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 #else
  21 # include "crush_compat.h"
  22 # include "crush.h"
  23 # include "hash.h"
  24 #endif
  25 #include "crush_ln_table.h"
  26
  27 #define dprintk(args...) /* printf(args) */
  28
  29 /*
  30  * Implement the core CRUSH mapping algorithm.
  31  */
  32
  33 /**
  34  * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
  35  * @map: the crush_map
  36  * @ruleset: the storage ruleset id (user defined)
  37  * @type: storage ruleset type (user defined)
  38  * @size: output set size
  39  */
  40 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
  41 {
  42         __u32 i;
  43
  44         for (i = 0; i < map->max_rules; i++) {
  45                 if (map->rules[i] &&
  46                     map->rules[i]->mask.ruleset == ruleset &&
  47                     map->rules[i]->mask.type == type &&
  48                     map->rules[i]->mask.min_size <= size &&
  49                     map->rules[i]->mask.max_size >= size)
  50                         return i;
  51         }
  52         return -1;
  53 }
  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(struct crush_bucket *bucket,
  74                               int x, int r)
  75 {
  76         unsigned int pr = r % bucket->size;
  77         unsigned int i, s;
  78
  79         /* start a new permutation if @x has changed */
  80         if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) {
  81                 dprintk("bucket %d new x=%d\n", bucket->id, x);
  82                 bucket->perm_x = x;
  83
  84                 /* optimize common r=0 case */
  85                 if (pr == 0) {
  86                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
  87                                 bucket->size;
  88                         bucket->perm[0] = s;
  89                         bucket->perm_n = 0xffff;   /* magic value, see below */
  90                         goto out;
  91                 }
  92
  93                 for (i = 0; i < bucket->size; i++)
  94                         bucket->perm[i] = i;
  95                 bucket->perm_n = 0;
  96         } else if (bucket->perm_n == 0xffff) {
  97                 /* clean up after the r=0 case above */
  98                 for (i = 1; i < bucket->size; i++)
  99                         bucket->perm[i] = i;
 100                 bucket->perm[bucket->perm[0]] = 0;
 101                 bucket->perm_n = 1;
 102         }
 103
 104         /* calculate permutation up to pr */
 105         for (i = 0; i < bucket->perm_n; i++)
 106                 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]);
 107         while (bucket->perm_n <= pr) {
 108                 unsigned int p = bucket->perm_n;
 109                 /* no point in swapping the final entry */
 110                 if (p < bucket->size - 1) {
 111                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
 112                                 (bucket->size - p);
 113                         if (i) {
 114                                 unsigned int t = bucket->perm[p + i];
 115                                 bucket->perm[p + i] = bucket->perm[p];
 116                                 bucket->perm[p] = t;
 117                         }
 118                         dprintk(" perm_choose swap %d with %d\n", p, p+i);
 119                 }
 120                 bucket->perm_n++;
 121         }
 122         for (i = 0; i < bucket->size; i++)
 123                 dprintk(" perm_choose  %d: %d\n", i, bucket->perm[i]);
 124
 125         s = bucket->perm[pr];
 126 out:
 127         dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
 128                 bucket->size, x, r, pr, s);
 129         return bucket->items[s];
 130 }
 131
 132 /* uniform */
 133 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
 134                                  int x, int r)
 135 {
 136         return bucket_perm_choose(&bucket->h, x, r);
 137 }
 138
 139 /* list */
 140 static int bucket_list_choose(struct crush_bucket_list *bucket,
 141                               int x, int r)
 142 {
 143         int i;
 144
 145         for (i = bucket->h.size-1; i >= 0; i--) {
 146                 __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i],
 147                                          r, bucket->h.id);
 148                 w &= 0xffff;
 149                 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
 150                         "sw %x rand %llx",
 151                         i, x, r, bucket->h.items[i], bucket->item_weights[i],
