blob: e3ade074541c88752dabe4975d74ce6714ce4c6b [file] [log] [blame]
Sage Weil5ecc0a02009-10-06 11:31:11 -07001
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
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070021#include <linux/crush/crush.h>
22#include <linux/crush/hash.h>
hartleysfeb50ac2012-04-24 14:38:37 +000023#include <linux/crush/mapper.h>
Sage Weil5ecc0a02009-10-06 11:31:11 -070024
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 */
Sage Weil8b12d472012-05-07 15:38:35 -070036int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
Sage Weil5ecc0a02009-10-06 11:31:11 -070037{
Sage Weil8b12d472012-05-07 15:38:35 -070038 __u32 i;
Sage Weil5ecc0a02009-10-06 11:31:11 -070039
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 */
69static int bucket_perm_choose(struct crush_bucket *bucket,
70 int x, int r)
71{
Eric Dumazet95c96172012-04-15 05:58:06 +000072 unsigned int pr = r % bucket->size;
73 unsigned int i, s;
Sage Weil5ecc0a02009-10-06 11:31:11 -070074
75 /* start a new permutation if @x has changed */
Sage Weil8b12d472012-05-07 15:38:35 -070076 if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) {
Sage Weil5ecc0a02009-10-06 11:31:11 -070077 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) {
Sage Weilfb690392009-11-07 20:18:22 -080082 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
Sage Weil5ecc0a02009-10-06 11:31:11 -070083 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) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000104 unsigned int p = bucket->perm_n;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700105 /* no point in swapping the final entry */
106 if (p < bucket->size - 1) {
Sage Weilfb690392009-11-07 20:18:22 -0800107 i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
Sage Weil5ecc0a02009-10-06 11:31:11 -0700108 (bucket->size - p);
109 if (i) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000110 unsigned int t = bucket->perm[p + i];
Sage Weil5ecc0a02009-10-06 11:31:11 -0700111 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];
122out:
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 */
129static 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 */
136static 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--) {
Sage Weilfb690392009-11-07 20:18:22 -0800142 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
143 r, bucket->h.id);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700144 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
Sage Weila1f48952012-05-07 15:35:24 -0700156 dprintk("bad list sums for bucket %d\n", bucket->h.id);
157 return bucket->h.items[0];
Sage Weil5ecc0a02009-10-06 11:31:11 -0700158}
159
160
161/* (binary) tree */
162static 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
172static int left(int x)
173{
174 int h = height(x);
175 return x - (1 << (h-1));
176}
177
178static int right(int x)
179{
180 int h = height(x);
181 return x + (1 << (h-1));
182}
183
184static int terminal(int x)
185{
186 return x & 1;
187}
188
189static int bucket_tree_choose(struct crush_bucket_tree *bucket,
190 int x, int r)
191{
Ilya Dryomov8f99c852013-12-24 21:19:24 +0200192 int n;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700193 __u32 w;
194 __u64 t;
195
196 /* start at root */
197 n = bucket->num_nodes >> 1;
198
199 while (!terminal(n)) {
Ilya Dryomov8f99c852013-12-24 21:19:24 +0200200 int l;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700201 /* pick point in [0, w) */
202 w = bucket->node_weights[n];
Sage Weilfb690392009-11-07 20:18:22 -0800203 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
204 bucket->h.id) * (__u64)w;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700205 t = t >> 32;
206
207 /* descend to the left or right? */
208 l = left(n);
209 if (t < bucket->node_weights[l])
210 n = l;
211 else
212 n = right(n);
213 }
214
215 return bucket->h.items[n >> 1];
216}
217
218
219/* straw */
220
221static int bucket_straw_choose(struct crush_bucket_straw *bucket,
222 int x, int r)
223{
Sage Weil8b12d472012-05-07 15:38:35 -0700224 __u32 i;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700225 int high = 0;
226 __u64 high_draw = 0;
227 __u64 draw;
228
229 for (i = 0; i < bucket->h.size; i++) {
Sage Weilfb690392009-11-07 20:18:22 -0800230 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700231 draw &= 0xffff;
232 draw *= bucket->straws[i];
233 if (i == 0 || draw > high_draw) {
234 high = i;
235 high_draw = draw;
236 }
237 }
238 return bucket->h.items[high];
239}
240
241static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
242{
Sage Weila1a31e72010-06-24 12:58:14 -0700243 dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
Sage Weila1f48952012-05-07 15:35:24 -0700244 BUG_ON(in->size == 0);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700245 switch (in->alg) {
246 case CRUSH_BUCKET_UNIFORM:
247 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
248 x, r);
249 case CRUSH_BUCKET_LIST:
