]>
Commit | Line | Data |
---|---|---|
b459be73 ID |
1 | /* |
2 | * Ceph - scalable distributed file system | |
3 | * | |
4 | * Copyright (C) 2015 Intel Corporation All Rights Reserved | |
5 | * | |
6 | * This is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU Lesser General Public | |
8 | * License version 2.1, as published by the Free Software | |
9 | * Foundation. See file COPYING. | |
10 | * | |
11 | */ | |
5ecc0a0f SW |
12 | |
13 | #ifdef __KERNEL__ | |
14 | # include <linux/string.h> | |
15 | # include <linux/slab.h> | |
16 | # include <linux/bug.h> | |
17 | # include <linux/kernel.h> | |
b459be73 ID |
18 | # include <linux/crush/crush.h> |
19 | # include <linux/crush/hash.h> | |
5ecc0a0f | 20 | #else |
b459be73 ID |
21 | # include "crush_compat.h" |
22 | # include "crush.h" | |
23 | # include "hash.h" | |
5ecc0a0f | 24 | #endif |
958a2765 | 25 | #include "crush_ln_table.h" |
5ecc0a0f | 26 | |
b459be73 ID |
27 | #define dprintk(args...) /* printf(args) */ |
28 | ||
5ecc0a0f SW |
29 | /* |
30 | * Implement the core CRUSH mapping algorithm. | |
31 | */ | |
32 | ||
33 | /** | |
34 | * crush_find_rule - find a crush_rule id for a given ruleset, type, and size. | |
35 | * @map: the crush_map | |
36 | * @ruleset: the storage ruleset id (user defined) | |
37 | * @type: storage ruleset type (user defined) | |
38 | * @size: output set size | |
39 | */ | |
8b12d47b | 40 | int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size) |
5ecc0a0f | 41 | { |
8b12d47b | 42 | __u32 i; |
5ecc0a0f SW |
43 | |
44 | for (i = 0; i < map->max_rules; i++) { | |
45 | if (map->rules[i] && | |
46 | map->rules[i]->mask.ruleset == ruleset && | |
47 | map->rules[i]->mask.type == type && | |
48 | map->rules[i]->mask.min_size <= size && | |
49 | map->rules[i]->mask.max_size >= size) | |
50 | return i; | |
51 | } | |
52 | return -1; | |
53 | } | |
54 | ||
55 | ||
56 | /* | |
57 | * bucket choose methods | |
58 | * | |
59 | * For each bucket algorithm, we have a "choose" method that, given a | |
60 | * crush input @x and replica position (usually, position in output set) @r, | |
61 | * will produce an item in the bucket. | |
62 | */ | |
63 | ||
64 | /* | |
65 | * Choose based on a random permutation of the bucket. | |
66 | * | |
67 | * We used to use some prime number arithmetic to do this, but it | |
68 | * wasn't very random, and had some other bad behaviors. Instead, we | |
69 | * calculate an actual random permutation of the bucket members. | |
70 | * Since this is expensive, we optimize for the r=0 case, which | |
71 | * captures the vast majority of calls. | |
72 | */ | |
73 | static int bucket_perm_choose(struct crush_bucket *bucket, | |
74 | int x, int r) | |
75 | { | |
95c96174 ED |
76 | unsigned int pr = r % bucket->size; |
77 | unsigned int i, s; | |
5ecc0a0f SW |
78 | |
79 | /* start a new permutation if @x has changed */ | |
8b12d47b | 80 | if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) { |
5ecc0a0f SW |
81 | dprintk("bucket %d new x=%d\n", bucket->id, x); |
82 | bucket->perm_x = x; | |
83 | ||
84 | /* optimize common r=0 case */ | |
85 | if (pr == 0) { | |
fb690390 | 86 | s = crush_hash32_3(bucket->hash, x, bucket->id, 0) % |
5ecc0a0f SW |
87 | bucket->size; |
88 | bucket->perm[0] = s; | |
89 | bucket->perm_n = 0xffff; /* magic value, see below */ | |
90 | goto out; | |
91 | } | |
92 | ||
93 | for (i = 0; i < bucket->size; i++) | |
94 | bucket->perm[i] = i; | |
95 | bucket->perm_n = 0; | |
96 | } else if (bucket->perm_n == 0xffff) { | |
97 | /* clean up after the r=0 case above */ | |
98 | for (i = 1; i < bucket->size; i++) | |
99 | bucket->perm[i] = i; | |
100 | bucket->perm[bucket->perm[0]] = 0; | |
101 | bucket->perm_n = 1; | |
102 | } | |
103 | ||
104 | /* calculate permutation up to pr */ | |
105 | for (i = 0; i < bucket->perm_n; i++) | |
106 | dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]); | |
107 | while (bucket->perm_n <= pr) { | |
95c96174 | 108 | unsigned int p = bucket->perm_n; |
5ecc0a0f SW |
109 | /* no point in swapping the final entry */ |
110 | if (p < bucket->size - 1) { | |
