net/ceph/crush/mapper.c

   1
   2 #ifdef __KERNEL__
   3 # include <linux/string.h>
   4 # include <linux/slab.h>
   5 # include <linux/bug.h>
   6 # include <linux/kernel.h>
   7 # ifndef dprintk
   8 #  define dprintk(args...)
   9 # endif
  10 #else
  11 # include <string.h>
  12 # include <stdio.h>
  13 # include <stdlib.h>
  14 # include <assert.h>
  15 # define BUG_ON(x) assert(!(x))
  16 # define dprintk(args...) /* printf(args) */
  17 # define kmalloc(x, f) malloc(x)
  18 # define kfree(x) free(x)
  19 #endif
  20
  21 #include <linux/crush/crush.h>
  22 #include <linux/crush/hash.h>
  23
  24 /*
  25  * Implement the core CRUSH mapping algorithm.
  26  */
  27
  28 /**
  29  * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
  30  * @map: the crush_map
  31  * @ruleset: the storage ruleset id (user defined)
  32  * @type: storage ruleset type (user defined)
  33  * @size: output set size
  34  */
  35 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
  36 {
  37         __u32 i;
  38
  39         for (i = 0; i < map->max_rules; i++) {
  40                 if (map->rules[i] &&
  41                     map->rules[i]->mask.ruleset == ruleset &&
  42                     map->rules[i]->mask.type == type &&
  43                     map->rules[i]->mask.min_size <= size &&
  44                     map->rules[i]->mask.max_size >= size)
  45                         return i;
  46         }
  47         return -1;
  48 }
  49
  50
  51 /*
  52  * bucket choose methods
  53  *
  54  * For each bucket algorithm, we have a "choose" method that, given a
  55  * crush input @x and replica position (usually, position in output set) @r,
  56  * will produce an item in the bucket.
  57  */
  58
  59 /*
  60  * Choose based on a random permutation of the bucket.
  61  *
  62  * We used to use some prime number arithmetic to do this, but it
  63  * wasn't very random, and had some other bad behaviors.  Instead, we
  64  * calculate an actual random permutation of the bucket members.
  65  * Since this is expensive, we optimize for the r=0 case, which
  66  * captures the vast majority of calls.
  67  */
  68 static int bucket_perm_choose(struct crush_bucket *bucket,
  69                               int x, int r)
  70 {
  71         unsigned pr = r % bucket->size;
  72         unsigned i, s;
  73
  74         /* start a new permutation if @x has changed */
  75         if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) {
  76                 dprintk("bucket %d new x=%d\n", bucket->id, x);
  77                 bucket->perm_x = x;
  78
  79                 /* optimize common r=0 case */
  80                 if (pr == 0) {
  81                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
  82                                 bucket->size;
  83                         bucket->perm[0] = s;
  84                         bucket->perm_n = 0xffff;   /* magic value, see below */
  85                         goto out;
  86                 }
  87
  88                 for (i = 0; i < bucket->size; i++)
  89                         bucket->perm[i] = i;
  90                 bucket->perm_n = 0;
  91         } else if (bucket->perm_n == 0xffff) {
  92                 /* clean up after the r=0 case above */
  93                 for (i = 1; i < bucket->size; i++)
  94                         bucket->perm[i] = i;
  95                 bucket->perm[bucket->perm[0]] = 0;
  96                 bucket->perm_n = 1;
  97         }
  98
  99         /* calculate permutation up to pr */
 100         for (i = 0; i < bucket->perm_n; i++)
 101                 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]);
 102         while (bucket->perm_n <= pr) {
 103                 unsigned p = bucket->perm_n;
 104                 /* no point in swapping the final entry */
 105                 if (p < bucket->size - 1) {
 106                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
 107                                 (bucket->size - p);
 108                         if (i) {
 109                                 unsigned t = bucket->perm[p + i];
 110                                 bucket->perm[p + i] = bucket->perm[p];
 111                                 bucket->perm[p] = t;
 112                         }
 113                         dprintk(" perm_choose swap %d with %d\n", p, p+i);
 114                 }
 115                 bucket->perm_n++;
 116         }
 117         for (i = 0; i < bucket->size; i++)
 118                 dprintk(" perm_choose  %d: %d\n", i, bucket->perm[i]);
 119
 120         s = bucket->perm[pr];
 121 out:
 122         dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
 123                 bucket->size, x, r, pr, s);
 124         return bucket->items[s];
 125 }
 126
 127 /* uniform */
 128 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
 129                                  int x, int r)
 130 {
 131         return bucket_perm_choose(&bucket->h, x, r);
 132 }
 133
 134 /* list */
 135 static int bucket_list_choose(struct crush_bucket_list *bucket,
 136                               int x, int r)
 137 {
 138         int i;
 139
 140         for (i = bucket->h.size-1; i >= 0; i--) {