250 return bucket_list_choose((struct crush_bucket_list *)in,
251 x, r);
252 case CRUSH_BUCKET_TREE:
253 return bucket_tree_choose((struct crush_bucket_tree *)in,
254 x, r);
255 case CRUSH_BUCKET_STRAW:
256 return bucket_straw_choose((struct crush_bucket_straw *)in,
257 x, r);
258 default:
Sage Weila1f48952012-05-07 15:35:24 -0700259 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
Sage Weil50b885b2009-12-01 14:12:07 -0800260 return in->items[0];
Sage Weil5ecc0a02009-10-06 11:31:11 -0700261 }
262}
263
264/*
265 * true if device is marked "out" (failed, fully offloaded)
266 * of the cluster
267 */
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200268static int is_out(const struct crush_map *map,
269 const __u32 *weight, int weight_max,
270 int item, int x)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700271{
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200272 if (item >= weight_max)
273 return 1;
Sage Weil153a1092010-07-05 09:44:17 -0700274 if (weight[item] >= 0x10000)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700275 return 0;
276 if (weight[item] == 0)
277 return 1;
Sage Weilfb690392009-11-07 20:18:22 -0800278 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
279 < weight[item])
Sage Weil5ecc0a02009-10-06 11:31:11 -0700280 return 0;
281 return 1;
282}
283
284/**
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200285 * crush_choose_firstn - choose numrep distinct items of given type
Sage Weil5ecc0a02009-10-06 11:31:11 -0700286 * @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
Sage Weil5ecc0a02009-10-06 11:31:11 -0700293 * @recurse_to_leaf: true if we want one device under each item of given type
Jim Schutt1604f482012-11-30 09:15:25 -0700294 * @descend_once: true if we should only try one descent before giving up
Sage Weil5ecc0a02009-10-06 11:31:11 -0700295 * @out2: second output vector for leaf items (if @recurse_to_leaf)
296 */
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200297static int crush_choose_firstn(const struct crush_map *map,
298 struct crush_bucket *bucket,
299 const __u32 *weight, int weight_max,
300 int x, int numrep, int type,
301 int *out, int outpos,
302 int recurse_to_leaf,
303 int descend_once, int *out2)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700304{
305 int rep;
Sage Weil8b12d472012-05-07 15:38:35 -0700306 unsigned int ftotal, flocal;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700307 int retry_descent, retry_bucket, skip_rep;
308 struct crush_bucket *in = bucket;
309 int r;
310 int i;
Sage Weilb28813a2009-10-07 10:59:34 -0700311 int item = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700312 int itemtype;
313 int collide, reject;
Sage Weila1a31e72010-06-24 12:58:14 -0700314
315 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
316 bucket->id, x, outpos, numrep);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700317
318 for (rep = outpos; rep < numrep; rep++) {
319 /* keep trying until we get a non-out, non-colliding item */
320 ftotal = 0;
321 skip_rep = 0;
322 do {
323 retry_descent = 0;
324 in = bucket; /* initial bucket */
325
326 /* choose through intervening buckets */
327 flocal = 0;
328 do {
Sage Weilb28813a2009-10-07 10:59:34 -0700329 collide = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700330 retry_bucket = 0;
331 r = rep;
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200332 /* r' = r + f_total */
333 r += ftotal;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700334
335 /* bucket choose */
Sage Weilb28813a2009-10-07 10:59:34 -0700336 if (in->size == 0) {
337 reject = 1;
338 goto reject;
339 }
Sage Weil546f04e2012-07-30 18:15:23 -0700340 if (map->choose_local_fallback_tries > 0 &&
341 flocal >= (in->size>>1) &&
342 flocal > map->choose_local_fallback_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700343 item = bucket_perm_choose(in, x, r);
344 else
345 item = crush_bucket_choose(in, x, r);
Sage Weila1f48952012-05-07 15:35:24 -0700346 if (item >= map->max_devices) {
347 dprintk(" bad item %d\n", item);
348 skip_rep = 1;
349 break;
350 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700351
352 /* desired type? */
353 if (item < 0)
354 itemtype = map->buckets[-1-item]->type;
355 else
356 itemtype = 0;
357 dprintk(" item %d type %d\n", item, itemtype);
358
359 /* keep going? */
360 if (itemtype != type) {
Sage Weila1f48952012-05-07 15:35:24 -0700361 if (item >= 0 ||
362 (-1-item) >= map->max_buckets) {
363 dprintk(" bad item type %d\n", type);
364 skip_rep = 1;
365 break;
366 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700367 in = map->buckets[-1-item];
Sage Weil55bda7a2010-06-24 12:55:48 -0700368 retry_bucket = 1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700369 continue;
370 }
371
372 /* collision? */
Sage Weil5ecc0a02009-10-06 11:31:11 -0700373 for (i = 0; i < outpos; i++) {
374 if (out[i] == item) {
375 collide = 1;
376 break;
377 }
378 }
379
Sage Weila1a31e72010-06-24 12:58:14 -0700380 reject = 0;
