1 2 #ifdef __KERNEL__ 3 # include <linux/string.h> 4 # include <linux/slab.h> 5 # include <linux/bug.h> 6 # include <linux/kernel.h> 7 # ifndef dprintk 8 # define dprintk(args...) 9 # endif 10 #else 11 # include <string.h> 12 # include <stdio.h> 13 # include <stdlib.h> 14 # include <assert.h> 15 # define BUG_ON(x) assert(!(x)) 16 # define dprintk(args...) /* printf(args) */ 17 # define kmalloc(x, f) malloc(x) 18 # define kfree(x) free(x) 19 #endif 20 21 #include <linux/crush/crush.h> 22 #include <linux/crush/hash.h> 23 #include <linux/crush/mapper.h> 24 25 /* 26 * Implement the core CRUSH mapping algorithm. 27 */ 28 29 /** 30 * crush_find_rule - find a crush_rule id for a given ruleset, type, and size. 31 * @map: the crush_map 32 * @ruleset: the storage ruleset id (user defined) 33 * @type: storage ruleset type (user defined) 34 * @size: output set size 35 */ 36 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size) 37 { 38 __u32 i; 39 40 for (i = 0; i < map->max_rules; i++) { 41 if (map->rules[i] && 42 map->rules[i]->mask.ruleset == ruleset && 43 map->rules[i]->mask.type == type && 44 map->rules[i]->mask.min_size <= size && 45 map->rules[i]->mask.max_size >= size) 46 return i; 47 } 48 return -1; 49 } 50 51 52 /* 53 * bucket choose methods 54 * 55 * For each bucket algorithm, we have a "choose" method that, given a 56 * crush input @x and replica position (usually, position in output set) @r, 57 * will produce an item in the bucket. 58 */ 59 60 /* 61 * Choose based on a random permutation of the bucket. 62 * 63 * We used to use some prime number arithmetic to do this, but it 64 * wasn't very random, and had some other bad behaviors. Instead, we 65 * calculate an actual random permutation of the bucket members. 66 * Since this is expensive, we optimize for the r=0 case, which 67 * captures the vast majority of calls. 68 */ 69 static int bucket_perm_choose(struct crush_bucket *bucket, 70 int x, int r) 71 { 72 unsigned int pr = r % bucket->size; 73 unsigned int i, s; 74 75 /* start a new permutation if @x has changed */ 76 if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) { 77 dprintk("bucket %d new x=%d\n", bucket->id, x); 78 bucket->perm_x = x; 79 80 /* optimize common r=0 case */ 81 if (pr == 0) { 82 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) % 83 bucket->size; 84 bucket->perm[0] = s; 85 bucket->perm_n = 0xffff; /* magic value, see below */ 86 goto out; 87 } 88 89 for (i = 0; i < bucket->size; i++) 90 bucket->perm[i] = i; 91 bucket->perm_n = 0; 92 } else if (bucket->perm_n == 0xffff) { 93 /* clean up after the r=0 case above */ 94 for (i = 1; i < bucket->size; i++) 95 bucket->perm[i] = i; 96 bucket->perm[bucket->perm[0]] = 0; 97 bucket->perm_n = 1; 98 } 99 100 /* calculate permutation up to pr */ 101 for (i = 0; i < bucket->perm_n; i++) 102 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]); 103 while (bucket->perm_n <= pr) { 104 unsigned int p = bucket->perm_n; 105 /* no point in swapping the final entry */ 106 if (p < bucket->size - 1) { 107 i = crush_hash32_3(bucket->hash, x, bucket->id, p) % 108 (bucket->size - p); 109 if (i) { 110 unsigned int t = bucket->perm[p + i]; 111 bucket->perm[p + i] = bucket->perm[p]; 112 bucket->perm[p] = t; 113 } 114 dprintk(" perm_choose swap %d with %d\n", p, p+i); 115 } 116 