net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  *      This program is free software; you can redistribute it and/or
   7  *      modify it under the terms of the GNU General Public License
   8  *      as published by the Free Software Foundation; either version
   9  *      2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/types.h>
  15 #include <linux/fcntl.h>
  16 #include <linux/random.h>
  17 #include <linux/slab.h>
  18 #include <linux/cache.h>
  19 #include <linux/init.h>
  20 #include <linux/time.h>
  21
  22 #include <net/icmp.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/inet_common.h>
  26 #include <net/inet_connection_sock.h>
  27 #include <net/inet_hashtables.h>
  28 #include <net/inet_timewait_sock.h>
  29 #include <net/inet6_hashtables.h>
  30 #include <net/netlink.h>
  31
  32 #include <linux/inet.h>
  33 #include <linux/stddef.h>
  34
  35 #include <linux/inet_diag.h>
  36
  37 static const struct inet_diag_handler **inet_diag_table;
  38
  39 struct inet_diag_entry {
  40         __be32 *saddr;
  41         __be32 *daddr;
  42         u16 sport;
  43         u16 dport;
  44         u16 family;
  45         u16 userlocks;
  46 };
  47
  48 static struct sock *sdiagnl;
  49
  50 #define INET_DIAG_PUT(skb, attrtype, attrlen) \
  51         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  52
  53 static DEFINE_MUTEX(inet_diag_table_mutex);
  54
  55 static const struct inet_diag_handler *inet_diag_lock_handler(int type)
  56 {
  57         if (!inet_diag_table[type])
  58                 request_module("net-pf-%d-proto-%d-type-%d", PF_NETLINK,
  59                                NETLINK_SOCK_DIAG, type);
  60
  61         mutex_lock(&inet_diag_table_mutex);
  62         if (!inet_diag_table[type])
  63                 return ERR_PTR(-ENOENT);
  64
  65         return inet_diag_table[type];
  66 }
  67
  68 static inline void inet_diag_unlock_handler(
  69         const struct inet_diag_handler *handler)
  70 {
  71         mutex_unlock(&inet_diag_table_mutex);
  72 }
  73
  74 static int inet_csk_diag_fill(struct sock *sk,
  75                               struct sk_buff *skb,
  76                               int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  77                               const struct nlmsghdr *unlh)
  78 {
  79         const struct inet_sock *inet = inet_sk(sk);
  80         const struct inet_connection_sock *icsk = inet_csk(sk);
  81         struct inet_diag_msg *r;
  82         struct nlmsghdr  *nlh;
  83         void *info = NULL;
  84         struct inet_diag_meminfo  *minfo = NULL;
  85         unsigned char    *b = skb_tail_pointer(skb);
  86         const struct inet_diag_handler *handler;
  87
  88         handler = inet_diag_table[unlh->nlmsg_type];
  89         BUG_ON(handler == NULL);
  90
  91         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
  92         nlh->nlmsg_flags = nlmsg_flags;
  93
  94         r = NLMSG_DATA(nlh);
  95         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  96
  97         if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
  98                 minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO, sizeof(*minfo));
  99
 100         if (ext & (1 << (INET_DIAG_INFO - 1)))
 101                 info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
 102                                      handler->idiag_info_size);
 103
 104         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
 105                 const size_t len = strlen(icsk->icsk_ca_ops->name);
 106
 107                 strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
 108                        icsk->icsk_ca_ops->name);
 109         }
 110
 111         r->idiag_family = sk->sk_family;
 112         r->idiag_state = sk->sk_state;
 113         r->idiag_timer = 0;
 114         r->idiag_retrans = 0;
 115
 116         r->id.idiag_if = sk->sk_bound_dev_if;
 117         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
 118         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 119
 120         r->id.idiag_sport = inet->inet_sport;
 121         r->id.idiag_dport = inet->inet_dport;
 122         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 123         r->id.idiag_dst[0] = inet->inet_daddr;
 124
 125         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 126          * hence this needs to be included regardless of socket family.