fb690390 | 111 | i = crush_hash32_3(bucket->hash, x, bucket->id, p) % |
5ecc0a0f SW |
112 | (bucket->size - p); |
113 | if (i) { | |
95c96174 | 114 | unsigned int t = bucket->perm[p + i]; |
5ecc0a0f SW |
115 | bucket->perm[p + i] = bucket->perm[p]; |
116 | bucket->perm[p] = t; | |
117 | } | |
118 | dprintk(" perm_choose swap %d with %d\n", p, p+i); | |
119 | } | |
120 | bucket->perm_n++; | |
121 | } | |
122 | for (i = 0; i < bucket->size; i++) | |
123 | dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]); | |
124 | ||
125 | s = bucket->perm[pr]; | |
126 | out: | |
127 | dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id, | |
128 | bucket->size, x, r, pr, s); | |
129 | return bucket->items[s]; | |
130 | } | |
131 | ||
132 | /* uniform */ | |
133 | static int bucket_uniform_choose(struct crush_bucket_uniform *bucket, | |
134 | int x, int r) | |
135 | { | |
136 | return bucket_perm_choose(&bucket->h, x, r); | |
137 | } | |
138 | ||
139 | /* list */ | |
140 | static int bucket_list_choose(struct crush_bucket_list *bucket, | |
141 | int x, int r) | |
142 | { | |
143 | int i; | |
144 | ||
145 | for (i = bucket->h.size-1; i >= 0; i--) { | |
b459be73 | 146 | __u64 w = crush_hash32_4(bucket->h.hash, x, bucket->h.items[i], |
fb690390 | 147 | r, bucket->h.id); |
5ecc0a0f SW |
148 | w &= 0xffff; |
149 | dprintk("list_choose i=%d x=%d r=%d item %d weight %x " | |
150 | "sw %x rand %llx", | |
151 | i, x, r, bucket->h.items[i], bucket->item_weights[i], | |
152 | bucket->sum_weights[i], w); | |
153 | w *= bucket->sum_weights[i]; | |
154 | w = w >> 16; | |
155 | /*dprintk(" scaled %llx\n", w);*/ | |
156 | if (w < bucket->item_weights[i]) | |
157 | return bucket->h.items[i]; | |
158 | } | |
159 | ||
a1f4895b SW |
160 | dprintk("bad list sums for bucket %d\n", bucket->h.id); |
161 | return bucket->h.items[0]; | |
5ecc0a0f SW |
162 | } |
163 | ||
164 | ||
165 | /* (binary) tree */ | |
166 | static int height(int n) | |
167 | { | |
168 | int h = 0; | |
169 | while ((n & 1) == 0) { | |
170 | h++; | |
171 | n = n >> 1; | |
172 | } | |
173 | return h; | |
174 | } | |
175 | ||
176 | static int left(int x) | |
177 | { | |
178 | int h = height(x); | |
179 | return x - (1 << (h-1)); | |
180 | } | |
181 | ||
182 | static int right(int x) | |
183 | { | |
184 | int h = height(x); | |
185 | return x + (1 << (h-1)); | |
186 | } | |
187 | ||
188 | static int terminal(int x) | |
189 | { | |
190 | return x & 1; | |
191 | } | |
192 | ||
193 | static int bucket_tree_choose(struct crush_bucket_tree *bucket, | |
194 | int x, int r) | |
195 | { | |
8f99c85b | 196 | int n; |
5ecc0a0f SW |
197 | __u32 w; |
198 | __u64 t; | |
199 | ||
200 | /* start at root */ | |
201 | n = bucket->num_nodes >> 1; | |
202 | ||
203 | while (!terminal(n)) { | |
8f99c85b | 204 | int l; |
5ecc0a0f SW |
205 | /* pick point in [0, w) */ |
206 | w = bucket->node_weights[n]; | |
fb690390 SW |
207 | t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r, |
208 | bucket->h.id) * (__u64)w; | |
5ecc0a0f SW |
209 | t = t >> 32; |
210 | ||
211 | /* descend to the left or right? */ | |
212 | l = left(n); | |
213 | if (t < bucket->node_weights[l]) | |
214 | n = l; | |
215 | else | |
216 | n = right(n); | |
217 | } | |
218 | ||
219 | return bucket->h.items[n >> 1]; | |
220 | } | |
221 | ||
222 | ||
223 | /* straw */ | |
224 | ||
225 | static int bucket_straw_choose(struct crush_bucket_straw *bucket, | |
226 | int x, int r) | |
227 | { | |
8b12d47b | 228 | __u32 i; |
5ecc0a0f SW |
229 | int high = 0; |
230 | __u64 high_draw = 0; | |
231 | __u64 draw; | |
232 | ||
233 | for (i = 0; i < bucket->h.size; i++) { | |
fb690390 | 234 | draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r); |
5ecc0a0f SW |
235 | draw &= 0xffff; |
236 | draw *= bucket->straws[i]; | |
237 | if (i == 0 || draw > high_draw) { | |
238 | high = i; | |
239 | high_draw = draw; | |
240 | } | |
241 | } | |
242 | return bucket->h.items[high]; | |
243 | } | |
244 | ||
b459be73 ID |
245 | /* compute 2^44*log2(input+1) */ |
246 | static __u64 crush_ln(unsigned int xin) | |
