net/ceph/ceph_common.c

   1
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/backing-dev.h>
   4 #include <linux/ctype.h>
   5 #include <linux/fs.h>
   6 #include <linux/inet.h>
   7 #include <linux/in6.h>
   8 #include <linux/key.h>
   9 #include <keys/ceph-type.h>
  10 #include <linux/module.h>
  11 #include <linux/mount.h>
  12 #include <linux/nsproxy.h>
  13 #include <linux/parser.h>
  14 #include <linux/sched.h>
  15 #include <linux/seq_file.h>
  16 #include <linux/slab.h>
  17 #include <linux/statfs.h>
  18 #include <linux/string.h>
  19 #include <linux/vmalloc.h>
  20
  21
  22 #include <linux/ceph/ceph_features.h>
  23 #include <linux/ceph/libceph.h>
  24 #include <linux/ceph/debugfs.h>
  25 #include <linux/ceph/decode.h>
  26 #include <linux/ceph/mon_client.h>
  27 #include <linux/ceph/auth.h>
  28 #include "crypto.h"
  29
  30
  31 /*
  32  * Module compatibility interface.  For now it doesn't do anything,
  33  * but its existence signals a certain level of functionality.
  34  *
  35  * The data buffer is used to pass information both to and from
  36  * libceph.  The return value indicates whether libceph determines
  37  * it is compatible with the caller (from another kernel module),
  38  * given the provided data.
  39  *
  40  * The data pointer can be null.
  41  */
  42 bool libceph_compatible(void *data)
  43 {
  44         return true;
  45 }
  46 EXPORT_SYMBOL(libceph_compatible);
  47
  48 static int param_get_supported_features(char *buffer,
  49                                         const struct kernel_param *kp)
  50 {
  51         return sprintf(buffer, "0x%llx", CEPH_FEATURES_SUPPORTED_DEFAULT);
  52 }
  53 static const struct kernel_param_ops param_ops_supported_features = {
  54         .get = param_get_supported_features,
  55 };
  56 module_param_cb(supported_features, &param_ops_supported_features, NULL,
  57                 0444);
  58
  59 const char *ceph_msg_type_name(int type)
  60 {
  61         switch (type) {
  62         case CEPH_MSG_SHUTDOWN: return "shutdown";
  63         case CEPH_MSG_PING: return "ping";
  64         case CEPH_MSG_AUTH: return "auth";
  65         case CEPH_MSG_AUTH_REPLY: return "auth_reply";
  66         case CEPH_MSG_MON_MAP: return "mon_map";
  67         case CEPH_MSG_MON_GET_MAP: return "mon_get_map";
  68         case CEPH_MSG_MON_SUBSCRIBE: return "mon_subscribe";
  69         case CEPH_MSG_MON_SUBSCRIBE_ACK: return "mon_subscribe_ack";
  70         case CEPH_MSG_STATFS: return "statfs";
  71         case CEPH_MSG_STATFS_REPLY: return "statfs_reply";
  72         case CEPH_MSG_MON_GET_VERSION: return "mon_get_version";
  73         case CEPH_MSG_MON_GET_VERSION_REPLY: return "mon_get_version_reply";
  74         case CEPH_MSG_MDS_MAP: return "mds_map";
  75         case CEPH_MSG_FS_MAP_USER: return "fs_map_user";
  76         case CEPH_MSG_CLIENT_SESSION: return "client_session";
  77         case CEPH_MSG_CLIENT_RECONNECT: return "client_reconnect";
  78         case CEPH_MSG_CLIENT_REQUEST: return "client_request";
  79         case CEPH_MSG_CLIENT_REQUEST_FORWARD: return "client_request_forward";
  80         case CEPH_MSG_CLIENT_REPLY: return "client_reply";
  81         case CEPH_MSG_CLIENT_CAPS: return "client_caps";
  82         case CEPH_MSG_CLIENT_CAPRELEASE: return "client_cap_release";
  83         case CEPH_MSG_CLIENT_QUOTA: return "client_quota";
  84         case CEPH_MSG_CLIENT_SNAP: return "client_snap";
  85         case CEPH_MSG_CLIENT_LEASE: return "client_lease";
  86         case CEPH_MSG_POOLOP_REPLY: return "poolop_reply";
  87         case CEPH_MSG_POOLOP: return "poolop";
  88         case CEPH_MSG_MON_COMMAND: return "mon_command";
  89         case CEPH_MSG_MON_COMMAND_ACK: return "mon_command_ack";
  90         case CEPH_MSG_OSD_MAP: return "osd_map";
  91         case CEPH_MSG_OSD_OP: return "osd_op";
  92         case CEPH_MSG_OSD_OPREPLY: return "osd_opreply";
  93         case CEPH_MSG_WATCH_NOTIFY: return "watch_notify";
  94         case CEPH_MSG_OSD_BACKOFF: return "osd_backoff";
  95         default: return "unknown";
  96         }
  97 }
  98 EXPORT_SYMBOL(ceph_msg_type_name);
  99
 100 /*
 101  * Initially learn our fsid, or verify an fsid matches.
