btree.c revision c18536a72ddd7fe30d63e6c1500b5c930ac14594
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 "writeback.h" 27 28#include <linux/slab.h> 29#include <linux/bitops.h> 30#include <linux/freezer.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 <trace/events/bcache.h> 37 38/* 39 * Todo: 40 * register_bcache: Return errors out to userspace correctly 41 * 42 * Writeback: don't undirty key until after a cache flush 43 * 44 * Create an iterator for key pointers 45 * 46 * On btree write error, mark bucket such that it won't be freed from the cache 47 * 48 * Journalling: 49 * Check for bad keys in replay 50 * Propagate barriers 51 * Refcount journal entries in journal_replay 52 * 53 * Garbage collection: 54 * Finish incremental gc 55 * Gc should free old UUIDs, data for invalid UUIDs 56 * 57 * Provide a way to list backing device UUIDs we have data cached for, and 58 * probably how long it's been since we've seen them, and a way to invalidate 59 * dirty data for devices that will never be attached again 60 * 61 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so 62 * that based on that and how much dirty data we have we can keep writeback 63 * from being starved 64 * 65 * Add a tracepoint or somesuch to watch for writeback starvation 66 * 67 * When btree depth > 1 and splitting an interior node, we have to make sure 68 * alloc_bucket() cannot fail. This should be true but is not completely 69 * obvious. 70 * 71 * Make sure all allocations get charged to the root cgroup 72 * 73 * Plugging? 74 * 75 * If data write is less than hard sector size of ssd, round up offset in open 76 * bucket to the next whole sector 77 * 78 * Also lookup by cgroup in get_open_bucket() 79 * 80 * Superblock needs to be fleshed out for multiple cache devices 81 * 82 * Add a sysfs tunable for the number of writeback IOs in flight 83 * 84 * Add a sysfs tunable for the number of open data buckets 85 * 86 * IO tracking: Can we track when one process is doing io on behalf of another? 87 * IO tracking: Don't use just an average, weigh more recent stuff higher 88 * 89 * Test module load/unload 90 */ 91 92static const char * const op_types[] = { 93 "insert", "replace" 94}; 95 96static const char *op_type(struct btree_op *op) 97{ 98 return op_types[op->type]; 99} 100 101enum { 102 BTREE_INSERT_STATUS_INSERT, 103 BTREE_INSERT_STATUS_BACK_MERGE, 104 BTREE_INSERT_STATUS_OVERWROTE, 105 BTREE_INSERT_STATUS_FRONT_MERGE, 106}; 107 108#define MAX_NEED_GC 64 109#define MAX_SAVE_PRIO 72 110 111#define PTR_DIRTY_BIT (((uint64_t) 1 << 36)) 112 113#define PTR_HASH(c, k) \ 114 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0)) 115 116static struct workqueue_struct *btree_io_wq; 117 118void bch_btree_op_init_stack(struct btree_op *op) 119{ 120 memset(op, 0, sizeof(struct btree_op)); 121 closure_init_stack(&op->cl); 122 op->lock = -1; 123} 124 125static inline bool should_split(struct btree *b) 126{ 127 struct bset *i = write_block(b); 128 return b->written >= btree_blocks(b) || 129 (b->written + __set_blocks(i, i->keys + 15, b->c) 130 > btree_blocks(b)); 131} 132 133#define insert_lock(s, b) ((b)->level <= (s)->lock) 134 135/* 136 * These macros are for recursing down the btree - they handle the details of 137 * locking and looking up nodes in the cache for you. They're best treated as 138 * mere syntax when reading code that uses them. 139 * 140 * op->lock determines whether we take a read or a write lock at a given depth. 141 * If you've got a read lock and find that you need a write lock (i.e. you're 142 * going to have to split), set op->lock and return -EINTR; btree_root() will 143 * call you again and you'll have the correct lock. 144 */ 145 146/** 147 * btree - recurse down the btree on a specified key 148 * @fn: function to call, which will be passed the child node 149 * @key: key to recurse on 150 * @b: parent btree node 151 * @op: pointer to struct btree_op 152 */ 153#define btree(fn, key, b, op, ...) \ 154({ \ 155 int _r, l = (b)->level - 1; \ 156 bool _w = l <= (op)->lock; \ 157 struct btree *_child = bch_btree_node_get((b)->c, key, l, _w); \ 158 if (!IS_ERR(_child)) { \ 159 _child->parent = (b); \ 160 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \ 161 rw_unlock(_w, _child); \ 162 } else \ 163 _r = PTR_ERR(_child); \ 164 _r; \ 165}) 166 167/** 168 * btree_root - call a function on the root of the btree 169 * @fn: function to call, which will be passed the child node 170 * @c: cache set 171 * @op: pointer to struct btree_op 172 */ 173#define btree_root(fn, c, op, ...) \ 174({ \ 175 int _r = -EINTR; \ 176 do { \ 177 struct btree *_b = (c)->root; \ 178 bool _w = insert_lock(op, _b); \ 179 rw_lock(_w, _b, _b->level); \ 180 if (_b == (c)->root && \ 181 _w == insert_lock(op, _b)) { \ 182 _b->parent = NULL; \ 183 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \ 184 } \ 185 rw_unlock(_w, _b); \ 186 bch_cannibalize_unlock(c); \ 187 if (_r == -ENOSPC) { \ 188 wait_event((c)->try_wait, \ 189 !(c)->try_harder); \ 190 _r = -EINTR; \ 191 } \ 192 } while (_r == -EINTR); \ 193 \ 194 _r; \ 195}) 196 197/* Btree key manipulation */ 198 199void __bkey_put(struct cache_set *c, struct bkey *k) 200{ 201 unsigned i; 202 203 for (i = 0; i < KEY_PTRS(k); i++) 204 if (ptr_available(c, k, i)) 205 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin); 206} 207 208static void bkey_put(struct cache_set *c, struct bkey *k, int level) 209{ 210 if ((level && KEY_OFFSET(k)) || !level) 211 __bkey_put(c, k); 212} 213 214/* Btree IO */ 215 216static uint64_t btree_csum_set(struct btree *b, struct bset *i) 217{ 218 uint64_t crc = b->key.ptr[0]; 219 void *data = (void *) i + 8, *end = end(i); 220 221 crc = bch_crc64_update(crc, data, end - data); 222 return crc ^ 0xffffffffffffffffULL; 223} 224 225static void bch_btree_node_read_done(struct btree *b) 226{ 227 const char *err = "bad btree header"; 228 struct bset *i = b->sets[0].data; 229 struct btree_iter *iter; 230 231 iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT); 232 iter->size = b->c->sb.bucket_size / b->c->sb.block_size; 233 iter->used = 0; 234 235 if (!i->seq) 236 goto err; 237 238 for (; 239 b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq; 240 i = write_block(b)) { 241 err = "unsupported bset version"; 242 if (i->version > BCACHE_BSET_VERSION) 243 goto err; 244 245 err = "bad btree header"; 246 if (b->written + set_blocks(i, b->c) > btree_blocks(b)) 247 goto err; 248 249 err = "bad magic"; 250 if (i->magic != bset_magic(b->c)) 251 goto err; 252 253 err = "bad checksum"; 254 switch (i->version) { 255 case 0: 256 if (i->csum != csum_set(i)) 257 goto err; 258 break; 259 case BCACHE_BSET_VERSION: 260 if (i->csum != btree_csum_set(b, i)) 261 goto err; 262 break; 263 } 264 265 err = "empty set"; 266 if (i != b->sets[0].data && !i->keys) 267 goto err; 268 269 bch_btree_iter_push(iter, i->start, end(i)); 270 271 b->written += set_blocks(i, b->c); 272 } 273 274 err = "corrupted btree"; 275 for (i = write_block(b); 276 index(i, b) < btree_blocks(b); 277 i = ((void *) i) + block_bytes(b->c)) 278 if (i->seq == b->sets[0].data->seq) 279 goto err; 280 281 bch_btree_sort_and_fix_extents(b, iter); 282 283 i = b->sets[0].data; 284 err = "short btree key"; 285 if (b->sets[0].size && 286 bkey_cmp(&b->key, &b->sets[0].end) < 0) 287 goto err; 288 289 if (b->written < btree_blocks(b)) 290 bch_bset_init_next(b); 291out: 292 mempool_free(iter, b->c->fill_iter); 293 return; 294err: 295 set_btree_node_io_error(b); 296 bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys", 297 err, PTR_BUCKET_NR(b->c, &b->key, 0), 298 index(i, b), i->keys); 299 goto out; 300} 301 302static void btree_node_read_endio(struct bio *bio, int error) 303{ 304 struct closure *cl = bio->bi_private; 305 closure_put(cl); 306} 307 308void bch_btree_node_read(struct btree *b) 309{ 310 uint64_t start_time = local_clock(); 311 struct closure cl; 312 struct bio *bio; 313 314 trace_bcache_btree_read(b); 315 316 closure_init_stack(&cl); 