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 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1245 {
1246         uint8_t stale = 0;
1247         unsigned keys = 0, good_keys = 0;
1248         struct bkey *k;
1249         struct btree_iter iter;
1250         struct bset_tree *t;
1251
1252         gc->nodes++;
1253
1254         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1255                 stale = max(stale, btree_mark_key(b, k));
1256                 keys++;
1257
1258                 if (bch_ptr_bad(&b->keys, k))
1259                         continue;
1260
1261                 gc->key_bytes += bkey_u64s(k);
1262                 gc->nkeys++;
1263                 good_keys++;
1264
1265                 gc->data += KEY_SIZE(k);
1266         }
1267
1268         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1269                 btree_bug_on(t->size &&
1270                              bset_written(&b->keys, t) &&
1271                              bkey_cmp(&b->key, &t->end) < 0,
1272                              b, "found short btree key in gc");
1273
1274         if (b->c->gc_always_rewrite)
1275                 return true;
1276
1277         if (stale > 10)
1278                 return true;
1279
1280         if ((keys - good_keys) * 2 > keys)
1281                 return true;
1282
1283         return false;
1284 }
1285
1286 #define GC_MERGE_NODES  4U
1287
1288 struct gc_merge_info {
1289         struct btree    *b;
1290         unsigned        keys;
1291 };
1292
1293 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1294                                  struct keylist *, atomic_t *, struct bkey *);
1295
1296 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1297                              struct gc_stat *gc, struct gc_merge_info *r)
1298 {
1299         unsigned i, nodes = 0, keys = 0, blocks;
1300         struct btree *new_nodes[GC_MERGE_NODES];
1301         struct keylist keylist;
1302         struct closure cl;
1303         struct bkey *k;
1304
1305         bch_keylist_init(&keylist);
1306
1307         if (btree_check_reserve(b, NULL))
1308                 return 0;
1309
1310         memset(new_nodes, 0, sizeof(new_nodes));
1311         closure_init_stack(&cl);
1312
1313         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1314                 keys += r[nodes++].keys;
1315
1316         blocks = btree_default_blocks(b->c) * 2 / 3;
1317
1318         if (nodes < 2 ||
1319             __set_blocks(b->keys.set[0].data, keys,
1320                          block_bytes(b->c)) > blocks * (nodes - 1))
1321                 return 0;
1322
1323         for (i = 0; i < nodes; i++) {
1324                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1325                 if (IS_ERR_OR_NULL(new_nodes[i]))
1326                         goto out_nocoalesce;
1327         }
1328
1329         /*
1330          * We have to check the reserve here, after we've allocated our new
1331          * nodes, to make sure the insert below will succeed - we also check
1332          * before as an optimization to potentially avoid a bunch of expensive
1333          * allocs/sorts
1334          */
1335         if (btree_check_reserve(b, NULL))
1336                 goto out_nocoalesce;
1337
1338         for (i = 0; i < nodes; i++)
1339                 mutex_lock(&new_nodes[i]->write_lock);
1340
1341         for (i = nodes - 1; i > 0; --i) {
1342                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1343                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1344                 struct bkey *k, *last = NULL;
1345
1346                 keys = 0;
1347
1348                 if (i > 1) {
1349                         for (k = n2->start;
1350                              k < bset_bkey_last(n2);
1351                              k = bkey_next(k)) {
1352                                 if (__set_blocks(n1, n1->keys + keys +
1353                                                  bkey_u64s(k),
1354                                                  block_bytes(b->c)) > blocks)
1355                                         break;
1356
1357                                 last = k;
1358                                 keys += bkey_u64s(k);
1359                         }
1360                 } else {
1361                         /*
1362                          * Last node we're not getting rid of - we're getting
1363                          * rid of the node at r[0]. Have to try and fit all of
1364                          * the remaining keys into this node; we can't ensure
1365                          * they will always fit due to rounding and variable
1366                          * length keys (shouldn't be possible in practice,
1367                          * though)
1368                          */
1369                         if (__set_blocks(n1, n1->keys + n2->keys,
1370                                          block_bytes(b->c)) >
1371                             btree_blocks(new_nodes[i]))
1372                                 goto out_nocoalesce;
1373
1374                         keys = n2->keys;
1375                         /* Take the key of the node we're getting rid of */
1376                         last = &r->b->key;
1377                 }
1378
1379                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1380                        btree_blocks(new_nodes[i]));
1381
1382                 if (last)
1383                         bkey_copy_key(&new_nodes[i]->key, last);