958a2765 | 247 | { |
b459be73 ID |
248 | unsigned int x = xin, x1; |
249 | int iexpon, index1, index2; | |
250 | __u64 RH, LH, LL, xl64, result; | |
958a2765 | 251 | |
b459be73 | 252 | x++; |
958a2765 | 253 | |
b459be73 ID |
254 | /* normalize input */ |
255 | iexpon = 15; | |
256 | while (!(x & 0x18000)) { | |
257 | x <<= 1; | |
258 | iexpon--; | |
259 | } | |
958a2765 | 260 | |
b459be73 ID |
261 | index1 = (x >> 8) << 1; |
262 | /* RH ~ 2^56/index1 */ | |
263 | RH = __RH_LH_tbl[index1 - 256]; | |
264 | /* LH ~ 2^48 * log2(index1/256) */ | |
265 | LH = __RH_LH_tbl[index1 + 1 - 256]; | |
958a2765 | 266 | |
b459be73 ID |
267 | /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */ |
268 | xl64 = (__s64)x * RH; | |
269 | xl64 >>= 48; | |
270 | x1 = xl64; | |
958a2765 | 271 | |
b459be73 ID |
272 | result = iexpon; |
273 | result <<= (12 + 32); | |
958a2765 | 274 | |
b459be73 ID |
275 | index2 = x1 & 0xff; |
276 | /* LL ~ 2^48*log2(1.0+index2/2^15) */ | |
277 | LL = __LL_tbl[index2]; | |
958a2765 | 278 | |
b459be73 | 279 | LH = LH + LL; |
958a2765 | 280 | |
b459be73 ID |
281 | LH >>= (48 - 12 - 32); |
282 | result += LH; | |
958a2765 | 283 | |
b459be73 | 284 | return result; |
958a2765 ID |
285 | } |
286 | ||
287 | ||
288 | /* | |
289 | * straw2 | |
290 | * | |
291 | * for reference, see: | |
292 | * | |
293 | * http://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables | |
294 | * | |
295 | */ | |
296 | ||
297 | static int bucket_straw2_choose(struct crush_bucket_straw2 *bucket, | |
298 | int x, int r) | |
299 | { | |
b459be73 ID |
300 | unsigned int i, high = 0; |
301 | unsigned int u; | |
302 | unsigned int w; | |
958a2765 ID |
303 | __s64 ln, draw, high_draw = 0; |
304 | ||
305 | for (i = 0; i < bucket->h.size; i++) { | |
306 | w = bucket->item_weights[i]; | |
307 | if (w) { | |
308 | u = crush_hash32_3(bucket->h.hash, x, | |
309 | bucket->h.items[i], r); | |
310 | u &= 0xffff; | |
311 | ||
312 | /* | |
313 | * for some reason slightly less than 0x10000 produces | |
314 | * a slightly more accurate distribution... probably a | |
315 | * rounding effect. | |
316 | * | |
317 | * the natural log lookup table maps [0,0xffff] | |
318 | * (corresponding to real numbers [1/0x10000, 1] to | |
319 | * [0, 0xffffffffffff] (corresponding to real numbers | |
320 | * [-11.090355,0]). | |
321 | */ | |
322 | ln = crush_ln(u) - 0x1000000000000ll; | |
323 | ||
324 | /* | |
325 | * divide by 16.16 fixed-point weight. note | |
326 | * that the ln value is negative, so a larger | |
327 | * weight means a larger (less negative) value | |
328 | * for draw. | |
329 | */ | |
330 | draw = div64_s64(ln, w); | |
331 | } else { | |
332 | draw = S64_MIN; | |
333 | } | |
334 | ||
335 | if (i == 0 || draw > high_draw) { | |
336 | high = i; | |
337 | high_draw = draw; | |
338 | } | |
339 | } | |
340 | return bucket->h.items[high]; | |
341 | } | |
342 | ||
343 | ||
5ecc0a0f SW |
344 | static int crush_bucket_choose(struct crush_bucket *in, int x, int r) |
345 | { | |
a1a31e73 | 346 | dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r); |
a1f4895b | 347 | BUG_ON(in->size == 0); |
5ecc0a0f SW |
348 | switch (in->alg) { |
349 | case CRUSH_BUCKET_UNIFORM: | |
350 | return bucket_uniform_choose((struct crush_bucket_uniform *)in, | |
351 | x, r); | |
352 | case CRUSH_BUCKET_LIST: | |
353 | return bucket_list_choose((struct crush_bucket_list *)in, | |
354 | x, r); | |
355 | case CRUSH_BUCKET_TREE: | |
356 | return bucket_tree_choose((struct crush_bucket_tree *)in, | |
357 | x, r); | |
358 | case CRUSH_BUCKET_STRAW: | |
359 | return bucket_straw_choose((struct crush_bucket_straw *)in, | |
360 | x, r); | |
958a2765 ID |
361 | case CRUSH_BUCKET_STRAW2: |
362 | return bucket_straw2_choose((struct crush_bucket_straw2 *)in, | |
363 | x, r); | |
5ecc0a0f | 364 | default: |
a1f4895b | 365 | dprintk("unknown bucket %d alg %d\n", in->id, in->alg); |
50b885b9 | 366 | return in->items[0]; |
5ecc0a0f SW |
367 | } |
368 | } | |
369 | ||
958a2765 | 370 | |
5ecc0a0f SW |
371 | /* |
372 | * true if device is marked "out" (failed, fully offloaded) | |
