1 /* 2 * Ceph - scalable distributed file system 3 * 4 * Copyright (C) 2015 Intel Corporation All Rights Reserved 5 * 6 * This is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU Lesser General Public 8 * License version 2.1, as published by the Free Software 9 * Foundation. See file COPYING. 10 * 11 */ 12 13 #ifdef __KERNEL__ 14 # include <linux/string.h> 15 # include <linux/slab.h> 16 # include <linux/bug.h> 17 # include <linux/kernel.h> 18 # include <linux/crush/crush.h> 19 # include <linux/crush/hash.h> 20 #else 21 # include "crush_compat.h" 22 # include "crush.h" 23 # include "hash.h" 24 #endif 25 #include "crush_ln_table.h" 26 27 #define dprintk(args...) /* printf(args) */ 28 29 /* 30 * Implement the core CRUSH mapping algorithm. 31 */ 32 33 /** 34 * crush_find_rule - find a crush_rule id for a given ruleset, type, and size. 35 * @map: the crush_map 36 * @ruleset: the storage ruleset id (user defined) 37 * @type: storage ruleset type (user defined) 38 * @size: output set size 39 */ 40 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size) 41 { 42 __u32 i; 43 44 for (i = 0; i < map->max_rules; i++) { 45 if (map->rules[i] && 46 map->rules[i]->mask.ruleset == ruleset && 47 map->rules[i]->mask.type == type && 48 map->rules[i]->mask.min_size <= size && 49 map->rules[i]->mask.max_size >= size) 50 return i; 51 } 52 return -1; 53 } 54 55 56 /* 57 * bucket choose methods 58 * 59 * For each bucket algorithm, we have a "choose" method that, given a 60 * crush input @x and replica position (usually, position in output set) @r, 61 * will produce an item in the bucket. 62 */ 63 64 /* 65 * Choose based on a random permutation of the bucket. 66 * 67 * We used to use some prime number arithmetic to do this, but it 68 * wasn't very random, and had some other bad behaviors. Instead, we 69 * calculate an actual random permutation of the bucket members. 70 * Since this is expensive, we optimize for the r=0 case, which 71 * captures the vast majority of calls. 72 */ 73 static int bucket_perm_choose(struct crush_bucket *bucket, 74 int x, int r) 75 { 76 unsigned int pr = r % bucket->size; 77 unsigned int i, s; 78 79 /* start a new permutation if @x has changed */ 80 if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) { 81 dprintk("bucket %d new x=%d\n", bucket->id, x); 82 bucket->perm_x = x; 83 84 /* optimize common r=0 case */ 85 if (pr == 0) { 86 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) % 87 bucket->size; 88 bucket->perm[0] = s; 89 bucket->perm_n = 0xffff; /* magic value, see below */ 90 goto out; 91 } 92 93 for (i = 0; i < bucket->size; i++) 94 bucket->perm[i] = i; 95 bucket->perm_n = 0; 96 } else if (bucket->perm_n == 0xffff) { 97 /* clean up after the r=0 case above */ 98 for (i = 1; i < bucket->size; i++) 99 bucket->perm[i] = i; 100 bucket->perm[bucket->perm[0]] = 0; 101 bucket->perm_n = 1; 102 } 103 104 /* calculate permutation up to pr */ 105 for (i = 0; i < bucket->perm_n; i++) 106 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]); 107 while (bucket->perm_n <= pr) { 108 unsigned int p = bucket->perm_n; 109 /* no point in swapping the final entry */ 110 if (p < bucket->size - 1) { 111 i = crush_hash32_3(bucket->hash, x, bucket->id, p) % 112 (bucket->size - p); 113 if (i) { 114 unsigned int t = bucket->perm[p + i]; 115 bucket->perm[p + i] = bucket->perm[p]; 116 bucket->perm[p] = t; 117 } 118 dprintk(" perm_choose swap %d with %d\n", p, p+i); 119 } 120 bucket->perm_n++; 121 } 