net/ceph/crush/mapper.c

   1 /*
   2  * Ceph - scalable distributed file system
   3  *
   4  * Copyright (C) 2015 Intel Corporation All Rights Reserved
   5  *
   6  * This is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License version 2.1, as published by the Free Software
   9  * Foundation.  See file COPYING.
  10  *
  11  */
  12
  13 #ifdef __KERNEL__
  14 # include <linux/string.h>
  15 # include <linux/slab.h>
  16 # include <linux/bug.h>
  17 # include <linux/kernel.h>
  18 # include <linux/crush/crush.h>
  19 # include <linux/crush/hash.h>
  20 # include <linux/crush/mapper.h>
  21 #else
  22 # include "crush_compat.h"
  23 # include "crush.h"
  24 # include "hash.h"
  25 # include "mapper.h"
  26 #endif
  27 #include "crush_ln_table.h"
  28
  29 #define dprintk(args...) /* printf(args) */
  30
  31 /*
  32  * Implement the core CRUSH mapping algorithm.
  33  */
  34
  35 /**
  36  * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
  37  * @map: the crush_map
  38  * @ruleset: the storage ruleset id (user defined)
  39  * @type: storage ruleset type (user defined)
  40  * @size: output set size
  41  */
  42 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
  43 {
  44         __u32 i;
  45
  46         for (i = 0; i < map->max_rules; i++) {
  47                 if (map->rules[i] &&
  48                     map->rules[i]->mask.ruleset == ruleset &&
  49                     map->rules[i]->mask.type == type &&
  50                     map->rules[i]->mask.min_size <= size &&
  51                     map->rules[i]->mask.max_size >= size)
  52                         return i;
  53         }
  54         return -1;
  55 }
  56
  57
  58 /*
  59  * bucket choose methods
  60  *
  61  * For each bucket algorithm, we have a "choose" method that, given a
  62  * crush input @x and replica position (usually, position in output set) @r,
  63  * will produce an item in the bucket.
  64  */
  65
  66 /*
  67  * Choose based on a random permutation of the bucket.
  68  *
  69  * We used to use some prime number arithmetic to do this, but it
  70  * wasn't very random, and had some other bad behaviors.  Instead, we
  71  * calculate an actual random permutation of the bucket members.
  72  * Since this is expensive, we optimize for the r=0 case, which
  73  * captures the vast majority of calls.
  74  */
  75 static int bucket_perm_choose(struct crush_bucket *bucket,
  76                               int x, int r)
  77 {
  78         unsigned int pr = r % bucket->size;
  79         unsigned int i, s;
  80
  81         /* start a new permutation if @x has changed */
  82         if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) {
  83                 dprintk("bucket %d new x=%d\n", bucket->id, x);
  84                 bucket->perm_x = x;
  85
  86                 /* optimize common r=0 case */
  87                 if (pr == 0) {
  88                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
  89                                 bucket->size;
  90                         bucket->perm[0] = s;
  91                         bucket->perm_n = 0xffff;   /* magic value, see below */
  92                         goto out;
  93                 }
  94
  95                 for (i = 0; i < bucket->size; i++)
  96                         bucket->perm[i] = i;
  97                 bucket->perm_n = 0;
  98         } else if (bucket->perm_n == 0xffff) {
  99                 /* clean up after the r=0 case above */
 100                 for (i = 1; i < bucket->size; i++)
 101                         bucket->perm[i] = i;
 102                 bucket->perm[bucket->perm[0]] = 0;
 103                 bucket->perm_n = 1;
 104         }
 105
 106         /* calculate permutation up to pr */
 107         for (i = 0; i < bucket->perm_n; i++)
 108                 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]);
 109         while (bucket->perm_n <= pr) {
 110                 unsigned int p = bucket->perm_n;
 111                 /* no point in swapping the final entry */
 112                 if (p < bucket->size - 1) {
 113                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
 114                                 (bucket->size - p);
 115                         if (i) {
 116                                 unsigned int t = bucket->perm[p + i];
 117                                 bucket->perm[p + i] = bucket->perm[p];
 118                                 bucket->perm[p] = t;
 119                         }
 120                         dprintk(" perm_choose swap %d with %d\n", p, p+i);
 121                 }
 122                 bucket->perm_n++;
 123         }
 124         for (i = 0; i < bucket->size; i++)
 125                 dprintk(" perm_choose  %d: %d\n", i, bucket->perm[i]);
 126
 127         s = bucket->perm[pr];
 128 out:
 129         dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
 130                 bucket->size, x, r, pr, s);
 131         return bucket->items[s];
 132 }
 133
 134 /* uniform */
 135 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
 136                                  int x, int r)
 137 {
 138         return bucket_perm_choose(&bucket->h, x, r);
 139 }
 140
 141 /* list */
 142 static int bucket_list_choose(struct crush_bucket_list *bucket,
 143                               int x, int r)
 144 {
 145         int i;
 146
 147         for (i = bucket->h.size-1; i >= 0; i--) {
 148                 __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i],
