]>
Commit | Line | Data |
---|---|---|
5ecc0a0f SW |
1 | |
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 | ||
3d14c5d2 YS |
21 | #include <linux/crush/crush.h> |
22 | #include <linux/crush/hash.h> | |
958a2765 | 23 | #include "crush_ln_table.h" |
5ecc0a0f SW |
24 | |
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 | */ | |
8b12d47b | 36 | int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size) |
5ecc0a0f | 37 | { |
8b12d47b | 38 | __u32 i; |
5ecc0a0f SW |
39 | |
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 | */ | |
69 | static int bucket_perm_choose(struct crush_bucket *bucket, | |
70 | int x, int r) | |
71 | { | |
95c96174 ED |
72 | unsigned int pr = r % bucket->size; |
73 | unsigned int i, s; | |
5ecc0a0f SW |
74 | |
75 | /* start a new permutation if @x has changed */ | |
8b12d47b | 76 | if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) { |
5ecc0a0f SW |
77 | 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) { | |
fb690390 | 82 | s = crush_hash32_3(bucket->hash, x, bucket->id, 0) % |
5ecc0a0f SW |
83 | 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) { | |
95c96174 | 104 | unsigned int p = bucket->perm_n; |
5ecc0a0f SW |
105 | /* no point in swapping the final entry */ |
106 | if (p < bucket->size - 1) { | |
fb690390 | 107 | i = crush_hash32_3(bucket->hash, x, bucket->id, p) % |
5ecc0a0f SW |
108 | (bucket->size - p); |
109 | if (i) { | |
95c96174 | 110 | unsigned int t = bucket->perm[p + i]; |
5ecc0a0f SW |
111 | 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]; | |
122 | out: | |
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 */ | |
129 | static 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 */ | |
136 | static 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--) { | |
fb690390 SW |
142 | __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i], |
143 | r, bucket->h.id); | |
5ecc0a0f SW |
144 | 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 | ||
a1f4895b SW |
156 | dprintk("bad list sums for bucket %d\n", bucket->h.id); |
157 | return bucket->h.items[0]; | |
5ecc0a0f SW |
158 | } |
159 | ||
160 | ||
161 | /* (binary) tree */ | |
162 | static 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 | ||
172 | static int left(int x) | |
173 | { | |
174 | int h = height(x); | |
175 | return x - (1 << (h-1)); | |
176 | } | |
177 | ||
178 | static int right(int x) | |
179 | { | |
180 | int h = height(x); | |
181 | return x + (1 << (h-1)); | |
182 | } | |
183 | ||
184 | static int terminal(int x) | |
185 | { | |
186 | return x & 1; | |
187 | } | |
188 | ||
189 | static int bucket_tree_choose(struct crush_bucket_tree *bucket, | |
190 | int x, int r) | |
191 | { | |
8f99c85b | 192 | int n; |
5ecc0a0f SW |
193 | __u32 w; |
194 | __u64 t; | |
195 | ||
196 | /* start at root */ | |
197 | n = bucket->num_nodes >> 1; | |
198 | ||
199 | while (!terminal(n)) { | |
8f99c85b | 200 | int l; |
5ecc0a0f SW |
201 | /* pick point in [0, w) */ |
202 | w = bucket->node_weights[n]; | |
fb690390 SW |
203 | t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r, |
204 | bucket->h.id) * (__u64)w; | |
5ecc0a0f SW |
205 | 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 | ||
221 | static int bucket_straw_choose(struct crush_bucket_straw *bucket, | |
222 | int x, int r) | |
