net/ceph/crush/mapper.c

   1 /*
   2  * Ceph - scalable distributed file system
   3  *
   4  * Copyright (C) 2015 Intel Corporation All Rights Reserved
   5  *
   6  * This is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License version 2.1, as published by the Free Software
   9  * Foundation.  See file COPYING.
  10  *
  11  */
  12
  13 #ifdef __KERNEL__
  14 # include <linux/string.h>
  15 # include <linux/slab.h>
  16 # include <linux/bug.h>
  17 # include <linux/kernel.h>
  18 # include <linux/crush/crush.h>
  19 # include <linux/crush/hash.h>
  20 # include <linux/crush/mapper.h>
  21 #else
  22 # include "crush_compat.h"
  23 # include "crush.h"
  24 # include "hash.h"
  25 # include "mapper.h"
  26 #endif
  27 #include "crush_ln_table.h"
  28
  29 #define dprintk(args...) /* printf(args) */
  30
  31 /*
  32  * Implement the core CRUSH mapping algorithm.
  33  */
  34
  35 /**
  36  * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
  37  * @map: the crush_map
  38  * @ruleset: the storage ruleset id (user defined)
  39  * @type: storage ruleset type (user defined)
  40  * @size: output set size
  41  */
  42 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
  43 {
  44         __u32 i;
  45
  46         for (i = 0; i < map->max_rules; i++) {
  47                 if (map->rules[i] &&
  48                     map->rules[i]->mask.ruleset == ruleset &&
  49                     map->rules[i]->mask.type == type &&
  50                     map->rules[i]->mask.min_size <= size &&
  51                     map->rules[i]->mask.max_size >= size)
  52                         return i;
  53         }
  54         return -1;
  55 }
  56
  57 /*
  58  * bucket choose methods
  59  *
  60  * For each bucket algorithm, we have a "choose" method that, given a
  61  * crush input @x and replica position (usually, position in output set) @r,
  62  * will produce an item in the bucket.
  63  */
  64
  65 /*
  66  * Choose based on a random permutation of the bucket.
  67  *
  68  * We used to use some prime number arithmetic to do this, but it
  69  * wasn't very random, and had some other bad behaviors.  Instead, we
  70  * calculate an actual random permutation of the bucket members.
  71  * Since this is expensive, we optimize for the r=0 case, which
  72  * captures the vast majority of calls.
  73  */
  74 static int bucket_perm_choose(const struct crush_bucket *bucket,
  75                               struct crush_work_bucket *work,
  76                               int x, int r)
  77 {
  78         unsigned int pr = r % bucket->size;
  79         unsigned int i, s;
  80
  81         /* start a new permutation if @x has changed */
  82         if (work->perm_x != (__u32)x || work->perm_n == 0) {
  83                 dprintk("bucket %d new x=%d\n", bucket->id, x);
  84                 work->perm_x = x;
  85
  86                 /* optimize common r=0 case */
  87                 if (pr == 0) {
  88                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
  89                                 bucket->size;
  90                         work->perm[0] = s;
  91                         work->perm_n = 0xffff;   /* magic value, see below */
  92                         goto out;
  93                 }
  94
  95                 for (i = 0; i < bucket->size; i++)
  96                         work->perm[i] = i;
  97                 work->perm_n = 0;
  98         } else if (work->perm_n == 0xffff) {
  99                 /* clean up after the r=0 case above */
 100                 for (i = 1; i < bucket->size; i++)
 101                         work->perm[i] = i;
 102                 work->perm[work->perm[0]] = 0;
 103                 work->perm_n = 1;
 104         }
 105
 106         /* calculate permutation up to pr */
 107         for (i = 0; i < work->perm_n; i++)
 108                 dprintk(" perm_choose have %d: %d\n", i, work->perm[i]);
 109         while (work->perm_n <= pr) {
 110                 unsigned int p = work->perm_n;
 111                 /* no point in swapping the final entry */
 112                 if (p < bucket->size - 1) {
 113                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
 114                                 (bucket->size - p);
 115                         if (i) {
 116                                 unsigned int t = work->perm[p + i];
 117                                 work->perm[p + i] = work->perm[p];
 118                                 work->perm[p] = t;
 119                         }
 120                         dprintk(" perm_choose swap %d with %d\n", p, p+i);
 121                 }
 122                 work->perm_n++;
 123         }
 124         for (i = 0; i < bucket->size; i++)
 125                 dprintk(" perm_choose  %d: %d\n", i, work->perm[i]);
 126
 127         s = work->perm[pr];
 128 out:
 129         dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
 130                 bucket->size, x, r, pr, s);
 131         return bucket->items[s];
 132 }
 133
 134 /* uniform */
 135 static int bucket_uniform_choose(const struct crush_bucket_uniform *bucket,
 136                                  struct crush_work_bucket *work, int x, int r)
 137 {
 138         return bucket_perm_choose(&bucket->h, work, x, r);
 139 }
 140
 141 /* list */
 142 static int bucket_list_choose(const struct crush_bucket_list *bucket,
 143                               int x, int r)
 144 {
 145         int i;
 146
