]> git.karo-electronics.de Git - karo-tx-linux.git/blob - net/ceph/crush/mapper.c
crush: clarify numrep vs endpos
[karo-tx-linux.git] / net / ceph / crush / mapper.c
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 <linux/crush/crush.h>
22 #include <linux/crush/hash.h>
23 #include <linux/crush/mapper.h>
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  */
36 int crush_find_rule(const struct crush_map *map, int ruleset, int type, int size)
37 {
38         __u32 i;
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 {
72         unsigned int pr = r % bucket->size;
73         unsigned int i, s;
74
75         /* start a new permutation if @x has changed */
76         if (bucket->perm_x != (__u32)x || bucket->perm_n == 0) {
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) {
82                         s = crush_hash32_3(bucket->hash, x, bucket->id, 0) %
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) {
104                 unsigned int p = bucket->perm_n;
105                 /* no point in swapping the final entry */
106                 if (p < bucket->size - 1) {
107                         i = crush_hash32_3(bucket->hash, x, bucket->id, p) %
108                                 (bucket->size - p);
109                         if (i) {
110                                 unsigned int t = bucket->perm[p + i];
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--) {
142                 __u64 w = crush_hash32_4(bucket->h.hash,x, bucket->h.items[i],
143                                          r, bucket->h.id);
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
156         dprintk("bad list sums for bucket %d\n", bucket->h.id);
157         return bucket->h.items[0];
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 {
192         int n;
193         __u32 w;
194         __u64 t;
195
196         /* start at root */
197         n = bucket->num_nodes >> 1;
198
199         while (!terminal(n)) {
200                 int l;
201                 /* pick point in [0, w) */
202                 w = bucket->node_weights[n];
203                 t = (__u64)crush_hash32_4(bucket->h.hash, x, n, r,
204                                           bucket->h.id) * (__u64)w;
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 {
224         __u32 i;
225         int high = 0;
226         __u64 high_draw = 0;
227         __u64 draw;
228
229         for (i = 0; i < bucket->h.size; i++) {
230                 draw = crush_hash32_3(bucket->h.hash, x, bucket->h.items[i], r);
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
241 static int crush_bucket_choose(struct crush_bucket *in, int x, int r)
242 {
243         dprintk(" crush_bucket_choose %d x=%d r=%d\n", in->id, x, r);
244         BUG_ON(in->size == 0);
245         switch (in->alg) {
246         case CRUSH_BUCKET_UNIFORM:
247                 return bucket_uniform_choose((struct crush_bucket_uniform *)in,
248                                           x, r);
249         case CRUSH_BUCKET_LIST:
250                 return bucket_list_choose((struct crush_bucket_list *)in,
251                                           x, r);
252         case CRUSH_BUCKET_TREE:
253                 return bucket_tree_choose((struct crush_bucket_tree *)in,
254                                           x, r);
255         case CRUSH_BUCKET_STRAW:
256                 return bucket_straw_choose((struct crush_bucket_straw *)in,
257                                            x, r);
258         default:
259                 dprintk("unknown bucket %d alg %d\n", in->id, in->alg);
260                 return in->items[0];
261         }
262 }
263
264 /*
265  * true if device is marked "out" (failed, fully offloaded)
266  * of the cluster
267  */
268 static int is_out(const struct crush_map *map,
269                   const __u32 *weight, int weight_max,
270                   int item, int x)
271 {
272         if (item >= weight_max)
273                 return 1;
274         if (weight[item] >= 0x10000)
275                 return 0;
276         if (weight[item] == 0)
277                 return 1;
278         if ((crush_hash32_2(CRUSH_HASH_RJENKINS1, x, item) & 0xffff)
279             < weight[item])
280                 return 0;
281         return 1;
282 }
283
284 /**
285  * crush_choose_firstn - choose numrep distinct items of given type
286  * @map: the crush_map
287  * @bucket: the bucket we are choose an item from
288  * @x: crush input value
289  * @numrep: the number of items to choose
290  * @type: the type of item to choose
291  * @out: pointer to output vector
292  * @outpos: our position in that vector
293  * @recurse_to_leaf: true if we want one device under each item of given type
294  * @descend_once: true if we should only try one descent before giving up
295  * @out2: second output vector for leaf items (if @recurse_to_leaf)
296  */
297 static int crush_choose_firstn(const struct crush_map *map,
298                                struct crush_bucket *bucket,
299                                const __u32 *weight, int weight_max,
300                                int x, int numrep, int type,
301                                int *out, int outpos,
302                                int recurse_to_leaf,
303                                int descend_once, int *out2)
304 {
305         int rep;
306         unsigned int ftotal, flocal;
307         int retry_descent, retry_bucket, skip_rep;
308         struct crush_bucket *in = bucket;
309         int r;
310         int i;
311         int item = 0;
312         int itemtype;
313         int collide, reject;
314
315         dprintk("CHOOSE%s bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
316                 bucket->id, x, outpos, numrep);
317
318         for (rep = outpos; rep < numrep; rep++) {
319                 /* keep trying until we get a non-out, non-colliding item */
320                 ftotal = 0;
321                 skip_rep = 0;
322                 do {
323                         retry_descent = 0;
324                         in = bucket;               /* initial bucket */
325
326                         /* choose through intervening buckets */
327                         flocal = 0;
328                         do {
329                                 collide = 0;
330                                 retry_bucket = 0;
331                                 r = rep;
332                                 /* r' = r + f_total */
333                                 r += ftotal;
334