 141                 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
 142                                          r, bucket->h.id);
 143                 w &= 0xffff;
 144                 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
 145                         "sw %x rand %llx",
 146                         i, x, r, bucket->h.items[i], bucket->item_weights[i],
 147                         bucket->sum_weights[i], w);
 148                 w *= bucket->sum_weights[i];
 149                 w = w >> 16;
 150                 /*dprintk(" scaled %llx\n", w);*/
 151                 if (w < bucket->item_weights[i])
 152                         return bucket->h.items[i];
 153         }
 154
 155         dprintk("bad list sums for bucket %d\n", bucket->h.id);
 156         return bucket->h.items[0];
 157 }
 158
 159
 160 /* (binary) tree */
 161 static int height(int n)
 162 {
 163         int h = 0;
 164         while ((n & 1) == 0) {
 165                 h++;
 166                 n = n >> 1;
 167         }
 168         return h;
 169 }
 170
 171 static int left(int x)
 172 {
 173         int h = height(x);
 174         return x - (1 << (h-1));
 175 }
 176
 177 static int right(int x)
 178 {
 179         int h = height(x);
 180         return x + (1 << (h-1));
 181 }
 182
 183 static int terminal(int x)
 184 {
 185         return x & 1;
 186 }
 187
 188 static int bucket_tree_choose(struct crush_bucket_tree *bucket,
 189                               int x, int r)
 190 {
 191         int n, l;
 192         __u32 w;
 193         __u64 t;
 194
 195         /* start at root */
 196         n = bucket->num_nodes >> 1;
 197
 198         while (!terminal(n)) {
 199                 /* pick point in [0, w) */
 200                 w = bucket->node_weights[n];
 201                 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
 202                                           bucket->h.id) * (__u64)w;
 203                 t = t >> 32;
 204
 205                 /* descend to the left or right? */
 206                 l = left(n);
 207                 if (t < bucket->node_weights[l])
 208                         n = l;
 209                 else
 210                         n = right(n);
 211         }
 212
 213         return bucket->h.items[n >> 1];
 214 }
 215
 216
 217 /* straw */
 218
 219 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
 220                                int x, int r)
 221 {
 222         __u32 i;
 223         int high = 0;
 224         __u64 high_draw = 0;
 225         __u64 draw;
 226
 227         for (i = 0; i < bucket->h.size; i++) {
 228                 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
 229                 draw &= 0xffff;
 230                 draw *= bucket->straws[i];
 231                 if (i == 0 || draw > high_draw) {
 232                         high = i;
 233                         high_draw = draw;
 234                 }
 235         }
 236         return bucket->h.items[high];
 237 }
 238
 239 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
 240 {
 241         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
 242         BUG_ON(in->size == 0);
 243         switch (in->alg) {
 244         case CRUSH_BUCKET_UNIFORM:
 245                 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
 246                                           x, r);
 247         case CRUSH_BUCKET_LIST:
 248                 return bucket_list_choose((struct crush_bucket_list *)in,
 249                                           x, r);
 250         case CRUSH_BUCKET_TREE:
 251                 return bucket_tree_choose((struct crush_bucket_tree *)in,
 252                                           x, r);
 253         case CRUSH_BUCKET_STRAW:
 254                 return bucket_straw_choose((struct crush_bucket_straw *)in,
 255                                            x, r);
 256         default:
 257                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
 258                 return in->items[0];
 259         }
 260 }
 261
 262 /*
 263  * true if device is marked "out" (failed, fully offloaded)
 264  * of the cluster
 265  */
 266 static int is_out(const struct crush_map *map, const __u32 *weight, int item, int x)
 267 {
 268         if (weight[item] >= 0x10000)
 269                 return 0;
 270         if (weight[item] == 0)
 271                 return 1;
 272         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
 273             < weight[item])
 274                 return 0;
 275         return 1;
 276 }
 277
 278 /**
 279  * crush_choose - choose numrep distinct items of given type
 280  * @map: the crush_map
 281  * @bucket: the bucket we are choose an item from
 282  * @x: crush input value
 283  * @numrep: the number of items to choose
 284  * @type: the type of item to choose
 285  * @out: pointer to output vector
 286  * @outpos: our position in that vector
 287  * @firstn: true if choosing "first n" items, false if choosing "indep"
 288  * @recurse_to_leaf: true if we want one device under each item of given type
 289  * @out2: second output vector for leaf items (if @recurse_to_leaf)
 290  */
 291 static int crush_choose(const struct crush_map *map,