 147         for (i = bucket->h.size-1; i >= 0; i--) {
 148                 __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i],
 149                                          r, bucket->h.id);
 150                 w &= 0xffff;
 151                 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
 152                         "sw %x rand %llx",
 153                         i, x, r, bucket->h.items[i], bucket->item_weights[i],
 154                         bucket->sum_weights[i], w);
 155                 w *= bucket->sum_weights[i];
 156                 w = w >> 16;
 157                 /*dprintk(" scaled %llx\n", w);*/
 158                 if (w < bucket->item_weights[i]) {
 159                         return bucket->h.items[i];
 160                 }
 161         }
 162
 163         dprintk("bad list sums for bucket %d\n", bucket->h.id);
 164         return bucket->h.items[0];
 165 }
 166
 167
 168 /* (binary) tree */
 169 static int height(int n)
 170 {
 171         int h = 0;
 172         while ((n & 1) == 0) {
 173                 h++;
 174                 n = n >> 1;
 175         }
 176         return h;
 177 }
 178
 179 static int left(int x)
 180 {
 181         int h = height(x);
 182         return x - (1 << (h-1));
 183 }
 184
 185 static int right(int x)
 186 {
 187         int h = height(x);
 188         return x + (1 << (h-1));
 189 }
 190
 191 static int terminal(int x)
 192 {
 193         return x & 1;
 194 }
 195
 196 static int bucket_tree_choose(const struct crush_bucket_tree *bucket,
 197                               int x, int r)
 198 {
 199         int n;
 200         __u32 w;
 201         __u64 t;
 202
 203         /* start at root */
 204         n = bucket->num_nodes >> 1;
 205
 206         while (!terminal(n)) {
 207                 int l;
 208                 /* pick point in [0, w) */
 209                 w = bucket->node_weights[n];
 210                 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
 211                                           bucket->h.id) * (__u64)w;
 212                 t = t >> 32;
 213
 214                 /* descend to the left or right? */
 215                 l = left(n);
 216                 if (t < bucket->node_weights[l])
 217                         n = l;
 218                 else
 219                         n = right(n);
 220         }
 221
 222         return bucket->h.items[n >> 1];
 223 }
 224
 225
 226 /* straw */
 227
 228 static int bucket_straw_choose(const struct crush_bucket_straw *bucket,
 229                                int x, int r)
 230 {
 231         __u32 i;
 232         int high = 0;
 233         __u64 high_draw = 0;
 234         __u64 draw;
 235
 236         for (i = 0; i < bucket->h.size; i++) {
 237                 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
 238                 draw &= 0xffff;
 239                 draw *= bucket->straws[i];
 240                 if (i == 0 || draw > high_draw) {
 241                         high = i;
 242                         high_draw = draw;
 243                 }
 244         }
 245         return bucket->h.items[high];
 246 }
 247
 248 /* compute 2^44*log2(input+1) */
 249 static __u64 crush_ln(unsigned int xin)
 250 {
 251         unsigned int x = xin;
 252         int iexpon, index1, index2;
 253         __u64 RH, LH, LL, xl64, result;
 254
 255         x++;
 256
 257         /* normalize input */
 258         iexpon = 15;
 259
 260         /*
 261          * figure out number of bits we need to shift and
 262          * do it in one step instead of iteratively
 263          */
 264         if (!(x & 0x18000)) {
 265                 int bits = __builtin_clz(x & 0x1FFFF) - 16;
 266                 x <<= bits;
 267                 iexpon = 15 - bits;
 268         }
 269
 270         index1 = (x >> 8) << 1;
 271         /* RH ~ 2^56/index1 */
 272         RH = __RH_LH_tbl[index1 - 256];
 273         /* LH ~ 2^48 * log2(index1/256) */
 274         LH = __RH_LH_tbl[index1 + 1 - 256];
 275
 276         /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */
 277         xl64 = (__s64)x * RH;
 278         xl64 >>= 48;
 279
 280         result = iexpon;
 281         result <<= (12 + 32);
 282
 283         index2 = xl64 & 0xff;
 284         /* LL ~ 2^48*log2(1.0+index2/2^15) */
 285         LL = __LL_tbl[index2];
 286
 287         LH = LH + LL;
 288
 289         LH >>= (48 - 12 - 32);
 290         result += LH;
 291
 292         return result;
 293 }
 294
 295
 296 /*
 297  * straw2
 298  *
 299  * for reference, see:
 300  *
 301  * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables
 302  *
 303  */
 304
 305 static int bucket_straw2_choose(const struct crush_bucket_straw2 *bucket,
 306                                 int x, int r)
 307 {
 308         unsigned int i, high = 0;
 309         unsigned int u;
 310         unsigned int w;
 311         __s64 ln, draw, high_draw = 0;
 312
 313         for (i = 0; i < bucket->h.size; i++) {
 314                 w = bucket->item_weights[i];
 315                 if (w) {
 316                         u = crush_hash32_3(bucket->h.hash, x,
 317                                            bucket->h.items[i], r);
 318                         u &= 0xffff;
 319
 320                         /*
 321                          * for some reason slightly less than 0x10000 produces
 322                          * a slightly more accurate distribution... probably a
 323                          * rounding effect.