 152                         bucket->sum_weights[i], w);
 153                 w *= bucket->sum_weights[i];
 154                 w = w >> 16;
 155                 /*dprintk(" scaled %llx\n", w);*/
 156                 if (w < bucket->item_weights[i])
 157                         return bucket->h.items[i];
 158         }
 159
 160         dprintk("bad list sums for bucket %d\n", bucket->h.id);
 161         return bucket->h.items[0];
 162 }
 163
 164
 165 /* (binary) tree */
 166 static int height(int n)
 167 {
 168         int h = 0;
 169         while ((n & 1) == 0) {
 170                 h++;
 171                 n = n >> 1;
 172         }
 173         return h;
 174 }
 175
 176 static int left(int x)
 177 {
 178         int h = height(x);
 179         return x - (1 << (h-1));
 180 }
 181
 182 static int right(int x)
 183 {
 184         int h = height(x);
 185         return x + (1 << (h-1));
 186 }
 187
 188 static int terminal(int x)
 189 {
 190         return x & 1;
 191 }
 192
 193 static int bucket_tree_choose(struct crush_bucket_tree *bucket,
 194                               int x, int r)
 195 {
 196         int n;
 197         __u32 w;
 198         __u64 t;
 199
 200         /* start at root */
 201         n = bucket->num_nodes >> 1;
 202
 203         while (!terminal(n)) {
 204                 int l;
 205                 /* pick point in [0, w) */
 206                 w = bucket->node_weights[n];
 207                 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
 208                                           bucket->h.id) * (__u64)w;
 209                 t = t >> 32;
 210
 211                 /* descend to the left or right? */
 212                 l = left(n);
 213                 if (t < bucket->node_weights[l])
 214                         n = l;
 215                 else
 216                         n = right(n);
 217         }
 218
 219         return bucket->h.items[n >> 1];
 220 }
 221
 222
 223 /* straw */
 224
 225 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
 226                                int x, int r)
 227 {
 228         __u32 i;
 229         int high = 0;
 230         __u64 high_draw = 0;
 231         __u64 draw;
 232
 233         for (i = 0; i < bucket->h.size; i++) {
 234                 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
 235                 draw &= 0xffff;
 236                 draw *= bucket->straws[i];
 237                 if (i == 0 || draw > high_draw) {
 238                         high = i;
 239                         high_draw = draw;
 240                 }
 241         }
 242         return bucket->h.items[high];
 243 }
 244
 245 /* compute 2^44*log2(input+1) */
 246 static __u64 crush_ln(unsigned int xin)
 247 {
 248         unsigned int x = xin, x1;
 249         int iexpon, index1, index2;
 250         __u64 RH, LH, LL, xl64, result;
 251
 252         x++;
 253
 254         /* normalize input */
 255         iexpon = 15;
 256
 257         /*
 258          * figure out number of bits we need to shift and
 259          * do it in one step instead of iteratively
 260          */
 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         x1 = xl64;
 277
 278         result = iexpon;
 279         result <<= (12 + 32);
 280
 281         index2 = x1 & 0xff;
 282         /* LL ~ 2^48*log2(1.0+index2/2^15) */
 283         LL = __LL_tbl[index2];
 284
 285         LH = LH + LL;
 286
 287         LH >>= (48 - 12 - 32);
 288         result += LH;
 289
 290         return result;
 291 }
 292
 293
 294 /*
 295  * straw2
 296  *
 297  * for reference, see:
 298  *
 299  * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables
 300  *
 301  */
 302
 303 static int bucket_straw2_choose(struct crush_bucket_straw2 *bucket,
 304                                 int x, int r)
 305 {
 306         unsigned int i, high = 0;
 307         unsigned int u;
 308         unsigned int w;
 309         __s64 ln, draw, high_draw = 0;
 310
 311         for (i = 0; i < bucket->h.size; i++) {
 312                 w = bucket->item_weights[i];
 313                 if (w) {
 314                         u = crush_hash32_3(bucket->h.hash, x,
 315                                            bucket->h.items[i], r);
 316                         u &= 0xffff;
 317
 318                         /*
 319                          * for some reason slightly less than 0x10000 produces
 320                          * a slightly more accurate distribution... probably a
 321                          * rounding effect.
