net/ceph/osdmap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/ceph/ceph_debug.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/slab.h>
   7
   8 #include <linux/ceph/libceph.h>
   9 #include <linux/ceph/osdmap.h>
  10 #include <linux/ceph/decode.h>
  11 #include <linux/crush/hash.h>
  12 #include <linux/crush/mapper.h>
  13
  14 static __printf(2, 3)
  15 void osdmap_info(const struct ceph_osdmap *map, const char *fmt, ...)
  16 {
  17         struct va_format vaf;
  18         va_list args;
  19
  20         va_start(args, fmt);
  21         vaf.fmt = fmt;
  22         vaf.va = &args;
  23
  24         printk(KERN_INFO "%s (%pU e%u): %pV", KBUILD_MODNAME, &map->fsid,
  25                map->epoch, &vaf);
  26
  27         va_end(args);
  28 }
  29
  30 char *ceph_osdmap_state_str(char *str, int len, u32 state)
  31 {
  32         if (!len)
  33                 return str;
  34
  35         if ((state & CEPH_OSD_EXISTS) && (state & CEPH_OSD_UP))
  36                 snprintf(str, len, "exists, up");
  37         else if (state & CEPH_OSD_EXISTS)
  38                 snprintf(str, len, "exists");
  39         else if (state & CEPH_OSD_UP)
  40                 snprintf(str, len, "up");
  41         else
  42                 snprintf(str, len, "doesn't exist");
  43
  44         return str;
  45 }
  46
  47 /* maps */
  48
  49 static int calc_bits_of(unsigned int t)
  50 {
  51         int b = 0;
  52         while (t) {
  53                 t = t >> 1;
  54                 b++;
  55         }
  56         return b;
  57 }
  58
  59 /*
  60  * the foo_mask is the smallest value 2^n-1 that is >= foo.
  61  */
  62 static void calc_pg_masks(struct ceph_pg_pool_info *pi)
  63 {
  64         pi->pg_num_mask = (1 << calc_bits_of(pi->pg_num-1)) - 1;
  65         pi->pgp_num_mask = (1 << calc_bits_of(pi->pgp_num-1)) - 1;
  66 }
  67
  68 /*
  69  * decode crush map
  70  */
  71 static int crush_decode_uniform_bucket(void **p, void *end,
  72                                        struct crush_bucket_uniform *b)
  73 {
  74         dout("crush_decode_uniform_bucket %p to %p\n", *p, end);
  75         ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad);
  76         b->item_weight = ceph_decode_32(p);
  77         return 0;
  78 bad:
  79         return -EINVAL;
  80 }
  81
  82 static int crush_decode_list_bucket(void **p, void *end,
  83                                     struct crush_bucket_list *b)
  84 {
  85         int j;
  86         dout("crush_decode_list_bucket %p to %p\n", *p, end);
  87         b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
  88         if (b->item_weights == NULL)
  89                 return -ENOMEM;
  90         b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
  91         if (b->sum_weights == NULL)
  92                 return -ENOMEM;
  93         ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
  94         for (j = 0; j < b->h.size; j++) {
  95                 b->item_weights[j] = ceph_decode_32(p);
  96                 b->sum_weights[j] = ceph_decode_32(p);
  97         }
  98         return 0;
  99 bad:
 100         return -EINVAL;
 101 }
 102
 103 static int crush_decode_tree_bucket(void **p, void *end,
 104                                     struct crush_bucket_tree *b)
 105 {
 106         int j;
 107         dout("crush_decode_tree_bucket %p to %p\n", *p, end);
 108         ceph_decode_8_safe(p, end, b->num_nodes, bad);
 109         b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS);
 110         if (b->node_weights == NULL)
 111                 return -ENOMEM;
 112         ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad);
 113         for (j = 0; j < b->num_nodes; j++)
 114                 b->node_weights[j] = ceph_decode_32(p);
 115         return 0;
 116 bad:
 117         return -EINVAL;
 118 }
 119
 120 static int crush_decode_straw_bucket(void **p, void *end,
 121                                      struct crush_bucket_straw *b)
 122 {
 123         int j;
 124         dout("crush_decode_straw_bucket %p to %p\n", *p, end);
 125         b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
 126         if (b->item_weights == NULL)
 127                 return -ENOMEM;
 128         b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
 129         if (b->straws == NULL)
 130                 return -ENOMEM;
 131         ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
 132         for (j = 0; j < b->h.size; j++) {
 133                 b->item_weights[j] = ceph_decode_32(p);
 134                 b->straws[j] = ceph_decode_32(p);
 135         }
 136         return 0;
 137 bad:
 138         return -EINVAL;
 139 }
 140
 141 static int crush_decode_straw2_bucket(void **p, void *end,
 142                                       struct crush_bucket_straw2 *b)
 143 {
 144         int j;
 145         dout("crush_decode_straw2_bucket %p to %p\n", *p, end);
 146         b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
 147         if (b->item_weights == NULL)
 148                 return -ENOMEM;
 149         ceph_decode_need(p, end, b->h.size * sizeof(u32), bad);
 150         for (j = 0; j < b->h.size; j++)
 151                 b->item_weights[j] = ceph_decode_32(p);
 152         return 0;
 153 bad:
 154         return -EINVAL;
 155 }
 156
 157 struct crush_name_node {
 158         struct rb_node cn_node;
 159         int cn_id;
 160         char cn_name[];
 161 };
 162
 163 static struct crush_name_node *alloc_crush_name(size_t name_len)
 164 {
 165         struct crush_name_node *cn;
 166
 167         cn = kmalloc(sizeof(*cn) + name_len + 1, GFP_NOIO);
 168         if (!cn)
 169                 return NULL;
 170
 171         RB_CLEAR_NODE(&cn->cn_node);
 172         return cn;
 173 }
 174
 175 static void free_crush_name(struct crush_name_node *cn)
 176 {
 177         WARN_ON(!RB_EMPTY_NODE(&cn->cn_node));
 178
 179         kfree(cn);
 180 }
 181
 182 DEFINE_RB_FUNCS(crush_name, struct crush_name_node, cn_id, cn_node)
 183
 184 static int decode_crush_names(void **p, void *end, struct rb_root *root)
 185 {
 186         u32 n;
 187
 188         ceph_decode_32_safe(p, end, n, e_inval);
 189         while (n--) {
 190                 struct crush_name_node *cn;
 191                 int id;
 192                 u32 name_len;
 193
 194                 ceph_decode_32_safe(p, end, id, e_inval);
 195                 ceph_decode_32_safe(p, end, name_len, e_inval);
 196                 ceph_decode_need(p, end, name_len, e_inval);
 197
 198                 cn = alloc_crush_name(name_len);
 199                 if (!cn)
 200                         return -ENOMEM;
 201
 202                 cn->cn_id = id;
 203                 memcpy(cn->cn_name, *p, name_len);
 204                 cn->cn_name[name_len] = '\0';
 205                 *p += name_len;
 206
 207                 if (!__insert_crush_name(root, cn)) {
 208                         free_crush_name(cn);
 209                         return -EEXIST;
 210                 }
 211         }
 212
 213         return 0;
 214
 215 e_inval:
 216         return -EINVAL;
 217 }
 218
 219 void clear_crush_names(struct rb_root *root)
 220 {
 221         while (!RB_EMPTY_ROOT(root)) {
 222                 struct crush_name_node *cn =
 223                     rb_entry(rb_first(root), struct crush_name_node, cn_node);
 224
 225                 erase_crush_name(root, cn);
 226                 free_crush_name(cn);
 227         }
 228 }
 229
 230 static struct crush_choose_arg_map *alloc_choose_arg_map(void)
 231 {
 232         struct crush_choose_arg_map *arg_map;
 233
 234         arg_map = kzalloc(sizeof(*arg_map), GFP_NOIO);
 235         if (!arg_map)
 236                 return NULL;
 237
 238         RB_CLEAR_NODE(&arg_map->node);
 239         return arg_map;
 240 }
 241
 242 static void free_choose_arg_map(struct crush_choose_arg_map *arg_map)
 243 {
 244         if (arg_map) {
 245                 int i, j;
 246
 247                 WARN_ON(!RB_EMPTY_NODE(&arg_map->node));
 248
 249                 for (i = 0; i < arg_map->size; i++) {
 250                         struct crush_choose_arg *arg = &arg_map->args[i];
 251
 252                         for (j = 0; j < arg->weight_set_size; j++)
 253                                 kfree(arg->weight_set[j].weights);
 254                         kfree(arg->weight_set);
 255                         kfree(arg->ids);
 256                 }
 257                 kfree(arg_map->args);
 258                 kfree(arg_map);
 259         }
 260 }
 261
 262 DEFINE_RB_FUNCS(choose_arg_map, struct crush_choose_arg_map, choose_args_index,
 263                 node);
 264
 265 void clear_choose_args(struct crush_map *c)
 266 {
 267         while (!RB_EMPTY_ROOT(&c->choose_args)) {
 268                 struct crush_choose_arg_map *arg_map =
 269                     rb_entry(rb_first(&c->choose_args),
 270                              struct crush_choose_arg_map, node);
 271
 272                 erase_choose_arg_map(&c->choose_args, arg_map);
 273                 free_choose_arg_map(arg_map);
 274         }
 275 }
 276
 277 static u32 *decode_array_32_alloc(void **p, void *end, u32 *plen)
 278 {
 279         u32 *a = NULL;
 280         u32 len;
 281         int ret;
 282
 283         ceph_decode_32_safe(p, end, len, e_inval);
 284         if (len) {
 285                 u32 i;
 286
 287                 a = kmalloc_array(len, sizeof(u32), GFP_NOIO);
 288                 if (!a) {
 289                         ret = -ENOMEM;
 290                         goto fail;
 291                 }
 292
 293                 ceph_decode_need(p, end, len * sizeof(u32), e_inval);
 294                 for (i = 0; i < len; i++)
 295                         a[i] = ceph_decode_32(p);
 296         }
 297
 298         *plen = len;
 299         return a;
 300
 301 e_inval:
 302         ret = -EINVAL;
 303 fail:
 304         kfree(a);
 305         return ERR_PTR(ret);
 306 }
 307
 308 /*
 309  * Assumes @arg is zero-initialized.
 310  */
 311 static int decode_choose_arg(void **p, void *end, struct crush_choose_arg *arg)
 312 {
 313         int ret;
 314
 315         ceph_decode_32_safe(p, end, arg->weight_set_size, e_inval);
 316         if (arg->weight_set_size) {
 317                 u32 i;
 318
 319                 arg->weight_set = kmalloc_array(arg->weight_set_size,
 320                                                 sizeof(*arg->weight_set),
 321                                                 GFP_NOIO);
 322                 if (!arg->weight_set)
 323                         return -ENOMEM;
 324
 325                 for (i = 0; i < arg->weight_set_size; i++) {
 326                         struct crush_weight_set *w = &arg->weight_set[i];
 327
 328                         w->weights = decode_array_32_alloc(p, end, &w->size);
 329                         if (IS_ERR(w->weights)) {
 330                                 ret = PTR_ERR(w->weights);
 331                                 w->weights = NULL;
 332                                 return ret;
 333                         }
 334                 }
 335         }
 336
 337         arg->ids = decode_array_32_alloc(p, end, &arg->ids_size);
 338         if (IS_ERR(arg->ids)) {
 339                 ret = PTR_ERR(arg->ids);
 340                 arg->ids = NULL;
 341                 return ret;
 342         }
 343
 344         return 0;
 345
 346 e_inval:
 347         return -EINVAL;
 348 }
 349
 350 static int decode_choose_args(void **p, void *end, struct crush_map *c)
 351 {
 352         struct crush_choose_arg_map *arg_map = NULL;
 353         u32 num_choose_arg_maps, num_buckets;
 354         int ret;
 355
 356         ceph_decode_32_safe(p, end, num_choose_arg_maps, e_inval);
 357         while (num_choose_arg_maps--) {
 358                 arg_map = alloc_choose_arg_map();
 359                 if (!arg_map) {
 360                         ret = -ENOMEM;
 361                         goto fail;
 362                 }
 363
 364                 ceph_decode_64_safe(p, end, arg_map->choose_args_index,
 365                                     e_inval);
 366                 arg_map->size = c->max_buckets;
 367                 arg_map->args = kcalloc(arg_map->size, sizeof(*arg_map->args),
 368                                         GFP_NOIO);
 369                 if (!arg_map->args) {
 370                         ret = -ENOMEM;
 371                         goto fail;
 372                 }
 373
 374                 ceph_decode_32_safe(p, end, num_buckets, e_inval);
 375                 while (num_buckets--) {
 376                         struct crush_choose_arg *arg;
 377                         u32 bucket_index;
 378
 379                         ceph_decode_32_safe(p, end, bucket_index, e_inval);
 380                         if (bucket_index >= arg_map->size)
 381                                 goto e_inval;
 382
 383                         arg = &arg_map->args[bucket_index];
 384                         ret = decode_choose_arg(p, end, arg);
 385                         if (ret)
 386                                 goto fail;
 387
 388                         if (arg->ids_size &&
 389                             arg->ids_size != c->buckets[bucket_index]->size)
 390                                 goto e_inval;
 391                 }
 392
 393                 insert_choose_arg_map(&c->choose_args, arg_map);
 394         }
 395
 396         return 0;
 397
 398 e_inval:
 399         ret = -EINVAL;
 400 fail:
 401         free_choose_arg_map(arg_map);
 402         return ret;
 403 }
 404
 405 static void crush_finalize(struct crush_map *c)
 406 {
 407         __s32 b;
 408
 409         /* Space for the array of pointers to per-bucket workspace */
 410         c->working_size = sizeof(struct crush_work) +
 411             c->max_buckets * sizeof(struct crush_work_bucket *);
 412
 413         for (b = 0; b < c->max_buckets; b++) {
 414                 if (!c->buckets[b])
 415                         continue;
 416
 417                 switch (c->buckets[b]->alg) {
 418                 default:
 419                         /*
 420                          * The base case, permutation variables and
 421                          * the pointer to the permutation array.
