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
 811         if (register_shrinker(&c->shrink))
 812                 pr_warn("bcache: %s: could not register shrinker",
 813                                 __func__);
 814
 815         return 0;
 816 }
 817
 818 /* Btree in memory cache - hash table */
 819
 820 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 821 {
 822         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 823 }
 824
 825 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 826 {
 827         struct btree *b;
 828
 829         rcu_read_lock();
 830         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 831                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 832                         goto out;
 833         b = NULL;
 834 out:
 835         rcu_read_unlock();
 836         return b;
 837 }
 838
 839 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 840 {
 841         struct task_struct *old;
 842
 843         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 844         if (old && old != current) {
 845                 if (op)
 846                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 847                                         TASK_UNINTERRUPTIBLE);
 848                 return -EINTR;
 849         }
 850
 851         return 0;
 852 }
 853
 854 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 855                                      struct bkey *k)
 856 {
 857         struct btree *b;
 858
 859         trace_bcache_btree_cache_cannibalize(c);
 860
 861         if (mca_cannibalize_lock(c, op))
 862                 return ERR_PTR(-EINTR);
 863
 864         list_for_each_entry_reverse(b, &c->btree_cache, list)
 865                 if (!mca_reap(b, btree_order(k), false))
 866                         return b;
 867
 868         list_for_each_entry_reverse(b, &c->btree_cache, list)
 869                 if (!mca_reap(b, btree_order(k), true))
 870                         return b;
 871
 872         WARN(1, "btree cache cannibalize failed\n");
 873         return ERR_PTR(-ENOMEM);
 874 }
 875
 876 /*
 877  * We can only have one thread cannibalizing other cached btree nodes at a time,
 878  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 879  * cannibalize_bucket() will take. This means every time we unlock the root of
 880  * the btree, we need to release this lock if we have it held.
 881  */
 882 static void bch_cannibalize_unlock(struct cache_set *c)
 883 {
 884         if (c->btree_cache_alloc_lock == current) {
 885                 c->btree_cache_alloc_lock = NULL;
 886                 wake_up(&c->btree_cache_wait);
 887         }
 888 }
 889
 890 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 891                                struct bkey *k, int level)
 892 {
 893         struct btree *b;
 894
 895         BUG_ON(current->bio_list);
 896
 897         lockdep_assert_held(&c->bucket_lock);
 898
 899         if (mca_find(c, k))
 900                 return NULL;
 901
 902         /* btree_free() doesn't free memory; it sticks the node on the end of
 903          * the list. Check if there's any freed nodes there:
 904          */
 905         list_for_each_entry(b, &c->btree_cache_freeable, list)
 906                 if (!mca_reap(b, btree_order(k), false))
 907                         goto out;
 908
 909         /* We never free struct btree itself, just the memory that holds the on
 910          * disk node. Check the freed list before allocating a new one:
 911          */
 912         list_for_each_entry(b, &c->btree_cache_freed, list)
 913                 if (!mca_reap(b, 0, false)) {
 914                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 915                         if (!b->keys.set[0].data)
 916                                 goto err;
 917                         else
 918                                 goto out;
 919                 }
 920
 921         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 922         if (!b)
 923                 goto err;
 924
 925         BUG_ON(!down_write_trylock(&b->lock));
 926         if (!b->keys.set->data)
 927                 goto err;
 928 out:
 929         BUG_ON(b->io_mutex.count != 1);
 930
 931         bkey_copy(&b->key, k);
 932         list_move(&b->list, &c->btree_cache);
 933         hlist_del_init_rcu(&b->hash);
 934         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 935
 936         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 937         b->parent       = (void *) ~0UL;
 938         b->flags        = 0;
 939         b->written      = 0;
 940         b->level        = level;
 941
 942         if (!b->level)
 943                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 944                                     &b->c->expensive_debug_checks);
 945         else
 946                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 947                                     &b->c->expensive_debug_checks);
 948
 949         return b;
 950 err:
 951         if (b)
 952                 rw_unlock(true, b);
 953
 954         b = mca_cannibalize(c, op, k);
 955         if (!IS_ERR(b))
 956                 goto out;
 957
 958         return b;
 959 }
 960
 961 /**
 962  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 963  * in from disk if necessary.
 964  *
 965  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 966  *
 967  * The btree node will have either a read or a write lock held, depending on
 968  * level and op->lock.
 969  */
 970 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 971                                  struct bkey *k, int level, bool write,
 972                                  struct btree *parent)
 973 {
 974         int i = 0;
 975         struct btree *b;
 976
 977         BUG_ON(level < 0);
 978 retry:
 979         b = mca_find(c, k);
 980
 981         if (!b) {
 982                 if (current->bio_list)
 983                         return ERR_PTR(-EAGAIN);
 984
 985                 mutex_lock(&c->bucket_lock);
 986                 b = mca_alloc(c, op, k, level);
 987                 mutex_unlock(&c->bucket_lock);
 988
 989                 if (!b)
 990                         goto retry;
 991                 if (IS_ERR(b))
 992                         return b;
 993
 994                 bch_btree_node_read(b);
 995
 996                 if (!write)
 997                         downgrade_write(&b->lock);
 998         } else {
 999                 rw_lock(write, b, level);
1000                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1001                         rw_unlock(write, b);
1002                         goto retry;
1003                 }
1004                 BUG_ON(b->level != level);
1005         }
1006
1007         b->parent = parent;
1008         b->accessed = 1;
1009
1010         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1011                 prefetch(b->keys.set[i].tree);
1012                 prefetch(b->keys.set[i].data);
1013         }
1014
1015         for (; i <= b->keys.nsets; i++)
1016                 prefetch(b->keys.set[i].data);
1017
1018         if (btree_node_io_error(b)) {
1019                 rw_unlock(write, b);
1020                 return ERR_PTR(-EIO);
1021         }
1022
1023         BUG_ON(!b->written);
1024
1025         return b;
1026 }
1027
