net/ceph/crush/mapper.c

   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  * @recurse_to_leaf: true if we want one device under each item of given type
 294  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 295  */
 296 static int crush_choose_firstn(const struct crush_map *map,
 297                                struct crush_bucket *bucket,
 298                                const __u32 *weight, int weight_max,
 299                                int x, int numrep, int type,
 300                                int *out, int outpos,
 301                                unsigned int attempts,
 302                                unsigned int recurse_attempts,
 303                                unsigned int local_tries,
 304                                unsigned int local_fallback_tries,
 305                                int recurse_to_leaf,
 306                                int *out2)
 307 {
 308         int rep;
 309         unsigned int ftotal, flocal;
 310         int retry_descent, retry_bucket, skip_rep;
 311         struct crush_bucket *in = bucket;
 312         int r;
 313         int i;
 314         int item = 0;
 315         int itemtype;
 316         int collide, reject;
 317
 318         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 319                 bucket->id, x, outpos, numrep);
 320
 321         for (rep = outpos; rep < numrep; rep++) {
 322                 /* keep trying until we get a non-out, non-colliding item */
 323                 ftotal = 0;
 324                 skip_rep = 0;
 325                 do {
 326                         retry_descent = 0;
 327                         in = bucket;               /* initial bucket */
 328
 329                         /* choose through intervening buckets */
 330                         flocal = 0;
 331                         do {
 332                                 collide = 0;
 333                                 retry_bucket = 0;
 334                                 r = rep;
 335                                 /* r' = r + f_total */
 336                                 r += ftotal;
 337
 338                                 /* bucket choose */
 339                                 if (in->size == 0) {
 340                                         reject = 1;
 341                                         goto reject;
 342                                 }
 343                                 if (local_fallback_tries > 0 &&
 344                                     flocal >= (in->size>>1) &&
 345                                     flocal > local_fallback_tries)
 346                                         item = bucket_perm_choose(in, x, r);
 347                                 else
 348                                         item = crush_bucket_choose(in, x, r);
 349                                 if (item >= map->max_devices) {
 350                                         dprintk("   bad item %d\n", item);
 351                                         skip_rep = 1;
 352                                         break;
 353                                 }
 354
 355                                 /* desired type? */
 356                                 if (item < 0)
 357                                         itemtype = map->buckets[-1-item]->type;
 358                                 else
 359                                         itemtype = 0;
 360                                 dprintk("  item %d type %d\n", item, itemtype);
 361
 362                                 /* keep going? */
 363                                 if (itemtype != type) {
 364                                         if (item >= 0 ||
 365                                             (-1-item) >= map->max_buckets) {
 366                                                 dprintk("   bad item type %d\n", type);
 367                                                 skip_rep = 1;
 368                                                 break;
 369                                         }
 370                                         in = map->buckets[-1-item];
 371                                         retry_bucket = 1;
 372                                         continue;
 373                                 }
 374
 375                                 /* collision? */
 376                                 for (i = 0; i < outpos; i++) {
 377                                         if (out[i] == item) {
 378                                                 collide = 1;
 379                                                 break;
 380                                         }
 381                                 }
 382
 383                                 reject = 0;
 384                                 if (!collide && recurse_to_leaf) {
 385                                         if (item < 0) {
 386                                                 if (crush_choose_firstn(map,
 387                                                          map->buckets[-1-item],
 388                                                          weight, weight_max,
 389                                                          x, outpos+1, 0,
 390                                                          out2, outpos,
 391                                                          recurse_attempts, 0,
 392                                                          local_tries,
 393                                                          local_fallback_tries,
 394                                                          0,