 422                          */
 423                         c->working_size += sizeof(struct crush_work_bucket);
 424                         break;
 425                 }
 426                 /* Every bucket has a permutation array. */
 427                 c->working_size += c->buckets[b]->size * sizeof(__u32);
 428         }
 429 }
 430
 431 static struct crush_map *crush_decode(void *pbyval, void *end)
 432 {
 433         struct crush_map *c;
 434         int err;
 435         int i, j;
 436         void **p = &pbyval;
 437         void *start = pbyval;
 438         u32 magic;
 439
 440         dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p));
 441
 442         c = kzalloc(sizeof(*c), GFP_NOFS);
 443         if (c == NULL)
 444                 return ERR_PTR(-ENOMEM);
 445
 446         c->type_names = RB_ROOT;
 447         c->names = RB_ROOT;
 448         c->choose_args = RB_ROOT;
 449
 450         /* set tunables to default values */
 451         c->choose_local_tries = 2;
 452         c->choose_local_fallback_tries = 5;
 453         c->choose_total_tries = 19;
 454         c->chooseleaf_descend_once = 0;
 455
 456         ceph_decode_need(p, end, 4*sizeof(u32), bad);
 457         magic = ceph_decode_32(p);
 458         if (magic != CRUSH_MAGIC) {
 459                 pr_err("crush_decode magic %x != current %x\n",
 460                        (unsigned int)magic, (unsigned int)CRUSH_MAGIC);
 461                 goto bad;
 462         }
 463         c->max_buckets = ceph_decode_32(p);
 464         c->max_rules = ceph_decode_32(p);
 465         c->max_devices = ceph_decode_32(p);
 466
 467         c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS);
 468         if (c->buckets == NULL)
 469                 goto badmem;
 470         c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS);
 471         if (c->rules == NULL)
 472                 goto badmem;
 473
 474         /* buckets */
 475         for (i = 0; i < c->max_buckets; i++) {
 476                 int size = 0;
 477                 u32 alg;
 478                 struct crush_bucket *b;
 479
 480                 ceph_decode_32_safe(p, end, alg, bad);
 481                 if (alg == 0) {
 482                         c->buckets[i] = NULL;
 483                         continue;
 484                 }
 485                 dout("crush_decode bucket %d off %x %p to %p\n",
 486                      i, (int)(*p-start), *p, end);
 487
 488                 switch (alg) {
 489                 case CRUSH_BUCKET_UNIFORM:
 490                         size = sizeof(struct crush_bucket_uniform);
 491                         break;
 492                 case CRUSH_BUCKET_LIST:
 493                         size = sizeof(struct crush_bucket_list);
 494                         break;
 495                 case CRUSH_BUCKET_TREE:
 496                         size = sizeof(struct crush_bucket_tree);
 497                         break;
 498                 case CRUSH_BUCKET_STRAW:
 499                         size = sizeof(struct crush_bucket_straw);
 500                         break;
 501                 case CRUSH_BUCKET_STRAW2:
 502                         size = sizeof(struct crush_bucket_straw2);
 503                         break;
 504                 default:
 505                         goto bad;
 506                 }
 507                 BUG_ON(size == 0);
 508                 b = c->buckets[i] = kzalloc(size, GFP_NOFS);
 509                 if (b == NULL)
 510                         goto badmem;
 511
 512                 ceph_decode_need(p, end, 4*sizeof(u32), bad);
 513                 b->id = ceph_decode_32(p);
 514                 b->type = ceph_decode_16(p);
 515                 b->alg = ceph_decode_8(p);
 516                 b->hash = ceph_decode_8(p);
 517                 b->weight = ceph_decode_32(p);
 518                 b->size = ceph_decode_32(p);
 519
 520                 dout("crush_decode bucket size %d off %x %p to %p\n",
 521                      b->size, (int)(*p-start), *p, end);
 522
 523                 b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS);
 524                 if (b->items == NULL)
 525                         goto badmem;
 526
 527                 ceph_decode_need(p, end, b->size*sizeof(u32), bad);
 528                 for (j = 0; j < b->size; j++)
 529                         b->items[j] = ceph_decode_32(p);
 530
 531                 switch (b->alg) {
 532                 case CRUSH_BUCKET_UNIFORM:
 533                         err = crush_decode_uniform_bucket(p, end,
 534                                   (struct crush_bucket_uniform *)b);
 535                         if (err < 0)
 536                                 goto fail;
 537                         break;
 538                 case CRUSH_BUCKET_LIST:
 539                         err = crush_decode_list_bucket(p, end,
 540                                (struct crush_bucket_list *)b);
 541                         if (err < 0)
 542                                 goto fail;
 543                         break;
 544                 case CRUSH_BUCKET_TREE:
 545                         err = crush_decode_tree_bucket(p, end,
 546                                 (struct crush_bucket_tree *)b);
 547                         if (err < 0)
 548                                 goto fail;
 549                         break;
 550                 case CRUSH_BUCKET_STRAW:
 551                         err = crush_decode_straw_bucket(p, end,
 552                                 (struct crush_bucket_straw *)b);
 553                         if (err < 0)
 554                                 goto fail;
 555                         break;
 556                 case CRUSH_BUCKET_STRAW2:
 557                         err = crush_decode_straw2_bucket(p, end,
 558                                 (struct crush_bucket_straw2 *)b);
 559                         if (err < 0)
 560                                 goto fail;
 561                         break;
 562                 }
 563         }
 564
 565         /* rules */
 566         dout("rule vec is %p\n", c->rules);
 567         for (i = 0; i < c->max_rules; i++) {
 568                 u32 yes;
 569                 struct crush_rule *r;
 570
 571                 ceph_decode_32_safe(p, end, yes, bad);
 572                 if (!yes) {
 573                         dout("crush_decode NO rule %d off %x %p to %p\n",
 574                              i, (int)(*p-start), *p, end);
 575                         c->rules[i] = NULL;
 576                         continue;
 577                 }
 578
 579                 dout("crush_decode rule %d off %x %p to %p\n",
 580                      i, (int)(*p-start), *p, end);
 581
 582                 /* len */
 583                 ceph_decode_32_safe(p, end, yes, bad);
 584 #if BITS_PER_LONG == 32
 585                 if (yes > (ULONG_MAX - sizeof(*r))
 586                           / sizeof(struct crush_rule_step))
 587                         goto bad;
 588 #endif
 589                 r = kmalloc(struct_size(r, steps, yes), GFP_NOFS);
 590                 if (r == NULL)
 591                         goto badmem;
 592                 dout(" rule %d is at %p\n", i, r);
 593                 c->rules[i] = r;
 594                 r->len = yes;
 595                 ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */
 596                 ceph_decode_need(p, end, r->len*3*sizeof(u32), bad);
 597                 for (j = 0; j < r->len; j++) {
 598                         r->steps[j].op = ceph_decode_32(p);
 599                         r->steps[j].arg1 = ceph_decode_32(p);
 600                         r->steps[j].arg2 = ceph_decode_32(p);
 601                 }
 602         }
 603
 604         err = decode_crush_names(p, end, &c->type_names);
 605         if (err)
 606                 goto fail;
 607
 608         err = decode_crush_names(p, end, &c->names);
 609         if (err)
 610                 goto fail;
 611
 612         ceph_decode_skip_map(p, end, 32, string, bad); /* rule_name_map */
 613
 614         /* tunables */
 615         ceph_decode_need(p, end, 3*sizeof(u32), done);
 616         c->choose_local_tries = ceph_decode_32(p);
 617         c->choose_local_fallback_tries =  ceph_decode_32(p);
 618         c->choose_total_tries = ceph_decode_32(p);
 619         dout("crush decode tunable choose_local_tries = %d\n",
 620              c->choose_local_tries);
 621         dout("crush decode tunable choose_local_fallback_tries = %d\n",
 622              c->choose_local_fallback_tries);
 623         dout("crush decode tunable choose_total_tries = %d\n",
 624              c->choose_total_tries);
 625
 626         ceph_decode_need(p, end, sizeof(u32), done);
 627         c->chooseleaf_descend_once = ceph_decode_32(p);
 628         dout("crush decode tunable chooseleaf_descend_once = %d\n",
 629              c->chooseleaf_descend_once);
 630
 631         ceph_decode_need(p, end, sizeof(u8), done);
 632         c->chooseleaf_vary_r = ceph_decode_8(p);
 633         dout("crush decode tunable chooseleaf_vary_r = %d\n",
 634              c->chooseleaf_vary_r);
 635
 636         /* skip straw_calc_version, allowed_bucket_algs */
 637         ceph_decode_need(p, end, sizeof(u8) + sizeof(u32), done);
 638         *p += sizeof(u8) + sizeof(u32);
 639
 640         ceph_decode_need(p, end, sizeof(u8), done);
 641         c->chooseleaf_stable = ceph_decode_8(p);
 642         dout("crush decode tunable chooseleaf_stable = %d\n",
 643              c->chooseleaf_stable);
 644
 645         if (*p != end) {
 646                 /* class_map */
 647                 ceph_decode_skip_map(p, end, 32, 32, bad);
 648                 /* class_name */
 649                 ceph_decode_skip_map(p, end, 32, string, bad);
 650                 /* class_bucket */
 651                 ceph_decode_skip_map_of_map(p, end, 32, 32, 32, bad);
 652         }
 653
 654         if (*p != end) {
 655                 err = decode_choose_args(p, end, c);
 656                 if (err)
 657                         goto fail;
 658         }
 659
 660 done:
 661         crush_finalize(c);
 662         dout("crush_decode success\n");
 663         return c;
 664
 665 badmem:
 666         err = -ENOMEM;
 667 fail:
 668         dout("crush_decode fail %d\n", err);
 669         crush_destroy(c);
 670         return ERR_PTR(err);
 671
 672 bad:
 673         err = -EINVAL;
 674         goto fail;
 675 }
 676
 677 int ceph_pg_compare(const struct ceph_pg *lhs, const struct ceph_pg *rhs)
 678 {
 679         if (lhs->pool < rhs->pool)
 680                 return -1;
 681         if (lhs->pool > rhs->pool)
 682                 return 1;
 683         if (lhs->seed < rhs->seed)
 684                 return -1;
 685         if (lhs->seed > rhs->seed)
 686                 return 1;
 687
 688         return 0;
 689 }
 690
 691 int ceph_spg_compare(const struct ceph_spg *lhs, const struct ceph_spg *rhs)
 692 {
 693         int ret;
 694
 695         ret = ceph_pg_compare(&lhs->pgid, &rhs->pgid);
 696         if (ret)
 697                 return ret;
 698
 699         if (lhs->shard < rhs->shard)
 700                 return -1;
 701         if (lhs->shard > rhs->shard)
 702                 return 1;
 703
 704         return 0;
 705 }
 706
 707 static struct ceph_pg_mapping *alloc_pg_mapping(size_t payload_len)
 708 {
 709         struct ceph_pg_mapping *pg;
 710
 711         pg = kmalloc(sizeof(*pg) + payload_len, GFP_NOIO);
 712         if (!pg)
 713                 return NULL;
 714
 715         RB_CLEAR_NODE(&pg->node);
 716         return pg;
 717 }
 718
 719 static void free_pg_mapping(struct ceph_pg_mapping *pg)
 720 {
 721         WARN_ON(!RB_EMPTY_NODE(&pg->node));
 722
 723         kfree(pg);
 724 }
 725
 726 /*
 727  * rbtree of pg_mapping for handling pg_temp (explicit mapping of pgid
 728  * to a set of osds) and primary_temp (explicit primary setting)
 729  */
 730 DEFINE_RB_FUNCS2(pg_mapping, struct ceph_pg_mapping, pgid, ceph_pg_compare,
 731                  RB_BYPTR, const struct ceph_pg *, node)
 732
 733 /*
 734  * rbtree of pg pool info
 735  */
 736 DEFINE_RB_FUNCS(pg_pool, struct ceph_pg_pool_info, id, node)
 737
 738 struct ceph_pg_pool_info *ceph_pg_pool_by_id(struct ceph_osdmap *map, u64 id)
 739 {
 740         return lookup_pg_pool(&map->pg_pools, id);
 741 }
 742
 743 const char *ceph_pg_pool_name_by_id(struct ceph_osdmap *map, u64 id)
 744 {
 745         struct ceph_pg_pool_info *pi;
 746
 747         if (id == CEPH_NOPOOL)
 748                 return NULL;
 749
 750         if (WARN_ON_ONCE(id > (u64) INT_MAX))
 751                 return NULL;
 752
 753         pi = lookup_pg_pool(&map->pg_pools, id);
 754         return pi ? pi->name : NULL;
 755 }
 756 EXPORT_SYMBOL(ceph_pg_pool_name_by_id);
 757
 758 int ceph_pg_poolid_by_name(struct ceph_osdmap *map, const char *name)
 759 {
 760         struct rb_node *rbp;
 761
 762         for (rbp = rb_first(&map->pg_pools); rbp; rbp = rb_next(rbp)) {
 763                 struct ceph_pg_pool_info *pi =
 764                         rb_entry(rbp, struct ceph_pg_pool_info, node);
 765                 if (pi->name && strcmp(pi->name, name) == 0)
 766                         return pi->id;
 767         }
 768         return -ENOENT;
 769 }
 770 EXPORT_SYMBOL(ceph_pg_poolid_by_name);
 771
 772 u64 ceph_pg_pool_flags(struct ceph_osdmap *map, u64 id)
 773 {
 774         struct ceph_pg_pool_info *pi;
 775
 776         pi = lookup_pg_pool(&map->pg_pools, id);
 777         return pi ? pi->flags : 0;
 778 }
 779 EXPORT_SYMBOL(ceph_pg_pool_flags);
 780
 781 static void __remove_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *pi)
 782 {
 783         erase_pg_pool(root, pi);
 784         kfree(pi->name);
 785         kfree(pi);
 786 }
 787
 788 static int decode_pool(void **p, void *end, struct ceph_pg_pool_info *pi)
 789 {
 790         u8 ev, cv;
 791         unsigned len, num;
 792         void *pool_end;
 793
 794         ceph_decode_need(p, end, 2 + 4, bad);
