bad checksum";
 238                 switch (i->version) {
 239                 case 0:
 240                         if (i->csum != csum_set(i))
 241                                 goto err;
 242                         break;
 243                 case BCACHE_BSET_VERSION:
 244                         if (i->csum != btree_csum_set(b, i))
 245                                 goto err;
 246                         break;
 247                 }
 248
 249                 err = "empty set";
 250                 if (i != b->keys.set[0].data && !i->keys)
 251                         goto err;
 252
 253                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 254
 255                 b->written += set_blocks(i, block_bytes(b->c));
 256         }
 257
 258         err = "corrupted btree";
 259         for (i = write_block(b);
 260              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 261              i = ((void *) i) + block_bytes(b->c))
 262                 if (i->seq == b->keys.set[0].data->seq)
 263                         goto err;
 264
 265         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 266
 267         i = b->keys.set[0].data;
 268         err = "short btree key";
 269         if (b->keys.set[0].size &&
 270             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 271                 goto err;
 272
 273         if (b->written < btree_blocks(b))
 274                 bch_bset_init_next(&b->keys, write_block(b),
 275                                    bset_magic(&b->c->sb));
 276 out:
 277         mempool_free(iter, b->c->fill_iter);
 278         return;
 279 err:
 280         set_btree_node_io_error(b);
 281         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 282                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 283                             bset_block_offset(b, i), i->keys);
 284         goto out;
 285 }
 286
 287 static void btree_node_read_endio(struct bio *bio, int error)
 288 {
 289         struct closure *cl = bio->bi_private;
 290         closure_put(cl);
 291 }
 292
 293 static void bch_btree_node_read(struct btree *b)
 294 {
 295         uint64_t start_time = local_clock();
 296         struct closure cl;
 297         struct bio *bio;
 298
 299         trace_bcache_btree_read(b);
 300
 301         closure_init_stack(&cl);
 302
 303         bio = bch_bbio_alloc(b->c);
 304         bio->bi_rw      = REQ_META|READ_SYNC;
 305         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 306         bio->bi_end_io  = btree_node_read_endio;
 307         bio->bi_private = &cl;
 308
 309         bch_bio_map(bio, b->keys.set[0].data);
 310
 311         bch_submit_bbio(bio, b->c, &b->key, 0);
 312         closure_sync(&cl);
 313
 314         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 315                 set_btree_node_io_error(b);
 316
 317         bch_bbio_free(bio, b->c);
 318
 319         if (btree_node_io_error(b))
 320                 goto err;
 321
 322         bch_btree_node_read_done(b);
 323         bch_time_stats_update(&b->c->btree_read_time, start_time);
 324
 325         return;
 326 err:
 327         bch_cache_set_error(b->c, "io error reading bucket %zu",
 328                             PTR_BUCKET_NR(b->c, &b->key, 0));
 329 }
 330
 331 static void btree_complete_write(struct btree *b, struct btree_write *w)
 332 {
 333         if (w->prio_blocked &&
 334             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 335                 wake_up_allocators(b->c);
 336
 337         if (w->journal) {
 338                 atomic_dec_bug(w->journal);
 339                 __closure_wake_up(&b->c->journal.wait);
 340         }
 341
 342         w->prio_blocked = 0;
 343         w->journal      = NULL;
 344 }
 345
 346 static void btree_node_write_unlock(struct closure *cl)
 347 {
 348         struct btree *b = container_of(cl, struct btree, io);
 349
 350         up(&b->io_mutex);
 351 }
 352
 353 static void __btree_node_write_done(struct closure *cl)
 354 {
 355         struct btree *b = container_of(cl, struct btree, io);
 356         struct btree_write *w = btree_prev_write(b);
 357
 358         bch_bbio_free(b->bio, b->c);
 359         b->bio = NULL;
 360         btree_complete_write(b, w);
 361
 362         if (btree_node_dirty(b))
 363                 schedule_delayed_work(&b->work, 30 * HZ);
 364
 365         closure_return_with_destructor(cl, btree_node_write_unlock);
 366 }
 367
 368 static void btree_node_write_done(struct closure *cl)
 369 {
 370         struct btree *b = container_of(cl, struct btree, io);
 371         struct bio_vec *bv;
 372         int n;
 373
 374         bio_for_each_segment_all(bv, b->bio, n)
 375                 __free_page(bv->bv_page);
 376
 377         __btree_node_write_done(cl);
 378 }
 379
 380 static void btree_node_write_endio(struct bio *bio, int error)
 381 {
 382         struct closure *cl = bio->bi_private;
 383         struct btree *b = container_of(cl, struct btree, io);
 384
 385         if (error)
 386                 set_btree_node_io_error(b);
 387
 388         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 389         closure_put(cl);
 390 }
 391
 392 static void do_btree_node_write(struct btree *b)
 393 {
 394         struct closure *cl = &b->io;
 395         struct bset *i = btree_bset_last(b);
 396         BKEY_PADDED(key) k;
 397
 398         i->version      = BCACHE_BSET_VERSION;
 399         i->csum         = btree_csum_set(b, i);
 400
 401         BUG_ON(b->bio);
 402         b->bio = bch_bbio_alloc(b->c);
 403
 404         b->bio->bi_end_io       = btree_node_write_endio;
 405         b->bio->bi_private      = cl;
 406         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 407         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 408         bch_bio_map(b->bio, i);
 409
 410         /*
 411          * If we're appending to a leaf node, we don't technically need FUA -
 412          * this write just needs to be persisted before the next journal write,
 413          * which will be marked FLUSH|FUA.
 414          *
 415          * Similarly if we're writing a new btree root - the pointer is going to
 416          * be in the next journal entry.
 417          *
 418          * But if we're writing a new btree node (that isn't a root) or
 419          * appending to a non leaf btree node, we need either FUA or a flush
 420          * when we write the parent with the new pointer. FUA is cheaper than a
 421          * flush, and writes appending to leaf nodes aren't blocking anything so
 422          * just make all btree node writes FUA to keep things sane.
