blob: 243574c8cf33807fcaf9374530358f1e44080764 [file] [log] [blame]
Sage Weilf24e9982009-10-06 11:31:10 -07001
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07002#include <linux/ceph/ceph_debug.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09003
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07004#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09005#include <linux/slab.h>
Sage Weilf24e9982009-10-06 11:31:10 -07006#include <asm/div64.h>
7
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07008#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>
Sage Weilf24e9982009-10-06 11:31:10 -070013
14char *ceph_osdmap_state_str(char *str, int len, int state)
15{
Sage Weilf24e9982009-10-06 11:31:10 -070016 if (!len)
Cong Ding1ec39112013-01-25 17:48:59 -060017 return str;
Sage Weilf24e9982009-10-06 11:31:10 -070018
Cong Ding1ec39112013-01-25 17:48:59 -060019 if ((state & CEPH_OSD_EXISTS) && (state & CEPH_OSD_UP))
20 snprintf(str, len, "exists, up");
21 else if (state & CEPH_OSD_EXISTS)
22 snprintf(str, len, "exists");
23 else if (state & CEPH_OSD_UP)
24 snprintf(str, len, "up");
25 else
Sage Weilf24e9982009-10-06 11:31:10 -070026 snprintf(str, len, "doesn't exist");
Cong Ding1ec39112013-01-25 17:48:59 -060027
Sage Weilf24e9982009-10-06 11:31:10 -070028 return str;
29}
30
31/* maps */
32
Eric Dumazet95c96172012-04-15 05:58:06 +000033static int calc_bits_of(unsigned int t)
Sage Weilf24e9982009-10-06 11:31:10 -070034{
35 int b = 0;
36 while (t) {
37 t = t >> 1;
38 b++;
39 }
40 return b;
41}
42
43/*
44 * the foo_mask is the smallest value 2^n-1 that is >= foo.
45 */
46static void calc_pg_masks(struct ceph_pg_pool_info *pi)
47{
Sage Weil4f6a7e52013-02-23 10:41:09 -080048 pi->pg_num_mask = (1 << calc_bits_of(pi->pg_num-1)) - 1;
49 pi->pgp_num_mask = (1 << calc_bits_of(pi->pgp_num-1)) - 1;
Sage Weilf24e9982009-10-06 11:31:10 -070050}
51
52/*
53 * decode crush map
54 */
55static int crush_decode_uniform_bucket(void **p, void *end,
56 struct crush_bucket_uniform *b)
57{
58 dout("crush_decode_uniform_bucket %p to %p\n", *p, end);
59 ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad);
Sage Weilc89136e2009-10-14 09:59:09 -070060 b->item_weight = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -070061 return 0;
62bad:
63 return -EINVAL;
64}
65
66static int crush_decode_list_bucket(void **p, void *end,
67 struct crush_bucket_list *b)
68{
69 int j;
70 dout("crush_decode_list_bucket %p to %p\n", *p, end);
71 b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
72 if (b->item_weights == NULL)
73 return -ENOMEM;
74 b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
75 if (b->sum_weights == NULL)
76 return -ENOMEM;
77 ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
78 for (j = 0; j < b->h.size; j++) {
Sage Weilc89136e2009-10-14 09:59:09 -070079 b->item_weights[j] = ceph_decode_32(p);
80 b->sum_weights[j] = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -070081 }
82 return 0;
83bad:
84 return -EINVAL;
85}
86
87static int crush_decode_tree_bucket(void **p, void *end,
88 struct crush_bucket_tree *b)
89{
90 int j;
91 dout("crush_decode_tree_bucket %p to %p\n", *p, end);
Ilya Dryomov82cd0032015-06-29 19:30:23 +030092 ceph_decode_8_safe(p, end, b->num_nodes, bad);
Sage Weilf24e9982009-10-06 11:31:10 -070093 b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS);
94 if (b->node_weights == NULL)
95 return -ENOMEM;
96 ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad);
97 for (j = 0; j < b->num_nodes; j++)
Sage Weilc89136e2009-10-14 09:59:09 -070098 b->node_weights[j] = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -070099 return 0;
100bad:
101 return -EINVAL;
102}
103
104static int crush_decode_straw_bucket(void **p, void *end,
105 struct crush_bucket_straw *b)
106{
107 int j;
108 dout("crush_decode_straw_bucket %p to %p\n", *p, end);
109 b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
110 if (b->item_weights == NULL)
111 return -ENOMEM;
112 b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
113 if (b->straws == NULL)
114 return -ENOMEM;
115 ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
116 for (j = 0; j < b->h.size; j++) {
Sage Weilc89136e2009-10-14 09:59:09 -0700117 b->item_weights[j] = ceph_decode_32(p);
118 b->straws[j] = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700119 }
120 return 0;
121bad:
122 return -EINVAL;
123}
124
Ilya Dryomov958a2762015-04-14 16:54:52 +0300125static int crush_decode_straw2_bucket(void **p, void *end,
126 struct crush_bucket_straw2 *b)
127{
128 int j;
129 dout("crush_decode_straw2_bucket %p to %p\n", *p, end);
130 b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
131 if (b->item_weights == NULL)
132 return -ENOMEM;
133 ceph_decode_need(p, end, b->h.size * sizeof(u32), bad);
134 for (j = 0; j < b->h.size; j++)
135 b->item_weights[j] = ceph_decode_32(p);
136 return 0;
137bad:
138 return -EINVAL;
139}
140
Sage Weil546f04e2012-07-30 18:15:23 -0700141static int skip_name_map(void **p, void *end)
142{
143 int len;
144 ceph_decode_32_safe(p, end, len ,bad);
145 while (len--) {
146 int strlen;
147 *p += sizeof(u32);
148 ceph_decode_32_safe(p, end, strlen, bad);
149 *p += strlen;
150}
151 return 0;
152bad:
153 return -EINVAL;
154}
155
Sage Weilf24e9982009-10-06 11:31:10 -0700156static struct crush_map *crush_decode(void *pbyval, void *end)
157{
158 struct crush_map *c;
159 int err = -EINVAL;
160 int i, j;
161 void **p = &pbyval;
162 void *start = pbyval;
163 u32 magic;
Sage Weil546f04e2012-07-30 18:15:23 -0700164 u32 num_name_maps;
Sage Weilf24e9982009-10-06 11:31:10 -0700165
166 dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p));
167
168 c = kzalloc(sizeof(*c), GFP_NOFS);
169 if (c == NULL)
170 return ERR_PTR(-ENOMEM);
171
Sage Weil546f04e2012-07-30 18:15:23 -0700172 /* set tunables to default values */
173 c->choose_local_tries = 2;
174 c->choose_local_fallback_tries = 5;
175 c->choose_total_tries = 19;
Jim Schutt1604f482012-11-30 09:15:25 -0700176 c->chooseleaf_descend_once = 0;
Sage Weil546f04e2012-07-30 18:15:23 -0700177
Sage Weilf24e9982009-10-06 11:31:10 -0700178 ceph_decode_need(p, end, 4*sizeof(u32), bad);
Sage Weilc89136e2009-10-14 09:59:09 -0700179 magic = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700180 if (magic != CRUSH_MAGIC) {
181 pr_err("crush_decode magic %x != current %x\n",
Eric Dumazet95c96172012-04-15 05:58:06 +0000182 (unsigned int)magic, (unsigned int)CRUSH_MAGIC);
Sage Weilf24e9982009-10-06 11:31:10 -0700183 goto bad;
184 }
Sage Weilc89136e2009-10-14 09:59:09 -0700185 c->max_buckets = ceph_decode_32(p);
186 c->max_rules = ceph_decode_32(p);
187 c->max_devices = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700188
Sage Weilf24e9982009-10-06 11:31:10 -0700189 c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS);
190 if (c->buckets == NULL)
191 goto badmem;
192 c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS);
193 if (c->rules == NULL)
194 goto badmem;
195
196 /* buckets */
197 for (i = 0; i < c->max_buckets; i++) {
198 int size = 0;
199 u32 alg;
200 struct crush_bucket *b;
201
202 ceph_decode_32_safe(p, end, alg, bad);
203 if (alg == 0) {
204 c->buckets[i] = NULL;
205 continue;
206 }
207 dout("crush_decode bucket %d off %x %p to %p\n",
208 i, (int)(*p-start), *p, end);
209
210 switch (alg) {
211 case CRUSH_BUCKET_UNIFORM:
212 size = sizeof(struct crush_bucket_uniform);
213 break;
214 case CRUSH_BUCKET_LIST:
215 size = sizeof(struct crush_bucket_list);
216 break;
217 case CRUSH_BUCKET_TREE:
218 size = sizeof(struct crush_bucket_tree);
219 break;
220 case CRUSH_BUCKET_STRAW:
221 size = sizeof(struct crush_bucket_straw);
222 break;
Ilya Dryomov958a2762015-04-14 16:54:52 +0300223 case CRUSH_BUCKET_STRAW2:
224 size = sizeof(struct crush_bucket_straw2);
225 break;
Sage Weilf24e9982009-10-06 11:31:10 -0700226 default:
Sage Weil30dc6382009-12-21 14:49:37 -0800227 err = -EINVAL;
Sage Weilf24e9982009-10-06 11:31:10 -0700228 goto bad;
229 }
230 BUG_ON(size == 0);
231 b = c->buckets[i] = kzalloc(size, GFP_NOFS);
232 if (b == NULL)
233 goto badmem;
234
235 ceph_decode_need(p, end, 4*sizeof(u32), bad);
Sage Weilc89136e2009-10-14 09:59:09 -0700236 b->id = ceph_decode_32(p);
237 b->type = ceph_decode_16(p);
Sage Weilfb690392009-11-07 20:18:22 -0800238 b->alg = ceph_decode_8(p);
239 b->hash = ceph_decode_8(p);