223 | { | |
8b12d47b | 224 | __u32 i; |
5ecc0a0f SW |
225 | int high = 0; |
226 | __u64 high_draw = 0; | |
227 | __u64 draw; | |
228 | ||
229 | for (i = 0; i < bucket->h.size; i++) { | |
fb690390 | 230 | draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r); |
5ecc0a0f SW |
231 | 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 | ||
958a2765 ID |
241 | // compute 2^44*log2(input+1) |
242 | uint64_t crush_ln(unsigned xin) | |
243 | { | |
244 | unsigned x=xin, x1; | |
245 | int iexpon, index1, index2; | |
246 | uint64_t RH, LH, LL, xl64, result; | |
247 | ||
248 | x++; | |
249 | ||
250 | // normalize input | |
251 | iexpon = 15; | |
252 | while(!(x&0x18000)) { x<<=1; iexpon--; } | |
253 | ||
254 | index1 = (x>>8)<<1; | |
255 | // RH ~ 2^56/index1 | |
256 | RH = __RH_LH_tbl[index1 - 256]; | |
257 | // LH ~ 2^48 * log2(index1/256) | |
258 | LH = __RH_LH_tbl[index1 + 1 - 256]; | |
259 | ||
260 | // RH*x ~ 2^48 * (2^15 + xf), xf<2^8 | |
261 | xl64 = (int64_t)x * RH; | |
262 | xl64 >>= 48; | |
263 | x1 = xl64; | |
264 | ||
265 | result = iexpon; | |
266 | result <<= (12 + 32); | |
267 | ||
268 | index2 = x1 & 0xff; | |
269 | // LL ~ 2^48*log2(1.0+index2/2^15) | |
270 | LL = __LL_tbl[index2]; | |
271 | ||
272 | LH = LH + LL; | |
273 | ||
274 | LH >>= (48-12 - 32); | |
275 | result += LH; | |
276 | ||
277 | return result; | |
278 | } | |
279 | ||
280 | ||
281 | /* | |
282 | * straw2 | |
283 | * | |
284 | * for reference, see: | |
285 | * | |
286 | * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables | |
287 | * | |
288 | */ | |
289 | ||
290 | static int bucket_straw2_choose(struct crush_bucket_straw2 *bucket, | |
291 | int x, int r) | |
292 | { | |
293 | unsigned i, high = 0; | |
294 | unsigned u; | |
295 | unsigned w; | |
296 | __s64 ln, draw, high_draw = 0; | |
297 | ||
298 | for (i = 0; i < bucket->h.size; i++) { | |
299 | w = bucket->item_weights[i]; | |
300 | if (w) { | |
301 | u = crush_hash32_3(bucket->h.hash, x, | |
302 | bucket->h.items[i], r); | |
303 | u &= 0xffff; | |
304 | ||
305 | /* | |
306 | * for some reason slightly less than 0x10000 produces | |
307 | * a slightly more accurate distribution... probably a | |
308 | * rounding effect. | |
309 | * | |
310 | * the natural log lookup table maps [0,0xffff] | |
311 | * (corresponding to real numbers [1/0x10000, 1] to | |
312 | * [0, 0xffffffffffff] (corresponding to real numbers | |
313 | * [-11.090355,0]). | |
314 | */ | |
315 | ln = crush_ln(u) - 0x1000000000000ll; | |
316 | ||
317 | /* | |
318 | * divide by 16.16 fixed-point weight. note | |
319 | * that the ln value is negative, so a larger | |
320 | * weight means a larger (less negative) value | |
321 | * for draw. | |
322 | */ | |
323 | draw = div64_s64(ln, w); | |
324 | } else { | |
325 | draw = S64_MIN; | |
326 | } | |
327 | ||
328 | if (i == 0 || draw > high_draw) { | |
329 | high = i; | |
330 | high_draw = draw; | |
331 | } | |
332 | } | |
333 | return bucket->h.items[high]; | |
334 | } | |
335 | ||
336 | ||
5ecc0a0f SW |
337 | static int crush_bucket_choose(struct crush_bucket *in, int x, int r) |
338 | { | |
a1a31e73 | 339 | dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r); |
a1f4895b | 340 | BUG_ON(in->size == 0); |
5ecc0a0f SW |
341 | switch (in->alg) { |
342 | case CRUSH_BUCKET_UNIFORM: | |