 102  */
 103 int ceph_check_fsid(struct ceph_client *client, struct ceph_fsid *fsid)
 104 {
 105         if (client->have_fsid) {
 106                 if (ceph_fsid_compare(&client->fsid, fsid)) {
 107                         pr_err("bad fsid, had %pU got %pU",
 108                                &client->fsid, fsid);
 109                         return -1;
 110                 }
 111         } else {
 112                 memcpy(&client->fsid, fsid, sizeof(*fsid));
 113         }
 114         return 0;
 115 }
 116 EXPORT_SYMBOL(ceph_check_fsid);
 117
 118 static int strcmp_null(const char *s1, const char *s2)
 119 {
 120         if (!s1 && !s2)
 121                 return 0;
 122         if (s1 && !s2)
 123                 return -1;
 124         if (!s1 && s2)
 125                 return 1;
 126         return strcmp(s1, s2);
 127 }
 128
 129 int ceph_compare_options(struct ceph_options *new_opt,
 130                          struct ceph_client *client)
 131 {
 132         struct ceph_options *opt1 = new_opt;
 133         struct ceph_options *opt2 = client->options;
 134         int ofs = offsetof(struct ceph_options, mon_addr);
 135         int i;
 136         int ret;
 137
 138         /*
 139          * Don't bother comparing options if network namespaces don't
 140          * match.
 141          */
 142         if (!net_eq(current->nsproxy->net_ns, read_pnet(&client->msgr.net)))
 143                 return -1;
 144
 145         ret = memcmp(opt1, opt2, ofs);
 146         if (ret)
 147                 return ret;
 148
 149         ret = strcmp_null(opt1->name, opt2->name);
 150         if (ret)
 151                 return ret;
 152
 153         if (opt1->key && !opt2->key)
 154                 return -1;
 155         if (!opt1->key && opt2->key)
 156                 return 1;
 157         if (opt1->key && opt2->key) {
 158                 if (opt1->key->type != opt2->key->type)
 159                         return -1;
 160                 if (opt1->key->created.tv_sec != opt2->key->created.tv_sec)
 161                         return -1;
 162                 if (opt1->key->created.tv_nsec != opt2->key->created.tv_nsec)
 163                         return -1;
 164                 if (opt1->key->len != opt2->key->len)
 165                         return -1;
 166                 if (opt1->key->key && !opt2->key->key)
 167                         return -1;
 168                 if (!opt1->key->key && opt2->key->key)
 169                         return 1;
 170                 if (opt1->key->key && opt2->key->key) {
 171                         ret = memcmp(opt1->key->key, opt2->key->key, opt1->key->len);
 172                         if (ret)
 173                                 return ret;
 174                 }
 175         }
 176
 177         /* any matching mon ip implies a match */
 178         for (i = 0; i < opt1->num_mon; i++) {
 179                 if (ceph_monmap_contains(client->monc.monmap,
 180                                  &opt1->mon_addr[i]))
 181                         return 0;
 182         }
 183         return -1;
 184 }
 185 EXPORT_SYMBOL(ceph_compare_options);
 186
 187 void *ceph_kvmalloc(size_t size, gfp_t flags)
 188 {
 189         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
 190                 void *ptr = kmalloc(size, flags | __GFP_NOWARN);
 191                 if (ptr)
 192                         return ptr;
 193         }
 194
 195         return __vmalloc(size, flags, PAGE_KERNEL);
 196 }
 197
 198
 199 static int parse_fsid(const char *str, struct ceph_fsid *fsid)
 200 {
 201         int i = 0;
 202         char tmp[3];
 203         int err = -EINVAL;
 204         int d;
 205
 206         dout("parse_fsid '%s'\n", str);
 207         tmp[2] = 0;
 208         while (*str && i < 16) {
 209                 if (ispunct(*str)) {
 210                         str++;
 211                         continue;