 324                          *
 325                          * the natural log lookup table maps [0,0xffff]
 326                          * (corresponding to real numbers [1/0x10000, 1] to
 327                          * [0, 0xffffffffffff] (corresponding to real numbers
 328                          * [-11.090355,0]).
 329                          */
 330                         ln = crush_ln(u) - 0x1000000000000ll;
 331
 332                         /*
 333                          * divide by 16.16 fixed-point weight.  note
 334                          * that the ln value is negative, so a larger
 335                          * weight means a larger (less negative) value
 336                          * for draw.
 337                          */
 338                         draw = div64_s64(ln, w);
 339                 } else {
 340                         draw = S64_MIN;
 341                 }
 342
 343                 if (i == 0 || draw > high_draw) {
 344                         high = i;
 345                         high_draw = draw;
 346                 }
 347         }
 348
 349         return bucket->h.items[high];
 350 }
 351
 352
 353 static int crush_bucket_choose(const struct crush_bucket *in,
 354                                struct crush_work_bucket *work,
 355                                int x, int r)
 356 {
 357         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
 358         BUG_ON(in->size == 0);
 359         switch (in->alg) {
 360         case CRUSH_BUCKET_UNIFORM:
 361                 return bucket_uniform_choose(
 362                         (const struct crush_bucket_uniform *)in,
 363                         work, x, r);
 364         case CRUSH_BUCKET_LIST:
 365                 return bucket_list_choose((const struct crush_bucket_list *)in,
 366                                           x, r);
 367         case CRUSH_BUCKET_TREE:
 368                 return bucket_tree_choose((const struct crush_bucket_tree *)in,
 369                                           x, r);
 370         case CRUSH_BUCKET_STRAW:
 371                 return bucket_straw_choose(
 372                         (const struct crush_bucket_straw *)in,
 373                         x, r);
 374         case CRUSH_BUCKET_STRAW2:
 375                 return bucket_straw2_choose(
 376                         (const struct crush_bucket_straw2 *)in,
 377                         x, r);
 378         default:
 379                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
 380                 return in->items[0];
 381         }
 382 }
 383
 384 /*
 385  * true if device is marked "out" (failed, fully offloaded)
 386  * of the cluster
 387  */
 388 static int is_out(const struct crush_map *map,
 389                   const __u32 *weight, int weight_max,
 390                   int item, int x)
 391 {
 392         if (item >= weight_max)
 393                 return 1;
 394         if (weight[item] >= 0x10000)
 395                 return 0;
 396         if (weight[item] == 0)
 397                 return 1;
 398         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
 399             < weight[item])
 400                 return 0;
 401         return 1;
 402 }
 403
 404 /**
 405  * crush_choose_firstn - choose numrep distinct items of given type
 406  * @map: the crush_map
 407  * @bucket: the bucket we are choose an item from
 408  * @x: crush input value
 409  * @numrep: the number of items to choose
 410  * @type: the type of item to choose
 411  * @out: pointer to output vector
 412  * @outpos: our position in that vector
 413  * @out_size: size of the out vector
 414  * @tries: number of attempts to make
 415  * @recurse_tries: number of attempts to have recursive chooseleaf make
 416  * @local_retries: localized retries
 417  * @local_fallback_retries: localized fallback retries
 418  * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose)
 419  * @stable: stable mode starts rep=0 in the recursive call for all replicas
 420  * @vary_r: pass r to recursive calls
 421  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 422  * @parent_r: r value passed from the parent
 423  */
 424 static int crush_choose_firstn(const struct crush_map *map,
 425                                struct crush_work *work,
 426                                const struct crush_bucket *bucket,
 427                                const __u32 *weight, int weight_max,
 428                                int x, int numrep, int type,
 429                                int *out, int outpos,
 430                                int out_size,
 431                                unsigned int tries,
 432                                unsigned int recurse_tries,
 433                                unsigned int local_retries,
 434                                unsigned int local_fallback_retries,
 435                                int recurse_to_leaf,
 436                                unsigned int vary_r,
 437                                unsigned int stable,
 438                                int *out2,
 439                                int parent_r)
 440 {
 441         int rep;
 442         unsigned int ftotal, flocal;
 443         int retry_descent, retry_bucket, skip_rep;
 444         const struct crush_bucket *in = bucket;
 445         int r;
 446         int i;
 447         int item = 0;
 448         int itemtype;
 449         int collide, reject;
 450         int count = out_size;
 451
 452         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",
 453                 recurse_to_leaf ? "_LEAF" : "",
 454                 bucket->id, x, outpos, numrep,
 455                 tries, recurse_tries, local_retries, local_fallback_retries,
 456                 parent_r, stable);
 457