 322                          *
 323                          * the natural log lookup table maps [0,0xffff]
 324                          * (corresponding to real numbers [1/0x10000, 1] to
 325                          * [0, 0xffffffffffff] (corresponding to real numbers
 326                          * [-11.090355,0]).
 327                          */
 328                         ln = crush_ln(u) - 0x1000000000000ll;
 329
 330                         /*
 331                          * divide by 16.16 fixed-point weight.  note
 332                          * that the ln value is negative, so a larger
 333                          * weight means a larger (less negative) value
 334                          * for draw.
 335                          */
 336                         draw = div64_s64(ln, w);
 337                 } else {
 338                         draw = S64_MIN;
 339                 }
 340
 341                 if (i == 0 || draw > high_draw) {
 342                         high = i;
 343                         high_draw = draw;
 344                 }
 345         }
 346         return bucket->h.items[high];
 347 }
 348
 349
 350 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
 351 {
 352         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
 353         BUG_ON(in->size == 0);
 354         switch (in->alg) {
 355         case CRUSH_BUCKET_UNIFORM:
 356                 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
 357                                           x, r);
 358         case CRUSH_BUCKET_LIST:
 359                 return bucket_list_choose((struct crush_bucket_list *)in,
 360                                           x, r);
 361         case CRUSH_BUCKET_TREE:
 362                 return bucket_tree_choose((struct crush_bucket_tree *)in,
 363                                           x, r);
 364         case CRUSH_BUCKET_STRAW:
 365                 return bucket_straw_choose((struct crush_bucket_straw *)in,
 366                                            x, r);
 367         case CRUSH_BUCKET_STRAW2:
 368                 return bucket_straw2_choose((struct crush_bucket_straw2 *)in,
 369                                             x, r);
 370         default:
 371                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
 372                 return in->items[0];
 373         }
 374 }
 375
 376
 377 /*
 378  * true if device is marked "out" (failed, fully offloaded)
 379  * of the cluster
 380  */
 381 static int is_out(const struct crush_map *map,
 382                   const __u32 *weight, int weight_max,
 383                   int item, int x)
 384 {
 385         if (item >= weight_max)
 386                 return 1;
 387         if (weight[item] >= 0x10000)
 388                 return 0;
 389         if (weight[item] == 0)
 390                 return 1;
 391         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
 392             < weight[item])
 393                 return 0;
 394         return 1;
 395 }
 396
 397 /**
 398  * crush_choose_firstn - choose numrep distinct items of given type
 399  * @map: the crush_map
 400  * @bucket: the bucket we are choose an item from
 401  * @x: crush input value
 402  * @numrep: the number of items to choose
 403  * @type: the type of item to choose
 404  * @out: pointer to output vector
 405  * @outpos: our position in that vector
 406  * @out_size: size of the out vector
 407  * @tries: number of attempts to make
 408  * @recurse_tries: number of attempts to have recursive chooseleaf make
 409  * @local_retries: localized retries
 410  * @local_fallback_retries: localized fallback retries
 411  * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose)
 412  * @stable: stable mode starts rep=0 in the recursive call for all replicas
 413  * @vary_r: pass r to recursive calls
 414  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 415  * @parent_r: r value passed from the parent
 416  */
 417 static int crush_choose_firstn(const struct crush_map *map,
 418                                struct crush_bucket *bucket,
 419                                const __u32 *weight, int weight_max,
 420                                int x, int numrep, int type,
 421                                int *out, int outpos,
 422                                int out_size,
 423                                unsigned int tries,
 424                                unsigned int recurse_tries,
 425                                unsigned int local_retries,
 426                                unsigned int local_fallback_retries,
 427                                int recurse_to_leaf,
 428                                unsigned int vary_r,
 429                                unsigned int stable,
 430                                int *out2,
 431                                int parent_r)
 432 {
 433         int rep;
 434         unsigned int ftotal, flocal;
 435         int retry_descent, retry_bucket, skip_rep;
 436         struct crush_bucket *in = bucket;
 437         int r;
 438         int i;
 439         int item = 0;
 440         int itemtype;
 441         int collide, reject;
 442         int count = out_size;
 443
 444         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",
 445                 recurse_to_leaf ? "_LEAF" : "",
 446                 bucket->id, x, outpos, numrep,
 447                 tries, recurse_tries, local_retries, local_fallback_retries,
 448                 parent_r, stable);
 449
