drivers/md/bcache/btree.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  *
   5  * Uses a block device as cache for other block devices; optimized for SSDs.
   6  * All allocation is done in buckets, which should match the erase block size
   7  * of the device.
   8  *
   9  * Buckets containing cached data are kept on a heap sorted by priority;
  10  * bucket priority is increased on cache hit, and periodically all the buckets
  11  * on the heap have their priority scaled down. This currently is just used as
  12  * an LRU but in the future should allow for more intelligent heuristics.
  13  *
  14  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  15  * counter. Garbage collection is used to remove stale pointers.
  16  *
  17  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  18  * as keys are inserted we only sort the pages that have not yet been written.
  19  * When garbage collection is run, we resort the entire node.
  20  *
  21  * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
  22  */
  23
  24 #include "bcache.h"
  25 #include "btree.h"
  26 #include "debug.h"
  27 #include "extents.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/random.h>
  35 #include <linux/rcupdate.h>
  36 #include <linux/sched/clock.h>
  37 #include <linux/rculist.h>
  38
  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 (!bch_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                 /* No problem for multipage bvec since the bio is just allocated */
 436                 b->bio->bi_vcnt = 0;
 437                 bch_bio_map(b->bio, i);
 438
 439                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 440
 441                 closure_sync(cl);
 442                 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
 443         }
 444 }
 445
 446 void __bch_btree_node_write(struct btree *b, struct closure *parent)
 447 {
 448         struct bset *i = btree_bset_last(b);
 449
 450         lockdep_assert_held(&b->write_lock);
 451
 452         trace_bcache_btree_write(b);
 453
 454         BUG_ON(current->bio_list);
 455         BUG_ON(b->written >= btree_blocks(b));
 456         BUG_ON(b->written && !i->keys);
 457         BUG_ON(btree_bset_first(b)->seq != i->seq);
 458         bch_check_keys(&b->keys, "writing");
 459
 460         cancel_delayed_work(&b->work);
 461
 462         /* If caller isn't waiting for write, parent refcount is cache set */
 463         down(&b->io_mutex);
 464         closure_init(&b->io, parent ?: &b->c->cl);
 465
 466         clear_bit(BTREE_NODE_dirty,      &b->flags);
 467         change_bit(BTREE_NODE_write_idx, &b->flags);
 468
 469         do_btree_node_write(b);
 470
 471         atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
 472                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 473
 474         b->written += set_blocks(i, block_bytes(b->c));
 475 }
 476
 477 void bch_btree_node_write(struct btree *b, struct closure *parent)
 478 {
 479         unsigned nsets = b->keys.nsets;
 480
 481         lockdep_assert_held(&b->lock);
 482
 483         __bch_btree_node_write(b, parent);
 484
 485         /*
 486          * do verify if there was more than one set initially (i.e. we did a
 487          * sort) and we sorted down to a single set:
 488          */
 489         if (nsets && !b->keys.nsets)
 490                 bch_btree_verify(b);
 491
 492         bch_btree_init_next(b);
 493 }
 494
 495 static void bch_btree_node_write_sync(struct btree *b)
 496 {
 497         struct closure cl;
 498
 499         closure_init_stack(&cl);
 500
 501         mutex_lock(&b->write_lock);
 502         bch_btree_node_write(b, &cl);
 503         mutex_unlock(&b->write_lock);
 504
 505         closure_sync(&cl);
 506 }
 507
 508 static void btree_node_write_work(struct work_struct *w)
 509 {
 510         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 511
 512         mutex_lock(&b->write_lock);
 513         if (btree_node_dirty(b))
 514                 __bch_btree_node_write(b, NULL);
 515         mutex_unlock(&b->write_lock);
 516 }
 517
 518 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
 519 {
 520         struct bset *i = btree_bset_last(b);
 521         struct btree_write *w = btree_current_write(b);
 522
 523         lockdep_assert_held(&b->write_lock);
 524
 525         BUG_ON(!b->written);
 526         BUG_ON(!i->keys);
 527
 528         if (!btree_node_dirty(b))
 529                 schedule_delayed_work(&b->work, 30 * HZ);
 530
 531         set_btree_node_dirty(b);
 532
 533         if (journal_ref) {
 534                 if (w->journal &&
 535                     journal_pin_cmp(b->c, w->journal, journal_ref)) {
 536                         atomic_dec_bug(w->journal);
 537                         w->journal = NULL;
 538                 }
 539
 540                 if (!w->journal) {
 541                         w->journal = journal_ref;
 542                         atomic_inc(w->journal);
 543                 }
 544         }
 545
 546         /* Force write if set is too big */
 547         if (set_bytes(i) > PAGE_SIZE - 48 &&
 548             !current->bio_list)
 549                 bch_btree_node_write(b, NULL);
 550 }
 551
 552 /*
 553  * Btree in memory cache - allocation/freeing
 554  * mca -> memory cache
 555  */
 556
 557 #define mca_reserve(c)  (((c->root && c->root->level)           \
 558                           ? c->root->level : 1) * 8 + 16)
 559 #define mca_can_free(c)                                         \
 560         max_t(int, 0, c->btree_cache_used - mca_reserve(c))
 561
 562 static void mca_data_free(struct btree *b)
 563 {
 564         BUG_ON(b->io_mutex.count != 1);
 565
 566         bch_btree_keys_free(&b->keys);
 567
 568         b->c->btree_cache_used--;
 569         list_move(&b->list, &b->c->btree_cache_freed);
 570 }
 571
 572 static void mca_bucket_free(struct btree *b)
 573 {
 574         BUG_ON(btree_node_dirty(b));
 575
 576         b->key.ptr[0] = 0;
 577         hlist_del_init_rcu(&b->hash);
 578         list_move(&b->list, &b->c->btree_cache_freeable);
 579 }
 580
 581 static unsigned btree_order(struct bkey *k)
 582 {
 583         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 584 }
 585
 586 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 587 {
 588         if (!bch_btree_keys_alloc(&b->keys,
 589                                   max_t(unsigned,
 590                                         ilog2(b->c->btree_pages),
 591                                         btree_order(k)),
 592                                   gfp)) {
 593                 b->c->btree_cache_used++;
 594                 list_move(&b->list, &b->c->btree_cache);
 595         } else {
 596                 list_move(&b->list, &b->c->btree_cache_freed);
 597         }
 598 }
 599
 600 static struct btree *mca_bucket_alloc(struct cache_set *c,
 601                                       struct bkey *k, gfp_t gfp)
 602 {
 603         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 604         if (!b)
 605                 return NULL;
 606
 607         init_rwsem(&b->lock);
 608         lockdep_set_novalidate_class(&b->lock);
 609         mutex_init(&b->write_lock);
 610         lockdep_set_novalidate_class(&b->write_lock);
 611         INIT_LIST_HEAD(&b->list);
 612         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 613         b->c = c;
 614         sema_init(&b->io_mutex, 1);
 615
 616         mca_data_alloc(b, k, gfp);
 617         return b;
 618 }
 619
 620 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
 621 {
 622         struct closure cl;
 623
 624         closure_init_stack(&cl);
 625         lockdep_assert_held(&b->c->bucket_lock);
 626
 627         if (!down_write_trylock(&b->lock))
 628                 return -ENOMEM;
 629
 630         BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
 631
 632         if (b->keys.page_order < min_order)
 633                 goto out_unlock;
 634
 635         if (!flush) {
 636                 if (btree_node_dirty(b))
 637                         goto out_unlock;
 638
 639                 if (down_trylock(&b->io_mutex))
 640                         goto out_unlock;
 641                 up(&b->io_mutex);
 642         }
 643
 644         mutex_lock(&b->write_lock);
