ext/raindrops/linux_inet_diag.c

   1 #include <ruby.h>
   2 #ifdef HAVE_RUBY_ST_H
   3 #  include <ruby/st.h>
   4 #else
   5 #  include <st.h>
   6 #endif
   7 #ifdef __linux__
   8
   9 /* Ruby 1.8.6+ macros (for compatibility with Ruby 1.9) */
  10 #ifndef RSTRING_LEN
  11 #  define RSTRING_LEN(s) (RSTRING(s)->len)
  12 #endif
  13
  14 /* partial emulation of the 1.9 rb_thread_blocking_region under 1.8 */
  15 #ifndef HAVE_RB_THREAD_BLOCKING_REGION
  16 #  include <rubysig.h>
  17 typedef void rb_unblock_function_t(void *);
  18 typedef VALUE rb_blocking_function_t(void *);
  19 static VALUE
  20 rb_thread_blocking_region(
  21         rb_blocking_function_t *func, void *data1,
  22         rb_unblock_function_t *ubf, void *data2)
  23 {
  24         VALUE rv;
  25
  26         TRAP_BEG;
  27         rv = func(data1);
  28         TRAP_END;
  29
  30         return rv;
  31 }
  32 #endif /* ! HAVE_RB_THREAD_BLOCKING_REGION */
  33
  34 #include <assert.h>
  35 #include <errno.h>
  36 #include <sys/socket.h>
  37 #include <sys/types.h>
  38 #include <netdb.h>
  39 #include <unistd.h>
  40 #include <string.h>
  41 #include <asm/types.h>
  42 #include <netinet/in.h>
  43 #include <arpa/inet.h>
  44 #include <netinet/tcp.h>
  45 #include <linux/netlink.h>
  46 #include <linux/rtnetlink.h>
  47 #include <linux/inet_diag.h>
  48
  49 static size_t page_size;
  50 static unsigned g_seq;
  51 static VALUE cListenStats;
  52
  53 struct listen_stats {
  54         uint32_t active;
  55         uint32_t listener_p:1;
  56         uint32_t queued:31;
  57 };
  58
  59 #define OPLEN (sizeof(struct inet_diag_bc_op) + \
  60                sizeof(struct inet_diag_hostcond) + \
  61                sizeof(struct sockaddr_storage))
  62
  63 struct nogvl_args {
  64         st_table *table;
  65         struct iovec iov[3]; /* last iov holds inet_diag bytecode */
  66         struct listen_stats stats;
  67 };
  68
  69 /* creates a Ruby ListenStats Struct based on our internal listen_stats */
  70 static VALUE rb_listen_stats(struct listen_stats *stats)
  71 {
  72         VALUE active = UINT2NUM(stats->active);
  73         VALUE queued = UINT2NUM(stats->queued);
  74
  75         return rb_struct_new(cListenStats, active, queued);
  76 }
  77
  78 static int st_free_data(st_data_t key, st_data_t value, st_data_t ignored)
  79 {
  80         xfree((void *)key);
  81         xfree((void *)value);
  82
  83         return ST_DELETE;
  84 }
  85
  86 static int st_to_hash(st_data_t key, st_data_t value, VALUE hash)
  87 {
  88         struct listen_stats *stats = (struct listen_stats *)value;
  89
  90         if (stats->listener_p) {
  91                 VALUE k = rb_str_new2((const char *)key);
  92                 VALUE v = rb_listen_stats(stats);
  93
  94                 OBJ_FREEZE(k);
  95                 rb_hash_aset(hash, k, v);
  96         }
  97         return st_free_data(key, value, 0);
  98 }
  99
 100 static struct listen_stats *stats_for(st_table *table, struct inet_diag_msg *r)
 101 {
 102         char *key, *port;
 103         struct listen_stats *stats;
 104         size_t keylen;
 105         size_t portlen = sizeof("65535");
 106         struct sockaddr_storage ss = { 0 };
 107         socklen_t len = sizeof(struct sockaddr_storage);
 108         int rc;
 109         int flags = NI_NUMERICHOST | NI_NUMERICSERV;
 110
 111         switch ((ss.ss_family = r->idiag_family)) {
 112         case AF_INET: {
 113                 struct sockaddr_in *in = (struct sockaddr_in *)&ss;
 114                 in->sin_port = r->id.idiag_sport;
 115                 in->sin_addr.s_addr = r->id.idiag_src[0];
 116                 keylen = INET_ADDRSTRLEN;
 117                 break;
 118                 }
 119         case AF_INET6: {
 120                 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *)&ss;
 121                 in6->sin6_port = r->id.idiag_sport;
 122                 memcpy(&in6->sin6_addr.in6_u.u6_addr32,