Sage Weilc89136e2009-10-14 09:59:09 -0700240 b->weight = ceph_decode_32(p);
241 b->size = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700242
243 dout("crush_decode bucket size %d off %x %p to %p\n",
244 b->size, (int)(*p-start), *p, end);
245
246 b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS);
247 if (b->items == NULL)
248 goto badmem;
249 b->perm = kcalloc(b->size, sizeof(u32), GFP_NOFS);
250 if (b->perm == NULL)
251 goto badmem;
252 b->perm_n = 0;
253
254 ceph_decode_need(p, end, b->size*sizeof(u32), bad);
255 for (j = 0; j < b->size; j++)
Sage Weilc89136e2009-10-14 09:59:09 -0700256 b->items[j] = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700257
258 switch (b->alg) {
259 case CRUSH_BUCKET_UNIFORM:
260 err = crush_decode_uniform_bucket(p, end,
261 (struct crush_bucket_uniform *)b);
262 if (err < 0)
263 goto bad;
264 break;
265 case CRUSH_BUCKET_LIST:
266 err = crush_decode_list_bucket(p, end,
267 (struct crush_bucket_list *)b);
268 if (err < 0)
269 goto bad;
270 break;
271 case CRUSH_BUCKET_TREE:
272 err = crush_decode_tree_bucket(p, end,
273 (struct crush_bucket_tree *)b);
274 if (err < 0)
275 goto bad;
276 break;
277 case CRUSH_BUCKET_STRAW:
278 err = crush_decode_straw_bucket(p, end,
279 (struct crush_bucket_straw *)b);
280 if (err < 0)
281 goto bad;
282 break;
Ilya Dryomov958a2762015-04-14 16:54:52 +0300283 case CRUSH_BUCKET_STRAW2:
284 err = crush_decode_straw2_bucket(p, end,
285 (struct crush_bucket_straw2 *)b);
286 if (err < 0)
287 goto bad;
288 break;
Sage Weilf24e9982009-10-06 11:31:10 -0700289 }
290 }
291
292 /* rules */
293 dout("rule vec is %p\n", c->rules);
294 for (i = 0; i < c->max_rules; i++) {
295 u32 yes;
296 struct crush_rule *r;
297
298 ceph_decode_32_safe(p, end, yes, bad);
299 if (!yes) {
300 dout("crush_decode NO rule %d off %x %p to %p\n",
301 i, (int)(*p-start), *p, end);
302 c->rules[i] = NULL;
303 continue;
304 }
305
306 dout("crush_decode rule %d off %x %p to %p\n",
307 i, (int)(*p-start), *p, end);
308
309 /* len */
310 ceph_decode_32_safe(p, end, yes, bad);
311#if BITS_PER_LONG == 32
Sage Weil30dc6382009-12-21 14:49:37 -0800312 err = -EINVAL;
Xi Wang64486692012-02-16 11:55:48 -0500313 if (yes > (ULONG_MAX - sizeof(*r))
314 / sizeof(struct crush_rule_step))
Sage Weilf24e9982009-10-06 11:31:10 -0700315 goto bad;
316#endif
317 r = c->rules[i] = kmalloc(sizeof(*r) +
318 yes*sizeof(struct crush_rule_step),
319 GFP_NOFS);
320 if (r == NULL)
321 goto badmem;
322 dout(" rule %d is at %p\n", i, r);
323 r->len = yes;
324 ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */
325 ceph_decode_need(p, end, r->len*3*sizeof(u32), bad);
326 for (j = 0; j < r->len; j++) {
Sage Weilc89136e2009-10-14 09:59:09 -0700327 r->steps[j].op = ceph_decode_32(p);
328 r->steps[j].arg1 = ceph_decode_32(p);
329 r->steps[j].arg2 = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -0700330 }
331 }
332
333 /* ignore trailing name maps. */
Sage Weil546f04e2012-07-30 18:15:23 -0700334 for (num_name_maps = 0; num_name_maps < 3; num_name_maps++) {
335 err = skip_name_map(p, end);
336 if (err < 0)
337 goto done;
338 }
Sage Weilf24e9982009-10-06 11:31:10 -0700339
Sage Weil546f04e2012-07-30 18:15:23 -0700340 /* tunables */
341 ceph_decode_need(p, end, 3*sizeof(u32), done);
342 c->choose_local_tries = ceph_decode_32(p);
343 c->choose_local_fallback_tries = ceph_decode_32(p);
344 c->choose_total_tries = ceph_decode_32(p);
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100345 dout("crush decode tunable choose_local_tries = %d\n",
Sage Weil546f04e2012-07-30 18:15:23 -0700346 c->choose_local_tries);
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100347 dout("crush decode tunable choose_local_fallback_tries = %d\n",
Sage Weil546f04e2012-07-30 18:15:23 -0700348 c->choose_local_fallback_tries);
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100349 dout("crush decode tunable choose_total_tries = %d\n",
Sage Weil546f04e2012-07-30 18:15:23 -0700350 c->choose_total_tries);
351
Jim Schutt1604f482012-11-30 09:15:25 -0700352 ceph_decode_need(p, end, sizeof(u32), done);
353 c->chooseleaf_descend_once = ceph_decode_32(p);
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100354 dout("crush decode tunable chooseleaf_descend_once = %d\n",
Jim Schutt1604f482012-11-30 09:15:25 -0700355 c->chooseleaf_descend_once);
356
Ilya Dryomovf1406622014-05-09 18:27:34 +0400357 ceph_decode_need(p, end, sizeof(u8), done);
358 c->chooseleaf_vary_r = ceph_decode_8(p);
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100359 dout("crush decode tunable chooseleaf_vary_r = %d\n",
Ilya Dryomovf1406622014-05-09 18:27:34 +0400360 c->chooseleaf_vary_r);
361
Ilya Dryomovb9b519b2016-02-01 16:57:16 +0100362 /* skip straw_calc_version, allowed_bucket_algs */
363 ceph_decode_need(p, end, sizeof(u8) + sizeof(u32), done);
364 *p += sizeof(u8) + sizeof(u32);
365
366 ceph_decode_need(p, end, sizeof(u8), done);
367 c->chooseleaf_stable = ceph_decode_8(p);
368 dout("crush decode tunable chooseleaf_stable = %d\n",
369 c->chooseleaf_stable);
370
Sage Weil546f04e2012-07-30 18:15:23 -0700371done:
Sage Weilf24e9982009-10-06 11:31:10 -0700372 dout("crush_decode success\n");
373 return c;
374
375badmem:
376 err = -ENOMEM;
377bad:
378 dout("crush_decode fail %d\n", err);
379 crush_destroy(c);
380 return ERR_PTR(err);
381}
382
Sage Weilf24e9982009-10-06 11:31:10 -0700383/*
Sage Weil9794b142010-02-16 15:53:32 -0800384 * rbtree of pg_mapping for handling pg_temp (explicit mapping of pgid
Ilya Dryomov9686f942014-03-21 19:05:29 +0200385 * to a set of osds) and primary_temp (explicit primary setting)
Sage Weilf24e9982009-10-06 11:31:10 -0700386 */
Sage Weil5b191d92013-02-23 10:38:16 -0800387static int pgid_cmp(struct ceph_pg l, struct ceph_pg r)
Sage Weil51042122009-11-04 11:39:12 -0800388{
Sage Weil5b191d92013-02-23 10:38:16 -0800389 if (l.pool < r.pool)
Sage Weil51042122009-11-04 11:39:12 -0800390 return -1;
Sage Weil5b191d92013-02-23 10:38:16 -0800391 if (l.pool > r.pool)
392 return 1;
393 if (l.seed < r.seed)
394 return -1;
395 if (l.seed > r.seed)
Sage Weil51042122009-11-04 11:39:12 -0800396 return 1;
397 return 0;
398}
399
Sage Weil991abb62009-10-08 22:22:37 -0700400static int __insert_pg_mapping(struct ceph_pg_mapping *new,
401 struct rb_root *root)
Sage Weilf24e9982009-10-06 11:31:10 -0700402{
403 struct rb_node **p = &root->rb_node;
404 struct rb_node *parent = NULL;
405 struct ceph_pg_mapping *pg = NULL;
Sage Weil51042122009-11-04 11:39:12 -0800406 int c;
Sage Weilf24e9982009-10-06 11:31:10 -0700407
Sage Weil8adc8b32011-09-28 10:11:04 -0700408 dout("__insert_pg_mapping %llx %p\n", *(u64 *)&new->pgid, new);
Sage Weilf24e9982009-10-06 11:31:10 -0700409 while (*p) {
410 parent = *p;
411 pg = rb_entry(parent, struct ceph_pg_mapping, node);
Sage Weil51042122009-11-04 11:39:12 -0800412 c = pgid_cmp(new->pgid, pg->pgid);
413 if (c < 0)
Sage Weilf24e9982009-10-06 11:31:10 -0700414 p = &(*p)->rb_left;
Sage Weil51042122009-11-04 11:39:12 -0800415 else if (c > 0)
Sage Weilf24e9982009-10-06 11:31:10 -0700416 p = &(*p)->rb_right;
417 else
Sage Weil991abb62009-10-08 22:22:37 -0700418 return -EEXIST;
Sage Weilf24e9982009-10-06 11:31:10 -0700419 }
420
421 rb_link_node(&new->node, parent, p);
422 rb_insert_color(&new->node, root);
Sage Weil991abb62009-10-08 22:22:37 -0700423 return 0;
Sage Weilf24e9982009-10-06 11:31:10 -0700424}
425
Sage Weil9794b142010-02-16 15:53:32 -0800426static struct ceph_pg_mapping *__lookup_pg_mapping(struct rb_root *root,
Sage Weil5b191d92013-02-23 10:38:16 -0800427 struct ceph_pg pgid)
Sage Weil9794b142010-02-16 15:53:32 -0800428{
429 struct rb_node *n = root->rb_node;
430 struct ceph_pg_mapping *pg;
431 int c;
432
433 while (n) {
434 pg = rb_entry(n, struct ceph_pg_mapping, node);
435 c = pgid_cmp(pgid, pg->pgid);
Sage Weil8adc8b32011-09-28 10:11:04 -0700436 if (c < 0) {
Sage Weil9794b142010-02-16 15:53:32 -0800437 n = n->rb_left;
Sage Weil8adc8b32011-09-28 10:11:04 -0700438 } else if (c > 0) {
Sage Weil9794b142010-02-16 15:53:32 -0800439 n = n->rb_right;
Sage Weil8adc8b32011-09-28 10:11:04 -0700440 } else {
Sage Weil5b191d92013-02-23 10:38:16 -0800441 dout("__lookup_pg_mapping %lld.%x got %p\n",
442 pgid.pool, pgid.seed, pg);
Sage Weil9794b142010-02-16 15:53:32 -0800443 return pg;
Sage Weil8adc8b32011-09-28 10:11:04 -0700444 }
Sage Weil9794b142010-02-16 15:53:32 -0800445 }
446 return NULL;