 795         ev = ceph_decode_8(p);  /* encoding version */
 796         cv = ceph_decode_8(p); /* compat version */
 797         if (ev < 5) {
 798                 pr_warn("got v %d < 5 cv %d of ceph_pg_pool\n", ev, cv);
 799                 return -EINVAL;
 800         }
 801         if (cv > 9) {
 802                 pr_warn("got v %d cv %d > 9 of ceph_pg_pool\n", ev, cv);
 803                 return -EINVAL;
 804         }
 805         len = ceph_decode_32(p);
 806         ceph_decode_need(p, end, len, bad);
 807         pool_end = *p + len;
 808
 809         pi->type = ceph_decode_8(p);
 810         pi->size = ceph_decode_8(p);
 811         pi->crush_ruleset = ceph_decode_8(p);
 812         pi->object_hash = ceph_decode_8(p);
 813
 814         pi->pg_num = ceph_decode_32(p);
 815         pi->pgp_num = ceph_decode_32(p);
 816
 817         *p += 4 + 4;  /* skip lpg* */
 818         *p += 4;      /* skip last_change */
 819         *p += 8 + 4;  /* skip snap_seq, snap_epoch */
 820
 821         /* skip snaps */
 822         num = ceph_decode_32(p);
 823         while (num--) {
 824                 *p += 8;  /* snapid key */
 825                 *p += 1 + 1; /* versions */
 826                 len = ceph_decode_32(p);
 827                 *p += len;
 828         }
 829
 830         /* skip removed_snaps */
 831         num = ceph_decode_32(p);
 832         *p += num * (8 + 8);
 833
 834         *p += 8;  /* skip auid */
 835         pi->flags = ceph_decode_64(p);
 836         *p += 4;  /* skip crash_replay_interval */
 837
 838         if (ev >= 7)
 839                 pi->min_size = ceph_decode_8(p);
 840         else
 841                 pi->min_size = pi->size - pi->size / 2;
 842
 843         if (ev >= 8)
 844                 *p += 8 + 8;  /* skip quota_max_* */
 845
 846         if (ev >= 9) {
 847                 /* skip tiers */
 848                 num = ceph_decode_32(p);
 849                 *p += num * 8;
 850
 851                 *p += 8;  /* skip tier_of */
 852                 *p += 1;  /* skip cache_mode */
 853
 854                 pi->read_tier = ceph_decode_64(p);
 855                 pi->write_tier = ceph_decode_64(p);
 856         } else {
 857                 pi->read_tier = -1;
 858                 pi->write_tier = -1;
 859         }
 860
 861         if (ev >= 10) {
 862                 /* skip properties */
 863                 num = ceph_decode_32(p);
 864                 while (num--) {
 865                         len = ceph_decode_32(p);
 866                         *p += len; /* key */
 867                         len = ceph_decode_32(p);
 868                         *p += len; /* val */
 869                 }
 870         }
 871
 872         if (ev >= 11) {
 873                 /* skip hit_set_params */
 874                 *p += 1 + 1; /* versions */
 875                 len = ceph_decode_32(p);
 876                 *p += len;
 877
 878                 *p += 4; /* skip hit_set_period */
 879                 *p += 4; /* skip hit_set_count */
 880         }
 881
 882         if (ev >= 12)
 883                 *p += 4; /* skip stripe_width */
 884
 885         if (ev >= 13) {
 886                 *p += 8; /* skip target_max_bytes */
 887                 *p += 8; /* skip target_max_objects */
 888                 *p += 4; /* skip cache_target_dirty_ratio_micro */
 889                 *p += 4; /* skip cache_target_full_ratio_micro */
 890                 *p += 4; /* skip cache_min_flush_age */
 891                 *p += 4; /* skip cache_min_evict_age */
 892         }
 893
 894         if (ev >=  14) {
 895                 /* skip erasure_code_profile */
 896                 len = ceph_decode_32(p);
 897                 *p += len;
 898         }
 899
 900         /*
 901          * last_force_op_resend_preluminous, will be overridden if the
 902          * map was encoded with RESEND_ON_SPLIT
 903          */
 904         if (ev >= 15)
 905                 pi->last_force_request_resend = ceph_decode_32(p);
 906         else
 907                 pi->last_force_request_resend = 0;
 908
 909         if (ev >= 16)
 910                 *p += 4; /* skip min_read_recency_for_promote */
 911
 912         if (ev >= 17)
 913                 *p += 8; /* skip expected_num_objects */
 914
 915         if (ev >= 19)
 916                 *p += 4; /* skip cache_target_dirty_high_ratio_micro */
 917
 918         if (ev >= 20)
 919                 *p += 4; /* skip min_write_recency_for_promote */
 920
 921         if (ev >= 21)
 922                 *p += 1; /* skip use_gmt_hitset */
 923
 924         if (ev >= 22)
 925                 *p += 1; /* skip fast_read */
 926
 927         if (ev >= 23) {
 928                 *p += 4; /* skip hit_set_grade_decay_rate */
 929                 *p += 4; /* skip hit_set_search_last_n */
 930         }
 931
 932         if (ev >= 24) {
 933                 /* skip opts */
 934                 *p += 1 + 1; /* versions */
 935                 len = ceph_decode_32(p);
 936                 *p += len;
 937         }
 938
 939         if (ev >= 25)
 940                 pi->last_force_request_resend = ceph_decode_32(p);
 941
 942         /* ignore the rest */
 943
 944         *p = pool_end;
 945         calc_pg_masks(pi);
 946         return 0;
 947
 948 bad:
 949         return -EINVAL;
 950 }
 951
 952 static int decode_pool_names(void **p, void *end, struct ceph_osdmap *map)
 953 {
 954         struct ceph_pg_pool_info *pi;
 955         u32 num, len;
 956         u64 pool;
 957
 958         ceph_decode_32_safe(p, end, num, bad);
 959         dout(" %d pool names\n", num);
 960         while (num--) {
 961                 ceph_decode_64_safe(p, end, pool, bad);
 962                 ceph_decode_32_safe(p, end, len, bad);
 963                 dout("  pool %llu len %d\n", pool, len);
 964                 ceph_decode_need(p, end, len, bad);
 965                 pi = lookup_pg_pool(&map->pg_pools, pool);
 966                 if (pi) {
 967                         char *name = kstrndup(*p, len, GFP_NOFS);
 968
 969                         if (!name)
 970                                 return -ENOMEM;
 971                         kfree(pi->name);
 972                         pi->name = name;
 973                         dout("  name is %s\n", pi->name);
 974                 }
 975                 *p += len;
 976         }
 977         return 0;
 978
 979 bad:
 980         return -EINVAL;
 981 }
 982
 983 /*
 984  * CRUSH workspaces
 985  *
 986  * workspace_manager framework borrowed from fs/btrfs/compression.c.
 987  * Two simplifications: there is only one type of workspace and there
 988  * is always at least one workspace.
 989  */
 990 static struct crush_work *alloc_workspace(const struct crush_map *c)
 991 {
 992         struct crush_work *work;
 993         size_t work_size;
 994
 995         WARN_ON(!c->working_size);
 996         work_size = crush_work_size(c, CEPH_PG_MAX_SIZE);
 997         dout("%s work_size %zu bytes\n", __func__, work_size);
 998
 999         work = kvmalloc(work_size, GFP_NOIO);
1000         if (!work)
1001                 return NULL;
1002
1003         INIT_LIST_HEAD(&work->item);
1004         crush_init_workspace(c, work);
1005         return work;
1006 }
1007
1008 static void free_workspace(struct crush_work *work)
1009 {
1010         WARN_ON(!list_empty(&work->item));
1011         kvfree(work);
1012 }
1013
1014 static void init_workspace_manager(struct workspace_manager *wsm)
1015 {
1016         INIT_LIST_HEAD(&wsm->idle_ws);
1017         spin_lock_init(&wsm->ws_lock);
1018         atomic_set(&wsm->total_ws, 0);
1019         wsm->free_ws = 0;
1020         init_waitqueue_head(&wsm->ws_wait);
1021 }
1022
1023 static void add_initial_workspace(struct workspace_manager *wsm,
1024                                   struct crush_work *work)
1025 {
1026         WARN_ON(!list_empty(&wsm->idle_ws));
1027
1028         list_add(&work->item, &wsm->idle_ws);
1029         atomic_set(&wsm->total_ws, 1);
1030         wsm->free_ws = 1;
1031 }
1032
1033 static void cleanup_workspace_manager(struct workspace_manager *wsm)
1034 {
1035         struct crush_work *work;
1036
1037         while (!list_empty(&wsm->idle_ws)) {
1038                 work = list_first_entry(&wsm->idle_ws, struct crush_work,
1039                                         item);
1040                 list_del_init(&work->item);
1041                 free_workspace(work);
1042         }
1043         atomic_set(&wsm->total_ws, 0);
1044         wsm->free_ws = 0;
1045 }
1046
1047 /*
1048  * Finds an available workspace or allocates a new one.  If it's not
1049  * possible to allocate a new one, waits until there is one.
1050  */
1051 static struct crush_work *get_workspace(struct workspace_manager *wsm,
1052                                         const struct crush_map *c)
1053 {
1054         struct crush_work *work;
1055         int cpus = num_online_cpus();
1056
1057 again:
1058         spin_lock(&wsm->ws_lock);
1059         if (!list_empty(&wsm->idle_ws)) {
1060                 work = list_first_entry(&wsm->idle_ws, struct crush_work,
1061                                         item);
1062                 list_del_init(&work->item);
1063                 wsm->free_ws--;
1064                 spin_unlock(&wsm->ws_lock);
1065                 return work;
1066
1067         }
1068         if (atomic_read(&wsm->total_ws) > cpus) {
1069                 DEFINE_WAIT(wait);
1070
1071                 spin_unlock(&wsm->ws_lock);
1072                 prepare_to_wait(&wsm->ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1073                 if (atomic_read(&wsm->total_ws) > cpus && !wsm->free_ws)
1074                         schedule();
1075                 finish_wait(&wsm->ws_wait, &wait);
1076                 goto again;
1077         }
1078         atomic_inc(&wsm->total_ws);
1079         spin_unlock(&wsm->ws_lock);
1080
1081         work = alloc_workspace(c);
1082         if (!work) {
1083                 atomic_dec(&wsm->total_ws);
1084                 wake_up(&wsm->ws_wait);
1085
1086                 /*
1087                  * Do not return the error but go back to waiting.  We
1088                  * have the initial workspace and the CRUSH computation
1089                  * time is bounded so we will get it eventually.
1090                  */
1091                 WARN_ON(atomic_read(&wsm->total_ws) < 1);
1092                 goto again;
1093         }
1094         return work;
1095 }
1096
1097 /*
1098  * Puts a workspace back on the list or frees it if we have enough
1099  * idle ones sitting around.
1100  */
1101 static void put_workspace(struct workspace_manager *wsm,
1102                           struct crush_work *work)
1103 {
1104         spin_lock(&wsm->ws_lock);
1105         if (wsm->free_ws <= num_online_cpus()) {
1106                 list_add(&work->item, &wsm->idle_ws);
1107                 wsm->free_ws++;
1108                 spin_unlock(&wsm->ws_lock);
1109                 goto wake;
1110         }
1111         spin_unlock(&wsm->ws_lock);
1112
1113         free_workspace(work);
1114         atomic_dec(&wsm->total_ws);
1115 wake:
1116         if (wq_has_sleeper(&wsm->ws_wait))
1117                 wake_up(&wsm->ws_wait);
1118 }
1119
1120 /*
1121  * osd map
1122  */
1123 struct ceph_osdmap *ceph_osdmap_alloc(void)
1124 {
1125         struct ceph_osdmap *map;
1126
1127         map = kzalloc(sizeof(*map), GFP_NOIO);
1128         if (!map)
1129                 return NULL;
1130
1131         map->pg_pools = RB_ROOT;
1132         map->pool_max = -1;
1133         map->pg_temp = RB_ROOT;
1134         map->primary_temp = RB_ROOT;
1135         map->pg_upmap = RB_ROOT;
1136         map->pg_upmap_items = RB_ROOT;
1137
1138         init_workspace_manager(&map->crush_wsm);
1139
1140         return map;
1141 }
1142
1143 void ceph_osdmap_destroy(struct ceph_osdmap *map)
1144 {
1145         dout("osdmap_destroy %p\n", map);
1146
1147         if (map->crush)
1148                 crush_destroy(map->crush);
1149         cleanup_workspace_manager(&map->crush_wsm);
1150
1151         while (!RB_EMPTY_ROOT(&map->pg_temp)) {
1152                 struct ceph_pg_mapping *pg =
1153                         rb_entry(rb_first(&map->pg_temp),
1154                                  struct ceph_pg_mapping, node);
1155                 erase_pg_mapping(&map->pg_temp, pg);
1156                 free_pg_mapping(pg);
1157         }
1158         while (!RB_EMPTY_ROOT(&map->primary_temp)) {
1159                 struct ceph_pg_mapping *pg =
1160                         rb_entry(rb_first(&map->primary_temp),
1161                                  struct ceph_pg_mapping, node);
1162                 erase_pg_mapping(&map->primary_temp, pg);
1163                 free_pg_mapping(pg);
1164         }
1165         while (!RB_EMPTY_ROOT(&map->pg_upmap)) {
1166                 struct ceph_pg_mapping *pg =
1167                         rb_entry(rb_first(&map->pg_upmap),
1168                                  struct ceph_pg_mapping, node);
1169                 rb_erase(&pg->node, &map->pg_upmap);
1170                 kfree(pg);
1171         }
1172         while (!RB_EMPTY_ROOT(&map->pg_upmap_items)) {
1173                 struct ceph_pg_mapping *pg =
1174                         rb_entry(rb_first(&map->pg_upmap_items),
1175                                  struct ceph_pg_mapping, node);
1176                 rb_erase(&pg->node, &map->pg_upmap_items);
1177                 kfree(pg);
1178         }
1179         while (!RB_EMPTY_ROOT(&map->pg_pools)) {
1180                 struct ceph_pg_pool_info *pi =
1181                         rb_entry(rb_first(&map->pg_pools),
1182                                  struct ceph_pg_pool_info, node);
1183                 __remove_pg_pool(&map->pg_pools, pi);
1184         }
1185         kvfree(map->osd_state);
1186         kvfree(map->osd_weight);
1187         kvfree(map->osd_addr);
1188         kvfree(map->osd_primary_affinity);
1189         kfree(map);
1190 }
1191
1192 /*
1193  * Adjust max_osd value, (re)allocate arrays.
