aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ceph/crush/mapper.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ceph/crush/mapper.c')
-rw-r--r--fs/ceph/crush/mapper.c596
1 files changed, 596 insertions, 0 deletions
diff --git a/fs/ceph/crush/mapper.c b/fs/ceph/crush/mapper.c
new file mode 100644
index 000000000000..9ba54efb6543
--- /dev/null
+++ b/fs/ceph/crush/mapper.c
@@ -0,0 +1,596 @@
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
21#include "crush.h"
22#include "hash.h"
23
24/*
25 * Implement the core CRUSH mapping algorithm.
26 */
27
28/**
29 * crush_find_rule - find a crush_rule id for a given ruleset, type, and size.
30 * @map: the crush_map
31 * @ruleset: the storage ruleset id (user defined)
32 * @type: storage ruleset type (user defined)
33 * @size: output set size
34 */
35int crush_find_rule(struct crush_map *map, int ruleset, int type, int size)
36{
37 int i;
38
39 for (i = 0; i < map->max_rules; i++) {
40 if (map->rules[i] &&
41 map->rules[i]->mask.ruleset == ruleset &&
42 map->rules[i]->mask.type == type &&
43 map->rules[i]->mask.min_size <= size &&
44 map->rules[i]->mask.max_size >= size)
45 return i;
46 }
47 return -1;
48}
49
50
51/*
52 * bucket choose methods
53 *
54 * For each bucket algorithm, we have a "choose" method that, given a
55 * crush input @x and replica position (usually, position in output set) @r,
56 * will produce an item in the bucket.
57 */
58
59/*
60 * Choose based on a random permutation of the bucket.
61 *
62 * We used to use some prime number arithmetic to do this, but it
63 * wasn't very random, and had some other bad behaviors. Instead, we
64 * calculate an actual random permutation of the bucket members.
65 * Since this is expensive, we optimize for the r=0 case, which
66 * captures the vast majority of calls.
67 */
68static int bucket_perm_choose(struct crush_bucket *bucket,
69 int x, int r)
70{
71 unsigned pr = r % bucket->size;
72 unsigned i, s;
73
74 /* start a new permutation if @x has changed */
75 if (bucket->perm_x != x || bucket->perm_n == 0) {
76 dprintk("bucket %d new x=%d\n", bucket->id, x);
77 bucket->perm_x = x;
78
79 /* optimize common r=0 case */
80 if (pr == 0) {
81 s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
82 bucket->size;
83 bucket->perm[0] = s;
84 bucket->perm_n = 0xffff; /* magic value, see below */
85 goto out;
86 }
87
88 for (i = 0; i < bucket->size; i++)
89 bucket->perm[i] = i;
90 bucket->perm_n = 0;
91 } else if (bucket->perm_n == 0xffff) {
92 /* clean up after the r=0 case above */
93 for (i = 1; i < bucket->size; i++)
94 bucket->perm[i] = i;
95 bucket->perm[bucket->perm[0]] = 0;
96 bucket->perm_n = 1;
97 }
98
99 /* calculate permutation up to pr */
100 for (i = 0; i < bucket->perm_n; i++)
101 dprintk(" perm_choose have %d: %d\n", i, bucket->perm[i]);
102 while (bucket->perm_n <= pr) {
103 unsigned p = bucket->perm_n;
104 /* no point in swapping the final entry */
105 if (p < bucket->size - 1) {
106 i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
107 (bucket->size - p);
108 if (i) {
109 unsigned t = bucket->perm[p + i];
110 bucket->perm[p + i] = bucket->perm[p];
111 bucket->perm[p] = t;
112 }
113 dprintk(" perm_choose swap %d with %d\n", p, p+i);
114 }
115 bucket->perm_n++;
116 }
117 for (i = 0; i < bucket->size; i++)
118 dprintk(" perm_choose %d: %d\n", i, bucket->perm[i]);
119
120 s = bucket->perm[pr];
121out:
122 dprintk(" perm_choose %d sz=%d x=%d r=%d (%d) s=%d\n", bucket->id,
123 bucket->size, x, r, pr, s);
124 return bucket->items[s];
125}
126
127/* uniform */
128static int bucket_uniform_choose(struct crush_bucket_uniform *bucket,
129 int x, int r)
130{
131 return bucket_perm_choose(&bucket->h, x, r);
132}
133
134/* list */
135static int bucket_list_choose(struct crush_bucket_list *bucket,
136 int x, int r)
137{
138 int i;
139
140 for (i = bucket->h.size-1; i >= 0; i--) {
141 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
