drivers/md/bcache/btree.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  *
   4  * Uses a block device as cache for other block devices; optimized for SSDs.
   5  * All allocation is done in buckets, which should match the erase block size
   6  * of the device.
   7  *
   8  * Buckets containing cached data are kept on a heap sorted by priority;
   9  * bucket priority is increased on cache hit, and periodically all the buckets
  10  * on the heap have their priority scaled down. This currently is just used as
  11  * an LRU but in the future should allow for more intelligent heuristics.
  12  *
  13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
  14  * counter. Garbage collection is used to remove stale pointers.
  15  *
  16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
  17  * as keys are inserted we only sort the pages that have not yet been written.
  18  * When garbage collection is run, we resort the entire node.
  19  *
  20  * All configuration is done via sysfs; see Documentation/bcache.txt.
  21  */
  22
  23 #include "bcache.h"
  24 #include "btree.h"
  25 #include "debug.h"
  26 #include "request.h"
  27 #include "writeback.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/hash.h>
  32 #include <linux/prefetch.h>
  33 #include <linux/random.h>
  34 #include <linux/rcupdate.h>
  35 #include <trace/events/bcache.h>
  36
  37 /*
  38  * Todo:
  39  * register_bcache: Return errors out to userspace correctly
  40  *
  41  * Writeback: don't undirty key until after a cache flush
  42  *
  43  * Create an iterator for key pointers
  44  *
  45  * On btree write error, mark bucket such that it won't be freed from the cache
  46  *
  47  * Journalling:
  48  *   Check for bad keys in replay
  49  *   Propagate barriers
  50  *   Refcount journal entries in journal_replay
  51  *
  52  * Garbage collection:
  53  *   Finish incremental gc
  54  *   Gc should free old UUIDs, data for invalid UUIDs
  55  *
  56  * Provide a way to list backing device UUIDs we have data cached for, and
  57  * probably how long it's been since we've seen them, and a way to invalidate
  58  * dirty data for devices that will never be attached again
  59  *
  60  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
  61  * that based on that and how much dirty data we have we can keep writeback
  62  * from being starved
  63  *
  64  * Add a tracepoint or somesuch to watch for writeback starvation
  65  *
  66  * When btree depth > 1 and splitting an interior node, we have to make sure
  67  * alloc_bucket() cannot fail. This should be true but is not completely
  68  * obvious.
  69  *
  70  * Make sure all allocations get charged to the root cgroup
  71  *
  72  * Plugging?
  73  *
  74  * If data write is less than hard sector size of ssd, round up offset in open
  75  * bucket to the next whole sector
  76  *
  77  * Also lookup by cgroup in get_open_bucket()
  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 static const char * const op_types[] = {
  92         "insert", "replace"
  93 };
  94
  95 static const char *op_type(struct btree_op *op)
  96 {
  97         return op_types[op->type];
  98 }
  99
 100 #define MAX_NEED_GC             64
 101 #define MAX_SAVE_PRIO           72
 102
 103 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
 104
 105 #define PTR_HASH(c, k)                                                  \
 106         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
 107
 108 struct workqueue_struct *bch_gc_wq;
 109 static struct workqueue_struct *btree_io_wq;
 110
 111 void bch_btree_op_init_stack(struct btree_op *op)
 112 {
 113         memset(op, 0, sizeof(struct btree_op));
 114         closure_init_stack(&op->cl);
 115         op->lock = -1;
 116         bch_keylist_init(&op->keys);
 117 }
 118
 119 /* Btree key manipulation */
 120
 121 static void bkey_put(struct cache_set *c, struct bkey *k, int level)
 122 {
 123         if ((level && KEY_OFFSET(k)) || !level)
 124                 __bkey_put(c, k);
 125 }
 126
 127 /* Btree IO */
 128
 129 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
 130 {
 131         uint64_t crc = b->key.ptr[0];
 132         void *data = (void *) i + 8, *end = end(i);
 133
 134         crc = bch_crc64_update(crc, data, end - data);
 135         return crc ^ 0xffffffffffffffffULL;
 136 }
 137
 138 static void bch_btree_node_read_done(struct btree *b)
 139 {
 140         const char *err = "bad btree header";
 141         struct bset *i = b->sets[0].data;
 142         struct btree_iter *iter;
 143
 144         iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
 145         iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
 146         iter->used = 0;
 147
 148         if (!i->seq)
 149                 goto err;
 150
 151         for (;
 152              b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
 153              i = write_block(b)) {
 154                 err = "unsupported bset version";
 155                 if (i->version > BCACHE_BSET_VERSION)
 156                         goto err;
 157
 158                 err = "bad btree header";
 159                 if (b->written + set_blocks(i, b->c) > btree_blocks(b))
 160                         goto err;
 161
 162                 err = "bad magic";
 163                 if (i->magic != bset_magic(b->c))
 164                         goto err;
 165
 166                 err = "bad checksum";
 167                 switch (i->version) {
 168                 case 0:
 169                         if (i->csum != csum_set(i))
 170                                 goto err;
 171                         break;
 172                 case BCACHE_BSET_VERSION:
 173                         if (i->csum != btree_csum_set(b, i))
 174                                 goto err;
 175                         break;
 176                 }
 177
 178                 err = "empty set";
 179                 if (i != b->sets[0].data && !i->keys)
 180                         goto err;
 181
 182                 bch_btree_iter_push(iter, i->start, end(i));
 183
 184                 b->written += set_blocks(i, b->c);
 185         }
 186
 187         err = "corrupted btree";
 188         for (i = write_block(b);
 189              index(i, b) < btree_blocks(b);
 190              i = ((void *) i) + block_bytes(b->c))
 191                 if (i->seq == b->sets[0].data->seq)
 192                         goto err;
 193
 194         bch_btree_sort_and_fix_extents(b, iter);
 195
 196         i = b->sets[0].data;
 197         err = "short btree key";
 198         if (b->sets[0].size &&
 199             bkey_cmp(&b->key, &b->sets[0].end) < 0)
 200                 goto err;
 201
 202         if (b->written < btree_blocks(b))
 203                 bch_bset_init_next(b);
 204 out:
 205         mempool_free(iter, b->c->fill_iter);
 206         return;
 207 err:
 208         set_btree_node_io_error(b);
 209         bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys",
 210                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
 211                             index(i, b), i->keys);
 212         goto out;
 213 }
 214
 215 static void btree_node_read_endio(struct bio *bio, int error)
 216 {
 217         struct closure *cl = bio->bi_private;
 218         closure_put(cl);
 219 }
 220
 221 void bch_btree_node_read(struct btree *b)
 222 {
 223         uint64_t start_time = local_clock();
 224         struct closure cl;
 225         struct bio *bio;
 226
 227         trace_bcache_btree_read(b);
 228
 229         closure_init_stack(&cl);
 230
 231         bio = bch_bbio_alloc(b->c);
 232         bio->bi_rw      = REQ_META|READ_SYNC;
 233         bio->bi_size    = KEY_SIZE(&b->key) << 9;
 234         bio->bi_end_io  = btree_node_read_endio;
 235         bio->bi_private = &cl;
 236
 237         bch_bio_map(bio, b->sets[0].data);
 238
 239         bch_submit_bbio(bio, b->c, &b->key, 0);
 240         closure_sync(&cl);
 241
 242         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 243                 set_btree_node_io_error(b);
 244
 245         bch_bbio_free(bio, b->c);
 246
 247         if (btree_node_io_error(b))
 248                 goto err;
 249
 250         bch_btree_node_read_done(b);
 251
 252         spin_lock(&b->c->btree_read_time_lock);
 253         bch_time_stats_update(&b->c->btree_read_time, start_time);
 254         spin_unlock(&b->c->btree_read_time_lock);
 255
 256         return;
 257 err:
 258         bch_cache_set_error(b->c, "io error reading bucket %lu",
 259                             PTR_BUCKET_NR(b->c, &b->key, 0));
 260 }
 261
 262 static void btree_complete_write(struct btree *b, struct btree_write *w)
 263 {
 264         if (w->prio_blocked &&
 265             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
 266                 wake_up_allocators(b->c);
 267
 268         if (w->journal) {
 269                 atomic_dec_bug(w->journal);
 270                 __closure_wake_up(&b->c->journal.wait);
 271         }
 272
 273         w->prio_blocked = 0;
 274         w->journal      = NULL;
 275 }
 276
 277 static void __btree_node_write_done(struct closure *cl)
 278 {
 279         struct btree *b = container_of(cl, struct btree, io.cl);
 280         struct btree_write *w = btree_prev_write(b);
 281
 282         bch_bbio_free(b->bio, b->c);
 283         b->bio = NULL;
 284         btree_complete_write(b, w);
 285
 286         if (btree_node_dirty(b))
 287                 queue_delayed_work(btree_io_wq, &b->work,
 288                                    msecs_to_jiffies(30000));
 289
 290         closure_return(cl);
 291 }
 292
 293 static void btree_node_write_done(struct closure *cl)
 294 {
 295         struct btree *b = container_of(cl, struct btree, io.cl);
 296         struct bio_vec *bv;
 297         int n;
 298
 299         __bio_for_each_segment(bv, b->bio, n, 0)
 300                 __free_page(bv->bv_page);
 301
 302         __btree_node_write_done(cl);
 303 }
 304
 305 static void btree_node_write_endio(struct bio *bio, int error)
 306 {
 307         struct closure *cl = bio->bi_private;
 308         struct btree *b = container_of(cl, struct btree, io.cl);
 309
 310         if (error)
 311                 set_btree_node_io_error(b);
 312
 313         bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
 314         closure_put(cl);
 315 }
 316
 317 static void do_btree_node_write(struct btree *b)
 318 {
 319         struct closure *cl = &b->io.cl;
 320         struct bset *i = b->sets[b->nsets].data;
 321         BKEY_PADDED(key) k;
 322
 323         i->version      = BCACHE_BSET_VERSION;
 324         i->csum         = btree_csum_set(b, i);
 325
 326         BUG_ON(b->bio);
 327         b->bio = bch_bbio_alloc(b->c);
 328
 329         b->bio->bi_end_io       = btree_node_write_endio;
 330         b->bio->bi_private      = &b->io.cl;
 331         b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
 332         b->bio->bi_size         = set_blocks(i, b->c) * block_bytes(b->c);
 333         bch_bio_map(b->bio, i);
 334
 335         /*
 336          * If we're appending to a leaf node, we don't technically need FUA -
 337          * this write just needs to be persisted before the next journal write,
 338          * which will be marked FLUSH|FUA.