1384
1385                 memcpy(bset_bkey_last(n1),
1386                        n2->start,
1387                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1388
1389                 n1->keys += keys;
1390                 r[i].keys = n1->keys;
1391
1392                 memmove(n2->start,
1393                         bset_bkey_idx(n2, keys),
1394                         (void *) bset_bkey_last(n2) -
1395                         (void *) bset_bkey_idx(n2, keys));
1396
1397                 n2->keys -= keys;
1398
1399                 if (__bch_keylist_realloc(&keylist,
1400                                           bkey_u64s(&new_nodes[i]->key)))
1401                         goto out_nocoalesce;
1402
1403                 bch_btree_node_write(new_nodes[i], &cl);
1404                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1405         }
1406
1407         for (i = 0; i < nodes; i++)
1408                 mutex_unlock(&new_nodes[i]->write_lock);
1409
1410         closure_sync(&cl);
1411
1412         /* We emptied out this node */
1413         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1414         btree_node_free(new_nodes[0]);
1415         rw_unlock(true, new_nodes[0]);
1416         new_nodes[0] = NULL;
1417
1418         for (i = 0; i < nodes; i++) {
1419                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1420                         goto out_nocoalesce;
1421
1422                 make_btree_freeing_key(r[i].b, keylist.top);
1423                 bch_keylist_push(&keylist);
1424         }
1425
1426         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1427         BUG_ON(!bch_keylist_empty(&keylist));
1428
1429         for (i = 0; i < nodes; i++) {
1430                 btree_node_free(r[i].b);
1431                 rw_unlock(true, r[i].b);
1432
1433                 r[i].b = new_nodes[i];
1434         }
1435
1436         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1437         r[nodes - 1].b = ERR_PTR(-EINTR);
1438
1439         trace_bcache_btree_gc_coalesce(nodes);
1440         gc->nodes--;
1441
1442         bch_keylist_free(&keylist);
1443
1444         /* Invalidated our iterator */
1445         return -EINTR;
1446
1447 out_nocoalesce:
1448         closure_sync(&cl);
1449         bch_keylist_free(&keylist);
1450
1451         while ((k = bch_keylist_pop(&keylist)))
1452                 if (!bkey_cmp(k, &ZERO_KEY))
1453                         atomic_dec(&b->c->prio_blocked);
1454
1455         for (i = 0; i < nodes; i++)
1456                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1457                         btree_node_free(new_nodes[i]);
1458                         rw_unlock(true, new_nodes[i]);
1459                 }
1460         return 0;
1461 }
1462
1463 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1464                                  struct btree *replace)
1465 {
1466         struct keylist keys;
1467         struct btree *n;
1468
1469         if (btree_check_reserve(b, NULL))
1470                 return 0;
1471
1472         n = btree_node_alloc_replacement(replace, NULL);
1473
1474         /* recheck reserve after allocating replacement node */
1475         if (btree_check_reserve(b, NULL)) {
1476                 btree_node_free(n);
1477                 rw_unlock(true, n);
1478                 return 0;
1479         }
1480
1481         bch_btree_node_write_sync(n);
1482
1483         bch_keylist_init(&keys);
1484         bch_keylist_add(&keys, &n->key);
1485
1486         make_btree_freeing_key(replace, keys.top);
1487         bch_keylist_push(&keys);
1488
1489         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1490         BUG_ON(!bch_keylist_empty(&keys));
1491
1492         btree_node_free(replace);
1493         rw_unlock(true, n);
1494
1495         /* Invalidated our iterator */
1496         return -EINTR;
1497 }
1498
1499 static unsigned btree_gc_count_keys(struct btree *b)
1500 {
1501         struct bkey *k;
1502         struct btree_iter iter;
1503         unsigned ret = 0;
1504
1505         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1506                 ret += bkey_u64s(k);
1507
1508         return ret;
1509 }
1510
1511 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1512                             struct closure *writes, struct gc_stat *gc)
1513 {
1514         int ret = 0;
1515         bool should_rewrite;
1516         struct bkey *k;
1517         struct btree_iter iter;
1518         struct gc_merge_info r[GC_MERGE_NODES];
1519         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1520
1521         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1522
1523         for (i = r; i < r + ARRAY_SIZE(r); i++)
1524                 i->b = ERR_PTR(-EINTR);
1525
1526         while (1) {
1527                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1528                 if (k) {
1529                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1530                                                   true, b);
1531                         if (IS_ERR(r->b)) {
1532                                 ret = PTR_ERR(r->b);
1533                                 break;
1534                         }
1535
1536                         r->keys = btree_gc_count_keys(r->b);
1537
1538                         ret = btree_gc_coalesce(b, op, gc, r);
1539                         if (ret)
1540                                 break;
1541                 }
1542
