src/ceph/src/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 #else
  21 # include "crush_compat.h"
  22 # include "crush.h"
  23 # include "hash.h"
  24 #endif
  25 #include "crush_ln_table.h"
  26 #include "mapper.h"
  27
  28 #define dprintk(args...) /* printf(args) */
  29
  30 /*
  31  * Implement the core CRUSH mapping algorithm.
  32  */
  33
  34 /**
  35  * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
  36  * @map: the crush_map
  37  * @ruleset: the storage ruleset id (user defined)
  38  * @type: storage ruleset type (user defined)
  39  * @size: output set size
  40  */
  41 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
  42 {
  43         __u32 i;
  44
  45         for (i = 0; i < map->max_rules; i++) {
  46                 if (map->rules[i] &&
  47                     map->rules[i]->mask.ruleset == ruleset &&
  48                     map->rules[i]->mask.type == type &&
  49                     map->rules[i]->mask.min_size <= size &&
  50                     map->rules[i]->mask.max_size >= size)
  51                         return i;
  52         }
  53         return -1;
  54 }
  55
  56 /*
  57  * bucket choose methods
  58  *
  59  * For each bucket algorithm, we have a "choose" method that, given a
  60  * crush input @x and replica position (usually, position in output set) @r,
  61  * will produce an item in the bucket.
  62  */
  63
  64 /*
  65  * Choose based on a random permutation of the bucket.
  66  *
  67  * We used to use some prime number arithmetic to do this, but it
  68  * wasn't very random, and had some other bad behaviors.  Instead, we
  69  * calculate an actual random permutation of the bucket members.
  70  * Since this is expensive, we optimize for the r=0 case, which
  71  * captures the vast majority of calls.
  72  */
  73 static int bucket_perm_choose(const struct crush_bucket *bucket,
  74                               struct crush_work_bucket *work,
  75                               int x, int r)
  76 {
  77         unsigned int pr = r % bucket->size;
  78         unsigned int i, s;
  79
  80         /* start a new permutation if @x has changed */
  81         if (work->perm_x != (__u32)x || work->perm_n == 0) {
  82                 dprintk("bucket %d new x=%d\n", bucket->id, x);
  83                 work->perm_x = x;
  84
  85                 /* optimize common r=0 case */
  86                 if (pr == 0) {
  87                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
  88                                 bucket->size;
  89                         work->perm[0] = s;
  90                         work->perm_n = 0xffff;   /* magic value, see below */
  91                         goto out;
  92                 }
  93
  94                 for (i = 0; i < bucket->size; i++)
  95                         work->perm[i] = i;
  96                 work->perm_n = 0;
  97         } else if (work->perm_n == 0xffff) {
  98                 /* clean up after the r=0 case above */
  99                 for (i = 1; i < bucket->size; i++)
 100                         work->perm[i] = i;
 101                 work->perm[work->perm[0]] = 0;
 102                 work->perm_n = 1;
 103         }
 104
 105         /* calculate permutation up to pr */
 106         for (i = 0; i < work->perm_n; i++)
 107                 dprintk(" perm_choose have %d: %d\n", i, work->perm[i]);
 108         while (work->perm_n <= pr) {
 109                 unsigned int p = work->perm_n;
 110                 /* no point in swapping the final entry */
 111                 if (p < bucket->size - 1) {
 112                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
 113                                 (bucket->size - p);
 114                         if (i) {
 115                                 unsigned int t = work->perm[p + i];
 116                                 work->perm[p + i] = work->perm[p];
 117                                 work->perm[p] = t;
 118                         }
 119                         dprintk(" perm_choose swap %d with %d\n", p, p+i);
 120                 }
 121                 work->perm_n++;
 122         }
 123         for (i = 0; i < bucket->size; i++)
 124                 dprintk(" perm_choose  %d: %d\n", i, work->perm[i]);
 125
 126         s = work->perm[pr];
 127 out:
 128         dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
 129                 bucket->size, x, r, pr, s);
 130         return bucket->items[s];
 131 }
 132
 133 /* uniform */
 134 static int bucket_uniform_choose(const struct crush_bucket_uniform *bucket,
 135                                  struct crush_work_bucket *work, int x, int r)
 136 {
 137         return bucket_perm_choose(&bucket->h, work, x, r);
 138 }
 139
 140 /* list */