343 | return bucket_uniform_choose((struct crush_bucket_uniform *)in, | |
344 | x, r); | |
345 | case CRUSH_BUCKET_LIST: | |
346 | return bucket_list_choose((struct crush_bucket_list *)in, | |
347 | x, r); | |
348 | case CRUSH_BUCKET_TREE: | |
349 | return bucket_tree_choose((struct crush_bucket_tree *)in, | |
350 | x, r); | |
351 | case CRUSH_BUCKET_STRAW: | |
352 | return bucket_straw_choose((struct crush_bucket_straw *)in, | |
353 | x, r); | |
958a2765 ID |
354 | case CRUSH_BUCKET_STRAW2: |
355 | return bucket_straw2_choose((struct crush_bucket_straw2 *)in, | |
356 | x, r); | |
5ecc0a0f | 357 | default: |
a1f4895b | 358 | dprintk("unknown bucket %d alg %d\n", in->id, in->alg); |
50b885b9 | 359 | return in->items[0]; |
5ecc0a0f SW |
360 | } |
361 | } | |
362 | ||
958a2765 | 363 | |
5ecc0a0f SW |
364 | /* |
365 | * true if device is marked "out" (failed, fully offloaded) | |
366 | * of the cluster | |
367 | */ | |
b3b33b0e ID |
368 | static int is_out(const struct crush_map *map, |
369 | const __u32 *weight, int weight_max, | |
370 | int item, int x) | |
5ecc0a0f | 371 | { |
b3b33b0e ID |
372 | if (item >= weight_max) |
373 | return 1; | |
153a1093 | 374 | if (weight[item] >= 0x10000) |
5ecc0a0f SW |
375 | return 0; |
376 | if (weight[item] == 0) | |
377 | return 1; | |
fb690390 SW |
378 | if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff) |
379 | < weight[item]) | |
5ecc0a0f SW |
380 | return 0; |
381 | return 1; | |
382 | } | |
383 | ||
384 | /** | |
9fe07182 | 385 | * crush_choose_firstn - choose numrep distinct items of given type |
5ecc0a0f SW |
386 | * @map: the crush_map |
387 | * @bucket: the bucket we are choose an item from | |
388 | * @x: crush input value | |
389 | * @numrep: the number of items to choose | |
390 | * @type: the type of item to choose | |
391 | * @out: pointer to output vector | |
392 | * @outpos: our position in that vector | |
45002267 | 393 | * @out_size: size of the out vector |
0e32d712 ID |
394 | * @tries: number of attempts to make |
395 | * @recurse_tries: number of attempts to have recursive chooseleaf make | |
48a163db ID |
396 | * @local_retries: localized retries |
397 | * @local_fallback_retries: localized fallback retries | |
0e32d712 | 398 | * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) |
e2b149cc | 399 | * @vary_r: pass r to recursive calls |
5ecc0a0f | 400 | * @out2: second output vector for leaf items (if @recurse_to_leaf) |
e2b149cc | 401 | * @parent_r: r value passed from the parent |
5ecc0a0f | 402 | */ |
9fe07182 ID |
403 | static int crush_choose_firstn(const struct crush_map *map, |
404 | struct crush_bucket *bucket, | |
405 | const __u32 *weight, int weight_max, | |
406 | int x, int numrep, int type, | |
407 | int *out, int outpos, | |
45002267 | 408 | int out_size, |
2d8be0bc ID |
409 | unsigned int tries, |
410 | unsigned int recurse_tries, | |
48a163db ID |
411 | unsigned int local_retries, |
412 | unsigned int local_fallback_retries, | |
9fe07182 | 413 | int recurse_to_leaf, |
e2b149cc ID |
414 | unsigned int vary_r, |
415 | int *out2, | |
416 | int parent_r) | |
5ecc0a0f SW |
417 | { |
418 | int rep; | |
8b12d47b | 419 | unsigned int ftotal, flocal; |
5ecc0a0f SW |
420 | int retry_descent, retry_bucket, skip_rep; |
421 | struct crush_bucket *in = bucket; | |
