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>
23 #include <linux/crush/mapper.h>
26 * Implement the core CRUSH mapping algorithm.
30 * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
32 * @ruleset: the storage ruleset id (user defined)
33 * @type: storage ruleset type (user defined)
34 * @size: output set size
36 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
40 for (i = 0; i < map->max_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)
53 * bucket choose methods
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.
61 * Choose based on a random permutation of the bucket.
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.
69 static int bucket_perm_choose(struct crush_bucket *bucket,
72 unsigned int pr = r % bucket->size;
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);
80 /* optimize common r=0 case */
82 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
85 bucket->perm_n = 0xffff; /* magic value, see below */
89 for (i = 0; i < bucket->size; i++)
92 } else if (bucket->perm_n == 0xffff) {
93 /* clean up after the r=0 case above */
94 for (i = 1; i < bucket->size; i++)
96 bucket->perm[bucket->perm[0]] = 0;
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) %
110 unsigned int t = bucket->perm[p + i];
111 bucket->perm[p + i] = bucket->perm[p];
114 dprintk(" perm_choose swap %d with %d\n", p, p+i);
118 for (i = 0; i < bucket->size; i++)
119 dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]);
121 s = bucket->perm[pr];
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];
129 static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
132 return bucket_perm_choose(&bucket->h, x, r);
136 static int bucket_list_choose(struct crush_bucket_list *bucket,
141 for (i = bucket->h.size-1; i >= 0; i--) {
142 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
145 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
147 i, x, r, bucket->h.items[i], bucket->item_weights[i],
148 bucket->sum_weights[i], w);
149 w *= bucket->sum_weights[i];
151 /*dprintk(" scaled %llx\n", w);*/
152 if (w < bucket->item_weights[i])
153 return bucket->h.items[i];
156 dprintk("bad list sums for bucket %d\n", bucket->h.id);
157 return bucket->h.items[0];
162 static int height(int n)
165 while ((n & 1) == 0) {
172 static int left(int x)
175 return x - (1 << (h-1));
178 static int right(int x)
181 return x + (1 << (h-1));
184 static int terminal(int x)
189 static int bucket_tree_choose(struct crush_bucket_tree *bucket,
197 n = bucket->num_nodes >> 1;
199 while (!terminal(n)) {
201 /* pick point in [0, w) */
202 w = bucket->node_weights[n];
203 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
204 bucket->h.id) * (__u64)w;
207 /* descend to the left or right? */
209 if (t < bucket->node_weights[l])
215 return bucket->h.items[n >> 1];
221 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
229 for (i = 0; i < bucket->h.size; i++) {
230 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
232 draw *= bucket->straws[i];
233 if (i == 0 || draw > high_draw) {
238 return bucket->h.items[high];
241 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
243 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
244 BUG_ON(in->size == 0);
246 case CRUSH_BUCKET_UNIFORM:
247 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
249 case CRUSH_BUCKET_LIST:
250 return bucket_list_choose((struct crush_bucket_list *)in,
252 case CRUSH_BUCKET_TREE:
253 return bucket_tree_choose((struct crush_bucket_tree *)in,
255 case CRUSH_BUCKET_STRAW:
256 return bucket_straw_choose((struct crush_bucket_straw *)in,
259 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
265 * true if device is marked "out" (failed, fully offloaded)
268 static int is_out(const struct crush_map *map,
269 const __u32 *weight, int weight_max,
272 if (item >= weight_max)
274 if (weight[item] >= 0x10000)
