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         /*
 211          * c->fill_iter can allocate an iterator with more memory space
 212          * than static MAX_BSETS.
 213          * See the comment arount cache_set->fill_iter.
 214          */
 215         iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
 216         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 217         iter->used = 0;
 218
 219 #ifdef CONFIG_BCACHE_DEBUG
 220         iter->b = &b->keys;
 221 #endif
 222
 223         if (!i->seq)
 224                 goto err;
 225
 226         for (;
 227              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 228              i = write_block(b)) {
 229                 err = "unsupported bset version";
 230                 if (i->version > BCACHE_BSET_VERSION)
 231                         goto err;
 232
 233                 err = "bad btree header";
 234                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 235                     btree_blocks(b))
 236                         goto err;
 237
 238                 err = "bad magic";
 239                 if (i->magic != bset_magic(&b->c->sb))
 240                         goto err;
 241
 242                 err = "bad checksum";
 243                 switch (i->version) {
 244                 case 0:
 245                         if (i->csum != csum_set(i))
 246                                 goto err;
 247                         break;
 248                 case BCACHE_BSET_VERSION:
 249                         if (i->csum != btree_csum_set(b, i))
 250                                 goto err;
 251                         break;
 252                 }
 253
 254                 err = "empty set";
 255                 if (i != b->keys.set[0].data && !i->keys)
 256                         goto err;
 257
 258                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 259
 260                 b->written += set_blocks(i, block_bytes(b->c));
 261         }
 262
 263         err = "corrupted btree";
 264         for (i = write_block(b);
 265              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 266              i = ((void *) i) + block_bytes(b->c))
 267                 if (i->seq == b->keys.set[0].data->seq)
 268                         goto err;
 269
 270         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 271
 272         i = b->keys.set[0].data;
 273         err = "short btree key";
 274         if (b->keys.set[0].size &&
 275             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 276                 goto err;
 277
 278         if (b->written < btree_blocks(b))
 279                 bch_bset_init_next(&b->keys, write_block(b),
 280                                    bset_magic(&b->c->sb));
 281 out:
 282         mempool_free(iter, &b->c->fill_iter);
 283         return;
 284 err:
 285         set_btree_node_io_error(b);
 286         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 287                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 288                             bset_block_offset(b, i), i->keys);
 289         goto out;
 290 }
 291
 292 static void btree_node_read_endio(struct bio *bio)
 293 {
 294         struct closure *cl = bio->bi_private;
 295
 296         closure_put(cl);
 297 }
 298
 299 static void bch_btree_node_read(struct btree *b)
 300 {
 301         uint64_t start_time = local_clock();
 302         struct closure cl;
 303         struct bio *bio;
 304
 305         trace_bcache_btree_read(b);
 306
 307         closure_init_stack(&cl);
 308
 309         bio = bch_bbio_alloc(b->c);
 310         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 311         bio->bi_end_io  = btree_node_read_endio;
 312         bio->bi_private = &cl;
 313         bio->bi_opf = REQ_OP_READ | REQ_META;
 314
 315         bch_bio_map(bio, b->keys.set[0].data);
 316
 317         bch_submit_bbio(bio, b->c, &b->key, 0);
 318         closure_sync(&cl);
 319
 320         if (bio->bi_status)
 321                 set_btree_node_io_error(b);
 322
 323         bch_bbio_free(bio, b->c);
 324
 325         if (btree_node_io_error(b))
 326                 goto err;
 327
 328         bch_btree_node_read_done(b);
 329         bch_time_stats_update(&b->c->btree_read_time, start_time);
 330
 331         return;
 332 err:
 333         bch_cache_set_error(b->c, "io error reading bucket %zu",
 334                             PTR_BUCKET_NR(b->c, &b->key, 0));
 335 }
 336
 337 static void btree_complete_write(struct btree *b, struct btree_write *w)
 338 {
 339         if (w->prio_blocked &&
 340             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 341                 wake_up_allocators(b->c);
 342
 343         if (w->journal) {
 344                 atomic_dec_bug(w->journal);
 345                 __closure_wake_up(&b->c->journal.wait);
 346         }
 347
 348         w->prio_blocked = 0;
 349         w->journal      = NULL;
 350 }
 351
 352 static void btree_node_write_unlock(struct closure *cl)
 353 {
 354         struct btree *b = container_of(cl, struct btree, io);
 355
 356         up(&b->io_mutex);
 357 }
 358
 359 static void __btree_node_write_done(struct closure *cl)
 360 {
 361         struct btree *b = container_of(cl, struct btree, io);
 362         struct btree_write *w = btree_prev_write(b);
 363
 364         bch_bbio_free(b->bio, b->c);
 365         b->bio = NULL;
 366         btree_complete_write(b, w);
 367
 368         if (btree_node_dirty(b))
 369                 schedule_delayed_work(&b->work, 30 * HZ);
 370
 371         closure_return_with_destructor(cl, btree_node_write_unlock);
 372 }
 373
 374 static void btree_node_write_done(struct closure *cl)
 375 {
 376         struct btree *b = container_of(cl, struct btree, io);
 377
 378         bio_free_pages(b->bio);
 379         __btree_node_write_done(cl);
 380 }
 381
 382 static void btree_node_write_endio(struct bio *bio)
 383 {
 384         struct closure *cl = bio->bi_private;
 385         struct btree *b = container_of(cl, struct btree, io);
 386
 387         if (bio->bi_status)
 388                 set_btree_node_io_error(b);
 389
 390         bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 391         closure_put(cl);
 392 }
 393
 394 static void do_btree_node_write(struct btree *b)
 395 {
 396         struct closure *cl = &b->io;
 397         struct bset *i = btree_bset_last(b);
 398         BKEY_PADDED(key) k;
 399
 400         i->version      = BCACHE_BSET_VERSION;
 401         i->csum         = btree_csum_set(b, i);
 402
 403         BUG_ON(b->bio);
 404         b->bio = bch_bbio_alloc(b->c);
 405
 406         b->bio->bi_end_io       = btree_node_write_endio;
 407         b->bio->bi_private      = cl;
 408         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 409         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 410         bch_bio_map(b->bio, i);
 411
 412         /*
 413          * If we're appending to a leaf node, we don't technically need FUA -
 414          * this write just needs to be persisted before the next journal write,
 415          * which will be marked FLUSH|FUA.
 416          *
 417          * Similarly if we're writing a new btree root - the pointer is going to
 418          * be in the next journal entry.
 419          *
 420          * But if we're writing a new btree node (that isn't a root) or
 421          * appending to a non leaf btree node, we need either FUA or a flush
 422          * when we write the parent with the new pointer. FUA is cheaper than a
 423          * flush, and writes appending to leaf nodes aren't blocking anything so
 424          * just make all btree node writes FUA to keep things sane.
 425          */
 426
 427         bkey_copy(&k.key, &b->key);
 428         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 429                        bset_sector_offset(&b->keys, i));
 430
 431         if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 432                 struct bio_vec *bv;
 433                 void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 434                 struct bvec_iter_all iter_all;
 435
 436                 bio_for_each_segment_all(bv, b->bio, iter_all) {
 437                         memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
 438                         addr += PAGE_SIZE;
 439                 }
 440
 441                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 442
 443                 continue_at(cl, btree_node_write_done, NULL);
 444         } else {
 445                 /*
 446                  * No problem for multipage bvec since the bio is
 447                  * just allocated
 448                  */
 449                 b->bio->bi_vcnt = 0;
 450                 bch_bio_map(b->bio, i);
 451
 452                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 453
 454                 closure_sync(cl);
 455                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 456         }
 457 }
 458
 459 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 460 {
 461         struct bset *i = btree_bset_last(b);
 462
 463         lockdep_assert_held(&b->write_lock);
 464
 465         trace_bcache_btree_write(b);
 466
 467         BUG_ON(current->bio_list);
 468         BUG_ON(b->written >= btree_blocks(b));
 469         BUG_ON(b->written && !i->keys);
 470         BUG_ON(btree_bset_first(b)->seq != i->seq);
 471         bch_check_keys(&b->keys, "writing");
 472
 473         cancel_delayed_work(&b->work);
 474
 475         /* If caller isn't waiting for write, parent refcount is cache set */
 476         down(&b->io_mutex);
 477         closure_init(&b->io, parent ?: &b->c->cl);
 478
 479         clear_bit(BTREE_NODE_dirty,      &b->flags);
 480         change_bit(BTREE_NODE_write_idx, &b->flags);
 481
 482         do_btree_node_write(b);
 483
 484         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 485                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 486
 487         b->written += set_blocks(i, block_bytes(b->c));
 488 }
 489
 490 void bch_btree_node_write(struct btree *b, struct closure *parent)
 491 {
 492         unsigned int nsets = b->keys.nsets;
 493
 494         lockdep_assert_held(&b->lock);
 495
 496         __bch_btree_node_write(b, parent);
 497
 498         /*
 499          * do verify if there was more than one set initially (i.e. we did a
 500          * sort) and we sorted down to a single set:
 501          */
 502         if (nsets && !b->keys.nsets)
 503                 bch_btree_verify(b);
 504
 505         bch_btree_init_next(b);
 506 }
 507
 508 static void bch_btree_node_write_sync(struct btree *b)
 509 {
 510         struct closure cl;
 511
 512         closure_init_stack(&cl);
 513
 514         mutex_lock(&b->write_lock);
 515         bch_btree_node_write(b, &cl);
 516         mutex_unlock(&b->write_lock);
 517
 518         closure_sync(&cl);
 519 }
 520
 521 static void btree_node_write_work(struct work_struct *w)
 522 {
 523         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 524
 525         mutex_lock(&b->write_lock);
 526         if (btree_node_dirty(b))
 527                 __bch_btree_node_write(b, NULL);
 528         mutex_unlock(&b->write_lock);
 529 }
 530
 531 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 532 {
 533         struct bset *i = btree_bset_last(b);
 534         struct btree_write *w = btree_current_write(b);
 535
 536         lockdep_assert_held(&b->write_lock);
 537
 538         BUG_ON(!b->written);
 539         BUG_ON(!i->keys);
 540
 541         if (!btree_node_dirty(b))
 542                 schedule_delayed_work(&b->work, 30 * HZ);
 543
 544         set_btree_node_dirty(b);
 545
 546         /*
 547          * w->journal is always the oldest journal pin of all bkeys
 548          * in the leaf node, to make sure the oldest jset seq won't
 549          * be increased before this btree node is flushed.