142 r, bucket->h.id);
143 w &= 0xffff;
144 dprintk("list_choose i=%d x=%d r=%d item %d weight %x "
145 "sw %x rand %llx",
146 i, x, r, bucket->h.items[i], bucket->item_weights[i],
147 bucket->sum_weights[i], w);
148 w *= bucket->sum_weights[i];
149 w = w >> 16;
150 /*dprintk(" scaled %llx\n", w);*/
151 if (w < bucket->item_weights[i])
152 return bucket->h.items[i];
153 }
154
155 BUG_ON(1);
156 return 0;
157}
158
159
160/* (binary) tree */
161static int height(int n)
162{
163 int h = 0;
164 while ((n & 1) == 0) {
165 h++;
166 n = n >> 1;
167 }
168 return h;
169}
170
171static int left(int x)
172{
173 int h = height(x);
174 return x - (1 << (h-1));
175}
176
177static int right(int x)
178{
179 int h = height(x);
180 return x + (1 << (h-1));
181}
182
183static int terminal(int x)
184{
185 return x & 1;
186}
187
188static int bucket_tree_choose(struct crush_bucket_tree *bucket,
189 int x, int r)
190{
191 int n, l;
192 __u32 w;
193 __u64 t;
194
195 /* start at root */
196 n = bucket->num_nodes >> 1;
197
198 while (!terminal(n)) {
199 /* pick point in [0, w) */
200 w = bucket->node_weights[n];
201 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
202 bucket->h.id) * (__u64)w;
203 t = t >> 32;
204
205 /* descend to the left or right? */
206 l = left(n);
207 if (t < bucket->node_weights[l])
208 n = l;
209 else
210 n = right(n);
211 }
212
213 return bucket->h.items[n >> 1];
214}
215
216
217/* straw */
218
219static int bucket_straw_choose(struct crush_bucket_straw *bucket,
220 int x, int r)
221{
222 int i;
223 int high = 0;
224 __u64 high_draw = 0;
225 __u64 draw;
226
227 for (i = 0; i < bucket->h.size; i++) {
228 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
229 draw &= 0xffff;
230 draw *= bucket->straws[i];
231 if (i == 0 || draw > high_draw) {
232 high = i;
233 high_draw = draw;
234 }
235 }
236 return bucket->h.items[high];
237}
238
239static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
240{
241 dprintk("choose %d x=%d r=%d\n", in->id, x, r);
242 switch (in->alg) {
243 case CRUSH_BUCKET_UNIFORM:
244 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
245 x, r);
246 case CRUSH_BUCKET_LIST:
247 return bucket_list_choose((struct crush_bucket_list *)in,
248 x, r);
249 case CRUSH_BUCKET_TREE:
250 return bucket_tree_choose((struct crush_bucket_tree *)in,
251 x, r);
252 case CRUSH_BUCKET_STRAW:
253 return bucket_straw_choose((struct crush_bucket_straw *)in,
254 x, r);
255 default:
256 BUG_ON(1);
257 return in->items[0];
258 }
259}
260
261/*
262 * true if device is marked "out" (failed, fully offloaded)
263 * of the cluster
264 */
265static int is_out(struct crush_map *map, __u32 *weight, int item, int x)
266{
267 if (weight[item] >= 0x1000)
268 return 0;
269 if (weight[item] == 0)
270 return 1;
271 if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
272 < weight[item])
273 return 0;
274 return 1;
275}
276
277/**
278 * crush_choose - choose numrep distinct items of given type
279 * @map: the crush_map
280 * @bucket: the bucket we are choose an item from
281 * @x: crush input value
282 * @numrep: the number of items to choose
283 * @type: the type of item to choose
284 * @out: pointer to output vector
285 * @outpos: our position in that vector
286 * @firstn: true if choosing "first n" items, false if choosing "indep"
287 * @recurse_to_leaf: true if we want one device under each item of given type
288 * @out2: second output vector for leaf items (if @recurse_to_leaf)
289 */
290static int crush_choose(struct crush_map *map,
291 struct crush_bucket *bucket,
292 __u32 *weight,
293 int x, int numrep, int type,
294 int *out, int outpos,
295 int firstn, int recurse_to_leaf,
296 int *out2)
297{
298 int rep;
299 int ftotal, flocal;
300 int retry_descent, retry_bucket, skip_rep;
301 struct crush_bucket *in = bucket;
302 int r;
303 int i;
304 int item = 0;
305 int itemtype;
306 int collide, reject;