 395                                                          NULL) <= outpos)
 396                                                         /* didn't get leaf */
 397                                                         reject = 1;
 398                                         } else {
 399                                                 /* we already have a leaf! */
 400                                                 out2[outpos] = item;
 401                                         }
 402                                 }
 403
 404                                 if (!reject) {
 405                                         /* out? */
 406                                         if (itemtype == 0)
 407                                                 reject = is_out(map, weight,
 408                                                                 weight_max,
 409                                                                 item, x);
 410                                         else
 411                                                 reject = 0;
 412                                 }
 413
 414 reject:
 415                                 if (reject || collide) {
 416                                         ftotal++;
 417                                         flocal++;
 418
 419                                         if (collide && flocal <= local_tries)
 420                                                 /* retry locally a few times */
 421                                                 retry_bucket = 1;
 422                                         else if (local_fallback_tries > 0 &&
 423                                                  flocal <= in->size + local_fallback_tries)
 424                                                 /* exhaustive bucket search */
 425                                                 retry_bucket = 1;
 426                                         else if (ftotal <= attempts)
 427                                                 /* then retry descent */
 428                                                 retry_descent = 1;
 429                                         else
 430                                                 /* else give up */
 431                                                 skip_rep = 1;
 432                                         dprintk("  reject %d  collide %d  "
 433                                                 "ftotal %u  flocal %u\n",
 434                                                 reject, collide, ftotal,
 435                                                 flocal);
 436                                 }
 437                         } while (retry_bucket);
 438                 } while (retry_descent);
 439
 440                 if (skip_rep) {
 441                         dprintk("skip rep\n");
 442                         continue;
 443                 }
 444
 445                 dprintk("CHOOSE got %d\n", item);
 446                 out[outpos] = item;
 447                 outpos++;
 448         }
 449
 450         dprintk("CHOOSE returns %d\n", outpos);
 451         return outpos;
 452 }
 453
 454
 455 /**
 456  * crush_choose_indep: alternative breadth-first positionally stable mapping
 457  *
 458  */
 459 static void crush_choose_indep(const struct crush_map *map,
 460                                struct crush_bucket *bucket,
 461                                const __u32 *weight, int weight_max,
 462                                int x, int left, int numrep, int type,
 463                                int *out, int outpos,
 464                                unsigned int attempts,
 465                                unsigned int recurse_attempts,
 466                                int recurse_to_leaf,
 467                                int *out2,
 468                                int parent_r)
 469 {
 470         struct crush_bucket *in = bucket;
 471         int endpos = outpos + left;
 472         int rep;
 473         unsigned int ftotal;
 474         int r;
 475         int i;
 476         int item = 0;
 477         int itemtype;
 478         int collide;
 479
 480         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 481                 bucket->id, x, outpos, numrep);
 482
 483         /* initially my result is undefined */
 484         for (rep = outpos; rep < endpos; rep++) {
 485                 out[rep] = CRUSH_ITEM_UNDEF;
 486                 if (out2)
 487                         out2[rep] = CRUSH_ITEM_UNDEF;
 488         }
 489
 490         for (ftotal = 0; left > 0 && ftotal < attempts; ftotal++) {
 491                 for (rep = outpos; rep < endpos; rep++) {
 492                         if (out[rep] != CRUSH_ITEM_UNDEF)
 493                                 continue;
 494
 495                         in = bucket;  /* initial bucket */
 496
 497                         /* choose through intervening buckets */
 498                         for (;;) {
 499                                 /* note: we base the choice on the position
 500                                  * even in the nested call.  that means that
 501                                  * if the first layer chooses the same bucket
 502                                  * in a different position, we will tend to
 503                                  * choose a different item in that bucket.
 504                                  * this will involve more devices in data
 505                                  * movement and tend to distribute the load.