 212                 }
 213                 if (!isxdigit(str[0]) || !isxdigit(str[1]))
 214                         break;
 215                 tmp[0] = str[0];
 216                 tmp[1] = str[1];
 217                 if (sscanf(tmp, "%x", &d) < 1)
 218                         break;
 219                 fsid->fsid[i] = d & 0xff;
 220                 i++;
 221                 str += 2;
 222         }
 223
 224         if (i == 16)
 225                 err = 0;
 226         dout("parse_fsid ret %d got fsid %pU\n", err, fsid);
 227         return err;
 228 }
 229
 230 /*
 231  * ceph options
 232  */
 233 enum {
 234         Opt_osdtimeout,
 235         Opt_osdkeepalivetimeout,
 236         Opt_mount_timeout,
 237         Opt_osd_idle_ttl,
 238         Opt_osd_request_timeout,
 239         Opt_last_int,
 240         /* int args above */
 241         Opt_fsid,
 242         Opt_name,
 243         Opt_secret,
 244         Opt_key,
 245         Opt_ip,
 246         Opt_last_string,
 247         /* string args above */
 248         Opt_share,
 249         Opt_noshare,
 250         Opt_crc,
 251         Opt_nocrc,
 252         Opt_cephx_require_signatures,
 253         Opt_nocephx_require_signatures,
 254         Opt_cephx_sign_messages,
 255         Opt_nocephx_sign_messages,
 256         Opt_tcp_nodelay,
 257         Opt_notcp_nodelay,
 258 };
 259
 260 static match_table_t opt_tokens = {
 261         {Opt_osdtimeout, "osdtimeout=%d"},
 262         {Opt_osdkeepalivetimeout, "osdkeepalive=%d"},
 263         {Opt_mount_timeout, "mount_timeout=%d"},
 264         {Opt_osd_idle_ttl, "osd_idle_ttl=%d"},
 265         {Opt_osd_request_timeout, "osd_request_timeout=%d"},
 266         /* int args above */
 267         {Opt_fsid, "fsid=%s"},
 268         {Opt_name, "name=%s"},
 269         {Opt_secret, "secret=%s"},
 270         {Opt_key, "key=%s"},
 271         {Opt_ip, "ip=%s"},
 272         /* string args above */
 273         {Opt_share, "share"},
 274         {Opt_noshare, "noshare"},
 275         {Opt_crc, "crc"},
 276         {Opt_nocrc, "nocrc"},
 277         {Opt_cephx_require_signatures, "cephx_require_signatures"},
 278         {Opt_nocephx_require_signatures, "nocephx_require_signatures"},
 279         {Opt_cephx_sign_messages, "cephx_sign_messages"},
 280         {Opt_nocephx_sign_messages, "nocephx_sign_messages"},
 281         {Opt_tcp_nodelay, "tcp_nodelay"},
 282         {Opt_notcp_nodelay, "notcp_nodelay"},
 283         {-1, NULL}
 284 };
 285
 286 void ceph_destroy_options(struct ceph_options *opt)
 287 {
 288         dout("destroy_options %p\n", opt);
 289         kfree(opt->name);
 290         if (opt->key) {
 291                 ceph_crypto_key_destroy(opt->key);
 292                 kfree(opt->key);
 293         }
 294         kfree(opt->mon_addr);
 295         kfree(opt);
 296 }
 297 EXPORT_SYMBOL(ceph_destroy_options);
 298
 299 /* get secret from key store */
 300 static int get_secret(struct ceph_crypto_key *dst, const char *name) {
 301         struct key *ukey;
 302         int key_err;
 303         int err = 0;
 304         struct ceph_crypto_key *ckey;
 305
 306         ukey = request_key(&key_type_ceph, name, NULL);
 307         if (IS_ERR(ukey)) {
 308                 /* request_key errors don't map nicely to mount(2)
 309                    errors; don't even try, but still printk */
 310                 key_err = PTR_ERR(ukey);
 311                 switch (key_err) {
 312                 case -ENOKEY:
 313                         pr_warn("ceph: Mount failed due to key not found: %s\n",
 314                                 name);
 315                         break;
 316                 case -EKEYEXPIRED:
 317                         pr_warn("ceph: Mount failed due to expired key: %s\n",
 318                                 name);
 319                         break;
 320                 case -EKEYREVOKED:
 321                         pr_warn("ceph: Mount failed due to revoked key: %s\n",
 322                                 name);
 323                         break;
 324                 default:
 325                         pr_warn("ceph: Mount failed due to unknown key error %d: %s\n",
 326                                 key_err, name);
 327                 }
 328                 err = -EPERM;
 329                 goto out;
 330         }
 331
 332         ckey = ukey->payload.data[0];
 333         err = ceph_crypto_key_clone(dst, ckey);
 334         if (err)
 335                 goto out_key;
 336         /* pass through, err is 0 */
 337
 338 out_key:
 339         key_put(ukey);
 340 out:
 341         return err;
 342 }
 343
 344 struct ceph_options *
 345 ceph_parse_options(char *options, const char *dev_name,
 346                         const char *dev_name_end,
 347                         int (*parse_extra_token)(char *c, void *private),
 348                         void *private)
 349 {
 350         struct ceph_options *opt;
 351         const char *c;
 352         int err = -ENOMEM;
 353         substring_t argstr[MAX_OPT_ARGS];
 354
 355         opt = kzalloc(sizeof(*opt), GFP_KERNEL);
 356         if (!opt)
 357                 return ERR_PTR(-ENOMEM);
 358         opt->mon_addr = kcalloc(CEPH_MAX_MON, sizeof(*opt->mon_addr),
 359                                 GFP_KERNEL);
 360         if (!opt->mon_addr)
 361                 goto out;
 362
 363         dout("parse_options %p options '%s' dev_name '%s'\n", opt, options,
 364              dev_name);
 365
 366         /* start with defaults */
 367         opt->flags = CEPH_OPT_DEFAULT;
 368         opt->osd_keepalive_timeout = CEPH_OSD_KEEPALIVE_DEFAULT;
 369         opt->mount_timeout = CEPH_MOUNT_TIMEOUT_DEFAULT;
 370         opt->osd_idle_ttl = CEPH_OSD_IDLE_TTL_DEFAULT;
 371         opt->osd_request_timeout = CEPH_OSD_REQUEST_TIMEOUT_DEFAULT;
 372
 373         /* get mon ip(s) */
 374         /* ip1[:port1][,ip2[:port2]...] */
 375         err = ceph_parse_ips(dev_name, dev_name_end, opt->mon_addr,
 376                              CEPH_MAX_MON, &opt->num_mon);
 377         if (err < 0)
 378                 goto out;
 379
 380         /* parse mount options */
 381         while ((c = strsep(&options, ",")) != NULL) {
 382                 int token, intval;
 383                 if (!*c)
 384                         continue;
 385                 err = -EINVAL;
 386                 token = match_token((char *)c, opt_tokens, argstr);
 387                 if (token < 0 && parse_extra_token) {
 388                         /* extra? */
 389                         err = parse_extra_token((char *)c, private);
 390                         if (err < 0) {
 391                                 pr_err("bad option at '%s'\n", c);
 392                                 goto out;
 393                         }
 394                         continue;
 395                 }
 396                 if (token < Opt_last_int) {
 397                         err = match_int(&argstr[0], &intval);
 398                         if (err < 0) {
 399                                 pr_err("bad option arg (not int) at '%s'\n", c);
 400                                 goto out;
 401                         }
 402                         dout("got int token %d val %d\n", token, intval);
 403                 } else if (token > Opt_last_int && token < Opt_last_string) {
 404                         dout("got string token %d val %s\n", token,
 405                              argstr[0].from);
 406                 } else {
 407                         dout("got token %d\n", token);
 408                 }
 409                 switch (token) {
 410                 case Opt_ip:
 411                         err = ceph_parse_ips(argstr[0].from,
 412                                              argstr[0].to,
