Commit | Line | Data |
---|---|---|
f24e9980 | 1 | |
3d14c5d2 | 2 | #include <linux/ceph/ceph_debug.h> |
5a0e3ad6 | 3 | |
3d14c5d2 | 4 | #include <linux/module.h> |
5a0e3ad6 | 5 | #include <linux/slab.h> |
f24e9980 SW |
6 | #include <asm/div64.h> |
7 | ||
3d14c5d2 YS |
8 | #include <linux/ceph/libceph.h> |
9 | #include <linux/ceph/osdmap.h> | |
10 | #include <linux/ceph/decode.h> | |
11 | #include <linux/crush/hash.h> | |
12 | #include <linux/crush/mapper.h> | |
f24e9980 SW |
13 | |
14 | char *ceph_osdmap_state_str(char *str, int len, int state) | |
15 | { | |
16 | int flag = 0; | |
17 | ||
18 | if (!len) | |
19 | goto done; | |
20 | ||
21 | *str = '\0'; | |
22 | if (state) { | |
23 | if (state & CEPH_OSD_EXISTS) { | |
24 | snprintf(str, len, "exists"); | |
25 | flag = 1; | |
26 | } | |
27 | if (state & CEPH_OSD_UP) { | |
28 | snprintf(str, len, "%s%s%s", str, (flag ? ", " : ""), | |
29 | "up"); | |
30 | flag = 1; | |
31 | } | |
32 | } else { | |
33 | snprintf(str, len, "doesn't exist"); | |
34 | } | |
35 | done: | |
36 | return str; | |
37 | } | |
38 | ||
39 | /* maps */ | |
40 | ||
95c96174 | 41 | static int calc_bits_of(unsigned int t) |
f24e9980 SW |
42 | { |
43 | int b = 0; | |
44 | while (t) { | |
45 | t = t >> 1; | |
46 | b++; | |
47 | } | |
48 | return b; | |
49 | } | |
50 | ||
51 | /* | |
52 | * the foo_mask is the smallest value 2^n-1 that is >= foo. | |
53 | */ | |
54 | static void calc_pg_masks(struct ceph_pg_pool_info *pi) | |
55 | { | |
56 | pi->pg_num_mask = (1 << calc_bits_of(le32_to_cpu(pi->v.pg_num)-1)) - 1; | |
57 | pi->pgp_num_mask = | |
58 | (1 << calc_bits_of(le32_to_cpu(pi->v.pgp_num)-1)) - 1; | |
59 | pi->lpg_num_mask = | |
60 | (1 << calc_bits_of(le32_to_cpu(pi->v.lpg_num)-1)) - 1; | |
61 | pi->lpgp_num_mask = | |
62 | (1 << calc_bits_of(le32_to_cpu(pi->v.lpgp_num)-1)) - 1; | |
63 | } | |
64 | ||
65 | /* | |
66 | * decode crush map | |
67 | */ | |
68 | static int crush_decode_uniform_bucket(void **p, void *end, | |
69 | struct crush_bucket_uniform *b) | |
70 | { | |
71 | dout("crush_decode_uniform_bucket %p to %p\n", *p, end); | |
72 | ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad); | |
c89136ea | 73 | b->item_weight = ceph_decode_32(p); |
f24e9980 SW |
74 | return 0; |
75 | bad: | |
76 | return -EINVAL; | |
77 | } | |
78 | ||
79 | static int crush_decode_list_bucket(void **p, void *end, | |
80 | struct crush_bucket_list *b) | |
81 | { | |
82 | int j; | |
83 | dout("crush_decode_list_bucket %p to %p\n", *p, end); | |
84 | b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); | |
85 | if (b->item_weights == NULL) | |
86 | return -ENOMEM; | |
87 | b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); | |
88 | if (b->sum_weights == NULL) | |
89 | return -ENOMEM; | |
90 | ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad); | |
91 | for (j = 0; j < b->h.size; j++) { | |
c89136ea SW |
92 | b->item_weights[j] = ceph_decode_32(p); |
93 | b->sum_weights[j] = ceph_decode_32(p); | |
f24e9980 SW |
94 | } |
95 | return 0; | |
96 | bad: | |
97 | return -EINVAL; | |
98 | } | |
99 | ||
100 | static int crush_decode_tree_bucket(void **p, void *end, | |
101 | struct crush_bucket_tree *b) | |
102 | { | |
103 | int j; | |
104 | dout("crush_decode_tree_bucket %p to %p\n", *p, end); | |
105 | ceph_decode_32_safe(p, end, b->num_nodes, bad); | |
106 | b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS); | |
107 | if (b->node_weights == NULL) | |
108 | return -ENOMEM; | |
109 | ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad); | |
110 | for (j = 0; j < b->num_nodes; j++) | |
c89136ea | 111 | b->node_weights[j] = ceph_decode_32(p); |
f24e9980 SW |
112 | return 0; |
113 | bad: | |
114 | return -EINVAL; | |
115 | } | |
116 | ||
117 | static int crush_decode_straw_bucket(void **p, void *end, | |
118 | struct crush_bucket_straw *b) | |
119 | { | |
120 | int j; | |
121 | dout("crush_decode_straw_bucket %p to %p\n", *p, end); | |
122 | b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); | |
123 | if (b->item_weights == NULL) | |
124 | return -ENOMEM; | |
125 | b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); | |
126 | if (b->straws == NULL) | |
127 | return -ENOMEM; | |
128 | ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad); | |
129 | for (j = 0; j < b->h.size; j++) { | |
c89136ea SW |
130 | b->item_weights[j] = ceph_decode_32(p); |
131 | b->straws[j] = ceph_decode_32(p); | |
f24e9980 SW |
132 | } |
133 | return 0; | |
134 | bad: | |
135 | return -EINVAL; | |
136 | } | |
137 | ||
546f04ef SW |
138 | static int skip_name_map(void **p, void *end) |
139 | { | |
140 | int len; | |
141 | ceph_decode_32_safe(p, end, len ,bad); | |
142 | while (len--) { | |
143 | int strlen; | |
144 | *p += sizeof(u32); | |
145 | ceph_decode_32_safe(p, end, strlen, bad); | |
146 | *p += strlen; | |
147 | } | |
148 | return 0; | |
149 | bad: | |
150 | return -EINVAL; | |
151 | } | |
152 | ||
f24e9980 SW |
153 | static struct crush_map *crush_decode(void *pbyval, void *end) |
154 | { | |
155 | struct crush_map *c; | |
156 | int err = -EINVAL; | |
157 | int i, j; | |
158 | void **p = &pbyval; | |
159 | void *start = pbyval; | |
160 | u32 magic; | |
546f04ef | 161 | u32 num_name_maps; |
f24e9980 SW |
162 | |
163 | dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p)); | |