373 | * of the cluster | |
374 | */ | |
b3b33b0e ID |
375 | static int is_out(const struct crush_map *map, |
376 | const __u32 *weight, int weight_max, | |
377 | int item, int x) | |
5ecc0a0f | 378 | { |
b3b33b0e ID |
379 | if (item >= weight_max) |
380 | return 1; | |
153a1093 | 381 | if (weight[item] >= 0x10000) |
5ecc0a0f SW |
382 | return 0; |
383 | if (weight[item] == 0) | |
384 | return 1; | |
fb690390 SW |
385 | if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff) |
386 | < weight[item]) | |
5ecc0a0f SW |
387 | return 0; |
388 | return 1; | |
389 | } | |
390 | ||
391 | /** | |
9fe07182 | 392 | * crush_choose_firstn - choose numrep distinct items of given type |
5ecc0a0f SW |
393 | * @map: the crush_map |
394 | * @bucket: the bucket we are choose an item from | |
395 | * @x: crush input value | |
396 | * @numrep: the number of items to choose | |
397 | * @type: the type of item to choose | |
398 | * @out: pointer to output vector | |
399 | * @outpos: our position in that vector | |
45002267 | 400 | * @out_size: size of the out vector |
0e32d712 ID |
401 | * @tries: number of attempts to make |
402 | * @recurse_tries: number of attempts to have recursive chooseleaf make | |
48a163db ID |
403 | * @local_retries: localized retries |
404 | * @local_fallback_retries: localized fallback retries | |
0e32d712 | 405 | * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) |
e2b149cc | 406 | * @vary_r: pass r to recursive calls |
5ecc0a0f | 407 | * @out2: second output vector for leaf items (if @recurse_to_leaf) |
e2b149cc | 408 | * @parent_r: r value passed from the parent |
5ecc0a0f | 409 | */ |
9fe07182 ID |
410 | static int crush_choose_firstn(const struct crush_map *map, |
411 | struct crush_bucket *bucket, | |
412 | const __u32 *weight, int weight_max, | |
413 | int x, int numrep, int type, | |
414 | int *out, int outpos, | |
45002267 | 415 | int out_size, |
2d8be0bc ID |
416 | unsigned int tries, |
417 | unsigned int recurse_tries, | |
48a163db ID |
418 | unsigned int local_retries, |
419 | unsigned int local_fallback_retries, | |
9fe07182 | 420 | int recurse_to_leaf, |
e2b149cc ID |
421 | unsigned int vary_r, |
422 | int *out2, | |
423 | int parent_r) | |
5ecc0a0f SW |
424 | { |
425 | int rep; | |
8b12d47b | 426 | unsigned int ftotal, flocal; |
5ecc0a0f SW |
427 | int retry_descent, retry_bucket, skip_rep; |
428 | struct crush_bucket *in = bucket; | |
429 | int r; | |
430 | int i; | |
b28813a6 | 431 | int item = 0; |
5ecc0a0f SW |
432 | int itemtype; |
433 | int collide, reject; | |
45002267 | 434 | int count = out_size; |
a1a31e73 | 435 | |
e2b149cc ID |
436 | 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", |
437 | recurse_to_leaf ? "_LEAF" : "", | |
438 | bucket->id, x, outpos, numrep, | |
439 | tries, recurse_tries, local_retries, local_fallback_retries, | |
440 | parent_r); | |
5ecc0a0f | 441 | |
45002267 | 442 | for (rep = outpos; rep < numrep && count > 0 ; rep++) { |
5ecc0a0f SW |
443 | /* keep trying until we get a non-out, non-colliding item */ |
444 | ftotal = 0; | |
445 | skip_rep = 0; | |
446 | do { | |
447 | retry_descent = 0; | |
448 | in = bucket; /* initial bucket */ | |
449 | ||
450 | /* choose through intervening buckets */ | |
451 | flocal = 0; | |
452 | do { | |
b28813a6 | 453 | collide = 0; |
5ecc0a0f | 454 | retry_bucket = 0; |
e2b149cc | 455 | r = rep + parent_r; |
9fe07182 ID |
456 | /* r' = r + f_total */ |
457 | r += ftotal; | |
5ecc0a0f SW |
458 | |
459 | /* bucket choose */ | |
b28813a6 SW |
460 | if (in->size == 0) { |
461 | reject = 1; | |
462 | goto reject; | |
463 | } | |
48a163db | 464 | if (local_fallback_retries > 0 && |
546f04ef | 465 | flocal >= (in->size>>1) && |
48a163db | 466 | flocal > local_fallback_retries) |
5ecc0a0f SW |
467 | item = bucket_perm_choose(in, x, r); |
468 | else | |
469 | item = crush_bucket_choose(in, x, r); | |
a1f4895b SW |
470 | if (item >= map->max_devices) { |
471 | dprintk(" bad item %d\n", item); | |
472 | skip_rep = 1; | |
473 | break; | |
474 | } | |
5ecc0a0f SW |
475 | |
476 | /* desired type? */ | |
