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/bcache.txt.
  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
  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
  94 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
  95
  96 #define PTR_HASH(c, k)                                                  \
  97         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
  98
  99 #define insert_lock(s, b)       ((b)->level <= (s)->lock)
 100
 101 /*
 102  * These macros are for recursing down the btree - they handle the details of
 103  * locking and looking up nodes in the cache for you. They're best treated as
 104  * mere syntax when reading code that uses them.
 105  *
 106  * op->lock determines whether we take a read or a write lock at a given depth.
 107  * If you've got a read lock and find that you need a write lock (i.e. you're
 108  * going to have to split), set op->lock and return -EINTR; btree_root() will
 109  * call you again and you'll have the correct lock.
 110  */
 111
 112 /**
 113  * btree - recurse down the btree on a specified key
 114  * @fn:         function to call, which will be passed the child node
 115  * @key:        key to recurse on
 116  * @b:          parent btree node
 117  * @op:         pointer to struct btree_op
 118  */
 119 #define btree(fn, key, b, op, ...)                                      \
 120 ({                                                                      \
 121         int _r, l = (b)->level - 1;                                     \
 122         bool _w = l <= (op)->lock;                                      \
 123         struct btree *_child = bch_btree_node_get((b)->c, op, key, l,   \
 124                                                   _w, b);               \
 125         if (!IS_ERR(_child)) {                                          \
 126                 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);       \
 127                 rw_unlock(_w, _child);                                  \
 128         } else                                                          \
 129                 _r = PTR_ERR(_child);                                   \
 130         _r;                                                             \
 131 })
 132
 133 /**
 134  * btree_root - call a function on the root of the btree
 135  * @fn:         function to call, which will be passed the child node
 136  * @c:          cache set
 137  * @op:         pointer to struct btree_op
 138  */
 139 #define btree_root(fn, c, op, ...)                                      \
 140 ({                                                                      \
 141         int _r = -EINTR;                                                \
 142         do {                                                            \
 143                 struct btree *_b = (c)->root;                           \
 144                 bool _w = insert_lock(op, _b);                          \
 145                 rw_lock(_w, _b, _b->level);                             \
 146                 if (_b == (c)->root &&                                  \
 147                     _w == insert_lock(op, _b)) {                        \
 148                         _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);   \
 149                 }                                                       \
 150                 rw_unlock(_w, _b);                                      \
 151                 bch_cannibalize_unlock(c);                              \
 152                 if (_r == -EINTR)                                       \
 153                         schedule();                                     \
 154         } while (_r == -EINTR);                                         \
 155                                                                         \
 156         finish_wait(&(c)->btree_cache_wait, &(op)->wait);               \
 157         _r;                                                             \
 158 })
 159
 160 static inline struct bset *write_block(struct btree *b)
 161 {
 162         return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
 163 }
 164
 165 static void bch_btree_init_next(struct btree *b)
 166 {
 167         /* If not a leaf node, always sort */
 168         if (b->level && b->keys.nsets)
 169                 bch_btree_sort(&b->keys, &b->c->sort);
 170         else
 171                 bch_btree_sort_lazy(&b->keys, &b->c->sort);
 172
 173         if (b->written < btree_blocks(b))
 174                 bch_bset_init_next(&b->keys, write_block(b),
 175                                    bset_magic(&b->c->sb));
 176
 177 }
 178
 179 /* Btree key manipulation */
 180
 181 void bkey_put(struct cache_set *c, struct bkey *k)
 182 {
 183         unsigned i;
 184
 185         for (i = 0; i < KEY_PTRS(k); i++)
 186                 if (ptr_available(c, k, i))
 187                         atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
 188 }
 189
 190 /* Btree IO */
 191
 192 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 193 {
 194         uint64_t crc = b->key.ptr[0];
 195         void *data = (void *) i + 8, *end = bset_bkey_last(i);
 196
 197         crc = bch_crc64_update(crc, data, end - data);
 198         return crc ^ 0xffffffffffffffffULL;
 199 }
 200
 201 void bch_btree_node_read_done(struct btree *b)
 202 {
 203         const char *err = "bad btree header";
 204         struct bset *i = btree_bset_first(b);
 205         struct btree_iter *iter;
 206
 207         iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
 208         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 209         iter->used = 0;
 210
 211 #ifdef CONFIG_BCACHE_DEBUG
 212         iter->b = &b->keys;
 213 #endif
 214
 215         if (!i->seq)
 216                 goto err;
 217
 218         for (;
 219              b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
 220              i = write_block(b)) {
 221                 err = "unsupported bset version";
 222                 if (i->version > BCACHE_BSET_VERSION)
 223                         goto err;
 224
 225                 err = "bad btree header";
 226                 if (b->written + set_blocks(i, block_bytes(b->c)) >
 227                     btree_blocks(b))
 228                         goto err;
 229
 230                 err = "bad magic";
 231                 if (i->magic != bset_magic(&b->c->sb))
 232                         goto err;
 233
 234                 err = "bad checksum";
 235                 switch (i->version) {
 236                 case 0:
 237                         if (i->csum != csum_set(i))
 238                                 goto err;
 239                         break;
 240                 case BCACHE_BSET_VERSION:
 241                         if (i->csum != btree_csum_set(b, i))
 242                                 goto err;
 243                         break;
 244                 }
 245
 246                 err = "empty set";
 247                 if (i != b->keys.set[0].data && !i->keys)
 248                         goto err;
 249
 250                 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
 251
 252                 b->written += set_blocks(i, block_bytes(b->c));
 253         }
 254
 255         err = "corrupted btree";
 256         for (i = write_block(b);
 257              bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
 258              i = ((void *) i) + block_bytes(b->c))
 259                 if (i->seq == b->keys.set[0].data->seq)
 260                         goto err;
 261
 262         bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
 263
 264         i = b->keys.set[0].data;
 265         err = "short btree key";
 266         if (b->keys.set[0].size &&
 267             bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
 268                 goto err;
 269
 270         if (b->written < btree_blocks(b))
 271                 bch_bset_init_next(&b->keys, write_block(b),
 272                                    bset_magic(&b->c->sb));
 273 out:
 274         mempool_free(iter, b->c->fill_iter);
 275         return;
 276 err:
 277         set_btree_node_io_error(b);
 278         bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
 279                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 280                             bset_block_offset(b, i), i->keys);
 281         goto out;
 282 }
 283
 284 static void btree_node_read_endio(struct bio *bio)
 285 {
 286         struct closure *cl = bio->bi_private;
 287         closure_put(cl);
 288 }
 289
 290 static void bch_btree_node_read(struct btree *b)
 291 {
 292         uint64_t start_time = local_clock();
 293         struct closure cl;
 294         struct bio *bio;
 295
 296         trace_bcache_btree_read(b);
 297
 298         closure_init_stack(&cl);
 299
 300         bio = bch_bbio_alloc(b->c);
 301         bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
 302         bio->bi_end_io  = btree_node_read_endio;
 303         bio->bi_private = &cl;
 304         bio->bi_opf = REQ_OP_READ | REQ_META;
 305
 306         bch_bio_map(bio, b->keys.set[0].data);
 307
 308         bch_submit_bbio(bio, b->c, &b->key, 0);
 309         closure_sync(&cl);
 310
 311         if (bio->bi_status)
 312                 set_btree_node_io_error(b);
 313
 314         bch_bbio_free(bio, b->c);
 315
 316         if (btree_node_io_error(b))
 317                 goto err;
 318
 319         bch_btree_node_read_done(b);
 320         bch_time_stats_update(&b->c->btree_read_time, start_time);
 321
 322         return;
 323 err:
 324         bch_cache_set_error(b->c, "io error reading bucket %zu",
 325                             PTR_BUCKET_NR(b->c, &b->key, 0));
 326 }
 327
 328 static void btree_complete_write(struct btree *b, struct btree_write *w)
 329 {
 330         if (w->prio_blocked &&
 331             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 332                 wake_up_allocators(b->c);
 333
 334         if (w->journal) {
 335                 atomic_dec_bug(w->journal);
 336                 __closure_wake_up(&b->c->journal.wait);
 337         }
 338
 339         w->prio_blocked = 0;
 340         w->journal      = NULL;
 341 }
 342
 343 static void btree_node_write_unlock(struct closure *cl)
 344 {
 345         struct btree *b = container_of(cl, struct btree, io);
 346
 347         up(&b->io_mutex);
 348 }
 349
 350 static void __btree_node_write_done(struct closure *cl)
 351 {
 352         struct btree *b = container_of(cl, struct btree, io);
 353         struct btree_write *w = btree_prev_write(b);
 354
 355         bch_bbio_free(b->bio, b->c);
 356         b->bio = NULL;
 357         btree_complete_write(b, w);
 358
 359         if (btree_node_dirty(b))
 360                 schedule_delayed_work(&b->work, 30 * HZ);
 361
 362         closure_return_with_destructor(cl, btree_node_write_unlock);
 363 }
 364
 365 static void btree_node_write_done(struct closure *cl)
 366 {
 367         struct btree *b = container_of(cl, struct btree, io);
 368
 369         bio_free_pages(b->bio);
 370         __btree_node_write_done(cl);
 371 }
 372
 373 static void btree_node_write_endio(struct bio *bio)
 374 {
 375         struct closure *cl = bio->bi_private;
 376         struct btree *b = container_of(cl, struct btree, io);
 377
 378         if (bio->bi_status)
 379                 set_btree_node_io_error(b);
 380
 381         bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
 382         closure_put(cl);
 383 }
 384
 385 static void do_btree_node_write(struct btree *b)
 386 {
 387         struct closure *cl = &b->io;
 388         struct bset *i = btree_bset_last(b);
 389         BKEY_PADDED(key) k;
 390
 391         i->version      = BCACHE_BSET_VERSION;
 392         i->csum         = btree_csum_set(b, i);
 393
 394         BUG_ON(b->bio);
 395         b->bio = bch_bbio_alloc(b->c);
 396
 397         b->bio->bi_end_io       = btree_node_write_endio;
 398         b->bio->bi_private      = cl;
 399         b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
 400         b->bio->bi_opf          = REQ_OP_WRITE | REQ_META | REQ_FUA;
 401         bch_bio_map(b->bio, i);
 402
 403         /*
 404          * If we're appending to a leaf node, we don't technically need FUA -
 405          * this write just needs to be persisted before the next journal write,
 406          * which will be marked FLUSH|FUA.