 506                                  */
 507                                 r = rep + parent_r;
 508
 509                                 /* be careful */
 510                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
 511                                     in->size % numrep == 0)
 512                                         /* r'=r+(n+1)*f_total */
 513                                         r += (numrep+1) * ftotal;
 514                                 else
 515                                         /* r' = r + n*f_total */
 516                                         r += numrep * ftotal;
 517
 518                                 /* bucket choose */
 519                                 if (in->size == 0) {
 520                                         dprintk("   empty bucket\n");
 521                                         break;
 522                                 }
 523
 524                                 item = crush_bucket_choose(in, x, r);
 525                                 if (item >= map->max_devices) {
 526                                         dprintk("   bad item %d\n", item);
 527                                         out[rep] = CRUSH_ITEM_NONE;
 528                                         if (out2)
 529                                                 out2[rep] = CRUSH_ITEM_NONE;
 530                                         left--;
 531                                         break;
 532                                 }
 533
 534                                 /* desired type? */
 535                                 if (item < 0)
 536                                         itemtype = map->buckets[-1-item]->type;
 537                                 else
 538                                         itemtype = 0;
 539                                 dprintk("  item %d type %d\n", item, itemtype);
 540
 541                                 /* keep going? */
 542                                 if (itemtype != type) {
 543                                         if (item >= 0 ||
 544                                             (-1-item) >= map->max_buckets) {
 545                                                 dprintk("   bad item type %d\n", type);
 546                                                 out[rep] = CRUSH_ITEM_NONE;
 547                                                 if (out2)
 548                                                         out2[rep] =
 549                                                                 CRUSH_ITEM_NONE;
 550                                                 left--;
 551                                                 break;
 552                                         }
 553                                         in = map->buckets[-1-item];
 554                                         continue;
 555                                 }
 556
 557                                 /* collision? */
 558                                 collide = 0;
 559                                 for (i = outpos; i < endpos; i++) {
 560                                         if (out[i] == item) {
 561                                                 collide = 1;
 562                                                 break;
 563                                         }
 564                                 }
 565                                 if (collide)
 566                                         break;
 567
 568                                 if (recurse_to_leaf) {
 569                                         if (item < 0) {
 570                                                 crush_choose_indep(map,
 571                                                    map->buckets[-1-item],
 572                                                    weight, weight_max,
 573                                                    x, 1, numrep, 0,
 574                                                    out2, rep,
 575                                                    recurse_attempts, 0,
 576                                                    0, NULL, r);
 577                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 578                                                         /* placed nothing; no leaf */
 579                                                         break;
 580                                                 }
 581                                         } else {
 582                                                 /* we already have a leaf! */
 583                                                 out2[rep] = item;
 584                                         }
 585                                 }
 586
 587                                 /* out? */
 588                                 if (itemtype == 0 &&
 589                                     is_out(map, weight, weight_max, item, x))
 590                                         break;
 591
 592                                 /* yay! */
 593                                 out[rep] = item;
 594                                 left--;
 595                                 break;
 596                         }
 597                 }
 598         }
 599         for (rep = outpos; rep < endpos; rep++) {
 600                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 601                         out[rep] = CRUSH_ITEM_NONE;
 602                 }
 603                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 604                         out2[rep] = CRUSH_ITEM_NONE;
 605                 }
 606         }
 607 }
 608
 609 /**
 610  * crush_do_rule - calculate a mapping with the given input and rule
 611  * @map: the crush_map
 612  * @ruleno: the rule id
 613  * @x: hash input
 614  * @result: pointer to result vector
 615  * @result_max: maximum result size
 616  * @weight: weight vector (for map leaves)
 617  * @weight_max: size of weight vector
 618  * @scratch: scratch vector for private use; must be >= 3 * result_max
 619  */
 620 int crush_do_rule(const struct crush_map *map,
 621                   int ruleno, int x, int *result, int result_max,
 622                   const __u32 *weight, int weight_max,
 623                   int *scratch)
 624 {
 625         int result_len;
 626         int *a = scratch;
 627         int *b = scratch + result_max;
 628         int *c = scratch + result_max*2;
 629         int recurse_to_leaf;
 630         int *w;
 631         int wsize = 0;
 632         int *o;
 633         int osize;
 634         int *tmp;
 635         struct crush_rule *rule;
 636         __u32 step;
 637         int i, j;
 638         int numrep;
 639         int choose_tries = map->choose_total_tries;
 640         int choose_local_tries = map->choose_local_tries;
 641         int choose_local_fallback_tries = map->choose_local_fallback_tries;
 642         int choose_leaf_tries = 0;