422 | int r; | |
423 | int i; | |
b28813a6 | 424 | int item = 0; |
5ecc0a0f SW |
425 | int itemtype; |
426 | int collide, reject; | |
45002267 | 427 | int count = out_size; |
a1a31e73 | 428 | |
e2b149cc ID |
429 | dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d tries %d recurse_tries %d local_retries %d local_fallback_retries %d parent_r %d\n", |
430 | recurse_to_leaf ? "_LEAF" : "", | |
431 | bucket->id, x, outpos, numrep, | |
432 | tries, recurse_tries, local_retries, local_fallback_retries, | |
433 | parent_r); | |
5ecc0a0f | 434 | |
45002267 | 435 | for (rep = outpos; rep < numrep && count > 0 ; rep++) { |
5ecc0a0f SW |
436 | /* keep trying until we get a non-out, non-colliding item */ |
437 | ftotal = 0; | |
438 | skip_rep = 0; | |
439 | do { | |
440 | retry_descent = 0; | |
441 | in = bucket; /* initial bucket */ | |
442 | ||
443 | /* choose through intervening buckets */ | |
444 | flocal = 0; | |
445 | do { | |
b28813a6 | 446 | collide = 0; |
5ecc0a0f | 447 | retry_bucket = 0; |
e2b149cc | 448 | r = rep + parent_r; |
9fe07182 ID |
449 | /* r' = r + f_total */ |
450 | r += ftotal; | |
5ecc0a0f SW |
451 | |
452 | /* bucket choose */ | |
b28813a6 SW |
453 | if (in->size == 0) { |
454 | reject = 1; | |
455 | goto reject; | |
456 | } | |
48a163db | 457 | if (local_fallback_retries > 0 && |
546f04ef | 458 | flocal >= (in->size>>1) && |
48a163db | 459 | flocal > local_fallback_retries) |
5ecc0a0f SW |
460 | item = bucket_perm_choose(in, x, r); |
461 | else | |
462 | item = crush_bucket_choose(in, x, r); | |
a1f4895b SW |
463 | if (item >= map->max_devices) { |
464 | dprintk(" bad item %d\n", item); | |
465 | skip_rep = 1; | |
466 | break; | |
467 | } | |
5ecc0a0f SW |
468 | |
469 | /* desired type? */ | |
470 | if (item < 0) | |
471 | itemtype = map->buckets[-1-item]->type; | |
472 | else | |
473 | itemtype = 0; | |
474 | dprintk(" item %d type %d\n", item, itemtype); | |
475 | ||
476 | /* keep going? */ | |
477 | if (itemtype != type) { | |
a1f4895b SW |
478 | if (item >= 0 || |
479 | (-1-item) >= map->max_buckets) { | |
480 | dprintk(" bad item type %d\n", type); | |
481 | skip_rep = 1; | |
482 | break; | |
483 | } | |
5ecc0a0f | 484 | in = map->buckets[-1-item]; |
55bda7aa | 485 | retry_bucket = 1; |
5ecc0a0f SW |
486 | continue; |
487 | } | |
488 | ||
489 | /* collision? */ | |
5ecc0a0f SW |
490 | for (i = 0; i < outpos; i++) { |
491 | if (out[i] == item) { | |
492 | collide = 1; | |
493 | break; | |
494 | } | |
495 | } | |
496 | ||
a1a31e73 | 497 | reject = 0; |
7d7c1f61 | 498 | if (!collide && recurse_to_leaf) { |
a1a31e73 | 499 | if (item < 0) { |
e2b149cc ID |
500 | int sub_r; |
501 | if (vary_r) | |
502 | sub_r = r >> (vary_r-1); | |
503 | else | |
504 | sub_r = 0; | |
9fe07182 | 505 | if (crush_choose_firstn(map, |
a1a31e73 | 506 | map->buckets[-1-item], |
b3b33b0e | 507 | weight, weight_max, |
a1a31e73 | 508 | x, outpos+1, 0, |
45002267 | 509 | out2, outpos, count, |
2d8be0bc | 510 | recurse_tries, 0, |
48a163db ID |
511 | local_retries, |
512 | local_fallback_retries, | |
9fe07182 | 513 | 0, |
e2b149cc ID |
514 | vary_r, |
515 | NULL, | |
516 | sub_r) <= outpos) | |
a1a31e73 SW |
517 | /* didn't get leaf */ |
518 | reject = 1; | |
519 | } else { | |
520 | /* we already have a leaf! */ | |
521 | out2[outpos] = item; | |