1543                 if (!last->b)
1544                         break;
1545
1546                 if (!IS_ERR(last->b)) {
1547                         should_rewrite = btree_gc_mark_node(last->b, gc);
1548                         if (should_rewrite) {
1549                                 ret = btree_gc_rewrite_node(b, op, last->b);
1550                                 if (ret)
1551                                         break;
1552                         }
1553
1554                         if (last->b->level) {
1555                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1556                                 if (ret)
1557                                         break;
1558                         }
1559
1560                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1561
1562                         /*
1563                          * Must flush leaf nodes before gc ends, since replace
1564                          * operations aren't journalled
1565                          */
1566                         mutex_lock(&last->b->write_lock);
1567                         if (btree_node_dirty(last->b))
1568                                 bch_btree_node_write(last->b, writes);
1569                         mutex_unlock(&last->b->write_lock);
1570                         rw_unlock(true, last->b);
1571                 }
1572
1573                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1574                 r->b = NULL;
1575
1576                 if (need_resched()) {
1577                         ret = -EAGAIN;
1578                         break;
1579                 }
1580         }
1581
1582         for (i = r; i < r + ARRAY_SIZE(r); i++)
1583                 if (!IS_ERR_OR_NULL(i->b)) {
1584                         mutex_lock(&i->b->write_lock);
1585                         if (btree_node_dirty(i->b))
1586                                 bch_btree_node_write(i->b, writes);
1587                         mutex_unlock(&i->b->write_lock);
1588                         rw_unlock(true, i->b);
1589                 }
1590
1591         return ret;
1592 }
1593
1594 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1595                              struct closure *writes, struct gc_stat *gc)
1596 {
1597         struct btree *n = NULL;
1598         int ret = 0;
1599         bool should_rewrite;
1600
1601         should_rewrite = btree_gc_mark_node(b, gc);
1602         if (should_rewrite) {
1603                 n = btree_node_alloc_replacement(b, NULL);
1604
1605                 if (!IS_ERR_OR_NULL(n)) {
1606                         bch_btree_node_write_sync(n);
1607
1608                         bch_btree_set_root(n);
1609                         btree_node_free(b);
1610                         rw_unlock(true, n);
1611
1612                         return -EINTR;
1613                 }
1614         }
1615
1616         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1617
1618         if (b->level) {
1619                 ret = btree_gc_recurse(b, op, writes, gc);
1620                 if (ret)
1621                         return ret;
1622         }
1623
1624         bkey_copy_key(&b->c->gc_done, &b->key);
1625
1626         return ret;
1627 }
1628
1629 static void btree_gc_start(struct cache_set *c)
1630 {
1631         struct cache *ca;
1632         struct bucket *b;
1633         unsigned i;
1634
1635         if (!c->gc_mark_valid)
1636                 return;
1637
1638         mutex_lock(&c->bucket_lock);
1639
1640         c->gc_mark_valid = 0;
1641         c->gc_done = ZERO_KEY;
1642
1643         for_each_cache(ca, c, i)
1644                 for_each_bucket(b, ca) {
1645                         b->last_gc = b->gen;
1646                         if (!atomic_read(&b->pin)) {
1647                                 SET_GC_MARK(b, 0);
1648                                 SET_GC_SECTORS_USED(b, 0);
1649                         }
1650                 }
1651
1652         mutex_unlock(&c->bucket_lock);
1653 }
1654
1655 static size_t bch_btree_gc_finish(struct cache_set *c)
1656 {
1657         size_t available = 0;
1658         struct bucket *b;
1659         struct cache *ca;
1660         unsigned i;
1661
1662         mutex_lock(&c->bucket_lock);
1663
1664         set_gc_sectors(c);
1665         c->gc_mark_valid = 1;
1666         c->need_gc      = 0;
1667
1668         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1669                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1670                             GC_MARK_METADATA);
1671
1672         /* don't reclaim buckets to which writeback keys point */
1673         rcu_read_lock();
1674         for (i = 0; i < c->nr_uuids; i++) {
1675                 struct bcache_device *d = c->devices[i];
1676                 struct cached_dev *dc;
1677                 struct keybuf_key *w, *n;
1678                 unsigned j;
1679
1680                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1681                         continue;
1682                 dc = container_of(d, struct cached_dev, disk);
1683
1684                 spin_lock(&dc->writeback_keys.lock);
1685                 rbtree_postorder_for_each_entry_safe(w, n,
1686                                         &dc->writeback_keys.keys, node)
1687                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1688                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1689                                             GC_MARK_DIRTY);