 423          */
 424
 425         bkey_copy(&k.key, &b->key);
 426         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 427                        bset_sector_offset(&b->keys, i));
 428
 429         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 430                 int j;
 431                 struct bio_vec *bv;
 432                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 433
 434                 bio_for_each_segment_all(bv, b->bio, j)
 435                         memcpy(page_address(bv->bv_page),
 436                                base + j * PAGE_SIZE, PAGE_SIZE);
 437
 438                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 439
 440                 continue_at(cl, btree_node_write_done, NULL);
 441         } else {
 442                 b->bio->bi_vcnt = 0;
 443                 bch_bio_map(b->bio, i);
 444
 445                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 446
 447                 closure_sync(cl);
 448                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 449         }
 450 }
 451
 452 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 453 {
 454         struct bset *i = btree_bset_last(b);
 455
 456         lockdep_assert_held(&b->write_lock);
 457
 458         trace_bcache_btree_write(b);
 459
 460         BUG_ON(current->bio_list);
 461         BUG_ON(b->written >= btree_blocks(b));
 462         BUG_ON(b->written && !i->keys);
 463         BUG_ON(btree_bset_first(b)->seq != i->seq);
 464         bch_check_keys(&b->keys, "writing");
 465
 466         cancel_delayed_work(&b->work);
 467
 468         /* If caller isn't waiting for write, parent refcount is cache set */
 469         down(&b->io_mutex);
 470         closure_init(&b->io, parent ?: &b->c->cl);
 471
 472         clear_bit(BTREE_NODE_dirty,      &b->flags);
 473         change_bit(BTREE_NODE_write_idx, &b->flags);
 474
 475         do_btree_node_write(b);
 476
 477         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 478                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 479
 480         b->written += set_blocks(i, block_bytes(b->c));
 481 }
 482
 483 void bch_btree_node_write(struct btree *b, struct closure *parent)
 484 {
 485         unsigned nsets = b->keys.nsets;
 486
 487         lockdep_assert_held(&b->lock);
 488
 489         __bch_btree_node_write(b, parent);
 490
 491         /*
 492          * do verify if there was more than one set initially (i.e. we did a
 493          * sort) and we sorted down to a single set:
 494          */
 495         if (nsets && !b->keys.nsets)
 496                 bch_btree_verify(b);
 497
 498         bch_btree_init_next(b);
 499 }
 500
 501 static void bch_btree_node_write_sync(struct btree *b)
 502 {
 503         struct closure cl;
 504
 505         closure_init_stack(&cl);
 506
 507         mutex_lock(&b->write_lock);
 508         bch_btree_node_write(b, &cl);
 509         mutex_unlock(&b->write_lock);
 510
 511         closure_sync(&cl);
 512 }
 513
 514 static void btree_node_write_work(struct work_struct *w)
 515 {
 516         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 517
 518         mutex_lock(&b->write_lock);
 519         if (btree_node_dirty(b))
 520                 __bch_btree_node_write(b, NULL);
 521         mutex_unlock(&b->write_lock);
 522 }
 523
 524 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 525 {
 526         struct bset *i = btree_bset_last(b);
 527         struct btree_write *w = btree_current_write(b);
 528
 529         lockdep_assert_held(&b->write_lock);
 530
 531         BUG_ON(!b->written);
 532         BUG_ON(!i->keys);
 533
 534         if (!btree_node_dirty(b))
 535                 schedule_delayed_work(&b->work, 30 * HZ);
 536
 537         set_btree_node_dirty(b);
 538
 539         if (journal_ref) {
 540                 if (w->journal &&
 541                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 542                         atomic_dec_bug(w->journal);
 543                         w->journal = NULL;
 544                 }
 545
 546                 if (!w->journal) {
 547                         w->journal = journal_ref;
 548                         atomic_inc(w->journal);
 549                 }
 550         }
 551
 552         /* Force write if set is too big */
 553         if (set_bytes(i) > PAGE_SIZE - 48 &&
 554             !current->bio_list)
 555                 bch_btree_node_write(b, NULL);
 556 }
 557
 558 /*
 559  * Btree in memory cache - allocation/freeing
 560  * mca -> memory cache
 561  */
 562
 563 #define mca_reserve(c)  (((c->root && c->root->level)           \
 564                           ? c->root->level : 1) * 8 + 16)
 565 #define mca_can_free(c)                                         \
 566         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 567
 568 static void mca_data_free(struct btree *b)
 569 {
 570         BUG_ON(b->io_mutex.count != 1);
 571
 572         bch_btree_keys_free(&b->keys);
 573
 574         b->c->bucket_cache_used--;
 575         list_move(&b->list, &b->c->btree_cache_freed);
 576 }
 577
 578 static void mca_bucket_free(struct btree *b)
 579 {
 580         BUG_ON(btree_node_dirty(b));
 581
 582         b->key.ptr[0] = 0;
 583         hlist_del_init_rcu(&b->hash);
 584         list_move(&b->list, &b->c->btree_cache_freeable);
 585 }
 586
 587 static unsigned btree_order(struct bkey *k)
 588 {
 589         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 590 }
 591
 592 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 593 {
 594         if (!bch_btree_keys_alloc(&b->keys,
 595                                   max_t(unsigned,
 596                                         ilog2(b->c->btree_pages),
 597                                         btree_order(k)),
 598                                   gfp)) {
 599                 b->c->bucket_cache_used++;
 600                 list_move(&b->list, &b->c->btree_cache);
 601         } else {
 602                 list_move(&b->list, &b->c->btree_cache_freed);
 603         }
 604 }
 605
 606 static struct btree *mca_bucket_alloc(struct cache_set *c,
 607                                       struct bkey *k, gfp_t gfp)
 608 {
 609         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 610         if (!b)
 611                 return NULL;
 612
 613         init_rwsem(&b->lock);
 614         lockdep_set_novalidate_class(&b->lock);
 615         mutex_init(&b->write_lock);
 616         lockdep_set_novalidate_class(&b->write_lock);
 617         INIT_LIST_HEAD(&b->list);
 618         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 619         b->c = c;
 620         sema_init(&b->io_mutex, 1);
 621
 622         mca_data_alloc(b, k, gfp);
 623         return b;
 624 }
 625
 626 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 627 {
 628         struct closure cl;
 629
 630         closure_init_stack(&cl);
 631         lockdep_assert_held(&b->c->bucket_lock);
 632
 633         if (!down_write_trylock(&b->lock))
 634                 return -ENOMEM;
 635
 636         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 637
 638         if (b->keys.page_order < min_order)
 639                 goto out_unlock;
 640
 641         if (!flush) {
 642                 if (btree_node_dirty(b))
 643                         goto out_unlock;
 644
 645                 if (down_trylock(&b->io_mutex))
 646                         goto out_unlock;
 647                 up(&b->io_mutex);
 648         }
 649
 650         mutex_lock(&b->write_lock);
 651         if (btree_node_dirty(b))
 652                 __bch_btree_node_write(b, &cl);
 653         mutex_unlock(&b->write_lock);
 654
 655         closure_sync(&cl);
 656
 657         /* wait for any in flight btree write */
 658         down(&b->io_mutex);
 659         up(&b->io_mutex);
 660
 661         return 0;
 662 out_unlock:
 663         rw_unlock(true, b);
 664         return -ENOMEM;
 665 }
 666
 667 static unsigned long bch_mca_scan(struct shrinker *shrink,
 668                                   struct shrink_control *sc)
 669 {
 670         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 671         struct btree *b, *t;
 672         unsigned long i, nr = sc->nr_to_scan;
 673         unsigned long freed = 0;
 674
 675         if (c->shrinker_disabled)
 676                 return SHRINK_STOP;
 677
 678         if (c->try_harder)
 679                 return SHRINK_STOP;
 680
 681         /* Return -1 if we can't do anything right now */
 682         if (sc->gfp_mask & __GFP_IO)
 683                 mutex_lock(&c->bucket_lock);
 684         else if (!mutex_trylock(&c->bucket_lock))
 685                 return -1;
 686
 687         /*
 688          * It's _really_ critical that we don't free too many btree nodes - we
 689          * have to always leave ourselves a reserve. The reserve is how we
 690          * guarantee that allocating memory for a new btree node can always
 691          * succeed, so that inserting keys into the btree can always succeed and
 692          * IO can always make forward progress:
 693          */
 694         nr /= c->btree_pages;
 695         nr = min_t(unsigned long, nr, mca_can_free(c));
 696
 697         i = 0;
 698         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 699                 if (freed >= nr)
 700                         break;
 701
 702                 if (++i > 3 &&
 703                     !mca_reap(b, 0, false)) {
 704                         mca_data_free(b);
 705                         rw_unlock(true, b);
 706                         freed++;
 707                 }
 708         }
 709
 710         for (i = 0; (nr--) && i < c->bucket_cache_used; i++) {
 711                 if (list_empty(&c->btree_cache))
 712                         goto out;
 713
 714                 b = list_first_entry(&c->btree_cache, struct btree, list);
 715                 list_rotate_left(&c->btree_cache);
 716
 717                 if (!b->accessed &&