122 for (i = 0; i < bucket->size; i++) 123 dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]); 124 125 s = bucket->perm[pr]; 126 out: 127 dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id, 128 bucket->size, x, r, pr, s); 129 return bucket->items[s]; 130 } 131 132 /* uniform */ 133 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket, 134 int x, int r) 135 { 136 return bucket_perm_choose(&bucket->h, x, r); 137 } 138 139 /* list */ 140 static int bucket_list_choose(struct crush_bucket_list *bucket, 141 int x, int r) 142 { 143 int i; 144 145 for (i = bucket->h.size-1; i >= 0; i--) { 146 __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i], 147 r, bucket->h.id); 148 w &= 0xffff; 149 dprintk("list_choose i=%d x=%d r=%d item %d weight %x " 150 "sw %x rand %llx", 151 i, x, r, bucket->h.items[i], bucket->item_weights[i], 152 bucket->sum_weights[i], w); 153 w *= bucket->sum_weights[i]; 154 w = w >> 16; 155 /*dprintk(" scaled %llx\n", w);*/ 156 if (w < bucket->item_weights[i]) 157 return bucket->h.items[i]; 158 } 159 160 dprintk("bad list sums for bucket %d\n", bucket->h.id); 161 return bucket->h.items[0]; 162 } 163 164 165 /* (binary) tree */ 166 static int height(int n) 167 { 168 int h = 0; 169 while ((n & 1) == 0) { 170 h++; 171 n = n >> 1; 172 } 173 return h; 174 } 175 176 static int left(int x) 177 { 178 int h = height(x); 179 return x - (1 << (h-1)); 180 } 181 182 static int right(int x) 183 { 184 int h = height(x); 185 return x + (1 << (h-1)); 186 } 187 188 static int terminal(int x) 189 { 190 return x & 1; 191 } 192 193 static int bucket_tree_choose(struct crush_bucket_tree *bucket, 194 int x, int r) 195 { 196 int n; 197 __u32 w; 198 __u64 t; 199 200 /* start at root */ 201 n = bucket->num_nodes >> 1; 202 203 while (!terminal(n)) { 204 int l; 205 /* pick point in [0, w) */ 206 w = bucket->node_weights[n]; 207 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r, 208 bucket->h.id) * (__u64)w; 209 t = t >> 32; 210 211 /* descend to the left or right? */ 212 l = left(n); 213 if (t < bucket->node_weights[l]) 214 n = l; 215 else 216 n = right(n); 217 } 218 219 return bucket->h.items[n >> 1]; 220 } 221 222 223 /* straw */ 224 225 static int bucket_straw_choose(struct crush_bucket_straw *bucket, 226 int x, int r) 227 { 228 __u32 i; 229 int high = 0; 230 __u64 high_draw = 0; 231 __u64 draw; 232 233 for (i = 0; i < bucket->h.size; i++) { 234 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r); 235 draw &= 0xffff; 236 draw *= bucket->straws[i]; 237 if (i == 0 || draw > high_draw) { 238 high = i; 239 high_draw = draw; 240 } 241 } 242 return bucket->h.items[high]; 243 } 244 245 /* compute 2^44*log2(input+1) */ 246 static __u64 crush_ln(unsigned int xin) 247 { 248 unsigned int x = xin; 249 int iexpon, index1, index2; 250 __u64 RH, LH, LL, xl64, result; 251 252 x++; 253 254 /* normalize input */ 255 iexpon = 15; 256 257 /* 258 * figure out number of bits we need to shift and 259 * do it in one step instead of iteratively 260 */ 261 if (!