1028 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1029 {
1030         struct btree *b;
1031
1032         mutex_lock(&parent->c->bucket_lock);
1033         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1034         mutex_unlock(&parent->c->bucket_lock);
1035
1036         if (!IS_ERR_OR_NULL(b)) {
1037                 b->parent = parent;
1038                 bch_btree_node_read(b);
1039                 rw_unlock(true, b);
1040         }
1041 }
1042
1043 /* Btree alloc */
1044
1045 static void btree_node_free(struct btree *b)
1046 {
1047         trace_bcache_btree_node_free(b);
1048
1049         BUG_ON(b == b->c->root);
1050
1051         mutex_lock(&b->write_lock);
1052
1053         if (btree_node_dirty(b))
1054                 btree_complete_write(b, btree_current_write(b));
1055         clear_bit(BTREE_NODE_dirty, &b->flags);
1056
1057         mutex_unlock(&b->write_lock);
1058
1059         cancel_delayed_work(&b->work);
1060
1061         mutex_lock(&b->c->bucket_lock);
1062         bch_bucket_free(b->c, &b->key);
1063         mca_bucket_free(b);
1064         mutex_unlock(&b->c->bucket_lock);
1065 }
1066
1067 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1068                                      int level, bool wait,
1069                                      struct btree *parent)
1070 {
1071         BKEY_PADDED(key) k;
1072         struct btree *b = ERR_PTR(-EAGAIN);
1073
1074         mutex_lock(&c->bucket_lock);
1075 retry:
1076         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1077                 goto err;
1078
1079         bkey_put(c, &k.key);
1080         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1081
1082         b = mca_alloc(c, op, &k.key, level);
1083         if (IS_ERR(b))
1084                 goto err_free;
1085
1086         if (!b) {
1087                 cache_bug(c,
1088                         "Tried to allocate bucket that was in btree cache");
1089                 goto retry;
1090         }
1091
1092         b->accessed = 1;
1093         b->parent = parent;
1094         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1095
1096         mutex_unlock(&c->bucket_lock);
1097
1098         trace_bcache_btree_node_alloc(b);
1099         return b;
1100 err_free:
1101         bch_bucket_free(c, &k.key);
1102 err:
1103         mutex_unlock(&c->bucket_lock);
1104
1105         trace_bcache_btree_node_alloc_fail(c);
1106         return b;
1107 }
1108
1109 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1110                                           struct btree_op *op, int level,
1111                                           struct btree *parent)
1112 {
1113         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1114 }
1115
1116 static struct btree *btree_node_alloc_replacement(struct btree *b,
1117                                                   struct btree_op *op)
1118 {
1119         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1120         if (!IS_ERR_OR_NULL(n)) {
1121                 mutex_lock(&n->write_lock);
1122                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1123                 bkey_copy_key(&n->key, &b->key);
1124                 mutex_unlock(&n->write_lock);
1125         }
1126
1127         return n;
1128 }
1129
1130 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1131 {
1132         unsigned i;
1133
1134         mutex_lock(&b->c->bucket_lock);
1135
1136         atomic_inc(&b->c->prio_blocked);
1137
1138         bkey_copy(k, &b->key);
1139         bkey_copy_key(k, &ZERO_KEY);
1140
1141         for (i = 0; i < KEY_PTRS(k); i++)
1142                 SET_PTR_GEN(k, i,
1143                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1144                                         PTR_BUCKET(b->c, &b->key, i)));
1145
1146         mutex_unlock(&b->c->bucket_lock);
1147 }
1148
1149 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1150 {
1151         struct cache_set *c = b->c;
1152         struct cache *ca;
1153         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1154
1155         mutex_lock(&c->bucket_lock);
1156
1157         for_each_cache(ca, c, i)
1158                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1159                         if (op)
1160                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1161                                                 TASK_UNINTERRUPTIBLE);
1162                         mutex_unlock(&c->bucket_lock);
1163                         return -EINTR;
1164                 }
1165
1166         mutex_unlock(&c->bucket_lock);
1167
1168         return mca_cannibalize_lock(b->c, op);
1169 }
1170
1171 /* Garbage collection */
1172
1173 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1174                                     struct bkey *k)
1175 {
1176         uint8_t stale = 0;
1177         unsigned i;
1178         struct bucket *g;
1179
1180         /*
1181          * ptr_invalid() can't return true for the keys that mark btree nodes as
1182          * freed, but since ptr_bad() returns true we'll never actually use them
1183          * for anything and thus we don't want mark their pointers here
1184          */
1185         if (!bkey_cmp(k, &ZERO_KEY))
1186                 return stale;
1187
1188         for (i = 0; i < KEY_PTRS(k); i++) {
1189                 if (!ptr_available(c, k, i))
1190                         continue;
1191
1192                 g = PTR_BUCKET(c, k, i);
1193
1194                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1195                         g->last_gc = PTR_GEN(k, i);
1196
1197                 if (ptr_stale(c, k, i)) {
1198                         stale = max(stale, ptr_stale(c, k, i));
1199                         continue;
1200                 }
1201
1202                 cache_bug_on(GC_MARK(g) &&
1203                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1204                              c, "inconsistent ptrs: mark = %llu, level = %i",
1205                              GC_MARK(g), level);
1206
1207                 if (level)
1208                         SET_GC_MARK(g, GC_MARK_METADATA);
1209                 else if (KEY_DIRTY(k))
1210                         SET_GC_MARK(g, GC_MARK_DIRTY);
1211                 else if (!GC_MARK(g))
1212                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1213
1214                 /* guard against overflow */
1215                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1216                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1217                                              MAX_GC_SECTORS_USED));
1218
1219                 BUG_ON(!GC_SECTORS_USED(g));
1220         }
1221
1222         return stale;
1223 }
1224
1225 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1226
1227 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1228 {
1229         unsigned i;
1230
1231         for (i = 0; i < KEY_PTRS(k); i++)
1232                 if (ptr_available(c, k, i) &&
1233                     !ptr_stale(c, k, i)) {
1234                         struct bucket *b = PTR_BUCKET(c, k, i);
1235
1236                         b->gen = PTR_GEN(k, i);
1237
1238                         if (level && bkey_cmp(k, &ZERO_KEY))
1239                                 b->prio = BTREE_PRIO;
1240                         else if (!level && b->prio == BTREE_PRIO)