 450         for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) {
 451                 /* keep trying until we get a non-out, non-colliding item */
 452                 ftotal = 0;
 453                 skip_rep = 0;
 454                 do {
 455                         retry_descent = 0;
 456                         in = bucket;               /* initial bucket */
 457
 458                         /* choose through intervening buckets */
 459                         flocal = 0;
 460                         do {
 461                                 collide = 0;
 462                                 retry_bucket = 0;
 463                                 r = rep + parent_r;
 464                                 /* r' = r + f_total */
 465                                 r += ftotal;
 466
 467                                 /* bucket choose */
 468                                 if (in->size == 0) {
 469                                         reject = 1;
 470                                         goto reject;
 471                                 }
 472                                 if (local_fallback_retries > 0 &&
 473                                     flocal >= (in->size>>1) &&
 474                                     flocal > local_fallback_retries)
 475                                         item = bucket_perm_choose(in, x, r);
 476                                 else
 477                                         item = crush_bucket_choose(in, x, r);
 478                                 if (item >= map->max_devices) {
 479                                         dprintk("   bad item %d\n", item);
 480                                         skip_rep = 1;
 481                                         break;
 482                                 }
 483
 484                                 /* desired type? */
 485                                 if (item < 0)
 486                                         itemtype = map->buckets[-1-item]->type;
 487                                 else
 488                                         itemtype = 0;
 489                                 dprintk("  item %d type %d\n", item, itemtype);
 490
 491                                 /* keep going? */
 492                                 if (itemtype != type) {
 493                                         if (item >= 0 ||
 494                                             (-1-item) >= map->max_buckets) {
 495                                                 dprintk("   bad item type %d\n", type);
 496                                                 skip_rep = 1;
 497                                                 break;
 498                                         }
 499                                         in = map->buckets[-1-item];
 500                                         retry_bucket = 1;
 501                                         continue;
 502                                 }
 503
 504                                 /* collision? */
 505                                 for (i = 0; i < outpos; i++) {
 506                                         if (out[i] == item) {
 507                                                 collide = 1;
 508                                                 break;
 509                                         }
 510                                 }
 511
 512                                 reject = 0;
 513                                 if (!collide && recurse_to_leaf) {
 514                                         if (item < 0) {
 515                                                 int sub_r;
 516                                                 if (vary_r)
 517                                                         sub_r = r >> (vary_r-1);
 518                                                 else
 519                                                         sub_r = 0;
 520                                                 if (crush_choose_firstn(map,
 521                                                          map->buckets[-1-item],
 522                                                          weight, weight_max,
 523                                                          x, stable ? 1 : outpos+1, 0,
 524                                                          out2, outpos, count,
 525                                                          recurse_tries, 0,
 526                                                          local_retries,
 527                                                          local_fallback_retries,
 528                                                          0,
 529                                                          vary_r,
 530                                                          stable,
 531                                                          NULL,
 532                                                          sub_r) <= outpos)
 533                                                         /* didn't get leaf */
 534                                                         reject = 1;
 535                                         } else {
 536                                                 /* we already have a leaf! */
 537                                                 out2[outpos] = item;
 538                                         }
 539                                 }
 540
 541                                 if (!reject) {
 542                                         /* out? */
 543                                         if (itemtype == 0)
