13d14c5d2SYehuda Sadeh 23d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 33d14c5d2SYehuda Sadeh 43d14c5d2SYehuda Sadeh #include <linux/module.h> 53d14c5d2SYehuda Sadeh #include <linux/slab.h> 63d14c5d2SYehuda Sadeh #include <asm/div64.h> 73d14c5d2SYehuda Sadeh 83d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 93d14c5d2SYehuda Sadeh #include <linux/ceph/osdmap.h> 103d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 113d14c5d2SYehuda Sadeh #include <linux/crush/hash.h> 123d14c5d2SYehuda Sadeh #include <linux/crush/mapper.h> 133d14c5d2SYehuda Sadeh 143d14c5d2SYehuda Sadeh char *ceph_osdmap_state_str(char *str, int len, int state) 153d14c5d2SYehuda Sadeh { 163d14c5d2SYehuda Sadeh if (!len) 171ec3911dSCong Ding return str; 183d14c5d2SYehuda Sadeh 191ec3911dSCong Ding if ((state & CEPH_OSD_EXISTS) && (state & CEPH_OSD_UP)) 201ec3911dSCong Ding snprintf(str, len, "exists, up"); 211ec3911dSCong Ding else if (state & CEPH_OSD_EXISTS) 223d14c5d2SYehuda Sadeh snprintf(str, len, "exists"); 231ec3911dSCong Ding else if (state & CEPH_OSD_UP) 241ec3911dSCong Ding snprintf(str, len, "up"); 251ec3911dSCong Ding else 263d14c5d2SYehuda Sadeh snprintf(str, len, "doesn't exist"); 271ec3911dSCong Ding 283d14c5d2SYehuda Sadeh return str; 293d14c5d2SYehuda Sadeh } 303d14c5d2SYehuda Sadeh 313d14c5d2SYehuda Sadeh /* maps */ 323d14c5d2SYehuda Sadeh 3395c96174SEric Dumazet static int calc_bits_of(unsigned int t) 343d14c5d2SYehuda Sadeh { 353d14c5d2SYehuda Sadeh int b = 0; 363d14c5d2SYehuda Sadeh while (t) { 373d14c5d2SYehuda Sadeh t = t >> 1; 383d14c5d2SYehuda Sadeh b++; 393d14c5d2SYehuda Sadeh } 403d14c5d2SYehuda Sadeh return b; 413d14c5d2SYehuda Sadeh } 423d14c5d2SYehuda Sadeh 433d14c5d2SYehuda Sadeh /* 443d14c5d2SYehuda Sadeh * the foo_mask is the smallest value 2^n-1 that is >= foo. 453d14c5d2SYehuda Sadeh */ 463d14c5d2SYehuda Sadeh static void calc_pg_masks(struct ceph_pg_pool_info *pi) 473d14c5d2SYehuda Sadeh { 484f6a7e5eSSage Weil pi->pg_num_mask = (1 << calc_bits_of(pi->pg_num-1)) - 1; 494f6a7e5eSSage Weil pi->pgp_num_mask = (1 << calc_bits_of(pi->pgp_num-1)) - 1; 503d14c5d2SYehuda Sadeh } 513d14c5d2SYehuda Sadeh 523d14c5d2SYehuda Sadeh /* 533d14c5d2SYehuda Sadeh * decode crush map 543d14c5d2SYehuda Sadeh */ 553d14c5d2SYehuda Sadeh static int crush_decode_uniform_bucket(void **p, void *end, 563d14c5d2SYehuda Sadeh struct crush_bucket_uniform *b) 573d14c5d2SYehuda Sadeh { 583d14c5d2SYehuda Sadeh dout("crush_decode_uniform_bucket %p to %p\n", *p, end); 593d14c5d2SYehuda Sadeh ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad); 603d14c5d2SYehuda Sadeh b->item_weight = ceph_decode_32(p); 613d14c5d2SYehuda Sadeh return 0; 623d14c5d2SYehuda Sadeh bad: 633d14c5d2SYehuda Sadeh return -EINVAL; 643d14c5d2SYehuda Sadeh } 653d14c5d2SYehuda Sadeh 663d14c5d2SYehuda Sadeh static int crush_decode_list_bucket(void **p, void *end, 673d14c5d2SYehuda Sadeh struct crush_bucket_list *b) 683d14c5d2SYehuda Sadeh { 693d14c5d2SYehuda Sadeh int j; 703d14c5d2SYehuda Sadeh dout("crush_decode_list_bucket %p to %p\n", *p, end); 713d14c5d2SYehuda Sadeh b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); 723d14c5d2SYehuda Sadeh if (b->item_weights == NULL) 733d14c5d2SYehuda Sadeh return -ENOMEM; 743d14c5d2SYehuda Sadeh b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); 753d14c5d2SYehuda Sadeh if (b->sum_weights == NULL) 763d14c5d2SYehuda Sadeh return -ENOMEM; 773d14c5d2SYehuda Sadeh ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad); 783d14c5d2SYehuda Sadeh for (j = 0; j < b->h.size; j++) { 793d14c5d2SYehuda Sadeh b->item_weights[j] = ceph_decode_32(p); 803d14c5d2SYehuda Sadeh b->sum_weights[j] = ceph_decode_32(p); 813d14c5d2SYehuda Sadeh } 823d14c5d2SYehuda Sadeh return 0; 833d14c5d2SYehuda Sadeh bad: 843d14c5d2SYehuda Sadeh return -EINVAL; 853d14c5d2SYehuda Sadeh } 863d14c5d2SYehuda Sadeh 873d14c5d2SYehuda Sadeh static int crush_decode_tree_bucket(void **p, void *end, 883d14c5d2SYehuda Sadeh struct crush_bucket_tree *b) 893d14c5d2SYehuda Sadeh { 903d14c5d2SYehuda Sadeh int j; 913d14c5d2SYehuda Sadeh dout("crush_decode_tree_bucket %p to %p\n", *p, end); 9282cd003aSIlya Dryomov ceph_decode_8_safe(p, end, b->num_nodes, bad); 933d14c5d2SYehuda Sadeh b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS); 943d14c5d2SYehuda Sadeh if (b->node_weights == NULL) 953d14c5d2SYehuda Sadeh return -ENOMEM; 963d14c5d2SYehuda Sadeh ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad); 973d14c5d2SYehuda Sadeh for (j = 0; j < b->num_nodes; j++) 983d14c5d2SYehuda Sadeh b->node_weights[j] = ceph_decode_32(p); 993d14c5d2SYehuda Sadeh return 0; 1003d14c5d2SYehuda Sadeh bad: 1013d14c5d2SYehuda Sadeh return -EINVAL; 1023d14c5d2SYehuda Sadeh } 1033d14c5d2SYehuda Sadeh 1043d14c5d2SYehuda Sadeh static int crush_decode_straw_bucket(void **p, void *end, 1053d14c5d2SYehuda Sadeh struct crush_bucket_straw *b) 1063d14c5d2SYehuda Sadeh { 1073d14c5d2SYehuda Sadeh int j; 1083d14c5d2SYehuda Sadeh dout("crush_decode_straw_bucket %p to %p\n", *p, end); 1093d14c5d2SYehuda Sadeh b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); 1103d14c5d2SYehuda Sadeh if (b->item_weights == NULL) 1113d14c5d2SYehuda Sadeh return -ENOMEM; 1123d14c5d2SYehuda Sadeh b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); 1133d14c5d2SYehuda Sadeh if (b->straws == NULL) 1143d14c5d2SYehuda Sadeh return -ENOMEM; 1153d14c5d2SYehuda Sadeh ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad); 1163d14c5d2SYehuda Sadeh for (j = 0; j < b->h.size; j++) { 1173d14c5d2SYehuda Sadeh b->item_weights[j] = ceph_decode_32(p); 1183d14c5d2SYehuda Sadeh b->straws[j] = ceph_decode_32(p); 1193d14c5d2SYehuda Sadeh } 1203d14c5d2SYehuda Sadeh return 0; 1213d14c5d2SYehuda Sadeh bad: 1223d14c5d2SYehuda Sadeh return -EINVAL; 1233d14c5d2SYehuda Sadeh } 1243d14c5d2SYehuda Sadeh 125958a2765SIlya Dryomov static int crush_decode_straw2_bucket(void **p, void *end, 126958a2765SIlya Dryomov struct crush_bucket_straw2 *b) 127958a2765SIlya Dryomov { 128958a2765SIlya Dryomov int j; 129958a2765SIlya Dryomov dout("crush_decode_straw2_bucket %p to %p\n", *p, end); 130958a2765SIlya Dryomov b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS); 131958a2765SIlya Dryomov if (b->item_weights == NULL) 132958a2765SIlya Dryomov return -ENOMEM; 133958a2765SIlya Dryomov ceph_decode_need(p, end, b->h.size * sizeof(u32), bad); 134958a2765SIlya Dryomov for (j = 0; j < b->h.size; j++) 135958a2765SIlya Dryomov b->item_weights[j] = ceph_decode_32(p); 136958a2765SIlya Dryomov return 0; 137958a2765SIlya Dryomov bad: 138958a2765SIlya Dryomov return -EINVAL; 139958a2765SIlya Dryomov } 140958a2765SIlya Dryomov 141546f04efSSage Weil static int skip_name_map(void **p, void *end) 142546f04efSSage Weil { 143546f04efSSage Weil int len; 144546f04efSSage Weil ceph_decode_32_safe(p, end, len ,bad); 145546f04efSSage Weil while (len--) { 146546f04efSSage Weil int strlen; 147546f04efSSage Weil *p += sizeof(u32); 148546f04efSSage Weil ceph_decode_32_safe(p, end, strlen, bad); 149546f04efSSage Weil *p += strlen; 150546f04efSSage Weil } 151546f04efSSage Weil return 0; 152546f04efSSage Weil bad: 153546f04efSSage Weil return -EINVAL; 154546f04efSSage Weil } 155546f04efSSage Weil 15666a0e2d5SIlya Dryomov static void crush_finalize(struct crush_map *c) 15766a0e2d5SIlya Dryomov { 15866a0e2d5SIlya Dryomov __s32 b; 15966a0e2d5SIlya Dryomov 16066a0e2d5SIlya Dryomov /* Space for the array of pointers to per-bucket workspace */ 16166a0e2d5SIlya Dryomov c->working_size = sizeof(struct crush_work) + 16266a0e2d5SIlya Dryomov c->max_buckets * sizeof(struct crush_work_bucket *); 16366a0e2d5SIlya Dryomov 16466a0e2d5SIlya Dryomov for (b = 0; b < c->max_buckets; b++) { 16566a0e2d5SIlya Dryomov if (!c->buckets[b]) 16666a0e2d5SIlya Dryomov continue; 16766a0e2d5SIlya Dryomov 16866a0e2d5SIlya Dryomov switch (c->buckets[b]->alg) { 16966a0e2d5SIlya Dryomov default: 17066a0e2d5SIlya Dryomov /* 17166a0e2d5SIlya Dryomov * The base case, permutation variables and 17266a0e2d5SIlya Dryomov * the pointer to the permutation array. 17366a0e2d5SIlya Dryomov */ 17466a0e2d5SIlya Dryomov c->working_size += sizeof(struct crush_work_bucket); 17566a0e2d5SIlya Dryomov break; 17666a0e2d5SIlya Dryomov } 17766a0e2d5SIlya Dryomov /* Every bucket has a permutation array. */ 17866a0e2d5SIlya Dryomov c->working_size += c->buckets[b]->size * sizeof(__u32); 17966a0e2d5SIlya Dryomov } 18066a0e2d5SIlya Dryomov } 18166a0e2d5SIlya Dryomov 1823d14c5d2SYehuda Sadeh static struct crush_map *crush_decode(void *pbyval, void *end) 1833d14c5d2SYehuda Sadeh { 1843d14c5d2SYehuda Sadeh struct crush_map *c; 1853d14c5d2SYehuda Sadeh int err = -EINVAL; 1863d14c5d2SYehuda Sadeh int i, j; 1873d14c5d2SYehuda Sadeh void **p = &pbyval; 1883d14c5d2SYehuda Sadeh void *start = pbyval; 1893d14c5d2SYehuda Sadeh u32 magic; 190546f04efSSage Weil u32 num_name_maps; 1913d14c5d2SYehuda Sadeh 1923d14c5d2SYehuda Sadeh dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p)); 1933d14c5d2SYehuda Sadeh 1943d14c5d2SYehuda Sadeh c = kzalloc(sizeof(*c), GFP_NOFS); 1953d14c5d2SYehuda Sadeh if (c == NULL) 1963d14c5d2SYehuda Sadeh return ERR_PTR(-ENOMEM); 1973d14c5d2SYehuda Sadeh 198546f04efSSage Weil /* set tunables to default values */ 199546f04efSSage Weil c->choose_local_tries = 2; 200546f04efSSage Weil c->choose_local_fallback_tries = 5; 201546f04efSSage Weil c->choose_total_tries = 19; 2021604f488SJim Schutt c->chooseleaf_descend_once = 0; 203546f04efSSage Weil 2043d14c5d2SYehuda Sadeh ceph_decode_need(p, end, 4*sizeof(u32), bad); 2053d14c5d2SYehuda Sadeh magic = ceph_decode_32(p); 2063d14c5d2SYehuda Sadeh if (magic != CRUSH_MAGIC) { 2073d14c5d2SYehuda Sadeh pr_err("crush_decode magic %x != current %x\n", 20895c96174SEric Dumazet (unsigned int)magic, (unsigned int)CRUSH_MAGIC); 2093d14c5d2SYehuda Sadeh goto bad; 2103d14c5d2SYehuda Sadeh } 2113d14c5d2SYehuda Sadeh c->max_buckets = ceph_decode_32(p); 2123d14c5d2SYehuda Sadeh c->max_rules = ceph_decode_32(p); 2133d14c5d2SYehuda Sadeh c->max_devices = ceph_decode_32(p); 2143d14c5d2SYehuda Sadeh 2153d14c5d2SYehuda Sadeh c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS); 2163d14c5d2SYehuda Sadeh if (c->buckets == NULL) 2173d14c5d2SYehuda Sadeh goto badmem; 2183d14c5d2SYehuda Sadeh c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS); 2193d14c5d2SYehuda Sadeh if (c->rules == NULL) 2203d14c5d2SYehuda Sadeh goto badmem; 2213d14c5d2SYehuda Sadeh 2223d14c5d2SYehuda Sadeh /* buckets */ 2233d14c5d2SYehuda Sadeh for (i = 0; i < c->max_buckets; i++) { 2243d14c5d2SYehuda Sadeh int size = 0; 2253d14c5d2SYehuda Sadeh u32 alg; 2263d14c5d2SYehuda Sadeh struct crush_bucket *b; 2273d14c5d2SYehuda Sadeh 2283d14c5d2SYehuda Sadeh ceph_decode_32_safe(p, end, alg, bad); 2293d14c5d2SYehuda Sadeh if (alg == 0) { 2303d14c5d2SYehuda Sadeh c->buckets[i] = NULL; 2313d14c5d2SYehuda Sadeh continue; 2323d14c5d2SYehuda Sadeh } 2333d14c5d2SYehuda Sadeh dout("crush_decode bucket %d off %x %p to %p\n", 2343d14c5d2SYehuda Sadeh i, (int)(*p-start), *p, end); 2353d14c5d2SYehuda Sadeh 2363d14c5d2SYehuda Sadeh switch (alg) { 2373d14c5d2SYehuda Sadeh case CRUSH_BUCKET_UNIFORM: 2383d14c5d2SYehuda Sadeh size = sizeof(struct crush_bucket_uniform); 2393d14c5d2SYehuda Sadeh break; 2403d14c5d2SYehuda Sadeh case CRUSH_BUCKET_LIST: 2413d14c5d2SYehuda Sadeh size = sizeof(struct crush_bucket_list); 2423d14c5d2SYehuda Sadeh break; 2433d14c5d2SYehuda Sadeh case CRUSH_BUCKET_TREE: 2443d14c5d2SYehuda Sadeh size = sizeof(struct crush_bucket_tree); 2453d14c5d2SYehuda Sadeh break; 2463d14c5d2SYehuda Sadeh case CRUSH_BUCKET_STRAW: 2473d14c5d2SYehuda Sadeh size = sizeof(struct crush_bucket_straw); 2483d14c5d2SYehuda Sadeh break; 249958a2765SIlya Dryomov case CRUSH_BUCKET_STRAW2: 250958a2765SIlya Dryomov size = sizeof(struct crush_bucket_straw2); 251958a2765SIlya Dryomov break; 2523d14c5d2SYehuda Sadeh default: 2533d14c5d2SYehuda Sadeh err = -EINVAL; 2543d14c5d2SYehuda Sadeh goto bad; 2553d14c5d2SYehuda Sadeh } 2563d14c5d2SYehuda Sadeh BUG_ON(size == 0); 2573d14c5d2SYehuda Sadeh b = c->buckets[i] = kzalloc(size, GFP_NOFS); 2583d14c5d2SYehuda Sadeh if (b == NULL) 2593d14c5d2SYehuda Sadeh goto badmem; 2603d14c5d2SYehuda Sadeh 2613d14c5d2SYehuda Sadeh ceph_decode_need(p, end, 4*sizeof(u32), bad); 2623d14c5d2SYehuda Sadeh b->id = ceph_decode_32(p); 2633d14c5d2SYehuda Sadeh b->type = ceph_decode_16(p); 2643d14c5d2SYehuda Sadeh b->alg = ceph_decode_8(p); 2653d14c5d2SYehuda Sadeh b->hash = ceph_decode_8(p); 2663d14c5d2SYehuda Sadeh b->weight = ceph_decode_32(p); 2673d14c5d2SYehuda Sadeh b->size = ceph_decode_32(p); 2683d14c5d2SYehuda Sadeh 2693d14c5d2SYehuda Sadeh dout("crush_decode bucket size %d off %x %p to %p\n", 2703d14c5d2SYehuda Sadeh b->size, (int)(*p-start), *p, end); 2713d14c5d2SYehuda Sadeh 2723d14c5d2SYehuda Sadeh b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS); 2733d14c5d2SYehuda Sadeh if (b->items == NULL) 2743d14c5d2SYehuda Sadeh goto badmem; 2753d14c5d2SYehuda Sadeh 2763d14c5d2SYehuda Sadeh ceph_decode_need(p, end, b->size*sizeof(u32), bad); 2773d14c5d2SYehuda Sadeh for (j = 0; j < b->size; j++) 2783d14c5d2SYehuda Sadeh b->items[j] = ceph_decode_32(p); 2793d14c5d2SYehuda Sadeh 2803d14c5d2SYehuda Sadeh switch (b->alg) { 2813d14c5d2SYehuda Sadeh case CRUSH_BUCKET_UNIFORM: 2823d14c5d2SYehuda Sadeh err = crush_decode_uniform_bucket(p, end, 2833d14c5d2SYehuda Sadeh (struct crush_bucket_uniform *)b); 2843d14c5d2SYehuda Sadeh if (err < 0) 2853d14c5d2SYehuda Sadeh goto bad; 2863d14c5d2SYehuda