 413                                              &opt->my_addr,
 414                                              1, NULL);
 415                         if (err < 0)
 416                                 goto out;
 417                         opt->flags |= CEPH_OPT_MYIP;
 418                         break;
 419
 420                 case Opt_fsid:
 421                         err = parse_fsid(argstr[0].from, &opt->fsid);
 422                         if (err == 0)
 423                                 opt->flags |= CEPH_OPT_FSID;
 424                         break;
 425                 case Opt_name:
 426                         kfree(opt->name);
 427                         opt->name = kstrndup(argstr[0].from,
 428                                               argstr[0].to-argstr[0].from,
 429                                               GFP_KERNEL);
 430                         if (!opt->name) {
 431                                 err = -ENOMEM;
 432                                 goto out;
 433                         }
 434                         break;
 435                 case Opt_secret:
 436                         ceph_crypto_key_destroy(opt->key);
 437                         kfree(opt->key);
 438
 439                         opt->key = kzalloc(sizeof(*opt->key), GFP_KERNEL);
 440                         if (!opt->key) {
 441                                 err = -ENOMEM;
 442                                 goto out;
 443                         }
 444                         err = ceph_crypto_key_unarmor(opt->key, argstr[0].from);
 445                         if (err < 0)
 446                                 goto out;
 447                         break;
 448                 case Opt_key:
 449                         ceph_crypto_key_destroy(opt->key);
 450                         kfree(opt->key);
 451
 452                         opt->key = kzalloc(sizeof(*opt->key), GFP_KERNEL);
 453                         if (!opt->key) {
 454                                 err = -ENOMEM;
 455                                 goto out;
 456                         }
 457                         err = get_secret(opt->key, argstr[0].from);
 458                         if (err < 0)
 459                                 goto out;
 460                         break;
 461
 462                         /* misc */
 463                 case Opt_osdtimeout:
 464                         pr_warn("ignoring deprecated osdtimeout option\n");
 465                         break;
 466                 case Opt_osdkeepalivetimeout:
 467                         /* 0 isn't well defined right now, reject it */
 468                         if (intval < 1 || intval > INT_MAX / 1000) {
 469                                 pr_err("osdkeepalive out of range\n");
 470                                 err = -EINVAL;
 471                                 goto out;
 472                         }
 473                         opt->osd_keepalive_timeout =
 474                                         msecs_to_jiffies(intval * 1000);
 475                         break;
 476                 case Opt_osd_idle_ttl:
 477                         /* 0 isn't well defined right now, reject it */
 478                         if (intval < 1 || intval > INT_MAX / 1000) {
 479                                 pr_err("osd_idle_ttl out of range\n");
 480                                 err = -EINVAL;
 481                                 goto out;
 482                         }
 483                         opt->osd_idle_ttl = msecs_to_jiffies(intval * 1000);
 484                         break;
 485                 case Opt_mount_timeout:
 486                         /* 0 is "wait forever" (i.e. infinite timeout) */
 487                         if (intval < 0 || intval > INT_MAX / 1000) {
 488                                 pr_err("mount_timeout out of range\n");