 550          */
 551         if (journal_ref) {
 552                 if (w->journal &&
 553                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 554                         atomic_dec_bug(w->journal);
 555                         w->journal = NULL;
 556                 }
 557
 558                 if (!w->journal) {
 559                         w->journal = journal_ref;
 560                         atomic_inc(w->journal);
 561                 }
 562         }
 563
 564         /* Force write if set is too big */
 565         if (set_bytes(i) > PAGE_SIZE - 48 &&
 566             !current->bio_list)
 567                 bch_btree_node_write(b, NULL);
 568 }
 569
 570 /*
 571  * Btree in memory cache - allocation/freeing
 572  * mca -> memory cache
 573  */
 574
 575 #define mca_reserve(c)  (((c->root && c->root->level)           \
 576                           ? c->root->level : 1) * 8 + 16)
 577 #define mca_can_free(c)                                         \
 578         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 579
 580 static void mca_data_free(struct btree *b)
 581 {
 582         BUG_ON(b->io_mutex.count != 1);
 583
 584         bch_btree_keys_free(&b->keys);
 585
 586         b->c->btree_cache_used--;
 587         list_move(&b->list, &b->c->btree_cache_freed);
 588 }
 589
 590 static void mca_bucket_free(struct btree *b)
 591 {
 592         BUG_ON(btree_node_dirty(b));
 593
 594         b->key.ptr[0] = 0;
 595         hlist_del_init_rcu(&b->hash);
 596         list_move(&b->list, &b->c->btree_cache_freeable);
 597 }
 598
 599 static unsigned int btree_order(struct bkey *k)
 600 {
 601         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 602 }
 603
 604 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 605 {
 606         if (!bch_btree_keys_alloc(&b->keys,
 607                                   max_t(unsigned int,
 608                                         ilog2(b->c->btree_pages),
 609                                         btree_order(k)),
 610                                   gfp)) {
 611                 b->c->btree_cache_used++;
 612                 list_move(&b->list, &b->c->btree_cache);
 613         } else {
 614                 list_move(&b->list, &b->c->btree_cache_freed);
 615         }
 616 }
 617
 618 static struct btree *mca_bucket_alloc(struct cache_set *c,
 619                                       struct bkey *k, gfp_t gfp)
 620 {
 621         /*
 622          * kzalloc() is necessary here for initialization,
 623          * see code comments in bch_btree_keys_init().
 624          */
 625         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 626
 627         if (!b)
 628                 return NULL;
 629
 630         init_rwsem(&b->lock);
 631         lockdep_set_novalidate_class(&b->lock);
 632         mutex_init(&b->write_lock);
 633         lockdep_set_novalidate_class(&b->write_lock);
 634         INIT_LIST_HEAD(&b->list);
 635         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 636         b->c = c;
 637         sema_init(&b->io_mutex, 1);
 638
 639         mca_data_alloc(b, k, gfp);
 640         return b;
 641 }
 642
 643 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
 644 {
 645         struct closure cl;
 646
 647         closure_init_stack(&cl);
 648         lockdep_assert_held(&b->c->bucket_lock);
 649
 650         if (!down_write_trylock(&b->lock))
 651                 return -ENOMEM;
 652
 653         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 654
 655         if (b->keys.page_order < min_order)
 656                 goto out_unlock;
 657
 658         if (!flush) {
 659                 if (btree_node_dirty(b))
 660                         goto out_unlock;
 661
 662                 if (down_trylock(&b->io_mutex))
 663                         goto out_unlock;
 664                 up(&b->io_mutex);
 665         }
 666
 667 retry:
 668         /*
 669          * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
 670          * __bch_btree_node_write(). To avoid an extra flush, acquire
 671          * b->write_lock before checking BTREE_NODE_dirty bit.
 672          */
 673         mutex_lock(&b->write_lock);
 674         /*
 675          * If this btree node is selected in btree_flush_write() by journal
 676          * code, delay and retry until the node is flushed by journal code
 677          * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
 678          */
 679         if (btree_node_journal_flush(b)) {
 680                 pr_debug("bnode %p is flushing by journal, retry", b);
 681                 mutex_unlock(&b->write_lock);
 682                 udelay(1);
 683                 goto retry;
 684         }
 685
 686         if (btree_node_dirty(b))
 687                 __bch_btree_node_write(b, &cl);
 688         mutex_unlock(&b->write_lock);
 689
 690         closure_sync(&cl);
 691
 692         /* wait for any in flight btree write */
 693         down(&b->io_mutex);
 694         up(&b->io_mutex);
 695
 696         return 0;
 697 out_unlock:
 698         rw_unlock(true, b);
 699         return -ENOMEM;
 700 }
 701
 702 static unsigned long bch_mca_scan(struct shrinker *shrink,
 703                                   struct shrink_control *sc)
 704 {
 705         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 706         struct btree *b, *t;
 707         unsigned long i, nr = sc->nr_to_scan;
 708         unsigned long freed = 0;
 709         unsigned int btree_cache_used;
 710
 711         if (c->shrinker_disabled)
 712                 return SHRINK_STOP;
 713
 714         if (c->btree_cache_alloc_lock)
 715                 return SHRINK_STOP;
 716
 717         /* Return -1 if we can't do anything right now */
 718         if (sc->gfp_mask & __GFP_IO)
 719                 mutex_lock(&c->bucket_lock);
 720         else if (!mutex_trylock(&c->bucket_lock))
 721                 return -1;
 722
 723         /*
 724          * It's _really_ critical that we don't free too many btree nodes - we
 725          * have to always leave ourselves a reserve. The reserve is how we
 726          * guarantee that allocating memory for a new btree node can always
 727          * succeed, so that inserting keys into the btree can always succeed and
 728          * IO can always make forward progress:
 729          */
 730         nr /= c->btree_pages;
 731         if (nr == 0)
 732                 nr = 1;
 733         nr = min_t(unsigned long, nr, mca_can_free(c));
 734
 735         i = 0;
 736         btree_cache_used = c->btree_cache_used;
 737         list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
 738                 if (nr <= 0)
 739                         goto out;
 740
 741                 if (!mca_reap(b, 0, false)) {
 742                         mca_data_free(b);
 743                         rw_unlock(true, b);
 744                         freed++;
 745                 }
 746                 nr--;
 747                 i++;
 748         }
 749
 750         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 751                 if (nr <= 0 || i >= btree_cache_used)
 752                         goto out;
 753
 754                 if (!mca_reap(b, 0, false)) {
 755                         mca_bucket_free(b);
 756                         mca_data_free(b);
 757                         rw_unlock(true, b);
 758                         freed++;
 759                 }
 760
 761                 nr--;
 762                 i++;
 763         }
 764 out:
 765         mutex_unlock(&c->bucket_lock);
 766         return freed * c->btree_pages;
 767 }
 768
 769 static unsigned long bch_mca_count(struct shrinker *shrink,
 770                                    struct shrink_control *sc)
 771 {
 772         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 773
 774         if (c->shrinker_disabled)
 775                 return 0;
 776
 777         if (c->btree_cache_alloc_lock)
 778                 return 0;
 779
 780         return mca_can_free(c) * c->btree_pages;
 781 }
 782
 783 void bch_btree_cache_free(struct cache_set *c)
 784 {
 785         struct btree *b;
 786         struct closure cl;
 787
 788         closure_init_stack(&cl);
 789
 790         if (c->shrink.list.next)
 791                 unregister_shrinker(&c->shrink);
 792
 793         mutex_lock(&c->bucket_lock);
 794
 795 #ifdef CONFIG_BCACHE_DEBUG
 796         if (c->verify_data)
 797                 list_move(&c->verify_data->list, &c->btree_cache);
 798
 799         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 800 #endif
 801
 802         list_splice(&c->btree_cache_freeable,
 803                     &c->btree_cache);
 804
 805         while (!list_empty(&c->btree_cache)) {
 806                 b = list_first_entry(&c->btree_cache, struct btree, list);
 807
 808                 /*
 809                  * This function is called by cache_set_free(), no I/O
 810                  * request on cache now, it is unnecessary to acquire
 811                  * b->write_lock before clearing BTREE_NODE_dirty anymore.