Sadeh break; 2873d14c5d2SYehuda Sadeh case CRUSH_BUCKET_LIST: 2883d14c5d2SYehuda Sadeh err = crush_decode_list_bucket(p, end, 2893d14c5d2SYehuda Sadeh (struct crush_bucket_list *)b); 2903d14c5d2SYehuda Sadeh if (err < 0) 2913d14c5d2SYehuda Sadeh goto bad; 2923d14c5d2SYehuda Sadeh break; 2933d14c5d2SYehuda Sadeh case CRUSH_BUCKET_TREE: 2943d14c5d2SYehuda Sadeh err = crush_decode_tree_bucket(p, end, 2953d14c5d2SYehuda Sadeh (struct crush_bucket_tree *)b); 2963d14c5d2SYehuda Sadeh if (err < 0) 2973d14c5d2SYehuda Sadeh goto bad; 2983d14c5d2SYehuda Sadeh break; 2993d14c5d2SYehuda Sadeh case CRUSH_BUCKET_STRAW: 3003d14c5d2SYehuda Sadeh err = crush_decode_straw_bucket(p, end, 3013d14c5d2SYehuda Sadeh (struct crush_bucket_straw *)b); 3023d14c5d2SYehuda Sadeh if (err < 0) 3033d14c5d2SYehuda Sadeh goto bad; 3043d14c5d2SYehuda Sadeh break; 305958a2765SIlya Dryomov case CRUSH_BUCKET_STRAW2: 306958a2765SIlya Dryomov err = crush_decode_straw2_bucket(p, end, 307958a2765SIlya Dryomov (struct crush_bucket_straw2 *)b); 308958a2765SIlya Dryomov if (err < 0) 309958a2765SIlya Dryomov goto bad; 310958a2765SIlya Dryomov break; 3113d14c5d2SYehuda Sadeh } 3123d14c5d2SYehuda Sadeh } 3133d14c5d2SYehuda Sadeh 3143d14c5d2SYehuda Sadeh /* rules */ 3153d14c5d2SYehuda Sadeh dout("rule vec is %p\n", c->rules); 3163d14c5d2SYehuda Sadeh for (i = 0; i < c->max_rules; i++) { 3173d14c5d2SYehuda Sadeh u32 yes; 3183d14c5d2SYehuda Sadeh struct crush_rule *r; 3193d14c5d2SYehuda Sadeh 320293dffaaSDan Carpenter err = -EINVAL; 3213d14c5d2SYehuda Sadeh ceph_decode_32_safe(p, end, yes, bad); 3223d14c5d2SYehuda Sadeh if (!yes) { 3233d14c5d2SYehuda Sadeh dout("crush_decode NO rule %d off %x %p to %p\n", 3243d14c5d2SYehuda Sadeh i, (int)(*p-start), *p, end); 3253d14c5d2SYehuda Sadeh c->rules[i] = NULL; 3263d14c5d2SYehuda Sadeh continue; 3273d14c5d2SYehuda Sadeh } 3283d14c5d2SYehuda Sadeh 3293d14c5d2SYehuda Sadeh dout("crush_decode rule %d off %x %p to %p\n", 3303d14c5d2SYehuda Sadeh i, (int)(*p-start), *p, end); 3313d14c5d2SYehuda Sadeh 3323d14c5d2SYehuda Sadeh /* len */ 3333d14c5d2SYehuda Sadeh ceph_decode_32_safe(p, end, yes, bad); 3343d14c5d2SYehuda Sadeh #if BITS_PER_LONG == 32 3353d14c5d2SYehuda Sadeh err = -EINVAL; 33664486697SXi Wang if (yes > (ULONG_MAX - sizeof(*r)) 33764486697SXi Wang / sizeof(struct crush_rule_step)) 3383d14c5d2SYehuda Sadeh goto bad; 3393d14c5d2SYehuda Sadeh #endif 3403d14c5d2SYehuda Sadeh r = c->rules[i] = kmalloc(sizeof(*r) + 3413d14c5d2SYehuda Sadeh yes*sizeof(struct crush_rule_step), 3423d14c5d2SYehuda Sadeh GFP_NOFS); 3433d14c5d2SYehuda Sadeh if (r == NULL) 3443d14c5d2SYehuda Sadeh goto badmem; 3453d14c5d2SYehuda Sadeh dout(" rule %d is at %p\n", i, r); 3463d14c5d2SYehuda Sadeh r->len = yes; 3473d14c5d2SYehuda Sadeh ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */ 3483d14c5d2SYehuda Sadeh ceph_decode_need(p, end, r->len*3*sizeof(u32), bad); 3493d14c5d2SYehuda Sadeh for (j = 0; j < r->len; j++) { 3503d14c5d2SYehuda Sadeh r->steps[j].op = ceph_decode_32(p); 3513d14c5d2SYehuda Sadeh r->steps[j].arg1 = ceph_decode_32(p); 3523d14c5d2SYehuda Sadeh r->steps[j].arg2 = ceph_decode_32(p); 3533d14c5d2SYehuda Sadeh } 3543d14c5d2SYehuda Sadeh } 3553d14c5d2SYehuda Sadeh 3563d14c5d2SYehuda Sadeh /* ignore trailing name maps. */ 357546f04efSSage Weil for (num_name_maps = 0; num_name_maps < 3; num_name_maps++) { 358546f04efSSage Weil err = skip_name_map(p, end); 359546f04efSSage Weil if (err < 0) 360546f04efSSage Weil goto done; 361546f04efSSage Weil } 3623d14c5d2SYehuda Sadeh 363546f04efSSage Weil /* tunables */ 364546f04efSSage Weil ceph_decode_need(p, end, 3*sizeof(u32), done); 365546f04efSSage Weil c->choose_local_tries = ceph_decode_32(p); 366546f04efSSage Weil c->choose_local_fallback_tries = ceph_decode_32(p); 367546f04efSSage Weil c->choose_total_tries = ceph_decode_32(p); 368b9b519b7SIlya Dryomov dout("crush decode tunable choose_local_tries = %d\n", 369546f04efSSage Weil c->choose_local_tries); 370b9b519b7SIlya Dryomov dout("crush decode tunable choose_local_fallback_tries = %d\n", 371546f04efSSage Weil c->choose_local_fallback_tries); 372b9b519b7SIlya Dryomov dout("crush decode tunable choose_total_tries = %d\n", 373546f04efSSage Weil c->choose_total_tries); 374546f04efSSage Weil 3751604f488SJim Schutt ceph_decode_need(p, end, sizeof(u32), done); 3761604f488SJim Schutt c->chooseleaf_descend_once = ceph_decode_32(p); 377b9b519b7SIlya Dryomov dout("crush decode tunable chooseleaf_descend_once = %d\n", 3781604f488SJim Schutt c->chooseleaf_descend_once); 3791604f488SJim Schutt 380f140662fSIlya Dryomov ceph_decode_need(p, end, sizeof(u8), done); 381f140662fSIlya Dryomov c->chooseleaf_vary_r = ceph_decode_8(p); 382b9b519b7SIlya Dryomov dout("crush decode tunable chooseleaf_vary_r = %d\n", 383f140662fSIlya Dryomov c->chooseleaf_vary_r); 384f140662fSIlya Dryomov 385b9b519b7SIlya Dryomov /* skip straw_calc_version, allowed_bucket_algs */ 386b9b519b7SIlya Dryomov ceph_decode_need(p, end, sizeof(u8) + sizeof(u32), done); 387b9b519b7SIlya Dryomov *p += sizeof(u8) + sizeof(u32); 388b9b519b7SIlya Dryomov 389b9b519b7SIlya Dryomov ceph_decode_need(p, end, sizeof(u8), done); 390b9b519b7SIlya Dryomov c->chooseleaf_stable = ceph_decode_8(p); 391b9b519b7SIlya Dryomov dout("crush decode tunable chooseleaf_stable = %d\n", 392b9b519b7SIlya Dryomov c->chooseleaf_stable); 393b9b519b7SIlya Dryomov 394546f04efSSage Weil done: 3959afd30dbSIlya Dryomov crush_finalize(c); 3963d14c5d2SYehuda Sadeh dout("crush_decode success\n"); 3973d14c5d2SYehuda Sadeh return c; 3983d14c5d2SYehuda Sadeh 3993d14c5d2SYehuda Sadeh badmem: 4003d14c5d2SYehuda Sadeh err = -ENOMEM; 4013d14c5d2SYehuda Sadeh bad: 4023d14c5d2SYehuda Sadeh dout("crush_decode fail %d\n", err); 4033d14c5d2SYehuda Sadeh crush_destroy(c); 4043d14c5d2SYehuda Sadeh return ERR_PTR(err); 4053d14c5d2SYehuda Sadeh } 4063d14c5d2SYehuda Sadeh 407f984cb76SIlya Dryomov int ceph_pg_compare(const struct ceph_pg *lhs, const struct ceph_pg *rhs) 408f984cb76SIlya Dryomov { 409f984cb76SIlya Dryomov if (lhs->pool < rhs->pool) 410f984cb76SIlya Dryomov return -1; 411f984cb76SIlya Dryomov if (lhs->pool > rhs->pool) 412f984cb76SIlya Dryomov return 1; 413f984cb76SIlya Dryomov if (lhs->seed < rhs->seed) 414f984cb76SIlya Dryomov return -1; 415f984cb76SIlya Dryomov if (lhs->seed > rhs->seed) 416f984cb76SIlya Dryomov return 1; 417f984cb76SIlya Dryomov 418f984cb76SIlya Dryomov return 0; 419f984cb76SIlya Dryomov } 420f984cb76SIlya Dryomov 4213d14c5d2SYehuda Sadeh /* 4223d14c5d2SYehuda Sadeh * rbtree of pg_mapping for handling pg_temp (explicit mapping of pgid 4239686f94cSIlya Dryomov * to a set of osds) and primary_temp (explicit primary setting) 4243d14c5d2SYehuda Sadeh */ 4253d14c5d2SYehuda Sadeh static int __insert_pg_mapping(struct ceph_pg_mapping *new, 4263d14c5d2SYehuda Sadeh struct rb_root *root) 4273d14c5d2SYehuda Sadeh { 4283d14c5d2SYehuda Sadeh struct rb_node **p = &root->rb_node; 4293d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 4303d14c5d2SYehuda Sadeh struct ceph_pg_mapping *pg = NULL; 4313d14c5d2SYehuda Sadeh int c; 4323d14c5d2SYehuda Sadeh 4338adc8b3dSSage Weil dout("__insert_pg_mapping %llx %p\n", *(u64 *)&new->pgid, new); 4343d14c5d2SYehuda Sadeh while (*p) { 4353d14c5d2SYehuda Sadeh parent = *p; 4363d14c5d2SYehuda Sadeh pg = rb_entry(parent, struct ceph_pg_mapping, node); 437f984cb76SIlya Dryomov c = ceph_pg_compare(&new->pgid, &pg->pgid); 4383d14c5d2SYehuda Sadeh if (c < 0) 4393d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 4403d14c5d2SYehuda Sadeh else if (c > 0) 4413d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 4423d14c5d2SYehuda Sadeh else 4433d14c5d2SYehuda Sadeh return -EEXIST; 4443d14c5d2SYehuda Sadeh } 4453d14c5d2SYehuda Sadeh 4463d14c5d2SYehuda Sadeh rb_link_node(&new->node, parent, p); 4473d14c5d2SYehuda Sadeh rb_insert_color(&new->node, root); 4483d14c5d2SYehuda Sadeh return 0; 4493d14c5d2SYehuda Sadeh } 4503d14c5d2SYehuda Sadeh 4513d14c5d2SYehuda Sadeh static struct ceph_pg_mapping *__lookup_pg_mapping(struct rb_root *root, 4525b191d99SSage Weil struct ceph_pg pgid) 4533d14c5d2SYehuda Sadeh { 4543d14c5d2SYehuda Sadeh struct rb_node *n = root->rb_node; 4553d14c5d2SYehuda Sadeh struct ceph_pg_mapping *pg; 4563d14c5d2SYehuda Sadeh int c; 4573d14c5d2SYehuda Sadeh 4583d14c5d2SYehuda Sadeh while (n) { 4593d14c5d2SYehuda Sadeh pg = rb_entry(n, struct ceph_pg_mapping, node); 460f984cb76SIlya Dryomov c = ceph_pg_compare(&pgid, &pg->pgid); 4618adc8b3dSSage Weil if (c < 0) { 4623d14c5d2SYehuda Sadeh n = n->rb_left; 4638adc8b3dSSage Weil } else if (c > 0) { 4643d14c5d2SYehuda Sadeh n = n->rb_right; 4658adc8b3dSSage Weil } else { 4665b191d99SSage Weil dout("__lookup_pg_mapping %lld.%x got %p\n", 4675b191d99SSage Weil pgid.pool, pgid.seed, pg); 4683d14c5d2SYehuda Sadeh return pg; 4693d14c5d2SYehuda Sadeh } 4708adc8b3dSSage Weil } 4713d14c5d2SYehuda Sadeh return NULL; 4723d14c5d2SYehuda Sadeh } 4733d14c5d2SYehuda Sadeh 4745b191d99SSage Weil static int __remove_pg_mapping(struct rb_root *root, struct ceph_pg pgid) 4758adc8b3dSSage Weil { 4768adc8b3dSSage Weil struct ceph_pg_mapping *pg = __lookup_pg_mapping(root, pgid); 4778adc8b3dSSage Weil 4788adc8b3dSSage Weil if (pg) { 4795b191d99SSage Weil dout("__remove_pg_mapping %lld.%x %p\n", pgid.pool, pgid.seed, 4805b191d99SSage Weil pg); 4818adc8b3dSSage Weil rb_erase(&pg->node, root); 4828adc8b3dSSage Weil kfree(pg); 4838adc8b3dSSage Weil return 0; 4848adc8b3dSSage Weil } 4855b191d99SSage Weil dout("__remove_pg_mapping %lld.%x dne\n", pgid.pool, pgid.seed); 4868adc8b3dSSage Weil return -ENOENT; 4878adc8b3dSSage Weil } 4888adc8b3dSSage Weil 4893d14c5d2SYehuda Sadeh /* 4903d14c5d2SYehuda Sadeh * rbtree of pg pool info 4913d14c5d2SYehuda Sadeh */ 4923d14c5d2SYehuda Sadeh static int __insert_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *new) 4933d14c5d2SYehuda Sadeh { 4943d14c5d2SYehuda Sadeh struct rb_node **p = &root->rb_node; 4953d14c5d2SYehuda Sadeh struct rb_node *parent = NULL; 4963d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi = NULL; 4973d14c5d2SYehuda Sadeh 4983d14c5d2SYehuda Sadeh while (*p) { 4993d14c5d2SYehuda Sadeh parent = *p; 5003d14c5d2SYehuda Sadeh pi = rb_entry(parent, struct ceph_pg_pool_info, node); 5013d14c5d2SYehuda Sadeh if (new->id < pi->id) 5023d14c5d2SYehuda Sadeh p = &(*p)->rb_left; 5033d14c5d2SYehuda Sadeh else if (new->id > pi->id) 5043d14c5d2SYehuda Sadeh p = &(*p)->rb_right; 5053d14c5d2SYehuda Sadeh else 5063d14c5d2SYehuda Sadeh return -EEXIST; 5073d14c5d2SYehuda Sadeh } 5083d14c5d2SYehuda Sadeh 5093d14c5d2SYehuda Sadeh rb_link_node(&new->node, parent, p); 5103d14c5d2SYehuda Sadeh rb_insert_color(&new->node, root); 5113d14c5d2SYehuda Sadeh return 0; 5123d14c5d2SYehuda Sadeh } 5133d14c5d2SYehuda Sadeh 5144f6a7e5eSSage Weil static struct ceph_pg_pool_info *__lookup_pg_pool(struct rb_root *root, u64 id) 5153d14c5d2SYehuda Sadeh { 5163d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi; 5173d14c5d2SYehuda Sadeh struct rb_node *n = root->rb_node; 5183d14c5d2SYehuda Sadeh 5193d14c5d2SYehuda Sadeh while (n) { 5203d14c5d2SYehuda Sadeh pi = rb_entry(n, struct ceph_pg_pool_info, node); 5213d14c5d2SYehuda Sadeh if (id < pi->id) 5223d14c5d2SYehuda Sadeh n = n->rb_left; 5233d14c5d2SYehuda Sadeh else if (id > pi->id) 5243d14c5d2SYehuda Sadeh n = n->rb_right; 5253d14c5d2SYehuda Sadeh else 5263d14c5d2SYehuda Sadeh return pi; 5273d14c5d2SYehuda Sadeh } 5283d14c5d2SYehuda Sadeh return NULL; 5293d14c5d2SYehuda Sadeh } 5303d14c5d2SYehuda Sadeh 531ce7f6a27SIlya Dryomov struct ceph_pg_pool_info *ceph_pg_pool_by_id(struct ceph_osdmap *map, u64 id) 532ce7f6a27SIlya Dryomov { 533ce7f6a27SIlya Dryomov return __lookup_pg_pool(&map->pg_pools, id); 534ce7f6a27SIlya Dryomov } 535ce7f6a27SIlya Dryomov 53672afc71fSAlex Elder const char *ceph_pg_pool_name_by_id(struct ceph_osdmap *map, u64 id) 53772afc71fSAlex Elder { 53872afc71fSAlex Elder struct ceph_pg_pool_info *pi; 53972afc71fSAlex Elder 54072afc71fSAlex Elder if (id == CEPH_NOPOOL) 54172afc71fSAlex Elder return NULL; 54272afc71fSAlex Elder 54372afc71fSAlex Elder if (WARN_ON_ONCE(id > (u64) INT_MAX)) 54472afc71fSAlex Elder return NULL; 54572afc71fSAlex Elder 54672afc71fSAlex Elder pi = __lookup_pg_pool(&map->pg_pools, (int) id); 54772afc71fSAlex Elder 54872afc71fSAlex Elder return pi ? pi->name : NULL; 54972afc71fSAlex Elder } 55072afc71fSAlex Elder EXPORT_SYMBOL(ceph_pg_pool_name_by_id); 55172afc71fSAlex Elder 5523d14c5d2SYehuda Sadeh int ceph_pg_poolid_by_name(struct ceph_osdmap *map, const char *name) 5533d14c5d2SYehuda Sadeh { 5543d14c5d2SYehuda Sadeh struct rb_node *rbp; 5553d14c5d2SYehuda Sadeh 5563d14c5d2SYehuda Sadeh for (rbp = rb_first(&map->pg_pools); rbp; rbp = rb_next(rbp)) { 5573d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi = 5583d14c5d2SYehuda Sadeh rb_entry(rbp, struct ceph_pg_pool_info, node); 5593d14c5d2SYehuda Sadeh if (pi->name && strcmp(pi->name, name) == 0) 5603d14c5d2SYehuda Sadeh return pi->id; 5613d14c5d2SYehuda Sadeh } 5623d14c5d2SYehuda Sadeh return -ENOENT; 5633d14c5d2SYehuda Sadeh } 5643d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_pg_poolid_by_name); 5653d14c5d2SYehuda Sadeh 5663d14c5d2SYehuda Sadeh static void __remove_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *pi) 5673d14c5d2SYehuda Sadeh { 5683d14c5d2SYehuda Sadeh rb_erase(&pi->node, root); 5693d14c5d2SYehuda Sadeh kfree(pi->name); 5703d14c5d2SYehuda Sadeh kfree(pi); 5713d14c5d2SYehuda Sadeh } 5723d14c5d2SYehuda Sadeh 5730f70c7eeSIlya Dryomov