 718                     !mca_reap(b, 0, false)) {
 719                         mca_bucket_free(b);
 720                         mca_data_free(b);
 721                         rw_unlock(true, b);
 722                         freed++;
 723                 } else
 724                         b->accessed = 0;
 725         }
 726 out:
 727         mutex_unlock(&c->bucket_lock);
 728         return freed;
 729 }
 730
 731 static unsigned long bch_mca_count(struct shrinker *shrink,
 732                                    struct shrink_control *sc)
 733 {
 734         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 735
 736         if (c->shrinker_disabled)
 737                 return 0;
 738
 739         if (c->try_harder)
 740                 return 0;
 741
 742         return mca_can_free(c) * c->btree_pages;
 743 }
 744
 745 void bch_btree_cache_free(struct cache_set *c)
 746 {
 747         struct btree *b;
 748         struct closure cl;
 749         closure_init_stack(&cl);
 750
 751         if (c->shrink.list.next)
 752                 unregister_shrinker(&c->shrink);
 753
 754         mutex_lock(&c->bucket_lock);
 755
 756 #ifdef CONFIG_BCACHE_DEBUG
 757         if (c->verify_data)
 758                 list_move(&c->verify_data->list, &c->btree_cache);
 759
 760         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 761 #endif
 762
 763         list_splice(&c->btree_cache_freeable,
 764                     &c->btree_cache);
 765
 766         while (!list_empty(&c->btree_cache)) {
 767                 b = list_first_entry(&c->btree_cache, struct btree, list);
 768
 769                 if (btree_node_dirty(b))
 770                         btree_complete_write(b, btree_current_write(b));
 771                 clear_bit(BTREE_NODE_dirty, &b->flags);
 772
 773                 mca_data_free(b);
 774         }
 775
 776         while (!list_empty(&c->btree_cache_freed)) {
 777                 b = list_first_entry(&c->btree_cache_freed,
 778                                      struct btree, list);
 779                 list_del(&b->list);
 780                 cancel_delayed_work_sync(&b->work);
 781                 kfree(b);
 782         }
 783
 784         mutex_unlock(&c->bucket_lock);
 785 }
 786
 787 int bch_btree_cache_alloc(struct cache_set *c)
 788 {
 789         unsigned i;
 790
 791         for (i = 0; i < mca_reserve(c); i++)
 792                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 793                         return -ENOMEM;
 794
 795         list_splice_init(&c->btree_cache,
 796                          &c->btree_cache_freeable);
 797
 798 #ifdef CONFIG_BCACHE_DEBUG
 799         mutex_init(&c->verify_lock);
 800
 801         c->verify_ondisk = (void *)
 802                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 803
 804         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 805
 806         if (c->verify_data &&
 807             c->verify_data->keys.set->data)
 808                 list_del_init(&c->verify_data->list);
 809         else
 810                 c->verify_data = NULL;
 811 #endif
 812
 813         c->shrink.count_objects = bch_mca_count;
 814         c->shrink.scan_objects = bch_mca_scan;
 815         c->shrink.seeks = 4;
 816         c->shrink.batch = c->btree_pages * 2;
 817         register_shrinker(&c->shrink);
 818
 819         return 0;
 820 }
 821
 822 /* Btree in memory cache - hash table */
 823
 824 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 825 {
 826         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 827 }
 828
 829 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 830 {
 831         struct btree *b;
 832
 833         rcu_read_lock();
 834         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 835                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 836                         goto out;
 837         b = NULL;
 838 out:
 839         rcu_read_unlock();
 840         return b;
 841 }
 842
 843 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k)
 844 {
 845         struct btree *b;
 846
 847         trace_bcache_btree_cache_cannibalize(c);
 848
 849         if (!c->try_harder) {
 850                 c->try_harder = current;
 851                 c->try_harder_start = local_clock();
 852         } else if (c->try_harder != current)
 853                 return ERR_PTR(-ENOSPC);
 854
 855         list_for_each_entry_reverse(b, &c->btree_cache, list)
 856                 if (!mca_reap(b, btree_order(k), false))
 857                         return b;
 858
 859         list_for_each_entry_reverse(b, &c->btree_cache, list)
 860                 if (!mca_reap(b, btree_order(k), true))
 861                         return b;
 862
 863         return ERR_PTR(-ENOMEM);
 864 }
 865
 866 /*
 867  * We can only have one thread cannibalizing other cached btree nodes at a time,
 868  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 869  * cannibalize_bucket() will take. This means every time we unlock the root of
 870  * the btree, we need to release this lock if we have it held.
 871  */
 872 static void bch_cannibalize_unlock(struct cache_set *c)
 873 {
 874         if (c->try_harder == current) {
 875                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 876                 c->try_harder = NULL;
 877                 wake_up(&c->try_wait);
 878         }
 879 }
 880
 881 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level)
 882 {
 883         struct btree *b;
 884
 885         BUG_ON(current->bio_list);
 886
 887         lockdep_assert_held(&c->bucket_lock);
 888
 889         if (mca_find(c, k))
 890                 return NULL;
 891
 892         /* btree_free() doesn't free memory; it sticks the node on the end of
 893          * the list. Check if there's any freed nodes there:
 894          */
 895         list_for_each_entry(b, &c->btree_cache_freeable, list)
 896                 if (!mca_reap(b, btree_order(k), false))
 897                         goto out;
 898
 899         /* We never free struct btree itself, just the memory that holds the on
 900          * disk node. Check the freed list before allocating a new one:
 901          */
 902         list_for_each_entry(b, &c->btree_cache_freed, list)
 903                 if (!mca_reap(b, 0, false)) {
 904                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 905                         if (!b->keys.set[0].data)
 906                                 goto err;
 907                         else
 908                                 goto out;
 909                 }
 910
 911         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 912         if (!b)
 913                 goto err;
 914
 915         BUG_ON(!down_write_trylock(&b->lock));
 916         if (!b->keys.set->data)
 917                 goto err;
 918 out:
 919         BUG_ON(b->io_mutex.count != 1);
 920
 921         bkey_copy(&b->key, k);
 922         list_move(&b->list, &c->btree_cache);
 923         hlist_del_init_rcu(&b->hash);
 924         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 925
 926         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 927         b->parent       = (void *) ~0UL;
 928         b->flags        = 0;
 929         b->written      = 0;
 930         b->level        = level;
 931
 932         if (!b->level)
 933                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 934                                     &b->c->expensive_debug_checks);
 935         else
 936                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 937                                     &b->c->expensive_debug_checks);
 938
 939         return b;
 940 err:
 941         if (b)
 942                 rw_unlock(true, b);
 943
 944         b = mca_cannibalize(c, k);
 945         if (!IS_ERR(b))
 946                 goto out;
 947
 948         return b;
 949 }
 950
 951 /**
 952  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 953  * in from disk if necessary.
 954  *
 955  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 956  *
 957  * The btree node will have either a read or a write lock held, depending on
 958  * level and op->lock.
 959  */
 960 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 961                                  int level, bool write)
 962 {
 963         int i = 0;
 964         struct btree *b;
 965
 966         BUG_ON(level < 0);
 967 retry:
 968         b = mca_find(c, k);
 969
 970         if (!b) {
 971                 if (current->bio_list)
 972                         return ERR_PTR(-EAGAIN);
 973
 974                 mutex_lock(&c->bucket_lock);
 975                 b = mca_alloc(c, k, level);
 976                 mutex_unlock(&c->bucket_lock);
 977
 978                 if (!b)
 979                         goto retry;
 980                 if (IS_ERR(b))
 981                         return b;
 982
 983                 bch_btree_node_read(b);
 984
 985                 if (!write)
 986                         downgrade_write(&b->lock);
 987         } else {
 988                 rw_lock(write, b, level);
 989                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 990                         rw_unlock(write, b);
 991                         goto retry;
 992                 }
 993                 BUG_ON(b->level != level);
 994         }
 995
 996         b->accessed = 1;
 997
 998         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
 999                 prefetch(b->keys.set[i].tree);
1000                 prefetch(b->keys.set[i].data);
1001         }
1002
1003         for (; i <= b->keys.nsets; i++)
1004                 prefetch(b->keys.set[i].data);
1005
1006         if (btree_node_io_error(b)) {
1007                 rw_unlock(write, b);
1008                 return ERR_PTR(-EIO);
1009         }
1010
1011         BUG_ON(!b->written);
1012
1013         return b;
1014 }
1015
1016 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1017 {
1018         struct btree *b;