 141 static int bucket_list_choose(const struct crush_bucket_list *bucket,
 142                               int x, int r)
 143 {
 144         int i;
 145
 146         for (i = bucket->h.size-1; i >= 0; i--) {
 147                 __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i],
 148                                          r, bucket->h.id);
 149                 w &= 0xffff;
 150                 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
 151                         "sw %x rand %llx",
 152                         i, x, r, bucket->h.items[i], bucket->item_weights[i],
 153                         bucket->sum_weights[i], w);
 154                 w *= bucket->sum_weights[i];
 155                 w = w >> 16;
 156                 /*dprintk(" scaled %llx\n", w);*/
 157                 if (w < bucket->item_weights[i]) {
 158                         return bucket->h.items[i];
 159                 }
 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(const 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(const 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         // figure out number of bits we need to shift and
 260         // do it in one step instead of iteratively
 261         if (!(x & 0x18000)) {
 262           int bits = __builtin_clz(x & 0x1FFFF) - 16;
 263           x <<= bits;
 264           iexpon = 15 - bits;
 265         }
 266
 267         index1 = (x >> 8) << 1;
 268         /* RH ~ 2^56/index1 */
 269         RH = __RH_LH_tbl[index1 - 256];
 270         /* LH ~ 2^48 * log2(index1/256) */
 271         LH = __RH_LH_tbl[index1 + 1 - 256];
 272
 273         /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */
 274         xl64 = (__s64)x * RH;
 275         xl64 >>= 48;
 276
 277         result = iexpon;
 278         result <<= (12 + 32);
 279
 280         index2 = xl64 & 0xff;
 281         /* LL ~ 2^48*log2(1.0+index2/2^15) */
 282         LL = __LL_tbl[index2];
 283
 284         LH = LH + LL;
 285
 286         LH >>= (48 - 12 - 32);
 287         result += LH;
 288
 289         return result;
 290 }
 291
 292
 293 /*
 294  * straw2
 295  *
 296  * for reference, see:
 297  *
 298  * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables
 299  *
 300  */
 301
 302 static inline __u32 *get_choose_arg_weights(const struct crush_bucket_straw2 *bucket,
 303                                             const struct crush_choose_arg *arg,
 304                                             int position)
 305 {
 306   if ((arg == NULL) ||
 307       (arg->weight_set == NULL))
 308     return bucket->item_weights;
 309   if (position >= arg->weight_set_size)
 310     position = arg->weight_set_size - 1;
 311   return arg->weight_set[position].weights;
 312 }
 313
 314 static inline __s32 *get_choose_arg_ids(const struct crush_bucket_straw2 *bucket,
 315                                         const struct crush_choose_arg *arg)
 316 {
 317   if ((arg == NULL) || (arg->ids == NULL))
 318     return bucket->h.items;
 319   return arg->ids;
 320 }
 321
 322 static int bucket_straw2_choose(const struct crush_bucket_straw2 *bucket,
 323                                 int x, int r, const struct crush_choose_arg *arg,
 324                                 int position)
 325 {
 326         unsigned int i, high = 0;
 327         unsigned int u;
 328         __s64 ln, draw, high_draw = 0;
 329         __u32 *weights = get_choose_arg_weights(bucket, arg, position);
 330         __s32 *ids = get_choose_arg_ids(bucket, arg);
 331         for (i = 0; i < bucket->h.size; i++) {
 332                 dprintk("weight 0x%x item %d\n", weights[i], ids[i]);
 333                 if (weights[i]) {
 334                         u = crush_hash32_3(bucket->h.hash, x, ids[i], r);
 335                         u &= 0xffff;
 336
 337                         /*
 338                          * for some reason slightly less than 0x10000 produces
 339                          * a slightly more accurate distribution... probably a
 340                          * rounding effect.
 341                          *
 342                          * the natural log lookup table maps [0,0xffff]
 343                          * (corresponding to real numbers [1/0x10000, 1] to
 344                          * [0, 0xffffffffffff] (corresponding to real numbers
 345                          * [-11.090355,0]).
 346                          */
 347                         ln = crush_ln(u) - 0x1000000000000ll;
 348
 349                         /*
 350                          * divide by 16.16 fixed-point weight.  note
 351                          * that the ln value is negative, so a larger
 352                          * weight means a larger (less negative) value
 353                          * for draw.