164 | ||
165 | c = kzalloc(sizeof(*c), GFP_NOFS); | |
166 | if (c == NULL) | |
167 | return ERR_PTR(-ENOMEM); | |
168 | ||
546f04ef SW |
169 | /* set tunables to default values */ |
170 | c->choose_local_tries = 2; | |
171 | c->choose_local_fallback_tries = 5; | |
172 | c->choose_total_tries = 19; | |
1604f488 | 173 | c->chooseleaf_descend_once = 0; |
546f04ef | 174 | |
f24e9980 | 175 | ceph_decode_need(p, end, 4*sizeof(u32), bad); |
c89136ea | 176 | magic = ceph_decode_32(p); |
f24e9980 SW |
177 | if (magic != CRUSH_MAGIC) { |
178 | pr_err("crush_decode magic %x != current %x\n", | |
95c96174 | 179 | (unsigned int)magic, (unsigned int)CRUSH_MAGIC); |
f24e9980 SW |
180 | goto bad; |
181 | } | |
c89136ea SW |
182 | c->max_buckets = ceph_decode_32(p); |
183 | c->max_rules = ceph_decode_32(p); | |
184 | c->max_devices = ceph_decode_32(p); | |
f24e9980 | 185 | |
f24e9980 SW |
186 | c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS); |
187 | if (c->buckets == NULL) | |
188 | goto badmem; | |
189 | c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS); | |
190 | if (c->rules == NULL) | |
191 | goto badmem; | |
192 | ||
193 | /* buckets */ | |
194 | for (i = 0; i < c->max_buckets; i++) { | |
195 | int size = 0; | |
196 | u32 alg; | |
197 | struct crush_bucket *b; | |
198 | ||
199 | ceph_decode_32_safe(p, end, alg, bad); | |
200 | if (alg == 0) { | |
201 | c->buckets[i] = NULL; | |
202 | continue; | |
203 | } | |
204 | dout("crush_decode bucket %d off %x %p to %p\n", | |
205 | i, (int)(*p-start), *p, end); | |
206 | ||
207 | switch (alg) { | |
208 | case CRUSH_BUCKET_UNIFORM: | |
209 | size = sizeof(struct crush_bucket_uniform); | |
210 | break; | |
211 | case CRUSH_BUCKET_LIST: | |
212 | size = sizeof(struct crush_bucket_list); | |
213 | break; | |
214 | case CRUSH_BUCKET_TREE: | |
215 | size = sizeof(struct crush_bucket_tree); | |
216 | break; | |
217 | case CRUSH_BUCKET_STRAW: | |
218 | size = sizeof(struct crush_bucket_straw); | |
219 | break; | |
220 | default: | |
30dc6381 | 221 | err = -EINVAL; |
f24e9980 SW |
222 | goto bad; |
223 | } | |
224 | BUG_ON(size == 0); | |
225 | b = c->buckets[i] = kzalloc(size, GFP_NOFS); | |
226 | if (b == NULL) | |
227 | goto badmem; | |
228 | ||
229 | ceph_decode_need(p, end, 4*sizeof(u32), bad); | |
c89136ea SW |
230 | b->id = ceph_decode_32(p); |
231 | b->type = ceph_decode_16(p); | |
fb690390 SW |
232 | b->alg = ceph_decode_8(p); |
233 | b->hash = ceph_decode_8(p); | |
c89136ea SW |
234 | b->weight = ceph_decode_32(p); |
235 | b->size = ceph_decode_32(p); | |
f24e9980 SW |
236 | |
237 | dout("crush_decode bucket size %d off %x %p to %p\n", | |
238 | b->size, (int)(*p-start), *p, end); | |
239 | ||
240 | b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS); | |
241 | if (b->items == NULL) | |
242 | goto badmem; | |
243 | b->perm = kcalloc(b->size, sizeof(u32), GFP_NOFS); | |
244 | if (b->perm == NULL) | |
245 | goto badmem; | |
246 | b->perm_n = 0; | |
247 | ||
248 | ceph_decode_need(p, end, b->size*sizeof(u32), bad); | |
249 | for (j = 0; j < b->size; j++) | |
c89136ea | 250 | b->items[j] = ceph_decode_32(p); |
f24e9980 SW |
251 | |
252 | switch (b->alg) { | |
253 | case CRUSH_BUCKET_UNIFORM: | |
254 | err = crush_decode_uniform_bucket(p, end, | |
255 | (struct crush_bucket_uniform *)b); | |
256 | if (err < 0) | |
257 | goto bad; | |
258 | break; | |
259 | case CRUSH_BUCKET_LIST: | |
260 | err = crush_decode_list_bucket(p, end, | |
261 | (struct crush_bucket_list *)b); | |
262 | if (err < 0) | |
263 | goto bad; | |
264 | break; | |
265 | case CRUSH_BUCKET_TREE: | |
266 | err = crush_decode_tree_bucket(p, end, | |
267 | (struct crush_bucket_tree *)b); | |
268 | if (err < 0) | |
269 | goto bad; | |
270 | break; | |
271 | case CRUSH_BUCKET_STRAW: | |
272 | err = crush_decode_straw_bucket(p, end, | |
273 | (struct crush_bucket_straw *)b); | |
274 | if (err < 0) | |
275 | goto bad; | |
276 | break; | |
277 | } | |
278 | } | |
279 | ||
280 | /* rules */ | |
281 | dout("rule vec is %p\n", c->rules); | |
282 | for (i = 0; i < c->max_rules; i++) { | |
283 | u32 yes; | |
284 | struct crush_rule *r; | |
285 | ||
286 | ceph_decode_32_safe(p, end, yes, bad); | |
287 | if (!yes) { | |
288 | dout("crush_decode NO rule %d off %x %p to %p\n", | |
289 | i, (int)(*p-start), *p, end); | |
290 | c->rules[i] = NULL; | |
291 | continue; | |
292 | } | |
293 | ||
294 | dout("crush_decode rule %d off %x %p to %p\n", | |
295 | i, (int)(*p-start), *p, end); | |
296 | ||
297 | /* len */ | |
298 | ceph_decode_32_safe(p, end, yes, bad); | |
299 | #if BITS_PER_LONG == 32 | |
30dc6381 | 300 | err = -EINVAL; |
64486697 XW |
301 | if (yes > (ULONG_MAX - sizeof(*r)) |
302 | / sizeof(struct crush_rule_step)) | |
f24e9980 SW |
303 | goto bad; |
304 | #endif | |
305 | r = c->rules[i] = kmalloc(sizeof(*r) + | |
306 | yes*sizeof(struct crush_rule_step), | |
307 | GFP_NOFS); | |
308 | if (r == NULL) | |
309 | goto badmem; | |
310 | dout(" rule %d is at %p\n", i, r); | |
311 | r->len = yes; | |
312 | ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */ | |
313 | ceph_decode_need(p, end, r->len*3*sizeof(u32), bad); | |
314 | for (j = 0; j < r->len; j++) { | |
c89136ea SW |
315 | r->steps[j].op = ceph_decode_32(p); |
316 | r->steps[j].arg1 = ceph_decode_32(p); | |