1690                 spin_unlock(&dc->writeback_keys.lock);
1691         }
1692         rcu_read_unlock();
1693
1694         for_each_cache(ca, c, i) {
1695                 uint64_t *i;
1696
1697                 ca->invalidate_needs_gc = 0;
1698
1699                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1700                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1701
1702                 for (i = ca->prio_buckets;
1703                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1704                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1705
1706                 for_each_bucket(b, ca) {
1707                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1708
1709                         if (atomic_read(&b->pin))
1710                                 continue;
1711
1712                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1713
1714                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1715                                 available++;
1716                 }
1717         }
1718
1719         mutex_unlock(&c->bucket_lock);
1720         return available;
1721 }
1722
1723 static void bch_btree_gc(struct cache_set *c)
1724 {
1725         int ret;
1726         unsigned long available;
1727         struct gc_stat stats;
1728         struct closure writes;
1729         struct btree_op op;
1730         uint64_t start_time = local_clock();
1731
1732         trace_bcache_gc_start(c);
1733
1734         memset(&stats, 0, sizeof(struct gc_stat));
1735         closure_init_stack(&writes);
1736         bch_btree_op_init(&op, SHRT_MAX);
1737
1738         btree_gc_start(c);
1739
1740         do {
1741                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1742                 closure_sync(&writes);
1743                 cond_resched();
1744
1745                 if (ret && ret != -EAGAIN)
1746                         pr_warn("gc failed!");
1747         } while (ret);
1748
1749         available = bch_btree_gc_finish(c);
1750         wake_up_allocators(c);
1751
1752         bch_time_stats_update(&c->btree_gc_time, start_time);
1753
1754         stats.key_bytes *= sizeof(uint64_t);
1755         stats.data      <<= 9;
1756         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1757         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1758
1759         trace_bcache_gc_end(c);
1760
1761         bch_moving_gc(c);
1762 }
1763
1764 static bool gc_should_run(struct cache_set *c)
1765 {
1766         struct cache *ca;
1767         unsigned i;
1768
1769         for_each_cache(ca, c, i)
1770                 if (ca->invalidate_needs_gc)
1771                         return true;
1772
1773         if (atomic_read(&c->sectors_to_gc) < 0)
1774                 return true;
1775
1776         return false;
1777 }
1778
1779 static int bch_gc_thread(void *arg)
1780 {
1781         struct cache_set *c = arg;
1782
1783         while (1) {
1784                 wait_event_interruptible(c->gc_wait,
1785                            kthread_should_stop() || gc_should_run(c));
1786
1787                 if (kthread_should_stop())
1788                         break;
1789
1790                 set_gc_sectors(c);
1791                 bch_btree_gc(c);
1792         }
1793
1794         return 0;
1795 }
1796
1797 int bch_gc_thread_start(struct cache_set *c)
1798 {
1799         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1800         if (IS_ERR(c->gc_thread))
1801                 return PTR_ERR(c->gc_thread);
1802
1803         return 0;
1804 }
1805
1806 /* Initial partial gc */
1807
1808 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1809 {
1810         int ret = 0;
1811         struct bkey *k, *p = NULL;
1812         struct btree_iter iter;
1813
1814         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1815                 bch_initial_mark_key(b->c, b->level, k);
1816
1817         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1818
1819         if (b->level) {
1820                 bch_btree_iter_init(&b->keys, &iter, NULL);
1821
1822                 do {
1823                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1824                                                        bch_ptr_bad);
1825                         if (k)
1826                                 btree_node_prefetch(b, k);
1827
1828                         if (p)
1829                                 ret = btree(check_recurse, p, b, op);
1830
1831                         p = k;
1832                 } while (p && !ret);
1833         }
1834
1835         return ret;
1836 }
1837
1838 int bch_btree_check(struct cache_set *c)
1839 {
1840         struct btree_op op;
1841
1842         bch_btree_op_init(&op, SHRT_MAX);
1843
1844         return btree_root(check_recurse, c, &op);
1845 }
1846
1847 void bch_initial_gc_finish(struct cache_set *c)
1848 {
1849         struct cache *ca;
1850         struct bucket *b;
1851         unsigned i;
1852
1853         bch_btree_gc_finish(c);
1854
1855         mutex_lock(&c->bucket_lock);
1856
1857         /*
1858          * We need to put some unused buckets directly on the prio freelist in
1859          * order to get the allocator thread started - it needs freed buckets in
1860          * order to rewrite the prios and gens, and it needs to rewrite prios
1861          * and gens in order to free buckets.
1862          *
1863          * This is only safe for buckets that have no live data in them, which
1864          * there should always be some of.