 149                                          r, bucket->h.id);
 150                 w &= 0xffff;
 151                 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
 152                         "sw %x rand %llx",
 153                         i, x, r, bucket->h.items[i], bucket->item_weights[i],
 154                         bucket->sum_weights[i], w);
 155                 w *= bucket->sum_weights[i];
 156                 w = w >> 16;
 157                 /*dprintk(" scaled %llx\n", w);*/
 158                 if (w < bucket->item_weights[i])
 159                         return bucket->h.items[i];
 160         }
 161
 162         dprintk("bad list sums for bucket %d\n", bucket->h.id);
 163         return bucket->h.items[0];
 164 }
 165
 166
 167 /* (binary) tree */
 168 static int height(int n)
 169 {
 170         int h = 0;
 171         while ((n & 1) == 0) {
 172                 h++;
 173                 n = n >> 1;
 174         }
 175         return h;
 176 }
 177
 178 static int left(int x)
 179 {
 180         int h = height(x);
 181         return x - (1 << (h-1));
 182 }
 183
 184 static int right(int x)
 185 {
 186         int h = height(x);
 187         return x + (1 << (h-1));
 188 }
 189
 190 static int terminal(int x)
 191 {
 192         return x & 1;
 193 }
 194
 195 static int bucket_tree_choose(struct crush_bucket_tree *bucket,
 196                               int x, int r)
 197 {
 198         int n;
 199         __u32 w;
 200         __u64 t;
 201
 202         /* start at root */
 203         n = bucket->num_nodes >> 1;
 204
 205         while (!terminal(n)) {
 206                 int l;
 207                 /* pick point in [0, w) */
 208                 w = bucket->node_weights[n];
 209                 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
 210                                           bucket->h.id) * (__u64)w;
 211                 t = t >> 32;
 212
 213                 /* descend to the left or right? */
 214                 l = left(n);
 215                 if (t < bucket->node_weights[l])
 216                         n = l;
 217                 else
 218                         n = right(n);
 219         }
 220
 221         return bucket->h.items[n >> 1];
 222 }
 223
 224
 225 /* straw */
 226
 227 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
 228                                int x, int r)
 229 {
 230         __u32 i;
 231         int high = 0;
 232         __u64 high_draw = 0;
 233         __u64 draw;
 234
 235         for (i = 0; i < bucket->h.size; i++) {
 236                 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
 237                 draw &= 0xffff;
 238                 draw *= bucket->straws[i];
 239                 if (i == 0 || draw > high_draw) {
 240                         high = i;
 241                         high_draw = draw;
 242                 }
 243         }
 244         return bucket->h.items[high];
 245 }
 246
 247 /* compute 2^44*log2(input+1) */
 248 static __u64 crush_ln(unsigned int xin)
 249 {
 250         unsigned int x = xin;
 251         int iexpon, index1, index2;
 252         __u64 RH, LH, LL, xl64, result;
 253
 254         x++;
 255
 256         /* normalize input */
 257         iexpon = 15;
 258
 259         /*
 260          * figure out number of bits we need to shift and
 261          * do it in one step instead of iteratively
 262          */
 263         if (!(x & 0x18000)) {
 264                 int bits = __builtin_clz(x & 0x1FFFF) - 16;
 265                 x <<= bits;
 266                 iexpon = 15 - bits;
 267         }
 268
 269         index1 = (x >> 8) << 1;
 270         /* RH ~ 2^56/index1 */
 271         RH = __RH_LH_tbl[index1 - 256];
 272         /* LH ~ 2^48 * log2(index1/256) */
 273         LH = __RH_LH_tbl[index1 + 1 - 256];
 274
 275         /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */
 276         xl64 = (__s64)x * RH;
 277         xl64 >>= 48;
 278
 279         result = iexpon;
 280         result <<= (12 + 32);
 281
 282         index2 = xl64 & 0xff;
 283         /* LL ~ 2^48*log2(1.0+index2/2^15) */
 284         LL = __LL_tbl[index2];
 285
 286         LH = LH + LL;
 287
 288         LH >>= (48 - 12 - 32);
 289         result += LH;
 290
 291         return result;
 292 }
 293
 294
 295 /*
 296  * straw2
 297  *
 298  * for reference, see:
 299  *
 300  * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables
 301  *
 302  */
 303
 304 static int bucket_straw2_choose(struct crush_bucket_straw2 *bucket,
 305                                 int x, int r)
 306 {
 307         unsigned int i, high = 0;
 308         unsigned int u;
 309         unsigned int w;
 310         __s64 ln, draw, high_draw = 0;
 311
 312         for (i = 0; i < bucket->h.size; i++) {
 313                 w = bucket->item_weights[i];
 314                 if (w) {
 315                         u = crush_hash32_3(bucket->h.hash, x,
 316                                            bucket->h.items[i], r);
 317                         u &= 0xffff;
 318
 319                         /*
 320                          * for some reason slightly less than 0x10000 produces
 321                          * a slightly more accurate distribution... probably a
 322                          * rounding effect.
 323                          *
 324                          * the natural log lookup table maps [0,0xffff]
 325                          * (corresponding to real numbers [1/0x10000, 1] to