 354                          */
 355                         draw = div64_s64(ln, weights[i]);
 356                 } else {
 357                         draw = S64_MIN;
 358                 }
 359
 360                 if (i == 0 || draw > high_draw) {
 361                         high = i;
 362                         high_draw = draw;
 363                 }
 364         }
 365
 366         return bucket->h.items[high];
 367 }
 368
 369
 370 static int crush_bucket_choose(const struct crush_bucket *in,
 371                                struct crush_work_bucket *work,
 372                                int x, int r,
 373                                const struct crush_choose_arg *arg,
 374                                int position)
 375 {
 376         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
 377         BUG_ON(in->size == 0);
 378         switch (in->alg) {
 379         case CRUSH_BUCKET_UNIFORM:
 380                 return bucket_uniform_choose(
 381                         (const struct crush_bucket_uniform *)in,
 382                         work, x, r);
 383         case CRUSH_BUCKET_LIST:
 384                 return bucket_list_choose((const struct crush_bucket_list *)in,
 385                                           x, r);
 386         case CRUSH_BUCKET_TREE:
 387                 return bucket_tree_choose((const struct crush_bucket_tree *)in,
 388                                           x, r);
 389         case CRUSH_BUCKET_STRAW:
 390                 return bucket_straw_choose(
 391                         (const struct crush_bucket_straw *)in,
 392                         x, r);
 393         case CRUSH_BUCKET_STRAW2:
 394                 return bucket_straw2_choose(
 395                         (const struct crush_bucket_straw2 *)in,
 396                         x, r, arg, position);
 397         default:
 398                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
 399                 return in->items[0];
 400         }
 401 }
 402
 403 /*
 404  * true if device is marked "out" (failed, fully offloaded)
 405  * of the cluster
 406  */
 407 static int is_out(const struct crush_map *map,
 408                   const __u32 *weight, int weight_max,
 409                   int item, int x)
 410 {
 411         if (item >= weight_max)
 412                 return 1;
 413         if (weight[item] >= 0x10000)
 414                 return 0;
 415         if (weight[item] == 0)
 416                 return 1;
 417         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
 418             < weight[item])
 419                 return 0;
 420         return 1;
 421 }
 422
 423 /**
 424  * crush_choose_firstn - choose numrep distinct items of given type
 425  * @map: the crush_map
 426  * @bucket: the bucket we are choose an item from
 427  * @x: crush input value
 428  * @numrep: the number of items to choose
 429  * @type: the type of item to choose
 430  * @out: pointer to output vector
 431  * @outpos: our position in that vector
 432  * @out_size: size of the out vector
 433  * @tries: number of attempts to make
 434  * @recurse_tries: number of attempts to have recursive chooseleaf make
 435  * @local_retries: localized retries
 436  * @local_fallback_retries: localized fallback retries
 437  * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose)
 438  * @stable: stable mode starts rep=0 in the recursive call for all replicas
 439  * @vary_r: pass r to recursive calls
 440  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 441  * @parent_r: r value passed from the parent
 442  */
 443 static int crush_choose_firstn(const struct crush_map *map,
 444                                struct crush_work *work,
 445                                const struct crush_bucket *bucket,
 446                                const __u32 *weight, int weight_max,
 447                                int x, int numrep, int type,
 448                                int *out, int outpos,
 449                                int out_size,
 450                                unsigned int tries,
 451                                unsigned int recurse_tries,
 452                                unsigned int local_retries,
 453                                unsigned int local_fallback_retries,
 454                                int recurse_to_leaf,
 455                                unsigned int vary_r,
 456                                unsigned int stable,
 457                                int *out2,
 458                                int parent_r,
 459                                const struct crush_choose_arg *choose_args)
 460 {
 461         int rep;
 462         unsigned int ftotal, flocal;
 463         int retry_descent, retry_bucket, skip_rep;
 464         const struct crush_bucket *in = bucket;
 465         int r;
 466         int i;
 467         int item = 0;
 468         int itemtype;
 469         int collide, reject;
 470         int count = out_size;
 471
 472         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d tries %d \
 473 recurse_tries %d local_retries %d local_fallback_retries %d \
 474 parent_r %d stable %d\n",
 475                 recurse_to_leaf ? "_LEAF" : "",
 476                 bucket->id, x, outpos, numrep,
 477                 tries, recurse_tries, local_retries, local_fallback_retries,
 478                 parent_r, stable);
 479
 480         for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) {
 481                 /* keep trying until we get a non-out, non-colliding item */
 482                 ftotal = 0;
 483                 skip_rep = 0;
 484                 do {
 485                         retry_descent = 0;
 486                         in = bucket;              /* initial bucket */