 812                  */
 813                 if (btree_node_dirty(b)) {
 814                         btree_complete_write(b, btree_current_write(b));
 815                         clear_bit(BTREE_NODE_dirty, &b->flags);
 816                 }
 817                 mca_data_free(b);
 818         }
 819
 820         while (!list_empty(&c->btree_cache_freed)) {
 821                 b = list_first_entry(&c->btree_cache_freed,
 822                                      struct btree, list);
 823                 list_del(&b->list);
 824                 cancel_delayed_work_sync(&b->work);
 825                 kfree(b);
 826         }
 827
 828         mutex_unlock(&c->bucket_lock);
 829 }
 830
 831 int bch_btree_cache_alloc(struct cache_set *c)
 832 {
 833         unsigned int i;
 834
 835         for (i = 0; i < mca_reserve(c); i++)
 836                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 837                         return -ENOMEM;
 838
 839         list_splice_init(&c->btree_cache,
 840                          &c->btree_cache_freeable);
 841
 842 #ifdef CONFIG_BCACHE_DEBUG
 843         mutex_init(&c->verify_lock);
 844
 845         c->verify_ondisk = (void *)
 846                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 847
 848         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 849
 850         if (c->verify_data &&
 851             c->verify_data->keys.set->data)
 852                 list_del_init(&c->verify_data->list);
 853         else
 854                 c->verify_data = NULL;
 855 #endif
 856
 857         c->shrink.count_objects = bch_mca_count;
 858         c->shrink.scan_objects = bch_mca_scan;
 859         c->shrink.seeks = 4;
 860         c->shrink.batch = c->btree_pages * 2;
 861
 862         if (register_shrinker(&c->shrink))
 863                 pr_warn("bcache: %s: could not register shrinker",
 864                                 __func__);
 865
 866         return 0;
 867 }
 868
 869 /* Btree in memory cache - hash table */
 870
 871 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 872 {
 873         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 874 }
 875
 876 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 877 {
 878         struct btree *b;
 879
 880         rcu_read_lock();
 881         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 882                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 883                         goto out;
 884         b = NULL;
 885 out:
 886         rcu_read_unlock();
 887         return b;
 888 }
 889
 890 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 891 {
 892         spin_lock(&c->btree_cannibalize_lock);
 893         if (likely(c->btree_cache_alloc_lock == NULL)) {
 894                 c->btree_cache_alloc_lock = current;
 895         } else if (c->btree_cache_alloc_lock != current) {
 896                 if (op)
 897                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 898                                         TASK_UNINTERRUPTIBLE);
 899                 spin_unlock(&c->btree_cannibalize_lock);
 900                 return -EINTR;
 901         }
 902         spin_unlock(&c->btree_cannibalize_lock);
 903
 904         return 0;
 905 }
 906
 907 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 908                                      struct bkey *k)
 909 {
 910         struct btree *b;
 911
 912         trace_bcache_btree_cache_cannibalize(c);
 913
 914         if (mca_cannibalize_lock(c, op))
 915                 return ERR_PTR(-EINTR);
 916
 917         list_for_each_entry_reverse(b, &c->btree_cache, list)
 918                 if (!mca_reap(b, btree_order(k), false))
 919                         return b;
 920
 921         list_for_each_entry_reverse(b, &c->btree_cache, list)
 922                 if (!mca_reap(b, btree_order(k), true))
 923                         return b;
 924
 925         WARN(1, "btree cache cannibalize failed\n");
 926         return ERR_PTR(-ENOMEM);
 927 }
 928
 929 /*
 930  * We can only have one thread cannibalizing other cached btree nodes at a time,
 931  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 932  * cannibalize_bucket() will take. This means every time we unlock the root of
 933  * the btree, we need to release this lock if we have it held.
 934  */
 935 static void bch_cannibalize_unlock(struct cache_set *c)
 936 {
 937         spin_lock(&c->btree_cannibalize_lock);
 938         if (c->btree_cache_alloc_lock == current) {
 939                 c->btree_cache_alloc_lock = NULL;
 940                 wake_up(&c->btree_cache_wait);
 941         }
 942         spin_unlock(&c->btree_cannibalize_lock);
 943 }
 944
 945 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 946                                struct bkey *k, int level)
 947 {
 948         struct btree *b;
 949
 950         BUG_ON(current->bio_list);
 951
 952         lockdep_assert_held(&c->bucket_lock);
 953
 954         if (mca_find(c, k))
 955                 return NULL;
 956
 957         /* btree_free() doesn't free memory; it sticks the node on the end of
 958          * the list. Check if there's any freed nodes there:
 959          */
 960         list_for_each_entry(b, &c->btree_cache_freeable, list)
 961                 if (!mca_reap(b, btree_order(k), false))
 962                         goto out;
 963
 964         /* We never free struct btree itself, just the memory that holds the on
 965          * disk node. Check the freed list before allocating a new one:
 966          */
 967         list_for_each_entry(b, &c->btree_cache_freed, list)
 968                 if (!mca_reap(b, 0, false)) {
 969                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 970                         if (!b->keys.set[0].data)
 971                                 goto err;
 972                         else
 973                                 goto out;
 974                 }
 975
 976         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 977         if (!b)
 978                 goto err;
 979
 980         BUG_ON(!down_write_trylock(&b->lock));
 981         if (!b->keys.set->data)
 982                 goto err;
 983 out:
 984         BUG_ON(b->io_mutex.count != 1);
 985
 986         bkey_copy(&b->key, k);
 987         list_move(&b->list, &c->btree_cache);
 988         hlist_del_init_rcu(&b->hash);
 989         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 990
 991         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 992         b->parent       = (void *) ~0UL;
 993         b->flags        = 0;
 994         b->written      = 0;
 995         b->level        = level;
 996
 997         if (!b->level)
 998                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 999                                     &b->c->expensive_debug_checks);
1000         else
1001                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
1002                                     &b->c->expensive_debug_checks);
1003
1004         return b;
1005 err:
1006         if (b)
1007                 rw_unlock(true, b);
1008
1009         b = mca_cannibalize(c, op, k);
1010         if (!IS_ERR(b))
1011                 goto out;
1012
1013         return b;
1014 }
1015
1016 /*
1017  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
1018  * in from disk if necessary.
1019  *
1020  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
1021  *
1022  * The btree node will have either a read or a write lock held, depending on
1023  * level and op->lock.
1024  */
1025 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1026                                  struct bkey *k, int level, bool write,
1027                                  struct btree *parent)
1028 {
1029         int i = 0;
1030         struct btree *b;
1031
1032         BUG_ON(level < 0);
1033 retry:
1034         b = mca_find(c, k);
1035
1036         if (!b) {
1037                 if (current->bio_list)
1038                         return ERR_PTR(-EAGAIN);
1039
1040                 mutex_lock(&c->bucket_lock);
1041                 b = mca_alloc(c, op, k, level);
1042                 mutex_unlock(&c->bucket_lock);
1043
1044                 if (!b)
1045                         goto retry;
1046                 if (IS_ERR(b))
1047                         return b;
1048
1049                 bch_btree_node_read(b);
1050
1051                 if (!write)
1052                         downgrade_write(&b->lock);
1053         } else {
1054                 rw_lock(write, b, level);
1055                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1056                         rw_unlock(write, b);
1057                         goto retry;
1058                 }
1059                 BUG_ON(b->level != level);
1060         }
1061
1062         if (btree_node_io_error(b)) {
1063                 rw_unlock(write, b);
1064                 return ERR_PTR(-EIO);
1065         }
1066
1067         BUG_ON(!b->written);
1068
1069         b->parent = parent;
1070
1071         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1072                 prefetch(b->keys.set[i].tree);
1073                 prefetch(b->keys.set[i].data);
1074         }
1075
1076         for (; i <= b->keys.nsets; i++)
1077                 prefetch(b->keys.set[i].data);
1078
1079         return b;
1080 }
1081
1082 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1083 {
1084         struct btree *b;
1085
1086         mutex_lock(&parent->c->bucket_lock);
1087         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1088         mutex_unlock(&parent->c->bucket_lock);
1089
1090         if (!IS_ERR_OR_NULL(b)) {
1091                 b->parent = parent;
1092                 bch_btree_node_read(b);
1093                 rw_unlock(true, b);
1094         }
1095 }
1096
1097 /* Btree alloc */
1098
1099 static void btree_node_free(struct btree *b)
1100 {
1101         trace_bcache_btree_node_free(b);
1102
1103         BUG_ON(b == b->c->root);
1104
1105 retry:
1106         mutex_lock(&b->write_lock);
1107         /*
1108          * If the btree node is selected and flushing in btree_flush_write(),
1109          * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1110          * then it is safe to free the btree node here. Otherwise this btree
1111          * node will be in race condition.