 123                        &r->id.idiag_src, sizeof(__be32[4]));
 124                 keylen = INET6_ADDRSTRLEN;
 125                 break;
 126                 }
 127         default:
 128                 assert(0 && "unsupported address family, could that be IPv7?!");
 129         }
 130         key = alloca(keylen + 1 + portlen);
 131         key[keylen] = 0; /* will be ':' later */
 132         port = key + keylen + 1;
 133         rc = getnameinfo((struct sockaddr *)&ss, len,
 134                          key, keylen, port, portlen, flags);
 135         if (rc != 0) {
 136                 fprintf(stderr, "BUG: getnameinfo: %s\n"
 137                         "Please report how you produced this at %s\n",
 138                         gai_strerror(rc), "raindrops@librelist.com");
 139                 fflush(stderr);
 140                 *key = 0;
 141         }
 142         keylen = strlen(key);
 143         portlen = strlen(port);
 144         key[keylen] = ':';
 145         memmove(key + keylen + 1, port, portlen + 1);
 146         if (!st_lookup(table, (st_data_t)key, (st_data_t *)&stats)) {
 147                 char *old_key = key;
 148
 149                 key = xmalloc(keylen + 1 + portlen + 1);
 150                 memcpy(key, old_key, keylen + 1 + portlen + 1);
 151                 stats = xcalloc(1, sizeof(struct listen_stats));
 152                 st_insert(table, (st_data_t)key, (st_data_t)stats);
 153         }
 154         return stats;
 155 }
 156
 157 static void table_incr_active(st_table *table, struct inet_diag_msg *r)
 158 {
 159         struct listen_stats *stats = stats_for(table, r);
 160         ++stats->active;
 161 }
 162
 163 static void table_set_queued(st_table *table, struct inet_diag_msg *r)
 164 {
 165         struct listen_stats *stats = stats_for(table, r);
 166         stats->listener_p = 1;
 167         stats->queued = r->idiag_rqueue;
 168 }
 169
 170 /* inner loop of inet_diag, called for every socket returned by netlink */
 171 static inline void r_acc(struct nogvl_args *args, struct inet_diag_msg *r)
 172 {
 173         /*
 174          * inode == 0 means the connection is still in the listen queue
 175          * and has not yet been accept()-ed by the server.  The
 176          * inet_diag bytecode cannot filter this for us.
 177          */
 178         if (r->idiag_inode == 0)
 179                 return;
 180         if (r->idiag_state == TCP_ESTABLISHED) {
 181                 if (args->table)
 182                         table_incr_active(args->table, r);
 183                 else
 184                         args->stats.active++;
 185         } else { /* if (r->idiag_state == TCP_LISTEN) */
 186                 if (args->table)
 187                         table_set_queued(args->table, r);
 188                 else
 189                         args->stats.queued = r->idiag_rqueue;
 190         }
 191         /*
 192          * we wont get anything else because of the idiag_states filter
 193          */
 194 }
 195
 196 static const char err_socket[] = "socket";
 197 static const char err_sendmsg[] = "sendmsg";
 198 static const char err_recvmsg[] = "recvmsg";
 199 static const char err_nlmsg[] = "nlmsg";
 200
 201 struct diag_req {
 202         struct nlmsghdr nlh;
 203         struct inet_diag_req r;
 204 };
 205
 206 static void prep_msghdr(
 207         struct msghdr *msg,
 208         struct nogvl_args *args,
 209         struct sockaddr_nl *nladdr,
 210         size_t iovlen)
 211 {
 212         memset(msg, 0, sizeof(struct msghdr));
 213         msg->msg_name = (void *)nladdr;
 214         msg->msg_namelen = sizeof(struct sockaddr_nl);
 215         msg->msg_iov = args->iov;
 216         msg->msg_iovlen = iovlen;
 217 }
 218
 219 static void prep_diag_args(
 220         struct nogvl_args *args,
 221         struct sockaddr_nl *nladdr,
 222         struct rtattr *rta,
 223         struct diag_req *req,
 224         struct msghdr *msg)
 225 {
 226         struct inet_diag_bc_op *op = args->iov[2].iov_base;