 292                         struct crush_bucket *bucket,
 293                         const __u32 *weight,
 294                         int x, int numrep, int type,
 295                         int *out, int outpos,
 296                         int firstn, int recurse_to_leaf,
 297                         int *out2)
 298 {
 299         int rep;
 300         unsigned int ftotal, flocal;
 301         int retry_descent, retry_bucket, skip_rep;
 302         struct crush_bucket *in = bucket;
 303         int r;
 304         int i;
 305         int item = 0;
 306         int itemtype;
 307         int collide, reject;
 308         const unsigned int orig_tries = 5; /* attempts before we fall back to search */
 309
 310         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
 311                 bucket->id, x, outpos, numrep);
 312
 313         for (rep = outpos; rep < numrep; rep++) {
 314                 /* keep trying until we get a non-out, non-colliding item */
 315                 ftotal = 0;
 316                 skip_rep = 0;
 317                 do {
 318                         retry_descent = 0;
 319                         in = bucket;               /* initial bucket */
 320
 321                         /* choose through intervening buckets */
 322                         flocal = 0;
 323                         do {
 324                                 collide = 0;
 325                                 retry_bucket = 0;
 326                                 r = rep;
 327                                 if (in->alg == CRUSH_BUCKET_UNIFORM) {
 328                                         /* be careful */
 329                                         if (firstn || (__u32)numrep >= in->size)
 330                                                 /* r' = r + f_total */
 331                                                 r += ftotal;
 332                                         else if (in->size % numrep == 0)
 333                                                 /* r'=r+(n+1)*f_local */
 334                                                 r += (numrep+1) *
 335                                                         (flocal+ftotal);
 336                                         else
 337                                                 /* r' = r + n*f_local */
 338                                                 r += numrep * (flocal+ftotal);
 339                                 } else {
 340                                         if (firstn)
 341                                                 /* r' = r + f_total */
 342                                                 r += ftotal;
 343                                         else
 344                                                 /* r' = r + n*f_local */
 345                                                 r += numrep * (flocal+ftotal);
 346                                 }
 347
 348                                 /* bucket choose */
 349                                 if (in->size == 0) {
 350                                         reject = 1;
 351                                         goto reject;
 352                                 }
 353                                 if (flocal >= (in->size>>1) &&
 354                                     flocal > orig_tries)
 355                                         item = bucket_perm_choose(in, x, r);
 356                                 else
 357                                         item = crush_bucket_choose(in, x, r);
 358                                 if (item >= map->max_devices) {
 359                                         dprintk("   bad item %d\n", item);
 360                                         skip_rep = 1;
 361                                         break;
 362                                 }
 363
 364                                 /* desired type? */
 365                                 if (item < 0)
 366                                         itemtype = map->buckets[-1-item]->type;
 367                                 else
 368                                         itemtype = 0;
 369                                 dprintk("  item %d type %d\n", item, itemtype);
 370
 371                                 /* keep going? */
 372                                 if (itemtype != type) {
 373                                         if (item >= 0 ||
 374                                             (-1-item) >= map->max_buckets) {
 375                                                 dprintk("   bad item type %d\n", type);
 376                                                 skip_rep = 1;
 377                                                 break;
 378                                         }
 379                                         in = map->buckets[-1-item];
 380                                         retry_bucket = 1;
 381                                         continue;
 382                                 }
 383
 384                                 /* collision? */
 385                                 for (i = 0; i < outpos; i++) {
 386                                         if (out[i] == item) {
 387                                                 collide = 1;
 388                                                 break;