 339          *
 340          * Similarly if we're writing a new btree root - the pointer is going to
 341          * be in the next journal entry.
 342          *
 343          * But if we're writing a new btree node (that isn't a root) or
 344          * appending to a non leaf btree node, we need either FUA or a flush
 345          * when we write the parent with the new pointer. FUA is cheaper than a
 346          * flush, and writes appending to leaf nodes aren't blocking anything so
 347          * just make all btree node writes FUA to keep things sane.
 348          */
 349
 350         bkey_copy(&k.key, &b->key);
 351         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
 352
 353         if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
 354                 int j;
 355                 struct bio_vec *bv;
 356                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
 357
 358                 bio_for_each_segment(bv, b->bio, j)
 359                         memcpy(page_address(bv->bv_page),
 360                                base + j * PAGE_SIZE, PAGE_SIZE);
 361
 362                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 363
 364                 continue_at(cl, btree_node_write_done, NULL);
 365         } else {
 366                 b->bio->bi_vcnt = 0;
 367                 bch_bio_map(b->bio, i);
 368
 369                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
 370
 371                 closure_sync(cl);
 372                 __btree_node_write_done(cl);
 373         }
 374 }
 375
 376 void bch_btree_node_write(struct btree *b, struct closure *parent)
 377 {
 378         struct bset *i = b->sets[b->nsets].data;
 379
 380         trace_bcache_btree_write(b);
 381
 382         BUG_ON(current->bio_list);
 383         BUG_ON(b->written >= btree_blocks(b));
 384         BUG_ON(b->written && !i->keys);
 385         BUG_ON(b->sets->data->seq != i->seq);
 386         bch_check_key_order(b, i);
 387
 388         cancel_delayed_work(&b->work);
 389
 390         /* If caller isn't waiting for write, parent refcount is cache set */
 391         closure_lock(&b->io, parent ?: &b->c->cl);
 392
 393         clear_bit(BTREE_NODE_dirty,      &b->flags);
 394         change_bit(BTREE_NODE_write_idx, &b->flags);
 395
 396         do_btree_node_write(b);
 397
 398         b->written += set_blocks(i, b->c);
 399         atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
 400                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
 401
 402         bch_btree_sort_lazy(b);
 403
 404         if (b->written < btree_blocks(b))
 405                 bch_bset_init_next(b);
 406 }
 407
 408 static void btree_node_write_work(struct work_struct *w)
 409 {
 410         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
 411
 412         rw_lock(true, b, b->level);
 413
 414         if (btree_node_dirty(b))
 415                 bch_btree_node_write(b, NULL);
 416         rw_unlock(true, b);
 417 }
 418
 419 static void bch_btree_leaf_dirty(struct btree *b, struct btree_op *op)
 420 {
 421         struct bset *i = b->sets[b->nsets].data;
 422         struct btree_write *w = btree_current_write(b);
 423
 424         BUG_ON(!b->written);
 425         BUG_ON(!i->keys);
 426
 427         if (!btree_node_dirty(b))
 428                 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
 429
 430         set_btree_node_dirty(b);
 431
 432         if (op && op->journal) {
 433                 if (w->journal &&
 434                     journal_pin_cmp(b->c, w, op)) {
 435                         atomic_dec_bug(w->journal);
 436                         w->journal = NULL;
 437                 }
 438
 439                 if (!w->journal) {
 440                         w->journal = op->journal;
 441                         atomic_inc(w->journal);
 442                 }
 443         }
 444
 445         /* Force write if set is too big */
 446         if (set_bytes(i) > PAGE_SIZE - 48 &&
 447             !current->bio_list)
 448                 bch_btree_node_write(b, NULL);
 449 }
 450
 451 /*
 452  * Btree in memory cache - allocation/freeing
 453  * mca -> memory cache
 454  */
 455
 456 static void mca_reinit(struct btree *b)
 457 {
 458         unsigned i;
 459
 460         b->flags        = 0;
 461         b->written      = 0;
 462         b->nsets        = 0;
 463
 464         for (i = 0; i < MAX_BSETS; i++)
 465                 b->sets[i].size = 0;
 466         /*
 467          * Second loop starts at 1 because b->sets[0]->data is the memory we
 468          * allocated
 469          */
 470         for (i = 1; i < MAX_BSETS; i++)
 471                 b->sets[i].data = NULL;
 472 }
 473
 474 #define mca_reserve(c)  (((c->root && c->root->level)           \
 475                           ? c->root->level : 1) * 8 + 16)
 476 #define mca_can_free(c)                                         \
 477         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
 478
 479 static void mca_data_free(struct btree *b)
 480 {
 481         struct bset_tree *t = b->sets;
 482         BUG_ON(!closure_is_unlocked(&b->io.cl));
 483
 484         if (bset_prev_bytes(b) < PAGE_SIZE)
 485                 kfree(t->prev);
 486         else
 487                 free_pages((unsigned long) t->prev,
 488                            get_order(bset_prev_bytes(b)));
 489
 490         if (bset_tree_bytes(b) < PAGE_SIZE)
 491                 kfree(t->tree);
 492         else
 493                 free_pages((unsigned long) t->tree,
 494                            get_order(bset_tree_bytes(b)));
 495
 496         free_pages((unsigned long) t->data, b->page_order);
 497
 498         t->prev = NULL;
 499         t->tree = NULL;
 500         t->data = NULL;
 501         list_move(&b->list, &b->c->btree_cache_freed);
 502         b->c->bucket_cache_used--;
 503 }
 504
 505 static void mca_bucket_free(struct btree *b)
 506 {
 507         BUG_ON(btree_node_dirty(b));
 508
 509         b->key.ptr[0] = 0;
 510         hlist_del_init_rcu(&b->hash);
 511         list_move(&b->list, &b->c->btree_cache_freeable);
 512 }
 513
 514 static unsigned btree_order(struct bkey *k)
 515 {
 516         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
 517 }
 518
 519 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
 520 {
 521         struct bset_tree *t = b->sets;
 522         BUG_ON(t->data);
 523
 524         b->page_order = max_t(unsigned,
 525                               ilog2(b->c->btree_pages),
 526                               btree_order(k));
 527
 528         t->data = (void *) __get_free_pages(gfp, b->page_order);
 529         if (!t->data)
 530                 goto err;
 531
 532         t->tree = bset_tree_bytes(b) < PAGE_SIZE
 533                 ? kmalloc(bset_tree_bytes(b), gfp)
 534                 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
 535         if (!t->tree)
 536                 goto err;
 537
 538         t->prev = bset_prev_bytes(b) < PAGE_SIZE
 539                 ? kmalloc(bset_prev_bytes(b), gfp)
 540                 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
 541         if (!t->prev)
 542                 goto err;
 543
 544         list_move(&b->list, &b->c->btree_cache);
 545         b->c->bucket_cache_used++;
 546         return;
 547 err:
 548         mca_data_free(b);
 549 }
 550
 551 static struct btree *mca_bucket_alloc(struct cache_set *c,
 552                                       struct bkey *k, gfp_t gfp)
 553 {
 554         struct btree *b = kzalloc(sizeof(struct btree), gfp);
 555         if (!b)
 556                 return NULL;
 557
 558         init_rwsem(&b->lock);
 559         lockdep_set_novalidate_class(&b->lock);
 560         INIT_LIST_HEAD(&b->list);
 561         INIT_DELAYED_WORK(&b->work, btree_node_write_work);
 562         b->c = c;
 563         closure_init_unlocked(&b->io);
 564
 565         mca_data_alloc(b, k, gfp);
 566         return b;
 567 }
 568
 569 static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order)
 570 {
 571         lockdep_assert_held(&b->c->bucket_lock);
 572
 573         if (!down_write_trylock(&b->lock))
 574                 return -ENOMEM;
 575
 576         if (b->page_order < min_order) {
 577                 rw_unlock(true, b);
 578                 return -ENOMEM;
 579         }
 580
 581         BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
 582
 583         if (cl && btree_node_dirty(b))
 584                 bch_btree_node_write(b, NULL);
 585
 586         if (cl)
 587                 closure_wait_event_async(&b->io.wait, cl,
 588                          atomic_read(&b->io.cl.remaining) == -1);
 589
 590         if (btree_node_dirty(b) ||
 591             !closure_is_unlocked(&b->io.cl) ||
 592             work_pending(&b->work.work)) {
 593                 rw_unlock(true, b);
 594                 return -EAGAIN;
 595         }
 596
 597         return 0;
 598 }
 599
 600 static int bch_mca_shrink(struct shrinker *shrink, struct shrink_control *sc)
 601 {
 602         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
 603         struct btree *b, *t;
 604         unsigned long i, nr = sc->nr_to_scan;
 605
 606         if (c->shrinker_disabled)
 607                 return 0;
 608
 609         if (c->try_harder)
 610                 return 0;
 611
 612         /*
 613          * If nr == 0, we're supposed to return the number of items we have
 614          * cached. Not allowed to return -1.