 227         struct inet_diag_hostcond *cond = (struct inet_diag_hostcond *)(op + 1);
 228
 229         memset(req, 0, sizeof(struct diag_req));
 230         memset(nladdr, 0, sizeof(struct sockaddr_nl));
 231
 232         nladdr->nl_family = AF_NETLINK;
 233
 234         req->nlh.nlmsg_len = sizeof(struct diag_req) +
 235                             RTA_LENGTH(args->iov[2].iov_len);
 236         req->nlh.nlmsg_type = TCPDIAG_GETSOCK;
 237         req->nlh.nlmsg_flags = NLM_F_ROOT | NLM_F_MATCH | NLM_F_REQUEST;
 238         req->nlh.nlmsg_pid = getpid();
 239         req->r.idiag_states = (1<<TCP_ESTABLISHED) | (1<<TCP_LISTEN);
 240         req->r.idiag_family = cond->family;
 241         rta->rta_type = INET_DIAG_REQ_BYTECODE;
 242         rta->rta_len = RTA_LENGTH(args->iov[2].iov_len);
 243
 244         args->iov[0].iov_base = req;
 245         args->iov[0].iov_len = sizeof(struct diag_req);
 246         args->iov[1].iov_base = rta;
 247         args->iov[1].iov_len = sizeof(struct rtattr);
 248
 249         prep_msghdr(msg, args, nladdr, 3);
 250 }
 251
 252 static void prep_recvmsg_buf(struct nogvl_args *args)
 253 {
 254         /* reuse buffer that was allocated for bytecode */
 255         args->iov[0].iov_len = page_size;
 256         args->iov[0].iov_base = args->iov[2].iov_base;
 257 }
 258
 259 /* does the inet_diag stuff with netlink(), this is called w/o GVL */
 260 static VALUE diag(void *ptr)
 261 {
 262         struct nogvl_args *args = ptr;
 263         struct sockaddr_nl nladdr;
 264         struct rtattr rta;
 265         struct diag_req req;
 266         struct msghdr msg;
 267         const char *err = NULL;
 268         unsigned seq = ++g_seq;
 269         int fd = socket(AF_NETLINK, SOCK_RAW, NETLINK_INET_DIAG);
 270
 271         if (fd < 0)
 272                 return (VALUE)err_socket;
 273
 274         prep_diag_args(args, &nladdr, &rta, &req, &msg);
 275         req.nlh.nlmsg_seq = seq;
 276
 277         if (sendmsg(fd, &msg, 0) < 0) {
 278                 err = err_sendmsg;
 279                 goto out;
 280         }
 281
 282         prep_recvmsg_buf(args);
 283
 284         while (1) {
 285                 ssize_t readed;
 286                 struct nlmsghdr *h = (struct nlmsghdr *)args->iov[0].iov_base;
 287
 288                 prep_msghdr(&msg, args, &nladdr, 1);
 289                 readed = recvmsg(fd, &msg, 0);
 290                 if (readed < 0) {
 291                         if (errno == EINTR)
 292                                 continue;
 293                         err = err_recvmsg;
 294                         goto out;
 295                 }
 296                 if (readed == 0)
 297                         goto out;
 298
 299                 for ( ; NLMSG_OK(h, readed); h = NLMSG_NEXT(h, readed)) {
 300                         if (h->nlmsg_seq != seq)
 301                                 continue;
 302                         if (h->nlmsg_type == NLMSG_DONE)
 303                                 goto out;
 304                         if (h->nlmsg_type == NLMSG_ERROR) {
 305                                 err = err_nlmsg;
 306                                 goto out;
 307                         }
 308                         r_acc(args, NLMSG_DATA(h));
 309                 }
 310         }
 311 out:
 312         {
 313                 int save_errno = errno;
 314                 close(fd);
 315                 if (err && args->table) {
 316                         st_foreach(args->table, st_free_data, 0);
 317                         st_free_table(args->table);
 318                 }
 319                 errno = save_errno;
 320         }
 321         return (VALUE)err;
 322 }
 323
 324 /* populates sockaddr_storage struct by parsing +addr+ */
 325 static void parse_addr(struct sockaddr_storage *inet, VALUE addr)
 326 {
 327         char *host_ptr;
 328         char *colon = NULL;
 329         char *rbracket = NULL;
 330         long host_len;