1019
1020         mutex_lock(&c->bucket_lock);
1021         b = mca_alloc(c, k, level);
1022         mutex_unlock(&c->bucket_lock);
1023
1024         if (!IS_ERR_OR_NULL(b)) {
1025                 bch_btree_node_read(b);
1026                 rw_unlock(true, b);
1027         }
1028 }
1029
1030 /* Btree alloc */
1031
1032 static void btree_node_free(struct btree *b)
1033 {
1034         trace_bcache_btree_node_free(b);
1035
1036         BUG_ON(b == b->c->root);
1037
1038         mutex_lock(&b->write_lock);
1039
1040         if (btree_node_dirty(b))
1041                 btree_complete_write(b, btree_current_write(b));
1042         clear_bit(BTREE_NODE_dirty, &b->flags);
1043
1044         mutex_unlock(&b->write_lock);
1045
1046         cancel_delayed_work(&b->work);
1047
1048         mutex_lock(&b->c->bucket_lock);
1049         bch_bucket_free(b->c, &b->key);
1050         mca_bucket_free(b);
1051         mutex_unlock(&b->c->bucket_lock);
1052 }
1053
1054 struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait)
1055 {
1056         BKEY_PADDED(key) k;
1057         struct btree *b = ERR_PTR(-EAGAIN);
1058
1059         mutex_lock(&c->bucket_lock);
1060 retry:
1061         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1062                 goto err;
1063
1064         bkey_put(c, &k.key);
1065         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1066
1067         b = mca_alloc(c, &k.key, level);
1068         if (IS_ERR(b))
1069                 goto err_free;
1070
1071         if (!b) {
1072                 cache_bug(c,
1073                         "Tried to allocate bucket that was in btree cache");
1074                 goto retry;
1075         }
1076
1077         b->accessed = 1;
1078         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1079
1080         mutex_unlock(&c->bucket_lock);
1081
1082         trace_bcache_btree_node_alloc(b);
1083         return b;
1084 err_free:
1085         bch_bucket_free(c, &k.key);
1086 err:
1087         mutex_unlock(&c->bucket_lock);
1088
1089         trace_bcache_btree_node_alloc_fail(b);
1090         return b;
1091 }
1092
1093 static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait)
1094 {
1095         struct btree *n = bch_btree_node_alloc(b->c, b->level, wait);
1096         if (!IS_ERR_OR_NULL(n)) {
1097                 mutex_lock(&n->write_lock);
1098                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1099                 bkey_copy_key(&n->key, &b->key);
1100                 mutex_unlock(&n->write_lock);
1101         }
1102
1103         return n;
1104 }
1105
1106 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1107 {
1108         unsigned i;
1109
1110         mutex_lock(&b->c->bucket_lock);
1111
1112         atomic_inc(&b->c->prio_blocked);
1113
1114         bkey_copy(k, &b->key);
1115         bkey_copy_key(k, &ZERO_KEY);
1116
1117         for (i = 0; i < KEY_PTRS(k); i++)
1118                 SET_PTR_GEN(k, i,
1119                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1120                                         PTR_BUCKET(b->c, &b->key, i)));
1121
1122         mutex_unlock(&b->c->bucket_lock);
1123 }
1124
1125 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1126 {
1127         struct cache_set *c = b->c;
1128         struct cache *ca;
1129         unsigned i, reserve = c->root->level * 2 + 1;
1130         int ret = 0;
1131
1132         mutex_lock(&c->bucket_lock);
1133
1134         for_each_cache(ca, c, i)
1135                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1136                         if (op)
1137                                 prepare_to_wait(&c->bucket_wait, &op->wait,
1138                                                 TASK_UNINTERRUPTIBLE);
1139                         ret = -EINTR;
1140                         break;
1141                 }
1142
1143         mutex_unlock(&c->bucket_lock);
1144         return ret;
1145 }
1146
1147 /* Garbage collection */
1148
1149 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1150                                     struct bkey *k)
1151 {
1152         uint8_t stale = 0;
1153         unsigned i;
1154         struct bucket *g;
1155
1156         /*
1157          * ptr_invalid() can't return true for the keys that mark btree nodes as
1158          * freed, but since ptr_bad() returns true we'll never actually use them
1159          * for anything and thus we don't want mark their pointers here
1160          */
1161         if (!bkey_cmp(k, &ZERO_KEY))
1162                 return stale;
1163
1164         for (i = 0; i < KEY_PTRS(k); i++) {
1165                 if (!ptr_available(c, k, i))
1166                         continue;
1167
1168                 g = PTR_BUCKET(c, k, i);
1169
1170                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1171                         g->gc_gen = PTR_GEN(k, i);
1172
1173                 if (ptr_stale(c, k, i)) {
1174                         stale = max(stale, ptr_stale(c, k, i));
1175                         continue;
1176                 }
1177
1178                 cache_bug_on(GC_MARK(g) &&
1179                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1180                              c, "inconsistent ptrs: mark = %llu, level = %i",
1181                              GC_MARK(g), level);
1182
1183                 if (level)
1184                         SET_GC_MARK(g, GC_MARK_METADATA);
1185                 else if (KEY_DIRTY(k))
1186                         SET_GC_MARK(g, GC_MARK_DIRTY);
1187                 else if (!GC_MARK(g))
1188                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1189
1190                 /* guard against overflow */
1191                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1192                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1193                                              MAX_GC_SECTORS_USED));
1194
1195                 BUG_ON(!GC_SECTORS_USED(g));
1196         }
1197
1198         return stale;
1199 }
1200
1201 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1202
1203 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1204 {
1205         unsigned i;
1206
1207         for (i = 0; i < KEY_PTRS(k); i++)
1208                 if (ptr_available(c, k, i) &&
1209                     !ptr_stale(c, k, i)) {
1210                         struct bucket *b = PTR_BUCKET(c, k, i);
1211
1212                         b->gen = PTR_GEN(k, i);
1213
1214                         if (level && bkey_cmp(k, &ZERO_KEY))
1215                                 b->prio = BTREE_PRIO;
1216                         else if (!level && b->prio == BTREE_PRIO)
1217                                 b->prio = INITIAL_PRIO;
1218                 }
1219
1220         __bch_btree_mark_key(c, level, k);
1221 }
1222
1223 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1224 {
1225         uint8_t stale = 0;
1226         unsigned keys = 0, good_keys = 0;
1227         struct bkey *k;
1228         struct btree_iter iter;
1229         struct bset_tree *t;
1230
1231         gc->nodes++;
1232
1233         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1234                 stale = max(stale, btree_mark_key(b, k));
1235                 keys++;
1236
1237                 if (bch_ptr_bad(&b->keys, k))
1238                         continue;
1239
1240                 gc->key_bytes += bkey_u64s(k);
1241                 gc->nkeys++;
1242                 good_keys++;
1243
1244                 gc->data += KEY_SIZE(k);
1245         }
1246
1247         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1248                 btree_bug_on(t->size &&
1249                              bset_written(&b->keys, t) &&
1250                              bkey_cmp(&b->key, &t->end) < 0,
1251                              b, "found short btree key in gc");
1252
1253         if (b->c->gc_always_rewrite)
1254                 return true;
1255
1256         if (stale > 10)
1257                 return true;
1258
1259         if ((keys - good_keys) * 2 > keys)
1260                 return true;
1261
1262         return false;
1263 }
1264
1265 #define GC_MERGE_NODES  4U
1266
1267 struct gc_merge_info {
1268         struct btree    *b;
1269         unsigned        keys;
1270 };
1271
1272 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1273                                  struct keylist *, atomic_t *, struct bkey *);
1274
1275 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1276                              struct keylist *keylist, struct gc_stat *gc,
1277                              struct gc_merge_info *r)
1278 {
1279         unsigned i, nodes = 0, keys = 0, blocks;
1280         struct btree *new_nodes[GC_MERGE_NODES];
1281         struct closure cl;
1282         struct bkey *k;
1283
1284         memset(new_nodes, 0, sizeof(new_nodes));
1285         closure_init_stack(&cl);
1286
1287         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1288                 keys += r[nodes++].keys;
1289
1290         blocks = btree_default_blocks(b->c) * 2 / 3;
1291
1292         if (nodes < 2 ||
1293             __set_blocks(b->keys.set[0].data, keys,
1294                          block_bytes(b->c)) > blocks * (nodes - 1))
1295                 return 0;
1296
1297         for (i = 0; i < nodes; i++) {
1298                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, false);
1299                 if (IS_ERR_OR_NULL(new_nodes[i]))
1300                         goto out_nocoalesce;
1301         }
1302
1303         for (i = 0; i < nodes; i++)
1304                 mutex_lock(&new_nodes[i]->write_lock);
1305
1306         for (i = nodes - 1; i > 0; --i) {
1307                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1308                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1309                 struct bkey *k, *last = NULL;
1310
1311                 keys = 0;
1312
1313                 if (i > 1) {
1314                         for (k = n2->start;