 645         if (btree_node_dirty(b))
 646                 __bch_btree_node_write(b, &cl);
 647         mutex_unlock(&b->write_lock);
 648
 649         closure_sync(&cl);
 650
 651         /* wait for any in flight btree write */
 652         down(&b->io_mutex);
 653         up(&b->io_mutex);
 654
 655         return 0;
 656 out_unlock:
 657         rw_unlock(true, b);
 658         return -ENOMEM;
 659 }
 660
 661 static unsigned long bch_mca_scan(struct shrinker *shrink,
 662                                   struct shrink_control *sc)
 663 {
 664         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 665         struct btree *b, *t;
 666         unsigned long i, nr = sc->nr_to_scan;
 667         unsigned long freed = 0;
 668         unsigned int btree_cache_used;
 669
 670         if (c->shrinker_disabled)
 671                 return SHRINK_STOP;
 672
 673         if (c->btree_cache_alloc_lock)
 674                 return SHRINK_STOP;
 675
 676         /* Return -1 if we can't do anything right now */
 677         if (sc->gfp_mask & __GFP_IO)
 678                 mutex_lock(&c->bucket_lock);
 679         else if (!mutex_trylock(&c->bucket_lock))
 680                 return -1;
 681
 682         /*
 683          * It's _really_ critical that we don't free too many btree nodes - we
 684          * have to always leave ourselves a reserve. The reserve is how we
 685          * guarantee that allocating memory for a new btree node can always
 686          * succeed, so that inserting keys into the btree can always succeed and
 687          * IO can always make forward progress:
 688          */
 689         nr /= c->btree_pages;
 690         nr = min_t(unsigned long, nr, mca_can_free(c));
 691
 692         i = 0;
 693         btree_cache_used = c->btree_cache_used;
 694         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 695                 if (nr <= 0)
 696                         goto out;
 697
 698                 if (++i > 3 &&
 699                     !mca_reap(b, 0, false)) {
 700                         mca_data_free(b);
 701                         rw_unlock(true, b);
 702                         freed++;
 703                 }
 704                 nr--;
 705         }
 706
 707         for (;  (nr--) && i < btree_cache_used; i++) {
 708                 if (list_empty(&c->btree_cache))
 709                         goto out;
 710
 711                 b = list_first_entry(&c->btree_cache, struct btree, list);
 712                 list_rotate_left(&c->btree_cache);
 713
 714                 if (!b->accessed &&
 715                     !mca_reap(b, 0, false)) {
 716                         mca_bucket_free(b);
 717                         mca_data_free(b);
 718                         rw_unlock(true, b);
 719                         freed++;
 720                 } else
 721                         b->accessed = 0;
 722         }
 723 out:
 724         mutex_unlock(&c->bucket_lock);
 725         return freed * c->btree_pages;
 726 }
 727
 728 static unsigned long bch_mca_count(struct shrinker *shrink,
 729                                    struct shrink_control *sc)
 730 {
 731         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 732
 733         if (c->shrinker_disabled)
 734                 return 0;
 735
 736         if (c->btree_cache_alloc_lock)
 737                 return 0;
 738
 739         return mca_can_free(c) * c->btree_pages;
 740 }
 741
 742 void bch_btree_cache_free(struct cache_set *c)
 743 {
 744         struct btree *b;
 745         struct closure cl;
 746         closure_init_stack(&cl);
 747
 748         if (c->shrink.list.next)
 749                 unregister_shrinker(&c->shrink);
 750
 751         mutex_lock(&c->bucket_lock);
 752
 753 #ifdef CONFIG_BCACHE_DEBUG
 754         if (c->verify_data)
 755                 list_move(&c->verify_data->list, &c->btree_cache);
 756
 757         free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
 758 #endif
 759
 760         list_splice(&c->btree_cache_freeable,
 761                     &c->btree_cache);
 762
 763         while (!list_empty(&c->btree_cache)) {
 764                 b = list_first_entry(&c->btree_cache, struct btree, list);
 765
 766                 if (btree_node_dirty(b))
 767                         btree_complete_write(b, btree_current_write(b));
 768                 clear_bit(BTREE_NODE_dirty, &b->flags);
 769
 770                 mca_data_free(b);
 771         }
 772
 773         while (!list_empty(&c->btree_cache_freed)) {
 774                 b = list_first_entry(&c->btree_cache_freed,
 775                                      struct btree, list);
 776                 list_del(&b->list);
 777                 cancel_delayed_work_sync(&b->work);
 778                 kfree(b);
 779         }
 780
 781         mutex_unlock(&c->bucket_lock);
 782 }
 783
 784 int bch_btree_cache_alloc(struct cache_set *c)
 785 {
 786         unsigned i;
 787
 788         for (i = 0; i < mca_reserve(c); i++)
 789                 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
 790                         return -ENOMEM;
 791
 792         list_splice_init(&c->btree_cache,
 793                          &c->btree_cache_freeable);
 794
 795 #ifdef CONFIG_BCACHE_DEBUG
 796         mutex_init(&c->verify_lock);
 797
 798         c->verify_ondisk = (void *)
 799                 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
 800
 801         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 802
 803         if (c->verify_data &&
 804             c->verify_data->keys.set->data)
 805                 list_del_init(&c->verify_data->list);
 806         else
 807                 c->verify_data = NULL;
 808 #endif
 809
 810         c->shrink.count_objects = bch_mca_count;
 811         c->shrink.scan_objects = bch_mca_scan;
 812         c->shrink.seeks = 4;
 813         c->shrink.batch = c->btree_pages * 2;
 814
 815         if (register_shrinker(&c->shrink))
 816                 pr_warn("bcache: %s: could not register shrinker",
 817                                 __func__);
 818
 819         return 0;
 820 }
 821
 822 /* Btree in memory cache - hash table */
 823
 824 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 825 {
 826         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 827 }
 828
 829 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 830 {
 831         struct btree *b;
 832
 833         rcu_read_lock();
 834         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 835                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 836                         goto out;
 837         b = NULL;
 838 out:
 839         rcu_read_unlock();
 840         return b;
 841 }
 842
 843 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
 844 {
 845         struct task_struct *old;
 846
 847         old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
 848         if (old && old != current) {
 849                 if (op)
 850                         prepare_to_wait(&c->btree_cache_wait, &op->wait,
 851                                         TASK_UNINTERRUPTIBLE);
 852                 return -EINTR;
 853         }
 854
 855         return 0;
 856 }
 857
 858 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
 859                                      struct bkey *k)
 860 {
 861         struct btree *b;
 862
 863         trace_bcache_btree_cache_cannibalize(c);
 864
 865         if (mca_cannibalize_lock(c, op))
 866                 return ERR_PTR(-EINTR);
 867
 868         list_for_each_entry_reverse(b, &c->btree_cache, list)
 869                 if (!mca_reap(b, btree_order(k), false))
 870                         return b;
 871
 872         list_for_each_entry_reverse(b, &c->btree_cache, list)
 873                 if (!mca_reap(b, btree_order(k), true))
 874                         return b;
 875
 876         WARN(1, "btree cache cannibalize failed\n");
 877         return ERR_PTR(-ENOMEM);
 878 }
 879
 880 /*
 881  * We can only have one thread cannibalizing other cached btree nodes at a time,
 882  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 883  * cannibalize_bucket() will take. This means every time we unlock the root of
 884  * the btree, we need to release this lock if we have it held.