307 const int orig_tries = 5; /* attempts before we fall back to search */
308 dprintk("choose bucket %d x %d outpos %d\n", bucket->id, x, outpos);
309
310 for (rep = outpos; rep < numrep; rep++) {
311 /* keep trying until we get a non-out, non-colliding item */
312 ftotal = 0;
313 skip_rep = 0;
314 do {
315 retry_descent = 0;
316 in = bucket; /* initial bucket */
317
318 /* choose through intervening buckets */
319 flocal = 0;
320 do {
321 collide = 0;
322 retry_bucket = 0;
323 r = rep;
324 if (in->alg == CRUSH_BUCKET_UNIFORM) {
325 /* be careful */
326 if (firstn || numrep >= in->size)
327 /* r' = r + f_total */
328 r += ftotal;
329 else if (in->size % numrep == 0)
330 /* r'=r+(n+1)*f_local */
331 r += (numrep+1) *
332 (flocal+ftotal);
333 else
334 /* r' = r + n*f_local */
335 r += numrep * (flocal+ftotal);
336 } else {
337 if (firstn)
338 /* r' = r + f_total */
339 r += ftotal;
340 else
341 /* r' = r + n*f_local */
342 r += numrep * (flocal+ftotal);
343 }
344
345 /* bucket choose */
346 if (in->size == 0) {
347 reject = 1;
348 goto reject;
349 }
350 if (flocal >= (in->size>>1) &&
351 flocal > orig_tries)
352 item = bucket_perm_choose(in, x, r);
353 else
354 item = crush_bucket_choose(in, x, r);
355 BUG_ON(item >= map->max_devices);
356
357 /* desired type? */
358 if (item < 0)
359 itemtype = map->buckets[-1-item]->type;
360 else
361 itemtype = 0;
362 dprintk(" item %d type %d\n", item, itemtype);
363
364 /* keep going? */
365 if (itemtype != type) {
366 BUG_ON(item >= 0 ||
367 (-1-item) >= map->max_buckets);
368 in = map->buckets[-1-item];
369 continue;
370 }
371
372 /* collision? */
373 for (i = 0; i < outpos; i++) {
374 if (out[i] == item) {
375 collide = 1;
376 break;
377 }
378 }
379
380 if (recurse_to_leaf &&
381 item < 0 &&
382 crush_choose(map, map->buckets[-1-item],
383 weight,
384 x, outpos+1, 0,
385 out2, outpos,
386 firstn, 0, NULL) <= outpos) {
387 reject = 1;
388 } else {
389 /* out? */
390 if (itemtype == 0)
391 reject = is_out(map, weight,
392 item, x);
393 else
394 reject = 0;
395 }
396
397reject:
398 if (reject || collide) {
399 ftotal++;
400 flocal++;
401
402 if (collide && flocal < 3)
403 /* retry locally a few times */
404 retry_bucket = 1;
405 else if (flocal < in->size + orig_tries)
406 /* exhaustive bucket search */
407 retry_bucket = 1;
408 else if (ftotal < 20)
409 /* then retry descent */
410 retry_descent = 1;
411 else
412 /* else give up */
413 skip_rep = 1;
414 dprintk(" reject %d collide %d "
415 "ftotal %d flocal %d\n",
416 reject, collide, ftotal,
417 flocal);
418 }
419 } while (retry_bucket);
420 } while (retry_descent);
421
422 if (skip_rep) {
423 dprintk("skip rep\n");
424 continue;
425 }
426
427 dprintk("choose got %d\n", item);
428 out[outpos] = item;
429 outpos++;
430 }
431
432 dprintk("choose returns %d\n", outpos);
433 return outpos;
434}
435
436
437/**
438 * crush_do_rule - calculate a mapping with the given input and rule
439 * @map: the crush_map
440 * @ruleno: the rule id
441 * @x: hash input
442 * @result: pointer to result vector
443 * @result_max: maximum result size
444 * @force: force initial replica choice; -1 for none
445 */
446int crush_do_rule(struct crush_map *map,
447 int ruleno, int x, int *result, int result_max,
448 int force, __u32 *weight)
449{
450 int result_len;
451 int force_context[CRUSH_MAX_DEPTH];
452 int force_pos = -1;
453 int a[CRUSH_MAX_SET];
454 int b[CRUSH_MAX_SET];
455 int c[CRUSH_MAX_SET];
456 int recurse_to_leaf;
457 int *w;
458 int wsize = 0;
459 int *o;
460 int osize;
461 int *tmp;
462 struct crush_rule *rule;
463 int step;
464 int i, j;
465 int numrep;
466 int firstn;
467 int rc = -1;
468
469 BUG_ON(ruleno >= map->max_rules);
470
471 rule = map->rules[ruleno];
472 result_len = 0;
473 w = a;
474 o = b;
475
476 /*
477 * determine hierarchical context of force, if any. note
478 * that this may or may not correspond to the specific types
479 * referenced by the crush rule.