1315                              k < bset_bkey_last(n2);
1316                              k = bkey_next(k)) {
1317                                 if (__set_blocks(n1, n1->keys + keys +
1318                                                  bkey_u64s(k),
1319                                                  block_bytes(b->c)) > blocks)
1320                                         break;
1321
1322                                 last = k;
1323                                 keys += bkey_u64s(k);
1324                         }
1325                 } else {
1326                         /*
1327                          * Last node we're not getting rid of - we're getting
1328                          * rid of the node at r[0]. Have to try and fit all of
1329                          * the remaining keys into this node; we can't ensure
1330                          * they will always fit due to rounding and variable
1331                          * length keys (shouldn't be possible in practice,
1332                          * though)
1333                          */
1334                         if (__set_blocks(n1, n1->keys + n2->keys,
1335                                          block_bytes(b->c)) >
1336                             btree_blocks(new_nodes[i]))
1337                                 goto out_nocoalesce;
1338
1339                         keys = n2->keys;
1340                         /* Take the key of the node we're getting rid of */
1341                         last = &r->b->key;
1342                 }
1343
1344                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1345                        btree_blocks(new_nodes[i]));
1346
1347                 if (last)
1348                         bkey_copy_key(&new_nodes[i]->key, last);
1349
1350                 memcpy(bset_bkey_last(n1),
1351                        n2->start,
1352                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1353
1354                 n1->keys += keys;
1355                 r[i].keys = n1->keys;
1356
1357                 memmove(n2->start,
1358                         bset_bkey_idx(n2, keys),
1359                         (void *) bset_bkey_last(n2) -
1360                         (void *) bset_bkey_idx(n2, keys));
1361
1362                 n2->keys -= keys;
1363
1364                 if (__bch_keylist_realloc(keylist,
1365                                           bkey_u64s(&new_nodes[i]->key)))
1366                         goto out_nocoalesce;
1367
1368                 bch_btree_node_write(new_nodes[i], &cl);
1369                 bch_keylist_add(keylist, &new_nodes[i]->key);
1370         }
1371
1372         for (i = 0; i < nodes; i++)
1373                 mutex_unlock(&new_nodes[i]->write_lock);
1374
1375         closure_sync(&cl);
1376
1377         /* We emptied out this node */
1378         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1379         btree_node_free(new_nodes[0]);
1380         rw_unlock(true, new_nodes[0]);
1381
1382         for (i = 0; i < nodes; i++) {
1383                 if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key)))
1384                         goto out_nocoalesce;
1385
1386                 make_btree_freeing_key(r[i].b, keylist->top);
1387                 bch_keylist_push(keylist);
1388         }
1389
1390         bch_btree_insert_node(b, op, keylist, NULL, NULL);
1391         BUG_ON(!bch_keylist_empty(keylist));
1392
1393         for (i = 0; i < nodes; i++) {
1394                 btree_node_free(r[i].b);
1395                 rw_unlock(true, r[i].b);
1396
1397                 r[i].b = new_nodes[i];
1398         }
1399
1400         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1401         r[nodes - 1].b = ERR_PTR(-EINTR);
1402
1403         trace_bcache_btree_gc_coalesce(nodes);
1404         gc->nodes--;
1405
1406         /* Invalidated our iterator */
1407         return -EINTR;
1408
1409 out_nocoalesce:
1410         closure_sync(&cl);
1411
1412         while ((k = bch_keylist_pop(keylist)))
1413                 if (!bkey_cmp(k, &ZERO_KEY))
1414                         atomic_dec(&b->c->prio_blocked);
1415
1416         for (i = 0; i < nodes; i++)
1417                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1418                         btree_node_free(new_nodes[i]);
1419                         rw_unlock(true, new_nodes[i]);
1420                 }
1421         return 0;
1422 }
1423
1424 static unsigned btree_gc_count_keys(struct btree *b)
1425 {
1426         struct bkey *k;
1427         struct btree_iter iter;
1428         unsigned ret = 0;
1429
1430         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1431                 ret += bkey_u64s(k);
1432
1433         return ret;
1434 }
1435
1436 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1437                             struct closure *writes, struct gc_stat *gc)
1438 {
1439         int ret = 0;
1440         bool should_rewrite;
1441         struct btree *n;
1442         struct bkey *k;
1443         struct keylist keys;
1444         struct btree_iter iter;
1445         struct gc_merge_info r[GC_MERGE_NODES];
1446         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1447
1448         bch_keylist_init(&keys);
1449         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1450
1451         for (i = r; i < r + ARRAY_SIZE(r); i++)
1452                 i->b = ERR_PTR(-EINTR);
1453
1454         while (1) {
1455                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1456                 if (k) {
1457                         r->b = bch_btree_node_get(b->c, k, b->level - 1, true);
1458                         if (IS_ERR(r->b)) {
1459                                 ret = PTR_ERR(r->b);
1460                                 break;
1461                         }
1462
1463                         r->keys = btree_gc_count_keys(r->b);
1464
1465                         ret = btree_gc_coalesce(b, op, &keys, gc, r);
1466                         if (ret)
1467                                 break;
1468                 }
1469
1470                 if (!last->b)
1471                         break;
1472
1473                 if (!IS_ERR(last->b)) {
1474                         should_rewrite = btree_gc_mark_node(last->b, gc);
1475                         if (should_rewrite &&
1476                             !btree_check_reserve(b, NULL)) {
1477                                 n = btree_node_alloc_replacement(last->b,
1478                                                                  false);
1479
1480                                 if (!IS_ERR_OR_NULL(n)) {
1481                                         bch_btree_node_write_sync(n);
1482
1483                                         bch_keylist_add(&keys, &n->key);
1484
1485                                         make_btree_freeing_key(last->b,
1486                                                                keys.top);
1487                                         bch_keylist_push(&keys);
1488
1489                                         bch_btree_insert_node(b, op, &keys,
1490                                                               NULL, NULL);
1491                                         BUG_ON(!bch_keylist_empty(&keys));
1492
1493                                         btree_node_free(last->b);
1494                                         rw_unlock(true, last->b);
1495                                         last->b = n;
1496
1497                                         /* Invalidated our iterator */
1498                                         ret = -EINTR;
1499                                         break;
1500                                 }
1501                         }
1502
1503                         if (last->b->level) {
1504                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1505                                 if (ret)
1506                                         break;
1507                         }
1508
1509                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1510
1511                         /*
1512                          * Must flush leaf nodes before gc ends, since replace
1513                          * operations aren't journalled
1514                          */
1515                         mutex_lock(&last->b->write_lock);
1516                         if (btree_node_dirty(last->b))
1517                                 bch_btree_node_write(last->b, writes);
1518                         mutex_unlock(&last->b->write_lock);
1519                         rw_unlock(true, last->b);
1520                 }
1521
1522                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1523                 r->b = NULL;
1524
1525                 if (need_resched()) {
1526                         ret = -EAGAIN;
1527                         break;
1528                 }
1529         }
1530
1531         for (i = r; i < r + ARRAY_SIZE(r); i++)
1532                 if (!IS_ERR_OR_NULL(i->b)) {
1533                         mutex_lock(&i->b->write_lock);
1534                         if (btree_node_dirty(i->b))
1535                                 bch_btree_node_write(i->b, writes);
1536                         mutex_unlock(&i->b->write_lock);
1537                         rw_unlock(true, i->b);
1538                 }
1539
1540         bch_keylist_free(&keys);
1541
1542         return ret;
1543 }
1544
1545 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1546                              struct closure *writes, struct gc_stat *gc)
1547 {
1548         struct btree *n = NULL;
1549         int ret = 0;
1550         bool should_rewrite;
1551
1552         should_rewrite = btree_gc_mark_node(b, gc);
1553         if (should_rewrite) {
1554                 n = btree_node_alloc_replacement(b, false);
1555
1556                 if (!IS_ERR_OR_NULL(n)) {
1557                         bch_btree_node_write_sync(n);
1558
1559                         bch_btree_set_root(n);
1560                         btree_node_free(b);