1194  *
1195  * The new elements are properly initialized.
1196  */
1197 static int osdmap_set_max_osd(struct ceph_osdmap *map, u32 max)
1198 {
1199         u32 *state;
1200         u32 *weight;
1201         struct ceph_entity_addr *addr;
1202         u32 to_copy;
1203         int i;
1204
1205         dout("%s old %u new %u\n", __func__, map->max_osd, max);
1206         if (max == map->max_osd)
1207                 return 0;
1208
1209         state = kvmalloc(array_size(max, sizeof(*state)), GFP_NOFS);
1210         weight = kvmalloc(array_size(max, sizeof(*weight)), GFP_NOFS);
1211         addr = kvmalloc(array_size(max, sizeof(*addr)), GFP_NOFS);
1212         if (!state || !weight || !addr) {
1213                 kvfree(state);
1214                 kvfree(weight);
1215                 kvfree(addr);
1216                 return -ENOMEM;
1217         }
1218
1219         to_copy = min(map->max_osd, max);
1220         if (map->osd_state) {
1221                 memcpy(state, map->osd_state, to_copy * sizeof(*state));
1222                 memcpy(weight, map->osd_weight, to_copy * sizeof(*weight));
1223                 memcpy(addr, map->osd_addr, to_copy * sizeof(*addr));
1224                 kvfree(map->osd_state);
1225                 kvfree(map->osd_weight);
1226                 kvfree(map->osd_addr);
1227         }
1228
1229         map->osd_state = state;
1230         map->osd_weight = weight;
1231         map->osd_addr = addr;
1232         for (i = map->max_osd; i < max; i++) {
1233                 map->osd_state[i] = 0;
1234                 map->osd_weight[i] = CEPH_OSD_OUT;
1235                 memset(map->osd_addr + i, 0, sizeof(*map->osd_addr));
1236         }
1237
1238         if (map->osd_primary_affinity) {
1239                 u32 *affinity;
1240
1241                 affinity = kvmalloc(array_size(max, sizeof(*affinity)),
1242                                          GFP_NOFS);
1243                 if (!affinity)
1244                         return -ENOMEM;
1245
1246                 memcpy(affinity, map->osd_primary_affinity,
1247                        to_copy * sizeof(*affinity));
1248                 kvfree(map->osd_primary_affinity);
1249
1250                 map->osd_primary_affinity = affinity;
1251                 for (i = map->max_osd; i < max; i++)
1252                         map->osd_primary_affinity[i] =
1253                             CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
1254         }
1255
1256         map->max_osd = max;
1257
1258         return 0;
1259 }
1260
1261 static int osdmap_set_crush(struct ceph_osdmap *map, struct crush_map *crush)
1262 {
1263         struct crush_work *work;
1264
1265         if (IS_ERR(crush))
1266                 return PTR_ERR(crush);
1267
1268         work = alloc_workspace(crush);
1269         if (!work) {
1270                 crush_destroy(crush);
1271                 return -ENOMEM;
1272         }
1273
1274         if (map->crush)
1275                 crush_destroy(map->crush);
1276         cleanup_workspace_manager(&map->crush_wsm);
1277         map->crush = crush;
1278         add_initial_workspace(&map->crush_wsm, work);
1279         return 0;
1280 }
1281
1282 #define OSDMAP_WRAPPER_COMPAT_VER       7
1283 #define OSDMAP_CLIENT_DATA_COMPAT_VER   1
1284
1285 /*
1286  * Return 0 or error.  On success, *v is set to 0 for old (v6) osdmaps,
1287  * to struct_v of the client_data section for new (v7 and above)
1288  * osdmaps.
1289  */
1290 static int get_osdmap_client_data_v(void **p, void *end,
1291                                     const char *prefix, u8 *v)
1292 {
1293         u8 struct_v;
1294
1295         ceph_decode_8_safe(p, end, struct_v, e_inval);
1296         if (struct_v >= 7) {
1297                 u8 struct_compat;
1298
1299                 ceph_decode_8_safe(p, end, struct_compat, e_inval);
1300                 if (struct_compat > OSDMAP_WRAPPER_COMPAT_VER) {
1301                         pr_warn("got v %d cv %d > %d of %s ceph_osdmap\n",
1302                                 struct_v, struct_compat,
1303                                 OSDMAP_WRAPPER_COMPAT_VER, prefix);
1304                         return -EINVAL;
1305                 }
1306                 *p += 4; /* ignore wrapper struct_len */
1307
1308                 ceph_decode_8_safe(p, end, struct_v, e_inval);
1309                 ceph_decode_8_safe(p, end, struct_compat, e_inval);
1310                 if (struct_compat > OSDMAP_CLIENT_DATA_COMPAT_VER) {
1311                         pr_warn("got v %d cv %d > %d of %s ceph_osdmap client data\n",
1312                                 struct_v, struct_compat,
1313                                 OSDMAP_CLIENT_DATA_COMPAT_VER, prefix);
1314                         return -EINVAL;
1315                 }
1316                 *p += 4; /* ignore client data struct_len */
1317         } else {
1318                 u16 version;
1319
1320                 *p -= 1;
1321                 ceph_decode_16_safe(p, end, version, e_inval);
1322                 if (version < 6) {
1323                         pr_warn("got v %d < 6 of %s ceph_osdmap\n",
1324                                 version, prefix);
1325                         return -EINVAL;
1326                 }
1327
1328                 /* old osdmap encoding */
1329                 struct_v = 0;
1330         }
1331
1332         *v = struct_v;
1333         return 0;
1334
1335 e_inval:
1336         return -EINVAL;
1337 }
1338
1339 static int __decode_pools(void **p, void *end, struct ceph_osdmap *map,
1340                           bool incremental)
1341 {
1342         u32 n;
1343
1344         ceph_decode_32_safe(p, end, n, e_inval);
1345         while (n--) {
1346                 struct ceph_pg_pool_info *pi;
1347                 u64 pool;
1348                 int ret;
1349
1350                 ceph_decode_64_safe(p, end, pool, e_inval);
1351
1352                 pi = lookup_pg_pool(&map->pg_pools, pool);
1353                 if (!incremental || !pi) {
1354                         pi = kzalloc(sizeof(*pi), GFP_NOFS);
1355                         if (!pi)
1356                                 return -ENOMEM;
1357
1358                         RB_CLEAR_NODE(&pi->node);
1359                         pi->id = pool;
1360
1361                         if (!__insert_pg_pool(&map->pg_pools, pi)) {
1362                                 kfree(pi);
1363                                 return -EEXIST;
1364                         }
1365                 }
1366
1367                 ret = decode_pool(p, end, pi);
1368                 if (ret)
1369                         return ret;
1370         }
1371
1372         return 0;
1373
1374 e_inval:
1375         return -EINVAL;
1376 }
1377
1378 static int decode_pools(void **p, void *end, struct ceph_osdmap *map)
1379 {
1380         return __decode_pools(p, end, map, false);
1381 }
1382
1383 static int decode_new_pools(void **p, void *end, struct ceph_osdmap *map)
1384 {
1385         return __decode_pools(p, end, map, true);
1386 }
1387
1388 typedef struct ceph_pg_mapping *(*decode_mapping_fn_t)(void **, void *, bool);
1389
1390 static int decode_pg_mapping(void **p, void *end, struct rb_root *mapping_root,
1391                              decode_mapping_fn_t fn, bool incremental)
1392 {
1393         u32 n;
1394
1395         WARN_ON(!incremental && !fn);
1396
1397         ceph_decode_32_safe(p, end, n, e_inval);
1398         while (n--) {
1399                 struct ceph_pg_mapping *pg;
1400                 struct ceph_pg pgid;
1401                 int ret;
1402
1403                 ret = ceph_decode_pgid(p, end, &pgid);
1404                 if (ret)
1405                         return ret;
1406
1407                 pg = lookup_pg_mapping(mapping_root, &pgid);
1408                 if (pg) {
1409                         WARN_ON(!incremental);
1410                         erase_pg_mapping(mapping_root, pg);
1411                         free_pg_mapping(pg);
1412                 }
1413
1414                 if (fn) {
1415                         pg = fn(p, end, incremental);
1416                         if (IS_ERR(pg))
1417                                 return PTR_ERR(pg);
1418
1419                         if (pg) {
1420                                 pg->pgid = pgid; /* struct */
1421                                 insert_pg_mapping(mapping_root, pg);
1422                         }
1423                 }
1424         }
1425
1426         return 0;
1427
1428 e_inval:
1429         return -EINVAL;
1430 }
1431
1432 static struct ceph_pg_mapping *__decode_pg_temp(void **p, void *end,
1433                                                 bool incremental)
1434 {
1435         struct ceph_pg_mapping *pg;
1436         u32 len, i;
1437
1438         ceph_decode_32_safe(p, end, len, e_inval);
1439         if (len == 0 && incremental)
1440                 return NULL;    /* new_pg_temp: [] to remove */
1441         if (len > (SIZE_MAX - sizeof(*pg)) / sizeof(u32))
1442                 return ERR_PTR(-EINVAL);
1443
1444         ceph_decode_need(p, end, len * sizeof(u32), e_inval);
1445         pg = alloc_pg_mapping(len * sizeof(u32));
1446         if (!pg)
1447                 return ERR_PTR(-ENOMEM);
1448
1449         pg->pg_temp.len = len;
1450         for (i = 0; i < len; i++)
1451                 pg->pg_temp.osds[i] = ceph_decode_32(p);
1452
1453         return pg;
1454
1455 e_inval:
1456         return ERR_PTR(-EINVAL);
1457 }
1458
1459 static int decode_pg_temp(void **p, void *end, struct ceph_osdmap *map)
1460 {
1461         return decode_pg_mapping(p, end, &map->pg_temp, __decode_pg_temp,
1462                                  false);
1463 }
1464
1465 static int decode_new_pg_temp(void **p, void *end, struct ceph_osdmap *map)
1466 {
1467         return decode_pg_mapping(p, end, &map->pg_temp, __decode_pg_temp,
1468                                  true);
1469 }
1470
1471 static struct ceph_pg_mapping *__decode_primary_temp(void **p, void *end,
1472                                                      bool incremental)
1473 {
1474         struct ceph_pg_mapping *pg;
1475         u32 osd;
1476
1477         ceph_decode_32_safe(p, end, osd, e_inval);
1478         if (osd == (u32)-1 && incremental)
1479                 return NULL;    /* new_primary_temp: -1 to remove */
1480
1481         pg = alloc_pg_mapping(0);
1482         if (!pg)
1483                 return ERR_PTR(-ENOMEM);
1484
1485         pg->primary_temp.osd = osd;
1486         return pg;
1487
1488 e_inval:
1489         return ERR_PTR(-EINVAL);
1490 }
1491
1492 static int decode_primary_temp(void **p, void *end, struct ceph_osdmap *map)
1493 {
1494         return decode_pg_mapping(p, end, &map->primary_temp,
1495                                  __decode_primary_temp, false);
1496 }
1497
1498 static int decode_new_primary_temp(void **p, void *end,
1499                                    struct ceph_osdmap *map)
1500 {
1501         return decode_pg_mapping(p, end, &map->primary_temp,
1502                                  __decode_primary_temp, true);
1503 }
1504
1505 u32 ceph_get_primary_affinity(struct ceph_osdmap *map, int osd)
1506 {
1507         BUG_ON(osd >= map->max_osd);
1508
1509         if (!map->osd_primary_affinity)
1510                 return CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
1511
1512         return map->osd_primary_affinity[osd];
1513 }
1514
1515 static int set_primary_affinity(struct ceph_osdmap *map, int osd, u32 aff)
1516 {
1517         BUG_ON(osd >= map->max_osd);
1518
1519         if (!map->osd_primary_affinity) {
1520                 int i;
1521
1522                 map->osd_primary_affinity = kvmalloc(
1523                     array_size(map->max_osd, sizeof(*map->osd_primary_affinity)),
1524                     GFP_NOFS);
1525                 if (!map->osd_primary_affinity)
1526                         return -ENOMEM;
1527
1528                 for (i = 0; i < map->max_osd; i++)
1529                         map->osd_primary_affinity[i] =
1530                             CEPH_OSD_DEFAULT_PRIMARY_AFFINITY;
1531         }
1532
1533         map->osd_primary_affinity[osd] = aff;
1534
1535         return 0;
1536 }
1537
1538 static int decode_primary_affinity(void **p, void *end,
1539                                    struct ceph_osdmap *map)
1540 {
1541         u32 len, i;
1542
1543         ceph_decode_32_safe(p, end, len, e_inval);
1544         if (len == 0) {
1545                 kvfree(map->osd_primary_affinity);
1546                 map->osd_primary_affinity = NULL;
1547                 return 0;
1548         }
1549         if (len != map->max_osd)
1550                 goto e_inval;
1551
1552         ceph_decode_need(p, end, map->max_osd*sizeof(u32), e_inval);
1553
1554         for (i = 0; i < map->max_osd; i++) {
1555                 int ret;
1556
1557                 ret = set_primary_affinity(map, i, ceph_decode_32(p));
1558                 if (ret)
1559                         return ret;
1560         }
1561
1562         return 0;
1563
1564 e_inval:
1565         return -EINVAL;
1566 }
1567
1568 static int decode_new_primary_affinity(void **p, void *end,
1569                                        struct ceph_osdmap *map)
1570 {
1571         u32 n;
1572
1573         ceph_decode_32_safe(p, end, n, e_inval);
1574         while (n--) {
1575                 u32 osd, aff;
1576                 int ret;
1577
1578                 ceph_decode_32_safe(p, end, osd, e_inval);
1579                 ceph_decode_32_safe(p, end, aff, e_inval);
1580
1581                 ret = set_primary_affinity(map, osd, aff);
1582                 if (ret)
1583                         return ret;
1584
1585                 osdmap_info(map, "osd%d primary-affinity 0x%x\n", osd, aff);
1586         }
1587
1588         return 0;
1589
1590 e_inval:
1591         return -EINVAL;
1592 }
1593
1594 static struct ceph_pg_mapping *__decode_pg_upmap(void **p, void *end,
1595                                                  bool __unused)
1596 {
1597         return __decode_pg_temp(p, end, false);
1598 }
1599
1600 static int decode_pg_upmap(void **p, void *end, struct ceph_osdmap *map)
1601 {
1602         return decode_pg_mapping(p, end, &map->pg_upmap, __decode_pg_upmap,
1603                                  false);
1604 }
1605
1606 static int decode_new_pg_upmap(void **p, void *end, struct ceph_osdmap *map)
1607 {
1608         return decode_pg_mapping(p, end, &map->pg_upmap, __decode_pg_upmap,
1609                                  true);
1610 }
1611
1612 static int decode_old_pg_upmap(void **p, void *end, struct ceph_osdmap *map)
1613 {
1614         return decode_pg_mapping(p, end, &map->pg_upmap, NULL, true);
1615 }
1616
1617 static struct ceph_pg_mapping *__decode_pg_upmap_items(void **p, void *end,
1618                                                        bool __unused)
1619 {
1620         struct ceph_pg_mapping *pg;
1621         u32 len, i;
1622
1623         ceph_decode_32_safe(p, end, len, e_inval);
1624         if (len > (SIZE_MAX - sizeof(*pg)) / (2 * sizeof(u32)))
1625                 return ERR_PTR(-EINVAL);
1626
1627         ceph_decode_need(p, end, 2 * len * sizeof(u32), e_inval);
1628         pg = alloc_pg_mapping(2 * len * sizeof(u32));
1629         if (!pg)
1630                 return ERR_PTR(-ENOMEM);
1631
1632         pg->pg_upmap_items.len = len;
1633         for (i = 0; i < len; i++) {
1634                 pg->pg_upmap_items.from_to[i][0] = ceph_decode_32(p);
1635                 pg->pg_upmap_items.from_to[i][1] = ceph_decode_32(p);
1636         }
1637
1638         return pg;
1639
1640 e_inval:
1641         return ERR_PTR(-EINVAL);
1642 }
1643
1644 static int decode_pg_upmap_items(void **p, void *end, struct ceph_osdmap *map)
1645 {
1646         return decode_pg_mapping(p, end, &map->pg_upmap_items,
1647                                  __decode_pg_upmap_items, false);
1648 }
1649
1650 static int decode_new_pg_upmap_items(void **p, void *end,
1651                                      struct ceph_osdmap *map)
1652 {
1653         return decode_pg_mapping(p, end, &map->pg_upmap_items,
1654                                  __decode_pg_upmap_items, true);
1655 }
1656
1657 static int decode_old_pg_upmap_items(void **p, void *end,
1658                                      struct ceph_osdmap *map)
1659 {
1660         return decode_pg_mapping(p, end, &map->pg_upmap_items, NULL, true);
1661 }
1662
1663 /*
1664  * decode a full map.
