3 # include <linux/string.h>
4 # include <linux/slab.h>
5 # include <linux/bug.h>
6 # include <linux/kernel.h>
8 # define dprintk(args...)
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)
21 #include <linux/crush/crush.h>
22 #include <linux/crush/hash.h>
25 * Implement the core CRUSH mapping algorithm.
29 * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
31 * @ruleset: the storage ruleset id (user defined)
32 * @type: storage ruleset type (user defined)
33 * @size: output set size
35 int crush_find_rule(struct crush_map *map, int ruleset, int type, int size)
39 for (i = 0; i < map->max_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)
52 * bucket choose methods
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.
60 * Choose based on a random permutation of the bucket.
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.
68 static int bucket_perm_choose(struct crush_bucket *bucket,
71 unsigned pr = r % bucket->size;
74 /* start a new permutation if @x has changed */
75 if (bucket->perm_x != x || bucket->perm_n == 0) {
76 dprintk("bucket %d new x=%d\n", bucket->id, x);
79 /* optimize common r=0 case */
81 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
84 bucket->perm_n = 0xffff; /* magic value, see below */
88 for (i = 0; i < bucket->size; i++)
91 } else if (bucket->perm_n == 0xffff) {
92 /* clean up after the r=0 case above */
93 for (i = 1; i < bucket->size; i++)
95 bucket->perm[bucket->perm[0]] = 0;
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) %
109 unsigned t = bucket->perm[p + i];
110 bucket->perm[p + i] = bucket->perm[p];
113 dprintk(" perm_choose swap %d with %d\n", p, p+i);
117 for (i = 0; i < bucket->size; i++)
118 dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]);
120 s = bucket->perm[pr];
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];
128 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
131 return bucket_perm_choose(&bucket->h, x, r);
135 static int bucket_list_choose(struct crush_bucket_list *bucket,
140 for (i = bucket->h.size-1; i >= 0; i--) {
141 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
144 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
146 i, x, r, bucket->h.items[i], bucket->item_weights[i],
147 bucket->sum_weights[i], w);
148 w *= bucket->sum_weights[i];
150 /*dprintk(" scaled %llx\n", w);*/
151 if (w < bucket->item_weights[i])
152 return bucket->h.items[i];
161 static int height(int n)
164 while ((n & 1) == 0) {
171 static int left(int x)
174 return x - (1 << (h-1));
177 static int right(int x)
180 return x + (1 << (h-1));
183 static int terminal(int x)
188 static int bucket_tree_choose(struct crush_bucket_tree *bucket,
196 n = bucket->num_nodes >> 1;
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;
205 /* descend to the left or right? */
207 if (t < bucket->node_weights[l])
213 return bucket->h.items[n >> 1];
219 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
227 for (i = 0; i < bucket->h.size; i++) {
228 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
230 draw *= bucket->straws[i];
231 if (i == 0 || draw > high_draw) {
236 return bucket->h.items[high];
239 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
241 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
243 case CRUSH_BUCKET_UNIFORM:
244 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
246 case CRUSH_BUCKET_LIST:
247 return bucket_list_choose((struct crush_bucket_list *)in,
249 case CRUSH_BUCKET_TREE:
250 return bucket_tree_choose((struct crush_bucket_tree *)in,
252 case CRUSH_BUCKET_STRAW:
253 return bucket_straw_choose((struct crush_bucket_straw *)in,
262 * true if device is marked "out" (failed, fully offloaded)
265 static int is_out(struct crush_map *map, __u32 *weight, int item, int x)
267 if (weight[item] >= 0x10000)
269 if (weight[item] == 0)
271 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
278 * crush_choose - choose numrep distinct items of given type
279 * @map: the crush_map
280 * @bucket: the bucket we are choose an item from
281 * @x: crush input value
282 * @numrep: the number of items to choose
283 * @type: the type of item to choose
284 * @out: pointer to output vector
285 * @outpos: our position in that vector
286 * @firstn: true if choosing "first n" items, false if choosing "indep"
287 * @recurse_to_leaf: true if we want one device under each item of given type
288 * @out2: second output vector for leaf items (if @recurse_to_leaf)
290 static int crush_choose(struct crush_map *map,