317 318 bio = bch_bbio_alloc(b->c); 319 bio->bi_rw = REQ_META|READ_SYNC; 320 bio->bi_size = KEY_SIZE(&b->key) << 9; 321 bio->bi_end_io = btree_node_read_endio; 322 bio->bi_private = &cl; 323 324 bch_bio_map(bio, b->sets[0].data); 325 326 bch_submit_bbio(bio, b->c, &b->key, 0); 327 closure_sync(&cl); 328 329 if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) 330 set_btree_node_io_error(b); 331 332 bch_bbio_free(bio, b->c); 333 334 if (btree_node_io_error(b)) 335 goto err; 336 337 bch_btree_node_read_done(b); 338 339 spin_lock(&b->c->btree_read_time_lock); 340 bch_time_stats_update(&b->c->btree_read_time, start_time); 341 spin_unlock(&b->c->btree_read_time_lock); 342 343 return; 344err: 345 bch_cache_set_error(b->c, "io error reading bucket %zu", 346 PTR_BUCKET_NR(b->c, &b->key, 0)); 347} 348 349static void btree_complete_write(struct btree *b, struct btree_write *w) 350{ 351 if (w->prio_blocked && 352 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) 353 wake_up_allocators(b->c); 354 355 if (w->journal) { 356 atomic_dec_bug(w->journal); 357 __closure_wake_up(&b->c->journal.wait); 358 } 359 360 w->prio_blocked = 0; 361 w->journal = NULL; 362} 363 364static void __btree_node_write_done(struct closure *cl) 365{ 366 struct btree *b = container_of(cl, struct btree, io.cl); 367 struct btree_write *w = btree_prev_write(b); 368 369 bch_bbio_free(b->bio, b->c); 370 b->bio = NULL; 371 btree_complete_write(b, w); 372 373 if (btree_node_dirty(b)) 374 queue_delayed_work(btree_io_wq, &b->work, 375 msecs_to_jiffies(30000)); 376 377 closure_return(cl); 378} 379 380static void btree_node_write_done(struct closure *cl) 381{ 382 struct btree *b = container_of(cl, struct btree, io.cl); 383 struct bio_vec *bv; 384 int n; 385 386 __bio_for_each_segment(bv, b->bio, n, 0) 387 __free_page(bv->bv_page); 388 389 __btree_node_write_done(cl); 390} 391 392static void btree_node_write_endio(struct bio *bio, int error) 393{ 394 struct closure *cl = bio->bi_private; 395 struct btree *b = container_of(cl, struct btree, io.cl); 396 397 if (error) 398 set_btree_node_io_error(b); 399 400 bch_bbio_count_io_errors(b->c, bio, error, "writing btree"); 401 closure_put(cl); 402} 403 404static void do_btree_node_write(struct btree *b) 405{ 406 struct closure *cl = &b->io.cl; 407 struct bset *i = b->sets[b->nsets].data; 408 BKEY_PADDED(key) k; 409 410 i->version = BCACHE_BSET_VERSION; 411 i->csum = btree_csum_set(b, i); 412 413 BUG_ON(b->bio); 414 b->bio = bch_bbio_alloc(b->c); 415 416 b->bio->bi_end_io = btree_node_write_endio; 417 b->bio->bi_private = &b->io.cl; 418 b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA; 419 b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c); 420 bch_bio_map(b->bio, i); 421 422 /* 423 * If we're appending to a leaf node, we don't technically need FUA - 424 * this write just needs to be persisted before the next journal write, 425 * which will be marked FLUSH|FUA. 426 * 427 * Similarly if we're writing a new btree root - the pointer is going to 428 * be in the next journal entry. 429 * 430 * But if we're writing a new btree node (that isn't a root) or 431 * appending to a non leaf btree node, we need either FUA or a flush 432 * when we write the parent with the new pointer. FUA is cheaper than a 433 * flush, and writes appending to leaf nodes aren't blocking anything so 434 * just make all btree node writes FUA to keep things sane. 435 */ 436 437 bkey_copy(&k.key, &b->key); 438 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); 439 440 if (!bio_alloc_pages(b->bio, GFP_NOIO)) { 441 int j; 442 struct bio_vec *bv; 443 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); 444 445 bio_for_each_segment(bv, b->bio, j) 446 memcpy(page_address(bv->bv_page), 447 base + j * PAGE_SIZE, PAGE_SIZE); 448 449 bch_submit_bbio(b->bio, b->c, &k.key, 0); 450 451 continue_at(cl, btree_node_write_done, NULL); 452 } else { 453 b->bio->bi_vcnt = 0; 454 bch_bio_map(b->bio, i); 455 456 bch_submit_bbio(b->bio, b->c, &k.key, 0); 457 458 closure_sync(cl); 459 __btree_node_write_done(cl); 460 } 461} 462 463void bch_btree_node_write(struct btree *b, struct closure *parent) 464{ 465 struct bset *i = b->sets[b->nsets].data; 466 467 trace_bcache_btree_write(b); 468 469 BUG_ON(current->bio_list); 470 BUG_ON(b->written >= btree_blocks(b)); 471 BUG_ON(b->written && !i->keys); 472 BUG_ON(b->sets->data->seq != i->seq); 473 bch_check_key_order(b, i); 474 475 cancel_delayed_work(&b->work); 476 477 /* If caller isn't waiting for write, parent refcount is cache set */ 478 closure_lock(&b->io, parent ?: &b->c->cl); 479 480 clear_bit(BTREE_NODE_dirty, &b->flags); 481 change_bit(BTREE_NODE_write_idx, &b->flags); 482 483 do_btree_node_write(b); 484 485 b->written += set_blocks(i, b->c); 486 atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, 487 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); 488 489 bch_btree_sort_lazy(b); 490 491 if (b->written < btree_blocks(b)) 492 bch_bset_init_next(b); 493} 494 495static void btree_node_write_work(struct work_struct *w) 496{ 497 struct btree *b = container_of(to_delayed_work(w), struct btree, work); 498 499 rw_lock(true, b, b->level); 500 501 if (btree_node_dirty(b)) 502 bch_btree_node_write(b, NULL); 503 rw_unlock(true, b); 504} 505 506static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref) 507{ 508 struct bset *i = b->sets[b->nsets].data; 509 struct btree_write *w = btree_current_write(b); 510 511 BUG_ON(!b->written); 512 BUG_ON(!i->keys); 513 514 if (!btree_node_dirty(b)) 515 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); 516 517 set_btree_node_dirty(b); 518 519 if (journal_ref) { 520 if (w->journal && 521 journal_pin_cmp(b->c, w->journal, journal_ref)) { 522 atomic_dec_bug(w->journal); 523 w->journal = NULL; 524 } 525 526 if (!w->journal) { 527 w->journal = journal_ref; 528 atomic_inc(w->journal); 529 } 530 } 531 532 /* Force write if set is too big */ 533 if (set_bytes(i) > PAGE_SIZE - 48 && 534 !current->bio_list) 535 bch_btree_node_write(b, NULL); 536} 537 538/* 539 * Btree in memory cache - allocation/freeing 540 * mca -> memory cache 541 */ 542 543static void mca_reinit(struct btree *b) 544{ 545 unsigned i; 546 547 b->flags = 0; 548 b->written = 0; 549 b->nsets = 0; 550 551 for (i = 0; i < MAX_BSETS; i++) 552 b->sets[i].size = 0; 553 /* 554 * Second loop starts at 1 because b->sets[0]->data is the memory we 555 * allocated 556 */ 557 for (i = 1; i < MAX_BSETS; i++) 558 b->sets[i].data = NULL; 559} 560 561#define mca_reserve(c) (((c->root && c->root->level) \ 562 ? c->root->level : 1) * 8 + 16) 563#define mca_can_free(c) \ 564 max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) 565 566static void mca_data_free(struct btree *b) 567{ 568 struct bset_tree *t = b->sets; 569 BUG_ON(!closure_is_unlocked(&b->io.cl)); 570 571 if (bset_prev_bytes(b) < PAGE_SIZE) 572 kfree(t->prev); 573 else 574 free_pages((unsigned long) t->prev, 575 get_order(bset_prev_bytes(b))); 576 577 if (bset_tree_bytes(b) < PAGE_SIZE) 578 kfree(t->tree); 579 else 580 free_pages((unsigned long) t->tree, 581 get_order(bset_tree_bytes(b))); 582 583 free_pages((unsigned long) t->data, b->page_order); 584 585 t->prev = NULL; 586 t->tree = NULL; 587 t->data = NULL; 588 list_move(&b->list, &b->c->btree_cache_freed); 589 b->c->bucket_cache_used--; 590} 591 592static void mca_bucket_free(struct btree *b) 593{ 594 BUG_ON(btree_node_dirty(b)); 595 596 b->key.ptr[0] = 0; 597 hlist_del_init_rcu(&b->hash); 598 list_move(&b->list, &b->c->btree_cache_freeable); 599} 600 601static unsigned btree_order(struct bkey *k) 602{ 603 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); 604} 605 606static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) 607{ 608 struct bset_tree *t = b->sets; 609 BUG_ON(t->data); 610 611 b->page_order = max_t(unsigned, 612 ilog2(b->c->btree_pages), 613 btree_order(k)); 614 615 t->data = (void *) __get_free_pages(gfp, b->page_order); 616 if (!t->data) 617 goto err; 618 619 t->tree = bset_tree_bytes(b) < PAGE_SIZE 620 ? kmalloc(bset_tree_bytes(b), gfp) 621 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); 622 if (!t->tree) 623 goto err; 624 625 t->prev = bset_prev_bytes(b) < PAGE_SIZE 626 ? kmalloc(bset_prev_bytes(b), gfp) 627 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); 628 if (!t->prev) 629 goto err; 630 631 list_move(&b->list, &b->c->btree_cache); 632 b->c->bucket_cache_used++; 633 return; 634err: 635 mca_data_free(b); 636} 637 638static struct btree *mca_bucket_alloc(struct cache_set *c, 639 struct bkey *k, gfp_t gfp) 640{ 641 struct btree *b = kzalloc(sizeof(struct btree), gfp); 642 if (!b) 643 return NULL; 644 645 init_rwsem(&b->lock); 646 lockdep_set_novalidate_class(&b->lock); 647 INIT_LIST_HEAD(&b->list); 648 INIT_DELAYED_WORK(&b->work, btree_node_write_work); 649 b->c = c; 650 closure_init_unlocked(&b->io); 651 652 mca_data_alloc(b, k, gfp); 653 return b; 654} 655 656static int mca_reap(struct btree *b, unsigned min_order, bool flush) 657{ 658 struct closure cl; 659 660 closure_init_stack(&cl); 661 lockdep_assert_held(&b->c->bucket_lock); 662 663 if (!down_write_trylock(&b->lock)) 664 return -ENOMEM; 665 666 BUG_ON(btree_node_dirty(b) && !b->sets[0].data); 667 668 if (b->page_order < min_order || 669 (!flush && 670 (btree_node_dirty(b) || 671 atomic_read(&b->io.cl.remaining) != -1))) { 672 rw_unlock(true, b); 673 return -ENOMEM; 674 } 675 676 if (btree_node_dirty(b)) { 677 bch_btree_node_write(b, &cl); 678 closure_sync(&cl); 679 } 680 681 /* wait for any in flight btree write */ 682 closure_wait_event_sync(&b->io.wait, &cl, 683 atomic_read(&b->io.cl.remaining) == -1); 684 685 return 0; 686} 687 688static unsigned long bch_mca_scan(struct shrinker *shrink, 689 struct shrink_control *sc) 690{ 691 struct cache_set *c = container_of(shrink, struct cache_set, shrink); 692 struct btree *b, *t; 693 unsigned long i, nr = sc->nr_to_scan; 694 unsigned long freed = 0; 695 696 if (c->shrinker_disabled) 697 return SHRINK_STOP; 698 699 if (c->try_harder) 700 return SHRINK_STOP; 701 702 /* Return -1 if we can't do anything right now */ 703 if (sc->gfp_mask & __GFP_IO) 704 mutex_lock(&c->bucket_lock); 705 else if (!mutex_trylock(&c->bucket_lock)) 706 return -1; 707 708 /* 709 * It's _really_ critical that we don't free too many btree nodes - we 710 * have to always leave ourselves a reserve. The reserve is how we 711 * guarantee that allocating memory for a new btree node can always 712 * succeed, so that inserting keys into the btree can always succeed and 713 * IO can always make forward progress: 714 */ 715 nr /= c->btree_pages; 716 nr = min_t(unsigned long, nr, mca_can_free(c)); 717 718 i = 0; 719 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { 720 if (freed >= nr) 721 break; 722 723 if (++i > 3 && 724 !mca_reap(b, 0, false)) { 725 mca_data_free(b); 726 rw_unlock(true, b); 727 freed++; 728 } 729 } 730 731 /* 732 * Can happen right when we first start up, before we've read in any 733 * btree nodes 734 */ 735 if (list_empty(&c->btree_cache)) 736 goto out; 737 738 for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { 739 b = list_first_entry(&c->btree_cache, struct btree, list); 740 list_rotate_left(&c->btree_cache); 741 742 if (!b->accessed && 743 !mca_reap(b, 0, false)) { 744 mca_bucket_free(b); 745 mca_data_free(b); 746 rw_unlock(true, b); 747 freed++; 748 } else 749 b->accessed = 0; 750 } 751out: 752 mutex_unlock(&c->bucket_lock); 753 return freed; 754} 755 756static unsigned long bch_mca_count(struct shrinker *shrink, 757 struct shrink_control *sc) 758{ 759 struct cache_set *c = container_of(shrink, struct cache_set, shrink); 760 761 if (c->shrinker_disabled) 762 return 0; 763 764 if (c->try_harder) 765 return 0; 766 767 return mca_can_free(c) * c->btree_pages; 768} 769 770void bch_btree_cache_free(struct cache_set *c) 771{ 772 struct btree *b; 773 struct closure cl; 774 closure_init_stack(&cl); 775 776 if (c->shrink.list.next) 777 unregister_shrinker(&c->shrink); 778 779 mutex_lock(&c->bucket_lock); 780 781#ifdef CONFIG_BCACHE_DEBUG 782 if (c->verify_data) 783 list_move(&c->verify_data->list, &c->btree_cache); 784#endif 785 786 list_splice(&c->btree_cache_freeable, 787 &c->btree_cache); 788 789 while (!list_empty(&c->btree_cache)) { 790 b = list_first_entry(&c->btree_cache, struct btree, list); 791 792 if (btree_node_dirty(b)) 793 btree_complete_write(b, btree_current_write(b)); 794 clear_bit(BTREE_NODE_dirty, &b->flags); 795 796 mca_data_free(b); 797 } 798 799 while (!list_empty(&c->btree_cache_freed)) { 800 b = list_first_entry(&c->btree_cache_freed, 801 struct btree, list); 802 list_del(&b->list); 803 cancel_delayed_work_sync(&b->work); 804 kfree(b); 805 } 806 807 mutex_unlock(&c->bucket_lock); 808} 809 810int bch_btree_cache_alloc(struct cache_set *c) 811{ 812 unsigned i; 813 814 for (i = 0; i < mca_reserve(c); i++) 815 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL)) 816 return -ENOMEM; 817 818 list_splice_init(&c->btree_cache, 819 &c->btree_cache_freeable); 820 821#ifdef CONFIG_BCACHE_DEBUG 822 mutex_init(&c->verify_lock); 823 824 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); 825 826 if (c->verify_data && 827 c->verify_data->sets[0].data) 828 list_del_init(&c->verify_data->list); 829 else 830 c->verify_data = NULL; 831#endif 832 833 c->shrink.count_objects = bch_mca_count; 834 c->shrink.scan_objects = bch_mca_scan; 835 c->shrink.seeks = 4; 836 c->shrink.batch = c->btree_pages * 2; 837 register_shrinker(&c->shrink); 838 839 return 0; 840} 841 842/* Btree in memory cache - hash table */ 843 844static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) 845{ 846 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; 847} 848 849static struct btree *mca_find(struct cache_set *c, struct bkey *k) 850{ 851 struct btree *b; 852 853 rcu_read_lock(); 854 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) 855 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) 856 goto out; 857 b = NULL; 858out: 859 rcu_read_unlock(); 860 return b; 861} 862 863static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k) 864{ 865 struct btree *b; 866 867 trace_bcache_btree_cache_cannibalize(c); 868 869 if (!c->try_harder) { 870 c->try_harder = current; 871 c->try_harder_start = local_clock(); 872 } else if (c->try_harder != current) 873 return ERR_PTR(-ENOSPC); 874 875 list_for_each_entry_reverse(b, &c->btree_cache, list) 876 if (!mca_reap(b, btree_order(k), false)) 877 return b; 878 879 list_for_each_entry_reverse(b, &c->btree_cache, list) 880 if (!mca_reap(b, btree_order(k), true)) 881 return b; 882 883 return ERR_PTR(-ENOMEM); 884} 885 886/* 887 * We can only have one thread cannibalizing other cached btree nodes at a time, 888 * or we'll deadlock. We use an open coded mutex to ensure that, which a 889 * cannibalize_bucket() will take. This means every time we unlock the root of 890 * the btree, we need to release this lock if we have it held. 891 */ 892static void bch_cannibalize_unlock(struct cache_set *c) 893{ 894 if (c->try_harder == current) { 895 bch_time_stats_update(&c->try_harder_time, c->try_harder_start); 896 c->try_harder = NULL; 897 wake_up(&c->try_wait); 898 } 899} 900 901static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level) 902{ 903 struct btree *b; 904 905 BUG_ON(current->bio_list); 906 907 lockdep_assert_held(&c->bucket_lock); 908 909 if (mca_find(c, k)) 910 return NULL; 911 912 /* btree_free() doesn't free memory; it sticks the node on the end of 913 * the list. Check if there's any freed nodes there: 914 */ 915 list_for_each_entry(b, &c->btree_cache_freeable, list) 916 if (!mca_reap(b, btree_order(k), false)) 917 goto out; 918 919 /* We never free struct btree itself, just the memory that holds the on 920 * disk node. Check the freed list before allocating a new one: 921 */ 922 list_for_each_entry(b, &c->btree_cache_freed, list) 923 if (!mca_reap(b, 0, false)) { 924 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); 925 if (!b->sets[0].data) 926 goto err; 927 else 928 goto out; 929 } 930 931 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); 932 if (!b) 933 goto err; 934 935 BUG_ON(!down_write_trylock(&b->lock)); 936 if (!b->sets->data) 937 goto err; 938out: 939 BUG_ON(!closure_is_unlocked(&b->io.cl)); 940 941 bkey_copy(&b->key, k); 942 list_move(&b->list, &c->btree_cache); 943 hlist_del_init_rcu(&b->hash); 944 hlist_add_head_rcu(&b->hash, mca_hash(c, k)); 945 946 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); 947 b->level = level; 948 b->parent = (void *) ~0UL; 949 950 mca_reinit(b); 951 952 return b; 953err: 954 if (b) 955 rw_unlock(true, b); 956 957 b = mca_cannibalize(c, k); 958 if (!IS_ERR(b)) 959 goto out; 960 961 return b; 962} 963 964/** 965 * bch_btree_node_get - find a btree node in the cache and lock it, reading it 966 * in from disk if necessary. 