1665  */
1666 static int osdmap_decode(void **p, void *end, bool msgr2,
1667                          struct ceph_osdmap *map)
1668 {
1669         u8 struct_v;
1670         u32 epoch = 0;
1671         void *start = *p;
1672         u32 max;
1673         u32 len, i;
1674         int err;
1675
1676         dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p));
1677
1678         err = get_osdmap_client_data_v(p, end, "full", &struct_v);
1679         if (err)
1680                 goto bad;
1681
1682         /* fsid, epoch, created, modified */
1683         ceph_decode_need(p, end, sizeof(map->fsid) + sizeof(u32) +
1684                          sizeof(map->created) + sizeof(map->modified), e_inval);
1685         ceph_decode_copy(p, &map->fsid, sizeof(map->fsid));
1686         epoch = map->epoch = ceph_decode_32(p);
1687         ceph_decode_copy(p, &map->created, sizeof(map->created));
1688         ceph_decode_copy(p, &map->modified, sizeof(map->modified));
1689
1690         /* pools */
1691         err = decode_pools(p, end, map);
1692         if (err)
1693                 goto bad;
1694
1695         /* pool_name */
1696         err = decode_pool_names(p, end, map);
1697         if (err)
1698                 goto bad;
1699
1700         ceph_decode_32_safe(p, end, map->pool_max, e_inval);
1701
1702         ceph_decode_32_safe(p, end, map->flags, e_inval);
1703
1704         /* max_osd */
1705         ceph_decode_32_safe(p, end, max, e_inval);
1706
1707         /* (re)alloc osd arrays */
1708         err = osdmap_set_max_osd(map, max);
1709         if (err)
1710                 goto bad;
1711
1712         /* osd_state, osd_weight, osd_addrs->client_addr */
1713         ceph_decode_need(p, end, 3*sizeof(u32) +
1714                          map->max_osd*(struct_v >= 5 ? sizeof(u32) :
1715                                                        sizeof(u8)) +
1716                                        sizeof(*map->osd_weight), e_inval);
1717         if (ceph_decode_32(p) != map->max_osd)
1718                 goto e_inval;
1719
1720         if (struct_v >= 5) {
1721                 for (i = 0; i < map->max_osd; i++)
1722                         map->osd_state[i] = ceph_decode_32(p);
1723         } else {
1724                 for (i = 0; i < map->max_osd; i++)
1725                         map->osd_state[i] = ceph_decode_8(p);
1726         }
1727
1728         if (ceph_decode_32(p) != map->max_osd)
1729                 goto e_inval;
1730
1731         for (i = 0; i < map->max_osd; i++)
1732                 map->osd_weight[i] = ceph_decode_32(p);
1733
1734         if (ceph_decode_32(p) != map->max_osd)
1735                 goto e_inval;
1736
1737         for (i = 0; i < map->max_osd; i++) {
1738                 struct ceph_entity_addr *addr = &map->osd_addr[i];
1739
1740                 if (struct_v >= 8)
1741                         err = ceph_decode_entity_addrvec(p, end, msgr2, addr);
1742                 else
1743                         err = ceph_decode_entity_addr(p, end, addr);
1744                 if (err)
1745                         goto bad;
1746
1747                 dout("%s osd%d addr %s\n", __func__, i, ceph_pr_addr(addr));
1748         }
1749
1750         /* pg_temp */
1751         err = decode_pg_temp(p, end, map);
1752         if (err)
1753                 goto bad;
1754
1755         /* primary_temp */
1756         if (struct_v >= 1) {
1757                 err = decode_primary_temp(p, end, map);
1758                 if (err)
1759                         goto bad;
1760         }
1761
1762         /* primary_affinity */
1763         if (struct_v >= 2) {
1764                 err = decode_primary_affinity(p, end, map);
1765                 if (err)
1766                         goto bad;
1767         } else {
1768                 WARN_ON(map->osd_primary_affinity);
1769         }
1770
1771         /* crush */
1772         ceph_decode_32_safe(p, end, len, e_inval);
1773         err = osdmap_set_crush(map, crush_decode(*p, min(*p + len, end)));
1774         if (err)
1775                 goto bad;
1776
1777         *p += len;
1778         if (struct_v >= 3) {
1779                 /* erasure_code_profiles */
1780                 ceph_decode_skip_map_of_map(p, end, string, string, string,
1781                                             e_inval);
1782         }
1783
1784         if (struct_v >= 4) {
1785                 err = decode_pg_upmap(p, end, map);
1786                 if (err)
1787                         goto bad;
1788
1789                 err = decode_pg_upmap_items(p, end, map);
1790                 if (err)
1791                         goto bad;
1792         } else {
1793                 WARN_ON(!RB_EMPTY_ROOT(&map->pg_upmap));
1794                 WARN_ON(!RB_EMPTY_ROOT(&map->pg_upmap_items));
1795         }
1796
1797         /* ignore the rest */
1798         *p = end;
1799
1800         dout("full osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd);
1801         return 0;
1802
1803 e_inval:
1804         err = -EINVAL;
1805 bad:
1806         pr_err("corrupt full osdmap (%d) epoch %d off %d (%p of %p-%p)\n",
1807                err, epoch, (int)(*p - start), *p, start, end);
1808         print_hex_dump(KERN_DEBUG, "osdmap: ",
1809                        DUMP_PREFIX_OFFSET, 16, 1,
1810                        start, end - start, true);
1811         return err;
1812 }
1813
1814 /*
1815  * Allocate and decode a full map.
1816  */
1817 struct ceph_osdmap *ceph_osdmap_decode(void **p, void *end, bool msgr2)
1818 {
1819         struct ceph_osdmap *map;
1820         int ret;
1821
1822         map = ceph_osdmap_alloc();
1823         if (!map)
1824                 return ERR_PTR(-ENOMEM);
1825
1826         ret = osdmap_decode(p, end, msgr2, map);
1827         if (ret) {
1828                 ceph_osdmap_destroy(map);
1829                 return ERR_PTR(ret);
1830         }
1831
1832         return map;
1833 }
1834
1835 /*
1836  * Encoding order is (new_up_client, new_state, new_weight).  Need to
1837  * apply in the (new_weight, new_state, new_up_client) order, because
1838  * an incremental map may look like e.g.
1839  *
1840  *     new_up_client: { osd=6, addr=... } # set osd_state and addr
1841  *     new_state: { osd=6, xorstate=EXISTS } # clear osd_state
1842  */
1843 static int decode_new_up_state_weight(void **p, void *end, u8 struct_v,
1844                                       bool msgr2, struct ceph_osdmap *map)
1845 {
1846         void *new_up_client;
1847         void *new_state;
1848         void *new_weight_end;
1849         u32 len;
1850         int ret;
1851         int i;
1852
1853         new_up_client = *p;
1854         ceph_decode_32_safe(p, end, len, e_inval);
1855         for (i = 0; i < len; ++i) {
1856                 struct ceph_entity_addr addr;
1857
1858                 ceph_decode_skip_32(p, end, e_inval);
1859                 if (struct_v >= 7)
1860                         ret = ceph_decode_entity_addrvec(p, end, msgr2, &addr);
1861                 else
1862                         ret = ceph_decode_entity_addr(p, end, &addr);
1863                 if (ret)
1864                         return ret;
1865         }
1866
1867         new_state = *p;
1868         ceph_decode_32_safe(p, end, len, e_inval);
1869         len *= sizeof(u32) + (struct_v >= 5 ? sizeof(u32) : sizeof(u8));
1870         ceph_decode_need(p, end, len, e_inval);
1871         *p += len;
1872
1873         /* new_weight */
1874         ceph_decode_32_safe(p, end, len, e_inval);
1875         while (len--) {
1876                 s32 osd;
1877                 u32 w;
1878
1879                 ceph_decode_need(p, end, 2*sizeof(u32), e_inval);
1880                 osd = ceph_decode_32(p);
1881                 w = ceph_decode_32(p);
1882                 BUG_ON(osd >= map->max_osd);
1883                 osdmap_info(map, "osd%d weight 0x%x %s\n", osd, w,
1884                             w == CEPH_OSD_IN ? "(in)" :
1885                             (w == CEPH_OSD_OUT ? "(out)" : ""));
1886                 map->osd_weight[osd] = w;
1887
1888                 /*
1889                  * If we are marking in, set the EXISTS, and clear the
1890                  * AUTOOUT and NEW bits.
1891                  */
1892                 if (w) {
1893                         map->osd_state[osd] |= CEPH_OSD_EXISTS;
1894                         map->osd_state[osd] &= ~(CEPH_OSD_AUTOOUT |
1895                                                  CEPH_OSD_NEW);
1896                 }
1897         }
1898         new_weight_end = *p;
1899
1900         /* new_state (up/down) */
1901         *p = new_state;
1902         len = ceph_decode_32(p);
1903         while (len--) {
1904                 s32 osd;
1905                 u32 xorstate;
1906
1907                 osd = ceph_decode_32(p);
1908                 if (struct_v >= 5)
1909                         xorstate = ceph_decode_32(p);
1910                 else
1911                         xorstate = ceph_decode_8(p);
1912                 if (xorstate == 0)
1913                         xorstate = CEPH_OSD_UP;
1914                 BUG_ON(osd >= map->max_osd);
1915                 if ((map->osd_state[osd] & CEPH_OSD_UP) &&
1916                     (xorstate & CEPH_OSD_UP))
1917                         osdmap_info(map, "osd%d down\n", osd);
1918                 if ((map->osd_state[osd] & CEPH_OSD_EXISTS) &&
1919                     (xorstate & CEPH_OSD_EXISTS)) {
1920                         osdmap_info(map, "osd%d does not exist\n", osd);
1921                         ret = set_primary_affinity(map, osd,
1922                                                    CEPH_OSD_DEFAULT_PRIMARY_AFFINITY);
1923                         if (ret)
1924                                 return ret;
1925                         memset(map->osd_addr + osd, 0, sizeof(*map->osd_addr));
1926                         map->osd_state[osd] = 0;
1927                 } else {
1928                         map->osd_state[osd] ^= xorstate;
1929                 }
1930         }
1931
1932         /* new_up_client */
1933         *p = new_up_client;
1934         len = ceph_decode_32(p);
1935         while (len--) {
1936                 s32 osd;
1937                 struct ceph_entity_addr addr;
1938
1939                 osd = ceph_decode_32(p);
1940                 BUG_ON(osd >= map->max_osd);
1941                 if (struct_v >= 7)
1942                         ret = ceph_decode_entity_addrvec(p, end, msgr2, &addr);
1943                 else
1944                         ret = ceph_decode_entity_addr(p, end, &addr);
1945                 if (ret)
1946                         return ret;
1947
1948                 dout("%s osd%d addr %s\n", __func__, osd, ceph_pr_addr(&addr));
1949
1950                 osdmap_info(map, "osd%d up\n", osd);
1951                 map->osd_state[osd] |= CEPH_OSD_EXISTS | CEPH_OSD_UP;
1952                 map->osd_addr[osd] = addr;
1953         }
1954
1955         *p = new_weight_end;
1956         return 0;
1957
1958 e_inval:
1959         return -EINVAL;
1960 }
1961
1962 /*
1963  * decode and apply an incremental map update.