 885  */
 886 static void bch_cannibalize_unlock(struct cache_set *c)
 887 {
 888         if (c->btree_cache_alloc_lock == current) {
 889                 c->btree_cache_alloc_lock = NULL;
 890                 wake_up(&c->btree_cache_wait);
 891         }
 892 }
 893
 894 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
 895                                struct bkey *k, int level)
 896 {
 897         struct btree *b;
 898
 899         BUG_ON(current->bio_list);
 900
 901         lockdep_assert_held(&c->bucket_lock);
 902
 903         if (mca_find(c, k))
 904                 return NULL;
 905
 906         /* btree_free() doesn't free memory; it sticks the node on the end of
 907          * the list. Check if there's any freed nodes there:
 908          */
 909         list_for_each_entry(b, &c->btree_cache_freeable, list)
 910                 if (!mca_reap(b, btree_order(k), false))
 911                         goto out;
 912
 913         /* We never free struct btree itself, just the memory that holds the on
 914          * disk node. Check the freed list before allocating a new one:
 915          */
 916         list_for_each_entry(b, &c->btree_cache_freed, list)
 917                 if (!mca_reap(b, 0, false)) {
 918                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 919                         if (!b->keys.set[0].data)
 920                                 goto err;
 921                         else
 922                                 goto out;
 923                 }
 924
 925         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 926         if (!b)
 927                 goto err;
 928
 929         BUG_ON(!down_write_trylock(&b->lock));
 930         if (!b->keys.set->data)
 931                 goto err;
 932 out:
 933         BUG_ON(b->io_mutex.count != 1);
 934
 935         bkey_copy(&b->key, k);
 936         list_move(&b->list, &c->btree_cache);
 937         hlist_del_init_rcu(&b->hash);
 938         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 939
 940         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 941         b->parent       = (void *) ~0UL;
 942         b->flags        = 0;
 943         b->written      = 0;
 944         b->level        = level;
 945
 946         if (!b->level)
 947                 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
 948                                     &b->c->expensive_debug_checks);
 949         else
 950                 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
 951                                     &b->c->expensive_debug_checks);
 952
 953         return b;
 954 err:
 955         if (b)
 956                 rw_unlock(true, b);
 957
 958         b = mca_cannibalize(c, op, k);
 959         if (!IS_ERR(b))
 960                 goto out;
 961
 962         return b;
 963 }
 964
 965 /*
 966  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 967  * in from disk if necessary.
 968  *
 969  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
 970  *
 971  * The btree node will have either a read or a write lock held, depending on
 972  * level and op->lock.
 973  */
 974 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
 975                                  struct bkey *k, int level, bool write,
 976                                  struct btree *parent)
 977 {
 978         int i = 0;
 979         struct btree *b;
 980
 981         BUG_ON(level < 0);
 982 retry:
 983         b = mca_find(c, k);
 984
 985         if (!b) {
 986                 if (current->bio_list)
 987                         return ERR_PTR(-EAGAIN);
 988
 989                 mutex_lock(&c->bucket_lock);
 990                 b = mca_alloc(c, op, k, level);
 991                 mutex_unlock(&c->bucket_lock);
 992
 993                 if (!b)
 994                         goto retry;
 995                 if (IS_ERR(b))
 996                         return b;
 997
 998                 bch_btree_node_read(b);
 999
1000                 if (!write)
1001                         downgrade_write(&b->lock);
1002         } else {
1003                 rw_lock(write, b, level);
1004                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1005                         rw_unlock(write, b);
1006                         goto retry;
1007                 }
1008                 BUG_ON(b->level != level);
1009         }
1010
1011         b->parent = parent;
1012         b->accessed = 1;
1013
1014         for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1015                 prefetch(b->keys.set[i].tree);
1016                 prefetch(b->keys.set[i].data);
1017         }
1018
1019         for (; i <= b->keys.nsets; i++)
1020                 prefetch(b->keys.set[i].data);
1021
1022         if (btree_node_io_error(b)) {
1023                 rw_unlock(write, b);
1024                 return ERR_PTR(-EIO);
1025         }
1026
1027         BUG_ON(!b->written);
1028
1029         return b;
1030 }
1031
1032 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1033 {
1034         struct btree *b;
1035
1036         mutex_lock(&parent->c->bucket_lock);
1037         b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1038         mutex_unlock(&parent->c->bucket_lock);
1039
1040         if (!IS_ERR_OR_NULL(b)) {
1041                 b->parent = parent;
1042                 bch_btree_node_read(b);
1043                 rw_unlock(true, b);
1044         }
1045 }
1046
1047 /* Btree alloc */
1048
1049 static void btree_node_free(struct btree *b)
1050 {
1051         trace_bcache_btree_node_free(b);
1052
1053         BUG_ON(b == b->c->root);
1054
1055         mutex_lock(&b->write_lock);
1056
1057         if (btree_node_dirty(b))
1058                 btree_complete_write(b, btree_current_write(b));
1059         clear_bit(BTREE_NODE_dirty, &b->flags);
1060
1061         mutex_unlock(&b->write_lock);
1062
1063         cancel_delayed_work(&b->work);
1064
1065         mutex_lock(&b->c->bucket_lock);
1066         bch_bucket_free(b->c, &b->key);
1067         mca_bucket_free(b);
1068         mutex_unlock(&b->c->bucket_lock);
1069 }
1070
1071 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1072                                      int level, bool wait,
1073                                      struct btree *parent)
1074 {
1075         BKEY_PADDED(key) k;
1076         struct btree *b = ERR_PTR(-EAGAIN);
1077
1078         mutex_lock(&c->bucket_lock);
1079 retry:
1080         if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1081                 goto err;
1082
1083         bkey_put(c, &k.key);
1084         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1085
1086         b = mca_alloc(c, op, &k.key, level);
1087         if (IS_ERR(b))
1088                 goto err_free;
1089
1090         if (!b) {
1091                 cache_bug(c,
1092                         "Tried to allocate bucket that was in btree cache");
1093                 goto retry;
1094         }
1095
1096         b->accessed = 1;
1097         b->parent = parent;
1098         bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1099
1100         mutex_unlock(&c->bucket_lock);
1101
1102         trace_bcache_btree_node_alloc(b);
1103         return b;
1104 err_free:
1105         bch_bucket_free(c, &k.key);
1106 err:
1107         mutex_unlock(&c->bucket_lock);
1108
1109         trace_bcache_btree_node_alloc_fail(c);
1110         return b;
1111 }
1112
1113 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1114                                           struct btree_op *op, int level,
1115                                           struct btree *parent)
1116 {
1117         return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1118 }
1119
1120 static struct btree *btree_node_alloc_replacement(struct btree *b,
1121                                                   struct btree_op *op)
1122 {
1123         struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1124         if (!IS_ERR_OR_NULL(n)) {
1125                 mutex_lock(&n->write_lock);
1126                 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1127                 bkey_copy_key(&n->key, &b->key);
1128                 mutex_unlock(&n->write_lock);
1129         }
1130
1131         return n;
1132 }
1133
1134 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1135 {
1136         unsigned i;
1137
1138         mutex_lock(&b->c->bucket_lock);
1139
1140         atomic_inc(&b->c->prio_blocked);
1141
1142         bkey_copy(k, &b->key);
1143         bkey_copy_key(k, &ZERO_KEY);
1144
1145         for (i = 0; i < KEY_PTRS(k); i++)
1146                 SET_PTR_GEN(k, i,
1147                             bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1148                                         PTR_BUCKET(b->c, &b->key, i)));
1149
1150         mutex_unlock(&b->c->bucket_lock);
1151 }
1152
1153 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1154 {
1155         struct cache_set *c = b->c;
1156         struct cache *ca;
1157         unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1158
1159         mutex_lock(&c->bucket_lock);
1160
1161         for_each_cache(ca, c, i)
1162                 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1163                         if (op)
1164                                 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1165                                                 TASK_UNINTERRUPTIBLE);
1166                         mutex_unlock(&c->bucket_lock);
1167                         return -EINTR;
1168                 }
1169
1170         mutex_unlock(&c->bucket_lock);
1171
1172         return mca_cannibalize_lock(b->c, op);
1173 }
1174
1175 /* Garbage collection */
1176