 127          */
 128         if (ext & (1 << (INET_DIAG_TOS - 1)))
 129                 RTA_PUT_U8(skb, INET_DIAG_TOS, inet->tos);
 130
 131 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 132         if (r->idiag_family == AF_INET6) {
 133                 const struct ipv6_pinfo *np = inet6_sk(sk);
 134
 135                 *(struct in6_addr *)r->id.idiag_src = np->rcv_saddr;
 136                 *(struct in6_addr *)r->id.idiag_dst = np->daddr;
 137                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 138                         RTA_PUT_U8(skb, INET_DIAG_TCLASS, np->tclass);
 139         }
 140 #endif
 141
 142 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 143
 144         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 145                 r->idiag_timer = 1;
 146                 r->idiag_retrans = icsk->icsk_retransmits;
 147                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 148         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 149                 r->idiag_timer = 4;
 150                 r->idiag_retrans = icsk->icsk_probes_out;
 151                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 152         } else if (timer_pending(&sk->sk_timer)) {
 153                 r->idiag_timer = 2;
 154                 r->idiag_retrans = icsk->icsk_probes_out;
 155                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 156         } else {
 157                 r->idiag_timer = 0;
 158                 r->idiag_expires = 0;
 159         }
 160 #undef EXPIRES_IN_MS
 161
 162         r->idiag_uid = sock_i_uid(sk);
 163         r->idiag_inode = sock_i_ino(sk);
 164
 165         if (minfo) {
 166                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 167                 minfo->idiag_wmem = sk->sk_wmem_queued;
 168                 minfo->idiag_fmem = sk->sk_forward_alloc;
 169                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 170         }
 171
 172         handler->idiag_get_info(sk, r, info);
 173
 174         if (sk->sk_state < TCP_TIME_WAIT &&
 175             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 176                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 177
 178         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 179         return skb->len;
 180
 181 rtattr_failure:
 182 nlmsg_failure:
 183         nlmsg_trim(skb, b);
 184         return -EMSGSIZE;
 185 }
 186
 187 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 188                                struct sk_buff *skb, int ext, u32 pid,
 189                                u32 seq, u16 nlmsg_flags,
 190                                const struct nlmsghdr *unlh)
 191 {
 192         long tmo;
 193         struct inet_diag_msg *r;
 194         const unsigned char *previous_tail = skb_tail_pointer(skb);
 195         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 196                                          unlh->nlmsg_type, sizeof(*r));
 197
 198         r = NLMSG_DATA(nlh);
 199         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 200
 201         nlh->nlmsg_flags = nlmsg_flags;
 202
 203         tmo = tw->tw_ttd - jiffies;
 204         if (tmo < 0)
 205                 tmo = 0;
 206
 207         r->idiag_family       = tw->tw_family;
 208         r->idiag_retrans      = 0;
 209         r->id.idiag_if        = tw->tw_bound_dev_if;
 210         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 211         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 212         r->id.idiag_sport     = tw->tw_sport;
 213         r->id.idiag_dport     = tw->tw_dport;
 214         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 215         r->id.idiag_dst[0]    = tw->tw_daddr;
 216         r->idiag_state        = tw->tw_substate;
 217         r->idiag_timer        = 3;
 218         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 219         r->idiag_rqueue       = 0;
 220         r->idiag_wqueue       = 0;
 221         r->idiag_uid          = 0;
 222         r->idiag_inode        = 0;
 223 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 224         if (tw->tw_family == AF_INET6) {
 225                 const struct inet6_timewait_sock *tw6 =
 226                                                 inet6_twsk((struct sock *)tw);
 227
 228                 *(struct in6_addr *)r->id.idiag_src = tw6->tw_v6_rcv_saddr;
 229                 *(struct in6_addr *)r->id.idiag_dst = tw6->tw_v6_daddr;
 230         }
 231 #endif
 232         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 233         return skb->len;
 234 nlmsg_failure:
 235         nlmsg_trim(skb, previous_tail);
 236         return -EMSGSIZE;
 237 }
 238