 458         for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) {
 459                 /* keep trying until we get a non-out, non-colliding item */
 460                 ftotal = 0;
 461                 skip_rep = 0;
 462                 do {
 463                         retry_descent = 0;
 464                         in = bucket;               /* initial bucket */
 465
 466                         /* choose through intervening buckets */
 467                         flocal = 0;
 468                         do {
 469                                 collide = 0;
 470                                 retry_bucket = 0;
 471                                 r = rep + parent_r;
 472                                 /* r' = r + f_total */
 473                                 r += ftotal;
 474
 475                                 /* bucket choose */
 476                                 if (in->size == 0) {
 477                                         reject = 1;
 478                                         goto reject;
 479                                 }
 480                                 if (local_fallback_retries > 0 &&
 481                                     flocal >= (in->size>>1) &&
 482                                     flocal > local_fallback_retries)
 483                                         item = bucket_perm_choose(
 484                                                 in, work->work[-1-in->id],
 485                                                 x, r);
 486                                 else
 487                                         item = crush_bucket_choose(
 488                                                 in, work->work[-1-in->id],
 489                                                 x, r);
 490                                 if (item >= map->max_devices) {
 491                                         dprintk("   bad item %d\n", item);
 492                                         skip_rep = 1;
 493                                         break;
 494                                 }
 495
 496                                 /* desired type? */
 497                                 if (item < 0)
 498                                         itemtype = map->buckets[-1-item]->type;
 499                                 else
 500                                         itemtype = 0;
 501                                 dprintk("  item %d type %d\n", item, itemtype);
 502
 503                                 /* keep going? */
 504                                 if (itemtype != type) {
 505                                         if (item >= 0 ||
 506                                             (-1-item) >= map->max_buckets) {
 507                                                 dprintk("   bad item type %d\n", type);
 508                                                 skip_rep = 1;
 509                                                 break;
 510                                         }
 511                                         in = map->buckets[-1-item];
 512                                         retry_bucket = 1;
 513                                         continue;
 514                                 }
 515
 516                                 /* collision? */
 517                                 for (i = 0; i < outpos; i++) {
 518                                         if (out[i] == item) {
 519                                                 collide = 1;
 520                                                 break;
 521                                         }
 522                                 }
 523
 524                                 reject = 0;
 525                                 if (!collide && recurse_to_leaf) {
 526                                         if (item < 0) {
 527                                                 int sub_r;
 528                                                 if (vary_r)
 529                                                         sub_r = r >> (vary_r-1);
 530                                                 else
 531                                                         sub_r = 0;
 532                                                 if (crush_choose_firstn(
 533                                                             map,
 534                                                             work,
 535                                                             map->buckets[-1-item],
 536                                                             weight, weight_max,
 537                                                             x, stable ? 1 : outpos+1, 0,
 538                                                             out2, outpos, count,
 539                                                             recurse_tries, 0,
 540                                                             local_retries,
 541                                                             local_fallback_retries,
 542                                                             0,
 543                                                             vary_r,
 544                                                             stable,
 545                                                             NULL,
 546                                                             sub_r) <= outpos)
 547                                                         /* didn't get leaf */
 548                                                         reject = 1;
 549                                         } else {
 550                                                 /* we already have a leaf! */
 551                                                 out2[outpos] = item;
 552                                         }
 553                                 }
 554