static int decode_pool(void **p, void *end, struct ceph_pg_pool_info *pi) 5743d14c5d2SYehuda Sadeh { 5754f6a7e5eSSage Weil u8 ev, cv; 5764f6a7e5eSSage Weil unsigned len, num; 5774f6a7e5eSSage Weil void *pool_end; 5783d14c5d2SYehuda Sadeh 5794f6a7e5eSSage Weil ceph_decode_need(p, end, 2 + 4, bad); 5804f6a7e5eSSage Weil ev = ceph_decode_8(p); /* encoding version */ 5814f6a7e5eSSage Weil cv = ceph_decode_8(p); /* compat version */ 5824f6a7e5eSSage Weil if (ev < 5) { 583b9a67899SJoe Perches pr_warn("got v %d < 5 cv %d of ceph_pg_pool\n", ev, cv); 5844f6a7e5eSSage Weil return -EINVAL; 5854f6a7e5eSSage Weil } 58617a13e40SIlya Dryomov if (cv > 9) { 587b9a67899SJoe Perches pr_warn("got v %d cv %d > 9 of ceph_pg_pool\n", ev, cv); 5884f6a7e5eSSage Weil return -EINVAL; 5894f6a7e5eSSage Weil } 5904f6a7e5eSSage Weil len = ceph_decode_32(p); 5914f6a7e5eSSage Weil ceph_decode_need(p, end, len, bad); 5924f6a7e5eSSage Weil pool_end = *p + len; 5933d14c5d2SYehuda Sadeh 5944f6a7e5eSSage Weil pi->type = ceph_decode_8(p); 5954f6a7e5eSSage Weil pi->size = ceph_decode_8(p); 5964f6a7e5eSSage Weil pi->crush_ruleset = ceph_decode_8(p); 5974f6a7e5eSSage Weil pi->object_hash = ceph_decode_8(p); 5984f6a7e5eSSage Weil 5994f6a7e5eSSage Weil pi->pg_num = ceph_decode_32(p); 6004f6a7e5eSSage Weil pi->pgp_num = ceph_decode_32(p); 6014f6a7e5eSSage Weil 6024f6a7e5eSSage Weil *p += 4 + 4; /* skip lpg* */ 6034f6a7e5eSSage Weil *p += 4; /* skip last_change */ 6044f6a7e5eSSage Weil *p += 8 + 4; /* skip snap_seq, snap_epoch */ 6054f6a7e5eSSage Weil 6064f6a7e5eSSage Weil /* skip snaps */ 6074f6a7e5eSSage Weil num = ceph_decode_32(p); 6084f6a7e5eSSage Weil while (num--) { 6094f6a7e5eSSage Weil *p += 8; /* snapid key */ 6104f6a7e5eSSage Weil *p += 1 + 1; /* versions */ 6114f6a7e5eSSage Weil len = ceph_decode_32(p); 6124f6a7e5eSSage Weil *p += len; 6133d14c5d2SYehuda Sadeh } 6143d14c5d2SYehuda Sadeh 61517a13e40SIlya Dryomov /* skip removed_snaps */ 6164f6a7e5eSSage Weil num = ceph_decode_32(p); 6174f6a7e5eSSage Weil *p += num * (8 + 8); 6184f6a7e5eSSage Weil 6194f6a7e5eSSage Weil *p += 8; /* skip auid */ 6204f6a7e5eSSage Weil pi->flags = ceph_decode_64(p); 62117a13e40SIlya Dryomov *p += 4; /* skip crash_replay_interval */ 62217a13e40SIlya Dryomov 62317a13e40SIlya Dryomov if (ev >= 7) 62404812acfSIlya Dryomov pi->min_size = ceph_decode_8(p); 62504812acfSIlya Dryomov else 62604812acfSIlya Dryomov pi->min_size = pi->size - pi->size / 2; 62717a13e40SIlya Dryomov 62817a13e40SIlya Dryomov if (ev >= 8) 62917a13e40SIlya Dryomov *p += 8 + 8; /* skip quota_max_* */ 63017a13e40SIlya Dryomov 63117a13e40SIlya Dryomov if (ev >= 9) { 63217a13e40SIlya Dryomov /* skip tiers */ 63317a13e40SIlya Dryomov num = ceph_decode_32(p); 63417a13e40SIlya Dryomov *p += num * 8; 63517a13e40SIlya Dryomov 63617a13e40SIlya Dryomov *p += 8; /* skip tier_of */ 63717a13e40SIlya Dryomov *p += 1; /* skip cache_mode */ 63817a13e40SIlya Dryomov 63917a13e40SIlya Dryomov pi->read_tier = ceph_decode_64(p); 64017a13e40SIlya Dryomov pi->write_tier = ceph_decode_64(p); 64117a13e40SIlya Dryomov } else { 64217a13e40SIlya Dryomov pi->read_tier = -1; 64317a13e40SIlya Dryomov pi->write_tier = -1; 64417a13e40SIlya Dryomov } 6454f6a7e5eSSage Weil 64604812acfSIlya Dryomov if (ev >= 10) { 64704812acfSIlya Dryomov /* skip properties */ 64804812acfSIlya Dryomov num = ceph_decode_32(p); 64904812acfSIlya Dryomov while (num--) { 65004812acfSIlya Dryomov len = ceph_decode_32(p); 65104812acfSIlya Dryomov *p += len; /* key */ 65204812acfSIlya Dryomov len = ceph_decode_32(p); 65304812acfSIlya Dryomov *p += len; /* val */ 65404812acfSIlya Dryomov } 65504812acfSIlya Dryomov } 65604812acfSIlya Dryomov 65704812acfSIlya Dryomov if (ev >= 11) { 65804812acfSIlya Dryomov /* skip hit_set_params */ 65904812acfSIlya Dryomov *p += 1 + 1; /* versions */ 66004812acfSIlya Dryomov len = ceph_decode_32(p); 66104812acfSIlya Dryomov *p += len; 66204812acfSIlya Dryomov 66304812acfSIlya Dryomov *p += 4; /* skip hit_set_period */ 66404812acfSIlya Dryomov *p += 4; /* skip hit_set_count */ 66504812acfSIlya Dryomov } 66604812acfSIlya Dryomov 66704812acfSIlya Dryomov if (ev >= 12) 66804812acfSIlya Dryomov *p += 4; /* skip stripe_width */ 66904812acfSIlya Dryomov 67004812acfSIlya Dryomov if (ev >= 13) { 67104812acfSIlya Dryomov *p += 8; /* skip target_max_bytes */ 67204812acfSIlya Dryomov *p += 8; /* skip target_max_objects */ 67304812acfSIlya Dryomov *p += 4; /* skip cache_target_dirty_ratio_micro */ 67404812acfSIlya Dryomov *p += 4; /* skip cache_target_full_ratio_micro */ 67504812acfSIlya Dryomov *p += 4; /* skip cache_min_flush_age */ 67604812acfSIlya Dryomov *p += 4; /* skip cache_min_evict_age */ 67704812acfSIlya Dryomov } 67804812acfSIlya Dryomov 67904812acfSIlya Dryomov if (ev >= 14) { 68004812acfSIlya Dryomov /* skip erasure_code_profile */ 68104812acfSIlya Dryomov len = ceph_decode_32(p); 68204812acfSIlya Dryomov *p += len; 68304812acfSIlya Dryomov } 68404812acfSIlya Dryomov 6858e48cf00SIlya Dryomov /* 6868e48cf00SIlya Dryomov * last_force_op_resend_preluminous, will be overridden if the 6878e48cf00SIlya Dryomov * map was encoded with RESEND_ON_SPLIT 6888e48cf00SIlya Dryomov */ 68904812acfSIlya Dryomov if (ev >= 15) 69004812acfSIlya Dryomov pi->last_force_request_resend = ceph_decode_32(p); 69104812acfSIlya Dryomov else 69204812acfSIlya Dryomov pi->last_force_request_resend = 0; 69304812acfSIlya Dryomov 6948e48cf00SIlya Dryomov if (ev >= 16) 6958e48cf00SIlya Dryomov *p += 4; /* skip min_read_recency_for_promote */ 6968e48cf00SIlya Dryomov 6978e48cf00SIlya Dryomov if (ev >= 17) 6988e48cf00SIlya Dryomov *p += 8; /* skip expected_num_objects */ 6998e48cf00SIlya Dryomov 7008e48cf00SIlya Dryomov if (ev >= 19) 7018e48cf00SIlya Dryomov *p += 4; /* skip cache_target_dirty_high_ratio_micro */ 7028e48cf00SIlya Dryomov 7038e48cf00SIlya Dryomov if (ev >= 20) 7048e48cf00SIlya Dryomov *p += 4; /* skip min_write_recency_for_promote */ 7058e48cf00SIlya Dryomov 7068e48cf00SIlya Dryomov if (ev >= 21) 7078e48cf00SIlya Dryomov *p += 1; /* skip use_gmt_hitset */ 7088e48cf00SIlya Dryomov 7098e48cf00SIlya Dryomov if (ev >= 22) 7108e48cf00SIlya Dryomov *p += 1; /* skip fast_read */ 7118e48cf00SIlya Dryomov 7128e48cf00SIlya Dryomov if (ev >= 23) { 7138e48cf00SIlya Dryomov *p += 4; /* skip hit_set_grade_decay_rate */ 7148e48cf00SIlya Dryomov *p += 4; /* skip hit_set_search_last_n */ 7158e48cf00SIlya Dryomov } 7168e48cf00SIlya Dryomov 7178e48cf00SIlya Dryomov if (ev >= 24) { 7188e48cf00SIlya Dryomov /* skip opts */ 7198e48cf00SIlya Dryomov *p += 1 + 1; /* versions */ 7208e48cf00SIlya Dryomov len = ceph_decode_32(p); 7218e48cf00SIlya Dryomov *p += len; 7228e48cf00SIlya Dryomov } 7238e48cf00SIlya Dryomov 7248e48cf00SIlya Dryomov if (ev >= 25) 7258e48cf00SIlya Dryomov pi->last_force_request_resend = ceph_decode_32(p); 7268e48cf00SIlya Dryomov 7274f6a7e5eSSage Weil /* ignore the rest */ 7284f6a7e5eSSage Weil 7294f6a7e5eSSage Weil *p = pool_end; 7304f6a7e5eSSage Weil calc_pg_masks(pi); 7313d14c5d2SYehuda Sadeh return 0; 7323d14c5d2SYehuda Sadeh 7333d14c5d2SYehuda Sadeh bad: 7343d14c5d2SYehuda Sadeh return -EINVAL; 7353d14c5d2SYehuda Sadeh } 7363d14c5d2SYehuda Sadeh 7370f70c7eeSIlya Dryomov static int decode_pool_names(void **p, void *end, struct ceph_osdmap *map) 7383d14c5d2SYehuda Sadeh { 7393d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi; 7404f6a7e5eSSage Weil u32 num, len; 7414f6a7e5eSSage Weil u64 pool; 7423d14c5d2SYehuda Sadeh 7433d14c5d2SYehuda Sadeh ceph_decode_32_safe(p, end, num, bad); 7443d14c5d2SYehuda Sadeh dout(" %d pool names\n", num); 7453d14c5d2SYehuda Sadeh while (num--) { 7464f6a7e5eSSage Weil ceph_decode_64_safe(p, end, pool, bad); 7473d14c5d2SYehuda Sadeh ceph_decode_32_safe(p, end, len, bad); 7484f6a7e5eSSage Weil dout(" pool %llu len %d\n", pool, len); 749ad3b904cSXi Wang ceph_decode_need(p, end, len, bad); 7503d14c5d2SYehuda Sadeh pi = __lookup_pg_pool(&map->pg_pools, pool); 7513d14c5d2SYehuda Sadeh if (pi) { 752ad3b904cSXi Wang char *name = kstrndup(*p, len, GFP_NOFS); 753ad3b904cSXi Wang 754ad3b904cSXi Wang if (!name) 755ad3b904cSXi Wang return -ENOMEM; 7563d14c5d2SYehuda Sadeh kfree(pi->name); 757ad3b904cSXi Wang pi->name = name; 7583d14c5d2SYehuda Sadeh dout(" name is %s\n", pi->name); 7593d14c5d2SYehuda Sadeh } 7603d14c5d2SYehuda Sadeh *p += len; 7613d14c5d2SYehuda Sadeh } 7623d14c5d2SYehuda Sadeh return 0; 7633d14c5d2SYehuda Sadeh 7643d14c5d2SYehuda Sadeh bad: 7653d14c5d2SYehuda Sadeh return -EINVAL; 7663d14c5d2SYehuda Sadeh } 7673d14c5d2SYehuda Sadeh 7683d14c5d2SYehuda Sadeh /* 7693d14c5d2SYehuda Sadeh * osd map 7703d14c5d2SYehuda Sadeh */ 771e5253a7bSIlya Dryomov struct ceph_osdmap *ceph_osdmap_alloc(void) 772e5253a7bSIlya Dryomov { 773e5253a7bSIlya Dryomov struct ceph_osdmap *map; 774e5253a7bSIlya Dryomov 775e5253a7bSIlya Dryomov map = kzalloc(sizeof(*map), GFP_NOIO); 776e5253a7bSIlya Dryomov if (!map) 777e5253a7bSIlya Dryomov return NULL; 778e5253a7bSIlya Dryomov 779e5253a7bSIlya Dryomov map->pg_pools = RB_ROOT; 780e5253a7bSIlya Dryomov map->pool_max = -1; 781e5253a7bSIlya Dryomov map->pg_temp = RB_ROOT; 782e5253a7bSIlya Dryomov map->primary_temp = RB_ROOT; 783743efcffSIlya Dryomov mutex_init(&map->crush_workspace_mutex); 784e5253a7bSIlya Dryomov 785e5253a7bSIlya Dryomov return map; 786e5253a7bSIlya Dryomov } 787e5253a7bSIlya Dryomov 7883d14c5d2SYehuda Sadeh void ceph_osdmap_destroy(struct ceph_osdmap *map) 7893d14c5d2SYehuda Sadeh { 7903d14c5d2SYehuda Sadeh dout("osdmap_destroy %p\n", map); 7913d14c5d2SYehuda Sadeh if (map->crush) 7923d14c5d2SYehuda Sadeh crush_destroy(map->crush); 7933d14c5d2SYehuda Sadeh while (!RB_EMPTY_ROOT(&map->pg_temp)) { 7943d14c5d2SYehuda Sadeh struct ceph_pg_mapping *pg = 7953d14c5d2SYehuda Sadeh rb_entry(rb_first(&map->pg_temp), 7963d14c5d2SYehuda Sadeh struct ceph_pg_mapping, node); 7973d14c5d2SYehuda Sadeh rb_erase(&pg->node, &map->pg_temp); 7983d14c5d2SYehuda Sadeh kfree(pg); 7993d14c5d2SYehuda Sadeh } 8009686f94cSIlya Dryomov while (!RB_EMPTY_ROOT(&map->primary_temp)) { 8019686f94cSIlya Dryomov struct ceph_pg_mapping *pg = 8029686f94cSIlya Dryomov rb_entry(rb_first(&map->primary_temp), 8039686f94cSIlya Dryomov struct ceph_pg_mapping, node); 8049686f94cSIlya Dryomov rb_erase(&pg->node, &map->primary_temp); 8059686f94cSIlya Dryomov kfree(pg); 8069686f94cSIlya Dryomov } 8073d14c5d2SYehuda Sadeh while (!RB_EMPTY_ROOT(&map->pg_pools)) { 8083d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi = 8093d14c5d2SYehuda Sadeh rb_entry(rb_first(&map->pg_pools), 8103d14c5d2SYehuda Sadeh struct ceph_pg_pool_info, node); 8113d14c5d2SYehuda Sadeh __remove_pg_pool(&map->pg_pools, pi); 8123d14c5d2SYehuda Sadeh } 8133d14c5d2SYehuda Sadeh kfree(map->osd_state); 8143d14c5d2SYehuda Sadeh kfree(map->osd_weight); 8153d14c5d2SYehuda Sadeh kfree(map->osd_addr); 8162cfa34f2SIlya Dryomov kfree(map->osd_primary_affinity); 81766a0e2d5SIlya Dryomov kfree(map->crush_workspace); 8183d14c5d2SYehuda Sadeh kfree(map); 8193d14c5d2SYehuda Sadeh } 8203d14c5d2SYehuda Sadeh 8213d14c5d2SYehuda Sadeh /* 8224d60351fSIlya Dryomov * Adjust max_osd value, (re)allocate arrays. 8234d60351fSIlya Dryomov * 8244d60351fSIlya Dryomov * The new elements are properly initialized. 8253d14c5d2SYehuda Sadeh */ 8263d14c5d2SYehuda Sadeh static int osdmap_set_max_osd(struct ceph_osdmap *map, int max) 8273d14c5d2SYehuda Sadeh { 8283d14c5d2SYehuda Sadeh u8 *state; 8293d14c5d2SYehuda Sadeh u32 *weight; 8304d60351fSIlya Dryomov struct ceph_entity_addr *addr; 8314d60351fSIlya Dryomov int i; 8323d14c5d2SYehuda Sadeh 8334d60351fSIlya Dryomov state = krealloc(map->osd_state, max*sizeof(*state), GFP_NOFS); 834589506f1SLi RongQing if (!state) 8353d14c5d2SYehuda Sadeh return -ENOMEM; 836589506f1SLi RongQing map->osd_state = state; 837589506f1SLi RongQing 838589506f1SLi RongQing weight = krealloc(map->osd_weight, max*sizeof(*weight), GFP_NOFS); 839589506f1SLi RongQing if (!weight) 840589506f1SLi RongQing return -ENOMEM; 841589506f1SLi RongQing map->osd_weight = weight; 842589506f1SLi RongQing 843589506f1SLi RongQing addr = krealloc(map->osd_addr, max*sizeof(*addr), GFP_NOFS); 844589506f1SLi RongQing if (!addr) 845589506f1SLi RongQing return -ENOMEM; 846589506f1SLi RongQing map->osd_addr = addr; 8473d14c5d2SYehuda Sadeh 8484d60351fSIlya Dryomov for (i = map->max_osd; i < max; i++) { 849589506f1SLi RongQing map->osd_state[i] = 0; 850589506f1SLi RongQing map->osd_weight[i] = CEPH_OSD_OUT; 851589506f1SLi RongQing memset(map->osd_addr + i, 0, sizeof(*map->osd_addr)); 8523d14c5d2SYehuda Sadeh } 8533d14c5d2SYehuda Sadeh 8542cfa34f2SIlya Dryomov if (map->osd_primary_affinity) { 8552cfa34f2SIlya Dryomov u32 *affinity; 8562cfa34f2SIlya Dryomov 8572cfa34f2SIlya Dryomov affinity = krealloc(map->osd_primary_affinity, 8582cfa34f2SIlya Dryomov max*sizeof(*affinity), GFP_NOFS); 8592cfa34f2SIlya Dryomov if (!affinity) 8602cfa34f2SIlya Dryomov return -ENOMEM; 861589506f1SLi RongQing map->osd_primary_affinity = affinity; 8622cfa34f2SIlya Dryomov 8632cfa34f2SIlya Dryomov for (i = map->max_osd; i < max; i++) 864589506f1SLi RongQing map->osd_primary_affinity[i] = 865589506f1SLi RongQing CEPH_OSD_DEFAULT_PRIMARY_AFFINITY; 8662cfa34f2SIlya Dryomov } 8672cfa34f2SIlya Dryomov 8683d14c5d2SYehuda Sadeh map->max_osd = max; 8694d60351fSIlya Dryomov 8703d14c5d2SYehuda Sadeh return 0; 8713d14c5d2SYehuda Sadeh } 8723d14c5d2SYehuda Sadeh 8731b6a78b5SIlya Dryomov static int osdmap_set_crush(struct ceph_osdmap *map, struct crush_map *crush) 8741b6a78b5SIlya Dryomov { 87566a0e2d5SIlya Dryomov void *workspace; 876743efcffSIlya Dryomov size_t work_size; 87766a0e2d5SIlya Dryomov 8781b6a78b5SIlya Dryomov if (IS_ERR(crush)) 8791b6a78b5SIlya Dryomov return PTR_ERR(crush); 8801b6a78b5SIlya Dryomov 881743efcffSIlya Dryomov work_size = crush_work_size(crush, CEPH_PG_MAX_SIZE); 882743efcffSIlya Dryomov dout("%s work_size %zu bytes\n", __func__, work_size); 883743efcffSIlya Dryomov workspace = kmalloc(work_size, GFP_NOIO); 88466a0e2d5SIlya Dryomov if (!workspace) { 88566a0e2d5SIlya Dryomov crush_destroy(crush); 88666a0e2d5SIlya Dryomov return -ENOMEM; 88766a0e2d5SIlya Dryomov } 88866a0e2d5SIlya Dryomov crush_init_workspace(crush, workspace); 88966a0e2d5SIlya Dryomov 8901b6a78b5SIlya Dryomov if (map->crush) 8911b6a78b5SIlya Dryomov crush_destroy(map->crush); 89266a0e2d5SIlya Dryomov kfree(map->crush_workspace); 8931b6a78b5SIlya Dryomov map->crush = crush; 89466a0e2d5SIlya Dryomov map->crush_workspace = workspace; 8951b6a78b5SIlya Dryomov return 0; 8961b6a78b5SIlya Dryomov } 8971b6a78b5SIlya Dryomov 898ec7af972SIlya Dryomov #define OSDMAP_WRAPPER_COMPAT_VER 7 899ec7af972SIlya Dryomov #define OSDMAP_CLIENT_DATA_COMPAT_VER 1 900ec7af972SIlya Dryomov 901ec7af972SIlya Dryomov /* 902ec7af972SIlya Dryomov * Return 0 or error. On success, *v is set to 0 for old (v6) osdmaps, 903ec7af972SIlya Dryomov * to struct_v of the client_data section for new (v7 and above) 904ec7af972SIlya Dryomov * osdmaps. 