 407          *
 408          * Similarly if we're writing a new btree root - the pointer is going to
 409          * be in the next journal entry.
 410          *
 411          * But if we're writing a new btree node (that isn't a root) or
 412          * appending to a non leaf btree node, we need either FUA or a flush
 413          * when we write the parent with the new pointer. FUA is cheaper than a
 414          * flush, and writes appending to leaf nodes aren't blocking anything so
 415          * just make all btree node writes FUA to keep things sane.
 416          */
 417
 418         bkey_copy(&k.key, &b->key);
 419         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
 420                        bset_sector_offset(&b->keys, i));
 421
 422         if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
 423                 int j;
 424                 struct bio_vec *bv;
 425                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 426
 427                 bio_for_each_segment_all(bv, b->bio, j)
 428                         memcpy(page_address(bv->bv_page),
 429                                base + j * PAGE_SIZE, PAGE_SIZE);
 430
 431                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 432
 433                 continue_at(cl, btree_node_write_done, NULL);
 434         } else {
 435                 b->bio->bi_vcnt = 0;
 436                 bch_bio_map(b->bio, i);
 437
 438                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 439
 440                 closure_sync(cl);
 441                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 442         }
 443 }
 444
 445 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 446 {
 447         struct bset *i = btree_bset_last(b);
 448
 449         lockdep_assert_held(&b->write_lock);
 450
 451         trace_bcache_btree_write(b);
 452
 453         BUG_ON(current->bio_list);
 454         BUG_ON(b->written >= btree_blocks(b));
 455         BUG_ON(b->written && !i->keys);
 456         BUG_ON(btree_bset_first(b)->seq != i->seq);
 457         bch_check_keys(&b->keys, "writing");
 458
 459         cancel_delayed_work(&b->work);
 460
 461         /* If caller isn't waiting for write, parent refcount is cache set */
 462         down(&b->io_mutex);
 463         closure_init(&b->io, parent ?: &b->c->cl);
 464
 465         clear_bit(BTREE_NODE_dirty,      &b->flags);
 466         change_bit(BTREE_NODE_write_idx, &b->flags);
 467
 468         do_btree_node_write(b);
 469
 470         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 471                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 472
 473         b->written += set_blocks(i, block_bytes(b->c));
 474 }
 475
 476 void bch_btree_node_write(struct btree *b, struct closure *parent)
 477 {
 478         unsigned nsets = b->keys.nsets;
 479
 480         lockdep_assert_held(&b->lock);
 481
 482         __bch_btree_node_write(b, parent);
 483
 484         /*
 485          * do verify if there was more than one set initially (i.e. we did a
 486          * sort) and we sorted down to a single set:
 487          */
 488         if (nsets && !b->keys.nsets)
 489                 bch_btree_verify(b);
 490
 491         bch_btree_init_next(b);
 492 }
 493
 494 static void bch_btree_node_write_sync(struct btree *b)
 495 {
 496         struct closure cl;
 497
 498         closure_init_stack(&cl);
 499
 500         mutex_lock(&b->write_lock);
 501         bch_btree_node_write(b, &cl);
 502         mutex_unlock(&b->write_lock);
 503
 504         closure_sync(&cl);
 505 }
 506
 507 static void btree_node_write_work(struct work_struct *w)
 508 {
 509         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 510
 511         mutex_lock(&b->write_lock);
 512         if (btree_node_dirty(b))
 513                 __bch_btree_node_write(b, NULL);
 514         mutex_unlock(&b->write_lock);
 515 }
 516
 517 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 518 {
 519         struct bset *i = btree_bset_last(b);
 520         struct btree_write *w = btree_current_write(b);
 521
 522         lockdep_assert_held(&b->write_lock);
 523
 524         BUG_ON(!b->written);
 525         BUG_ON(!i->keys);
 526
 527         if (!btree_node_dirty(b))
 528                 schedule_delayed_work(&b->work, 30 * HZ);
 529
 530         set_btree_node_dirty(b);
 531
 532         if (journal_ref) {
 533                 if (w->journal &&
 534                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 535                         atomic_dec_bug(w->journal);
 536                         w->journal = NULL;
 537                 }
 538
 539                 if (!w->journal) {
 540                         w->journal = journal_ref;
 541                         atomic_inc(w->journal);
 542                 }
 543         }
 544
 545         /* Force write if set is too big */
 546         if (set_bytes(i) > PAGE_SIZE - 48 &&
 547             !current->bio_list)
 548                 bch_btree_node_write(b, NULL);
 549 }
 550
 551 /*
 552  * Btree in memory cache - allocation/freeing
 553  * mca -> memory cache
 554  */
 555
 556 #define mca_reserve(c)  (((c->root && c->root->level)           \
 557                           ? c->root->level : 1) * 8 + 16)
 558 #define mca_can_free(c)                                         \
 559         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 560
 561 static void mca_data_free(struct btree *b)
 562 {
 563         BUG_ON(b->io_mutex.count != 1);
 564
 565         bch_btree_keys_free(&b->keys);
 566
 567         b->c->btree_cache_used--;
 568         list_move(&b->list, &b->c->btree_cache_freed);
 569 }
 570
 571 static void mca_bucket_free(struct btree *b)
 572 {
 573         BUG_ON(btree_node_dirty(b));
 574
 575         b->key.ptr[0] = 0;
 576         hlist_del_init_rcu(&b->hash);
 577         list_move(&b->list, &b->c->btree_cache_freeable);
 578 }
 579
 580 static unsigned btree_order(struct bkey *k)
 581 {
 582         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 583 }
 584
 585 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 586 {
 587         if (!bch_btree_keys_alloc(&b->keys,
 588                                   max_t(unsigned,
 589                                         ilog2(b->c->btree_pages),
 590                                         btree_order(k)),
 591                                   gfp)) {
 592                 b->c->btree_cache_used++;
 593                 list_move(&b->list, &b->c->btree_cache);
 594         } else {
 595                 list_move(&b->list, &b->c->btree_cache_freed);
 596         }
 597 }
 598
 599 static struct btree *mca_bucket_alloc(struct cache_set *c,
 600                                       struct bkey *k, gfp_t gfp)
 601 {
 602         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 603         if (!b)
 604                 return NULL;
 605
 606         init_rwsem(&b->lock);
 607         lockdep_set_novalidate_class(&b->lock);
 608         mutex_init(&b->write_lock);
 609         lockdep_set_novalidate_class(&b->write_lock);
 610         INIT_LIST_HEAD(&b->list);
 611         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 612         b->c = c;
 613         sema_init(&b->io_mutex, 1);
 614
 615         mca_data_alloc(b, k, gfp);
 616         return b;
 617 }
 618
 619 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 620 {
 621         struct closure cl;
 622
 623         closure_init_stack(&cl);
 624         lockdep_assert_held(&b->c->bucket_lock);
 625
 626         if (!down_write_trylock(&b->lock))
 627                 return -ENOMEM;
 628
 629         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 630
 631         if (b->keys.page_order < min_order)
 632                 goto out_unlock;
 633
 634         if (!flush) {
 635                 if (btree_node_dirty(b))
 636                         goto out_unlock;
 637
 638                 if (down_trylock(&b->io_mutex))
 639                         goto out_unlock;
 640                 up(&b->io_mutex);
 641         }
 642
 643         mutex_lock(&b->write_lock);
 644         if (btree_node_dirty(b))
 645                 __bch_btree_node_write(b, &cl);
 646         mutex_unlock(&b->write_lock);
 647
 648         closure_sync(&cl);
 649
 650         /* wait for any in flight btree write */
 651         down(&b->io_mutex);
 652         up(&b->io_mutex);
 653
 654         return 0;
 655 out_unlock:
 656         rw_unlock(true, b);
 657         return -ENOMEM;
 658 }
 659
 660 static unsigned long bch_mca_scan(struct shrinker *shrink,
 661                                   struct shrink_control *sc)
 662 {
 663         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 664         struct btree *b, *t;
 665         unsigned long i, nr = sc->nr_to_scan;
 666         unsigned long freed = 0;
 667
 668         if (c->shrinker_disabled)
 669                 return SHRINK_STOP;
 670
 671         if (c->btree_cache_alloc_lock)
 672                 return SHRINK_STOP;
 673
 674         /* Return -1 if we can't do anything right now */
 675         if (sc->gfp_mask & __GFP_IO)
 676                 mutex_lock(&c->bucket_lock);
 677         else if (!mutex_trylock(&c->bucket_lock))
 678                 return -1;
 679
 680         /*
 681          * It's _really_ critical that we don't free too many btree nodes - we
 682          * have to always leave ourselves a reserve. The reserve is how we
 683          * guarantee that allocating memory for a new btree node can always
