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 - 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  * @firstn: true if choosing "first n" items, false if choosing "indep"
 294  * @recurse_to_leaf: true if we want one device under each item of given type
 295  * @descend_once: true if we should only try one descent before giving up
 296  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 297  */
 298 static int crush_choose(const struct crush_map *map,
 299                         struct crush_bucket *bucket,
 300                         const __u32 *weight, int weight_max,
 301                         int x, int numrep, int type,
 302                         int *out, int outpos,
 303                         int firstn, int recurse_to_leaf,
 304                         int descend_once, int *out2)
 305 {
 306         int rep;
 307         unsigned int ftotal, flocal;
 308         int retry_descent, retry_bucket, skip_rep;
 309         struct crush_bucket *in = bucket;
 310         int r;
 311         int i;
 312         int item = 0;
 313         int itemtype;
 314         int collide, reject;
 315
 316         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 317                 bucket->id, x, outpos, numrep);
 318
 319         for (rep = outpos; rep < numrep; rep++) {
 320                 /* keep trying until we get a non-out, non-colliding item */
 321                 ftotal = 0;
 322                 skip_rep = 0;
 323                 do {
 324                         retry_descent = 0;
 325                         in = bucket;               /* initial bucket */
 326
 327                         /* choose through intervening buckets */
 328                         flocal = 0;
 329                         do {
 330                                 collide = 0;
 331                                 retry_bucket = 0;
 332                                 r = rep;
 333                                 if (in->alg == CRUSH_BUCKET_UNIFORM) {
 334                                         /* be careful */
 335                                         if (firstn || (__u32)numrep >= in->size)
 336                                                 /* r' = r + f_total */
 337                                                 r += ftotal;
 338                                         else if (in->size % numrep == 0)
 339                                                 /* r'=r+(n+1)*f_local */
 340                                                 r += (numrep+1) *
 341                                                         (flocal+ftotal);
 342                                         else
 343                                                 /* r' = r + n*f_local */
 344                                                 r += numrep * (flocal+ftotal);
 345                                 } else {
 346                                         if (firstn)
 347                                                 /* r' = r + f_total */
 348                                                 r += ftotal;
 349                                         else
 350                                                 /* r' = r + n*f_local */
 351                                                 r += numrep * (flocal+ftotal);
 352                                 }
 353
 354                                 /* bucket choose */
 355                                 if (in->size == 0) {
 356                                         reject = 1;
 357                                         goto reject;
 358                                 }
 359                                 if (map->choose_local_fallback_tries > 0 &&
 360                                     flocal >= (in->size>>1) &&
 361                                     flocal > map->choose_local_fallback_tries)
 362                                         item = bucket_perm_choose(in, x, r);
 363                                 else
 364                                         item = crush_bucket_choose(in, x, r);
 365                                 if (item >= map->max_devices) {
 366                                         dprintk("   bad item %d\n", item);
 367                                         skip_rep = 1;
 368                                         break;
 369                                 }
 370
 371                                 /* desired type? */
 372                                 if (item < 0)
 373                                         itemtype = map->buckets[-1-item]->type;
 374                                 else
 375                                         itemtype = 0;
 376                                 dprintk("  item %d type %d\n", item, itemtype);
 377
 378                                 /* keep going? */
 379                                 if (itemtype != type) {
 380                                         if (item >= 0 ||
 381                                             (-1-item) >= map->max_buckets) {
 382                                                 dprintk("   bad item type %d\n", type);
 383                                                 skip_rep = 1;
 384                                                 break;
 385                                         }
 386                                         in = map->buckets[-1-item];
 387                                         retry_bucket = 1;
 388                                         continue;
 389                                 }
 390
 391                                 /* collision? */
 392                                 for (i = 0; i < outpos; i++) {
 393                                         if (out[i] == item) {
 394                                                 collide = 1;
 395                                                 break;
 396                                         }
 397                                 }
 398
 399                                 reject = 0;
 400                                 if (!collide && recurse_to_leaf) {
 401                                         if (item < 0) {
 402                                                 if (crush_choose(map,
 403                                                          map->buckets[-1-item],
 404                                                          weight, weight_max,