Sage Weil7d7c1f62013-01-15 18:49:09 -0800381 if (!collide && recurse_to_leaf) {
Sage Weila1a31e72010-06-24 12:58:14 -0700382 if (item < 0) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200383 if (crush_choose_firstn(map,
Sage Weila1a31e72010-06-24 12:58:14 -0700384 map->buckets[-1-item],
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200385 weight, weight_max,
Sage Weila1a31e72010-06-24 12:58:14 -0700386 x, outpos+1, 0,
387 out2, outpos,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200388 0,
Jim Schutt1604f482012-11-30 09:15:25 -0700389 map->chooseleaf_descend_once,
Sage Weila1a31e72010-06-24 12:58:14 -0700390 NULL) <= outpos)
391 /* didn't get leaf */
392 reject = 1;
393 } else {
394 /* we already have a leaf! */
395 out2[outpos] = item;
396 }
397 }
398
399 if (!reject) {
Sage Weil5ecc0a02009-10-06 11:31:11 -0700400 /* out? */
401 if (itemtype == 0)
402 reject = is_out(map, weight,
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200403 weight_max,
Sage Weil5ecc0a02009-10-06 11:31:11 -0700404 item, x);
405 else
406 reject = 0;
407 }
408
Sage Weilb28813a2009-10-07 10:59:34 -0700409reject:
Sage Weil5ecc0a02009-10-06 11:31:11 -0700410 if (reject || collide) {
411 ftotal++;
412 flocal++;
413
Jim Schutt1604f482012-11-30 09:15:25 -0700414 if (reject && descend_once)
415 /* let outer call try again */
416 skip_rep = 1;
417 else if (collide && flocal <= map->choose_local_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700418 /* retry locally a few times */
419 retry_bucket = 1;
Sage Weil546f04e2012-07-30 18:15:23 -0700420 else if (map->choose_local_fallback_tries > 0 &&
421 flocal <= in->size + map->choose_local_fallback_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700422 /* exhaustive bucket search */
423 retry_bucket = 1;
Sage Weil546f04e2012-07-30 18:15:23 -0700424 else if (ftotal <= map->choose_total_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700425 /* then retry descent */
426 retry_descent = 1;
427 else
428 /* else give up */
429 skip_rep = 1;
430 dprintk(" reject %d collide %d "
Sage Weil8b12d472012-05-07 15:38:35 -0700431 "ftotal %u flocal %u\n",
Sage Weil5ecc0a02009-10-06 11:31:11 -0700432 reject, collide, ftotal,
433 flocal);
434 }
435 } while (retry_bucket);
436 } while (retry_descent);
437
438 if (skip_rep) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200439 dprintk("skip rep\n");
440 continue;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700441 }
442
Sage Weila1a31e72010-06-24 12:58:14 -0700443 dprintk("CHOOSE got %d\n", item);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700444 out[outpos] = item;
445 outpos++;
446 }
447
Sage Weila1a31e72010-06-24 12:58:14 -0700448 dprintk("CHOOSE returns %d\n", outpos);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700449 return outpos;
450}
451
452
453/**
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200454 * crush_choose_indep: alternative breadth-first positionally stable mapping
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200455 *
456 */
457static void crush_choose_indep(const struct crush_map *map,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200458 struct crush_bucket *bucket,
459 const __u32 *weight, int weight_max,
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200460 int x, int left, int numrep, int type,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200461 int *out, int outpos,
462 unsigned int attempts,
463 unsigned int recurse_attempts,
464 int recurse_to_leaf,
Ilya Dryomov41586082013-12-24 21:19:25 +0200465 int *out2,
466 int parent_r)
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200467{
468 struct crush_bucket *in = bucket;
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200469 int endpos = outpos + left;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200470 int rep;
471 unsigned int ftotal;
472 int r;
473 int i;
474 int item = 0;
475 int itemtype;
476 int collide;
477
478 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
479 bucket->id, x, outpos, numrep);
480
481 /* initially my result is undefined */
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200482 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200483 out[rep] = CRUSH_ITEM_UNDEF;
484 if (out2)
485 out2[rep] = CRUSH_ITEM_UNDEF;
486 }
487
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200488 for (ftotal = 0; left > 0 && ftotal < attempts; ftotal++) {
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200489 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200490 if (out[rep] != CRUSH_ITEM_UNDEF)
491 continue;
492
493 in = bucket; /* initial bucket */
494
495 /* choose through intervening buckets */
496 for (;;) {
Ilya Dryomov3102b0a2013-12-24 21:19:25 +0200497 /* note: we base the choice on the position
498 * even in the nested call. that means that
499 * if the first layer chooses the same bucket
500 * in a different position, we will tend to
501 * choose a different item in that bucket.