967 * 968 * If IO is necessary, it uses the closure embedded in struct btree_op to wait; 969 * if that closure is in non blocking mode, will return -EAGAIN. 970 * 971 * The btree node will have either a read or a write lock held, depending on 972 * level and op->lock. 973 */ 974struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, 975 int level, bool write) 976{ 977 int i = 0; 978 struct btree *b; 979 980 BUG_ON(level < 0); 981retry: 982 b = mca_find(c, k); 983 984 if (!b) { 985 if (current->bio_list) 986 return ERR_PTR(-EAGAIN); 987 988 mutex_lock(&c->bucket_lock); 989 b = mca_alloc(c, k, level); 990 mutex_unlock(&c->bucket_lock); 991 992 if (!b) 993 goto retry; 994 if (IS_ERR(b)) 995 return b; 996 997 bch_btree_node_read(b); 998 999 if (!write) 1000 downgrade_write(&b->lock); 1001 } else { 1002 rw_lock(write, b, level); 1003 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { 1004 rw_unlock(write, b); 1005 goto retry; 1006 } 1007 BUG_ON(b->level != level); 1008 } 1009 1010 b->accessed = 1; 1011 1012 for (; i <= b->nsets && b->sets[i].size; i++) { 1013 prefetch(b->sets[i].tree); 1014 prefetch(b->sets[i].data); 1015 } 1016 1017 for (; i <= b->nsets; i++) 1018 prefetch(b->sets[i].data); 1019 1020 if (btree_node_io_error(b)) { 1021 rw_unlock(write, b); 1022 return ERR_PTR(-EIO); 1023 } 1024 1025 BUG_ON(!b->written); 1026 1027 return b; 1028} 1029 1030static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) 1031{ 1032 struct btree *b; 1033 1034 mutex_lock(&c->bucket_lock); 1035 b = mca_alloc(c, k, level); 1036 mutex_unlock(&c->bucket_lock); 1037 1038 if (!IS_ERR_OR_NULL(b)) { 1039 bch_btree_node_read(b); 1040 rw_unlock(true, b); 1041 } 1042} 1043 1044/* Btree alloc */ 1045 1046static void btree_node_free(struct btree *b) 1047{ 1048 unsigned i; 1049 1050 trace_bcache_btree_node_free(b); 1051 1052 BUG_ON(b == b->c->root); 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 cancel_delayed_work(&b->work); 1059 1060 mutex_lock(&b->c->bucket_lock); 1061 1062 for (i = 0; i < KEY_PTRS(&b->key); i++) { 1063 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); 1064 1065 bch_inc_gen(PTR_CACHE(b->c, &b->key, i), 1066 PTR_BUCKET(b->c, &b->key, i)); 1067 } 1068 1069 bch_bucket_free(b->c, &b->key); 1070 mca_bucket_free(b); 1071 mutex_unlock(&b->c->bucket_lock); 1072} 1073 1074struct btree *bch_btree_node_alloc(struct cache_set *c, int level) 1075{ 1076 BKEY_PADDED(key) k; 1077 struct btree *b = ERR_PTR(-EAGAIN); 1078 1079 mutex_lock(&c->bucket_lock); 1080retry: 1081 if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, true)) 1082 goto err; 1083 1084 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); 1085 1086 b = mca_alloc(c, &k.key, level); 1087 if (IS_ERR(b)) 1088 goto err_free; 1089 1090 if (!b) { 1091 cache_bug(c, 1092 "Tried to allocate bucket that was in btree cache"); 1093 __bkey_put(c, &k.key); 1094 goto retry; 1095 } 1096 1097 b->accessed = 1; 1098 bch_bset_init_next(b); 1099 1100 mutex_unlock(&c->bucket_lock); 1101 1102 trace_bcache_btree_node_alloc(b); 1103 return b; 1104err_free: 1105 bch_bucket_free(c, &k.key); 1106 __bkey_put(c, &k.key); 1107err: 1108 mutex_unlock(&c->bucket_lock); 1109 1110 trace_bcache_btree_node_alloc_fail(b); 1111 return b; 1112} 1113 1114static struct btree *btree_node_alloc_replacement(struct btree *b) 1115{ 1116 struct btree *n = bch_btree_node_alloc(b->c, b->level); 1117 if (!IS_ERR_OR_NULL(n)) 1118 bch_btree_sort_into(b, n); 1119 1120 return n; 1121} 1122 1123/* Garbage collection */ 1124 1125uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) 1126{ 1127 uint8_t stale = 0; 1128 unsigned i; 1129 struct bucket *g; 1130 1131 /* 1132 * ptr_invalid() can't return true for the keys that mark btree nodes as 1133 * freed, but since ptr_bad() returns true we'll never actually use them 1134 * for anything and thus we don't want mark their pointers here 1135 */ 1136 if (!bkey_cmp(k, &ZERO_KEY)) 1137 return stale; 1138 1139 for (i = 0; i < KEY_PTRS(k); i++) { 1140 if (!ptr_available(c, k, i)) 1141 continue; 1142 1143 g = PTR_BUCKET(c, k, i); 1144 1145 if (gen_after(g->gc_gen, PTR_GEN(k, i))) 1146 g->gc_gen = PTR_GEN(k, i); 1147 1148 if (ptr_stale(c, k, i)) { 1149 stale = max(stale, ptr_stale(c, k, i)); 1150 continue; 1151 } 1152 1153 cache_bug_on(GC_MARK(g) && 1154 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), 1155 c, "inconsistent ptrs: mark = %llu, level = %i", 1156 GC_MARK(g), level); 1157 1158 if (level) 1159 SET_GC_MARK(g, GC_MARK_METADATA); 1160 else if (KEY_DIRTY(k)) 1161 SET_GC_MARK(g, GC_MARK_DIRTY); 1162 1163 /* guard against overflow */ 1164 SET_GC_SECTORS_USED(g, min_t(unsigned, 1165 GC_SECTORS_USED(g) + KEY_SIZE(k), 1166 (1 << 14) - 1)); 1167 1168 BUG_ON(!GC_SECTORS_USED(g)); 1169 } 1170 1171 return stale; 1172} 1173 1174#define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) 1175 1176static int btree_gc_mark_node(struct btree *b, unsigned *keys, 1177 struct gc_stat *gc) 1178{ 1179 uint8_t stale = 0; 1180 unsigned last_dev = -1; 1181 struct bcache_device *d = NULL; 1182 struct bkey *k; 1183 struct btree_iter iter; 1184 struct bset_tree *t; 1185 1186 gc->nodes++; 1187 1188 for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1189 if (last_dev != KEY_INODE(k)) { 1190 last_dev = KEY_INODE(k); 1191 1192 d = KEY_INODE(k) < b->c->nr_uuids 1193 ? b->c->devices[last_dev] 1194 : NULL; 1195 } 1196 1197 stale = max(stale, btree_mark_key(b, k)); 1198 1199 if (bch_ptr_bad(b, k)) 1200 continue; 1201 1202 *keys += bkey_u64s(k); 1203 1204 gc->key_bytes += bkey_u64s(k); 1205 gc->nkeys++; 1206 1207 gc->data += KEY_SIZE(k); 1208 if (KEY_DIRTY(k)) 1209 gc->dirty += KEY_SIZE(k); 1210 } 1211 1212 for (t = b->sets; t <= &b->sets[b->nsets]; t++) 1213 btree_bug_on(t->size && 1214 bset_written(b, t) && 1215 bkey_cmp(&b->key, &t->end) < 0, 1216 b, "found short btree key in gc"); 1217 1218 return stale; 1219} 1220 1221static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k) 1222{ 1223 /* 1224 * We block priorities from being written for the duration of garbage 1225 * collection, so we can't sleep in btree_alloc() -> 1226 * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it 1227 * our closure. 1228 */ 1229 struct btree *n = btree_node_alloc_replacement(b); 1230 1231 if (!IS_ERR_OR_NULL(n)) { 1232 swap(b, n); 1233 __bkey_put(b->c, &b->key); 1234 1235 memcpy(k->ptr, b->key.ptr, 1236 sizeof(uint64_t) * KEY_PTRS(&b->key)); 1237 1238 btree_node_free(n); 1239 up_write(&n->lock); 1240 } 1241 1242 return b; 1243} 1244 1245/* 1246 * Leaving this at 2 until we've got incremental garbage collection done; it 1247 * could be higher (and has been tested with 4) except that garbage collection 1248 * could take much longer, adversely affecting latency. 