 326                          * [0, 0xffffffffffff] (corresponding to real numbers
 327                          * [-11.090355,0]).
 328                          */
 329                         ln = crush_ln(u) - 0x1000000000000ll;
 330
 331                         /*
 332                          * divide by 16.16 fixed-point weight.  note
 333                          * that the ln value is negative, so a larger
 334                          * weight means a larger (less negative) value
 335                          * for draw.
 336                          */
 337                         draw = div64_s64(ln, w);
 338                 } else {
 339                         draw = S64_MIN;
 340                 }
 341
 342                 if (i == 0 || draw > high_draw) {
 343                         high = i;
 344                         high_draw = draw;
 345                 }
 346         }
 347         return bucket->h.items[high];
 348 }
 349
 350
 351 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
 352 {
 353         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
 354         BUG_ON(in->size == 0);
 355         switch (in->alg) {
 356         case CRUSH_BUCKET_UNIFORM:
 357                 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
 358                                           x, r);
 359         case CRUSH_BUCKET_LIST:
 360                 return bucket_list_choose((struct crush_bucket_list *)in,
 361                                           x, r);
 362         case CRUSH_BUCKET_TREE:
 363                 return bucket_tree_choose((struct crush_bucket_tree *)in,
 364                                           x, r);
 365         case CRUSH_BUCKET_STRAW:
 366                 return bucket_straw_choose((struct crush_bucket_straw *)in,
 367                                            x, r);
 368         case CRUSH_BUCKET_STRAW2:
 369                 return bucket_straw2_choose((struct crush_bucket_straw2 *)in,
 370                                             x, r);
 371         default:
 372                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
 373                 return in->items[0];
 374         }
 375 }
 376
 377
 378 /*
 379  * true if device is marked "out" (failed, fully offloaded)
 380  * of the cluster
 381  */
 382 static int is_out(const struct crush_map *map,
 383                   const __u32 *weight, int weight_max,
 384                   int item, int x)
 385 {
 386         if (item >= weight_max)
 387                 return 1;
 388         if (weight[item] >= 0x10000)
 389                 return 0;
 390         if (weight[item] == 0)
 391                 return 1;
 392         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
 393             < weight[item])
 394                 return 0;
 395         return 1;
 396 }
 397
 398 /**
 399  * crush_choose_firstn - choose numrep distinct items of given type
 400  * @map: the crush_map
 401  * @bucket: the bucket we are choose an item from
 402  * @x: crush input value
 403  * @numrep: the number of items to choose
 404  * @type: the type of item to choose
 405  * @out: pointer to output vector
 406  * @outpos: our position in that vector
 407  * @out_size: size of the out vector
 408  * @tries: number of attempts to make
 409  * @recurse_tries: number of attempts to have recursive chooseleaf make
 410  * @local_retries: localized retries
 411  * @local_fallback_retries: localized fallback retries
 412  * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose)
 413  * @stable: stable mode starts rep=0 in the recursive call for all replicas
 414  * @vary_r: pass r to recursive calls
 415  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 416  * @parent_r: r value passed from the parent
 417  */
 418 static int crush_choose_firstn(const struct crush_map *map,
 419                                struct crush_bucket *bucket,
 420                                const __u32 *weight, int weight_max,
 421                                int x, int numrep, int type,
 422                                int *out, int outpos,
 423                                int out_size,
 424                                unsigned int tries,
 425                                unsigned int recurse_tries,
 426                                unsigned int local_retries,
 427                                unsigned int local_fallback_retries,
 428                                int recurse_to_leaf,
 429                                unsigned int vary_r,
 430                                unsigned int stable,
 431                                int *out2,
 432                                int parent_r)
 433 {
 434         int rep;
 435         unsigned int ftotal, flocal;
 436         int retry_descent, retry_bucket, skip_rep;
 437         struct crush_bucket *in = bucket;
 438         int r;
 439         int i;
 440         int item = 0;
 441         int itemtype;
 442         int collide, reject;
 443         int count = out_size;
 444
 445         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d tries %d recurse_tries %d local_retries %d local_fallback_retries %d parent_r %d stable %d\n",
 446                 recurse_to_leaf ? "_LEAF" : "",
 447                 bucket->id, x, outpos, numrep,
 448                 tries, recurse_tries, local_retries, local_fallback_retries,
 449                 parent_r, stable);
 450
 451         for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) {