 544                                                 reject = is_out(map, weight,
 545                                                                 weight_max,
 546                                                                 item, x);
 547                                         else
 548                                                 reject = 0;
 549                                 }
 550
 551 reject:
 552                                 if (reject || collide) {
 553                                         ftotal++;
 554                                         flocal++;
 555
 556                                         if (collide && flocal <= local_retries)
 557                                                 /* retry locally a few times */
 558                                                 retry_bucket = 1;
 559                                         else if (local_fallback_retries > 0 &&
 560                                                  flocal <= in->size + local_fallback_retries)
 561                                                 /* exhaustive bucket search */
 562                                                 retry_bucket = 1;
 563                                         else if (ftotal < tries)
 564                                                 /* then retry descent */
 565                                                 retry_descent = 1;
 566                                         else
 567                                                 /* else give up */
 568                                                 skip_rep = 1;
 569                                         dprintk("  reject %d  collide %d  "
 570                                                 "ftotal %u  flocal %u\n",
 571                                                 reject, collide, ftotal,
 572                                                 flocal);
 573                                 }
 574                         } while (retry_bucket);
 575                 } while (retry_descent);
 576
 577                 if (skip_rep) {
 578                         dprintk("skip rep\n");
 579                         continue;
 580                 }
 581
 582                 dprintk("CHOOSE got %d\n", item);
 583                 out[outpos] = item;
 584                 outpos++;
 585                 count--;
 586 #ifndef __KERNEL__
 587                 if (map->choose_tries && ftotal <= map->choose_total_tries)
 588                         map->choose_tries[ftotal]++;
 589 #endif
 590         }
 591
 592         dprintk("CHOOSE returns %d\n", outpos);
 593         return outpos;
 594 }
 595
 596
 597 /**
 598  * crush_choose_indep: alternative breadth-first positionally stable mapping
 599  *
 600  */
 601 static void crush_choose_indep(const struct crush_map *map,
 602                                struct crush_bucket *bucket,
 603                                const __u32 *weight, int weight_max,
 604                                int x, int left, int numrep, int type,
 605                                int *out, int outpos,
 606                                unsigned int tries,
 607                                unsigned int recurse_tries,
 608                                int recurse_to_leaf,
 609                                int *out2,
 610                                int parent_r)
 611 {
 612         struct crush_bucket *in = bucket;
 613         int endpos = outpos + left;
 614         int rep;
 615         unsigned int ftotal;
 616         int r;
 617         int i;
 618         int item = 0;
 619         int itemtype;
 620         int collide;
 621
 622         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 623                 bucket->id, x, outpos, numrep);
 624
 625         /* initially my result is undefined */
 626         for (rep = outpos; rep < endpos; rep++) {
 627                 out[rep] = CRUSH_ITEM_UNDEF;
 628                 if (out2)
 629                         out2[rep] = CRUSH_ITEM_UNDEF;
 630         }
 631
 632         for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) {
 633 #ifdef DEBUG_INDEP
 634                 if (out2 && ftotal) {
 635                         dprintk("%u %d a: ", ftotal, left);
 636                         for (rep = outpos; rep < endpos; rep++) {
 637                                 dprintk(" %d", out[rep]);
 638                         }
 639                         dprintk("\n");
 640                         dprintk("%u %d b: ", ftotal, left);
 641                         for (rep = outpos; rep < endpos; rep++) {
 642                                 dprintk(" %d", out2[rep]);
 643                         }
 644                         dprintk("\n");
 645                 }
 646 #endif
 647                 for (rep = outpos; rep < endpos; rep++) {
 648                         if (out[rep] != CRUSH_ITEM_UNDEF)
 649                                 continue;
 650
 651                         in = bucket;  /* initial bucket */
 652
 653                         /* choose through intervening buckets */
 654                         for (;;) {
 655                                 /* note: we base the choice on the position
 656                                  * even in the nested call.  that means that
 657                                  * if the first layer chooses the same bucket
 658                                  * in a different position, we will tend to
 659                                  * choose a different item in that bucket.