317 | r->steps[j].arg2 = ceph_decode_32(p); | |
f24e9980 SW |
318 | } |
319 | } | |
320 | ||
321 | /* ignore trailing name maps. */ | |
546f04ef SW |
322 | for (num_name_maps = 0; num_name_maps < 3; num_name_maps++) { |
323 | err = skip_name_map(p, end); | |
324 | if (err < 0) | |
325 | goto done; | |
326 | } | |
327 | ||
328 | /* tunables */ | |
329 | ceph_decode_need(p, end, 3*sizeof(u32), done); | |
330 | c->choose_local_tries = ceph_decode_32(p); | |
331 | c->choose_local_fallback_tries = ceph_decode_32(p); | |
332 | c->choose_total_tries = ceph_decode_32(p); | |
333 | dout("crush decode tunable choose_local_tries = %d", | |
334 | c->choose_local_tries); | |
335 | dout("crush decode tunable choose_local_fallback_tries = %d", | |
336 | c->choose_local_fallback_tries); | |
337 | dout("crush decode tunable choose_total_tries = %d", | |
338 | c->choose_total_tries); | |
f24e9980 | 339 | |
1604f488 JS |
340 | ceph_decode_need(p, end, sizeof(u32), done); |
341 | c->chooseleaf_descend_once = ceph_decode_32(p); | |
342 | dout("crush decode tunable chooseleaf_descend_once = %d", | |
343 | c->chooseleaf_descend_once); | |
344 | ||
546f04ef | 345 | done: |
f24e9980 SW |
346 | dout("crush_decode success\n"); |
347 | return c; | |
348 | ||
349 | badmem: | |
350 | err = -ENOMEM; | |
351 | bad: | |
352 | dout("crush_decode fail %d\n", err); | |
353 | crush_destroy(c); | |
354 | return ERR_PTR(err); | |
355 | } | |
356 | ||
f24e9980 | 357 | /* |
9794b146 SW |
358 | * rbtree of pg_mapping for handling pg_temp (explicit mapping of pgid |
359 | * to a set of osds) | |
f24e9980 | 360 | */ |
51042122 SW |
361 | static int pgid_cmp(struct ceph_pg l, struct ceph_pg r) |
362 | { | |
363 | u64 a = *(u64 *)&l; | |
364 | u64 b = *(u64 *)&r; | |
365 | ||
366 | if (a < b) | |
367 | return -1; | |
368 | if (a > b) | |
369 | return 1; | |
370 | return 0; | |
371 | } | |
372 | ||
991abb6e SW |
373 | static int __insert_pg_mapping(struct ceph_pg_mapping *new, |
374 | struct rb_root *root) | |
f24e9980 SW |
375 | { |
376 | struct rb_node **p = &root->rb_node; | |
377 | struct rb_node *parent = NULL; | |
378 | struct ceph_pg_mapping *pg = NULL; | |
51042122 | 379 | int c; |
f24e9980 | 380 | |
8adc8b3d | 381 | dout("__insert_pg_mapping %llx %p\n", *(u64 *)&new->pgid, new); |
f24e9980 SW |
382 | while (*p) { |
383 | parent = *p; | |
384 | pg = rb_entry(parent, struct ceph_pg_mapping, node); | |
51042122 SW |
385 | c = pgid_cmp(new->pgid, pg->pgid); |
386 | if (c < 0) | |
f24e9980 | 387 | p = &(*p)->rb_left; |
51042122 | 388 | else if (c > 0) |
f24e9980 SW |
389 | p = &(*p)->rb_right; |
390 | else | |
991abb6e | 391 | return -EEXIST; |
f24e9980 SW |
392 | } |
393 | ||
394 | rb_link_node(&new->node, parent, p); | |
395 | rb_insert_color(&new->node, root); | |
991abb6e | 396 | return 0; |
f24e9980 SW |
397 | } |
398 | ||
9794b146 SW |
399 | static struct ceph_pg_mapping *__lookup_pg_mapping(struct rb_root *root, |
400 | struct ceph_pg pgid) | |
401 | { | |
402 | struct rb_node *n = root->rb_node; | |
403 | struct ceph_pg_mapping *pg; | |
404 | int c; | |
405 | ||
406 | while (n) { | |
407 | pg = rb_entry(n, struct ceph_pg_mapping, node); | |
408 | c = pgid_cmp(pgid, pg->pgid); | |
8adc8b3d | 409 | if (c < 0) { |
9794b146 | 410 | n = n->rb_left; |
8adc8b3d | 411 | } else if (c > 0) { |
9794b146 | 412 | n = n->rb_right; |
8adc8b3d SW |
413 | } else { |
414 | dout("__lookup_pg_mapping %llx got %p\n", | |
415 | *(u64 *)&pgid, pg); | |
9794b146 | 416 | return pg; |
8adc8b3d | 417 | } |
9794b146 SW |
418 | } |
419 | return NULL; | |
420 | } | |
421 | ||
8adc8b3d SW |
422 | static int __remove_pg_mapping(struct rb_root *root, struct ceph_pg pgid) |
423 | { | |
424 | struct ceph_pg_mapping *pg = __lookup_pg_mapping(root, pgid); | |
425 | ||
426 | if (pg) { | |
427 | dout("__remove_pg_mapping %llx %p\n", *(u64 *)&pgid, pg); | |
428 | rb_erase(&pg->node, root); | |
429 | kfree(pg); | |
430 | return 0; | |
431 | } | |
432 | dout("__remove_pg_mapping %llx dne\n", *(u64 *)&pgid); | |
433 | return -ENOENT; | |
434 | } | |
435 | ||
4fc51be8 SW |
436 | /* |
437 | * rbtree of pg pool info | |
438 | */ | |
439 | static int __insert_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *new) | |
440 | { | |
441 | struct rb_node **p = &root->rb_node; | |
442 | struct rb_node *parent = NULL; | |
443 | struct ceph_pg_pool_info *pi = NULL; | |
444 | ||
445 | while (*p) { | |
446 | parent = *p; | |
447 | pi = rb_entry(parent, struct ceph_pg_pool_info, node); | |
448 | if (new->id < pi->id) | |
449 | p = &(*p)->rb_left; | |
450 | else if (new->id > pi->id) | |
451 | p = &(*p)->rb_right; | |
452 | else | |
453 | return -EEXIST; | |
454 | } | |
455 | ||
456 | rb_link_node(&new->node, parent, p); | |
457 | rb_insert_color(&new->node, root); | |
458 | return 0; | |
459 | } | |
460 | ||
461 | static struct ceph_pg_pool_info *__lookup_pg_pool(struct rb_root *root, int id) | |
462 | { | |
463 | struct ceph_pg_pool_info *pi; | |
464 | struct rb_node *n = root->rb_node; | |
465 | ||
466 | while (n) { | |
467 | pi = rb_entry(n, struct ceph_pg_pool_info, node); | |
468 | if (id < pi->id) | |
469 | n = n->rb_left; | |
470 | else if (id > pi->id) | |
471 | n = n->rb_right; | |
472 | else | |
473 | return pi; | |
474 | } | |
475 | return NULL; | |
476 | } | |
477 | ||