447}
448
Sage Weil5b191d92013-02-23 10:38:16 -0800449static int __remove_pg_mapping(struct rb_root *root, struct ceph_pg pgid)
Sage Weil8adc8b32011-09-28 10:11:04 -0700450{
451 struct ceph_pg_mapping *pg = __lookup_pg_mapping(root, pgid);
452
453 if (pg) {
Sage Weil5b191d92013-02-23 10:38:16 -0800454 dout("__remove_pg_mapping %lld.%x %p\n", pgid.pool, pgid.seed,
455 pg);
Sage Weil8adc8b32011-09-28 10:11:04 -0700456 rb_erase(&pg->node, root);
457 kfree(pg);
458 return 0;
459 }
Sage Weil5b191d92013-02-23 10:38:16 -0800460 dout("__remove_pg_mapping %lld.%x dne\n", pgid.pool, pgid.seed);
Sage Weil8adc8b32011-09-28 10:11:04 -0700461 return -ENOENT;
462}
463
Sage Weilf24e9982009-10-06 11:31:10 -0700464/*
Sage Weil4fc51be2010-02-16 15:55:03 -0800465 * rbtree of pg pool info
466 */
467static int __insert_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *new)
468{
469 struct rb_node **p = &root->rb_node;
470 struct rb_node *parent = NULL;
471 struct ceph_pg_pool_info *pi = NULL;
472
473 while (*p) {
474 parent = *p;
475 pi = rb_entry(parent, struct ceph_pg_pool_info, node);
476 if (new->id < pi->id)
477 p = &(*p)->rb_left;
478 else if (new->id > pi->id)
479 p = &(*p)->rb_right;
480 else
481 return -EEXIST;
482 }
483
484 rb_link_node(&new->node, parent, p);
485 rb_insert_color(&new->node, root);
486 return 0;
487}
488
Sage Weil4f6a7e52013-02-23 10:41:09 -0800489static struct ceph_pg_pool_info *__lookup_pg_pool(struct rb_root *root, u64 id)
Sage Weil4fc51be2010-02-16 15:55:03 -0800490{
491 struct ceph_pg_pool_info *pi;
492 struct rb_node *n = root->rb_node;
493
494 while (n) {
495 pi = rb_entry(n, struct ceph_pg_pool_info, node);
496 if (id < pi->id)
497 n = n->rb_left;
498 else if (id > pi->id)
499 n = n->rb_right;
500 else
501 return pi;
502 }
503 return NULL;
504}
505
Ilya Dryomovce7f6a22014-01-27 17:40:19 +0200506struct ceph_pg_pool_info *ceph_pg_pool_by_id(struct ceph_osdmap *map, u64 id)
507{
508 return __lookup_pg_pool(&map->pg_pools, id);
509}
510
Alex Elder72afc712012-10-30 19:40:33 -0500511const char *ceph_pg_pool_name_by_id(struct ceph_osdmap *map, u64 id)
512{
513 struct ceph_pg_pool_info *pi;
514
515 if (id == CEPH_NOPOOL)
516 return NULL;
517
518 if (WARN_ON_ONCE(id > (u64) INT_MAX))
519 return NULL;
520
521 pi = __lookup_pg_pool(&map->pg_pools, (int) id);
522
523 return pi ? pi->name : NULL;
524}
525EXPORT_SYMBOL(ceph_pg_pool_name_by_id);
526
Yehuda Sadeh7669a2c2010-05-17 12:31:35 -0700527int ceph_pg_poolid_by_name(struct ceph_osdmap *map, const char *name)
528{
529 struct rb_node *rbp;
530
531 for (rbp = rb_first(&map->pg_pools); rbp; rbp = rb_next(rbp)) {
532 struct ceph_pg_pool_info *pi =
533 rb_entry(rbp, struct ceph_pg_pool_info, node);
534 if (pi->name && strcmp(pi->name, name) == 0)
535 return pi->id;
536 }
537 return -ENOENT;
538}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700539EXPORT_SYMBOL(ceph_pg_poolid_by_name);
Yehuda Sadeh7669a2c2010-05-17 12:31:35 -0700540
Sage Weil2844a762010-04-09 15:46:42 -0700541static void __remove_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *pi)
542{
543 rb_erase(&pi->node, root);
544 kfree(pi->name);
545 kfree(pi);
546}
547
Ilya Dryomov0f70c7e2014-03-21 19:05:27 +0200548static int decode_pool(void **p, void *end, struct ceph_pg_pool_info *pi)
Sage Weilefd75762010-03-17 10:05:28 -0700549{
Sage Weil4f6a7e52013-02-23 10:41:09 -0800550 u8 ev, cv;
551 unsigned len, num;
552 void *pool_end;
Sage Weil73a7e692010-08-02 11:00:55 -0700553
Sage Weil4f6a7e52013-02-23 10:41:09 -0800554 ceph_decode_need(p, end, 2 + 4, bad);
555 ev = ceph_decode_8(p); /* encoding version */
556 cv = ceph_decode_8(p); /* compat version */
557 if (ev < 5) {
Joe Perchesb9a67892014-09-09 21:17:29 -0700558 pr_warn("got v %d < 5 cv %d of ceph_pg_pool\n", ev, cv);
Sage Weil4f6a7e52013-02-23 10:41:09 -0800559 return -EINVAL;
560 }
Ilya Dryomov17a13e42014-01-27 17:40:19 +0200561 if (cv > 9) {
Joe Perchesb9a67892014-09-09 21:17:29 -0700562 pr_warn("got v %d cv %d > 9 of ceph_pg_pool\n", ev, cv);
Sage Weil4f6a7e52013-02-23 10:41:09 -0800563 return -EINVAL;
564 }
565 len = ceph_decode_32(p);
566 ceph_decode_need(p, end, len, bad);
567 pool_end = *p + len;
Sage Weil73a7e692010-08-02 11:00:55 -0700568
Sage Weil4f6a7e52013-02-23 10:41:09 -0800569 pi->type = ceph_decode_8(p);
570 pi->size = ceph_decode_8(p);
571 pi->crush_ruleset = ceph_decode_8(p);
572 pi->object_hash = ceph_decode_8(p);
573
574 pi->pg_num = ceph_decode_32(p);
575 pi->pgp_num = ceph_decode_32(p);
576
577 *p += 4 + 4; /* skip lpg* */
578 *p += 4; /* skip last_change */
579 *p += 8 + 4; /* skip snap_seq, snap_epoch */
580
581 /* skip snaps */
582 num = ceph_decode_32(p);
583 while (num--) {
584 *p += 8; /* snapid key */
585 *p += 1 + 1; /* versions */
586 len = ceph_decode_32(p);
587 *p += len;
Sage Weil73a7e692010-08-02 11:00:55 -0700588 }
589
Ilya Dryomov17a13e42014-01-27 17:40:19 +0200590 /* skip removed_snaps */
Sage Weil4f6a7e52013-02-23 10:41:09 -0800591 num = ceph_decode_32(p);
592 *p += num * (8 + 8);
593
594 *p += 8; /* skip auid */
595 pi->flags = ceph_decode_64(p);
Ilya Dryomov17a13e42014-01-27 17:40:19 +0200596 *p += 4; /* skip crash_replay_interval */
597
598 if (ev >= 7)
599 *p += 1; /* skip min_size */
600
601 if (ev >= 8)
602 *p += 8 + 8; /* skip quota_max_* */
603
604 if (ev >= 9) {
605 /* skip tiers */
606 num = ceph_decode_32(p);
607 *p += num * 8;
608
609 *p += 8; /* skip tier_of */
610 *p += 1; /* skip cache_mode */
611
612 pi->read_tier = ceph_decode_64(p);
613 pi->write_tier = ceph_decode_64(p);
614 } else {
615 pi->read_tier = -1;
616 pi->write_tier = -1;
617 }
Sage Weil4f6a7e52013-02-23 10:41:09 -0800618
619 /* ignore the rest */
620
621 *p = pool_end;
622 calc_pg_masks(pi);
Sage Weil73a7e692010-08-02 11:00:55 -0700623 return 0;
624
625bad:
626 return -EINVAL;
Sage Weilefd75762010-03-17 10:05:28 -0700627}
628
Ilya Dryomov0f70c7e2014-03-21 19:05:27 +0200629static int decode_pool_names(void **p, void *end, struct ceph_osdmap *map)
Sage Weil2844a762010-04-09 15:46:42 -0700630{
631 struct ceph_pg_pool_info *pi;
Sage Weil4f6a7e52013-02-23 10:41:09 -0800632 u32 num, len;
633 u64 pool;
Sage Weil2844a762010-04-09 15:46:42 -0700634
635 ceph_decode_32_safe(p, end, num, bad);
636 dout(" %d pool names\n", num);
637 while (num--) {
Sage Weil4f6a7e52013-02-23 10:41:09 -0800638 ceph_decode_64_safe(p, end, pool, bad);
Sage Weil2844a762010-04-09 15:46:42 -0700639 ceph_decode_32_safe(p, end, len, bad);
Sage Weil4f6a7e52013-02-23 10:41:09 -0800640 dout(" pool %llu len %d\n", pool, len);
Xi Wangad3b9042012-06-06 19:35:55 -0500641 ceph_decode_need(p, end, len, bad);
Sage Weil2844a762010-04-09 15:46:42 -0700642 pi = __lookup_pg_pool(&map->pg_pools, pool);
643 if (pi) {
Xi Wangad3b9042012-06-06 19:35:55 -0500644 char *name = kstrndup(*p, len, GFP_NOFS);
645
646 if (!name)
647 return -ENOMEM;
Sage Weil2844a762010-04-09 15:46:42 -0700648 kfree(pi->name);
Xi Wangad3b9042012-06-06 19:35:55 -0500649 pi->name = name;
650 dout(" name is %s\n", pi->name);
Sage Weil2844a762010-04-09 15:46:42 -0700651 }
652 *p += len;
653 }
654 return 0;
655
656bad:
657 return -EINVAL;
658}
659
660/*
661 * osd map
662 */
663void ceph_osdmap_destroy(struct ceph_osdmap *map)
664{
665 dout("osdmap_destroy %p\n", map);
666 if (map->crush)
667 crush_destroy(map->crush);
668 while (!RB_EMPTY_ROOT(&map->pg_temp)) {
669 struct ceph_pg_mapping *pg =
670 rb_entry(rb_first(&map->pg_temp),
671 struct ceph_pg_mapping, node);
672 rb_erase(&pg->node, &map->pg_temp);
673 kfree(pg);
674 }
Ilya Dryomov9686f942014-03-21 19:05:29 +0200675 while (!RB_EMPTY_ROOT(&map->primary_temp)) {
676 struct ceph_pg_mapping *pg =
677 rb_entry(rb_first(&map->primary_temp),
678 struct ceph_pg_mapping, node);
679 rb_erase(&pg->node, &map->primary_temp);
680 kfree(pg);
681 }
Sage Weil2844a762010-04-09 15:46:42 -0700682 while (!RB_EMPTY_ROOT(&map->pg_pools)) {
683 struct ceph_pg_pool_info *pi =
684 rb_entry(rb_first(&map->pg_pools),
685 struct ceph_pg_pool_info, node);
686 __remove_pg_pool(&map->pg_pools, pi);
687 }
688 kfree(map->osd_state);
689 kfree(map->osd_weight);
690 kfree(map->osd_addr);
Ilya Dryomov2cfa34f2014-03-21 19:05:30 +0200691 kfree(map->osd_primary_affinity);
Sage Weil2844a762010-04-09 15:46:42 -0700692 kfree(map);
693}
694
695/*
Ilya Dryomov4d603512014-03-21 19:05:28 +0200696 * Adjust max_osd value, (re)allocate arrays.