 489                                 err = -EINVAL;
 490                                 goto out;
 491                         }
 492                         opt->mount_timeout = msecs_to_jiffies(intval * 1000);
 493                         break;
 494                 case Opt_osd_request_timeout:
 495                         /* 0 is "wait forever" (i.e. infinite timeout) */
 496                         if (intval < 0 || intval > INT_MAX / 1000) {
 497                                 pr_err("osd_request_timeout out of range\n");
 498                                 err = -EINVAL;
 499                                 goto out;
 500                         }
 501                         opt->osd_request_timeout = msecs_to_jiffies(intval * 1000);
 502                         break;
 503
 504                 case Opt_share:
 505                         opt->flags &= ~CEPH_OPT_NOSHARE;
 506                         break;
 507                 case Opt_noshare:
 508                         opt->flags |= CEPH_OPT_NOSHARE;
 509                         break;
 510
 511                 case Opt_crc:
 512                         opt->flags &= ~CEPH_OPT_NOCRC;
 513                         break;
 514                 case Opt_nocrc:
 515                         opt->flags |= CEPH_OPT_NOCRC;
 516                         break;
 517
 518                 case Opt_cephx_require_signatures:
 519                         opt->flags &= ~CEPH_OPT_NOMSGAUTH;
 520                         break;
 521                 case Opt_nocephx_require_signatures:
 522                         opt->flags |= CEPH_OPT_NOMSGAUTH;
 523                         break;
 524                 case Opt_cephx_sign_messages:
 525                         opt->flags &= ~CEPH_OPT_NOMSGSIGN;
 526                         break;
 527                 case Opt_nocephx_sign_messages:
 528                         opt->flags |= CEPH_OPT_NOMSGSIGN;
 529                         break;
 530
 531                 case Opt_tcp_nodelay:
 532                         opt->flags |= CEPH_OPT_TCP_NODELAY;
 533                         break;
 534                 case Opt_notcp_nodelay:
 535                         opt->flags &= ~CEPH_OPT_TCP_NODELAY;
 536                         break;
 537
 538                 default:
 539                         BUG_ON(token);
 540                 }
 541         }
 542
 543         /* success */
 544         return opt;
 545
 546 out:
 547         ceph_destroy_options(opt);
 548         return ERR_PTR(err);
 549 }
 550 EXPORT_SYMBOL(ceph_parse_options);
 551
 552 int ceph_print_client_options(struct seq_file *m, struct ceph_client *client)
 553 {
 554         struct ceph_options *opt = client->options;
 555         size_t pos = m->count;
 556
 557         if (opt->name) {
 558                 seq_puts(m, "name=");
 559                 seq_escape(m, opt->name, ", \t\n\\");
 560                 seq_putc(m, ',');
 561         }
 562         if (opt->key)
 563                 seq_puts(m, "secret=<hidden>,");
 564
 565         if (opt->flags & CEPH_OPT_FSID)
 566                 seq_printf(m, "fsid=%pU,", &opt->fsid);
 567         if (opt->flags & CEPH_OPT_NOSHARE)
 568                 seq_puts(m, "noshare,");
 569         if (opt->flags & CEPH_OPT_NOCRC)
 570                 seq_puts(m, "nocrc,");
 571         if (opt->flags & CEPH_OPT_NOMSGAUTH)
 572                 seq_puts(m, "nocephx_require_signatures,");
 573         if (opt->flags & CEPH_OPT_NOMSGSIGN)
 574                 seq_puts(m, "nocephx_sign_messages,");
 575         if ((opt->flags & CEPH_OPT_TCP_NODELAY) == 0)
 576                 seq_puts(m, "notcp_nodelay,");
 577
 578         if (opt->mount_timeout != CEPH_MOUNT_TIMEOUT_DEFAULT)
 579                 seq_printf(m, "mount_timeout=%d,",
 580                            jiffies_to_msecs(opt->mount_timeout) / 1000);
 581         if (opt->osd_idle_ttl != CEPH_OSD_IDLE_TTL_DEFAULT)