72afc71f AE |
478 | const char *ceph_pg_pool_name_by_id(struct ceph_osdmap *map, u64 id) |
479 | { | |
480 | struct ceph_pg_pool_info *pi; | |
481 | ||
482 | if (id == CEPH_NOPOOL) | |
483 | return NULL; | |
484 | ||
485 | if (WARN_ON_ONCE(id > (u64) INT_MAX)) | |
486 | return NULL; | |
487 | ||
488 | pi = __lookup_pg_pool(&map->pg_pools, (int) id); | |
489 | ||
490 | return pi ? pi->name : NULL; | |
491 | } | |
492 | EXPORT_SYMBOL(ceph_pg_pool_name_by_id); | |
493 | ||
7669a2c9 YS |
494 | int ceph_pg_poolid_by_name(struct ceph_osdmap *map, const char *name) |
495 | { | |
496 | struct rb_node *rbp; | |
497 | ||
498 | for (rbp = rb_first(&map->pg_pools); rbp; rbp = rb_next(rbp)) { | |
499 | struct ceph_pg_pool_info *pi = | |
500 | rb_entry(rbp, struct ceph_pg_pool_info, node); | |
501 | if (pi->name && strcmp(pi->name, name) == 0) | |
502 | return pi->id; | |
503 | } | |
504 | return -ENOENT; | |
505 | } | |
3d14c5d2 | 506 | EXPORT_SYMBOL(ceph_pg_poolid_by_name); |
7669a2c9 | 507 | |
2844a76a SW |
508 | static void __remove_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *pi) |
509 | { | |
510 | rb_erase(&pi->node, root); | |
511 | kfree(pi->name); | |
512 | kfree(pi); | |
513 | } | |
514 | ||
73a7e693 | 515 | static int __decode_pool(void **p, void *end, struct ceph_pg_pool_info *pi) |
efd7576b | 516 | { |
95c96174 | 517 | unsigned int n, m; |
73a7e693 | 518 | |
efd7576b SW |
519 | ceph_decode_copy(p, &pi->v, sizeof(pi->v)); |
520 | calc_pg_masks(pi); | |
73a7e693 SW |
521 | |
522 | /* num_snaps * snap_info_t */ | |
523 | n = le32_to_cpu(pi->v.num_snaps); | |
524 | while (n--) { | |
525 | ceph_decode_need(p, end, sizeof(u64) + 1 + sizeof(u64) + | |
526 | sizeof(struct ceph_timespec), bad); | |
527 | *p += sizeof(u64) + /* key */ | |
528 | 1 + sizeof(u64) + /* u8, snapid */ | |
529 | sizeof(struct ceph_timespec); | |
530 | m = ceph_decode_32(p); /* snap name */ | |
531 | *p += m; | |
532 | } | |
533 | ||
efd7576b | 534 | *p += le32_to_cpu(pi->v.num_removed_snap_intervals) * sizeof(u64) * 2; |
73a7e693 SW |
535 | return 0; |
536 | ||
537 | bad: | |
538 | return -EINVAL; | |
efd7576b SW |
539 | } |
540 | ||
2844a76a SW |
541 | static int __decode_pool_names(void **p, void *end, struct ceph_osdmap *map) |
542 | { | |
543 | struct ceph_pg_pool_info *pi; | |
544 | u32 num, len, pool; | |
545 | ||
546 | ceph_decode_32_safe(p, end, num, bad); | |
547 | dout(" %d pool names\n", num); | |
548 | while (num--) { | |
549 | ceph_decode_32_safe(p, end, pool, bad); | |
550 | ceph_decode_32_safe(p, end, len, bad); | |
551 | dout(" pool %d len %d\n", pool, len); | |
ad3b904c | 552 | ceph_decode_need(p, end, len, bad); |
2844a76a SW |
553 | pi = __lookup_pg_pool(&map->pg_pools, pool); |
554 | if (pi) { | |
ad3b904c XW |
555 | char *name = kstrndup(*p, len, GFP_NOFS); |
556 | ||
557 | if (!name) | |
558 | return -ENOMEM; | |
2844a76a | 559 | kfree(pi->name); |
ad3b904c XW |
560 | pi->name = name; |
561 | dout(" name is %s\n", pi->name); | |
2844a76a SW |
562 | } |
563 | *p += len; | |
564 | } | |
565 | return 0; | |
566 | ||
567 | bad: | |
568 | return -EINVAL; | |
569 | } | |
570 | ||
571 | /* | |
572 | * osd map | |
573 | */ | |
574 | void ceph_osdmap_destroy(struct ceph_osdmap *map) | |
575 | { | |
576 | dout("osdmap_destroy %p\n", map); | |
577 | if (map->crush) | |
578 | crush_destroy(map->crush); | |
579 | while (!RB_EMPTY_ROOT(&map->pg_temp)) { | |
580 | struct ceph_pg_mapping *pg = | |
581 | rb_entry(rb_first(&map->pg_temp), | |
582 | struct ceph_pg_mapping, node); | |
583 | rb_erase(&pg->node, &map->pg_temp); | |
584 | kfree(pg); | |
585 | } | |
586 | while (!RB_EMPTY_ROOT(&map->pg_pools)) { | |
587 | struct ceph_pg_pool_info *pi = | |
588 | rb_entry(rb_first(&map->pg_pools), | |
589 | struct ceph_pg_pool_info, node); | |
590 | __remove_pg_pool(&map->pg_pools, pi); | |
591 | } | |
592 | kfree(map->osd_state); | |
593 | kfree(map->osd_weight); | |
594 | kfree(map->osd_addr); | |
595 | kfree(map); | |
596 | } | |
597 | ||
598 | /* | |
599 | * adjust max osd value. reallocate arrays. | |
600 | */ | |
601 | static int osdmap_set_max_osd(struct ceph_osdmap *map, int max) | |
602 | { | |
603 | u8 *state; | |
604 | struct ceph_entity_addr *addr; | |
605 | u32 *weight; | |
606 | ||
607 | state = kcalloc(max, sizeof(*state), GFP_NOFS); | |
608 | addr = kcalloc(max, sizeof(*addr), GFP_NOFS); | |
609 | weight = kcalloc(max, sizeof(*weight), GFP_NOFS); | |
610 | if (state == NULL || addr == NULL || weight == NULL) { | |
611 | kfree(state); | |
612 | kfree(addr); | |
613 | kfree(weight); | |
614 | return -ENOMEM; | |
615 | } | |
616 | ||
617 | /* copy old? */ | |
618 | if (map->osd_state) { | |
619 | memcpy(state, map->osd_state, map->max_osd*sizeof(*state)); | |
620 | memcpy(addr, map->osd_addr, map->max_osd*sizeof(*addr)); | |
621 | memcpy(weight, map->osd_weight, map->max_osd*sizeof(*weight)); | |
622 | kfree(map->osd_state); | |
623 | kfree(map->osd_addr); | |
624 | kfree(map->osd_weight); | |
625 | } | |
626 | ||
627 | map->osd_state = state; | |
628 | map->osd_weight = weight; | |
629 | map->osd_addr = addr; | |
630 | map->max_osd = max; | |
631 | return 0; | |
632 | } | |
633 | ||
f24e9980 SW |
634 | /* |
635 | * decode a full map. | |
636 | */ | |
637 | struct ceph_osdmap *osdmap_decode(void **p, void *end) | |