1112          */
1113         if (btree_node_journal_flush(b)) {
1114                 mutex_unlock(&b->write_lock);
1115                 pr_debug("bnode %p journal_flush set, retry", b);
1116                 udelay(1);
1117                 goto retry;
1118         }
1119
1120         if (btree_node_dirty(b)) {
1121                 btree_complete_write(b, btree_current_write(b));
1122                 clear_bit(BTREE_NODE_dirty, &b->flags);
1123         }
1124
1125         mutex_unlock(&b->write_lock);
1126
1127         cancel_delayed_work(&b->work);
1128
1129         mutex_lock(&b->c->bucket_lock);
1130         bch_bucket_free(b->c, &b->key);
1131         mca_bucket_free(b);
1132         mutex_unlock(&b->c->bucket_lock);
1133 }
1134
1135 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1136                                      int level, bool wait,
1137                                      struct btree *parent)
1138 {
1139         BKEY_PADDED(key) k;
1140         struct btree *b = ERR_PTR(-EAGAIN);
1141
1142         mutex_lock(&c->bucket_lock);
1143 retry:
1144         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1145                 goto err;
1146
1147         bkey_put(c, &k.key);
1148         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1149
1150         b = mca_alloc(c, op, &k.key, level);
1151         if (IS_ERR(b))
1152                 goto err_free;
1153
1154         if (!b) {
1155                 cache_bug(c,
1156                         "Tried to allocate bucket that was in btree cache");
1157                 goto retry;
1158         }
1159
1160         b->parent = parent;
1161         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1162
1163         mutex_unlock(&c->bucket_lock);
1164
1165         trace_bcache_btree_node_alloc(b);
1166         return b;
1167 err_free:
1168         bch_bucket_free(c, &k.key);
1169 err:
1170         mutex_unlock(&c->bucket_lock);
1171
1172         trace_bcache_btree_node_alloc_fail(c);
1173         return b;
1174 }
1175
1176 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1177                                           struct btree_op *op, int level,
1178                                           struct btree *parent)
1179 {
1180         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1181 }
1182
1183 static struct btree *btree_node_alloc_replacement(struct btree *b,
1184                                                   struct btree_op *op)
1185 {
1186         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1187
1188         if (!IS_ERR_OR_NULL(n)) {
1189                 mutex_lock(&n->write_lock);
1190                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1191                 bkey_copy_key(&n->key, &b->key);
1192                 mutex_unlock(&n->write_lock);
1193         }
1194
1195         return n;
1196 }
1197
1198 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1199 {
1200         unsigned int i;
1201
1202         mutex_lock(&b->c->bucket_lock);
1203
1204         atomic_inc(&b->c->prio_blocked);
1205
1206         bkey_copy(k, &b->key);
1207         bkey_copy_key(k, &ZERO_KEY);
1208
1209         for (i = 0; i < KEY_PTRS(k); i++)
1210                 SET_PTR_GEN(k, i,
1211                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1212                                         PTR_BUCKET(b->c, &b->key, i)));
1213
1214         mutex_unlock(&b->c->bucket_lock);
1215 }
1216
1217 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1218 {
1219         struct cache_set *c = b->c;
1220         struct cache *ca;
1221         unsigned int i, reserve = (c->root->level - b->level) * 2 + 1;
1222
1223         mutex_lock(&c->bucket_lock);
1224
1225         for_each_cache(ca, c, i)
1226                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1227                         if (op)
1228                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1229                                                 TASK_UNINTERRUPTIBLE);
1230                         mutex_unlock(&c->bucket_lock);
1231                         return -EINTR;
1232                 }
1233
1234         mutex_unlock(&c->bucket_lock);
1235
1236         return mca_cannibalize_lock(b->c, op);
1237 }
1238
1239 /* Garbage collection */
1240
1241 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1242                                     struct bkey *k)
1243 {
1244         uint8_t stale = 0;
1245         unsigned int i;
1246         struct bucket *g;
1247
1248         /*
1249          * ptr_invalid() can't return true for the keys that mark btree nodes as
1250          * freed, but since ptr_bad() returns true we'll never actually use them
1251          * for anything and thus we don't want mark their pointers here
1252          */
1253         if (!bkey_cmp(k, &ZERO_KEY))
1254                 return stale;
1255
1256         for (i = 0; i < KEY_PTRS(k); i++) {
1257                 if (!ptr_available(c, k, i))
1258                         continue;
1259
1260                 g = PTR_BUCKET(c, k, i);
1261
1262                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1263                         g->last_gc = PTR_GEN(k, i);
1264
1265                 if (ptr_stale(c, k, i)) {
1266                         stale = max(stale, ptr_stale(c, k, i));
1267                         continue;
1268                 }
1269
1270                 cache_bug_on(GC_MARK(g) &&
1271                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1272                              c, "inconsistent ptrs: mark = %llu, level = %i",
1273                              GC_MARK(g), level);
1274
1275                 if (level)
1276                         SET_GC_MARK(g, GC_MARK_METADATA);
1277                 else if (KEY_DIRTY(k))
1278                         SET_GC_MARK(g, GC_MARK_DIRTY);
1279                 else if (!GC_MARK(g))
1280                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1281
1282                 /* guard against overflow */
1283                 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1284                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1285                                              MAX_GC_SECTORS_USED));
1286
1287                 BUG_ON(!GC_SECTORS_USED(g));
1288         }
1289
1290         return stale;
1291 }
1292
1293 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1294
1295 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1296 {
1297         unsigned int i;
1298
1299         for (i = 0; i < KEY_PTRS(k); i++)
1300                 if (ptr_available(c, k, i) &&
1301                     !ptr_stale(c, k, i)) {
1302                         struct bucket *b = PTR_BUCKET(c, k, i);
1303
1304                         b->gen = PTR_GEN(k, i);
1305
1306                         if (level && bkey_cmp(k, &ZERO_KEY))
1307                                 b->prio = BTREE_PRIO;
1308                         else if (!level && b->prio == BTREE_PRIO)
1309                                 b->prio = INITIAL_PRIO;
1310                 }
1311
1312         __bch_btree_mark_key(c, level, k);
1313 }
1314
1315 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1316 {
1317         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1318 }
1319
1320 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1321 {
1322         uint8_t stale = 0;
1323         unsigned int keys = 0, good_keys = 0;
1324         struct bkey *k;
1325         struct btree_iter iter;
1326         struct bset_tree *t;
1327
1328         gc->nodes++;
1329
1330         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1331                 stale = max(stale, btree_mark_key(b, k));
1332                 keys++;
1333
1334                 if (bch_ptr_bad(&b->keys, k))
1335                         continue;
1336
1337                 gc->key_bytes += bkey_u64s(k);
1338                 gc->nkeys++;
1339                 good_keys++;
1340
1341                 gc->data += KEY_SIZE(k);
1342         }
1343
1344         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1345                 btree_bug_on(t->size &&
1346                              bset_written(&b->keys, t) &&
1347                              bkey_cmp(&b->key, &t->end) < 0,
1348                              b, "found short btree key in gc");
1349
1350         if (b->c->gc_always_rewrite)
1351                 return true;
1352
1353         if (stale > 10)
1354                 return true;
1355
1356         if ((keys - good_keys) * 2 > keys)
1357                 return true;
1358
1359         return false;
1360 }
1361
1362 #define GC_MERGE_NODES  4U
1363
1364 struct gc_merge_info {
1365         struct btree    *b;
1366         unsigned int    keys;
1367 };
1368
1369 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1370                                  struct keylist *insert_keys,
1371                                  atomic_t *journal_ref,
1372                                  struct bkey *replace_key);
1373
1374 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1375                              struct gc_stat *gc, struct gc_merge_info *r)
1376 {
1377         unsigned int i, nodes = 0, keys = 0, blocks;
1378         struct btree *new_nodes[GC_MERGE_NODES];
1379         struct keylist keylist;
1380         struct closure cl;
1381         struct bkey *k;
1382
1383         bch_keylist_init(&keylist);
1384
1385         if (btree_check_reserve(b, NULL))
1386                 return 0;
1387
1388         memset(new_nodes, 0, sizeof(new_nodes));
1389         closure_init_stack(&cl);
1390
1391         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1392                 keys += r[nodes++].keys;
1393
1394         blocks = btree_default_blocks(b->c) * 2 / 3;
1395
1396         if (nodes < 2 ||
1397             __set_blocks(b->keys.set[0].data, keys,
1398                          block_bytes(b->c)) > blocks * (nodes - 1))
1399                 return 0;
1400
1401         for (i = 0; i < nodes; i++) {
1402                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1403                 if (IS_ERR_OR_NULL(new_nodes[i]))
1404                         goto out_nocoalesce;
1405         }
1406
1407         /*
1408          * We have to check the reserve here, after we've allocated our new
1409          * nodes, to make sure the insert below will succeed - we also check
1410          * before as an optimization to potentially avoid a bunch of expensive
1411          * allocs/sorts