 487
 488                         /* choose through intervening buckets */
 489                         flocal = 0;
 490                         do {
 491                                 collide = 0;
 492                                 retry_bucket = 0;
 493                                 r = rep + parent_r;
 494                                 /* r' = r + f_total */
 495                                 r += ftotal;
 496
 497                                 /* bucket choose */
 498                                 if (in->size == 0) {
 499                                         reject = 1;
 500                                         goto reject;
 501                                 }
 502                                 if (local_fallback_retries > 0 &&
 503                                     flocal >= (in->size>>1) &&
 504                                     flocal > local_fallback_retries)
 505                                         item = bucket_perm_choose(
 506                                                 in, work->work[-1-in->id],
 507                                                 x, r);
 508                                 else
 509                                         item = crush_bucket_choose(
 510                                                 in, work->work[-1-in->id],
 511                                                 x, r,
 512                                                 (choose_args ? &choose_args[-1-in->id] : 0),
 513                                                 outpos);
 514                                 if (item >= map->max_devices) {
 515                                         dprintk("   bad item %d\n", item);
 516                                         skip_rep = 1;
 517                                         break;
 518                                 }
 519
 520                                 /* desired type? */
 521                                 if (item < 0)
 522                                         itemtype = map->buckets[-1-item]->type;
 523                                 else
 524                                         itemtype = 0;
 525                                 dprintk("  item %d type %d\n", item, itemtype);
 526
 527                                 /* keep going? */
 528                                 if (itemtype != type) {
 529                                         if (item >= 0 ||
 530                                             (-1-item) >= map->max_buckets) {
 531                                                 dprintk("   bad item type %d\n", type);
 532                                                 skip_rep = 1;
 533                                                 break;
 534                                         }
 535                                         in = map->buckets[-1-item];
 536                                         retry_bucket = 1;
 537                                         continue;
 538                                 }
 539
 540                                 /* collision? */
 541                                 for (i = 0; i < outpos; i++) {
 542                                         if (out[i] == item) {
 543                                                 collide = 1;
 544                                                 break;
 545                                         }
 546                                 }
 547
 548                                 reject = 0;
 549                                 if (!collide && recurse_to_leaf) {
 550                                         if (item < 0) {
 551                                                 int sub_r;
 552                                                 if (vary_r)
 553                                                         sub_r = r >> (vary_r-1);
 554                                                 else
 555                                                         sub_r = 0;
 556                                                 if (crush_choose_firstn(
 557                                                             map,
 558                                                             work,
 559                                                             map->buckets[-1-item],
 560                                                             weight, weight_max,
 561                                                             x, stable ? 1 : outpos+1, 0,
 562                                                             out2, outpos, count,
 563                                                             recurse_tries, 0,
 564                                                             local_retries,
 565                                                             local_fallback_retries,
 566                                                             0,
 567                                                             vary_r,
 568                                                             stable,
 569                                                             NULL,
 570                                                             sub_r,
 571                                                             choose_args) <= outpos)
 572                                                         /* didn't get leaf */
 573                                                         reject = 1;
 574                                         } else {
 575                                                 /* we already have a leaf! */
 576                                                 out2[outpos] = item;
 577                                         }
 578                                 }
 579
 580                                 if (!reject && !collide) {
 581                                         /* out? */