1177 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1178                                     struct bkey *k)
1179 {
1180         uint8_t stale = 0;
1181         unsigned i;
1182         struct bucket *g;
1183
1184         /*
1185          * ptr_invalid() can't return true for the keys that mark btree nodes as
1186          * freed, but since ptr_bad() returns true we'll never actually use them
1187          * for anything and thus we don't want mark their pointers here
1188          */
1189         if (!bkey_cmp(k, &ZERO_KEY))
1190                 return stale;
1191
1192         for (i = 0; i < KEY_PTRS(k); i++) {
1193                 if (!ptr_available(c, k, i))
1194                         continue;
1195
1196                 g = PTR_BUCKET(c, k, i);
1197
1198                 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1199                         g->last_gc = PTR_GEN(k, i);
1200
1201                 if (ptr_stale(c, k, i)) {
1202                         stale = max(stale, ptr_stale(c, k, i));
1203                         continue;
1204                 }
1205
1206                 cache_bug_on(GC_MARK(g) &&
1207                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1208                              c, "inconsistent ptrs: mark = %llu, level = %i",
1209                              GC_MARK(g), level);
1210
1211                 if (level)
1212                         SET_GC_MARK(g, GC_MARK_METADATA);
1213                 else if (KEY_DIRTY(k))
1214                         SET_GC_MARK(g, GC_MARK_DIRTY);
1215                 else if (!GC_MARK(g))
1216                         SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1217
1218                 /* guard against overflow */
1219                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1220                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1221                                              MAX_GC_SECTORS_USED));
1222
1223                 BUG_ON(!GC_SECTORS_USED(g));
1224         }
1225
1226         return stale;
1227 }
1228
1229 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1230
1231 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1232 {
1233         unsigned i;
1234
1235         for (i = 0; i < KEY_PTRS(k); i++)
1236                 if (ptr_available(c, k, i) &&
1237                     !ptr_stale(c, k, i)) {
1238                         struct bucket *b = PTR_BUCKET(c, k, i);
1239
1240                         b->gen = PTR_GEN(k, i);
1241
1242                         if (level && bkey_cmp(k, &ZERO_KEY))
1243                                 b->prio = BTREE_PRIO;
1244                         else if (!level && b->prio == BTREE_PRIO)
1245                                 b->prio = INITIAL_PRIO;
1246                 }
1247
1248         __bch_btree_mark_key(c, level, k);
1249 }
1250
1251 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1252 {
1253         stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1254 }
1255
1256 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1257 {
1258         uint8_t stale = 0;
1259         unsigned keys = 0, good_keys = 0;
1260         struct bkey *k;
1261         struct btree_iter iter;
1262         struct bset_tree *t;
1263
1264         gc->nodes++;
1265
1266         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1267                 stale = max(stale, btree_mark_key(b, k));
1268                 keys++;
1269
1270                 if (bch_ptr_bad(&b->keys, k))
1271                         continue;
1272
1273                 gc->key_bytes += bkey_u64s(k);
1274                 gc->nkeys++;
1275                 good_keys++;
1276
1277                 gc->data += KEY_SIZE(k);
1278         }
1279
1280         for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1281                 btree_bug_on(t->size &&
1282                              bset_written(&b->keys, t) &&
1283                              bkey_cmp(&b->key, &t->end) < 0,
1284                              b, "found short btree key in gc");
1285
1286         if (b->c->gc_always_rewrite)
1287                 return true;
1288
1289         if (stale > 10)
1290                 return true;
1291
1292         if ((keys - good_keys) * 2 > keys)
1293                 return true;
1294
1295         return false;
1296 }
1297
1298 #define GC_MERGE_NODES  4U
1299
1300 struct gc_merge_info {
1301         struct btree    *b;
1302         unsigned        keys;
1303 };
1304
1305 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1306                                  struct keylist *, atomic_t *, struct bkey *);
1307
1308 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1309                              struct gc_stat *gc, struct gc_merge_info *r)
1310 {
1311         unsigned i, nodes = 0, keys = 0, blocks;
1312         struct btree *new_nodes[GC_MERGE_NODES];
1313         struct keylist keylist;
1314         struct closure cl;
1315         struct bkey *k;
1316
1317         bch_keylist_init(&keylist);
1318
1319         if (btree_check_reserve(b, NULL))
1320                 return 0;
1321
1322         memset(new_nodes, 0, sizeof(new_nodes));
1323         closure_init_stack(&cl);
1324
1325         while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1326                 keys += r[nodes++].keys;
1327
1328         blocks = btree_default_blocks(b->c) * 2 / 3;
1329
1330         if (nodes < 2 ||
1331             __set_blocks(b->keys.set[0].data, keys,
1332                          block_bytes(b->c)) > blocks * (nodes - 1))
1333                 return 0;
1334
1335         for (i = 0; i < nodes; i++) {
1336                 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1337                 if (IS_ERR_OR_NULL(new_nodes[i]))
1338                         goto out_nocoalesce;
1339         }
1340
1341         /*
1342          * We have to check the reserve here, after we've allocated our new
1343          * nodes, to make sure the insert below will succeed - we also check
1344          * before as an optimization to potentially avoid a bunch of expensive
1345          * allocs/sorts
1346          */
1347         if (btree_check_reserve(b, NULL))
1348                 goto out_nocoalesce;
1349
1350         for (i = 0; i < nodes; i++)
1351                 mutex_lock(&new_nodes[i]->write_lock);
1352
1353         for (i = nodes - 1; i > 0; --i) {
1354                 struct bset *n1 = btree_bset_first(new_nodes[i]);
1355                 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1356                 struct bkey *k, *last = NULL;
1357
1358                 keys = 0;
1359
1360                 if (i > 1) {
1361                         for (k = n2->start;
1362                              k < bset_bkey_last(n2);
1363                              k = bkey_next(k)) {
1364                                 if (__set_blocks(n1, n1->keys + keys +
1365                                                  bkey_u64s(k),
1366                                                  block_bytes(b->c)) > blocks)
1367                                         break;
1368
1369                                 last = k;
1370                                 keys += bkey_u64s(k);
1371                         }
1372                 } else {
1373                         /*
1374                          * Last node we're not getting rid of - we're getting
1375                          * rid of the node at r[0]. Have to try and fit all of
1376                          * the remaining keys into this node; we can't ensure
1377                          * they will always fit due to rounding and variable
1378                          * length keys (shouldn't be possible in practice,
1379                          * though)
1380                          */
1381                         if (__set_blocks(n1, n1->keys + n2->keys,
1382                                          block_bytes(b->c)) >
1383                             btree_blocks(new_nodes[i]))
1384                                 goto out_nocoalesce;
1385
1386                         keys = n2->keys;
1387                         /* Take the key of the node we're getting rid of */
1388                         last = &r->b->key;
1389                 }
1390
1391                 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1392                        btree_blocks(new_nodes[i]));
1393
1394                 if (last)
1395                         bkey_copy_key(&new_nodes[i]->key, last);
1396
1397                 memcpy(bset_bkey_last(n1),
1398                        n2->start,
1399                        (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1400
1401                 n1->keys += keys;
1402                 r[i].keys = n1->keys;
1403
1404                 memmove(n2->start,
1405                         bset_bkey_idx(n2, keys),
1406                         (void *) bset_bkey_last(n2) -
1407                         (void *) bset_bkey_idx(n2, keys));
1408
1409                 n2->keys -= keys;
1410
1411                 if (__bch_keylist_realloc(&keylist,
1412                                           bkey_u64s(&new_nodes[i]->key)))
1413                         goto out_nocoalesce;
1414
1415                 bch_btree_node_write(new_nodes[i], &cl);
1416                 bch_keylist_add(&keylist, &new_nodes[i]->key);
1417         }
1418
1419         for (i = 0; i < nodes; i++)
1420                 mutex_unlock(&new_nodes[i]->write_lock);
1421
1422         closure_sync(&cl);
1423
1424         /* We emptied out this node */
1425         BUG_ON(btree_bset_first(new_nodes[0])->keys);
1426         btree_node_free(new_nodes[0]);
1427         rw_unlock(true, new_nodes[0]);
1428         new_nodes[0] = NULL;
1429
