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 | */ |
12 | |
13 | #ifdef __KERNEL__ |
14 | # include <linux/string.h> |
15 | # include <linux/slab.h> |
16 | # include <linux/bug.h> |
17 | # include <linux/kernel.h> |
18 | # include <linux/crush/crush.h> |
19 | # include <linux/crush/hash.h> |
20 | # include <linux/crush/mapper.h> |
21 | #else |
22 | # include "crush_compat.h" |
23 | # include "crush.h" |
24 | # include "hash.h" |
25 | # include "mapper.h" |
26 | #endif |
27 | #include "crush_ln_table.h" |
28 | |
29 | #define dprintk(args...) /* printf(args) */ |
30 | |
31 | /* |
32 | * Implement the core CRUSH mapping algorithm. |
33 | */ |
34 | |
35 | /** |
36 | * crush_find_rule - find a crush_rule id for a given ruleset, type, and size. |
37 | * @map: the crush_map |
38 | * @ruleset: the storage ruleset id (user defined) |
39 | * @type: storage ruleset type (user defined) |
40 | * @size: output set size |
41 | */ |
42 | int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size) |
43 | { |
44 | __u32 i; |
45 | |
46 | for (i = 0; i < map->max_rules; i++) { |
47 | if (map->rules[i] && |
48 | map->rules[i]->mask.ruleset == ruleset && |
49 | map->rules[i]->mask.type == type && |
50 | map->rules[i]->mask.min_size <= size && |
51 | map->rules[i]->mask.max_size >= size) |
52 | return i; |
53 | } |
54 | return -1; |
55 | } |
56 | |
57 | /* |
58 | * bucket choose methods |
59 | * |
60 | * For each bucket algorithm, we have a "choose" method that, given a |
61 | * crush input @x and replica position (usually, position in output set) @r, |
62 | * will produce an item in the bucket. |
63 | */ |
64 | |
65 | /* |
66 | * Choose based on a random permutation of the bucket. |
67 | * |
68 | * We used to use some prime number arithmetic to do this, but it |
69 | * wasn't very random, and had some other bad behaviors. Instead, we |
70 | * calculate an actual random permutation of the bucket members. |
71 | * Since this is expensive, we optimize for the r=0 case, which |
72 | * captures the vast majority of calls. |
73 | */ |
74 | static int bucket_perm_choose(const struct crush_bucket *bucket, |
75 | struct crush_work_bucket *work, |
76 | int x, int r) |
77 | { |
78 | unsigned int pr = r % bucket->size; |
79 | unsigned int i, s; |
80 | |
81 | /* start a new permutation if @x has changed */ |
82 | if (work->perm_x != (__u32)x || work->perm_n == 0) { |
83 | dprintk("bucket %d new x=%d\n" , bucket->id, x); |
84 | work->perm_x = x; |
85 | |
86 | /* optimize common r=0 case */ |
87 | if (pr == 0) { |
88 | s = crush_hash32_3(type: bucket->hash, a: x, b: bucket->id, c: 0) % |
89 | bucket->size; |
90 | work->perm[0] = s; |
91 | work->perm_n = 0xffff; /* magic value, see below */ |
92 | goto out; |
93 | } |
94 | |
95 | for (i = 0; i < bucket->size; i++) |
96 | work->perm[i] = i; |
97 | work->perm_n = 0; |
98 | } else if (work->perm_n == 0xffff) { |
99 | /* clean up after the r=0 case above */ |
100 | for (i = 1; i < bucket->size; i++) |
101 | work->perm[i] = i; |
102 | work->perm[work->perm[0]] = 0; |
103 | work->perm_n = 1; |
104 | } |
105 | |
106 | /* calculate permutation up to pr */ |
107 | for (i = 0; i < work->perm_n; i++) |
108 | dprintk(" perm_choose have %d: %d\n" , i, work->perm[i]); |
109 | while (work->perm_n <= pr) { |
110 | unsigned int p = work->perm_n; |
111 | /* no point in swapping the final entry */ |
112 | if (p < bucket->size - 1) { |
113 | i = crush_hash32_3(type: bucket->hash, a: x, b: bucket->id, c: p) % |
114 | (bucket->size - p); |
115 | if (i) { |
116 | unsigned int t = work->perm[p + i]; |
117 | work->perm[p + i] = work->perm[p]; |
118 | work->perm[p] = t; |
119 | } |
120 | dprintk(" perm_choose swap %d with %d\n" , p, p+i); |
121 | } |
122 | work->perm_n++; |
123 | } |
124 | for (i = 0; i < bucket->size; i++) |
125 | dprintk(" perm_choose %d: %d\n" , i, work->perm[i]); |
126 | |
127 | s = work->perm[pr]; |
128 | out: |
129 | dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n" , bucket->id, |
130 | bucket->size, x, r, pr, s); |
131 | return bucket->items[s]; |