905ec7af972SIlya Dryomov */ 906ec7af972SIlya Dryomov static int get_osdmap_client_data_v(void **p, void *end, 907ec7af972SIlya Dryomov const char *prefix, u8 *v) 908ec7af972SIlya Dryomov { 909ec7af972SIlya Dryomov u8 struct_v; 910ec7af972SIlya Dryomov 911ec7af972SIlya Dryomov ceph_decode_8_safe(p, end, struct_v, e_inval); 912ec7af972SIlya Dryomov if (struct_v >= 7) { 913ec7af972SIlya Dryomov u8 struct_compat; 914ec7af972SIlya Dryomov 915ec7af972SIlya Dryomov ceph_decode_8_safe(p, end, struct_compat, e_inval); 916ec7af972SIlya Dryomov if (struct_compat > OSDMAP_WRAPPER_COMPAT_VER) { 917b9a67899SJoe Perches pr_warn("got v %d cv %d > %d of %s ceph_osdmap\n", 918ec7af972SIlya Dryomov struct_v, struct_compat, 919ec7af972SIlya Dryomov OSDMAP_WRAPPER_COMPAT_VER, prefix); 920ec7af972SIlya Dryomov return -EINVAL; 921ec7af972SIlya Dryomov } 922ec7af972SIlya Dryomov *p += 4; /* ignore wrapper struct_len */ 923ec7af972SIlya Dryomov 924ec7af972SIlya Dryomov ceph_decode_8_safe(p, end, struct_v, e_inval); 925ec7af972SIlya Dryomov ceph_decode_8_safe(p, end, struct_compat, e_inval); 926ec7af972SIlya Dryomov if (struct_compat > OSDMAP_CLIENT_DATA_COMPAT_VER) { 927b9a67899SJoe Perches pr_warn("got v %d cv %d > %d of %s ceph_osdmap client data\n", 928ec7af972SIlya Dryomov struct_v, struct_compat, 929ec7af972SIlya Dryomov OSDMAP_CLIENT_DATA_COMPAT_VER, prefix); 930ec7af972SIlya Dryomov return -EINVAL; 931ec7af972SIlya Dryomov } 932ec7af972SIlya Dryomov *p += 4; /* ignore client data struct_len */ 933ec7af972SIlya Dryomov } else { 934ec7af972SIlya Dryomov u16 version; 935ec7af972SIlya Dryomov 936ec7af972SIlya Dryomov *p -= 1; 937ec7af972SIlya Dryomov ceph_decode_16_safe(p, end, version, e_inval); 938ec7af972SIlya Dryomov if (version < 6) { 939b9a67899SJoe Perches pr_warn("got v %d < 6 of %s ceph_osdmap\n", 940b9a67899SJoe Perches version, prefix); 941ec7af972SIlya Dryomov return -EINVAL; 942ec7af972SIlya Dryomov } 943ec7af972SIlya Dryomov 944ec7af972SIlya Dryomov /* old osdmap enconding */ 945ec7af972SIlya Dryomov struct_v = 0; 946ec7af972SIlya Dryomov } 947ec7af972SIlya Dryomov 948ec7af972SIlya Dryomov *v = struct_v; 949ec7af972SIlya Dryomov return 0; 950ec7af972SIlya Dryomov 951ec7af972SIlya Dryomov e_inval: 952ec7af972SIlya Dryomov return -EINVAL; 953ec7af972SIlya Dryomov } 954ec7af972SIlya Dryomov 955433fbdd3SIlya Dryomov static int __decode_pools(void **p, void *end, struct ceph_osdmap *map, 956433fbdd3SIlya Dryomov bool incremental) 957433fbdd3SIlya Dryomov { 958433fbdd3SIlya Dryomov u32 n; 959433fbdd3SIlya Dryomov 960433fbdd3SIlya Dryomov ceph_decode_32_safe(p, end, n, e_inval); 961433fbdd3SIlya Dryomov while (n--) { 962433fbdd3SIlya Dryomov struct ceph_pg_pool_info *pi; 963433fbdd3SIlya Dryomov u64 pool; 964433fbdd3SIlya Dryomov int ret; 965433fbdd3SIlya Dryomov 966433fbdd3SIlya Dryomov ceph_decode_64_safe(p, end, pool, e_inval); 967433fbdd3SIlya Dryomov 968433fbdd3SIlya Dryomov pi = __lookup_pg_pool(&map->pg_pools, pool); 969433fbdd3SIlya Dryomov if (!incremental || !pi) { 970433fbdd3SIlya Dryomov pi = kzalloc(sizeof(*pi), GFP_NOFS); 971433fbdd3SIlya Dryomov if (!pi) 972433fbdd3SIlya Dryomov return -ENOMEM; 973433fbdd3SIlya Dryomov 974433fbdd3SIlya Dryomov pi->id = pool; 975433fbdd3SIlya Dryomov 976433fbdd3SIlya Dryomov ret = __insert_pg_pool(&map->pg_pools, pi); 977433fbdd3SIlya Dryomov if (ret) { 978433fbdd3SIlya Dryomov kfree(pi); 979433fbdd3SIlya Dryomov return ret; 980433fbdd3SIlya Dryomov } 981433fbdd3SIlya Dryomov } 982433fbdd3SIlya Dryomov 983433fbdd3SIlya Dryomov ret = decode_pool(p, end, pi); 984433fbdd3SIlya Dryomov if (ret) 985433fbdd3SIlya Dryomov return ret; 986433fbdd3SIlya Dryomov } 987433fbdd3SIlya Dryomov 988433fbdd3SIlya Dryomov return 0; 989433fbdd3SIlya Dryomov 990433fbdd3SIlya Dryomov e_inval: 991433fbdd3SIlya Dryomov return -EINVAL; 992433fbdd3SIlya Dryomov } 993433fbdd3SIlya Dryomov 994433fbdd3SIlya Dryomov static int decode_pools(void **p, void *end, struct ceph_osdmap *map) 995433fbdd3SIlya Dryomov { 996433fbdd3SIlya Dryomov return __decode_pools(p, end, map, false); 997433fbdd3SIlya Dryomov } 998433fbdd3SIlya Dryomov 999433fbdd3SIlya Dryomov static int decode_new_pools(void **p, void *end, struct ceph_osdmap *map) 1000433fbdd3SIlya Dryomov { 1001433fbdd3SIlya Dryomov return __decode_pools(p, end, map, true); 1002433fbdd3SIlya Dryomov } 1003433fbdd3SIlya Dryomov 100410db634eSIlya Dryomov static int __decode_pg_temp(void **p, void *end, struct ceph_osdmap *map, 100510db634eSIlya Dryomov bool incremental) 100610db634eSIlya Dryomov { 100710db634eSIlya Dryomov u32 n; 100810db634eSIlya Dryomov 100910db634eSIlya Dryomov ceph_decode_32_safe(p, end, n, e_inval); 101010db634eSIlya Dryomov while (n--) { 101110db634eSIlya Dryomov struct ceph_pg pgid; 101210db634eSIlya Dryomov u32 len, i; 101310db634eSIlya Dryomov int ret; 101410db634eSIlya Dryomov 101510db634eSIlya Dryomov ret = ceph_decode_pgid(p, end, &pgid); 101610db634eSIlya Dryomov if (ret) 101710db634eSIlya Dryomov return ret; 101810db634eSIlya Dryomov 101910db634eSIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 102010db634eSIlya Dryomov 102110db634eSIlya Dryomov ret = __remove_pg_mapping(&map->pg_temp, pgid); 102210db634eSIlya Dryomov BUG_ON(!incremental && ret != -ENOENT); 102310db634eSIlya Dryomov 102410db634eSIlya Dryomov if (!incremental || len > 0) { 102510db634eSIlya Dryomov struct ceph_pg_mapping *pg; 102610db634eSIlya Dryomov 102710db634eSIlya Dryomov ceph_decode_need(p, end, len*sizeof(u32), e_inval); 102810db634eSIlya Dryomov 102910db634eSIlya Dryomov if (len > (UINT_MAX - sizeof(*pg)) / sizeof(u32)) 103010db634eSIlya Dryomov return -EINVAL; 103110db634eSIlya Dryomov 103210db634eSIlya Dryomov pg = kzalloc(sizeof(*pg) + len*sizeof(u32), GFP_NOFS); 103310db634eSIlya Dryomov if (!pg) 103410db634eSIlya Dryomov return -ENOMEM; 103510db634eSIlya Dryomov 103610db634eSIlya Dryomov pg->pgid = pgid; 103735a935d7SIlya Dryomov pg->pg_temp.len = len; 103810db634eSIlya Dryomov for (i = 0; i < len; i++) 103935a935d7SIlya Dryomov pg->pg_temp.osds[i] = ceph_decode_32(p); 104010db634eSIlya Dryomov 104110db634eSIlya Dryomov ret = __insert_pg_mapping(pg, &map->pg_temp); 104210db634eSIlya Dryomov if (ret) { 104310db634eSIlya Dryomov kfree(pg); 104410db634eSIlya Dryomov return ret; 104510db634eSIlya Dryomov } 104610db634eSIlya Dryomov } 104710db634eSIlya Dryomov } 104810db634eSIlya Dryomov 104910db634eSIlya Dryomov return 0; 105010db634eSIlya Dryomov 105110db634eSIlya Dryomov e_inval: 105210db634eSIlya Dryomov return -EINVAL; 105310db634eSIlya Dryomov } 105410db634eSIlya Dryomov 105510db634eSIlya Dryomov static int decode_pg_temp(void **p, void *end, struct ceph_osdmap *map) 105610db634eSIlya Dryomov { 105710db634eSIlya Dryomov return __decode_pg_temp(p, end, map, false); 105810db634eSIlya Dryomov } 105910db634eSIlya Dryomov 106010db634eSIlya Dryomov static int decode_new_pg_temp(void **p, void *end, struct ceph_osdmap *map) 106110db634eSIlya Dryomov { 106210db634eSIlya Dryomov return __decode_pg_temp(p, end, map, true); 106310db634eSIlya Dryomov } 106410db634eSIlya Dryomov 1065d286de79SIlya Dryomov static int __decode_primary_temp(void **p, void *end, struct ceph_osdmap *map, 1066d286de79SIlya Dryomov bool incremental) 1067d286de79SIlya Dryomov { 1068d286de79SIlya Dryomov u32 n; 1069d286de79SIlya Dryomov 1070d286de79SIlya Dryomov ceph_decode_32_safe(p, end, n, e_inval); 1071d286de79SIlya Dryomov while (n--) { 1072d286de79SIlya Dryomov struct ceph_pg pgid; 1073d286de79SIlya Dryomov u32 osd; 1074d286de79SIlya Dryomov int ret; 1075d286de79SIlya Dryomov 1076d286de79SIlya Dryomov ret = ceph_decode_pgid(p, end, &pgid); 1077d286de79SIlya Dryomov if (ret) 1078d286de79SIlya Dryomov return ret; 1079d286de79SIlya Dryomov 1080d286de79SIlya Dryomov ceph_decode_32_safe(p, end, osd, e_inval); 1081d286de79SIlya Dryomov 1082d286de79SIlya Dryomov ret = __remove_pg_mapping(&map->primary_temp, pgid); 1083d286de79SIlya Dryomov BUG_ON(!incremental && ret != -ENOENT); 1084d286de79SIlya Dryomov 1085d286de79SIlya Dryomov if (!incremental || osd != (u32)-1) { 1086d286de79SIlya Dryomov struct ceph_pg_mapping *pg; 1087d286de79SIlya Dryomov 1088d286de79SIlya Dryomov pg = kzalloc(sizeof(*pg), GFP_NOFS); 1089d286de79SIlya Dryomov if (!pg) 1090d286de79SIlya Dryomov return -ENOMEM; 1091d286de79SIlya Dryomov 1092d286de79SIlya Dryomov pg->pgid = pgid; 1093d286de79SIlya Dryomov pg->primary_temp.osd = osd; 1094d286de79SIlya Dryomov 1095d286de79SIlya Dryomov ret = __insert_pg_mapping(pg, &map->primary_temp); 1096d286de79SIlya Dryomov if (ret) { 1097d286de79SIlya Dryomov kfree(pg); 1098d286de79SIlya Dryomov return ret; 1099d286de79SIlya Dryomov } 1100d286de79SIlya Dryomov } 1101d286de79SIlya Dryomov } 1102d286de79SIlya Dryomov 1103d286de79SIlya Dryomov return 0; 1104d286de79SIlya Dryomov 1105d286de79SIlya Dryomov e_inval: 1106d286de79SIlya Dryomov return -EINVAL; 1107d286de79SIlya Dryomov } 1108d286de79SIlya Dryomov 1109d286de79SIlya Dryomov static int decode_primary_temp(void **p, void *end, struct ceph_osdmap *map) 1110d286de79SIlya Dryomov { 1111d286de79SIlya Dryomov return __decode_primary_temp(p, end, map, false); 1112d286de79SIlya Dryomov } 1113d286de79SIlya Dryomov 1114d286de79SIlya Dryomov static int decode_new_primary_temp(void **p, void *end, 1115d286de79SIlya Dryomov struct ceph_osdmap *map) 1116d286de79SIlya Dryomov { 1117d286de79SIlya Dryomov return __decode_primary_temp(p, end, map, true); 1118d286de79SIlya Dryomov } 1119d286de79SIlya Dryomov 11202cfa34f2SIlya Dryomov u32 ceph_get_primary_affinity(struct ceph_osdmap *map, int osd) 11212cfa34f2SIlya Dryomov { 11222cfa34f2SIlya Dryomov BUG_ON(osd >= map->max_osd); 11232cfa34f2SIlya Dryomov 11242cfa34f2SIlya Dryomov if (!map->osd_primary_affinity) 11252cfa34f2SIlya Dryomov return CEPH_OSD_DEFAULT_PRIMARY_AFFINITY; 11262cfa34f2SIlya Dryomov 11272cfa34f2SIlya Dryomov return map->osd_primary_affinity[osd]; 11282cfa34f2SIlya Dryomov } 11292cfa34f2SIlya Dryomov 11302cfa34f2SIlya Dryomov static int set_primary_affinity(struct ceph_osdmap *map, int osd, u32 aff) 11312cfa34f2SIlya Dryomov { 11322cfa34f2SIlya Dryomov BUG_ON(osd >= map->max_osd); 11332cfa34f2SIlya Dryomov 11342cfa34f2SIlya Dryomov if (!map->osd_primary_affinity) { 11352cfa34f2SIlya Dryomov int i; 11362cfa34f2SIlya Dryomov 11372cfa34f2SIlya Dryomov map->osd_primary_affinity = kmalloc(map->max_osd*sizeof(u32), 11382cfa34f2SIlya Dryomov GFP_NOFS); 11392cfa34f2SIlya Dryomov if (!map->osd_primary_affinity) 11402cfa34f2SIlya Dryomov return -ENOMEM; 11412cfa34f2SIlya Dryomov 11422cfa34f2SIlya Dryomov for (i = 0; i < map->max_osd; i++) 11432cfa34f2SIlya Dryomov map->osd_primary_affinity[i] = 11442cfa34f2SIlya Dryomov CEPH_OSD_DEFAULT_PRIMARY_AFFINITY; 11452cfa34f2SIlya Dryomov } 11462cfa34f2SIlya Dryomov 11472cfa34f2SIlya Dryomov map->osd_primary_affinity[osd] = aff; 11482cfa34f2SIlya Dryomov 11492cfa34f2SIlya Dryomov return 0; 11502cfa34f2SIlya Dryomov } 11512cfa34f2SIlya Dryomov 115263a6993fSIlya Dryomov static int decode_primary_affinity(void **p, void *end, 115363a6993fSIlya Dryomov struct ceph_osdmap *map) 115463a6993fSIlya Dryomov { 115563a6993fSIlya Dryomov u32 len, i; 115663a6993fSIlya Dryomov 115763a6993fSIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 115863a6993fSIlya Dryomov if (len == 0) { 115963a6993fSIlya Dryomov kfree(map->osd_primary_affinity); 116063a6993fSIlya Dryomov map->osd_primary_affinity = NULL; 116163a6993fSIlya Dryomov return 0; 116263a6993fSIlya Dryomov } 116363a6993fSIlya Dryomov if (len != map->max_osd) 116463a6993fSIlya Dryomov goto e_inval; 116563a6993fSIlya Dryomov 116663a6993fSIlya Dryomov ceph_decode_need(p, end, map->max_osd*sizeof(u32), e_inval); 116763a6993fSIlya Dryomov 116863a6993fSIlya Dryomov for (i = 0; i < map->max_osd; i++) { 116963a6993fSIlya Dryomov int ret; 117063a6993fSIlya Dryomov 117163a6993fSIlya Dryomov ret = set_primary_affinity(map, i, ceph_decode_32(p)); 117263a6993fSIlya Dryomov if (ret) 117363a6993fSIlya Dryomov return ret; 117463a6993fSIlya Dryomov } 117563a6993fSIlya Dryomov 117663a6993fSIlya Dryomov