 684          * succeed, so that inserting keys into the btree can always succeed and
 685          * IO can always make forward progress:
 686          */
 687         nr /= c->btree_pages;
 688         nr = min_t(unsigned long, nr, mca_can_free(c));
 689
 690         i = 0;
 691         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 692                 if (freed >= nr)
 693                         break;
 694
 695                 if (++i > 3 &&
 696                     !mca_reap(b, 0, false)) {
 697                         mca_data_free(b);
 698                         rw_unlock(true, b);
 699                         freed++;
 700                 }
 701         }
 702
 703         for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
 704                 if (list_empty(&c->btree_cache))
 705                         goto out;
 706
 707                 b = list_first_entry(&c->btree_cache, struct btree, list);
 708                 list_rotate_left(&c->btree_cache);
 709
 710                 if (!b->accessed &&
 711                     !mca_reap(b, 0, false)) {
 712                         mca_bucket_free(b);
 713                         mca_data_free(b);
 714                         rw_unlock(true, b);
 715                         freed++;
 716                 } else
 717                         b->accessed = 0;
 718         }
 719 out:
 720         mutex_unlock(&c->bucket_lock);
 721         return freed;
 722 }
 723
 724 static unsigned long bch_mca_count(struct shrinker *shrink,
 725                                    struct shrink_control *sc)
 726 {
 727         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 728
 729         if (c->shrinker_disabled)
 730                 return 0;
 731
 732         if (c->btree_cache_alloc_lock)
 733                 return 0;
 734
 735         return mca_can_free(c) * c->btree_pages;
 736 }
 737
 738 void bch_btree_cache_free(struct cache_set *c)
 739 {
 740         struct btree *b;
 741         struct closure cl;
 742         closure_init_stack(&cl);
 743
 744         if (c->shrink.list.next)
 745                 unregister_shrinker(&c->shrink);
 746
 747         mutex_lock(&c->bucket_lock);
 748
 749 #ifdef CONFIG_BCACHE_DEBUG
 750         if (c->verify_data)
 751                 list_move(&c->verify_data->list, &c->btree_cache);
 752
 753         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 754 #endif
 755
 756         list_splice(&c->btree_cache_freeable,
 757                     &c->btree_cache);
 758
 759         while (!list_empty(&c->btree_cache)) {
 760                 b = list_first_entry(&c->btree_cache, struct btree, list);
 761
 762                 if (btree_node_dirty(b))
 763                         btree_complete_write(b, btree_current_write(b));
 764                 clear_bit(BTREE_NODE_dirty, &b->flags);
 765
 766                 mca_data_free(b);
 767         }
 768
 769         while (!list_empty(&c->btree_cache_freed)) {
 770                 b = list_first_entry(&c->btree_cache_freed,
 771                                      struct btree, list);
 772                 list_del(&b->list);
 773                 cancel_delayed_work_sync(&b->work);
 774                 kfree(b);
 775         }
 776
 777         mutex_unlock(&c->bucket_lock);
 778 }
 779
 780 int bch_btree_cache_alloc(struct cache_set *c)
 781 {
 782         unsigned i;
 783
 784         for (i = 0; i < mca_reserve(c); i++)
 785                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 786                         return -ENOMEM;
 787
 788         list_splice_init(&c->btree_cache,
 789                          &c->btree_cache_freeable);
 790
 791 #ifdef CONFIG_BCACHE_DEBUG
 792         mutex_init(&c->verify_lock);
 793
 794         c->verify_ondisk = (void *)
 795                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 796
 797         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 798
 799         if (c->verify_data &&
 800             c->verify_data->keys.set->data)
 801                 list_del_init(&c->verify_data->list);
 802         else
 803                 c->verify_data = NULL;
 804 #endif
 805
 806         c->shrink.count_objects = bch_mca_count;
 807         c->shrink.scan_objects = bch_mca_scan;
 808         c->shrink.seeks = 4;
 809         c->shrink.batch = c->btree_pages * 2;
 810         register_shrinker(&c->shrink);
 811
 812         return 0;
 813 }
 814
 815 /* Btree in memory cache - hash table */
 816
 817 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 818 {
 819         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 820 }
 821
 822 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 823 {
 824         struct btree *b;
 825
 826         rcu_read_lock();
 827         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 828                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 829                         goto out;
 830         b = NULL;
 831 out:
 832         rcu_read_unlock();
 833         return b;
 834 }
 835
 836 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 837 {
 838         struct task_struct *old;
 839
 840         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 841         if (old && old != current) {
 842                 if (op)
 843                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 844                                         TASK_UNINTERRUPTIBLE);
 845                 return -EINTR;
 846         }
 847
 848         return 0;
 849 }
 850
 851 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 852                                      struct bkey *k)
 853 {
 854         struct btree *b;
 855
 856         trace_bcache_btree_cache_cannibalize(c);
 857
 858         if (mca_cannibalize_lock(c, op))
 859                 return ERR_PTR(-EINTR);
 860
 861         list_for_each_entry_reverse(b, &c->btree_cache, list)
 862                 if (!mca_reap(b, btree_order(k), false))
 863                         return b;
 864
 865         list_for_each_entry_reverse(b, &c->btree_cache, list)
 866                 if (!mca_reap(b, btree_order(k), true))
 867                         return b;
 868
 869         WARN(1, "btree cache cannibalize failed\n");
 870         return ERR_PTR(-ENOMEM);
 871 }
 872
 873 /*
 874  * We can only have one thread cannibalizing other cached btree nodes at a time,
 875  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 876  * cannibalize_bucket() will take. This means every time we unlock the root of
 877  * the btree, we need to release this lock if we have it held.
 878  */
 879 static void bch_cannibalize_unlock(struct cache_set *c)
 880 {
 881         if (c->btree_cache_alloc_lock == current) {
 882                 c->btree_cache_alloc_lock = NULL;
 883                 wake_up(&c->btree_cache_wait);
 884         }
 885 }
 886
 887 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 888                                struct bkey *k, int level)
 889 {
 890         struct btree *b;
 891
 892         BUG_ON(current->bio_list);
 893
 894         lockdep_assert_held(&c->bucket_lock);
 895
 896         if (mca_find(c, k))
 897                 return NULL;
 898
 899         /* btree_free() doesn't free memory; it sticks the node on the end of
 900          * the list. Check if there's any freed nodes there:
 901          */
 902         list_for_each_entry(b, &c->btree_cache_freeable, list)
 903                 if (!mca_reap(b, btree_order(k), false))
 904                         goto out;
 905
 906         /* We never free struct btree itself, just the memory that holds the on
 907          * disk node. Check the freed list before allocating a new one:
 908          */
 909         list_for_each_entry(b, &c->btree_cache_freed, list)
 910                 if (!mca_reap(b, 0, false)) {
 911                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 912                         if (!b->keys.set[0].data)
 913                                 goto err;
 914                         else
 915                                 goto out;
 916                 }
 917
 918         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 919         if (!b)
 920                 goto err;
 921
 922         BUG_ON(!down_write_trylock(&b->lock));
 923         if (!b->keys.set->data)
 924                 goto err;
 925 out:
 926         BUG_ON(b->io_mutex.count != 1);
 927
 928         bkey_copy(&b->key, k);
 929         list_move(&b->list, &c->btree_cache);
 930         hlist_del_init_rcu(&b->hash);
 931         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 932
 933         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 934         b->parent       = (void *) ~0UL;
 935         b->flags        = 0;
 936         b->written      = 0;
 937         b->level        = level;
 938
 939         if (!b->level)
 940                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 941                                     &b->c->expensive_debug_checks);
 942         else
 943                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 944                                     &b->c->expensive_debug_checks);
 945
 946         return b;
 947 err:
 948         if (b)
 949                 rw_unlock(true, b);
 950
 951         b = mca_cannibalize(c, op, k);
 952         if (!IS_ERR(b))
 953                 goto out;
 954
 955         return b;
 956 }
 957
 958 /**
 959  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 960  * in from disk if necessary.
