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)) {
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;
206 /* descend to the left or right? */
208 if (t < bucket->node_weights[l])
214 return bucket->h.items[n >> 1];
220 static int bucket_straw_choose(struct crush_bucket_straw *bucket,
228 for (i = 0; i < bucket->h.size; i++) {
229 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
231 draw *= bucket->straws[i];
232 if (i == 0 || draw > high_draw) {
237 return bucket->h.items[high];
240 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
242 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
243 BUG_ON(in->size == 0);
245 case CRUSH_BUCKET_UNIFORM:
246 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
248 case CRUSH_BUCKET_LIST:
249 return bucket_list_choose((struct crush_bucket_list *)in,
251 case CRUSH_BUCKET_TREE:
252 return bucket_tree_choose((struct crush_bucket_tree *)in,
254 case CRUSH_BUCKET_STRAW:
255 return bucket_straw_choose((struct crush_bucket_straw *)in,
258 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
264 * true if device is marked "out" (failed, fully offloaded)
267 static int is_out(const struct crush_map *map, const __u32 *weight, int item, int x)
269 if (weight[item] >= 0x10000)
271 if (weight[item] == 0)
273 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
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)
293 static int crush_choose(const struct crush_map *map,
294 struct crush_bucket *bucket,
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)
302 unsigned int ftotal, flocal;
303 int retry_descent, retry_bucket, skip_rep;
304 struct crush_bucket *in = bucket;
311 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
312 bucket->id, x, outpos, numrep);
314 for (rep = outpos; rep < numrep; rep++) {
315 /* keep trying until we get a non-out, non-colliding item */
320 in = bucket; /* initial bucket */
322 /* choose through intervening buckets */
328 if (in->alg == CRUSH_BUCKET_UNIFORM) {
330 if (firstn || (__u32)numrep >= in->size)
331 /* r' = r + f_total */
333 else if (in->size % numrep == 0)
334 /* r'=r+(n+1)*f_local */
338 /* r' = r + n*f_local */
339 r += numrep * (flocal+ftotal);
342 /* r' = r + f_total */
345 /* r' = r + n*f_local */
346 r += numrep * (flocal+ftotal);
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);
359 item = crush_bucket_choose(in, x, r);
360 if (item >= map->max_devices) {
361 dprintk(" bad item %d\n", item);
368 itemtype = map->buckets[-1-item]->type;
371 dprintk(" item %d type %d\n", item, itemtype);
374 if (itemtype != type) {
376 (-1-item) >= map->max_buckets) {
377 dprintk(" bad item type %d\n", type);
381 in = map->buckets[-1-item];
387 for (i = 0; i < outpos; i++) {
388 if (out[i] == item) {
395 if (!collide && recurse_to_leaf) {
397 if (crush_choose(map,
398 map->buckets[-1-item],
403 map->chooseleaf_descend_once,
405 /* didn't get leaf */
408 /* we already have a leaf! */
416 reject = is_out(map, weight,
423 if (reject || collide) {
427 if (reject && descend_once)
428 /* let outer call try again */
430 else if (collide && flocal <= map->choose_local_tries)
431 /* retry locally a few times */
433 else if (map->choose_local_fallback_tries > 0 &&
434 flocal <= in->size + map->choose_local_fallback_tries)
435 /* exhaustive bucket search */
437 else if (ftotal <= map->choose_total_tries)
438 /* then retry descent */
443 dprintk(" reject %d collide %d "
444 "ftotal %u flocal %u\n",
445 reject, collide, ftotal,
448 } while (retry_bucket);
449 } while (retry_descent);
452 dprintk("skip rep\n");
456 dprintk("CHOOSE got %d\n", item);
461 dprintk("CHOOSE returns %d\n", outpos);
467 * crush_do_rule - calculate a mapping with the given input and rule
468 * @map: the crush_map
469 * @ruleno: the rule id
471 * @result: pointer to result vector
472 * @result_max: maximum result size
474 int crush_do_rule(const struct crush_map *map,
475 int ruleno, int x, int *result, int result_max,
479 int a[CRUSH_MAX_SET];
480 int b[CRUSH_MAX_SET];
481 int c[CRUSH_MAX_SET];
488 struct crush_rule *rule;
493 const int descend_once = 0;
495 if ((__u32)ruleno >= map->max_rules) {
496 dprintk(" bad ruleno %d\n", ruleno);
500 rule = map->rules[ruleno];
505 for (step = 0; step < rule->len; step++) {
506 struct crush_rule_step *curstep = &rule->steps[step];
509 switch (curstep->op) {
510 case CRUSH_RULE_TAKE:
511 w[0] = curstep->arg1;
515 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
516 case CRUSH_RULE_CHOOSE_FIRSTN:
519 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
520 case CRUSH_RULE_CHOOSE_INDEP:
526 CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
528 CRUSH_RULE_CHOOSE_LEAF_INDEP;
533 for (i = 0; i < wsize; i++) {
535 * see CRUSH_N, CRUSH_N_MINUS macros.
536 * basically, numrep <= 0 means relative to
537 * the provided result_max
539 numrep = curstep->arg1;
541 numrep += result_max;
546 osize += crush_choose(map,
547 map->buckets[-1-w[i]],
554 descend_once, c+osize);
558 /* copy final _leaf_ values to output set */
559 memcpy(o, c, osize*sizeof(*o));
561 /* swap t and w arrays */
569 case CRUSH_RULE_EMIT:
570 for (i = 0; i < wsize && result_len < result_max; i++) {
571 result[result_len] = w[i];
578 dprintk(" unknown op %d at step %d\n",