522 | } | |
523 | } | |
524 | ||
525 | if (!reject) { | |
5ecc0a0f SW |
526 | /* out? */ |
527 | if (itemtype == 0) | |
528 | reject = is_out(map, weight, | |
b3b33b0e | 529 | weight_max, |
5ecc0a0f SW |
530 | item, x); |
531 | else | |
532 | reject = 0; | |
533 | } | |
534 | ||
b28813a6 | 535 | reject: |
5ecc0a0f SW |
536 | if (reject || collide) { |
537 | ftotal++; | |
538 | flocal++; | |
539 | ||
48a163db | 540 | if (collide && flocal <= local_retries) |
5ecc0a0f SW |
541 | /* retry locally a few times */ |
542 | retry_bucket = 1; | |
48a163db ID |
543 | else if (local_fallback_retries > 0 && |
544 | flocal <= in->size + local_fallback_retries) | |
5ecc0a0f SW |
545 | /* exhaustive bucket search */ |
546 | retry_bucket = 1; | |
48a163db | 547 | else if (ftotal < tries) |
5ecc0a0f SW |
548 | /* then retry descent */ |
549 | retry_descent = 1; | |
550 | else | |
551 | /* else give up */ | |
552 | skip_rep = 1; | |
553 | dprintk(" reject %d collide %d " | |
8b12d47b | 554 | "ftotal %u flocal %u\n", |
5ecc0a0f SW |
555 | reject, collide, ftotal, |
556 | flocal); | |
557 | } | |
558 | } while (retry_bucket); | |
559 | } while (retry_descent); | |
560 | ||
561 | if (skip_rep) { | |
9fe07182 ID |
562 | dprintk("skip rep\n"); |
563 | continue; | |
5ecc0a0f SW |
564 | } |
565 | ||
a1a31e73 | 566 | dprintk("CHOOSE got %d\n", item); |
5ecc0a0f SW |
567 | out[outpos] = item; |
568 | outpos++; | |
45002267 | 569 | count--; |
5ecc0a0f SW |
570 | } |
571 | ||
a1a31e73 | 572 | dprintk("CHOOSE returns %d\n", outpos); |
5ecc0a0f SW |
573 | return outpos; |
574 | } | |
575 | ||
576 | ||
9a3b490a | 577 | /** |
9fe07182 | 578 | * crush_choose_indep: alternative breadth-first positionally stable mapping |
9a3b490a ID |
579 | * |
580 | */ | |
581 | static void crush_choose_indep(const struct crush_map *map, | |
be3226ac ID |
582 | struct crush_bucket *bucket, |
583 | const __u32 *weight, int weight_max, | |
ab4ce2b5 | 584 | int x, int left, int numrep, int type, |
be3226ac | 585 | int *out, int outpos, |
2d8be0bc ID |
586 | unsigned int tries, |
587 | unsigned int recurse_tries, | |
be3226ac | 588 | int recurse_to_leaf, |
41586081 ID |
589 | int *out2, |
590 | int parent_r) | |
9a3b490a ID |
591 | { |
592 | struct crush_bucket *in = bucket; | |
ab4ce2b5 | 593 | int endpos = outpos + left; |
9a3b490a ID |
594 | int rep; |
595 | unsigned int ftotal; | |
596 | int r; | |
597 | int i; | |
598 | int item = 0; | |
599 | int itemtype; | |
600 | int collide; | |
601 | ||
602 | dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", | |
603 | bucket->id, x, outpos, numrep); | |
604 | ||
605 | /* initially my result is undefined */ | |
ab4ce2b5 | 606 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
607 | out[rep] = CRUSH_ITEM_UNDEF; |
608 | if (out2) | |
609 | out2[rep] = CRUSH_ITEM_UNDEF; | |
610 | } | |
611 | ||
2d8be0bc | 612 | for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { |
ab4ce2b5 | 613 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
614 | if (out[rep] != CRUSH_ITEM_UNDEF) |
615 | continue; | |
616 | ||
617 | in = bucket; /* initial bucket */ | |
618 | ||
619 | /* choose through intervening buckets */ | |
620 | for (;;) { | |
3102b0a5 ID |
621 | /* note: we base the choice on the position |
622 | * even in the nested call. that means that | |