 961  *
 962  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 963  *
 964  * The btree node will have either a read or a write lock held, depending on
 965  * level and op->lock.
 966  */
 967 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 968                                  struct bkey *k, int level, bool write,
 969                                  struct btree *parent)
 970 {
 971         int i = 0;
 972         struct btree *b;
 973
 974         BUG_ON(level < 0);
 975 retry:
 976         b = mca_find(c, k);
 977
 978         if (!b) {
 979                 if (current->bio_list)
 980                         return ERR_PTR(-EAGAIN);
 981
 982                 mutex_lock(&c->bucket_lock);
 983                 b = mca_alloc(c, op, k, level);
 984                 mutex_unlock(&c->bucket_lock);
 985
 986                 if (!b)
 987                         goto retry;
 988                 if (IS_ERR(b))
 989                         return b;
 990
 991                 bch_btree_node_read(b);
 992
 993                 if (!write)
 994                         downgrade_write(&b->lock);
 995         } else {
 996                 rw_lock(write, b, level);
 997                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 998                         rw_unlock(write, b);
 999                         goto retry;
1000                 }
1001                 BUG_ON(b->level != level);
1002         }
1003
1004         b->parent = parent;
1005         b->accessed = 1;
1006
1007         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1008                 prefetch(b->keys.set[i].tree);
1009                 prefetch(b->keys.set[i].data);
1010         }
1011
1012         for (; i <= b->keys.nsets; i++)
1013                 prefetch(b->keys.set[i].data);
1014
1015         if (btree_node_io_error(b)) {
1016                 rw_unlock(write, b);
1017                 return ERR_PTR(-EIO);
1018         }
1019
1020         BUG_ON(!b->written);
1021
1022         return b;
1023 }
1024
1025 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1026 {
1027         struct btree *b;
1028
1029         mutex_lock(&parent->c->bucket_lock);
1030         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1031         mutex_unlock(&parent->c->bucket_lock);
1032
1033         if (!IS_ERR_OR_NULL(b)) {
1034                 b->parent = parent;
1035                 bch_btree_node_read(b);
1036                 rw_unlock(true, b);
1037         }
1038 }
1039
1040 /* Btree alloc */
1041
1042 static void btree_node_free(struct btree *b)
1043 {
1044         trace_bcache_btree_node_free(b);
1045
1046         BUG_ON(b == b->c->root);
1047
1048         mutex_lock(&b->write_lock);
1049
1050         if (btree_node_dirty(b))
1051                 btree_complete_write(b, btree_current_write(b));
1052         clear_bit(BTREE_NODE_dirty, &b->flags);
1053
1054         mutex_unlock(&b->write_lock);
1055
1056         cancel_delayed_work(&b->work);
1057
1058         mutex_lock(&b->c->bucket_lock);
1059         bch_bucket_free(b->c, &b->key);
1060         mca_bucket_free(b);
1061         mutex_unlock(&b->c->bucket_lock);
1062 }
1063
1064 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1065                                      int level, bool wait,
1066                                      struct btree *parent)
1067 {
1068         BKEY_PADDED(key) k;
1069         struct btree *b = ERR_PTR(-EAGAIN);
1070
1071         mutex_lock(&c->bucket_lock);
1072 retry:
1073         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1074                 goto err;
1075
1076         bkey_put(c, &k.key);
1077         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1078
1079         b = mca_alloc(c, op, &k.key, level);
1080         if (IS_ERR(b))
1081                 goto err_free;
1082
1083         if (!b) {
1084                 cache_bug(c,
1085                         "Tried to allocate bucket that was in btree cache");
1086                 goto retry;
1087         }
1088
1089         b->accessed = 1;
1090         b->parent = parent;
1091         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1092
1093         mutex_unlock(&c->bucket_lock);
1094
1095         trace_bcache_btree_node_alloc(b);
1096         return b;
1097 err_free:
1098         bch_bucket_free(c, &k.key);
1099 err:
1100         mutex_unlock(&c->bucket_lock);
1101
1102         trace_bcache_btree_node_alloc_fail(c);
1103         return b;
1104 }
1105
1106 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1107                                           struct btree_op *op, int level,
1108                                           struct btree *parent)
1109 {
1110         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1111 }
1112
1113 static struct btree *btree_node_alloc_replacement(struct btree *b,
1114                                                   struct btree_op *op)
1115 {
1116         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1117         if (!IS_ERR_OR_NULL(n)) {
1118                 mutex_lock(&n->write_lock);
1119                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1120                 bkey_copy_key(&n->key, &b->key);
1121                 mutex_unlock(&n->write_lock);
1122         }
1123
1124         return n;
1125 }
1126
1127 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1128 {
1129         unsigned i;
1130
1131         mutex_lock(&b->c->bucket_lock);
1132
1133         atomic_inc(&b->c->prio_blocked);
1134
1135         bkey_copy(k, &b->key);
1136         bkey_copy_key(k, &ZERO_KEY);
1137
1138         for (i = 0; i < KEY_PTRS(k); i++)
1139                 SET_PTR_GEN(k, i,
1140                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1141                                         PTR_BUCKET(b->c, &b->key, i)));
1142
1143         mutex_unlock(&b->c->bucket_lock);
1144 }
1145
1146 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1147 {
1148         struct cache_set *c = b->c;
1149         struct cache *ca;
1150         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1151
1152         mutex_lock(&c->bucket_lock);
1153
1154         for_each_cache(ca, c, i)
1155                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1156                         if (op)
1157                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1158                                                 TASK_UNINTERRUPTIBLE);
1159                         mutex_unlock(&c->bucket_lock);
1160                         return -EINTR;
1161                 }
1162
1163         mutex_unlock(&c->bucket_lock);
1164
1165         return mca_cannibalize_lock(b->c, op);
1166 }
1167
1168 /* Garbage collection */
1169
1170 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1171                                     struct bkey *k)
1172 {
1173         uint8_t stale = 0;
1174         unsigned i;
1175         struct bucket *g;
1176
1177         /*
1178          * ptr_invalid() can't return true for the keys that mark btree nodes as
1179          * freed, but since ptr_bad() returns true we'll never actually use them
1180          * for anything and thus we don't want mark their pointers here
1181          */
1182         if (!bkey_cmp(k, &ZERO_KEY))
1183                 return stale;
1184
1185         for (i = 0; i < KEY_PTRS(k); i++) {
1186                 if (!ptr_available(c, k, i))
1187                         continue;
1188
1189                 g = PTR_BUCKET(c, k, i);
1190
1191                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1192                         g->last_gc = PTR_GEN(k, i);
1193
1194                 if (ptr_stale(c, k, i)) {
1195                         stale = max(stale, ptr_stale(c, k, i));
1196                         continue;
1197                 }
1198
1199                 cache_bug_on(GC_MARK(g) &&
1200                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1201                              c, "inconsistent ptrs: mark = %llu, level = %i",
1202                              GC_MARK(g), level);
1203
1204                 if (level)
1205                         SET_GC_MARK(g, GC_MARK_METADATA);
1206                 else if (KEY_DIRTY(k))
1207                         SET_GC_MARK(g, GC_MARK_DIRTY);
1208                 else if (!GC_MARK(g))
1209                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1210
1211                 /* guard against overflow */
1212                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1213                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1214                                              MAX_GC_SECTORS_USED));
1215
1216                 BUG_ON(!GC_SECTORS_USED(g));
1217         }
1218
1219         return stale;
1220 }
1221
1222 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1223
1224 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1225 {
1226         unsigned i;
1227
1228         for (i = 0; i < KEY_PTRS(k); i++)
1229                 if (ptr_available(c, k, i) &&
1230                     !ptr_stale(c, k, i)) {
1231                         struct bucket *b = PTR_BUCKET(c, k, i);
1232
1233                         b->gen = PTR_GEN(k, i);
1234
1235                         if (level && bkey_cmp(k, &ZERO_KEY))
1236                                 b->prio = BTREE_PRIO;
1237                         else if (!level && b->prio == BTREE_PRIO)
1238                                 b->prio = INITIAL_PRIO;
1239                 }
1240
1241         __bch_btree_mark_key(c, level, k);
1242 }
1243
1244 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1245 {
1246         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1247 }
1248
1249 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1250 {
1251         uint8_t stale = 0;
1252         unsigned keys = 0, good_keys = 0;
1253         struct bkey *k;
1254         struct btree_iter iter;
1255         struct bset_tree *t;
1256
1257         gc->nodes++;
1258
1259         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1260                 stale = max(stale, btree_mark_key(b, k));
1261                 keys++;
1262
1263                 if (bch_ptr_bad(&b->keys, k))
1264                         continue;
1265
1266                 gc->key_bytes += bkey_u64s(k);
1267                 gc->nkeys++;
1268                 good_keys++;
1269
1270                 gc->data += KEY_SIZE(k);
1271         }
1272
1273         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1274                 btree_bug_on(t->size &&