1241                                 b->prio = INITIAL_PRIO;
1242                 }
1243
1244         __bch_btree_mark_key(c, level, k);
1245 }
1246
1247 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1248 {
1249         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1250 }
1251
1252 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1253 {
1254         uint8_t stale = 0;
1255         unsigned keys = 0, good_keys = 0;
1256         struct bkey *k;
1257         struct btree_iter iter;
1258         struct bset_tree *t;
1259
1260         gc->nodes++;
1261
1262         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1263                 stale = max(stale, btree_mark_key(b, k));
1264                 keys++;
1265
1266                 if (bch_ptr_bad(&b->keys, k))
1267                         continue;
1268
1269                 gc->key_bytes += bkey_u64s(k);
1270                 gc->nkeys++;
1271                 good_keys++;
1272
1273                 gc->data += KEY_SIZE(k);
1274         }
1275
1276         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1277                 btree_bug_on(t->size &&
1278                              bset_written(&b->keys, t) &&
1279                              bkey_cmp(&b->key, &t->end) < 0,
1280                              b, "found short btree key in gc");
1281
1282         if (b->c->gc_always_rewrite)
1283                 return true;
1284
1285         if (stale > 10)
1286                 return true;
1287
1288         if ((keys - good_keys) * 2 > keys)
1289                 return true;
1290
1291         return false;
1292 }
1293
1294 #define GC_MERGE_NODES  4U
1295
1296 struct gc_merge_info {
1297         struct btree    *b;
1298         unsigned        keys;
1299 };
1300
1301 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1302                                  struct keylist *, atomic_t *, struct bkey *);
1303
1304 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1305                              struct gc_stat *gc, struct gc_merge_info *r)
1306 {
1307         unsigned i, nodes = 0, keys = 0, blocks;
1308         struct btree *new_nodes[GC_MERGE_NODES];
1309         struct keylist keylist;
1310         struct closure cl;
1311         struct bkey *k;
1312
1313         bch_keylist_init(&keylist);
1314
1315         if (btree_check_reserve(b, NULL))
1316                 return 0;
1317
1318         memset(new_nodes, 0, sizeof(new_nodes));
1319         closure_init_stack(&cl);
1320
1321         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1322                 keys += r[nodes++].keys;
1323
1324         blocks = btree_default_blocks(b->c) * 2 / 3;
1325
1326         if (nodes < 2 ||
1327             __set_blocks(b->keys.set[0].data, keys,
1328                          block_bytes(b->c)) > blocks * (nodes - 1))
1329                 return 0;
1330
1331         for (i = 0; i < nodes; i++) {
1332                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1333                 if (IS_ERR_OR_NULL(new_nodes[i]))
1334                         goto out_nocoalesce;
1335         }
1336
1337         /*
1338          * We have to check the reserve here, after we've allocated our new
1339          * nodes, to make sure the insert below will succeed - we also check
1340          * before as an optimization to potentially avoid a bunch of expensive
1341          * allocs/sorts
1342          */
1343         if (btree_check_reserve(b, NULL))
1344                 goto out_nocoalesce;
1345
1346         for (i = 0; i < nodes; i++)
1347                 mutex_lock(&new_nodes[i]->write_lock);
1348
1349         for (i = nodes - 1; i > 0; --i) {
1350                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1351                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1352                 struct bkey *k, *last = NULL;
1353
1354                 keys = 0;
1355
1356                 if (i > 1) {
1357                         for (k = n2->start;
1358                              k < bset_bkey_last(n2);
1359                              k = bkey_next(k)) {
1360                                 if (__set_blocks(n1, n1->keys + keys +
1361                                                  bkey_u64s(k),
1362                                                  block_bytes(b->c)) > blocks)
1363                                         break;
1364
1365                                 last = k;
1366                                 keys += bkey_u64s(k);
1367                         }
1368                 } else {
1369                         /*
1370                          * Last node we're not getting rid of - we're getting
1371                          * rid of the node at r[0]. Have to try and fit all of
1372                          * the remaining keys into this node; we can't ensure
1373                          * they will always fit due to rounding and variable
1374                          * length keys (shouldn't be possible in practice,
1375                          * though)
1376                          */
1377                         if (__set_blocks(n1, n1->keys + n2->keys,
1378                                          block_bytes(b->c)) >
1379                             btree_blocks(new_nodes[i]))
1380                                 goto out_nocoalesce;
1381
1382                         keys = n2->keys;
1383                         /* Take the key of the node we're getting rid of */
1384                         last = &r->b->key;
1385                 }
1386
1387                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1388                        btree_blocks(new_nodes[i]));
1389
1390                 if (last)
1391                         bkey_copy_key(&new_nodes[i]->key, last);
1392
1393                 memcpy(bset_bkey_last(n1),
1394                        n2->start,
1395                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1396
1397                 n1->keys += keys;
1398                 r[i].keys = n1->keys;
1399
1400                 memmove(n2->start,
1401                         bset_bkey_idx(n2, keys),
1402                         (void *) bset_bkey_last(n2) -
1403                         (void *) bset_bkey_idx(n2, keys));
1404
1405                 n2->keys -= keys;
1406
1407                 if (__bch_keylist_realloc(&keylist,
1408                                           bkey_u64s(&new_nodes[i]->key)))
1409                         goto out_nocoalesce;
1410
1411                 bch_btree_node_write(new_nodes[i], &cl);
1412                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1413         }
1414
1415         for (i = 0; i < nodes; i++)
1416                 mutex_unlock(&new_nodes[i]->write_lock);
1417
1418         closure_sync(&cl);
1419
1420         /* We emptied out this node */
1421         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1422         btree_node_free(new_nodes[0]);
1423         rw_unlock(true, new_nodes[0]);
1424         new_nodes[0] = NULL;
1425
1426         for (i = 0; i < nodes; i++) {
1427                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1428                         goto out_nocoalesce;
1429
1430                 make_btree_freeing_key(r[i].b, keylist.top);
1431                 bch_keylist_push(&keylist);
1432         }
1433
1434         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1435         BUG_ON(!bch_keylist_empty(&keylist));