1964  */
1965 struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end, bool msgr2,
1966                                              struct ceph_osdmap *map)
1967 {
1968         struct ceph_fsid fsid;
1969         u32 epoch = 0;
1970         struct ceph_timespec modified;
1971         s32 len;
1972         u64 pool;
1973         __s64 new_pool_max;
1974         __s32 new_flags, max;
1975         void *start = *p;
1976         int err;
1977         u8 struct_v;
1978
1979         dout("%s %p to %p len %d\n", __func__, *p, end, (int)(end - *p));
1980
1981         err = get_osdmap_client_data_v(p, end, "inc", &struct_v);
1982         if (err)
1983                 goto bad;
1984
1985         /* fsid, epoch, modified, new_pool_max, new_flags */
1986         ceph_decode_need(p, end, sizeof(fsid) + sizeof(u32) + sizeof(modified) +
1987                          sizeof(u64) + sizeof(u32), e_inval);
1988         ceph_decode_copy(p, &fsid, sizeof(fsid));
1989         epoch = ceph_decode_32(p);
1990         BUG_ON(epoch != map->epoch+1);
1991         ceph_decode_copy(p, &modified, sizeof(modified));
1992         new_pool_max = ceph_decode_64(p);
1993         new_flags = ceph_decode_32(p);
1994
1995         /* full map? */
1996         ceph_decode_32_safe(p, end, len, e_inval);
1997         if (len > 0) {
1998                 dout("apply_incremental full map len %d, %p to %p\n",
1999                      len, *p, end);
2000                 return ceph_osdmap_decode(p, min(*p+len, end), msgr2);
2001         }
2002
2003         /* new crush? */
2004         ceph_decode_32_safe(p, end, len, e_inval);
2005         if (len > 0) {
2006                 err = osdmap_set_crush(map,
2007                                        crush_decode(*p, min(*p + len, end)));
2008                 if (err)
2009                         goto bad;
2010                 *p += len;
2011         }
2012
2013         /* new flags? */
2014         if (new_flags >= 0)
2015                 map->flags = new_flags;
2016         if (new_pool_max >= 0)
2017                 map->pool_max = new_pool_max;
2018
2019         /* new max? */
2020         ceph_decode_32_safe(p, end, max, e_inval);
2021         if (max >= 0) {
2022                 err = osdmap_set_max_osd(map, max);
2023                 if (err)
2024                         goto bad;
2025         }
2026
2027         map->epoch++;
2028         map->modified = modified;
2029
2030         /* new_pools */
2031         err = decode_new_pools(p, end, map);
2032         if (err)
2033                 goto bad;
2034
2035         /* new_pool_names */
2036         err = decode_pool_names(p, end, map);
2037         if (err)
2038                 goto bad;
2039
2040         /* old_pool */
2041         ceph_decode_32_safe(p, end, len, e_inval);
2042         while (len--) {
2043                 struct ceph_pg_pool_info *pi;
2044
2045                 ceph_decode_64_safe(p, end, pool, e_inval);
2046                 pi = lookup_pg_pool(&map->pg_pools, pool);
2047                 if (pi)
2048                         __remove_pg_pool(&map->pg_pools, pi);
2049         }
2050
2051         /* new_up_client, new_state, new_weight */
2052         err = decode_new_up_state_weight(p, end, struct_v, msgr2, map);
2053         if (err)
2054                 goto bad;
2055
2056         /* new_pg_temp */
2057         err = decode_new_pg_temp(p, end, map);
2058         if (err)
2059                 goto bad;
2060
2061         /* new_primary_temp */
2062         if (struct_v >= 1) {
2063                 err = decode_new_primary_temp(p, end, map);
2064                 if (err)
2065                         goto bad;
2066         }
2067
2068         /* new_primary_affinity */
2069         if (struct_v >= 2) {
2070                 err = decode_new_primary_affinity(p, end, map);
2071                 if (err)
2072                         goto bad;
2073         }
2074
2075         if (struct_v >= 3) {
2076                 /* new_erasure_code_profiles */
2077                 ceph_decode_skip_map_of_map(p, end, string, string, string,
2078                                             e_inval);
2079                 /* old_erasure_code_profiles */
2080                 ceph_decode_skip_set(p, end, string, e_inval);
2081         }
2082
2083         if (struct_v >= 4) {
2084                 err = decode_new_pg_upmap(p, end, map);
2085                 if (err)
2086                         goto bad;
2087
2088                 err = decode_old_pg_upmap(p, end, map);
2089                 if (err)
2090                         goto bad;
2091
2092                 err = decode_new_pg_upmap_items(p, end, map);
2093                 if (err)
2094                         goto bad;
2095
2096                 err = decode_old_pg_upmap_items(p, end, map);
2097                 if (err)
2098                         goto bad;
2099         }
2100
2101         /* ignore the rest */
2102         *p = end;
2103
2104         dout("inc osdmap epoch %d max_osd %d\n", map->epoch, map->max_osd);
2105         return map;
2106
2107 e_inval:
2108         err = -EINVAL;
2109 bad:
2110         pr_err("corrupt inc osdmap (%d) epoch %d off %d (%p of %p-%p)\n",
2111                err, epoch, (int)(*p - start), *p, start, end);
2112         print_hex_dump(KERN_DEBUG, "osdmap: ",
2113                        DUMP_PREFIX_OFFSET, 16, 1,
2114                        start, end - start, true);
2115         return ERR_PTR(err);
2116 }
2117
2118 void ceph_oloc_copy(struct ceph_object_locator *dest,
2119                     const struct ceph_object_locator *src)
2120 {
2121         ceph_oloc_destroy(dest);
2122
2123         dest->pool = src->pool;
2124         if (src->pool_ns)
2125                 dest->pool_ns = ceph_get_string(src->pool_ns);
2126         else
2127                 dest->pool_ns = NULL;
2128 }
2129 EXPORT_SYMBOL(ceph_oloc_copy);
2130
2131 void ceph_oloc_destroy(struct ceph_object_locator *oloc)
2132 {
2133         ceph_put_string(oloc->pool_ns);
2134 }
2135 EXPORT_SYMBOL(ceph_oloc_destroy);
2136
2137 void ceph_oid_copy(struct ceph_object_id *dest,
2138                    const struct ceph_object_id *src)
2139 {
2140         ceph_oid_destroy(dest);
2141
2142         if (src->name != src->inline_name) {
2143                 /* very rare, see ceph_object_id definition */
2144                 dest->name = kmalloc(src->name_len + 1,
2145                                      GFP_NOIO | __GFP_NOFAIL);
2146         } else {
2147                 dest->name = dest->inline_name;
2148         }
2149         memcpy(dest->name, src->name, src->name_len + 1);
2150         dest->name_len = src->name_len;
2151 }
2152 EXPORT_SYMBOL(ceph_oid_copy);
2153
2154 static __printf(2, 0)
2155 int oid_printf_vargs(struct ceph_object_id *oid, const char *fmt, va_list ap)
2156 {
2157         int len;
2158
2159         WARN_ON(!ceph_oid_empty(oid));
2160
2161         len = vsnprintf(oid->inline_name, sizeof(oid->inline_name), fmt, ap);
2162         if (len >= sizeof(oid->inline_name))
2163                 return len;
2164
2165         oid->name_len = len;
2166         return 0;
2167 }
2168
2169 /*
2170  * If oid doesn't fit into inline buffer, BUG.
2171  */
2172 void ceph_oid_printf(struct ceph_object_id *oid, const char *fmt, ...)
2173 {
2174         va_list ap;
2175
2176         va_start(ap, fmt);
2177         BUG_ON(oid_printf_vargs(oid, fmt, ap));
2178         va_end(ap);
2179 }
2180 EXPORT_SYMBOL(ceph_oid_printf);
2181
2182 static __printf(3, 0)
2183 int oid_aprintf_vargs(struct ceph_object_id *oid, gfp_t gfp,
2184                       const char *fmt, va_list ap)
2185 {
2186         va_list aq;
2187         int len;
2188
2189         va_copy(aq, ap);
2190         len = oid_printf_vargs(oid, fmt, aq);
2191         va_end(aq);
2192
2193         if (len) {
2194                 char *external_name;
2195
2196                 external_name = kmalloc(len + 1, gfp);
2197                 if (!external_name)
2198                         return -ENOMEM;
2199
2200                 oid->name = external_name;
2201                 WARN_ON(vsnprintf(oid->name, len + 1, fmt, ap) != len);
2202                 oid->name_len = len;
2203         }
2204
2205         return 0;
2206 }
2207
2208 /*
2209  * If oid doesn't fit into inline buffer, allocate.
2210  */
2211 int ceph_oid_aprintf(struct ceph_object_id *oid, gfp_t gfp,
2212                      const char *fmt, ...)
2213 {
2214         va_list ap;
2215         int ret;
2216
2217         va_start(ap, fmt);
2218         ret = oid_aprintf_vargs(oid, gfp, fmt, ap);
2219         va_end(ap);
2220
2221         return ret;
2222 }
2223 EXPORT_SYMBOL(ceph_oid_aprintf);
2224
2225 void ceph_oid_destroy(struct ceph_object_id *oid)
2226 {
2227         if (oid->name != oid->inline_name)
2228                 kfree(oid->name);
2229 }
2230 EXPORT_SYMBOL(ceph_oid_destroy);
2231
2232 /*
2233  * osds only
2234  */
2235 static bool __osds_equal(const struct ceph_osds *lhs,
2236                          const struct ceph_osds *rhs)
2237 {
2238         if (lhs->size == rhs->size &&
2239             !memcmp(lhs->osds, rhs->osds, rhs->size * sizeof(rhs->osds[0])))
2240                 return true;
2241
2242         return false;
2243 }
2244
2245 /*
2246  * osds + primary
2247  */
2248 static bool osds_equal(const struct ceph_osds *lhs,
2249                        const struct ceph_osds *rhs)
2250 {
2251         if (__osds_equal(lhs, rhs) &&
2252             lhs->primary == rhs->primary)
2253                 return true;
2254
2255         return false;
2256 }
2257
2258 static bool osds_valid(const struct ceph_osds *set)
2259 {
2260         /* non-empty set */
2261         if (set->size > 0 && set->primary >= 0)
2262                 return true;
2263
2264         /* empty can_shift_osds set */
2265         if (!set->size && set->primary == -1)
2266                 return true;
2267
2268         /* empty !can_shift_osds set - all NONE */
2269         if (set->size > 0 && set->primary == -1) {
2270                 int i;
2271
2272                 for (i = 0; i < set->size; i++) {
2273                         if (set->osds[i] != CRUSH_ITEM_NONE)
2274                                 break;
2275                 }
2276                 if (i == set->size)
2277                         return true;
2278         }
2279
2280         return false;
2281 }
2282
2283 void ceph_osds_copy(struct ceph_osds *dest, const struct ceph_osds *src)
2284 {
2285         memcpy(dest->osds, src->osds, src->size * sizeof(src->osds[0]));
2286         dest->size = src->size;
2287         dest->primary = src->primary;
2288 }
2289
2290 bool ceph_pg_is_split(const struct ceph_pg *pgid, u32 old_pg_num,
2291                       u32 new_pg_num)
2292 {
2293         int old_bits = calc_bits_of(old_pg_num);
2294         int old_mask = (1 << old_bits) - 1;
2295         int n;
2296
2297         WARN_ON(pgid->seed >= old_pg_num);
2298         if (new_pg_num <= old_pg_num)
2299                 return false;
2300
2301         for (n = 1; ; n++) {
2302                 int next_bit = n << (old_bits - 1);
2303                 u32 s = next_bit | pgid->seed;
2304
2305                 if (s < old_pg_num || s == pgid->seed)
2306                         continue;
2307                 if (s >= new_pg_num)
2308                         break;
2309
2310                 s = ceph_stable_mod(s, old_pg_num, old_mask);
2311                 if (s == pgid->seed)
2312                         return true;
2313         }
2314
2315         return false;
2316 }
2317
2318 bool ceph_is_new_interval(const struct ceph_osds *old_acting,
2319                           const struct ceph_osds *new_acting,
2320                           const struct ceph_osds *old_up,
2321                           const struct ceph_osds *new_up,
2322                           int old_size,
2323                           int new_size,
2324                           int old_min_size,
2325                           int new_min_size,
2326                           u32 old_pg_num,
2327                           u32 new_pg_num,
2328                           bool old_sort_bitwise,
2329                           bool new_sort_bitwise,
2330                           bool old_recovery_deletes,
2331                           bool new_recovery_deletes,
2332                           const struct ceph_pg *pgid)
2333 {
2334         return !osds_equal(old_acting, new_acting) ||
2335                !osds_equal(old_up, new_up) ||
2336                old_size != new_size ||
2337                old_min_size != new_min_size ||
2338                ceph_pg_is_split(pgid, old_pg_num, new_pg_num) ||
2339                old_sort_bitwise != new_sort_bitwise ||
2340                old_recovery_deletes != new_recovery_deletes;
2341 }
2342
2343 static int calc_pg_rank(int osd, const struct ceph_osds *acting)
2344 {
2345         int i;
2346
2347         for (i = 0; i < acting->size; i++) {
2348                 if (acting->osds[i] == osd)
2349                         return i;
2350         }
2351
2352         return -1;
2353 }
2354
2355 static bool primary_changed(const struct ceph_osds *old_acting,
2356                             const struct ceph_osds *new_acting)
2357 {
2358         if (!old_acting->size && !new_acting->size)
2359                 return false; /* both still empty */
2360
2361         if (!old_acting->size ^ !new_acting->size)
2362                 return true; /* was empty, now not, or vice versa */
2363
2364         if (old_acting->primary != new_acting->primary)
2365                 return true; /* primary changed */
2366
2367         if (calc_pg_rank(old_acting->primary, old_acting) !=
2368             calc_pg_rank(new_acting->primary, new_acting))
2369                 return true;
2370
2371         return false; /* same primary (tho replicas may have changed) */
2372 }
2373
2374 bool ceph_osds_changed(const struct ceph_osds *old_acting,
2375                        const struct ceph_osds *new_acting,
2376                        bool any_change)
2377 {
2378         if (primary_changed(old_acting, new_acting))
2379                 return true;
2380
2381         if (any_change && !__osds_equal(old_acting, new_acting))
2382                 return true;
2383
2384         return false;
2385 }
2386
2387 /*
2388  * Map an object into a PG.