 452                 /* keep trying until we get a non-out, non-colliding item */
 453                 ftotal = 0;
 454                 skip_rep = 0;
 455                 do {
 456                         retry_descent = 0;
 457                         in = bucket;               /* initial bucket */
 458
 459                         /* choose through intervening buckets */
 460                         flocal = 0;
 461                         do {
 462                                 collide = 0;
 463                                 retry_bucket = 0;
 464                                 r = rep + parent_r;
 465                                 /* r' = r + f_total */
 466                                 r += ftotal;
 467
 468                                 /* bucket choose */
 469                                 if (in->size == 0) {
 470                                         reject = 1;
 471                                         goto reject;
 472                                 }
 473                                 if (local_fallback_retries > 0 &&
 474                                     flocal >= (in->size>>1) &&
 475                                     flocal > local_fallback_retries)
 476                                         item = bucket_perm_choose(in, x, r);
 477                                 else
 478                                         item = crush_bucket_choose(in, x, r);
 479                                 if (item >= map->max_devices) {
 480                                         dprintk("   bad item %d\n", item);
 481                                         skip_rep = 1;
 482                                         break;
 483                                 }
 484
 485                                 /* desired type? */
 486                                 if (item < 0)
 487                                         itemtype = map->buckets[-1-item]->type;
 488                                 else
 489                                         itemtype = 0;
 490                                 dprintk("  item %d type %d\n", item, itemtype);
 491
 492                                 /* keep going? */
 493                                 if (itemtype != type) {
 494                                         if (item >= 0 ||
 495                                             (-1-item) >= map->max_buckets) {
 496                                                 dprintk("   bad item type %d\n", type);
 497                                                 skip_rep = 1;
 498                                                 break;
 499                                         }
 500                                         in = map->buckets[-1-item];
 501                                         retry_bucket = 1;
 502                                         continue;
 503                                 }
 504
 505                                 /* collision? */
 506                                 for (i = 0; i < outpos; i++) {
 507                                         if (out[i] == item) {
 508                                                 collide = 1;
 509                                                 break;
 510                                         }
 511                                 }
 512
 513                                 reject = 0;
 514                                 if (!collide && recurse_to_leaf) {
 515                                         if (item < 0) {
 516                                                 int sub_r;
 517                                                 if (vary_r)
 518                                                         sub_r = r >> (vary_r-1);
 519                                                 else
 520                                                         sub_r = 0;
 521                                                 if (crush_choose_firstn(map,
 522                                                          map->buckets[-1-item],
 523                                                          weight, weight_max,
 524                                                          x, stable ? 1 : outpos+1, 0,
 525                                                          out2, outpos, count,
 526                                                          recurse_tries, 0,
 527                                                          local_retries,
 528                                                          local_fallback_retries,
 529                                                          0,
 530                                                          vary_r,
 531                                                          stable,
 532                                                          NULL,
 533                                                          sub_r) <= outpos)
 534                                                         /* didn't get leaf */
 535                                                         reject = 1;
 536                                         } else {
 537                                                 /* we already have a leaf! */
 538                                                 out2[outpos] = item;
 539                                         }
 540                                 }
 541
 542                                 if (!reject) {
 543                                         /* out? */
 544                                         if (itemtype == 0)