276 if (weight[item] == 0)
278 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
285 * crush_choose_firstn - choose numrep distinct items of given type
286 * @map: the crush_map
287 * @bucket: the bucket we are choose an item from
288 * @x: crush input value
289 * @numrep: the number of items to choose
290 * @type: the type of item to choose
291 * @out: pointer to output vector
292 * @outpos: our position in that vector
293 * @recurse_to_leaf: true if we want one device under each item of given type
294 * @out2: second output vector for leaf items (if @recurse_to_leaf)
296 static int crush_choose_firstn(const struct crush_map *map,
297 struct crush_bucket *bucket,
298 const __u32 *weight, int weight_max,
299 int x, int numrep, int type,
300 int *out, int outpos,
302 unsigned int recurse_tries,
303 unsigned int local_tries,
304 unsigned int local_fallback_tries,
309 unsigned int ftotal, flocal;
310 int retry_descent, retry_bucket, skip_rep;
311 struct crush_bucket *in = bucket;
318 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
319 bucket->id, x, outpos, numrep);
321 for (rep = outpos; rep < numrep; rep++) {
322 /* keep trying until we get a non-out, non-colliding item */
327 in = bucket; /* initial bucket */
329 /* choose through intervening buckets */
335 /* r' = r + f_total */
343 if (local_fallback_tries > 0 &&
344 flocal >= (in->size>>1) &&
345 flocal > local_fallback_tries)
346 item = bucket_perm_choose(in, x, r);
348 item = crush_bucket_choose(in, x, r);
349 if (item >= map->max_devices) {
350 dprintk(" bad item %d\n", item);
357 itemtype = map->buckets[-1-item]->type;
360 dprintk(" item %d type %d\n", item, itemtype);
363 if (itemtype != type) {
365 (-1-item) >= map->max_buckets) {
366 dprintk(" bad item type %d\n", type);
370 in = map->buckets[-1-item];
376 for (i = 0; i < outpos; i++) {
377 if (out[i] == item) {
384 if (!collide && recurse_to_leaf) {
386 if (crush_choose_firstn(map,
387 map->buckets[-1-item],
393 local_fallback_tries,
396 /* didn't get leaf */
399 /* we already have a leaf! */
407 reject = is_out(map, weight,
415 if (reject || collide) {
419 if (collide && flocal <= local_tries)
420 /* retry locally a few times */
422 else if (local_fallback_tries > 0 &&
423 flocal <= in->size + local_fallback_tries)
424 /* exhaustive bucket search */
426 else if (ftotal <= tries)
427 /* then retry descent */
432 dprintk(" reject %d collide %d "
433 "ftotal %u flocal %u\n",
434 reject, collide, ftotal,
437 } while (retry_bucket);
438 } while (retry_descent);
441 dprintk("skip rep\n");
445 dprintk("CHOOSE got %d\n", item);
450 dprintk("CHOOSE returns %d\n", outpos);
456 * crush_choose_indep: alternative breadth-first positionally stable mapping
459 static void crush_choose_indep(const struct crush_map *map,
460 struct crush_bucket *bucket,
461 const __u32 *weight, int weight_max,
462 int x, int left, int numrep, int type,
463 int *out, int outpos,
465 unsigned int recurse_tries,
470 struct crush_bucket *in = bucket;
471 int endpos = outpos + left;
480 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
481 bucket->id, x, outpos, numrep);
483 /* initially my result is undefined */
484 for (rep = outpos; rep < endpos; rep++) {
485 out[rep] = CRUSH_ITEM_UNDEF;
487 out2[rep] = CRUSH_ITEM_UNDEF;
490 for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) {
491 for (rep = outpos; rep < endpos; rep++) {
492 if (out[rep] != CRUSH_ITEM_UNDEF)
495 in = bucket; /* initial bucket */
497 /* choose through intervening buckets */
499 /* note: we base the choice on the position
500 * even in the nested call. that means that
501 * if the first layer chooses the same bucket
502 * in a different position, we will tend to
503 * choose a different item in that bucket.
504 * this will involve more devices in data
505 * movement and tend to distribute the load.