 582                 seq_printf(m, "osd_idle_ttl=%d,",
 583                            jiffies_to_msecs(opt->osd_idle_ttl) / 1000);
 584         if (opt->osd_keepalive_timeout != CEPH_OSD_KEEPALIVE_DEFAULT)
 585                 seq_printf(m, "osdkeepalivetimeout=%d,",
 586                     jiffies_to_msecs(opt->osd_keepalive_timeout) / 1000);
 587         if (opt->osd_request_timeout != CEPH_OSD_REQUEST_TIMEOUT_DEFAULT)
 588                 seq_printf(m, "osd_request_timeout=%d,",
 589                            jiffies_to_msecs(opt->osd_request_timeout) / 1000);
 590
 591         /* drop redundant comma */
 592         if (m->count != pos)
 593                 m->count--;
 594
 595         return 0;
 596 }
 597 EXPORT_SYMBOL(ceph_print_client_options);
 598
 599 struct ceph_entity_addr *ceph_client_addr(struct ceph_client *client)
 600 {
 601         return &client->msgr.inst.addr;
 602 }
 603 EXPORT_SYMBOL(ceph_client_addr);
 604
 605 u64 ceph_client_gid(struct ceph_client *client)
 606 {
 607         return client->monc.auth->global_id;
 608 }
 609 EXPORT_SYMBOL(ceph_client_gid);
 610
 611 /*
 612  * create a fresh client instance
 613  */
 614 struct ceph_client *ceph_create_client(struct ceph_options *opt, void *private)
 615 {
 616         struct ceph_client *client;
 617         struct ceph_entity_addr *myaddr = NULL;
 618         int err;
 619
 620         err = wait_for_random_bytes();
 621         if (err < 0)
 622                 return ERR_PTR(err);
 623
 624         client = kzalloc(sizeof(*client), GFP_KERNEL);
 625         if (client == NULL)
 626                 return ERR_PTR(-ENOMEM);
 627
 628         client->private = private;
 629         client->options = opt;
 630
 631         mutex_init(&client->mount_mutex);
 632         init_waitqueue_head(&client->auth_wq);
 633         client->auth_err = 0;
 634
 635         client->extra_mon_dispatch = NULL;
 636         client->supported_features = CEPH_FEATURES_SUPPORTED_DEFAULT;
 637         client->required_features = CEPH_FEATURES_REQUIRED_DEFAULT;
 638
 639         if (!ceph_test_opt(client, NOMSGAUTH))
 640                 client->required_features |= CEPH_FEATURE_MSG_AUTH;
 641
 642         /* msgr */
 643         if (ceph_test_opt(client, MYIP))
 644                 myaddr = &client->options->my_addr;
 645
 646         ceph_messenger_init(&client->msgr, myaddr);
 647
 648         /* subsystems */
 649         err = ceph_monc_init(&client->monc, client);
 650         if (err < 0)
 651                 goto fail;
 652         err = ceph_osdc_init(&client->osdc, client);
 653         if (err < 0)
 654                 goto fail_monc;
 655
 656         return client;
 657
 658 fail_monc:
 659         ceph_monc_stop(&client->monc);
 660 fail:
 661         ceph_messenger_fini(&client->msgr);
 662         kfree(client);
 663         return ERR_PTR(err);
 664 }
 665 EXPORT_SYMBOL(ceph_create_client);
 666
 667 void ceph_destroy_client(struct ceph_client *client)
 668 {
 669         dout("destroy_client %p\n", client);
 670
 671         atomic_set(&client->msgr.stopping, 1);
 672
 673         /* unmount */
 674         ceph_osdc_stop(&client->osdc);
 675         ceph_monc_stop(&client->monc);
 676         ceph_messenger_fini(&client->msgr);
 677
 678         ceph_debugfs_client_cleanup(client);
 679
 680         ceph_destroy_options(client->options);
 681
 682         kfree(client);
 683         dout("destroy_client %p done\n", client);
 684 }
 685 EXPORT_SYMBOL(ceph_destroy_client);
 686
 687 /*
 688  * true if we have the mon map (and have thus joined the cluster)
 689  */
 690 static bool have_mon_and_osd_map(struct ceph_client *client)
 691 {
 692         return client->monc.monmap && client->monc.monmap->epoch &&
 693                client->osdc.osdmap && client->osdc.osdmap->epoch;
 694 }
 695
 696 /*
 697  * mount: join the ceph cluster, and open root directory.