 389                                         }
 390                                 }
 391
 392                                 reject = 0;
 393                                 if (recurse_to_leaf) {
 394                                         if (item < 0) {
 395                                                 if (crush_choose(map,
 396                                                          map->buckets[-1-item],
 397                                                          weight,
 398                                                          x, outpos+1, 0,
 399                                                          out2, outpos,
 400                                                          firstn, 0,
 401                                                          NULL) <= outpos)
 402                                                         /* didn't get leaf */
 403                                                         reject = 1;
 404                                         } else {
 405                                                 /* we already have a leaf! */
 406                                                 out2[outpos] = item;
 407                                         }
 408                                 }
 409
 410                                 if (!reject) {
 411                                         /* out? */
 412                                         if (itemtype == 0)
 413                                                 reject = is_out(map, weight,
 414                                                                 item, x);
 415                                         else
 416                                                 reject = 0;
 417                                 }
 418
 419 reject:
 420                                 if (reject || collide) {
 421                                         ftotal++;
 422                                         flocal++;
 423
 424                                         if (collide && flocal < 3)
 425                                                 /* retry locally a few times */
 426                                                 retry_bucket = 1;
 427                                         else if (flocal <= in->size + orig_tries)
 428                                                 /* exhaustive bucket search */
 429                                                 retry_bucket = 1;
 430                                         else if (ftotal < 20)
 431                                                 /* then retry descent */
 432                                                 retry_descent = 1;
 433                                         else
 434                                                 /* else give up */
 435                                                 skip_rep = 1;
 436                                         dprintk("  reject %d  collide %d  "
 437                                                 "ftotal %u  flocal %u\n",
 438                                                 reject, collide, ftotal,
 439                                                 flocal);
 440                                 }
 441                         } while (retry_bucket);
 442                 } while (retry_descent);
 443
 444                 if (skip_rep) {
 445                         dprintk("skip rep\n");
 446                         continue;
 447                 }
 448
 449                 dprintk("CHOOSE got %d\n", item);
 450                 out[outpos] = item;
 451                 outpos++;
 452         }
 453
 454         dprintk("CHOOSE returns %d\n", outpos);
 455         return outpos;
 456 }
 457
 458
 459 /**
 460  * crush_do_rule - calculate a mapping with the given input and rule
 461  * @map: the crush_map
 462  * @ruleno: the rule id
 463  * @x: hash input
 464  * @result: pointer to result vector
 465  * @result_max: maximum result size
 466  * @force: force initial replica choice; -1 for none
 467  */
 468 int crush_do_rule(const struct crush_map *map,
 469                   int ruleno, int x, int *result, int result_max,
 470                   int force, const __u32 *weight)
 471 {
 472         int result_len;
 473         int force_context[CRUSH_MAX_DEPTH];
 474         int force_pos = -1;
 475         int a[CRUSH_MAX_SET];
 476         int b[CRUSH_MAX_SET];
 477         int c[CRUSH_MAX_SET];
 478         int recurse_to_leaf;
 479         int *w;
 480         int wsize = 0;
 481         int *o;
 482         int osize;
 483         int *tmp;
 484         struct crush_rule *rule;
 485         __u32 step;
 486         int i, j;
 487         int numrep;
 488         int firstn;
 489
 490         if ((__u32)ruleno >= map->max_rules) {
 491                 dprintk(" bad ruleno %d\n", ruleno);
 492                 return 0;
 493         }
 494
 495         rule = map->rules[ruleno];
 496         result_len = 0;
 497         w = a;
 498         o = b;
 499
 500         /*
 501          * determine hierarchical context of force, if any.  note
 502          * that this may or may not correspond to the specific types
 503          * referenced by the crush rule.  it will also only affect
 504          * the first descent (TAKE).