132 | } |
133 | |
134 | /* uniform */ |
135 | static int bucket_uniform_choose(const struct crush_bucket_uniform *bucket, |
136 | struct crush_work_bucket *work, int x, int r) |
137 | { |
138 | return bucket_perm_choose(bucket: &bucket->h, work, x, r); |
139 | } |
140 | |
141 | /* list */ |
142 | static int bucket_list_choose(const struct crush_bucket_list *bucket, |
143 | int x, int r) |
144 | { |
145 | int i; |
146 | |
147 | for (i = bucket->h.size-1; i >= 0; i--) { |
148 | __u64 w = crush_hash32_4(type: bucket->h.hash, a: x, b: bucket->h.items[i], |
149 | c: r, d: bucket->h.id); |
150 | w &= 0xffff; |
151 | dprintk("list_choose i=%d x=%d r=%d item %d weight %x " |
152 | "sw %x rand %llx" , |
153 | i, x, r, bucket->h.items[i], bucket->item_weights[i], |
154 | bucket->sum_weights[i], w); |
155 | w *= bucket->sum_weights[i]; |
156 | w = w >> 16; |
157 | /*dprintk(" scaled %llx\n", w);*/ |
158 | if (w < bucket->item_weights[i]) { |
159 | return bucket->h.items[i]; |
160 | } |
161 | } |
162 | |
163 | dprintk("bad list sums for bucket %d\n" , bucket->h.id); |
164 | return bucket->h.items[0]; |
165 | } |
166 | |
167 | |
168 | /* (binary) tree */ |
169 | static int height(int n) |
170 | { |
171 | int h = 0; |
172 | while ((n & 1) == 0) { |
173 | h++; |
174 | n = n >> 1; |
175 | } |
176 | return h; |
177 | } |
178 | |
179 | static int left(int x) |
180 | { |
181 | int h = height(n: x); |
182 | return x - (1 << (h-1)); |
183 | } |
184 | |
185 | static int right(int x) |
186 | { |
187 | int h = height(n: x); |
188 | return x + (1 << (h-1)); |
189 | } |
190 | |
191 | static int terminal(int x) |
192 | { |
193 | return x & 1; |
194 | } |
195 | |
196 | static int bucket_tree_choose(const struct crush_bucket_tree *bucket, |
197 | int x, int r) |
198 | { |
199 | int n; |
200 | __u32 w; |
201 | __u64 t; |
202 | |
203 | /* start at root */ |
204 | n = bucket->num_nodes >> 1; |
205 | |
206 | while (!terminal(x: n)) { |
207 | int l; |
208 | /* pick point in [0, w) */ |
209 | w = bucket->node_weights[n]; |
210 | t = (__u64)crush_hash32_4(type: bucket->h.hash, a: x, b: n, c: r, |
211 | d: bucket->h.id) * (__u64)w; |
212 | t = t >> 32; |
213 | |
214 | /* descend to the left or right? */ |
215 | l = left(x: n); |
216 | if (t < bucket->node_weights[l]) |
217 | n = l; |
218 | else |
219 | n = right(x: n); |
220 | } |
221 | |
222 | return bucket->h.items[n >> 1]; |
223 | } |
224 | |
225 | |
226 | /* straw */ |
227 | |
228 | static int bucket_straw_choose(const struct crush_bucket_straw *bucket, |
229 | int x, int r) |
230 | { |
231 | __u32 i; |
232 | int high = 0; |
233 | __u64 high_draw = 0; |
234 | __u64 draw; |
235 | |
236 | for (i = 0; i < bucket->h.size; i++) { |
237 | draw = crush_hash32_3(type: bucket->h.hash, a: x, b: bucket->h.items[i], c: r); |
238 | draw &= 0xffff; |
239 | draw *= bucket->straws[i]; |
240 | if (i == 0 || draw > high_draw) { |
241 | high = i; |
242 | high_draw = draw; |
243 | } |
244 | } |
245 | return bucket->h.items[high]; |
246 | } |
247 | |
248 | /* compute 2^44*log2(input+1) */ |
249 | static __u64 crush_ln(unsigned int xin) |
250 | { |
251 | unsigned int x = xin; |
252 | int iexpon, index1, index2; |
253 | __u64 RH, LH, LL, xl64, result; |
254 | |
255 | x++; |
256 | |
257 | /* normalize input */ |
258 | iexpon = 15; |
259 | |
260 | /* |
261 | * figure out number of bits we need to shift and |
262 | * do it in one step instead of iteratively |
263 | */ |
264 | if (!(x & 0x18000)) { |
265 | int bits = __builtin_clz(x & 0x1FFFF) - 16; |
266 | x <<= bits; |
267 | iexpon = 15 - bits; |
268 | } |
269 | |
270 | index1 = (x >> 8) << 1; |
271 | /* RH ~ 2^56/index1 */ |
272 | RH = __RH_LH_tbl[index1 - 256]; |
273 | /* LH ~ 2^48 * log2(index1/256) */ |
274 | LH = __RH_LH_tbl[index1 + 1 - 256]; |
275 | |
276 | /* RH*x ~ 2^48 * (2^15 + xf), xf<2^8 */ |
277 | xl64 = (__s64)x * RH; |
278 | xl64 >>= 48; |
279 | |
280 | result = iexpon; |
281 | result <<= (12 + 32); |
282 | |
283 | index2 = xl64 & 0xff; |
284 | /* LL ~ 2^48*log2(1.0+index2/2^15) */ |
