drivers/md/bcache/btree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  *
   5  * Uses a block device as cache for other block devices; optimized for SSDs.
   6  * All allocation is done in buckets, which should match the erase block size
   7  * of the device.
   8  *
   9  * Buckets containing cached data are kept on a heap sorted by priority;
  10  * bucket priority is increased on cache hit, and periodically all the buckets
  11  * on the heap have their priority scaled down. This currently is just used as
  12  * an LRU but in the future should allow for more intelligent heuristics.
  13  *
  14  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  15  * counter. Garbage collection is used to remove stale pointers.
  16  *
  17  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  18  * as keys are inserted we only sort the pages that have not yet been written.
  19  * When garbage collection is run, we resort the entire node.
  20  *
  21  * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
  22  */
  23
  24 #include "bcache.h"
  25 #include "btree.h"
  26 #include "debug.h"
  27 #include "extents.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.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 <linux/sched/clock.h>
  37 #include <linux/rculist.h>
  38 #include <linux/delay.h>
  39 #include <trace/events/bcache.h>
  40
  41 /*
  42  * Todo:
  43  * register_bcache: Return errors out to userspace correctly
  44  *
  45  * Writeback: don't undirty key until after a cache flush
  46  *
  47  * Create an iterator for key pointers
  48  *
  49  * On btree write error, mark bucket such that it won't be freed from the cache
  50  *
  51  * Journalling:
  52  *   Check for bad keys in replay
  53  *   Propagate barriers
  54  *   Refcount journal entries in journal_replay
  55  *
  56  * Garbage collection:
  57  *   Finish incremental gc
  58  *   Gc should free old UUIDs, data for invalid UUIDs
  59  *
  60  * Provide a way to list backing device UUIDs we have data cached for, and
  61  * probably how long it's been since we've seen them, and a way to invalidate
  62  * dirty data for devices that will never be attached again
  63  *
  64  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  65  * that based on that and how much dirty data we have we can keep writeback
  66  * from being starved
  67  *
  68  * Add a tracepoint or somesuch to watch for writeback starvation
  69  *
  70  * When btree depth > 1 and splitting an interior node, we have to make sure
  71  * alloc_bucket() cannot fail. This should be true but is not completely
  72  * obvious.
  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  * Superblock needs to be fleshed out for multiple cache devices
  80  *
  81  * Add a sysfs tunable for the number of writeback IOs in flight
  82  *
  83  * Add a sysfs tunable for the number of open data buckets
  84  *
  85  * IO tracking: Can we track when one process is doing io on behalf of another?
  86  * IO tracking: Don't use just an average, weigh more recent stuff higher
  87  *
  88  * Test module load/unload
  89  */
  90
  91 #define MAX_NEED_GC             64
  92 #define MAX_SAVE_PRIO           72
  93 #define MAX_GC_TIMES            100
  94 #define MIN_GC_NODES            100
  95 #define GC_SLEEP_MS             100
  96
  97 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  98
  99 #define PTR_HASH(c, k)                                                  \
 100         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 101
 102 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 103
 104 /*
 105  * These macros are for recursing down the btree - they handle the details of
 106  * locking and looking up nodes in the cache for you. They're best treated as
 107  * mere syntax when reading code that uses them.
 108  *
 109  * op->lock determines whether we take a read or a write lock at a given depth.
 110  * If you've got a read lock and find that you need a write lock (i.e. you're
 111  * going to have to split), set op->lock and return -EINTR; btree_root() will
 112  * call you again and you'll have the correct lock.
 113  */
 114
 115 /**
 116  * btree - recurse down the btree on a specified key
 117  * @fn:         function to call, which will be passed the child node
 118  * @key:        key to recurse on
 119  * @b:          parent btree node
 120  * @op:         pointer to struct btree_op
 121  */
 122 #define btree(fn, key, b, op, ...)                                      \
 123 ({                                                                      \
 124         int _r, l = (b)->level - 1;                                     \
 125         bool _w = l <= (op)->lock;                                      \
 126         struct btree *_child = bch_btree_node_get((b)->c, op, key, l,   \
 127                                                   _w, b);               \
 128         if (!IS_ERR(_child)) {                                          \
 129                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 130                 rw_unlock(_w, _child);                                  \
 131         } else                                                          \
 132                 _r = PTR_ERR(_child);                                   \
 133         _r;                                                             \
 134 })
 135
 136 /**
 137  * btree_root - call a function on the root of the btree
 138  * @fn:         function to call, which will be passed the child node
 139  * @c:          cache set
 140  * @op:         pointer to struct btree_op
 141  */
 142 #define btree_root(fn, c, op, ...)                                      \
 143 ({                                                                      \
 144         int _r = -EINTR;                                                \
 145         do {                                                            \
 146                 struct btree *_b = (c)->root;                           \
 147                 bool _w = insert_lock(op, _b);                          \
 148                 rw_lock(_w, _b, _b->level);                             \
 149                 if (_b == (c)->root &&                                  \
 150                     _w == insert_lock(op, _b)) {                        \
 151                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 152                 }                                                       \
 153                 rw_unlock(_w, _b);                                      \
 154                 bch_cannibalize_unlock(c);                              \
 155                 if (_r == -EINTR)                                       \
 156                         schedule();                                     \
 157         } while (_r == -EINTR);                                         \
 158                                                                         \
 159         finish_wait(&(c)->btree_cache_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 int 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_NOIO);
 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)
 288 {
 289         struct closure *cl = bio->bi_private;
 290
 291         closure_put(cl);
 292 }
 293
 294 static void bch_btree_node_read(struct btree *b)
 295 {
 296         uint64_t start_time = local_clock();
 297         struct closure cl;
 298         struct bio *bio;
 299
 300         trace_bcache_btree_read(b);
 301
 302         closure_init_stack(&cl);
 303
 304         bio = bch_bbio_alloc(b->c);
 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         bio->bi_opf = REQ_OP_READ | REQ_META;
 309
 310         bch_bio_map(bio, b->keys.set[0].data);
 311
 312         bch_submit_bbio(bio, b->c, &b->key, 0);
 313         closure_sync(&cl);
 314
 315         if (bio->bi_status)
 316                 set_btree_node_io_error(b);
 317
 318         bch_bbio_free(bio, b->c);
 319
 320         if (btree_node_io_error(b))
 321                 goto err;
 322
 323         bch_btree_node_read_done(b);
 324         bch_time_stats_update(&b->c->btree_read_time, start_time);
 325
 326         return;
 327 err:
 328         bch_cache_set_error(b->c, "io error reading bucket %zu",
 329                             PTR_BUCKET_NR(b->c, &b->key, 0));
 330 }
 331
 332 static void btree_complete_write(struct btree *b, struct btree_write *w)
 333 {
 334         if (w->prio_blocked &&
 335             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 336                 wake_up_allocators(b->c);
 337
 338         if (w->journal) {
 339                 atomic_dec_bug(w->journal);
 340                 __closure_wake_up(&b->c->journal.wait);
 341         }
 342
 343         w->prio_blocked = 0;
 344         w->journal      = NULL;
 345 }
 346
 347 static void btree_node_write_unlock(struct closure *cl)
 348 {
 349         struct btree *b = container_of(cl, struct btree, io);
 350
 351         up(&b->io_mutex);
 352 }
 353
 354 static void __btree_node_write_done(struct closure *cl)
 355 {
 356         struct btree *b = container_of(cl, struct btree, io);
 357         struct btree_write *w = btree_prev_write(b);
 358
 359         bch_bbio_free(b->bio, b->c);
 360         b->bio = NULL;
 361         btree_complete_write(b, w);
 362
 363         if (btree_node_dirty(b))
 364                 schedule_delayed_work(&b->work, 30 * HZ);
 365
 366         closure_return_with_destructor(cl, btree_node_write_unlock);
 367 }
 368
 369 static void btree_node_write_done(struct closure *cl)
 370 {
 371         struct btree *b = container_of(cl, struct btree, io);
 372
 373         bio_free_pages(b->bio);
 374         __btree_node_write_done(cl);
 375 }
 376
 377 static void btree_node_write_endio(struct bio *bio)
 378 {
 379         struct closure *cl = bio->bi_private;
 380         struct btree *b = container_of(cl, struct btree, io);
 381
 382         if (bio->bi_status)
 383                 set_btree_node_io_error(b);
 384
 385         bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 386         closure_put(cl);
 387 }
 388
 389 static void do_btree_node_write(struct btree *b)
 390 {
 391         struct closure *cl = &b->io;
 392         struct bset *i = btree_bset_last(b);
 393         BKEY_PADDED(key) k;
 394
 395         i->version      = BCACHE_BSET_VERSION;
 396         i->csum         = btree_csum_set(b, i);
 397
 398         BUG_ON(b->bio);
 399         b->bio = bch_bbio_alloc(b->c);
 400
 401         b->bio->bi_end_io       = btree_node_write_endio;
 402         b->bio->bi_private      = cl;
 403         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 404         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 405         bch_bio_map(b->bio, i);
 406
 407         /*
 408          * If we're appending to a leaf node, we don't technically need FUA -
 409          * this write just needs to be persisted before the next journal write,
 410          * which will be marked FLUSH|FUA.
 411          *
 412          * Similarly if we're writing a new btree root - the pointer is going to
 413          * be in the next journal entry.