1412          */
1413         if (btree_check_reserve(b, NULL))
1414                 goto out_nocoalesce;
1415
1416         for (i = 0; i < nodes; i++)
1417                 mutex_lock(&new_nodes[i]->write_lock);
1418
1419         for (i = nodes - 1; i > 0; --i) {
1420                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1421                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1422                 struct bkey *k, *last = NULL;
1423
1424                 keys = 0;
1425
1426                 if (i > 1) {
1427                         for (k = n2->start;
1428                              k < bset_bkey_last(n2);
1429                              k = bkey_next(k)) {
1430                                 if (__set_blocks(n1, n1->keys + keys +
1431                                                  bkey_u64s(k),
1432                                                  block_bytes(b->c)) > blocks)
1433                                         break;
1434
1435                                 last = k;
1436                                 keys += bkey_u64s(k);
1437                         }
1438                 } else {
1439                         /*
1440                          * Last node we're not getting rid of - we're getting
1441                          * rid of the node at r[0]. Have to try and fit all of
1442                          * the remaining keys into this node; we can't ensure
1443                          * they will always fit due to rounding and variable
1444                          * length keys (shouldn't be possible in practice,
1445                          * though)
1446                          */
1447                         if (__set_blocks(n1, n1->keys + n2->keys,
1448                                          block_bytes(b->c)) >
1449                             btree_blocks(new_nodes[i]))
1450                                 goto out_nocoalesce;
1451
1452                         keys = n2->keys;
1453                         /* Take the key of the node we're getting rid of */
1454                         last = &r->b->key;
1455                 }
1456
1457                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1458                        btree_blocks(new_nodes[i]));
1459
1460                 if (last)
1461                         bkey_copy_key(&new_nodes[i]->key, last);
1462
1463                 memcpy(bset_bkey_last(n1),
1464                        n2->start,
1465                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1466
1467                 n1->keys += keys;
1468                 r[i].keys = n1->keys;
1469
1470                 memmove(n2->start,
1471                         bset_bkey_idx(n2, keys),
1472                         (void *) bset_bkey_last(n2) -
1473                         (void *) bset_bkey_idx(n2, keys));
1474
1475                 n2->keys -= keys;
1476
1477                 if (__bch_keylist_realloc(&keylist,
1478                                           bkey_u64s(&new_nodes[i]->key)))
1479                         goto out_nocoalesce;
1480
1481                 bch_btree_node_write(new_nodes[i], &cl);
1482                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1483         }
1484
1485         for (i = 0; i < nodes; i++)
1486                 mutex_unlock(&new_nodes[i]->write_lock);
1487
1488         closure_sync(&cl);
1489
1490         /* We emptied out this node */
1491         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1492         btree_node_free(new_nodes[0]);
1493         rw_unlock(true, new_nodes[0]);
1494         new_nodes[0] = NULL;
1495
1496         for (i = 0; i < nodes; i++) {
1497                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1498                         goto out_nocoalesce;
1499
1500                 make_btree_freeing_key(r[i].b, keylist.top);
1501                 bch_keylist_push(&keylist);
1502         }
1503
1504         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1505         BUG_ON(!bch_keylist_empty(&keylist));
1506
1507         for (i = 0; i < nodes; i++) {
1508                 btree_node_free(r[i].b);
1509                 rw_unlock(true, r[i].b);
1510
1511                 r[i].b = new_nodes[i];
1512         }
1513
1514         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1515         r[nodes - 1].b = ERR_PTR(-EINTR);
1516
1517         trace_bcache_btree_gc_coalesce(nodes);
1518         gc->nodes--;
1519
1520         bch_keylist_free(&keylist);
1521
1522         /* Invalidated our iterator */
1523         return -EINTR;
1524
1525 out_nocoalesce:
1526         closure_sync(&cl);
1527
1528         while ((k = bch_keylist_pop(&keylist)))
1529                 if (!bkey_cmp(k, &ZERO_KEY))
1530                         atomic_dec(&b->c->prio_blocked);
1531         bch_keylist_free(&keylist);
1532
1533         for (i = 0; i < nodes; i++)
1534                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1535                         btree_node_free(new_nodes[i]);
1536                         rw_unlock(true, new_nodes[i]);
1537                 }
1538         return 0;
1539 }
1540
1541 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1542                                  struct btree *replace)
1543 {
1544         struct keylist keys;
1545         struct btree *n;
1546
1547         if (btree_check_reserve(b, NULL))
1548                 return 0;
1549
1550         n = btree_node_alloc_replacement(replace, NULL);
1551
1552         /* recheck reserve after allocating replacement node */
1553         if (btree_check_reserve(b, NULL)) {
1554                 btree_node_free(n);
1555                 rw_unlock(true, n);
1556                 return 0;
1557         }
1558
1559         bch_btree_node_write_sync(n);
1560
1561         bch_keylist_init(&keys);
1562         bch_keylist_add(&keys, &n->key);
1563
1564         make_btree_freeing_key(replace, keys.top);
1565         bch_keylist_push(&keys);
1566
1567         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1568         BUG_ON(!bch_keylist_empty(&keys));
1569
1570         btree_node_free(replace);
1571         rw_unlock(true, n);
1572
1573         /* Invalidated our iterator */
1574         return -EINTR;
1575 }
1576
1577 static unsigned int btree_gc_count_keys(struct btree *b)
1578 {
1579         struct bkey *k;
1580         struct btree_iter iter;
1581         unsigned int ret = 0;
1582
1583         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1584                 ret += bkey_u64s(k);
1585
1586         return ret;
1587 }
1588
1589 static size_t btree_gc_min_nodes(struct cache_set *c)
1590 {
1591         size_t min_nodes;
1592
1593         /*
1594          * Since incremental GC would stop 100ms when front
1595          * side I/O comes, so when there are many btree nodes,
1596          * if GC only processes constant (100) nodes each time,
1597          * GC would last a long time, and the front side I/Os
1598          * would run out of the buckets (since no new bucket
1599          * can be allocated during GC), and be blocked again.
1600          * So GC should not process constant nodes, but varied
1601          * nodes according to the number of btree nodes, which
1602          * realized by dividing GC into constant(100) times,
1603          * so when there are many btree nodes, GC can process
1604          * more nodes each time, otherwise, GC will process less
1605          * nodes each time (but no less than MIN_GC_NODES)
1606          */
1607         min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1608         if (min_nodes < MIN_GC_NODES)
1609                 min_nodes = MIN_GC_NODES;
1610
1611         return min_nodes;
1612 }
1613
1614
1615 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1616                             struct closure *writes, struct gc_stat *gc)
1617 {
1618         int ret = 0;
1619         bool should_rewrite;
1620         struct bkey *k;
1621         struct btree_iter iter;
1622         struct gc_merge_info r[GC_MERGE_NODES];
1623         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1624
1625         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1626
1627         for (i = r; i < r + ARRAY_SIZE(r); i++)
1628                 i->b = ERR_PTR(-EINTR);
1629
1630         while (1) {
1631                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1632                 if (k) {
1633                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1634                                                   true, b);
1635                         if (IS_ERR(r->b)) {
1636                                 ret = PTR_ERR(r->b);
1637                                 break;
1638                         }
1639
1640                         r->keys = btree_gc_count_keys(r->b);
1641
1642                         ret = btree_gc_coalesce(b, op, gc, r);
1643                         if (ret)
1644                                 break;
1645                 }
1646
1647                 if (!last->b)
1648                         break;
1649
1650                 if (!IS_ERR(last->b)) {
1651                         should_rewrite = btree_gc_mark_node(last->b, gc);
1652                         if (should_rewrite) {
1653                                 ret = btree_gc_rewrite_node(b, op, last->b);
1654                                 if (ret)
1655                                         break;
1656                         }
1657
1658                         if (last->b->level) {
1659                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1660                                 if (ret)
1661                                         break;
1662                         }
1663
1664                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1665
1666                         /*
1667                          * Must flush leaf nodes before gc ends, since replace
1668                          * operations aren't journalled
1669                          */
1670                         mutex_lock(&last->b->write_lock);
1671                         if (btree_node_dirty(last->b))
1672                                 bch_btree_node_write(last->b, writes);
1673                         mutex_unlock(&last->b->write_lock);
1674                         rw_unlock(true, last->b);
1675                 }
1676
1677                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1678                 r->b = NULL;
1679
1680                 if (atomic_read(&b->c->search_inflight) &&