 643
 644         if ((__u32)ruleno >= map->max_rules) {
 645                 dprintk(" bad ruleno %d\n", ruleno);
 646                 return 0;
 647         }
 648
 649         rule = map->rules[ruleno];
 650         result_len = 0;
 651         w = a;
 652         o = b;
 653
 654         for (step = 0; step < rule->len; step++) {
 655                 int firstn = 0;
 656                 struct crush_rule_step *curstep = &rule->steps[step];
 657
 658                 switch (curstep->op) {
 659                 case CRUSH_RULE_TAKE:
 660                         w[0] = curstep->arg1;
 661                         wsize = 1;
 662                         break;
 663
 664                 case CRUSH_RULE_SET_CHOOSE_TRIES:
 665                         if (curstep->arg1 > 0)
 666                                 choose_tries = curstep->arg1;
 667                         break;
 668
 669                 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
 670                         if (curstep->arg1 > 0)
 671                                 choose_leaf_tries = curstep->arg1;
 672                         break;
 673
 674                 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
 675                         if (curstep->arg1 > 0)
 676                                 choose_local_tries = curstep->arg1;
 677                         break;
 678
 679                 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
 680                         if (curstep->arg1 > 0)
 681                                 choose_local_fallback_tries = curstep->arg1;
 682                         break;
 683
 684                 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
 685                 case CRUSH_RULE_CHOOSE_FIRSTN:
 686                         firstn = 1;
 687                         /* fall through */
 688                 case CRUSH_RULE_CHOOSELEAF_INDEP:
 689                 case CRUSH_RULE_CHOOSE_INDEP:
 690                         if (wsize == 0)
 691                                 break;
 692
 693                         recurse_to_leaf =
 694                                 curstep->op ==
 695                                  CRUSH_RULE_CHOOSELEAF_FIRSTN ||
 696                                 curstep->op ==
 697                                 CRUSH_RULE_CHOOSELEAF_INDEP;
 698
 699                         /* reset output */
 700                         osize = 0;
 701
 702                         for (i = 0; i < wsize; i++) {
 703                                 /*
 704                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 705                                  * basically, numrep <= 0 means relative to
 706                                  * the provided result_max
 707                                  */
 708                                 numrep = curstep->arg1;
 709                                 if (numrep <= 0) {
 710                                         numrep += result_max;
 711                                         if (numrep <= 0)
 712                                                 continue;
 713                                 }
 714                                 j = 0;
 715                                 if (firstn) {
 716                                         int recurse_tries;
 717                                         if (choose_leaf_tries)
 718                                                 recurse_tries =
 719                                                         choose_leaf_tries;
 720                                         else if (map->chooseleaf_descend_once)
 721                                                 recurse_tries = 1;
 722                                         else
 723                                                 recurse_tries = choose_tries;
 724                                         osize += crush_choose_firstn(
 725                                                 map,
 726                                                 map->buckets[-1-w[i]],
 727                                                 weight, weight_max,
 728                                                 x, numrep,
 729                                                 curstep->arg2,
 730                                                 o+osize, j,
 731                                                 choose_tries,
 732                                                 recurse_tries,
 733                                                 choose_local_tries,
 734                                                 choose_local_fallback_tries,
 735                                                 recurse_to_leaf,
 736                                                 c+osize);
 737                                 } else {
 738                                         crush_choose_indep(
 739                                                 map,
 740                                                 map->buckets[-1-w[i]],
 741                                                 weight, weight_max,
 742                                                 x, numrep, numrep,
 743                                                 curstep->arg2,
 744                                                 o+osize, j,
 745                                                 choose_tries,
 746                                                 choose_leaf_tries ?
 747                                                    choose_leaf_tries : 1,
 748                                                 recurse_to_leaf,
 749                                                 c+osize,
 750                                                 0);
 751                                         osize += numrep;
 752                                 }
 753                         }
 754
 755                         if (recurse_to_leaf)
 756                                 /* copy final _leaf_ values to output set */
 757                                 memcpy(o, c, osize*sizeof(*o));
 758
 759                         /* swap o and w arrays */
 760                         tmp = o;
 761                         o = w;
 762                         w = tmp;
 763                         wsize = osize;
 764                         break;
 765
 766
 767                 case CRUSH_RULE_EMIT:
 768                         for (i = 0; i < wsize && result_len < result_max; i++) {
 769                                 result[result_len] = w[i];
 770                                 result_len++;
 771                         }
 772                         wsize = 0;
 773                         break;
 774
 775                 default:
 776                         dprintk(" unknown op %d at step %d\n",
 777                                 curstep->op, step);
 778                         break;
 779                 }
 780         }
 781         return result_len;
 782 }
 783
 784