 615          */
 616         if (!nr)
 617                 return mca_can_free(c) * c->btree_pages;
 618
 619         /* Return -1 if we can't do anything right now */
 620         if (sc->gfp_mask & __GFP_WAIT)
 621                 mutex_lock(&c->bucket_lock);
 622         else if (!mutex_trylock(&c->bucket_lock))
 623                 return -1;
 624
 625         /*
 626          * It's _really_ critical that we don't free too many btree nodes - we
 627          * have to always leave ourselves a reserve. The reserve is how we
 628          * guarantee that allocating memory for a new btree node can always
 629          * succeed, so that inserting keys into the btree can always succeed and
 630          * IO can always make forward progress:
 631          */
 632         nr /= c->btree_pages;
 633         nr = min_t(unsigned long, nr, mca_can_free(c));
 634
 635         i = 0;
 636         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
 637                 if (!nr)
 638                         break;
 639
 640                 if (++i > 3 &&
 641                     !mca_reap(b, NULL, 0)) {
 642                         mca_data_free(b);
 643                         rw_unlock(true, b);
 644                         --nr;
 645                 }
 646         }
 647
 648         /*
 649          * Can happen right when we first start up, before we've read in any
 650          * btree nodes
 651          */
 652         if (list_empty(&c->btree_cache))
 653                 goto out;
 654
 655         for (i = 0; nr && i < c->bucket_cache_used; i++) {
 656                 b = list_first_entry(&c->btree_cache, struct btree, list);
 657                 list_rotate_left(&c->btree_cache);
 658
 659                 if (!b->accessed &&
 660                     !mca_reap(b, NULL, 0)) {
 661                         mca_bucket_free(b);
 662                         mca_data_free(b);
 663                         rw_unlock(true, b);
 664                         --nr;
 665                 } else
 666                         b->accessed = 0;
 667         }
 668 out:
 669         nr = mca_can_free(c) * c->btree_pages;
 670         mutex_unlock(&c->bucket_lock);
 671         return nr;
 672 }
 673
 674 void bch_btree_cache_free(struct cache_set *c)
 675 {
 676         struct btree *b;
 677         struct closure cl;
 678         closure_init_stack(&cl);
 679
 680         if (c->shrink.list.next)
 681                 unregister_shrinker(&c->shrink);
 682
 683         mutex_lock(&c->bucket_lock);
 684
 685 #ifdef CONFIG_BCACHE_DEBUG
 686         if (c->verify_data)
 687                 list_move(&c->verify_data->list, &c->btree_cache);
 688 #endif
 689
 690         list_splice(&c->btree_cache_freeable,
 691                     &c->btree_cache);
 692
 693         while (!list_empty(&c->btree_cache)) {
 694                 b = list_first_entry(&c->btree_cache, struct btree, list);
 695
 696                 if (btree_node_dirty(b))
 697                         btree_complete_write(b, btree_current_write(b));
 698                 clear_bit(BTREE_NODE_dirty, &b->flags);
 699
 700                 mca_data_free(b);
 701         }
 702
 703         while (!list_empty(&c->btree_cache_freed)) {
 704                 b = list_first_entry(&c->btree_cache_freed,
 705                                      struct btree, list);
 706                 list_del(&b->list);
 707                 cancel_delayed_work_sync(&b->work);
 708                 kfree(b);
 709         }
 710
 711         mutex_unlock(&c->bucket_lock);
 712 }
 713
 714 int bch_btree_cache_alloc(struct cache_set *c)
 715 {
 716         unsigned i;
 717
 718         /* XXX: doesn't check for errors */
 719
 720         closure_init_unlocked(&c->gc);
 721
 722         for (i = 0; i < mca_reserve(c); i++)
 723                 mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 724
 725         list_splice_init(&c->btree_cache,
 726                          &c->btree_cache_freeable);
 727
 728 #ifdef CONFIG_BCACHE_DEBUG
 729         mutex_init(&c->verify_lock);
 730
 731         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
 732
 733         if (c->verify_data &&
 734             c->verify_data->sets[0].data)
 735                 list_del_init(&c->verify_data->list);
 736         else
 737                 c->verify_data = NULL;
 738 #endif
 739
 740         c->shrink.shrink = bch_mca_shrink;
 741         c->shrink.seeks = 4;
 742         c->shrink.batch = c->btree_pages * 2;
 743         register_shrinker(&c->shrink);
 744
 745         return 0;
 746 }
 747
 748 /* Btree in memory cache - hash table */
 749
 750 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
 751 {
 752         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
 753 }
 754
 755 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
 756 {
 757         struct btree *b;
 758
 759         rcu_read_lock();
 760         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
 761                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
 762                         goto out;
 763         b = NULL;
 764 out:
 765         rcu_read_unlock();
 766         return b;
 767 }
 768
 769 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k,
 770                                      int level, struct closure *cl)
 771 {
 772         int ret = -ENOMEM;
 773         struct btree *i;
 774
 775         trace_bcache_btree_cache_cannibalize(c);
 776
 777         if (!cl)
 778                 return ERR_PTR(-ENOMEM);
 779
 780         /*
 781          * Trying to free up some memory - i.e. reuse some btree nodes - may
 782          * require initiating IO to flush the dirty part of the node. If we're
 783          * running under generic_make_request(), that IO will never finish and
 784          * we would deadlock. Returning -EAGAIN causes the cache lookup code to
 785          * punt to workqueue and retry.
 786          */
 787         if (current->bio_list)
 788                 return ERR_PTR(-EAGAIN);
 789
 790         if (c->try_harder && c->try_harder != cl) {
 791                 closure_wait_event_async(&c->try_wait, cl, !c->try_harder);
 792                 return ERR_PTR(-EAGAIN);
 793         }
 794
 795         c->try_harder = cl;
 796         c->try_harder_start = local_clock();
 797 retry:
 798         list_for_each_entry_reverse(i, &c->btree_cache, list) {
 799                 int r = mca_reap(i, cl, btree_order(k));
 800                 if (!r)
 801                         return i;
 802                 if (r != -ENOMEM)
 803                         ret = r;
 804         }
 805
 806         if (ret == -EAGAIN &&
 807             closure_blocking(cl)) {
 808                 mutex_unlock(&c->bucket_lock);
 809                 closure_sync(cl);
 810                 mutex_lock(&c->bucket_lock);
 811                 goto retry;
 812         }
 813
 814         return ERR_PTR(ret);
 815 }
 816
 817 /*
 818  * We can only have one thread cannibalizing other cached btree nodes at a time,
 819  * or we'll deadlock. We use an open coded mutex to ensure that, which a
 820  * cannibalize_bucket() will take. This means every time we unlock the root of
 821  * the btree, we need to release this lock if we have it held.
 822  */
 823 void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl)
 824 {
 825         if (c->try_harder == cl) {
 826                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
 827                 c->try_harder = NULL;
 828                 __closure_wake_up(&c->try_wait);
 829         }
 830 }
 831
 832 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k,
 833                                int level, struct closure *cl)
 834 {
 835         struct btree *b;
 836
 837         lockdep_assert_held(&c->bucket_lock);
 838
 839         if (mca_find(c, k))
 840                 return NULL;
 841
 842         /* btree_free() doesn't free memory; it sticks the node on the end of
 843          * the list. Check if there's any freed nodes there:
 844          */
 845         list_for_each_entry(b, &c->btree_cache_freeable, list)
 846                 if (!mca_reap(b, NULL, btree_order(k)))
 847                         goto out;
 848
 849         /* We never free struct btree itself, just the memory that holds the on
 850          * disk node. Check the freed list before allocating a new one:
 851          */
 852         list_for_each_entry(b, &c->btree_cache_freed, list)
 853                 if (!mca_reap(b, NULL, 0)) {
 854                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
 855                         if (!b->sets[0].data)
 856                                 goto err;
 857                         else
 858                                 goto out;
 859                 }
 860
 861         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
 862         if (!b)
 863                 goto err;
 864
 865         BUG_ON(!down_write_trylock(&b->lock));
 866         if (!b->sets->data)
 867                 goto err;
 868 out:
 869         BUG_ON(!closure_is_unlocked(&b->io.cl));
 870
 871         bkey_copy(&b->key, k);
 872         list_move(&b->list, &c->btree_cache);
 873         hlist_del_init_rcu(&b->hash);
 874         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
 875
 876         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
 877         b->level        = level;
 878
 879         mca_reinit(b);
 880
 881         return b;
 882 err:
 883         if (b)
 884                 rw_unlock(true, b);
 885
 886         b = mca_cannibalize(c, k, level, cl);
 887         if (!IS_ERR(b))
 888                 goto out;
 889
 890         return b;
 891 }
 892
 893 /**
 894  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
 895  * in from disk if necessary.
 896  *
 897  * If IO is necessary, it uses the closure embedded in struct btree_op to wait;
 898  * if that closure is in non blocking mode, will return -EAGAIN.
 899  *
 900  * The btree node will have either a read or a write lock held, depending on
 901  * level and op->lock.