1561                         rw_unlock(true, n);
1562
1563                         return -EINTR;
1564                 }
1565         }
1566
1567         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1568
1569         if (b->level) {
1570                 ret = btree_gc_recurse(b, op, writes, gc);
1571                 if (ret)
1572                         return ret;
1573         }
1574
1575         bkey_copy_key(&b->c->gc_done, &b->key);
1576
1577         return ret;
1578 }
1579
1580 static void btree_gc_start(struct cache_set *c)
1581 {
1582         struct cache *ca;
1583         struct bucket *b;
1584         unsigned i;
1585
1586         if (!c->gc_mark_valid)
1587                 return;
1588
1589         mutex_lock(&c->bucket_lock);
1590
1591         c->gc_mark_valid = 0;
1592         c->gc_done = ZERO_KEY;
1593
1594         for_each_cache(ca, c, i)
1595                 for_each_bucket(b, ca) {
1596                         b->gc_gen = b->gen;
1597                         if (!atomic_read(&b->pin)) {
1598                                 SET_GC_MARK(b, 0);
1599                                 SET_GC_SECTORS_USED(b, 0);
1600                         }
1601                 }
1602
1603         mutex_unlock(&c->bucket_lock);
1604 }
1605
1606 size_t bch_btree_gc_finish(struct cache_set *c)
1607 {
1608         size_t available = 0;
1609         struct bucket *b;
1610         struct cache *ca;
1611         unsigned i;
1612
1613         mutex_lock(&c->bucket_lock);
1614
1615         set_gc_sectors(c);
1616         c->gc_mark_valid = 1;
1617         c->need_gc      = 0;
1618
1619         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1620                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1621                             GC_MARK_METADATA);
1622
1623         /* don't reclaim buckets to which writeback keys point */
1624         rcu_read_lock();
1625         for (i = 0; i < c->nr_uuids; i++) {
1626                 struct bcache_device *d = c->devices[i];
1627                 struct cached_dev *dc;
1628                 struct keybuf_key *w, *n;
1629                 unsigned j;
1630
1631                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1632                         continue;
1633                 dc = container_of(d, struct cached_dev, disk);
1634
1635                 spin_lock(&dc->writeback_keys.lock);
1636                 rbtree_postorder_for_each_entry_safe(w, n,
1637                                         &dc->writeback_keys.keys, node)
1638                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1639                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1640                                             GC_MARK_DIRTY);
1641                 spin_unlock(&dc->writeback_keys.lock);
1642         }
1643         rcu_read_unlock();
1644
1645         for_each_cache(ca, c, i) {
1646                 uint64_t *i;
1647
1648                 ca->invalidate_needs_gc = 0;
1649
1650                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1651                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1652
1653                 for (i = ca->prio_buckets;
1654                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1655                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1656
1657                 for_each_bucket(b, ca) {
1658                         b->last_gc      = b->gc_gen;
1659                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1660
1661                         if (atomic_read(&b->pin))
1662                                 continue;
1663
1664                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1665
1666                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1667                                 available++;
1668
1669                         if (!GC_MARK(b))
1670                                 bch_bucket_add_unused(ca, b);
1671                 }
1672         }
1673
1674         mutex_unlock(&c->bucket_lock);
1675         return available;
1676 }
1677
1678 static void bch_btree_gc(struct cache_set *c)
1679 {
1680         int ret;
1681         unsigned long available;
1682         struct gc_stat stats;
1683         struct closure writes;
1684         struct btree_op op;
1685         uint64_t start_time = local_clock();
1686
1687         trace_bcache_gc_start(c);
1688
1689         memset(&stats, 0, sizeof(struct gc_stat));
1690         closure_init_stack(&writes);
1691         bch_btree_op_init(&op, SHRT_MAX);
1692
1693         btree_gc_start(c);
1694
1695         do {
1696                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1697                 closure_sync(&writes);
1698
1699                 if (ret && ret != -EAGAIN)
1700                         pr_warn("gc failed!");
1701         } while (ret);
1702
1703         available = bch_btree_gc_finish(c);
1704         wake_up_allocators(c);
1705
1706         bch_time_stats_update(&c->btree_gc_time, start_time);
1707
1708         stats.key_bytes *= sizeof(uint64_t);
1709         stats.data      <<= 9;
1710         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1711         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1712
1713         trace_bcache_gc_end(c);
1714
1715         bch_moving_gc(c);
1716 }
1717
1718 static int bch_gc_thread(void *arg)
1719 {
1720         struct cache_set *c = arg;
1721         struct cache *ca;
1722         unsigned i;
1723
1724         while (1) {
1725 again:
1726                 bch_btree_gc(c);
1727
1728                 set_current_state(TASK_INTERRUPTIBLE);
1729                 if (kthread_should_stop())
1730                         break;
1731
1732                 mutex_lock(&c->bucket_lock);
1733
1734                 for_each_cache(ca, c, i)
1735                         if (ca->invalidate_needs_gc) {
1736                                 mutex_unlock(&c->bucket_lock);
1737                                 set_current_state(TASK_RUNNING);
1738                                 goto again;
1739                         }
1740
1741                 mutex_unlock(&c->bucket_lock);
1742
1743                 try_to_freeze();
1744                 schedule();
1745         }
1746
1747         return 0;
1748 }
1749
1750 int bch_gc_thread_start(struct cache_set *c)
1751 {
1752         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1753         if (IS_ERR(c->gc_thread))
1754                 return PTR_ERR(c->gc_thread);
1755
1756         set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1757         return 0;
1758 }
1759
1760 /* Initial partial gc */
1761
1762 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1763 {
1764         int ret = 0;
1765         struct bkey *k, *p = NULL;
1766         struct btree_iter iter;
1767
1768         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1769                 bch_initial_mark_key(b->c, b->level, k);
1770
1771         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1772
1773         if (b->level) {
1774                 bch_btree_iter_init(&b->keys, &iter, NULL);
1775
1776                 do {
1777                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1778                                                        bch_ptr_bad);
1779                         if (k)
1780                                 btree_node_prefetch(b->c, k, b->level - 1);
1781
1782                         if (p)
1783                                 ret = btree(check_recurse, p, b, op);
1784
1785                         p = k;
1786                 } while (p && !ret);
1787         }
1788
1789         return ret;
1790 }
1791
1792 int bch_btree_check(struct cache_set *c)
1793 {
1794         struct btree_op op;
1795
1796         bch_btree_op_init(&op, SHRT_MAX);
1797
1798         return btree_root(check_recurse, c, &op);
1799 }
1800
1801 /* Btree insertion */
1802
1803 static bool btree_insert_key(struct btree *b, struct bkey *k,
1804                              struct bkey *replace_key)
1805 {
1806         unsigned status;
1807
1808         BUG_ON(bkey_cmp(k, &b->key) > 0);
1809
1810         status = bch_btree_insert_key(&b->keys, k, replace_key);
1811         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1812                 bch_check_keys(&b->keys, "%u for %s", status,
1813                                replace_key ? "replace" : "insert");
1814
1815                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1816                                               status);
1817                 return true;
1818         } else
1819                 return false;
1820 }
1821
1822 static size_t insert_u64s_remaining(struct btree *b)
1823 {
1824         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1825
1826         /*
1827          * Might land in the middle of an existing extent and have to split it
1828          */
1829         if (b->keys.ops->is_extents)
1830                 ret -= KEY_MAX_U64S;
1831
1832         return max(ret, 0L);
1833 }
1834
1835 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1836                                   struct keylist *insert_keys,
1837                                   struct bkey *replace_key)
1838 {
1839         bool ret = false;
1840         int oldsize = bch_count_data(&b->keys);
1841
1842         while (!bch_keylist_empty(insert_keys)) {
1843                 struct bkey *k = insert_keys->keys;
1844
1845                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1846                         break;
1847
1848                 if (bkey_cmp(k, &b->key) <= 0) {
1849                         if (!b->level)
1850                                 bkey_put(b->c, k);
1851
1852                         ret |= btree_insert_key(b, k, replace_key);
1853                         bch_keylist_pop_front(insert_keys);
1854                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1855                         BKEY_PADDED(key) temp;
1856                         bkey_copy(&temp.key, insert_keys->keys);