 414          *
 415          * But if we're writing a new btree node (that isn't a root) or
 416          * appending to a non leaf btree node, we need either FUA or a flush
 417          * when we write the parent with the new pointer. FUA is cheaper than a
 418          * flush, and writes appending to leaf nodes aren't blocking anything so
 419          * just make all btree node writes FUA to keep things sane.
 420          */
 421
 422         bkey_copy(&k.key, &b->key);
 423         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 424                        bset_sector_offset(&b->keys, i));
 425
 426         if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 427                 int j;
 428                 struct bio_vec *bv;
 429                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 430
 431                 bio_for_each_segment_all(bv, b->bio, j)
 432                         memcpy(page_address(bv->bv_page),
 433                                base + j * PAGE_SIZE, PAGE_SIZE);
 434
 435                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 436
 437                 continue_at(cl, btree_node_write_done, NULL);
 438         } else {
 439                 /*
 440                  * No problem for multipage bvec since the bio is
 441                  * just allocated
 442                  */
 443                 b->bio->bi_vcnt = 0;
 444                 bch_bio_map(b->bio, i);
 445
 446                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 447
 448                 closure_sync(cl);
 449                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 450         }
 451 }
 452
 453 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 454 {
 455         struct bset *i = btree_bset_last(b);
 456
 457         lockdep_assert_held(&b->write_lock);
 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->keys, "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         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 479                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 480
 481         b->written += set_blocks(i, block_bytes(b->c));
 482 }
 483
 484 void bch_btree_node_write(struct btree *b, struct closure *parent)
 485 {
 486         unsigned int nsets = b->keys.nsets;
 487
 488         lockdep_assert_held(&b->lock);
 489
 490         __bch_btree_node_write(b, parent);
 491
 492         /*
 493          * do verify if there was more than one set initially (i.e. we did a
 494          * sort) and we sorted down to a single set:
 495          */
 496         if (nsets && !b->keys.nsets)
 497                 bch_btree_verify(b);
 498
 499         bch_btree_init_next(b);
 500 }
 501
 502 static void bch_btree_node_write_sync(struct btree *b)
 503 {
 504         struct closure cl;
 505
 506         closure_init_stack(&cl);
 507
 508         mutex_lock(&b->write_lock);
 509         bch_btree_node_write(b, &cl);
 510         mutex_unlock(&b->write_lock);
 511
 512         closure_sync(&cl);
 513 }
 514
 515 static void btree_node_write_work(struct work_struct *w)
 516 {
 517         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 518
 519         mutex_lock(&b->write_lock);
 520         if (btree_node_dirty(b))
 521                 __bch_btree_node_write(b, NULL);
 522         mutex_unlock(&b->write_lock);
 523 }
 524
 525 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 526 {
 527         struct bset *i = btree_bset_last(b);
 528         struct btree_write *w = btree_current_write(b);
 529
 530         lockdep_assert_held(&b->write_lock);
 531
 532         BUG_ON(!b->written);
 533         BUG_ON(!i->keys);
 534
 535         if (!btree_node_dirty(b))
 536                 schedule_delayed_work(&b->work, 30 * HZ);
 537
 538         set_btree_node_dirty(b);
 539
 540         if (journal_ref) {
 541                 if (w->journal &&
 542                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 543                         atomic_dec_bug(w->journal);
 544                         w->journal = NULL;
 545                 }
 546
 547                 if (!w->journal) {
 548                         w->journal = journal_ref;
 549                         atomic_inc(w->journal);
 550                 }
 551         }
 552
 553         /* Force write if set is too big */
 554         if (set_bytes(i) > PAGE_SIZE - 48 &&
 555             !current->bio_list)
 556                 bch_btree_node_write(b, NULL);
 557 }
 558
 559 /*
 560  * Btree in memory cache - allocation/freeing
 561  * mca -> memory cache
 562  */
 563
 564 #define mca_reserve(c)  (((c->root && c->root->level)           \
 565                           ? c->root->level : 1) * 8 + 16)
 566 #define mca_can_free(c)                                         \
 567         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 568
 569 static void mca_data_free(struct btree *b)
 570 {
 571         BUG_ON(b->io_mutex.count != 1);
 572
 573         bch_btree_keys_free(&b->keys);
 574
 575         b->c->btree_cache_used--;
 576         list_move(&b->list, &b->c->btree_cache_freed);
 577 }
 578
 579 static void mca_bucket_free(struct btree *b)
 580 {
 581         BUG_ON(btree_node_dirty(b));
 582
 583         b->key.ptr[0] = 0;
 584         hlist_del_init_rcu(&b->hash);
 585         list_move(&b->list, &b->c->btree_cache_freeable);
 586 }
 587
 588 static unsigned int btree_order(struct bkey *k)
 589 {
 590         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 591 }
 592
 593 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 594 {
 595         if (!bch_btree_keys_alloc(&b->keys,
 596                                   max_t(unsigned int,
 597                                         ilog2(b->c->btree_pages),
 598                                         btree_order(k)),
 599                                   gfp)) {
 600                 b->c->btree_cache_used++;
 601                 list_move(&b->list, &b->c->btree_cache);
 602         } else {
 603                 list_move(&b->list, &b->c->btree_cache_freed);
 604         }
 605 }
 606
 607 static struct btree *mca_bucket_alloc(struct cache_set *c,
 608                                       struct bkey *k, gfp_t gfp)
 609 {
 610         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 611
 612         if (!b)
 613                 return NULL;
 614
 615         init_rwsem(&b->lock);
 616         lockdep_set_novalidate_class(&b->lock);
 617         mutex_init(&b->write_lock);
 618         lockdep_set_novalidate_class(&b->write_lock);
 619         INIT_LIST_HEAD(&b->list);
 620         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 621         b->c = c;
 622         sema_init(&b->io_mutex, 1);
 623
 624         mca_data_alloc(b, k, gfp);
 625         return b;
 626 }
 627
 628 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
 629 {
 630         struct closure cl;
 631
 632         closure_init_stack(&cl);
 633         lockdep_assert_held(&b->c->bucket_lock);
 634
 635         if (!down_write_trylock(&b->lock))
 636                 return -ENOMEM;
 637
 638         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 639
 640         if (b->keys.page_order < min_order)
 641                 goto out_unlock;
 642
 643         if (!flush) {
 644                 if (btree_node_dirty(b))
 645                         goto out_unlock;
 646
 647                 if (down_trylock(&b->io_mutex))
 648                         goto out_unlock;
 649                 up(&b->io_mutex);
 650         }
 651
 652 retry:
 653         /*
 654          * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
 655          * __bch_btree_node_write(). To avoid an extra flush, acquire
 656          * b->write_lock before checking BTREE_NODE_dirty bit.
 657          */
 658         mutex_lock(&b->write_lock);
 659         /*
 660          * If this btree node is selected in btree_flush_write() by journal
 661          * code, delay and retry until the node is flushed by journal code
 662          * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
 663          */
 664         if (btree_node_journal_flush(b)) {
 665                 pr_debug("bnode %p is flushing by journal, retry", b);
 666                 mutex_unlock(&b->write_lock);
 667                 udelay(1);
 668                 goto retry;
 669         }
 670
 671         if (btree_node_dirty(b))
 672                 __bch_btree_node_write(b, &cl);
 673         mutex_unlock(&b->write_lock);
 674
 675         closure_sync(&cl);
 676
 677         /* wait for any in flight btree write */
 678         down(&b->io_mutex);
 679         up(&b->io_mutex);
 680
 681         return 0;
 682 out_unlock:
 683         rw_unlock(true, b);
 684         return -ENOMEM;
 685 }
 686
 687 static unsigned long bch_mca_scan(struct shrinker *shrink,
 688                                   struct shrink_control *sc)
 689 {
 690         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 691         struct btree *b, *t;
 692         unsigned long i, nr = sc->nr_to_scan;
 693         unsigned long freed = 0;
 694         unsigned int btree_cache_used;
 695
 696         if (c->shrinker_disabled)
 697                 return SHRINK_STOP;
 698
 699         if (c->btree_cache_alloc_lock)
 700                 return SHRINK_STOP;
 701
 702         /* Return -1 if we can't do anything right now */
 703         if (sc->gfp_mask & __GFP_IO)
 704                 mutex_lock(&c->bucket_lock);
 705         else if (!mutex_trylock(&c->bucket_lock))
 706                 return -1;
 707
 708         /*
 709          * It's _really_ critical that we don't free too many btree nodes - we
 710          * have to always leave ourselves a reserve. The reserve is how we
 711          * guarantee that allocating memory for a new btree node can always
 712          * succeed, so that inserting keys into the btree can always succeed and
 713          * IO can always make forward progress:
 714          */
 715         nr /= c->btree_pages;
 716         if (nr == 0)
 717                 nr = 1;
 718         nr = min_t(unsigned long, nr, mca_can_free(c));
 719
 720         i = 0;
 721         btree_cache_used = c->btree_cache_used;
 722         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 723                 if (nr <= 0)
 724                         goto out;
 725
 726                 if (++i > 3 &&
 727                     !mca_reap(b, 0, false)) {
 728                         mca_data_free(b);
 729                         rw_unlock(true, b);
 730                         freed++;
 731                 }
 732                 nr--;
 733         }
 734
 735         for (;  (nr--) && i < btree_cache_used; i++) {
 736                 if (list_empty(&c->btree_cache))
 737                         goto out;
 738
 739                 b = list_first_entry(&c->btree_cache, struct btree, list);
 740                 list_rotate_left(&c->btree_cache);
 741
 742                 if (!b->accessed &&
 743                     !mca_reap(b, 0, false)) {
 744                         mca_bucket_free(b);
 745                         mca_data_free(b);
 746                         rw_unlock(true, b);
 747                         freed++;
 748                 } else
 749                         b->accessed = 0;
 750         }
 751 out:
 752         mutex_unlock(&c->bucket_lock);
 753         return freed * c->btree_pages;
 754 }
 755
 756 static unsigned long bch_mca_count(struct shrinker *shrink,
 757                                    struct shrink_control *sc)
 758 {
 759         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 760
 761         if (c->shrinker_disabled)
 762                 return 0;
 763
 764         if (c->btree_cache_alloc_lock)
 765                 return 0;
 766
 767         return mca_can_free(c) * c->btree_pages;
 768 }
 769
 770 void bch_btree_cache_free(struct cache_set *c)
 771 {
 772         struct btree *b;
 773         struct closure cl;
 774
 775         closure_init_stack(&cl);
 776
 777         if (c->shrink.list.next)
 778                 unregister_shrinker(&c->shrink);
 779
 780         mutex_lock(&c->bucket_lock);
 781
 782 #ifdef CONFIG_BCACHE_DEBUG
 783         if (c->verify_data)
 784                 list_move(&c->verify_data->list, &c->btree_cache);
 785
 786         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 787 #endif
 788
 789         list_splice(&c->btree_cache_freeable,
 790                     &c->btree_cache);
 791
 792         while (!list_empty(&c->btree_cache)) {
 793                 b = list_first_entry(&c->btree_cache, struct btree, list);
 794
 795                 /*
 796                  * This function is called by cache_set_free(), no I/O
 797                  * request on cache now, it is unnecessary to acquire
 798                  * b->write_lock before clearing BTREE_NODE_dirty anymore.