 545                                                 reject = is_out(map, weight,
 546                                                                 weight_max,
 547                                                                 item, x);
 548                                         else
 549                                                 reject = 0;
 550                                 }
 551
 552 reject:
 553                                 if (reject || collide) {
 554                                         ftotal++;
 555                                         flocal++;
 556
 557                                         if (collide && flocal <= local_retries)
 558                                                 /* retry locally a few times */
 559                                                 retry_bucket = 1;
 560                                         else if (local_fallback_retries > 0 &&
 561                                                  flocal <= in->size + local_fallback_retries)
 562                                                 /* exhaustive bucket search */
 563                                                 retry_bucket = 1;
 564                                         else if (ftotal < tries)
 565                                                 /* then retry descent */
 566                                                 retry_descent = 1;
 567                                         else
 568                                                 /* else give up */
 569                                                 skip_rep = 1;
 570                                         dprintk("  reject %d  collide %d  "
 571                                                 "ftotal %u  flocal %u\n",
 572                                                 reject, collide, ftotal,
 573                                                 flocal);
 574                                 }
 575                         } while (retry_bucket);
 576                 } while (retry_descent);
 577
 578                 if (skip_rep) {
 579                         dprintk("skip rep\n");
 580                         continue;
 581                 }
 582
 583                 dprintk("CHOOSE got %d\n", item);
 584                 out[outpos] = item;
 585                 outpos++;
 586                 count--;
 587 #ifndef __KERNEL__
 588                 if (map->choose_tries && ftotal <= map->choose_total_tries)
 589                         map->choose_tries[ftotal]++;
 590 #endif
 591         }
 592
 593         dprintk("CHOOSE returns %d\n", outpos);
 594         return outpos;
 595 }
 596
 597
 598 /**
 599  * crush_choose_indep: alternative breadth-first positionally stable mapping
 600  *
 601  */
 602 static void crush_choose_indep(const struct crush_map *map,
 603                                struct crush_bucket *bucket,
 604                                const __u32 *weight, int weight_max,
 605                                int x, int left, int numrep, int type,
 606                                int *out, int outpos,
 607                                unsigned int tries,
 608                                unsigned int recurse_tries,
 609                                int recurse_to_leaf,
 610                                int *out2,
 611                                int parent_r)
 612 {
 613         struct crush_bucket *in = bucket;
 614         int endpos = outpos + left;
 615         int rep;
 616         unsigned int ftotal;
 617         int r;
 618         int i;
 619         int item = 0;
 620         int itemtype;
 621         int collide;
 622
 623         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 624                 bucket->id, x, outpos, numrep);
 625
 626         /* initially my result is undefined */
 627         for (rep = outpos; rep < endpos; rep++) {
 628                 out[rep] = CRUSH_ITEM_UNDEF;
 629                 if (out2)
 630                         out2[rep] = CRUSH_ITEM_UNDEF;
 631         }
 632
 633         for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) {
 634 #ifdef DEBUG_INDEP
 635                 if (out2 && ftotal) {
 636                         dprintk("%u %d a: ", ftotal, left);
 637                         for (rep = outpos; rep < endpos; rep++) {
 638                                 dprintk(" %d", out[rep]);
 639                         }
 640                         dprintk("\n");
 641                         dprintk("%u %d b: ", ftotal, left);
 642                         for (rep = outpos; rep < endpos; rep++) {
 643                                 dprintk(" %d", out2[rep]);
 644                         }
 645                         dprintk("\n");
 646                 }
 647 #endif
 648                 for (rep = outpos; rep < endpos; rep++) {
 649                         if (out[rep] != CRUSH_ITEM_UNDEF)
 650                                 continue;
 651
 652                         in = bucket;  /* initial bucket */
 653
 654                         /* choose through intervening buckets */
 655                         for (;;) {
 656                                 /* note: we base the choice on the position
 657                                  * even in the nested call.  that means that
 658                                  * if the first layer chooses the same bucket
 659                                  * in a different position, we will tend to
 660                                  * choose a different item in that bucket.