 582                                         if (itemtype == 0)
 583                                                 reject = is_out(map, weight,
 584                                                                 weight_max,
 585                                                                 item, x);
 586                                 }
 587
 588 reject:
 589                                 if (reject || collide) {
 590                                         ftotal++;
 591                                         flocal++;
 592
 593                                         if (collide && flocal <= local_retries)
 594                                                 /* retry locally a few times */
 595                                                 retry_bucket = 1;
 596                                         else if (local_fallback_retries > 0 &&
 597                                                  flocal <= in->size + local_fallback_retries)
 598                                                 /* exhaustive bucket search */
 599                                                 retry_bucket = 1;
 600                                         else if (ftotal < tries)
 601                                                 /* then retry descent */
 602                                                 retry_descent = 1;
 603                                         else
 604                                                 /* else give up */
 605                                                 skip_rep = 1;
 606                                         dprintk("  reject %d  collide %d  "
 607                                                 "ftotal %u  flocal %u\n",
 608                                                 reject, collide, ftotal,
 609                                                 flocal);
 610                                 }
 611                         } while (retry_bucket);
 612                 } while (retry_descent);
 613
 614                 if (skip_rep) {
 615                         dprintk("skip rep\n");
 616                         continue;
 617                 }
 618
 619                 dprintk("CHOOSE got %d\n", item);
 620                 out[outpos] = item;
 621                 outpos++;
 622                 count--;
 623 #ifndef __KERNEL__
 624                 if (map->choose_tries && ftotal <= map->choose_total_tries)
 625                         map->choose_tries[ftotal]++;
 626 #endif
 627         }
 628
 629         dprintk("CHOOSE returns %d\n", outpos);
 630         return outpos;
 631 }
 632
 633
 634 /**
 635  * crush_choose_indep: alternative breadth-first positionally stable mapping
 636  *
 637  */
 638 static void crush_choose_indep(const struct crush_map *map,
 639                                struct crush_work *work,
 640                                const struct crush_bucket *bucket,
 641                                const __u32 *weight, int weight_max,
 642                                int x, int left, int numrep, int type,
 643                                int *out, int outpos,
 644                                unsigned int tries,
 645                                unsigned int recurse_tries,
 646                                int recurse_to_leaf,
 647                                int *out2,
 648                                int parent_r,
 649                                const struct crush_choose_arg *choose_args)
 650 {
 651         const struct crush_bucket *in = bucket;
 652         int endpos = outpos + left;
 653         int rep;
 654         unsigned int ftotal;
 655         int r;
 656         int i;
 657         int item = 0;
 658         int itemtype;
 659         int collide;
 660
 661         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 662                 bucket->id, x, outpos, numrep);
 663
 664         /* initially my result is undefined */
 665         for (rep = outpos; rep < endpos; rep++) {
 666                 out[rep] = CRUSH_ITEM_UNDEF;
 667                 if (out2)
 668                         out2[rep] = CRUSH_ITEM_UNDEF;
 669         }
 670
 671         for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) {
 672 #ifdef DEBUG_INDEP
 673                 if (out2 && ftotal) {
 674                         dprintk("%u %d a: ", ftotal, left);
 675                         for (rep = outpos; rep < endpos; rep++) {
 676                                 dprintk(" %d", out[rep]);
 677                         }
 678                         dprintk("\n");
 679                         dprintk("%u %d b: ", ftotal, left);
 680                         for (rep = outpos; rep < endpos; rep++) {
 681                                 dprintk(" %d", out2[rep]);
 682                         }
 683                         dprintk("\n");
 684                 }
 685 #endif
 686                 for (rep = outpos; rep < endpos; rep++) {
 687                         if (out[rep] != CRUSH_ITEM_UNDEF)
 688                                 continue;
 689
 690                         in = bucket;  /* initial bucket */
 691
 692                         /* choose through intervening buckets */
 693                         for (;;) {
 694                                 /* note: we base the choice on the position
 695                                  * even in the nested call.  that means that
 696                                  * if the first layer chooses the same bucket
 697                                  * in a different position, we will tend to
 698                                  * choose a different item in that bucket.
 699                                  * this will involve more devices in data
 700                                  * movement and tend to distribute the load.