 799                  */
 800                 if (btree_node_dirty(b)) {
 801                         btree_complete_write(b, btree_current_write(b));
 802                         clear_bit(BTREE_NODE_dirty, &b->flags);
 803                 }
 804                 mca_data_free(b);
 805         }
 806
 807         while (!list_empty(&c->btree_cache_freed)) {
 808                 b = list_first_entry(&c->btree_cache_freed,
 809                                      struct btree, list);
 810                 list_del(&b->list);
 811                 cancel_delayed_work_sync(&b->work);
 812                 kfree(b);
 813         }
 814
 815         mutex_unlock(&c->bucket_lock);
 816 }
 817
 818 int bch_btree_cache_alloc(struct cache_set *c)
 819 {
 820         unsigned int i;
 821
 822         for (i = 0; i < mca_reserve(c); i++)
 823                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 824                         return -ENOMEM;
 825
 826         list_splice_init(&c->btree_cache,
 827                          &c->btree_cache_freeable);
 828
 829 #ifdef CONFIG_BCACHE_DEBUG
 830         mutex_init(&c->verify_lock);
 831
 832         c->verify_ondisk = (void *)
 833                 __get_free_pages(GFP_KERNEL|__GFP_COMP, ilog2(bucket_pages(c)));
 834
 835         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 836
 837         if (c->verify_data &&
 838             c->verify_data->keys.set->data)
 839                 list_del_init(&c->verify_data->list);
 840         else
 841                 c->verify_data = NULL;
 842 #endif
 843
 844         c->shrink.count_objects = bch_mca_count;
 845         c->shrink.scan_objects = bch_mca_scan;
 846         c->shrink.seeks = 4;
 847         c->shrink.batch = c->btree_pages * 2;
 848
 849         if (register_shrinker(&c->shrink))
 850                 pr_warn("bcache: %s: could not register shrinker",
 851                                 __func__);
 852
 853         return 0;
 854 }
 855
 856 /* Btree in memory cache - hash table */
 857
 858 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 859 {
 860         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 861 }
 862
 863 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 864 {
 865         struct btree *b;
 866
 867         rcu_read_lock();
 868         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 869                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 870                         goto out;
 871         b = NULL;
 872 out:
 873         rcu_read_unlock();
 874         return b;
 875 }
 876
 877 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 878 {
 879         spin_lock(&c->btree_cannibalize_lock);
 880         if (likely(c->btree_cache_alloc_lock == NULL)) {
 881                 c->btree_cache_alloc_lock = current;
 882         } else if (c->btree_cache_alloc_lock != current) {
 883                 if (op)
 884                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 885                                         TASK_UNINTERRUPTIBLE);
 886                 spin_unlock(&c->btree_cannibalize_lock);
 887                 return -EINTR;
 888         }
 889         spin_unlock(&c->btree_cannibalize_lock);
 890
 891         return 0;
 892 }
 893
 894 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 895                                      struct bkey *k)
 896 {
 897         struct btree *b;
 898
 899         trace_bcache_btree_cache_cannibalize(c);
 900
 901         if (mca_cannibalize_lock(c, op))
 902                 return ERR_PTR(-EINTR);
 903
 904         list_for_each_entry_reverse(b, &c->btree_cache, list)
 905                 if (!mca_reap(b, btree_order(k), false))
 906                         return b;
 907
 908         list_for_each_entry_reverse(b, &c->btree_cache, list)
 909                 if (!mca_reap(b, btree_order(k), true))
 910                         return b;
 911
 912         WARN(1, "btree cache cannibalize failed\n");
 913         return ERR_PTR(-ENOMEM);
 914 }
 915
 916 /*
 917  * We can only have one thread cannibalizing other cached btree nodes at a time,
 918  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 919  * cannibalize_bucket() will take. This means every time we unlock the root of
 920  * the btree, we need to release this lock if we have it held.
 921  */
 922 static void bch_cannibalize_unlock(struct cache_set *c)
 923 {
 924         spin_lock(&c->btree_cannibalize_lock);
 925         if (c->btree_cache_alloc_lock == current) {
 926                 c->btree_cache_alloc_lock = NULL;
 927                 wake_up(&c->btree_cache_wait);
 928         }
 929         spin_unlock(&c->btree_cannibalize_lock);
 930 }
 931
 932 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 933                                struct bkey *k, int level)
 934 {
 935         struct btree *b;
 936
 937         BUG_ON(current->bio_list);
 938
 939         lockdep_assert_held(&c->bucket_lock);
 940
 941         if (mca_find(c, k))
 942                 return NULL;
 943
 944         /* btree_free() doesn't free memory; it sticks the node on the end of
 945          * the list. Check if there's any freed nodes there:
 946          */
 947         list_for_each_entry(b, &c->btree_cache_freeable, list)
 948                 if (!mca_reap(b, btree_order(k), false))
 949                         goto out;
 950
 951         /* We never free struct btree itself, just the memory that holds the on
 952          * disk node. Check the freed list before allocating a new one:
 953          */
 954         list_for_each_entry(b, &c->btree_cache_freed, list)
 955                 if (!mca_reap(b, 0, false)) {
 956                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 957                         if (!b->keys.set[0].data)
 958                                 goto err;
 959                         else
 960                                 goto out;
 961                 }
 962
 963         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 964         if (!b)
 965                 goto err;
 966
 967         BUG_ON(!down_write_trylock(&b->lock));
 968         if (!b->keys.set->data)
 969                 goto err;
 970 out:
 971         BUG_ON(b->io_mutex.count != 1);
 972
 973         bkey_copy(&b->key, k);
 974         list_move(&b->list, &c->btree_cache);
 975         hlist_del_init_rcu(&b->hash);
 976         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 977
 978         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 979         b->parent       = (void *) ~0UL;
 980         b->flags        = 0;
 981         b->written      = 0;
 982         b->level        = level;
 983
 984         if (!b->level)
 985                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 986                                     &b->c->expensive_debug_checks);
 987         else
 988                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 989                                     &b->c->expensive_debug_checks);
 990
 991         return b;
 992 err:
 993         if (b)
 994                 rw_unlock(true, b);
 995
 996         b = mca_cannibalize(c, op, k);
 997         if (!IS_ERR(b))
 998                 goto out;
 999
1000         return b;
1001 }
1002
1003 /*
1004  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
1005  * in from disk if necessary.
1006  *
1007  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
1008  *
1009  * The btree node will have either a read or a write lock held, depending on
1010  * level and op->lock.
