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_tries: localized retries 296 * @local_fallback_tries: localized fallback retries 297 * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) 298 * @out2: second output vector for leaf items (if @recurse_to_leaf) 299 */ 300 static int crush_choose_firstn(const struct crush_map *map, 301 struct crush_bucket *bucket, 302 const __u32 *weight, int weight_max, 303 int x, int numrep, int type, 304 int *out, int outpos, 305 unsigned int tries, 306 unsigned int recurse_tries, 307 unsigned int local_tries, 308 unsigned int local_fallback_tries, 309 int recurse_to_leaf, 310 int *out2) 311 { 312 int rep; 313 unsigned int ftotal, flocal; 314 int retry_descent, retry_bucket, skip_rep; 315 struct crush_bucket *in = bucket; 316 int r; 317 int i; 318 int item = 0; 319 int itemtype; 320 int collide, reject; 321 322 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", 323 bucket->id, x, outpos, numrep); 324 325 for (rep = outpos; rep < numrep; rep++) { 326 /* keep trying until we get a non-out, non-colliding item */ 327 ftotal = 0; 328 skip_rep = 0; 329 do { 330 retry_descent = 0; 331 in = bucket; /* initial bucket */ 332 333 /* choose through intervening buckets */ 334 flocal = 0; 335 do { 336 collide = 0; 337 retry_bucket = 0; 338 r = rep; 339 /* r' = r + f_total */ 340 r += ftotal; 341 342 /* bucket choose */ 343 if (in->size == 0) { 344 reject = 1; 345 goto reject; 346 } 347 if (local_fallback_tries > 0 && 348 flocal >= (in->size>>1) && 349 flocal > local_fallback_tries) 350 item = bucket_perm_choose(in, x, r); 351 else 352 item = crush_bucket_choose(in, x, r); 353 if (item >= map->max_devices) { 354 dprintk(" bad item %d\n", item); 355 skip_rep = 1; 356 break; 357 } 358 359 /* desired type? */ 360 if (item < 0) 361 itemtype = map->buckets[-1-item]->type; 362 else 363 itemtype = 0; 364 dprintk(" item %d type %d\n", item, itemtype); 365 366 /* keep going? */ 367 if (itemtype != type) { 368 if (item >= 0 || 369 (-1-item) >= map->max_buckets) { 370 dprintk(" bad item type %d\n", type); 371 skip_rep = 1; 372 break; 373 } 374 in = map->buckets[-1-item]; 375 retry_bucket = 1; 376 continue; 377 } 378 379 /* collision? */ 380 for (i = 0; i < outpos; i++) { 381 if (out[i] == item) { 382 collide = 1; 383 break; 384 } 385 } 386 387 reject = 0; 388 if (!collide && recurse_to_leaf) { 389 if (item < 0) { 390 if (crush_choose_firstn(map, 391 map->buckets[-1-item], 392 weight, weight_max, 393 x, outpos+1, 0, 394 out2, outpos, 395 recurse_tries, 0, 396 local_tries, 397 local_fallback_tries, 398 0, 399 NULL) <= outpos) 400 /* didn't get leaf */ 401 reject = 1; 402 } else { 403 /* we already have a leaf! */ 404 out2[outpos] = item; 405 } 406 } 407 408 if (!reject) { 409 /* out? */ 410 if (itemtype == 0) 411 reject = is_out(map, weight, 412 weight_max, 413 item, x); 414 else 415 reject = 0; 416 } 417 418 reject: 419 if (reject || collide) { 420 ftotal++; 421 flocal++; 422 423 if (collide && flocal <= local_tries) 424 /* retry locally a few times */ 425 retry_bucket = 1; 426 else if (local_fallback_tries > 0 && 427 flocal <= in->size + local_fallback_tries) 428 /* exhaustive bucket search */ 429 retry_bucket = 1; 430 else if (ftotal <= tries) 431 /* then retry descent */ 432 retry_descent = 1; 433 else 434 /* else give up */ 435 skip_rep = 1; 436 dprintk(" reject %d collide %d " 437 "ftotal %u flocal %u\n", 438 reject, collide, ftotal, 439 flocal); 440 } 441 } while (retry_bucket); 442 } while (retry_descent); 443 444 if (skip_rep) { 445 dprintk("skip rep\n"); 446 continue; 447 } 448 449 dprintk("CHOOSE got %d\n", item); 450 out[outpos] = item; 451 outpos++; 452 } 453 454 dprintk("CHOOSE returns %d\n", outpos); 455 return outpos; 456 } 457 458 459 /** 460 * crush_choose_indep: alternative breadth-first positionally stable mapping 461 * 462 */ 463 static void crush_choose_indep(const struct crush_map *map, 464 struct crush_bucket *bucket, 465 const __u32 *weight, int weight_max, 466 int x, int left, int numrep, int type, 467 int *out, int outpos, 468 unsigned int tries, 469 unsigned int recurse_tries, 470 int recurse_to_leaf, 471 int *out2, 472 int parent_r) 473 { 474 struct crush_bucket *in = bucket; 475 int endpos = outpos + left; 476 int rep; 477 unsigned int ftotal; 478 int r; 479 int i; 480 int item = 0; 481 int itemtype; 482 int collide; 483 484 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", 485 bucket->id, x, outpos, numrep); 486 487 /* initially my result is undefined */ 488 for (rep = outpos; rep < endpos; rep++) { 489 out[rep] = CRUSH_ITEM_UNDEF; 490 if (out2) 491 out2[rep] = CRUSH_ITEM_UNDEF; 492 } 493 494 for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { 495 for (rep = outpos; rep < endpos; rep++) { 496 if (out[rep] != CRUSH_ITEM_UNDEF) 497 continue; 498 499 in = bucket; /* initial bucket */ 500 501 /* choose through intervening buckets */ 502 for (;;) { 503 /* note: we base the choice on the position 504 * even in the nested call. that means that 505 * if the first layer chooses the same bucket 506 * in a different position, we will tend to 507 * choose a different item in that bucket. 508 * this will involve more devices in data 509 * movement and tend to distribute the load. 510 */ 511 r = rep + parent_r; 512 513 /* be careful */ 514 if (in->alg == CRUSH_BUCKET_UNIFORM && 515 in->size % numrep == 0) 516 /* r'=r+(n+1)*f_total */ 517 r += (numrep+1) * ftotal; 518 else 519 /* r' = r + n*f_total */ 520 r += numrep * ftotal; 521 522 /* bucket choose */ 523 if (in->size == 0) { 524 dprintk(" empty bucket\n"); 525 break; 526 } 527 528 item = crush_bucket_choose(in, x, r); 529 if (item >= map->max_devices) { 530 dprintk(" bad item %d\n", item); 531 out[rep] = CRUSH_ITEM_NONE; 532 if (out2) 533 out2[rep] = CRUSH_ITEM_NONE; 534 left--; 535 break; 536 } 537 538 /* desired type? */ 539 if (item < 0) 540 itemtype = map->buckets[-1-item]->type; 541 else 542 itemtype = 0; 543 dprintk(" item %d type %d\n", item, itemtype); 544 545 /* keep going? */ 546 if (itemtype != type) { 547 if (item >= 0 || 548 (-1-item) >= map->max_buckets) { 549 dprintk(" bad item type %d\n", type); 550 out[rep] = CRUSH_ITEM_NONE; 551 if (out2) 552 out2[rep] = 553 CRUSH_ITEM_NONE; 554 left--; 555 break; 556 } 557 in = map->buckets[-1-item]; 558 continue; 559 } 560 561 /* collision? */ 562 collide = 0; 563 for (i = outpos; i < endpos; i++) { 564 if (out[i] == item) { 565 collide = 1; 566 break; 567 } 568 } 569 if (collide) 570 break; 571 572 if (recurse_to_leaf) { 573 if (item < 0) { 574 crush_choose_indep(map, 575 map->buckets[-1-item], 576 weight, weight_max, 577 x, 1, numrep, 0, 578 out2, rep, 579 recurse_tries, 0, 580 0, NULL, r); 581 if (out2[rep] == CRUSH_ITEM_NONE) { 582 /* placed nothing; no leaf */ 583 break; 584 } 585 } else { 586 /* we already have a leaf! */ 587 out2[rep] = item; 588 } 589 } 590 591 /* out? */ 592 if (itemtype == 0 && 593 is_out(map, weight, weight_max, item, x)) 594 break; 595 596 /* yay! */ 597 out[rep] = item; 598 left--; 599 break; 600 } 601 } 602 } 603 for (rep = outpos; rep < endpos; rep++) { 604 if (out[rep] == CRUSH_ITEM_UNDEF) { 605 out[rep] = CRUSH_ITEM_NONE; 606 } 607 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { 608 out2[rep] = CRUSH_ITEM_NONE; 609 } 610 } 611 } 612 613 /** 614 * crush_do_rule - calculate a mapping with the given input and rule 615 * @map: the crush_map 616 * @ruleno: the rule id 617 * @x: hash input 618 * @result: pointer to result vector 619 * @result_max: maximum result size 620 * @weight: weight vector (for map