bucket->perm_n++; 117 } 118 for (i = 0; i < bucket->size; i++) 119 dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]); 120 121 s = bucket->perm[pr]; 122 out: 123 dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id, 124 bucket->size, x, r, pr, s); 125 return bucket->items[s]; 126 } 127 128 /* uniform */ 129 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket, 130 int x, int r) 131 { 132 return bucket_perm_choose(&bucket->h, x, r); 133 } 134 135 /* list */ 136 static int bucket_list_choose(struct crush_bucket_list *bucket, 137 int x, int r) 138 { 139 int i; 140 141 for (i = bucket->h.size-1; i >= 0; i--) { 142 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i], 143 r, bucket->h.id); 144 w &= 0xffff; 145 dprintk("list_choose i=%d x=%d r=%d item %d weight %x " 146 "sw %x rand %llx", 147 i, x, r, bucket->h.items[i], bucket->item_weights[i], 148 bucket->sum_weights[i], w); 149 w *= bucket->sum_weights[i]; 150 w = w >> 16; 151 /*dprintk(" scaled %llx\n", w);*/ 152 if (w < bucket->item_weights[i]) 153 return bucket->h.items[i]; 154 } 155 156 dprintk("bad list sums for bucket %d\n", bucket->h.id); 157 return bucket->h.items[0]; 158 } 159 160 161 /* (binary) tree */ 162 static int height(int n) 163 { 164 int h = 0; 165 while ((n & 1) == 0) { 166 h++; 167 n = n >> 1; 168 } 169 return h; 170 } 171 172 static int left(int x) 173 { 174 int h = height(x); 175 return x - (1 << (h-1)); 176 } 177 178 static int right(int x) 179 { 180 int h = height(x); 181 return x + (1 << (h-1)); 182 } 183 184 static int terminal(int x) 185 { 186 return x & 1; 187 } 188 189 static int bucket_tree_choose(struct crush_bucket_tree *bucket, 190 int x, int r) 191 { 192 int n; 193 __u32 w; 194 __u64 t; 195 196 /* start at root */ 197 n = bucket->num_nodes >> 1; 198 199 while (!terminal(n)) { 200 int l; 201 /* pick point in [0, w) */ 202 w = bucket->node_weights[n]; 203 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r, 204 bucket->h.id) * (__u64)w; 205 t = t >> 32; 206 207 /* descend to the left or right? */ 208 l = left(n); 209 if (t < bucket->node_weights[l]) 210 n = l; 211 else 212 n = right(n); 213 } 214 215 return bucket->h.items[n >> 1]; 216 } 217 218 219 /* straw */ 220 221 static int bucket_straw_choose(struct crush_bucket_straw *bucket, 222 int x, int r) 223 { 224 __u32 i; 225 int high = 0; 226 __u64 high_draw = 0; 227 __u64 draw; 228 229 for (i = 0; i < bucket->h.size; i++) { 230 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r); 231 draw &= 0xffff; 232 draw *= bucket->straws[i]; 233 if (i == 0 || draw > high_draw) { 234 high = i; 235 high_draw = draw; 236 } 237 } 238 return bucket->h.items[high]; 239 } 240 241 static int crush_bucket_choose(struct crush_bucket *in, int x, int r) 242 { 243 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r); 244 BUG_ON(in->size == 0); 245 switch (in->alg) { 246 case CRUSH_BUCKET_UNIFORM: 247 return bucket_uniform_choose((struct crush_bucket_uniform *)in, 248 x, r); 249 case CRUSH_BUCKET_LIST: 250 return bucket_list_choose((struct crush_bucket_list *)in, 251 x, r); 252 case CRUSH_BUCKET_TREE: 253 return bucket_tree_choose((struct crush_bucket_tree *)in, 254 x, r); 255 case CRUSH_BUCKET_STRAW: 256 return bucket_straw_choose((struct crush_bucket_straw *)in, 257 x, r); 258 default: 259 dprintk("unknown