 660                                  * this will involve more devices in data
 661                                  * movement and tend to distribute the load.
 662                                  */
 663                                 r = rep + parent_r;
 664
 665                                 /* be careful */
 666                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
 667                                     in->size % numrep == 0)
 668                                         /* r'=r+(n+1)*f_total */
 669                                         r += (numrep+1) * ftotal;
 670                                 else
 671                                         /* r' = r + n*f_total */
 672                                         r += numrep * ftotal;
 673
 674                                 /* bucket choose */
 675                                 if (in->size == 0) {
 676                                         dprintk("   empty bucket\n");
 677                                         break;
 678                                 }
 679
 680                                 item = crush_bucket_choose(in, x, r);
 681                                 if (item >= map->max_devices) {
 682                                         dprintk("   bad item %d\n", item);
 683                                         out[rep] = CRUSH_ITEM_NONE;
 684                                         if (out2)
 685                                                 out2[rep] = CRUSH_ITEM_NONE;
 686                                         left--;
 687                                         break;
 688                                 }
 689
 690                                 /* desired type? */
 691                                 if (item < 0)
 692                                         itemtype = map->buckets[-1-item]->type;
 693                                 else
 694                                         itemtype = 0;
 695                                 dprintk("  item %d type %d\n", item, itemtype);
 696
 697                                 /* keep going? */
 698                                 if (itemtype != type) {
 699                                         if (item >= 0 ||
 700                                             (-1-item) >= map->max_buckets) {
 701                                                 dprintk("   bad item type %d\n", type);
 702                                                 out[rep] = CRUSH_ITEM_NONE;
 703                                                 if (out2)
 704                                                         out2[rep] =
 705                                                                 CRUSH_ITEM_NONE;
 706                                                 left--;
 707                                                 break;
 708                                         }
 709                                         in = map->buckets[-1-item];
 710                                         continue;
 711                                 }
 712
 713                                 /* collision? */
 714                                 collide = 0;
 715                                 for (i = outpos; i < endpos; i++) {
 716                                         if (out[i] == item) {
 717                                                 collide = 1;
 718                                                 break;
 719                                         }
 720                                 }
 721                                 if (collide)
 722                                         break;
 723
 724                                 if (recurse_to_leaf) {
 725                                         if (item < 0) {
 726                                                 crush_choose_indep(map,
 727                                                    map->buckets[-1-item],
 728                                                    weight, weight_max,
 729                                                    x, 1, numrep, 0,
 730                                                    out2, rep,
 731                                                    recurse_tries, 0,
 732                                                    0, NULL, r);
 733                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 734                                                         /* placed nothing; no leaf */
 735                                                         break;
 736                                                 }
 737                                         } else {
 738                                                 /* we already have a leaf! */
 739                                                 out2[rep] = item;
 740                                         }
 741                                 }
 742
 743                                 /* out? */
 744                                 if (itemtype == 0 &&
 745                                     is_out(map, weight, weight_max, item, x))
 746                                         break;
 747
 748                                 /* yay! */
 749                                 out[rep] = item;
 750                                 left--;
 751                                 break;
 752                         }
 753                 }
 754         }
 755         for (rep = outpos; rep < endpos; rep++) {
 756                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 757                         out[rep] = CRUSH_ITEM_NONE;
 758                 }
 759                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 760                         out2[rep] = CRUSH_ITEM_NONE;