502 * this will involve more devices in data
503 * movement and tend to distribute the load.
504 */
Ilya Dryomov41586082013-12-24 21:19:25 +0200505 r = rep + parent_r;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200506
507 /* be careful */
508 if (in->alg == CRUSH_BUCKET_UNIFORM &&
509 in->size % numrep == 0)
510 /* r'=r+(n+1)*f_total */
511 r += (numrep+1) * ftotal;
512 else
513 /* r' = r + n*f_total */
514 r += numrep * ftotal;
515
516 /* bucket choose */
517 if (in->size == 0) {
518 dprintk(" empty bucket\n");
519 break;
520 }
521
522 item = crush_bucket_choose(in, x, r);
523 if (item >= map->max_devices) {
524 dprintk(" bad item %d\n", item);
525 out[rep] = CRUSH_ITEM_NONE;
526 if (out2)
527 out2[rep] = CRUSH_ITEM_NONE;
528 left--;
529 break;
530 }
531
532 /* desired type? */
533 if (item < 0)
534 itemtype = map->buckets[-1-item]->type;
535 else
536 itemtype = 0;
537 dprintk(" item %d type %d\n", item, itemtype);
538
539 /* keep going? */
540 if (itemtype != type) {
541 if (item >= 0 ||
542 (-1-item) >= map->max_buckets) {
543 dprintk(" bad item type %d\n", type);
544 out[rep] = CRUSH_ITEM_NONE;
545 if (out2)
546 out2[rep] =
547 CRUSH_ITEM_NONE;
548 left--;
549 break;
550 }
551 in = map->buckets[-1-item];
552 continue;
553 }
554
555 /* collision? */
556 collide = 0;
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200557 for (i = outpos; i < endpos; i++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200558 if (out[i] == item) {
559 collide = 1;
560 break;
561 }
562 }
563 if (collide)
564 break;
565
566 if (recurse_to_leaf) {
567 if (item < 0) {
568 crush_choose_indep(map,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200569 map->buckets[-1-item],
570 weight, weight_max,
571 x, 1, numrep, 0,
572 out2, rep,
573 recurse_attempts, 0,
574 0, NULL, r);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200575 if (out2[rep] == CRUSH_ITEM_NONE) {
576 /* placed nothing; no leaf */
577 break;
578 }
579 } else {
580 /* we already have a leaf! */
581 out2[rep] = item;
582 }
583 }
584
585 /* out? */
586 if (itemtype == 0 &&
587 is_out(map, weight, weight_max, item, x))
588 break;
589
590 /* yay! */
591 out[rep] = item;
592 left--;
593 break;
594 }
595 }
596 }
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200597 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200598 if (out[rep] == CRUSH_ITEM_UNDEF) {
599 out[rep] = CRUSH_ITEM_NONE;
600 }
601 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
602 out2[rep] = CRUSH_ITEM_NONE;
603 }
604 }
605}
606
607/**
Sage Weil5ecc0a02009-10-06 11:31:11 -0700608 * crush_do_rule - calculate a mapping with the given input and rule
609 * @map: the crush_map
610 * @ruleno: the rule id
611 * @x: hash input
612 * @result: pointer to result vector
613 * @result_max: maximum result size
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200614 * @weight: weight vector (for map leaves)
615 * @weight_max: size of weight vector
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200616 * @scratch: scratch vector for private use; must be >= 3 * result_max
Sage Weil5ecc0a02009-10-06 11:31:11 -0700617 */
Sage Weil8b12d472012-05-07 15:38:35 -0700618int crush_do_rule(const struct crush_map *map,
Sage Weil5ecc0a02009-10-06 11:31:11 -0700619 int ruleno, int x, int *result, int result_max,
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200620 const __u32 *weight, int weight_max,
621 int *scratch)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700622{
623 int result_len;
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200624 int *a = scratch;
625 int *b = scratch + result_max;
626 int *c = scratch + result_max*2;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700627 int recurse_to_leaf;
628 int *w;
629 int wsize = 0;
630 int *o;
631 int osize;
632 int *tmp;
633 struct crush_rule *rule;
Sage Weil8b12d472012-05-07 15:38:35 -0700634 __u32 step;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700635 int i, j;
636 int numrep;
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200637 int choose_leaf_tries = 1;
Jim Schutt1604f482012-11-30 09:15:25 -0700638 const int descend_once = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700639
