blob: e9256a30e60d3f12442fdcfd6496ebffa3c62d27 [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,
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200302 unsigned int attempts,
303 unsigned int recurse_attempts,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200304 int recurse_to_leaf,
305 int descend_once, int *out2)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700306{
307 int rep;
Sage Weil8b12d472012-05-07 15:38:35 -0700308 unsigned int ftotal, flocal;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700309 int retry_descent, retry_bucket, skip_rep;
310 struct crush_bucket *in = bucket;
311 int r;
312 int i;
Sage Weilb28813a2009-10-07 10:59:34 -0700313 int item = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700314 int itemtype;
315 int collide, reject;
Sage Weila1a31e72010-06-24 12:58:14 -0700316
317 dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
318 bucket->id, x, outpos, numrep);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700319
320 for (rep = outpos; rep < numrep; rep++) {
321 /* keep trying until we get a non-out, non-colliding item */
322 ftotal = 0;
323 skip_rep = 0;
324 do {
325 retry_descent = 0;
326 in = bucket; /* initial bucket */
327
328 /* choose through intervening buckets */
329 flocal = 0;
330 do {
Sage Weilb28813a2009-10-07 10:59:34 -0700331 collide = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700332 retry_bucket = 0;
333 r = rep;
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200334 /* r' = r + f_total */
335 r += ftotal;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700336
337 /* bucket choose */
Sage Weilb28813a2009-10-07 10:59:34 -0700338 if (in->size == 0) {
339 reject = 1;
340 goto reject;
341 }
Sage Weil546f04e2012-07-30 18:15:23 -0700342 if (map->choose_local_fallback_tries > 0 &&
343 flocal >= (in->size>>1) &&
344 flocal > map->choose_local_fallback_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700345 item = bucket_perm_choose(in, x, r);
346 else
347 item = crush_bucket_choose(in, x, r);
Sage Weila1f48952012-05-07 15:35:24 -0700348 if (item >= map->max_devices) {
349 dprintk(" bad item %d\n", item);
350 skip_rep = 1;
351 break;
352 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700353
354 /* desired type? */
355 if (item < 0)
356 itemtype = map->buckets[-1-item]->type;
357 else
358 itemtype = 0;
359 dprintk(" item %d type %d\n", item, itemtype);
360
361 /* keep going? */
362 if (itemtype != type) {
Sage Weila1f48952012-05-07 15:35:24 -0700363 if (item >= 0 ||
364 (-1-item) >= map->max_buckets) {
365 dprintk(" bad item type %d\n", type);
366 skip_rep = 1;
367 break;
368 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700369 in = map->buckets[-1-item];
Sage Weil55bda7a2010-06-24 12:55:48 -0700370 retry_bucket = 1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700371 continue;
372 }
373
374 /* collision? */
Sage Weil5ecc0a02009-10-06 11:31:11 -0700375 for (i = 0; i < outpos; i++) {
376 if (out[i] == item) {
377 collide = 1;
378 break;
379 }
380 }
381
Sage Weila1a31e72010-06-24 12:58:14 -0700382 reject = 0;
Sage Weil7d7c1f62013-01-15 18:49:09 -0800383 if (!collide && recurse_to_leaf) {
Sage Weila1a31e72010-06-24 12:58:14 -0700384 if (item < 0) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200385 if (crush_choose_firstn(map,
Sage Weila1a31e72010-06-24 12:58:14 -0700386 map->buckets[-1-item],
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200387 weight, weight_max,
Sage Weila1a31e72010-06-24 12:58:14 -0700388 x, outpos+1, 0,
389 out2, outpos,
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200390 recurse_attempts, 0,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200391 0,
Jim Schutt1604f482012-11-30 09:15:25 -0700392 map->chooseleaf_descend_once,
Sage Weila1a31e72010-06-24 12:58:14 -0700393 NULL) <= outpos)
394 /* didn't get leaf */
395 reject = 1;
396 } else {
397 /* we already have a leaf! */
398 out2[outpos] = item;
399 }
400 }
401
402 if (!reject) {
Sage Weil5ecc0a02009-10-06 11:31:11 -0700403 /* out? */
404 if (itemtype == 0)
405 reject = is_out(map, weight,