480 */
481 if (force >= 0) {
482 if (force >= map->max_devices ||
483 map->device_parents[force] == 0) {
484 /*dprintk("CRUSH: forcefed device dne\n");*/
485 rc = -1; /* force fed device dne */
486 goto out;
487 }
488 if (!is_out(map, weight, force, x)) {
489 while (1) {
490 force_context[++force_pos] = force;
491 if (force >= 0)
492 force = map->device_parents[force];
493 else
494 force = map->bucket_parents[-1-force];
495 if (force == 0)
496 break;
497 }
498 }
499 }
500
501 for (step = 0; step < rule->len; step++) {
502 firstn = 0;
503 switch (rule->steps[step].op) {
504 case CRUSH_RULE_TAKE:
505 w[0] = rule->steps[step].arg1;
506 if (force_pos >= 0) {
507 BUG_ON(force_context[force_pos] != w[0]);
508 force_pos--;
509 }
510 wsize = 1;
511 break;
512
513 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
514 case CRUSH_RULE_CHOOSE_FIRSTN:
515 firstn = 1;
516 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
517 case CRUSH_RULE_CHOOSE_INDEP:
518 BUG_ON(wsize == 0);
519
520 recurse_to_leaf =
521 rule->steps[step].op ==
522 CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
523 rule->steps[step].op ==
524 CRUSH_RULE_CHOOSE_LEAF_INDEP;
525
526 /* reset output */
527 osize = 0;
528
529 for (i = 0; i < wsize; i++) {
530 /*
531 * see CRUSH_N, CRUSH_N_MINUS macros.
532 * basically, numrep <= 0 means relative to
533 * the provided result_max
534 */
535 numrep = rule->steps[step].arg1;
536 if (numrep <= 0) {
537 numrep += result_max;
538 if (numrep <= 0)
539 continue;
540 }
541 j = 0;
542 if (osize == 0 && force_pos >= 0) {
543 /* skip any intermediate types */
544 while (force_pos &&
545 force_context[force_pos] < 0 &&
546 rule->steps[step].arg2 !=
547 map->buckets[-1 -
548 force_context[force_pos]]->type)
549 force_pos--;
550 o[osize] = force_context[force_pos];
551 if (recurse_to_leaf)
552 c[osize] = force_context[0];
553 j++;
554 force_pos--;
555 }
556 osize += crush_choose(map,
557 map->buckets[-1-w[i]],
558 weight,
559 x, numrep,
560 rule->steps[step].arg2,
561 o+osize, j,
562 firstn,
563 recurse_to_leaf, c+osize);
564 }
565
566 if (recurse_to_leaf)
567 /* copy final _leaf_ values to output set */
568 memcpy(o, c, osize*sizeof(*o));
569
570 /* swap t and w arrays */
571 tmp = o;
572 o = w;
573 w = tmp;
574 wsize = osize;
575 break;
576
577
578 case CRUSH_RULE_EMIT:
579 for (i = 0; i < wsize && result_len < result_max; i++) {
580 result[result_len] = w[i];
581 result_len++;
582 }
583 wsize = 0;
584 break;
585
586 default:
587 BUG_ON(1);
588 }
589 }
590 rc = result_len;
591
592out:
593 return rc;
594}
595
596
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-03 00:48:46 -0500