return 0; 117763a6993fSIlya Dryomov 117863a6993fSIlya Dryomov e_inval: 117963a6993fSIlya Dryomov return -EINVAL; 118063a6993fSIlya Dryomov } 118163a6993fSIlya Dryomov 118263a6993fSIlya Dryomov static int decode_new_primary_affinity(void **p, void *end, 118363a6993fSIlya Dryomov struct ceph_osdmap *map) 118463a6993fSIlya Dryomov { 118563a6993fSIlya Dryomov u32 n; 118663a6993fSIlya Dryomov 118763a6993fSIlya Dryomov ceph_decode_32_safe(p, end, n, e_inval); 118863a6993fSIlya Dryomov while (n--) { 118963a6993fSIlya Dryomov u32 osd, aff; 119063a6993fSIlya Dryomov int ret; 119163a6993fSIlya Dryomov 119263a6993fSIlya Dryomov ceph_decode_32_safe(p, end, osd, e_inval); 119363a6993fSIlya Dryomov ceph_decode_32_safe(p, end, aff, e_inval); 119463a6993fSIlya Dryomov 119563a6993fSIlya Dryomov ret = set_primary_affinity(map, osd, aff); 119663a6993fSIlya Dryomov if (ret) 119763a6993fSIlya Dryomov return ret; 1198f31da0f3SIlya Dryomov 1199f31da0f3SIlya Dryomov pr_info("osd%d primary-affinity 0x%x\n", osd, aff); 120063a6993fSIlya Dryomov } 120163a6993fSIlya Dryomov 120263a6993fSIlya Dryomov return 0; 120363a6993fSIlya Dryomov 120463a6993fSIlya Dryomov e_inval: 120563a6993fSIlya Dryomov return -EINVAL; 120663a6993fSIlya Dryomov } 120763a6993fSIlya Dryomov 12083d14c5d2SYehuda Sadeh /* 12093d14c5d2SYehuda Sadeh * decode a full map. 12103d14c5d2SYehuda Sadeh */ 1211a2505d63SIlya Dryomov static int osdmap_decode(void **p, void *end, struct ceph_osdmap *map) 12123d14c5d2SYehuda Sadeh { 1213ec7af972SIlya Dryomov u8 struct_v; 121438a8d560SIlya Dryomov u32 epoch = 0; 12153d14c5d2SYehuda Sadeh void *start = *p; 12163977058cSIlya Dryomov u32 max; 12173977058cSIlya Dryomov u32 len, i; 1218597b52f6SIlya Dryomov int err; 12193d14c5d2SYehuda Sadeh 1220a2505d63SIlya Dryomov dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p)); 12213d14c5d2SYehuda Sadeh 1222ec7af972SIlya Dryomov err = get_osdmap_client_data_v(p, end, "full", &struct_v); 1223ec7af972SIlya Dryomov if (err) 1224ec7af972SIlya Dryomov goto bad; 12253d14c5d2SYehuda Sadeh 122653bbaba9SIlya Dryomov /* fsid, epoch, created, modified */ 122753bbaba9SIlya Dryomov ceph_decode_need(p, end, sizeof(map->fsid) + sizeof(u32) + 122853bbaba9SIlya Dryomov sizeof(map->created) + sizeof(map->modified), e_inval); 12293d14c5d2SYehuda Sadeh ceph_decode_copy(p, &map->fsid, sizeof(map->fsid)); 123038a8d560SIlya Dryomov epoch = map->epoch = ceph_decode_32(p); 12313d14c5d2SYehuda Sadeh ceph_decode_copy(p, &map->created, sizeof(map->created)); 12323d14c5d2SYehuda Sadeh ceph_decode_copy(p, &map->modified, sizeof(map->modified)); 12333d14c5d2SYehuda Sadeh 1234433fbdd3SIlya Dryomov /* pools */ 1235433fbdd3SIlya Dryomov err = decode_pools(p, end, map); 1236433fbdd3SIlya Dryomov if (err) 12373d14c5d2SYehuda Sadeh goto bad; 12383d14c5d2SYehuda Sadeh 12390f70c7eeSIlya Dryomov /* pool_name */ 12400f70c7eeSIlya Dryomov err = decode_pool_names(p, end, map); 1241597b52f6SIlya Dryomov if (err) 12423d14c5d2SYehuda Sadeh goto bad; 12433d14c5d2SYehuda Sadeh 1244597b52f6SIlya Dryomov ceph_decode_32_safe(p, end, map->pool_max, e_inval); 12453d14c5d2SYehuda Sadeh 1246597b52f6SIlya Dryomov ceph_decode_32_safe(p, end, map->flags, e_inval); 12473d14c5d2SYehuda Sadeh 12483977058cSIlya Dryomov /* max_osd */ 12493977058cSIlya Dryomov ceph_decode_32_safe(p, end, max, e_inval); 12503d14c5d2SYehuda Sadeh 12513d14c5d2SYehuda Sadeh /* (re)alloc osd arrays */ 12523d14c5d2SYehuda Sadeh err = osdmap_set_max_osd(map, max); 1253597b52f6SIlya Dryomov if (err) 12543d14c5d2SYehuda Sadeh goto bad; 12553d14c5d2SYehuda Sadeh 12562d88b2e0SIlya Dryomov /* osd_state, osd_weight, osd_addrs->client_addr */ 12573d14c5d2SYehuda Sadeh ceph_decode_need(p, end, 3*sizeof(u32) + 12583d14c5d2SYehuda Sadeh map->max_osd*(1 + sizeof(*map->osd_weight) + 1259597b52f6SIlya Dryomov sizeof(*map->osd_addr)), e_inval); 1260597b52f6SIlya Dryomov 12612d88b2e0SIlya Dryomov if (ceph_decode_32(p) != map->max_osd) 12622d88b2e0SIlya Dryomov goto e_inval; 12632d88b2e0SIlya Dryomov 12643d14c5d2SYehuda Sadeh ceph_decode_copy(p, map->osd_state, map->max_osd); 12653d14c5d2SYehuda Sadeh 12662d88b2e0SIlya Dryomov if (ceph_decode_32(p) != map->max_osd) 12672d88b2e0SIlya Dryomov goto e_inval; 12682d88b2e0SIlya Dryomov 12693d14c5d2SYehuda Sadeh for (i = 0; i < map->max_osd; i++) 12703d14c5d2SYehuda Sadeh map->osd_weight[i] = ceph_decode_32(p); 12713d14c5d2SYehuda Sadeh 12722d88b2e0SIlya Dryomov if (ceph_decode_32(p) != map->max_osd) 12732d88b2e0SIlya Dryomov goto e_inval; 12742d88b2e0SIlya Dryomov 12753d14c5d2SYehuda Sadeh ceph_decode_copy(p, map->osd_addr, map->max_osd*sizeof(*map->osd_addr)); 12763d14c5d2SYehuda Sadeh for (i = 0; i < map->max_osd; i++) 12773d14c5d2SYehuda Sadeh ceph_decode_addr(&map->osd_addr[i]); 12783d14c5d2SYehuda Sadeh 12793d14c5d2SYehuda Sadeh /* pg_temp */ 128010db634eSIlya Dryomov err = decode_pg_temp(p, end, map); 1281d6c0dd6bSSage Weil if (err) 1282d6c0dd6bSSage Weil goto bad; 12833d14c5d2SYehuda Sadeh 1284d286de79SIlya Dryomov /* primary_temp */ 1285d286de79SIlya Dryomov if (struct_v >= 1) { 1286d286de79SIlya Dryomov err = decode_primary_temp(p, end, map); 1287d286de79SIlya Dryomov if (err) 1288d286de79SIlya Dryomov goto bad; 1289d286de79SIlya Dryomov } 1290d286de79SIlya Dryomov 129163a6993fSIlya Dryomov /* primary_affinity */ 129263a6993fSIlya Dryomov if (struct_v >= 2) { 129363a6993fSIlya Dryomov err = decode_primary_affinity(p, end, map); 129463a6993fSIlya Dryomov if (err) 129563a6993fSIlya Dryomov goto bad; 129663a6993fSIlya Dryomov } else { 129763a6993fSIlya Dryomov /* XXX can this happen? */ 129863a6993fSIlya Dryomov kfree(map->osd_primary_affinity); 129963a6993fSIlya Dryomov map->osd_primary_affinity = NULL; 130063a6993fSIlya Dryomov } 130163a6993fSIlya Dryomov 13023d14c5d2SYehuda Sadeh /* crush */ 1303597b52f6SIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 13041b6a78b5SIlya Dryomov err = osdmap_set_crush(map, crush_decode(*p, min(*p + len, end))); 13051b6a78b5SIlya Dryomov if (err) 13063d14c5d2SYehuda Sadeh goto bad; 13073d14c5d2SYehuda Sadeh 130838a8d560SIlya Dryomov /* ignore the rest */ 13093d14c5d2SYehuda Sadeh *p = end; 13103d14c5d2SYehuda Sadeh 131138a8d560SIlya Dryomov dout("full osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd); 1312a2505d63SIlya Dryomov return 0; 13133d14c5d2SYehuda Sadeh 1314597b52f6SIlya Dryomov e_inval: 1315597b52f6SIlya Dryomov err = -EINVAL; 13163d14c5d2SYehuda Sadeh bad: 131738a8d560SIlya Dryomov pr_err("corrupt full osdmap (%d) epoch %d off %d (%p of %p-%p)\n", 131838a8d560SIlya Dryomov err, epoch, (int)(*p - start), *p, start, end); 131938a8d560SIlya Dryomov print_hex_dump(KERN_DEBUG, "osdmap: ", 132038a8d560SIlya Dryomov DUMP_PREFIX_OFFSET, 16, 1, 132138a8d560SIlya Dryomov start, end - start, true); 1322a2505d63SIlya Dryomov return err; 1323a2505d63SIlya Dryomov } 1324a2505d63SIlya Dryomov 1325a2505d63SIlya Dryomov /* 1326a2505d63SIlya Dryomov * Allocate and decode a full map. 1327a2505d63SIlya Dryomov */ 1328a2505d63SIlya Dryomov struct ceph_osdmap *ceph_osdmap_decode(void **p, void *end) 1329a2505d63SIlya Dryomov { 1330a2505d63SIlya Dryomov struct ceph_osdmap *map; 1331a2505d63SIlya Dryomov int ret; 1332a2505d63SIlya Dryomov 1333e5253a7bSIlya Dryomov map = ceph_osdmap_alloc(); 1334a2505d63SIlya Dryomov if (!map) 1335a2505d63SIlya Dryomov return ERR_PTR(-ENOMEM); 1336a2505d63SIlya Dryomov 1337a2505d63SIlya Dryomov ret = osdmap_decode(p, end, map); 1338a2505d63SIlya Dryomov if (ret) { 13393d14c5d2SYehuda Sadeh ceph_osdmap_destroy(map); 1340a2505d63SIlya Dryomov return ERR_PTR(ret); 1341a2505d63SIlya Dryomov } 1342a2505d63SIlya Dryomov 1343a2505d63SIlya Dryomov return map; 13443d14c5d2SYehuda Sadeh } 13453d14c5d2SYehuda Sadeh 13463d14c5d2SYehuda Sadeh /* 1347930c5328SIlya Dryomov * Encoding order is (new_up_client, new_state, new_weight). Need to 1348930c5328SIlya Dryomov * apply in the (new_weight, new_state, new_up_client) order, because 1349930c5328SIlya Dryomov * an incremental map may look like e.g. 1350930c5328SIlya Dryomov * 1351930c5328SIlya Dryomov * new_up_client: { osd=6, addr=... } # set osd_state and addr 1352930c5328SIlya Dryomov * new_state: { osd=6, xorstate=EXISTS } # clear osd_state 1353930c5328SIlya Dryomov */ 1354930c5328SIlya Dryomov static int decode_new_up_state_weight(void **p, void *end, 1355930c5328SIlya Dryomov struct ceph_osdmap *map) 1356930c5328SIlya Dryomov { 1357930c5328SIlya Dryomov void *new_up_client; 1358930c5328SIlya Dryomov void *new_state; 1359930c5328SIlya Dryomov void *new_weight_end; 1360930c5328SIlya Dryomov u32 len; 1361930c5328SIlya Dryomov 1362930c5328SIlya Dryomov new_up_client = *p; 1363930c5328SIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 1364930c5328SIlya Dryomov len *= sizeof(u32) + sizeof(struct ceph_entity_addr); 1365930c5328SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1366930c5328SIlya Dryomov *p += len; 1367930c5328SIlya Dryomov 1368930c5328SIlya Dryomov new_state = *p; 1369930c5328SIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 1370930c5328SIlya Dryomov len *= sizeof(u32) + sizeof(u8); 1371930c5328SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 1372930c5328SIlya Dryomov *p += len; 1373930c5328SIlya Dryomov 1374930c5328SIlya Dryomov /* new_weight */ 1375930c5328SIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 1376930c5328SIlya Dryomov while (len--) { 1377930c5328SIlya Dryomov s32 osd; 1378930c5328SIlya Dryomov u32 w; 1379930c5328SIlya Dryomov 1380930c5328SIlya Dryomov ceph_decode_need(p, end, 2*sizeof(u32), e_inval); 1381930c5328SIlya Dryomov osd = ceph_decode_32(p); 1382930c5328SIlya Dryomov w = ceph_decode_32(p); 1383930c5328SIlya Dryomov BUG_ON(osd >= map->max_osd); 1384930c5328SIlya Dryomov pr_info("osd%d weight 0x%x %s\n", osd, w, 1385930c5328SIlya Dryomov w == CEPH_OSD_IN ? "(in)" : 1386930c5328SIlya Dryomov (w == CEPH_OSD_OUT ? "(out)" : "")); 1387930c5328SIlya Dryomov map->osd_weight[osd] = w; 1388930c5328SIlya Dryomov 1389930c5328SIlya Dryomov /* 1390930c5328SIlya Dryomov * If we are marking in, set the EXISTS, and clear the 1391930c5328SIlya Dryomov * AUTOOUT and NEW bits. 1392930c5328SIlya Dryomov */ 1393930c5328SIlya Dryomov if (w) { 1394930c5328SIlya Dryomov map->osd_state[osd] |= CEPH_OSD_EXISTS; 1395930c5328SIlya Dryomov map->osd_state[osd] &= ~(CEPH_OSD_AUTOOUT | 1396930c5328SIlya Dryomov CEPH_OSD_NEW); 1397930c5328SIlya Dryomov } 1398930c5328SIlya Dryomov } 1399930c5328SIlya Dryomov new_weight_end = *p; 1400930c5328SIlya Dryomov 1401930c5328SIlya Dryomov /* new_state (up/down) */ 1402930c5328SIlya Dryomov *p = new_state; 1403930c5328SIlya Dryomov len = ceph_decode_32(p); 1404930c5328SIlya Dryomov while (len--) { 1405930c5328SIlya Dryomov s32 osd; 1406930c5328SIlya Dryomov u8 xorstate; 1407930c5328SIlya Dryomov int ret; 1408930c5328SIlya Dryomov 1409930c5328SIlya Dryomov osd = ceph_decode_32(p); 1410930c5328SIlya Dryomov xorstate = ceph_decode_8(p); 1411930c5328SIlya Dryomov if (xorstate == 0) 1412930c5328SIlya Dryomov xorstate = CEPH_OSD_UP; 1413930c5328SIlya Dryomov BUG_ON(osd >= map->max_osd); 1414930c5328SIlya Dryomov if ((map->osd_state[osd] & CEPH_OSD_UP) && 1415930c5328SIlya Dryomov (xorstate & CEPH_OSD_UP)) 1416930c5328SIlya Dryomov pr_info("osd%d down\n", osd); 1417930c5328SIlya Dryomov if ((map->osd_state[osd] & CEPH_OSD_EXISTS) && 1418930c5328SIlya Dryomov (xorstate & CEPH_OSD_EXISTS)) { 1419930c5328SIlya Dryomov pr_info("osd%d does not exist\n", osd); 1420930c5328SIlya Dryomov ret = set_primary_affinity(map, osd, 1421930c5328SIlya Dryomov CEPH_OSD_DEFAULT_PRIMARY_AFFINITY); 1422930c5328SIlya Dryomov if (ret) 1423930c5328SIlya Dryomov return ret; 1424930c5328SIlya Dryomov memset(map->osd_addr + osd, 0, sizeof(*map->osd_addr)); 1425930c5328SIlya Dryomov map->osd_state[osd] = 0; 1426930c5328SIlya Dryomov } else { 1427930c5328SIlya Dryomov map->osd_state[osd] ^= xorstate; 1428930c5328SIlya Dryomov } 1429930c5328SIlya Dryomov } 1430930c5328SIlya Dryomov 1431930c5328SIlya Dryomov /* new_up_client */ 1432930c5328SIlya Dryomov *p = new_up_client; 1433930c5328SIlya Dryomov len = ceph_decode_32(p); 1434930c5328SIlya Dryomov while (len--) { 1435930c5328SIlya Dryomov s32 osd; 1436930c5328SIlya Dryomov struct ceph_entity_addr addr; 1437930c5328SIlya Dryomov 1438930c5328SIlya Dryomov osd = ceph_decode_32(p); 1439930c5328SIlya Dryomov ceph_decode_copy(p, &addr, sizeof(addr)); 1440930c5328SIlya Dryomov ceph_decode_addr(&addr); 1441930c5328SIlya Dryomov BUG_ON(osd >= map->max_osd); 1442930c5328SIlya Dryomov pr_info("osd%d up\n", osd); 1443930c5328SIlya Dryomov map->osd_state[osd] |= CEPH_OSD_EXISTS | CEPH_OSD_UP; 1444930c5328SIlya Dryomov map->osd_addr[osd] = addr; 1445930c5328SIlya Dryomov } 1446930c5328SIlya Dryomov 1447930c5328SIlya Dryomov *p = new_weight_end; 1448930c5328SIlya Dryomov return 0; 1449930c5328SIlya Dryomov 1450930c5328SIlya Dryomov e_inval: 1451930c5328SIlya Dryomov return -EINVAL; 1452930c5328SIlya Dryomov } 1453930c5328SIlya Dryomov 1454930c5328SIlya Dryomov /* 14553d14c5d2SYehuda Sadeh * decode and apply an incremental map update. 