 902  */
 903 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
 904                                  int level, struct btree_op *op)
 905 {
 906         int i = 0;
 907         bool write = level <= op->lock;
 908         struct btree *b;
 909
 910         BUG_ON(level < 0);
 911 retry:
 912         b = mca_find(c, k);
 913
 914         if (!b) {
 915                 if (current->bio_list)
 916                         return ERR_PTR(-EAGAIN);
 917
 918                 mutex_lock(&c->bucket_lock);
 919                 b = mca_alloc(c, k, level, &op->cl);
 920                 mutex_unlock(&c->bucket_lock);
 921
 922                 if (!b)
 923                         goto retry;
 924                 if (IS_ERR(b))
 925                         return b;
 926
 927                 bch_btree_node_read(b);
 928
 929                 if (!write)
 930                         downgrade_write(&b->lock);
 931         } else {
 932                 rw_lock(write, b, level);
 933                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
 934                         rw_unlock(write, b);
 935                         goto retry;
 936                 }
 937                 BUG_ON(b->level != level);
 938         }
 939
 940         b->accessed = 1;
 941
 942         for (; i <= b->nsets && b->sets[i].size; i++) {
 943                 prefetch(b->sets[i].tree);
 944                 prefetch(b->sets[i].data);
 945         }
 946
 947         for (; i <= b->nsets; i++)
 948                 prefetch(b->sets[i].data);
 949
 950         if (btree_node_io_error(b)) {
 951                 rw_unlock(write, b);
 952                 return ERR_PTR(-EIO);
 953         }
 954
 955         BUG_ON(!b->written);
 956
 957         return b;
 958 }
 959
 960 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
 961 {
 962         struct btree *b;
 963
 964         mutex_lock(&c->bucket_lock);
 965         b = mca_alloc(c, k, level, NULL);
 966         mutex_unlock(&c->bucket_lock);
 967
 968         if (!IS_ERR_OR_NULL(b)) {
 969                 bch_btree_node_read(b);
 970                 rw_unlock(true, b);
 971         }
 972 }
 973
 974 /* Btree alloc */
 975
 976 static void btree_node_free(struct btree *b, struct btree_op *op)
 977 {
 978         unsigned i;
 979
 980         trace_bcache_btree_node_free(b);
 981
 982         /*
 983          * The BUG_ON() in btree_node_get() implies that we must have a write
 984          * lock on parent to free or even invalidate a node
 985          */
 986         BUG_ON(op->lock <= b->level);
 987         BUG_ON(b == b->c->root);
 988
 989         if (btree_node_dirty(b))
 990                 btree_complete_write(b, btree_current_write(b));
 991         clear_bit(BTREE_NODE_dirty, &b->flags);
 992
 993         cancel_delayed_work(&b->work);
 994
 995         mutex_lock(&b->c->bucket_lock);
 996
 997         for (i = 0; i < KEY_PTRS(&b->key); i++) {
 998                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
 999
1000                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1001                             PTR_BUCKET(b->c, &b->key, i));
1002         }
1003
1004         bch_bucket_free(b->c, &b->key);
1005         mca_bucket_free(b);
1006         mutex_unlock(&b->c->bucket_lock);
1007 }
1008
1009 struct btree *bch_btree_node_alloc(struct cache_set *c, int level,
1010                                    struct closure *cl)
1011 {
1012         BKEY_PADDED(key) k;
1013         struct btree *b = ERR_PTR(-EAGAIN);
1014
1015         mutex_lock(&c->bucket_lock);
1016 retry:
1017         if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl))
1018                 goto err;
1019
1020         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1021
1022         b = mca_alloc(c, &k.key, level, cl);
1023         if (IS_ERR(b))
1024                 goto err_free;
1025
1026         if (!b) {
1027                 cache_bug(c,
1028                         "Tried to allocate bucket that was in btree cache");
1029                 __bkey_put(c, &k.key);
1030                 goto retry;
1031         }
1032
1033         b->accessed = 1;
1034         bch_bset_init_next(b);
1035
1036         mutex_unlock(&c->bucket_lock);
1037
1038         trace_bcache_btree_node_alloc(b);
1039         return b;
1040 err_free:
1041         bch_bucket_free(c, &k.key);
1042         __bkey_put(c, &k.key);
1043 err:
1044         mutex_unlock(&c->bucket_lock);
1045
1046         trace_bcache_btree_node_alloc_fail(b);
1047         return b;
1048 }
1049
1050 static struct btree *btree_node_alloc_replacement(struct btree *b,
1051                                                   struct closure *cl)
1052 {
1053         struct btree *n = bch_btree_node_alloc(b->c, b->level, cl);
1054         if (!IS_ERR_OR_NULL(n))
1055                 bch_btree_sort_into(b, n);
1056
1057         return n;
1058 }
1059
1060 /* Garbage collection */
1061
1062 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1063 {
1064         uint8_t stale = 0;
1065         unsigned i;
1066         struct bucket *g;
1067
1068         /*
1069          * ptr_invalid() can't return true for the keys that mark btree nodes as
1070          * freed, but since ptr_bad() returns true we'll never actually use them
1071          * for anything and thus we don't want mark their pointers here
1072          */
1073         if (!bkey_cmp(k, &ZERO_KEY))
1074                 return stale;
1075
1076         for (i = 0; i < KEY_PTRS(k); i++) {
1077                 if (!ptr_available(c, k, i))
1078                         continue;
1079
1080                 g = PTR_BUCKET(c, k, i);
1081
1082                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1083                         g->gc_gen = PTR_GEN(k, i);
1084
1085                 if (ptr_stale(c, k, i)) {
1086                         stale = max(stale, ptr_stale(c, k, i));
1087                         continue;
1088                 }
1089
1090                 cache_bug_on(GC_MARK(g) &&
1091                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1092                              c, "inconsistent ptrs: mark = %llu, level = %i",
1093                              GC_MARK(g), level);
1094
1095                 if (level)
1096                         SET_GC_MARK(g, GC_MARK_METADATA);
1097                 else if (KEY_DIRTY(k))
1098                         SET_GC_MARK(g, GC_MARK_DIRTY);
1099
1100                 /* guard against overflow */
1101                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1102                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1103                                              (1 << 14) - 1));
1104
1105                 BUG_ON(!GC_SECTORS_USED(g));
1106         }
1107
1108         return stale;
1109 }
1110
1111 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1112
1113 static int btree_gc_mark_node(struct btree *b, unsigned *keys,
1114                               struct gc_stat *gc)
1115 {
1116         uint8_t stale = 0;
1117         unsigned last_dev = -1;
1118         struct bcache_device *d = NULL;
1119         struct bkey *k;
1120         struct btree_iter iter;
1121         struct bset_tree *t;
1122
1123         gc->nodes++;
1124
1125         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1126                 if (last_dev != KEY_INODE(k)) {
1127                         last_dev = KEY_INODE(k);
1128
1129                         d = KEY_INODE(k) < b->c->nr_uuids
1130                                 ? b->c->devices[last_dev]
1131                                 : NULL;
1132                 }
1133
1134                 stale = max(stale, btree_mark_key(b, k));
1135
1136                 if (bch_ptr_bad(b, k))
1137                         continue;
1138
1139                 *keys += bkey_u64s(k);
1140
1141                 gc->key_bytes += bkey_u64s(k);
1142                 gc->nkeys++;
1143
1144                 gc->data += KEY_SIZE(k);
1145                 if (KEY_DIRTY(k))
1146                         gc->dirty += KEY_SIZE(k);
1147         }
1148
1149         for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1150                 btree_bug_on(t->size &&
1151                              bset_written(b, t) &&
1152                              bkey_cmp(&b->key, &t->end) < 0,
1153                              b, "found short btree key in gc");
1154
1155         return stale;
1156 }
1157
1158 static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k,
1159                                     struct btree_op *op)
1160 {
1161         /*
1162          * We block priorities from being written for the duration of garbage
1163          * collection, so we can't sleep in btree_alloc() ->
1164          * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it
1165          * our closure.
1166          */
1167         struct btree *n = btree_node_alloc_replacement(b, NULL);
1168
1169         if (!IS_ERR_OR_NULL(n)) {
1170                 swap(b, n);
1171                 __bkey_put(b->c, &b->key);
1172
1173                 memcpy(k->ptr, b->key.ptr,
1174                        sizeof(uint64_t) * KEY_PTRS(&b->key));
1175
1176                 btree_node_free(n, op);
1177                 up_write(&n->lock);
1178         }
1179
1180         return b;
1181 }
1182
1183 /*
1184  * Leaving this at 2 until we've got incremental garbage collection done; it
1185  * could be higher (and has been tested with 4) except that garbage collection
1186  * could take much longer, adversely affecting latency.