2389  *
2390  * Should only be called with target_oid and target_oloc (as opposed to
2391  * base_oid and base_oloc), since tiering isn't taken into account.
2392  */
2393 void __ceph_object_locator_to_pg(struct ceph_pg_pool_info *pi,
2394                                  const struct ceph_object_id *oid,
2395                                  const struct ceph_object_locator *oloc,
2396                                  struct ceph_pg *raw_pgid)
2397 {
2398         WARN_ON(pi->id != oloc->pool);
2399
2400         if (!oloc->pool_ns) {
2401                 raw_pgid->pool = oloc->pool;
2402                 raw_pgid->seed = ceph_str_hash(pi->object_hash, oid->name,
2403                                              oid->name_len);
2404                 dout("%s %s -> raw_pgid %llu.%x\n", __func__, oid->name,
2405                      raw_pgid->pool, raw_pgid->seed);
2406         } else {
2407                 char stack_buf[256];
2408                 char *buf = stack_buf;
2409                 int nsl = oloc->pool_ns->len;
2410                 size_t total = nsl + 1 + oid->name_len;
2411
2412                 if (total > sizeof(stack_buf))
2413                         buf = kmalloc(total, GFP_NOIO | __GFP_NOFAIL);
2414                 memcpy(buf, oloc->pool_ns->str, nsl);
2415                 buf[nsl] = '\037';
2416                 memcpy(buf + nsl + 1, oid->name, oid->name_len);
2417                 raw_pgid->pool = oloc->pool;
2418                 raw_pgid->seed = ceph_str_hash(pi->object_hash, buf, total);
2419                 if (buf != stack_buf)
2420                         kfree(buf);
2421                 dout("%s %s ns %.*s -> raw_pgid %llu.%x\n", __func__,
2422                      oid->name, nsl, oloc->pool_ns->str,
2423                      raw_pgid->pool, raw_pgid->seed);
2424         }
2425 }
2426
2427 int ceph_object_locator_to_pg(struct ceph_osdmap *osdmap,
2428                               const struct ceph_object_id *oid,
2429                               const struct ceph_object_locator *oloc,
2430                               struct ceph_pg *raw_pgid)
2431 {
2432         struct ceph_pg_pool_info *pi;
2433
2434         pi = ceph_pg_pool_by_id(osdmap, oloc->pool);
2435         if (!pi)
2436                 return -ENOENT;
2437
2438         __ceph_object_locator_to_pg(pi, oid, oloc, raw_pgid);
2439         return 0;
2440 }
2441 EXPORT_SYMBOL(ceph_object_locator_to_pg);
2442
2443 /*
2444  * Map a raw PG (full precision ps) into an actual PG.
2445  */
2446 static void raw_pg_to_pg(struct ceph_pg_pool_info *pi,
2447                          const struct ceph_pg *raw_pgid,
2448                          struct ceph_pg *pgid)
2449 {
2450         pgid->pool = raw_pgid->pool;
2451         pgid->seed = ceph_stable_mod(raw_pgid->seed, pi->pg_num,
2452                                      pi->pg_num_mask);
2453 }
2454
2455 /*
2456  * Map a raw PG (full precision ps) into a placement ps (placement
2457  * seed).  Include pool id in that value so that different pools don't
2458  * use the same seeds.
2459  */
2460 static u32 raw_pg_to_pps(struct ceph_pg_pool_info *pi,
2461                          const struct ceph_pg *raw_pgid)
2462 {
2463         if (pi->flags & CEPH_POOL_FLAG_HASHPSPOOL) {
2464                 /* hash pool id and seed so that pool PGs do not overlap */
2465                 return crush_hash32_2(CRUSH_HASH_RJENKINS1,
2466                                       ceph_stable_mod(raw_pgid->seed,
2467                                                       pi->pgp_num,
2468                                                       pi->pgp_num_mask),
2469                                       raw_pgid->pool);
2470         } else {
2471                 /*
2472                  * legacy behavior: add ps and pool together.  this is
2473                  * not a great approach because the PGs from each pool
2474                  * will overlap on top of each other: 0.5 == 1.4 ==
2475                  * 2.3 == ...
2476                  */
2477                 return ceph_stable_mod(raw_pgid->seed, pi->pgp_num,
2478                                        pi->pgp_num_mask) +
2479                        (unsigned)raw_pgid->pool;
2480         }
2481 }
2482
2483 /*
2484  * Magic value used for a "default" fallback choose_args, used if the
2485  * crush_choose_arg_map passed to do_crush() does not exist.  If this
2486  * also doesn't exist, fall back to canonical weights.
2487  */
2488 #define CEPH_DEFAULT_CHOOSE_ARGS        -1
2489
2490 static int do_crush(struct ceph_osdmap *map, int ruleno, int x,
2491                     int *result, int result_max,
2492                     const __u32 *weight, int weight_max,
2493                     s64 choose_args_index)
2494 {
2495         struct crush_choose_arg_map *arg_map;
2496         struct crush_work *work;
2497         int r;
2498
2499         BUG_ON(result_max > CEPH_PG_MAX_SIZE);
2500
2501         arg_map = lookup_choose_arg_map(&map->crush->choose_args,
2502                                         choose_args_index);
2503         if (!arg_map)
2504                 arg_map = lookup_choose_arg_map(&map->crush->choose_args,
2505                                                 CEPH_DEFAULT_CHOOSE_ARGS);
2506
2507         work = get_workspace(&map->crush_wsm, map->crush);
2508         r = crush_do_rule(map->crush, ruleno, x, result, result_max,
2509                           weight, weight_max, work,
2510                           arg_map ? arg_map->args : NULL);
2511         put_workspace(&map->crush_wsm, work);
2512         return r;
2513 }
2514
2515 static void remove_nonexistent_osds(struct ceph_osdmap *osdmap,
2516                                     struct ceph_pg_pool_info *pi,
2517                                     struct ceph_osds *set)
2518 {
2519         int i;
2520
2521         if (ceph_can_shift_osds(pi)) {
2522                 int removed = 0;
2523
2524                 /* shift left */
2525                 for (i = 0; i < set->size; i++) {
2526                         if (!ceph_osd_exists(osdmap, set->osds[i])) {
2527                                 removed++;
2528                                 continue;
2529                         }
2530                         if (removed)
2531                                 set->osds[i - removed] = set->osds[i];
2532                 }
2533                 set->size -= removed;
2534         } else {
2535                 /* set dne devices to NONE */
2536                 for (i = 0; i < set->size; i++) {
2537                         if (!ceph_osd_exists(osdmap, set->osds[i]))
2538                                 set->osds[i] = CRUSH_ITEM_NONE;
2539                 }
2540         }
2541 }
2542
2543 /*
2544  * Calculate raw set (CRUSH output) for given PG and filter out
2545  * nonexistent OSDs.  ->primary is undefined for a raw set.
2546  *
2547  * Placement seed (CRUSH input) is returned through @ppps.
2548  */
2549 static void pg_to_raw_osds(struct ceph_osdmap *osdmap,
2550                            struct ceph_pg_pool_info *pi,
2551                            const struct ceph_pg *raw_pgid,
2552                            struct ceph_osds *raw,
2553                            u32 *ppps)
2554 {
2555         u32 pps = raw_pg_to_pps(pi, raw_pgid);
2556         int ruleno;
2557         int len;
2558
2559         ceph_osds_init(raw);
2560         if (ppps)
2561                 *ppps = pps;
2562
2563         ruleno = crush_find_rule(osdmap->crush, pi->crush_ruleset, pi->type,
2564                                  pi->size);
2565         if (ruleno < 0) {
2566                 pr_err("no crush rule: pool %lld ruleset %d type %d size %d\n",
2567                        pi->id, pi->crush_ruleset, pi->type, pi->size);
2568                 return;
2569         }
2570
2571         if (pi->size > ARRAY_SIZE(raw->osds)) {
2572                 pr_err_ratelimited("pool %lld ruleset %d type %d too wide: size %d > %zu\n",
2573                        pi->id, pi->crush_ruleset, pi->type, pi->size,
2574                        ARRAY_SIZE(raw->osds));
2575                 return;
2576         }
2577
2578         len = do_crush(osdmap, ruleno, pps, raw->osds, pi->size,
2579                        osdmap->osd_weight, osdmap->max_osd, pi->id);
2580         if (len < 0) {
2581                 pr_err("error %d from crush rule %d: pool %lld ruleset %d type %d size %d\n",
2582                        len, ruleno, pi->id, pi->crush_ruleset, pi->type,
2583                        pi->size);
2584                 return;
2585         }
2586
2587         raw->size = len;
2588         remove_nonexistent_osds(osdmap, pi, raw);
2589 }
2590
2591 /* apply pg_upmap[_items] mappings */
2592 static void apply_upmap(struct ceph_osdmap *osdmap,
2593                         const struct ceph_pg *pgid,
2594                         struct ceph_osds *raw)
2595 {
2596         struct ceph_pg_mapping *pg;
2597         int i, j;
2598
2599         pg = lookup_pg_mapping(&osdmap->pg_upmap, pgid);
2600         if (pg) {
2601                 /* make sure targets aren't marked out */
2602                 for (i = 0; i < pg->pg_upmap.len; i++) {
2603                         int osd = pg->pg_upmap.osds[i];
2604
2605                         if (osd != CRUSH_ITEM_NONE &&
2606                             osd < osdmap->max_osd &&
2607                             osdmap->osd_weight[osd] == 0) {
2608                                 /* reject/ignore explicit mapping */
2609                                 return;
2610                         }
2611                 }
2612                 for (i = 0; i < pg->pg_upmap.len; i++)
2613                         raw->osds[i] = pg->pg_upmap.osds[i];
2614                 raw->size = pg->pg_upmap.len;
2615                 /* check and apply pg_upmap_items, if any */
2616         }
2617
2618         pg = lookup_pg_mapping(&osdmap->pg_upmap_items, pgid);
2619         if (pg) {
2620                 /*
2621                  * Note: this approach does not allow a bidirectional swap,
2622                  * e.g., [[1,2],[2,1]] applied to [0,1,2] -> [0,2,1].
2623                  */
2624                 for (i = 0; i < pg->pg_upmap_items.len; i++) {
2625                         int from = pg->pg_upmap_items.from_to[i][0];
2626                         int to = pg->pg_upmap_items.from_to[i][1];
2627                         int pos = -1;
2628                         bool exists = false;
2629
2630                         /* make sure replacement doesn't already appear */
2631                         for (j = 0; j < raw->size; j++) {
2632                                 int osd = raw->osds[j];
2633
2634                                 if (osd == to) {
2635                                         exists = true;
2636                                         break;
2637                                 }
2638                                 /* ignore mapping if target is marked out */
2639                                 if (osd == from && pos < 0 &&
2640                                     !(to != CRUSH_ITEM_NONE &&
2641                                       to < osdmap->max_osd &&
2642                                       osdmap->osd_weight[to] == 0)) {
2643                                         pos = j;
2644                                 }
2645                         }
2646                         if (!exists && pos >= 0)
2647                                 raw->osds[pos] = to;
2648                 }
2649         }
2650 }
2651
2652 /*
2653  * Given raw set, calculate up set and up primary.  By definition of an
2654  * up set, the result won't contain nonexistent or down OSDs.
2655  *
2656  * This is done in-place - on return @set is the up set.  If it's
2657  * empty, ->primary will remain undefined.
2658  */
2659 static void raw_to_up_osds(struct ceph_osdmap *osdmap,
2660                            struct ceph_pg_pool_info *pi,
2661                            struct ceph_osds *set)
2662 {
2663         int i;
2664
2665         /* ->primary is undefined for a raw set */
2666         BUG_ON(set->primary != -1);
2667
2668         if (ceph_can_shift_osds(pi)) {
2669                 int removed = 0;
2670
2671                 /* shift left */
2672                 for (i = 0; i < set->size; i++) {
2673                         if (ceph_osd_is_down(osdmap, set->osds[i])) {
2674                                 removed++;
2675                                 continue;
2676                         }
2677                         if (removed)
2678                                 set->osds[i - removed] = set->osds[i];
2679                 }
2680                 set->size -= removed;
2681                 if (set->size > 0)
2682                         set->primary = set->osds[0];
2683         } else {
2684                 /* set down/dne devices to NONE */
2685                 for (i = set->size - 1; i >= 0; i--) {
2686                         if (ceph_osd_is_down(osdmap, set->osds[i]))
2687                                 set->osds[i] = CRUSH_ITEM_NONE;
2688                         else
2689                                 set->primary = set->osds[i];
2690                 }
2691         }
2692 }
2693
2694 static void apply_primary_affinity(struct ceph_osdmap *osdmap,
2695                                    struct ceph_pg_pool_info *pi,
2696                                    u32 pps,
2697                                    struct ceph_osds *up)
2698 {
2699         int i;
2700         int pos = -1;
2701
2702         /*
2703          * Do we have any non-default primary_affinity values for these
2704          * osds?
2705          */
2706         if (!osdmap->osd_primary_affinity)
2707                 return;
2708
2709         for (i = 0; i < up->size; i++) {
2710                 int osd = up->osds[i];
2711
2712                 if (osd != CRUSH_ITEM_NONE &&
2713                     osdmap->osd_primary_affinity[osd] !=
2714                                         CEPH_OSD_DEFAULT_PRIMARY_AFFINITY) {
2715                         break;
2716                 }
2717         }
2718         if (i == up->size)
2719                 return;
2720
2721         /*
2722          * Pick the primary.  Feed both the seed (for the pg) and the
2723          * osd into the hash/rng so that a proportional fraction of an
2724          * osd's pgs get rejected as primary.
2725          */
2726         for (i = 0; i < up->size; i++) {
2727                 int osd = up->osds[i];
2728                 u32 aff;
2729
2730                 if (osd == CRUSH_ITEM_NONE)
2731                         continue;
2732
2733                 aff = osdmap->osd_primary_affinity[osd];
2734                 if (aff < CEPH_OSD_MAX_PRIMARY_AFFINITY &&
2735                     (crush_hash32_2(CRUSH_HASH_RJENKINS1,
2736                                     pps, osd) >> 16) >= aff) {
2737                         /*
2738                          * We chose not to use this primary.  Note it
2739                          * anyway as a fallback in case we don't pick
2740                          * anyone else, but keep looking.