(x & 0x18000)) { 262 int bits = __builtin_clz(x & 0x1FFFF) - 16; 263 x <<= bits; 264 iexpon = 15 - bits; 265 } 266 267 index1 = (x >> 8) << 1; 268 /* RH ~ 2^56/index1 */ 269 RH = __RH_LH_tbl[index1 - 256]; 270 /* LH ~ 2^48 * log2(index1/256) */ 271 LH = __RH_LH_tbl[index1 + 1 - 256]; 272 273 /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */ 274 xl64 = (__s64)x * RH; 275 xl64 >>= 48; 276 277 result = iexpon; 278 result <<= (12 + 32); 279 280 index2 = xl64 & 0xff; 281 /* LL ~ 2^48*log2(1.0+index2/2^15) */ 282 LL = __LL_tbl[index2]; 283 284 LH = LH + LL; 285 286 LH >>= (48 - 12 - 32); 287 result += LH; 288 289 return result; 290 } 291 292 293 /* 294 * straw2 295 * 296 * for reference, see: 297 * 298 * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables 299 * 300 */ 301 302 static int bucket_straw2_choose(struct crush_bucket_straw2 *bucket, 303 int x, int r) 304 { 305 unsigned int i, high = 0; 306 unsigned int u; 307 unsigned int w; 308 __s64 ln, draw, high_draw = 0; 309 310 for (i = 0; i < bucket->h.size; i++) { 311 w = bucket->item_weights[i]; 312 if (w) { 313 u = crush_hash32_3(bucket->h.hash, x, 314 bucket->h.items[i], r); 315 u &= 0xffff; 316 317 /* 318 * for some reason slightly less than 0x10000 produces 319 * a slightly more accurate distribution... probably a 320 * rounding effect. 321 * 322 * the natural log lookup table maps [0,0xffff] 323 * (corresponding to real numbers [1/0x10000, 1] to 324 * [0, 0xffffffffffff] (corresponding to real numbers 325 * [-11.090355,0]). 326 */ 327 ln = crush_ln(u) - 0x1000000000000ll; 328 329 /* 330 * divide by 16.16 fixed-point weight. note 331 * that the ln value is negative, so a larger 332 * weight means a larger (less negative) value 333 * for draw. 334 */ 335 draw = div64_s64(ln, w); 336 } else { 337 draw = S64_MIN; 338 } 339 340 if (i == 0 || draw > high_draw) { 341 high = i; 342 high_draw = draw; 343 } 344 } 345 return bucket->h.items[high]; 346 } 347 348 349 static int crush_bucket_choose(struct crush_bucket *in, int x, int r) 350 { 351 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r); 352 BUG_ON(in->size == 0); 353 switch (in->alg) { 354 case CRUSH_BUCKET_UNIFORM: 355 return bucket_uniform_choose((struct crush_bucket_uniform *)in, 356 x, r); 357 case CRUSH_BUCKET_LIST: 358 return bucket_list_choose((struct crush_bucket_list *)in, 359 x, r); 360 case CRUSH_BUCKET_TREE: 361 return bucket_tree_choose((struct crush_bucket_tree *)in, 362 x, r); 363 case CRUSH_BUCKET_STRAW: 364 return bucket_straw_choose((struct crush_bucket_straw *)in, 365 x, r); 366 case CRUSH_BUCKET_STRAW2: 367 return bucket_straw2_choose((struct crush_bucket_straw2 *)in, 368 x, r); 369 default: 370 dprintk("unknown bucket %d alg %d\n", in->id, in->alg); 371 return in->items[0]; 372 } 373 } 374 375 376 /* 377 * true if device is marked "out" (failed, fully offloaded) 378 * of the cluster 379 */ 380 static int is_out(const struct crush_map *map, 381 const __u32 *weight, int weight_max, 382 int item, int x) 383 { 384 if (item >= weight_max) 385 return 1; 386 if (weight[item] >= 0x10000) 387 return 0; 388 if (weight[item] == 0) 389 return 1; 390 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff) 391 < weight[item]) 392 return 0; 393 return 1; 394 } 395 396 /** 397 * crush_choose_firstn - choose numrep distinct items of given type 398 * @map: the crush_map 399 * @bucket: the bucket we are choose an item from 400 * @x: crush input value 401 * @numrep: the number of items to choose 402 * @type: the type of item to choose 403 * @out: pointer to output vector 404 * @outpos: our position in that vector 405 * @out_size: size of the out vector 406 * @tries: number of attempts to make 407 * @recurse_tries: number of attempts to have recursive chooseleaf make 408 * @local_retries: localized retries 409 * @local_fallback_retries: localized fallback retries 410 * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) 411 * @stable: stable mode starts rep=0 in the recursive call for all replicas 412 * @vary_r: pass r to recursive calls 413 * @out2: second output vector for leaf items (if @recurse_to_leaf) 414 * @parent_r: r value passed from the parent 415 */ 416 static int crush_choose_firstn(const struct crush_map *map, 417 struct crush_bucket *bucket, 418 const __u32 *weight, int weight_max, 419 int x, int numrep, int type, 420 int *out, int outpos, 421 int out_size, 422 unsigned int tries, 423 unsigned int recurse_tries, 424 unsigned int local_retries, 425 unsigned int local_fallback_retries, 426 int recurse_to_leaf, 427 unsigned int vary_r, 428 unsigned int stable, 429 int *out2, 430 int parent_r) 431 { 432 int rep; 433 unsigned int ftotal, flocal; 434 int retry_descent, retry_bucket, skip_rep; 435 struct crush_bucket *in = bucket; 436 int r; 437 int i; 438 int item = 0; 439 int itemtype; 440 int collide, reject; 441 int count = out_size; 442 443 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 stable %d\n", 444 recurse_to_leaf ? "_LEAF" : "", 445 bucket->id, x, outpos, numrep, 446 tries, recurse_tries, local_retries, local_fallback_retries, 447 parent_r, stable); 448 449 for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) { 450 /* keep trying until we get a non-out, non-colliding item */ 451 ftotal = 0; 452 skip_rep = 0; 453 do { 454 retry_descent = 0; 455 in = bucket; /* initial bucket */ 456 457 /* choose through intervening buckets */ 458 flocal = 0; 459 do { 460 collide = 0; 461 retry_bucket = 0; 462 r = rep + parent_r; 463 /* r' = r + f_total */ 464 r += ftotal; 465 466 /* bucket choose */ 467 if (in->size == 0) { 468 reject = 1; 469 goto reject; 470 } 471 if (local_fallback_retries > 0 && 472 flocal >= (in->size>>1) && 473 flocal > local_fallback_retries) 474 item = bucket_perm_choose(in, x, r); 475 else 476 item = crush_bucket_choose(in, x, r); 477 if (item >= map->max_devices) { 478 dprintk(" bad item %d\n", item); 479 skip_rep = 1; 480 break; 481 } 482 483 /* desired type? */ 484 if (item < 0) 485 itemtype = map->buckets[-1-item]->type; 486 else 487 itemtype = 0; 488 dprintk(" item %d type %d\n", item, itemtype); 489 490 /* keep going? */ 491 if (itemtype != type) { 492 if (item >= 0 || 493 (-1-item) >= map->max_buckets) { 494 dprintk(" bad item type %d\n", type); 495 skip_rep = 1; 496 break; 497 } 498 in = map->buckets[-1-item]; 499 retry_bucket = 1; 500 continue; 501 } 502 503 /* collision? */ 504 for (i = 0; i < outpos; i++) { 505 if (out[i] == item) { 506 collide = 1; 507 break; 508 } 509 } 510 511 reject = 0; 512 if (!collide && recurse_to_leaf) { 513 if (item < 0) { 514 int sub_r; 515 if (vary_r) 516 sub_r = r >> (vary_r-1); 517 else 518 sub_r = 0; 519 if (crush_choose_firstn(map, 520 map->buckets[-1-item], 521 weight, weight_max, 522 x, stable ? 1 : outpos+1, 0, 523 out2, outpos, count, 524 recurse_tries, 0, 525 local_retries, 526 local_fallback_retries, 527 0, 528 vary_r, 529 stable, 530 NULL, 531 sub_r) <= outpos) 532 /* didn't get leaf */ 533 reject = 1; 534 } else { 535 /* we already have a leaf! */ 536 out2[outpos] = item; 537 } 538 } 539 540 if (!reject) { 541 /* out? */ 542 if (itemtype == 0) 543 reject = is_out(map, weight, 544 weight_max, 545 item, x); 546 else 547 reject = 0; 548 } 549 550 reject: 551 if (reject || collide) { 552 ftotal++; 553 flocal++; 554 555 if (collide && flocal <= local_retries) 556 /* retry locally a few times */ 557 retry_bucket = 1; 558 else if (local_fallback_retries > 0 && 559 flocal <= in->size + local_fallback_retries) 560 /* exhaustive bucket search */ 561 retry_bucket = 1; 562 else if (ftotal < tries) 563 /* then retry descent */ 564 retry_descent = 1; 565 else 566 /* else give up */ 567 skip_rep = 1; 568 dprintk(" reject %d collide %d " 569 "ftotal %u flocal %u\n", 570 reject, collide, ftotal, 571 flocal); 572 } 573 } while (retry_bucket); 574 } while (retry_descent); 575 576 if (skip_rep) { 577 dprintk("skip rep\n"); 578 continue; 579 } 580 581 dprintk("CHOOSE got %d\n", item); 582 out[outpos] = item; 583 outpos++; 584 count--; 585 #ifndef __KERNEL__ 586 if (map->choose_tries && ftotal <= map->choose_total_tries) 587 map->choose_tries[ftotal]++; 588 #endif 589 } 590 591 dprintk("CHOOSE returns %d\n", outpos); 592 return outpos; 593 } 594 595 596 /** 597 * crush_choose_indep: alternative breadth-first positionally stable mapping 598 * 599 */ 600 static void crush_choose_indep(const struct crush_map *map, 601 struct crush_bucket *bucket, 602 const __u32 *weight, int weight_max, 603 int x, int left, int numrep, int type, 604 int *out, int outpos, 605 unsigned int tries, 606 unsigned int recurse_tries, 607 int recurse_to_leaf, 608 int *out2, 609 int parent_r) 610 { 611 struct crush_bucket *in = bucket; 612 int endpos = outpos + left; 613 int rep; 614 unsigned int ftotal; 615 int r; 616 int i; 617 int item = 0; 618 int itemtype; 619 int collide; 620 621 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", 622 bucket->id, x, outpos, numrep); 623 624 /* initially my result is undefined */ 625 for (rep = outpos; rep < endpos; rep++) { 626 out[rep] = CRUSH_ITEM_UNDEF; 627 if (out2) 628 out2[rep] = CRUSH_ITEM_UNDEF; 629 } 630 631 for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { 632 #ifdef DEBUG_INDEP 633 if (out2 && ftotal) { 634 dprintk("%u %d a: ", ftotal, left); 635 for (rep = outpos; rep < endpos; rep++) { 636 dprintk(" %d", out[rep]); 637 } 638 dprintk("\n"); 639 dprintk("%u %d b: ", ftotal, left); 640 for (rep = outpos; rep < endpos; rep++) { 641 dprintk(" %d", out2[rep]); 642 } 643 dprintk("\n"); 644 } 645 #endif 646 for (rep = outpos; rep < endpos; rep++) { 647 if (out[rep] != CRUSH_ITEM_UNDEF) 648 continue; 649 650 in = bucket; /* initial bucket */ 651 652 /* choose through intervening buckets */ 653 for (;;) { 654 /* note: we base the choice on the position 655 * even in the nested call. that means that 656 * if the first layer chooses the same bucket 657 * in a different position, we will tend to 658 * choose a different item in that bucket. 659 * this will involve more devices in data 660 * movement and tend to distribute the load. 661 */ 662 r = rep + parent_r; 663 664 /* be careful */ 665 if (in->alg == CRUSH_BUCKET_UNIFORM && 666 in->size % numrep == 0) 667 /* r'=r+(n+1)*f_total */ 668 r += (numrep+1) * ftotal; 669 else 670 /* r' = r + n*f_total */ 671 r += numrep * ftotal; 672 673 /* bucket choose */ 674 if (in->size == 0) { 675 dprintk(" empty bucket\n"); 676 break; 677 } 678 679 item = crush_bucket_choose(in, x, r); 680 if (item >= map->max_devices) { 681 dprintk(" bad item %d\n", item); 682 out[rep] = CRUSH_ITEM_NONE; 683 if (out2) 684 out2[rep] = CRUSH_ITEM_NONE; 685 left--; 686 break; 687 } 688 689 /* desired type? */ 690 if (item < 0) 691 itemtype = map->buckets[-1-item]->type; 692 else 693 itemtype = 0; 694 dprintk(" item %d type %d\n", item, itemtype); 695 696 /* keep going? */ 697 if (itemtype != type) { 698 if (item >= 