1011  */
1012 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1013                                  struct bkey *k, int level, bool write,
1014                                  struct btree *parent)
1015 {
1016         int i = 0;
1017         struct btree *b;
1018
1019         BUG_ON(level < 0);
1020 retry:
1021         b = mca_find(c, k);
1022
1023         if (!b) {
1024                 if (current->bio_list)
1025                         return ERR_PTR(-EAGAIN);
1026
1027                 mutex_lock(&c->bucket_lock);
1028                 b = mca_alloc(c, op, k, level);
1029                 mutex_unlock(&c->bucket_lock);
1030
1031                 if (!b)
1032                         goto retry;
1033                 if (IS_ERR(b))
1034                         return b;
1035
1036                 bch_btree_node_read(b);
1037
1038                 if (!write)
1039                         downgrade_write(&b->lock);
1040         } else {
1041                 rw_lock(write, b, level);
1042                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1043                         rw_unlock(write, b);
1044                         goto retry;
1045                 }
1046                 BUG_ON(b->level != level);
1047         }
1048
1049         if (btree_node_io_error(b)) {
1050                 rw_unlock(write, b);
1051                 return ERR_PTR(-EIO);
1052         }
1053
1054         BUG_ON(!b->written);
1055
1056         b->parent = parent;
1057         b->accessed = 1;
1058
1059         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1060                 prefetch(b->keys.set[i].tree);
1061                 prefetch(b->keys.set[i].data);
1062         }
1063
1064         for (; i <= b->keys.nsets; i++)
1065                 prefetch(b->keys.set[i].data);
1066
1067         return b;
1068 }
1069
1070 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1071 {
1072         struct btree *b;
1073
1074         mutex_lock(&parent->c->bucket_lock);
1075         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1076         mutex_unlock(&parent->c->bucket_lock);
1077
1078         if (!IS_ERR_OR_NULL(b)) {
1079                 b->parent = parent;
1080                 bch_btree_node_read(b);
1081                 rw_unlock(true, b);
1082         }
1083 }
1084
1085 /* Btree alloc */
1086
1087 static void btree_node_free(struct btree *b)
1088 {
1089         trace_bcache_btree_node_free(b);
1090
1091         BUG_ON(b == b->c->root);
1092
1093 retry:
1094         mutex_lock(&b->write_lock);
1095         /*
1096          * If the btree node is selected and flushing in btree_flush_write(),
1097          * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1098          * then it is safe to free the btree node here. Otherwise this btree
1099          * node will be in race condition.
1100          */
1101         if (btree_node_journal_flush(b)) {
1102                 mutex_unlock(&b->write_lock);
1103                 pr_debug("bnode %p journal_flush set, retry", b);
1104                 udelay(1);
1105                 goto retry;
1106         }
1107
1108         if (btree_node_dirty(b)) {
1109                 btree_complete_write(b, btree_current_write(b));
1110                 clear_bit(BTREE_NODE_dirty, &b->flags);
1111         }
1112
1113         mutex_unlock(&b->write_lock);
1114
1115         cancel_delayed_work(&b->work);
1116
1117         mutex_lock(&b->c->bucket_lock);
1118         bch_bucket_free(b->c, &b->key);
1119         mca_bucket_free(b);
1120         mutex_unlock(&b->c->bucket_lock);
1121 }
1122
1123 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1124                                      int level, bool wait,
1125                                      struct btree *parent)
1126 {
1127         BKEY_PADDED(key) k;
1128         struct btree *b;
1129
1130         mutex_lock(&c->bucket_lock);
1131 retry:
1132         /* return ERR_PTR(-EAGAIN) when it fails */
1133         b = ERR_PTR(-EAGAIN);
1134         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1135                 goto err;
1136
1137         bkey_put(c, &k.key);
1138         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1139
1140         b = mca_alloc(c, op, &k.key, level);
1141         if (IS_ERR(b))
1142                 goto err_free;
1143
1144         if (!b) {
1145                 cache_bug(c,
1146                         "Tried to allocate bucket that was in btree cache");
1147                 goto retry;
1148         }
1149
1150         b->accessed = 1;
1151         b->parent = parent;
1152         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1153
1154         mutex_unlock(&c->bucket_lock);
1155
1156         trace_bcache_btree_node_alloc(b);
1157         return b;
1158 err_free:
1159         bch_bucket_free(c, &k.key);
1160 err:
1161         mutex_unlock(&c->bucket_lock);
1162
1163         trace_bcache_btree_node_alloc_fail(c);
1164         return b;
1165 }
1166
1167 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1168                                           struct btree_op *op, int level,
1169                                           struct btree *parent)
1170 {
1171         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1172 }
1173
1174 static struct btree *btree_node_alloc_replacement(struct btree *b,
1175                                                   struct btree_op *op)
1176 {
1177         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1178
1179         if (!IS_ERR(n)) {
1180                 mutex_lock(&n->write_lock);
1181                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1182                 bkey_copy_key(&n->key, &b->key);
1183                 mutex_unlock(&n->write_lock);
1184         }
1185
1186         return n;
1187 }
1188
1189 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1190 {
1191         unsigned int i;
1192
1193         mutex_lock(&b->c->bucket_lock);
1194
1195         atomic_inc(&b->c->prio_blocked);
1196
1197         bkey_copy(k, &b->key);
1198         bkey_copy_key(k, &ZERO_KEY);
1199
1200         for (i = 0; i < KEY_PTRS(k); i++)
1201                 SET_PTR_GEN(k, i,
1202                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1203                                         PTR_BUCKET(b->c, &b->key, i)));
1204
1205         mutex_unlock(&b->c->bucket_lock);
1206 }
1207
1208 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1209 {
1210         struct cache_set *c = b->c;
1211         struct cache *ca;
1212         unsigned int i, reserve = (c->root->level - b->level) * 2 + 1;
1213
1214         mutex_lock(&c->bucket_lock);
1215
1216         for_each_cache(ca, c, i)
1217                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1218                         if (op)
1219                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1220                                                 TASK_UNINTERRUPTIBLE);
1221                         mutex_unlock(&c->bucket_lock);
1222                         return -EINTR;
1223                 }
1224
1225         mutex_unlock(&c->bucket_lock);
1226
1227         return mca_cannibalize_lock(b->c, op);
1228 }
1229
1230 /* Garbage collection */
1231
1232 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1233                                     struct bkey *k)
1234 {
1235         uint8_t stale = 0;
1236         unsigned int i;
1237         struct bucket *g;
1238
1239         /*
1240          * ptr_invalid() can't return true for the keys that mark btree nodes as
1241          * freed, but since ptr_bad() returns true we'll never actually use them
1242          * for anything and thus we don't want mark their pointers here
1243          */
1244         if (!bkey_cmp(k, &ZERO_KEY))
1245                 return stale;
1246
1247         for (i = 0; i < KEY_PTRS(k); i++) {
1248                 if (!ptr_available(c, k, i))
1249                         continue;
1250
1251                 g = PTR_BUCKET(c, k, i);
1252
1253                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1254                         g->last_gc = PTR_GEN(k, i);
1255
1256                 if (ptr_stale(c, k, i)) {
1257                         stale = max(stale, ptr_stale(c, k, i));
1258                         continue;
1259                 }
1260
1261                 cache_bug_on(GC_MARK(g) &&
1262                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1263                              c, "inconsistent ptrs: mark = %llu, level = %i",
1264                              GC_MARK(g), level);
1265
1266                 if (level)
1267                         SET_GC_MARK(g, GC_MARK_METADATA);
1268                 else if (KEY_DIRTY(k))
1269                         SET_GC_MARK(g, GC_MARK_DIRTY);
1270                 else if (!GC_MARK(g))
1271                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1272
1273                 /* guard against overflow */
1274                 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1275                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1276                                              MAX_GC_SECTORS_USED));
1277
1278                 BUG_ON(!GC_SECTORS_USED(g));
1279         }
1280
1281         return stale;
1282 }
1283
1284 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1285
1286 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1287 {
1288         unsigned int i;
1289
1290         for (i = 0; i < KEY_PTRS(k); i++)
1291                 if (ptr_available(c, k, i) &&
1292                     !ptr_stale(c, k, i)) {
1293                         struct bucket *b = PTR_BUCKET(c, k, i);
1294
1295                         b->gen = PTR_GEN(k, i);
1296
1297                         if (level && bkey_cmp(k, &ZERO_KEY))
1298                                 b->prio = BTREE_PRIO;
1299                         else if (!level && b->prio == BTREE_PRIO)
1300                                 b->prio = INITIAL_PRIO;
1301                 }
1302
1303         __bch_btree_mark_key(c, level, k);
1304 }
1305
1306 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1307 {
1308         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1309 }
1310
1311 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1312 {
1313         uint8_t stale = 0;