 555                                 if (!reject && !collide) {
 556                                         /* out? */
 557                                         if (itemtype == 0)
 558                                                 reject = is_out(map, weight,
 559                                                                 weight_max,
 560                                                                 item, x);
 561                                 }
 562
 563 reject:
 564                                 if (reject || collide) {
 565                                         ftotal++;
 566                                         flocal++;
 567
 568                                         if (collide && flocal <= local_retries)
 569                                                 /* retry locally a few times */
 570                                                 retry_bucket = 1;
 571                                         else if (local_fallback_retries > 0 &&
 572                                                  flocal <= in->size + local_fallback_retries)
 573                                                 /* exhaustive bucket search */
 574                                                 retry_bucket = 1;
 575                                         else if (ftotal < tries)
 576                                                 /* then retry descent */
 577                                                 retry_descent = 1;
 578                                         else
 579                                                 /* else give up */
 580                                                 skip_rep = 1;
 581                                         dprintk("  reject %d  collide %d  "
 582                                                 "ftotal %u  flocal %u\n",
 583                                                 reject, collide, ftotal,
 584                                                 flocal);
 585                                 }
 586                         } while (retry_bucket);
 587                 } while (retry_descent);
 588
 589                 if (skip_rep) {
 590                         dprintk("skip rep\n");
 591                         continue;
 592                 }
 593
 594                 dprintk("CHOOSE got %d\n", item);
 595                 out[outpos] = item;
 596                 outpos++;
 597                 count--;
 598 #ifndef __KERNEL__
 599                 if (map->choose_tries && ftotal <= map->choose_total_tries)
 600                         map->choose_tries[ftotal]++;
 601 #endif
 602         }
 603
 604         dprintk("CHOOSE returns %d\n", outpos);
 605         return outpos;
 606 }
 607
 608
 609 /**
 610  * crush_choose_indep: alternative breadth-first positionally stable mapping
 611  *
 612  */
 613 static void crush_choose_indep(const struct crush_map *map,
 614                                struct crush_work *work,
 615                                const struct crush_bucket *bucket,
 616                                const __u32 *weight, int weight_max,
 617                                int x, int left, int numrep, int type,
 618                                int *out, int outpos,
 619                                unsigned int tries,
 620                                unsigned int recurse_tries,
 621                                int recurse_to_leaf,
 622                                int *out2,
 623                                int parent_r)
 624 {
 625         const struct crush_bucket *in = bucket;
 626         int endpos = outpos + left;
 627         int rep;
 628         unsigned int ftotal;
 629         int r;
 630         int i;
 631         int item = 0;
 632         int itemtype;
 633         int collide;
 634
 635         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 636                 bucket->id, x, outpos, numrep);
 637
 638         /* initially my result is undefined */
 639         for (rep = outpos; rep < endpos; rep++) {
 640                 out[rep] = CRUSH_ITEM_UNDEF;
 641                 if (out2)
 642                         out2[rep] = CRUSH_ITEM_UNDEF;
 643         }
 644
 645         for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) {
 646 #ifdef DEBUG_INDEP
 647                 if (out2 && ftotal) {
 648                         dprintk("%u %d a: ", ftotal, left);
 649                         for (rep = outpos; rep < endpos; rep++) {
 650                                 dprintk(" %d", out[rep]);
 651                         }
 652                         dprintk("\n");
 653                         dprintk("%u %d b: ", ftotal, left);
 654                         for (rep = outpos; rep < endpos; rep++) {
 655                                 dprintk(" %d", out2[rep]);
 656                         }
 657                         dprintk("\n");
 658                 }
 659 #endif
 660                 for (rep = outpos; rep < endpos; rep++) {
 661                         if (out[rep] != CRUSH_ITEM_UNDEF)
 662                                 continue;
 663
 664                         in = bucket;  /* initial bucket */
 665
 666                         /* choose through intervening buckets */
 667                         for (;;) {
 668                                 /* note: we base the choice on the position
 669                                  * even in the nested call.  that means that
 670                                  * if the first layer chooses the same bucket
 671                                  * in a different position, we will tend to
 672                                  * choose a different item in that bucket.
 673                                  * this will involve more devices in data
 674                                  * movement and tend to distribute the load.
