1/* 2 * Code for working with individual keys, and sorted sets of keys with in a 3 * btree node 4 * 5 * Copyright 2012 Google, Inc. 6 */ 7 8#define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__ 9 10#include "util.h" 11#include "bset.h" 12 13#include <linux/console.h> 14#include <linux/random.h> 15#include <linux/prefetch.h> 16 17#ifdef CONFIG_BCACHE_DEBUG 18 19void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned set) 20{ 21 struct bkey *k, *next; 22 23 for (k = i->start; k < bset_bkey_last(i); k = next) { 24 next = bkey_next(k); 25 26 printk(KERN_ERR "block %u key %u/%u: ", set, 27 (unsigned) ((u64 *) k - i->d), i->keys); 28 29 if (b->ops->key_dump) 30 b->ops->key_dump(b, k); 31 else 32 printk("%llu:%llu\n", KEY_INODE(k), KEY_OFFSET(k)); 33 34 if (next < bset_bkey_last(i) && 35 bkey_cmp(k, b->ops->is_extents ? 36 &START_KEY(next) : next) > 0) 37 printk(KERN_ERR "Key skipped backwards\n"); 38 } 39} 40 41void bch_dump_bucket(struct btree_keys *b) 42{ 43 unsigned i; 44 45 console_lock(); 46 for (i = 0; i <= b->nsets; i++) 47 bch_dump_bset(b, b->set[i].data, 48 bset_sector_offset(b, b->set[i].data)); 49 console_unlock(); 50} 51 52int __bch_count_data(struct btree_keys *b) 53{ 54 unsigned ret = 0; 55 struct btree_iter iter; 56 struct bkey *k; 57 58 if (b->ops->is_extents) 59 for_each_key(b, k, &iter) 60 ret += KEY_SIZE(k); 61 return ret; 62} 63 64void __bch_check_keys(struct btree_keys *b, const char *fmt, ...) 65{ 66 va_list args; 67 struct bkey *k, *p = NULL; 68 struct btree_iter iter; 69 const char *err; 70 71 for_each_key(b, k, &iter) { 72 if (b->ops->is_extents) { 73 err = "Keys out of order"; 74 if (p && bkey_cmp(&START_KEY(p), &START_KEY(k)) > 0) 75 goto bug; 76 77 if (bch_ptr_invalid(b, k)) 78 continue; 79 80 err = "Overlapping keys"; 81 if (p && bkey_cmp(p, &START_KEY(k)) > 0) 82 goto bug; 83 } else { 84 if (bch_ptr_bad(b, k)) 85 continue; 86 87 err = "Duplicate keys"; 88 if (p && !bkey_cmp(p, k)) 89 goto bug; 90 } 91 p = k; 92 } 93#if 0 94 err = "Key larger than btree node key"; 95 if (p && bkey_cmp(p, &b->key) > 0) 96 goto bug; 97#endif 98 return; 99bug: 100 bch_dump_bucket(b); 101 102 va_start(args, fmt); 103 vprintk(fmt, args); 104 va_end(args); 105 106 panic("bch_check_keys error: %s:\n", err); 107} 108 109static void bch_btree_iter_next_check(struct btree_iter *iter) 110{ 111 struct bkey *k = iter->data->k, *next = bkey_next(k); 112 113 if (next < iter->data->end && 114 bkey_cmp(k, iter->b->ops->is_extents ? 115 &START_KEY(next) : next) > 0) { 116 bch_dump_bucket(iter->b); 117 panic("Key skipped backwards\n"); 118 } 119} 120 121#else 122 123static inline void bch_btree_iter_next_check(struct btree_iter *iter) {} 124 125#endif 126 127/* Keylists */ 128 129int __bch_keylist_realloc(struct keylist *l, unsigned u64s) 130{ 131 size_t oldsize = bch_keylist_nkeys(l); 132 size_t newsize = oldsize + u64s; 133 uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p; 134 uint64_t *new_keys; 135 136 newsize = roundup_pow_of_two(newsize); 137 138 if (newsize <= KEYLIST_INLINE || 139 roundup_pow_of_two(oldsize) == newsize) 140 return 0; 141 142 new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO); 143 144 if (!new_keys) 145 return -ENOMEM; 146 147 if (!old_keys) 148 memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize); 149 150 l->keys_p = new_keys; 151 l->top_p = new_keys + oldsize; 152 153 return 0; 154} 155 156struct bkey *bch_keylist_pop(struct keylist *l) 157{ 158 struct bkey *k = l->keys; 159 160 if (k == l->top) 161 return NULL; 162 163 while (bkey_next(k) != l->top) 164 k = bkey_next(k); 165 166 return l->top = k; 167} 168 169void bch_keylist_pop_front(struct keylist *l) 170{ 171 l->top_p -= bkey_u64s(l->keys); 172 173 memmove(l->keys, 174 bkey_next(l->keys), 175 bch_keylist_bytes(l)); 176} 177 178/* Key/pointer manipulation */ 179 180void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src, 181 unsigned i) 182{ 183 BUG_ON(i > KEY_PTRS(src)); 184 185 /* Only copy the header, key, and one pointer. */ 186 memcpy(dest, src, 2 * sizeof(uint64_t)); 187 dest->ptr[0] = src->ptr[i]; 188 SET_KEY_PTRS(dest, 1); 189 /* We didn't copy the checksum so clear that bit. */ 190 SET_KEY_CSUM(dest, 0); 191} 192 193bool __bch_cut_front(const struct bkey *where, struct bkey *k) 194{ 195 unsigned i, len = 0; 196 197 if (bkey_cmp(where, &START_KEY(k)) <= 0) 198 return false; 199 200 if (bkey_cmp(where, k) < 0) 201 len = KEY_OFFSET(k) - KEY_OFFSET(where); 202 else 203 bkey_copy_key(k, where); 204 205 for (i = 0; i < KEY_PTRS(k); i++) 206 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len); 207 208 BUG_ON(len > KEY_SIZE(k)); 209 SET_KEY_SIZE(k, len); 210 return true; 211} 212 213bool __bch_cut_back(const struct bkey *where, struct bkey *k) 214{ 215 unsigned len = 0; 216 217 if (bkey_cmp(where, k) >= 0) 218 return false; 219 220 BUG_ON(KEY_INODE(where) != KEY_INODE(k)); 221 222 if (bkey_cmp(where, &START_KEY(k)) > 0) 223 len = KEY_OFFSET(where) - KEY_START(k); 224 225 bkey_copy_key(k, where); 226 227 BUG_ON(len > KEY_SIZE(k)); 228 SET_KEY_SIZE(k, len); 229 return true; 230} 231 232/* Auxiliary search trees */ 233 234/* 32 bits total: */ 235#define BKEY_MID_BITS 3 236#define BKEY_EXPONENT_BITS 7 237#define BKEY_MANTISSA_BITS (32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS) 238#define BKEY_MANTISSA_MASK ((1 << BKEY_MANTISSA_BITS) - 1) 239 240struct bkey_float { 241 unsigned exponent:BKEY_EXPONENT_BITS; 242 unsigned m:BKEY_MID_BITS; 243 unsigned mantissa:BKEY_MANTISSA_BITS; 244} __packed; 245 246/* 247 * BSET_CACHELINE was originally intended to match the hardware cacheline size - 248 * it used to be 64, but I realized the lookup code would touch slightly less 249 * memory if it was 128. 250 * 251 * It definites the number of bytes (in struct bset) per struct bkey_float in 252 * the auxiliar search tree - when we're done searching the bset_float tree we 253 * have this many bytes left that we do a linear search over. 254 * 255 * Since (after level 5) every level of the bset_tree is on a new cacheline, 256 * we're touching one fewer cacheline in the bset tree in exchange for one more 257 * cacheline in the linear search - but the linear search might stop before it 258 * gets to the second cacheline. 259 */ 260 261#define BSET_CACHELINE 128 262 263/* Space required for the btree node keys */ 264static inline size_t btree_keys_bytes(struct btree_keys *b) 265{ 266 return PAGE_SIZE << b->page_order; 267} 268 269static inline size_t btree_keys_cachelines(struct btree_keys *b) 270{ 271 return btree_keys_bytes(b) / BSET_CACHELINE; 272} 273 274/* Space required for the auxiliary search trees */ 275static inline size_t bset_tree_bytes(struct btree_keys *b) 276{ 277 return btree_keys_cachelines(b) * sizeof(struct bkey_float); 278} 279 280/* Space required for the prev pointers */ 281static inline size_t bset_prev_bytes(struct btree_keys *b) 282{ 283 return btree_keys_cachelines(b) * sizeof(uint8_t); 284} 285 286/* Memory allocation */ 287 288void bch_btree_keys_free(struct btree_keys *b) 289{ 290 struct bset_tree *t = b->set; 291 292 if (bset_prev_bytes(b) < PAGE_SIZE) 293 kfree(t->prev); 294 else 295 free_pages((unsigned long) t->prev, 296 get_order(bset_prev_bytes(b))); 297 298 if (bset_tree_bytes(b) < PAGE_SIZE) 299 kfree(t->tree); 300 else 301 free_pages((unsigned long) t->tree, 302 get_order(bset_tree_bytes(b))); 303 304 free_pages((unsigned long) t->data, b->page_order); 305 306 t->prev = NULL; 307 t->tree = NULL; 308 t->data = NULL; 309} 310EXPORT_SYMBOL(bch_btree_keys_free); 311 312int bch_btree_keys_alloc(struct btree_keys *b, unsigned page_order, gfp_t gfp) 313{ 314 struct bset_tree *t = b->set; 315 316 BUG_ON(t->data); 317 318 b->page_order = page_order; 319 320 t->data = (void *) __get_free_pages(gfp, b->page_order); 321 if (!t->data) 322 goto err; 323 324 t->tree = bset_tree_bytes(b) < PAGE_SIZE 325 ? kmalloc(bset_tree_bytes(b), gfp) 326 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); 327 if (!t->tree) 328 goto err; 329 330 t->prev = bset_prev_bytes(b) < PAGE_SIZE 331 ? kmalloc(bset_prev_bytes(b), gfp) 332 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); 333 if (!t->prev) 334 goto err; 335 336 return 0; 337err: 338 bch_btree_keys_free(b); 339 return -ENOMEM; 340} 341EXPORT_SYMBOL(bch_btree_keys_alloc); 342 343void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops, 344 bool *expensive_debug_checks) 345{ 346 unsigned i; 347 348 b->ops = ops; 349 b->expensive_debug_checks = expensive_debug_checks; 