 331         struct addrinfo hints;
 332         struct addrinfo *res;
 333         int rc;
 334
 335         if (TYPE(addr) != T_STRING)
 336                 rb_raise(rb_eArgError, "addrs must be an Array of Strings");
 337
 338         host_ptr = StringValueCStr(addr);
 339         host_len = RSTRING_LEN(addr);
 340         if (*host_ptr == '[') { /* ipv6 address format (rfc2732) */
 341                 rbracket = memchr(host_ptr + 1, ']', host_len - 1);
 342
 343                 if (rbracket == NULL)
 344                         rb_raise(rb_eArgError, "']' not found in IPv6 addr=%s",
 345                                  host_ptr);
 346                 if (rbracket[1] != ':')
 347                         rb_raise(rb_eArgError, "':' not found in IPv6 addr=%s",
 348                                  host_ptr);
 349                 colon = rbracket + 1;
 350                 host_ptr++;
 351                 *rbracket = 0;
 352         } else { /* ipv4 */
 353                 colon = memchr(host_ptr, ':', host_len);
 354         }
 355
 356         if (!colon)
 357                 rb_raise(rb_eArgError, "port not found in: `%s'", host_ptr);
 358
 359         hints.ai_family = AF_UNSPEC;
 360         hints.ai_socktype = SOCK_STREAM;
 361         hints.ai_protocol = IPPROTO_TCP;
 362         hints.ai_flags = AI_NUMERICHOST | AI_NUMERICSERV;
 363
 364         *colon = 0;
 365         rc = getaddrinfo(host_ptr, colon + 1, &hints, &res);
 366         *colon = ':';
 367         if (rbracket) *rbracket = ']';
 368         if (rc != 0)
 369                 rb_raise(rb_eArgError, "getaddrinfo(%s): %s",
 370                          host_ptr, gai_strerror(rc));
 371
 372         memcpy(inet, res->ai_addr, res->ai_addrlen);
 373         freeaddrinfo(res);
 374 }
 375
 376 /* generates inet_diag bytecode to match all addrs for a given family */
 377 static void gen_bytecode_all(struct iovec *iov, sa_family_t family)
 378 {
 379         struct inet_diag_bc_op *op;
 380         struct inet_diag_hostcond *cond;
 381
 382         /* iov_len was already set and base allocated in a parent function */
 383         assert(iov->iov_len == OPLEN && iov->iov_base && "iov invalid");
 384         op = iov->iov_base;
 385         op->code = INET_DIAG_BC_S_COND;
 386         op->yes = OPLEN;
 387         op->no = sizeof(struct inet_diag_bc_op) + OPLEN;
 388         cond = (struct inet_diag_hostcond *)(op + 1);
 389         cond->family = family;
 390         cond->port = -1;
 391         cond->prefix_len = 0;
 392 }
 393
 394 /* generates inet_diag bytecode to match a single addr */
 395 static void gen_bytecode(struct iovec *iov, struct sockaddr_storage *inet)
 396 {
 397         struct inet_diag_bc_op *op;
 398         struct inet_diag_hostcond *cond;
 399
 400         /* iov_len was already set and base allocated in a parent function */
 401         assert(iov->iov_len == OPLEN && iov->iov_base && "iov invalid");
 402         op = iov->iov_base;
 403         op->code = INET_DIAG_BC_S_COND;
 404         op->yes = OPLEN;
 405         op->no = sizeof(struct inet_diag_bc_op) + OPLEN;
 406
 407         cond = (struct inet_diag_hostcond *)(op + 1);
 408         cond->family = inet->ss_family;
 409         switch (inet->ss_family) {
 410         case AF_INET: {
 411                 struct sockaddr_in *in = (struct sockaddr_in *)inet;
 412
 413                 cond->port = ntohs(in->sin_port);
 414                 cond->prefix_len = in->sin_addr.s_addr == 0 ? 0 :
 415                                    sizeof(in->sin_addr.s_addr) * CHAR_BIT;
 416                 *cond->addr = in->sin_addr.s_addr;
 417                 }
 418                 break;
 419         case AF_INET6: {
 420                 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *)inet;
 421
 422                 cond->port = ntohs(in6->sin6_port);
 423                 cond->prefix_len = memcmp(&in6addr_any, &in6->sin6_addr,
 424                                           sizeof(struct in6_addr)) == 0 ?