1430         for (i = 0; i < nodes; i++) {
1431                 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1432                         goto out_nocoalesce;
1433
1434                 make_btree_freeing_key(r[i].b, keylist.top);
1435                 bch_keylist_push(&keylist);
1436         }
1437
1438         bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1439         BUG_ON(!bch_keylist_empty(&keylist));
1440
1441         for (i = 0; i < nodes; i++) {
1442                 btree_node_free(r[i].b);
1443                 rw_unlock(true, r[i].b);
1444
1445                 r[i].b = new_nodes[i];
1446         }
1447
1448         memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1449         r[nodes - 1].b = ERR_PTR(-EINTR);
1450
1451         trace_bcache_btree_gc_coalesce(nodes);
1452         gc->nodes--;
1453
1454         bch_keylist_free(&keylist);
1455
1456         /* Invalidated our iterator */
1457         return -EINTR;
1458
1459 out_nocoalesce:
1460         closure_sync(&cl);
1461         bch_keylist_free(&keylist);
1462
1463         while ((k = bch_keylist_pop(&keylist)))
1464                 if (!bkey_cmp(k, &ZERO_KEY))
1465                         atomic_dec(&b->c->prio_blocked);
1466
1467         for (i = 0; i < nodes; i++)
1468                 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1469                         btree_node_free(new_nodes[i]);
1470                         rw_unlock(true, new_nodes[i]);
1471                 }
1472         return 0;
1473 }
1474
1475 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1476                                  struct btree *replace)
1477 {
1478         struct keylist keys;
1479         struct btree *n;
1480
1481         if (btree_check_reserve(b, NULL))
1482                 return 0;
1483
1484         n = btree_node_alloc_replacement(replace, NULL);
1485
1486         /* recheck reserve after allocating replacement node */
1487         if (btree_check_reserve(b, NULL)) {
1488                 btree_node_free(n);
1489                 rw_unlock(true, n);
1490                 return 0;
1491         }
1492
1493         bch_btree_node_write_sync(n);
1494
1495         bch_keylist_init(&keys);
1496         bch_keylist_add(&keys, &n->key);
1497
1498         make_btree_freeing_key(replace, keys.top);
1499         bch_keylist_push(&keys);
1500
1501         bch_btree_insert_node(b, op, &keys, NULL, NULL);
1502         BUG_ON(!bch_keylist_empty(&keys));
1503
1504         btree_node_free(replace);
1505         rw_unlock(true, n);
1506
1507         /* Invalidated our iterator */
1508         return -EINTR;
1509 }
1510
1511 static unsigned btree_gc_count_keys(struct btree *b)
1512 {
1513         struct bkey *k;
1514         struct btree_iter iter;
1515         unsigned ret = 0;
1516
1517         for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1518                 ret += bkey_u64s(k);
1519
1520         return ret;
1521 }
1522
1523 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1524                             struct closure *writes, struct gc_stat *gc)
1525 {
1526         int ret = 0;
1527         bool should_rewrite;
1528         struct bkey *k;
1529         struct btree_iter iter;
1530         struct gc_merge_info r[GC_MERGE_NODES];
1531         struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1532
1533         bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1534
1535         for (i = r; i < r + ARRAY_SIZE(r); i++)
1536                 i->b = ERR_PTR(-EINTR);
1537
1538         while (1) {
1539                 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1540                 if (k) {
1541                         r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1542                                                   true, b);
1543                         if (IS_ERR(r->b)) {
1544                                 ret = PTR_ERR(r->b);
1545                                 break;
1546                         }
1547
1548                         r->keys = btree_gc_count_keys(r->b);
1549
1550                         ret = btree_gc_coalesce(b, op, gc, r);
1551                         if (ret)
1552                                 break;
1553                 }
1554
1555                 if (!last->b)
1556                         break;
1557
1558                 if (!IS_ERR(last->b)) {
1559                         should_rewrite = btree_gc_mark_node(last->b, gc);
1560                         if (should_rewrite) {
1561                                 ret = btree_gc_rewrite_node(b, op, last->b);
1562                                 if (ret)
1563                                         break;
1564                         }
1565
1566                         if (last->b->level) {
1567                                 ret = btree_gc_recurse(last->b, op, writes, gc);
1568                                 if (ret)
1569                                         break;
1570                         }
1571
1572                         bkey_copy_key(&b->c->gc_done, &last->b->key);
1573
1574                         /*
1575                          * Must flush leaf nodes before gc ends, since replace
1576                          * operations aren't journalled
1577                          */
1578                         mutex_lock(&last->b->write_lock);
1579                         if (btree_node_dirty(last->b))
1580                                 bch_btree_node_write(last->b, writes);
1581                         mutex_unlock(&last->b->write_lock);
1582                         rw_unlock(true, last->b);
1583                 }
1584
1585                 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1586                 r->b = NULL;
1587
1588                 if (need_resched()) {
1589                         ret = -EAGAIN;
1590                         break;
1591                 }
1592         }
1593
1594         for (i = r; i < r + ARRAY_SIZE(r); i++)
1595                 if (!IS_ERR_OR_NULL(i->b)) {
1596                         mutex_lock(&i->b->write_lock);
1597                         if (btree_node_dirty(i->b))
1598                                 bch_btree_node_write(i->b, writes);
1599                         mutex_unlock(&i->b->write_lock);
1600                         rw_unlock(true, i->b);
1601                 }
1602
1603         return ret;
1604 }
1605
1606 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1607                              struct closure *writes, struct gc_stat *gc)
1608 {
1609         struct btree *n = NULL;
1610         int ret = 0;
1611         bool should_rewrite;
1612
1613         should_rewrite = btree_gc_mark_node(b, gc);
1614         if (should_rewrite) {
1615                 n = btree_node_alloc_replacement(b, NULL);
1616
1617                 if (!IS_ERR_OR_NULL(n)) {
1618                         bch_btree_node_write_sync(n);
1619
1620                         bch_btree_set_root(n);
1621                         btree_node_free(b);
1622                         rw_unlock(true, n);
1623
1624                         return -EINTR;
1625                 }
1626         }
1627
1628         __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1629
1630         if (b->level) {
1631                 ret = btree_gc_recurse(b, op, writes, gc);
1632                 if (ret)
1633                         return ret;
1634         }
1635
1636         bkey_copy_key(&b->c->gc_done, &b->key);
1637
1638         return ret;
1639 }
1640
1641 static void btree_gc_start(struct cache_set *c)
1642 {
1643         struct cache *ca;
1644         struct bucket *b;
1645         unsigned i;
1646
1647         if (!c->gc_mark_valid)
1648                 return;
1649
1650         mutex_lock(&c->bucket_lock);
1651
1652         c->gc_mark_valid = 0;
1653         c->gc_done = ZERO_KEY;
1654
1655         for_each_cache(ca, c, i)
1656                 for_each_bucket(b, ca) {
1657                         b->last_gc = b->gen;
1658                         if (!atomic_read(&b->pin)) {
1659                                 SET_GC_MARK(b, 0);
1660                                 SET_GC_SECTORS_USED(b, 0);
1661                         }
1662                 }
1663
1664         mutex_unlock(&c->bucket_lock);
1665 }
1666
1667 static void bch_btree_gc_finish(struct cache_set *c)
1668 {
1669         struct bucket *b;
1670         struct cache *ca;
1671         unsigned i;
1672
1673         mutex_lock(&c->bucket_lock);
1674
1675         set_gc_sectors(c);
1676         c->gc_mark_valid = 1;
1677         c->need_gc      = 0;
1678
1679         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1680                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1681                             GC_MARK_METADATA);
1682
1683         /* don't reclaim buckets to which writeback keys point */
1684         rcu_read_lock();
1685         for (i = 0; i < c->devices_max_used; i++) {
1686                 struct bcache_device *d = c->devices[i];
1687                 struct cached_dev *dc;
1688                 struct keybuf_key *w, *n;
1689                 unsigned j;
1690
1691                 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1692                         continue;
1693                 dc = container_of(d, struct cached_dev, disk);
1694
1695                 spin_lock(&dc->writeback_keys.lock);
1696                 rbtree_postorder_for_each_entry_safe(w, n,
1697                                         &dc->writeback_keys.keys, node)
1698                         for (j = 0; j < KEY_PTRS(&w->key); j++)
1699                                 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1700                                             GC_MARK_DIRTY);