1187  */
1188 #define GC_MERGE_NODES  2U
1189
1190 struct gc_merge_info {
1191         struct btree    *b;
1192         struct bkey     *k;
1193         unsigned        keys;
1194 };
1195
1196 static void btree_gc_coalesce(struct btree *b, struct btree_op *op,
1197                               struct gc_stat *gc, struct gc_merge_info *r)
1198 {
1199         unsigned nodes = 0, keys = 0, blocks;
1200         int i;
1201
1202         while (nodes < GC_MERGE_NODES && r[nodes].b)
1203                 keys += r[nodes++].keys;
1204
1205         blocks = btree_default_blocks(b->c) * 2 / 3;
1206
1207         if (nodes < 2 ||
1208             __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
1209                 return;
1210
1211         for (i = nodes - 1; i >= 0; --i) {
1212                 if (r[i].b->written)
1213                         r[i].b = btree_gc_alloc(r[i].b, r[i].k, op);
1214
1215                 if (r[i].b->written)
1216                         return;
1217         }
1218
1219         for (i = nodes - 1; i > 0; --i) {
1220                 struct bset *n1 = r[i].b->sets->data;
1221                 struct bset *n2 = r[i - 1].b->sets->data;
1222                 struct bkey *k, *last = NULL;
1223
1224                 keys = 0;
1225
1226                 if (i == 1) {
1227                         /*
1228                          * Last node we're not getting rid of - we're getting
1229                          * rid of the node at r[0]. Have to try and fit all of
1230                          * the remaining keys into this node; we can't ensure
1231                          * they will always fit due to rounding and variable
1232                          * length keys (shouldn't be possible in practice,
1233                          * though)
1234                          */
1235                         if (__set_blocks(n1, n1->keys + r->keys,
1236                                          b->c) > btree_blocks(r[i].b))
1237                                 return;
1238
1239                         keys = n2->keys;
1240                         last = &r->b->key;
1241                 } else
1242                         for (k = n2->start;
1243                              k < end(n2);
1244                              k = bkey_next(k)) {
1245                                 if (__set_blocks(n1, n1->keys + keys +
1246                                                  bkey_u64s(k), b->c) > blocks)
1247                                         break;
1248
1249                                 last = k;
1250                                 keys += bkey_u64s(k);
1251                         }
1252
1253                 BUG_ON(__set_blocks(n1, n1->keys + keys,
1254                                     b->c) > btree_blocks(r[i].b));
1255
1256                 if (last) {
1257                         bkey_copy_key(&r[i].b->key, last);
1258                         bkey_copy_key(r[i].k, last);
1259                 }
1260
1261                 memcpy(end(n1),
1262                        n2->start,
1263                        (void *) node(n2, keys) - (void *) n2->start);
1264
1265                 n1->keys += keys;
1266
1267                 memmove(n2->start,
1268                         node(n2, keys),
1269                         (void *) end(n2) - (void *) node(n2, keys));
1270
1271                 n2->keys -= keys;
1272
1273                 r[i].keys       = n1->keys;
1274                 r[i - 1].keys   = n2->keys;
1275         }
1276
1277         btree_node_free(r->b, op);
1278         up_write(&r->b->lock);
1279
1280         trace_bcache_btree_gc_coalesce(nodes);
1281
1282         gc->nodes--;
1283         nodes--;
1284
1285         memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes);
1286         memset(&r[nodes], 0, sizeof(struct gc_merge_info));
1287 }
1288
1289 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1290                             struct closure *writes, struct gc_stat *gc)
1291 {
1292         void write(struct btree *r)
1293         {
1294                 if (!r->written)
1295                         bch_btree_node_write(r, &op->cl);
1296                 else if (btree_node_dirty(r))
1297                         bch_btree_node_write(r, writes);
1298
1299                 up_write(&r->lock);
1300         }
1301
1302         int ret = 0, stale;
1303         unsigned i;
1304         struct gc_merge_info r[GC_MERGE_NODES];
1305
1306         memset(r, 0, sizeof(r));
1307
1308         while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) {
1309                 r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op);
1310
1311                 if (IS_ERR(r->b)) {
1312                         ret = PTR_ERR(r->b);
1313                         break;
1314                 }
1315
1316                 r->keys = 0;
1317                 stale = btree_gc_mark_node(r->b, &r->keys, gc);
1318
1319                 if (!b->written &&
1320                     (r->b->level || stale > 10 ||
1321                      b->c->gc_always_rewrite))
1322                         r->b = btree_gc_alloc(r->b, r->k, op);
1323
1324                 if (r->b->level)
1325                         ret = btree_gc_recurse(r->b, op, writes, gc);
1326
1327                 if (ret) {
1328                         write(r->b);
1329                         break;
1330                 }
1331
1332                 bkey_copy_key(&b->c->gc_done, r->k);
1333
1334                 if (!b->written)
1335                         btree_gc_coalesce(b, op, gc, r);
1336
1337                 if (r[GC_MERGE_NODES - 1].b)
1338                         write(r[GC_MERGE_NODES - 1].b);
1339
1340                 memmove(&r[1], &r[0],
1341                         sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1));
1342
1343                 /* When we've got incremental GC working, we'll want to do
1344                  * if (should_resched())
1345                  *      return -EAGAIN;
1346                  */
1347                 cond_resched();
1348 #if 0
1349                 if (need_resched()) {
1350                         ret = -EAGAIN;
1351                         break;
1352                 }
1353 #endif
1354         }
1355
1356         for (i = 1; i < GC_MERGE_NODES && r[i].b; i++)
1357                 write(r[i].b);
1358
1359         /* Might have freed some children, must remove their keys */
1360         if (!b->written)
1361                 bch_btree_sort(b);
1362
1363         return ret;
1364 }
1365
1366 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1367                              struct closure *writes, struct gc_stat *gc)
1368 {
1369         struct btree *n = NULL;
1370         unsigned keys = 0;
1371         int ret = 0, stale = btree_gc_mark_node(b, &keys, gc);
1372
1373         if (b->level || stale > 10)
1374                 n = btree_node_alloc_replacement(b, NULL);
1375
1376         if (!IS_ERR_OR_NULL(n))
1377                 swap(b, n);
1378
1379         if (b->level)
1380                 ret = btree_gc_recurse(b, op, writes, gc);
1381
1382         if (!b->written || btree_node_dirty(b)) {
1383                 bch_btree_node_write(b, n ? &op->cl : NULL);
1384         }
1385
1386         if (!IS_ERR_OR_NULL(n)) {
1387                 closure_sync(&op->cl);
1388                 bch_btree_set_root(b);
1389                 btree_node_free(n, op);
1390                 rw_unlock(true, b);
1391         }
1392
1393         return ret;
1394 }
1395
1396 static void btree_gc_start(struct cache_set *c)
1397 {
1398         struct cache *ca;
1399         struct bucket *b;
1400         unsigned i;
1401
1402         if (!c->gc_mark_valid)
1403                 return;
1404
1405         mutex_lock(&c->bucket_lock);
1406
1407         c->gc_mark_valid = 0;
1408         c->gc_done = ZERO_KEY;
1409
1410         for_each_cache(ca, c, i)
1411                 for_each_bucket(b, ca) {
1412                         b->gc_gen = b->gen;
1413                         if (!atomic_read(&b->pin)) {
1414                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1415                                 SET_GC_SECTORS_USED(b, 0);
1416                         }
1417                 }
1418
1419         mutex_unlock(&c->bucket_lock);
1420 }
1421
1422 size_t bch_btree_gc_finish(struct cache_set *c)
1423 {
1424         size_t available = 0;
1425         struct bucket *b;
1426         struct cache *ca;
1427         unsigned i;
1428
1429         mutex_lock(&c->bucket_lock);
1430
1431         set_gc_sectors(c);
1432         c->gc_mark_valid = 1;
1433         c->need_gc      = 0;
1434
1435         if (c->root)
1436                 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1437                         SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1438                                     GC_MARK_METADATA);
1439
1440         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1441                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1442                             GC_MARK_METADATA);
1443
1444         for_each_cache(ca, c, i) {
1445                 uint64_t *i;
1446
1447                 ca->invalidate_needs_gc = 0;
1448
1449                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1450                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1451
1452                 for (i = ca->prio_buckets;
1453                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1454                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1455
1456                 for_each_bucket(b, ca) {
1457                         b->last_gc      = b->gc_gen;
1458                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1459
1460                         if (!atomic_read(&b->pin) &&
1461                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1462                                 available++;
1463                                 if (!GC_SECTORS_USED(b))
1464                                         bch_bucket_add_unused(ca, b);
1465                         }
1466                 }
1467         }
1468
1469         mutex_unlock(&c->bucket_lock);
1470         return available;
1471 }
1472
1473 static void bch_btree_gc(struct closure *cl)
1474 {
1475         struct cache_set *c = container_of(cl, struct cache_set, gc.cl);
1476         int ret;
1477         unsigned long available;
1478         struct gc_stat stats;
1479         struct closure writes;
1480         struct btree_op op;
1481         uint64_t start_time = local_clock();
1482
1483         trace_bcache_gc_start(c);
1484
1485         memset(&stats, 0, sizeof(struct gc_stat));
1486         closure_init_stack(&writes);
1487         bch_btree_op_init_stack(&op);
1488         op.lock = SHRT_MAX;
1489
1490         btree_gc_start(c);
1491
1492         atomic_inc(&c->prio_blocked);
1493
1494         ret = btree_root(gc_root, c, &op, &writes, &stats);
1495         closure_sync(&op.cl);
1496         closure_sync(&writes);
1497
1498         if (ret) {
1499                 pr_warn("gc failed!");
1500                 continue_at(cl, bch_btree_gc, bch_gc_wq);
1501         }
1502
1503         /* Possibly wait for new UUIDs or whatever to hit disk */
1504         bch_journal_meta(c, &op.cl);
1505         closure_sync(&op.cl);
1506
1507         available = bch_btree_gc_finish(c);
1508
1509         atomic_dec(&c->prio_blocked);
1510         wake_up_allocators(c);
1511
1512         bch_time_stats_update(&c->btree_gc_time, start_time);
1513
1514         stats.key_bytes *= sizeof(uint64_t);
1515         stats.dirty     <<= 9;
1516         stats.data      <<= 9;
1517         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1518         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1519
1520         trace_bcache_gc_end(c);
1521
1522         continue_at(cl, bch_moving_gc, bch_gc_wq);
1523 }
1524
1525 void bch_queue_gc(struct cache_set *c)
1526 {
1527         closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl);
1528 }
1529
1530 /* Initial partial gc */
1531
1532 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1533                                    unsigned long **seen)
1534 {
1535         int ret;
1536         unsigned i;
1537         struct bkey *k;
1538         struct bucket *g;
1539         struct btree_iter iter;
1540
1541         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1542                 for (i = 0; i < KEY_PTRS(k); i++) {
1543                         if (!ptr_available(b->c, k, i))
1544                                 continue;
1545
1546                         g = PTR_BUCKET(b->c, k, i);
1547
1548                         if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1549                                                 seen[PTR_DEV(k, i)]) ||
1550                             !ptr_stale(b->c, k, i)) {
1551                                 g->gen = PTR_GEN(k, i);
1552
1553                                 if (b->level)
1554                                         g->prio = BTREE_PRIO;
1555                                 else if (g->prio == BTREE_PRIO)
1556                                         g->prio = INITIAL_PRIO;
1557                         }
1558                 }
1559
1560                 btree_mark_key(b, k);
1561         }
1562
1563         if (b->level) {
1564                 k = bch_next_recurse_key(b, &ZERO_KEY);
1565
1566                 while (k) {
1567                         struct bkey *p = bch_next_recurse_key(b, k);
1568                         if (p)
1569                                 btree_node_prefetch(b->c, p, b->level - 1);
1570
1571                         ret = btree(check_recurse, k, b, op, seen);
1572                         if (ret)
1573                                 return ret;
1574
1575                         k = p;
1576                 }
1577         }
1578
1579         return 0;
1580 }
1581
1582 int bch_btree_check(struct cache_set *c, struct btree_op *op)
1583 {
1584         int ret = -ENOMEM;
1585         unsigned i;
1586         unsigned long *seen[MAX_CACHES_PER_SET];
1587
1588         memset(seen, 0, sizeof(seen));
1589
1590         for (i = 0; c->cache[i]; i++) {
1591                 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1592                 seen[i] = kmalloc(n, GFP_KERNEL);
1593                 if (!seen[i])
1594                         goto err;
1595
1596                 /* Disables the seen array until prio_read() uses it too */
1597                 memset(seen[i], 0xFF, n);
1598         }
1599
1600         ret = btree_root(check_recurse, c, op, seen);
1601 err:
1602         for (i = 0; i < MAX_CACHES_PER_SET; i++)
1603                 kfree(seen[i]);
1604         return ret;
1605 }
1606
1607 /* Btree insertion */
1608
1609 static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
1610 {
1611         struct bset *i = b->sets[b->nsets].data;
1612
1613         memmove((uint64_t *) where + bkey_u64s(insert),
1614                 where,
1615                 (void *) end(i) - (void *) where);
1616
1617         i->keys += bkey_u64s(insert);
1618         bkey_copy(where, insert);
1619         bch_bset_fix_lookup_table(b, where);
1620 }
1621
1622 static bool fix_overlapping_extents(struct btree *b,
1623                                     struct bkey *insert,
1624                                     struct btree_iter *iter,
1625                                     struct btree_op *op)
1626 {
1627         void subtract_dirty(struct bkey *k, uint64_t offset, int sectors)
1628         {
1629                 if (KEY_DIRTY(k))
1630                         bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1631                                                      offset, -sectors);
1632         }
1633
1634         uint64_t old_offset;
1635         unsigned old_size, sectors_found = 0;
1636
1637         while (1) {
1638                 struct bkey *k = bch_btree_iter_next(iter);
1639                 if (!k ||
1640                     bkey_cmp(&START_KEY(k), insert) >= 0)
1641                         break;
1642
1643                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1644                         continue;
1645
1646                 old_offset = KEY_START(k);
1647                 old_size = KEY_SIZE(k);
1648
1649                 /*
1650                  * We might overlap with 0 size extents; we can't skip these
1651                  * because if they're in the set we're inserting to we have to
1652                  * adjust them so they don't overlap with the key we're
1653                  * inserting. But we don't want to check them for BTREE_REPLACE
1654                  * operations.