bucket %d alg %d\n", in->id, in->alg); 260 return in->items[0]; 261 } 262 } 263 264 /* 265 * true if device is marked "out" (failed, fully offloaded) 266 * of the cluster 267 */ 268 static int is_out(const struct crush_map *map, 269 const __u32 *weight, int weight_max, 270 int item, int x) 271 { 272 if (item >= weight_max) 273 return 1; 274 if (weight[item] >= 0x10000) 275 return 0; 276 if (weight[item] == 0) 277 return 1; 278 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff) 279 < weight[item]) 280 return 0; 281 return 1; 282 } 283 284 /** 285 * crush_choose_firstn - choose numrep distinct items of given type 286 * @map: the crush_map 287 * @bucket: the bucket we are choose an item from 288 * @x: crush input value 289 * @numrep: the number of items to choose 290 * @type: the type of item to choose 291 * @out: pointer to output vector 292 * @outpos: our position in that vector 293 * @tries: number of attempts to make 294 * @recurse_tries: number of attempts to have recursive chooseleaf make 295 * @local_retries: localized retries 296 * @local_fallback_retries: localized fallback retries 297 * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) 298 * @vary_r: pass r to recursive calls 299 * @out2: second output vector for leaf items (if @recurse_to_leaf) 300 * @parent_r: r value passed from the parent 301 */ 302 static int crush_choose_firstn(const struct crush_map *map, 303 struct crush_bucket *bucket, 304 const __u32 *weight, int weight_max, 305 int x, int numrep, int type, 306 int *out, int outpos, 307 unsigned int tries, 308 unsigned int recurse_tries, 309 unsigned int local_retries, 310 unsigned int local_fallback_retries, 311 int recurse_to_leaf, 312 unsigned int vary_r, 313 int *out2, 314 int parent_r) 315 { 316 int rep; 317 unsigned int ftotal, flocal; 318 int retry_descent, retry_bucket, skip_rep; 319 struct crush_bucket *in = bucket; 320 int r; 321 int i; 322 int item = 0; 323 int itemtype; 324 int collide, reject; 325 326 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d tries %d recurse_tries %d local_retries %d local_fallback_retries %d parent_r %d\n", 327 recurse_to_leaf ? "_LEAF" : "", 328 bucket->id, x, outpos, numrep, 329 tries, recurse_tries, local_retries, local_fallback_retries, 330 parent_r); 331 332 for (rep = outpos; rep < numrep; rep++) { 333 /* keep trying until we get a non-out, non-colliding item */ 334 ftotal = 0; 335 skip_rep = 0; 336 do { 337 retry_descent = 0; 338 in = bucket; /* initial bucket */ 339 340 /* choose through intervening buckets */ 341 flocal = 0; 342 do { 343 collide = 0; 344 retry_bucket = 0; 345 r = rep + parent_r; 346 /* r' = r + f_total */ 347 r += ftotal; 348 349 /* bucket choose */ 350 if (in->size == 0) { 351 reject = 1; 352 goto reject; 353 } 354 if (local_fallback_retries > 0 && 355 flocal >= (in->size>>1) && 356 flocal > local_fallback_retries) 357 item = bucket_perm_choose(in, x, r); 358 else 359 item = crush_bucket_choose(in, x, r); 360 if (item >= map->max_devices) { 361 dprintk(" bad item %d\n", item); 362 skip_rep = 1; 363 break; 364 } 365 366 /* desired type? */ 367 if (item < 0) 368 itemtype = map->buckets[-1-item]->type; 369 else 370 itemtype = 0; 371 dprintk(" item %d type %d\n", item, itemtype); 372 373 /* keep going? */ 374 if (itemtype != type) { 375 if (item >= 0 || 376 (-1-item) >= map->max_buckets) { 377 dprintk(" bad item type %d\n", type); 378 skip_rep = 1; 379 break; 380 } 381 in = map->buckets[-1-item]; 382 retry_bucket = 1; 383 continue; 384 } 385 386 /* collision? */ 387 for (i = 0; i < outpos; i++) { 388 if (out[i] == item) { 389 collide = 1; 390 break; 391 } 392 } 393 394 reject = 0; 395 if (!collide && recurse_to_leaf) { 396 if (item < 0) { 397 int sub_r; 398 if (vary_r) 399 sub_r = r >> (vary_r-1); 400 else 401 sub_r = 0; 402 if (crush_choose_firstn(map, 403 map->buckets[-1-item], 404 weight, weight_max, 405 x, outpos+1, 0, 406 out2, outpos, 407 recurse_tries, 0, 408 local_retries, 409 local_fallback_retries, 410 0, 411 vary_r, 412 NULL, 413 sub_r) <= outpos) 414 /* didn't get leaf */ 415 reject = 1; 416 } else { 417 /* we already have a leaf! */ 418 out2[outpos] = item; 419 } 420 } 421 422 if (!reject) { 423 /* out? */ 424 if (itemtype == 0) 425 reject = is_out(map, weight, 426 weight_max, 427 item, x); 428 else 429 reject = 0; 430 } 431 432 reject: 433 if (reject || collide) { 434 ftotal++; 435 flocal++; 436 437 if (collide && flocal <= local_retries) 438 /* retry locally a few times */ 439 retry_bucket = 1; 440 else if (local_fallback_retries > 0 && 441 flocal <= in->size + local_fallback_retries) 442 /* exhaustive bucket search */ 443 retry_bucket = 1; 444 else if (ftotal < tries) 445 /* then retry descent */ 446 retry_descent = 1; 447 else 448 /* else give up */ 449 skip_rep = 1; 450 dprintk(" reject %d collide %d " 451 "ftotal %u flocal %u\n", 452 reject, collide, ftotal, 453 flocal); 454 } 455 } while (retry_bucket); 456 } while (retry_descent); 457 458 if (skip_rep) { 459 dprintk("skip rep\n"); 460 continue; 461 } 462 463 dprintk("CHOOSE got %d\n", item); 464 out[outpos] = item; 465 outpos++; 466 } 467 468 dprintk("CHOOSE returns %d\n", outpos); 469 return outpos; 470 } 471 472 473 /** 474 * crush_choose_indep: alternative breadth-first positionally stable mapping 475 * 476 */ 477 static void crush_choose_indep(const struct crush_map *map, 478 struct crush_bucket *bucket, 479 const __u32 *weight, int weight_max, 480 int x, int left, int numrep, int type, 481 int *out, int outpos, 482 unsigned int tries, 483 unsigned int recurse_tries, 484 int recurse_to_leaf, 485 int *out2, 486 int parent_r) 487 { 488 struct crush_bucket *in = bucket; 489 int endpos = outpos + left; 490 int rep; 491 unsigned int ftotal; 492 int r; 493 int i; 494 int item = 0; 495 int itemtype; 496 int collide; 497 498 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", 499 bucket->id, x, outpos, numrep); 500 501 /* initially my result is undefined */ 502 for (rep = outpos; rep < endpos; rep++) { 503 out[rep] = CRUSH_ITEM_UNDEF; 504 if (out2) 505 out2[rep] = CRUSH_ITEM_UNDEF; 506 } 507 508 for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { 509 for (rep = outpos; rep < endpos; rep++) { 510 if (out[rep] != CRUSH_ITEM_UNDEF) 511 continue; 512 513 in = bucket; /* initial bucket */ 514 515 /* choose through intervening buckets */ 516 for (;;) { 517 /* note: we base the choice on the position 518 * even in the nested call. that means that 519 * if the first layer chooses the same bucket 520 * in a different position, we will tend to 521 * choose a different item in that bucket. 