697 *
698 * The new elements are properly initialized.
Sage Weil2844a762010-04-09 15:46:42 -0700699 */
700static int osdmap_set_max_osd(struct ceph_osdmap *map, int max)
701{
702 u8 *state;
Sage Weil2844a762010-04-09 15:46:42 -0700703 u32 *weight;
Ilya Dryomov4d603512014-03-21 19:05:28 +0200704 struct ceph_entity_addr *addr;
705 int i;
Sage Weil2844a762010-04-09 15:46:42 -0700706
Ilya Dryomov4d603512014-03-21 19:05:28 +0200707 state = krealloc(map->osd_state, max*sizeof(*state), GFP_NOFS);
Li RongQing589506f2014-09-07 18:10:51 +0800708 if (!state)
Sage Weil2844a762010-04-09 15:46:42 -0700709 return -ENOMEM;
Li RongQing589506f2014-09-07 18:10:51 +0800710 map->osd_state = state;
711
712 weight = krealloc(map->osd_weight, max*sizeof(*weight), GFP_NOFS);
713 if (!weight)
714 return -ENOMEM;
715 map->osd_weight = weight;
716
717 addr = krealloc(map->osd_addr, max*sizeof(*addr), GFP_NOFS);
718 if (!addr)
719 return -ENOMEM;
720 map->osd_addr = addr;
Sage Weil2844a762010-04-09 15:46:42 -0700721
Ilya Dryomov4d603512014-03-21 19:05:28 +0200722 for (i = map->max_osd; i < max; i++) {
Li RongQing589506f2014-09-07 18:10:51 +0800723 map->osd_state[i] = 0;
724 map->osd_weight[i] = CEPH_OSD_OUT;
725 memset(map->osd_addr + i, 0, sizeof(*map->osd_addr));
Sage Weil2844a762010-04-09 15:46:42 -0700726 }
727
Ilya Dryomov2cfa34f2014-03-21 19:05:30 +0200728 if (map->osd_primary_affinity) {
729 u32 *affinity;
730
731 affinity = krealloc(map->osd_primary_affinity,
732 max*sizeof(*affinity), GFP_NOFS);
733 if (!affinity)
734 return -ENOMEM;
Li RongQing589506f2014-09-07 18:10:51 +0800735 map->osd_primary_affinity = affinity;
Ilya Dryomov2cfa34f2014-03-21 19:05:30 +0200736
737 for (i = map->max_osd; i < max; i++)
Li RongQing589506f2014-09-07 18:10:51 +0800738 map->osd_primary_affinity[i] =
739 CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
Ilya Dryomov2cfa34f2014-03-21 19:05:30 +0200740 }
741
Sage Weil2844a762010-04-09 15:46:42 -0700742 map->max_osd = max;
Ilya Dryomov4d603512014-03-21 19:05:28 +0200743
Sage Weil2844a762010-04-09 15:46:42 -0700744 return 0;
745}
746
Ilya Dryomovec7af972014-03-21 19:05:29 +0200747#define OSDMAP_WRAPPER_COMPAT_VER 7
748#define OSDMAP_CLIENT_DATA_COMPAT_VER 1
749
750/*
751 * Return 0 or error. On success, *v is set to 0 for old (v6) osdmaps,
752 * to struct_v of the client_data section for new (v7 and above)
753 * osdmaps.
754 */
755static int get_osdmap_client_data_v(void **p, void *end,
756 const char *prefix, u8 *v)
757{
758 u8 struct_v;
759
760 ceph_decode_8_safe(p, end, struct_v, e_inval);
761 if (struct_v >= 7) {
762 u8 struct_compat;
763
764 ceph_decode_8_safe(p, end, struct_compat, e_inval);
765 if (struct_compat > OSDMAP_WRAPPER_COMPAT_VER) {
Joe Perchesb9a67892014-09-09 21:17:29 -0700766 pr_warn("got v %d cv %d > %d of %s ceph_osdmap\n",
767 struct_v, struct_compat,
768 OSDMAP_WRAPPER_COMPAT_VER, prefix);
Ilya Dryomovec7af972014-03-21 19:05:29 +0200769 return -EINVAL;
770 }
771 *p += 4; /* ignore wrapper struct_len */
772
773 ceph_decode_8_safe(p, end, struct_v, e_inval);
774 ceph_decode_8_safe(p, end, struct_compat, e_inval);
775 if (struct_compat > OSDMAP_CLIENT_DATA_COMPAT_VER) {
Joe Perchesb9a67892014-09-09 21:17:29 -0700776 pr_warn("got v %d cv %d > %d of %s ceph_osdmap client data\n",
777 struct_v, struct_compat,
778 OSDMAP_CLIENT_DATA_COMPAT_VER, prefix);
Ilya Dryomovec7af972014-03-21 19:05:29 +0200779 return -EINVAL;
780 }
781 *p += 4; /* ignore client data struct_len */
782 } else {
783 u16 version;
784
785 *p -= 1;
786 ceph_decode_16_safe(p, end, version, e_inval);
787 if (version < 6) {
Joe Perchesb9a67892014-09-09 21:17:29 -0700788 pr_warn("got v %d < 6 of %s ceph_osdmap\n",
789 version, prefix);
Ilya Dryomovec7af972014-03-21 19:05:29 +0200790 return -EINVAL;
791 }
792
793 /* old osdmap enconding */
794 struct_v = 0;
795 }
796
797 *v = struct_v;
798 return 0;
799
800e_inval:
801 return -EINVAL;
802}
803
Ilya Dryomov433fbdd2014-03-21 19:05:27 +0200804static int __decode_pools(void **p, void *end, struct ceph_osdmap *map,
805 bool incremental)
806{
807 u32 n;
808
809 ceph_decode_32_safe(p, end, n, e_inval);
810 while (n--) {
811 struct ceph_pg_pool_info *pi;
812 u64 pool;
813 int ret;
814
815 ceph_decode_64_safe(p, end, pool, e_inval);
816
817 pi = __lookup_pg_pool(&map->pg_pools, pool);
818 if (!incremental || !pi) {
819 pi = kzalloc(sizeof(*pi), GFP_NOFS);
820 if (!pi)
821 return -ENOMEM;
822
823 pi->id = pool;
824
825 ret = __insert_pg_pool(&map->pg_pools, pi);
826 if (ret) {
827 kfree(pi);
828 return ret;
829 }
830 }
831
832 ret = decode_pool(p, end, pi);
833 if (ret)
834 return ret;
835 }
836
837 return 0;
838
839e_inval:
840 return -EINVAL;
841}
842
843static int decode_pools(void **p, void *end, struct ceph_osdmap *map)
844{
845 return __decode_pools(p, end, map, false);
846}
847
848static int decode_new_pools(void **p, void *end, struct ceph_osdmap *map)
849{
850 return __decode_pools(p, end, map, true);
851}
852
Ilya Dryomov10db6342014-03-21 19:05:28 +0200853static int __decode_pg_temp(void **p, void *end, struct ceph_osdmap *map,
854 bool incremental)
855{
856 u32 n;
857
858 ceph_decode_32_safe(p, end, n, e_inval);
859 while (n--) {
860 struct ceph_pg pgid;
861 u32 len, i;
862 int ret;
863
864 ret = ceph_decode_pgid(p, end, &pgid);
865 if (ret)
866 return ret;
867
868 ceph_decode_32_safe(p, end, len, e_inval);
869
870 ret = __remove_pg_mapping(&map->pg_temp, pgid);
871 BUG_ON(!incremental && ret != -ENOENT);
872
873 if (!incremental || len > 0) {
874 struct ceph_pg_mapping *pg;
875
876 ceph_decode_need(p, end, len*sizeof(u32), e_inval);
877
878 if (len > (UINT_MAX - sizeof(*pg)) / sizeof(u32))
879 return -EINVAL;
880
881 pg = kzalloc(sizeof(*pg) + len*sizeof(u32), GFP_NOFS);
882 if (!pg)
883 return -ENOMEM;
884
885 pg->pgid = pgid;
Ilya Dryomov35a935d2014-03-21 19:05:29 +0200886 pg->pg_temp.len = len;
Ilya Dryomov10db6342014-03-21 19:05:28 +0200887 for (i = 0; i < len; i++)
Ilya Dryomov35a935d2014-03-21 19:05:29 +0200888 pg->pg_temp.osds[i] = ceph_decode_32(p);
Ilya Dryomov10db6342014-03-21 19:05:28 +0200889
890 ret = __insert_pg_mapping(pg, &map->pg_temp);
891 if (ret) {
892 kfree(pg);
893 return ret;
894 }
895 }
896 }
897
898 return 0;
899
900e_inval:
901 return -EINVAL;
902}
903
904static int decode_pg_temp(void **p, void *end, struct ceph_osdmap *map)
905{
906 return __decode_pg_temp(p, end, map, false);
907}
908
909static int decode_new_pg_temp(void **p, void *end, struct ceph_osdmap *map)
910{
911 return __decode_pg_temp(p, end, map, true);
912}
913
Ilya Dryomovd286de72014-03-21 19:05:30 +0200914static int __decode_primary_temp(void **p, void *end, struct ceph_osdmap *map,
915 bool incremental)
916{
917 u32 n;
918
919 ceph_decode_32_safe(p, end, n, e_inval);
920 while (n--) {
921 struct ceph_pg pgid;
922 u32 osd;
923 int ret;
924
925 ret = ceph_decode_pgid(p, end, &pgid);
926 if (ret)
927 return ret;
928
929 ceph_decode_32_safe(p, end, osd, e_inval);
930
931 ret = __remove_pg_mapping(&map->primary_temp, pgid);
932 BUG_ON(!incremental && ret != -ENOENT);
933
934 if (!incremental || osd != (u32)-1) {
935 struct ceph_pg_mapping *pg;
936
937 pg = kzalloc(sizeof(*pg), GFP_NOFS);
938 if (!pg)
939 return -ENOMEM;
940
941 pg->pgid = pgid;
942 pg->primary_temp.osd = osd;
943
944 ret = __insert_pg_mapping(pg, &map->primary_temp);
945 if (ret) {
946 kfree(pg);
947 return ret;
948 }
949 }
950 }
951
952 return 0;
953
954e_inval:
955 return -EINVAL;
956}
957
958static int decode_primary_temp(void **p, void *end, struct ceph_osdmap *map)
959{
960 return __decode_primary_temp(p, end, map, false);
961}
962
963static int decode_new_primary_temp(void **p, void *end,
964 struct ceph_osdmap *map)
965{
966 return __decode_primary_temp(p, end, map, true);
967}
968
Ilya Dryomov2cfa34f2014-03-21 19:05:30 +0200969u32 ceph_get_primary_affinity(struct ceph_osdmap *map, int osd)
970{
971 BUG_ON(osd >= map->max_osd);
972
973 if (!map->osd_primary_affinity)
974 return CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
975
976 return map->osd_primary_affinity[osd];