350 b->nsets = 0; 351 b->last_set_unwritten = 0; 352 353 /* XXX: shouldn't be needed */ 354 for (i = 0; i < MAX_BSETS; i++) 355 b->set[i].size = 0; 356 /* 357 * Second loop starts at 1 because b->keys[0]->data is the memory we 358 * allocated 359 */ 360 for (i = 1; i < MAX_BSETS; i++) 361 b->set[i].data = NULL; 362} 363EXPORT_SYMBOL(bch_btree_keys_init); 364 365/* Binary tree stuff for auxiliary search trees */ 366 367static unsigned inorder_next(unsigned j, unsigned size) 368{ 369 if (j * 2 + 1 < size) { 370 j = j * 2 + 1; 371 372 while (j * 2 < size) 373 j *= 2; 374 } else 375 j >>= ffz(j) + 1; 376 377 return j; 378} 379 380static unsigned inorder_prev(unsigned j, unsigned size) 381{ 382 if (j * 2 < size) { 383 j = j * 2; 384 385 while (j * 2 + 1 < size) 386 j = j * 2 + 1; 387 } else 388 j >>= ffs(j); 389 390 return j; 391} 392 393/* I have no idea why this code works... and I'm the one who wrote it 394 * 395 * However, I do know what it does: 396 * Given a binary tree constructed in an array (i.e. how you normally implement 397 * a heap), it converts a node in the tree - referenced by array index - to the 398 * index it would have if you did an inorder traversal. 399 * 400 * Also tested for every j, size up to size somewhere around 6 million. 401 * 402 * The binary tree starts at array index 1, not 0 403 * extra is a function of size: 404 * extra = (size - rounddown_pow_of_two(size - 1)) << 1; 405 */ 406static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra) 407{ 408 unsigned b = fls(j); 409 unsigned shift = fls(size - 1) - b; 410 411 j ^= 1U << (b - 1); 412 j <<= 1; 413 j |= 1; 414 j <<= shift; 415 416 if (j > extra) 417 j -= (j - extra) >> 1; 418 419 return j; 420} 421 422static unsigned to_inorder(unsigned j, struct bset_tree *t) 423{ 424 return __to_inorder(j, t->size, t->extra); 425} 426 427static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra) 428{ 429 unsigned shift; 430 431 if (j > extra) 432 j += j - extra; 433 434 shift = ffs(j); 435 436 j >>= shift; 437 j |= roundup_pow_of_two(size) >> shift; 438 439 return j; 440} 441 442static unsigned inorder_to_tree(unsigned j, struct bset_tree *t) 443{ 444 return __inorder_to_tree(j, t->size, t->extra); 445} 446 447#if 0 448void inorder_test(void) 449{ 450 unsigned long done = 0; 451 ktime_t start = ktime_get(); 452 453 for (unsigned size = 2; 454 size < 65536000; 455 size++) { 456 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1; 457 unsigned i = 1, j = rounddown_pow_of_two(size - 1); 458 459 if (!(size % 4096)) 460 printk(KERN_NOTICE "loop %u, %llu per us\n", size, 461 done / ktime_us_delta(ktime_get(), start)); 462 463 while (1) { 464 if (__inorder_to_tree(i, size, extra) != j) 465 panic("size %10u j %10u i %10u", size, j, i); 466 467 if (__to_inorder(j, size, extra) != i) 468 panic("size %10u j %10u i %10u", size, j, i); 469 470 if (j == rounddown_pow_of_two(size) - 1) 471 break; 472 473 BUG_ON(inorder_prev(inorder_next(j, size), size) != j); 474 475 j = inorder_next(j, size); 476 i++; 477 } 478 479 done += size - 1; 480 } 481} 482#endif 483 484/* 485 * Cacheline/offset <-> bkey pointer arithmetic: 486 * 487 * t->tree is a binary search tree in an array; each node corresponds to a key 488 * in one cacheline in t->set (BSET_CACHELINE bytes). 489 * 490 * This means we don't have to store the full index of the key that a node in 491 * the binary tree points to; to_inorder() gives us the cacheline, and then 492 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes. 493 * 494 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to 495 * make this work. 496 * 497 * To construct the bfloat for an arbitrary key we need to know what the key 498 * immediately preceding it is: we have to check if the two keys differ in the 499 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size 500 * of the previous key so we can walk backwards to it from t->tree[j]'s key. 501 */ 502 503static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline, 504 unsigned offset) 505{ 506 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8; 507} 508 509static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k) 510{ 511 return ((void *) k - (void *) t->data) / BSET_CACHELINE; 512} 513 514static unsigned