 405                                                          x, outpos+1, 0,
 406                                                          out2, outpos,
 407                                                          firstn, 0,
 408                                                          map->chooseleaf_descend_once,
 409                                                          NULL) <= outpos)
 410                                                         /* didn't get leaf */
 411                                                         reject = 1;
 412                                         } else {
 413                                                 /* we already have a leaf! */
 414                                                 out2[outpos] = item;
 415                                         }
 416                                 }
 417
 418                                 if (!reject) {
 419                                         /* out? */
 420                                         if (itemtype == 0)
 421                                                 reject = is_out(map, weight,
 422                                                                 weight_max,
 423                                                                 item, x);
 424                                         else
 425                                                 reject = 0;
 426                                 }
 427
 428 reject:
 429                                 if (reject || collide) {
 430                                         ftotal++;
 431                                         flocal++;
 432
 433                                         if (reject && descend_once)
 434                                                 /* let outer call try again */
 435                                                 skip_rep = 1;
 436                                         else if (collide && flocal <= map->choose_local_tries)
 437                                                 /* retry locally a few times */
 438                                                 retry_bucket = 1;
 439                                         else if (map->choose_local_fallback_tries > 0 &&
 440                                                  flocal <= in->size + map->choose_local_fallback_tries)
 441                                                 /* exhaustive bucket search */
 442                                                 retry_bucket = 1;
 443                                         else if (ftotal <= map->choose_total_tries)
 444                                                 /* then retry descent */
 445                                                 retry_descent = 1;
 446                                         else
 447                                                 /* else give up */
 448                                                 skip_rep = 1;
 449                                         dprintk("  reject %d  collide %d  "
 450                                                 "ftotal %u  flocal %u\n",
 451                                                 reject, collide, ftotal,
 452                                                 flocal);
 453                                 }
 454                         } while (retry_bucket);
 455                 } while (retry_descent);
 456
 457                 if (skip_rep) {
 458                         if (firstn) {
 459                                 dprintk("skip rep\n");
 460                                 continue;
 461                         }
 462                         dprintk("undef rep, continuing\n");
 463                         item = CRUSH_ITEM_UNDEF;
 464                 }
 465
 466                 dprintk("CHOOSE got %d\n", item);
 467                 out[outpos] = item;
 468                 outpos++;
 469         }
 470
 471         dprintk("CHOOSE returns %d\n", outpos);
 472         return outpos;
 473 }
 474
 475
 476 /**
 477  * choose indep: alternative breadth-first positionally stable mapping
 478  *
 479  */
 480 static void crush_choose_indep(const struct crush_map *map,
 481                               struct crush_bucket *bucket,
 482                               const __u32 *weight, int weight_max,
 483                               int x, int numrep, int type,
 484                               int *out, int outpos,
 485                               int recurse_to_leaf,
 486                               int *out2)
 487 {
 488         struct crush_bucket *in = bucket;
 489         int left = numrep - outpos;
 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 < numrep; 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 < map->choose_total_tries; ftotal++) {
 509                 for (rep = outpos; rep < numrep; 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;
 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 < numrep; 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, rep+1, 0,
 592                                                                    out2, rep,
 593                                                                    0, NULL);
 594                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 595                                                         /* placed nothing; no leaf */
 596                                                         break;
 597                                                 }
 598                                         } else {
 599                                                 /* we already have a leaf! */
 600                                                 out2[rep] = item;
 601                                         }
 602                                 }
 603
 604                                 /* out? */
 605                                 if (itemtype == 0 &&
 606                                     is_out(map, weight, weight_max, item, x))
 607                                         break;
 608
 609                                 /* yay! */
 610                                 out[rep] = item;
 611                                 left--;
 612                                 break;
 613                         }
 614                 }
 615         }
 616         for (rep = outpos; rep < numrep; rep++) {
 617                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 618                         out[rep] = CRUSH_ITEM_NONE;