1275                              bset_written(&b->keys, t) &&
1276                              bkey_cmp(&b->key, &t->end) < 0,
1277                              b, "found short btree key in gc");
1278
1279         if (b->c->gc_always_rewrite)
1280                 return true;
1281
1282         if (stale > 10)
1283                 return true;
1284
1285         if ((keys - good_keys) * 2 > keys)
1286                 return true;
1287
1288         return false;
1289 }
1290
1291 #define GC_MERGE_NODES  4U
1292
1293 struct gc_merge_info {
1294         struct btree    *b;
1295         unsigned        keys;
1296 };
1297
1298 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1299                                  struct keylist *, atomic_t *, struct bkey *);
1300
1301 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1302                              struct gc_stat *gc, struct gc_merge_info *r)
1303 {
1304         unsigned i, nodes = 0, keys = 0, blocks;
1305         struct btree *new_nodes[GC_MERGE_NODES];
1306         struct keylist keylist;
1307         struct closure cl;
1308         struct bkey *k;
1309
1310         bch_keylist_init(&keylist);
1311
1312         if (btree_check_reserve(b, NULL))
1313                 return 0;
1314
1315         memset(new_nodes, 0, sizeof(new_nodes));
1316         closure_init_stack(&cl);
1317
1318         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1319                 keys += r[nodes++].keys;
1320
1321         blocks = btree_default_blocks(b->c) * 2 / 3;
1322
1323         if (nodes < 2 ||
1324             __set_blocks(b->keys.set[0].data, keys,
1325                          block_bytes(b->c)) > blocks * (nodes - 1))
1326                 return 0;
1327
1328         for (i = 0; i < nodes; i++) {
1329                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1330                 if (IS_ERR_OR_NULL(new_nodes[i]))
1331                         goto out_nocoalesce;
1332         }
1333
1334         /*
1335          * We have to check the reserve here, after we've allocated our new
1336          * nodes, to make sure the insert below will succeed - we also check
1337          * before as an optimization to potentially avoid a bunch of expensive
1338          * allocs/sorts
1339          */
1340         if (btree_check_reserve(b, NULL))
1341                 goto out_nocoalesce;
1342
1343         for (i = 0; i < nodes; i++)
1344                 mutex_lock(&new_nodes[i]->write_lock);
1345
1346         for (i = nodes - 1; i > 0; --i) {
1347                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1348                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1349                 struct bkey *k, *last = NULL;
1350
1351                 keys = 0;
1352
1353                 if (i > 1) {
1354                         for (k = n2->start;
1355                              k < bset_bkey_last(n2);
1356                              k = bkey_next(k)) {
1357                                 if (__set_blocks(n1, n1->keys + keys +
1358                                                  bkey_u64s(k),
1359                                                  block_bytes(b->c)) > blocks)
1360                                         break;
1361
1362                                 last = k;
1363                                 keys += bkey_u64s(k);
1364                         }
1365                 } else {
1366                         /*
1367                          * Last node we're not getting rid of - we're getting
1368                          * rid of the node at r[0]. Have to try and fit all of
1369                          * the remaining keys into this node; we can't ensure
1370                          * they will always fit due to rounding and variable
1371                          * length keys (shouldn't be possible in practice,
1372                          * though)
1373                          */
1374                         if (__set_blocks(n1, n1->keys + n2->keys,
1375                                          block_bytes(b->c)) >
1376                             btree_blocks(new_nodes[i]))
1377                                 goto out_nocoalesce;
1378
1379                         keys = n2->keys;
1380                         /* Take the key of the node we're getting rid of */
1381                         last = &r->b->key;
1382                 }
1383
1384                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1385                        btree_blocks(new_nodes[i]));
1386
1387                 if (last)
1388                         bkey_copy_key(&new_nodes[i]->key, last);
1389
1390                 memcpy(bset_bkey_last(n1),
1391                        n2->start,
1392                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1393
1394                 n1->keys += keys;
1395                 r[i].keys = n1->keys;
1396
1397                 memmove(n2->start,
1398                         bset_bkey_idx(n2, keys),
1399                         (void *) bset_bkey_last(n2) -
1400                         (void *) bset_bkey_idx(n2, keys));
1401
1402                 n2->keys -= keys;
1403
1404                 if (__bch_keylist_realloc(&keylist,
1405                                           bkey_u64s(&new_nodes[i]->key)))
1406                         goto out_nocoalesce;
1407
1408                 bch_btree_node_write(new_nodes[i], &cl);
1409                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1410         }
1411
1412         for (i = 0; i < nodes; i++)
1413                 mutex_unlock(&new_nodes[i]->write_lock);
1414
1415         closure_sync(&cl);
1416
1417         /* We emptied out this node */
1418         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1419         btree_node_free(new_nodes[0]);
1420         rw_unlock(true, new_nodes[0]);
1421         new_nodes[0] = NULL;
1422
1423         for (i = 0; i < nodes; i++) {
1424                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1425                         goto out_nocoalesce;
1426
1427                 make_btree_freeing_key(r[i].b, keylist.top);
1428                 bch_keylist_push(&keylist);
1429         }
1430
1431         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1432         BUG_ON(!bch_keylist_empty(&keylist));
1433
1434         for (i = 0; i < nodes; i++) {
1435                 btree_node_free(r[i].b);
1436                 rw_unlock(true, r[i].b);
1437
1438                 r[i].b = new_nodes[i];
1439         }
1440
1441         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1442         r[nodes - 1].b = ERR_PTR(-EINTR);
1443
1444         trace_bcache_btree_gc_coalesce(nodes);
1445         gc->nodes--;
1446
1447         bch_keylist_free(&keylist);
1448
1449         /* Invalidated our iterator */
1450         return -EINTR;
1451
1452 out_nocoalesce:
1453         closure_sync(&cl);
1454         bch_keylist_free(&keylist);
1455
1456         while ((k = bch_keylist_pop(&keylist)))
1457                 if (!bkey_cmp(k, &ZERO_KEY))
1458                         atomic_dec(&b->c->prio_blocked);
1459
1460         for (i = 0; i < nodes; i++)
1461                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1462                         btree_node_free(new_nodes[i]);
1463                         rw_unlock(true, new_nodes[i]);
1464                 }
1465         return 0;
1466 }
1467
1468 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1469                                  struct btree *replace)
1470 {
1471         struct keylist keys;
1472         struct btree *n;
1473
1474         if (btree_check_reserve(b, NULL))
1475                 return 0;
1476
1477         n = btree_node_alloc_replacement(replace, NULL);
1478
1479         /* recheck reserve after allocating replacement node */
1480         if (btree_check_reserve(b, NULL)) {
1481                 btree_node_free(n);
1482                 rw_unlock(true, n);
1483                 return 0;
1484         }
1485
1486         bch_btree_node_write_sync(n);
1487
1488         bch_keylist_init(&keys);
1489         bch_keylist_add(&keys, &n->key);
1490
1491         make_btree_freeing_key(replace, keys.top);
1492         bch_keylist_push(&keys);
1493
1494         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1495         BUG_ON(!bch_keylist_empty(&keys));
1496
1497         btree_node_free(replace);
1498         rw_unlock(true, n);
1499
1500         /* Invalidated our iterator */
1501         return -EINTR;
1502 }
1503
1504 static unsigned btree_gc_count_keys(struct btree *b)
1505 {
1506         struct bkey *k;
1507         struct btree_iter iter;
1508         unsigned ret = 0;
1509
1510         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1511                 ret += bkey_u64s(k);
1512
1513         return ret;
1514 }
1515
1516 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1517                             struct closure *writes, struct gc_stat *gc)
1518 {
1519         int ret = 0;
1520         bool should_rewrite;
1521         struct bkey *k;
1522         struct btree_iter iter;
1523         struct gc_merge_info r[GC_MERGE_NODES];
1524         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1525
1526         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1527
1528         for (i = r; i < r + ARRAY_SIZE(r); i++)
1529                 i->b = ERR_PTR(-EINTR);
1530
1531         while (1) {
1532                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1533                 if (k) {
1534                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1535                                                   true, b);
1536                         if (IS_ERR(r->b)) {
1537                                 ret = PTR_ERR(r->b);
1538                                 break;
1539                         }
1540
1541                         r->keys = btree_gc_count_keys(r->b);
1542
1543                         ret = btree_gc_coalesce(b, op, gc, r);
1544                         if (ret)
1545                                 break;
1546                 }
1547
1548                 if (!last->b)
1549                         break;
1550
1551                 if (!IS_ERR(last->b)) {
1552                         should_rewrite = btree_gc_mark_node(last->b, gc);
1553                         if (should_rewrite) {
1554                                 ret = btree_gc_rewrite_node(b, op, last->b);
1555                                 if (ret)