638 | { | |
639 | struct ceph_osdmap *map; | |
640 | u16 version; | |
641 | u32 len, max, i; | |
361be860 | 642 | u8 ev; |
f24e9980 SW |
643 | int err = -EINVAL; |
644 | void *start = *p; | |
4fc51be8 | 645 | struct ceph_pg_pool_info *pi; |
f24e9980 SW |
646 | |
647 | dout("osdmap_decode %p to %p len %d\n", *p, end, (int)(end - *p)); | |
648 | ||
649 | map = kzalloc(sizeof(*map), GFP_NOFS); | |
650 | if (map == NULL) | |
651 | return ERR_PTR(-ENOMEM); | |
652 | map->pg_temp = RB_ROOT; | |
653 | ||
654 | ceph_decode_16_safe(p, end, version, bad); | |
02f90c61 SW |
655 | if (version > CEPH_OSDMAP_VERSION) { |
656 | pr_warning("got unknown v %d > %d of osdmap\n", version, | |
657 | CEPH_OSDMAP_VERSION); | |
658 | goto bad; | |
659 | } | |
f24e9980 SW |
660 | |
661 | ceph_decode_need(p, end, 2*sizeof(u64)+6*sizeof(u32), bad); | |
662 | ceph_decode_copy(p, &map->fsid, sizeof(map->fsid)); | |
c89136ea | 663 | map->epoch = ceph_decode_32(p); |
f24e9980 SW |
664 | ceph_decode_copy(p, &map->created, sizeof(map->created)); |
665 | ceph_decode_copy(p, &map->modified, sizeof(map->modified)); | |
666 | ||
f24e9980 SW |
667 | ceph_decode_32_safe(p, end, max, bad); |
668 | while (max--) { | |
4fc51be8 | 669 | ceph_decode_need(p, end, 4 + 1 + sizeof(pi->v), bad); |
0ed7285e | 670 | err = -ENOMEM; |
2844a76a | 671 | pi = kzalloc(sizeof(*pi), GFP_NOFS); |
4fc51be8 | 672 | if (!pi) |
f24e9980 | 673 | goto bad; |
4fc51be8 | 674 | pi->id = ceph_decode_32(p); |
0ed7285e | 675 | err = -EINVAL; |
361be860 | 676 | ev = ceph_decode_8(p); /* encoding version */ |
02f90c61 SW |
677 | if (ev > CEPH_PG_POOL_VERSION) { |
678 | pr_warning("got unknown v %d > %d of ceph_pg_pool\n", | |
679 | ev, CEPH_PG_POOL_VERSION); | |
b0bbb0be | 680 | kfree(pi); |
02f90c61 SW |
681 | goto bad; |
682 | } | |
73a7e693 | 683 | err = __decode_pool(p, end, pi); |
b0aee351 JJ |
684 | if (err < 0) { |
685 | kfree(pi); | |
73a7e693 | 686 | goto bad; |
b0aee351 | 687 | } |
4fc51be8 | 688 | __insert_pg_pool(&map->pg_pools, pi); |
f24e9980 | 689 | } |
2844a76a | 690 | |
0ed7285e SW |
691 | if (version >= 5) { |
692 | err = __decode_pool_names(p, end, map); | |
693 | if (err < 0) { | |
694 | dout("fail to decode pool names"); | |
695 | goto bad; | |
696 | } | |
697 | } | |
2844a76a | 698 | |
4fc51be8 | 699 | ceph_decode_32_safe(p, end, map->pool_max, bad); |
f24e9980 SW |
700 | |
701 | ceph_decode_32_safe(p, end, map->flags, bad); | |
702 | ||
c89136ea | 703 | max = ceph_decode_32(p); |
f24e9980 SW |
704 | |
705 | /* (re)alloc osd arrays */ | |
706 | err = osdmap_set_max_osd(map, max); | |
707 | if (err < 0) | |
708 | goto bad; | |
709 | dout("osdmap_decode max_osd = %d\n", map->max_osd); | |
710 | ||
711 | /* osds */ | |
712 | err = -EINVAL; | |
713 | ceph_decode_need(p, end, 3*sizeof(u32) + | |
714 | map->max_osd*(1 + sizeof(*map->osd_weight) + | |
715 | sizeof(*map->osd_addr)), bad); | |
716 | *p += 4; /* skip length field (should match max) */ | |
717 | ceph_decode_copy(p, map->osd_state, map->max_osd); | |
718 | ||
719 | *p += 4; /* skip length field (should match max) */ | |
720 | for (i = 0; i < map->max_osd; i++) | |
c89136ea | 721 | map->osd_weight[i] = ceph_decode_32(p); |
f24e9980 SW |
722 | |
723 | *p += 4; /* skip length field (should match max) */ | |
724 | ceph_decode_copy(p, map->osd_addr, map->max_osd*sizeof(*map->osd_addr)); | |
63f2d211 SW |
725 | for (i = 0; i < map->max_osd; i++) |
726 | ceph_decode_addr(&map->osd_addr[i]); | |
f24e9980 SW |
727 | |
728 | /* pg_temp */ | |
729 | ceph_decode_32_safe(p, end, len, bad); | |
730 | for (i = 0; i < len; i++) { | |
731 | int n, j; | |
51042122 | 732 | struct ceph_pg pgid; |
f24e9980 SW |
733 | struct ceph_pg_mapping *pg; |
734 | ||
735 | ceph_decode_need(p, end, sizeof(u32) + sizeof(u64), bad); | |
51042122 | 736 | ceph_decode_copy(p, &pgid, sizeof(pgid)); |
c89136ea | 737 | n = ceph_decode_32(p); |
e91a9b63 XW |
738 | err = -EINVAL; |
739 | if (n > (UINT_MAX - sizeof(*pg)) / sizeof(u32)) | |
740 | goto bad; | |
f24e9980 | 741 | ceph_decode_need(p, end, n * sizeof(u32), bad); |
30dc6381 | 742 | err = -ENOMEM; |
f24e9980 | 743 | pg = kmalloc(sizeof(*pg) + n*sizeof(u32), GFP_NOFS); |
30dc6381 | 744 | if (!pg) |
f24e9980 | 745 | goto bad; |
f24e9980 SW |
746 | pg->pgid = pgid; |
747 | pg->len = n; | |
748 | for (j = 0; j < n; j++) | |
c89136ea | 749 | pg->osds[j] = ceph_decode_32(p); |
f24e9980 | 750 | |
991abb6e SW |
751 | err = __insert_pg_mapping(pg, &map->pg_temp); |
752 | if (err) | |
753 | goto bad; | |
51042122 | 754 | dout(" added pg_temp %llx len %d\n", *(u64 *)&pgid, len); |
f24e9980 SW |
755 | } |
756 | ||
757 | /* crush */ | |
758 | ceph_decode_32_safe(p, end, len, bad); | |
759 | dout("osdmap_decode crush len %d from off 0x%x\n", len, | |
760 | (int)(*p - start)); | |
761 | ceph_decode_need(p, end, len, bad); | |
762 | map->crush = crush_decode(*p, end); | |
763 | *p += len; | |
764 | if (IS_ERR(map->crush)) { | |
765 | err = PTR_ERR(map->crush); | |
766 | map->crush = NULL; | |
767 | goto bad; | |
768 | } | |
769 | ||
770 | /* ignore the rest of the map */ | |
771 | *p = end; | |
772 | ||
773 | dout("osdmap_decode done %p %p\n", *p, end); | |
774 | return map; | |
775 | ||
776 | bad: | |
0ed7285e | 777 | dout("osdmap_decode fail err %d\n", err); |
f24e9980 SW |