1701                 spin_unlock(&dc->writeback_keys.lock);
1702         }
1703         rcu_read_unlock();
1704
1705         c->avail_nbuckets = 0;
1706         for_each_cache(ca, c, i) {
1707                 uint64_t *i;
1708
1709                 ca->invalidate_needs_gc = 0;
1710
1711                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1712                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1713
1714                 for (i = ca->prio_buckets;
1715                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1716                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1717
1718                 for_each_bucket(b, ca) {
1719                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1720
1721                         if (atomic_read(&b->pin))
1722                                 continue;
1723
1724                         BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1725
1726                         if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1727                                 c->avail_nbuckets++;
1728                 }
1729         }
1730
1731         mutex_unlock(&c->bucket_lock);
1732 }
1733
1734 static void bch_btree_gc(struct cache_set *c)
1735 {
1736         int ret;
1737         struct gc_stat stats;
1738         struct closure writes;
1739         struct btree_op op;
1740         uint64_t start_time = local_clock();
1741
1742         trace_bcache_gc_start(c);
1743
1744         memset(&stats, 0, sizeof(struct gc_stat));
1745         closure_init_stack(&writes);
1746         bch_btree_op_init(&op, SHRT_MAX);
1747
1748         btree_gc_start(c);
1749
1750         /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1751         do {
1752                 ret = btree_root(gc_root, c, &op, &writes, &stats);
1753                 closure_sync(&writes);
1754                 cond_resched();
1755
1756                 if (ret && ret != -EAGAIN)
1757                         pr_warn("gc failed!");
1758         } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1759
1760         bch_btree_gc_finish(c);
1761         wake_up_allocators(c);
1762
1763         bch_time_stats_update(&c->btree_gc_time, start_time);
1764
1765         stats.key_bytes *= sizeof(uint64_t);
1766         stats.data      <<= 9;
1767         bch_update_bucket_in_use(c, &stats);
1768         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1769
1770         trace_bcache_gc_end(c);
1771
1772         bch_moving_gc(c);
1773 }
1774
1775 static bool gc_should_run(struct cache_set *c)
1776 {
1777         struct cache *ca;
1778         unsigned i;
1779
1780         for_each_cache(ca, c, i)
1781                 if (ca->invalidate_needs_gc)
1782                         return true;
1783
1784         if (atomic_read(&c->sectors_to_gc) < 0)
1785                 return true;
1786
1787         return false;
1788 }
1789
1790 static int bch_gc_thread(void *arg)
1791 {
1792         struct cache_set *c = arg;
1793
1794         while (1) {
1795                 wait_event_interruptible(c->gc_wait,
1796                            kthread_should_stop() ||
1797                            test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1798                            gc_should_run(c));
1799
1800                 if (kthread_should_stop() ||
1801                     test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1802                         break;
1803
1804                 set_gc_sectors(c);
1805                 bch_btree_gc(c);
1806         }
1807
1808         wait_for_kthread_stop();
1809         return 0;
1810 }
1811
1812 int bch_gc_thread_start(struct cache_set *c)
1813 {
1814         c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1815         return PTR_ERR_OR_ZERO(c->gc_thread);
1816 }
1817
1818 /* Initial partial gc */
1819
1820 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1821 {
1822         int ret = 0;
1823         struct bkey *k, *p = NULL;
1824         struct btree_iter iter;
1825
1826         for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1827                 bch_initial_mark_key(b->c, b->level, k);
1828
1829         bch_initial_mark_key(b->c, b->level + 1, &b->key);
1830
1831         if (b->level) {
1832                 bch_btree_iter_init(&b->keys, &iter, NULL);
1833
1834                 do {
1835                         k = bch_btree_iter_next_filter(&iter, &b->keys,
1836                                                        bch_ptr_bad);
1837                         if (k)
1838                                 btree_node_prefetch(b, k);
1839
1840                         if (p)
1841                                 ret = btree(check_recurse, p, b, op);
1842
1843                         p = k;
1844                 } while (p && !ret);
1845         }
1846
1847         return ret;
1848 }
1849
1850 int bch_btree_check(struct cache_set *c)
1851 {
1852         struct btree_op op;
1853
1854         bch_btree_op_init(&op, SHRT_MAX);
1855
1856         return btree_root(check_recurse, c, &op);
1857 }
1858
1859 void bch_initial_gc_finish(struct cache_set *c)
1860 {
1861         struct cache *ca;
1862         struct bucket *b;
1863         unsigned i;
1864
1865         bch_btree_gc_finish(c);
1866
1867         mutex_lock(&c->bucket_lock);
1868
1869         /*
1870          * We need to put some unused buckets directly on the prio freelist in
1871          * order to get the allocator thread started - it needs freed buckets in
1872          * order to rewrite the prios and gens, and it needs to rewrite prios
1873          * and gens in order to free buckets.
1874          *
1875          * This is only safe for buckets that have no live data in them, which
1876          * there should always be some of.
1877          */
1878         for_each_cache(ca, c, i) {
1879                 for_each_bucket(b, ca) {
1880                         if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1881                             fifo_full(&ca->free[RESERVE_BTREE]))
1882                                 break;
1883
1884                         if (bch_can_invalidate_bucket(ca, b) &&
1885                             !GC_MARK(b)) {
1886                                 __bch_invalidate_one_bucket(ca, b);
1887                                 if (!fifo_push(&ca->free[RESERVE_PRIO],
1888                                    b - ca->buckets))
1889                                         fifo_push(&ca->free[RESERVE_BTREE],
1890                                                   b - ca->buckets);
1891                         }
1892                 }
1893         }
1894
1895         mutex_unlock(&c->bucket_lock);
1896 }
1897
1898 /* Btree insertion */
1899
1900 static bool btree_insert_key(struct btree *b, struct bkey *k,
1901                              struct bkey *replace_key)
1902 {
1903         unsigned status;
1904
1905         BUG_ON(bkey_cmp(k, &b->key) > 0);
1906
1907         status = bch_btree_insert_key(&b->keys, k, replace_key);
1908         if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1909                 bch_check_keys(&b->keys, "%u for %s", status,
1910                                replace_key ? "replace" : "insert");
1911
1912                 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1913                                               status);
1914                 return true;
1915         } else
1916                 return false;
1917 }
1918
1919 static size_t insert_u64s_remaining(struct btree *b)
1920 {
1921         long ret = bch_btree_keys_u64s_remaining(&b->keys);
1922
1923         /*
1924          * Might land in the middle of an existing extent and have to split it
1925          */
1926         if (b->keys.ops->is_extents)
1927                 ret -= KEY_MAX_U64S;
1928
1929         return max(ret, 0L);
1930 }
1931
1932 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1933                                   struct keylist *insert_keys,
1934                                   struct bkey *replace_key)
1935 {
1936         bool ret = false;
1937         int oldsize = bch_count_data(&b->keys);
1938
1939         while (!bch_keylist_empty(insert_keys)) {
1940                 struct bkey *k = insert_keys->keys;
1941
1942                 if (bkey_u64s(k) > insert_u64s_remaining(b))
1943                         break;
1944
1945                 if (bkey_cmp(k, &b->key) <= 0) {
1946                         if (!b->level)
1947                                 bkey_put(b->c, k);
1948
1949                         ret |= btree_insert_key(b, k, replace_key);
1950                         bch_keylist_pop_front(insert_keys);
1951                 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1952                         BKEY_PADDED(key) temp;
1953                         bkey_copy(&temp.key, insert_keys->keys);
1954
1955                         bch_cut_back(&b->key, &temp.key);
1956                         bch_cut_front(&b->key, insert_keys->keys);
1957
1958                         ret |= btree_insert_key(b, &temp.key, replace_key);
1959                         break;
1960                 } else {
1961                         break;
1962                 }
1963         }
1964
1965         if (!ret)
1966                 op->insert_collision = true;
1967
1968         BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1969
1970         BUG_ON(bch_count_data(&b->keys) < oldsize);
1971         return ret;
1972 }
1973
1974 static int btree_split(struct btree *b, struct btree_op *op,
1975                        struct keylist *insert_keys,
1976                        struct bkey *replace_key)
1977 {
1978         bool split;
1979         struct btree *n1, *n2 = NULL, *n3 = NULL;