623 | * if the first layer chooses the same bucket | |
624 | * in a different position, we will tend to | |
625 | * choose a different item in that bucket. | |
626 | * this will involve more devices in data | |
627 | * movement and tend to distribute the load. | |
628 | */ | |
41586081 | 629 | r = rep + parent_r; |
9a3b490a ID |
630 | |
631 | /* be careful */ | |
632 | if (in->alg == CRUSH_BUCKET_UNIFORM && | |
633 | in->size % numrep == 0) | |
634 | /* r'=r+(n+1)*f_total */ | |
635 | r += (numrep+1) * ftotal; | |
636 | else | |
637 | /* r' = r + n*f_total */ | |
638 | r += numrep * ftotal; | |
639 | ||
640 | /* bucket choose */ | |
641 | if (in->size == 0) { | |
642 | dprintk(" empty bucket\n"); | |
643 | break; | |
644 | } | |
645 | ||
646 | item = crush_bucket_choose(in, x, r); | |
647 | if (item >= map->max_devices) { | |
648 | dprintk(" bad item %d\n", item); | |
649 | out[rep] = CRUSH_ITEM_NONE; | |
650 | if (out2) | |
651 | out2[rep] = CRUSH_ITEM_NONE; | |
652 | left--; | |
653 | break; | |
654 | } | |
655 | ||
656 | /* desired type? */ | |
657 | if (item < 0) | |
658 | itemtype = map->buckets[-1-item]->type; | |
659 | else | |
660 | itemtype = 0; | |
661 | dprintk(" item %d type %d\n", item, itemtype); | |
662 | ||
663 | /* keep going? */ | |
664 | if (itemtype != type) { | |
665 | if (item >= 0 || | |
666 | (-1-item) >= map->max_buckets) { | |
667 | dprintk(" bad item type %d\n", type); | |
668 | out[rep] = CRUSH_ITEM_NONE; | |
669 | if (out2) | |
670 | out2[rep] = | |
671 | CRUSH_ITEM_NONE; | |
672 | left--; | |
673 | break; | |
674 | } | |
675 | in = map->buckets[-1-item]; | |
676 | continue; | |
677 | } | |
678 | ||
679 | /* collision? */ | |
680 | collide = 0; | |
ab4ce2b5 | 681 | for (i = outpos; i < endpos; i++) { |
9a3b490a ID |
682 | if (out[i] == item) { |
683 | collide = 1; | |
684 | break; | |
685 | } | |
686 | } | |
687 | if (collide) | |
688 | break; | |
689 | ||
690 | if (recurse_to_leaf) { | |
691 | if (item < 0) { | |
692 | crush_choose_indep(map, | |
be3226ac ID |
693 | map->buckets[-1-item], |
694 | weight, weight_max, | |
695 | x, 1, numrep, 0, | |
696 | out2, rep, | |
2d8be0bc | 697 | recurse_tries, 0, |
be3226ac | 698 | 0, NULL, r); |
9a3b490a ID |
699 | if (out2[rep] == CRUSH_ITEM_NONE) { |
700 | /* placed nothing; no leaf */ | |
701 | break; | |
702 | } | |
703 | } else { | |
704 | /* we already have a leaf! */ | |
705 | out2[rep] = item; | |
706 | } | |
707 | } | |
708 | ||
709 | /* out? */ | |
710 | if (itemtype == 0 && | |
711 | is_out(map, weight, weight_max, item, x)) | |
712 | break; | |
713 | ||
714 | /* yay! */ | |
715 | out[rep] = item; | |
716 | left--; | |
717 | break; | |
718 | } | |
719 | } | |
720 | } | |
ab4ce2b5 | 721 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
722 | if (out[rep] == CRUSH_ITEM_UNDEF) { |
723 | out[rep] = CRUSH_ITEM_NONE; | |
724 | } | |
725 | if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { | |
726 | out2[rep] = CRUSH_ITEM_NONE; | |
727 | } | |
728 | } | |
729 | } | |
730 | ||
5ecc0a0f SW |
731 | /** |
732 | * crush_do_rule - calculate a mapping with the given input and rule | |
733 | * @map: the crush_map | |
734 | * @ruleno: the rule id | |
735 | * @x: hash input | |
736 | * @result: pointer to result vector | |
737 | * @result_max: maximum result size | |