1436
1437         for (i = 0; i < nodes; i++) {
1438                 btree_node_free(r[i].b);
1439                 rw_unlock(true, r[i].b);
1440
1441                 r[i].b = new_nodes[i];
1442         }
1443
1444         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1445         r[nodes - 1].b = ERR_PTR(-EINTR);
1446
1447         trace_bcache_btree_gc_coalesce(nodes);
1448         gc->nodes--;
1449
1450         bch_keylist_free(&keylist);
1451
1452         /* Invalidated our iterator */
1453         return -EINTR;
1454
1455 out_nocoalesce:
1456         closure_sync(&cl);
1457         bch_keylist_free(&keylist);
1458
1459         while ((k = bch_keylist_pop(&keylist)))
1460                 if (!bkey_cmp(k, &ZERO_KEY))
1461                         atomic_dec(&b->c->prio_blocked);
1462
1463         for (i = 0; i < nodes; i++)
1464                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1465                         btree_node_free(new_nodes[i]);
1466                         rw_unlock(true, new_nodes[i]);
1467                 }
1468         return 0;
1469 }
1470
1471 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1472                                  struct btree *replace)
1473 {
1474         struct keylist keys;
1475         struct btree *n;
1476
1477         if (btree_check_reserve(b, NULL))
1478                 return 0;
1479
1480         n = btree_node_alloc_replacement(replace, NULL);
1481
1482         /* recheck reserve after allocating replacement node */
1483         if (btree_check_reserve(b, NULL)) {
1484                 btree_node_free(n);
1485                 rw_unlock(true, n);
1486                 return 0;
1487         }
1488
1489         bch_btree_node_write_sync(n);
1490
1491         bch_keylist_init(&keys);
1492         bch_keylist_add(&keys, &n->key);
1493
1494         make_btree_freeing_key(replace, keys.top);
1495         bch_keylist_push(&keys);
1496
1497         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1498         BUG_ON(!bch_keylist_empty(&keys));
1499
1500         btree_node_free(replace);
1501         rw_unlock(true, n);
1502
1503         /* Invalidated our iterator */
1504         return -EINTR;
1505 }
1506
1507 static unsigned btree_gc_count_keys(struct btree *b)
1508 {
1509         struct bkey *k;
1510         struct btree_iter iter;
1511         unsigned ret = 0;
1512
1513         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1514                 ret += bkey_u64s(k);
1515
1516         return ret;
1517 }
1518
1519 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1520                             struct closure *writes, struct gc_stat *gc)
1521 {
1522         int ret = 0;
1523         bool should_rewrite;
1524         struct bkey *k;
1525         struct btree_iter iter;
1526         struct gc_merge_info r[GC_MERGE_NODES];
1527         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1528
1529         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1530
1531         for (i = r; i < r + ARRAY_SIZE(r); i++)
1532                 i->b = ERR_PTR(-EINTR);
1533
1534         while (1) {
1535                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1536                 if (k) {
1537                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1538                                                   true, b);
1539                         if (IS_ERR(r->b)) {
1540                                 ret = PTR_ERR(r->b);
1541                                 break;
1542                         }
1543
1544                         r->keys = btree_gc_count_keys(r->b);
1545
1546                         ret = btree_gc_coalesce(b, op, gc, r);
1547                         if (ret)
1548                                 break;
1549                 }
1550
1551                 if (!last->b)
1552                         break;
1553
1554                 if (!IS_ERR(last->b)) {
1555                         should_rewrite = btree_gc_mark_node(last->b, gc);
1556                         if (should_rewrite) {
1557                                 ret = btree_gc_rewrite_node(b, op, last->b);
1558                                 if (ret)
1559                                         break;
1560                         }
1561
1562                         if (last->b->level) {
1563                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1564                                 if (ret)
1565                                         break;
1566                         }
1567
1568                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1569
1570                         /*
1571                          * Must flush leaf nodes before gc ends, since replace
1572                          * operations aren't journalled
1573                          */
1574                         mutex_lock(&last->b->write_lock);
1575                         if (btree_node_dirty(last->b))
1576                                 bch_btree_node_write(last->b, writes);
1577                         mutex_unlock(&last->b->write_lock);
1578                         rw_unlock(true, last->b);
1579                 }
1580
1581                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1582                 r->b = NULL;
1583
1584                 if (need_resched()) {
1585                         ret = -EAGAIN;
1586                         break;
1587                 }
1588         }
1589
1590         for (i = r; i < r + ARRAY_SIZE(r); i++)
1591                 if (!IS_ERR_OR_NULL(i->b)) {
1592                         mutex_lock(&i->b->write_lock);
1593                         if (btree_node_dirty(i->b))
1594                                 bch_btree_node_write(i->b, writes);
1595                         mutex_unlock(&i->b->write_lock);
1596                         rw_unlock(true, i->b);
1597                 }
1598
1599         return ret;
1600 }
1601
1602 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1603                              struct closure *writes, struct gc_stat *gc)
1604 {
1605         struct btree *n = NULL;
1606         int ret = 0;
1607         bool should_rewrite;
1608
1609         should_rewrite = btree_gc_mark_node(b, gc);
1610         if (should_rewrite) {
1611                 n = btree_node_alloc_replacement(b, NULL);
1612
1613                 if (!IS_ERR_OR_NULL(n)) {
1614                         bch_btree_node_write_sync(n);
1615
1616                         bch_btree_set_root(n);
1617                         btree_node_free(b);
1618                         rw_unlock(true, n);
1619
1620                         return -EINTR;
1621                 }
1622         }
1623
1624         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1625
1626         if (b->level) {
1627                 ret = btree_gc_recurse(b, op, writes, gc);
1628                 if (ret)
1629                         return ret;
1630         }
1631
1632         bkey_copy_key(&b->c->gc_done, &b->key);
1633
1634         return ret;
1635 }
1636
1637 static void btree_gc_start(struct cache_set *c)
1638 {
1639         struct cache *ca;
1640         struct bucket *b;
1641         unsigned i;
1642
1643         if (!c->gc_mark_valid)
1644                 return;
1645
1646         mutex_lock(&c->bucket_lock);
1647
1648         c->gc_mark_valid = 0;