1556                                         break;
1557                         }
1558
1559                         if (last->b->level) {
1560                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1561                                 if (ret)
1562                                         break;
1563                         }
1564
1565                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1566
1567                         /*
1568                          * Must flush leaf nodes before gc ends, since replace
1569                          * operations aren't journalled
1570                          */
1571                         mutex_lock(&last->b->write_lock);
1572                         if (btree_node_dirty(last->b))
1573                                 bch_btree_node_write(last->b, writes);
1574                         mutex_unlock(&last->b->write_lock);
1575                         rw_unlock(true, last->b);
1576                 }
1577
1578                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1579                 r->b = NULL;
1580
1581                 if (need_resched()) {
1582                         ret = -EAGAIN;
1583                         break;
1584                 }
1585         }
1586
1587         for (i = r; i < r + ARRAY_SIZE(r); i++)
1588                 if (!IS_ERR_OR_NULL(i->b)) {
1589                         mutex_lock(&i->b->write_lock);
1590                         if (btree_node_dirty(i->b))
1591                                 bch_btree_node_write(i->b, writes);
1592                         mutex_unlock(&i->b->write_lock);
1593                         rw_unlock(true, i->b);
1594                 }
1595
1596         return ret;
1597 }
1598
1599 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1600                              struct closure *writes, struct gc_stat *gc)
1601 {
1602         struct btree *n = NULL;
1603         int ret = 0;
1604         bool should_rewrite;
1605
1606         should_rewrite = btree_gc_mark_node(b, gc);
1607         if (should_rewrite) {
1608                 n = btree_node_alloc_replacement(b, NULL);
1609
1610                 if (!IS_ERR_OR_NULL(n)) {
1611                         bch_btree_node_write_sync(n);
1612
1613                         bch_btree_set_root(n);
1614                         btree_node_free(b);
1615                         rw_unlock(true, n);
1616
1617                         return -EINTR;
1618                 }
1619         }
1620
1621         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1622
1623         if (b->level) {
1624                 ret = btree_gc_recurse(b, op, writes, gc);
1625                 if (ret)
1626                         return ret;
1627         }
1628
1629         bkey_copy_key(&b->c->gc_done, &b->key);
1630
1631         return ret;
1632 }
1633
1634 static void btree_gc_start(struct cache_set *c)
1635 {
1636         struct cache *ca;
1637         struct bucket *b;
1638         unsigned i;
1639
1640         if (!c->gc_mark_valid)
1641                 return;
1642
1643         mutex_lock(&c->bucket_lock);
1644
1645         c->gc_mark_valid = 0;
1646         c->gc_done = ZERO_KEY;
1647
1648         for_each_cache(ca, c, i)
1649                 for_each_bucket(b, ca) {
1650                         b->last_gc = b->gen;
1651                         if (!atomic_read(&b->pin)) {
1652                                 SET_GC_MARK(b, 0);
1653                                 SET_GC_SECTORS_USED(b, 0);
1654                         }
1655                 }
1656
1657         mutex_unlock(&c->bucket_lock);
1658 }
1659
1660 static void bch_btree_gc_finish(struct cache_set *c)
1661 {
1662         struct bucket *b;
1663         struct cache *ca;
1664         unsigned i;
1665
1666         mutex_lock(&c->bucket_lock);
1667
1668         set_gc_sectors(c);
1669         c->gc_mark_valid = 1;
1670         c->need_gc      = 0;
1671
1672         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1673                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1674                             GC_MARK_METADATA);
1675
1676         /* don't reclaim buckets to which writeback keys point */
1677         rcu_read_lock();
1678         for (i = 0; i < c->nr_uuids; i++) {
1679                 struct bcache_device *d = c->devices[i];
1680                 struct cached_dev *dc;
1681                 struct keybuf_key *w, *n;
1682                 unsigned j;
1683
1684                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1685                         continue;
1686                 dc = container_of(d, struct cached_dev, disk);
1687
1688                 spin_lock(&dc->writeback_keys.lock);
1689                 rbtree_postorder_for_each_entry_safe(w, n,
1690                                         &dc->writeback_keys.keys, node)
1691                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1692                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1693                                             GC_MARK_DIRTY);
1694                 spin_unlock(&dc->writeback_keys.lock);
1695         }
1696         rcu_read_unlock();
1697
1698         c->avail_nbuckets = 0;
1699         for_each_cache(ca, c, i) {
1700                 uint64_t *i;
1701
1702                 ca->invalidate_needs_gc = 0;
1703
1704                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1705                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1706
1707                 for (i = ca->prio_buckets;
1708                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1709                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1710
1711                 for_each_bucket(b, ca) {
1712                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1713
1714                         if (atomic_read(&b->pin))
1715                                 continue;
1716
1717                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1718
1719                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1720                                 c->avail_nbuckets++;
1721                 }
1722         }
1723
1724         mutex_unlock(&c->bucket_lock);
1725 }
1726
1727 static void bch_btree_gc(struct cache_set *c)
1728 {
1729         int ret;
1730         struct gc_stat stats;
1731         struct closure writes;
1732         struct btree_op op;
1733         uint64_t start_time = local_clock();
1734
1735         trace_bcache_gc_start(c);
1736
1737         memset(&stats, 0, sizeof(struct gc_stat));
1738         closure_init_stack(&writes);
1739         bch_btree_op_init(&op, SHRT_MAX);
1740
1741         btree_gc_start(c);
1742
1743         do {
1744                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1745                 closure_sync(&writes);
1746                 cond_resched();
1747
1748                 if (ret && ret != -EAGAIN)
1749                         pr_warn("gc failed!");
1750         } while (ret);
1751
1752         bch_btree_gc_finish(c);
1753         wake_up_allocators(c);
1754
1755         bch_time_stats_update(&c->btree_gc_time, start_time);
1756
1757         stats.key_bytes *= sizeof(uint64_t);
1758         stats.data      <<= 9;
1759         bch_update_bucket_in_use(c, &stats);
1760         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1761
1762         trace_bcache_gc_end(c);
1763
1764         bch_moving_gc(c);
1765 }
1766
1767 static bool gc_should_run(struct cache_set *c)
1768 {
1769         struct cache *ca;
1770         unsigned i;
1771
1772         for_each_cache(ca, c, i)
1773                 if (ca->invalidate_needs_gc)
1774                         return true;
1775
1776         if (atomic_read(&c->sectors_to_gc) < 0)
1777                 return true;
1778
1779         return false;
1780 }
1781
1782 static int bch_gc_thread(void *arg)
1783 {
1784         struct cache_set *c = arg;
1785
1786         while (1) {
1787                 wait_event_interruptible(c->gc_wait,
1788                            kthread_should_stop() || gc_should_run(c));
1789
1790                 if (kthread_should_stop())
1791                         break;
1792
1793                 set_gc_sectors(c);
1794                 bch_btree_gc(c);
1795         }
1796
1797         return 0;
1798 }
1799
1800 int bch_gc_thread_start(struct cache_set *c)
1801 {
1802         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1803         if (IS_ERR(c->gc_thread))
1804                 return PTR_ERR(c->gc_thread);
1805
1806         return 0;
1807 }
1808
1809 /* Initial partial gc */
1810
1811 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1812 {
1813         int ret = 0;
1814         struct bkey *k, *p = NULL;
1815         struct btree_iter iter;
1816
1817         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1818                 bch_initial_mark_key(b->c, b->level, k);
1819
1820         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1821
1822         if (b->level) {
1823                 bch_btree_iter_init(&b->keys, &iter, NULL);
1824
1825                 do {
1826                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1827                                                        bch_ptr_bad);
1828                         if (k)
1829                                 btree_node_prefetch(b, k);
1830
1831                         if (p)
1832                                 ret = btree(check_recurse, p, b, op);
1833
1834                         p = k;
1835                 } while (p && !ret);
1836         }
1837
1838         return ret;
1839 }
1840
1841 int bch_btree_check(struct cache_set *c)
1842 {
1843         struct btree_op op;
1844
1845         bch_btree_op_init(&op, SHRT_MAX);
1846
1847         return btree_root(check_recurse, c, &op);
1848 }
1849
1850 void bch_initial_gc_finish(struct cache_set *c)
1851 {
1852         struct cache *ca;
1853         struct bucket *b;
1854         unsigned i;
1855
1856         bch_btree_gc_finish(c);
1857
1858         mutex_lock(&c->bucket_lock);
1859
1860         /*
1861          * We need to put some unused buckets directly on the prio freelist in
1862          * order to get the allocator thread started - it needs freed buckets in
1863          * order to rewrite the prios and gens, and it needs to rewrite prios
1864          * and gens in order to free buckets.
1865          *
1866          * This is only safe for buckets that have no live data in them, which
1867          * there should always be some of.