 661                                  * this will involve more devices in data
 662                                  * movement and tend to distribute the load.
 663                                  */
 664                                 r = rep + parent_r;
 665
 666                                 /* be careful */
 667                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
 668                                     in->size % numrep == 0)
 669                                         /* r'=r+(n+1)*f_total */
 670                                         r += (numrep+1) * ftotal;
 671                                 else
 672                                         /* r' = r + n*f_total */
 673                                         r += numrep * ftotal;
 674
 675                                 /* bucket choose */
 676                                 if (in->size == 0) {
 677                                         dprintk("   empty bucket\n");
 678                                         break;
 679                                 }
 680
 681                                 item = crush_bucket_choose(in, x, r);
 682                                 if (item >= map->max_devices) {
 683                                         dprintk("   bad item %d\n", item);
 684                                         out[rep] = CRUSH_ITEM_NONE;
 685                                         if (out2)
 686                                                 out2[rep] = CRUSH_ITEM_NONE;
 687                                         left--;
 688                                         break;
 689                                 }
 690
 691                                 /* desired type? */
 692                                 if (item < 0)
 693                                         itemtype = map->buckets[-1-item]->type;
 694                                 else
 695                                         itemtype = 0;
 696                                 dprintk("  item %d type %d\n", item, itemtype);
 697
 698                                 /* keep going? */
 699                                 if (itemtype != type) {
 700                                         if (item >= 0 ||
 701                                             (-1-item) >= map->max_buckets) {
 702                                                 dprintk("   bad item type %d\n", type);
 703                                                 out[rep] = CRUSH_ITEM_NONE;
 704                                                 if (out2)
 705                                                         out2[rep] =
 706                                                                 CRUSH_ITEM_NONE;
 707                                                 left--;
 708                                                 break;
 709                                         }
 710                                         in = map->buckets[-1-item];
 711                                         continue;
 712                                 }
 713
 714                                 /* collision? */
 715                                 collide = 0;
 716                                 for (i = outpos; i < endpos; i++) {
 717                                         if (out[i] == item) {
 718                                                 collide = 1;
 719                                                 break;
 720                                         }
 721                                 }
 722                                 if (collide)
 723                                         break;
 724
 725                                 if (recurse_to_leaf) {
 726                                         if (item < 0) {
 727                                                 crush_choose_indep(map,
 728                                                    map->buckets[-1-item],
 729                                                    weight, weight_max,
 730                                                    x, 1, numrep, 0,
 731                                                    out2, rep,
 732                                                    recurse_tries, 0,
 733                                                    0, NULL, r);
 734                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 735                                                         /* placed nothing; no leaf */
 736                                                         break;
 737                                                 }
 738                                         } else {
 739                                                 /* we already have a leaf! */
 740                                                 out2[rep] = item;
 741                                         }
 742                                 }
 743
 744                                 /* out? */
 745                                 if (itemtype == 0 &&
 746                                     is_out(map, weight, weight_max, item, x))
 747                                         break;
 748
 749                                 /* yay! */
 750                                 out[rep] = item;
 751                                 left--;
 752                                 break;
 753                         }
 754                 }
 755         }
 756         for (rep = outpos; rep < endpos; rep++) {
 757                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 758                         out[rep] = CRUSH_ITEM_NONE;
 759                 }
 760                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 761                         out2[rep] = CRUSH_ITEM_NONE;