285 | LL = __LL_tbl[index2]; |
286 | |
287 | LH = LH + LL; |
288 | |
289 | LH >>= (48 - 12 - 32); |
290 | result += LH; |
291 | |
292 | return result; |
293 | } |
294 | |
295 | |
296 | /* |
297 | * straw2 |
298 | * |
299 | * for reference, see: |
300 | * |
301 | * https://en.wikipedia.org/wiki/Exponential_distribution#Distribution_of_the_minimum_of_exponential_random_variables |
302 | * |
303 | */ |
304 | |
305 | static __u32 *get_choose_arg_weights(const struct crush_bucket_straw2 *bucket, |
306 | const struct crush_choose_arg *arg, |
307 | int position) |
308 | { |
309 | if (!arg || !arg->weight_set) |
310 | return bucket->item_weights; |
311 | |
312 | if (position >= arg->weight_set_size) |
313 | position = arg->weight_set_size - 1; |
314 | return arg->weight_set[position].weights; |
315 | } |
316 | |
317 | static __s32 *get_choose_arg_ids(const struct crush_bucket_straw2 *bucket, |
318 | const struct crush_choose_arg *arg) |
319 | { |
320 | if (!arg || !arg->ids) |
321 | return bucket->h.items; |
322 | |
323 | return arg->ids; |
324 | } |
325 | |
326 | static int bucket_straw2_choose(const struct crush_bucket_straw2 *bucket, |
327 | int x, int r, |
328 | const struct crush_choose_arg *arg, |
329 | int position) |
330 | { |
331 | unsigned int i, high = 0; |
332 | unsigned int u; |
333 | __s64 ln, draw, high_draw = 0; |
334 | __u32 *weights = get_choose_arg_weights(bucket, arg, position); |
335 | __s32 *ids = get_choose_arg_ids(bucket, arg); |
336 | |
337 | for (i = 0; i < bucket->h.size; i++) { |
338 | dprintk("weight 0x%x item %d\n" , weights[i], ids[i]); |
339 | if (weights[i]) { |
340 | u = crush_hash32_3(type: bucket->h.hash, a: x, b: ids[i], c: r); |
341 | u &= 0xffff; |
342 | |
343 | /* |
344 | * for some reason slightly less than 0x10000 produces |
345 | * a slightly more accurate distribution... probably a |
346 | * rounding effect. |
347 | * |
348 | * the natural log lookup table maps [0,0xffff] |
349 | * (corresponding to real numbers [1/0x10000, 1] to |
350 | * [0, 0xffffffffffff] (corresponding to real numbers |
351 | * [-11.090355,0]). |
352 | */ |
353 | ln = crush_ln(xin: u) - 0x1000000000000ll; |
354 | |
355 | /* |
356 | * divide by 16.16 fixed-point weight. note |
357 | * that the ln value is negative, so a larger |
358 | * weight means a larger (less negative) value |
359 | * for draw. |
360 | */ |
361 | draw = div64_s64(dividend: ln, divisor: weights[i]); |
362 | } else { |
363 | draw = S64_MIN; |
364 | } |
365 | |
366 | if (i == 0 || draw > high_draw) { |
367 | high = i; |
368 | high_draw = draw; |
369 | } |
370 | } |
371 | |
372 | return bucket->h.items[high]; |
373 | } |
374 | |
375 | |
376 | static int crush_bucket_choose(const struct crush_bucket *in, |
377 | struct crush_work_bucket *work, |
378 | int x, int r, |
379 | const struct crush_choose_arg *arg, |
380 | int position) |
381 | { |
382 | dprintk(" crush_bucket_choose %d x=%d r=%d\n" , in->id, x, r); |
383 | BUG_ON(in->size == 0); |
384 | switch (in->alg) { |
385 | case CRUSH_BUCKET_UNIFORM: |
386 | return bucket_uniform_choose( |
387 | bucket: (const struct crush_bucket_uniform *)in, |
388 | work, x, r); |
389 | case CRUSH_BUCKET_LIST: |
390 | return bucket_list_choose(bucket: (const struct crush_bucket_list *)in, |
391 | x, r); |
392 | case CRUSH_BUCKET_TREE: |
393 | return bucket_tree_choose(bucket: (const struct crush_bucket_tree *)in, |
394 | x, r); |
395 | case CRUSH_BUCKET_STRAW: |
396 | return bucket_straw_choose( |
397 | bucket: (const struct crush_bucket_straw *)in, |
398 | x, r); |
399 | case CRUSH_BUCKET_STRAW2: |
400 | return bucket_straw2_choose( |
401 | bucket: (const struct crush_bucket_straw2 *)in, |
402 | x, r, arg, position); |
403 | default: |
404 | dprintk("unknown bucket %d alg %d\n" , in->id, in->alg); |
405 | return in->items[0]; |
406 | } |
407 | } |
408 | |
409 | /* |
410 | * true if device is marked "out" (failed, fully offloaded) |
411 | * of the cluster |
412 | */ |
413 | static int is_out(const struct crush_map *map, |
414 | const __u32 *weight, int weight_max, |