 701                                  */
 702                                 r = rep + parent_r;
 703
 704                                 /* be careful */
 705                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
 706                                     in->size % numrep == 0)
 707                                         /* r'=r+(n+1)*f_total */
 708                                         r += (numrep+1) * ftotal;
 709                                 else
 710                                         /* r' = r + n*f_total */
 711                                         r += numrep * ftotal;
 712
 713                                 /* bucket choose */
 714                                 if (in->size == 0) {
 715                                         dprintk("   empty bucket\n");
 716                                         break;
 717                                 }
 718
 719                                 item = crush_bucket_choose(
 720                                         in, work->work[-1-in->id],
 721                                         x, r,
 722                                         (choose_args ? &choose_args[-1-in->id] : 0),
 723                                         outpos);
 724                                 if (item >= map->max_devices) {
 725                                         dprintk("   bad item %d\n", item);
 726                                         out[rep] = CRUSH_ITEM_NONE;
 727                                         if (out2)
 728                                                 out2[rep] = CRUSH_ITEM_NONE;
 729                                         left--;
 730                                         break;
 731                                 }
 732
 733                                 /* desired type? */
 734                                 if (item < 0)
 735                                         itemtype = map->buckets[-1-item]->type;
 736                                 else
 737                                         itemtype = 0;
 738                                 dprintk("  item %d type %d\n", item, itemtype);
 739
 740                                 /* keep going? */
 741                                 if (itemtype != type) {
 742                                         if (item >= 0 ||
 743                                             (-1-item) >= map->max_buckets) {
 744                                                 dprintk("   bad item type %d\n", type);
 745                                                 out[rep] = CRUSH_ITEM_NONE;
 746                                                 if (out2)
 747                                                         out2[rep] =
 748                                                                 CRUSH_ITEM_NONE;
 749                                                 left--;
 750                                                 break;
 751                                         }
 752                                         in = map->buckets[-1-item];
 753                                         continue;
 754                                 }
 755
 756                                 /* collision? */
 757                                 collide = 0;
 758                                 for (i = outpos; i < endpos; i++) {
 759                                         if (out[i] == item) {
 760                                                 collide = 1;
 761                                                 break;
 762                                         }
 763                                 }
 764                                 if (collide)
 765                                         break;
 766
 767                                 if (recurse_to_leaf) {
 768                                         if (item < 0) {
 769                                                 crush_choose_indep(
 770                                                         map,
 771                                                         work,
 772                                                         map->buckets[-1-item],
 773                                                         weight, weight_max,
 774                                                         x, 1, numrep, 0,
 775                                                         out2, rep,
 776                                                         recurse_tries, 0,
 777                                                         0, NULL, r, choose_args);
 778                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 779                                                         /* placed nothing; no leaf */
 780                                                         break;
 781                                                 }
 782                                         } else {
 783                                                 /* we already have a leaf! */
 784                                                 out2[rep] = item;
 785                                         }
 786                                 }
 787
 788                                 /* out? */
 789                                 if (itemtype == 0 &&
 790                                     is_out(map, weight, weight_max, item, x))
 791                                         break;
 792
 793                                 /* yay! */
 794                                 out[rep] = item;
 795                                 left--;
 796                                 break;
 797                         }
 798                 }
 799         }
 800         for (rep = outpos; rep < endpos; rep++) {
 801                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 802                         out[rep] = CRUSH_ITEM_NONE;
 803                 }
 804                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 805                         out2[rep] = CRUSH_ITEM_NONE;
 806                 }
 807         }
 808 #ifndef __KERNEL__
 809         if (map->choose_tries && ftotal <= map->choose_total_tries)
 810                 map->choose_tries[ftotal]++;
 811 #endif
 812 #ifdef DEBUG_INDEP
 813         if (out2) {
 814                 dprintk("%u %d a: ", ftotal, left);
 815                 for (rep = outpos; rep < endpos; rep++) {
 816                         dprintk(" %d", out[rep]);
 817                 }
 818                 dprintk("\n");
 819                 dprintk("%u %d b: ", ftotal, left);
 820                 for (rep = outpos; rep < endpos; rep++) {
 821                         dprintk(" %d", out2[rep]);
 822                 }
 823                 dprintk("\n");
 824         }
 825 #endif
 826 }
 827
 828
 829 /* This takes a chunk of memory and sets it up to be a shiny new
 830    working area for a CRUSH placement computation. It must be called
 831    on any newly allocated memory before passing it in to
 832    crush_do_rule. It may be used repeatedly after that, so long as the
 833    map has not changed. If the map /has/ changed, you must make sure
 834    the working size is no smaller than what was allocated and re-run
 835    crush_init_workspace.
 836
 837    If you do retain the working space between calls to crush, make it
 838    thread-local. If you reinstitute the locking I've spent so much
 839    time getting rid of, I will be very unhappy with you. */
 840
 841 void crush_init_workspace(const struct crush_map *m, void *v) {
 842         /* We work by moving through the available space and setting
 843            values and pointers as we go.