14563d14c5d2SYehuda Sadeh */ 14573d14c5d2SYehuda Sadeh struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end, 14580c0a8de1SIlya Dryomov struct ceph_osdmap *map) 14593d14c5d2SYehuda Sadeh { 14603d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 14613d14c5d2SYehuda Sadeh u32 epoch = 0; 14623d14c5d2SYehuda Sadeh struct ceph_timespec modified; 14634f6a7e5eSSage Weil s32 len; 14644f6a7e5eSSage Weil u64 pool; 14654f6a7e5eSSage Weil __s64 new_pool_max; 14664f6a7e5eSSage Weil __s32 new_flags, max; 14673d14c5d2SYehuda Sadeh void *start = *p; 146886f1742bSIlya Dryomov int err; 1469ec7af972SIlya Dryomov u8 struct_v; 14703d14c5d2SYehuda Sadeh 147138a8d560SIlya Dryomov dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p)); 147238a8d560SIlya Dryomov 1473ec7af972SIlya Dryomov err = get_osdmap_client_data_v(p, end, "inc", &struct_v); 1474ec7af972SIlya Dryomov if (err) 1475ec7af972SIlya Dryomov goto bad; 14763d14c5d2SYehuda Sadeh 147753bbaba9SIlya Dryomov /* fsid, epoch, modified, new_pool_max, new_flags */ 147853bbaba9SIlya Dryomov ceph_decode_need(p, end, sizeof(fsid) + sizeof(u32) + sizeof(modified) + 147953bbaba9SIlya Dryomov sizeof(u64) + sizeof(u32), e_inval); 14803d14c5d2SYehuda Sadeh ceph_decode_copy(p, &fsid, sizeof(fsid)); 14813d14c5d2SYehuda Sadeh epoch = ceph_decode_32(p); 14823d14c5d2SYehuda Sadeh BUG_ON(epoch != map->epoch+1); 14833d14c5d2SYehuda Sadeh ceph_decode_copy(p, &modified, sizeof(modified)); 14844f6a7e5eSSage Weil new_pool_max = ceph_decode_64(p); 14853d14c5d2SYehuda Sadeh new_flags = ceph_decode_32(p); 14863d14c5d2SYehuda Sadeh 14873d14c5d2SYehuda Sadeh /* full map? */ 148886f1742bSIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 14893d14c5d2SYehuda Sadeh if (len > 0) { 14903d14c5d2SYehuda Sadeh dout("apply_incremental full map len %d, %p to %p\n", 14913d14c5d2SYehuda Sadeh len, *p, end); 1492a2505d63SIlya Dryomov return ceph_osdmap_decode(p, min(*p+len, end)); 14933d14c5d2SYehuda Sadeh } 14943d14c5d2SYehuda Sadeh 14953d14c5d2SYehuda Sadeh /* new crush? */ 149686f1742bSIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 14973d14c5d2SYehuda Sadeh if (len > 0) { 14981b6a78b5SIlya Dryomov err = osdmap_set_crush(map, 14991b6a78b5SIlya Dryomov crush_decode(*p, min(*p + len, end))); 15001b6a78b5SIlya Dryomov if (err) 150186f1742bSIlya Dryomov goto bad; 15023d14c5d2SYehuda Sadeh *p += len; 15033d14c5d2SYehuda Sadeh } 15043d14c5d2SYehuda Sadeh 15053d14c5d2SYehuda Sadeh /* new flags? */ 15063d14c5d2SYehuda Sadeh if (new_flags >= 0) 15073d14c5d2SYehuda Sadeh map->flags = new_flags; 15083d14c5d2SYehuda Sadeh if (new_pool_max >= 0) 15093d14c5d2SYehuda Sadeh map->pool_max = new_pool_max; 15103d14c5d2SYehuda Sadeh 15113d14c5d2SYehuda Sadeh /* new max? */ 151253bbaba9SIlya Dryomov ceph_decode_32_safe(p, end, max, e_inval); 15133d14c5d2SYehuda Sadeh if (max >= 0) { 15143d14c5d2SYehuda Sadeh err = osdmap_set_max_osd(map, max); 151586f1742bSIlya Dryomov if (err) 15163d14c5d2SYehuda Sadeh goto bad; 15173d14c5d2SYehuda Sadeh } 15183d14c5d2SYehuda Sadeh 15193d14c5d2SYehuda Sadeh map->epoch++; 152031456665SSage Weil map->modified = modified; 15213d14c5d2SYehuda Sadeh 1522433fbdd3SIlya Dryomov /* new_pools */ 1523433fbdd3SIlya Dryomov err = decode_new_pools(p, end, map); 1524433fbdd3SIlya Dryomov if (err) 15253d14c5d2SYehuda Sadeh goto bad; 15269464d008SIlya Dryomov 15270f70c7eeSIlya Dryomov /* new_pool_names */ 15280f70c7eeSIlya Dryomov err = decode_pool_names(p, end, map); 152986f1742bSIlya Dryomov if (err) 15303d14c5d2SYehuda Sadeh goto bad; 15313d14c5d2SYehuda Sadeh 15323d14c5d2SYehuda Sadeh /* old_pool */ 153386f1742bSIlya Dryomov ceph_decode_32_safe(p, end, len, e_inval); 15343d14c5d2SYehuda Sadeh while (len--) { 15353d14c5d2SYehuda Sadeh struct ceph_pg_pool_info *pi; 15363d14c5d2SYehuda Sadeh 153786f1742bSIlya Dryomov ceph_decode_64_safe(p, end, pool, e_inval); 15383d14c5d2SYehuda Sadeh pi = __lookup_pg_pool(&map->pg_pools, pool); 15393d14c5d2SYehuda Sadeh if (pi) 15403d14c5d2SYehuda Sadeh __remove_pg_pool(&map->pg_pools, pi); 15413d14c5d2SYehuda Sadeh } 15423d14c5d2SYehuda Sadeh 1543930c5328SIlya Dryomov /* new_up_client, new_state, new_weight */ 1544930c5328SIlya Dryomov err = decode_new_up_state_weight(p, end, map); 1545930c5328SIlya Dryomov if (err) 1546930c5328SIlya Dryomov goto bad; 15473d14c5d2SYehuda Sadeh 15483d14c5d2SYehuda Sadeh /* new_pg_temp */ 154910db634eSIlya Dryomov err = decode_new_pg_temp(p, end, map); 1550d6c0dd6bSSage Weil if (err) 1551d6c0dd6bSSage Weil goto bad; 15523d14c5d2SYehuda Sadeh 1553d286de79SIlya Dryomov /* new_primary_temp */ 1554d286de79SIlya Dryomov if (struct_v >= 1) { 1555d286de79SIlya Dryomov err = decode_new_primary_temp(p, end, map); 1556d286de79SIlya Dryomov if (err) 1557d286de79SIlya Dryomov goto bad; 1558d286de79SIlya Dryomov } 1559d286de79SIlya Dryomov 156063a6993fSIlya Dryomov /* new_primary_affinity */ 156163a6993fSIlya Dryomov if (struct_v >= 2) { 156263a6993fSIlya Dryomov err = decode_new_primary_affinity(p, end, map); 156363a6993fSIlya Dryomov if (err) 156463a6993fSIlya Dryomov goto bad; 156563a6993fSIlya Dryomov } 156663a6993fSIlya Dryomov 15673d14c5d2SYehuda Sadeh /* ignore the rest */ 15683d14c5d2SYehuda Sadeh *p = end; 156938a8d560SIlya Dryomov 157038a8d560SIlya Dryomov dout("inc osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd); 15713d14c5d2SYehuda Sadeh return map; 15723d14c5d2SYehuda Sadeh 157386f1742bSIlya Dryomov e_inval: 157486f1742bSIlya Dryomov err = -EINVAL; 15753d14c5d2SYehuda Sadeh bad: 157638a8d560SIlya Dryomov pr_err("corrupt inc osdmap (%d) epoch %d off %d (%p of %p-%p)\n", 157738a8d560SIlya Dryomov err, epoch, (int)(*p - start), *p, start, end); 15783d14c5d2SYehuda Sadeh print_hex_dump(KERN_DEBUG, "osdmap: ", 15793d14c5d2SYehuda Sadeh DUMP_PREFIX_OFFSET, 16, 1, 15803d14c5d2SYehuda Sadeh start, end - start, true); 15813d14c5d2SYehuda Sadeh return ERR_PTR(err); 15823d14c5d2SYehuda Sadeh } 15833d14c5d2SYehuda Sadeh 158430c156d9SYan, Zheng void ceph_oloc_copy(struct ceph_object_locator *dest, 158530c156d9SYan, Zheng const struct ceph_object_locator *src) 158630c156d9SYan, Zheng { 1587ca35ffeaSIlya Dryomov ceph_oloc_destroy(dest); 158830c156d9SYan, Zheng 158930c156d9SYan, Zheng dest->pool = src->pool; 159030c156d9SYan, Zheng if (src->pool_ns) 159130c156d9SYan, Zheng dest->pool_ns = ceph_get_string(src->pool_ns); 1592ca35ffeaSIlya Dryomov else 1593ca35ffeaSIlya Dryomov dest->pool_ns = NULL; 159430c156d9SYan, Zheng } 159530c156d9SYan, Zheng EXPORT_SYMBOL(ceph_oloc_copy); 159630c156d9SYan, Zheng 159730c156d9SYan, Zheng void ceph_oloc_destroy(struct ceph_object_locator *oloc) 159830c156d9SYan, Zheng { 159930c156d9SYan, Zheng ceph_put_string(oloc->pool_ns); 160030c156d9SYan, Zheng } 160130c156d9SYan, Zheng EXPORT_SYMBOL(ceph_oloc_destroy); 160230c156d9SYan, Zheng 1603d30291b9SIlya Dryomov void ceph_oid_copy(struct ceph_object_id *dest, 1604d30291b9SIlya Dryomov const struct ceph_object_id *src) 1605d30291b9SIlya Dryomov { 1606ca35ffeaSIlya Dryomov ceph_oid_destroy(dest); 16073d14c5d2SYehuda Sadeh 1608d30291b9SIlya Dryomov if (src->name != src->inline_name) { 1609d30291b9SIlya Dryomov /* very rare, see ceph_object_id definition */ 1610d30291b9SIlya Dryomov dest->name = kmalloc(src->name_len + 1, 1611d30291b9SIlya Dryomov GFP_NOIO | __GFP_NOFAIL); 1612ca35ffeaSIlya Dryomov } else { 1613ca35ffeaSIlya Dryomov dest->name = dest->inline_name; 1614d30291b9SIlya Dryomov } 1615d30291b9SIlya Dryomov memcpy(dest->name, src->name, src->name_len + 1); 1616d30291b9SIlya Dryomov dest->name_len = src->name_len; 1617d30291b9SIlya Dryomov } 1618d30291b9SIlya Dryomov EXPORT_SYMBOL(ceph_oid_copy); 1619d30291b9SIlya Dryomov 1620d30291b9SIlya Dryomov static __printf(2, 0) 1621d30291b9SIlya Dryomov int oid_printf_vargs(struct ceph_object_id *oid, const char *fmt, va_list ap) 1622d30291b9SIlya Dryomov { 1623d30291b9SIlya Dryomov int len; 1624d30291b9SIlya Dryomov 1625d30291b9SIlya Dryomov WARN_ON(!ceph_oid_empty(oid)); 1626d30291b9SIlya Dryomov 1627d30291b9SIlya Dryomov len = vsnprintf(oid->inline_name, sizeof(oid->inline_name), fmt, ap); 1628d30291b9SIlya Dryomov if (len >= sizeof(oid->inline_name)) 1629d30291b9SIlya Dryomov return len; 1630d30291b9SIlya Dryomov 1631d30291b9SIlya Dryomov oid->name_len = len; 1632d30291b9SIlya Dryomov return 0; 1633d30291b9SIlya Dryomov } 1634d30291b9SIlya Dryomov 1635d30291b9SIlya Dryomov /* 1636d30291b9SIlya Dryomov * If oid doesn't fit into inline buffer, BUG. 1637d30291b9SIlya Dryomov */ 1638d30291b9SIlya Dryomov void ceph_oid_printf(struct ceph_object_id *oid, const char *fmt, ...) 1639d30291b9SIlya Dryomov { 1640d30291b9SIlya Dryomov va_list ap; 1641d30291b9SIlya Dryomov 1642d30291b9SIlya Dryomov va_start(ap, fmt); 1643d30291b9SIlya Dryomov BUG_ON(oid_printf_vargs(oid, fmt, ap)); 1644d30291b9SIlya Dryomov va_end(ap); 1645d30291b9SIlya Dryomov } 1646d30291b9SIlya Dryomov EXPORT_SYMBOL(ceph_oid_printf); 1647d30291b9SIlya Dryomov 1648d30291b9SIlya Dryomov static __printf(3, 0) 1649d30291b9SIlya Dryomov int oid_aprintf_vargs(struct ceph_object_id *oid, gfp_t gfp, 1650d30291b9SIlya Dryomov const char *fmt, va_list ap) 1651d30291b9SIlya Dryomov { 1652d30291b9SIlya Dryomov va_list aq; 1653d30291b9SIlya Dryomov int len; 1654d30291b9SIlya Dryomov 1655d30291b9SIlya Dryomov va_copy(aq, ap); 1656d30291b9SIlya Dryomov len = oid_printf_vargs(oid, fmt, aq); 1657d30291b9SIlya Dryomov va_end(aq); 1658d30291b9SIlya Dryomov 1659d30291b9SIlya Dryomov if (len) { 1660d30291b9SIlya Dryomov char *external_name; 1661d30291b9SIlya Dryomov 1662d30291b9SIlya Dryomov external_name = kmalloc(len + 1, gfp); 1663d30291b9SIlya Dryomov if (!external_name) 1664d30291b9SIlya Dryomov return -ENOMEM; 1665d30291b9SIlya Dryomov 1666d30291b9SIlya Dryomov oid->name = external_name; 1667d30291b9SIlya Dryomov WARN_ON(vsnprintf(oid->name, len + 1, fmt, ap) != len); 1668d30291b9SIlya Dryomov oid->name_len = len; 1669d30291b9SIlya Dryomov } 1670d30291b9SIlya Dryomov 1671d30291b9SIlya Dryomov return 0; 1672d30291b9SIlya Dryomov } 1673d30291b9SIlya Dryomov 1674d30291b9SIlya Dryomov /* 1675d30291b9SIlya Dryomov * If oid doesn't fit into inline buffer, allocate. 1676d30291b9SIlya Dryomov */ 1677d30291b9SIlya Dryomov int ceph_oid_aprintf(struct ceph_object_id *oid, gfp_t gfp, 1678d30291b9SIlya Dryomov const char *fmt, ...) 1679d30291b9SIlya Dryomov { 1680d30291b9SIlya Dryomov va_list ap; 1681d30291b9SIlya Dryomov int ret; 1682d30291b9SIlya Dryomov 1683d30291b9SIlya Dryomov va_start(ap, fmt); 1684d30291b9SIlya Dryomov ret = oid_aprintf_vargs(oid, gfp, fmt, ap); 1685d30291b9SIlya Dryomov va_end(ap); 1686d30291b9SIlya Dryomov 1687d30291b9SIlya Dryomov return ret; 1688d30291b9SIlya Dryomov } 1689d30291b9SIlya Dryomov EXPORT_SYMBOL(ceph_oid_aprintf); 1690d30291b9SIlya Dryomov 1691d30291b9SIlya Dryomov void ceph_oid_destroy(struct ceph_object_id *oid) 1692d30291b9SIlya Dryomov { 1693d30291b9SIlya Dryomov if (oid->name != oid->inline_name) 1694d30291b9SIlya Dryomov kfree(oid->name); 1695d30291b9SIlya Dryomov } 1696d30291b9SIlya Dryomov EXPORT_SYMBOL(ceph_oid_destroy); 16973d14c5d2SYehuda Sadeh 169863244fa1SIlya Dryomov /* 169963244fa1SIlya Dryomov * osds only 170063244fa1SIlya Dryomov */ 170163244fa1SIlya Dryomov static bool __osds_equal(const struct ceph_osds *lhs, 170263244fa1SIlya Dryomov const struct ceph_osds *rhs) 170363244fa1SIlya Dryomov { 170463244fa1SIlya Dryomov if (lhs->size == rhs->size && 170563244fa1SIlya Dryomov !memcmp(lhs->osds, rhs->osds, rhs->size * sizeof(rhs->osds[0]))) 170663244fa1SIlya Dryomov return true; 170763244fa1SIlya Dryomov 170863244fa1SIlya Dryomov return false; 170963244fa1SIlya Dryomov } 171063244fa1SIlya Dryomov 171163244fa1SIlya Dryomov /* 171263244fa1SIlya Dryomov * osds + primary 171363244fa1SIlya Dryomov */ 171463244fa1SIlya Dryomov static bool osds_equal(const struct ceph_osds *lhs, 171563244fa1SIlya Dryomov const struct ceph_osds *rhs) 171663244fa1SIlya Dryomov { 171763244fa1SIlya Dryomov if (__osds_equal(lhs, rhs) && 171863244fa1SIlya Dryomov lhs->primary == rhs->primary) 171963244fa1SIlya Dryomov return true; 172063244fa1SIlya Dryomov 172163244fa1SIlya Dryomov return false; 172263244fa1SIlya Dryomov } 172363244fa1SIlya Dryomov 17246f3bfd45SIlya Dryomov static bool osds_valid(const struct ceph_osds *set) 17256f3bfd45SIlya Dryomov { 17266f3bfd45SIlya Dryomov /* non-empty set */ 17276f3bfd45SIlya Dryomov if (set->size > 0 && set->primary >= 0) 17286f3bfd45SIlya Dryomov return true; 17296f3bfd45SIlya Dryomov 17306f3bfd45SIlya Dryomov /* empty can_shift_osds set */ 17316f3bfd45SIlya Dryomov if (!set->size && set->primary == -1) 17326f3bfd45SIlya Dryomov return true; 17336f3bfd45SIlya Dryomov 17346f3bfd45SIlya Dryomov /* empty !can_shift_osds set - all NONE */ 17356f3bfd45SIlya Dryomov if (set->size > 0 && set->primary == -1) { 17366f3bfd45SIlya Dryomov int i; 17376f3bfd45SIlya Dryomov 17386f3bfd45SIlya Dryomov for (i = 0; i < set->size; i++) { 17396f3bfd45SIlya Dryomov if (set->osds[i] != CRUSH_ITEM_NONE) 17406f3bfd45SIlya Dryomov break; 17416f3bfd45SIlya Dryomov } 17426f3bfd45SIlya Dryomov if (i == set->size) 17436f3bfd45SIlya Dryomov return true; 17446f3bfd45SIlya Dryomov } 17456f3bfd45SIlya Dryomov 17466f3bfd45SIlya Dryomov return false; 17476f3bfd45SIlya Dryomov } 17486f3bfd45SIlya Dryomov 17496f3bfd45SIlya Dryomov void ceph_osds_copy(struct ceph_osds *dest, const struct ceph_osds *src) 17506f3bfd45SIlya Dryomov { 17516f3bfd45SIlya Dryomov memcpy(dest->osds, src->osds, src->size * sizeof(src->osds[0])); 17526f3bfd45SIlya Dryomov dest->size = src->size; 17536f3bfd45SIlya Dryomov dest->primary = src->primary; 17546f3bfd45SIlya Dryomov } 17556f3bfd45SIlya Dryomov 175663244fa1SIlya Dryomov static bool is_split(const struct ceph_pg *pgid, 175763244fa1SIlya Dryomov u32 old_pg_num, 175863244fa1SIlya Dryomov u32 new_pg_num) 175963244fa1SIlya Dryomov { 176063244fa1SIlya Dryomov int old_bits = calc_bits_of(old_pg_num); 176163244fa1SIlya