bkey_to_cacheline_offset(struct bset_tree *t, 515 unsigned cacheline, 516 struct bkey *k) 517{ 518 return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0); 519} 520 521static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j) 522{ 523 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m); 524} 525 526static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j) 527{ 528 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]); 529} 530 531/* 532 * For the write set - the one we're currently inserting keys into - we don't 533 * maintain a full search tree, we just keep a simple lookup table in t->prev. 534 */ 535static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline) 536{ 537 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]); 538} 539 540static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift) 541{ 542 low >>= shift; 543 low |= (high << 1) << (63U - shift); 544 return low; 545} 546 547static inline unsigned bfloat_mantissa(const struct bkey *k, 548 struct bkey_float *f) 549{ 550 const uint64_t *p = &k->low - (f->exponent >> 6); 551 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK; 552} 553 554static void make_bfloat(struct bset_tree *t, unsigned j) 555{ 556 struct bkey_float *f = &t->tree[j]; 557 struct bkey *m = tree_to_bkey(t, j); 558 struct bkey *p = tree_to_prev_bkey(t, j); 559 560 struct bkey *l = is_power_of_2(j) 561 ? t->data->start 562 : tree_to_prev_bkey(t, j >> ffs(j)); 563 564 struct bkey *r = is_power_of_2(j + 1) 565 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end)) 566 : tree_to_bkey(t, j >> (ffz(j) + 1)); 567 568 BUG_ON(m < l || m > r); 569 BUG_ON(bkey_next(p) != m); 570 571 if (KEY_INODE(l) != KEY_INODE(r)) 572 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64; 573 else 574 f->exponent = fls64(r->low ^ l->low); 575 576 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0); 577 578 /* 579 * Setting f->exponent = 127 flags this node as failed, and causes the 580 * lookup code to fall back to comparing against the original key. 581 */ 582 583 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f)) 584 f->mantissa = bfloat_mantissa(m, f) - 1; 585 else 586 f->exponent = 127; 587} 588 589static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t) 590{ 591 if (t != b->set) { 592 unsigned j = roundup(t[-1].size, 593 64 / sizeof(struct bkey_float)); 594 595 t->tree = t[-1].tree + j; 596 t->prev = t[-1].prev + j; 597 } 598 599 while (t < b->set + MAX_BSETS) 600 t++->size = 0; 601} 602 603static void bch_bset_build_unwritten_tree(struct btree_keys *b) 604{ 605 struct bset_tree *t = bset_tree_last(b); 606 607 BUG_ON(b->last_set_unwritten); 608 b->last_set_unwritten = 1; 609 610 bset_alloc_tree(b, t); 611 612 if (t->tree != b->set->tree + btree_keys_cachelines(b)) { 613 t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start); 614 t->size = 1; 615 } 616} 617 618void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic) 619{ 620 if (i != b->set->data) { 621 b->set[++b->nsets].data = i; 622 i->seq = b->set->data->seq; 623 } else 624 get_random_bytes(&i->seq, sizeof(uint64_t)); 625 626 i->magic = magic; 627 i->version = 0; 628 i->keys = 0; 629 630 bch_bset_build_unwritten_tree(b); 631} 632EXPORT_SYMBOL(bch_bset_init_next); 633 634void bch_bset_build_written_tree(struct btree_keys *b) 635{ 636 struct bset_tree *t = bset_tree_last(b); 637 struct bkey *prev = NULL, *k = t->data->start; 638 unsigned j, cacheline = 1; 639 640 b->last_set_unwritten = 0; 641 642 bset_alloc_tree(b, t); 643 644 t->size = min_t(unsigned, 645 bkey_to_cacheline(t, bset_bkey_last(t->data)), 646 b->set->tree + btree_keys_cachelines(b) - t->tree); 647 648 if (t->size < 2) { 649 t->size = 0; 650 return; 651 } 652 653 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1; 654 655 /* First we figure out where the first key in each cacheline is */ 656 for (j = inorder_next(0, t->size); 657 j; 658 j = inorder_next(j, t->size)) { 659 while (bkey_to_cacheline(t, k) < cacheline) 660 prev = k, k = bkey_next(k); 661 662 t->prev[j] = bkey_u64s(prev); 663 t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k); 664 } 665 666 while (bkey_next(k) != bset_bkey_last(t->data)) 667 k = bkey_next(k); 668 669 t->end = *k; 670 671 /* Then we build the tree */ 672 for (j = inorder_next(0, t->size); 673 j; 674 j = inorder_next(j, t->size)) 675 make_bfloat(t, j); 676} 677EXPORT_SYMBOL(bch_bset_build_written_tree); 678 679/* Insert */ 680 681void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k) 682{ 683 struct bset_tree *t; 684 unsigned inorder, j = 1; 685 686 for (t = b->set; t <= bset_tree_last(b); t++) 687 if (k < bset_bkey_last(t->data)) 688 goto found_set; 689 690 BUG(); 691found_set: 692 if (!t->size || !bset_written(b, t)) 693 return; 694 695 inorder = bkey_to_cacheline(t, k); 696 697 if (k == t->data->start) 698 goto fix_left; 699 700 if (bkey_next(k) == bset_bkey_last(t->data)) { 701 t->end = *k; 702 goto fix_right; 703 } 704 705 j = inorder_to_tree(inorder, t); 706 707 if (j && 708 j < t->size && 709 k == tree_to_bkey(t, j)) 710fix_left: do { 711 make_bfloat(t, j); 712 j = j * 2; 713 } while (j < t->size); 714 715 j = inorder_to_tree(inorder + 1, t); 716 717 if (j && 718 j < t->size && 719 k == tree_to_prev_bkey(t, j)) 720fix_right: do { 721 make_bfloat(t, j); 722 j = j * 2 + 1; 723 } while (j < t->size); 724} 725EXPORT_SYMBOL(bch_bset_fix_invalidated_key); 726 727static void bch_bset_fix_lookup_table(struct btree_keys *b, 728 struct bset_tree *t, 729 struct bkey *k) 730{ 731 unsigned shift = bkey_u64s(k); 732 unsigned j = bkey_to_cacheline(t, k); 733 734 /* We're getting called from btree_split() or btree_gc, just bail out */ 735 if (!t->size) 736 return; 737 738 /* k is the key we just inserted; we need to find the entry in the 739 * lookup table for the first key that is strictly greater than k: 740 * it's either k's cacheline or the next one 741 */ 742 while (j < t->size && 743 table_to_bkey(t, j) <= k) 744 j++; 745 746 /* Adjust all the lookup table entries, and find a new key for any that 747 * have gotten too big 748 */ 749 for (; j < t->size; j++) { 750 t->prev[j] += shift; 751 752 if (t->prev[j] > 7) { 753 k = table_to_bkey(t, j - 1); 754 755 while (k < cacheline_to_bkey(t, j, 0)) 756 k = bkey_next(k); 757 758 t->prev[j] = bkey_to_cacheline_offset(t, j, k); 759 } 760 } 761 762 if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree) 763 return; 764 765 /* Possibly add a new entry to the end of the lookup table */ 766 767 for (k = table_to_bkey(t, t->size - 1); 768 k != bset_bkey_last(t->data); 769 k = bkey_next(k)) 770 if (t->size == bkey_to_cacheline(t, k)) { 771 t->prev[t->size] = bkey_to_cacheline_offset(t, t->size, k); 772 t->size++; 773 } 774} 775 776/* 777 * Tries to merge l and r: l should be lower than r 778 * Returns true if we were able to merge. If we did merge, l will be the merged 779 * key, r will be untouched. 780 */ 781bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r) 782{ 783 if (!b->ops->key_merge) 784 return false; 785 786 /* 787 * Generic header checks 788 * Assumes left and right are in order 789 * Left and right must be exactly aligned 790 */ 791 if (!bch_bkey_equal_header(l, r) || 792 bkey_cmp(l, &START_KEY(r))) 793 return false; 794 795 return b->ops->key_merge(b, l, r); 796} 797EXPORT_SYMBOL(bch_bkey_try_merge); 798 799void bch_bset_insert(struct btree_keys *b, struct bkey *where, 800 struct bkey *insert) 801{ 802 struct bset_tree *t = bset_tree_last(b); 803 804 BUG_ON(!b->last_set_unwritten); 805 BUG_ON(bset_byte_offset(b, t->data) + 806 __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) > 807 PAGE_SIZE << b->page_order); 808 809 memmove((uint64_t *) where + bkey_u64s(insert), 810 where, 811 (void *) bset_bkey_last(t->data) - (void *) where); 812 813 t->data->keys += bkey_u64s(insert); 814 bkey_copy(where, insert); 815 bch_bset_fix_lookup_table(b, t, where); 816} 817EXPORT_SYMBOL(bch_bset_insert); 818 819unsigned bch_btree_insert_key(struct btree_keys *b, struct bkey *k, 820 struct bkey *replace_key) 821{ 822 unsigned status = BTREE_INSERT_STATUS_NO_INSERT; 823 struct bset *i = bset_tree_last(b)->data; 824 struct bkey *m, *prev = NULL; 825 struct btree_iter iter; 826 827 BUG_ON(b->ops->is_extents && !KEY_SIZE(k)); 828 829 m = bch_btree_iter_init(b, &iter, b->ops->is_extents 830 ? PRECEDING_KEY(&START_KEY(k)) 831 : PRECEDING_KEY(k)); 832 833 if (b->ops->insert_fixup(b, k, &iter, replace_key)) 834 return status; 835 836 status = BTREE_INSERT_STATUS_INSERT; 837 838 while (m != bset_bkey_last(i) && 839 bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0) 840 prev = m, m = bkey_next(m); 841 842 /* prev is in the tree, if we merge we're done */ 843 status = BTREE_INSERT_STATUS_BACK_MERGE; 844 if (prev && 845 bch_bkey_try_merge(b, prev, k)) 846 goto merged; 847#if 0 848 status = BTREE_INSERT_STATUS_OVERWROTE; 849 if (m != bset_bkey_last(i) && 850 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) 851 goto copy; 852#endif 853 status = BTREE_INSERT_STATUS_FRONT_MERGE; 854 if (m != bset_bkey_last(i) && 855 bch_bkey_try_merge(b, k, m)) 856 goto copy; 857 858 bch_bset_insert(b, m, k); 859copy: bkey_copy(m, k); 860merged: 861 return status; 862} 863EXPORT_SYMBOL(bch_btree_insert_key); 864 865/* Lookup */ 866 867struct bset_search_iter { 868 struct bkey *l, *r; 869}; 870 871static struct bset_search_iter bset_search_write_set(struct bset_tree *t, 872 const struct bkey *search) 873{ 874 unsigned li = 0, ri = t->size; 875 876 while (li + 1 != ri) { 877 unsigned m = (li + ri) >> 1; 878 879 if (bkey_cmp(table_to_bkey(t, m), search) > 0) 880 ri = m; 881 else 882 li = m; 883 } 884 885 return (struct bset_search_iter) { 886 table_to_bkey(t, li), 887 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data) 888 }; 889} 890 891static struct bset_search_iter bset_search_tree(struct bset_tree *t, 892 const struct bkey *search) 893{ 894 struct bkey *l, *r; 895 struct bkey_float *f; 896 unsigned inorder, j, n = 1; 897 898 do { 899 unsigned p = n << 4; 900 p &= ((int) (p - t->size)) >> 31; 901 902 prefetch(&t->tree[p]); 903 904 j = n; 905 f = &t->tree[j]; 906 907 /* 908 * n = (f->mantissa > bfloat_mantissa()) 909 * ? j * 2 910 * : j * 2 + 1; 911 * 912 * We need to subtract 1 from f->mantissa for the sign bit trick 913 * to work - that's done in make_bfloat() 914 */ 915 if (likely(f->exponent != 127)) 916 n = j * 2 + (((unsigned) 917 (f->mantissa - 918 bfloat_mantissa(search, f))) >> 31); 919 else 920 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0) 921 ? j * 2 922 : j * 2 + 1; 923 } while (n < t->size); 924 925 inorder = to_inorder(j, t); 926 927 /* 928 * n would have been the node we recursed to - the low bit tells us if 929 * we recursed left or recursed right. 930 */ 931 if (n & 1) { 932 l = cacheline_to_bkey(t, inorder, f->m); 933 934 if (++inorder != t->size) { 935 f = &t->tree[inorder_next(j, t->size)]; 936 r = cacheline_to_bkey(t, inorder, f->m); 937 } else 938 r = bset_bkey_last(t->data); 939 } else { 940 r = cacheline_to_bkey(t, inorder, f->m); 941 942 if (--inorder) { 943 f = &t->tree[inorder_prev(j, t->size)]; 944 l = cacheline_to_bkey(t, inorder, f->m); 945 } else 946 l = t->data->start; 947 } 948 949 return (struct bset_search_iter) {l, r}; 950} 951 952struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t, 953 const struct bkey *search) 954{ 955 struct bset_search_iter i; 956 957 /* 958 * First, we search for a cacheline, then lastly we do a linear search 959 * within that cacheline. 960 * 961 * To search for the cacheline, there's three different possibilities: 962 * * The set is too small to have a search tree, so we just do a linear 963 * search over the whole set. 964 * * The set is the one we're currently inserting into; keeping a full 965 * auxiliary search tree up to date would be too expensive, so we 966 * use a much simpler lookup table to do a binary search - 967 * bset_search_write_set(). 968 * * Or we use the auxiliary search tree we constructed earlier - 969 * bset_search_tree() 970 */ 971 972 if (unlikely(!t->size)) { 973 i.l = t->data->start; 974 i.r = bset_bkey_last(t->data); 975 } else if (bset_written(b, t)) { 976 /* 977 * Each node in the auxiliary search tree covers a certain range 978 * of bits, and keys above and below the set it covers might 979 * differ outside those bits - so we have to special case the 980 * start and end - handle that here: 981 */ 982 983 if (unlikely(bkey_cmp(search, &t->end) >= 0)) 984 return bset_bkey_last(t->data); 985 986 if (unlikely(bkey_cmp(search, t->data->start) < 0)) 987 return t->data->start; 988 989 i = bset_search_tree(t, search); 990 } else { 991 BUG_ON(!b->nsets && 992 t->size < bkey_to_cacheline(t, bset_bkey_last(t->data))); 993 994 i = bset_search_write_set(t, search); 995 } 996 