1865          */
1866         for_each_cache(ca, c, i) {
1867                 for_each_bucket(b, ca) {
1868                         if (fifo_full(&ca->free[RESERVE_PRIO]))
1869                                 break;
1870
1871                         if (bch_can_invalidate_bucket(ca, b) &&
1872                             !GC_MARK(b)) {
1873                                 __bch_invalidate_one_bucket(ca, b);
1874                                 fifo_push(&ca->free[RESERVE_PRIO],
1875                                           b - ca->buckets);
1876                         }
1877                 }
1878         }
1879
1880         mutex_unlock(&c->bucket_lock);
1881 }
1882
1883 /* Btree insertion */
1884
1885 static bool btree_insert_key(struct btree *b, struct bkey *k,
1886                              struct bkey *replace_key)
1887 {
1888         unsigned status;
1889
1890         BUG_ON(bkey_cmp(k, &b->key) > 0);
1891
1892         status = bch_btree_insert_key(&b->keys, k, replace_key);
1893         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1894                 bch_check_keys(&b->keys, "%u for %s", status,
1895                                replace_key ? "replace" : "insert");
1896
1897                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1898                                               status);
1899                 return true;
1900         } else
1901                 return false;
1902 }
1903
1904 static size_t insert_u64s_remaining(struct btree *b)
1905 {
1906         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1907
1908         /*
1909          * Might land in the middle of an existing extent and have to split it
1910          */
1911         if (b->keys.ops->is_extents)
1912                 ret -= KEY_MAX_U64S;
1913
1914         return max(ret, 0L);
1915 }
1916
1917 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1918                                   struct keylist *insert_keys,
1919                                   struct bkey *replace_key)
1920 {
1921         bool ret = false;
1922         int oldsize = bch_count_data(&b->keys);
1923
1924         while (!bch_keylist_empty(insert_keys)) {
1925                 struct bkey *k = insert_keys->keys;
1926
1927                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1928                         break;
1929
1930                 if (bkey_cmp(k, &b->key) <= 0) {
1931                         if (!b->level)
1932                                 bkey_put(b->c, k);
1933
1934                         ret |= btree_insert_key(b, k, replace_key);
1935                         bch_keylist_pop_front(insert_keys);
1936                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1937                         BKEY_PADDED(key) temp;
1938                         bkey_copy(&temp.key, insert_keys->keys);
1939
1940                         bch_cut_back(&b->key, &temp.key);
1941                         bch_cut_front(&b->key, insert_keys->keys);
1942
1943                         ret |= btree_insert_key(b, &temp.key, replace_key);
1944                         break;
1945                 } else {
1946                         break;
1947                 }
1948         }
1949
1950         if (!ret)
1951                 op->insert_collision = true;
1952
1953         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1954
1955         BUG_ON(bch_count_data(&b->keys) < oldsize);
1956         return ret;
1957 }
1958
1959 static int btree_split(struct btree *b, struct btree_op *op,
1960                        struct keylist *insert_keys,
1961                        struct bkey *replace_key)
1962 {
1963         bool split;
1964         struct btree *n1, *n2 = NULL, *n3 = NULL;
1965         uint64_t start_time = local_clock();
1966         struct closure cl;
1967         struct keylist parent_keys;
1968
1969         closure_init_stack(&cl);
1970         bch_keylist_init(&parent_keys);
1971
1972         if (btree_check_reserve(b, op)) {
1973                 if (!b->level)
1974                         return -EINTR;
1975                 else
1976                         WARN(1, "insufficient reserve for split\n");
1977         }
1978
1979         n1 = btree_node_alloc_replacement(b, op);
1980         if (IS_ERR(n1))
1981                 goto err;
1982
1983         split = set_blocks(btree_bset_first(n1),
1984                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1985
1986         if (split) {
1987                 unsigned keys = 0;
1988
1989                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1990
1991                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1992                 if (IS_ERR(n2))
1993                         goto err_free1;
1994
1995                 if (!b->parent) {
1996                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
1997                         if (IS_ERR(n3))
1998                                 goto err_free2;
1999                 }
2000
2001                 mutex_lock(&n1->write_lock);
2002                 mutex_lock(&n2->write_lock);
2003
2004                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2005
2006                 /*
2007                  * Has to be a linear search because we don't have an auxiliary
2008                  * search tree yet
2009                  */
2010
2011                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2012                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2013                                                         keys));
2014
2015                 bkey_copy_key(&n1->key,
2016                               bset_bkey_idx(btree_bset_first(n1), keys));
2017                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2018