 761                 }
 762         }
 763 #ifndef __KERNEL__
 764         if (map->choose_tries && ftotal <= map->choose_total_tries)
 765                 map->choose_tries[ftotal]++;
 766 #endif
 767 #ifdef DEBUG_INDEP
 768         if (out2) {
 769                 dprintk("%u %d a: ", ftotal, left);
 770                 for (rep = outpos; rep < endpos; rep++) {
 771                         dprintk(" %d", out[rep]);
 772                 }
 773                 dprintk("\n");
 774                 dprintk("%u %d b: ", ftotal, left);
 775                 for (rep = outpos; rep < endpos; rep++) {
 776                         dprintk(" %d", out2[rep]);
 777                 }
 778                 dprintk("\n");
 779         }
 780 #endif
 781 }
 782
 783 /**
 784  * crush_do_rule - calculate a mapping with the given input and rule
 785  * @map: the crush_map
 786  * @ruleno: the rule id
 787  * @x: hash input
 788  * @result: pointer to result vector
 789  * @result_max: maximum result size
 790  * @weight: weight vector (for map leaves)
 791  * @weight_max: size of weight vector
 792  * @scratch: scratch vector for private use; must be >= 3 * result_max
 793  */
 794 int crush_do_rule(const struct crush_map *map,
 795                   int ruleno, int x, int *result, int result_max,
 796                   const __u32 *weight, int weight_max,
 797                   int *scratch)
 798 {
 799         int result_len;
 800         int *a = scratch;
 801         int *b = scratch + result_max;
 802         int *c = scratch + result_max*2;
 803         int recurse_to_leaf;
 804         int *w;
 805         int wsize = 0;
 806         int *o;
 807         int osize;
 808         int *tmp;
 809         struct crush_rule *rule;
 810         __u32 step;
 811         int i, j;
 812         int numrep;
 813         int out_size;
 814         /*
 815          * the original choose_total_tries value was off by one (it
 816          * counted "retries" and not "tries").  add one.
 817          */
 818         int choose_tries = map->choose_total_tries + 1;
 819         int choose_leaf_tries = 0;
 820         /*
 821          * the local tries values were counted as "retries", though,
 822          * and need no adjustment
 823          */
 824         int choose_local_retries = map->choose_local_tries;
 825         int choose_local_fallback_retries = map->choose_local_fallback_tries;
 826
 827         int vary_r = map->chooseleaf_vary_r;
 828         int stable = map->chooseleaf_stable;
 829
 830         if ((__u32)ruleno >= map->max_rules) {
 831                 dprintk(" bad ruleno %d\n", ruleno);
 832                 return 0;
 833         }
 834
 835         rule = map->rules[ruleno];
 836         result_len = 0;
 837         w = a;
 838         o = b;
 839
 840         for (step = 0; step < rule->len; step++) {
 841                 int firstn = 0;
 842                 struct crush_rule_step *curstep = &rule->steps[step];
 843
 844                 switch (curstep->op) {
 845                 case CRUSH_RULE_TAKE:
 846                         if ((curstep->arg1 >= 0 &&
 847                              curstep->arg1 < map->max_devices) ||
 848                             (-1-curstep->arg1 >= 0 &&
 849                              -1-curstep->arg1 < map->max_buckets &&
 850                              map->buckets[-1-curstep->arg1])) {
 851                                 w[0] = curstep->arg1;
 852                                 wsize = 1;
 853                         } else {
 854                                 dprintk(" bad take value %d\n", curstep->arg1);
 855                         }
 856                         break;
 857
 858                 case CRUSH_RULE_SET_CHOOSE_TRIES:
 859                         if (curstep->arg1 > 0)
 860                                 choose_tries = curstep->arg1;
 861                         break;
 862
 863                 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
 864                         if (curstep->arg1 > 0)
 865                                 choose_leaf_tries = curstep->arg1;
 866                         break;
 867
 868                 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
 869                         if (curstep->arg1 >= 0)
 870                                 choose_local_retries = curstep->arg1;
 871                         break;
 872
 873                 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
 874                         if (curstep->arg1 >= 0)
 875                                 choose_local_fallback_retries = curstep->arg1;
 876                         break;
 877
 878                 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R:
 879                         if (curstep->arg1 >= 0)
 880                                 vary_r = curstep->arg1;
 881                         break;
 882
 883                 case CRUSH_RULE_SET_CHOOSELEAF_STABLE:
 884                         if (curstep->arg1 >= 0)
 885                                 stable = curstep->arg1;
 886                         break;
 887
 888                 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
 889                 case CRUSH_RULE_CHOOSE_FIRSTN:
 890                         firstn = 1;
 891                         /* fall through */
 892                 case CRUSH_RULE_CHOOSELEAF_INDEP:
 893                 case CRUSH_RULE_CHOOSE_INDEP:
 894                         if (wsize == 0)
 895                                 break;
 896
 897                         recurse_to_leaf =
 898                                 curstep->op ==