 675                                  */
 676                                 r = rep + parent_r;
 677
 678                                 /* be careful */
 679                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
 680                                     in->size % numrep == 0)
 681                                         /* r'=r+(n+1)*f_total */
 682                                         r += (numrep+1) * ftotal;
 683                                 else
 684                                         /* r' = r + n*f_total */
 685                                         r += numrep * ftotal;
 686
 687                                 /* bucket choose */
 688                                 if (in->size == 0) {
 689                                         dprintk("   empty bucket\n");
 690                                         break;
 691                                 }
 692
 693                                 item = crush_bucket_choose(
 694                                         in, work->work[-1-in->id],
 695                                         x, r);
 696                                 if (item >= map->max_devices) {
 697                                         dprintk("   bad item %d\n", item);
 698                                         out[rep] = CRUSH_ITEM_NONE;
 699                                         if (out2)
 700                                                 out2[rep] = CRUSH_ITEM_NONE;
 701                                         left--;
 702                                         break;
 703                                 }
 704
 705                                 /* desired type? */
 706                                 if (item < 0)
 707                                         itemtype = map->buckets[-1-item]->type;
 708                                 else
 709                                         itemtype = 0;
 710                                 dprintk("  item %d type %d\n", item, itemtype);
 711
 712                                 /* keep going? */
 713                                 if (itemtype != type) {
 714                                         if (item >= 0 ||
 715                                             (-1-item) >= map->max_buckets) {
 716                                                 dprintk("   bad item type %d\n", type);
 717                                                 out[rep] = CRUSH_ITEM_NONE;
 718                                                 if (out2)
 719                                                         out2[rep] =
 720                                                                 CRUSH_ITEM_NONE;
 721                                                 left--;
 722                                                 break;
 723                                         }
 724                                         in = map->buckets[-1-item];
 725                                         continue;
 726                                 }
 727
 728                                 /* collision? */
 729                                 collide = 0;
 730                                 for (i = outpos; i < endpos; i++) {
 731                                         if (out[i] == item) {
 732                                                 collide = 1;
 733                                                 break;
 734                                         }
 735                                 }
 736                                 if (collide)
 737                                         break;
 738
 739                                 if (recurse_to_leaf) {
 740                                         if (item < 0) {
 741                                                 crush_choose_indep(
 742                                                         map,
 743                                                         work,
 744                                                         map->buckets[-1-item],
 745                                                         weight, weight_max,
 746                                                         x, 1, numrep, 0,
 747                                                         out2, rep,
 748                                                         recurse_tries, 0,
 749                                                         0, NULL, r);
 750                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
 751                                                         /* placed nothing; no leaf */
 752                                                         break;
 753                                                 }
 754                                         } else {
 755                                                 /* we already have a leaf! */
 756                                                 out2[rep] = item;
 757                                         }
 758                                 }
 759
 760                                 /* out? */
 761                                 if (itemtype == 0 &&
 762                                     is_out(map, weight, weight_max, item, x))
 763                                         break;
 764
 765                                 /* yay! */
 766                                 out[rep] = item;
 767                                 left--;
 768                                 break;
 769                         }
 770                 }
 771         }
 772         for (rep = outpos; rep < endpos; rep++) {
 773                 if (out[rep] == CRUSH_ITEM_UNDEF) {
 774                         out[rep] = CRUSH_ITEM_NONE;
 775                 }
 776                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
 777                         out2[rep] = CRUSH_ITEM_NONE;
 778                 }
 779         }
 780 #ifndef __KERNEL__
 781         if (map->choose_tries && ftotal <= map->choose_total_tries)
 782                 map->choose_tries[ftotal]++;
 783 #endif
 784 #ifdef DEBUG_INDEP
 785         if (out2) {
 786                 dprintk("%u %d a: ", ftotal, left);
 787                 for (rep = outpos; rep < endpos; rep++) {
 788                         dprintk(" %d", out[rep]);
 789                 }
 790                 dprintk("\n");
 791                 dprintk("%u %d b: ", ftotal, left);
 792                 for (rep = outpos; rep < endpos; rep++) {
 793                         dprintk(" %d", out2[rep]);
 794                 }
 795                 dprintk("\n");
 796         }
 797 #endif
 798 }
 799
 800
 801 /*
 802  * This takes a chunk of memory and sets it up to be a shiny new
 803  * working area for a CRUSH placement computation. It must be called
 804  * on any newly allocated memory before passing it in to
 805  * crush_do_rule. It may be used repeatedly after that, so long as the
 806  * map has not changed. If the map /has/ changed, you must make sure
 807  * the working size is no smaller than what was allocated and re-run
 808  * crush_init_workspace.
 809  *
 810  * If you do retain the working space between calls to crush, make it
 811  * thread-local.