1649         c->gc_done = ZERO_KEY;
1650
1651         for_each_cache(ca, c, i)
1652                 for_each_bucket(b, ca) {
1653                         b->last_gc = b->gen;
1654                         if (!atomic_read(&b->pin)) {
1655                                 SET_GC_MARK(b, 0);
1656                                 SET_GC_SECTORS_USED(b, 0);
1657                         }
1658                 }
1659
1660         mutex_unlock(&c->bucket_lock);
1661 }
1662
1663 static void bch_btree_gc_finish(struct cache_set *c)
1664 {
1665         struct bucket *b;
1666         struct cache *ca;
1667         unsigned i;
1668
1669         mutex_lock(&c->bucket_lock);
1670
1671         set_gc_sectors(c);
1672         c->gc_mark_valid = 1;
1673         c->need_gc      = 0;
1674
1675         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1676                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1677                             GC_MARK_METADATA);
1678
1679         /* don't reclaim buckets to which writeback keys point */
1680         rcu_read_lock();
1681         for (i = 0; i < c->nr_uuids; i++) {
1682                 struct bcache_device *d = c->devices[i];
1683                 struct cached_dev *dc;
1684                 struct keybuf_key *w, *n;
1685                 unsigned j;
1686
1687                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1688                         continue;
1689                 dc = container_of(d, struct cached_dev, disk);
1690
1691                 spin_lock(&dc->writeback_keys.lock);
1692                 rbtree_postorder_for_each_entry_safe(w, n,
1693                                         &dc->writeback_keys.keys, node)
1694                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1695                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1696                                             GC_MARK_DIRTY);
1697                 spin_unlock(&dc->writeback_keys.lock);
1698         }
1699         rcu_read_unlock();
1700
1701         c->avail_nbuckets = 0;
1702         for_each_cache(ca, c, i) {
1703                 uint64_t *i;
1704
1705                 ca->invalidate_needs_gc = 0;
1706
1707                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1708                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1709
1710                 for (i = ca->prio_buckets;
1711                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1712                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1713
1714                 for_each_bucket(b, ca) {
1715                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1716
1717                         if (atomic_read(&b->pin))
1718                                 continue;
1719
1720                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1721
1722                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1723                                 c->avail_nbuckets++;
1724                 }
1725         }
1726
1727         mutex_unlock(&c->bucket_lock);
1728 }
1729
1730 static void bch_btree_gc(struct cache_set *c)
1731 {
1732         int ret;
1733         struct gc_stat stats;
1734         struct closure writes;
1735         struct btree_op op;
1736         uint64_t start_time = local_clock();
1737
1738         trace_bcache_gc_start(c);
1739
1740         memset(&stats, 0, sizeof(struct gc_stat));
1741         closure_init_stack(&writes);
1742         bch_btree_op_init(&op, SHRT_MAX);
1743
1744         btree_gc_start(c);
1745
1746         do {
1747                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1748                 closure_sync(&writes);
1749                 cond_resched();
1750
1751                 if (ret && ret != -EAGAIN)
1752                         pr_warn("gc failed!");
1753         } while (ret);
1754
1755         bch_btree_gc_finish(c);
1756         wake_up_allocators(c);
1757
1758         bch_time_stats_update(&c->btree_gc_time, start_time);
1759
1760         stats.key_bytes *= sizeof(uint64_t);
1761         stats.data      <<= 9;
1762         bch_update_bucket_in_use(c, &stats);
1763         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1764
1765         trace_bcache_gc_end(c);
1766
1767         bch_moving_gc(c);
1768 }
1769
1770 static bool gc_should_run(struct cache_set *c)
1771 {
1772         struct cache *ca;
1773         unsigned i;
1774
1775         for_each_cache(ca, c, i)
1776                 if (ca->invalidate_needs_gc)
1777                         return true;
1778
1779         if (atomic_read(&c->sectors_to_gc) < 0)
1780                 return true;
1781
1782         return false;
1783 }
1784
1785 static int bch_gc_thread(void *arg)
1786 {
1787         struct cache_set *c = arg;
1788
1789         while (1) {
1790                 wait_event_interruptible(c->gc_wait,
1791                            kthread_should_stop() || gc_should_run(c));
1792
1793                 if (kthread_should_stop())
1794                         break;
1795
1796                 set_gc_sectors(c);
1797                 bch_btree_gc(c);
1798         }
1799
1800         return 0;
1801 }
1802
1803 int bch_gc_thread_start(struct cache_set *c)
1804 {
1805         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1806         if (IS_ERR(c->gc_thread))
1807                 return PTR_ERR(c->gc_thread);
1808
1809         return 0;
1810 }
1811
1812 /* Initial partial gc */
1813
1814 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1815 {
1816         int ret = 0;
1817         struct bkey *k, *p = NULL;
1818         struct btree_iter iter;
1819
1820         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1821                 bch_initial_mark_key(b->c, b->level, k);
1822
1823         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1824
1825         if (b->level) {
1826                 bch_btree_iter_init(&b->keys, &iter, NULL);
1827
1828                 do {
1829                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1830                                                        bch_ptr_bad);
1831                         if (k)
1832                                 btree_node_prefetch(b, k);
1833
1834                         if (p)
1835                                 ret = btree(check_recurse, p, b, op);
1836
1837                         p = k;
1838                 } while (p && !ret);
1839         }
1840
1841         return ret;
1842 }
1843
1844 int bch_btree_check(struct cache_set *c)
1845 {
1846         struct btree_op op;
1847
1848         bch_btree_op_init(&op, SHRT_MAX);
1849
1850         return btree_root(check_recurse, c, &op);
1851 }
1852
1853 void bch_initial_gc_finish(struct cache_set *c)
1854 {
1855         struct cache *ca;
1856         struct bucket *b;
1857         unsigned i;
1858
1859         bch_btree_gc_finish(c);
1860
1861         mutex_lock(&c->bucket_lock);
1862
1863         /*
1864          * We need to put some unused buckets directly on the prio freelist in
1865          * order to get the allocator thread started - it needs freed buckets in
1866          * order to rewrite the prios and gens, and it needs to rewrite prios
1867          * and gens in order to free buckets.
1868          *
1869          * This is only safe for buckets that have no live data in them, which
1870          * there should always be some of.