335                                 /* bucket choose */
336                                 if (in->size == 0) {
337                                         reject = 1;
338                                         goto reject;
339                                 }
340                                 if (map->choose_local_fallback_tries > 0 &&
341                                     flocal >= (in->size>>1) &&
342                                     flocal > map->choose_local_fallback_tries)
343                                         item = bucket_perm_choose(in, x, r);
344                                 else
345                                         item = crush_bucket_choose(in, x, r);
346                                 if (item >= map->max_devices) {
347                                         dprintk("   bad item %d\n", item);
348                                         skip_rep = 1;
349                                         break;
350                                 }
351
352                                 /* desired type? */
353                                 if (item < 0)
354                                         itemtype = map->buckets[-1-item]->type;
355                                 else
356                                         itemtype = 0;
357                                 dprintk("  item %d type %d\n", item, itemtype);
358
359                                 /* keep going? */
360                                 if (itemtype != type) {
361                                         if (item >= 0 ||
362                                             (-1-item) >= map->max_buckets) {
363                                                 dprintk("   bad item type %d\n", type);
364                                                 skip_rep = 1;
365                                                 break;
366                                         }
367                                         in = map->buckets[-1-item];
368                                         retry_bucket = 1;
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                                 reject = 0;
381                                 if (!collide && recurse_to_leaf) {
382                                         if (item < 0) {
383                                                 if (crush_choose_firstn(map,
384                                                          map->buckets[-1-item],
385                                                          weight, weight_max,
386                                                          x, outpos+1, 0,
387                                                          out2, outpos,
388                                                          0,
389                                                          map->chooseleaf_descend_once,
390                                                          NULL) <= outpos)
391                                                         /* didn't get leaf */
392                                                         reject = 1;
393                                         } else {
394                                                 /* we already have a leaf! */
395                                                 out2[outpos] = item;
396                                         }
397                                 }
398
399                                 if (!reject) {
400                                         /* out? */
401                                         if (itemtype == 0)
402                                                 reject = is_out(map, weight,
403                                                                 weight_max,
404                                                                 item, x);
405                                         else
406                                                 reject = 0;
407                                 }
408
409 reject:
410                                 if (reject || collide) {
411                                         ftotal++;
412                                         flocal++;
413
414                                         if (reject && descend_once)
415                                                 /* let outer call try again */
416                                                 skip_rep = 1;
417                                         else if (collide && flocal <= map->choose_local_tries)
418                                                 /* retry locally a few times */
419                                                 retry_bucket = 1;
420                                         else if (map->choose_local_fallback_tries > 0 &&
421                                                  flocal <= in->size + map->choose_local_fallback_tries)
422                                                 /* exhaustive bucket search */
423                                                 retry_bucket = 1;
424                                         else if (ftotal <= map->choose_total_tries)
425                                                 /* then retry descent */
426                                                 retry_descent = 1;
427                                         else
428                                                 /* else give up */
429                                                 skip_rep = 1;
430                                         dprintk("  reject %d  collide %d  "
431                                                 "ftotal %u  flocal %u\n",
432                                                 reject, collide, ftotal,
433                                                 flocal);
434                                 }
435                         } while (retry_bucket);
436                 } while (retry_descent);
437
438                 if (skip_rep) {
439                         dprintk("skip rep\n");
440                         continue;
441                 }
442
443                 dprintk("CHOOSE got %d\n", item);
444                 out[outpos] = item;
445                 outpos++;
446         }
447
448         dprintk("CHOOSE returns %d\n", outpos);
449         return outpos;
450 }
451
452
453 /**
454  * crush_choose_indep: alternative breadth-first positionally stable mapping
455  *
456  */
457 static void crush_choose_indep(const struct crush_map *map,
458                               struct crush_bucket *bucket,
459                               const __u32 *weight, int weight_max,
460                                int x, int left, int numrep, int type,