 844
 845            It's a bit like Forth's use of the 'allot' word since we
 846            set the pointer first and then reserve the space for it to
 847            point to by incrementing the point. */
 848         struct crush_work *w = (struct crush_work *)v;
 849         char *point = (char *)v;
 850         __s32 b;
 851         point += sizeof(struct crush_work);
 852         w->work = (struct crush_work_bucket **)point;
 853         point += m->max_buckets * sizeof(struct crush_work_bucket *);
 854         for (b = 0; b < m->max_buckets; ++b) {
 855                 if (m->buckets[b] == 0)
 856                         continue;
 857
 858                 w->work[b] = (struct crush_work_bucket *) point;
 859                 switch (m->buckets[b]->alg) {
 860                 default:
 861                         point += sizeof(struct crush_work_bucket);
 862                         break;
 863                 }
 864                 w->work[b]->perm_x = 0;
 865                 w->work[b]->perm_n = 0;
 866                 w->work[b]->perm = (__u32 *)point;
 867                 point += m->buckets[b]->size * sizeof(__u32);
 868         }
 869         BUG_ON((char *)point - (char *)w != m->working_size);
 870 }
 871
 872 /**
 873  * crush_do_rule - calculate a mapping with the given input and rule
 874  * @map: the crush_map
 875  * @ruleno: the rule id
 876  * @x: hash input
 877  * @result: pointer to result vector
 878  * @result_max: maximum result size
 879  * @weight: weight vector (for map leaves)
 880  * @weight_max: size of weight vector
 881  * @cwin: Pointer to at least map->working_size bytes of memory or NULL.
 882  */
 883 int crush_do_rule(const struct crush_map *map,
 884                   int ruleno, int x, int *result, int result_max,
 885                   const __u32 *weight, int weight_max,
 886                   void *cwin, const struct crush_choose_arg *choose_args)
 887 {
 888         int result_len;
 889         struct crush_work *cw = cwin;
 890         int *a = (int *)((char *)cw + map->working_size);
 891         int *b = a + result_max;
 892         int *c = b + result_max;
 893         int *w = a;
 894         int *o = b;
 895         int recurse_to_leaf;
 896         int wsize = 0;
 897         int osize;
 898         int *tmp;
 899         const struct crush_rule *rule;
 900         __u32 step;
 901         int i, j;
 902         int numrep;
 903         int out_size;
 904         /*
 905          * the original choose_total_tries value was off by one (it
 906          * counted "retries" and not "tries").  add one.
 907          */
 908         int choose_tries = map->choose_total_tries + 1;
 909         int choose_leaf_tries = 0;
 910         /*
 911          * the local tries values were counted as "retries", though,
 912          * and need no adjustment
 913          */
 914         int choose_local_retries = map->choose_local_tries;
 915         int choose_local_fallback_retries = map->choose_local_fallback_tries;
 916
 917         int vary_r = map->chooseleaf_vary_r;
 918         int stable = map->chooseleaf_stable;
 919
 920         if ((__u32)ruleno >= map->max_rules) {
 921                 dprintk(" bad ruleno %d\n", ruleno);
 922                 return 0;
 923         }
 924
 925         rule = map->rules[ruleno];
 926         result_len = 0;
 927
 928         for (step = 0; step < rule->len; step++) {
 929                 int firstn = 0;
 930                 const struct crush_rule_step *curstep = &rule->steps[step];
 931
 932                 switch (curstep->op) {
 933                 case CRUSH_RULE_TAKE:
 934                         if ((curstep->arg1 >= 0 &&
 935                              curstep->arg1 < map->max_devices) ||
 936                             (-1-curstep->arg1 >= 0 &&
 937                              -1-curstep->arg1 < map->max_buckets &&
 938                              map->buckets[-1-curstep->arg1])) {
 939                                 w[0] = curstep->arg1;
 940                                 wsize = 1;
 941                         } else {
 942                                 dprintk(" bad take value %d\n", curstep->arg1);
 943                         }
 944                         break;
 945
 946                 case CRUSH_RULE_SET_CHOOSE_TRIES:
 947                         if (curstep->arg1 > 0)
 948                                 choose_tries = curstep->arg1;
 949                         break;
 950
 951                 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
 952                         if (curstep->arg1 > 0)
 953                                 choose_leaf_tries = curstep->arg1;
 954                         break;
 955
 956                 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
 957                         if (curstep->arg1 >= 0)
 958                                 choose_local_retries = curstep->arg1;
 959                         break;
 960
 961                 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
 962                         if (curstep->arg1 >= 0)
 963                                 choose_local_fallback_retries = curstep->arg1;
 964                         break;
 965
 966                 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R:
 967                         if (curstep->arg1 >= 0)
 968                                 vary_r = curstep->arg1;
 969                         break;
 970
 971                 case CRUSH_RULE_SET_CHOOSELEAF_STABLE:
 972                         if (curstep->arg1 >= 0)
 973                                 stable = curstep->arg1;
 974                         break;
 975
 976                 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
 977                 case CRUSH_RULE_CHOOSE_FIRSTN:
 978                         firstn = 1;
 979                         /* fall through */