1871          */
1872         for_each_cache(ca, c, i) {
1873                 for_each_bucket(b, ca) {
1874                         if (fifo_full(&ca->free[RESERVE_PRIO]))
1875                                 break;
1876
1877                         if (bch_can_invalidate_bucket(ca, b) &&
1878                             !GC_MARK(b)) {
1879                                 __bch_invalidate_one_bucket(ca, b);
1880                                 fifo_push(&ca->free[RESERVE_PRIO],
1881                                           b - ca->buckets);
1882                         }
1883                 }
1884         }
1885
1886         mutex_unlock(&c->bucket_lock);
1887 }
1888
1889 /* Btree insertion */
1890
1891 static bool btree_insert_key(struct btree *b, struct bkey *k,
1892                              struct bkey *replace_key)
1893 {
1894         unsigned status;
1895
1896         BUG_ON(bkey_cmp(k, &b->key) > 0);
1897
1898         status = bch_btree_insert_key(&b->keys, k, replace_key);
1899         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1900                 bch_check_keys(&b->keys, "%u for %s", status,
1901                                replace_key ? "replace" : "insert");
1902
1903                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1904                                               status);
1905                 return true;
1906         } else
1907                 return false;
1908 }
1909
1910 static size_t insert_u64s_remaining(struct btree *b)
1911 {
1912         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1913
1914         /*
1915          * Might land in the middle of an existing extent and have to split it
1916          */
1917         if (b->keys.ops->is_extents)
1918                 ret -= KEY_MAX_U64S;
1919
1920         return max(ret, 0L);
1921 }
1922
1923 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1924                                   struct keylist *insert_keys,
1925                                   struct bkey *replace_key)
1926 {
1927         bool ret = false;
1928         int oldsize = bch_count_data(&b->keys);
1929
1930         while (!bch_keylist_empty(insert_keys)) {
1931                 struct bkey *k = insert_keys->keys;
1932
1933                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1934                         break;
1935
1936                 if (bkey_cmp(k, &b->key) <= 0) {
1937                         if (!b->level)
1938                                 bkey_put(b->c, k);
1939
1940                         ret |= btree_insert_key(b, k, replace_key);
1941                         bch_keylist_pop_front(insert_keys);
1942                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1943                         BKEY_PADDED(key) temp;
1944                         bkey_copy(&temp.key, insert_keys->keys);
1945
1946                         bch_cut_back(&b->key, &temp.key);
1947                         bch_cut_front(&b->key, insert_keys->keys);
1948
1949                         ret |= btree_insert_key(b, &temp.key, replace_key);
1950                         break;
1951                 } else {
1952                         break;
1953                 }
1954         }
1955
1956         if (!ret)
1957                 op->insert_collision = true;
1958
1959         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1960
1961         BUG_ON(bch_count_data(&b->keys) < oldsize);
1962         return ret;
1963 }
1964
1965 static int btree_split(struct btree *b, struct btree_op *op,
1966                        struct keylist *insert_keys,
1967                        struct bkey *replace_key)
1968 {
1969         bool split;
1970         struct btree *n1, *n2 = NULL, *n3 = NULL;
1971         uint64_t start_time = local_clock();
1972         struct closure cl;
1973         struct keylist parent_keys;
1974
1975         closure_init_stack(&cl);
1976         bch_keylist_init(&parent_keys);
1977
1978         if (btree_check_reserve(b, op)) {
1979                 if (!b->level)
1980                         return -EINTR;
1981                 else
1982                         WARN(1, "insufficient reserve for split\n");
1983         }
1984
1985         n1 = btree_node_alloc_replacement(b, op);
1986         if (IS_ERR(n1))
1987                 goto err;
1988
1989         split = set_blocks(btree_bset_first(n1),
1990                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1991
1992         if (split) {
1993                 unsigned keys = 0;
1994
1995                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1996
1997                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1998                 if (IS_ERR(n2))
1999                         goto err_free1;
2000
2001                 if (!b->parent) {
2002                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2003                         if (IS_ERR(n3))
2004                                 goto err_free2;
2005                 }
2006
2007                 mutex_lock(&n1->write_lock);
2008                 mutex_lock(&n2->write_lock);
2009
2010                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2011
2012                 /*
2013                  * Has to be a linear search because we don't have an auxiliary
2014                  * search tree yet
2015                  */
2016
2017                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2018                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2019                                                         keys));
2020
2021                 bkey_copy_key(&n1->key,
2022                               bset_bkey_idx(btree_bset_first(n1), keys));
2023                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2024
2025                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2026                 btree_bset_first(n1)->keys = keys;
2027
2028                 memcpy(btree_bset_first(n2)->start,
2029                        bset_bkey_last(btree_bset_first(n1)),
2030                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2031
2032                 bkey_copy_key(&n2->key, &b->key);
2033
2034                 bch_keylist_add(&parent_keys, &n2->key);
2035                 bch_btree_node_write(n2, &cl);
2036                 mutex_unlock(&n2->write_lock);
2037                 rw_unlock(true, n2);
2038         } else {
2039                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2040
2041                 mutex_lock(&n1->write_lock);
2042                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2043         }
2044
2045         bch_keylist_add(&parent_keys, &n1->key);
2046         bch_btree_node_write(n1, &cl);
2047         mutex_unlock(&n1->write_lock);
2048
2049         if (n3) {
2050                 /* Depth increases, make a new root */
2051                 mutex_lock(&n3->write_lock);
2052                 bkey_copy_key(&n3->key, &MAX_KEY);
2053                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2054                 bch_btree_node_write(n3, &cl);
2055                 mutex_unlock(&n3->write_lock);
2056
2057                 closure_sync(&cl);
2058                 bch_btree_set_root(n3);
2059                 rw_unlock(true, n3);
2060         } else if (!b->parent) {
2061                 /* Root filled up but didn't need to be split */
2062                 closure_sync(&cl);
2063                 bch_btree_set_root(n1);
2064         } else {
2065                 /* Split a non root node */
2066                 closure_sync(&cl);
2067                 make_btree_freeing_key(b, parent_keys.top);
2068                 bch_keylist_push(&parent_keys);
2069