477 | if (item < 0) | |
478 | itemtype = map->buckets[-1-item]->type; | |
479 | else | |
480 | itemtype = 0; | |
481 | dprintk(" item %d type %d\n", item, itemtype); | |
482 | ||
483 | /* keep going? */ | |
484 | if (itemtype != type) { | |
a1f4895b SW |
485 | if (item >= 0 || |
486 | (-1-item) >= map->max_buckets) { | |
487 | dprintk(" bad item type %d\n", type); | |
488 | skip_rep = 1; | |
489 | break; | |
490 | } | |
5ecc0a0f | 491 | in = map->buckets[-1-item]; |
55bda7aa | 492 | retry_bucket = 1; |
5ecc0a0f SW |
493 | continue; |
494 | } | |
495 | ||
496 | /* collision? */ | |
5ecc0a0f SW |
497 | for (i = 0; i < outpos; i++) { |
498 | if (out[i] == item) { | |
499 | collide = 1; | |
500 | break; | |
501 | } | |
502 | } | |
503 | ||
a1a31e73 | 504 | reject = 0; |
7d7c1f61 | 505 | if (!collide && recurse_to_leaf) { |
a1a31e73 | 506 | if (item < 0) { |
e2b149cc ID |
507 | int sub_r; |
508 | if (vary_r) | |
509 | sub_r = r >> (vary_r-1); | |
510 | else | |
511 | sub_r = 0; | |
9fe07182 | 512 | if (crush_choose_firstn(map, |
a1a31e73 | 513 | map->buckets[-1-item], |
b3b33b0e | 514 | weight, weight_max, |
a1a31e73 | 515 | x, outpos+1, 0, |
45002267 | 516 | out2, outpos, count, |
2d8be0bc | 517 | recurse_tries, 0, |
48a163db ID |
518 | local_retries, |
519 | local_fallback_retries, | |
9fe07182 | 520 | 0, |
e2b149cc ID |
521 | vary_r, |
522 | NULL, | |
523 | sub_r) <= outpos) | |
a1a31e73 SW |
524 | /* didn't get leaf */ |
525 | reject = 1; | |
526 | } else { | |
527 | /* we already have a leaf! */ | |
528 | out2[outpos] = item; | |
529 | } | |
530 | } | |
531 | ||
532 | if (!reject) { | |
5ecc0a0f SW |
533 | /* out? */ |
534 | if (itemtype == 0) | |
535 | reject = is_out(map, weight, | |
b3b33b0e | 536 | weight_max, |
5ecc0a0f SW |
537 | item, x); |
538 | else | |
539 | reject = 0; | |
540 | } | |
541 | ||
b28813a6 | 542 | reject: |
5ecc0a0f SW |
543 | if (reject || collide) { |
544 | ftotal++; | |
545 | flocal++; | |
546 | ||
48a163db | 547 | if (collide && flocal <= local_retries) |
5ecc0a0f SW |
548 | /* retry locally a few times */ |
549 | retry_bucket = 1; | |
48a163db ID |
550 | else if (local_fallback_retries > 0 && |
551 | flocal <= in->size + local_fallback_retries) | |
5ecc0a0f SW |
552 | /* exhaustive bucket search */ |
553 | retry_bucket = 1; | |
48a163db | 554 | else if (ftotal < tries) |
5ecc0a0f SW |
555 | /* then retry descent */ |
556 | retry_descent = 1; | |
557 | else | |
558 | /* else give up */ | |
559 | skip_rep = 1; | |
560 | dprintk(" reject %d collide %d " | |
8b12d47b | 561 | "ftotal %u flocal %u\n", |
5ecc0a0f SW |
562 | reject, collide, ftotal, |
563 | flocal); | |
564 | } | |
565 | } while (retry_bucket); | |
566 | } while (retry_descent); | |
567 | ||
568 | if (skip_rep) { | |
9fe07182 ID |
569 | dprintk("skip rep\n"); |
570 | continue; | |
5ecc0a0f SW |
571 | } |
572 | ||
a1a31e73 | 573 | dprintk("CHOOSE got %d\n", item); |
5ecc0a0f SW |
574 | out[outpos] = item; |
575 | outpos++; | |
45002267 | 576 | count--; |
b459be73 ID |
577 | #ifndef __KERNEL__ |
578 | if (map->choose_tries && ftotal <= map->choose_total_tries) | |
579 | map->choose_tries[ftotal]++; | |
580 | #endif | |
5ecc0a0f SW |
581 | } |
582 | ||
a1a31e73 | 583 | dprintk("CHOOSE returns %d\n", outpos); |
5ecc0a0f SW |
584 | return outpos; |
585 | } | |
586 | ||
587 | ||
9a3b490a | 588 | /** |
9fe07182 | 589 | * crush_choose_indep: alternative breadth-first positionally stable mapping |
9a3b490a ID |
590 | * |
591 | */ | |
592 | static void crush_choose_indep(const struct crush_map *map, | |
be3226ac ID |
593 | struct crush_bucket *bucket, |
594 | const __u32 *weight, int weight_max, | |
ab4ce2b5 | 595 | int x, int left, int numrep, int type, |
be3226ac | 596 | int *out, int outpos, |
2d8be0bc ID |
597 | unsigned int tries, |
598 | unsigned int recurse_tries, | |
be3226ac | 599 | int recurse_to_leaf, |
41586081 ID |
600 | int *out2, |
601 | int parent_r) | |
9a3b490a ID |
602 | { |
603 | struct crush_bucket *in = bucket; | |
ab4ce2b5 | 604 | int endpos = outpos + left; |
9a3b490a ID |
605 | int rep; |