 698  */
 699 int __ceph_open_session(struct ceph_client *client, unsigned long started)
 700 {
 701         unsigned long timeout = client->options->mount_timeout;
 702         long err;
 703
 704         /* open session, and wait for mon and osd maps */
 705         err = ceph_monc_open_session(&client->monc);
 706         if (err < 0)
 707                 return err;
 708
 709         while (!have_mon_and_osd_map(client)) {
 710                 if (timeout && time_after_eq(jiffies, started + timeout))
 711                         return -ETIMEDOUT;
 712
 713                 /* wait */
 714                 dout("mount waiting for mon_map\n");
 715                 err = wait_event_interruptible_timeout(client->auth_wq,
 716                         have_mon_and_osd_map(client) || (client->auth_err < 0),
 717                         ceph_timeout_jiffies(timeout));
 718                 if (err < 0)
 719                         return err;
 720                 if (client->auth_err < 0)
 721                         return client->auth_err;
 722         }
 723
 724         pr_info("client%llu fsid %pU\n", ceph_client_gid(client),
 725                 &client->fsid);
 726         ceph_debugfs_client_init(client);
 727
 728         return 0;
 729 }
 730 EXPORT_SYMBOL(__ceph_open_session);
 731
 732 int ceph_open_session(struct ceph_client *client)
 733 {
 734         int ret;
 735         unsigned long started = jiffies;  /* note the start time */
 736
 737         dout("open_session start\n");
 738         mutex_lock(&client->mount_mutex);
 739
 740         ret = __ceph_open_session(client, started);
 741
 742         mutex_unlock(&client->mount_mutex);
 743         return ret;
 744 }
 745 EXPORT_SYMBOL(ceph_open_session);
 746
 747 int ceph_wait_for_latest_osdmap(struct ceph_client *client,
 748                                 unsigned long timeout)
 749 {
 750         u64 newest_epoch;
 751         int ret;
 752
 753         ret = ceph_monc_get_version(&client->monc, "osdmap", &newest_epoch);
 754         if (ret)
 755                 return ret;
 756
 757         if (client->osdc.osdmap->epoch >= newest_epoch)
 758                 return 0;
 759
 760         ceph_osdc_maybe_request_map(&client->osdc);
 761         return ceph_monc_wait_osdmap(&client->monc, newest_epoch, timeout);
 762 }
 763 EXPORT_SYMBOL(ceph_wait_for_latest_osdmap);
 764
 765 static int __init init_ceph_lib(void)
 766 {
 767         int ret = 0;
 768
 769         ret = ceph_debugfs_init();
 770         if (ret < 0)
 771                 goto out;
 772
 773         ret = ceph_crypto_init();
 774         if (ret < 0)
 775                 goto out_debugfs;
 776
 777         ret = ceph_msgr_init();
 778         if (ret < 0)
 779                 goto out_crypto;
 780
 781         ret = ceph_osdc_setup();
 782         if (ret < 0)
 783                 goto out_msgr;
 784
 785         pr_info("loaded (mon/osd proto %d/%d)\n",
 786                 CEPH_MONC_PROTOCOL, CEPH_OSDC_PROTOCOL);
 787
 788         return 0;
 789
 790 out_msgr:
 791         ceph_msgr_exit();
 792 out_crypto:
 793         ceph_crypto_shutdown();
 794 out_debugfs:
 795         ceph_debugfs_cleanup();
 796 out:
 797         return ret;
 798 }
 799
 800 static void __exit exit_ceph_lib(void)
 801 {
 802         dout("exit_ceph_lib\n");
 803         WARN_ON(!ceph_strings_empty());
 804
 805         ceph_osdc_cleanup();
 806         ceph_msgr_exit();
 807         ceph_crypto_shutdown();
 808         ceph_debugfs_cleanup();
 809 }
 810
 811 module_init(init_ceph_lib);
 812 module_exit(exit_ceph_lib);
 813
 814 MODULE_AUTHOR("Sage Weil <sage@newdream.net>");
 815 MODULE_AUTHOR("Yehuda Sadeh <yehuda@hq.newdream.net>");
 816 MODULE_AUTHOR("Patience Warnick <patience@newdream.net>");
 817 MODULE_DESCRIPTION("Ceph core library");
 818 MODULE_LICENSE("GPL");