291 struct crush_bucket *bucket,
293 int x, int numrep, int type,
294 int *out, int outpos,
295 int firstn, int recurse_to_leaf,
300 int retry_descent, retry_bucket, skip_rep;
301 struct crush_bucket *in = bucket;
307 const int orig_tries = 5; /* attempts before we fall back to search */
309 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
310 bucket->id, x, outpos, numrep);
312 for (rep = outpos; rep < numrep; rep++) {
313 /* keep trying until we get a non-out, non-colliding item */
318 in = bucket; /* initial bucket */
320 /* choose through intervening buckets */
326 if (in->alg == CRUSH_BUCKET_UNIFORM) {
328 if (firstn || numrep >= in->size)
329 /* r' = r + f_total */
331 else if (in->size % numrep == 0)
332 /* r'=r+(n+1)*f_local */
336 /* r' = r + n*f_local */
337 r += numrep * (flocal+ftotal);
340 /* r' = r + f_total */
343 /* r' = r + n*f_local */
344 r += numrep * (flocal+ftotal);
352 if (flocal >= (in->size>>1) &&
354 item = bucket_perm_choose(in, x, r);
356 item = crush_bucket_choose(in, x, r);
357 BUG_ON(item >= map->max_devices);
361 itemtype = map->buckets[-1-item]->type;
364 dprintk(" item %d type %d\n", item, itemtype);
367 if (itemtype != type) {
369 (-1-item) >= map->max_buckets);
370 in = map->buckets[-1-item];
376 for (i = 0; i < outpos; i++) {
377 if (out[i] == item) {
384 if (recurse_to_leaf) {
386 if (crush_choose(map,
387 map->buckets[-1-item],
393 /* didn't get leaf */
396 /* we already have a leaf! */
404 reject = is_out(map, weight,
411 if (reject || collide) {
415 if (collide && flocal < 3)
416 /* retry locally a few times */
418 else if (flocal < in->size + orig_tries)
419 /* exhaustive bucket search */
421 else if (ftotal < 20)
422 /* then retry descent */
427 dprintk(" reject %d collide %d "
428 "ftotal %d flocal %d\n",
429 reject, collide, ftotal,
432 } while (retry_bucket);
433 } while (retry_descent);
436 dprintk("skip rep\n");
440 dprintk("CHOOSE got %d\n", item);
445 dprintk("CHOOSE returns %d\n", outpos);
451 * crush_do_rule - calculate a mapping with the given input and rule
452 * @map: the crush_map
453 * @ruleno: the rule id
455 * @result: pointer to result vector
456 * @result_max: maximum result size
457 * @force: force initial replica choice; -1 for none
459 int crush_do_rule(struct crush_map *map,
460 int ruleno, int x, int *result, int result_max,
461 int force, __u32 *weight)
464 int force_context[CRUSH_MAX_DEPTH];
466 int a[CRUSH_MAX_SET];
467 int b[CRUSH_MAX_SET];
468 int c[CRUSH_MAX_SET];
475 struct crush_rule *rule;
482 BUG_ON(ruleno >= map->max_rules);
484 rule = map->rules[ruleno];
490 * determine hierarchical context of force, if any. note
491 * that this may or may not correspond to the specific types
492 * referenced by the crush rule.
495 if (force >= map->max_devices ||
496 map->device_parents[force] == 0) {
497 /*dprintk("CRUSH: forcefed device dne\n");*/
498 rc = -1; /* force fed device dne */
501 if (!is_out(map, weight, force, x)) {
503 force_context[++force_pos] = force;
505 force = map->device_parents[force];
507 force = map->bucket_parents[-1-force];
514 for (step = 0; step < rule->len; step++) {
516 switch (rule->steps[step].op) {
517 case CRUSH_RULE_TAKE:
518 w[0] = rule->steps[step].arg1;
519 if (force_pos >= 0) {
520 BUG_ON(force_context[force_pos] != w[0]);
526 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
527 case CRUSH_RULE_CHOOSE_FIRSTN:
529 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
530 case CRUSH_RULE_CHOOSE_INDEP:
534 rule->steps[step].op ==
535 CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
536 rule->steps[step].op ==
537 CRUSH_RULE_CHOOSE_LEAF_INDEP;
542 for (i = 0; i < wsize; i++) {
544 * see CRUSH_N, CRUSH_N_MINUS macros.
545 * basically, numrep <= 0 means relative to
546 * the provided result_max
548 numrep = rule->steps[step].arg1;
550 numrep += result_max;
555 if (osize == 0 && force_pos >= 0) {
556 /* skip any intermediate types */
558 force_context[force_pos] < 0 &&
559 rule->steps[step].arg2 !=
561 force_context[force_pos]]->type)
563 o[osize] = force_context[force_pos];
565 c[osize] = force_context[0];
569 osize += crush_choose(map,
570 map->buckets[-1-w[i]],
573 rule->steps[step].arg2,
576 recurse_to_leaf, c+osize);
580 /* copy final _leaf_ values to output set */
581 memcpy(o, c, osize*sizeof(*o));
583 /* swap t and w arrays */
591 case CRUSH_RULE_EMIT:
592 for (i = 0; i < wsize && result_len < result_max; i++) {
593 result[result_len] = w[i];