977}
978
979static int set_primary_affinity(struct ceph_osdmap *map, int osd, u32 aff)
980{
981 BUG_ON(osd >= map->max_osd);
982
983 if (!map->osd_primary_affinity) {
984 int i;
985
986 map->osd_primary_affinity = kmalloc(map->max_osd*sizeof(u32),
987 GFP_NOFS);
988 if (!map->osd_primary_affinity)
989 return -ENOMEM;
990
991 for (i = 0; i < map->max_osd; i++)
992 map->osd_primary_affinity[i] =
993 CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
994 }
995
996 map->osd_primary_affinity[osd] = aff;
997
998 return 0;
999}
1000
Ilya Dryomov63a69932014-03-21 19:05:30 +02001001static int decode_primary_affinity(void **p, void *end,
1002 struct ceph_osdmap *map)
1003{
1004 u32 len, i;
1005
1006 ceph_decode_32_safe(p, end, len, e_inval);
1007 if (len == 0) {
1008 kfree(map->osd_primary_affinity);
1009 map->osd_primary_affinity = NULL;
1010 return 0;
1011 }
1012 if (len != map->max_osd)
1013 goto e_inval;
1014
1015 ceph_decode_need(p, end, map->max_osd*sizeof(u32), e_inval);
1016
1017 for (i = 0; i < map->max_osd; i++) {
1018 int ret;
1019
1020 ret = set_primary_affinity(map, i, ceph_decode_32(p));
1021 if (ret)
1022 return ret;
1023 }
1024
1025 return 0;
1026
1027e_inval:
1028 return -EINVAL;
1029}
1030
1031static int decode_new_primary_affinity(void **p, void *end,
1032 struct ceph_osdmap *map)
1033{
1034 u32 n;
1035
1036 ceph_decode_32_safe(p, end, n, e_inval);
1037 while (n--) {
1038 u32 osd, aff;
1039 int ret;
1040
1041 ceph_decode_32_safe(p, end, osd, e_inval);
1042 ceph_decode_32_safe(p, end, aff, e_inval);
1043
1044 ret = set_primary_affinity(map, osd, aff);
1045 if (ret)
1046 return ret;
Ilya Dryomovf31da0f2014-04-02 20:34:04 +04001047
1048 pr_info("osd%d primary-affinity 0x%x\n", osd, aff);
Ilya Dryomov63a69932014-03-21 19:05:30 +02001049 }
1050
1051 return 0;
1052
1053e_inval:
1054 return -EINVAL;
1055}
1056
Sage Weil4fc51be2010-02-16 15:55:03 -08001057/*
Sage Weilf24e9982009-10-06 11:31:10 -07001058 * decode a full map.
1059 */
Ilya Dryomova2505d62014-03-13 16:36:13 +02001060static int osdmap_decode(void **p, void *end, struct ceph_osdmap *map)
Sage Weilf24e9982009-10-06 11:31:10 -07001061{
Ilya Dryomovec7af972014-03-21 19:05:29 +02001062 u8 struct_v;
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001063 u32 epoch = 0;
Sage Weilf24e9982009-10-06 11:31:10 -07001064 void *start = *p;
Ilya Dryomov39770582014-03-13 16:36:14 +02001065 u32 max;
1066 u32 len, i;
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001067 int err;
Sage Weilf24e9982009-10-06 11:31:10 -07001068
Ilya Dryomova2505d62014-03-13 16:36:13 +02001069 dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p));
Sage Weilf24e9982009-10-06 11:31:10 -07001070
Ilya Dryomovec7af972014-03-21 19:05:29 +02001071 err = get_osdmap_client_data_v(p, end, "full", &struct_v);
1072 if (err)
1073 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001074
Ilya Dryomov53bbaba2014-03-13 16:36:16 +02001075 /* fsid, epoch, created, modified */
1076 ceph_decode_need(p, end, sizeof(map->fsid) + sizeof(u32) +
1077 sizeof(map->created) + sizeof(map->modified), e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001078 ceph_decode_copy(p, &map->fsid, sizeof(map->fsid));
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001079 epoch = map->epoch = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -07001080 ceph_decode_copy(p, &map->created, sizeof(map->created));
1081 ceph_decode_copy(p, &map->modified, sizeof(map->modified));
1082
Ilya Dryomov433fbdd2014-03-21 19:05:27 +02001083 /* pools */
1084 err = decode_pools(p, end, map);
1085 if (err)
1086 goto bad;
Sage Weil2844a762010-04-09 15:46:42 -07001087
Ilya Dryomov0f70c7e2014-03-21 19:05:27 +02001088 /* pool_name */
1089 err = decode_pool_names(p, end, map);
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001090 if (err)
Sage Weil4f6a7e52013-02-23 10:41:09 -08001091 goto bad;
Sage Weil2844a762010-04-09 15:46:42 -07001092
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001093 ceph_decode_32_safe(p, end, map->pool_max, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001094
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001095 ceph_decode_32_safe(p, end, map->flags, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001096
Ilya Dryomov39770582014-03-13 16:36:14 +02001097 /* max_osd */
1098 ceph_decode_32_safe(p, end, max, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001099
1100 /* (re)alloc osd arrays */
1101 err = osdmap_set_max_osd(map, max);
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001102 if (err)
Sage Weilf24e9982009-10-06 11:31:10 -07001103 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001104
Ilya Dryomov2d88b2e2014-03-13 16:36:14 +02001105 /* osd_state, osd_weight, osd_addrs->client_addr */
Sage Weilf24e9982009-10-06 11:31:10 -07001106 ceph_decode_need(p, end, 3*sizeof(u32) +
1107 map->max_osd*(1 + sizeof(*map->osd_weight) +
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001108 sizeof(*map->osd_addr)), e_inval);
1109
Ilya Dryomov2d88b2e2014-03-13 16:36:14 +02001110 if (ceph_decode_32(p) != map->max_osd)
1111 goto e_inval;
1112
Sage Weilf24e9982009-10-06 11:31:10 -07001113 ceph_decode_copy(p, map->osd_state, map->max_osd);
1114
Ilya Dryomov2d88b2e2014-03-13 16:36:14 +02001115 if (ceph_decode_32(p) != map->max_osd)
1116 goto e_inval;
1117
Sage Weilf24e9982009-10-06 11:31:10 -07001118 for (i = 0; i < map->max_osd; i++)
Sage Weilc89136e2009-10-14 09:59:09 -07001119 map->osd_weight[i] = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -07001120
Ilya Dryomov2d88b2e2014-03-13 16:36:14 +02001121 if (ceph_decode_32(p) != map->max_osd)
1122 goto e_inval;
1123
Sage Weilf24e9982009-10-06 11:31:10 -07001124 ceph_decode_copy(p, map->osd_addr, map->max_osd*sizeof(*map->osd_addr));
Sage Weil63f2d212009-11-03 15:17:56 -08001125 for (i = 0; i < map->max_osd; i++)
1126 ceph_decode_addr(&map->osd_addr[i]);
Sage Weilf24e9982009-10-06 11:31:10 -07001127
1128 /* pg_temp */
Ilya Dryomov10db6342014-03-21 19:05:28 +02001129 err = decode_pg_temp(p, end, map);
1130 if (err)
1131 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001132
Ilya Dryomovd286de72014-03-21 19:05:30 +02001133 /* primary_temp */
1134 if (struct_v >= 1) {
1135 err = decode_primary_temp(p, end, map);
1136 if (err)
1137 goto bad;
1138 }
1139
Ilya Dryomov63a69932014-03-21 19:05:30 +02001140 /* primary_affinity */
1141 if (struct_v >= 2) {
1142 err = decode_primary_affinity(p, end, map);
1143 if (err)
1144 goto bad;
1145 } else {
1146 /* XXX can this happen? */
1147 kfree(map->osd_primary_affinity);
1148 map->osd_primary_affinity = NULL;
1149 }
1150
Sage Weilf24e9982009-10-06 11:31:10 -07001151 /* crush */
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001152 ceph_decode_32_safe(p, end, len, e_inval);
Ilya Dryomov9902e682014-03-13 16:36:15 +02001153 map->crush = crush_decode(*p, min(*p + len, end));
Sage Weilf24e9982009-10-06 11:31:10 -07001154 if (IS_ERR(map->crush)) {
1155 err = PTR_ERR(map->crush);
1156 map->crush = NULL;
1157 goto bad;
1158 }
Ilya Dryomov9902e682014-03-13 16:36:15 +02001159 *p += len;
Sage Weilf24e9982009-10-06 11:31:10 -07001160
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001161 /* ignore the rest */
Sage Weilf24e9982009-10-06 11:31:10 -07001162 *p = end;
1163
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001164 dout("full osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd);
Ilya Dryomova2505d62014-03-13 16:36:13 +02001165 return 0;
Sage Weilf24e9982009-10-06 11:31:10 -07001166
Ilya Dryomov597b52f2014-03-13 16:36:14 +02001167e_inval:
1168 err = -EINVAL;
Sage Weilf24e9982009-10-06 11:31:10 -07001169bad:
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001170 pr_err("corrupt full osdmap (%d) epoch %d off %d (%p of %p-%p)\n",
1171 err, epoch, (int)(*p - start), *p, start, end);
1172 print_hex_dump(KERN_DEBUG, "osdmap: ",
1173 DUMP_PREFIX_OFFSET, 16, 1,
1174 start, end - start, true);
Ilya Dryomova2505d62014-03-13 16:36:13 +02001175 return err;
1176}
1177
1178/*
1179 * Allocate and decode a full map.