997 if (btree_keys_expensive_checks(b)) { 998 BUG_ON(bset_written(b, t) && 999 i.l != t->data->start && 1000 bkey_cmp(tree_to_prev_bkey(t, 1001 inorder_to_tree(bkey_to_cacheline(t, i.l), t)), 1002 search) > 0); 1003 1004 BUG_ON(i.r != bset_bkey_last(t->data) && 1005 bkey_cmp(i.r, search) <= 0); 1006 } 1007 1008 while (likely(i.l != i.r) && 1009 bkey_cmp(i.l, search) <= 0) 1010 i.l = bkey_next(i.l); 1011 1012 return i.l; 1013} 1014EXPORT_SYMBOL(__bch_bset_search); 1015 1016/* Btree iterator */ 1017 1018typedef bool (btree_iter_cmp_fn)(struct btree_iter_set, 1019 struct btree_iter_set); 1020 1021static inline bool btree_iter_cmp(struct btree_iter_set l, 1022 struct btree_iter_set r) 1023{ 1024 return bkey_cmp(l.k, r.k) > 0; 1025} 1026 1027static inline bool btree_iter_end(struct btree_iter *iter) 1028{ 1029 return !iter->used; 1030} 1031 1032void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k, 1033 struct bkey *end) 1034{ 1035 if (k != end) 1036 BUG_ON(!heap_add(iter, 1037 ((struct btree_iter_set) { k, end }), 1038 btree_iter_cmp)); 1039} 1040 1041static struct bkey *__bch_btree_iter_init(struct btree_keys *b, 1042 struct btree_iter *iter, 1043 struct bkey *search, 1044 struct bset_tree *start) 1045{ 1046 struct bkey *ret = NULL; 1047 iter->size = ARRAY_SIZE(iter->data); 1048 iter->used = 0; 1049 1050#ifdef CONFIG_BCACHE_DEBUG 1051 iter->b = b; 1052#endif 1053 1054 for (; start <= bset_tree_last(b); start++) { 1055 ret = bch_bset_search(b, start, search); 1056 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data)); 1057 } 1058 1059 return ret; 1060} 1061 1062struct bkey *bch_btree_iter_init(struct btree_keys *b, 1063 struct btree_iter *iter, 1064 struct bkey *search) 1065{ 1066 return __bch_btree_iter_init(b, iter, search, b->set); 1067} 1068EXPORT_SYMBOL(bch_btree_iter_init); 1069 1070static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter, 1071 btree_iter_cmp_fn *cmp) 1072{ 1073 struct btree_iter_set unused; 1074 struct bkey *ret = NULL; 1075 1076 if (!btree_iter_end(iter)) { 1077 bch_btree_iter_next_check(iter); 1078 1079 ret = iter->data->k; 1080 iter->data->k = bkey_next(iter->data->k); 1081 1082 if (iter->data->k > iter->data->end) { 1083 WARN_ONCE(1, "bset was corrupt!\n"); 1084 iter->data->k = iter->data->end; 1085 } 1086 1087 if (iter->data->k == iter->data->end) 1088 heap_pop(iter, unused, cmp); 1089 else 1090 heap_sift(iter, 0, cmp); 1091 } 1092 1093 return ret; 1094} 1095 1096struct bkey *bch_btree_iter_next(struct btree_iter *iter) 1097{ 1098 return __bch_btree_iter_next(iter, btree_iter_cmp); 1099 1100} 1101EXPORT_SYMBOL(bch_btree_iter_next); 1102 1103struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter, 1104 struct btree_keys *b, ptr_filter_fn fn) 1105{ 1106 struct bkey *ret; 1107 1108 do { 1109 ret = bch_btree_iter_next(iter); 1110 } while (ret && fn(b, ret)); 1111 1112 return ret; 1113} 1114 1115/* Mergesort */ 1116 1117void bch_bset_sort_state_free(struct bset_sort_state *state) 1118{ 1119 if (state->pool) 1120 mempool_destroy(state->pool); 1121} 1122 1123int bch_bset_sort_state_init(struct bset_sort_state *state, unsigned page_order) 1124{ 1125 spin_lock_init(&state->time.lock); 1126 1127 state->page_order = page_order; 1128 state->crit_factor = int_sqrt(1 << page_order); 1129 1130 state->pool = mempool_create_page_pool(1, page_order); 1131 if (!state->pool) 1132 return -ENOMEM; 1133 1134 return 0; 1135} 1136EXPORT_SYMBOL(bch_bset_sort_state_init); 1137 1138static void btree_mergesort(struct btree_keys *b, struct bset *out, 1139 struct btree_iter *iter, 1140 bool fixup, bool remove_stale) 1141{ 1142 int i; 1143 struct bkey *k, *last = NULL; 1144 BKEY_PADDED(k) tmp; 1145 bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale 1146 ? bch_ptr_bad 1147 : bch_ptr_invalid; 1148 1149 /* Heapify the iterator, using our comparison function */ 1150 for (i = iter->used / 2 - 1; i >= 0; --i) 1151 heap_sift(iter, i, b->ops->sort_cmp); 1152 1153 while (!btree_iter_end(iter)) { 1154 if (b->ops->sort_fixup && fixup) 1155 k = b->ops->sort_fixup(iter, &tmp.k); 1156 else 1157 k = NULL; 1158 1159 if (!k) 1160 k = __bch_btree_iter_next(iter, b->ops->sort_cmp); 1161 1162 if (bad(b, k)) 1163 continue; 1164 1165 if (!last) { 1166 last = out->start; 