1249 */ 1250#define GC_MERGE_NODES 2U 1251 1252struct gc_merge_info { 1253 struct btree *b; 1254 struct bkey *k; 1255 unsigned keys; 1256}; 1257 1258static void btree_gc_coalesce(struct btree *b, struct gc_stat *gc, 1259 struct gc_merge_info *r) 1260{ 1261 unsigned nodes = 0, keys = 0, blocks; 1262 int i; 1263 1264 while (nodes < GC_MERGE_NODES && r[nodes].b) 1265 keys += r[nodes++].keys; 1266 1267 blocks = btree_default_blocks(b->c) * 2 / 3; 1268 1269 if (nodes < 2 || 1270 __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) 1271 return; 1272 1273 for (i = nodes - 1; i >= 0; --i) { 1274 if (r[i].b->written) 1275 r[i].b = btree_gc_alloc(r[i].b, r[i].k); 1276 1277 if (r[i].b->written) 1278 return; 1279 } 1280 1281 for (i = nodes - 1; i > 0; --i) { 1282 struct bset *n1 = r[i].b->sets->data; 1283 struct bset *n2 = r[i - 1].b->sets->data; 1284 struct bkey *k, *last = NULL; 1285 1286 keys = 0; 1287 1288 if (i == 1) { 1289 /* 1290 * Last node we're not getting rid of - we're getting 1291 * rid of the node at r[0]. Have to try and fit all of 1292 * the remaining keys into this node; we can't ensure 1293 * they will always fit due to rounding and variable 1294 * length keys (shouldn't be possible in practice, 1295 * though) 1296 */ 1297 if (__set_blocks(n1, n1->keys + r->keys, 1298 b->c) > btree_blocks(r[i].b)) 1299 return; 1300 1301 keys = n2->keys; 1302 last = &r->b->key; 1303 } else 1304 for (k = n2->start; 1305 k < end(n2); 1306 k = bkey_next(k)) { 1307 if (__set_blocks(n1, n1->keys + keys + 1308 bkey_u64s(k), b->c) > blocks) 1309 break; 1310 1311 last = k; 1312 keys += bkey_u64s(k); 1313 } 1314 1315 BUG_ON(__set_blocks(n1, n1->keys + keys, 1316 b->c) > btree_blocks(r[i].b)); 1317 1318 if (last) { 1319 bkey_copy_key(&r[i].b->key, last); 1320 bkey_copy_key(r[i].k, last); 1321 } 1322 1323 memcpy(end(n1), 1324 n2->start, 1325 (void *) node(n2, keys) - (void *) n2->start); 1326 1327 n1->keys += keys; 1328 1329 memmove(n2->start, 1330 node(n2, keys), 1331 (void *) end(n2) - (void *) node(n2, keys)); 1332 1333 n2->keys -= keys; 1334 1335 r[i].keys = n1->keys; 1336 r[i - 1].keys = n2->keys; 1337 } 1338 1339 btree_node_free(r->b); 1340 up_write(&r->b->lock); 1341 1342 trace_bcache_btree_gc_coalesce(nodes); 1343 1344 gc->nodes--; 1345 nodes--; 1346 1347 memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes); 1348 memset(&r[nodes], 0, sizeof(struct gc_merge_info)); 1349} 1350 1351static int btree_gc_recurse(struct btree *b, struct btree_op *op, 1352 struct closure *writes, struct gc_stat *gc) 1353{ 1354 void write(struct btree *r) 1355 { 1356 if (!r->written) 1357 bch_btree_node_write(r, &op->cl); 1358 else if (btree_node_dirty(r)) 1359 bch_btree_node_write(r, writes); 1360 1361 up_write(&r->lock); 1362 } 1363 1364 int ret = 0, stale; 1365 unsigned i; 1366 struct gc_merge_info r[GC_MERGE_NODES]; 1367 1368 memset(r, 0, sizeof(r)); 1369 1370 while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) { 1371 r->b = bch_btree_node_get(b->c, r->k, b->level - 1, true); 1372 1373 if (IS_ERR(r->b)) { 1374 ret = PTR_ERR(r->b); 1375 break; 1376 } 1377 1378 r->keys = 0; 1379 stale = btree_gc_mark_node(r->b, &r->keys, gc); 1380 1381 if (!b->written && 1382 (r->b->level || stale > 10 || 1383 b->c->gc_always_rewrite)) 1384 r->b = btree_gc_alloc(r->b, r->k); 1385 1386 if (r->b->level) 1387 ret = btree_gc_recurse(r->b, op, writes, gc); 1388 1389 if (ret) { 1390 write(r->b); 1391 break; 1392 } 1393 1394 bkey_copy_key(&b->c->gc_done, r->k); 1395 1396 if (!b->written) 1397 btree_gc_coalesce(b, gc, r); 1398 1399 if (r[GC_MERGE_NODES - 1].b) 1400 write(r[GC_MERGE_NODES - 1].b); 1401 1402 memmove(&r[1], &r[0], 1403 sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1)); 1404 1405 /* When we've got incremental GC working, we'll want to do 1406 * if (should_resched()) 1407 * return -EAGAIN; 1408 */ 1409 cond_resched(); 1410#if 0 1411 if (need_resched()) { 1412 ret = -EAGAIN; 1413 break; 1414 } 1415#endif 1416 } 1417 1418 for (i = 1; i < GC_MERGE_NODES && r[i].b; i++) 1419 write(r[i].b); 1420 1421 /* Might have freed some children, must remove their keys */ 1422 if (!b->written) 1423 bch_btree_sort(b); 1424 1425 return ret; 1426} 1427 1428static int bch_btree_gc_root(struct btree *b, struct btree_op *op, 1429 struct closure *writes, struct gc_stat *gc) 1430{ 1431 struct btree *n = NULL; 1432 unsigned keys = 0; 1433 int ret = 0, stale = btree_gc_mark_node(b, &keys, gc); 1434 1435 if (b->level || stale > 10) 1436 n = btree_node_alloc_replacement(b); 1437 1438 if (!IS_ERR_OR_NULL(n)) 1439 swap(b, n); 1440 1441 if (b->level) 1442 ret = btree_gc_recurse(b, op, writes, gc); 1443 1444 if (!b->written || btree_node_dirty(b)) { 1445 bch_btree_node_write(b, n ? &op->cl : NULL); 1446 } 1447 1448 if (!IS_ERR_OR_NULL(n)) { 1449 closure_sync(&op->cl); 1450 bch_btree_set_root(b); 1451 btree_node_free(n); 1452 rw_unlock(true, b); 1453 } 1454 1455 return ret; 1456} 1457 1458static void btree_gc_start(struct cache_set *c) 1459{ 1460 struct cache *ca; 1461 struct bucket *b; 1462 unsigned i; 1463 1464 if (!c->gc_mark_valid) 1465 return; 1466 1467 mutex_lock(&c->bucket_lock); 1468 1469 c->gc_mark_valid = 0; 1470 c->gc_done = ZERO_KEY; 1471 1472 for_each_cache(ca, c, i) 1473 for_each_bucket(b, ca) { 1474 b->gc_gen = b->gen; 1475 if (!atomic_read(&b->pin)) { 1476 SET_GC_MARK(b, GC_MARK_RECLAIMABLE); 1477 SET_GC_SECTORS_USED(b, 0); 1478 } 1479 } 1480 1481 mutex_unlock(&c->bucket_lock); 1482} 1483 1484size_t bch_btree_gc_finish(struct cache_set *c) 1485{ 1486 size_t available = 0; 1487 struct bucket *b; 1488 struct cache *ca; 1489 unsigned i; 1490 1491 mutex_lock(&c->bucket_lock); 1492 1493 set_gc_sectors(c); 1494 c->gc_mark_valid = 1; 1495 c->need_gc = 0; 1496 1497 if (c->root) 1498 for (i = 0; i < KEY_PTRS(&c->root->key); i++) 1499 SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), 1500 GC_MARK_METADATA); 1501 1502 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) 1503 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), 1504 GC_MARK_METADATA); 1505 1506 for_each_cache(ca, c, i) { 1507 uint64_t *i; 1508 1509 ca->invalidate_needs_gc = 0; 1510 1511 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) 1512 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1513 1514 for (i = ca->prio_buckets; 1515 i < ca->prio_buckets + prio_buckets(ca) * 2; i++) 1516 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1517 1518 for_each_bucket(b, ca) { 1519 b->last_gc = b->gc_gen; 1520 c->need_gc = max(c->need_gc, bucket_gc_gen(b)); 1521 1522 if (!atomic_read(&b->pin) && 1523 GC_MARK(b) == GC_MARK_RECLAIMABLE) { 1524 available++; 1525 if (!GC_SECTORS_USED(b)) 1526 bch_bucket_add_unused(ca, b); 1527 } 1528 } 1529 } 1530 1531 mutex_unlock(&c->bucket_lock); 1532 return available; 1533} 1534 1535static void bch_btree_gc(struct cache_set *c) 1536{ 1537 int ret; 1538 unsigned long available; 1539 struct gc_stat stats; 1540 struct closure writes; 1541 struct btree_op op; 1542 uint64_t start_time = local_clock(); 1543 1544 trace_bcache_gc_start(c); 1545 1546 memset(&stats, 0, sizeof(struct gc_stat)); 1547 closure_init_stack(&writes); 1548 bch_btree_op_init_stack(&op); 1549 op.lock = SHRT_MAX; 1550 1551 btree_gc_start(c); 1552 1553 atomic_inc(&c->prio_blocked); 1554 1555 ret = btree_root(gc_root, c, &op, &writes, &stats); 1556 closure_sync(&op.cl); 1557 closure_sync(&writes); 1558 1559 if (ret) { 1560 pr_warn("gc failed!"); 1561 return; 1562 } 1563 1564 /* Possibly wait for new UUIDs or whatever to hit disk */ 1565 bch_journal_meta(c, &op.cl); 1566 closure_sync(&op.cl); 1567 1568 available = bch_btree_gc_finish(c); 1569 1570 atomic_dec(&c->prio_blocked); 1571 wake_up_allocators(c); 1572 1573 bch_time_stats_update(&c->btree_gc_time, start_time); 1574 1575 stats.key_bytes *= sizeof(uint64_t); 1576 stats.dirty <<= 9; 1577 stats.data <<= 9; 1578 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; 1579 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); 1580 1581 trace_bcache_gc_end(c); 1582 1583 bch_moving_gc(c); 1584} 1585 1586static int bch_gc_thread(void *arg) 1587{ 1588 struct cache_set *c = arg; 1589 1590 while (1) { 1591 bch_btree_gc(c); 1592 1593 set_current_state(TASK_INTERRUPTIBLE); 1594 if (kthread_should_stop()) 1595 break; 1596 1597 try_to_freeze(); 1598 schedule(); 1599 } 1600 1601 return 0; 1602} 1603 1604int bch_gc_thread_start(struct