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200406 weight_max,
Sage Weil5ecc0a02009-10-06 11:31:11 -0700407 item, x);
408 else
409 reject = 0;
410 }
411
Sage Weilb28813a2009-10-07 10:59:34 -0700412reject:
Sage Weil5ecc0a02009-10-06 11:31:11 -0700413 if (reject || collide) {
414 ftotal++;
415 flocal++;
416
Jim Schutt1604f482012-11-30 09:15:25 -0700417 if (reject && descend_once)
418 /* let outer call try again */
419 skip_rep = 1;
420 else if (collide && flocal <= map->choose_local_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700421 /* retry locally a few times */
422 retry_bucket = 1;
Sage Weil546f04e2012-07-30 18:15:23 -0700423 else if (map->choose_local_fallback_tries > 0 &&
424 flocal <= in->size + map->choose_local_fallback_tries)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700425 /* exhaustive bucket search */
426 retry_bucket = 1;
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200427 else if (ftotal <= attempts)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700428 /* then retry descent */
429 retry_descent = 1;
430 else
431 /* else give up */
432 skip_rep = 1;
433 dprintk(" reject %d collide %d "
Sage Weil8b12d472012-05-07 15:38:35 -0700434 "ftotal %u flocal %u\n",
Sage Weil5ecc0a02009-10-06 11:31:11 -0700435 reject, collide, ftotal,
436 flocal);
437 }
438 } while (retry_bucket);
439 } while (retry_descent);
440
441 if (skip_rep) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200442 dprintk("skip rep\n");
443 continue;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700444 }
445
Sage Weila1a31e72010-06-24 12:58:14 -0700446 dprintk("CHOOSE got %d\n", item);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700447 out[outpos] = item;
448 outpos++;
449 }
450
Sage Weila1a31e72010-06-24 12:58:14 -0700451 dprintk("CHOOSE returns %d\n", outpos);
Sage Weil5ecc0a02009-10-06 11:31:11 -0700452 return outpos;
453}
454
455
456/**
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200457 * crush_choose_indep: alternative breadth-first positionally stable mapping
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200458 *
459 */
460static void crush_choose_indep(const struct crush_map *map,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200461 struct crush_bucket *bucket,
462 const __u32 *weight, int weight_max,
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200463 int x, int left, int numrep, int type,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200464 int *out, int outpos,
465 unsigned int attempts,
466 unsigned int recurse_attempts,
467 int recurse_to_leaf,
Ilya Dryomov41586082013-12-24 21:19:25 +0200468 int *out2,
469 int parent_r)
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200470{
471 struct crush_bucket *in = bucket;
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200472 int endpos = outpos + left;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200473 int rep;
474 unsigned int ftotal;
475 int r;
476 int i;
477 int item = 0;
478 int itemtype;
479 int collide;
480
481 dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
482 bucket->id, x, outpos, numrep);
483
484 /* initially my result is undefined */
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200485 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200486 out[rep] = CRUSH_ITEM_UNDEF;
487 if (out2)
488 out2[rep] = CRUSH_ITEM_UNDEF;
489 }
490
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200491 for (ftotal = 0; left > 0 && ftotal < attempts; ftotal++) {
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200492 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200493 if (out[rep] != CRUSH_ITEM_UNDEF)
494 continue;
495
496 in = bucket; /* initial bucket */
497
498 /* choose through intervening buckets */
499 for (;;) {
Ilya Dryomov3102b0a2013-12-24 21:19:25 +0200500 /* note: we base the choice on the position
501 * even in the nested call. that means that
502 * if the first layer chooses the same bucket
503 * in a different position, we will tend to
504 * choose a different item in that bucket.
505 * this will involve more devices in data
506 * movement and tend to distribute the load.