1681                     gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1682                         gc->nodes_pre =  gc->nodes;
1683                         ret = -EAGAIN;
1684                         break;
1685                 }
1686
1687                 if (need_resched()) {
1688                         ret = -EAGAIN;
1689                         break;
1690                 }
1691         }
1692
1693         for (i = r; i < r + ARRAY_SIZE(r); i++)
1694                 if (!IS_ERR_OR_NULL(i->b)) {
1695                         mutex_lock(&i->b->write_lock);
1696                         if (btree_node_dirty(i->b))
1697                                 bch_btree_node_write(i->b, writes);
1698                         mutex_unlock(&i->b->write_lock);
1699                         rw_unlock(true, i->b);
1700                 }
1701
1702         return ret;
1703 }
1704
1705 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1706                              struct closure *writes, struct gc_stat *gc)
1707 {
1708         struct btree *n = NULL;
1709         int ret = 0;
1710         bool should_rewrite;
1711
1712         should_rewrite = btree_gc_mark_node(b, gc);
1713         if (should_rewrite) {
1714                 n = btree_node_alloc_replacement(b, NULL);
1715
1716                 if (!IS_ERR_OR_NULL(n)) {
1717                         bch_btree_node_write_sync(n);
1718
1719                         bch_btree_set_root(n);
1720                         btree_node_free(b);
1721                         rw_unlock(true, n);
1722
1723                         return -EINTR;
1724                 }
1725         }
1726
1727         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1728
1729         if (b->level) {
1730                 ret = btree_gc_recurse(b, op, writes, gc);
1731                 if (ret)
1732                         return ret;
1733         }
1734
1735         bkey_copy_key(&b->c->gc_done, &b->key);
1736
1737         return ret;
1738 }
1739
1740 static void btree_gc_start(struct cache_set *c)
1741 {
1742         struct cache *ca;
1743         struct bucket *b;
1744         unsigned int i;
1745
1746         if (!c->gc_mark_valid)
1747                 return;
1748
1749         mutex_lock(&c->bucket_lock);
1750
1751         c->gc_mark_valid = 0;
1752         c->gc_done = ZERO_KEY;
1753
1754         for_each_cache(ca, c, i)
1755                 for_each_bucket(b, ca) {
1756                         b->last_gc = b->gen;
1757                         if (!atomic_read(&b->pin)) {
1758                                 SET_GC_MARK(b, 0);
1759                                 SET_GC_SECTORS_USED(b, 0);
1760                         }
1761                 }
1762
1763         mutex_unlock(&c->bucket_lock);
1764 }
1765
1766 static void bch_btree_gc_finish(struct cache_set *c)
1767 {
1768         struct bucket *b;
1769         struct cache *ca;
1770         unsigned int i;
1771
1772         mutex_lock(&c->bucket_lock);
1773
1774         set_gc_sectors(c);
1775         c->gc_mark_valid = 1;
1776         c->need_gc      = 0;
1777
1778         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1779                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1780                             GC_MARK_METADATA);
1781
1782         /* don't reclaim buckets to which writeback keys point */
1783         rcu_read_lock();
1784         for (i = 0; i < c->devices_max_used; i++) {
1785                 struct bcache_device *d = c->devices[i];
1786                 struct cached_dev *dc;
1787                 struct keybuf_key *w, *n;
1788                 unsigned int j;
1789
1790                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1791                         continue;
1792                 dc = container_of(d, struct cached_dev, disk);
1793
1794                 spin_lock(&dc->writeback_keys.lock);
1795                 rbtree_postorder_for_each_entry_safe(w, n,
1796                                         &dc->writeback_keys.keys, node)
1797                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1798                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1799                                             GC_MARK_DIRTY);
1800                 spin_unlock(&dc->writeback_keys.lock);
1801         }
1802         rcu_read_unlock();
1803
1804         c->avail_nbuckets = 0;
1805         for_each_cache(ca, c, i) {
1806                 uint64_t *i;
1807
1808                 ca->invalidate_needs_gc = 0;
1809
1810                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1811                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1812
1813                 for (i = ca->prio_buckets;
1814                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1815                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1816
1817                 for_each_bucket(b, ca) {
1818                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1819
1820                         if (atomic_read(&b->pin))
1821                                 continue;
1822
1823                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1824
1825                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1826                                 c->avail_nbuckets++;
1827                 }
1828         }
1829
1830         mutex_unlock(&c->bucket_lock);
1831 }
1832
1833 static void bch_btree_gc(struct cache_set *c)
1834 {
1835         int ret;
1836         struct gc_stat stats;
1837         struct closure writes;
1838         struct btree_op op;
1839         uint64_t start_time = local_clock();
1840
1841         trace_bcache_gc_start(c);
1842
1843         memset(&stats, 0, sizeof(struct gc_stat));
1844         closure_init_stack(&writes);
1845         bch_btree_op_init(&op, SHRT_MAX);
1846
1847         btree_gc_start(c);
1848
1849         /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1850         do {
1851                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1852                 closure_sync(&writes);
1853                 cond_resched();
1854
1855                 if (ret == -EAGAIN)
1856                         schedule_timeout_interruptible(msecs_to_jiffies
1857                                                        (GC_SLEEP_MS));
1858                 else if (ret)
1859                         pr_warn("gc failed!");
1860         } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1861
1862         bch_btree_gc_finish(c);
1863         wake_up_allocators(c);
1864
1865         bch_time_stats_update(&c->btree_gc_time, start_time);
1866
1867         stats.key_bytes *= sizeof(uint64_t);
1868         stats.data      <<= 9;
1869         bch_update_bucket_in_use(c, &stats);
1870         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1871
1872         trace_bcache_gc_end(c);
1873
1874         bch_moving_gc(c);
1875 }
1876
1877 static bool gc_should_run(struct cache_set *c)
1878 {
1879         struct cache *ca;
1880         unsigned int i;
1881
1882         for_each_cache(ca, c, i)
1883                 if (ca->invalidate_needs_gc)
1884                         return true;
1885
1886         if (atomic_read(&c->sectors_to_gc) < 0)
1887                 return true;
1888
1889         return false;
1890 }
1891
1892 static int bch_gc_thread(void *arg)
1893 {
1894         struct cache_set *c = arg;
1895
1896         while (1) {
1897                 wait_event_interruptible(c->gc_wait,
1898                            kthread_should_stop() ||
1899                            test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1900                            gc_should_run(c));
1901
1902                 if (kthread_should_stop() ||
1903                     test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1904                         break;
1905
1906                 set_gc_sectors(c);
1907                 bch_btree_gc(c);
1908         }
1909
1910         wait_for_kthread_stop();
1911         return 0;
1912 }
1913
1914 int bch_gc_thread_start(struct cache_set *c)
1915 {
1916         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1917         return PTR_ERR_OR_ZERO(c->gc_thread);
1918 }
1919
1920 /* Initial partial gc */
1921
1922 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1923 {
1924         int ret = 0;
1925         struct bkey *k, *p = NULL;
1926         struct btree_iter iter;
1927
1928         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1929                 bch_initial_mark_key(b->c, b->level, k);
1930
1931         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1932
1933         if (b->level) {
1934                 bch_btree_iter_init(&b->keys, &iter, NULL);
1935
1936                 do {
1937                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1938                                                        bch_ptr_bad);
1939                         if (k) {
1940                                 btree_node_prefetch(b, k);
1941                                 /*
1942                                  * initiallize c->gc_stats.nodes
1943                                  * for incremental GC
1944                                  */
1945                                 b->c->gc_stats.nodes++;
1946                         }
1947
1948                         if (p)
1949                                 ret = btree(check_recurse, p, b, op);
1950
1951                         p = k;
1952                 } while (p && !ret);
1953         }
1954
1955         return ret;
1956 }
1957
1958 int bch_btree_check(struct cache_set *c)
1959 {
1960         struct btree_op op;
1961
1962         bch_btree_op_init(&op, SHRT_MAX);
1963
1964         return btree_root(check_recurse, c, &op);
1965 }
1966
1967 void bch_initial_gc_finish(struct cache_set *c)
1968 {
1969         struct cache *ca;
1970         struct bucket *b;
1971         unsigned int i;
1972
1973         bch_btree_gc_finish(c);
1974
1975         mutex_lock(&c->bucket_lock);
1976
1977         /*
1978          * We need to put some unused buckets directly on the prio freelist in
1979          * order to get the allocator thread started - it needs freed buckets in
1980          * order to rewrite the prios and gens, and it needs to rewrite prios
1981          * and gens in order to free buckets.
1982          *
1983          * This is only safe for buckets that have no live data in them, which
1984          * there should always be some of.