778 | ceph_osdmap_destroy(map); |
779 | return ERR_PTR(err); | |
780 | } | |
781 | ||
782 | /* | |
783 | * decode and apply an incremental map update. | |
784 | */ | |
785 | struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end, | |
786 | struct ceph_osdmap *map, | |
787 | struct ceph_messenger *msgr) | |
788 | { | |
f24e9980 SW |
789 | struct crush_map *newcrush = NULL; |
790 | struct ceph_fsid fsid; | |
791 | u32 epoch = 0; | |
792 | struct ceph_timespec modified; | |
793 | u32 len, pool; | |
4fc51be8 | 794 | __s32 new_pool_max, new_flags, max; |
f24e9980 SW |
795 | void *start = *p; |
796 | int err = -EINVAL; | |
797 | u16 version; | |
f24e9980 SW |
798 | |
799 | ceph_decode_16_safe(p, end, version, bad); | |
02f90c61 SW |
800 | if (version > CEPH_OSDMAP_INC_VERSION) { |
801 | pr_warning("got unknown v %d > %d of inc osdmap\n", version, | |
802 | CEPH_OSDMAP_INC_VERSION); | |
803 | goto bad; | |
804 | } | |
f24e9980 SW |
805 | |
806 | ceph_decode_need(p, end, sizeof(fsid)+sizeof(modified)+2*sizeof(u32), | |
807 | bad); | |
808 | ceph_decode_copy(p, &fsid, sizeof(fsid)); | |
c89136ea | 809 | epoch = ceph_decode_32(p); |
f24e9980 SW |
810 | BUG_ON(epoch != map->epoch+1); |
811 | ceph_decode_copy(p, &modified, sizeof(modified)); | |
4fc51be8 | 812 | new_pool_max = ceph_decode_32(p); |
c89136ea | 813 | new_flags = ceph_decode_32(p); |
f24e9980 SW |
814 | |
815 | /* full map? */ | |
816 | ceph_decode_32_safe(p, end, len, bad); | |
817 | if (len > 0) { | |
818 | dout("apply_incremental full map len %d, %p to %p\n", | |
819 | len, *p, end); | |
30dc6381 | 820 | return osdmap_decode(p, min(*p+len, end)); |
f24e9980 SW |
821 | } |
822 | ||
823 | /* new crush? */ | |
824 | ceph_decode_32_safe(p, end, len, bad); | |
825 | if (len > 0) { | |
826 | dout("apply_incremental new crush map len %d, %p to %p\n", | |
827 | len, *p, end); | |
828 | newcrush = crush_decode(*p, min(*p+len, end)); | |
829 | if (IS_ERR(newcrush)) | |
7e34bc52 | 830 | return ERR_CAST(newcrush); |
cebc5be6 | 831 | *p += len; |
f24e9980 SW |
832 | } |
833 | ||
834 | /* new flags? */ | |
835 | if (new_flags >= 0) | |
836 | map->flags = new_flags; | |
4fc51be8 SW |
837 | if (new_pool_max >= 0) |
838 | map->pool_max = new_pool_max; | |
f24e9980 SW |
839 | |
840 | ceph_decode_need(p, end, 5*sizeof(u32), bad); | |
841 | ||
842 | /* new max? */ | |
c89136ea | 843 | max = ceph_decode_32(p); |
f24e9980 SW |
844 | if (max >= 0) { |
845 | err = osdmap_set_max_osd(map, max); | |
846 | if (err < 0) | |
847 | goto bad; | |
848 | } | |
849 | ||
850 | map->epoch++; | |
31456665 | 851 | map->modified = modified; |
f24e9980 SW |
852 | if (newcrush) { |
853 | if (map->crush) | |
854 | crush_destroy(map->crush); | |
855 | map->crush = newcrush; | |
856 | newcrush = NULL; | |
857 | } | |
858 | ||
859 | /* new_pool */ | |
860 | ceph_decode_32_safe(p, end, len, bad); | |
861 | while (len--) { | |
361be860 | 862 | __u8 ev; |
4fc51be8 | 863 | struct ceph_pg_pool_info *pi; |
361be860 | 864 | |
f24e9980 | 865 | ceph_decode_32_safe(p, end, pool, bad); |
4fc51be8 | 866 | ceph_decode_need(p, end, 1 + sizeof(pi->v), bad); |
361be860 | 867 | ev = ceph_decode_8(p); /* encoding version */ |
02f90c61 SW |
868 | if (ev > CEPH_PG_POOL_VERSION) { |
869 | pr_warning("got unknown v %d > %d of ceph_pg_pool\n", | |
870 | ev, CEPH_PG_POOL_VERSION); | |
0ed7285e | 871 | err = -EINVAL; |
02f90c61 SW |
872 | goto bad; |
873 | } | |
4fc51be8 SW |
874 | pi = __lookup_pg_pool(&map->pg_pools, pool); |
875 | if (!pi) { | |
2844a76a | 876 | pi = kzalloc(sizeof(*pi), GFP_NOFS); |
4fc51be8 SW |
877 | if (!pi) { |
878 | err = -ENOMEM; | |
879 | goto bad; | |
880 | } | |
881 | pi->id = pool; | |
882 | __insert_pg_pool(&map->pg_pools, pi); | |
883 | } | |
73a7e693 SW |
884 | err = __decode_pool(p, end, pi); |
885 | if (err < 0) | |
886 | goto bad; | |
f24e9980 | 887 | } |
0ed7285e SW |
888 | if (version >= 5) { |
889 | err = __decode_pool_names(p, end, map); | |
890 | if (err < 0) | |
891 | goto bad; | |
892 | } | |
f24e9980 | 893 | |
4fc51be8 | 894 | /* old_pool */ |
f24e9980 | 895 | ceph_decode_32_safe(p, end, len, bad); |
4fc51be8 SW |
896 | while (len--) { |
897 | struct ceph_pg_pool_info *pi; | |
898 | ||
899 | ceph_decode_32_safe(p, end, pool, bad); | |
900 | pi = __lookup_pg_pool(&map->pg_pools, pool); | |
2844a76a SW |
901 | if (pi) |
902 | __remove_pg_pool(&map->pg_pools, pi); | |
4fc51be8 | 903 | } |
f24e9980 SW |
904 | |
905 | /* new_up */ | |
906 | err = -EINVAL; | |
907 | ceph_decode_32_safe(p, end, len, bad); | |
908 | while (len--) { | |
909 | u32 osd; | |
910 | struct ceph_entity_addr addr; | |
911 | ceph_decode_32_safe(p, end, osd, bad); | |
912 | ceph_decode_copy_safe(p, end, &addr, sizeof(addr), bad); | |
63f2d211 | 913 | ceph_decode_addr(&addr); |
f24e9980 SW |
914 | pr_info("osd%d up\n", osd); |
915 | BUG_ON(osd >= map->max_osd); | |
916 | map->osd_state[osd] |= CEPH_OSD_UP; | |
917 | map->osd_addr[osd] = addr; | |
918 | } | |
919 | ||
7662d8ff | 920 | /* new_state */ |
f24e9980 SW |
921 | ceph_decode_32_safe(p, end, len, bad); |
922 | while (len--) { | |
923 | u32 osd; | |
7662d8ff | 924 | u8 xorstate; |
f24e9980 | 925 | ceph_decode_32_safe(p, end, osd, bad); |
7662d8ff | 926 | xorstate = **(u8 **)p; |
f24e9980 | 927 | (*p)++; /* clean flag */ |
7662d8ff SW |