507 */
Ilya Dryomov41586082013-12-24 21:19:25 +0200508 r = rep + parent_r;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200509
510 /* be careful */
511 if (in->alg == CRUSH_BUCKET_UNIFORM &&
512 in->size % numrep == 0)
513 /* r'=r+(n+1)*f_total */
514 r += (numrep+1) * ftotal;
515 else
516 /* r' = r + n*f_total */
517 r += numrep * ftotal;
518
519 /* bucket choose */
520 if (in->size == 0) {
521 dprintk(" empty bucket\n");
522 break;
523 }
524
525 item = crush_bucket_choose(in, x, r);
526 if (item >= map->max_devices) {
527 dprintk(" bad item %d\n", item);
528 out[rep] = CRUSH_ITEM_NONE;
529 if (out2)
530 out2[rep] = CRUSH_ITEM_NONE;
531 left--;
532 break;
533 }
534
535 /* desired type? */
536 if (item < 0)
537 itemtype = map->buckets[-1-item]->type;
538 else
539 itemtype = 0;
540 dprintk(" item %d type %d\n", item, itemtype);
541
542 /* keep going? */
543 if (itemtype != type) {
544 if (item >= 0 ||
545 (-1-item) >= map->max_buckets) {
546 dprintk(" bad item type %d\n", type);
547 out[rep] = CRUSH_ITEM_NONE;
548 if (out2)
549 out2[rep] =
550 CRUSH_ITEM_NONE;
551 left--;
552 break;
553 }
554 in = map->buckets[-1-item];
555 continue;
556 }
557
558 /* collision? */
559 collide = 0;
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200560 for (i = outpos; i < endpos; i++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200561 if (out[i] == item) {
562 collide = 1;
563 break;
564 }
565 }
566 if (collide)
567 break;
568
569 if (recurse_to_leaf) {
570 if (item < 0) {
571 crush_choose_indep(map,
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200572 map->buckets[-1-item],
573 weight, weight_max,
574 x, 1, numrep, 0,
575 out2, rep,
576 recurse_attempts, 0,
577 0, NULL, r);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200578 if (out2[rep] == CRUSH_ITEM_NONE) {
579 /* placed nothing; no leaf */
580 break;
581 }
582 } else {
583 /* we already have a leaf! */
584 out2[rep] = item;
585 }
586 }
587
588 /* out? */
589 if (itemtype == 0 &&
590 is_out(map, weight, weight_max, item, x))
591 break;
592
593 /* yay! */
594 out[rep] = item;
595 left--;
596 break;
597 }
598 }
599 }
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200600 for (rep = outpos; rep < endpos; rep++) {
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200601 if (out[rep] == CRUSH_ITEM_UNDEF) {
602 out[rep] = CRUSH_ITEM_NONE;
603 }
604 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
605 out2[rep] = CRUSH_ITEM_NONE;
606 }
607 }
608}
609
610/**
Sage Weil5ecc0a02009-10-06 11:31:11 -0700611 * crush_do_rule - calculate a mapping with the given input and rule
612 * @map: the crush_map
613 * @ruleno: the rule id
614 * @x: hash input
615 * @result: pointer to result vector
616 * @result_max: maximum result size
Ilya Dryomovb3b33b02013-12-24 21:19:24 +0200617 * @weight: weight vector (for map leaves)
618 * @weight_max: size of weight vector
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200619 * @scratch: scratch vector for private use; must be >= 3 * result_max
Sage Weil5ecc0a02009-10-06 11:31:11 -0700620 */
Sage Weil8b12d472012-05-07 15:38:35 -0700621int crush_do_rule(const struct crush_map *map,
Sage Weil5ecc0a02009-10-06 11:31:11 -0700622 int ruleno, int x, int *result, int result_max,
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200623 const __u32 *weight, int weight_max,
624 int *scratch)
Sage Weil5ecc0a02009-10-06 11:31:11 -0700625{
626 int result_len;
Ilya Dryomove8ef19c2013-12-24 21:19:24 +0200627 int *a = scratch;
628 int *b = scratch + result_max;
629 int *c = scratch + result_max*2;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700630 int recurse_to_leaf;
631 int *w;
632 int wsize = 0;
633 int *o;
634 int osize;
635 int *tmp;
636 struct crush_rule *rule;
Sage Weil8b12d472012-05-07 15:38:35 -0700637 __u32 step;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700638 int i, j;
639 int numrep;
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200640 int choose_tries = map->choose_total_tries;
641 int choose_leaf_tries = 0;
Jim Schutt1604f482012-11-30 09:15:25 -0700642 const int descend_once = 0;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700643
Sage Weila1f48952012-05-07 15:35:24 -0700644 if ((__u32)ruleno >= map->max_rules) {
645 dprintk(" bad ruleno %d\n", ruleno);
646 return 0;
647 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700648