1868          */
1869         for_each_cache(ca, c, i) {
1870                 for_each_bucket(b, ca) {
1871                         if (fifo_full(&ca->free[RESERVE_PRIO]))
1872                                 break;
1873
1874                         if (bch_can_invalidate_bucket(ca, b) &&
1875                             !GC_MARK(b)) {
1876                                 __bch_invalidate_one_bucket(ca, b);
1877                                 fifo_push(&ca->free[RESERVE_PRIO],
1878                                           b - ca->buckets);
1879                         }
1880                 }
1881         }
1882
1883         mutex_unlock(&c->bucket_lock);
1884 }
1885
1886 /* Btree insertion */
1887
1888 static bool btree_insert_key(struct btree *b, struct bkey *k,
1889                              struct bkey *replace_key)
1890 {
1891         unsigned status;
1892
1893         BUG_ON(bkey_cmp(k, &b->key) > 0);
1894
1895         status = bch_btree_insert_key(&b->keys, k, replace_key);
1896         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1897                 bch_check_keys(&b->keys, "%u for %s", status,
1898                                replace_key ? "replace" : "insert");
1899
1900                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1901                                               status);
1902                 return true;
1903         } else
1904                 return false;
1905 }
1906
1907 static size_t insert_u64s_remaining(struct btree *b)
1908 {
1909         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1910
1911         /*
1912          * Might land in the middle of an existing extent and have to split it
1913          */
1914         if (b->keys.ops->is_extents)
1915                 ret -= KEY_MAX_U64S;
1916
1917         return max(ret, 0L);
1918 }
1919
1920 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1921                                   struct keylist *insert_keys,
1922                                   struct bkey *replace_key)
1923 {
1924         bool ret = false;
1925         int oldsize = bch_count_data(&b->keys);
1926
1927         while (!bch_keylist_empty(insert_keys)) {
1928                 struct bkey *k = insert_keys->keys;
1929
1930                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1931                         break;
1932
1933                 if (bkey_cmp(k, &b->key) <= 0) {
1934                         if (!b->level)
1935                                 bkey_put(b->c, k);
1936
1937                         ret |= btree_insert_key(b, k, replace_key);
1938                         bch_keylist_pop_front(insert_keys);
1939                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1940                         BKEY_PADDED(key) temp;
1941                         bkey_copy(&temp.key, insert_keys->keys);
1942
1943                         bch_cut_back(&b->key, &temp.key);
1944                         bch_cut_front(&b->key, insert_keys->keys);
1945
1946                         ret |= btree_insert_key(b, &temp.key, replace_key);
1947                         break;
1948                 } else {
1949                         break;
1950                 }
1951         }
1952
1953         if (!ret)
1954                 op->insert_collision = true;
1955
1956         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1957
1958         BUG_ON(bch_count_data(&b->keys) < oldsize);
1959         return ret;
1960 }
1961
1962 static int btree_split(struct btree *b, struct btree_op *op,
1963                        struct keylist *insert_keys,
1964                        struct bkey *replace_key)
1965 {
1966         bool split;
1967         struct btree *n1, *n2 = NULL, *n3 = NULL;
1968         uint64_t start_time = local_clock();
1969         struct closure cl;
1970         struct keylist parent_keys;
1971
1972         closure_init_stack(&cl);
1973         bch_keylist_init(&parent_keys);
1974
1975         if (btree_check_reserve(b, op)) {
1976                 if (!b->level)
1977                         return -EINTR;
1978                 else
1979                         WARN(1, "insufficient reserve for split\n");
1980         }
1981
1982         n1 = btree_node_alloc_replacement(b, op);
1983         if (IS_ERR(n1))
1984                 goto err;
1985
1986         split = set_blocks(btree_bset_first(n1),
1987                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1988
1989         if (split) {
1990                 unsigned keys = 0;
1991
1992                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1993
1994                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1995                 if (IS_ERR(n2))
1996                         goto err_free1;
1997
1998                 if (!b->parent) {
1999                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2000                         if (IS_ERR(n3))
2001                                 goto err_free2;
2002                 }
2003
2004                 mutex_lock(&n1->write_lock);
2005                 mutex_lock(&n2->write_lock);
2006
2007                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2008
2009                 /*
2010                  * Has to be a linear search because we don't have an auxiliary
2011                  * search tree yet
2012                  */
2013
2014                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2015                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2016                                                         keys));
2017
2018                 bkey_copy_key(&n1->key,
2019                               bset_bkey_idx(btree_bset_first(n1), keys));
2020                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2021
2022                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2023                 btree_bset_first(n1)->keys = keys;
2024
2025                 memcpy(btree_bset_first(n2)->start,
2026                        bset_bkey_last(btree_bset_first(n1)),
2027                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2028
2029                 bkey_copy_key(&n2->key, &b->key);
2030
2031                 bch_keylist_add(&parent_keys, &n2->key);
2032                 bch_btree_node_write(n2, &cl);
2033                 mutex_unlock(&n2->write_lock);
2034                 rw_unlock(true, n2);
2035         } else {
2036                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2037
2038                 mutex_lock(&n1->write_lock);
2039                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2040         }
2041
2042         bch_keylist_add(&parent_keys, &n1->key);
2043         bch_btree_node_write(n1, &cl);
2044         mutex_unlock(&n1->write_lock);
2045
2046         if (n3) {
2047                 /* Depth increases, make a new root */
2048                 mutex_lock(&n3->write_lock);
2049                 bkey_copy_key(&n3->key, &MAX_KEY);
2050                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2051                 bch_btree_node_write(n3, &cl);
2052                 mutex_unlock(&n3->write_lock);
2053
2054                 closure_sync(&cl);
2055                 bch_btree_set_root(n3);
2056                 rw_unlock(true, n3);
2057         } else if (!b->parent) {
2058                 /* Root filled up but didn't need to be split */
2059                 closure_sync(&cl);
2060                 bch_btree_set_root(n1);
2061         } else {
2062                 /* Split a non root node */
2063                 closure_sync(&cl);
2064                 make_btree_freeing_key(b, parent_keys.top);
2065                 bch_keylist_push(&parent_keys);
2066
2067                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2068                 BUG_ON(!bch_keylist_empty(&parent_keys));
2069         }
2070
2071         btree_node_free(b);
2072         rw_unlock(true, n1);
2073
2074         bch_time_stats_update(&b->c->btree_split_time, start_time);
2075
2076         return 0;
2077 err_free2:
2078         bkey_put(b->c, &n2->key);
2079         btree_node_free(n2);
2080         rw_unlock(true, n2);
2081 err_free1:
2082         bkey_put(b->c, &n1->key);
2083         btree_node_free(n1);
2084         rw_unlock(true, n1);
2085 err:
2086         WARN(1, "bcache: btree split failed (level %u)", b->level);
2087
2088         if (n3 == ERR_PTR(-EAGAIN) ||
2089             n2 == ERR_PTR(-EAGAIN) ||
2090             n1 == ERR_PTR(-EAGAIN))
2091                 return -EAGAIN;
2092
2093         return -ENOMEM;
2094 }
2095
2096 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2097                                  struct keylist *insert_keys,
2098                                  atomic_t *journal_ref,
2099                                  struct bkey *replace_key)
2100 {
2101         struct closure cl;
2102
2103         BUG_ON(b->level && replace_key);
2104
2105         closure_init_stack(&cl);
2106
2107         mutex_lock(&b->write_lock);
2108
2109         if (write_block(b) != btree_bset_last(b) &&
2110             b->keys.last_set_unwritten)
2111                 bch_btree_init_next(b); /* just wrote a set */
2112
2113         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2114                 mutex_unlock(&b->write_lock);
2115                 goto split;
2116         }
2117
2118         BUG_ON(write_block(b) != btree_bset_last(b));
2119
2120         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2121                 if (!b->level)
2122                         bch_btree_leaf_dirty(b, journal_ref);
2123                 else
2124                         bch_btree_node_write(b, &cl);
2125         }
2126
2127         mutex_unlock(&b->write_lock);
2128
2129         /* wait for btree node write if necessary, after unlock */
2130         closure_sync(&cl);
2131
2132         return 0;
2133 split:
2134         if (current->bio_list) {
2135                 op->lock = b->c->root->level + 1;
2136                 return -EAGAIN;
2137         } else if (op->lock <= b->c->root->level) {
2138                 op->lock = b->c->root->level + 1;
2139                 return -EINTR;
2140         } else {
2141                 /* Invalidated all iterators */
2142                 int ret = btree_split(b, op, insert_keys, replace_key);
2143
2144                 if (bch_keylist_empty(insert_keys))
2145                         return 0;
2146                 else if (!ret)
2147                         return -EINTR;
2148                 return ret;
2149         }
2150 }
2151
2152 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2153                                struct bkey *check_key)
2154 {
2155         int ret = -EINTR;
2156         uint64_t btree_ptr = b->key.ptr[0];
2157         unsigned long seq = b->seq;
2158         struct keylist insert;
2159         bool upgrade = op->lock == -1;
2160
2161         bch_keylist_init(&insert);
2162
2163         if (upgrade) {
2164                 rw_unlock(false, b);
2165                 rw_lock(true, b, b->level);
2166
2167                 if (b->key.ptr[0] != btree_ptr ||
2168                    b->seq != seq + 1) {
2169                        op->lock = b->level;
2170                         goto out;
2171                }
2172         }
2173
2174         SET_KEY_PTRS(check_key, 1);
2175         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2176
2177         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2178
2179         bch_keylist_add(&insert, check_key);
2180