415 | int item, int x) |
416 | { |
417 | if (item >= weight_max) |
418 | return 1; |
419 | if (weight[item] >= 0x10000) |
420 | return 0; |
421 | if (weight[item] == 0) |
422 | return 1; |
423 | if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, a: x, b: item) & 0xffff) |
424 | < weight[item]) |
425 | return 0; |
426 | return 1; |
427 | } |
428 | |
429 | /** |
430 | * crush_choose_firstn - choose numrep distinct items of given type |
431 | * @map: the crush_map |
432 | * @bucket: the bucket we are choose an item from |
433 | * @x: crush input value |
434 | * @numrep: the number of items to choose |
435 | * @type: the type of item to choose |
436 | * @out: pointer to output vector |
437 | * @outpos: our position in that vector |
438 | * @out_size: size of the out vector |
439 | * @tries: number of attempts to make |
440 | * @recurse_tries: number of attempts to have recursive chooseleaf make |
441 | * @local_retries: localized retries |
442 | * @local_fallback_retries: localized fallback retries |
443 | * @recurse_to_leaf: true if we want one device under each item of given type (chooseleaf instead of choose) |
444 | * @stable: stable mode starts rep=0 in the recursive call for all replicas |
445 | * @vary_r: pass r to recursive calls |
446 | * @out2: second output vector for leaf items (if @recurse_to_leaf) |
447 | * @parent_r: r value passed from the parent |
448 | */ |
449 | static int crush_choose_firstn(const struct crush_map *map, |
450 | struct crush_work *work, |
451 | const struct crush_bucket *bucket, |
452 | const __u32 *weight, int weight_max, |
453 | int x, int numrep, int type, |
454 | int *out, int outpos, |
455 | int out_size, |
456 | unsigned int tries, |
457 | unsigned int recurse_tries, |
458 | unsigned int local_retries, |
459 | unsigned int local_fallback_retries, |
460 | int recurse_to_leaf, |
461 | unsigned int vary_r, |
462 | unsigned int stable, |
463 | int *out2, |
464 | int parent_r, |
465 | const struct crush_choose_arg *choose_args) |
466 | { |
467 | int rep; |
468 | unsigned int ftotal, flocal; |
469 | int retry_descent, retry_bucket, skip_rep; |
470 | const struct crush_bucket *in = bucket; |
471 | int r; |
472 | int i; |
473 | int item = 0; |
474 | int itemtype; |
475 | int collide, reject; |
476 | int count = out_size; |
477 | |
478 | 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 stable %d\n" , |
479 | recurse_to_leaf ? "_LEAF" : "" , |
480 | bucket->id, x, outpos, numrep, |
481 | tries, recurse_tries, local_retries, local_fallback_retries, |
482 | parent_r, stable); |
483 | |
484 | for (rep = stable ? 0 : outpos; rep < numrep && count > 0 ; rep++) { |
485 | /* keep trying until we get a non-out, non-colliding item */ |
486 | ftotal = 0; |
487 | skip_rep = 0; |
488 | do { |
489 | retry_descent = 0; |
490 | in = bucket; /* initial bucket */ |
491 | |
492 | /* choose through intervening buckets */ |
493 | flocal = 0; |
494 | do { |
495 | collide = 0; |
496 | retry_bucket = 0; |
497 | r = rep + parent_r; |
498 | /* r' = r + f_total */ |
499 | r += ftotal; |
500 | |
501 | /* bucket choose */ |
502 | if (in->size == 0) { |
503 | reject = 1; |
504 | goto reject; |
505 | } |
506 | if (local_fallback_retries > 0 && |
507 | flocal >= (in->size>>1) && |
508 | flocal > local_fallback_retries) |
509 | item = bucket_perm_choose( |
510 | bucket: in, work: work->work[-1-in->id], |
511 | x, r); |
512 | else |
513 | item = crush_bucket_choose( |
514 | in, work: work->work[-1-in->id], |
515 | x, r, |
516 | arg: (choose_args ? |
517 | &choose_args[-1-in->id] : NULL), |
518 | position: outpos); |
519 | if (item >= map->max_devices) { |
520 | dprintk(" bad item %d\n" , item); |
521 | skip_rep = 1; |
522 | break; |
523 | } |
524 | |
525 | /* desired type? */ |
526 | if (item < 0) |
527 | itemtype = map->buckets[-1-item]->type; |
528 | else |
529 | itemtype = 0; |
530 | dprintk(" item %d type %d\n" , item, itemtype); |
531 | |
532 | /* keep going? */ |
533 | if (itemtype != type) { |
534 | if (item >= 0 || |