2070                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2071                 BUG_ON(!bch_keylist_empty(&parent_keys));
2072         }
2073
2074         btree_node_free(b);
2075         rw_unlock(true, n1);
2076
2077         bch_time_stats_update(&b->c->btree_split_time, start_time);
2078
2079         return 0;
2080 err_free2:
2081         bkey_put(b->c, &n2->key);
2082         btree_node_free(n2);
2083         rw_unlock(true, n2);
2084 err_free1:
2085         bkey_put(b->c, &n1->key);
2086         btree_node_free(n1);
2087         rw_unlock(true, n1);
2088 err:
2089         WARN(1, "bcache: btree split failed (level %u)", b->level);
2090
2091         if (n3 == ERR_PTR(-EAGAIN) ||
2092             n2 == ERR_PTR(-EAGAIN) ||
2093             n1 == ERR_PTR(-EAGAIN))
2094                 return -EAGAIN;
2095
2096         return -ENOMEM;
2097 }
2098
2099 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2100                                  struct keylist *insert_keys,
2101                                  atomic_t *journal_ref,
2102                                  struct bkey *replace_key)
2103 {
2104         struct closure cl;
2105
2106         BUG_ON(b->level && replace_key);
2107
2108         closure_init_stack(&cl);
2109
2110         mutex_lock(&b->write_lock);
2111
2112         if (write_block(b) != btree_bset_last(b) &&
2113             b->keys.last_set_unwritten)
2114                 bch_btree_init_next(b); /* just wrote a set */
2115
2116         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2117                 mutex_unlock(&b->write_lock);
2118                 goto split;
2119         }
2120
2121         BUG_ON(write_block(b) != btree_bset_last(b));
2122
2123         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2124                 if (!b->level)
2125                         bch_btree_leaf_dirty(b, journal_ref);
2126                 else
2127                         bch_btree_node_write(b, &cl);
2128         }
2129
2130         mutex_unlock(&b->write_lock);
2131
2132         /* wait for btree node write if necessary, after unlock */
2133         closure_sync(&cl);
2134
2135         return 0;
2136 split:
2137         if (current->bio_list) {
2138                 op->lock = b->c->root->level + 1;
2139                 return -EAGAIN;
2140         } else if (op->lock <= b->c->root->level) {
2141                 op->lock = b->c->root->level + 1;
2142                 return -EINTR;
2143         } else {
2144                 /* Invalidated all iterators */
2145                 int ret = btree_split(b, op, insert_keys, replace_key);
2146
2147                 if (bch_keylist_empty(insert_keys))
2148                         return 0;
2149                 else if (!ret)
2150                         return -EINTR;
2151                 return ret;
2152         }
2153 }
2154
2155 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2156                                struct bkey *check_key)
2157 {
2158         int ret = -EINTR;
2159         uint64_t btree_ptr = b->key.ptr[0];
2160         unsigned long seq = b->seq;
2161         struct keylist insert;
2162         bool upgrade = op->lock == -1;
2163
2164         bch_keylist_init(&insert);
2165
2166         if (upgrade) {
2167                 rw_unlock(false, b);
2168                 rw_lock(true, b, b->level);
2169
2170                 if (b->key.ptr[0] != btree_ptr ||
2171                    b->seq != seq + 1) {
2172                        op->lock = b->level;
2173                         goto out;
2174                }
2175         }
2176
2177         SET_KEY_PTRS(check_key, 1);
2178         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2179
2180         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2181
2182         bch_keylist_add(&insert, check_key);
2183
2184         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2185
2186         BUG_ON(!ret && !bch_keylist_empty(&insert));
2187 out:
2188         if (upgrade)
2189                 downgrade_write(&b->lock);
2190         return ret;
2191 }
2192
2193 struct btree_insert_op {
2194         struct btree_op op;
2195         struct keylist  *keys;
2196         atomic_t        *journal_ref;
2197         struct bkey     *replace_key;
2198 };
2199
2200 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2201 {
2202         struct btree_insert_op *op = container_of(b_op,
2203                                         struct btree_insert_op, op);
2204
2205         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2206                                         op->journal_ref, op->replace_key);
2207         if (ret && !bch_keylist_empty(op->keys))
2208                 return ret;
2209         else
2210                 return MAP_DONE;
2211 }
2212
2213 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2214                      atomic_t *journal_ref, struct bkey *replace_key)
2215 {
2216         struct btree_insert_op op;
2217         int ret = 0;
2218
2219         BUG_ON(current->bio_list);
2220         BUG_ON(bch_keylist_empty(keys));
2221
2222         bch_btree_op_init(&op.op, 0);
2223         op.keys         = keys;
2224         op.journal_ref  = journal_ref;
2225         op.replace_key  = replace_key;
2226
2227         while (!ret && !bch_keylist_empty(keys)) {
2228                 op.op.lock = 0;
2229                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2230                                                &START_KEY(keys->keys),
2231                                                btree_insert_fn);
2232         }
2233
2234         if (ret) {
2235                 struct bkey *k;
2236
2237                 pr_err("error %i", ret);
2238
2239                 while ((k = bch_keylist_pop(keys)))
2240                         bkey_put(c, k);
2241         } else if (op.op.insert_collision)
2242                 ret = -ESRCH;
2243
2244         return ret;
2245 }
2246
2247 void bch_btree_set_root(struct btree *b)
2248 {
2249         unsigned i;
2250         struct closure cl;
2251
2252         closure_init_stack(&cl);
2253
2254         trace_bcache_btree_set_root(b);
2255
2256         BUG_ON(!b->written);
2257
2258         for (i = 0; i < KEY_PTRS(&b->key); i++)
2259                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2260
2261         mutex_lock(&b->c->bucket_lock);
2262         list_del_init(&b->list);
2263         mutex_unlock(&b->c->bucket_lock);
2264
2265         b->c->root = b;
2266
2267         bch_journal_meta(b->c, &cl);
2268         closure_sync(&cl);
2269 }
2270
2271 /* Map across nodes or keys */
2272
2273 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2274                                        struct bkey *from,
2275                                        btree_map_nodes_fn *fn, int flags)
2276 {
2277         int ret = MAP_CONTINUE;
2278
2279         if (b->level) {
2280                 struct bkey *k;
2281                 struct btree_iter iter;
2282
2283                 bch_btree_iter_init(&b->keys, &iter, from);
2284
2285                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2286                                                        bch_ptr_bad))) {
2287                         ret = btree(map_nodes_recurse, k, b,
2288                                     op, from, fn, flags);
2289                         from = NULL;
2290
2291                         if (ret != MAP_CONTINUE)
2292                                 return ret;
2293                 }
2294         }
2295
2296         if (!b->level || flags == MAP_ALL_NODES)
2297                 ret = fn(op, b);
2298
2299         return ret;
2300 }
2301