928 | if (xorstate == 0) |
929 | xorstate = CEPH_OSD_UP; | |
930 | if (xorstate & CEPH_OSD_UP) | |
931 | pr_info("osd%d down\n", osd); | |
f24e9980 | 932 | if (osd < map->max_osd) |
7662d8ff | 933 | map->osd_state[osd] ^= xorstate; |
f24e9980 SW |
934 | } |
935 | ||
936 | /* new_weight */ | |
937 | ceph_decode_32_safe(p, end, len, bad); | |
938 | while (len--) { | |
939 | u32 osd, off; | |
940 | ceph_decode_need(p, end, sizeof(u32)*2, bad); | |
c89136ea SW |
941 | osd = ceph_decode_32(p); |
942 | off = ceph_decode_32(p); | |
f24e9980 SW |
943 | pr_info("osd%d weight 0x%x %s\n", osd, off, |
944 | off == CEPH_OSD_IN ? "(in)" : | |
945 | (off == CEPH_OSD_OUT ? "(out)" : "")); | |
946 | if (osd < map->max_osd) | |
947 | map->osd_weight[osd] = off; | |
948 | } | |
949 | ||
950 | /* new_pg_temp */ | |
f24e9980 SW |
951 | ceph_decode_32_safe(p, end, len, bad); |
952 | while (len--) { | |
953 | struct ceph_pg_mapping *pg; | |
954 | int j; | |
51042122 | 955 | struct ceph_pg pgid; |
f24e9980 SW |
956 | u32 pglen; |
957 | ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad); | |
51042122 | 958 | ceph_decode_copy(p, &pgid, sizeof(pgid)); |
c89136ea | 959 | pglen = ceph_decode_32(p); |
f24e9980 | 960 | |
f24e9980 | 961 | if (pglen) { |
f24e9980 | 962 | ceph_decode_need(p, end, pglen*sizeof(u32), bad); |
6bd9adbd SW |
963 | |
964 | /* removing existing (if any) */ | |
965 | (void) __remove_pg_mapping(&map->pg_temp, pgid); | |
966 | ||
967 | /* insert */ | |
0ed7285e SW |
968 | err = -EINVAL; |
969 | if (pglen > (UINT_MAX - sizeof(*pg)) / sizeof(u32)) | |
a5506049 | 970 | goto bad; |
0ed7285e | 971 | err = -ENOMEM; |
f24e9980 | 972 | pg = kmalloc(sizeof(*pg) + sizeof(u32)*pglen, GFP_NOFS); |
0ed7285e | 973 | if (!pg) |
f24e9980 | 974 | goto bad; |
f24e9980 SW |
975 | pg->pgid = pgid; |
976 | pg->len = pglen; | |
7067f797 | 977 | for (j = 0; j < pglen; j++) |
c89136ea | 978 | pg->osds[j] = ceph_decode_32(p); |
991abb6e | 979 | err = __insert_pg_mapping(pg, &map->pg_temp); |
bc4fdca8 SW |
980 | if (err) { |
981 | kfree(pg); | |
991abb6e | 982 | goto bad; |
bc4fdca8 | 983 | } |
51042122 SW |
984 | dout(" added pg_temp %llx len %d\n", *(u64 *)&pgid, |
985 | pglen); | |
8adc8b3d SW |
986 | } else { |
987 | /* remove */ | |
988 | __remove_pg_mapping(&map->pg_temp, pgid); | |
f24e9980 SW |
989 | } |
990 | } | |
f24e9980 SW |
991 | |
992 | /* ignore the rest */ | |
993 | *p = end; | |
994 | return map; | |
995 | ||
996 | bad: | |
997 | pr_err("corrupt inc osdmap epoch %d off %d (%p of %p-%p)\n", | |
998 | epoch, (int)(*p - start), *p, start, end); | |
9ec7cab1 SW |
999 | print_hex_dump(KERN_DEBUG, "osdmap: ", |
1000 | DUMP_PREFIX_OFFSET, 16, 1, | |
1001 | start, end - start, true); | |
f24e9980 SW |
1002 | if (newcrush) |
1003 | crush_destroy(newcrush); | |
1004 | return ERR_PTR(err); | |
1005 | } | |
1006 | ||
1007 | ||
1008 | ||
1009 | ||
1010 | /* | |
1011 | * calculate file layout from given offset, length. | |
1012 | * fill in correct oid, logical length, and object extent | |
1013 | * offset, length. | |
1014 | * | |
1015 | * for now, we write only a single su, until we can | |
1016 | * pass a stride back to the caller. | |
1017 | */ | |
d63b77f4 | 1018 | int ceph_calc_file_object_mapping(struct ceph_file_layout *layout, |
e8afad65 | 1019 | u64 off, u64 len, |
645a1025 | 1020 | u64 *ono, |
f24e9980 SW |
1021 | u64 *oxoff, u64 *oxlen) |
1022 | { | |
1023 | u32 osize = le32_to_cpu(layout->fl_object_size); | |
1024 | u32 su = le32_to_cpu(layout->fl_stripe_unit); | |
1025 | u32 sc = le32_to_cpu(layout->fl_stripe_count); | |
1026 | u32 bl, stripeno, stripepos, objsetno; | |
1027 | u32 su_per_object; | |
ff1d1f71 | 1028 | u64 t, su_offset; |
f24e9980 | 1029 | |
e8afad65 | 1030 | dout("mapping %llu~%llu osize %u fl_su %u\n", off, len, |
f24e9980 | 1031 | osize, su); |
d63b77f4 SW |
1032 | if (su == 0 || sc == 0) |
1033 | goto invalid; | |
35e054a6 | 1034 | su_per_object = osize / su; |
d63b77f4 SW |
1035 | if (su_per_object == 0) |
1036 | goto invalid; | |
f24e9980 SW |
1037 | dout("osize %u / su %u = su_per_object %u\n", osize, su, |
1038 | su_per_object); | |
1039 | ||
d63b77f4 SW |
1040 | if ((su & ~PAGE_MASK) != 0) |
1041 | goto invalid; | |
1042 | ||
f24e9980 SW |
1043 | /* bl = *off / su; */ |
1044 | t = off; | |
1045 | do_div(t, su); | |
1046 | bl = t; | |
1047 | dout("off %llu / su %u = bl %u\n", off, su, bl); | |
1048 | ||
1049 | stripeno = bl / sc; | |
1050 | stripepos = bl % sc; | |
1051 | objsetno = stripeno / su_per_object; | |
1052 | ||
645a1025 | 1053 | *ono = objsetno * sc + stripepos; |
95c96174 | 1054 | dout("objset %u * sc %u = ono %u\n", objsetno, sc, (unsigned int)*ono); |
645a1025 SW |
1055 | |
1056 | /* *oxoff = *off % layout->fl_stripe_unit; # offset in su */ | |
f24e9980 | 1057 | t = off; |
ff1d1f71 NW |
1058 | su_offset = do_div(t, su); |
1059 | *oxoff = su_offset + (stripeno % su_per_object) * su; | |
1060 | ||
1061 | /* | |
1062 | * Calculate the length of the extent being written to the selected | |
e8afad65 | 1063 | * object. This is the minimum of the full length requested (len) or |
ff1d1f71 NW |
1064 | * the remainder of the current stripe being written to. |
1065 | */ | |
e8afad65 | 1066 | *oxlen = min_t(u64, len, su - su_offset); |
f24e9980 SW |
1067 | |