535 | (-1-item) >= map->max_buckets) { |
536 | dprintk(" bad item type %d\n" , type); |
537 | skip_rep = 1; |
538 | break; |
539 | } |
540 | in = map->buckets[-1-item]; |
541 | retry_bucket = 1; |
542 | continue; |
543 | } |
544 | |
545 | /* collision? */ |
546 | for (i = 0; i < outpos; i++) { |
547 | if (out[i] == item) { |
548 | collide = 1; |
549 | break; |
550 | } |
551 | } |
552 | |
553 | reject = 0; |
554 | if (!collide && recurse_to_leaf) { |
555 | if (item < 0) { |
556 | int sub_r; |
557 | if (vary_r) |
558 | sub_r = r >> (vary_r-1); |
559 | else |
560 | sub_r = 0; |
561 | if (crush_choose_firstn( |
562 | map, |
563 | work, |
564 | bucket: map->buckets[-1-item], |
565 | weight, weight_max, |
566 | x, numrep: stable ? 1 : outpos+1, type: 0, |
567 | out: out2, outpos, out_size: count, |
568 | tries: recurse_tries, recurse_tries: 0, |
569 | local_retries, |
570 | local_fallback_retries, |
571 | recurse_to_leaf: 0, |
572 | vary_r, |
573 | stable, |
574 | NULL, |
575 | parent_r: sub_r, |
576 | choose_args) <= outpos) |
577 | /* didn't get leaf */ |
578 | reject = 1; |
579 | } else { |
580 | /* we already have a leaf! */ |
581 | out2[outpos] = item; |
582 | } |
583 | } |
584 | |
585 | if (!reject && !collide) { |
586 | /* out? */ |
587 | if (itemtype == 0) |
588 | reject = is_out(map, weight, |
589 | weight_max, |
590 | item, x); |
591 | } |
592 | |
593 | reject: |
594 | if (reject || collide) { |
595 | ftotal++; |
596 | flocal++; |
597 | |
598 | if (collide && flocal <= local_retries) |
599 | /* retry locally a few times */ |
600 | retry_bucket = 1; |
601 | else if (local_fallback_retries > 0 && |
602 | flocal <= in->size + local_fallback_retries) |
603 | /* exhaustive bucket search */ |
604 | retry_bucket = 1; |
605 | else if (ftotal < tries) |
606 | /* then retry descent */ |
607 | retry_descent = 1; |
608 | else |
609 | /* else give up */ |
610 | skip_rep = 1; |
611 | dprintk(" reject %d collide %d " |
612 | "ftotal %u flocal %u\n" , |
613 | reject, collide, ftotal, |
614 | flocal); |
615 | } |
616 | } while (retry_bucket); |
617 | } while (retry_descent); |
618 | |
619 | if (skip_rep) { |
620 | dprintk("skip rep\n" ); |
621 | continue; |
622 | } |
623 | |
624 | dprintk("CHOOSE got %d\n" , item); |
625 | out[outpos] = item; |
626 | outpos++; |
627 | count--; |
628 | #ifndef __KERNEL__ |
629 | if (map->choose_tries && ftotal <= map->choose_total_tries) |
630 | map->choose_tries[ftotal]++; |
631 | #endif |
632 | } |
633 | |
634 | dprintk("CHOOSE returns %d\n" , outpos); |
635 | return outpos; |
636 | } |
637 | |
638 | |
639 | /** |
640 | * crush_choose_indep: alternative breadth-first positionally stable mapping |
641 | * |
642 | */ |
643 | static void crush_choose_indep(const struct crush_map *map, |
644 | struct crush_work *work, |
645 | const struct crush_bucket *bucket, |
646 | const __u32 *weight, int weight_max, |
647 | int x, int left, int numrep, int type, |
648 | int *out, int outpos, |
649 | unsigned int tries, |
650 | unsigned int recurse_tries, |
651 | int recurse_to_leaf, |
652 | int *out2, |
653 | int parent_r, |
654 | const struct crush_choose_arg *choose_args) |
655 | { |
656 | const struct crush_bucket *in = bucket; |
657 | int endpos = outpos + left; |
658 | int rep; |
659 | unsigned int ftotal; |
660 | int r; |
661 | int i; |
662 | int item = 0; |
663 | int itemtype; |
664 | int collide; |
665 | |
666 | dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n" , recurse_to_leaf ? "_LEAF" : "" , |
667 | bucket->id, x, outpos, numrep); |
668 | |
669 | /* initially my result is undefined */ |
670 | for (rep = outpos; rep < endpos; rep++) { |
671 | out[rep] = CRUSH_ITEM_UNDEF; |
672 | if (out2) |
673 | out2[rep] = CRUSH_ITEM_UNDEF; |
674 | } |
675 | |
676 | for (ftotal = 0; left > 0 && ftotal < tries; ftotal++) { |
677 | #ifdef DEBUG_INDEP |
678 | if (out2 && ftotal) { |
679 | dprintk("%u %d a: " , ftotal, left); |
680 | for (rep = outpos; rep < endpos; rep++) { |
681 | dprintk(" %d" , out[rep]); |
682 | } |
683 | dprintk("\n" ); |
684 | dprintk("%u %d b: " , ftotal, left); |
