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