1980         uint64_t start_time = local_clock();
1981         struct closure cl;
1982         struct keylist parent_keys;
1983
1984         closure_init_stack(&cl);
1985         bch_keylist_init(&parent_keys);
1986
1987         if (btree_check_reserve(b, op)) {
1988                 if (!b->level)
1989                         return -EINTR;
1990                 else
1991                         WARN(1, "insufficient reserve for split\n");
1992         }
1993
1994         n1 = btree_node_alloc_replacement(b, op);
1995         if (IS_ERR(n1))
1996                 goto err;
1997
1998         split = set_blocks(btree_bset_first(n1),
1999                            block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2000
2001         if (split) {
2002                 unsigned keys = 0;
2003
2004                 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2005
2006                 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2007                 if (IS_ERR(n2))
2008                         goto err_free1;
2009
2010                 if (!b->parent) {
2011                         n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2012                         if (IS_ERR(n3))
2013                                 goto err_free2;
2014                 }
2015
2016                 mutex_lock(&n1->write_lock);
2017                 mutex_lock(&n2->write_lock);
2018
2019                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2020
2021                 /*
2022                  * Has to be a linear search because we don't have an auxiliary
2023                  * search tree yet
2024                  */
2025
2026                 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2027                         keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2028                                                         keys));
2029
2030                 bkey_copy_key(&n1->key,
2031                               bset_bkey_idx(btree_bset_first(n1), keys));
2032                 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2033
2034                 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2035                 btree_bset_first(n1)->keys = keys;
2036
2037                 memcpy(btree_bset_first(n2)->start,
2038                        bset_bkey_last(btree_bset_first(n1)),
2039                        btree_bset_first(n2)->keys * sizeof(uint64_t));
2040
2041                 bkey_copy_key(&n2->key, &b->key);
2042
2043                 bch_keylist_add(&parent_keys, &n2->key);
2044                 bch_btree_node_write(n2, &cl);
2045                 mutex_unlock(&n2->write_lock);
2046                 rw_unlock(true, n2);
2047         } else {
2048                 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2049
2050                 mutex_lock(&n1->write_lock);
2051                 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2052         }
2053
2054         bch_keylist_add(&parent_keys, &n1->key);
2055         bch_btree_node_write(n1, &cl);
2056         mutex_unlock(&n1->write_lock);
2057
2058         if (n3) {
2059                 /* Depth increases, make a new root */
2060                 mutex_lock(&n3->write_lock);
2061                 bkey_copy_key(&n3->key, &MAX_KEY);
2062                 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2063                 bch_btree_node_write(n3, &cl);
2064                 mutex_unlock(&n3->write_lock);
2065
2066                 closure_sync(&cl);
2067                 bch_btree_set_root(n3);
2068                 rw_unlock(true, n3);
2069         } else if (!b->parent) {
2070                 /* Root filled up but didn't need to be split */
2071                 closure_sync(&cl);
2072                 bch_btree_set_root(n1);
2073         } else {
2074                 /* Split a non root node */
2075                 closure_sync(&cl);
2076                 make_btree_freeing_key(b, parent_keys.top);
2077                 bch_keylist_push(&parent_keys);
2078
2079                 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2080                 BUG_ON(!bch_keylist_empty(&parent_keys));
2081         }
2082
2083         btree_node_free(b);
2084         rw_unlock(true, n1);
2085
2086         bch_time_stats_update(&b->c->btree_split_time, start_time);
2087
2088         return 0;
2089 err_free2:
2090         bkey_put(b->c, &n2->key);
2091         btree_node_free(n2);
2092         rw_unlock(true, n2);
2093 err_free1:
2094         bkey_put(b->c, &n1->key);
2095         btree_node_free(n1);
2096         rw_unlock(true, n1);
2097 err:
2098         WARN(1, "bcache: btree split failed (level %u)", b->level);
2099
2100         if (n3 == ERR_PTR(-EAGAIN) ||
2101             n2 == ERR_PTR(-EAGAIN) ||
2102             n1 == ERR_PTR(-EAGAIN))
2103                 return -EAGAIN;
2104
2105         return -ENOMEM;
2106 }
2107
2108 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2109                                  struct keylist *insert_keys,
2110                                  atomic_t *journal_ref,
2111                                  struct bkey *replace_key)
2112 {
2113         struct closure cl;
2114
2115         BUG_ON(b->level && replace_key);
2116
2117         closure_init_stack(&cl);
2118
2119         mutex_lock(&b->write_lock);
2120
2121         if (write_block(b) != btree_bset_last(b) &&
2122             b->keys.last_set_unwritten)
2123                 bch_btree_init_next(b); /* just wrote a set */
2124
2125         if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2126                 mutex_unlock(&b->write_lock);
2127                 goto split;
2128         }
2129
2130         BUG_ON(write_block(b) != btree_bset_last(b));
2131
2132         if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2133                 if (!b->level)
2134                         bch_btree_leaf_dirty(b, journal_ref);
2135                 else
2136                         bch_btree_node_write(b, &cl);
2137         }
2138
2139         mutex_unlock(&b->write_lock);
2140
2141         /* wait for btree node write if necessary, after unlock */
2142         closure_sync(&cl);
2143
2144         return 0;
2145 split:
2146         if (current->bio_list) {
2147                 op->lock = b->c->root->level + 1;
2148                 return -EAGAIN;
2149         } else if (op->lock <= b->c->root->level) {
2150                 op->lock = b->c->root->level + 1;
2151                 return -EINTR;
2152         } else {
2153                 /* Invalidated all iterators */
2154                 int ret = btree_split(b, op, insert_keys, replace_key);
2155
2156                 if (bch_keylist_empty(insert_keys))
2157                         return 0;
2158                 else if (!ret)
2159                         return -EINTR;
2160                 return ret;
2161         }
2162 }
2163
2164 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2165                                struct bkey *check_key)
2166 {
2167         int ret = -EINTR;
2168         uint64_t btree_ptr = b->key.ptr[0];
2169         unsigned long seq = b->seq;
2170         struct keylist insert;
2171         bool upgrade = op->lock == -1;
2172
2173         bch_keylist_init(&insert);
2174
2175         if (upgrade) {
2176                 rw_unlock(false, b);
2177                 rw_lock(true, b, b->level);
2178
2179                 if (b->key.ptr[0] != btree_ptr ||
2180                    b->seq != seq + 1) {
2181                         op->lock = b->level;
2182                         goto out;
2183                }
2184         }
2185
2186         SET_KEY_PTRS(check_key, 1);
2187         get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2188
2189         SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2190
2191         bch_keylist_add(&insert, check_key);
2192
2193         ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2194
2195         BUG_ON(!ret && !bch_keylist_empty(&insert));
2196 out:
2197         if (upgrade)
2198                 downgrade_write(&b->lock);
2199         return ret;
2200 }
2201
2202 struct btree_insert_op {
2203         struct btree_op op;
2204         struct keylist  *keys;
2205         atomic_t        *journal_ref;
2206         struct bkey     *replace_key;
2207 };
2208
2209 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2210 {
2211         struct btree_insert_op *op = container_of(b_op,
2212                                         struct btree_insert_op, op);
2213
2214         int ret = bch_btree_insert_node(b, &op->op, op->keys,
2215                                         op->journal_ref, op->replace_key);
2216         if (ret && !bch_keylist_empty(op->keys))
2217                 return ret;
2218         else
2219                 return MAP_DONE;
2220 }
2221
2222 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2223                      atomic_t *journal_ref, struct bkey *replace_key)
2224 {
2225         struct btree_insert_op op;
2226         int ret = 0;
2227
2228         BUG_ON(current->bio_list);
2229         BUG_ON(bch_keylist_empty(keys));
2230
2231         bch_btree_op_init(&op.op, 0);
2232         op.keys         = keys;
2233         op.journal_ref  = journal_ref;
2234         op.replace_key  = replace_key;
2235
2236         while (!ret && !bch_keylist_empty(keys)) {
2237                 op.op.lock = 0;
2238                 ret = bch_btree_map_leaf_nodes(&op.op, c,
2239                                                &START_KEY(keys->keys),
2240                                                btree_insert_fn);
2241         }
2242
2243         if (ret) {
2244                 struct bkey *k;
2245
2246                 pr_err("error %i", ret);
2247