 619                 }
 620                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 621                         out2[rep] = CRUSH_ITEM_NONE;
 622                 }
 623         }
 624 }
 625
 626 /**
 627  * crush_do_rule - calculate a mapping with the given input and rule
 628  * @map: the crush_map
 629  * @ruleno: the rule id
 630  * @x: hash input
 631  * @result: pointer to result vector
 632  * @result_max: maximum result size
 633  * @weight: weight vector (for map leaves)
 634  * @weight_max: size of weight vector
 635  * @scratch: scratch vector for private use; must be >= 3 * result_max
 636  */
 637 int crush_do_rule(const struct crush_map *map,
 638                   int ruleno, int x, int *result, int result_max,
 639                   const __u32 *weight, int weight_max,
 640                   int *scratch)
 641 {
 642         int result_len;
 643         int *a = scratch;
 644         int *b = scratch + result_max;
 645         int *c = scratch + result_max*2;
 646         int recurse_to_leaf;
 647         int *w;
 648         int wsize = 0;
 649         int *o;
 650         int osize;
 651         int *tmp;
 652         struct crush_rule *rule;
 653         __u32 step;
 654         int i, j;
 655         int numrep;
 656         const int descend_once = 0;
 657
 658         if ((__u32)ruleno >= map->max_rules) {
 659                 dprintk(" bad ruleno %d\n", ruleno);
 660                 return 0;
 661         }
 662
 663         rule = map->rules[ruleno];
 664         result_len = 0;
 665         w = a;
 666         o = b;
 667
 668         for (step = 0; step < rule->len; step++) {
 669                 int firstn = 0;
 670                 struct crush_rule_step *curstep = &rule->steps[step];
 671
 672                 switch (curstep->op) {
 673                 case CRUSH_RULE_TAKE:
 674                         w[0] = curstep->arg1;
 675                         wsize = 1;
 676                         break;
 677
 678                 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
 679                 case CRUSH_RULE_CHOOSE_FIRSTN:
 680                         firstn = 1;
 681                         /* fall through */
 682                 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
 683                 case CRUSH_RULE_CHOOSE_INDEP:
 684                         if (wsize == 0)
 685                                 break;
 686
 687                         recurse_to_leaf =
 688                                 curstep->op ==
 689                                  CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
 690                                 curstep->op ==
 691                                 CRUSH_RULE_CHOOSE_LEAF_INDEP;
 692
 693                         /* reset output */
 694                         osize = 0;
 695
 696                         for (i = 0; i < wsize; i++) {
 697                                 /*
 698                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 699                                  * basically, numrep <= 0 means relative to
 700                                  * the provided result_max
 701                                  */
 702                                 numrep = curstep->arg1;
 703                                 if (numrep <= 0) {
 704                                         numrep += result_max;
 705                                         if (numrep <= 0)
 706                                                 continue;
 707                                 }
 708                                 j = 0;
 709                                 if (firstn) {
 710                                         osize += crush_choose(map,
 711                                                               map->buckets[-1-w[i]],
 712                                                               weight, weight_max,
 713                                                               x, numrep,
 714                                                               curstep->arg2,
 715                                                               o+osize, j,
 716                                                               firstn,
 717                                                               recurse_to_leaf,
 718                                                               descend_once, c+osize);
 719                                 } else {
 720                                         crush_choose_indep(map,
 721                                                            map->buckets[-1-w[i]],
 722                                                            weight, weight_max,
 723                                                            x, numrep,
 724                                                            curstep->arg2,
 725                                                            o+osize, j,
 726                                                            recurse_to_leaf,
 727                                                            c+osize);
 728                                         osize += numrep;
 729                                 }
 730                         }
 731
 732                         if (recurse_to_leaf)
 733                                 /* copy final _leaf_ values to output set */
 734                                 memcpy(o, c, osize*sizeof(*o));
 735
 736                         /* swap o and w arrays */
 737                         tmp = o;
 738                         o = w;
 739                         w = tmp;
 740                         wsize = osize;
 741                         break;
 742
 743
 744                 case CRUSH_RULE_EMIT:
 745                         for (i = 0; i < wsize && result_len < result_max; i++) {
 746                                 result[result_len] = w[i];
 747                                 result_len++;
 748                         }
 749                         wsize = 0;
 750                         break;
 751
 752                 default:
 753                         dprintk(" unknown op %d at step %d\n",
 754                                 curstep->op, step);
 755                         break;
 756                 }
 757         }
 758         return result_len;
 759 }
 760
 761