 980                 case CRUSH_RULE_CHOOSELEAF_INDEP:
 981                 case CRUSH_RULE_CHOOSE_INDEP:
 982                         if (wsize == 0)
 983                                 break;
 984
 985                         recurse_to_leaf =
 986                                 curstep->op ==
 987                                  CRUSH_RULE_CHOOSELEAF_FIRSTN ||
 988                                 curstep->op ==
 989                                 CRUSH_RULE_CHOOSELEAF_INDEP;
 990
 991                         /* reset output */
 992                         osize = 0;
 993
 994                         for (i = 0; i < wsize; i++) {
 995                                 int bno;
 996                                 numrep = curstep->arg1;
 997                                 if (numrep <= 0) {
 998                                         numrep += result_max;
 999                                         if (numrep <= 0)
1000                                                 continue;
1001                                 }
1002                                 j = 0;
1003                                 /* make sure bucket id is valid */
1004                                 bno = -1 - w[i];
1005                                 if (bno < 0 || bno >= map->max_buckets) {
1006                                         // w[i] is probably CRUSH_ITEM_NONE
1007                                         dprintk("  bad w[i] %d\n", w[i]);
1008                                         continue;
1009                                 }
1010                                 if (firstn) {
1011                                         int recurse_tries;
1012                                         if (choose_leaf_tries)
1013                                                 recurse_tries =
1014                                                         choose_leaf_tries;
1015                                         else if (map->chooseleaf_descend_once)
1016                                                 recurse_tries = 1;
1017                                         else
1018                                                 recurse_tries = choose_tries;
1019                                         osize += crush_choose_firstn(
1020                                                 map,
1021                                                 cw,
1022                                                 map->buckets[bno],
1023                                                 weight, weight_max,
1024                                                 x, numrep,
1025                                                 curstep->arg2,
1026                                                 o+osize, j,
1027                                                 result_max-osize,
1028                                                 choose_tries,
1029                                                 recurse_tries,
1030                                                 choose_local_retries,
1031                                                 choose_local_fallback_retries,
1032                                                 recurse_to_leaf,
1033                                                 vary_r,
1034                                                 stable,
1035                                                 c+osize,
1036                                                 0,
1037                                                 choose_args);
1038                                 } else {
1039                                         out_size = ((numrep < (result_max-osize)) ?
1040                                                     numrep : (result_max-osize));
1041                                         crush_choose_indep(
1042                                                 map,
1043                                                 cw,
1044                                                 map->buckets[bno],
1045                                                 weight, weight_max,
1046                                                 x, out_size, numrep,
1047                                                 curstep->arg2,
1048                                                 o+osize, j,
1049                                                 choose_tries,
1050                                                 choose_leaf_tries ?
1051                                                    choose_leaf_tries : 1,
1052                                                 recurse_to_leaf,
1053                                                 c+osize,
1054                                                 0,
1055                                                 choose_args);
1056                                         osize += out_size;
1057                                 }
1058                         }
1059
1060                         if (recurse_to_leaf)
1061                                 /* copy final _leaf_ values to output set */
1062                                 memcpy(o, c, osize*sizeof(*o));
1063
1064                         /* swap o and w arrays */
1065                         tmp = o;
1066                         o = w;
1067                         w = tmp;
1068                         wsize = osize;
1069                         break;
1070
1071
1072                 case CRUSH_RULE_EMIT:
1073                         for (i = 0; i < wsize && result_len < result_max; i++) {
1074                                 result[result_len] = w[i];
1075                                 result_len++;
1076                         }
1077                         wsize = 0;
1078                         break;
1079
1080                 default:
1081                         dprintk(" unknown op %d at step %d\n",
1082                                 curstep->op, step);
1083                         break;
1084                 }
1085         }
1086
1087         return result_len;
1088 }