1655                  */
1656
1657                 if (op->type == BTREE_REPLACE &&
1658                     KEY_SIZE(k)) {
1659                         /*
1660                          * k might have been split since we inserted/found the
1661                          * key we're replacing
1662                          */
1663                         unsigned i;
1664                         uint64_t offset = KEY_START(k) -
1665                                 KEY_START(&op->replace);
1666
1667                         /* But it must be a subset of the replace key */
1668                         if (KEY_START(k) < KEY_START(&op->replace) ||
1669                             KEY_OFFSET(k) > KEY_OFFSET(&op->replace))
1670                                 goto check_failed;
1671
1672                         /* We didn't find a key that we were supposed to */
1673                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
1674                                 goto check_failed;
1675
1676                         if (KEY_PTRS(&op->replace) != KEY_PTRS(k))
1677                                 goto check_failed;
1678
1679                         /* skip past gen */
1680                         offset <<= 8;
1681
1682                         BUG_ON(!KEY_PTRS(&op->replace));
1683
1684                         for (i = 0; i < KEY_PTRS(&op->replace); i++)
1685                                 if (k->ptr[i] != op->replace.ptr[i] + offset)
1686                                         goto check_failed;
1687
1688                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1689                 }
1690
1691                 if (bkey_cmp(insert, k) < 0 &&
1692                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1693                         /*
1694                          * We overlapped in the middle of an existing key: that
1695                          * means we have to split the old key. But we have to do
1696                          * slightly different things depending on whether the
1697                          * old key has been written out yet.
1698                          */
1699
1700                         struct bkey *top;
1701
1702                         subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert));
1703
1704                         if (bkey_written(b, k)) {
1705                                 /*
1706                                  * We insert a new key to cover the top of the
1707                                  * old key, and the old key is modified in place
1708                                  * to represent the bottom split.
1709                                  *
1710                                  * It's completely arbitrary whether the new key
1711                                  * is the top or the bottom, but it has to match
1712                                  * up with what btree_sort_fixup() does - it
1713                                  * doesn't check for this kind of overlap, it
1714                                  * depends on us inserting a new key for the top
1715                                  * here.
1716                                  */
1717                                 top = bch_bset_search(b, &b->sets[b->nsets],
1718                                                       insert);
1719                                 shift_keys(b, top, k);
1720                         } else {
1721                                 BKEY_PADDED(key) temp;
1722                                 bkey_copy(&temp.key, k);
1723                                 shift_keys(b, k, &temp.key);
1724                                 top = bkey_next(k);
1725                         }
1726
1727                         bch_cut_front(insert, top);
1728                         bch_cut_back(&START_KEY(insert), k);
1729                         bch_bset_fix_invalidated_key(b, k);
1730                         return false;
1731                 }
1732
1733                 if (bkey_cmp(insert, k) < 0) {
1734                         bch_cut_front(insert, k);
1735                 } else {
1736                         if (bkey_written(b, k) &&
1737                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1738                                 /*
1739                                  * Completely overwrote, so we don't have to
1740                                  * invalidate the binary search tree
1741                                  */
1742                                 bch_cut_front(k, k);
1743                         } else {
1744                                 __bch_cut_back(&START_KEY(insert), k);
1745                                 bch_bset_fix_invalidated_key(b, k);
1746                         }
1747                 }
1748
1749                 subtract_dirty(k, old_offset, old_size - KEY_SIZE(k));
1750         }
1751
1752 check_failed:
1753         if (op->type == BTREE_REPLACE) {
1754                 if (!sectors_found) {
1755                         op->insert_collision = true;
1756                         return true;
1757                 } else if (sectors_found < KEY_SIZE(insert)) {
1758                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1759                                        (KEY_SIZE(insert) - sectors_found));
1760                         SET_KEY_SIZE(insert, sectors_found);
1761                 }
1762         }
1763
1764         return false;
1765 }
1766
1767 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1768                              struct bkey *k)
1769 {
1770         struct bset *i = b->sets[b->nsets].data;
1771         struct bkey *m, *prev;
1772         unsigned status = BTREE_INSERT_STATUS_INSERT;
1773
1774         BUG_ON(bkey_cmp(k, &b->key) > 0);
1775         BUG_ON(b->level && !KEY_PTRS(k));
1776         BUG_ON(!b->level && !KEY_OFFSET(k));
1777
1778         if (!b->level) {
1779                 struct btree_iter iter;
1780                 struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0);
1781
1782                 /*
1783                  * bset_search() returns the first key that is strictly greater
1784                  * than the search key - but for back merging, we want to find
1785                  * the first key that is greater than or equal to KEY_START(k) -
1786                  * unless KEY_START(k) is 0.