0 || 699 (-1-item) >= map->max_buckets) { 700 dprintk(" bad item type %d\n", type); 701 out[rep] = CRUSH_ITEM_NONE; 702 if (out2) 703 out2[rep] = 704 CRUSH_ITEM_NONE; 705 left--; 706 break; 707 } 708 in = map->buckets[-1-item]; 709 continue; 710 } 711 712 /* collision? */ 713 collide = 0; 714 for (i = outpos; i < endpos; i++) { 715 if (out[i] == item) { 716 collide = 1; 717 break; 718 } 719 } 720 if (collide) 721 break; 722 723 if (recurse_to_leaf) { 724 if (item < 0) { 725 crush_choose_indep(map, 726 map->buckets[-1-item], 727 weight, weight_max, 728 x, 1, numrep, 0, 729 out2, rep, 730 recurse_tries, 0, 731 0, NULL, r); 732 if (out2[rep] == CRUSH_ITEM_NONE) { 733 /* placed nothing; no leaf */ 734 break; 735 } 736 } else { 737 /* we already have a leaf! */ 738 out2[rep] = item; 739 } 740 } 741 742 /* out? */ 743 if (itemtype == 0 && 744 is_out(map, weight, weight_max, item, x)) 745 break; 746 747 /* yay! */ 748 out[rep] = item; 749 left--; 750 break; 751 } 752 } 753 } 754 for (rep = outpos; rep < endpos; rep++) { 755 if (out[rep] == CRUSH_ITEM_UNDEF) { 756 out[rep] = CRUSH_ITEM_NONE; 757 } 758 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { 759 out2[rep] = CRUSH_ITEM_NONE; 760 } 761 } 762 #ifndef __KERNEL__ 763 if (map->choose_tries && ftotal <= map->choose_total_tries) 764 map->choose_tries[ftotal]++; 765 #endif 766 #ifdef DEBUG_INDEP 767 if (out2) { 768 dprintk("%u %d a: ", ftotal, left); 769 for (rep = outpos; rep < endpos; rep++) { 770 dprintk(" %d", out[rep]); 771 } 772 dprintk("\n"); 773 dprintk("%u %d b: ", ftotal, left); 774 for (rep = outpos; rep < endpos; rep++) { 775 dprintk(" %d", out2[rep]); 776 } 777 dprintk("\n"); 778 } 779 #endif 780 } 781 782 /** 783 * crush_do_rule - calculate a mapping with the given input and rule 784 * @map: the crush_map 785 * @ruleno: the rule id 786 * @x: hash input 787 * @result: pointer to result vector 788 * @result_max: maximum result size 789 * @weight: weight vector (for map leaves) 790 * @weight_max: size of weight vector 791 * @scratch: scratch vector for private use; must be >= 3 * result_max 792 */ 793 int crush_do_rule(const struct crush_map *map, 794 int ruleno, int x, int *result, int result_max, 795 const __u32 *weight, int weight_max, 796 int *scratch) 797 { 798 int result_len; 799 int *a = scratch; 800 int *b = scratch + result_max; 801 int *c = scratch + result_max*2; 802 int recurse_to_leaf; 803 int *w; 804 int wsize = 0; 805 int *o; 806 int osize; 807 int *tmp; 808 struct crush_rule *rule; 809 __u32 step; 810 int i, j; 811 int numrep; 812 int out_size; 813 /* 814 * the original choose_total_tries value was off by one (it 815 * counted "retries" and not "tries"). add one. 816 */ 817 int choose_tries = map->choose_total_tries + 1; 818 int choose_leaf_tries = 0; 819 /* 820 * the local tries values were counted as "retries", though, 821 * and need no adjustment 822 */ 823 int choose_local_retries = map->choose_local_tries; 824 int choose_local_fallback_retries = map->choose_local_fallback_tries; 825 826 int vary_r = map->chooseleaf_vary_r; 827 int stable = map->chooseleaf_stable; 828 829 if ((__u32)ruleno >= map->max_rules) { 830 dprintk(" bad ruleno %d\n", ruleno); 831 return 0; 832 } 833 834 rule = map->rules[ruleno]; 835 result_len = 0; 836 w = a; 837 o = b; 838 839 for (step = 0; step < rule->len; step++) { 840 int firstn = 0; 841 struct crush_rule_step *curstep = &rule->steps[step]; 842 843 switch (curstep->op) { 844 case CRUSH_RULE_TAKE: 