522 * this will involve more devices in data 523 * movement and tend to distribute the load. 524 */ 525 r = rep + parent_r; 526 527 /* be careful */ 528 if (in->alg == CRUSH_BUCKET_UNIFORM && 529 in->size % numrep == 0) 530 /* r'=r+(n+1)*f_total */ 531 r += (numrep+1) * ftotal; 532 else 533 /* r' = r + n*f_total */ 534 r += numrep * ftotal; 535 536 /* bucket choose */ 537 if (in->size == 0) { 538 dprintk(" empty bucket\n"); 539 break; 540 } 541 542 item = crush_bucket_choose(in, x, r); 543 if (item >= map->max_devices) { 544 dprintk(" bad item %d\n", item); 545 out[rep] = CRUSH_ITEM_NONE; 546 if (out2) 547 out2[rep] = CRUSH_ITEM_NONE; 548 left--; 549 break; 550 } 551 552 /* desired type? */ 553 if (item < 0) 554 itemtype = map->buckets[-1-item]->type; 555 else 556 itemtype = 0; 557 dprintk(" item %d type %d\n", item, itemtype); 558 559 /* keep going? */ 560 if (itemtype != type) { 561 if (item >= 0 || 562 (-1-item) >= map->max_buckets) { 563 dprintk(" bad item type %d\n", type); 564 out[rep] = CRUSH_ITEM_NONE; 565 if (out2) 566 out2[rep] = 567 CRUSH_ITEM_NONE; 568 left--; 569 break; 570 } 571 in = map->buckets[-1-item]; 572 continue; 573 } 574 575 /* collision? */ 576 collide = 0; 577 for (i = outpos; i < endpos; i++) { 578 if (out[i] == item) { 579 collide = 1; 580 break; 581 } 582 } 583 if (collide) 584 break; 585 586 if (recurse_to_leaf) { 587 if (item < 0) { 588 crush_choose_indep(map, 589 map->buckets[-1-item], 590 weight, weight_max, 591 x, 1, numrep, 0, 592 out2, rep, 593 recurse_tries, 0, 594 0, NULL, r); 595 if (out2[rep] == CRUSH_ITEM_NONE) { 596 /* placed nothing; no leaf */ 597 break; 598 } 599 } else { 600 /* we already have a leaf! */ 601 out2[rep] = item; 602 } 603 } 604 605 /* out? */ 606 if (itemtype == 0 && 607 is_out(map, weight, weight_max, item, x)) 608 break; 609 610 /* yay! */ 611 out[rep] = item; 612 left--; 613 break; 614 } 615 } 616 } 617 for (rep = outpos; rep < endpos; rep++) { 618 if (out[rep] == CRUSH_ITEM_UNDEF) { 619 out[rep] = CRUSH_ITEM_NONE; 620 } 621 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { 622 out2[rep] = CRUSH_ITEM_NONE; 623 } 624 } 625 } 626 627 /** 628 * crush_do_rule - calculate a mapping with the given input and rule 629 * @map: the crush_map 630 * @ruleno: the rule id 631 * @x: hash input 632 * @result: pointer to result vector 633 * @result_max: maximum result size 634 * @weight: weight vector (for map leaves) 635 * @weight_max: size of weight vector 636 * @scratch: scratch vector for private use; must be >= 3 * result_max 637 */ 638 int crush_do_rule(const struct crush_map *map, 639 int ruleno, int x, int *result, int result_max, 640 const __u32 *weight, int weight_max, 641 int *scratch) 642 { 643 int result_len; 644 int *a = scratch; 645 int *b = scratch + result_max; 646 int *c = scratch + result_max*2; 647 int recurse_to_leaf; 648 int *w; 649 int wsize = 0; 650 int *o; 651 int osize; 652 int *tmp; 653 struct crush_rule *rule; 654 __u32 step; 655 int i, j; 656 int numrep; 657 /* 658 * the original choose_total_tries value was off by one (it 659 * counted "retries" and not "tries"). add one. 660 */ 661 int choose_tries = map->choose_total_tries + 1; 662 int choose_leaf_tries = 0; 663 /* 664 * the local tries values were counted as "retries", though, 665 * and need no adjustment 666 */ 667 int