1985          */
1986         for_each_cache(ca, c, i) {
1987                 for_each_bucket(b, ca) {
1988                         if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1989                             fifo_full(&ca->free[RESERVE_BTREE]))
1990                                 break;
1991
1992                         if (bch_can_invalidate_bucket(ca, b) &&
1993                             !GC_MARK(b)) {
1994                                 __bch_invalidate_one_bucket(ca, b);
1995                                 if (!fifo_push(&ca->free[RESERVE_PRIO],
1996                                    b - ca->buckets))
1997                                         fifo_push(&ca->free[RESERVE_BTREE],
1998                                                   b - ca->buckets);
1999                         }
2000                 }
2001         }
2002
2003         mutex_unlock(&c->bucket_lock);
2004 }
2005
2006 /* Btree insertion */
2007
2008 static bool btree_insert_key(struct btree *b, struct bkey *k,
2009                              struct bkey *replace_key)
2010 {
2011         unsigned int status;
2012
2013         BUG_ON(bkey_cmp(k, &b->key) > 0);
2014
2015         status = bch_btree_insert_key(&b->keys, k, replace_key);
2016         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2017                 bch_check_keys(&b->keys, "%u for %s", status,
2018                                replace_key ? "replace" : "insert");
2019
2020                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2021                                               status);
2022                 return true;
2023         } else
2024                 return false;
2025 }
2026
2027 static size_t insert_u64s_remaining(struct btree *b)
2028 {
2029         long ret = bch_btree_keys_u64s_remaining(&b->keys);
2030
2031         /*
2032          * Might land in the middle of an existing extent and have to split it
2033          */
2034         if (b->keys.ops->is_extents)
2035                 ret -= KEY_MAX_U64S;
2036
2037         return max(ret, 0L);
2038 }
2039
2040 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2041                                   struct keylist *insert_keys,
2042                                   struct bkey *replace_key)
2043 {
2044         bool ret = false;
2045         int oldsize = bch_count_data(&b->keys);
2046
2047         while (!bch_keylist_empty(insert_keys)) {
2048                 struct bkey *k = insert_keys->keys;
2049
2050                 if (bkey_u64s(k) > insert_u64s_remaining(b))
2051                         break;
2052
2053                 if (bkey_cmp(k, &b->key) <= 0) {
2054                         if (!b->level)
2055                                 bkey_put(b->c, k);
2056
2057                         ret |= btree_insert_key(b, k, replace_key);
2058                         bch_keylist_pop_front(insert_keys);
2059                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2060                         BKEY_PADDED(key) temp;
2061                         bkey_copy(&temp.key, insert_keys->keys);
2062
2063                         bch_cut_back(&b->key, &temp.key);
2064                         bch_cut_front(&b->key, insert_keys->keys);
2065
2066                         ret |= btree_insert_key(b, &temp.key, replace_key);
2067                         break;
2068                 } else {
2069                         break;
2070                 }
2071         }
2072
2073         if (!ret)
2074                 op->insert_collision = true;
2075
2076         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2077
2078         BUG_ON(bch_count_data(&b->keys) < oldsize);
2079         return ret;
2080 }
2081
2082 static int btree_split(struct btree *b, struct btree_op *op,
2083                        struct keylist *insert_keys,
2084                        struct bkey *replace_key)
2085 {
2086         bool split;
2087         struct btree *n1, *n2 = NULL, *n3 = NULL;
2088         uint64_t start_time = local_clock();
2089         struct closure cl;
2090         struct keylist parent_keys;
2091
2092         closure_init_stack(&cl);
2093         bch_keylist_init(&parent_keys);
2094
2095         if (btree_check_reserve(b, op)) {
2096                 if (!b->level)
2097                         return -EINTR;
2098                 else
2099                         WARN(1, "insufficient reserve for split\n");
2100         }
2101
2102         n1 = btree_node_alloc_replacement(b, op);
2103         if (IS_ERR(n1))
2104                 goto err;
2105
2106         split = set_blocks(btree_bset_first(n1),
2107                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2108
2109         if (split) {
2110                 unsigned int keys = 0;
2111
2112                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2113
2114                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2115                 if (IS_ERR(n2))
2116                         goto err_free1;
2117
2118                 if (!b->parent) {
2119                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2120                         if (IS_ERR(n3))
2121                                 goto err_free2;
2122                 }
2123
2124                 mutex_lock(&n1->write_lock);
2125                 mutex_lock(&n2->write_lock);
2126
2127                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2128
2129                 /*
2130                  * Has to be a linear search because we don't have an auxiliary
2131                  * search tree yet
2132                  */
2133
2134                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2135                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2136                                                         keys));
2137
2138                 bkey_copy_key(&n1->key,
2139                               bset_bkey_idx(btree_bset_first(n1), keys));
2140                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2141
2142                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2143                 btree_bset_first(n1)->keys = keys;
2144
2145                 memcpy(btree_bset_first(n2)->start,
2146                        bset_bkey_last(btree_bset_first(n1)),
2147                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2148
2149                 bkey_copy_key(&n2->key, &b->key);
2150
2151                 bch_keylist_add(&parent_keys, &n2->key);
2152                 bch_btree_node_write(n2, &cl);
2153                 mutex_unlock(&n2->write_lock);
2154                 rw_unlock(true, n2);
2155         } else {
2156                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2157
2158                 mutex_lock(&n1->write_lock);
2159                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2160         }
2161
2162         bch_keylist_add(&parent_keys, &n1->key);
2163         bch_btree_node_write(n1, &cl);
2164         mutex_unlock(&n1->write_lock);
2165
2166         if (n3) {
2167                 /* Depth increases, make a new root */
2168                 mutex_lock(&n3->write_lock);
2169                 bkey_copy_key(&n3->key, &MAX_KEY);
2170                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2171                 bch_btree_node_write(n3, &cl);
2172                 mutex_unlock(&n3->write_lock);
2173
2174                 closure_sync(&cl);
2175                 bch_btree_set_root(n3);
2176                 rw_unlock(true, n3);
2177         } else if (!b->parent) {
2178                 /* Root filled up but didn't need to be split */
2179                 closure_sync(&cl);
2180                 bch_btree_set_root(n1);
2181         } else {
2182                 /* Split a non root node */
2183                 closure_sync(&cl);
2184                 make_btree_freeing_key(b, parent_keys.top);
2185                 bch_keylist_push(&parent_keys);
2186
2187                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2188                 BUG_ON(!bch_keylist_empty(&parent_keys));
2189         }
2190
2191         btree_node_free(b);
2192         rw_unlock(true, n1);
2193
2194         bch_time_stats_update(&b->c->btree_split_time, start_time);
2195
2196         return 0;
2197 err_free2:
2198         bkey_put(b->c, &n2->key);
2199         btree_node_free(n2);
2200         rw_unlock(true, n2);
2201 err_free1:
2202         bkey_put(b->c, &n1->key);
2203         btree_node_free(n1);
2204         rw_unlock(true, n1);
2205 err:
2206         WARN(1, "bcache: btree split failed (level %u)", b->level);
2207
2208         if (n3 == ERR_PTR(-EAGAIN) ||
2209             n2 == ERR_PTR(-EAGAIN) ||
2210             n1 == ERR_PTR(-EAGAIN))
2211                 return -EAGAIN;
2212
2213         return -ENOMEM;
2214 }
2215
2216 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2217                                  struct keylist *insert_keys,
2218                                  atomic_t *journal_ref,
2219                                  struct bkey *replace_key)
2220 {
2221         struct closure cl;
2222
2223         BUG_ON(b->level && replace_key);
2224
2225         closure_init_stack(&cl);
2226
2227         mutex_lock(&b->write_lock);
2228
2229         if (write_block(b) != btree_bset_last(b) &&
2230             b->keys.last_set_unwritten)
2231                 bch_btree_init_next(b); /* just wrote a set */
2232
2233         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2234                 mutex_unlock(&b->write_lock);
2235                 goto split;
2236         }
2237
2238         BUG_ON(write_block(b) != btree_bset_last(b));
2239
2240         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2241                 if (!b->level)
2242                         bch_btree_leaf_dirty(b, journal_ref);
2243                 else
2244                         bch_btree_node_write(b, &cl);
2245         }
2246
2247         mutex_unlock(&b->write_lock);
2248
2249         /* wait for btree node write if necessary, after unlock */
2250         closure_sync(&cl);
2251
2252         return 0;
2253 split:
2254         if (current->bio_list) {
2255                 op->lock = b->c->root->level + 1;
2256                 return -EAGAIN;
2257         } else if (op->lock <= b->c->root->level) {
2258                 op->lock = b->c->root->level + 1;
2259                 return -EINTR;
2260         } else {
2261                 /* Invalidated all iterators */
2262                 int ret = btree_split(b, op, insert_keys, replace_key);
2263
2264                 if (bch_keylist_empty(insert_keys))
2265                         return 0;
2266                 else if (!ret)
2267                         return -EINTR;
2268                 return ret;
2269         }
2270 }
2271
2272 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2273                                struct bkey *check_key)
2274 {
2275         int ret = -EINTR;
2276         uint64_t btree_ptr = b->key.ptr[0];
2277         unsigned long seq = b->seq;
2278         struct keylist insert;
2279         bool upgrade = op->lock == -1;
2280
2281         bch_keylist_init(&insert);
2282
2283         if (upgrade) {
2284                 rw_unlock(false, b);
2285                 rw_lock(true, b, b->level);
2286
2287                 if (b->key.ptr[0] != btree_ptr ||
2288                     b->seq != seq + 1) {
2289                         op->lock = b->level;
2290                         goto out;
2291                 }
2292         }
2293
2294         SET_KEY_PTRS(check_key, 1);
2295         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2296
2297         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2298