606 | unsigned int ftotal; | |
607 | int r; | |
608 | int i; | |
609 | int item = 0; | |
610 | int itemtype; | |
611 | int collide; | |
612 | ||
613 | dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "", | |
614 | bucket->id, x, outpos, numrep); | |
615 | ||
616 | /* initially my result is undefined */ | |
ab4ce2b5 | 617 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
618 | out[rep] = CRUSH_ITEM_UNDEF; |
619 | if (out2) | |
620 | out2[rep] = CRUSH_ITEM_UNDEF; | |
621 | } | |
622 | ||
2d8be0bc | 623 | for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { |
b459be73 ID |
624 | #ifdef DEBUG_INDEP |
625 | if (out2 && ftotal) { | |
626 | dprintk("%u %d a: ", ftotal, left); | |
627 | for (rep = outpos; rep < endpos; rep++) { | |
628 | dprintk(" %d", out[rep]); | |
629 | } | |
630 | dprintk("\n"); | |
631 | dprintk("%u %d b: ", ftotal, left); | |
632 | for (rep = outpos; rep < endpos; rep++) { | |
633 | dprintk(" %d", out2[rep]); | |
634 | } | |
635 | dprintk("\n"); | |
636 | } | |
637 | #endif | |
ab4ce2b5 | 638 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
639 | if (out[rep] != CRUSH_ITEM_UNDEF) |
640 | continue; | |
641 | ||
642 | in = bucket; /* initial bucket */ | |
643 | ||
644 | /* choose through intervening buckets */ | |
645 | for (;;) { | |
3102b0a5 ID |
646 | /* note: we base the choice on the position |
647 | * even in the nested call. that means that | |
648 | * if the first layer chooses the same bucket | |
649 | * in a different position, we will tend to | |
650 | * choose a different item in that bucket. | |
651 | * this will involve more devices in data | |
652 | * movement and tend to distribute the load. | |
653 | */ | |
41586081 | 654 | r = rep + parent_r; |
9a3b490a ID |
655 | |
656 | /* be careful */ | |
657 | if (in->alg == CRUSH_BUCKET_UNIFORM && | |
658 | in->size % numrep == 0) | |
659 | /* r'=r+(n+1)*f_total */ | |
660 | r += (numrep+1) * ftotal; | |
661 | else | |
662 | /* r' = r + n*f_total */ | |
663 | r += numrep * ftotal; | |
664 | ||
665 | /* bucket choose */ | |
666 | if (in->size == 0) { | |
667 | dprintk(" empty bucket\n"); | |
668 | break; | |
669 | } | |
670 | ||
671 | item = crush_bucket_choose(in, x, r); | |
672 | if (item >= map->max_devices) { | |
673 | dprintk(" bad item %d\n", item); | |
674 | out[rep] = CRUSH_ITEM_NONE; | |
675 | if (out2) | |
676 | out2[rep] = CRUSH_ITEM_NONE; | |
677 | left--; | |
678 | break; | |
679 | } | |
680 | ||
681 | /* desired type? */ | |
682 | if (item < 0) | |
683 | itemtype = map->buckets[-1-item]->type; | |
684 | else | |
685 | itemtype = 0; | |
686 | dprintk(" item %d type %d\n", item, itemtype); | |
687 | ||
688 | /* keep going? */ | |
689 | if (itemtype != type) { | |
690 | if (item >= 0 || | |
691 | (-1-item) >= map->max_buckets) { | |
692 | dprintk(" bad item type %d\n", type); | |
693 | out[rep] = CRUSH_ITEM_NONE; | |
694 | if (out2) | |
695 | out2[rep] = | |
696 | CRUSH_ITEM_NONE; | |
697 | left--; | |
698 | break; | |
699 | } | |
700 | in = map->buckets[-1-item]; | |
701 | continue; | |
702 | } | |
703 | ||
704 | /* collision? */ | |
705 | collide = 0; | |
ab4ce2b5 | 706 | for (i = outpos; i < endpos; i++) { |
9a3b490a ID |
707 | if (out[i] == item) { |
708 | collide = 1; | |
709 | break; | |
710 | } | |
711 | } | |
712 | if (collide) | |
713 | break; | |
714 | ||
715 | if (recurse_to_leaf) { | |
716 | if (item < 0) { | |
717 | crush_choose_indep(map, | |
be3226ac ID |
718 | map->buckets[-1-item], |
719 | weight, weight_max, | |
720 | x, 1, numrep, 0, | |
721 | out2, rep, | |
2d8be0bc | 722 | recurse_tries, 0, |
be3226ac | 723 | 0, NULL, r); |
9a3b490a ID |
724 | if (out2[rep] == CRUSH_ITEM_NONE) { |
725 | /* placed nothing; no leaf */ | |
726 | break; | |
727 | } | |
728 | } else { | |
729 | /* we already have a leaf! */ | |
730 | out2[rep] = item; | |
731 | } | |
732 | } | |
733 | ||
734 | /* out? */ | |
735 | if (itemtype == 0 && | |
736 | is_out(map, weight, weight_max, item, x)) | |
737 | break; | |
738 | ||
739 | /* yay! */ | |
740 | out[rep] = item; | |
741 | left--; | |
742 | break; | |