b3b33b0e ID |
738 | * @weight: weight vector (for map leaves) |
739 | * @weight_max: size of weight vector | |
e8ef19c4 | 740 | * @scratch: scratch vector for private use; must be >= 3 * result_max |
5ecc0a0f | 741 | */ |
8b12d47b | 742 | int crush_do_rule(const struct crush_map *map, |
5ecc0a0f | 743 | int ruleno, int x, int *result, int result_max, |
e8ef19c4 ID |
744 | const __u32 *weight, int weight_max, |
745 | int *scratch) | |
5ecc0a0f SW |
746 | { |
747 | int result_len; | |
e8ef19c4 ID |
748 | int *a = scratch; |
749 | int *b = scratch + result_max; | |
750 | int *c = scratch + result_max*2; | |
5ecc0a0f SW |
751 | int recurse_to_leaf; |
752 | int *w; | |
753 | int wsize = 0; | |
754 | int *o; | |
755 | int osize; | |
756 | int *tmp; | |
757 | struct crush_rule *rule; | |
8b12d47b | 758 | __u32 step; |
5ecc0a0f SW |
759 | int i, j; |
760 | int numrep; | |
45002267 | 761 | int out_size; |
48a163db ID |
762 | /* |
763 | * the original choose_total_tries value was off by one (it | |
764 | * counted "retries" and not "tries"). add one. | |
765 | */ | |
766 | int choose_tries = map->choose_total_tries + 1; | |
f18650ac | 767 | int choose_leaf_tries = 0; |
48a163db ID |
768 | /* |
769 | * the local tries values were counted as "retries", though, | |
770 | * and need no adjustment | |
771 | */ | |
772 | int choose_local_retries = map->choose_local_tries; | |
773 | int choose_local_fallback_retries = map->choose_local_fallback_tries; | |
5ecc0a0f | 774 | |
e2b149cc ID |
775 | int vary_r = map->chooseleaf_vary_r; |
776 | ||
a1f4895b SW |
777 | if ((__u32)ruleno >= map->max_rules) { |
778 | dprintk(" bad ruleno %d\n", ruleno); | |
779 | return 0; | |
780 | } | |
5ecc0a0f SW |
781 | |
782 | rule = map->rules[ruleno]; | |
783 | result_len = 0; | |
784 | w = a; | |
785 | o = b; | |
786 | ||
5ecc0a0f | 787 | for (step = 0; step < rule->len; step++) { |
8f99c85b | 788 | int firstn = 0; |
0668216e SW |
789 | struct crush_rule_step *curstep = &rule->steps[step]; |
790 | ||
0668216e | 791 | switch (curstep->op) { |
5ecc0a0f | 792 | case CRUSH_RULE_TAKE: |
8f529795 ID |
793 | if ((curstep->arg1 >= 0 && |
794 | curstep->arg1 < map->max_devices) || | |
795 | (-1-curstep->arg1 < map->max_buckets && | |
796 | map->buckets[-1-curstep->arg1])) { | |
797 | w[0] = curstep->arg1; | |
798 | wsize = 1; | |
799 | } else { | |
800 | dprintk(" bad take value %d\n", curstep->arg1); | |
801 | } | |
5ecc0a0f SW |
802 | break; |
803 | ||
cc10df4a ID |
804 | case CRUSH_RULE_SET_CHOOSE_TRIES: |
805 | if (curstep->arg1 > 0) | |
806 | choose_tries = curstep->arg1; | |
807 | break; | |
808 | ||
917edad5 | 809 | case CRUSH_RULE_SET_CHOOSELEAF_TRIES: |
be3226ac ID |
810 | if (curstep->arg1 > 0) |
811 | choose_leaf_tries = curstep->arg1; | |
812 | break; | |
813 | ||
f046bf92 | 814 | case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: |
6ed1002f | 815 | if (curstep->arg1 >= 0) |
48a163db | 816 | choose_local_retries = curstep->arg1; |
f046bf92 ID |
817 | break; |
818 | ||
819 | case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: | |
6ed1002f | 820 | if (curstep->arg1 >= 0) |
48a163db | 821 | choose_local_fallback_retries = curstep->arg1; |
f046bf92 | 822 | break; |
d83ed858 ID |
823 | |
824 | case CRUSH_RULE_SET_CHOOSELEAF_VARY_R: | |