1167 bkey_copy(last, k); 1168 } else if (!bch_bkey_try_merge(b, last, k)) { 1169 last = bkey_next(last); 1170 bkey_copy(last, k); 1171 } 1172 } 1173 1174 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0; 1175 1176 pr_debug("sorted %i keys", out->keys); 1177} 1178 1179static void __btree_sort(struct btree_keys *b, struct btree_iter *iter, 1180 unsigned start, unsigned order, bool fixup, 1181 struct bset_sort_state *state) 1182{ 1183 uint64_t start_time; 1184 bool used_mempool = false; 1185 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, 1186 order); 1187 if (!out) { 1188 struct page *outp; 1189 1190 BUG_ON(order > state->page_order); 1191 1192 outp = mempool_alloc(state->pool, GFP_NOIO); 1193 out = page_address(outp); 1194 used_mempool = true; 1195 order = state->page_order; 1196 } 1197 1198 start_time = local_clock(); 1199 1200 btree_mergesort(b, out, iter, fixup, false); 1201 b->nsets = start; 1202 1203 if (!start && order == b->page_order) { 1204 /* 1205 * Our temporary buffer is the same size as the btree node's 1206 * buffer, we can just swap buffers instead of doing a big 1207 * memcpy() 1208 */ 1209 1210 out->magic = b->set->data->magic; 1211 out->seq = b->set->data->seq; 1212 out->version = b->set->data->version; 1213 swap(out, b->set->data); 1214 } else { 1215 b->set[start].data->keys = out->keys; 1216 memcpy(b->set[start].data->start, out->start, 1217 (void *) bset_bkey_last(out) - (void *) out->start); 1218 } 1219 1220 if (used_mempool) 1221 mempool_free(virt_to_page(out), state->pool); 1222 else 1223 free_pages((unsigned long) out, order); 1224 1225 bch_bset_build_written_tree(b); 1226 1227 if (!start) 1228 bch_time_stats_update(&state->time, start_time); 1229} 1230 1231void bch_btree_sort_partial(struct btree_keys *b, unsigned start, 1232 struct bset_sort_state *state) 1233{ 1234 size_t order = b->page_order, keys = 0; 1235 struct btree_iter iter; 1236 int oldsize = bch_count_data(b); 1237 1238 __bch_btree_iter_init(b, &iter, NULL, &b->set[start]); 1239 1240 if (start) { 1241 unsigned i; 1242 1243 for (i = start; i <= b->nsets; i++) 1244 keys += b->set[i].data->keys; 1245 1246 order = get_order(__set_bytes(b->set->data, keys)); 1247 } 1248 1249 __btree_sort(b, &iter, start, order, false, state); 1250 1251 EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize); 1252} 1253EXPORT_SYMBOL(bch_btree_sort_partial); 1254 1255void bch_btree_sort_and_fix_extents(struct btree_keys *b, 1256 struct btree_iter *iter, 1257 struct bset_sort_state *state) 1258{ 1259 __btree_sort(b, iter, 0, b->page_order, true, state); 1260} 1261 1262void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new, 1263 struct bset_sort_state *state) 1264{ 1265 uint64_t start_time = local_clock(); 1266 1267 struct btree_iter iter; 1268 bch_btree_iter_init(b, &iter, NULL); 1269 1270 btree_mergesort(b, new->set->data, &iter, false, true); 1271 1272 bch_time_stats_update(&state->time, start_time); 1273 1274 new->set->size = 0; // XXX: why? 1275} 1276 1277#define SORT_CRIT (4096 / sizeof(uint64_t)) 1278 1279void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state) 1280{ 1281 unsigned crit = SORT_CRIT; 1282 int i; 1283 1284 /* Don't sort if nothing to do */ 1285 if (!b->nsets) 1286 goto out; 1287 1288 for (i = b->nsets - 1; i >= 0; --i) { 1289 crit *= state->crit_factor; 1290 1291 if (b->set[i].data->keys < crit) { 1292 bch_btree_sort_partial(b, i, state); 1293 return; 1294 } 1295 } 1296 1297 /* Sort if we'd overflow */ 1298 if (b->nsets + 1 == MAX_BSETS) { 1299 bch_btree_sort(b, state); 1300 return; 1301 } 1302 1303out: 1304 bch_bset_build_written_tree(b); 1305} 1306EXPORT_SYMBOL(bch_btree_sort_lazy); 1307 1308void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats) 1309{ 1310 unsigned i; 1311 1312 for (i = 0; i <= b->nsets; i++) { 1313 struct bset_tree *t = &b->set[i]; 1314 size_t bytes = t->data->keys * sizeof(uint64_t); 1315 size_t j; 1316 1317 if (bset_written(b, t)) { 1318 stats->sets_written++; 1319 stats->bytes_written += bytes; 1320 1321 stats->floats += t->size - 1; 1322 1323 for (j = 1; j < t->size; j++) 1324 if (t->tree[j].exponent == 127) 1325 stats->failed++; 1326 } else { 1327 stats->sets_unwritten++; 1328 stats->bytes_unwritten += bytes; 1329 } 1330 } 1331} 1332