 762                 }
 763         }
 764 #ifndef __KERNEL__
 765         if (map->choose_tries && ftotal <= map->choose_total_tries)
 766                 map->choose_tries[ftotal]++;
 767 #endif
 768 #ifdef DEBUG_INDEP
 769         if (out2) {
 770                 dprintk("%u %d a: ", ftotal, left);
 771                 for (rep = outpos; rep < endpos; rep++) {
 772                         dprintk(" %d", out[rep]);
 773                 }
 774                 dprintk("\n");
 775                 dprintk("%u %d b: ", ftotal, left);
 776                 for (rep = outpos; rep < endpos; rep++) {
 777                         dprintk(" %d", out2[rep]);
 778                 }
 779                 dprintk("\n");
 780         }
 781 #endif
 782 }
 783
 784 /**
 785  * crush_do_rule - calculate a mapping with the given input and rule
 786  * @map: the crush_map
 787  * @ruleno: the rule id
 788  * @x: hash input
 789  * @result: pointer to result vector
 790  * @result_max: maximum result size
 791  * @weight: weight vector (for map leaves)
 792  * @weight_max: size of weight vector
 793  * @scratch: scratch vector for private use; must be >= 3 * result_max
 794  */
 795 int crush_do_rule(const struct crush_map *map,
 796                   int ruleno, int x, int *result, int result_max,
 797                   const __u32 *weight, int weight_max,
 798                   int *scratch)
 799 {
 800         int result_len;
 801         int *a = scratch;
 802         int *b = scratch + result_max;
 803         int *c = scratch + result_max*2;
 804         int recurse_to_leaf;
 805         int *w;
 806         int wsize = 0;
 807         int *o;
 808         int osize;
 809         int *tmp;
 810         struct crush_rule *rule;
 811         __u32 step;
 812         int i, j;
 813         int numrep;
 814         int out_size;
 815         /*
 816          * the original choose_total_tries value was off by one (it
 817          * counted "retries" and not "tries").  add one.
 818          */
 819         int choose_tries = map->choose_total_tries + 1;
 820         int choose_leaf_tries = 0;
 821         /*
 822          * the local tries values were counted as "retries", though,
 823          * and need no adjustment
 824          */
 825         int choose_local_retries = map->choose_local_tries;
 826         int choose_local_fallback_retries = map->choose_local_fallback_tries;
 827
 828         int vary_r = map->chooseleaf_vary_r;
 829         int stable = map->chooseleaf_stable;
 830
 831         if ((__u32)ruleno >= map->max_rules) {
 832                 dprintk(" bad ruleno %d\n", ruleno);
 833                 return 0;
 834         }
 835
 836         rule = map->rules[ruleno];
 837         result_len = 0;
 838         w = a;
 839         o = b;
 840
 841         for (step = 0; step < rule->len; step++) {
 842                 int firstn = 0;
 843                 struct crush_rule_step *curstep = &rule->steps[step];
 844
 845                 switch (curstep->op) {
 846                 case CRUSH_RULE_TAKE:
 847                         if ((curstep->arg1 >= 0 &&
 848                              curstep->arg1 < map->max_devices) ||
 849                             (-1-curstep->arg1 >= 0 &&
 850                              -1-curstep->arg1 < map->max_buckets &&
 851                              map->buckets[-1-curstep->arg1])) {
 852                                 w[0] = curstep->arg1;
 853                                 wsize = 1;
 854                         } else {
 855                                 dprintk(" bad take value %d\n", curstep->arg1);
 856                         }
 857                         break;
 858
 859                 case CRUSH_RULE_SET_CHOOSE_TRIES:
 860                         if (curstep->arg1 > 0)
 861                                 choose_tries = curstep->arg1;
 862                         break;
 863
 864                 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
 865                         if (curstep->arg1 > 0)
 866                                 choose_leaf_tries = curstep->arg1;
 867                         break;
 868
 869                 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
 870                         if (curstep->arg1 >= 0)
 871                                 choose_local_retries = curstep->arg1;
 872                         break;
 873
 874                 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
 875                         if (curstep->arg1 >= 0)
 876                                 choose_local_fallback_retries = curstep->arg1;
 877                         break;
 878
 879                 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R:
 880                         if (curstep->arg1 >= 0)
 881                                 vary_r = curstep->arg1;
 882                         break;
 883
 884                 case CRUSH_RULE_SET_CHOOSELEAF_STABLE:
 885                         if (curstep->arg1 >= 0)
 886                                 stable = curstep->arg1;
 887                         break;
 888
 889                 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
 890                 case CRUSH_RULE_CHOOSE_FIRSTN:
 891                         firstn = 1;
 892                         /* fall through */
 893                 case CRUSH_RULE_CHOOSELEAF_INDEP:
 894                 case CRUSH_RULE_CHOOSE_INDEP:
 895                         if (wsize == 0)
 896                                 break;
 897
 898                         recurse_to_leaf =
 899                                 curstep->op ==