685 | for (rep = outpos; rep < endpos; rep++) { |
686 | dprintk(" %d" , out2[rep]); |
687 | } |
688 | dprintk("\n" ); |
689 | } |
690 | #endif |
691 | for (rep = outpos; rep < endpos; rep++) { |
692 | if (out[rep] != CRUSH_ITEM_UNDEF) |
693 | continue; |
694 | |
695 | in = bucket; /* initial bucket */ |
696 | |
697 | /* choose through intervening buckets */ |
698 | for (;;) { |
699 | /* note: we base the choice on the position |
700 | * even in the nested call. that means that |
701 | * if the first layer chooses the same bucket |
702 | * in a different position, we will tend to |
703 | * choose a different item in that bucket. |
704 | * this will involve more devices in data |
705 | * movement and tend to distribute the load. |
706 | */ |
707 | r = rep + parent_r; |
708 | |
709 | /* be careful */ |
710 | if (in->alg == CRUSH_BUCKET_UNIFORM && |
711 | in->size % numrep == 0) |
712 | /* r'=r+(n+1)*f_total */ |
713 | r += (numrep+1) * ftotal; |
714 | else |
715 | /* r' = r + n*f_total */ |
716 | r += numrep * ftotal; |
717 | |
718 | /* bucket choose */ |
719 | if (in->size == 0) { |
720 | dprintk(" empty bucket\n" ); |
721 | break; |
722 | } |
723 | |
724 | item = crush_bucket_choose( |
725 | in, work: work->work[-1-in->id], |
726 | x, r, |
727 | arg: (choose_args ? |
728 | &choose_args[-1-in->id] : NULL), |
729 | position: outpos); |
730 | if (item >= map->max_devices) { |
731 | dprintk(" bad item %d\n" , item); |
732 | out[rep] = CRUSH_ITEM_NONE; |
733 | if (out2) |
734 | out2[rep] = CRUSH_ITEM_NONE; |
735 | left--; |
736 | break; |
737 | } |
738 | |
739 | /* desired type? */ |
740 | if (item < 0) |
741 | itemtype = map->buckets[-1-item]->type; |
742 | else |
743 | itemtype = 0; |
744 | dprintk(" item %d type %d\n" , item, itemtype); |
745 | |
746 | /* keep going? */ |
747 | if (itemtype != type) { |
748 | if (item >= 0 || |
749 | (-1-item) >= map->max_buckets) { |
750 | dprintk(" bad item type %d\n" , type); |
751 | out[rep] = CRUSH_ITEM_NONE; |
752 | if (out2) |
753 | out2[rep] = |
754 | CRUSH_ITEM_NONE; |
755 | left--; |
756 | break; |
757 | } |
758 | in = map->buckets[-1-item]; |
759 | continue; |
760 | } |
761 | |
762 | /* collision? */ |
763 | collide = 0; |
764 | for (i = outpos; i < endpos; i++) { |
765 | if (out[i] == item) { |
766 | collide = 1; |
767 | break; |
768 | } |
769 | } |
770 | if (collide) |
771 | break; |
772 | |
773 | if (recurse_to_leaf) { |
774 | if (item < 0) { |
775 | crush_choose_indep( |
776 | map, |
777 | work, |
778 | bucket: map->buckets[-1-item], |
779 | weight, weight_max, |
780 | x, left: 1, numrep, type: 0, |
781 | out: out2, outpos: rep, |
782 | tries: recurse_tries, recurse_tries: 0, |
783 | recurse_to_leaf: 0, NULL, parent_r: r, |
784 | choose_args); |
785 | if (out2[rep] == CRUSH_ITEM_NONE) { |
786 | /* placed nothing; no leaf */ |
787 | break; |
788 | } |
789 | } else { |
790 | /* we already have a leaf! */ |
791 | out2[rep] = item; |
792 | } |
793 | } |
794 | |
795 | /* out? */ |
796 | if (itemtype == 0 && |
797 | is_out(map, weight, weight_max, item, x)) |
798 | break; |
799 | |
800 | /* yay! */ |
801 | out[rep] = item; |
802 | left--; |
803 | break; |
804 | } |
805 | } |
806 | } |
807 | for (rep = outpos; rep < endpos; rep++) { |
808 | if (out[rep] == CRUSH_ITEM_UNDEF) { |
809 | out[rep] = CRUSH_ITEM_NONE; |
810 | } |
811 | if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) { |
812 | out2[rep] = CRUSH_ITEM_NONE; |
813 | } |
814 | } |
815 | #ifndef __KERNEL__ |
816 | if (map->choose_tries && ftotal <= map->choose_total_tries) |
817 | map->choose_tries[ftotal]++; |
818 | #endif |
819 | #ifdef DEBUG_INDEP |
820 | if (out2) { |
821 | dprintk("%u %d a: " , ftotal, left); |
822 | for (rep = outpos; rep < endpos; rep++) { |
823 | dprintk(" %d" , out[rep]); |
824 | } |
825 | dprintk("\n" ); |
826 | dprintk("%u %d b: " , ftotal, left); |
827 | for (rep = outpos; rep < endpos; rep++) { |
828 | dprintk(" %d" , out2[rep]); |
829 | } |
830 | dprintk("\n" ); |
831 | } |
832 | #endif |
833 | } |
834 | |
835 | |