1180 */
1181struct ceph_osdmap *ceph_osdmap_decode(void **p, void *end)
1182{
1183 struct ceph_osdmap *map;
1184 int ret;
1185
1186 map = kzalloc(sizeof(*map), GFP_NOFS);
1187 if (!map)
1188 return ERR_PTR(-ENOMEM);
1189
1190 map->pg_temp = RB_ROOT;
Ilya Dryomov9686f942014-03-21 19:05:29 +02001191 map->primary_temp = RB_ROOT;
Ilya Dryomova2505d62014-03-13 16:36:13 +02001192 mutex_init(&map->crush_scratch_mutex);
1193
1194 ret = osdmap_decode(p, end, map);
1195 if (ret) {
1196 ceph_osdmap_destroy(map);
1197 return ERR_PTR(ret);
1198 }
1199
1200 return map;
Sage Weilf24e9982009-10-06 11:31:10 -07001201}
1202
1203/*
1204 * decode and apply an incremental map update.
1205 */
1206struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end,
1207 struct ceph_osdmap *map,
1208 struct ceph_messenger *msgr)
1209{
Sage Weilf24e9982009-10-06 11:31:10 -07001210 struct crush_map *newcrush = NULL;
1211 struct ceph_fsid fsid;
1212 u32 epoch = 0;
1213 struct ceph_timespec modified;
Sage Weil4f6a7e52013-02-23 10:41:09 -08001214 s32 len;
1215 u64 pool;
1216 __s64 new_pool_max;
1217 __s32 new_flags, max;
Sage Weilf24e9982009-10-06 11:31:10 -07001218 void *start = *p;
Ilya Dryomov86f17422014-03-13 16:36:15 +02001219 int err;
Ilya Dryomovec7af972014-03-21 19:05:29 +02001220 u8 struct_v;
Sage Weilf24e9982009-10-06 11:31:10 -07001221
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001222 dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p));
1223
Ilya Dryomovec7af972014-03-21 19:05:29 +02001224 err = get_osdmap_client_data_v(p, end, "inc", &struct_v);
1225 if (err)
1226 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001227
Ilya Dryomov53bbaba2014-03-13 16:36:16 +02001228 /* fsid, epoch, modified, new_pool_max, new_flags */
1229 ceph_decode_need(p, end, sizeof(fsid) + sizeof(u32) + sizeof(modified) +
1230 sizeof(u64) + sizeof(u32), e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001231 ceph_decode_copy(p, &fsid, sizeof(fsid));
Sage Weilc89136e2009-10-14 09:59:09 -07001232 epoch = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -07001233 BUG_ON(epoch != map->epoch+1);
1234 ceph_decode_copy(p, &modified, sizeof(modified));
Sage Weil4f6a7e52013-02-23 10:41:09 -08001235 new_pool_max = ceph_decode_64(p);
Sage Weilc89136e2009-10-14 09:59:09 -07001236 new_flags = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -07001237
1238 /* full map? */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001239 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001240 if (len > 0) {
1241 dout("apply_incremental full map len %d, %p to %p\n",
1242 len, *p, end);
Ilya Dryomova2505d62014-03-13 16:36:13 +02001243 return ceph_osdmap_decode(p, min(*p+len, end));
Sage Weilf24e9982009-10-06 11:31:10 -07001244 }
1245
1246 /* new crush? */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001247 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001248 if (len > 0) {
Sage Weilf24e9982009-10-06 11:31:10 -07001249 newcrush = crush_decode(*p, min(*p+len, end));
Ilya Dryomov86f17422014-03-13 16:36:15 +02001250 if (IS_ERR(newcrush)) {
1251 err = PTR_ERR(newcrush);
1252 newcrush = NULL;
1253 goto bad;
1254 }
Sage Weilcebc5be2010-06-17 10:22:48 -07001255 *p += len;
Sage Weilf24e9982009-10-06 11:31:10 -07001256 }
1257
1258 /* new flags? */
1259 if (new_flags >= 0)
1260 map->flags = new_flags;
Sage Weil4fc51be2010-02-16 15:55:03 -08001261 if (new_pool_max >= 0)
1262 map->pool_max = new_pool_max;
Sage Weilf24e9982009-10-06 11:31:10 -07001263
Sage Weilf24e9982009-10-06 11:31:10 -07001264 /* new max? */
Ilya Dryomov53bbaba2014-03-13 16:36:16 +02001265 ceph_decode_32_safe(p, end, max, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001266 if (max >= 0) {
1267 err = osdmap_set_max_osd(map, max);
Ilya Dryomov86f17422014-03-13 16:36:15 +02001268 if (err)
Sage Weilf24e9982009-10-06 11:31:10 -07001269 goto bad;
1270 }
1271
1272 map->epoch++;
Sage Weil31456662011-05-12 15:18:43 -07001273 map->modified = modified;
Sage Weilf24e9982009-10-06 11:31:10 -07001274 if (newcrush) {
1275 if (map->crush)
1276 crush_destroy(map->crush);
1277 map->crush = newcrush;
1278 newcrush = NULL;
1279 }
1280
Ilya Dryomov433fbdd2014-03-21 19:05:27 +02001281 /* new_pools */
1282 err = decode_new_pools(p, end, map);
1283 if (err)
1284 goto bad;
Ilya Dryomov9464d002014-03-13 16:36:16 +02001285
Ilya Dryomov0f70c7e2014-03-21 19:05:27 +02001286 /* new_pool_names */
1287 err = decode_pool_names(p, end, map);
Ilya Dryomov9464d002014-03-13 16:36:16 +02001288 if (err)
1289 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001290
Sage Weil4fc51be2010-02-16 15:55:03 -08001291 /* old_pool */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001292 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weil4fc51be2010-02-16 15:55:03 -08001293 while (len--) {
1294 struct ceph_pg_pool_info *pi;
1295
Ilya Dryomov86f17422014-03-13 16:36:15 +02001296 ceph_decode_64_safe(p, end, pool, e_inval);
Sage Weil4fc51be2010-02-16 15:55:03 -08001297 pi = __lookup_pg_pool(&map->pg_pools, pool);
Sage Weil2844a762010-04-09 15:46:42 -07001298 if (pi)
1299 __remove_pg_pool(&map->pg_pools, pi);
Sage Weil4fc51be2010-02-16 15:55:03 -08001300 }
Sage Weilf24e9982009-10-06 11:31:10 -07001301
1302 /* new_up */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001303 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001304 while (len--) {
1305 u32 osd;
1306 struct ceph_entity_addr addr;
Ilya Dryomov86f17422014-03-13 16:36:15 +02001307 ceph_decode_32_safe(p, end, osd, e_inval);
1308 ceph_decode_copy_safe(p, end, &addr, sizeof(addr), e_inval);
Sage Weil63f2d212009-11-03 15:17:56 -08001309 ceph_decode_addr(&addr);
Sage Weilf24e9982009-10-06 11:31:10 -07001310 pr_info("osd%d up\n", osd);
1311 BUG_ON(osd >= map->max_osd);
Yan, Zheng6dd74e42015-08-28 17:59:35 +08001312 map->osd_state[osd] |= CEPH_OSD_UP | CEPH_OSD_EXISTS;
Sage Weilf24e9982009-10-06 11:31:10 -07001313 map->osd_addr[osd] = addr;
1314 }
1315
Sage Weil7662d8f2011-05-03 12:52:05 -07001316 /* new_state */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001317 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001318 while (len--) {
1319 u32 osd;
Sage Weil7662d8f2011-05-03 12:52:05 -07001320 u8 xorstate;
Ilya Dryomov86f17422014-03-13 16:36:15 +02001321 ceph_decode_32_safe(p, end, osd, e_inval);
Sage Weil7662d8f2011-05-03 12:52:05 -07001322 xorstate = **(u8 **)p;
Sage Weilf24e9982009-10-06 11:31:10 -07001323 (*p)++; /* clean flag */
Sage Weil7662d8f2011-05-03 12:52:05 -07001324 if (xorstate == 0)
1325 xorstate = CEPH_OSD_UP;
1326 if (xorstate & CEPH_OSD_UP)
1327 pr_info("osd%d down\n", osd);
Sage Weilf24e9982009-10-06 11:31:10 -07001328 if (osd < map->max_osd)
Sage Weil7662d8f2011-05-03 12:52:05 -07001329 map->osd_state[osd] ^= xorstate;
Sage Weilf24e9982009-10-06 11:31:10 -07001330 }
1331
1332 /* new_weight */
Ilya Dryomov86f17422014-03-13 16:36:15 +02001333 ceph_decode_32_safe(p, end, len, e_inval);
Sage Weilf24e9982009-10-06 11:31:10 -07001334 while (len--) {
1335 u32 osd, off;
Ilya Dryomov86f17422014-03-13 16:36:15 +02001336 ceph_decode_need(p, end, sizeof(u32)*2, e_inval);
Sage Weilc89136e2009-10-14 09:59:09 -07001337 osd = ceph_decode_32(p);
1338 off = ceph_decode_32(p);
Sage Weilf24e9982009-10-06 11:31:10 -07001339 pr_info("osd%d weight 0x%x %s\n", osd, off,
1340 off == CEPH_OSD_IN ? "(in)" :
1341 (off == CEPH_OSD_OUT ? "(out)" : ""));
1342 if (osd < map->max_osd)
1343 map->osd_weight[osd] = off;
1344 }
1345
1346 /* new_pg_temp */
Ilya Dryomov10db6342014-03-21 19:05:28 +02001347 err = decode_new_pg_temp(p, end, map);
1348 if (err)
1349 goto bad;
Sage Weilf24e9982009-10-06 11:31:10 -07001350
Ilya Dryomovd286de72014-03-21 19:05:30 +02001351 /* new_primary_temp */
1352 if (struct_v >= 1) {
1353 err = decode_new_primary_temp(p, end, map);
1354 if (err)
1355 goto bad;
1356 }
1357
Ilya Dryomov63a69932014-03-21 19:05:30 +02001358 /* new_primary_affinity */
1359 if (struct_v >= 2) {
1360 err = decode_new_primary_affinity(p, end, map);
1361 if (err)
1362 goto bad;
1363 }
1364
Sage Weilf24e9982009-10-06 11:31:10 -07001365 /* ignore the rest */
1366 *p = end;
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001367
1368 dout("inc osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd);
Sage Weilf24e9982009-10-06 11:31:10 -07001369 return map;
1370
Ilya Dryomov86f17422014-03-13 16:36:15 +02001371e_inval:
1372 err = -EINVAL;
Sage Weilf24e9982009-10-06 11:31:10 -07001373bad:
Ilya Dryomov38a8d562014-03-13 16:36:13 +02001374 pr_err("corrupt inc osdmap (%d) epoch %d off %d (%p of %p-%p)\n",
1375 err, epoch, (int)(*p - start), *p, start, end);
Sage Weil9ec7cab2009-12-14 15:13:47 -08001376 print_hex_dump(KERN_DEBUG, "osdmap: ",
1377 DUMP_PREFIX_OFFSET, 16, 1,
1378 start, end - start, true);
Sage Weilf24e9982009-10-06 11:31:10 -07001379 if (newcrush)
1380 crush_destroy(newcrush);
1381 return ERR_PTR(err);
1382}
1383
1384
1385
1386
1387/*
1388 * calculate file layout from given offset, length.