2019                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2020                 btree_bset_first(n1)->keys = keys;
2021
2022                 memcpy(btree_bset_first(n2)->start,
2023                        bset_bkey_last(btree_bset_first(n1)),
2024                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2025
2026                 bkey_copy_key(&n2->key, &b->key);
2027
2028                 bch_keylist_add(&parent_keys, &n2->key);
2029                 bch_btree_node_write(n2, &cl);
2030                 mutex_unlock(&n2->write_lock);
2031                 rw_unlock(true, n2);
2032         } else {
2033                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2034
2035                 mutex_lock(&n1->write_lock);
2036                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2037         }
2038
2039         bch_keylist_add(&parent_keys, &n1->key);
2040         bch_btree_node_write(n1, &cl);
2041         mutex_unlock(&n1->write_lock);
2042
2043         if (n3) {
2044                 /* Depth increases, make a new root */
2045                 mutex_lock(&n3->write_lock);
2046                 bkey_copy_key(&n3->key, &MAX_KEY);
2047                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2048                 bch_btree_node_write(n3, &cl);
2049                 mutex_unlock(&n3->write_lock);
2050
2051                 closure_sync(&cl);
2052                 bch_btree_set_root(n3);
2053                 rw_unlock(true, n3);
2054         } else if (!b->parent) {
2055                 /* Root filled up but didn't need to be split */
2056                 closure_sync(&cl);
2057                 bch_btree_set_root(n1);
2058         } else {
2059                 /* Split a non root node */
2060                 closure_sync(&cl);
2061                 make_btree_freeing_key(b, parent_keys.top);
2062                 bch_keylist_push(&parent_keys);
2063
2064                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2065                 BUG_ON(!bch_keylist_empty(&parent_keys));
2066         }
2067
2068         btree_node_free(b);
2069         rw_unlock(true, n1);
2070
2071         bch_time_stats_update(&b->c->btree_split_time, start_time);
2072
2073         return 0;
2074 err_free2:
2075         bkey_put(b->c, &n2->key);
2076         btree_node_free(n2);
2077         rw_unlock(true, n2);
2078 err_free1:
2079         bkey_put(b->c, &n1->key);
2080         btree_node_free(n1);
2081         rw_unlock(true, n1);
2082 err:
2083         WARN(1, "bcache: btree split failed (level %u)", b->level);
2084
2085         if (n3 == ERR_PTR(-EAGAIN) ||
2086             n2 == ERR_PTR(-EAGAIN) ||
2087             n1 == ERR_PTR(-EAGAIN))
2088                 return -EAGAIN;
2089
2090         return -ENOMEM;
2091 }
2092
2093 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2094                                  struct keylist *insert_keys,
2095                                  atomic_t *journal_ref,
2096                                  struct bkey *replace_key)
2097 {
2098         struct closure cl;
2099
2100         BUG_ON(b->level && replace_key);
2101
2102         closure_init_stack(&cl);
2103
2104         mutex_lock(&b->write_lock);
2105
2106         if (write_block(b) != btree_bset_last(b) &&
2107             b->keys.last_set_unwritten)
2108                 bch_btree_init_next(b); /* just wrote a set */
2109
2110         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2111                 mutex_unlock(&b->write_lock);
2112                 goto split;
2113         }
2114
2115         BUG_ON(write_block(b) != btree_bset_last(b));
2116
2117         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2118                 if (!b->level)
2119                         bch_btree_leaf_dirty(b, journal_ref);
2120                 else
2121                         bch_btree_node_write(b, &cl);
2122         }
2123
2124         mutex_unlock(&b->write_lock);
2125
2126         /* wait for btree node write if necessary, after unlock */
2127         closure_sync(&cl);
2128
2129         return 0;
2130 split:
2131         if (current->bio_list) {
2132                 op->lock = b->c->root->level + 1;
2133                 return -EAGAIN;
2134         } else if (op->lock <= b->c->root->level) {
2135                 op->lock = b->c->root->level + 1;
2136                 return -EINTR;
2137         } else {
2138                 /* Invalidated all iterators */
2139                 int ret = btree_split(b, op, insert_keys, replace_key);
2140
2141                 if (bch_keylist_empty(insert_keys))
2142                         return 0;
2143                 else if (!ret)
2144                         return -EINTR;
2145                 return ret;
2146         }
2147 }
2148
2149 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2150                                struct bkey *check_key)
2151 {
2152         int ret = -EINTR;
2153         uint64_t btree_ptr = b->key.ptr[0];
2154         unsigned long seq = b->seq;
2155         struct keylist insert;
2156         bool upgrade = op->lock == -1;
2157
2158         bch_keylist_init(&insert);
2159
2160         if (upgrade) {
2161                 rw_unlock(false, b);
2162                 rw_lock(true, b, b->level);
2163
2164                 if (b->key.ptr[0] != btree_ptr ||
2165                    b->seq != seq + 1) {
2166                        op->lock = b->level;
2167                         goto out;
2168                }
2169         }
2170
2171         SET_KEY_PTRS(check_key, 1);
2172         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2173
2174         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2175
2176         bch_keylist_add(&insert, check_key);
2177