1314         unsigned int keys = 0, good_keys = 0;
1315         struct bkey *k;
1316         struct btree_iter iter;
1317         struct bset_tree *t;
1318
1319         gc->nodes++;
1320
1321         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1322                 stale = max(stale, btree_mark_key(b, k));
1323                 keys++;
1324
1325                 if (bch_ptr_bad(&b->keys, k))
1326                         continue;
1327
1328                 gc->key_bytes += bkey_u64s(k);
1329                 gc->nkeys++;
1330                 good_keys++;
1331
1332                 gc->data += KEY_SIZE(k);
1333         }
1334
1335         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1336                 btree_bug_on(t->size &&
1337                              bset_written(&b->keys, t) &&
1338                              bkey_cmp(&b->key, &t->end) < 0,
1339                              b, "found short btree key in gc");
1340
1341         if (b->c->gc_always_rewrite)
1342                 return true;
1343
1344         if (stale > 10)
1345                 return true;
1346
1347         if ((keys - good_keys) * 2 > keys)
1348                 return true;
1349
1350         return false;
1351 }
1352
1353 #define GC_MERGE_NODES  4U
1354
1355 struct gc_merge_info {
1356         struct btree    *b;
1357         unsigned int    keys;
1358 };
1359
1360 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1361                                  struct keylist *insert_keys,
1362                                  atomic_t *journal_ref,
1363                                  struct bkey *replace_key);
1364
1365 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1366                              struct gc_stat *gc, struct gc_merge_info *r)
1367 {
1368         unsigned int i, nodes = 0, keys = 0, blocks;
1369         struct btree *new_nodes[GC_MERGE_NODES];
1370         struct keylist keylist;
1371         struct closure cl;
1372         struct bkey *k;
1373
1374         bch_keylist_init(&keylist);
1375
1376         if (btree_check_reserve(b, NULL))
1377                 return 0;
1378
1379         memset(new_nodes, 0, sizeof(new_nodes));
1380         closure_init_stack(&cl);
1381
1382         while (nodes < GC_MERGE_NODES && !IS_ERR(r[nodes].b))
1383                 keys += r[nodes++].keys;
1384
1385         blocks = btree_default_blocks(b->c) * 2 / 3;
1386
1387         if (nodes < 2 ||
1388             __set_blocks(b->keys.set[0].data, keys,
1389                          block_bytes(b->c)) > blocks * (nodes - 1))
1390                 return 0;
1391
1392         for (i = 0; i < nodes; i++) {
1393                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1394                 if (IS_ERR(new_nodes[i]))
1395                         goto out_nocoalesce;
1396         }
1397
1398         /*
1399          * We have to check the reserve here, after we've allocated our new
1400          * nodes, to make sure the insert below will succeed - we also check
1401          * before as an optimization to potentially avoid a bunch of expensive
1402          * allocs/sorts
1403          */
1404         if (btree_check_reserve(b, NULL))
1405                 goto out_nocoalesce;
1406
1407         for (i = 0; i < nodes; i++)
1408                 mutex_lock(&new_nodes[i]->write_lock);
1409
1410         for (i = nodes - 1; i > 0; --i) {
1411                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1412                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1413                 struct bkey *k, *last = NULL;
1414
1415                 keys = 0;
1416
1417                 if (i > 1) {
1418                         for (k = n2->start;
1419                              k < bset_bkey_last(n2);
1420                              k = bkey_next(k)) {
1421                                 if (__set_blocks(n1, n1->keys + keys +
1422                                                  bkey_u64s(k),
1423                                                  block_bytes(b->c)) > blocks)
1424                                         break;
1425
1426                                 last = k;
1427                                 keys += bkey_u64s(k);
1428                         }
1429                 } else {
1430                         /*
1431                          * Last node we're not getting rid of - we're getting
1432                          * rid of the node at r[0]. Have to try and fit all of
1433                          * the remaining keys into this node; we can't ensure
1434                          * they will always fit due to rounding and variable
1435                          * length keys (shouldn't be possible in practice,
1436                          * though)
1437                          */
1438                         if (__set_blocks(n1, n1->keys + n2->keys,
1439                                          block_bytes(b->c)) >
1440                             btree_blocks(new_nodes[i]))
1441                                 goto out_unlock_nocoalesce;
1442
1443                         keys = n2->keys;
1444                         /* Take the key of the node we're getting rid of */
1445                         last = &r->b->key;
1446                 }
1447
1448                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1449                        btree_blocks(new_nodes[i]));
1450
1451                 if (last)
1452                         bkey_copy_key(&new_nodes[i]->key, last);
1453
1454                 memcpy(bset_bkey_last(n1),
1455                        n2->start,
1456                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1457
1458                 n1->keys += keys;
1459                 r[i].keys = n1->keys;
1460
1461                 memmove(n2->start,
1462                         bset_bkey_idx(n2, keys),
1463                         (void *) bset_bkey_last(n2) -
1464                         (void *) bset_bkey_idx(n2, keys));
1465
1466                 n2->keys -= keys;
1467
1468                 if (__bch_keylist_realloc(&keylist,
1469                                           bkey_u64s(&new_nodes[i]->key)))
1470                         goto out_unlock_nocoalesce;
1471
1472                 bch_btree_node_write(new_nodes[i], &cl);
1473                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1474         }
1475
1476         for (i = 0; i < nodes; i++)
1477                 mutex_unlock(&new_nodes[i]->write_lock);
1478
1479         closure_sync(&cl);
1480
1481         /* We emptied out this node */
1482         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1483         btree_node_free(new_nodes[0]);
1484         rw_unlock(true, new_nodes[0]);
1485         new_nodes[0] = NULL;
1486
1487         for (i = 0; i < nodes; i++) {
1488                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1489                         goto out_nocoalesce;
1490
1491                 make_btree_freeing_key(r[i].b, keylist.top);
1492                 bch_keylist_push(&keylist);
1493         }
1494
1495         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1496         BUG_ON(!bch_keylist_empty(&keylist));
1497
1498         for (i = 0; i < nodes; i++) {
1499                 btree_node_free(r[i].b);
1500                 rw_unlock(true, r[i].b);
1501
1502                 r[i].b = new_nodes[i];
1503         }
1504
1505         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1506         r[nodes - 1].b = ERR_PTR(-EINTR);
1507
1508         trace_bcache_btree_gc_coalesce(nodes);
1509         gc->nodes--;
1510
1511         bch_keylist_free(&keylist);
1512
1513         /* Invalidated our iterator */
1514         return -EINTR;
1515
1516 out_unlock_nocoalesce:
1517         for (i = 0; i < nodes; i++)
1518                 mutex_unlock(&new_nodes[i]->write_lock);
1519
1520 out_nocoalesce:
1521         closure_sync(&cl);
1522         bch_keylist_free(&keylist);
1523
1524         while ((k = bch_keylist_pop(&keylist)))
1525                 if (!bkey_cmp(k, &ZERO_KEY))
1526                         atomic_dec(&b->c->prio_blocked);
1527
1528         for (i = 0; i < nodes; i++)
1529                 if (!IS_ERR(new_nodes[i])) {
1530                         btree_node_free(new_nodes[i]);
1531                         rw_unlock(true, new_nodes[i]);
1532                 }
1533         return 0;
1534 }
1535
1536 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1537                                  struct btree *replace)
1538 {
1539         struct keylist keys;
1540         struct btree *n;
1541
1542         if (btree_check_reserve(b, NULL))
1543                 return 0;
1544
1545         n = btree_node_alloc_replacement(replace, NULL);
1546
1547         /* recheck reserve after allocating replacement node */
1548         if (btree_check_reserve(b, NULL)) {
1549                 btree_node_free(n);
1550                 rw_unlock(true, n);
1551                 return 0;
1552         }
1553
1554         bch_btree_node_write_sync(n);
1555
1556         bch_keylist_init(&keys);
1557         bch_keylist_add(&keys, &n->key);
1558
1559         make_btree_freeing_key(replace, keys.top);
1560         bch_keylist_push(&keys);
1561
1562         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1563         BUG_ON(!bch_keylist_empty(&keys));
1564
1565         btree_node_free(replace);
1566         rw_unlock(true, n);
1567
1568         /* Invalidated our iterator */
1569         return -EINTR;
1570 }
1571
1572 static unsigned int btree_gc_count_keys(struct btree *b)
1573 {
1574         struct bkey *k;
1575         struct btree_iter iter;
1576         unsigned int ret = 0;
1577
1578         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1579                 ret += bkey_u64s(k);
1580
1581         return ret;
1582 }
1583
1584 static size_t btree_gc_min_nodes(struct cache_set *c)
1585 {
1586         size_t min_nodes;
1587
1588         /*
1589          * Since incremental GC would stop 100ms when front
1590          * side I/O comes, so when there are many btree nodes,
1591          * if GC only processes constant (100) nodes each time,
1592          * GC would last a long time, and the front side I/Os
1593          * would run out of the buckets (since no new bucket
1594          * can be allocated during GC), and be blocked again.