Dryomov int old_mask = (1 << old_bits) - 1; 176263244fa1SIlya Dryomov int n; 176363244fa1SIlya Dryomov 176463244fa1SIlya Dryomov WARN_ON(pgid->seed >= old_pg_num); 176563244fa1SIlya Dryomov if (new_pg_num <= old_pg_num) 176663244fa1SIlya Dryomov return false; 176763244fa1SIlya Dryomov 176863244fa1SIlya Dryomov for (n = 1; ; n++) { 176963244fa1SIlya Dryomov int next_bit = n << (old_bits - 1); 177063244fa1SIlya Dryomov u32 s = next_bit | pgid->seed; 177163244fa1SIlya Dryomov 177263244fa1SIlya Dryomov if (s < old_pg_num || s == pgid->seed) 177363244fa1SIlya Dryomov continue; 177463244fa1SIlya Dryomov if (s >= new_pg_num) 177563244fa1SIlya Dryomov break; 177663244fa1SIlya Dryomov 177763244fa1SIlya Dryomov s = ceph_stable_mod(s, old_pg_num, old_mask); 177863244fa1SIlya Dryomov if (s == pgid->seed) 177963244fa1SIlya Dryomov return true; 178063244fa1SIlya Dryomov } 178163244fa1SIlya Dryomov 178263244fa1SIlya Dryomov return false; 178363244fa1SIlya Dryomov } 178463244fa1SIlya Dryomov 178563244fa1SIlya Dryomov bool ceph_is_new_interval(const struct ceph_osds *old_acting, 178663244fa1SIlya Dryomov const struct ceph_osds *new_acting, 178763244fa1SIlya Dryomov const struct ceph_osds *old_up, 178863244fa1SIlya Dryomov const struct ceph_osds *new_up, 178963244fa1SIlya Dryomov int old_size, 179063244fa1SIlya Dryomov int new_size, 179163244fa1SIlya Dryomov int old_min_size, 179263244fa1SIlya Dryomov int new_min_size, 179363244fa1SIlya Dryomov u32 old_pg_num, 179463244fa1SIlya Dryomov u32 new_pg_num, 179563244fa1SIlya Dryomov bool old_sort_bitwise, 179663244fa1SIlya Dryomov bool new_sort_bitwise, 179763244fa1SIlya Dryomov const struct ceph_pg *pgid) 179863244fa1SIlya Dryomov { 179963244fa1SIlya Dryomov return !osds_equal(old_acting, new_acting) || 180063244fa1SIlya Dryomov !osds_equal(old_up, new_up) || 180163244fa1SIlya Dryomov old_size != new_size || 180263244fa1SIlya Dryomov old_min_size != new_min_size || 180363244fa1SIlya Dryomov is_split(pgid, old_pg_num, new_pg_num) || 180463244fa1SIlya Dryomov old_sort_bitwise != new_sort_bitwise; 180563244fa1SIlya Dryomov } 180663244fa1SIlya Dryomov 180763244fa1SIlya Dryomov static int calc_pg_rank(int osd, const struct ceph_osds *acting) 180863244fa1SIlya Dryomov { 180963244fa1SIlya Dryomov int i; 181063244fa1SIlya Dryomov 181163244fa1SIlya Dryomov for (i = 0; i < acting->size; i++) { 181263244fa1SIlya Dryomov if (acting->osds[i] == osd) 181363244fa1SIlya Dryomov return i; 181463244fa1SIlya Dryomov } 181563244fa1SIlya Dryomov 181663244fa1SIlya Dryomov return -1; 181763244fa1SIlya Dryomov } 181863244fa1SIlya Dryomov 181963244fa1SIlya Dryomov static bool primary_changed(const struct ceph_osds *old_acting, 182063244fa1SIlya Dryomov const struct ceph_osds *new_acting) 182163244fa1SIlya Dryomov { 182263244fa1SIlya Dryomov if (!old_acting->size && !new_acting->size) 182363244fa1SIlya Dryomov return false; /* both still empty */ 182463244fa1SIlya Dryomov 182563244fa1SIlya Dryomov if (!old_acting->size ^ !new_acting->size) 182663244fa1SIlya Dryomov return true; /* was empty, now not, or vice versa */ 182763244fa1SIlya Dryomov 182863244fa1SIlya Dryomov if (old_acting->primary != new_acting->primary) 182963244fa1SIlya Dryomov return true; /* primary changed */ 183063244fa1SIlya Dryomov 183163244fa1SIlya Dryomov if (calc_pg_rank(old_acting->primary, old_acting) != 183263244fa1SIlya Dryomov calc_pg_rank(new_acting->primary, new_acting)) 183363244fa1SIlya Dryomov return true; 183463244fa1SIlya Dryomov 183563244fa1SIlya Dryomov return false; /* same primary (tho replicas may have changed) */ 183663244fa1SIlya Dryomov } 183763244fa1SIlya Dryomov 183863244fa1SIlya Dryomov bool ceph_osds_changed(const struct ceph_osds *old_acting, 183963244fa1SIlya Dryomov const struct ceph_osds *new_acting, 184063244fa1SIlya Dryomov bool any_change) 184163244fa1SIlya Dryomov { 184263244fa1SIlya Dryomov if (primary_changed(old_acting, new_acting)) 184363244fa1SIlya Dryomov return true; 184463244fa1SIlya Dryomov 184563244fa1SIlya Dryomov if (any_change && !__osds_equal(old_acting, new_acting)) 184663244fa1SIlya Dryomov return true; 184763244fa1SIlya Dryomov 184863244fa1SIlya Dryomov return false; 184963244fa1SIlya Dryomov } 185063244fa1SIlya Dryomov 18513d14c5d2SYehuda Sadeh /* 18523d14c5d2SYehuda Sadeh * calculate file layout from given offset, length. 18533d14c5d2SYehuda Sadeh * fill in correct oid, logical length, and object extent 18543d14c5d2SYehuda Sadeh * offset, length. 18553d14c5d2SYehuda Sadeh * 18563d14c5d2SYehuda Sadeh * for now, we write only a single su, until we can 18573d14c5d2SYehuda Sadeh * pass a stride back to the caller. 18583d14c5d2SYehuda Sadeh */ 1859d63b77f4SSage Weil int ceph_calc_file_object_mapping(struct ceph_file_layout *layout, 1860e8afad65SAlex Elder u64 off, u64 len, 18613d14c5d2SYehuda Sadeh u64 *ono, 18623d14c5d2SYehuda Sadeh u64 *oxoff, u64 *oxlen) 18633d14c5d2SYehuda Sadeh { 18647627151eSYan, Zheng u32 osize = layout->object_size; 18657627151eSYan, Zheng u32 su = layout->stripe_unit; 18667627151eSYan, Zheng u32 sc = layout->stripe_count; 18673d14c5d2SYehuda Sadeh u32 bl, stripeno, stripepos, objsetno; 18683d14c5d2SYehuda Sadeh u32 su_per_object; 18693d14c5d2SYehuda Sadeh u64 t, su_offset; 18703d14c5d2SYehuda Sadeh 1871e8afad65SAlex Elder dout("mapping %llu~%llu osize %u fl_su %u\n", off, len, 18723d14c5d2SYehuda Sadeh osize, su); 1873d63b77f4SSage Weil if (su == 0 || sc == 0) 1874d63b77f4SSage Weil goto invalid; 18753d14c5d2SYehuda Sadeh su_per_object = osize / su; 1876d63b77f4SSage Weil if (su_per_object == 0) 1877d63b77f4SSage Weil goto invalid; 18783d14c5d2SYehuda Sadeh dout("osize %u / su %u = su_per_object %u\n", osize, su, 18793d14c5d2SYehuda Sadeh su_per_object); 18803d14c5d2SYehuda Sadeh 1881d63b77f4SSage Weil if ((su & ~PAGE_MASK) != 0) 1882d63b77f4SSage Weil goto invalid; 1883d63b77f4SSage Weil 18843d14c5d2SYehuda Sadeh /* bl = *off / su; */ 18853d14c5d2SYehuda Sadeh t = off; 18863d14c5d2SYehuda Sadeh do_div(t, su); 18873d14c5d2SYehuda Sadeh bl = t; 18883d14c5d2SYehuda Sadeh dout("off %llu / su %u = bl %u\n", off, su, bl); 18893d14c5d2SYehuda Sadeh 18903d14c5d2SYehuda Sadeh stripeno = bl / sc; 18913d14c5d2SYehuda Sadeh stripepos = bl % sc; 18923d14c5d2SYehuda Sadeh objsetno = stripeno / su_per_object; 18933d14c5d2SYehuda Sadeh 18943d14c5d2SYehuda Sadeh *ono = objsetno * sc + stripepos; 189595c96174SEric Dumazet dout("objset %u * sc %u = ono %u\n", objsetno, sc, (unsigned int)*ono); 18963d14c5d2SYehuda Sadeh 18973d14c5d2SYehuda Sadeh /* *oxoff = *off % layout->fl_stripe_unit; # offset in su */ 18983d14c5d2SYehuda Sadeh t = off; 18993d14c5d2SYehuda Sadeh su_offset = do_div(t, su); 19003d14c5d2SYehuda Sadeh *oxoff = su_offset + (stripeno % su_per_object) * su; 19013d14c5d2SYehuda Sadeh 19023d14c5d2SYehuda Sadeh /* 19033d14c5d2SYehuda Sadeh * Calculate the length of the extent being written to the selected 1904e8afad65SAlex Elder * object. This is the minimum of the full length requested (len) or 19053d14c5d2SYehuda Sadeh * the remainder of the current stripe being written to. 19063d14c5d2SYehuda Sadeh */ 1907e8afad65SAlex Elder *oxlen = min_t(u64, len, su - su_offset); 19083d14c5d2SYehuda Sadeh 19093d14c5d2SYehuda Sadeh dout(" obj extent %llu~%llu\n", *oxoff, *oxlen); 1910d63b77f4SSage Weil return 0; 1911d63b77f4SSage Weil 1912d63b77f4SSage Weil invalid: 1913d63b77f4SSage Weil dout(" invalid layout\n"); 1914d63b77f4SSage Weil *ono = 0; 1915d63b77f4SSage Weil *oxoff = 0; 1916d63b77f4SSage Weil *oxlen = 0; 1917d63b77f4SSage Weil return -EINVAL; 19183d14c5d2SYehuda Sadeh } 19193d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_calc_file_object_mapping); 19203d14c5d2SYehuda Sadeh 19213d14c5d2SYehuda Sadeh /* 1922d9591f5eSIlya Dryomov * Map an object into a PG. 1923d9591f5eSIlya Dryomov * 1924d9591f5eSIlya Dryomov * Should only be called with target_oid and target_oloc (as opposed to 1925d9591f5eSIlya Dryomov * base_oid and base_oloc), since tiering isn't taken into account. 19263d14c5d2SYehuda Sadeh */ 1927d9591f5eSIlya Dryomov int ceph_object_locator_to_pg(struct ceph_osdmap *osdmap, 19287c13cb64SIlya Dryomov struct ceph_object_id *oid, 1929d9591f5eSIlya Dryomov struct ceph_object_locator *oloc, 1930d9591f5eSIlya Dryomov struct ceph_pg *raw_pgid) 19313d14c5d2SYehuda Sadeh { 19327c13cb64SIlya Dryomov struct ceph_pg_pool_info *pi; 19333d14c5d2SYehuda Sadeh 1934d9591f5eSIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, oloc->pool); 19357c13cb64SIlya Dryomov if (!pi) 1936d9591f5eSIlya Dryomov return -ENOENT; 19373d14c5d2SYehuda Sadeh 193830c156d9SYan, Zheng if (!oloc->pool_ns) { 1939d9591f5eSIlya Dryomov raw_pgid->pool = oloc->pool; 1940d9591f5eSIlya Dryomov raw_pgid->seed = ceph_str_hash(pi->object_hash, oid->name, 19417c13cb64SIlya Dryomov oid->name_len); 19424a3262b1SIlya Dryomov dout("%s %s -> raw_pgid %llu.%x\n", __func__, oid->name, 19434a3262b1SIlya Dryomov raw_pgid->pool, raw_pgid->seed); 194430c156d9SYan, Zheng } else { 194530c156d9SYan, Zheng char stack_buf[256]; 194630c156d9SYan, Zheng char *buf = stack_buf; 194730c156d9SYan, Zheng int nsl = oloc->pool_ns->len; 194830c156d9SYan, Zheng size_t total = nsl + 1 + oid->name_len; 194930c156d9SYan, Zheng 195030c156d9SYan, Zheng if (total > sizeof(stack_buf)) { 195130c156d9SYan, Zheng buf = kmalloc(total, GFP_NOIO); 195230c156d9SYan, Zheng if (!buf) 195330c156d9SYan, Zheng return -ENOMEM; 195430c156d9SYan, Zheng } 195530c156d9SYan, Zheng memcpy(buf, oloc->pool_ns->str, nsl); 195630c156d9SYan, Zheng buf[nsl] = '\037'; 195730c156d9SYan, Zheng memcpy(buf + nsl + 1, oid->name, oid->name_len); 195830c156d9SYan, Zheng raw_pgid->pool = oloc->pool; 195930c156d9SYan, Zheng raw_pgid->seed = ceph_str_hash(pi->object_hash, buf, total); 196030c156d9SYan, Zheng if (buf != stack_buf) 196130c156d9SYan, Zheng kfree(buf); 196230c156d9SYan, Zheng dout("%s %s ns %.*s -> raw_pgid %llu.%x\n", __func__, 196330c156d9SYan, Zheng oid->name, nsl, oloc->pool_ns->str, 196430c156d9SYan, Zheng raw_pgid->pool, raw_pgid->seed); 196530c156d9SYan, Zheng } 19663d14c5d2SYehuda Sadeh return 0; 19673d14c5d2SYehuda Sadeh } 1968d9591f5eSIlya Dryomov EXPORT_SYMBOL(ceph_object_locator_to_pg); 19693d14c5d2SYehuda Sadeh 19706f3bfd45SIlya Dryomov /* 19716f3bfd45SIlya Dryomov * Map a raw PG (full precision ps) into an actual PG. 19726f3bfd45SIlya Dryomov */ 19736f3bfd45SIlya Dryomov static void raw_pg_to_pg(struct ceph_pg_pool_info *pi, 19746f3bfd45SIlya Dryomov const struct ceph_pg *raw_pgid, 19756f3bfd45SIlya Dryomov struct ceph_pg *pgid) 19766f3bfd45SIlya Dryomov { 19776f3bfd45SIlya Dryomov pgid->pool = raw_pgid->pool; 19786f3bfd45SIlya Dryomov pgid->seed = ceph_stable_mod(raw_pgid->seed, pi->pg_num, 19796f3bfd45SIlya Dryomov pi->pg_num_mask); 19806f3bfd45SIlya Dryomov } 19816f3bfd45SIlya Dryomov 19826f3bfd45SIlya Dryomov /* 19836f3bfd45SIlya Dryomov * Map a raw PG (full precision ps) into a placement ps (placement 19846f3bfd45SIlya Dryomov * seed). Include pool id in that value so that different pools don't 19856f3bfd45SIlya Dryomov * use the same seeds. 19866f3bfd45SIlya Dryomov */ 19876f3bfd45SIlya Dryomov static u32 raw_pg_to_pps(struct ceph_pg_pool_info *pi, 19886f3bfd45SIlya Dryomov const struct ceph_pg *raw_pgid) 19896f3bfd45SIlya Dryomov { 19906f3bfd45SIlya Dryomov if (pi->flags & CEPH_POOL_FLAG_HASHPSPOOL) { 19916f3bfd45SIlya Dryomov /* hash pool id and seed so that pool PGs do not overlap */ 19926f3bfd45SIlya Dryomov return crush_hash32_2(CRUSH_HASH_RJENKINS1, 19936f3bfd45SIlya Dryomov ceph_stable_mod(raw_pgid->seed, 19946f3bfd45SIlya Dryomov pi->pgp_num, 19956f3bfd45SIlya Dryomov pi->pgp_num_mask), 19966f3bfd45SIlya Dryomov raw_pgid->pool); 19976f3bfd45SIlya Dryomov } else { 19986f3bfd45SIlya Dryomov /* 19996f3bfd45SIlya Dryomov * legacy behavior: add ps and pool together. this is 20006f3bfd45SIlya Dryomov * not a great approach because the PGs from each pool 20016f3bfd45SIlya Dryomov * will overlap on top of each other: 0.5 == 1.4 == 20026f3bfd45SIlya Dryomov * 2.3 == ... 