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