 505          */
 506         if (force >= 0 &&
 507             force < map->max_devices &&
 508             map->device_parents[force] != 0 &&
 509             !is_out(map, weight, force, x)) {
 510                 while (1) {
 511                         force_context[++force_pos] = force;
 512                         if (force >= 0)
 513                                 force = map->device_parents[force];
 514                         else
 515                                 force = map->bucket_parents[-1-force];
 516                         if (force == 0)
 517                                 break;
 518                 }
 519         }
 520
 521         for (step = 0; step < rule->len; step++) {
 522                 firstn = 0;
 523                 switch (rule->steps[step].op) {
 524                 case CRUSH_RULE_TAKE:
 525                         w[0] = rule->steps[step].arg1;
 526
 527                         /* find position in force_context/hierarchy */
 528                         while (force_pos >= 0 &&
 529                                force_context[force_pos] != w[0])
 530                                 force_pos--;
 531                         /* and move past it */
 532                         if (force_pos >= 0)
 533                                 force_pos--;
 534
 535                         wsize = 1;
 536                         break;
 537
 538                 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
 539                 case CRUSH_RULE_CHOOSE_FIRSTN:
 540                         firstn = 1;
 541                 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
 542                 case CRUSH_RULE_CHOOSE_INDEP:
 543                         if (wsize == 0)
 544                                 break;
 545
 546                         recurse_to_leaf =
 547                                 rule->steps[step].op ==
 548                                  CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
 549                                 rule->steps[step].op ==
 550                                 CRUSH_RULE_CHOOSE_LEAF_INDEP;
 551
 552                         /* reset output */
 553                         osize = 0;
 554
 555                         for (i = 0; i < wsize; i++) {
 556                                 /*
 557                                  * see CRUSH_N, CRUSH_N_MINUS macros.
 558                                  * basically, numrep <= 0 means relative to
 559                                  * the provided result_max
 560                                  */
 561                                 numrep = rule->steps[step].arg1;
 562                                 if (numrep <= 0) {
 563                                         numrep += result_max;
 564                                         if (numrep <= 0)
 565                                                 continue;
 566                                 }
 567                                 j = 0;
 568                                 if (osize == 0 && force_pos >= 0) {
 569                                         /* skip any intermediate types */
 570                                         while (force_pos &&
 571                                                force_context[force_pos] < 0 &&
 572                                                rule->steps[step].arg2 !=
 573                                                map->buckets[-1 -
 574                                                force_context[force_pos]]->type)
 575                                                 force_pos--;
 576                                         o[osize] = force_context[force_pos];
 577                                         if (recurse_to_leaf)
 578                                                 c[osize] = force_context[0];
 579                                         j++;
 580                                         force_pos--;
 581                                 }
 582                                 osize += crush_choose(map,
 583                                                       map->buckets[-1-w[i]],
 584                                                       weight,
 585                                                       x, numrep,
 586                                                       rule->steps[step].arg2,
 587                                                       o+osize, j,
 588                                                       firstn,
 589                                                       recurse_to_leaf, c+osize);
 590                         }
 591
 592                         if (recurse_to_leaf)
 593                                 /* copy final _leaf_ values to output set */
 594                                 memcpy(o, c, osize*sizeof(*o));
 595
 596                         /* swap t and w arrays */
 597                         tmp = o;
 598                         o = w;
 599                         w = tmp;
 600                         wsize = osize;
 601                         break;
 602
 603
 604                 case CRUSH_RULE_EMIT:
 605                         for (i = 0; i < wsize && result_len < result_max; i++) {
 606                                 result[result_len] = w[i];
 607                                 result_len++;
 608                         }
 609                         wsize = 0;
 610                         break;
 611
 612                 default:
 613                         dprintk(" unknown op %d at step %d\n",
 614                                 curstep->op, step);
 615                         break;
 616                 }
 617         }
 618         return result_len;
 619 }
 620
 621