1595          * So GC should not process constant nodes, but varied
1596          * nodes according to the number of btree nodes, which
1597          * realized by dividing GC into constant(100) times,
1598          * so when there are many btree nodes, GC can process
1599          * more nodes each time, otherwise, GC will process less
1600          * nodes each time (but no less than MIN_GC_NODES)
1601          */
1602         min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1603         if (min_nodes < MIN_GC_NODES)
1604                 min_nodes = MIN_GC_NODES;
1605
1606         return min_nodes;
1607 }
1608
1609
1610 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1611                             struct closure *writes, struct gc_stat *gc)
1612 {
1613         int ret = 0;
1614         bool should_rewrite;
1615         struct bkey *k;
1616         struct btree_iter iter;
1617         struct gc_merge_info r[GC_MERGE_NODES];
1618         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1619
1620         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1621
1622         for (i = r; i < r + ARRAY_SIZE(r); i++)
1623                 i->b = ERR_PTR(-EINTR);
1624
1625         while (1) {
1626                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1627                 if (k) {
1628                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1629                                                   true, b);
1630                         if (IS_ERR(r->b)) {
1631                                 ret = PTR_ERR(r->b);
1632                                 break;
1633                         }
1634
1635                         r->keys = btree_gc_count_keys(r->b);
1636
1637                         ret = btree_gc_coalesce(b, op, gc, r);
1638                         if (ret)
1639                                 break;
1640                 }
1641
1642                 if (!last->b)
1643                         break;
1644
1645                 if (!IS_ERR(last->b)) {
1646                         should_rewrite = btree_gc_mark_node(last->b, gc);
1647                         if (should_rewrite) {
1648                                 ret = btree_gc_rewrite_node(b, op, last->b);
1649                                 if (ret)
1650                                         break;
1651                         }
1652
1653                         if (last->b->level) {
1654                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1655                                 if (ret)
1656                                         break;
1657                         }
1658
1659                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1660
1661                         /*
1662                          * Must flush leaf nodes before gc ends, since replace
1663                          * operations aren't journalled
1664                          */
1665                         mutex_lock(&last->b->write_lock);
1666                         if (btree_node_dirty(last->b))
1667                                 bch_btree_node_write(last->b, writes);
1668                         mutex_unlock(&last->b->write_lock);
1669                         rw_unlock(true, last->b);
1670                 }
1671
1672                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1673                 r->b = NULL;
1674
1675                 if (atomic_read(&b->c->search_inflight) &&
1676                     gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1677                         gc->nodes_pre =  gc->nodes;
1678                         ret = -EAGAIN;
1679                         break;
1680                 }
1681
1682                 if (need_resched()) {
1683                         ret = -EAGAIN;
1684                         break;
1685                 }
1686         }
1687
1688         for (i = r; i < r + ARRAY_SIZE(r); i++)
1689                 if (!IS_ERR_OR_NULL(i->b)) {
1690                         mutex_lock(&i->b->write_lock);
1691                         if (btree_node_dirty(i->b))
1692                                 bch_btree_node_write(i->b, writes);
1693                         mutex_unlock(&i->b->write_lock);
1694                         rw_unlock(true, i->b);
1695                 }
1696
1697         return ret;
1698 }
1699
1700 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1701                              struct closure *writes, struct gc_stat *gc)
1702 {
1703         struct btree *n = NULL;
1704         int ret = 0;
1705         bool should_rewrite;
1706
1707         should_rewrite = btree_gc_mark_node(b, gc);
1708         if (should_rewrite) {
1709                 n = btree_node_alloc_replacement(b, NULL);
1710
1711                 if (!IS_ERR(n)) {
1712                         bch_btree_node_write_sync(n);
1713
1714                         bch_btree_set_root(n);
1715                         btree_node_free(b);
1716                         rw_unlock(true, n);
1717
1718                         return -EINTR;
1719                 }
1720         }
1721
1722         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1723
1724         if (b->level) {
1725                 ret = btree_gc_recurse(b, op, writes, gc);
1726                 if (ret)
1727                         return ret;
1728         }
1729
1730         bkey_copy_key(&b->c->gc_done, &b->key);
1731
1732         return ret;
1733 }
1734
1735 static void btree_gc_start(struct cache_set *c)
1736 {
1737         struct cache *ca;
1738         struct bucket *b;
1739         unsigned int i;
1740
1741         if (!c->gc_mark_valid)
1742                 return;
1743
1744         mutex_lock(&c->bucket_lock);
1745
1746         c->gc_mark_valid = 0;
1747         c->gc_done = ZERO_KEY;
1748
1749         for_each_cache(ca, c, i)
1750                 for_each_bucket(b, ca) {
1751                         b->last_gc = b->gen;
1752                         if (!atomic_read(&b->pin)) {
1753                                 SET_GC_MARK(b, 0);
1754                                 SET_GC_SECTORS_USED(b, 0);
1755                         }
1756                 }
1757
1758         mutex_unlock(&c->bucket_lock);
1759 }
1760
1761 static void bch_btree_gc_finish(struct cache_set *c)
1762 {
1763         struct bucket *b;
1764         struct cache *ca;
1765         unsigned int i;
1766
1767         mutex_lock(&c->bucket_lock);
1768
1769         set_gc_sectors(c);
1770         c->gc_mark_valid = 1;
1771         c->need_gc      = 0;
1772
1773         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1774                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1775                             GC_MARK_METADATA);
1776
1777         /* don't reclaim buckets to which writeback keys point */
1778         rcu_read_lock();
1779         for (i = 0; i < c->devices_max_used; i++) {
1780                 struct bcache_device *d = c->devices[i];
1781                 struct cached_dev *dc;
1782                 struct keybuf_key *w, *n;
1783                 unsigned int j;
1784
1785                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1786                         continue;
1787                 dc = container_of(d, struct cached_dev, disk);
1788
1789                 spin_lock(&dc->writeback_keys.lock);
1790                 rbtree_postorder_for_each_entry_safe(w, n,
1791                                         &dc->writeback_keys.keys, node)
1792                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1793                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1794                                             GC_MARK_DIRTY);
1795                 spin_unlock(&dc->writeback_keys.lock);
1796         }
1797         rcu_read_unlock();
1798
1799         c->avail_nbuckets = 0;
1800         for_each_cache(ca, c, i) {
1801                 uint64_t *i;
1802
1803                 ca->invalidate_needs_gc = 0;
1804
1805                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1806                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1807
1808                 for (i = ca->prio_buckets;
1809                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1810                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1811
1812                 for_each_bucket(b, ca) {
1813                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1814
1815                         if (atomic_read(&b->pin))
1816                                 continue;
1817
1818                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1819
1820                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1821                                 c->avail_nbuckets++;
1822                 }
1823         }
1824
1825         mutex_unlock(&c->bucket_lock);
1826 }
1827
1828 static void bch_btree_gc(struct cache_set *c)
1829 {
1830         int ret;
1831         struct gc_stat stats;
1832         struct closure writes;
1833         struct btree_op op;
1834         uint64_t start_time = local_clock();
1835
1836         trace_bcache_gc_start(c);
1837
1838         memset(&stats, 0, sizeof(struct gc_stat));
1839         closure_init_stack(&writes);
1840         bch_btree_op_init(&op, SHRT_MAX);
1841
1842         btree_gc_start(c);
1843
1844         /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1845         do {
1846                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1847                 closure_sync(&writes);
1848                 cond_resched();
1849
1850                 if (ret == -EAGAIN)
1851                         schedule_timeout_interruptible(msecs_to_jiffies
1852                                                        (GC_SLEEP_MS));
1853                 else if (ret)
1854                         pr_warn("gc failed!");
1855         } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1856
1857         bch_btree_gc_finish(c);
1858         wake_up_allocators(c);
1859
1860         bch_time_stats_update(&c->btree_gc_time, start_time);
1861
1862         stats.key_bytes *= sizeof(uint64_t);
1863         stats.data      <<= 9;
1864         bch_update_bucket_in_use(c, &stats);
1865         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1866
1867         trace_bcache_gc_end(c);
1868
1869         bch_moving_gc(c);
1870 }
1871
1872 static bool gc_should_run(struct cache_set *c)
1873 {
1874         struct cache *ca;
1875         unsigned int i;
1876
1877         for_each_cache(ca, c, i)
1878                 if (ca->invalidate_needs_gc)
1879                         return true;
1880
1881         if (atomic_read(&c->sectors_to_gc) < 0)
1882                 return true;
1883
1884         return false;
1885 }
1886
1887 static int bch_gc_thread(void *arg)
1888 {
1889         struct cache_set *c = arg;
1890
1891         while (1) {
1892                 wait_event_interruptible(c->gc_wait,
1893                            kthread_should_stop() ||
1894                            test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1895                            gc_should_run(c));
1896
1897                 if (kthread_should_stop() ||
1898                     test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1899                         break;
1900
1901                 set_gc_sectors(c);
1902                 bch_btree_gc(c);
1903         }
1904
1905         wait_for_kthread_stop();
1906         return 0;
1907 }
1908
1909 int bch_gc_thread_start(struct cache_set *c)
1910 {
1911         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1912         return PTR_ERR_OR_ZERO(c->gc_thread);
1913 }
1914
1915 /* Initial partial gc */
1916
1917 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1918 {
1919         int ret = 0;
1920         struct bkey *k, *p = NULL;
1921         struct btree_iter iter;
1922
1923         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1924                 bch_initial_mark_key(b->c, b->level, k);
1925
1926         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1927
1928         if (b->level) {
1929                 bch_btree_iter_init(&b->keys, &iter, NULL);
1930
1931                 do {
1932                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1933                                                        bch_ptr_bad);
1934                         if (k) {
1935                                 btree_node_prefetch(b, k);
1936                                 /*
1937                                  * initiallize c->gc_stats.nodes
1938                                  * for incremental GC
1939                                  */
1940                                 b->c->gc_stats.nodes++;
1941                         }
1942
1943                         if (p)
1944                                 ret = btree(check_recurse, p, b, op);
1945
1946                         p = k;
1947                 } while (p && !ret);
1948         }
1949
1950         return ret;
1951 }
1952
1953 int bch_btree_check(struct cache_set *c)
1954 {
1955         struct btree_op op;
1956
1957         bch_btree_op_init(&op, SHRT_MAX);
1958
1959         return btree_root(check_recurse, c, &op);
1960 }
1961
1962 void bch_initial_gc_finish(struct cache_set *c)
1963 {
1964         struct cache *ca;
1965         struct bucket *b;
1966         unsigned int i;
1967
1968         bch_btree_gc_finish(c);
1969
1970         mutex_lock(&c->bucket_lock);
1971
1972         /*
1973          * We need to put some unused buckets directly on the prio freelist in
1974          * order to get the allocator thread started - it needs freed buckets in
1975          * order to rewrite the prios and gens, and it needs to rewrite prios
1976          * and gens in order to free buckets.
1977          *
1978          * This is only safe for buckets that have no live data in them, which
1979          * there should always be some of.