2302 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2303                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2304 {
2305         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2306 }
2307
2308 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2309                                       struct bkey *from, btree_map_keys_fn *fn,
2310                                       int flags)
2311 {
2312         int ret = MAP_CONTINUE;
2313         struct bkey *k;
2314         struct btree_iter iter;
2315
2316         bch_btree_iter_init(&b->keys, &iter, from);
2317
2318         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2319                 ret = !b->level
2320                         ? fn(op, b, k)
2321                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2322                 from = NULL;
2323
2324                 if (ret != MAP_CONTINUE)
2325                         return ret;
2326         }
2327
2328         if (!b->level && (flags & MAP_END_KEY))
2329                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2330                                      KEY_OFFSET(&b->key), 0));
2331
2332         return ret;
2333 }
2334
2335 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2336                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2337 {
2338         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2339 }
2340
2341 /* Keybuf code */
2342
2343 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2344 {
2345         /* Overlapping keys compare equal */
2346         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2347                 return -1;
2348         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2349                 return 1;
2350         return 0;
2351 }
2352
2353 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2354                                             struct keybuf_key *r)
2355 {
2356         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2357 }
2358
2359 struct refill {
2360         struct btree_op op;
2361         unsigned        nr_found;
2362         struct keybuf   *buf;
2363         struct bkey     *end;
2364         keybuf_pred_fn  *pred;
2365 };
2366
2367 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2368                             struct bkey *k)
2369 {
2370         struct refill *refill = container_of(op, struct refill, op);
2371         struct keybuf *buf = refill->buf;
2372         int ret = MAP_CONTINUE;
2373
2374         if (bkey_cmp(k, refill->end) >= 0) {
2375                 ret = MAP_DONE;
2376                 goto out;
2377         }
2378
2379         if (!KEY_SIZE(k)) /* end key */
2380                 goto out;
2381
2382         if (refill->pred(buf, k)) {
2383                 struct keybuf_key *w;
2384
2385                 spin_lock(&buf->lock);
2386
2387                 w = array_alloc(&buf->freelist);
2388                 if (!w) {
2389                         spin_unlock(&buf->lock);
2390                         return MAP_DONE;
2391                 }
2392
2393                 w->private = NULL;
2394                 bkey_copy(&w->key, k);
2395
2396                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2397                         array_free(&buf->freelist, w);
2398                 else
2399                         refill->nr_found++;
2400
2401                 if (array_freelist_empty(&buf->freelist))
2402                         ret = MAP_DONE;
2403
2404                 spin_unlock(&buf->lock);
2405         }
2406 out:
2407         buf->last_scanned = *k;
2408         return ret;
2409 }
2410
2411 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2412                        struct bkey *end, keybuf_pred_fn *pred)
2413 {
2414         struct bkey start = buf->last_scanned;
2415         struct refill refill;
2416
2417         cond_resched();
2418
2419         bch_btree_op_init(&refill.op, -1);
2420         refill.nr_found = 0;
2421         refill.buf      = buf;
2422         refill.end      = end;
2423         refill.pred     = pred;
2424
2425         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2426                            refill_keybuf_fn, MAP_END_KEY);
2427
2428         trace_bcache_keyscan(refill.nr_found,
2429                              KEY_INODE(&start), KEY_OFFSET(&start),
2430                              KEY_INODE(&buf->last_scanned),
2431                              KEY_OFFSET(&buf->last_scanned));
2432
2433         spin_lock(&buf->lock);
2434
2435         if (!RB_EMPTY_ROOT(&buf->keys)) {
2436                 struct keybuf_key *w;
2437                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2438                 buf->start      = START_KEY(&w->key);
2439
2440                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2441                 buf->end        = w->key;
2442         } else {
2443                 buf->start      = MAX_KEY;
2444                 buf->end        = MAX_KEY;
2445         }
2446
2447         spin_unlock(&buf->lock);
2448 }
2449
2450 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2451 {
2452         rb_erase(&w->node, &buf->keys);
2453         array_free(&buf->freelist, w);
2454 }
2455
2456 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2457 {
2458         spin_lock(&buf->lock);
2459         __bch_keybuf_del(buf, w);
2460         spin_unlock(&buf->lock);
2461 }
2462
2463 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2464                                   struct bkey *end)
2465 {
2466         bool ret = false;
2467         struct keybuf_key *p, *w, s;
2468         s.key = *start;
2469
2470         if (bkey_cmp(end, &buf->start) <= 0 ||
2471             bkey_cmp(start, &buf->end) >= 0)
2472                 return false;
2473
2474         spin_lock(&buf->lock);
2475         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2476
2477         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2478                 p = w;
2479                 w = RB_NEXT(w, node);
2480
2481                 if (p->private)
2482                         ret = true;
2483                 else
2484                         __bch_keybuf_del(buf, p);
2485         }
2486
2487         spin_unlock(&buf->lock);
2488         return ret;
2489 }
2490
2491 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2492 {
2493         struct keybuf_key *w;
2494         spin_lock(&buf->lock);
2495
2496         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2497
2498         while (w && w->private)
2499                 w = RB_NEXT(w, node);
2500
2501         if (w)
2502                 w->private = ERR_PTR(-EINTR);
2503
2504         spin_unlock(&buf->lock);
2505         return w;
2506 }
2507
2508 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2509                                           struct keybuf *buf,
2510                                           struct bkey *end,
2511                                           keybuf_pred_fn *pred)
2512 {
2513         struct keybuf_key *ret;
2514
2515         while (1) {
2516                 ret = bch_keybuf_next(buf);
2517                 if (ret)
2518                         break;
2519
2520                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2521                         pr_debug("scan finished");
2522                         break;
2523                 }
2524
2525                 bch_refill_keybuf(c, buf, end, pred);
2526         }
2527
2528         return ret;
2529 }
2530
2531 void bch_keybuf_init(struct keybuf *buf)
2532 {
2533         buf->last_scanned       = MAX_KEY;
2534         buf->keys               = RB_ROOT;
2535
2536         spin_lock_init(&buf->lock);
2537         array_allocator_init(&buf->freelist);
2538 }