 812  */
 813 void crush_init_workspace(const struct crush_map *map, void *v)
 814 {
 815         struct crush_work *w = v;
 816         __s32 b;
 817
 818         /*
 819          * We work by moving through the available space and setting
 820          * values and pointers as we go.
 821          *
 822          * It's a bit like Forth's use of the 'allot' word since we
 823          * set the pointer first and then reserve the space for it to
 824          * point to by incrementing the point.
 825          */
 826         v += sizeof(struct crush_work *);
 827         w->work = v;
 828         v += map->max_buckets * sizeof(struct crush_work_bucket *);
 829         for (b = 0; b < map->max_buckets; ++b) {
 830                 if (!map->buckets[b])
 831                         continue;
 832
 833                 w->work[b] = v;
 834                 switch (map->buckets[b]->alg) {
 835                 default:
 836                         v += sizeof(struct crush_work_bucket);
 837                         break;
 838                 }
 839                 w->work[b]->perm_x = 0;
 840                 w->work[b]->perm_n = 0;
 841                 w->work[b]->perm = v;
 842                 v += map->buckets[b]->size * sizeof(__u32);
 843         }
 844         BUG_ON(v - (void *)w != map->working_size);
 845 }
 846
 847 /**
 848  * crush_do_rule - calculate a mapping with the given input and rule
 849  * @map: the crush_map
 850  * @ruleno: the rule id
 851  * @x: hash input
 852  * @result: pointer to result vector
 853  * @result_max: maximum result size
 854  * @weight: weight vector (for map leaves)
 855  * @weight_max: size of weight vector
 856  * @cwin: pointer to at least crush_work_size() bytes of memory
 857  */
 858 int crush_do_rule(const struct crush_map *map,
 859                   int ruleno, int x, int *result, int result_max,
 860                   const __u32 *weight, int weight_max,
 861                   void *cwin)
 862 {
 863         int result_len;
 864         struct crush_work *cw = cwin;
 865         int *a = cwin + map->working_size;
 866         int *b = a + result_max;
 867         int *c = b + result_max;
 868         int *w = a;
 869         int *o = b;
 870         int recurse_to_leaf;
 871         int wsize = 0;
 872         int osize;
 873         int *tmp;
 874         const struct crush_rule *rule;
 875         __u32 step;
 876         int i, j;
 877         int numrep;
 878         int out_size;
 879         /*
 880          * the original choose_total_tries value was off by one (it
 881          * counted "retries" and not "tries").  add one.
 882          */
 883         int choose_tries = map->choose_total_tries + 1;
 884         int choose_leaf_tries = 0;
 885         /*
 886          * the local tries values were counted as "retries", though,
 887          * and need no adjustment
 888          */
 889         int choose_local_retries = map->choose_local_tries;
 890         int choose_local_fallback_retries = map->choose_local_fallback_tries;
 891
 892         int vary_r = map->chooseleaf_vary_r;
 893         int stable = map->chooseleaf_stable;
 894
 895         if ((__u32)ruleno >= map->max_rules) {
 896                 dprintk(" bad ruleno %d\n", ruleno);
 897                 return 0;
 898         }
 899
 900         rule = map->rules[ruleno];
 901         result_len = 0;
 902
 903         for (step = 0; step < rule->len; step++) {
 904                 int firstn = 0;
 905                 const struct crush_rule_step *curstep = &rule->steps[step];
 906
 907                 switch (curstep->op) {
 908                 case CRUSH_RULE_TAKE:
 909                         if ((curstep->arg1 >= 0 &&
 910                              curstep->arg1 < map->max_devices) ||
 911                             (-1-curstep->arg1 >= 0 &&
 912                              -1-curstep->arg1 < map->max_buckets &&
 913                              map->buckets[-1-curstep->arg1])) {
 914                                 w[0] = curstep->arg1;
 915                                 wsize = 1;
 916                         } else {
 917                                 dprintk(" bad take value %d\n", curstep->arg1);
 918                         }
 919                         break;
 920
 921                 case CRUSH_RULE_SET_CHOOSE_TRIES:
 922                         if (curstep->arg1 > 0)
 923                                 choose_tries = curstep->arg1;
 924                         break;
 925
 926                 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
 927                         if (curstep->arg1 > 0)
 928                                 choose_leaf_tries = curstep->arg1;
 929                         break;
 930
 931                 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
 932                         if (curstep->arg1 >= 0)
 933                                 choose_local_retries = curstep->arg1;
 934                         break;
 935
 936                 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
 937                         if (curstep->arg1 >= 0)
 938                                 choose_local_fallback_retries = curstep->arg1;
 939                         break;
 940
 941                 case CRUSH_RULE_SET_CHOOSELEAF_VARY_R:
 942                         if (curstep->arg1 >= 0)
 943                                 vary_r = curstep->arg1;
 944                         break;
 945
 946                 case CRUSH_RULE_SET_CHOOSELEAF_STABLE:
 947                         if (curstep->arg1 >= 0)
 948                                 stable = curstep->arg1;
 949                         break;
 950
 951                 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
 952                 case CRUSH_RULE_CHOOSE_FIRSTN:
 953                         firstn = 1;
 954                         /* fall through */
 955                 case CRUSH_RULE_CHOOSELEAF_INDEP:
 956                 case CRUSH_RULE_CHOOSE_INDEP:
 957                         if (wsize == 0)
 958                                 break;
 959
 960                         recurse_to_leaf =
 961                                 curstep->op ==
 962                                  CRUSH_RULE_CHOOSELEAF_FIRSTN ||
 963                                 curstep->op ==
 964                                 CRUSH_RULE_CHOOSELEAF_INDEP;
 965
 966                         /* reset output */
 967                         osize = 0;
 968
 969                         for (i = 0; i < wsize; i++) {
 970                                 int bno;
 971                                 /*
 972                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 973                                  * basically, numrep <= 0 means relative to
 974                                  * the provided result_max
 975                                  */
 976                                 numrep = curstep->arg1;
 977                                 if (numrep <= 0) {
 978                                         numrep += result_max;
 979                                         if (numrep <= 0)
 980                                                 continue;
 981                                 }
 982                                 j = 0;
 983                                 /* make sure bucket id is valid */
 984                                 bno = -1 - w[i];
 985                                 if (bno < 0 || bno >= map->max_buckets) {
 986                                         /* w[i] is probably CRUSH_ITEM_NONE */
 987                                         dprintk("  bad w[i] %d\n", w[i]);
 988                                         continue;
 989                                 }
 990                                 if (firstn) {
 991                                         int recurse_tries;
 992                                         if (choose_leaf_tries)
 993                                                 recurse_tries =
 994                                                         choose_leaf_tries;
 995                                         else if (map->chooseleaf_descend_once)
 996                                                 recurse_tries = 1;
 997                                         else
 998                                                 recurse_tries = choose_tries;
 999                                         osize += crush_choose_firstn(
1000                                                 map,
1001                                                 cw,
1002                                                 map->buckets[bno],
1003                                                 weight, weight_max,
1004                                                 x, numrep,
1005                                                 curstep->arg2,
1006                                                 o+osize, j,
1007                                                 result_max-osize,
1008                                                 choose_tries,
1009                                                 recurse_tries,
1010                                                 choose_local_retries,
1011                                                 choose_local_fallback_retries,
1012                                                 recurse_to_leaf,
1013                                                 vary_r,
1014                                                 stable,
1015                                                 c+osize,
1016                                                 0);
1017                                 } else {
1018                                         out_size = ((numrep < (result_max-osize)) ?
1019                                                     numrep : (result_max-osize));
1020                                         crush_choose_indep(
1021                                                 map,
1022                                                 cw,
1023                                                 map->buckets[bno],
1024                                                 weight, weight_max,
1025                                                 x, out_size, numrep,
1026                                                 curstep->arg2,
1027                                                 o+osize, j,
1028                                                 choose_tries,
1029                                                 choose_leaf_tries ?
1030                                                    choose_leaf_tries : 1,
1031                                                 recurse_to_leaf,
1032                                                 c+osize,
1033                                                 0);
1034                                         osize += out_size;
1035                                 }
1036                         }
1037
1038                         if (recurse_to_leaf)
1039                                 /* copy final _leaf_ values to output set */
1040                                 memcpy(o, c, osize*sizeof(*o));
1041
1042                         /* swap o and w arrays */
1043                         tmp = o;
1044                         o = w;
1045                         w = tmp;
1046                         wsize = osize;
1047                         break;
1048
1049
1050                 case CRUSH_RULE_EMIT:
1051                         for (i = 0; i < wsize && result_len < result_max; i++) {
1052                                 result[result_len] = w[i];
1053                                 result_len++;
1054                         }
1055                         wsize = 0;
1056                         break;
1057
1058                 default:
1059                         dprintk(" unknown op %d at step %d\n",
1060                                 curstep->op, step);
1061                         break;
1062                 }
1063         }
1064
1065         return result_len;
1066 }