choose_local_retries = map->choose_local_tries; 668 int choose_local_fallback_retries = map->choose_local_fallback_tries; 669 670 int vary_r = map->chooseleaf_vary_r; 671 672 if ((__u32)ruleno >= map->max_rules) { 673 dprintk(" bad ruleno %d\n", ruleno); 674 return 0; 675 } 676 677 rule = map->rules[ruleno]; 678 result_len = 0; 679 w = a; 680 o = b; 681 682 for (step = 0; step < rule->len; step++) { 683 int firstn = 0; 684 struct crush_rule_step *curstep = &rule->steps[step]; 685 686 switch (curstep->op) { 687 case CRUSH_RULE_TAKE: 688 w[0] = curstep->arg1; 689 wsize = 1; 690 break; 691 692 case CRUSH_RULE_SET_CHOOSE_TRIES: 693 if (curstep->arg1 > 0) 694 choose_tries = curstep->arg1; 695 break; 696 697 case CRUSH_RULE_SET_CHOOSELEAF_TRIES: 698 if (curstep->arg1 > 0) 699 choose_leaf_tries = curstep->arg1; 700 break; 701 702 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: 703 if (curstep->arg1 >= 0) 704 choose_local_retries = curstep->arg1; 705 break; 706 707 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: 708 if (curstep->arg1 >= 0) 709 choose_local_fallback_retries = curstep->arg1; 710 break; 711 712 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R: 713 if (curstep->arg1 >= 0) 714 vary_r = curstep->arg1; 715 break; 716 717 case CRUSH_RULE_CHOOSELEAF_FIRSTN: 718 case CRUSH_RULE_CHOOSE_FIRSTN: 719 firstn = 1; 720 /* fall through */ 721 case CRUSH_RULE_CHOOSELEAF_INDEP: 722 case CRUSH_RULE_CHOOSE_INDEP: 723 if (wsize == 0) 724 break; 725 726 recurse_to_leaf = 727 curstep->op == 728 CRUSH_RULE_CHOOSELEAF_FIRSTN || 729 curstep->op == 730 CRUSH_RULE_CHOOSELEAF_INDEP; 731 732 /* reset output */ 733 osize = 0; 734 735 for (i = 0; i < wsize; i++) { 736 /* 737 * see CRUSH_N, CRUSH_N_MINUS macros. 738 * basically, numrep <= 0 means relative to 739 * the provided result_max 740 */ 741 numrep = curstep->arg1; 742 if (numrep <= 0) { 743 numrep += result_max; 744 if (numrep <= 0) 745 continue; 746 } 747 j = 0; 748 if (firstn) { 749 int recurse_tries; 750 if (choose_leaf_tries) 751 recurse_tries = 752 choose_leaf_tries; 753 else if (map->chooseleaf_descend_once) 754 recurse_tries = 1; 755 else 756 recurse_tries = choose_tries; 757 osize += crush_choose_firstn( 758 map, 759 map->buckets[-1-w[i]], 760 weight, weight_max, 761 x, numrep, 762 curstep->arg2, 763 o+osize, j, 764 choose_tries, 765 recurse_tries, 766 choose_local_retries, 767 choose_local_fallback_retries, 768 recurse_to_leaf, 769 vary_r, 770 c+osize, 771 0); 772 } else { 773 crush_choose_indep( 774 map, 775 map->buckets[-1-w[i]], 776 weight, weight_max, 777 x, numrep, numrep, 778 curstep->arg2, 779 o+osize, j, 780 choose_tries, 781 choose_leaf_tries ? 782 choose_leaf_tries : 1, 783 recurse_to_leaf, 784 c+osize, 785 0); 786 osize += numrep; 787 } 788 } 789 790 if (recurse_to_leaf) 791 /* copy final _leaf_ values to output set */ 792 memcpy(o, c, osize*sizeof(*o)); 793 794 /* swap o and w arrays */ 795 tmp = o; 796 o = w; 797 w = tmp; 798 wsize = osize; 799 break; 800 801 802 case CRUSH_RULE_EMIT: 803 for (i = 0; i < wsize && result_len < result_max; i++) { 804 result[result_len] = w[i]; 805 result_len++; 806 } 807 wsize = 0; 808 break; 809 810 default: 811 dprintk(" unknown op %d at step %d\n", 812 curstep->op, step); 813 break; 814 } 815 } 816 return result_len; 817 } 818 819 820