2178         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2179
2180         BUG_ON(!ret && !bch_keylist_empty(&insert));
2181 out:
2182         if (upgrade)
2183                 downgrade_write(&b->lock);
2184         return ret;
2185 }
2186
2187 struct btree_insert_op {
2188         struct btree_op op;
2189         struct keylist  *keys;
2190         atomic_t        *journal_ref;
2191         struct bkey     *replace_key;
2192 };
2193
2194 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2195 {
2196         struct btree_insert_op *op = container_of(b_op,
2197                                         struct btree_insert_op, op);
2198
2199         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2200                                         op->journal_ref, op->replace_key);
2201         if (ret && !bch_keylist_empty(op->keys))
2202                 return ret;
2203         else
2204                 return MAP_DONE;
2205 }
2206
2207 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2208                      atomic_t *journal_ref, struct bkey *replace_key)
2209 {
2210         struct btree_insert_op op;
2211         int ret = 0;
2212
2213         BUG_ON(current->bio_list);
2214         BUG_ON(bch_keylist_empty(keys));
2215
2216         bch_btree_op_init(&op.op, 0);
2217         op.keys         = keys;
2218         op.journal_ref  = journal_ref;
2219         op.replace_key  = replace_key;
2220
2221         while (!ret && !bch_keylist_empty(keys)) {
2222                 op.op.lock = 0;
2223                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2224                                                &START_KEY(keys->keys),
2225                                                btree_insert_fn);
2226         }
2227
2228         if (ret) {
2229                 struct bkey *k;
2230
2231                 pr_err("error %i", ret);
2232
2233                 while ((k = bch_keylist_pop(keys)))
2234                         bkey_put(c, k);
2235         } else if (op.op.insert_collision)
2236                 ret = -ESRCH;
2237
2238         return ret;
2239 }
2240
2241 void bch_btree_set_root(struct btree *b)
2242 {
2243         unsigned i;
2244         struct closure cl;
2245
2246         closure_init_stack(&cl);
2247
2248         trace_bcache_btree_set_root(b);
2249
2250         BUG_ON(!b->written);
2251
2252         for (i = 0; i < KEY_PTRS(&b->key); i++)
2253                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2254
2255         mutex_lock(&b->c->bucket_lock);
2256         list_del_init(&b->list);
2257         mutex_unlock(&b->c->bucket_lock);
2258
2259         b->c->root = b;
2260
2261         bch_journal_meta(b->c, &cl);
2262         closure_sync(&cl);
2263 }
2264
2265 /* Map across nodes or keys */
2266
2267 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2268                                        struct bkey *from,
2269                                        btree_map_nodes_fn *fn, int flags)
2270 {
2271         int ret = MAP_CONTINUE;
2272
2273         if (b->level) {
2274                 struct bkey *k;
2275                 struct btree_iter iter;
2276
2277                 bch_btree_iter_init(&b->keys, &iter, from);
2278
2279                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2280                                                        bch_ptr_bad))) {
2281                         ret = btree(map_nodes_recurse, k, b,
2282                                     op, from, fn, flags);
2283                         from = NULL;
2284
2285                         if (ret != MAP_CONTINUE)
2286                                 return ret;
2287                 }
2288         }
2289
2290         if (!b->level || flags == MAP_ALL_NODES)
2291                 ret = fn(op, b);
2292
2293         return ret;
2294 }
2295
2296 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2297                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2298 {
2299         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2300 }
2301
2302 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2303                                       struct bkey *from, btree_map_keys_fn *fn,
2304                                       int flags)
2305 {
2306         int ret = MAP_CONTINUE;
2307         struct bkey *k;
2308         struct btree_iter iter;
2309
2310         bch_btree_iter_init(&b->keys, &iter, from);
2311
2312         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2313                 ret = !b->level
2314                         ? fn(op, b, k)
2315                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2316                 from = NULL;
2317
2318                 if (ret != MAP_CONTINUE)
2319                         return ret;
2320         }
2321
2322         if (!b->level && (flags & MAP_END_KEY))
2323                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2324                                      KEY_OFFSET(&b->key), 0));
2325
2326         return ret;
2327 }
2328
2329 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2330                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2331 {
2332         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2333 }
2334
2335 /* Keybuf code */
2336
2337 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2338 {
2339         /* Overlapping keys compare equal */
2340         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2341                 return -1;
2342         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2343                 return 1;
2344         return 0;
2345 }
2346
2347 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2348                                             struct keybuf_key *r)
2349 {
2350         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2351 }