836 | /* |
837 | * This takes a chunk of memory and sets it up to be a shiny new |
838 | * working area for a CRUSH placement computation. It must be called |
839 | * on any newly allocated memory before passing it in to |
840 | * crush_do_rule. It may be used repeatedly after that, so long as the |
841 | * map has not changed. If the map /has/ changed, you must make sure |
842 | * the working size is no smaller than what was allocated and re-run |
843 | * crush_init_workspace. |
844 | * |
845 | * If you do retain the working space between calls to crush, make it |
846 | * thread-local. |
847 | */ |
848 | void crush_init_workspace(const struct crush_map *map, void *v) |
849 | { |
850 | struct crush_work *w = v; |
851 | __s32 b; |
852 | |
853 | /* |
854 | * We work by moving through the available space and setting |
855 | * values and pointers as we go. |
856 | * |
857 | * It's a bit like Forth's use of the 'allot' word since we |
858 | * set the pointer first and then reserve the space for it to |
859 | * point to by incrementing the point. |
860 | */ |
861 | v += sizeof(struct crush_work); |
862 | w->work = v; |
863 | v += map->max_buckets * sizeof(struct crush_work_bucket *); |
864 | for (b = 0; b < map->max_buckets; ++b) { |
865 | if (!map->buckets[b]) |
866 | continue; |
867 | |
868 | w->work[b] = v; |
869 | switch (map->buckets[b]->alg) { |
870 | default: |
871 | v += sizeof(struct crush_work_bucket); |
872 | break; |
873 | } |
874 | w->work[b]->perm_x = 0; |
875 | w->work[b]->perm_n = 0; |
876 | w->work[b]->perm = v; |
877 | v += map->buckets[b]->size * sizeof(__u32); |
878 | } |
879 | BUG_ON(v - (void *)w != map->working_size); |
880 | } |
881 | |
882 | /** |
883 | * crush_do_rule - calculate a mapping with the given input and rule |
884 | * @map: the crush_map |
885 | * @ruleno: the rule id |
886 | * @x: hash input |
887 | * @result: pointer to result vector |
888 | * @result_max: maximum result size |
889 | * @weight: weight vector (for map leaves) |
890 | * @weight_max: size of weight vector |
891 | * @cwin: pointer to at least crush_work_size() bytes of memory |
892 | * @choose_args: weights and ids for each known bucket |
893 | */ |
894 | int crush_do_rule(const struct crush_map *map, |
895 | int ruleno, int x, int *result, int result_max, |
896 | const __u32 *weight, int weight_max, |
897 | void *cwin, const struct crush_choose_arg *choose_args) |
898 | { |
899 | int result_len; |
900 | struct crush_work *cw = cwin; |
901 | int *a = cwin + map->working_size; |
902 | int *b = a + result_max; |
903 | int *c = b + result_max; |
904 | int *w = a; |
905 | int *o = b; |
906 | int recurse_to_leaf; |
907 | int wsize = 0; |
908 | int osize; |
909 | const struct crush_rule *rule; |
910 | __u32 step; |
911 | int i, j; |
912 | int numrep; |
913 | int out_size; |
914 | /* |
915 | * the original choose_total_tries value was off by one (it |
916 | * counted "retries" and not "tries"). add one. |
917 | */ |
918 | int choose_tries = map->choose_total_tries + 1; |
919 | int choose_leaf_tries = 0; |
920 | /* |
921 | * the local tries values were counted as "retries", though, |
922 | * and need no adjustment |
923 | */ |
924 | int choose_local_retries = map->choose_local_tries; |
925 | int choose_local_fallback_retries = map->choose_local_fallback_tries; |
926 | |
927 | int vary_r = map->chooseleaf_vary_r; |
928 | int stable = map->chooseleaf_stable; |
929 | |
930 | if ((__u32)ruleno >= map->max_rules) { |
931 | dprintk(" bad ruleno %d\n" , ruleno); |
932 | return 0; |
933 | } |
934 | |
935 | rule = map->rules[ruleno]; |
936 | result_len = 0; |
937 | |
938 | for (step = 0; step < rule->len; step++) { |
939 | int firstn = 0; |
940 | const struct crush_rule_step *curstep = &rule->steps[step]; |
941 | |
942 | switch (curstep->op) { |
943 | case CRUSH_RULE_TAKE: |
944 | if ((curstep->arg1 >= 0 && |
945 | curstep->arg1 < map->max_devices) || |
946 | (-1-curstep->arg1 >= 0 && |
947 | -1-curstep->arg1 < map->max_buckets && |
948 | map->buckets[-1-curstep->arg1])) { |