1857
1858                         bch_cut_back(&b->key, &temp.key);
1859                         bch_cut_front(&b->key, insert_keys->keys);
1860
1861                         ret |= btree_insert_key(b, &temp.key, replace_key);
1862                         break;
1863                 } else {
1864                         break;
1865                 }
1866         }
1867
1868         if (!ret)
1869                 op->insert_collision = true;
1870
1871         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1872
1873         BUG_ON(bch_count_data(&b->keys) < oldsize);
1874         return ret;
1875 }
1876
1877 static int btree_split(struct btree *b, struct btree_op *op,
1878                        struct keylist *insert_keys,
1879                        struct bkey *replace_key)
1880 {
1881         bool split;
1882         struct btree *n1, *n2 = NULL, *n3 = NULL;
1883         uint64_t start_time = local_clock();
1884         struct closure cl;
1885         struct keylist parent_keys;
1886
1887         closure_init_stack(&cl);
1888         bch_keylist_init(&parent_keys);
1889
1890         if (!b->level &&
1891             btree_check_reserve(b, op))
1892                 return -EINTR;
1893
1894         n1 = btree_node_alloc_replacement(b, true);
1895         if (IS_ERR(n1))
1896                 goto err;
1897
1898         split = set_blocks(btree_bset_first(n1),
1899                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1900
1901         if (split) {
1902                 unsigned keys = 0;
1903
1904                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1905
1906                 n2 = bch_btree_node_alloc(b->c, b->level, true);
1907                 if (IS_ERR(n2))
1908                         goto err_free1;
1909
1910                 if (!b->parent) {
1911                         n3 = bch_btree_node_alloc(b->c, b->level + 1, true);
1912                         if (IS_ERR(n3))
1913                                 goto err_free2;
1914                 }
1915
1916                 mutex_lock(&n1->write_lock);
1917                 mutex_lock(&n2->write_lock);
1918
1919                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1920
1921                 /*
1922                  * Has to be a linear search because we don't have an auxiliary
1923                  * search tree yet
1924                  */
1925
1926                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
1927                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
1928                                                         keys));
1929
1930                 bkey_copy_key(&n1->key,
1931                               bset_bkey_idx(btree_bset_first(n1), keys));
1932                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
1933
1934                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
1935                 btree_bset_first(n1)->keys = keys;
1936
1937                 memcpy(btree_bset_first(n2)->start,
1938                        bset_bkey_last(btree_bset_first(n1)),
1939                        btree_bset_first(n2)->keys * sizeof(uint64_t));
1940
1941                 bkey_copy_key(&n2->key, &b->key);
1942
1943                 bch_keylist_add(&parent_keys, &n2->key);
1944                 bch_btree_node_write(n2, &cl);
1945                 mutex_unlock(&n2->write_lock);
1946                 rw_unlock(true, n2);
1947         } else {
1948                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
1949
1950                 mutex_lock(&n1->write_lock);
1951                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1952         }
1953
1954         bch_keylist_add(&parent_keys, &n1->key);
1955         bch_btree_node_write(n1, &cl);
1956         mutex_unlock(&n1->write_lock);
1957
1958         if (n3) {
1959                 /* Depth increases, make a new root */
1960                 mutex_lock(&n3->write_lock);
1961                 bkey_copy_key(&n3->key, &MAX_KEY);
1962                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
1963                 bch_btree_node_write(n3, &cl);
1964                 mutex_unlock(&n3->write_lock);
1965
1966                 closure_sync(&cl);
1967                 bch_btree_set_root(n3);
1968                 rw_unlock(true, n3);
1969         } else if (!b->parent) {
1970                 /* Root filled up but didn't need to be split */
1971                 closure_sync(&cl);
1972                 bch_btree_set_root(n1);
1973         } else {
1974                 /* Split a non root node */
1975                 closure_sync(&cl);
1976                 make_btree_freeing_key(b, parent_keys.top);
1977                 bch_keylist_push(&parent_keys);
1978
1979                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
1980                 BUG_ON(!bch_keylist_empty(&parent_keys));
1981         }
1982
1983         btree_node_free(b);
1984         rw_unlock(true, n1);
1985
1986         bch_time_stats_update(&b->c->btree_split_time, start_time);
1987
1988         return 0;
1989 err_free2:
1990         bkey_put(b->c, &n2->key);
1991         btree_node_free(n2);
1992         rw_unlock(true, n2);
1993 err_free1:
1994         bkey_put(b->c, &n1->key);
1995         btree_node_free(n1);
1996         rw_unlock(true, n1);
1997 err:
1998         WARN(1, "bcache: btree split failed");
1999
2000         if (n3 == ERR_PTR(-EAGAIN) ||
2001             n2 == ERR_PTR(-EAGAIN) ||
2002             n1 == ERR_PTR(-EAGAIN))
2003                 return -EAGAIN;
2004
2005         return -ENOMEM;
2006 }
2007
2008 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2009                                  struct keylist *insert_keys,
2010                                  atomic_t *journal_ref,
2011                                  struct bkey *replace_key)
2012 {
2013         struct closure cl;
2014
2015         BUG_ON(b->level && replace_key);
2016
2017         closure_init_stack(&cl);
2018
2019         mutex_lock(&b->write_lock);
2020
2021         if (write_block(b) != btree_bset_last(b) &&
2022             b->keys.last_set_unwritten)
2023                 bch_btree_init_next(b); /* just wrote a set */
2024
2025         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2026                 mutex_unlock(&b->write_lock);
2027                 goto split;
2028         }
2029
2030         BUG_ON(write_block(b) != btree_bset_last(b));
2031
2032         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2033                 if (!b->level)
2034                         bch_btree_leaf_dirty(b, journal_ref);
2035                 else
2036                         bch_btree_node_write(b, &cl);
2037         }
2038
2039         mutex_unlock(&b->write_lock);
2040
2041         /* wait for btree node write if necessary, after unlock */
2042         closure_sync(&cl);
2043
2044         return 0;
2045 split:
2046         if (current->bio_list) {
2047                 op->lock = b->c->root->level + 1;
2048                 return -EAGAIN;
2049         } else if (op->lock <= b->c->root->level) {
2050                 op->lock = b->c->root->level + 1;
2051                 return -EINTR;
2052         } else {
2053                 /* Invalidated all iterators */
2054                 int ret = btree_split(b, op, insert_keys, replace_key);
2055
2056                 if (bch_keylist_empty(insert_keys))
2057                         return 0;
2058                 else if (!ret)
2059                         return -EINTR;
2060                 return ret;
2061         }
2062 }
2063
2064 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2065                                struct bkey *check_key)
2066 {
2067         int ret = -EINTR;
2068         uint64_t btree_ptr = b->key.ptr[0];
2069         unsigned long seq = b->seq;
2070         struct keylist insert;
2071         bool upgrade = op->lock == -1;
2072
2073         bch_keylist_init(&insert);
2074
2075         if (upgrade) {
2076                 rw_unlock(false, b);
2077                 rw_lock(true, b, b->level);
2078
2079                 if (b->key.ptr[0] != btree_ptr ||
2080                     b->seq != seq + 1)
2081                         goto out;
2082         }
2083
2084         SET_KEY_PTRS(check_key, 1);
2085         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2086
2087         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2088
2089         bch_keylist_add(&insert, check_key);
2090
2091         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2092
2093         BUG_ON(!ret && !bch_keylist_empty(&insert));
2094 out:
2095         if (upgrade)
2096                 downgrade_write(&b->lock);
2097         return ret;
2098 }
2099
2100 struct btree_insert_op {
2101         struct btree_op op;
2102         struct keylist  *keys;
2103         atomic_t        *journal_ref;
2104         struct bkey     *replace_key;
2105 };
2106
2107 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2108 {
2109         struct btree_insert_op *op = container_of(b_op,
2110                                         struct btree_insert_op, op);
2111
2112         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2113                                         op->journal_ref, op->replace_key);
2114         if (ret && !bch_keylist_empty(op->keys))
2115                 return ret;
2116         else
2117                 return MAP_DONE;
2118 }
2119
2120 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2121                      atomic_t *journal_ref, struct bkey *replace_key)
2122 {
2123         struct btree_insert_op op;
2124         int ret = 0;
2125
2126         BUG_ON(current->bio_list);
2127         BUG_ON(bch_keylist_empty(keys));
2128
2129         bch_btree_op_init(&op.op, 0);
2130         op.keys         = keys;
2131         op.journal_ref  = journal_ref;
2132         op.replace_key  = replace_key;
2133
2134         while (!ret && !bch_keylist_empty(keys)) {
2135                 op.op.lock = 0;