1787                  */
1788                 if (KEY_OFFSET(&search))
1789                         SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1);
1790
1791                 prev = NULL;
1792                 m = bch_btree_iter_init(b, &iter, &search);
1793
1794                 if (fix_overlapping_extents(b, k, &iter, op))
1795                         return false;
1796
1797                 while (m != end(i) &&
1798                        bkey_cmp(k, &START_KEY(m)) > 0)
1799                         prev = m, m = bkey_next(m);
1800
1801                 if (key_merging_disabled(b->c))
1802                         goto insert;
1803
1804                 /* prev is in the tree, if we merge we're done */
1805                 status = BTREE_INSERT_STATUS_BACK_MERGE;
1806                 if (prev &&
1807                     bch_bkey_try_merge(b, prev, k))
1808                         goto merged;
1809
1810                 status = BTREE_INSERT_STATUS_OVERWROTE;
1811                 if (m != end(i) &&
1812                     KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
1813                         goto copy;
1814
1815                 status = BTREE_INSERT_STATUS_FRONT_MERGE;
1816                 if (m != end(i) &&
1817                     bch_bkey_try_merge(b, k, m))
1818                         goto copy;
1819         } else
1820                 m = bch_bset_search(b, &b->sets[b->nsets], k);
1821
1822 insert: shift_keys(b, m, k);
1823 copy:   bkey_copy(m, k);
1824 merged:
1825         if (KEY_DIRTY(k))
1826                 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
1827                                              KEY_START(k), KEY_SIZE(k));
1828
1829         bch_check_keys(b, "%u for %s", status, op_type(op));
1830
1831         if (b->level && !KEY_OFFSET(k))
1832                 btree_current_write(b)->prio_blocked++;
1833
1834         trace_bcache_btree_insert_key(b, k, op->type, status);
1835
1836         return true;
1837 }
1838
1839 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op)
1840 {
1841         bool ret = false;
1842         struct bkey *k;
1843         unsigned oldsize = bch_count_data(b);
1844
1845         while ((k = bch_keylist_pop(&op->keys))) {
1846                 bkey_put(b->c, k, b->level);
1847                 ret |= btree_insert_key(b, op, k);
1848         }
1849
1850         BUG_ON(bch_count_data(b) < oldsize);
1851         return ret;
1852 }
1853
1854 bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
1855                                    struct bio *bio)
1856 {
1857         bool ret = false;
1858         uint64_t btree_ptr = b->key.ptr[0];
1859         unsigned long seq = b->seq;
1860         BKEY_PADDED(k) tmp;
1861
1862         rw_unlock(false, b);
1863         rw_lock(true, b, b->level);
1864
1865         if (b->key.ptr[0] != btree_ptr ||
1866             b->seq != seq + 1 ||
1867             should_split(b))
1868                 goto out;
1869
1870         op->replace = KEY(op->inode, bio_end_sector(bio), bio_sectors(bio));
1871
1872         SET_KEY_PTRS(&op->replace, 1);
1873         get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t));
1874
1875         SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV);
1876
1877         bkey_copy(&tmp.k, &op->replace);
1878
1879         BUG_ON(op->type != BTREE_INSERT);
1880         BUG_ON(!btree_insert_key(b, op, &tmp.k));
1881         ret = true;
1882 out:
1883         downgrade_write(&b->lock);
1884         return ret;
1885 }
1886
1887 static int btree_split(struct btree *b, struct btree_op *op)
1888 {
1889         bool split, root = b == b->c->root;
1890         struct btree *n1, *n2 = NULL, *n3 = NULL;
1891         uint64_t start_time = local_clock();
1892
1893         if (b->level)
1894                 set_closure_blocking(&op->cl);
1895
1896         n1 = btree_node_alloc_replacement(b, &op->cl);
1897         if (IS_ERR(n1))
1898                 goto err;
1899
1900         split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
1901
1902         if (split) {
1903                 unsigned keys = 0;
1904
1905                 trace_bcache_btree_node_split(b, n1->sets[0].data->keys);
1906
1907                 n2 = bch_btree_node_alloc(b->c, b->level, &op->cl);
1908                 if (IS_ERR(n2))
1909                         goto err_free1;
1910
1911                 if (root) {
1912                         n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl);
1913                         if (IS_ERR(n3))
1914                                 goto err_free2;
1915                 }
1916
1917                 bch_btree_insert_keys(n1, op);
1918
1919                 /* Has to be a linear search because we don't have an auxiliary
1920                  * search tree yet
1921                  */
1922
1923                 while (keys < (n1->sets[0].data->keys * 3) / 5)
1924                         keys += bkey_u64s(node(n1->sets[0].data, keys));
1925
1926                 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
1927                 keys += bkey_u64s(node(n1->sets[0].data, keys));
1928
1929                 n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
1930                 n1->sets[0].data->keys = keys;
1931
1932                 memcpy(n2->sets[0].data->start,
1933                        end(n1->sets[0].data),
1934                        n2->sets[0].data->keys * sizeof(uint64_t));
1935
1936                 bkey_copy_key(&n2->key, &b->key);
1937
1938                 bch_keylist_add(&op->keys, &n2->key);
1939                 bch_btree_node_write(n2, &op->cl);
1940                 rw_unlock(true, n2);
1941         } else {
1942                 trace_bcache_btree_node_compact(b, n1->sets[0].data->keys);
1943
1944                 bch_btree_insert_keys(n1, op);
1945         }
1946
1947         bch_keylist_add(&op->keys, &n1->key);
1948         bch_btree_node_write(n1, &op->cl);
1949
1950         if (n3) {
1951                 bkey_copy_key(&n3->key, &MAX_KEY);
1952                 bch_btree_insert_keys(n3, op);
1953                 bch_btree_node_write(n3, &op->cl);
1954
1955                 closure_sync(&op->cl);
1956                 bch_btree_set_root(n3);
1957                 rw_unlock(true, n3);
1958         } else if (root) {
1959                 op->keys.top = op->keys.bottom;
1960                 closure_sync(&op->cl);
1961                 bch_btree_set_root(n1);
1962         } else {
1963                 unsigned i;
1964
1965                 bkey_copy(op->keys.top, &b->key);
1966                 bkey_copy_key(op->keys.top, &ZERO_KEY);
1967
1968                 for (i = 0; i < KEY_PTRS(&b->key); i++) {
1969                         uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1;
1970
1971                         SET_PTR_GEN(op->keys.top, i, g);
1972                 }
1973
1974                 bch_keylist_push(&op->keys);
1975                 closure_sync(&op->cl);
1976                 atomic_inc(&b->c->prio_blocked);
1977         }
1978
1979         rw_unlock(true, n1);
1980         btree_node_free(b, op);
1981
1982         bch_time_stats_update(&b->c->btree_split_time, start_time);
1983
1984         return 0;
1985 err_free2:
1986         __bkey_put(n2->c, &n2->key);
1987         btree_node_free(n2, op);
1988         rw_unlock(true, n2);
1989 err_free1:
1990         __bkey_put(n1->c, &n1->key);
1991         btree_node_free(n1, op);
1992         rw_unlock(true, n1);
1993 err:
1994         if (n3 == ERR_PTR(-EAGAIN) ||
1995             n2 == ERR_PTR(-EAGAIN) ||
1996             n1 == ERR_PTR(-EAGAIN))
1997                 return -EAGAIN;
1998
1999         pr_warn("couldn't split");
2000         return -ENOMEM;
2001 }
2002
2003 static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op,
2004                                     struct keylist *stack_keys)
2005 {
2006         if (b->level) {
2007                 int ret;
2008                 struct bkey *insert = op->keys.bottom;
2009                 struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert));
2010
2011                 if (!k) {
2012                         btree_bug(b, "no key to recurse on at level %i/%i",
2013                                   b->level, b->c->root->level);
2014
2015                         op->keys.top = op->keys.bottom;
2016                         return -EIO;
2017                 }
2018
2019                 if (bkey_cmp(insert, k) > 0) {
2020                         unsigned i;
2021
2022                         if (op->type == BTREE_REPLACE) {
2023                                 __bkey_put(b->c, insert);
2024                                 op->keys.top = op->keys.bottom;
2025                                 op->insert_collision = true;
2026                                 return 0;
2027                         }
2028
2029                         for (i = 0; i < KEY_PTRS(insert); i++)
2030                                 atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin);
2031
2032                         bkey_copy(stack_keys->top, insert);
2033
2034                         bch_cut_back(k, insert);
2035                         bch_cut_front(k, stack_keys->top);
2036
2037                         bch_keylist_push(stack_keys);
2038                 }
2039
2040                 ret = btree(insert_recurse, k, b, op, stack_keys);
2041                 if (ret)
2042                         return ret;
2043         }
2044
2045         if (!bch_keylist_empty(&op->keys)) {
2046                 if (should_split(b)) {
2047                         if (op->lock <= b->c->root->level) {
2048                                 BUG_ON(b->level);
2049                                 op->lock = b->c->root->level + 1;
2050                                 return -EINTR;
2051                         }
2052                         return btree_split(b, op);
2053                 }
2054
2055                 BUG_ON(write_block(b) != b->sets[b->nsets].data);
2056
2057                 if (bch_btree_insert_keys(b, op)) {
2058                         if (!b->level)
2059                                 bch_btree_leaf_dirty(b, op);
2060                         else
2061                                 bch_btree_node_write(b, &op->cl);
2062                 }
2063         }
2064
2065         return 0;
2066 }
2067
2068 int bch_btree_insert(struct btree_op *op, struct cache_set *c)
2069 {
2070         int ret = 0;
2071         struct keylist stack_keys;
2072
2073         /*
2074          * Don't want to block with the btree locked unless we have to,
2075          * otherwise we get deadlocks with try_harder and between split/gc
2076          */
2077         clear_closure_blocking(&op->cl);
2078
2079         BUG_ON(bch_keylist_empty(&op->keys));
2080         bch_keylist_copy(&stack_keys, &op->keys);
2081         bch_keylist_init(&op->keys);
2082
2083         while (!bch_keylist_empty(&stack_keys) ||
2084                !bch_keylist_empty(&op->keys)) {
2085                 if (bch_keylist_empty(&op->keys)) {
2086                         bch_keylist_add(&op->keys,
2087                                         bch_keylist_pop(&stack_keys));
2088                         op->lock = 0;
2089                 }
2090
2091                 ret = btree_root(insert_recurse, c, op, &stack_keys);
2092
2093                 if (ret == -EAGAIN) {
2094                         ret = 0;
2095                         closure_sync(&op->cl);
2096                 } else if (ret) {
2097                         struct bkey *k;
2098
2099                         pr_err("error %i trying to insert key for %s",
2100                                ret, op_type(op));
2101
2102                         while ((k = bch_keylist_pop(&stack_keys) ?:
2103                                     bch_keylist_pop(&op->keys)))
2104                                 bkey_put(c, k, 0);
2105                 }
2106         }
2107
2108         bch_keylist_free(&stack_keys);
2109
2110         if (op->journal)
2111                 atomic_dec_bug(op->journal);
2112         op->journal = NULL;
2113         return ret;
2114 }
2115
2116 void bch_btree_set_root(struct btree *b)
2117 {
2118         unsigned i;
2119         struct closure cl;
2120
2121         closure_init_stack(&cl);
2122
2123         trace_bcache_btree_set_root(b);
2124
2125         BUG_ON(!b->written);
2126
2127         for (i = 0; i < KEY_PTRS(&b->key); i++)
2128                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2129
2130         mutex_lock(&b->c->bucket_lock);
2131         list_del_init(&b->list);
2132         mutex_unlock(&b->c->bucket_lock);
2133
2134         b->c->root = b;
2135         __bkey_put(b->c, &b->key);
2136
2137         bch_journal_meta(b->c, &cl);
2138         closure_sync(&cl);
2139 }
2140
2141 /* Cache lookup */
2142
2143 static int submit_partial_cache_miss(struct btree *b, struct btree_op *op,
2144                                      struct bkey *k)
2145 {
2146         struct search *s = container_of(op, struct search, op);
2147         struct bio *bio = &s->bio.bio;
2148         int ret = 0;
2149
2150         while (!ret &&
2151                !op->lookup_done) {
2152                 unsigned sectors = INT_MAX;
2153
2154                 if (KEY_INODE(k) == op->inode) {
2155                         if (KEY_START(k) <= bio->bi_sector)
2156                                 break;
2157
2158                         sectors = min_t(uint64_t, sectors,
2159                                         KEY_START(k) - bio->bi_sector);
2160                 }
2161
2162                 ret = s->d->cache_miss(b, s, bio, sectors);
2163         }
2164
2165         return ret;
2166 }
2167
2168 /*
2169  * Read from a single key, handling the initial cache miss if the key starts in
2170  * the middle of the bio
2171  */
2172 static int submit_partial_cache_hit(struct btree *b, struct btree_op *op,
2173                                     struct bkey *k)
2174 {
2175         struct search *s = container_of(op, struct search, op);
2176         struct bio *bio = &s->bio.bio;
2177         unsigned ptr;
2178         struct bio *n;
2179
2180         int ret = submit_partial_cache_miss(b, op, k);
2181         if (ret || op->lookup_done)
2182                 return ret;
2183
2184         /* XXX: figure out best pointer - for multiple cache devices */
2185         ptr = 0;
2186
2187         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
2188
2189         while (!op->lookup_done &&
2190                KEY_INODE(k) == op->inode &&
2191                bio->bi_sector < KEY_OFFSET(k)) {
2192                 struct bkey *bio_key;
2193                 sector_t sector = PTR_OFFSET(k, ptr) +
2194                         (bio->bi_sector - KEY_START(k));
2195                 unsigned sectors = min_t(uint64_t, INT_MAX,
2196                                          KEY_OFFSET(k) - bio->bi_sector);
2197
2198                 n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split);
2199                 if (n == bio)
2200                         op->lookup_done = true;
2201
2202                 bio_key = &container_of(n, struct bbio, bio)->key;
2203
2204                 /*
2205                  * The bucket we're reading from might be reused while our bio
2206                  * is in flight, and we could then end up reading the wrong
2207                  * data.