845 if ((curstep->arg1 >= 0 && 846 curstep->arg1 < map->max_devices) || 847 (-1-curstep->arg1 >= 0 && 848 -1-curstep->arg1 < map->max_buckets && 849 map->buckets[-1-curstep->arg1])) { 850 w[0] = curstep->arg1; 851 wsize = 1; 852 } else { 853 dprintk(" bad take value %d\n", curstep->arg1); 854 } 855 break; 856 857 case CRUSH_RULE_SET_CHOOSE_TRIES: 858 if (curstep->arg1 > 0) 859 choose_tries = curstep->arg1; 860 break; 861 862 case CRUSH_RULE_SET_CHOOSELEAF_TRIES: 863 if (curstep->arg1 > 0) 864 choose_leaf_tries = curstep->arg1; 865 break; 866 867 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: 868 if (curstep->arg1 >= 0) 869 choose_local_retries = curstep->arg1; 870 break; 871 872 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: 873 if (curstep->arg1 >= 0) 874 choose_local_fallback_retries = curstep->arg1; 875 break; 876 877 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R: 878 if (curstep->arg1 >= 0) 879 vary_r = curstep->arg1; 880 break; 881 882 case CRUSH_RULE_SET_CHOOSELEAF_STABLE: 883 if (curstep->arg1 >= 0) 884 stable = curstep->arg1; 885 break; 886 887 case CRUSH_RULE_CHOOSELEAF_FIRSTN: 888 case CRUSH_RULE_CHOOSE_FIRSTN: 889 firstn = 1; 890 /* fall through */ 891 case CRUSH_RULE_CHOOSELEAF_INDEP: 892 case CRUSH_RULE_CHOOSE_INDEP: 893 if (wsize == 0) 894 break; 895 896 recurse_to_leaf = 897 curstep->op == 898 CRUSH_RULE_CHOOSELEAF_FIRSTN || 899 curstep->op == 900 CRUSH_RULE_CHOOSELEAF_INDEP; 901 902 /* reset output */ 903 osize = 0; 904 905 for (i = 0; i < wsize; i++) { 906 int bno; 907 /* 908 * see CRUSH_N, CRUSH_N_MINUS macros. 909 * basically, numrep <= 0 means relative to 910 * the provided result_max 911 */ 912 numrep = curstep->arg1; 913 if (numrep <= 0) { 914 numrep += result_max; 915 if (numrep <= 0) 916 continue; 917 } 918 j = 0; 919 /* make sure bucket id is valid */ 920 bno = -1 - w[i]; 921 if (bno < 0 || bno >= map->max_buckets) { 922 /* w[i] is probably CRUSH_ITEM_NONE */ 923 dprintk(" bad w[i] %d\n", w[i]); 924 continue; 925 } 926 if (firstn) { 927 int recurse_tries; 928 if (choose_leaf_tries) 929 recurse_tries = 930 choose_leaf_tries; 931 else if (map->chooseleaf_descend_once) 932 recurse_tries = 1; 933 else 934 recurse_tries = choose_tries; 935 osize += crush_choose_firstn( 936 map, 937 map->buckets[bno], 938 weight, weight_max, 939 x, numrep, 940 curstep->arg2, 941 o+osize, j, 942 result_max-osize, 943 choose_tries, 944 recurse_tries, 945 choose_local_retries, 946 choose_local_fallback_retries, 947 recurse_to_leaf, 948 vary_r, 949 stable, 950 c+osize, 951 0); 952 } else { 953 out_size = ((numrep < (result_max-osize)) ? 954 numrep : (result_max-osize)); 955 crush_choose_indep( 956 map, 957 map->buckets[bno], 958 weight, weight_max, 959 x, out_size, numrep, 960 curstep->arg2, 961 o+osize, j, 962 choose_tries, 963 choose_leaf_tries ? 964 choose_leaf_tries : 1, 965 recurse_to_leaf, 966 c+osize, 967 0); 968 osize += out_size; 969 } 970 } 971 972 if (recurse_to_leaf) 973 /* copy final _leaf_ values to output set */ 974 memcpy(o, c, osize*sizeof(*o)); 975 976 /* swap o and w arrays */ 977 tmp = o; 978 o = w; 979 w = tmp; 980 wsize = osize; 981 break; 982 983 984 case CRUSH_RULE_EMIT: 985 for (i = 0; i < wsize && result_len < result_max; i++) { 986 result[result_len] = w[i]; 987 result_len++; 988 } 989 wsize = 0; 990 break; 991 992 default: 993 dprintk(" unknown op %d at step %d\n", 994 curstep->op, step); 995 break; 996 } 997 } 998 return result_len; 999 } 1000