2248                 while ((k = bch_keylist_pop(keys)))
2249                         bkey_put(c, k);
2250         } else if (op.op.insert_collision)
2251                 ret = -ESRCH;
2252
2253         return ret;
2254 }
2255
2256 void bch_btree_set_root(struct btree *b)
2257 {
2258         unsigned i;
2259         struct closure cl;
2260
2261         closure_init_stack(&cl);
2262
2263         trace_bcache_btree_set_root(b);
2264
2265         BUG_ON(!b->written);
2266
2267         for (i = 0; i < KEY_PTRS(&b->key); i++)
2268                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2269
2270         mutex_lock(&b->c->bucket_lock);
2271         list_del_init(&b->list);
2272         mutex_unlock(&b->c->bucket_lock);
2273
2274         b->c->root = b;
2275
2276         bch_journal_meta(b->c, &cl);
2277         closure_sync(&cl);
2278 }
2279
2280 /* Map across nodes or keys */
2281
2282 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2283                                        struct bkey *from,
2284                                        btree_map_nodes_fn *fn, int flags)
2285 {
2286         int ret = MAP_CONTINUE;
2287
2288         if (b->level) {
2289                 struct bkey *k;
2290                 struct btree_iter iter;
2291
2292                 bch_btree_iter_init(&b->keys, &iter, from);
2293
2294                 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2295                                                        bch_ptr_bad))) {
2296                         ret = btree(map_nodes_recurse, k, b,
2297                                     op, from, fn, flags);
2298                         from = NULL;
2299
2300                         if (ret != MAP_CONTINUE)
2301                                 return ret;
2302                 }
2303         }
2304
2305         if (!b->level || flags == MAP_ALL_NODES)
2306                 ret = fn(op, b);
2307
2308         return ret;
2309 }
2310
2311 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2312                           struct bkey *from, btree_map_nodes_fn *fn, int flags)
2313 {
2314         return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2315 }
2316
2317 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2318                                       struct bkey *from, btree_map_keys_fn *fn,
2319                                       int flags)
2320 {
2321         int ret = MAP_CONTINUE;
2322         struct bkey *k;
2323         struct btree_iter iter;
2324
2325         bch_btree_iter_init(&b->keys, &iter, from);
2326
2327         while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2328                 ret = !b->level
2329                         ? fn(op, b, k)
2330                         : btree(map_keys_recurse, k, b, op, from, fn, flags);
2331                 from = NULL;
2332
2333                 if (ret != MAP_CONTINUE)
2334                         return ret;
2335         }
2336
2337         if (!b->level && (flags & MAP_END_KEY))
2338                 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2339                                      KEY_OFFSET(&b->key), 0));
2340
2341         return ret;
2342 }
2343
2344 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2345                        struct bkey *from, btree_map_keys_fn *fn, int flags)
2346 {
2347         return btree_root(map_keys_recurse, c, op, from, fn, flags);
2348 }
2349
2350 /* Keybuf code */
2351
2352 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2353 {
2354         /* Overlapping keys compare equal */
2355         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2356                 return -1;
2357         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2358                 return 1;
2359         return 0;
2360 }
2361
2362 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2363                                             struct keybuf_key *r)
2364 {
2365         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2366 }
2367
2368 struct refill {
2369         struct btree_op op;
2370         unsigned        nr_found;
2371         struct keybuf   *buf;
2372         struct bkey     *end;
2373         keybuf_pred_fn  *pred;
2374 };
2375
2376 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2377                             struct bkey *k)
2378 {
2379         struct refill *refill = container_of(op, struct refill, op);
2380         struct keybuf *buf = refill->buf;
2381         int ret = MAP_CONTINUE;
2382
2383         if (bkey_cmp(k, refill->end) >= 0) {
2384                 ret = MAP_DONE;
2385                 goto out;
2386         }
2387
2388         if (!KEY_SIZE(k)) /* end key */
2389                 goto out;
2390
2391         if (refill->pred(buf, k)) {
2392                 struct keybuf_key *w;
2393
2394                 spin_lock(&buf->lock);
2395
2396                 w = array_alloc(&buf->freelist);
2397                 if (!w) {
2398                         spin_unlock(&buf->lock);
2399                         return MAP_DONE;
2400                 }
2401
2402                 w->private = NULL;
2403                 bkey_copy(&w->key, k);
2404
2405                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2406                         array_free(&buf->freelist, w);
2407                 else
2408                         refill->nr_found++;
2409
2410                 if (array_freelist_empty(&buf->freelist))
2411                         ret = MAP_DONE;
2412
2413                 spin_unlock(&buf->lock);
2414         }
2415 out:
2416         buf->last_scanned = *k;
2417         return ret;
2418 }
2419
2420 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2421                        struct bkey *end, keybuf_pred_fn *pred)
2422 {
2423         struct bkey start = buf->last_scanned;
2424         struct refill refill;
2425
2426         cond_resched();
2427
2428         bch_btree_op_init(&refill.op, -1);
2429         refill.nr_found = 0;
2430         refill.buf      = buf;
2431         refill.end      = end;
2432         refill.pred     = pred;
2433
2434         bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2435                            refill_keybuf_fn, MAP_END_KEY);
2436
2437         trace_bcache_keyscan(refill.nr_found,
2438                              KEY_INODE(&start), KEY_OFFSET(&start),
2439                              KEY_INODE(&buf->last_scanned),
2440                              KEY_OFFSET(&buf->last_scanned));
2441
2442         spin_lock(&buf->lock);
2443
2444         if (!RB_EMPTY_ROOT(&buf->keys)) {
2445                 struct keybuf_key *w;
2446                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2447                 buf->start      = START_KEY(&w->key);
2448
2449                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2450                 buf->end        = w->key;
2451         } else {
2452                 buf->start      = MAX_KEY;
2453                 buf->end        = MAX_KEY;
2454         }
2455
2456         spin_unlock(&buf->lock);
2457 }
2458
2459 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2460 {
2461         rb_erase(&w->node, &buf->keys);
2462         array_free(&buf->freelist, w);
2463 }
2464
2465 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2466 {
2467         spin_lock(&buf->lock);
2468         __bch_keybuf_del(buf, w);
2469         spin_unlock(&buf->lock);
2470 }
2471
2472 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2473                                   struct bkey *end)
2474 {
2475         bool ret = false;
2476         struct keybuf_key *p, *w, s;
2477         s.key = *start;
2478
2479         if (bkey_cmp(end, &buf->start) <= 0 ||
2480             bkey_cmp(start, &buf->end) >= 0)
2481                 return false;
2482
2483         spin_lock(&buf->lock);
2484         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2485
2486         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2487                 p = w;
2488                 w = RB_NEXT(w, node);
2489
2490                 if (p->private)
2491                         ret = true;
2492                 else
2493                         __bch_keybuf_del(buf, p);
2494         }
2495
2496         spin_unlock(&buf->lock);
2497         return ret;
2498 }
2499
2500 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2501 {
2502         struct keybuf_key *w;
2503         spin_lock(&buf->lock);
2504
2505         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2506
2507         while (w && w->private)
2508                 w = RB_NEXT(w, node);
2509
2510         if (w)
2511                 w->private = ERR_PTR(-EINTR);
2512
2513         spin_unlock(&buf->lock);
2514         return w;
2515 }
2516
2517 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2518                                           struct keybuf *buf,
2519                                           struct bkey *end,
2520                                           keybuf_pred_fn *pred)
2521 {
2522         struct keybuf_key *ret;
2523
2524         while (1) {
2525                 ret = bch_keybuf_next(buf);
2526                 if (ret)
2527                         break;
2528
2529                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2530                         pr_debug("scan finished");
2531                         break;
2532                 }
2533
2534                 bch_refill_keybuf(c, buf, end, pred);
2535         }
2536
2537         return ret;
2538 }
2539
2540 void bch_keybuf_init(struct keybuf *buf)
2541 {
2542         buf->last_scanned       = MAX_KEY;
2543         buf->keys               = RB_ROOT;
2544
2545         spin_lock_init(&buf->lock);
2546         array_allocator_init(&buf->freelist);
2547 }