Sage Weila1f48952012-05-07 15:35:24 -0700640 if ((__u32)ruleno >= map->max_rules) {
641 dprintk(" bad ruleno %d\n", ruleno);
642 return 0;
643 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700644
645 rule = map->rules[ruleno];
646 result_len = 0;
647 w = a;
648 o = b;
649
Sage Weil5ecc0a02009-10-06 11:31:11 -0700650 for (step = 0; step < rule->len; step++) {
Ilya Dryomov8f99c852013-12-24 21:19:24 +0200651 int firstn = 0;
Sage Weil06682162012-05-07 15:35:48 -0700652 struct crush_rule_step *curstep = &rule->steps[step];
653
Sage Weil06682162012-05-07 15:35:48 -0700654 switch (curstep->op) {
Sage Weil5ecc0a02009-10-06 11:31:11 -0700655 case CRUSH_RULE_TAKE:
Sage Weil06682162012-05-07 15:35:48 -0700656 w[0] = curstep->arg1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700657 wsize = 1;
658 break;
659
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200660 case CRUSH_RULE_SET_CHOOSE_LEAF_TRIES:
661 if (curstep->arg1 > 0)
662 choose_leaf_tries = curstep->arg1;
663 break;
664
Sage Weil5ecc0a02009-10-06 11:31:11 -0700665 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
666 case CRUSH_RULE_CHOOSE_FIRSTN:
667 firstn = 1;
Sage Weil06682162012-05-07 15:35:48 -0700668 /* fall through */
Sage Weil5ecc0a02009-10-06 11:31:11 -0700669 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
670 case CRUSH_RULE_CHOOSE_INDEP:
Sage Weila1f48952012-05-07 15:35:24 -0700671 if (wsize == 0)
672 break;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700673
674 recurse_to_leaf =
Sage Weil06682162012-05-07 15:35:48 -0700675 curstep->op ==
Sage Weil5ecc0a02009-10-06 11:31:11 -0700676 CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
Sage Weil06682162012-05-07 15:35:48 -0700677 curstep->op ==
Sage Weil5ecc0a02009-10-06 11:31:11 -0700678 CRUSH_RULE_CHOOSE_LEAF_INDEP;
679
680 /* reset output */
681 osize = 0;
682
683 for (i = 0; i < wsize; i++) {
684 /*
685 * see CRUSH_N, CRUSH_N_MINUS macros.
686 * basically, numrep <= 0 means relative to
687 * the provided result_max
688 */
Sage Weil06682162012-05-07 15:35:48 -0700689 numrep = curstep->arg1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700690 if (numrep <= 0) {
691 numrep += result_max;
692 if (numrep <= 0)
693 continue;
694 }
695 j = 0;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200696 if (firstn) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200697 osize += crush_choose_firstn(
698 map,
699 map->buckets[-1-w[i]],
700 weight, weight_max,
701 x, numrep,
702 curstep->arg2,
703 o+osize, j,
704 recurse_to_leaf,
705 descend_once, c+osize);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200706 } else {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200707 crush_choose_indep(
708 map,
709 map->buckets[-1-w[i]],
710 weight, weight_max,
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200711 x, numrep, numrep,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200712 curstep->arg2,
713 o+osize, j,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200714 map->choose_total_tries,
715 choose_leaf_tries,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200716 recurse_to_leaf,
Ilya Dryomov41586082013-12-24 21:19:25 +0200717 c+osize,
718 0);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200719 osize += numrep;
720 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700721 }
722
723 if (recurse_to_leaf)
724 /* copy final _leaf_ values to output set */
725 memcpy(o, c, osize*sizeof(*o));
726
Ilya Dryomov2a4ba742013-12-24 21:19:24 +0200727 /* swap o and w arrays */
Sage Weil5ecc0a02009-10-06 11:31:11 -0700728 tmp = o;
729 o = w;
730 w = tmp;
731 wsize = osize;
732 break;
733
734
735 case CRUSH_RULE_EMIT:
736 for (i = 0; i < wsize && result_len < result_max; i++) {
737 result[result_len] = w[i];
738 result_len++;
739 }
740 wsize = 0;
741 break;
742
743 default:
Sage Weila1f48952012-05-07 15:35:24 -0700744 dprintk(" unknown op %d at step %d\n",
745 curstep->op, step);
746 break;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700747 }
748 }
Sage Weilf1932fc2011-12-07 09:10:26 -0800749 return result_len;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700750}
751
752