825 | if (curstep->arg1 >= 0) | |
826 | vary_r = curstep->arg1; | |
827 | break; | |
f046bf92 | 828 | |
917edad5 | 829 | case CRUSH_RULE_CHOOSELEAF_FIRSTN: |
5ecc0a0f SW |
830 | case CRUSH_RULE_CHOOSE_FIRSTN: |
831 | firstn = 1; | |
0668216e | 832 | /* fall through */ |
917edad5 | 833 | case CRUSH_RULE_CHOOSELEAF_INDEP: |
5ecc0a0f | 834 | case CRUSH_RULE_CHOOSE_INDEP: |
a1f4895b SW |
835 | if (wsize == 0) |
836 | break; | |
5ecc0a0f SW |
837 | |
838 | recurse_to_leaf = | |
0668216e | 839 | curstep->op == |
917edad5 | 840 | CRUSH_RULE_CHOOSELEAF_FIRSTN || |
0668216e | 841 | curstep->op == |
917edad5 | 842 | CRUSH_RULE_CHOOSELEAF_INDEP; |
5ecc0a0f SW |
843 | |
844 | /* reset output */ | |
845 | osize = 0; | |
846 | ||
847 | for (i = 0; i < wsize; i++) { | |
848 | /* | |
849 | * see CRUSH_N, CRUSH_N_MINUS macros. | |
850 | * basically, numrep <= 0 means relative to | |
851 | * the provided result_max | |
852 | */ | |
0668216e | 853 | numrep = curstep->arg1; |
5ecc0a0f SW |
854 | if (numrep <= 0) { |
855 | numrep += result_max; | |
856 | if (numrep <= 0) | |
857 | continue; | |
858 | } | |
859 | j = 0; | |
9a3b490a | 860 | if (firstn) { |
d390bb2a ID |
861 | int recurse_tries; |
862 | if (choose_leaf_tries) | |
863 | recurse_tries = | |
864 | choose_leaf_tries; | |
865 | else if (map->chooseleaf_descend_once) | |
866 | recurse_tries = 1; | |
867 | else | |
868 | recurse_tries = choose_tries; | |
9fe07182 ID |
869 | osize += crush_choose_firstn( |
870 | map, | |
871 | map->buckets[-1-w[i]], | |
872 | weight, weight_max, | |
873 | x, numrep, | |
874 | curstep->arg2, | |
875 | o+osize, j, | |
45002267 | 876 | result_max-osize, |
f18650ac | 877 | choose_tries, |
d390bb2a | 878 | recurse_tries, |
48a163db ID |
879 | choose_local_retries, |
880 | choose_local_fallback_retries, | |
9fe07182 | 881 | recurse_to_leaf, |
e2b149cc ID |
882 | vary_r, |
883 | c+osize, | |
884 | 0); | |
9a3b490a | 885 | } else { |
45002267 ID |
886 | out_size = ((numrep < (result_max-osize)) ? |
887 | numrep : (result_max-osize)); | |
9fe07182 ID |
888 | crush_choose_indep( |
889 | map, | |
890 | map->buckets[-1-w[i]], | |
891 | weight, weight_max, | |
45002267 | 892 | x, out_size, numrep, |
9fe07182 ID |
893 | curstep->arg2, |
894 | o+osize, j, | |
f18650ac | 895 | choose_tries, |
d390bb2a ID |
896 | choose_leaf_tries ? |
897 | choose_leaf_tries : 1, | |
9fe07182 | 898 | recurse_to_leaf, |
41586081 ID |
899 | c+osize, |
900 | 0); | |
45002267 | 901 | osize += out_size; |
9a3b490a | 902 | } |
5ecc0a0f SW |
903 | } |
904 | ||
905 | if (recurse_to_leaf) | |
906 | /* copy final _leaf_ values to output set */ | |
907 | memcpy(o, c, osize*sizeof(*o)); | |
908 | ||
2a4ba74e | 909 | /* swap o and w arrays */ |
5ecc0a0f SW |
910 | tmp = o; |
911 | o = w; | |
912 | w = tmp; | |
913 | wsize = osize; | |
914 | break; | |
915 | ||
916 | ||
917 | case CRUSH_RULE_EMIT: | |
918 | for (i = 0; i < wsize && result_len < result_max; i++) { | |
919 | result[result_len] = w[i]; | |
920 | result_len++; | |
921 | } | |
922 | wsize = 0; | |
923 | break; | |
924 | ||
925 | default: | |
a1f4895b SW |
926 | dprintk(" unknown op %d at step %d\n", |
927 | curstep->op, step); | |
928 | break; | |
5ecc0a0f SW |
929 | } |
930 | } | |
f1932fc1 | 931 | return result_len; |
5ecc0a0f SW |
932 | } |
933 | ||
934 |