2352
2353 struct refill {
2354         struct btree_op op;
2355         unsigned        nr_found;
2356         struct keybuf   *buf;
2357         struct bkey     *end;
2358         keybuf_pred_fn  *pred;
2359 };
2360
2361 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2362                             struct bkey *k)
2363 {
2364         struct refill *refill = container_of(op, struct refill, op);
2365         struct keybuf *buf = refill->buf;
2366         int ret = MAP_CONTINUE;
2367
2368         if (bkey_cmp(k, refill->end) >= 0) {
2369                 ret = MAP_DONE;
2370                 goto out;
2371         }
2372
2373         if (!KEY_SIZE(k)) /* end key */
2374                 goto out;
2375
2376         if (refill->pred(buf, k)) {
2377                 struct keybuf_key *w;
2378
2379                 spin_lock(&buf->lock);
2380
2381                 w = array_alloc(&buf->freelist);
2382                 if (!w) {
2383                         spin_unlock(&buf->lock);
2384                         return MAP_DONE;
2385                 }
2386
2387                 w->private = NULL;
2388                 bkey_copy(&w->key, k);
2389
2390                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2391                         array_free(&buf->freelist, w);
2392                 else
2393                         refill->nr_found++;
2394
2395                 if (array_freelist_empty(&buf->freelist))
2396                         ret = MAP_DONE;
2397
2398                 spin_unlock(&buf->lock);
2399         }
2400 out:
2401         buf->last_scanned = *k;
2402         return ret;
2403 }
2404
2405 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2406                        struct bkey *end, keybuf_pred_fn *pred)
2407 {
2408         struct bkey start = buf->last_scanned;
2409         struct refill refill;
2410
2411         cond_resched();
2412
2413         bch_btree_op_init(&refill.op, -1);
2414         refill.nr_found = 0;
2415         refill.buf      = buf;
2416         refill.end      = end;
2417         refill.pred     = pred;
2418
2419         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2420                            refill_keybuf_fn, MAP_END_KEY);
2421
2422         trace_bcache_keyscan(refill.nr_found,
2423                              KEY_INODE(&start), KEY_OFFSET(&start),
2424                              KEY_INODE(&buf->last_scanned),
2425                              KEY_OFFSET(&buf->last_scanned));
2426
2427         spin_lock(&buf->lock);
2428
2429         if (!RB_EMPTY_ROOT(&buf->keys)) {
2430                 struct keybuf_key *w;
2431                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2432                 buf->start      = START_KEY(&w->key);
2433
2434                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2435                 buf->end        = w->key;
2436         } else {
2437                 buf->start      = MAX_KEY;
2438                 buf->end        = MAX_KEY;
2439         }
2440
2441         spin_unlock(&buf->lock);
2442 }
2443
2444 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2445 {
2446         rb_erase(&w->node, &buf->keys);
2447         array_free(&buf->freelist, w);
2448 }
2449
2450 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2451 {
2452         spin_lock(&buf->lock);
2453         __bch_keybuf_del(buf, w);
2454         spin_unlock(&buf->lock);
2455 }
2456
2457 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2458                                   struct bkey *end)
2459 {
2460         bool ret = false;
2461         struct keybuf_key *p, *w, s;
2462         s.key = *start;
2463
2464         if (bkey_cmp(end, &buf->start) <= 0 ||
2465             bkey_cmp(start, &buf->end) >= 0)
2466                 return false;
2467
2468         spin_lock(&buf->lock);
2469         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2470
2471         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2472                 p = w;
2473                 w = RB_NEXT(w, node);
2474
2475                 if (p->private)
2476                         ret = true;
2477                 else
2478                         __bch_keybuf_del(buf, p);
2479         }
2480
2481         spin_unlock(&buf->lock);
2482         return ret;
2483 }
2484
2485 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2486 {
2487         struct keybuf_key *w;
2488         spin_lock(&buf->lock);
2489
2490         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2491
2492         while (w && w->private)
2493                 w = RB_NEXT(w, node);
2494
2495         if (w)
2496                 w->private = ERR_PTR(-EINTR);
2497
2498         spin_unlock(&buf->lock);
2499         return w;
2500 }
2501
2502 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2503                                           struct keybuf *buf,
2504                                           struct bkey *end,
2505                                           keybuf_pred_fn *pred)
2506 {
2507         struct keybuf_key *ret;
2508
2509         while (1) {
2510                 ret = bch_keybuf_next(buf);
2511                 if (ret)
2512                         break;
2513
2514                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2515                         pr_debug("scan finished");
2516                         break;
2517                 }
2518
2519                 bch_refill_keybuf(c, buf, end, pred);
2520         }
2521
2522         return ret;
2523 }
2524
2525 void bch_keybuf_init(struct keybuf *buf)
2526 {
2527         buf->last_scanned       = MAX_KEY;
2528         buf->keys               = RB_ROOT;
2529
2530         spin_lock_init(&buf->lock);
2531         array_allocator_init(&buf->freelist);
2532 }