461                               int *out, int outpos,
462                               int recurse_to_leaf,
463                               int *out2)
464 {
465         struct crush_bucket *in = bucket;
466         int endpos = outpos + left;
467         int rep;
468         unsigned int ftotal;
469         int r;
470         int i;
471         int item = 0;
472         int itemtype;
473         int collide;
474
475         dprintk("CHOOSE%s INDEP bucket %d x %d outpos %d numrep %d\n", recurse_to_leaf ? "_LEAF" : "",
476                 bucket->id, x, outpos, numrep);
477
478         /* initially my result is undefined */
479         for (rep = outpos; rep < endpos; rep++) {
480                 out[rep] = CRUSH_ITEM_UNDEF;
481                 if (out2)
482                         out2[rep] = CRUSH_ITEM_UNDEF;
483         }
484
485         for (ftotal = 0; left > 0 && ftotal < map->choose_total_tries; ftotal++) {
486                 for (rep = outpos; rep < endpos; rep++) {
487                         if (out[rep] != CRUSH_ITEM_UNDEF)
488                                 continue;
489
490                         in = bucket;  /* initial bucket */
491
492                         /* choose through intervening buckets */
493                         for (;;) {
494                                 /* note: we base the choice on the position
495                                  * even in the nested call.  that means that
496                                  * if the first layer chooses the same bucket
497                                  * in a different position, we will tend to
498                                  * choose a different item in that bucket.
499                                  * this will involve more devices in data
500                                  * movement and tend to distribute the load.
501                                  */
502                                 r = rep;
503
504                                 /* be careful */
505                                 if (in->alg == CRUSH_BUCKET_UNIFORM &&
506                                     in->size % numrep == 0)
507                                         /* r'=r+(n+1)*f_total */
508                                         r += (numrep+1) * ftotal;
509                                 else
510                                         /* r' = r + n*f_total */
511                                         r += numrep * ftotal;
512
513                                 /* bucket choose */
514                                 if (in->size == 0) {
515                                         dprintk("   empty bucket\n");
516                                         break;
517                                 }
518
519                                 item = crush_bucket_choose(in, x, r);
520                                 if (item >= map->max_devices) {
521                                         dprintk("   bad item %d\n", item);
522                                         out[rep] = CRUSH_ITEM_NONE;
523                                         if (out2)
524                                                 out2[rep] = CRUSH_ITEM_NONE;
525                                         left--;
526                                         break;
527                                 }
528
529                                 /* desired type? */
530                                 if (item < 0)
531                                         itemtype = map->buckets[-1-item]->type;
532                                 else
533                                         itemtype = 0;
534                                 dprintk("  item %d type %d\n", item, itemtype);
535
536                                 /* keep going? */
537                                 if (itemtype != type) {
538                                         if (item >= 0 ||
539                                             (-1-item) >= map->max_buckets) {
540                                                 dprintk("   bad item type %d\n", type);
541                                                 out[rep] = CRUSH_ITEM_NONE;
542                                                 if (out2)
543                                                         out2[rep] =
544                                                                 CRUSH_ITEM_NONE;
545                                                 left--;
546                                                 break;
547                                         }
548                                         in = map->buckets[-1-item];
549                                         continue;
550                                 }
551
552                                 /* collision? */
553                                 collide = 0;
554                                 for (i = outpos; i < endpos; i++) {
555                                         if (out[i] == item) {
556                                                 collide = 1;
557                                                 break;
558                                         }
559                                 }
560                                 if (collide)
561                                         break;
562
563                                 if (recurse_to_leaf) {
564                                         if (item < 0) {
565                                                 crush_choose_indep(map,
566                                                                    map->buckets[-1-item],
567                                                                    weight, weight_max,
568                                                                    x, 1, numrep, 0,
569                                                                    out2, rep,
570                                                                    0, NULL);
571                                                 if (out2[rep] == CRUSH_ITEM_NONE) {
572                                                         /* placed nothing; no leaf */
573                                                         break;
574                                                 }
575                                         } else {
576                                                 /* we already have a leaf! */
577                                                 out2[rep] = item;
578                                         }
579                                 }
580
581                                 /* out? */
582                                 if (itemtype == 0 &&
