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