1389 * fill in correct oid, logical length, and object extent
1390 * offset, length.
1391 *
1392 * for now, we write only a single su, until we can
1393 * pass a stride back to the caller.
1394 */
Sage Weild63b77f2012-09-24 20:59:48 -07001395int ceph_calc_file_object_mapping(struct ceph_file_layout *layout,
Alex Eldere8afad62012-11-14 09:38:19 -06001396 u64 off, u64 len,
Sage Weil645a1022009-10-28 15:15:05 -07001397 u64 *ono,
Sage Weilf24e9982009-10-06 11:31:10 -07001398 u64 *oxoff, u64 *oxlen)
1399{
1400 u32 osize = le32_to_cpu(layout->fl_object_size);
1401 u32 su = le32_to_cpu(layout->fl_stripe_unit);
1402 u32 sc = le32_to_cpu(layout->fl_stripe_count);
1403 u32 bl, stripeno, stripepos, objsetno;
1404 u32 su_per_object;
Noah Watkinsff1d1f72009-10-30 12:57:30 -07001405 u64 t, su_offset;
Sage Weilf24e9982009-10-06 11:31:10 -07001406
Alex Eldere8afad62012-11-14 09:38:19 -06001407 dout("mapping %llu~%llu osize %u fl_su %u\n", off, len,
Sage Weilf24e9982009-10-06 11:31:10 -07001408 osize, su);
Sage Weild63b77f2012-09-24 20:59:48 -07001409 if (su == 0 || sc == 0)
1410 goto invalid;
Noah Watkins35e054a2009-10-28 14:04:48 -07001411 su_per_object = osize / su;
Sage Weild63b77f2012-09-24 20:59:48 -07001412 if (su_per_object == 0)
1413 goto invalid;
Sage Weilf24e9982009-10-06 11:31:10 -07001414 dout("osize %u / su %u = su_per_object %u\n", osize, su,
1415 su_per_object);
1416
Sage Weild63b77f2012-09-24 20:59:48 -07001417 if ((su & ~PAGE_MASK) != 0)
1418 goto invalid;
1419
Sage Weilf24e9982009-10-06 11:31:10 -07001420 /* bl = *off / su; */
1421 t = off;
1422 do_div(t, su);
1423 bl = t;
1424 dout("off %llu / su %u = bl %u\n", off, su, bl);
1425
1426 stripeno = bl / sc;
1427 stripepos = bl % sc;
1428 objsetno = stripeno / su_per_object;
1429
Sage Weil645a1022009-10-28 15:15:05 -07001430 *ono = objsetno * sc + stripepos;
Eric Dumazet95c96172012-04-15 05:58:06 +00001431 dout("objset %u * sc %u = ono %u\n", objsetno, sc, (unsigned int)*ono);
Sage Weil645a1022009-10-28 15:15:05 -07001432
1433 /* *oxoff = *off % layout->fl_stripe_unit; # offset in su */
Sage Weilf24e9982009-10-06 11:31:10 -07001434 t = off;
Noah Watkinsff1d1f72009-10-30 12:57:30 -07001435 su_offset = do_div(t, su);
1436 *oxoff = su_offset + (stripeno % su_per_object) * su;
Sage Weil645a1022009-10-28 15:15:05 -07001437
Noah Watkinsff1d1f72009-10-30 12:57:30 -07001438 /*
1439 * Calculate the length of the extent being written to the selected
Alex Eldere8afad62012-11-14 09:38:19 -06001440 * object. This is the minimum of the full length requested (len) or
Noah Watkinsff1d1f72009-10-30 12:57:30 -07001441 * the remainder of the current stripe being written to.
1442 */
Alex Eldere8afad62012-11-14 09:38:19 -06001443 *oxlen = min_t(u64, len, su - su_offset);
Sage Weilf24e9982009-10-06 11:31:10 -07001444
1445 dout(" obj extent %llu~%llu\n", *oxoff, *oxlen);
Sage Weild63b77f2012-09-24 20:59:48 -07001446 return 0;
1447
1448invalid:
1449 dout(" invalid layout\n");
1450 *ono = 0;
1451 *oxoff = 0;
1452 *oxlen = 0;
1453 return -EINVAL;
Sage Weilf24e9982009-10-06 11:31:10 -07001454}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001455EXPORT_SYMBOL(ceph_calc_file_object_mapping);
Sage Weilf24e9982009-10-06 11:31:10 -07001456
1457/*
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001458 * Calculate mapping of a (oloc, oid) pair to a PG. Should only be
1459 * called with target's (oloc, oid), since tiering isn't taken into
1460 * account.
Sage Weilf24e9982009-10-06 11:31:10 -07001461 */
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001462int ceph_oloc_oid_to_pg(struct ceph_osdmap *osdmap,
1463 struct ceph_object_locator *oloc,
1464 struct ceph_object_id *oid,
1465 struct ceph_pg *pg_out)
Sage Weilf24e9982009-10-06 11:31:10 -07001466{
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001467 struct ceph_pg_pool_info *pi;
Sage Weilf24e9982009-10-06 11:31:10 -07001468
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001469 pi = __lookup_pg_pool(&osdmap->pg_pools, oloc->pool);
1470 if (!pi)
Sage Weil4fc51be2010-02-16 15:55:03 -08001471 return -EIO;
Sage Weilf24e9982009-10-06 11:31:10 -07001472
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001473 pg_out->pool = oloc->pool;
1474 pg_out->seed = ceph_str_hash(pi->object_hash, oid->name,
1475 oid->name_len);
1476
1477 dout("%s '%.*s' pgid %llu.%x\n", __func__, oid->name_len, oid->name,
1478 pg_out->pool, pg_out->seed);
Sage Weilf24e9982009-10-06 11:31:10 -07001479 return 0;
1480}
Ilya Dryomov7c13cb62014-01-27 17:40:19 +02001481EXPORT_SYMBOL(ceph_oloc_oid_to_pg);
Sage Weilf24e9982009-10-06 11:31:10 -07001482
Ilya Dryomov9d521472014-01-31 17:54:26 +02001483static int do_crush(struct ceph_osdmap *map, int ruleno, int x,
1484 int *result, int result_max,
1485 const __u32 *weight, int weight_max)
Ilya Dryomove8ef19c2013-12-24 21:19:24 +02001486{
Ilya Dryomov9d521472014-01-31 17:54:26 +02001487 int r;
Ilya Dryomove8ef19c2013-12-24 21:19:24 +02001488
Ilya Dryomov9d521472014-01-31 17:54:26 +02001489 BUG_ON(result_max > CEPH_PG_MAX_SIZE);
1490
1491 mutex_lock(&map->crush_scratch_mutex);
1492 r = crush_do_rule(map->crush, ruleno, x, result, result_max,
1493 weight, weight_max, map->crush_scratch_ary);
1494 mutex_unlock(&map->crush_scratch_mutex);
1495
1496 return r;
Ilya Dryomove8ef19c2013-12-24 21:19:24 +02001497}
1498
Sage Weilf24e9982009-10-06 11:31:10 -07001499/*
Ilya Dryomov2bd93d42014-03-24 17:12:47 +02001500 * Calculate raw (crush) set for given pgid.
1501 *
1502 * Return raw set length, or error.
1503 */
1504static int pg_to_raw_osds(struct ceph_osdmap *osdmap,
1505 struct ceph_pg_pool_info *pool,
1506 struct ceph_pg pgid, u32 pps, int *osds)
1507{
1508 int ruleno;
1509 int len;
1510
1511 /* crush */
1512 ruleno = crush_find_rule(osdmap->crush, pool->crush_ruleset,
1513 pool->type, pool->size);
1514 if (ruleno < 0) {
1515 pr_err("no crush rule: pool %lld ruleset %d type %d size %d\n",
1516 pgid.pool, pool->crush_ruleset, pool->type,
1517 pool->size);
1518 return -ENOENT;
1519 }
1520
1521 len = do_crush(osdmap, ruleno, pps, osds,
1522 min_t(int, pool->size, CEPH_PG_MAX_SIZE),
1523 osdmap->osd_weight, osdmap->max_osd);
1524 if (len < 0) {
1525 pr_err("error %d from crush rule %d: pool %lld ruleset %d type %d size %d\n",
1526 len, ruleno, pgid.pool, pool->crush_ruleset,
1527 pool->type, pool->size);
1528 return len;
1529 }
1530
1531 return len;
1532}
1533
1534/*
1535 * Given raw set, calculate up set and up primary.