2181         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2182
2183         BUG_ON(!ret && !bch_keylist_empty(&insert));
2184 out:
2185         if (upgrade)
2186                 downgrade_write(&b->lock);
2187         return ret;
2188 }
2189
2190 struct btree_insert_op {
2191         struct btree_op op;
2192         struct keylist  *keys;
2193         atomic_t        *journal_ref;
2194         struct bkey     *replace_key;
2195 };
2196
2197 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2198 {
2199         struct btree_insert_op *op = container_of(b_op,
2200                                         struct btree_insert_op, op);
2201
2202         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2203                                         op->journal_ref, op->replace_key);
2204         if (ret && !bch_keylist_empty(op->keys))
2205                 return ret;
2206         else
2207                 return MAP_DONE;
2208 }
2209
2210 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2211                      atomic_t *journal_ref, struct bkey *replace_key)
2212 {
2213         struct btree_insert_op op;
2214         int ret = 0;
2215
2216         BUG_ON(current->bio_list);
2217         BUG_ON(bch_keylist_empty(keys));
2218
2219         bch_btree_op_init(&op.op, 0);
2220         op.keys         = keys;
2221         op.journal_ref  = journal_ref;
2222         op.replace_key  = replace_key;
2223
2224         while (!ret && !bch_keylist_empty(keys)) {
2225                 op.op.lock = 0;
2226                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2227                                                &START_KEY(keys->keys),
2228                                                btree_insert_fn);
2229         }
2230
2231         if (ret) {
2232                 struct bkey *k;
2233
2234                 pr_err("error %i", ret);
2235
2236                 while ((k = bch_keylist_pop(keys)))
2237                         bkey_put(c, k);
2238         } else if (op.op.insert_collision)
2239                 ret = -ESRCH;
2240
2241         return ret;
2242 }
2243
2244 void bch_btree_set_root(struct btree *b)
2245 {
2246         unsigned i;
2247         struct closure cl;
2248
2249         closure_init_stack(&cl);
2250
2251         trace_bcache_btree_set_root(b);
2252
2253         BUG_ON(!b->written);
2254
2255         for (i = 0; i < KEY_PTRS(&b->key); i++)
2256                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2257
2258         mutex_lock(&b->c->bucket_lock);
2259         list_del_init(&b->list);
2260         mutex_unlock(&b->c->bucket_lock);
2261
2262         b->c->root = b;
2263
2264         bch_journal_meta(b->c, &cl);
2265         closure_sync(&cl);
2266 }
2267
2268 /* Map across nodes or keys */
2269
2270 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2271                                        struct bkey *from,
2272                                        btree_map_nodes_fn *fn, int flags)
2273 {
2274         int ret = MAP_CONTINUE;
2275
2276         if (b->level) {
2277                 struct bkey *k;
2278                 struct btree_iter iter;
2279
2280                 bch_btree_iter_init(&b->keys, &iter, from);
2281
2282                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2283                                                        bch_ptr_bad))) {
2284                         ret = btree(map_nodes_recurse, k, b,
2285                                     op, from, fn, flags);
2286                         from = NULL;
2287
2288                         if (ret != MAP_CONTINUE)
2289                                 return ret;
2290                 }
2291         }
2292
2293         if (!b->level || flags == MAP_ALL_NODES)
2294                 ret = fn(op, b);
2295
2296         return ret;
2297 }
2298
2299 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2300                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2301 {
2302         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2303 }
2304
2305 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2306                                       struct bkey *from, btree_map_keys_fn *fn,
2307                                       int flags)
2308 {
2309         int ret = MAP_CONTINUE;
2310         struct bkey *k;
2311         struct btree_iter iter;
2312
2313         bch_btree_iter_init(&b->keys, &iter, from);
2314
2315         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2316                 ret = !b->level
2317                         ? fn(op, b, k)
2318                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2319                 from = NULL;
2320
2321                 if (ret != MAP_CONTINUE)
2322                         return ret;
2323         }
2324
2325         if (!b->level && (flags & MAP_END_KEY))
2326                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2327                                      KEY_OFFSET(&b->key), 0));
2328
2329         return ret;
2330 }
2331
2332 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2333                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2334 {
2335         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2336 }
2337
2338 /* Keybuf code */
2339
2340 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2341 {
2342         /* Overlapping keys compare equal */
2343         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2344                 return -1;
2345         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2346                 return 1;
2347         return 0;
2348 }
2349
2350 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2351                                             struct keybuf_key *r)
2352 {
2353         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2354 }
2355
2356 struct refill {
2357         struct btree_op op;
2358         unsigned        nr_found;
2359         struct keybuf   *buf;
2360         struct bkey     *end;
2361         keybuf_pred_fn  *pred;
2362 };
2363
2364 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2365                             struct bkey *k)
2366 {
2367         struct refill *refill = container_of(op, struct refill, op);
2368         struct keybuf *buf = refill->buf;
2369         int ret = MAP_CONTINUE;
2370
2371         if (bkey_cmp(k, refill->end) >= 0) {
2372                 ret = MAP_DONE;
2373                 goto out;
2374         }
2375
2376         if (!KEY_SIZE(k)) /* end key */
2377                 goto out;
2378
2379         if (refill->pred(buf, k)) {
2380                 struct keybuf_key *w;
2381
2382                 spin_lock(&buf->lock);
2383
2384                 w = array_alloc(&buf->freelist);
2385                 if (!w) {
2386                         spin_unlock(&buf->lock);
2387                         return MAP_DONE;
2388                 }
2389
2390                 w->private = NULL;
2391                 bkey_copy(&w->key, k);
2392
2393                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2394                         array_free(&buf->freelist, w);
2395                 else
2396                         refill->nr_found++;
2397
2398                 if (array_freelist_empty(&buf->freelist))
2399                         ret = MAP_DONE;
2400
2401                 spin_unlock(&buf->lock);
2402         }
2403 out:
2404         buf->last_scanned = *k;
2405         return ret;
2406 }
2407
2408 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2409                        struct bkey *end, keybuf_pred_fn *pred)
2410 {
2411         struct bkey start = buf->last_scanned;
2412         struct refill refill;
2413
2414         cond_resched();
2415
2416         bch_btree_op_init(&refill.op, -1);
2417         refill.nr_found = 0;
2418         refill.buf      = buf;
2419         refill.end      = end;
2420         refill.pred     = pred;
2421
2422         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2423                            refill_keybuf_fn, MAP_END_KEY);
2424
2425         trace_bcache_keyscan(refill.nr_found,
2426                              KEY_INODE(&start), KEY_OFFSET(&start),
2427                              KEY_INODE(&buf->last_scanned),
2428                              KEY_OFFSET(&buf->last_scanned));
2429
2430         spin_lock(&buf->lock);
2431
2432         if (!RB_EMPTY_ROOT(&buf->keys)) {
2433                 struct keybuf_key *w;
2434                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2435                 buf->start      = START_KEY(&w->key);
2436
2437                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2438                 buf->end        = w->key;
2439         } else {
2440                 buf->start      = MAX_KEY;
2441                 buf->end        = MAX_KEY;
2442         }
2443
2444         spin_unlock(&buf->lock);
2445 }
2446
2447 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2448 {
2449         rb_erase(&w->node, &buf->keys);
2450         array_free(&buf->freelist, w);
2451 }
2452
2453 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2454 {
2455         spin_lock(&buf->lock);
2456         __bch_keybuf_del(buf, w);
2457         spin_unlock(&buf->lock);
2458 }
2459
2460 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2461                                   struct bkey *end)
2462 {
2463         bool ret = false;
2464         struct keybuf_key *p, *w, s;
2465         s.key = *start;
2466
2467         if (bkey_cmp(end, &buf->start) <= 0 ||
2468             bkey_cmp(start, &buf->end) >= 0)
2469                 return false;
2470
2471         spin_lock(&buf->lock);
2472         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2473
2474         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2475                 p = w;
2476                 w = RB_NEXT(w, node);
2477
2478                 if (p->private)
2479                         ret = true;
2480                 else
2481                         __bch_keybuf_del(buf, p);
2482         }
2483
2484         spin_unlock(&buf->lock);
2485         return ret;
2486 }
2487
2488 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2489 {
2490         struct keybuf_key *w;
2491         spin_lock(&buf->lock);
2492
2493         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2494
2495         while (w && w->private)
2496                 w = RB_NEXT(w, node);
2497
2498         if (w)
2499                 w->private = ERR_PTR(-EINTR);
2500
2501         spin_unlock(&buf->lock);
2502         return w;
2503 }
2504
2505 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2506                                           struct keybuf *buf,
2507                                           struct bkey *end,
2508                                           keybuf_pred_fn *pred)
2509 {
2510         struct keybuf_key *ret;
2511
2512         while (1) {
2513                 ret = bch_keybuf_next(buf);
2514                 if (ret)
2515                         break;
2516
2517                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2518                         pr_debug("scan finished");
2519                         break;
2520                 }
2521
2522                 bch_refill_keybuf(c, buf, end, pred);
2523         }
2524
2525         return ret;
2526 }
2527
2528 void bch_keybuf_init(struct keybuf *buf)
2529 {
2530         buf->last_scanned       = MAX_KEY;
2531         buf->keys               = RB_ROOT;
2532
2533         spin_lock_init(&buf->lock);
2534         array_allocator_init(&buf->freelist);
2535 }