1068 | dout(" obj extent %llu~%llu\n", *oxoff, *oxlen); | |
d63b77f4 SW |
1069 | return 0; |
1070 | ||
1071 | invalid: | |
1072 | dout(" invalid layout\n"); | |
1073 | *ono = 0; | |
1074 | *oxoff = 0; | |
1075 | *oxlen = 0; | |
1076 | return -EINVAL; | |
f24e9980 | 1077 | } |
3d14c5d2 | 1078 | EXPORT_SYMBOL(ceph_calc_file_object_mapping); |
f24e9980 SW |
1079 | |
1080 | /* | |
1081 | * calculate an object layout (i.e. pgid) from an oid, | |
1082 | * file_layout, and osdmap | |
1083 | */ | |
1084 | int ceph_calc_object_layout(struct ceph_object_layout *ol, | |
1085 | const char *oid, | |
1086 | struct ceph_file_layout *fl, | |
1087 | struct ceph_osdmap *osdmap) | |
1088 | { | |
95c96174 | 1089 | unsigned int num, num_mask; |
51042122 | 1090 | struct ceph_pg pgid; |
f24e9980 SW |
1091 | int poolid = le32_to_cpu(fl->fl_pg_pool); |
1092 | struct ceph_pg_pool_info *pool; | |
95c96174 | 1093 | unsigned int ps; |
f24e9980 | 1094 | |
30dc6381 | 1095 | BUG_ON(!osdmap); |
f24e9980 | 1096 | |
4fc51be8 SW |
1097 | pool = __lookup_pg_pool(&osdmap->pg_pools, poolid); |
1098 | if (!pool) | |
1099 | return -EIO; | |
1654dd0c | 1100 | ps = ceph_str_hash(pool->v.object_hash, oid, strlen(oid)); |
3469ac1a SW |
1101 | num = le32_to_cpu(pool->v.pg_num); |
1102 | num_mask = pool->pg_num_mask; | |
f24e9980 | 1103 | |
51042122 | 1104 | pgid.ps = cpu_to_le16(ps); |
3469ac1a | 1105 | pgid.preferred = cpu_to_le16(-1); |
51042122 | 1106 | pgid.pool = fl->fl_pg_pool; |
3469ac1a | 1107 | dout("calc_object_layout '%s' pgid %d.%x\n", oid, poolid, ps); |
f24e9980 | 1108 | |
51042122 | 1109 | ol->ol_pgid = pgid; |
f24e9980 | 1110 | ol->ol_stripe_unit = fl->fl_object_stripe_unit; |
f24e9980 SW |
1111 | return 0; |
1112 | } | |
3d14c5d2 | 1113 | EXPORT_SYMBOL(ceph_calc_object_layout); |
f24e9980 SW |
1114 | |
1115 | /* | |
1116 | * Calculate raw osd vector for the given pgid. Return pointer to osd | |
1117 | * array, or NULL on failure. | |
1118 | */ | |
51042122 | 1119 | static int *calc_pg_raw(struct ceph_osdmap *osdmap, struct ceph_pg pgid, |
f24e9980 SW |
1120 | int *osds, int *num) |
1121 | { | |
f24e9980 SW |
1122 | struct ceph_pg_mapping *pg; |
1123 | struct ceph_pg_pool_info *pool; | |
1124 | int ruleno; | |
af56e0aa | 1125 | unsigned int poolid, ps, pps, t, r; |
f24e9980 | 1126 | |
782e182e SW |
1127 | poolid = le32_to_cpu(pgid.pool); |
1128 | ps = le16_to_cpu(pgid.ps); | |
782e182e SW |
1129 | |
1130 | pool = __lookup_pg_pool(&osdmap->pg_pools, poolid); | |
1131 | if (!pool) | |
1132 | return NULL; | |
1133 | ||
f24e9980 | 1134 | /* pg_temp? */ |
3469ac1a SW |
1135 | t = ceph_stable_mod(ps, le32_to_cpu(pool->v.pg_num), |
1136 | pool->pgp_num_mask); | |
782e182e | 1137 | pgid.ps = cpu_to_le16(t); |
9794b146 SW |
1138 | pg = __lookup_pg_mapping(&osdmap->pg_temp, pgid); |
1139 | if (pg) { | |
1140 | *num = pg->len; | |
1141 | return pg->osds; | |
f24e9980 SW |
1142 | } |
1143 | ||
1144 | /* crush */ | |
f24e9980 SW |
1145 | ruleno = crush_find_rule(osdmap->crush, pool->v.crush_ruleset, |
1146 | pool->v.type, pool->v.size); | |
1147 | if (ruleno < 0) { | |
effcb9ed SW |
1148 | pr_err("no crush rule pool %d ruleset %d type %d size %d\n", |
1149 | poolid, pool->v.crush_ruleset, pool->v.type, | |
1150 | pool->v.size); | |
f24e9980 SW |
1151 | return NULL; |
1152 | } | |
1153 | ||
3469ac1a SW |
1154 | pps = ceph_stable_mod(ps, |
1155 | le32_to_cpu(pool->v.pgp_num), | |
1156 | pool->pgp_num_mask); | |
51042122 | 1157 | pps += poolid; |
8b393269 SW |
1158 | r = crush_do_rule(osdmap->crush, ruleno, pps, osds, |
1159 | min_t(int, pool->v.size, *num), | |
1160 | osdmap->osd_weight); | |
1161 | if (r < 0) { | |
1162 | pr_err("error %d from crush rule: pool %d ruleset %d type %d" | |
1163 | " size %d\n", r, poolid, pool->v.crush_ruleset, | |
1164 | pool->v.type, pool->v.size); | |
1165 | return NULL; | |
1166 | } | |
1167 | *num = r; | |
f24e9980 SW |
1168 | return osds; |
1169 | } | |
1170 | ||
d85b7056 SW |
1171 | /* |
1172 | * Return acting set for given pgid. | |
1173 | */ | |
1174 | int ceph_calc_pg_acting(struct ceph_osdmap *osdmap, struct ceph_pg pgid, | |
1175 | int *acting) | |
1176 | { | |
1177 | int rawosds[CEPH_PG_MAX_SIZE], *osds; | |
1178 | int i, o, num = CEPH_PG_MAX_SIZE; | |
1179 | ||
1180 | osds = calc_pg_raw(osdmap, pgid, rawosds, &num); | |
1181 | if (!osds) | |
1182 | return -1; | |
1183 | ||
1184 | /* primary is first up osd */ | |
1185 | o = 0; | |
1186 | for (i = 0; i < num; i++) | |
1187 | if (ceph_osd_is_up(osdmap, osds[i])) | |
1188 | acting[o++] = osds[i]; | |
1189 | return o; | |
1190 | } | |
1191 | ||
f24e9980 SW |
1192 | /* |
1193 | * Return primary osd for given pgid, or -1 if none. | |
1194 | */ | |
51042122 | 1195 | int ceph_calc_pg_primary(struct ceph_osdmap *osdmap, struct ceph_pg pgid) |
f24e9980 | 1196 | { |
d85b7056 SW |
1197 | int rawosds[CEPH_PG_MAX_SIZE], *osds; |
1198 | int i, num = CEPH_PG_MAX_SIZE; | |
f24e9980 SW |
1199 | |
1200 | osds = calc_pg_raw(osdmap, pgid, rawosds, &num); | |
1201 | if (!osds) | |
1202 | return -1; | |
1203 | ||
1204 | /* primary is first up osd */ | |
1205 | for (i = 0; i < num; i++) | |
d85b7056 | 1206 | if (ceph_osd_is_up(osdmap, osds[i])) |
f24e9980 | 1207 | return osds[i]; |
f24e9980 SW |
1208 | return -1; |
1209 | } | |
3d14c5d2 | 1210 | EXPORT_SYMBOL(ceph_calc_pg_primary); |