 900                                  CRUSH_RULE_CHOOSELEAF_FIRSTN ||
 901                                 curstep->op ==
 902                                 CRUSH_RULE_CHOOSELEAF_INDEP;
 903
 904                         /* reset output */
 905                         osize = 0;
 906
 907                         for (i = 0; i < wsize; i++) {
 908                                 int bno;
 909                                 /*
 910                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 911                                  * basically, numrep <= 0 means relative to
 912                                  * the provided result_max
 913                                  */
 914                                 numrep = curstep->arg1;
 915                                 if (numrep <= 0) {
 916                                         numrep += result_max;
 917                                         if (numrep <= 0)
 918                                                 continue;
 919                                 }
 920                                 j = 0;
 921                                 /* make sure bucket id is valid */
 922                                 bno = -1 - w[i];
 923                                 if (bno < 0 || bno >= map->max_buckets) {
 924                                         /* w[i] is probably CRUSH_ITEM_NONE */
 925                                         dprintk("  bad w[i] %d\n", w[i]);
 926                                         continue;
 927                                 }
 928                                 if (firstn) {
 929                                         int recurse_tries;
 930                                         if (choose_leaf_tries)
 931                                                 recurse_tries =
 932                                                         choose_leaf_tries;
 933                                         else if (map->chooseleaf_descend_once)
 934                                                 recurse_tries = 1;
 935                                         else
 936                                                 recurse_tries = choose_tries;
 937                                         osize += crush_choose_firstn(
 938                                                 map,
 939                                                 map->buckets[bno],
 940                                                 weight, weight_max,
 941                                                 x, numrep,
 942                                                 curstep->arg2,
 943                                                 o+osize, j,
 944                                                 result_max-osize,
 945                                                 choose_tries,
 946                                                 recurse_tries,
 947                                                 choose_local_retries,
 948                                                 choose_local_fallback_retries,
 949                                                 recurse_to_leaf,
 950                                                 vary_r,
 951                                                 stable,
 952                                                 c+osize,
 953                                                 0);
 954                                 } else {
 955                                         out_size = ((numrep < (result_max-osize)) ?
 956                                                     numrep : (result_max-osize));
 957                                         crush_choose_indep(
 958                                                 map,
 959                                                 map->buckets[bno],
 960                                                 weight, weight_max,
 961                                                 x, out_size, numrep,
 962                                                 curstep->arg2,
 963                                                 o+osize, j,
 964                                                 choose_tries,
 965                                                 choose_leaf_tries ?
 966                                                    choose_leaf_tries : 1,
 967                                                 recurse_to_leaf,
 968                                                 c+osize,
 969                                                 0);
 970                                         osize += out_size;
 971                                 }
 972                         }
 973
 974                         if (recurse_to_leaf)
 975                                 /* copy final _leaf_ values to output set */
 976                                 memcpy(o, c, osize*sizeof(*o));
 977
 978                         /* swap o and w arrays */
 979                         tmp = o;
 980                         o = w;
 981                         w = tmp;
 982                         wsize = osize;
 983                         break;
 984
 985
 986                 case CRUSH_RULE_EMIT:
 987                         for (i = 0; i < wsize && result_len < result_max; i++) {
 988                                 result[result_len] = w[i];
 989                                 result_len++;
 990                         }
 991                         wsize = 0;
 992                         break;
 993
 994                 default:
 995                         dprintk(" unknown op %d at step %d\n",
 996                                 curstep->op, step);
 997                         break;
 998                 }
 999         }
1000         return result_len;
1001 }