20036f3bfd45SIlya Dryomov */ 20046f3bfd45SIlya Dryomov return ceph_stable_mod(raw_pgid->seed, pi->pgp_num, 20056f3bfd45SIlya Dryomov pi->pgp_num_mask) + 20066f3bfd45SIlya Dryomov (unsigned)raw_pgid->pool; 20076f3bfd45SIlya Dryomov } 20086f3bfd45SIlya Dryomov } 20096f3bfd45SIlya Dryomov 20109d521470SIlya Dryomov static int do_crush(struct ceph_osdmap *map, int ruleno, int x, 2011e8ef19c4SIlya Dryomov int *result, int result_max, 2012e8ef19c4SIlya Dryomov const __u32 *weight, int weight_max) 2013e8ef19c4SIlya Dryomov { 20149d521470SIlya Dryomov int r; 2015e8ef19c4SIlya Dryomov 20169d521470SIlya Dryomov BUG_ON(result_max > CEPH_PG_MAX_SIZE); 20179d521470SIlya Dryomov 2018743efcffSIlya Dryomov mutex_lock(&map->crush_workspace_mutex); 20199d521470SIlya Dryomov r = crush_do_rule(map->crush, ruleno, x, result, result_max, 2020743efcffSIlya Dryomov weight, weight_max, map->crush_workspace); 2021743efcffSIlya Dryomov mutex_unlock(&map->crush_workspace_mutex); 20229d521470SIlya Dryomov 20239d521470SIlya Dryomov return r; 2024e8ef19c4SIlya Dryomov } 2025e8ef19c4SIlya Dryomov 20263d14c5d2SYehuda Sadeh /* 20276f3bfd45SIlya Dryomov * Calculate raw set (CRUSH output) for given PG. The result may 20286f3bfd45SIlya Dryomov * contain nonexistent OSDs. ->primary is undefined for a raw set. 20292bd93d4dSIlya Dryomov * 20306f3bfd45SIlya Dryomov * Placement seed (CRUSH input) is returned through @ppps. 20312bd93d4dSIlya Dryomov */ 20326f3bfd45SIlya Dryomov static void pg_to_raw_osds(struct ceph_osdmap *osdmap, 20336f3bfd45SIlya Dryomov struct ceph_pg_pool_info *pi, 20346f3bfd45SIlya Dryomov const struct ceph_pg *raw_pgid, 20356f3bfd45SIlya Dryomov struct ceph_osds *raw, 20366f3bfd45SIlya Dryomov u32 *ppps) 20372bd93d4dSIlya Dryomov { 20386f3bfd45SIlya Dryomov u32 pps = raw_pg_to_pps(pi, raw_pgid); 20392bd93d4dSIlya Dryomov int ruleno; 20402bd93d4dSIlya Dryomov int len; 20412bd93d4dSIlya Dryomov 20426f3bfd45SIlya Dryomov ceph_osds_init(raw); 20436f3bfd45SIlya Dryomov if (ppps) 20446f3bfd45SIlya Dryomov *ppps = pps; 20456f3bfd45SIlya Dryomov 20466f3bfd45SIlya Dryomov ruleno = crush_find_rule(osdmap->crush, pi->crush_ruleset, pi->type, 20476f3bfd45SIlya Dryomov pi->size); 20482bd93d4dSIlya Dryomov if (ruleno < 0) { 20492bd93d4dSIlya Dryomov pr_err("no crush rule: pool %lld ruleset %d type %d size %d\n", 20506f3bfd45SIlya Dryomov pi->id, pi->crush_ruleset, pi->type, pi->size); 20516f3bfd45SIlya Dryomov return; 20522bd93d4dSIlya Dryomov } 20532bd93d4dSIlya Dryomov 2054ef9324bbSIlya Dryomov if (pi->size > ARRAY_SIZE(raw->osds)) { 2055ef9324bbSIlya Dryomov pr_err_ratelimited("pool %lld ruleset %d type %d too wide: size %d > %zu\n", 2056ef9324bbSIlya Dryomov pi->id, pi->crush_ruleset, pi->type, pi->size, 2057ef9324bbSIlya Dryomov ARRAY_SIZE(raw->osds)); 2058ef9324bbSIlya Dryomov return; 2059ef9324bbSIlya Dryomov } 2060ef9324bbSIlya Dryomov 2061ef9324bbSIlya Dryomov len = do_crush(osdmap, ruleno, pps, raw->osds, pi->size, 20622bd93d4dSIlya Dryomov osdmap->osd_weight, osdmap->max_osd); 20632bd93d4dSIlya Dryomov if (len < 0) { 20642bd93d4dSIlya Dryomov pr_err("error %d from crush rule %d: pool %lld ruleset %d type %d size %d\n", 20656f3bfd45SIlya Dryomov len, ruleno, pi->id, pi->crush_ruleset, pi->type, 20666f3bfd45SIlya Dryomov pi->size); 20676f3bfd45SIlya Dryomov return; 20682bd93d4dSIlya Dryomov } 20692bd93d4dSIlya Dryomov 20706f3bfd45SIlya Dryomov raw->size = len; 20712bd93d4dSIlya Dryomov } 20722bd93d4dSIlya Dryomov 20732bd93d4dSIlya Dryomov /* 20746f3bfd45SIlya Dryomov * Given raw set, calculate up set and up primary. By definition of an 20756f3bfd45SIlya Dryomov * up set, the result won't contain nonexistent or down OSDs. 20762bd93d4dSIlya Dryomov * 20776f3bfd45SIlya Dryomov * This is done in-place - on return @set is the up set. If it's 20786f3bfd45SIlya Dryomov * empty, ->primary will remain undefined. 20792bd93d4dSIlya Dryomov */ 20806f3bfd45SIlya Dryomov static void raw_to_up_osds(struct ceph_osdmap *osdmap, 20816f3bfd45SIlya Dryomov struct ceph_pg_pool_info *pi, 20826f3bfd45SIlya Dryomov struct ceph_osds *set) 20832bd93d4dSIlya Dryomov { 20842bd93d4dSIlya Dryomov int i; 20852bd93d4dSIlya Dryomov 20866f3bfd45SIlya Dryomov /* ->primary is undefined for a raw set */ 20876f3bfd45SIlya Dryomov BUG_ON(set->primary != -1); 20886f3bfd45SIlya Dryomov 20896f3bfd45SIlya Dryomov if (ceph_can_shift_osds(pi)) { 20902bd93d4dSIlya Dryomov int removed = 0; 20912bd93d4dSIlya Dryomov 20926f3bfd45SIlya Dryomov /* shift left */ 20936f3bfd45SIlya Dryomov for (i = 0; i < set->size; i++) { 20946f3bfd45SIlya Dryomov if (ceph_osd_is_down(osdmap, set->osds[i])) { 20952bd93d4dSIlya Dryomov removed++; 20962bd93d4dSIlya Dryomov continue; 20972bd93d4dSIlya Dryomov } 20982bd93d4dSIlya Dryomov if (removed) 20996f3bfd45SIlya Dryomov set->osds[i - removed] = set->osds[i]; 21002bd93d4dSIlya Dryomov } 21016f3bfd45SIlya Dryomov set->size -= removed; 21026f3bfd45SIlya Dryomov if (set->size > 0) 21036f3bfd45SIlya Dryomov set->primary = set->osds[0]; 21042bd93d4dSIlya Dryomov } else { 21056f3bfd45SIlya Dryomov /* set down/dne devices to NONE */ 21066f3bfd45SIlya Dryomov for (i = set->size - 1; i >= 0; i--) { 21076f3bfd45SIlya Dryomov if (ceph_osd_is_down(osdmap, set->osds[i])) 21086f3bfd45SIlya Dryomov set->osds[i] = CRUSH_ITEM_NONE; 21092bd93d4dSIlya Dryomov else 21106f3bfd45SIlya Dryomov set->primary = set->osds[i]; 21116f3bfd45SIlya Dryomov } 21122bd93d4dSIlya Dryomov } 21132bd93d4dSIlya Dryomov } 21142bd93d4dSIlya Dryomov 21156f3bfd45SIlya Dryomov static void apply_primary_affinity(struct ceph_osdmap *osdmap, 21166f3bfd45SIlya Dryomov struct ceph_pg_pool_info *pi, 21176f3bfd45SIlya Dryomov u32 pps, 21186f3bfd45SIlya Dryomov struct ceph_osds *up) 211947ec1f3cSIlya Dryomov { 212047ec1f3cSIlya Dryomov int i; 212147ec1f3cSIlya Dryomov int pos = -1; 212247ec1f3cSIlya Dryomov 212347ec1f3cSIlya Dryomov /* 212447ec1f3cSIlya Dryomov * Do we have any non-default primary_affinity values for these 212547ec1f3cSIlya Dryomov * osds? 212647ec1f3cSIlya Dryomov */ 212747ec1f3cSIlya Dryomov if (!osdmap->osd_primary_affinity) 212847ec1f3cSIlya Dryomov return; 212947ec1f3cSIlya Dryomov 21306f3bfd45SIlya Dryomov for (i = 0; i < up->size; i++) { 21316f3bfd45SIlya Dryomov int osd = up->osds[i]; 213292b2e751SIlya Dryomov 213392b2e751SIlya Dryomov if (osd != CRUSH_ITEM_NONE && 213492b2e751SIlya Dryomov osdmap->osd_primary_affinity[osd] != 213547ec1f3cSIlya Dryomov CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) { 213647ec1f3cSIlya Dryomov break; 213747ec1f3cSIlya Dryomov } 213847ec1f3cSIlya Dryomov } 21396f3bfd45SIlya Dryomov if (i == up->size) 214047ec1f3cSIlya Dryomov return; 214147ec1f3cSIlya Dryomov 214247ec1f3cSIlya Dryomov /* 214347ec1f3cSIlya Dryomov * Pick the primary. Feed both the seed (for the pg) and the 214447ec1f3cSIlya Dryomov * osd into the hash/rng so that a proportional fraction of an 214547ec1f3cSIlya Dryomov * osd's pgs get rejected as primary. 214647ec1f3cSIlya Dryomov */ 21476f3bfd45SIlya Dryomov for (i = 0; i < up->size; i++) { 21486f3bfd45SIlya Dryomov int osd = up->osds[i]; 214947ec1f3cSIlya Dryomov u32 aff; 215047ec1f3cSIlya Dryomov 215147ec1f3cSIlya Dryomov if (osd == CRUSH_ITEM_NONE) 215247ec1f3cSIlya Dryomov continue; 215347ec1f3cSIlya Dryomov 215447ec1f3cSIlya Dryomov aff = osdmap->osd_primary_affinity[osd]; 215547ec1f3cSIlya Dryomov if (aff < CEPH_OSD_MAX_PRIMARY_AFFINITY && 215647ec1f3cSIlya Dryomov (crush_hash32_2(CRUSH_HASH_RJENKINS1, 215747ec1f3cSIlya Dryomov pps, osd) >> 16) >= aff) { 215847ec1f3cSIlya Dryomov /* 215947ec1f3cSIlya Dryomov * We chose not to use this primary. Note it 216047ec1f3cSIlya Dryomov * anyway as a fallback in case we don't pick 216147ec1f3cSIlya Dryomov * anyone else, but keep looking. 216247ec1f3cSIlya Dryomov */ 216347ec1f3cSIlya Dryomov if (pos < 0) 216447ec1f3cSIlya Dryomov pos = i; 216547ec1f3cSIlya Dryomov } else { 216647ec1f3cSIlya Dryomov pos = i; 216747ec1f3cSIlya Dryomov break; 216847ec1f3cSIlya Dryomov } 216947ec1f3cSIlya Dryomov } 217047ec1f3cSIlya Dryomov if (pos < 0) 217147ec1f3cSIlya Dryomov return; 217247ec1f3cSIlya Dryomov 21736f3bfd45SIlya Dryomov up->primary = up->osds[pos]; 217447ec1f3cSIlya Dryomov 21756f3bfd45SIlya Dryomov if (ceph_can_shift_osds(pi) && pos > 0) { 217647ec1f3cSIlya Dryomov /* move the new primary to the front */ 217747ec1f3cSIlya Dryomov for (i = pos; i > 0; i--) 21786f3bfd45SIlya Dryomov up->osds[i] = up->osds[i - 1]; 21796f3bfd45SIlya Dryomov up->osds[0] = up->primary; 218047ec1f3cSIlya Dryomov } 218147ec1f3cSIlya Dryomov } 218247ec1f3cSIlya Dryomov 21832bd93d4dSIlya Dryomov /* 21846f3bfd45SIlya Dryomov * Get pg_temp and primary_temp mappings for given PG. 218545966c34SIlya Dryomov * 21866f3bfd45SIlya Dryomov * Note that a PG may have none, only pg_temp, only primary_temp or 21876f3bfd45SIlya Dryomov * both pg_temp and primary_temp mappings. This means @temp isn't 21886f3bfd45SIlya Dryomov * always a valid OSD set on return: in the "only primary_temp" case, 21896f3bfd45SIlya Dryomov * @temp will have its ->primary >= 0 but ->size == 0. 219045966c34SIlya Dryomov */ 21916f3bfd45SIlya Dryomov static void get_temp_osds(struct ceph_osdmap *osdmap, 21926f3bfd45SIlya Dryomov struct ceph_pg_pool_info *pi, 21936f3bfd45SIlya Dryomov const struct ceph_pg *raw_pgid, 21946f3bfd45SIlya Dryomov struct ceph_osds *temp) 219545966c34SIlya Dryomov { 21966f3bfd45SIlya Dryomov struct ceph_pg pgid; 219745966c34SIlya Dryomov struct ceph_pg_mapping *pg; 219845966c34SIlya Dryomov int i; 219945966c34SIlya Dryomov 22006f3bfd45SIlya Dryomov raw_pg_to_pg(pi, raw_pgid, &pgid); 22016f3bfd45SIlya Dryomov ceph_osds_init(temp); 220245966c34SIlya Dryomov 220345966c34SIlya Dryomov /* pg_temp? */ 220445966c34SIlya Dryomov pg = __lookup_pg_mapping(&osdmap->pg_temp, pgid); 220545966c34SIlya Dryomov if (pg) { 220645966c34SIlya Dryomov for (i = 0; i < pg->pg_temp.len; i++) { 220745966c34SIlya Dryomov if (ceph_osd_is_down(osdmap, pg->pg_temp.osds[i])) { 22086f3bfd45SIlya Dryomov if (ceph_can_shift_osds(pi)) 220945966c34SIlya Dryomov continue; 22106f3bfd45SIlya Dryomov 22116f3bfd45SIlya Dryomov temp->osds[temp->size++] = CRUSH_ITEM_NONE; 221245966c34SIlya Dryomov } else { 22136f3bfd45SIlya Dryomov temp->osds[temp->size++] = pg->pg_temp.osds[i]; 221445966c34SIlya Dryomov } 221545966c34SIlya Dryomov } 221645966c34SIlya Dryomov 221745966c34SIlya Dryomov /* apply pg_temp's primary */ 22186f3bfd45SIlya Dryomov for (i = 0; i < temp->size; i++) { 22196f3bfd45SIlya Dryomov if (temp->osds[i] != CRUSH_ITEM_NONE) { 22206f3bfd45SIlya Dryomov temp->primary = temp->osds[i]; 222145966c34SIlya Dryomov break; 222245966c34SIlya Dryomov } 222345966c34SIlya Dryomov } 222445966c34SIlya Dryomov } 222545966c34SIlya Dryomov 22265e8d4d36SIlya Dryomov /* primary_temp? */ 22275e8d4d36SIlya Dryomov pg = __lookup_pg_mapping(&osdmap->primary_temp, pgid); 22285e8d4d36SIlya Dryomov if (pg) 22296f3bfd45SIlya Dryomov temp->primary = pg->primary_temp.osd; 223045966c34SIlya Dryomov } 223145966c34SIlya Dryomov 223245966c34SIlya Dryomov /* 22336f3bfd45SIlya Dryomov * Map a PG to its acting set as well as its up set. 2234ac972230SIlya Dryomov * 22356f3bfd45SIlya Dryomov * Acting set is used for data mapping purposes, while up set can be 22366f3bfd45SIlya Dryomov * recorded for detecting interval changes and deciding whether to 22376f3bfd45SIlya Dryomov * resend a request. 22383d14c5d2SYehuda Sadeh */ 22396f3bfd45SIlya Dryomov void ceph_pg_to_up_acting_osds(struct ceph_osdmap *osdmap, 22406f3bfd45SIlya Dryomov const struct ceph_pg *raw_pgid, 22416f3bfd45SIlya Dryomov struct ceph_osds *up, 22426f3bfd45SIlya Dryomov struct ceph_osds *acting) 22433d14c5d2SYehuda Sadeh { 22446f3bfd45SIlya Dryomov struct ceph_pg_pool_info *pi; 2245ac972230SIlya Dryomov u32 pps; 22463d14c5d2SYehuda Sadeh 22476f3bfd45SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, raw_pgid->pool); 22486f3bfd45SIlya Dryomov if (!pi) { 22496f3bfd45SIlya Dryomov ceph_osds_init(up); 22506f3bfd45SIlya Dryomov ceph_osds_init(acting); 22516f3bfd45SIlya Dryomov goto out; 22528008ab10SIlya Dryomov } 22533d14c5d2SYehuda Sadeh 22546f3bfd45SIlya Dryomov pg_to_raw_osds(osdmap, pi, raw_pgid, up, &pps); 22556f3bfd45SIlya Dryomov raw_to_up_osds(osdmap, pi, up); 22566f3bfd45SIlya Dryomov apply_primary_affinity(osdmap, pi, pps, up); 22576f3bfd45SIlya Dryomov get_temp_osds(osdmap, pi, raw_pgid, acting); 22586f3bfd45SIlya Dryomov if (!acting->size) { 22596f3bfd45SIlya Dryomov memcpy(acting->osds, up->osds, up->size * sizeof(up->osds[0])); 22606f3bfd45SIlya Dryomov acting->size = up->size; 22616f3bfd45SIlya Dryomov if (acting->primary == -1) 22626f3bfd45SIlya Dryomov acting->primary = up->primary; 2263ac972230SIlya Dryomov } 22646f3bfd45SIlya Dryomov out: 22656f3bfd45SIlya Dryomov WARN_ON(!osds_valid(up) || !osds_valid(acting)); 22663d14c5d2SYehuda Sadeh } 22673d14c5d2SYehuda Sadeh 2268dc98ff72SIlya Dryomov bool ceph_pg_to_primary_shard(struct ceph_osdmap *osdmap, 2269dc98ff72SIlya Dryomov const struct ceph_pg *raw_pgid, 2270dc98ff72SIlya Dryomov struct ceph_spg *spgid) 2271dc98ff72SIlya Dryomov { 2272dc98ff72SIlya Dryomov struct ceph_pg_pool_info *pi; 2273dc98ff72SIlya Dryomov struct ceph_pg pgid; 2274dc98ff72SIlya Dryomov struct ceph_osds up, acting; 2275dc98ff72SIlya Dryomov int i; 2276dc98ff72SIlya Dryomov 2277dc98ff72SIlya Dryomov pi = ceph_pg_pool_by_id(osdmap, raw_pgid->pool); 2278dc98ff72SIlya Dryomov if (!pi) 2279dc98ff72SIlya Dryomov return false; 2280dc98ff72SIlya Dryomov 2281dc98ff72SIlya Dryomov raw_pg_to_pg(pi, raw_pgid, &pgid); 2282dc98ff72SIlya Dryomov 2283dc98ff72SIlya Dryomov if (ceph_can_shift_osds(pi)) { 2284dc98ff72SIlya Dryomov spgid->pgid = pgid; /* struct */ 2285dc98ff72SIlya Dryomov spgid->shard = CEPH_SPG_NOSHARD; 2286dc98ff72SIlya Dryomov return true; 2287dc98ff72SIlya Dryomov } 2288dc98ff72SIlya Dryomov 2289dc98ff72SIlya Dryomov ceph_pg_to_up_acting_osds(osdmap, &pgid, &up, &acting); 2290dc98ff72SIlya Dryomov for (i = 0; i < acting.size; i++) { 2291dc98ff72SIlya Dryomov if (acting.osds[i] == acting.primary) { 2292dc98ff72SIlya Dryomov spgid->pgid = pgid; /* struct */ 2293dc98ff72SIlya Dryomov spgid->shard = i; 2294dc98ff72SIlya Dryomov return true; 2295dc98ff72SIlya Dryomov } 2296dc98ff72SIlya Dryomov } 2297dc98ff72SIlya Dryomov 2298dc98ff72SIlya Dryomov return false; 2299dc98ff72SIlya Dryomov } 2300dc98ff72SIlya Dryomov 23013d14c5d2SYehuda Sadeh /* 2302f81f1633SIlya Dryomov * Return acting primary for given PG, or -1 if none. 23033d14c5d2SYehuda Sadeh */ 2304f81f1633SIlya Dryomov int ceph_pg_to_acting_primary(struct ceph_osdmap *osdmap, 2305f81f1633SIlya Dryomov const struct ceph_pg *raw_pgid) 23063d14c5d2SYehuda Sadeh { 23076f3bfd45SIlya Dryomov struct ceph_osds up, acting; 23083d14c5d2SYehuda Sadeh 2309f81f1633SIlya Dryomov ceph_pg_to_up_acting_osds(osdmap, raw_pgid, &up, &acting); 23106f3bfd45SIlya Dryomov return acting.primary; 23113d14c5d2SYehuda Sadeh } 2312f81f1633SIlya Dryomov EXPORT_SYMBOL(ceph_pg_to_acting_primary); 2313