743 | } | |
744 | } | |
745 | } | |
ab4ce2b5 | 746 | for (rep = outpos; rep < endpos; rep++) { |
9a3b490a ID |
747 | if (out[rep] == CRUSH_ITEM_UNDEF) { |
748 | out[rep] = CRUSH_ITEM_NONE; | |
749 | } | |
750 | if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { | |
751 | out2[rep] = CRUSH_ITEM_NONE; | |
752 | } | |
753 | } | |
b459be73 ID |
754 | #ifndef __KERNEL__ |
755 | if (map->choose_tries && ftotal <= map->choose_total_tries) | |
756 | map->choose_tries[ftotal]++; | |
757 | #endif | |
758 | #ifdef DEBUG_INDEP | |
759 | if (out2) { | |
760 | dprintk("%u %d a: ", ftotal, left); | |
761 | for (rep = outpos; rep < endpos; rep++) { | |
762 | dprintk(" %d", out[rep]); | |
763 | } | |
764 | dprintk("\n"); | |
765 | dprintk("%u %d b: ", ftotal, left); | |
766 | for (rep = outpos; rep < endpos; rep++) { | |
767 | dprintk(" %d", out2[rep]); | |
768 | } | |
769 | dprintk("\n"); | |
770 | } | |
771 | #endif | |
9a3b490a ID |
772 | } |
773 | ||
5ecc0a0f SW |
774 | /** |
775 | * crush_do_rule - calculate a mapping with the given input and rule | |
776 | * @map: the crush_map | |
777 | * @ruleno: the rule id | |
778 | * @x: hash input | |
779 | * @result: pointer to result vector | |
780 | * @result_max: maximum result size | |
b3b33b0e ID |
781 | * @weight: weight vector (for map leaves) |
782 | * @weight_max: size of weight vector | |
e8ef19c4 | 783 | * @scratch: scratch vector for private use; must be >= 3 * result_max |
5ecc0a0f | 784 | */ |
8b12d47b | 785 | int crush_do_rule(const struct crush_map *map, |
5ecc0a0f | 786 | int ruleno, int x, int *result, int result_max, |
e8ef19c4 ID |
787 | const __u32 *weight, int weight_max, |
788 | int *scratch) | |
5ecc0a0f SW |
789 | { |
790 | int result_len; | |
e8ef19c4 ID |
791 | int *a = scratch; |
792 | int *b = scratch + result_max; | |
793 | int *c = scratch + result_max*2; | |
5ecc0a0f SW |
794 | int recurse_to_leaf; |
795 | int *w; | |
796 | int wsize = 0; | |
797 | int *o; | |
798 | int osize; | |
799 | int *tmp; | |
800 | struct crush_rule *rule; | |
8b12d47b | 801 | __u32 step; |
5ecc0a0f SW |
802 | int i, j; |
803 | int numrep; | |
45002267 | 804 | int out_size; |
48a163db ID |
805 | /* |
806 | * the original choose_total_tries value was off by one (it | |
807 | * counted "retries" and not "tries"). add one. | |
808 | */ | |
809 | int choose_tries = map->choose_total_tries + 1; | |
f18650ac | 810 | int choose_leaf_tries = 0; |
48a163db ID |
811 | /* |
812 | * the local tries values were counted as "retries", though, | |
813 | * and need no adjustment | |
814 | */ | |
815 | int choose_local_retries = map->choose_local_tries; | |
816 | int choose_local_fallback_retries = map->choose_local_fallback_tries; | |
5ecc0a0f | 817 | |
e2b149cc ID |
818 | int vary_r = map->chooseleaf_vary_r; |
819 | ||
a1f4895b SW |
820 | if ((__u32)ruleno >= map->max_rules) { |
821 | dprintk(" bad ruleno %d\n", ruleno); | |
822 | return 0; | |
823 | } | |
5ecc0a0f SW |
824 | |
825 | rule = map->rules[ruleno]; | |
826 | result_len = 0; | |
827 | w = a; | |
828 | o = b; | |
829 | ||
5ecc0a0f | 830 | for (step = 0; step < rule->len; step++) { |
8f99c85b | 831 | int firstn = 0; |
0668216e SW |
832 | struct crush_rule_step *curstep = &rule->steps[step]; |
833 | ||
0668216e | 834 | switch (curstep->op) { |
5ecc0a0f | 835 | case CRUSH_RULE_TAKE: |
8f529795 ID |
836 | if ((curstep->arg1 >= 0 && |
837 | curstep->arg1 < map->max_devices) || | |
838 | (-1-curstep->arg1 < map->max_buckets && | |
839 | map->buckets[-1-curstep->arg1])) { | |
840 | w[0] = curstep->arg1; | |
841 | wsize = 1; | |
842 | } else { | |
843 | dprintk(" bad take value %d\n", curstep->arg1); | |
844 | } | |
5ecc0a0f SW |
845 | break; |
846 | ||
cc10df4a ID |
847 | case CRUSH_RULE_SET_CHOOSE_TRIES: |
848 | if (curstep->arg1 > 0) | |
849 | choose_tries = curstep->arg1; | |
850 | break; | |
851 | ||
917edad5 | 852 | case CRUSH_RULE_SET_CHOOSELEAF_TRIES: |
be3226ac ID |
853 | if (curstep->arg1 > 0) |
854 | choose_leaf_tries = curstep->arg1; | |
855 | break; | |
856 | ||
f046bf92 | 857 | case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: |