2741                          */
2742                         if (pos < 0)
2743                                 pos = i;
2744                 } else {
2745                         pos = i;
2746                         break;
2747                 }
2748         }
2749         if (pos < 0)
2750                 return;
2751
2752         up->primary = up->osds[pos];
2753
2754         if (ceph_can_shift_osds(pi) && pos > 0) {
2755                 /* move the new primary to the front */
2756                 for (i = pos; i > 0; i--)
2757                         up->osds[i] = up->osds[i - 1];
2758                 up->osds[0] = up->primary;
2759         }
2760 }
2761
2762 /*
2763  * Get pg_temp and primary_temp mappings for given PG.
2764  *
2765  * Note that a PG may have none, only pg_temp, only primary_temp or
2766  * both pg_temp and primary_temp mappings.  This means @temp isn't
2767  * always a valid OSD set on return: in the "only primary_temp" case,
2768  * @temp will have its ->primary >= 0 but ->size == 0.
2769  */
2770 static void get_temp_osds(struct ceph_osdmap *osdmap,
2771                           struct ceph_pg_pool_info *pi,
2772                           const struct ceph_pg *pgid,
2773                           struct ceph_osds *temp)
2774 {
2775         struct ceph_pg_mapping *pg;
2776         int i;
2777
2778         ceph_osds_init(temp);
2779
2780         /* pg_temp? */
2781         pg = lookup_pg_mapping(&osdmap->pg_temp, pgid);
2782         if (pg) {
2783                 for (i = 0; i < pg->pg_temp.len; i++) {
2784                         if (ceph_osd_is_down(osdmap, pg->pg_temp.osds[i])) {
2785                                 if (ceph_can_shift_osds(pi))
2786                                         continue;
2787
2788                                 temp->osds[temp->size++] = CRUSH_ITEM_NONE;
2789                         } else {
2790                                 temp->osds[temp->size++] = pg->pg_temp.osds[i];
2791                         }
2792                 }
2793
2794                 /* apply pg_temp's primary */
2795                 for (i = 0; i < temp->size; i++) {
2796                         if (temp->osds[i] != CRUSH_ITEM_NONE) {
2797                                 temp->primary = temp->osds[i];
2798                                 break;
2799                         }
2800                 }
2801         }
2802
2803         /* primary_temp? */
2804         pg = lookup_pg_mapping(&osdmap->primary_temp, pgid);
2805         if (pg)
2806                 temp->primary = pg->primary_temp.osd;
2807 }
2808
2809 /*
2810  * Map a PG to its acting set as well as its up set.
2811  *
2812  * Acting set is used for data mapping purposes, while up set can be
2813  * recorded for detecting interval changes and deciding whether to
2814  * resend a request.
2815  */
2816 void ceph_pg_to_up_acting_osds(struct ceph_osdmap *osdmap,
2817                                struct ceph_pg_pool_info *pi,
2818                                const struct ceph_pg *raw_pgid,
2819                                struct ceph_osds *up,
2820                                struct ceph_osds *acting)
2821 {
2822         struct ceph_pg pgid;
2823         u32 pps;
2824
2825         WARN_ON(pi->id != raw_pgid->pool);
2826         raw_pg_to_pg(pi, raw_pgid, &pgid);
2827
2828         pg_to_raw_osds(osdmap, pi, raw_pgid, up, &pps);
2829         apply_upmap(osdmap, &pgid, up);
2830         raw_to_up_osds(osdmap, pi, up);
2831         apply_primary_affinity(osdmap, pi, pps, up);
2832         get_temp_osds(osdmap, pi, &pgid, acting);
2833         if (!acting->size) {
2834                 memcpy(acting->osds, up->osds, up->size * sizeof(up->osds[0]));
2835                 acting->size = up->size;
2836                 if (acting->primary == -1)
2837                         acting->primary = up->primary;
2838         }
2839         WARN_ON(!osds_valid(up) || !osds_valid(acting));
2840 }
2841
2842 bool ceph_pg_to_primary_shard(struct ceph_osdmap *osdmap,
2843                               struct ceph_pg_pool_info *pi,
2844                               const struct ceph_pg *raw_pgid,
2845                               struct ceph_spg *spgid)
2846 {
2847         struct ceph_pg pgid;
2848         struct ceph_osds up, acting;
2849         int i;
2850
2851         WARN_ON(pi->id != raw_pgid->pool);
2852         raw_pg_to_pg(pi, raw_pgid, &pgid);
2853
2854         if (ceph_can_shift_osds(pi)) {
2855                 spgid->pgid = pgid; /* struct */
2856                 spgid->shard = CEPH_SPG_NOSHARD;
2857                 return true;
2858         }
2859
2860         ceph_pg_to_up_acting_osds(osdmap, pi, &pgid, &up, &acting);
2861         for (i = 0; i < acting.size; i++) {
2862                 if (acting.osds[i] == acting.primary) {
2863                         spgid->pgid = pgid; /* struct */
2864                         spgid->shard = i;
2865                         return true;
2866                 }
2867         }
2868
2869         return false;
2870 }
2871
2872 /*
2873  * Return acting primary for given PG, or -1 if none.
2874  */
2875 int ceph_pg_to_acting_primary(struct ceph_osdmap *osdmap,
2876                               const struct ceph_pg *raw_pgid)
2877 {
2878         struct ceph_pg_pool_info *pi;
2879         struct ceph_osds up, acting;
2880
2881         pi = ceph_pg_pool_by_id(osdmap, raw_pgid->pool);
2882         if (!pi)
2883                 return -1;
2884
2885         ceph_pg_to_up_acting_osds(osdmap, pi, raw_pgid, &up, &acting);
2886         return acting.primary;
2887 }
2888 EXPORT_SYMBOL(ceph_pg_to_acting_primary);
2889
2890 static struct crush_loc_node *alloc_crush_loc(size_t type_name_len,
2891                                               size_t name_len)
2892 {
2893         struct crush_loc_node *loc;
2894
2895         loc = kmalloc(sizeof(*loc) + type_name_len + name_len + 2, GFP_NOIO);
2896         if (!loc)
2897                 return NULL;
2898
2899         RB_CLEAR_NODE(&loc->cl_node);
2900         return loc;
2901 }
2902
2903 static void free_crush_loc(struct crush_loc_node *loc)
2904 {
2905         WARN_ON(!RB_EMPTY_NODE(&loc->cl_node));
2906
2907         kfree(loc);
2908 }
2909
2910 static int crush_loc_compare(const struct crush_loc *loc1,
2911                              const struct crush_loc *loc2)
2912 {
2913         return strcmp(loc1->cl_type_name, loc2->cl_type_name) ?:
2914                strcmp(loc1->cl_name, loc2->cl_name);
2915 }
2916
2917 DEFINE_RB_FUNCS2(crush_loc, struct crush_loc_node, cl_loc, crush_loc_compare,
2918                  RB_BYPTR, const struct crush_loc *, cl_node)
2919
2920 /*
2921  * Parses a set of <bucket type name>':'<bucket name> pairs separated
2922  * by '|', e.g. "rack:foo1|rack:foo2|datacenter:bar".
2923  *
2924  * Note that @crush_location is modified by strsep().
2925  */
2926 int ceph_parse_crush_location(char *crush_location, struct rb_root *locs)
2927 {
2928         struct crush_loc_node *loc;
2929         const char *type_name, *name, *colon;
2930         size_t type_name_len, name_len;
2931
2932         dout("%s '%s'\n", __func__, crush_location);
2933         while ((type_name = strsep(&crush_location, "|"))) {
2934                 colon = strchr(type_name, ':');
2935                 if (!colon)
2936                         return -EINVAL;
2937
2938                 type_name_len = colon - type_name;
2939                 if (type_name_len == 0)
2940                         return -EINVAL;
2941
2942                 name = colon + 1;
2943                 name_len = strlen(name);
2944                 if (name_len == 0)
2945                         return -EINVAL;
2946
2947                 loc = alloc_crush_loc(type_name_len, name_len);
2948                 if (!loc)
2949                         return -ENOMEM;
2950
2951                 loc->cl_loc.cl_type_name = loc->cl_data;
2952                 memcpy(loc->cl_loc.cl_type_name, type_name, type_name_len);
2953                 loc->cl_loc.cl_type_name[type_name_len] = '\0';
2954
2955                 loc->cl_loc.cl_name = loc->cl_data + type_name_len + 1;
2956                 memcpy(loc->cl_loc.cl_name, name, name_len);
2957                 loc->cl_loc.cl_name[name_len] = '\0';
2958
2959                 if (!__insert_crush_loc(locs, loc)) {
2960                         free_crush_loc(loc);
2961                         return -EEXIST;
2962                 }
2963
2964                 dout("%s type_name '%s' name '%s'\n", __func__,
2965                      loc->cl_loc.cl_type_name, loc->cl_loc.cl_name);
2966         }
2967
2968         return 0;
2969 }
2970
2971 int ceph_compare_crush_locs(struct rb_root *locs1, struct rb_root *locs2)
2972 {
2973         struct rb_node *n1 = rb_first(locs1);
2974         struct rb_node *n2 = rb_first(locs2);
2975         int ret;
2976
2977         for ( ; n1 && n2; n1 = rb_next(n1), n2 = rb_next(n2)) {
2978                 struct crush_loc_node *loc1 =
2979                     rb_entry(n1, struct crush_loc_node, cl_node);
2980                 struct crush_loc_node *loc2 =
2981                     rb_entry(n2, struct crush_loc_node, cl_node);
2982
2983                 ret = crush_loc_compare(&loc1->cl_loc, &loc2->cl_loc);
2984                 if (ret)
2985                         return ret;
2986         }
2987
2988         if (!n1 && n2)
2989                 return -1;
2990         if (n1 && !n2)
2991                 return 1;
2992         return 0;
2993 }
2994
2995 void ceph_clear_crush_locs(struct rb_root *locs)
2996 {
2997         while (!RB_EMPTY_ROOT(locs)) {
2998                 struct crush_loc_node *loc =
2999                     rb_entry(rb_first(locs), struct crush_loc_node, cl_node);
3000
3001                 erase_crush_loc(locs, loc);
3002                 free_crush_loc(loc);
3003         }
3004 }
3005
3006 /*
3007  * [a-zA-Z0-9-_.]+
3008  */
3009 static bool is_valid_crush_name(const char *name)
3010 {
3011         do {
3012                 if (!('a' <= *name && *name <= 'z') &&
3013                     !('A' <= *name && *name <= 'Z') &&
3014                     !('0' <= *name && *name <= '9') &&
3015                     *name != '-' && *name != '_' && *name != '.')
3016                         return false;
3017         } while (*++name != '\0');
3018
3019         return true;
3020 }
3021
3022 /*
3023  * Gets the parent of an item.  Returns its id (<0 because the
3024  * parent is always a bucket), type id (>0 for the same reason,
3025  * via @parent_type_id) and location (via @parent_loc).  If no
3026  * parent, returns 0.
3027  *
3028  * Does a linear search, as there are no parent pointers of any
3029  * kind.  Note that the result is ambiguous for items that occur
3030  * multiple times in the map.
3031  */
3032 static int get_immediate_parent(struct crush_map *c, int id,
3033                                 u16 *parent_type_id,
3034                                 struct crush_loc *parent_loc)
3035 {
3036         struct crush_bucket *b;
3037         struct crush_name_node *type_cn, *cn;
3038         int i, j;
3039
3040         for (i = 0; i < c->max_buckets; i++) {
3041                 b = c->buckets[i];
3042                 if (!b)
3043                         continue;
3044
3045                 /* ignore per-class shadow hierarchy */
3046                 cn = lookup_crush_name(&c->names, b->id);
3047                 if (!cn || !is_valid_crush_name(cn->cn_name))
3048                         continue;
3049
3050                 for (j = 0; j < b->size; j++) {
3051                         if (b->items[j] != id)
3052                                 continue;
3053
3054                         *parent_type_id = b->type;
3055                         type_cn = lookup_crush_name(&c->type_names, b->type);
3056                         parent_loc->cl_type_name = type_cn->cn_name;
3057                         parent_loc->cl_name = cn->cn_name;
3058                         return b->id;
3059                 }
3060         }
3061
3062         return 0;  /* no parent */
3063 }
3064
3065 /*
3066  * Calculates the locality/distance from an item to a client
3067  * location expressed in terms of CRUSH hierarchy as a set of
3068  * (bucket type name, bucket name) pairs.  Specifically, looks
3069  * for the lowest-valued bucket type for which the location of
3070  * @id matches one of the locations in @locs, so for standard
3071  * bucket types (host = 1, rack = 3, datacenter = 8, zone = 9)
3072  * a matching host is closer than a matching rack and a matching
3073  * data center is closer than a matching zone.
3074  *
3075  * Specifying multiple locations (a "multipath" location) such
3076  * as "rack=foo1 rack=foo2 datacenter=bar" is allowed -- @locs
3077  * is a multimap.  The locality will be:
3078  *
3079  * - 3 for OSDs in racks foo1 and foo2
3080  * - 8 for OSDs in data center bar
3081  * - -1 for all other OSDs
3082  *
3083  * The lowest possible bucket type is 1, so the best locality
3084  * for an OSD is 1 (i.e. a matching host).  Locality 0 would be
3085  * the OSD itself.
3086  */
3087 int ceph_get_crush_locality(struct ceph_osdmap *osdmap, int id,
3088                             struct rb_root *locs)
3089 {
3090         struct crush_loc loc;
3091         u16 type_id;
3092
3093         /*
3094          * Instead of repeated get_immediate_parent() calls,
3095          * the location of @id could be obtained with a single
3096          * depth-first traversal.
3097          */
3098         for (;;) {
3099                 id = get_immediate_parent(osdmap->crush, id, &type_id, &loc);
3100                 if (id >= 0)
3101                         return -1;  /* not local */
3102
3103                 if (lookup_crush_loc(locs, &loc))
3104                         return type_id;
3105         }
3106 }