1980          */
1981         for_each_cache(ca, c, i) {
1982                 for_each_bucket(b, ca) {
1983                         if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1984                             fifo_full(&ca->free[RESERVE_BTREE]))
1985                                 break;
1986
1987                         if (bch_can_invalidate_bucket(ca, b) &&
1988                             !GC_MARK(b)) {
1989                                 __bch_invalidate_one_bucket(ca, b);
1990                                 if (!fifo_push(&ca->free[RESERVE_PRIO],
1991                                    b - ca->buckets))
1992                                         fifo_push(&ca->free[RESERVE_BTREE],
1993                                                   b - ca->buckets);
1994                         }
1995                 }
1996         }
1997
1998         mutex_unlock(&c->bucket_lock);
1999 }
2000
2001 /* Btree insertion */
2002
2003 static bool btree_insert_key(struct btree *b, struct bkey *k,
2004                              struct bkey *replace_key)
2005 {
2006         unsigned int status;
2007
2008         BUG_ON(bkey_cmp(k, &b->key) > 0);
2009
2010         status = bch_btree_insert_key(&b->keys, k, replace_key);
2011         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2012                 bch_check_keys(&b->keys, "%u for %s", status,
2013                                replace_key ? "replace" : "insert");
2014
2015                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2016                                               status);
2017                 return true;
2018         } else
2019                 return false;
2020 }
2021
2022 static size_t insert_u64s_remaining(struct btree *b)
2023 {
2024         long ret = bch_btree_keys_u64s_remaining(&b->keys);
2025
2026         /*
2027          * Might land in the middle of an existing extent and have to split it
2028          */
2029         if (b->keys.ops->is_extents)
2030                 ret -= KEY_MAX_U64S;
2031
2032         return max(ret, 0L);
2033 }
2034
2035 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2036                                   struct keylist *insert_keys,
2037                                   struct bkey *replace_key)
2038 {
2039         bool ret = false;
2040         int oldsize = bch_count_data(&b->keys);
2041
2042         while (!bch_keylist_empty(insert_keys)) {
2043                 struct bkey *k = insert_keys->keys;
2044
2045                 if (bkey_u64s(k) > insert_u64s_remaining(b))
2046                         break;
2047
2048                 if (bkey_cmp(k, &b->key) <= 0) {
2049                         if (!b->level)
2050                                 bkey_put(b->c, k);
2051
2052                         ret |= btree_insert_key(b, k, replace_key);
2053                         bch_keylist_pop_front(insert_keys);
2054                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2055                         BKEY_PADDED(key) temp;
2056                         bkey_copy(&temp.key, insert_keys->keys);
2057
2058                         bch_cut_back(&b->key, &temp.key);
2059                         bch_cut_front(&b->key, insert_keys->keys);
2060
2061                         ret |= btree_insert_key(b, &temp.key, replace_key);
2062                         break;
2063                 } else {
2064                         break;
2065                 }
2066         }
2067
2068         if (!ret)
2069                 op->insert_collision = true;
2070
2071         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2072
2073         BUG_ON(bch_count_data(&b->keys) < oldsize);
2074         return ret;
2075 }
2076
2077 static int btree_split(struct btree *b, struct btree_op *op,
2078                        struct keylist *insert_keys,
2079                        struct bkey *replace_key)
2080 {
2081         bool split;
2082         struct btree *n1, *n2 = NULL, *n3 = NULL;
2083         uint64_t start_time = local_clock();
2084         struct closure cl;
2085         struct keylist parent_keys;
2086
2087         closure_init_stack(&cl);
2088         bch_keylist_init(&parent_keys);
2089
2090         if (btree_check_reserve(b, op)) {
2091                 if (!b->level)
2092                         return -EINTR;
2093                 else
2094                         WARN(1, "insufficient reserve for split\n");
2095         }
2096
2097         n1 = btree_node_alloc_replacement(b, op);
2098         if (IS_ERR(n1))
2099                 goto err;
2100
2101         split = set_blocks(btree_bset_first(n1),
2102                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2103
2104         if (split) {
2105                 unsigned int keys = 0;
2106
2107                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2108
2109                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2110                 if (IS_ERR(n2))
2111                         goto err_free1;
2112
2113                 if (!b->parent) {
2114                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2115                         if (IS_ERR(n3))
2116                                 goto err_free2;
2117                 }
2118
2119                 mutex_lock(&n1->write_lock);
2120                 mutex_lock(&n2->write_lock);
2121
2122                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2123
2124                 /*
2125                  * Has to be a linear search because we don't have an auxiliary
2126                  * search tree yet
2127                  */
2128
2129                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2130                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2131                                                         keys));
2132
2133                 bkey_copy_key(&n1->key,
2134                               bset_bkey_idx(btree_bset_first(n1), keys));
2135                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2136
2137                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2138                 btree_bset_first(n1)->keys = keys;
2139
2140                 memcpy(btree_bset_first(n2)->start,
2141                        bset_bkey_last(btree_bset_first(n1)),
2142                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2143
2144                 bkey_copy_key(&n2->key, &b->key);
2145
2146                 bch_keylist_add(&parent_keys, &n2->key);
2147                 bch_btree_node_write(n2, &cl);
2148                 mutex_unlock(&n2->write_lock);
2149                 rw_unlock(true, n2);
2150         } else {
2151                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2152
2153                 mutex_lock(&n1->write_lock);
2154                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2155         }
2156
2157         bch_keylist_add(&parent_keys, &n1->key);
2158         bch_btree_node_write(n1, &cl);
2159         mutex_unlock(&n1->write_lock);
2160
2161         if (n3) {
2162                 /* Depth increases, make a new root */
2163                 mutex_lock(&n3->write_lock);
2164                 bkey_copy_key(&n3->key, &MAX_KEY);
2165                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2166                 bch_btree_node_write(n3, &cl);
2167                 mutex_unlock(&n3->write_lock);
2168
2169                 closure_sync(&cl);
2170                 bch_btree_set_root(n3);
2171                 rw_unlock(true, n3);
2172         } else if (!b->parent) {
2173                 /* Root filled up but didn't need to be split */
2174                 closure_sync(&cl);
2175                 bch_btree_set_root(n1);
2176         } else {
2177                 /* Split a non root node */
2178                 closure_sync(&cl);
2179                 make_btree_freeing_key(b, parent_keys.top);
2180                 bch_keylist_push(&parent_keys);
2181
2182                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2183                 BUG_ON(!bch_keylist_empty(&parent_keys));
2184         }
2185
2186         btree_node_free(b);
2187         rw_unlock(true, n1);
2188
2189         bch_time_stats_update(&b->c->btree_split_time, start_time);
2190
2191         return 0;
2192 err_free2:
2193         bkey_put(b->c, &n2->key);
2194         btree_node_free(n2);
2195         rw_unlock(true, n2);
2196 err_free1:
2197         bkey_put(b->c, &n1->key);
2198         btree_node_free(n1);
2199         rw_unlock(true, n1);
2200 err:
2201         WARN(1, "bcache: btree split failed (level %u)", b->level);
2202
2203         if (n3 == ERR_PTR(-EAGAIN) ||
2204             n2 == ERR_PTR(-EAGAIN) ||
2205             n1 == ERR_PTR(-EAGAIN))
2206                 return -EAGAIN;
2207
2208         return -ENOMEM;
2209 }
2210
2211 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2212                                  struct keylist *insert_keys,
2213                                  atomic_t *journal_ref,
2214                                  struct bkey *replace_key)
2215 {
2216         struct closure cl;
2217
2218         BUG_ON(b->level && replace_key);
2219
2220         closure_init_stack(&cl);
2221
2222         mutex_lock(&b->write_lock);
2223
2224         if (write_block(b) != btree_bset_last(b) &&
2225             b->keys.last_set_unwritten)
2226                 bch_btree_init_next(b); /* just wrote a set */
2227
2228         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2229                 mutex_unlock(&b->write_lock);
2230                 goto split;
2231         }
2232
2233         BUG_ON(write_block(b) != btree_bset_last(b));
2234
2235         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2236                 if (!b->level)
2237                         bch_btree_leaf_dirty(b, journal_ref);
2238                 else
2239                         bch_btree_node_write(b, &cl);
2240         }
2241
2242         mutex_unlock(&b->write_lock);
2243
2244         /* wait for btree node write if necessary, after unlock */
2245         closure_sync(&cl);
2246
2247         return 0;
2248 split:
2249         if (current->bio_list) {
2250                 op->lock = b->c->root->level + 1;
2251                 return -EAGAIN;
2252         } else if (op->lock <= b->c->root->level) {
2253                 op->lock = b->c->root->level + 1;
2254                 return -EINTR;
2255         } else {
2256                 /* Invalidated all iterators */
2257                 int ret = btree_split(b, op, insert_keys, replace_key);
2258
2259                 if (bch_keylist_empty(insert_keys))
2260                         return 0;
2261                 else if (!ret)
2262                         return -EINTR;
2263                 return ret;
2264         }
2265 }
2266
2267 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2268                                struct bkey *check_key)
2269 {
2270         int ret = -EINTR;
2271         uint64_t btree_ptr = b->key.ptr[0];
2272         unsigned long seq = b->seq;
2273         struct keylist insert;
2274         bool upgrade = op->lock == -1;
2275
2276         bch_keylist_init(&insert);
2277
2278         if (upgrade) {
2279                 rw_unlock(false, b);
2280                 rw_lock(true, b, b->level);
2281
2282                 if (b->key.ptr[0] != btree_ptr ||
2283                     b->seq != seq + 1) {
2284                         op->lock = b->level;
2285                         goto out;
2286                 }
2287         }
2288
2289         SET_KEY_PTRS(check_key, 1);
2290         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2291
2292         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2293