2136                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2137                                                &START_KEY(keys->keys),
2138                                                btree_insert_fn);
2139         }
2140
2141         if (ret) {
2142                 struct bkey *k;
2143
2144                 pr_err("error %i", ret);
2145
2146                 while ((k = bch_keylist_pop(keys)))
2147                         bkey_put(c, k);
2148         } else if (op.op.insert_collision)
2149                 ret = -ESRCH;
2150
2151         return ret;
2152 }
2153
2154 void bch_btree_set_root(struct btree *b)
2155 {
2156         unsigned i;
2157         struct closure cl;
2158
2159         closure_init_stack(&cl);
2160
2161         trace_bcache_btree_set_root(b);
2162
2163         BUG_ON(!b->written);
2164
2165         for (i = 0; i < KEY_PTRS(&b->key); i++)
2166                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2167
2168         mutex_lock(&b->c->bucket_lock);
2169         list_del_init(&b->list);
2170         mutex_unlock(&b->c->bucket_lock);
2171
2172         b->c->root = b;
2173
2174         bch_journal_meta(b->c, &cl);
2175         closure_sync(&cl);
2176 }
2177
2178 /* Map across nodes or keys */
2179
2180 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2181                                        struct bkey *from,
2182                                        btree_map_nodes_fn *fn, int flags)
2183 {
2184         int ret = MAP_CONTINUE;
2185
2186         if (b->level) {
2187                 struct bkey *k;
2188                 struct btree_iter iter;
2189
2190                 bch_btree_iter_init(&b->keys, &iter, from);
2191
2192                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2193                                                        bch_ptr_bad))) {
2194                         ret = btree(map_nodes_recurse, k, b,
2195                                     op, from, fn, flags);
2196                         from = NULL;
2197
2198                         if (ret != MAP_CONTINUE)
2199                                 return ret;
2200                 }
2201         }
2202
2203         if (!b->level || flags == MAP_ALL_NODES)
2204                 ret = fn(op, b);
2205
2206         return ret;
2207 }
2208
2209 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2210                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2211 {
2212         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2213 }
2214
2215 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2216                                       struct bkey *from, btree_map_keys_fn *fn,
2217                                       int flags)
2218 {
2219         int ret = MAP_CONTINUE;
2220         struct bkey *k;
2221         struct btree_iter iter;
2222
2223         bch_btree_iter_init(&b->keys, &iter, from);
2224
2225         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2226                 ret = !b->level
2227                         ? fn(op, b, k)
2228                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2229                 from = NULL;
2230
2231                 if (ret != MAP_CONTINUE)
2232                         return ret;
2233         }
2234
2235         if (!b->level && (flags & MAP_END_KEY))
2236                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2237                                      KEY_OFFSET(&b->key), 0));
2238
2239         return ret;
2240 }
2241
2242 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2243                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2244 {
2245         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2246 }
2247
2248 /* Keybuf code */
2249
2250 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2251 {
2252         /* Overlapping keys compare equal */
2253         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2254                 return -1;
2255         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2256                 return 1;
2257         return 0;
2258 }
2259
2260 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2261                                             struct keybuf_key *r)
2262 {
2263         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2264 }
2265
2266 struct refill {
2267         struct btree_op op;
2268         unsigned        nr_found;
2269         struct keybuf   *buf;
2270         struct bkey     *end;
2271         keybuf_pred_fn  *pred;
2272 };
2273
2274 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2275                             struct bkey *k)
2276 {
2277         struct refill *refill = container_of(op, struct refill, op);
2278         struct keybuf *buf = refill->buf;
2279         int ret = MAP_CONTINUE;
2280
2281         if (bkey_cmp(k, refill->end) >= 0) {
2282                 ret = MAP_DONE;
2283                 goto out;
2284         }
2285
2286         if (!KEY_SIZE(k)) /* end key */
2287                 goto out;
2288
2289         if (refill->pred(buf, k)) {
2290                 struct keybuf_key *w;
2291
2292                 spin_lock(&buf->lock);
2293
2294                 w = array_alloc(&buf->freelist);
2295                 if (!w) {
2296                         spin_unlock(&buf->lock);
2297                         return MAP_DONE;
2298                 }
2299
2300                 w->private = NULL;
2301                 bkey_copy(&w->key, k);
2302
2303                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2304                         array_free(&buf->freelist, w);
2305                 else
2306                         refill->nr_found++;
2307
2308                 if (array_freelist_empty(&buf->freelist))
2309                         ret = MAP_DONE;
2310
2311                 spin_unlock(&buf->lock);
2312         }
2313 out:
2314         buf->last_scanned = *k;
2315         return ret;
2316 }
2317
2318 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2319                        struct bkey *end, keybuf_pred_fn *pred)
2320 {
2321         struct bkey start = buf->last_scanned;
2322         struct refill refill;
2323
2324         cond_resched();
2325
2326         bch_btree_op_init(&refill.op, -1);
2327         refill.nr_found = 0;
2328         refill.buf      = buf;
2329         refill.end      = end;
2330         refill.pred     = pred;
2331
2332         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2333                            refill_keybuf_fn, MAP_END_KEY);
2334
2335         trace_bcache_keyscan(refill.nr_found,
2336                              KEY_INODE(&start), KEY_OFFSET(&start),
2337                              KEY_INODE(&buf->last_scanned),
2338                              KEY_OFFSET(&buf->last_scanned));
2339
2340         spin_lock(&buf->lock);
2341
2342         if (!RB_EMPTY_ROOT(&buf->keys)) {
2343                 struct keybuf_key *w;
2344                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2345                 buf->start      = START_KEY(&w->key);
2346
2347                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2348                 buf->end        = w->key;
2349         } else {
2350                 buf->start      = MAX_KEY;
2351                 buf->end        = MAX_KEY;
2352         }
2353
2354         spin_unlock(&buf->lock);
2355 }
2356
2357 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2358 {
2359         rb_erase(&w->node, &buf->keys);
2360         array_free(&buf->freelist, w);
2361 }
2362
2363 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2364 {
2365         spin_lock(&buf->lock);
2366         __bch_keybuf_del(buf, w);
2367         spin_unlock(&buf->lock);
2368 }
2369
2370 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2371                                   struct bkey *end)
2372 {
2373         bool ret = false;
2374         struct keybuf_key *p, *w, s;
2375         s.key = *start;
2376
2377         if (bkey_cmp(end, &buf->start) <= 0 ||
2378             bkey_cmp(start, &buf->end) >= 0)
2379                 return false;
2380
2381         spin_lock(&buf->lock);
2382         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2383
2384         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2385                 p = w;
2386                 w = RB_NEXT(w, node);
2387
2388                 if (p->private)
2389                         ret = true;
2390                 else
2391                         __bch_keybuf_del(buf, p);
2392         }
2393
2394         spin_unlock(&buf->lock);
2395         return ret;
2396 }
2397
2398 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2399 {
2400         struct keybuf_key *w;
2401         spin_lock(&buf->lock);
2402
2403         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2404
2405         while (w && w->private)
2406                 w = RB_NEXT(w, node);
2407
2408         if (w)
2409                 w->private = ERR_PTR(-EINTR);
2410
2411         spin_unlock(&buf->lock);
2412         return w;
2413 }
2414
2415 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2416                                           struct keybuf *buf,
2417                                           struct bkey *end,
2418                                           keybuf_pred_fn *pred)
2419 {
2420         struct keybuf_key *ret;
2421
2422         while (1) {
2423                 ret = bch_keybuf_next(buf);
2424                 if (ret)
2425                         break;
2426
2427                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2428                         pr_debug("scan finished");
2429                         break;
2430                 }
2431
2432                 bch_refill_keybuf(c, buf, end, pred);
2433         }
2434
2435         return ret;
2436 }
2437
2438 void bch_keybuf_init(struct keybuf *buf)
2439 {
2440         buf->last_scanned       = MAX_KEY;
2441         buf->keys               = RB_ROOT;
2442
2443         spin_lock_init(&buf->lock);
2444         array_allocator_init(&buf->freelist);
2445 }