2299         bch_keylist_add(&insert, check_key);
2300
2301         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2302
2303         BUG_ON(!ret && !bch_keylist_empty(&insert));
2304 out:
2305         if (upgrade)
2306                 downgrade_write(&b->lock);
2307         return ret;
2308 }
2309
2310 struct btree_insert_op {
2311         struct btree_op op;
2312         struct keylist  *keys;
2313         atomic_t        *journal_ref;
2314         struct bkey     *replace_key;
2315 };
2316
2317 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2318 {
2319         struct btree_insert_op *op = container_of(b_op,
2320                                         struct btree_insert_op, op);
2321
2322         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2323                                         op->journal_ref, op->replace_key);
2324         if (ret && !bch_keylist_empty(op->keys))
2325                 return ret;
2326         else
2327                 return MAP_DONE;
2328 }
2329
2330 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2331                      atomic_t *journal_ref, struct bkey *replace_key)
2332 {
2333         struct btree_insert_op op;
2334         int ret = 0;
2335
2336         BUG_ON(current->bio_list);
2337         BUG_ON(bch_keylist_empty(keys));
2338
2339         bch_btree_op_init(&op.op, 0);
2340         op.keys         = keys;
2341         op.journal_ref  = journal_ref;
2342         op.replace_key  = replace_key;
2343
2344         while (!ret && !bch_keylist_empty(keys)) {
2345                 op.op.lock = 0;
2346                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2347                                                &START_KEY(keys->keys),
2348                                                btree_insert_fn);
2349         }
2350
2351         if (ret) {
2352                 struct bkey *k;
2353
2354                 pr_err("error %i", ret);
2355
2356                 while ((k = bch_keylist_pop(keys)))
2357                         bkey_put(c, k);
2358         } else if (op.op.insert_collision)
2359                 ret = -ESRCH;
2360
2361         return ret;
2362 }
2363
2364 void bch_btree_set_root(struct btree *b)
2365 {
2366         unsigned int i;
2367         struct closure cl;
2368
2369         closure_init_stack(&cl);
2370
2371         trace_bcache_btree_set_root(b);
2372
2373         BUG_ON(!b->written);
2374
2375         for (i = 0; i < KEY_PTRS(&b->key); i++)
2376                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2377
2378         mutex_lock(&b->c->bucket_lock);
2379         list_del_init(&b->list);
2380         mutex_unlock(&b->c->bucket_lock);
2381
2382         b->c->root = b;
2383
2384         bch_journal_meta(b->c, &cl);
2385         closure_sync(&cl);
2386 }
2387
2388 /* Map across nodes or keys */
2389
2390 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2391                                        struct bkey *from,
2392                                        btree_map_nodes_fn *fn, int flags)
2393 {
2394         int ret = MAP_CONTINUE;
2395
2396         if (b->level) {
2397                 struct bkey *k;
2398                 struct btree_iter iter;
2399
2400                 bch_btree_iter_init(&b->keys, &iter, from);
2401
2402                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2403                                                        bch_ptr_bad))) {
2404                         ret = btree(map_nodes_recurse, k, b,
2405                                     op, from, fn, flags);
2406                         from = NULL;
2407
2408                         if (ret != MAP_CONTINUE)
2409                                 return ret;
2410                 }
2411         }
2412
2413         if (!b->level || flags == MAP_ALL_NODES)
2414                 ret = fn(op, b);
2415
2416         return ret;
2417 }
2418
2419 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2420                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2421 {
2422         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2423 }
2424
2425 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2426                                       struct bkey *from, btree_map_keys_fn *fn,
2427                                       int flags)
2428 {
2429         int ret = MAP_CONTINUE;
2430         struct bkey *k;
2431         struct btree_iter iter;
2432
2433         bch_btree_iter_init(&b->keys, &iter, from);
2434
2435         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2436                 ret = !b->level
2437                         ? fn(op, b, k)
2438                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2439                 from = NULL;
2440
2441                 if (ret != MAP_CONTINUE)
2442                         return ret;
2443         }
2444
2445         if (!b->level && (flags & MAP_END_KEY))
2446                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2447                                      KEY_OFFSET(&b->key), 0));
2448
2449         return ret;
2450 }
2451
2452 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2453                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2454 {
2455         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2456 }
2457
2458 /* Keybuf code */
2459
2460 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2461 {
2462         /* Overlapping keys compare equal */
2463         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2464                 return -1;
2465         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2466                 return 1;
2467         return 0;
2468 }
2469
2470 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2471                                             struct keybuf_key *r)
2472 {
2473         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2474 }
2475
2476 struct refill {
2477         struct btree_op op;
2478         unsigned int    nr_found;
2479         struct keybuf   *buf;
2480         struct bkey     *end;
2481         keybuf_pred_fn  *pred;
2482 };
2483
2484 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2485                             struct bkey *k)
2486 {
2487         struct refill *refill = container_of(op, struct refill, op);
2488         struct keybuf *buf = refill->buf;
2489         int ret = MAP_CONTINUE;
2490
2491         if (bkey_cmp(k, refill->end) > 0) {
2492                 ret = MAP_DONE;
2493                 goto out;
2494         }
2495
2496         if (!KEY_SIZE(k)) /* end key */
2497                 goto out;
2498
2499         if (refill->pred(buf, k)) {
2500                 struct keybuf_key *w;
2501
2502                 spin_lock(&buf->lock);
2503
2504                 w = array_alloc(&buf->freelist);
2505                 if (!w) {
2506                         spin_unlock(&buf->lock);
2507                         return MAP_DONE;
2508                 }
2509
2510                 w->private = NULL;
2511                 bkey_copy(&w->key, k);
2512
2513                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2514                         array_free(&buf->freelist, w);
2515                 else
2516                         refill->nr_found++;
2517
2518                 if (array_freelist_empty(&buf->freelist))
2519                         ret = MAP_DONE;
2520
2521                 spin_unlock(&buf->lock);
2522         }
2523 out:
2524         buf->last_scanned = *k;
2525         return ret;
2526 }
2527
2528 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2529                        struct bkey *end, keybuf_pred_fn *pred)
2530 {
2531         struct bkey start = buf->last_scanned;
2532         struct refill refill;
2533
2534         cond_resched();
2535
2536         bch_btree_op_init(&refill.op, -1);
2537         refill.nr_found = 0;
2538         refill.buf      = buf;
2539         refill.end      = end;
2540         refill.pred     = pred;
2541
2542         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2543                            refill_keybuf_fn, MAP_END_KEY);
2544
2545         trace_bcache_keyscan(refill.nr_found,
2546                              KEY_INODE(&start), KEY_OFFSET(&start),
2547                              KEY_INODE(&buf->last_scanned),
2548                              KEY_OFFSET(&buf->last_scanned));
2549
2550         spin_lock(&buf->lock);
2551
2552         if (!RB_EMPTY_ROOT(&buf->keys)) {
2553                 struct keybuf_key *w;
2554
2555                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2556                 buf->start      = START_KEY(&w->key);
2557
2558                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2559                 buf->end        = w->key;
2560         } else {
2561                 buf->start      = MAX_KEY;
2562                 buf->end        = MAX_KEY;
2563         }
2564
2565         spin_unlock(&buf->lock);
2566 }
2567
2568 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2569 {
2570         rb_erase(&w->node, &buf->keys);
2571         array_free(&buf->freelist, w);
2572 }
2573
2574 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2575 {
2576         spin_lock(&buf->lock);
2577         __bch_keybuf_del(buf, w);
2578         spin_unlock(&buf->lock);
2579 }
2580
2581 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2582                                   struct bkey *end)
2583 {
2584         bool ret = false;
2585         struct keybuf_key *p, *w, s;
2586
2587         s.key = *start;
2588
2589         if (bkey_cmp(end, &buf->start) <= 0 ||
2590             bkey_cmp(start, &buf->end) >= 0)
2591                 return false;
2592
2593         spin_lock(&buf->lock);
2594         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2595
2596         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2597                 p = w;
2598                 w = RB_NEXT(w, node);
2599
2600                 if (p->private)
2601                         ret = true;
2602                 else
2603                         __bch_keybuf_del(buf, p);
2604         }
2605
2606         spin_unlock(&buf->lock);
2607         return ret;
2608 }
2609
2610 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2611 {
2612         struct keybuf_key *w;
2613
2614         spin_lock(&buf->lock);
2615
2616         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2617
2618         while (w && w->private)
2619                 w = RB_NEXT(w, node);
2620
2621         if (w)
2622                 w->private = ERR_PTR(-EINTR);
2623
2624         spin_unlock(&buf->lock);
2625         return w;
2626 }
2627
2628 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2629                                           struct keybuf *buf,
2630                                           struct bkey *end,
2631                                           keybuf_pred_fn *pred)
2632 {
2633         struct keybuf_key *ret;
2634
2635         while (1) {
2636                 ret = bch_keybuf_next(buf);
2637                 if (ret)
2638                         break;
2639
2640                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2641                         pr_debug("scan finished");
2642                         break;
2643                 }
2644
2645                 bch_refill_keybuf(c, buf, end, pred);
2646         }
2647
2648         return ret;
2649 }
2650
2651 void bch_keybuf_init(struct keybuf *buf)
2652 {
2653         buf->last_scanned       = MAX_KEY;
2654         buf->keys               = RB_ROOT;
2655
2656         spin_lock_init(&buf->lock);
2657         array_allocator_init(&buf->freelist);
2658 }