2208                  *
2209                  * We guard against this by checking (in cache_read_endio()) if
2210                  * the pointer is stale again; if so, we treat it as an error
2211                  * and reread from the backing device (but we don't pass that
2212                  * error up anywhere).
2213                  */
2214
2215                 bch_bkey_copy_single_ptr(bio_key, k, ptr);
2216                 SET_PTR_OFFSET(bio_key, 0, sector);
2217
2218                 n->bi_end_io    = bch_cache_read_endio;
2219                 n->bi_private   = &s->cl;
2220
2221                 __bch_submit_bbio(n, b->c);
2222         }
2223
2224         return 0;
2225 }
2226
2227 int bch_btree_search_recurse(struct btree *b, struct btree_op *op)
2228 {
2229         struct search *s = container_of(op, struct search, op);
2230         struct bio *bio = &s->bio.bio;
2231
2232         int ret = 0;
2233         struct bkey *k;
2234         struct btree_iter iter;
2235         bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0));
2236
2237         do {
2238                 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
2239                 if (!k) {
2240                         /*
2241                          * b->key would be exactly what we want, except that
2242                          * pointers to btree nodes have nonzero size - we
2243                          * wouldn't go far enough
2244                          */
2245
2246                         ret = submit_partial_cache_miss(b, op,
2247                                         &KEY(KEY_INODE(&b->key),
2248                                              KEY_OFFSET(&b->key), 0));
2249                         break;
2250                 }
2251
2252                 ret = b->level
2253                         ? btree(search_recurse, k, b, op)
2254                         : submit_partial_cache_hit(b, op, k);
2255         } while (!ret &&
2256                  !op->lookup_done);
2257
2258         return ret;
2259 }
2260
2261 /* Keybuf code */
2262
2263 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2264 {
2265         /* Overlapping keys compare equal */
2266         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2267                 return -1;
2268         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2269                 return 1;
2270         return 0;
2271 }
2272
2273 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2274                                             struct keybuf_key *r)
2275 {
2276         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2277 }
2278
2279 static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op,
2280                                    struct keybuf *buf, struct bkey *end,
2281                                    keybuf_pred_fn *pred)
2282 {
2283         struct btree_iter iter;
2284         bch_btree_iter_init(b, &iter, &buf->last_scanned);
2285
2286         while (!array_freelist_empty(&buf->freelist)) {
2287                 struct bkey *k = bch_btree_iter_next_filter(&iter, b,
2288                                                             bch_ptr_bad);
2289
2290                 if (!b->level) {
2291                         if (!k) {
2292                                 buf->last_scanned = b->key;
2293                                 break;
2294                         }
2295
2296                         buf->last_scanned = *k;
2297                         if (bkey_cmp(&buf->last_scanned, end) >= 0)
2298                                 break;
2299
2300                         if (pred(buf, k)) {
2301                                 struct keybuf_key *w;
2302
2303                                 spin_lock(&buf->lock);
2304
2305                                 w = array_alloc(&buf->freelist);
2306
2307                                 w->private = NULL;
2308                                 bkey_copy(&w->key, k);
2309
2310                                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2311                                         array_free(&buf->freelist, w);
2312
2313                                 spin_unlock(&buf->lock);
2314                         }
2315                 } else {
2316                         if (!k)
2317                                 break;
2318
2319                         btree(refill_keybuf, k, b, op, buf, end, pred);
2320                         /*
2321                          * Might get an error here, but can't really do anything
2322                          * and it'll get logged elsewhere. Just read what we
2323                          * can.
2324                          */
2325
2326                         if (bkey_cmp(&buf->last_scanned, end) >= 0)
2327                                 break;
2328
2329                         cond_resched();
2330                 }
2331         }
2332
2333         return 0;
2334 }
2335
2336 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2337                        struct bkey *end, keybuf_pred_fn *pred)
2338 {
2339         struct bkey start = buf->last_scanned;
2340         struct btree_op op;
2341         bch_btree_op_init_stack(&op);
2342
2343         cond_resched();
2344
2345         btree_root(refill_keybuf, c, &op, buf, end, pred);
2346         closure_sync(&op.cl);
2347
2348         pr_debug("found %s keys from %llu:%llu to %llu:%llu",
2349                  RB_EMPTY_ROOT(&buf->keys) ? "no" :
2350                  array_freelist_empty(&buf->freelist) ? "some" : "a few",
2351                  KEY_INODE(&start), KEY_OFFSET(&start),
2352                  KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned));
2353
2354         spin_lock(&buf->lock);
2355
2356         if (!RB_EMPTY_ROOT(&buf->keys)) {
2357                 struct keybuf_key *w;
2358                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2359                 buf->start      = START_KEY(&w->key);
2360
2361                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2362                 buf->end        = w->key;
2363         } else {
2364                 buf->start      = MAX_KEY;
2365                 buf->end        = MAX_KEY;
2366         }
2367
2368         spin_unlock(&buf->lock);
2369 }
2370
2371 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2372 {
2373         rb_erase(&w->node, &buf->keys);
2374         array_free(&buf->freelist, w);
2375 }
2376
2377 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2378 {
2379         spin_lock(&buf->lock);
2380         __bch_keybuf_del(buf, w);
2381         spin_unlock(&buf->lock);
2382 }
2383
2384 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2385                                   struct bkey *end)
2386 {
2387         bool ret = false;
2388         struct keybuf_key *p, *w, s;
2389         s.key = *start;
2390
2391         if (bkey_cmp(end, &buf->start) <= 0 ||
2392             bkey_cmp(start, &buf->end) >= 0)
2393                 return false;
2394
2395         spin_lock(&buf->lock);
2396         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2397
2398         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2399                 p = w;
2400                 w = RB_NEXT(w, node);
2401
2402                 if (p->private)
2403                         ret = true;
2404                 else
2405                         __bch_keybuf_del(buf, p);
2406         }
2407
2408         spin_unlock(&buf->lock);
2409         return ret;
2410 }
2411
2412 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2413 {
2414         struct keybuf_key *w;
2415         spin_lock(&buf->lock);
2416
2417         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2418
2419         while (w && w->private)
2420                 w = RB_NEXT(w, node);
2421
2422         if (w)
2423                 w->private = ERR_PTR(-EINTR);
2424
2425         spin_unlock(&buf->lock);
2426         return w;
2427 }
2428
2429 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2430                                              struct keybuf *buf,
2431                                              struct bkey *end,
2432                                              keybuf_pred_fn *pred)
2433 {
2434         struct keybuf_key *ret;
2435
2436         while (1) {
2437                 ret = bch_keybuf_next(buf);
2438                 if (ret)
2439                         break;
2440
2441                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2442                         pr_debug("scan finished");
2443                         break;
2444                 }
2445
2446                 bch_refill_keybuf(c, buf, end, pred);
2447         }
2448
2449         return ret;
2450 }
2451
2452 void bch_keybuf_init(struct keybuf *buf)
2453 {
2454         buf->last_scanned       = MAX_KEY;
2455         buf->keys               = RB_ROOT;
2456
2457         spin_lock_init(&buf->lock);
2458         array_allocator_init(&buf->freelist);
2459 }
2460
2461 void bch_btree_exit(void)
2462 {
2463         if (btree_io_wq)
2464                 destroy_workqueue(btree_io_wq);
2465         if (bch_gc_wq)
2466                 destroy_workqueue(bch_gc_wq);
2467 }
2468
2469 int __init bch_btree_init(void)
2470 {
2471         if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) ||
2472             !(btree_io_wq = create_singlethread_workqueue("bch_btree_io")))
2473                 return -ENOMEM;
2474
2475         return 0;
2476 }