 899                                  CRUSH_RULE_CHOOSELEAF_FIRSTN ||
 900                                 curstep->op ==
 901                                 CRUSH_RULE_CHOOSELEAF_INDEP;
 902
 903                         /* reset output */
 904                         osize = 0;
 905
 906                         for (i = 0; i < wsize; i++) {
 907                                 int bno;
 908                                 /*
 909                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 910                                  * basically, numrep <= 0 means relative to
 911                                  * the provided result_max
 912                                  */
 913                                 numrep = curstep->arg1;
 914                                 if (numrep <= 0) {
 915                                         numrep += result_max;
 916                                         if (numrep <= 0)
 917                                                 continue;
 918                                 }
 919                                 j = 0;
 920                                 /* make sure bucket id is valid */
 921                                 bno = -1 - w[i];
 922                                 if (bno < 0 || bno >= map->max_buckets) {
 923                                         /* w[i] is probably CRUSH_ITEM_NONE */
 924                                         dprintk("  bad w[i] %d\n", w[i]);
 925                                         continue;
 926                                 }
 927                                 if (firstn) {
 928                                         int recurse_tries;
 929                                         if (choose_leaf_tries)
 930                                                 recurse_tries =
 931                                                         choose_leaf_tries;
 932                                         else if (map->chooseleaf_descend_once)
 933                                                 recurse_tries = 1;
 934                                         else
 935                                                 recurse_tries = choose_tries;
 936                                         osize += crush_choose_firstn(
 937                                                 map,
 938                                                 map->buckets[bno],
 939                                                 weight, weight_max,
 940                                                 x, numrep,
 941                                                 curstep->arg2,
 942                                                 o+osize, j,
 943                                                 result_max-osize,
 944                                                 choose_tries,
 945                                                 recurse_tries,
 946                                                 choose_local_retries,
 947                                                 choose_local_fallback_retries,
 948                                                 recurse_to_leaf,
 949                                                 vary_r,
 950                                                 stable,
 951                                                 c+osize,
 952                                                 0);
 953                                 } else {
 954                                         out_size = ((numrep < (result_max-osize)) ?
 955                                                     numrep : (result_max-osize));
 956                                         crush_choose_indep(
 957                                                 map,
 958                                                 map->buckets[bno],
 959                                                 weight, weight_max,
 960                                                 x, out_size, numrep,
 961                                                 curstep->arg2,
 962                                                 o+osize, j,
 963                                                 choose_tries,
 964                                                 choose_leaf_tries ?
 965                                                    choose_leaf_tries : 1,
 966                                                 recurse_to_leaf,
 967                                                 c+osize,
 968                                                 0);
 969                                         osize += out_size;
 970                                 }
 971                         }
 972
 973                         if (recurse_to_leaf)
 974                                 /* copy final _leaf_ values to output set */
 975                                 memcpy(o, c, osize*sizeof(*o));
 976
 977                         /* swap o and w arrays */
 978                         tmp = o;
 979                         o = w;
 980                         w = tmp;
 981                         wsize = osize;
 982                         break;
 983
 984
 985                 case CRUSH_RULE_EMIT:
 986                         for (i = 0; i < wsize && result_len < result_max; i++) {
 987                                 result[result_len] = w[i];
 988                                 result_len++;
 989                         }
 990                         wsize = 0;
 991                         break;
 992
 993                 default:
 994                         dprintk(" unknown op %d at step %d\n",
 995                                 curstep->op, step);
 996                         break;
 997                 }
 998         }
 999         return result_len;
1000 }