1536 *
1537 * Return up set length. *primary is set to up primary osd id, or -1
1538 * if up set is empty.
1539 */
1540static int raw_to_up_osds(struct ceph_osdmap *osdmap,
1541 struct ceph_pg_pool_info *pool,
1542 int *osds, int len, int *primary)
1543{
1544 int up_primary = -1;
1545 int i;
1546
1547 if (ceph_can_shift_osds(pool)) {
1548 int removed = 0;
1549
1550 for (i = 0; i < len; i++) {
1551 if (ceph_osd_is_down(osdmap, osds[i])) {
1552 removed++;
1553 continue;
1554 }
1555 if (removed)
1556 osds[i - removed] = osds[i];
1557 }
1558
1559 len -= removed;
1560 if (len > 0)
1561 up_primary = osds[0];
1562 } else {
1563 for (i = len - 1; i >= 0; i--) {
1564 if (ceph_osd_is_down(osdmap, osds[i]))
1565 osds[i] = CRUSH_ITEM_NONE;
1566 else
1567 up_primary = osds[i];
1568 }
1569 }
1570
1571 *primary = up_primary;
1572 return len;
1573}
1574
Ilya Dryomov47ec1f32014-03-24 17:12:49 +02001575static void apply_primary_affinity(struct ceph_osdmap *osdmap, u32 pps,
1576 struct ceph_pg_pool_info *pool,
1577 int *osds, int len, int *primary)
1578{
1579 int i;
1580 int pos = -1;
1581
1582 /*
1583 * Do we have any non-default primary_affinity values for these
1584 * osds?
1585 */
1586 if (!osdmap->osd_primary_affinity)
1587 return;
1588
1589 for (i = 0; i < len; i++) {
Ilya Dryomov92b2e752014-04-10 18:09:41 +04001590 int osd = osds[i];
1591
1592 if (osd != CRUSH_ITEM_NONE &&
1593 osdmap->osd_primary_affinity[osd] !=
Ilya Dryomov47ec1f32014-03-24 17:12:49 +02001594 CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) {
1595 break;
1596 }
1597 }
1598 if (i == len)
1599 return;
1600
1601 /*
1602 * Pick the primary. Feed both the seed (for the pg) and the
1603 * osd into the hash/rng so that a proportional fraction of an
1604 * osd's pgs get rejected as primary.
1605 */
1606 for (i = 0; i < len; i++) {
Ilya Dryomov92b2e752014-04-10 18:09:41 +04001607 int osd = osds[i];
Ilya Dryomov47ec1f32014-03-24 17:12:49 +02001608 u32 aff;
1609
Ilya Dryomov47ec1f32014-03-24 17:12:49 +02001610 if (osd == CRUSH_ITEM_NONE)
1611 continue;
1612
1613 aff = osdmap->osd_primary_affinity[osd];
1614 if (aff < CEPH_OSD_MAX_PRIMARY_AFFINITY &&
1615 (crush_hash32_2(CRUSH_HASH_RJENKINS1,
1616 pps, osd) >> 16) >= aff) {
1617 /*
1618 * We chose not to use this primary. Note it
1619 * anyway as a fallback in case we don't pick
1620 * anyone else, but keep looking.
1621 */
1622 if (pos < 0)
1623 pos = i;
1624 } else {
1625 pos = i;
1626 break;
1627 }
1628 }
1629 if (pos < 0)
1630 return;
1631
1632 *primary = osds[pos];
1633
1634 if (ceph_can_shift_osds(pool) && pos > 0) {
1635 /* move the new primary to the front */
1636 for (i = pos; i > 0; i--)
1637 osds[i] = osds[i - 1];
1638 osds[0] = *primary;
1639 }
1640}
1641
Ilya Dryomov2bd93d42014-03-24 17:12:47 +02001642/*
Ilya Dryomov5e8d4d32014-03-24 17:12:48 +02001643 * Given up set, apply pg_temp and primary_temp mappings.
Ilya Dryomov45966c32014-03-24 17:12:47 +02001644 *
1645 * Return acting set length. *primary is set to acting primary osd id,
1646 * or -1 if acting set is empty.
1647 */
1648static int apply_temps(struct ceph_osdmap *osdmap,
1649 struct ceph_pg_pool_info *pool, struct ceph_pg pgid,
1650 int *osds, int len, int *primary)
1651{
1652 struct ceph_pg_mapping *pg;
1653 int temp_len;
1654 int temp_primary;
1655 int i;
1656
1657 /* raw_pg -> pg */
1658 pgid.seed = ceph_stable_mod(pgid.seed, pool->pg_num,
1659 pool->pg_num_mask);
1660
1661 /* pg_temp? */
1662 pg = __lookup_pg_mapping(&osdmap->pg_temp, pgid);
1663 if (pg) {
1664 temp_len = 0;
1665 temp_primary = -1;
1666
1667 for (i = 0; i < pg->pg_temp.len; i++) {
1668 if (ceph_osd_is_down(osdmap, pg->pg_temp.osds[i])) {
1669 if (ceph_can_shift_osds(pool))
1670 continue;
1671 else
1672 osds[temp_len++] = CRUSH_ITEM_NONE;
1673 } else {
1674 osds[temp_len++] = pg->pg_temp.osds[i];
1675 }
1676 }
1677
1678 /* apply pg_temp's primary */
1679 for (i = 0; i < temp_len; i++) {
1680 if (osds[i] != CRUSH_ITEM_NONE) {
1681 temp_primary = osds[i];
1682 break;
1683 }
1684 }
1685 } else {
1686 temp_len = len;
1687 temp_primary = *primary;
1688 }
1689
Ilya Dryomov5e8d4d32014-03-24 17:12:48 +02001690 /* primary_temp? */
1691 pg = __lookup_pg_mapping(&osdmap->primary_temp, pgid);
1692 if (pg)
1693 temp_primary = pg->primary_temp.osd;
1694
Ilya Dryomov45966c32014-03-24 17:12:47 +02001695 *primary = temp_primary;
1696 return temp_len;
1697}
1698
1699/*
Ilya Dryomovac972232014-03-24 17:12:48 +02001700 * Calculate acting set for given pgid.
1701 *
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001702 * Return acting set length, or error. *primary is set to acting
1703 * primary osd id, or -1 if acting set is empty or on error.
Sage Weild85b7052010-05-10 10:24:48 -07001704 */
Sage Weil5b191d92013-02-23 10:38:16 -08001705int ceph_calc_pg_acting(struct ceph_osdmap *osdmap, struct ceph_pg pgid,
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001706 int *osds, int *primary)
Sage Weild85b7052010-05-10 10:24:48 -07001707{
Ilya Dryomovac972232014-03-24 17:12:48 +02001708 struct ceph_pg_pool_info *pool;
1709 u32 pps;
1710 int len;
Sage Weild85b7052010-05-10 10:24:48 -07001711
Ilya Dryomovac972232014-03-24 17:12:48 +02001712 pool = __lookup_pg_pool(&osdmap->pg_pools, pgid.pool);
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001713 if (!pool) {
1714 *primary = -1;
1715 return -ENOENT;
1716 }
Sage Weild85b7052010-05-10 10:24:48 -07001717
Ilya Dryomovac972232014-03-24 17:12:48 +02001718 if (pool->flags & CEPH_POOL_FLAG_HASHPSPOOL) {
1719 /* hash pool id and seed so that pool PGs do not overlap */
1720 pps = crush_hash32_2(CRUSH_HASH_RJENKINS1,
1721 ceph_stable_mod(pgid.seed, pool->pgp_num,
1722 pool->pgp_num_mask),
1723 pgid.pool);
1724 } else {
1725 /*
1726 * legacy behavior: add ps and pool together. this is
1727 * not a great approach because the PGs from each pool
1728 * will overlap on top of each other: 0.5 == 1.4 ==
1729 * 2.3 == ...
1730 */
1731 pps = ceph_stable_mod(pgid.seed, pool->pgp_num,
1732 pool->pgp_num_mask) +
1733 (unsigned)pgid.pool;
1734 }
1735
1736 len = pg_to_raw_osds(osdmap, pool, pgid, pps, osds);
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001737 if (len < 0) {
1738 *primary = -1;
Ilya Dryomovac972232014-03-24 17:12:48 +02001739 return len;
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001740 }
Ilya Dryomovac972232014-03-24 17:12:48 +02001741
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001742 len = raw_to_up_osds(osdmap, pool, osds, len, primary);
Ilya Dryomovac972232014-03-24 17:12:48 +02001743
Ilya Dryomov47ec1f32014-03-24 17:12:49 +02001744 apply_primary_affinity(osdmap, pps, pool, osds, len, primary);
1745
Ilya Dryomov8008ab12014-03-24 17:12:48 +02001746 len = apply_temps(osdmap, pool, pgid, osds, len, primary);
Ilya Dryomovac972232014-03-24 17:12:48 +02001747
1748 return len;
Sage Weild85b7052010-05-10 10:24:48 -07001749}
1750
1751/*
Sage Weilf24e9982009-10-06 11:31:10 -07001752 * Return primary osd for given pgid, or -1 if none.
1753 */
Sage Weil5b191d92013-02-23 10:38:16 -08001754int ceph_calc_pg_primary(struct ceph_osdmap *osdmap, struct ceph_pg pgid)
Sage Weilf24e9982009-10-06 11:31:10 -07001755{
Ilya Dryomovc4c12282014-03-24 17:12:49 +02001756 int osds[CEPH_PG_MAX_SIZE];
1757 int primary;
Sage Weilf24e9982009-10-06 11:31:10 -07001758
Ilya Dryomovc4c12282014-03-24 17:12:49 +02001759 ceph_calc_pg_acting(osdmap, pgid, osds, &primary);
Sage Weilf24e9982009-10-06 11:31:10 -07001760
Ilya Dryomovc4c12282014-03-24 17:12:49 +02001761 return primary;
Sage Weilf24e9982009-10-06 11:31:10 -07001762}
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001763EXPORT_SYMBOL(ceph_calc_pg_primary);