649 rule = map->rules[ruleno];
650 result_len = 0;
651 w = a;
652 o = b;
653
Sage Weil5ecc0a02009-10-06 11:31:11 -0700654 for (step = 0; step < rule->len; step++) {
Ilya Dryomov8f99c852013-12-24 21:19:24 +0200655 int firstn = 0;
Sage Weil06682162012-05-07 15:35:48 -0700656 struct crush_rule_step *curstep = &rule->steps[step];
657
Sage Weil06682162012-05-07 15:35:48 -0700658 switch (curstep->op) {
Sage Weil5ecc0a02009-10-06 11:31:11 -0700659 case CRUSH_RULE_TAKE:
Sage Weil06682162012-05-07 15:35:48 -0700660 w[0] = curstep->arg1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700661 wsize = 1;
662 break;
663
Ilya Dryomovcc10df42013-12-24 21:19:26 +0200664 case CRUSH_RULE_SET_CHOOSE_TRIES:
665 if (curstep->arg1 > 0)
666 choose_tries = curstep->arg1;
667 break;
668
Ilya Dryomov917edad2013-12-24 21:19:26 +0200669 case CRUSH_RULE_SET_CHOOSELEAF_TRIES:
Ilya Dryomovbe3226a2013-12-24 21:19:26 +0200670 if (curstep->arg1 > 0)
671 choose_leaf_tries = curstep->arg1;
672 break;
673
Ilya Dryomov917edad2013-12-24 21:19:26 +0200674 case CRUSH_RULE_CHOOSELEAF_FIRSTN:
Sage Weil5ecc0a02009-10-06 11:31:11 -0700675 case CRUSH_RULE_CHOOSE_FIRSTN:
676 firstn = 1;
Sage Weil06682162012-05-07 15:35:48 -0700677 /* fall through */
Ilya Dryomov917edad2013-12-24 21:19:26 +0200678 case CRUSH_RULE_CHOOSELEAF_INDEP:
Sage Weil5ecc0a02009-10-06 11:31:11 -0700679 case CRUSH_RULE_CHOOSE_INDEP:
Sage Weila1f48952012-05-07 15:35:24 -0700680 if (wsize == 0)
681 break;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700682
683 recurse_to_leaf =
Sage Weil06682162012-05-07 15:35:48 -0700684 curstep->op ==
Ilya Dryomov917edad2013-12-24 21:19:26 +0200685 CRUSH_RULE_CHOOSELEAF_FIRSTN ||
Sage Weil06682162012-05-07 15:35:48 -0700686 curstep->op ==
Ilya Dryomov917edad2013-12-24 21:19:26 +0200687 CRUSH_RULE_CHOOSELEAF_INDEP;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700688
689 /* reset output */
690 osize = 0;
691
692 for (i = 0; i < wsize; i++) {
693 /*
694 * see CRUSH_N, CRUSH_N_MINUS macros.
695 * basically, numrep <= 0 means relative to
696 * the provided result_max
697 */
Sage Weil06682162012-05-07 15:35:48 -0700698 numrep = curstep->arg1;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700699 if (numrep <= 0) {
700 numrep += result_max;
701 if (numrep <= 0)
702 continue;
703 }
704 j = 0;
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200705 if (firstn) {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200706 osize += crush_choose_firstn(
707 map,
708 map->buckets[-1-w[i]],
709 weight, weight_max,
710 x, numrep,
711 curstep->arg2,
712 o+osize, j,
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200713 choose_tries,
714 choose_leaf_tries ? choose_leaf_tries : choose_tries,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200715 recurse_to_leaf,
716 descend_once, c+osize);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200717 } else {
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200718 crush_choose_indep(
719 map,
720 map->buckets[-1-w[i]],
721 weight, weight_max,
Ilya Dryomovab4ce2b52013-12-24 21:19:25 +0200722 x, numrep, numrep,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200723 curstep->arg2,
724 o+osize, j,
Ilya Dryomovf18650a2013-12-24 21:19:26 +0200725 choose_tries,
726 choose_leaf_tries ? choose_leaf_tries : 1,
Ilya Dryomov9fe07182013-12-24 21:19:25 +0200727 recurse_to_leaf,
Ilya Dryomov41586082013-12-24 21:19:25 +0200728 c+osize,
729 0);
Ilya Dryomov9a3b4902013-12-24 21:19:25 +0200730 osize += numrep;
731 }
Sage Weil5ecc0a02009-10-06 11:31:11 -0700732 }
733
734 if (recurse_to_leaf)
735 /* copy final _leaf_ values to output set */
736 memcpy(o, c, osize*sizeof(*o));
737
Ilya Dryomov2a4ba742013-12-24 21:19:24 +0200738 /* swap o and w arrays */
Sage Weil5ecc0a02009-10-06 11:31:11 -0700739 tmp = o;
740 o = w;
741 w = tmp;
742 wsize = osize;
743 break;
744
745
746 case CRUSH_RULE_EMIT:
747 for (i = 0; i < wsize && result_len < result_max; i++) {
748 result[result_len] = w[i];
749 result_len++;
750 }
751 wsize = 0;
752 break;
753
754 default:
Sage Weila1f48952012-05-07 15:35:24 -0700755 dprintk(" unknown op %d at step %d\n",
756 curstep->op, step);
757 break;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700758 }
759 }
Sage Weilf1932fc2011-12-07 09:10:26 -0800760 return result_len;
Sage Weil5ecc0a02009-10-06 11:31:11 -0700761}
762
763