cache_set *c) 1605{ 1606 c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc"); 1607 if (IS_ERR(c->gc_thread)) 1608 return PTR_ERR(c->gc_thread); 1609 1610 set_task_state(c->gc_thread, TASK_INTERRUPTIBLE); 1611 return 0; 1612} 1613 1614/* Initial partial gc */ 1615 1616static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, 1617 unsigned long **seen) 1618{ 1619 int ret; 1620 unsigned i; 1621 struct bkey *k; 1622 struct bucket *g; 1623 struct btree_iter iter; 1624 1625 for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1626 for (i = 0; i < KEY_PTRS(k); i++) { 1627 if (!ptr_available(b->c, k, i)) 1628 continue; 1629 1630 g = PTR_BUCKET(b->c, k, i); 1631 1632 if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), 1633 seen[PTR_DEV(k, i)]) || 1634 !ptr_stale(b->c, k, i)) { 1635 g->gen = PTR_GEN(k, i); 1636 1637 if (b->level) 1638 g->prio = BTREE_PRIO; 1639 else if (g->prio == BTREE_PRIO) 1640 g->prio = INITIAL_PRIO; 1641 } 1642 } 1643 1644 btree_mark_key(b, k); 1645 } 1646 1647 if (b->level) { 1648 k = bch_next_recurse_key(b, &ZERO_KEY); 1649 1650 while (k) { 1651 struct bkey *p = bch_next_recurse_key(b, k); 1652 if (p) 1653 btree_node_prefetch(b->c, p, b->level - 1); 1654 1655 ret = btree(check_recurse, k, b, op, seen); 1656 if (ret) 1657 return ret; 1658 1659 k = p; 1660 } 1661 } 1662 1663 return 0; 1664} 1665 1666int bch_btree_check(struct cache_set *c) 1667{ 1668 int ret = -ENOMEM; 1669 unsigned i; 1670 unsigned long *seen[MAX_CACHES_PER_SET]; 1671 struct btree_op op; 1672 1673 memset(seen, 0, sizeof(seen)); 1674 bch_btree_op_init_stack(&op); 1675 op.lock = SHRT_MAX; 1676 1677 for (i = 0; c->cache[i]; i++) { 1678 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); 1679 seen[i] = kmalloc(n, GFP_KERNEL); 1680 if (!seen[i]) 1681 goto err; 1682 1683 /* Disables the seen array until prio_read() uses it too */ 1684 memset(seen[i], 0xFF, n); 1685 } 1686 1687 ret = btree_root(check_recurse, c, &op, seen); 1688err: 1689 for (i = 0; i < MAX_CACHES_PER_SET; i++) 1690 kfree(seen[i]); 1691 return ret; 1692} 1693 1694/* Btree insertion */ 1695 1696static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) 1697{ 1698 struct bset *i = b->sets[b->nsets].data; 1699 1700 memmove((uint64_t *) where + bkey_u64s(insert), 1701 where, 1702 (void *) end(i) - (void *) where); 1703 1704 i->keys += bkey_u64s(insert); 1705 bkey_copy(where, insert); 1706 bch_bset_fix_lookup_table(b, where); 1707} 1708 1709static bool fix_overlapping_extents(struct btree *b, 1710 struct bkey *insert, 1711 struct btree_iter *iter, 1712 struct btree_op *op) 1713{ 1714 void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) 1715 { 1716 if (KEY_DIRTY(k)) 1717 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 1718 offset, -sectors); 1719 } 1720 1721 uint64_t old_offset; 1722 unsigned old_size, sectors_found = 0; 1723 1724 while (1) { 1725 struct bkey *k = bch_btree_iter_next(iter); 1726 if (!k || 1727 bkey_cmp(&START_KEY(k), insert) >= 0) 1728 break; 1729 1730 if (bkey_cmp(k, &START_KEY(insert)) <= 0) 1731 continue; 1732 1733 old_offset = KEY_START(k); 1734 old_size = KEY_SIZE(k); 1735 1736 /* 1737 * We might overlap with 0 size extents; we can't skip these 1738 * because if they're in the set we're inserting to we have to 1739 * adjust them so they don't overlap with the key we're 1740 * inserting. But we don't want to check them for BTREE_REPLACE 1741 * operations. 1742 */ 1743 1744 if (op->type == BTREE_REPLACE && 1745 KEY_SIZE(k)) { 1746 /* 1747 * k might have been split since we inserted/found the 1748 * key we're replacing 1749 */ 1750 unsigned i; 1751 uint64_t offset = KEY_START(k) - 1752 KEY_START(&op->replace); 1753 1754 /* But it must be a subset of the replace key */ 1755 if (KEY_START(k) < KEY_START(&op->replace) || 1756 KEY_OFFSET(k) > KEY_OFFSET(&op->replace)) 1757 goto check_failed; 1758 1759 /* We didn't find a key that we were supposed to */ 1760 if (KEY_START(k) > KEY_START(insert) + sectors_found) 1761 goto check_failed; 1762 1763 if (KEY_PTRS(&op->replace) != KEY_PTRS(k)) 1764 goto check_failed; 1765 1766 /* skip past gen */ 1767 offset <<= 8; 1768 1769 BUG_ON(!KEY_PTRS(&op->replace)); 1770 1771 for (i = 0; i < KEY_PTRS(&op->replace); i++) 1772 if (k->ptr[i] != op->replace.ptr[i] + offset) 1773 goto check_failed; 1774 1775 sectors_found = KEY_OFFSET(k) - KEY_START(insert); 1776 } 1777 1778 if (bkey_cmp(insert, k) < 0 && 1779 bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { 1780 /* 1781 * We overlapped in the middle of an existing key: that 1782 * means we have to split the old key. But we have to do 1783 * slightly different things depending on whether the 1784 * old key has been written out yet. 1785 */ 1786 1787 struct bkey *top; 1788 1789 subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); 1790 1791 if (bkey_written(b, k)) { 1792 /* 1793 * We insert a new key to cover the top of the 1794 * old key, and the old key is modified in place 1795 * to represent the bottom split. 1796 * 1797 * It's completely arbitrary whether the new key 1798 * is the top or the bottom, but it has to match 1799 * up with what btree_sort_fixup() does - it 1800 * doesn't check for this kind of overlap, it 1801 * depends on us inserting a new key for the top 1802 * here. 1803 */ 1804 top = bch_bset_search(b, &b->sets[b->nsets], 1805 insert); 1806 shift_keys(b, top, k); 1807 } else { 1808 BKEY_PADDED(key) temp; 1809 bkey_copy(&temp.key, k); 1810 shift_keys(b, k, &temp.key); 1811 top = bkey_next(k); 1812 } 1813 1814 bch_cut_front(insert, top); 1815 bch_cut_back(&START_KEY(insert), k); 1816 bch_bset_fix_invalidated_key(b, k); 1817 return false; 1818 } 1819 1820 if (bkey_cmp(insert, k) < 0) { 1821 bch_cut_front(insert, k); 1822 } else { 1823 if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) 1824 old_offset = KEY_START(insert); 1825 1826 if (bkey_written(b, k) && 1827 bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { 1828 /* 1829 * Completely overwrote, so we don't have to 1830 * invalidate the binary search tree 1831 */ 1832 bch_cut_front(k, k); 1833 } else { 1834 __bch_cut_back(&START_KEY(insert), k); 1835 bch_bset_fix_invalidated_key(b, k); 1836 } 1837 } 1838 1839 subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); 1840 } 1841 1842check_failed: 1843 if (op->type == BTREE_REPLACE) { 1844 if (!sectors_found) { 1845 op->insert_collision = true; 1846 return true; 1847 } else if (sectors_found < KEY_SIZE(insert)) { 1848 SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - 1849 (KEY_SIZE(insert) - sectors_found)); 1850 SET_KEY_SIZE(insert, sectors_found); 1851 } 1852 } 1853 1854 return false; 1855} 1856 1857static bool btree_insert_key(struct btree *b, struct btree_op *op, 1858 struct bkey *k) 1859{ 1860 struct bset *i = b->sets[b->nsets].data; 1861 struct bkey *m, *prev; 1862 unsigned status = BTREE_INSERT_STATUS_INSERT; 1863 1864 BUG_ON(bkey_cmp(k, &b->key) > 0); 1865 BUG_ON(b->level && !KEY_PTRS(k)); 1866 BUG_ON(!b->level && !KEY_OFFSET(k)); 1867 1868 if (!b->level) { 1869 struct btree_iter iter; 1870 struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0); 1871 1872 /* 1873 * bset_search() returns the first key that is strictly greater 1874 * than the search key - but for back merging, we want to find 1875 * the first key that is greater than or equal to KEY_START(k) - 1876 * unless KEY_START(k) is 0. 