 425                                   0 : sizeof(in6->sin6_addr) * CHAR_BIT;
 426                 memcpy(&cond->addr, &in6->sin6_addr, sizeof(struct in6_addr));
 427                 }
 428                 break;
 429         default:
 430                 assert(0 && "unsupported address family, could that be IPv7?!");
 431         }
 432 }
 433
 434 static void nl_errcheck(VALUE r)
 435 {
 436         const char *err = (const char *)r;
 437
 438         if (err) {
 439                 if (err == err_nlmsg)
 440                         rb_raise(rb_eRuntimeError, "NLMSG_ERROR");
 441                 else
 442                         rb_sys_fail(err);
 443         }
 444 }
 445
 446 static VALUE tcp_stats(struct nogvl_args *args, VALUE addr)
 447 {
 448         struct sockaddr_storage query_addr;
 449
 450         parse_addr(&query_addr, addr);
 451         gen_bytecode(&args->iov[2], &query_addr);
 452
 453         memset(&args->stats, 0, sizeof(struct listen_stats));
 454         nl_errcheck(rb_thread_blocking_region(diag, args, 0, 0));
 455
 456         return rb_listen_stats(&args->stats);
 457 }
 458
 459 /*
 460  * call-seq:
 461  *      addrs = %w(0.0.0.0:80 127.0.0.1:8080)
 462  *      Raindrops::Linux.tcp_listener_stats(addrs) => hash
 463  *
 464  * Takes an array of strings representing listen addresses to filter for.
 465  * Returns a hash with given addresses as keys and ListenStats
 466  * objects as the values.
 467  */
 468 static VALUE tcp_listener_stats(VALUE obj, VALUE addrs)
 469 {
 470         VALUE *ary;
 471         long i;
 472         VALUE rv;
 473         struct nogvl_args args;
 474
 475         /*
 476          * allocating page_size instead of OP_LEN since we'll reuse the
 477          * buffer for recvmsg() later, we already checked for
 478          * OPLEN <= page_size at initialization
 479          */
 480         args.iov[2].iov_len = OPLEN;
 481         args.iov[2].iov_base = alloca(page_size);
 482         args.table = NULL;
 483
 484         if (TYPE(addrs) != T_ARRAY)
 485                 rb_raise(rb_eArgError, "addrs must be an Array of Strings");
 486
 487         rv = rb_hash_new();
 488         ary = RARRAY_PTR(addrs);
 489         for (i = RARRAY_LEN(addrs); --i >= 0; ary++)
 490                 rb_hash_aset(rv, *ary, tcp_stats(&args, *ary));
 491
 492         return rv;
 493 }
 494
 495 static VALUE all_tcp_listener_stats(VALUE obj)
 496 {
 497         VALUE rv;
 498         struct nogvl_args args;
 499
 500         /*
 501          * allocating page_size instead of OP_LEN since we'll reuse the
 502          * buffer for recvmsg() later, we already checked for
 503          * OPLEN <= page_size at initialization
 504          */
 505         args.iov[2].iov_len = OPLEN;
 506         args.iov[2].iov_base = alloca(page_size);
 507         args.table = st_init_strtable();
 508         gen_bytecode_all(&args.iov[2], AF_INET);
 509
 510         nl_errcheck(rb_thread_blocking_region(diag, &args, NULL, 0));
 511         rv = rb_hash_new();
 512         st_foreach(args.table, st_to_hash, rv);
 513         st_free_table(args.table);
 514         return rv;
 515 }
 516
 517 void Init_raindrops_linux_inet_diag(void)
 518 {
 519         VALUE cRaindrops = rb_const_get(rb_cObject, rb_intern("Raindrops"));
 520         VALUE mLinux = rb_define_module_under(cRaindrops, "Linux");
 521
 522         cListenStats = rb_const_get(cRaindrops, rb_intern("ListenStats"));
 523
 524         rb_define_module_function(mLinux, "tcp_listener_stats",
 525                                   tcp_listener_stats, 1);
 526         rb_define_module_function(mLinux, "all_tcp_listener_stats",
 527                                   all_tcp_listener_stats, 0);
 528
 529         page_size = getpagesize();
 530
 531         assert(OPLEN <= page_size && "bytecode OPLEN is not <= PAGE_SIZE");
 532 }
 533 #endif /* __linux__ */