949 | w[0] = curstep->arg1; |
950 | wsize = 1; |
951 | } else { |
952 | dprintk(" bad take value %d\n" , curstep->arg1); |
953 | } |
954 | break; |
955 | |
956 | case CRUSH_RULE_SET_CHOOSE_TRIES: |
957 | if (curstep->arg1 > 0) |
958 | choose_tries = curstep->arg1; |
959 | break; |
960 | |
961 | case CRUSH_RULE_SET_CHOOSELEAF_TRIES: |
962 | if (curstep->arg1 > 0) |
963 | choose_leaf_tries = curstep->arg1; |
964 | break; |
965 | |
966 | case CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES: |
967 | if (curstep->arg1 >= 0) |
968 | choose_local_retries = curstep->arg1; |
969 | break; |
970 | |
971 | case CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES: |
972 | if (curstep->arg1 >= 0) |
973 | choose_local_fallback_retries = curstep->arg1; |
974 | break; |
975 | |
976 | case CRUSH_RULE_SET_CHOOSELEAF_VARY_R: |
977 | if (curstep->arg1 >= 0) |
978 | vary_r = curstep->arg1; |
979 | break; |
980 | |
981 | case CRUSH_RULE_SET_CHOOSELEAF_STABLE: |
982 | if (curstep->arg1 >= 0) |
983 | stable = curstep->arg1; |
984 | break; |
985 | |
986 | case CRUSH_RULE_CHOOSELEAF_FIRSTN: |
987 | case CRUSH_RULE_CHOOSE_FIRSTN: |
988 | firstn = 1; |
989 | fallthrough; |
990 | case CRUSH_RULE_CHOOSELEAF_INDEP: |
991 | case CRUSH_RULE_CHOOSE_INDEP: |
992 | if (wsize == 0) |
993 | break; |
994 | |
995 | recurse_to_leaf = |
996 | curstep->op == |
997 | CRUSH_RULE_CHOOSELEAF_FIRSTN || |
998 | curstep->op == |
999 | CRUSH_RULE_CHOOSELEAF_INDEP; |
1000 | |
1001 | /* reset output */ |
1002 | osize = 0; |
1003 | |
1004 | for (i = 0; i < wsize; i++) { |
1005 | int bno; |
1006 | numrep = curstep->arg1; |
1007 | if (numrep <= 0) { |
1008 | numrep += result_max; |
1009 | if (numrep <= 0) |
1010 | continue; |
1011 | } |
1012 | j = 0; |
1013 | /* make sure bucket id is valid */ |
1014 | bno = -1 - w[i]; |
1015 | if (bno < 0 || bno >= map->max_buckets) { |
1016 | /* w[i] is probably CRUSH_ITEM_NONE */ |
1017 | dprintk(" bad w[i] %d\n" , w[i]); |
1018 | continue; |
1019 | } |
1020 | if (firstn) { |
1021 | int recurse_tries; |
1022 | if (choose_leaf_tries) |
1023 | recurse_tries = |
1024 | choose_leaf_tries; |
1025 | else if (map->chooseleaf_descend_once) |
1026 | recurse_tries = 1; |
1027 | else |
1028 | recurse_tries = choose_tries; |
1029 | osize += crush_choose_firstn( |
1030 | map, |
1031 | work: cw, |
1032 | bucket: map->buckets[bno], |
1033 | weight, weight_max, |
1034 | x, numrep, |
1035 | type: curstep->arg2, |
1036 | out: o+osize, outpos: j, |
1037 | out_size: result_max-osize, |
1038 | tries: choose_tries, |
1039 | recurse_tries, |
1040 | local_retries: choose_local_retries, |
1041 | local_fallback_retries: choose_local_fallback_retries, |
1042 | recurse_to_leaf, |
1043 | vary_r, |
1044 | stable, |
1045 | out2: c+osize, |
1046 | parent_r: 0, |
1047 | choose_args); |
1048 | } else { |
1049 | out_size = ((numrep < (result_max-osize)) ? |
1050 | numrep : (result_max-osize)); |
1051 | crush_choose_indep( |
1052 | map, |
1053 | work: cw, |
1054 | bucket: map->buckets[bno], |
1055 | weight, weight_max, |
1056 | x, left: out_size, numrep, |
1057 | type: curstep->arg2, |
1058 | out: o+osize, outpos: j, |
1059 | tries: choose_tries, |
1060 | recurse_tries: choose_leaf_tries ? |
1061 | choose_leaf_tries : 1, |
1062 | recurse_to_leaf, |
1063 | out2: c+osize, |
1064 | parent_r: 0, |
1065 | choose_args); |
1066 | osize += out_size; |
1067 | } |
1068 | } |
1069 | |
1070 | if (recurse_to_leaf) |
1071 | /* copy final _leaf_ values to output set */ |
1072 | memcpy(o, c, osize*sizeof(*o)); |
1073 | |
1074 | /* swap o and w arrays */ |
1075 | swap(o, w); |
1076 | wsize = osize; |
1077 | break; |
1078 | |
1079 | |
1080 | case CRUSH_RULE_EMIT: |
1081 | for (i = 0; i < wsize && result_len < result_max; i++) { |
1082 | result[result_len] = w[i]; |
1083 | result_len++; |
1084 | } |
1085 | wsize = 0; |
1086 | break; |
1087 | |
1088 | default: |
1089 | dprintk(" unknown op %d at step %d\n" , |
1090 | curstep->op, step); |
1091 | break; |
1092 | } |
1093 | } |
1094 | |
1095 | return result_len; |
1096 | } |
1097 | |