leaves) 621 * @weight_max: size of weight vector 622 * @scratch: scratch vector for private use; must be >= 3 * result_max 623 */ 624 int crush_do_rule(const struct crush_map *map, 625 int ruleno, int x, int *result, int result_max, 626 const __u32 *weight, int weight_max, 627 int *scratch) 628 { 629 int result_len; 630 int *a = scratch; 631 int *b = scratch + result_max; 632 int *c = scratch + result_max*2; 633 int recurse_to_leaf; 634 int *w; 635 int wsize = 0; 636 int *o; 637 int osize; 638 int *tmp; 639 struct crush_rule *rule; 640 __u32 step; 641 int i, j; 642 int numrep; 643 int choose_tries = map->choose_total_tries; 644 int choose_local_tries = map->choose_local_tries; 645 int choose_local_fallback_tries = map->choose_local_fallback_tries; 646 int choose_leaf_tries = 0; 647 648 if ((__u32)ruleno >= map->max_rules) { 649 dprintk(" bad ruleno %d\n", ruleno); 650 return 0; 651 } 652 653 rule = map->rules[ruleno]; 654 result_len = 0; 655 w = a; 656 o = b; 657 658 for (step = 0; step < rule->len; step++) { 659 int firstn = 0; 660 struct crush_rule_step *curstep = &rule->steps[step]; 661 662 switch (curstep->op) { 663 case CRUSH_RULE_TAKE: 664 w[0] = curstep->arg1; 665 wsize = 1; 666 break; 667 668 case CRUSH_RULE_SET_CHOOSE_TRIES: 669 if (curstep->arg1 > 0) 670 choose_tries = curstep->arg1; 671 break; 672 673 case CRUSH_RULE_SET_CHOOSELEAF_TRIES: 674 if (curstep->arg1 > 0) 675 choose_leaf_tries = curstep->arg1; 676 break; 677 678 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: 679 if (curstep->arg1 > 0) 680 choose_local_tries = curstep->arg1; 681 break; 682 683 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: 684 if (curstep->arg1 > 0) 685 choose_local_fallback_tries = curstep->arg1; 686 break; 687 688 case CRUSH_RULE_CHOOSELEAF_FIRSTN: 689 case CRUSH_RULE_CHOOSE_FIRSTN: 690 firstn = 1; 691 /* fall through */ 692 case CRUSH_RULE_CHOOSELEAF_INDEP: 693 case CRUSH_RULE_CHOOSE_INDEP: 694 if (wsize == 0) 695 break; 696 697 recurse_to_leaf = 698 curstep->op == 699 CRUSH_RULE_CHOOSELEAF_FIRSTN || 700 curstep->op == 701 CRUSH_RULE_CHOOSELEAF_INDEP; 702 703 /* reset output */ 704 osize = 0; 705 706 for (i = 0; i < wsize; i++) { 707 /* 708 * see CRUSH_N, CRUSH_N_MINUS macros. 709 * basically, numrep <= 0 means relative to 710 * the provided result_max 711 */ 712 numrep = curstep->arg1; 713 if (numrep <= 0) { 714 numrep += result_max; 715 if (numrep <= 0) 716 continue; 717 } 718 j = 0; 719 if (firstn) { 720 int recurse_tries; 721 if (choose_leaf_tries) 722 recurse_tries = 723 choose_leaf_tries; 724 else if (map->chooseleaf_descend_once) 725 recurse_tries = 1; 726 else 727 recurse_tries = choose_tries; 728 osize += crush_choose_firstn( 729 map, 730 map->buckets[-1-w[i]], 731 weight, weight_max, 732 x, numrep, 733 curstep->arg2, 734 o+osize, j, 735 choose_tries, 736 recurse_tries, 737 choose_local_tries, 738 choose_local_fallback_tries, 739 recurse_to_leaf, 740 c+osize); 741 } else { 742 crush_choose_indep( 743 map, 744 map->buckets[-1-w[i]], 745 weight, weight_max, 746 x, numrep, numrep, 747 curstep->arg2, 748 o+osize, j, 749 choose_tries, 750 choose_leaf_tries ? 751 choose_leaf_tries : 1, 752 recurse_to_leaf, 753 c+osize, 754 0); 755 osize += numrep; 756 } 757 } 758 759 if (recurse_to_leaf) 760 /* copy final _leaf_ values to output set */ 761 memcpy(o, c, osize*sizeof(*o)); 762 763 /* swap o and w arrays */ 764 tmp = o; 765 o = w; 766 w = tmp; 767 wsize = osize; 768 break; 769 770 771 case CRUSH_RULE_EMIT: 772 for (i = 0; i < wsize && result_len < result_max; i++) { 773 result[result_len] = w[i]; 774 result_len++; 775 } 776 wsize = 0; 777 break; 778 779 default: 780 dprintk(" unknown op %d at step %d\n", 781 curstep->op, step); 782 break; 783 } 784 } 785 return result_len; 786 } 787 788 789