1877 */ 1878 if (KEY_OFFSET(&search)) 1879 SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1); 1880 1881 prev = NULL; 1882 m = bch_btree_iter_init(b, &iter, &search); 1883 1884 if (fix_overlapping_extents(b, k, &iter, op)) 1885 return false; 1886 1887 if (KEY_DIRTY(k)) 1888 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 1889 KEY_START(k), KEY_SIZE(k)); 1890 1891 while (m != end(i) && 1892 bkey_cmp(k, &START_KEY(m)) > 0) 1893 prev = m, m = bkey_next(m); 1894 1895 if (key_merging_disabled(b->c)) 1896 goto insert; 1897 1898 /* prev is in the tree, if we merge we're done */ 1899 status = BTREE_INSERT_STATUS_BACK_MERGE; 1900 if (prev && 1901 bch_bkey_try_merge(b, prev, k)) 1902 goto merged; 1903 1904 status = BTREE_INSERT_STATUS_OVERWROTE; 1905 if (m != end(i) && 1906 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) 1907 goto copy; 1908 1909 status = BTREE_INSERT_STATUS_FRONT_MERGE; 1910 if (m != end(i) && 1911 bch_bkey_try_merge(b, k, m)) 1912 goto copy; 1913 } else 1914 m = bch_bset_search(b, &b->sets[b->nsets], k); 1915 1916insert: shift_keys(b, m, k); 1917copy: bkey_copy(m, k); 1918merged: 1919 bch_check_keys(b, "%u for %s", status, op_type(op)); 1920 1921 if (b->level && !KEY_OFFSET(k)) 1922 btree_current_write(b)->prio_blocked++; 1923 1924 trace_bcache_btree_insert_key(b, k, op->type, status); 1925 1926 return true; 1927} 1928 1929static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, 1930 struct keylist *insert_keys) 1931{ 1932 bool ret = false; 1933 unsigned oldsize = bch_count_data(b); 1934 1935 while (!bch_keylist_empty(insert_keys)) { 1936 struct bset *i = write_block(b); 1937 struct bkey *k = insert_keys->keys; 1938 1939 if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c) 1940 > btree_blocks(b)) 1941 break; 1942 1943 if (bkey_cmp(k, &b->key) <= 0) { 1944 bkey_put(b->c, k, b->level); 1945 1946 ret |= btree_insert_key(b, op, k); 1947 bch_keylist_pop_front(insert_keys); 1948 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { 1949#if 0 1950 if (op->type == BTREE_REPLACE) { 1951 bkey_put(b->c, k, b->level); 1952 bch_keylist_pop_front(insert_keys); 1953 op->insert_collision = true; 1954 break; 1955 } 1956#endif 1957 BKEY_PADDED(key) temp; 1958 bkey_copy(&temp.key, insert_keys->keys); 1959 1960 bch_cut_back(&b->key, &temp.key); 1961 bch_cut_front(&b->key, insert_keys->keys); 1962 1963 ret |= btree_insert_key(b, op, &temp.key); 1964 break; 1965 } else { 1966 break; 1967 } 1968 } 1969 1970 BUG_ON(!bch_keylist_empty(insert_keys) && b->level); 1971 1972 BUG_ON(bch_count_data(b) < oldsize); 1973 return ret; 1974} 1975 1976static int btree_split(struct btree *b, struct btree_op *op, 1977 struct keylist *insert_keys, 1978 struct keylist *parent_keys) 1979{ 1980 bool split; 1981 struct btree *n1, *n2 = NULL, *n3 = NULL; 1982 uint64_t start_time = local_clock(); 1983 1984 n1 = btree_node_alloc_replacement(b); 1985 if (IS_ERR(n1)) 1986 goto err; 1987 1988 split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; 1989 1990 if (split) { 1991 unsigned keys = 0; 1992 1993 trace_bcache_btree_node_split(b, n1->sets[0].data->keys); 1994 1995 n2 = bch_btree_node_alloc(b->c, b->level); 1996 if (IS_ERR(n2)) 1997 goto err_free1; 1998 1999 if (!b->parent) { 2000 n3 = bch_btree_node_alloc(b->c, b->level + 1); 2001 if (IS_ERR(n3)) 2002 goto err_free2; 2003 } 2004 2005 bch_btree_insert_keys(n1, op, insert_keys); 2006 2007 /* 2008 * Has to be a linear search because we don't have an auxiliary 2009 * search tree yet 2010 */ 2011 2012 while (keys < (n1->sets[0].data->keys * 3) / 5) 2013 keys += bkey_u64s(node(n1->sets[0].data, keys)); 2014 2015 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); 2016 keys += bkey_u64s(node(n1->sets[0].data, keys)); 2017 2018 n2->sets[0].data->keys = n1->sets[0].data->keys - keys; 2019 n1->sets[0].data->keys = keys; 2020 2021 memcpy(n2->sets[0].data->start, 2022 end(n1->sets[0].data), 2023 n2->sets[0].data->keys * sizeof(uint64_t)); 2024 2025 bkey_copy_key(&n2->key, &b->key); 2026 2027 bch_keylist_add(parent_keys, &n2->key); 2028 bch_btree_node_write(n2, &op->cl); 2029 rw_unlock(true, n2); 2030 } else { 2031 trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); 2032 2033 bch_btree_insert_keys(n1, op, insert_keys); 2034 } 2035 2036 bch_keylist_add(parent_keys, &n1->key); 2037 bch_btree_node_write(n1, &op->cl); 2038 2039 if (n3) { 2040 /* Depth increases, make a new root */ 2041 2042 bkey_copy_key(&n3->key, &MAX_KEY); 2043 bch_btree_insert_keys(n3, op, parent_keys); 2044 bch_btree_node_write(n3, &op->cl); 2045 2046 closure_sync(&op->cl); 2047 bch_btree_set_root(n3); 2048 rw_unlock(true, n3); 2049 } else if (!b->parent) { 2050 /* Root filled up but didn't need to be split */ 2051 2052 bch_keylist_reset(parent_keys); 2053 closure_sync(&op->cl); 2054 bch_btree_set_root(n1); 2055 } else { 2056 unsigned i; 2057 2058 bkey_copy(parent_keys->top, &b->key); 2059 bkey_copy_key(parent_keys->top, &ZERO_KEY); 2060 2061 for (i = 0; i < KEY_PTRS(&b->key); i++) { 2062 uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1; 2063 2064 SET_PTR_GEN(parent_keys->top, i, g); 2065 } 2066 2067 bch_keylist_push(parent_keys); 2068 closure_sync(&op->cl); 2069 atomic_inc(&b->c->prio_blocked); 2070 } 2071 2072 rw_unlock(true, n1); 2073 btree_node_free(b); 2074 2075 bch_time_stats_update(&b->c->btree_split_time, start_time); 2076 2077 return 0; 2078err_free2: 2079 __bkey_put(n2->c, &n2->key); 2080 btree_node_free(n2); 2081 rw_unlock(true, n2); 2082err_free1: 2083 __bkey_put(n1->c, &n1->key); 2084 btree_node_free(n1); 2085 rw_unlock(true, n1); 2086err: 2087 if (n3 == ERR_PTR(-EAGAIN) || 2088 n2 == ERR_PTR(-EAGAIN) || 2089 n1 == ERR_PTR(-EAGAIN)) 2090 return -EAGAIN; 2091 2092 pr_warn("couldn't split"); 2093 return -ENOMEM; 2094} 2095 2096static int bch_btree_insert_node(struct btree *b, struct btree_op *op, 2097 struct keylist *insert_keys, 2098 atomic_t *journal_ref) 2099{ 2100 int ret = 0; 2101 struct keylist split_keys; 2102 2103 bch_keylist_init(&split_keys); 2104 2105 BUG_ON(b->level); 2106 2107 do { 2108 if (should_split(b)) { 2109 if (current->bio_list) { 2110 op->lock = b->c->root->level + 1; 2111 ret = -EAGAIN; 2112 } else if (op->lock <= b->c->root->level) { 2113 op->lock = b->c->root->level + 1; 2114 ret = -EINTR; 2115 } else { 2116 struct btree *parent = b->parent; 2117 2118 ret = btree_split(b, op, insert_keys, 2119 &split_keys); 2120 insert_keys = &split_keys; 2121 b = parent; 2122 if (!ret) 2123 ret = -EINTR; 2124 } 2125 } else { 2126 BUG_ON(write_block(b) != b->sets[b->nsets].data); 2127 2128 if (bch_btree_insert_keys(b, op, insert_keys)) { 2129 if (!b->level) 2130 bch_btree_leaf_dirty(b, journal_ref); 2131 else 2132 bch_btree_node_write(b, &op->cl); 2133 } 2134 } 2135 } while (!bch_keylist_empty(&split_keys)); 2136 2137 return ret; 2138} 2139 2140int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, 2141 struct bkey *check_key) 2142{ 2143 int ret = -EINTR; 2144 uint64_t btree_ptr = b->key.ptr[0]; 2145 unsigned long seq = b->seq; 2146 struct keylist insert; 2147 bool upgrade = op->lock == -1; 2148 2149 bch_keylist_init(&insert); 2150 2151 if (upgrade) { 2152 rw_unlock(false, b); 2153 rw_lock(true, b, b->level); 2154 2155 if (b->key.ptr[0] != btree_ptr || 2156 b->seq != seq + 1) 2157 goto out; 2158 } 2159 2160 SET_KEY_PTRS(check_key, 1); 2161 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); 2162 2163 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); 2164 2165 bch_keylist_add(&insert, check_key); 2166 2167 BUG_ON(op->type != BTREE_INSERT); 2168 2169 ret = bch_btree_insert_node(b, op, &insert, NULL); 2170 2171 BUG_ON(!ret && !bch_keylist_empty(&insert)); 2172out: 2173 if (upgrade) 2174 downgrade_write(&b->lock); 2175 return ret; 2176} 2177 2178static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op, 2179 struct keylist *keys, atomic_t *journal_ref) 2180{ 2181 if (bch_keylist_empty(keys)) 2182 return 0; 2183 2184 if (b->level) { 2185 struct bkey *k; 2186 2187 k = bch_next_recurse_key(b, &START_KEY(keys->keys)); 2188 if (!k) { 2189 btree_bug(b, "no key to recurse on at level %i/%i", 2190 b->level, b->c->root->level); 2191 2192 bch_keylist_reset(keys); 2193 return -EIO; 2194 } 2195 2196 return btree(insert_recurse, k, b, op, keys, journal_ref); 2197 } else { 2198 return bch_btree_insert_node(b, op, keys, journal_ref); 2199 } 2200} 2201 2202int bch_btree_insert(struct btree_op *op, struct cache_set *c, 2203 struct keylist *keys, atomic_t *journal_ref) 2204{ 2205 int ret = 0; 2206 2207 /* 2208 * Don't want to block with the btree locked unless we have to, 2209 * otherwise we get deadlocks with try_harder and between split/gc 2210 */ 2211 clear_closure_blocking(&op->cl); 2212 2213 BUG_ON(bch_keylist_empty(keys)); 2214 2215 while (!bch_keylist_empty(keys)) { 2216 op->lock = 0; 2217 ret = btree_root(insert_recurse, c, op, keys, journal_ref); 2218 2219 if (ret == -EAGAIN) { 2220 ret = 0; 2221 closure_sync(&op->cl); 2222 } else if (ret) { 2223 struct bkey *k; 2224 2225 pr_err("error %i trying to insert key for %s", 2226 ret, op_type(op)); 2227 2228 while ((k = bch_keylist_pop(keys))) 2229 bkey_put(c, k, 0); 2230 } 2231 } 2232 2233 return ret; 2234} 2235 2236void bch_btree_set_root(struct btree *b) 2237{ 2238 unsigned i; 2239 struct closure cl; 2240 2241 closure_init_stack(&cl); 2242 2243 trace_bcache_btree_set_root(b); 2244 2245 BUG_ON(!b->written); 2246 2247 for (i = 0; i < KEY_PTRS(&b->key); i++) 2248 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); 2249 2250 mutex_lock(&b->c->bucket_lock); 2251 list_del_init(&b->list); 2252 mutex_unlock(&b->c->bucket_lock); 2253 2254 b->c->root = b; 2255 __bkey_put(b->c, &b->key); 2256 2257 bch_journal_meta(b->c, &cl); 2258 closure_sync(&cl); 2259} 2260 2261/* Map across nodes or keys */ 2262 2263static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, 2264 struct bkey *from, 2265 btree_map_nodes_fn *fn, int flags) 2266{ 2267 int ret = MAP_CONTINUE; 2268 2269 if (b->level) { 2270 struct bkey *k; 2271 struct btree_iter iter; 2272 2273 bch_btree_iter_init(b, &iter, from); 2274 2275 while ((k = bch_btree_iter_next_filter(&iter, b, 2276 bch_ptr_bad))) { 2277 ret = btree(map_nodes_recurse, k, b, 2278 op, from, fn, flags); 2279 from = NULL; 2280 2281 if (ret != MAP_CONTINUE) 2282 return ret; 2283 } 2284 } 2285 2286 if (!b->level || flags == MAP_ALL_NODES) 2287 ret = fn(op, b); 2288 2289 return ret; 2290} 2291 2292int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, 2293 struct bkey *from, btree_map_nodes_fn *fn, int flags) 2294{ 2295 int ret = btree_root(map_nodes_recurse, c, op, from, fn, flags); 2296 if (closure_blocking(&op->cl)) 2297 closure_sync(&op->cl); 2298 return ret; 2299} 2300 2301static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, 2302 struct bkey *from, btree_map_keys_fn *fn, 2303 int flags) 2304{ 2305 int ret = MAP_CONTINUE; 2306 struct bkey *k; 2307 struct btree_iter iter; 2308 2309 bch_btree_iter_init(b, &iter, from); 2310 2311 while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) { 2312 ret = !b->level 2313 ? fn(op, b, k) 2314 : btree(map_keys_recurse, k, b, op, from, fn, flags); 2315 from = NULL; 2316 2317 if (ret != MAP_CONTINUE) 2318 return ret; 2319 } 2320 2321 if (!b->level && (flags & MAP_END_KEY)) 2322 ret = fn(op, b, &KEY(KEY_INODE(&b->key), 2323 KEY_OFFSET(&b->key), 0)); 2324 2325 return ret; 2326} 2327 2328int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, 2329 struct bkey *from, btree_map_keys_fn *fn, int flags) 2330{ 2331 int ret = btree_root(map_keys_recurse, c, op, from, fn, flags); 2332 if (closure_blocking(&op->cl)) 2333 closure_sync(&op->cl); 2334 return ret; 2335} 2336 2337/* Keybuf code */ 2338 2339static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) 2340{ 2341 /* Overlapping keys compare equal */ 2342 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) 2343 return -1; 2344 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) 2345 return 1; 2346 return 0; 2347} 2348 2349static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, 2350 struct keybuf_key *r) 2351{ 2352 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); 2353} 2354 2355struct refill { 2356 struct btree_op op; 2357 struct keybuf *buf; 2358 struct bkey *end; 2359 keybuf_pred_fn *pred; 2360}; 2361 2362static int refill_keybuf_fn(struct btree_op *op, struct btree *b, 2363 struct bkey *k) 2364{ 2365 struct refill *refill = container_of(op, struct refill, op); 2366 struct keybuf *buf = refill->buf; 2367 int ret = MAP_CONTINUE; 2368 2369 if (bkey_cmp(k, refill->end) >= 0) { 2370 ret = MAP_DONE; 2371 goto out; 2372 } 2373 2374 if (!KEY_SIZE(k)) /* end key */ 2375 goto out; 2376 2377 if (refill->pred(buf, k)) { 2378 struct keybuf_key *w; 2379 2380 spin_lock(&buf->lock); 2381 2382 w = array_alloc(&buf->freelist); 2383 if (!w) { 2384 spin_unlock(&buf->lock); 2385 return MAP_DONE; 2386 } 2387 2388 w->private = NULL; 2389 bkey_copy(&w->key, k); 2390 2391 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) 2392 array_free(&buf->freelist, w); 2393 2394 if (array_freelist_empty(&buf->freelist)) 2395 ret = MAP_DONE; 2396 2397 spin_unlock(&buf->lock); 2398 } 2399out: 2400 buf->last_scanned = *k; 2401 return ret; 2402} 2403 2404void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, 2405 struct bkey *end, keybuf_pred_fn *pred) 2406{ 2407 struct bkey start = buf->last_scanned; 2408 struct refill refill; 2409 2410 cond_resched(); 2411 2412 bch_btree_op_init_stack(&refill.op); 2413 refill.buf = buf; 2414 refill.end = end; 2415 refill.pred = pred; 2416 2417 bch_btree_map_keys(&refill.op, c, &buf->last_scanned, 2418 refill_keybuf_fn, MAP_END_KEY); 2419 2420 pr_debug("found %s keys from %llu:%llu to %llu:%llu", 2421 RB_EMPTY_ROOT(&buf->keys) ? "no" : 2422 array_freelist_empty(&buf->freelist) ? "some" : "a few", 2423 KEY_INODE(&start), KEY_OFFSET(&start), 2424 KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned)); 2425 2426 spin_lock(&buf->lock); 2427 2428 if (!RB_EMPTY_ROOT(&buf->keys)) { 2429 struct keybuf_key *w; 2430 w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2431 buf->start = START_KEY(&w->key); 2432 2433 w = RB_LAST(&buf->keys, struct keybuf_key, node); 2434 buf->end = w->key; 2435 } else { 2436 buf->start = MAX_KEY; 2437 buf->end = MAX_KEY; 2438 } 2439 2440 spin_unlock(&buf->lock); 2441} 2442 2443static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2444{ 2445 rb_erase(&w->node, &buf->keys); 2446 array_free(&buf->freelist, w); 2447} 2448 2449void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2450{ 2451 spin_lock(&buf->lock); 2452 __bch_keybuf_del(buf, w); 2453 spin_unlock(&buf->lock); 2454} 2455 2456bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, 2457 struct bkey *end) 2458{ 2459 bool ret = false; 2460 struct keybuf_key *p, *w, s; 2461 s.key = *start; 2462 2463 if (bkey_cmp(end, &buf->start) <= 0 || 2464 bkey_cmp(start, &buf->end) >= 0) 2465 return false; 2466 2467 spin_lock(&buf->lock); 2468 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); 2469 2470 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { 2471 p = w; 2472 w = RB_NEXT(w, node); 2473 2474 if (p->private) 2475 ret = true; 2476 else 2477 __bch_keybuf_del(buf, p); 2478 } 2479 2480 spin_unlock(&buf->lock); 2481 return ret; 2482} 2483 2484struct keybuf_key *bch_keybuf_next(struct keybuf *buf) 2485{ 2486 struct keybuf_key *w; 2487 spin_lock(&buf->lock); 2488 2489 w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2490 2491 while (w && w->private) 2492 w = RB_NEXT(w, node); 2493 2494 if (w) 2495 w->private = ERR_PTR(-EINTR); 2496 2497 spin_unlock(&buf->lock); 2498 return w; 2499} 2500 2501struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, 2502 struct keybuf *buf, 2503 struct bkey *end, 2504 keybuf_pred_fn *pred) 2505{ 2506 struct keybuf_key *ret; 2507 2508 while (1) { 2509 ret = bch_keybuf_next(buf); 2510 if (ret) 2511 break; 2512 2513 if (bkey_cmp(&buf->last_scanned, end) >= 0) { 2514 pr_debug("scan finished"); 2515 break; 2516 } 2517 2518 bch_refill_keybuf(c, buf, end, pred); 2519 } 2520 2521 return ret; 2522} 2523 2524void bch_keybuf_init(struct keybuf *buf) 2525{ 2526 buf->last_scanned = MAX_KEY; 2527 buf->keys = RB_ROOT; 2528 2529 spin_lock_init(&buf->lock); 2530 array_allocator_init(&buf->freelist); 2531} 2532 2533void bch_btree_exit(void) 2534{ 2535 if (btree_io_wq) 2536 destroy_workqueue(btree_io_wq); 2537} 2538 2539int __init bch_btree_init(void) 2540{ 2541 btree_io_wq = create_singlethread_workqueue("bch_btree_io"); 2542 if (!btree_io_wq) 2543 return -ENOMEM; 2544 2545 return 0; 2546} 2547