2294         bch_keylist_add(&insert, check_key);
2295
2296         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2297
2298         BUG_ON(!ret && !bch_keylist_empty(&insert));
2299 out:
2300         if (upgrade)
2301                 downgrade_write(&b->lock);
2302         return ret;
2303 }
2304
2305 struct btree_insert_op {
2306         struct btree_op op;
2307         struct keylist  *keys;
2308         atomic_t        *journal_ref;
2309         struct bkey     *replace_key;
2310 };
2311
2312 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2313 {
2314         struct btree_insert_op *op = container_of(b_op,
2315                                         struct btree_insert_op, op);
2316
2317         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2318                                         op->journal_ref, op->replace_key);
2319         if (ret && !bch_keylist_empty(op->keys))
2320                 return ret;
2321         else
2322                 return MAP_DONE;
2323 }
2324
2325 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2326                      atomic_t *journal_ref, struct bkey *replace_key)
2327 {
2328         struct btree_insert_op op;
2329         int ret = 0;
2330
2331         BUG_ON(current->bio_list);
2332         BUG_ON(bch_keylist_empty(keys));
2333
2334         bch_btree_op_init(&op.op, 0);
2335         op.keys         = keys;
2336         op.journal_ref  = journal_ref;
2337         op.replace_key  = replace_key;
2338
2339         while (!ret && !bch_keylist_empty(keys)) {
2340                 op.op.lock = 0;
2341                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2342                                                &START_KEY(keys->keys),
2343                                                btree_insert_fn);
2344         }
2345
2346         if (ret) {
2347                 struct bkey *k;
2348
2349                 pr_err("error %i", ret);
2350
2351                 while ((k = bch_keylist_pop(keys)))
2352                         bkey_put(c, k);
2353         } else if (op.op.insert_collision)
2354                 ret = -ESRCH;
2355
2356         return ret;
2357 }
2358
2359 void bch_btree_set_root(struct btree *b)
2360 {
2361         unsigned int i;
2362         struct closure cl;
2363
2364         closure_init_stack(&cl);
2365
2366         trace_bcache_btree_set_root(b);
2367
2368         BUG_ON(!b->written);
2369
2370         for (i = 0; i < KEY_PTRS(&b->key); i++)
2371                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2372
2373         mutex_lock(&b->c->bucket_lock);
2374         list_del_init(&b->list);
2375         mutex_unlock(&b->c->bucket_lock);
2376
2377         b->c->root = b;
2378
2379         bch_journal_meta(b->c, &cl);
2380         closure_sync(&cl);
2381 }
2382
2383 /* Map across nodes or keys */
2384
2385 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2386                                        struct bkey *from,
2387                                        btree_map_nodes_fn *fn, int flags)
2388 {
2389         int ret = MAP_CONTINUE;
2390
2391         if (b->level) {
2392                 struct bkey *k;
2393                 struct btree_iter iter;
2394
2395                 bch_btree_iter_init(&b->keys, &iter, from);
2396
2397                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2398                                                        bch_ptr_bad))) {
2399                         ret = btree(map_nodes_recurse, k, b,
2400                                     op, from, fn, flags);
2401                         from = NULL;
2402
2403                         if (ret != MAP_CONTINUE)
2404                                 return ret;
2405                 }
2406         }
2407
2408         if (!b->level || flags == MAP_ALL_NODES)
2409                 ret = fn(op, b);
2410
2411         return ret;
2412 }
2413
2414 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2415                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2416 {
2417         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2418 }
2419
2420 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2421                                       struct bkey *from, btree_map_keys_fn *fn,
2422                                       int flags)
2423 {
2424         int ret = MAP_CONTINUE;
2425         struct bkey *k;
2426         struct btree_iter iter;
2427
2428         bch_btree_iter_init(&b->keys, &iter, from);
2429
2430         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2431                 ret = !b->level
2432                         ? fn(op, b, k)
2433                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2434                 from = NULL;
2435
2436                 if (ret != MAP_CONTINUE)
2437                         return ret;
2438         }
2439
2440         if (!b->level && (flags & MAP_END_KEY))
2441                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2442                                      KEY_OFFSET(&b->key), 0));
2443
2444         return ret;
2445 }
2446
2447 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2448                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2449 {
2450         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2451 }
2452
2453 /* Keybuf code */
2454
2455 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2456 {
2457         /* Overlapping keys compare equal */
2458         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2459                 return -1;
2460         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2461                 return 1;
2462         return 0;
2463 }
2464
2465 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2466                                             struct keybuf_key *r)
2467 {
2468         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2469 }
2470
2471 struct refill {
2472         struct btree_op op;
2473         unsigned int    nr_found;
2474         struct keybuf   *buf;
2475         struct bkey     *end;
2476         keybuf_pred_fn  *pred;
2477 };
2478
2479 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2480                             struct bkey *k)
2481 {
2482         struct refill *refill = container_of(op, struct refill, op);
2483         struct keybuf *buf = refill->buf;
2484         int ret = MAP_CONTINUE;
2485
2486         if (bkey_cmp(k, refill->end) > 0) {
2487                 ret = MAP_DONE;
2488                 goto out;
2489         }
2490
2491         if (!KEY_SIZE(k)) /* end key */
2492                 goto out;
2493
2494         if (refill->pred(buf, k)) {
2495                 struct keybuf_key *w;
2496
2497                 spin_lock(&buf->lock);
2498
2499                 w = array_alloc(&buf->freelist);
2500                 if (!w) {
2501                         spin_unlock(&buf->lock);
2502                         return MAP_DONE;
2503                 }
2504
2505                 w->private = NULL;
2506                 bkey_copy(&w->key, k);
2507
2508                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2509                         array_free(&buf->freelist, w);
2510                 else
2511                         refill->nr_found++;
2512
2513                 if (array_freelist_empty(&buf->freelist))
2514                         ret = MAP_DONE;
2515
2516                 spin_unlock(&buf->lock);
2517         }
2518 out:
2519         buf->last_scanned = *k;
2520         return ret;
2521 }
2522
2523 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2524                        struct bkey *end, keybuf_pred_fn *pred)
2525 {
2526         struct bkey start = buf->last_scanned;
2527         struct refill refill;
2528
2529         cond_resched();
2530
2531         bch_btree_op_init(&refill.op, -1);
2532         refill.nr_found = 0;
2533         refill.buf      = buf;
2534         refill.end      = end;
2535         refill.pred     = pred;
2536
2537         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2538                            refill_keybuf_fn, MAP_END_KEY);
2539
2540         trace_bcache_keyscan(refill.nr_found,
2541                              KEY_INODE(&start), KEY_OFFSET(&start),
2542                              KEY_INODE(&buf->last_scanned),
2543                              KEY_OFFSET(&buf->last_scanned));
2544
2545         spin_lock(&buf->lock);
2546
2547         if (!RB_EMPTY_ROOT(&buf->keys)) {
2548                 struct keybuf_key *w;
2549
2550                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2551                 buf->start      = START_KEY(&w->key);
2552
2553                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2554                 buf->end        = w->key;
2555         } else {
2556                 buf->start      = MAX_KEY;
2557                 buf->end        = MAX_KEY;
2558         }
2559
2560         spin_unlock(&buf->lock);
2561 }
2562
2563 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2564 {
2565         rb_erase(&w->node, &buf->keys);
2566         array_free(&buf->freelist, w);
2567 }
2568
2569 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2570 {
2571         spin_lock(&buf->lock);
2572         __bch_keybuf_del(buf, w);
2573         spin_unlock(&buf->lock);
2574 }
2575
2576 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2577                                   struct bkey *end)
2578 {
2579         bool ret = false;
2580         struct keybuf_key *p, *w, s;
2581
2582         s.key = *start;
2583
2584         if (bkey_cmp(end, &buf->start) <= 0 ||
2585             bkey_cmp(start, &buf->end) >= 0)
2586                 return false;
2587
2588         spin_lock(&buf->lock);
2589         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2590
2591         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2592                 p = w;
2593                 w = RB_NEXT(w, node);
2594
2595                 if (p->private)
2596                         ret = true;
2597                 else
2598                         __bch_keybuf_del(buf, p);
2599         }
2600
2601         spin_unlock(&buf->lock);
2602         return ret;
2603 }
2604
2605 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2606 {
2607         struct keybuf_key *w;
2608
2609         spin_lock(&buf->lock);
2610
2611         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2612
2613         while (w && w->private)
2614                 w = RB_NEXT(w, node);
2615
2616         if (w)
2617                 w->private = ERR_PTR(-EINTR);
2618
2619         spin_unlock(&buf->lock);
2620         return w;
2621 }
2622
2623 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2624                                           struct keybuf *buf,
2625                                           struct bkey *end,
2626                                           keybuf_pred_fn *pred)
2627 {
2628         struct keybuf_key *ret;
2629
2630         while (1) {
2631                 ret = bch_keybuf_next(buf);
2632                 if (ret)
2633                         break;
2634
2635                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2636                         pr_debug("scan finished");
2637                         break;
2638                 }
2639
2640                 bch_refill_keybuf(c, buf, end, pred);
2641         }
2642
2643         return ret;
2644 }
2645
2646 void bch_keybuf_init(struct keybuf *buf)
2647 {
2648         buf->last_scanned       = MAX_KEY;
2649         buf->keys               = RB_ROOT;
2650
2651         spin_lock_init(&buf->lock);
2652         array_allocator_init(&buf->freelist);
2653 }