6ed1002f | 858 | if (curstep->arg1 >= 0) |
48a163db | 859 | choose_local_retries = curstep->arg1; |
f046bf92 ID |
860 | break; |
861 | ||
862 | case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: | |
6ed1002f | 863 | if (curstep->arg1 >= 0) |
48a163db | 864 | choose_local_fallback_retries = curstep->arg1; |
f046bf92 | 865 | break; |
d83ed858 ID |
866 | |
867 | case CRUSH_RULE_SET_CHOOSELEAF_VARY_R: | |
868 | if (curstep->arg1 >= 0) | |
869 | vary_r = curstep->arg1; | |
870 | break; | |
f046bf92 | 871 | |
917edad5 | 872 | case CRUSH_RULE_CHOOSELEAF_FIRSTN: |
5ecc0a0f SW |
873 | case CRUSH_RULE_CHOOSE_FIRSTN: |
874 | firstn = 1; | |
0668216e | 875 | /* fall through */ |
917edad5 | 876 | case CRUSH_RULE_CHOOSELEAF_INDEP: |
5ecc0a0f | 877 | case CRUSH_RULE_CHOOSE_INDEP: |
a1f4895b SW |
878 | if (wsize == 0) |
879 | break; | |
5ecc0a0f SW |
880 | |
881 | recurse_to_leaf = | |
0668216e | 882 | curstep->op == |
917edad5 | 883 | CRUSH_RULE_CHOOSELEAF_FIRSTN || |
0668216e | 884 | curstep->op == |
917edad5 | 885 | CRUSH_RULE_CHOOSELEAF_INDEP; |
5ecc0a0f SW |
886 | |
887 | /* reset output */ | |
888 | osize = 0; | |
889 | ||
890 | for (i = 0; i < wsize; i++) { | |
f224a691 | 891 | int bno; |
5ecc0a0f SW |
892 | /* |
893 | * see CRUSH_N, CRUSH_N_MINUS macros. | |
894 | * basically, numrep <= 0 means relative to | |
895 | * the provided result_max | |
896 | */ | |
0668216e | 897 | numrep = curstep->arg1; |
5ecc0a0f SW |
898 | if (numrep <= 0) { |
899 | numrep += result_max; | |
900 | if (numrep <= 0) | |
901 | continue; | |
902 | } | |
903 | j = 0; | |
f224a691 ID |
904 | /* make sure bucket id is valid */ |
905 | bno = -1 - w[i]; | |
906 | if (bno < 0 || bno >= map->max_buckets) { | |
907 | /* w[i] is probably CRUSH_ITEM_NONE */ | |
908 | dprintk(" bad w[i] %d\n", w[i]); | |
909 | continue; | |
910 | } | |
9a3b490a | 911 | if (firstn) { |
d390bb2a ID |
912 | int recurse_tries; |
913 | if (choose_leaf_tries) | |
914 | recurse_tries = | |
915 | choose_leaf_tries; | |
916 | else if (map->chooseleaf_descend_once) | |
917 | recurse_tries = 1; | |
918 | else | |
919 | recurse_tries = choose_tries; | |
9fe07182 ID |
920 | osize += crush_choose_firstn( |
921 | map, | |
f224a691 | 922 | map->buckets[bno], |
9fe07182 ID |
923 | weight, weight_max, |
924 | x, numrep, | |
925 | curstep->arg2, | |
926 | o+osize, j, | |
45002267 | 927 | result_max-osize, |
f18650ac | 928 | choose_tries, |
d390bb2a | 929 | recurse_tries, |
48a163db ID |
930 | choose_local_retries, |
931 | choose_local_fallback_retries, | |
9fe07182 | 932 | recurse_to_leaf, |
e2b149cc ID |
933 | vary_r, |
934 | c+osize, | |
935 | 0); | |
9a3b490a | 936 | } else { |
45002267 | 937 | out_size = ((numrep < (result_max-osize)) ? |
b459be73 | 938 | numrep : (result_max-osize)); |
9fe07182 ID |
939 | crush_choose_indep( |
940 | map, | |
f224a691 | 941 | map->buckets[bno], |
9fe07182 | 942 | weight, weight_max, |
45002267 | 943 | x, out_size, numrep, |
9fe07182 ID |
944 | curstep->arg2, |
945 | o+osize, j, | |
f18650ac | 946 | choose_tries, |
d390bb2a ID |
947 | choose_leaf_tries ? |
948 | choose_leaf_tries : 1, | |
9fe07182 | 949 | recurse_to_leaf, |
41586081 ID |
950 | c+osize, |
951 | 0); | |
45002267 | 952 | osize += out_size; |
9a3b490a | 953 | } |
5ecc0a0f SW |
954 | } |
955 | ||
956 | if (recurse_to_leaf) | |
957 | /* copy final _leaf_ values to output set */ | |
958 | memcpy(o, c, osize*sizeof(*o)); | |
959 | ||
2a4ba74e | 960 | /* swap o and w arrays */ |
5ecc0a0f SW |
961 | tmp = o; |
962 | o = w; | |
963 | w = tmp; | |
964 | wsize = osize; | |
965 | break; | |
966 | ||
967 | ||
968 | case CRUSH_RULE_EMIT: | |
969 | for (i = 0; i < wsize && result_len < result_max; i++) { | |
970 | result[result_len] = w[i]; | |
971 | result_len++; | |
972 | } | |
973 | wsize = 0; | |
974 | break; | |
975 | ||
976 | default: | |
a1f4895b SW |
977 | dprintk(" unknown op %d at step %d\n", |
978 | curstep->op, step); | |
979 | break; | |
5ecc0a0f SW |
980 | } |
981 | } | |
f1932fc1 | 982 | return result_len; |
5ecc0a0f | 983 | } |