583                                     is_out(map, weight, weight_max, item, x))
584                                         break;
585
586                                 /* yay! */
587                                 out[rep] = item;
588                                 left--;
589                                 break;
590                         }
591                 }
592         }
593         for (rep = outpos; rep < endpos; rep++) {
594                 if (out[rep] == CRUSH_ITEM_UNDEF) {
595                         out[rep] = CRUSH_ITEM_NONE;
596                 }
597                 if (out2 && out2[rep] == CRUSH_ITEM_UNDEF) {
598                         out2[rep] = CRUSH_ITEM_NONE;
599                 }
600         }
601 }
602
603 /**
604  * crush_do_rule - calculate a mapping with the given input and rule
605  * @map: the crush_map
606  * @ruleno: the rule id
607  * @x: hash input
608  * @result: pointer to result vector
609  * @result_max: maximum result size
610  * @weight: weight vector (for map leaves)
611  * @weight_max: size of weight vector
612  * @scratch: scratch vector for private use; must be >= 3 * result_max
613  */
614 int crush_do_rule(const struct crush_map *map,
615                   int ruleno, int x, int *result, int result_max,
616                   const __u32 *weight, int weight_max,
617                   int *scratch)
618 {
619         int result_len;
620         int *a = scratch;
621         int *b = scratch + result_max;
622         int *c = scratch + result_max*2;
623         int recurse_to_leaf;
624         int *w;
625         int wsize = 0;
626         int *o;
627         int osize;
628         int *tmp;
629         struct crush_rule *rule;
630         __u32 step;
631         int i, j;
632         int numrep;
633         const int descend_once = 0;
634
635         if ((__u32)ruleno >= map->max_rules) {
636                 dprintk(" bad ruleno %d\n", ruleno);
637                 return 0;
638         }
639
640         rule = map->rules[ruleno];
641         result_len = 0;
642         w = a;
643         o = b;
644
645         for (step = 0; step < rule->len; step++) {
646                 int firstn = 0;
647                 struct crush_rule_step *curstep = &rule->steps[step];
648
649                 switch (curstep->op) {
650                 case CRUSH_RULE_TAKE:
651                         w[0] = curstep->arg1;
652                         wsize = 1;
653                         break;
654
655                 case CRUSH_RULE_CHOOSE_LEAF_FIRSTN:
656                 case CRUSH_RULE_CHOOSE_FIRSTN:
657                         firstn = 1;
658                         /* fall through */
659                 case CRUSH_RULE_CHOOSE_LEAF_INDEP:
660                 case CRUSH_RULE_CHOOSE_INDEP:
661                         if (wsize == 0)
662                                 break;
663
664                         recurse_to_leaf =
665                                 curstep->op ==
666                                  CRUSH_RULE_CHOOSE_LEAF_FIRSTN ||
667                                 curstep->op ==
668                                 CRUSH_RULE_CHOOSE_LEAF_INDEP;
669
670                         /* reset output */
671                         osize = 0;
672
673                         for (i = 0; i < wsize; i++) {
674                                 /*
675                                  * see CRUSH_N, CRUSH_N_MINUS macros.
676                                  * basically, numrep <= 0 means relative to
677                                  * the provided result_max
678                                  */
679                                 numrep = curstep->arg1;
680                                 if (numrep <= 0) {
681                                         numrep += result_max;
682                                         if (numrep <= 0)
683                                                 continue;
684                                 }
685                                 j = 0;
686                                 if (firstn) {
687                                         osize += crush_choose_firstn(
688                                                 map,
689                                                 map->buckets[-1-w[i]],
690                                                 weight, weight_max,
691                                                 x, numrep,
692                                                 curstep->arg2,
693                                                 o+osize, j,
694                                                 recurse_to_leaf,
695                                                 descend_once, c+osize);
696                                 } else {
697                                         crush_choose_indep(
698                                                 map,
699                                                 map->buckets[-1-w[i]],
700                                                 weight, weight_max,
701                                                 x, numrep, numrep,
702                                                 curstep->arg2,
703                                                 o+osize, j,
704                                                 recurse_to_leaf,
705                                                 c+osize);
706                                         osize += numrep;
707                                 }
708                         }
709
710                         if (recurse_to_leaf)
711                                 /* copy final _leaf_ values to output set */
712                                 memcpy(o, c, osize*sizeof(*o));
713
714                         /* swap o and w arrays */
715                         tmp = o;
716                         o = w;
717                         w = tmp;
718                         wsize = osize;
719                         break;
720
721
722                 case CRUSH_RULE_EMIT:
723                         for (i = 0; i < wsize && result_len < result_max; i++) {
724                                 result[result_len] = w[i];
725                                 result_len++;
726                         }
727                         wsize = 0;
728                         break;
729
730                 default:
731                         dprintk(" unknown op %d at step %d\n",
732                                 curstep->op, step);
733                         break;
734                 }
735         }
736         return result_len;
737 }
738
739