510 if (in->alg == CRUSH_BUCKET_UNIFORM &&
511 in->size % numrep == 0)
512 /* r'=r+(n+1)*f_total */
513 r += (numrep+1) * ftotal;
515 /* r' = r + n*f_total */
516 r += numrep * ftotal;
520 dprintk(" empty bucket\n");
524 item = crush_bucket_choose(in, x, r);
525 if (item >= map->max_devices) {
526 dprintk(" bad item %d\n", item);
527 out[rep] = CRUSH_ITEM_NONE;
529 out2[rep] = CRUSH_ITEM_NONE;
536 itemtype = map->buckets[-1-item]->type;
539 dprintk(" item %d type %d\n", item, itemtype);
542 if (itemtype != type) {
544 (-1-item) >= map->max_buckets) {
545 dprintk(" bad item type %d\n", type);
546 out[rep] = CRUSH_ITEM_NONE;
553 in = map->buckets[-1-item];
559 for (i = outpos; i < endpos; i++) {
560 if (out[i] == item) {
568 if (recurse_to_leaf) {
570 crush_choose_indep(map,
571 map->buckets[-1-item],
577 if (out2[rep] == CRUSH_ITEM_NONE) {
578 /* placed nothing; no leaf */
582 /* we already have a leaf! */
589 is_out(map, weight, weight_max, item, x))
599 for (rep = outpos; rep < endpos; rep++) {
600 if (out[rep] == CRUSH_ITEM_UNDEF) {
601 out[rep] = CRUSH_ITEM_NONE;
603 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
604 out2[rep] = CRUSH_ITEM_NONE;
610 * crush_do_rule - calculate a mapping with the given input and rule
611 * @map: the crush_map
612 * @ruleno: the rule id
614 * @result: pointer to result vector
615 * @result_max: maximum result size
616 * @weight: weight vector (for map leaves)
617 * @weight_max: size of weight vector
618 * @scratch: scratch vector for private use; must be >= 3 * result_max
620 int crush_do_rule(const struct crush_map *map,
621 int ruleno, int x, int *result, int result_max,
622 const __u32 *weight, int weight_max,
627 int *b = scratch + result_max;
628 int *c = scratch + result_max*2;
635 struct crush_rule *rule;
639 int choose_tries = map->choose_total_tries;
640 int choose_local_tries = map->choose_local_tries;
641 int choose_local_fallback_tries = map->choose_local_fallback_tries;
642 int choose_leaf_tries = 0;
644 if ((__u32)ruleno >= map->max_rules) {
645 dprintk(" bad ruleno %d\n", ruleno);
649 rule = map->rules[ruleno];
654 for (step = 0; step < rule->len; step++) {
656 struct crush_rule_step *curstep = &rule->steps[step];
658 switch (curstep->op) {
659 case CRUSH_RULE_TAKE:
660 w[0] = curstep->arg1;
664 case CRUSH_RULE_SET_CHOOSE_TRIES:
665 if (curstep->arg1 > 0)
666 choose_tries = curstep->arg1;
669 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
670 if (curstep->arg1 > 0)
671 choose_leaf_tries = curstep->arg1;
674 case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES:
675 if (curstep->arg1 > 0)
676 choose_local_tries = curstep->arg1;
679 case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES:
680 if (curstep->arg1 > 0)
681 choose_local_fallback_tries = curstep->arg1;
684 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
685 case CRUSH_RULE_CHOOSE_FIRSTN:
688 case CRUSH_RULE_CHOOSELEAF_INDEP:
689 case CRUSH_RULE_CHOOSE_INDEP:
695 CRUSH_RULE_CHOOSELEAF_FIRSTN ||
697 CRUSH_RULE_CHOOSELEAF_INDEP;
702 for (i = 0; i < wsize; i++) {
704 * see CRUSH_N, CRUSH_N_MINUS macros.
705 * basically, numrep <= 0 means relative to
706 * the provided result_max
708 numrep = curstep->arg1;
710 numrep += result_max;
717 if (choose_leaf_tries)
720 else if (map->chooseleaf_descend_once)
723 recurse_tries = choose_tries;
724 osize += crush_choose_firstn(
726 map->buckets[-1-w[i]],
734 choose_local_fallback_tries,
740 map->buckets[-1-w[i]],
747 choose_leaf_tries : 1,
756 /* copy final _leaf_ values to output set */
757 memcpy(o, c, osize*sizeof(*o));
759 /* swap o and w arrays */
767 case CRUSH_RULE_EMIT:
768 for (i = 0; i < wsize && result_len < result_max; i++) {
769 result[result_len] = w[i];
776 dprintk(" unknown op %d at step %d\n",