 239 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 240                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 241                         const struct nlmsghdr *unlh)
 242 {
 243         if (sk->sk_state == TCP_TIME_WAIT)
 244                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 245                                            skb, ext, pid, seq, nlmsg_flags,
 246                                            unlh);
 247         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 248 }
 249
 250 static int inet_diag_get_exact(struct sk_buff *in_skb,
 251                                const struct nlmsghdr *nlh)
 252 {
 253         int err;
 254         struct sock *sk;
 255         struct inet_diag_req *req = NLMSG_DATA(nlh);
 256         struct sk_buff *rep;
 257         struct inet_hashinfo *hashinfo;
 258         const struct inet_diag_handler *handler;
 259
 260         handler = inet_diag_lock_handler(nlh->nlmsg_type);
 261         if (IS_ERR(handler)) {
 262                 err = PTR_ERR(handler);
 263                 goto unlock;
 264         }
 265
 266         hashinfo = handler->idiag_hashinfo;
 267         err = -EINVAL;
 268
 269         if (req->idiag_family == AF_INET) {
 270                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 271                                  req->id.idiag_dport, req->id.idiag_src[0],
 272                                  req->id.idiag_sport, req->id.idiag_if);
 273         }
 274 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 275         else if (req->idiag_family == AF_INET6) {
 276                 sk = inet6_lookup(&init_net, hashinfo,
 277                                   (struct in6_addr *)req->id.idiag_dst,
 278                                   req->id.idiag_dport,
 279                                   (struct in6_addr *)req->id.idiag_src,
 280                                   req->id.idiag_sport,
 281                                   req->id.idiag_if);
 282         }
 283 #endif
 284         else {
 285                 goto unlock;
 286         }
 287
 288         err = -ENOENT;
 289         if (sk == NULL)
 290                 goto unlock;
 291
 292         err = -ESTALE;
 293         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 294              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 295             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 296              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 297                 goto out;
 298
 299         err = -ENOMEM;
 300         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 301                                      sizeof(struct inet_diag_meminfo) +
 302                                      handler->idiag_info_size + 64)),
 303                         GFP_KERNEL);
 304         if (!rep)
 305                 goto out;
 306
 307         err = sk_diag_fill(sk, rep, req->idiag_ext,
 308                            NETLINK_CB(in_skb).pid,
 309                            nlh->nlmsg_seq, 0, nlh);
 310         if (err < 0) {
 311                 WARN_ON(err == -EMSGSIZE);
 312                 kfree_skb(rep);
 313                 goto out;
 314         }
 315         err = netlink_unicast(sdiagnl, rep, NETLINK_CB(in_skb).pid,
 316                               MSG_DONTWAIT);
 317         if (err > 0)
 318                 err = 0;
 319
 320 out:
 321         if (sk) {
 322                 if (sk->sk_state == TCP_TIME_WAIT)
 323                         inet_twsk_put((struct inet_timewait_sock *)sk);
 324                 else
 325                         sock_put(sk);
 326         }
 327 unlock:
 328         inet_diag_unlock_handler(handler);
 329         return err;
 330 }
 331
 332 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 333 {
 334         int words = bits >> 5;
 335
 336         bits &= 0x1f;
 337
 338         if (words) {
 339                 if (memcmp(a1, a2, words << 2))
 340                         return 0;
 341         }
 342         if (bits) {
 343                 __be32 w1, w2;
 344                 __be32 mask;
 345
 346                 w1 = a1[words];
 347                 w2 = a2[words];
 348
 349                 mask = htonl((0xffffffff) << (32 - bits));
 350
 351                 if ((w1 ^ w2) & mask)
 352                         return 0;
 353         }
 354
 355         return 1;
 356 }
 357
 358
 359 static int inet_diag_bc_run(const void *bc, int len,
 360                             const struct inet_diag_entry *entry)
 361 {
 362         while (len > 0) {
 363                 int yes = 1;
 364                 const struct inet_diag_bc_op *op = bc;
 365
 366                 switch (op->code) {
 367                 case INET_DIAG_BC_NOP:
 368                         break;
 369                 case INET_DIAG_BC_JMP:
 370                         yes = 0;
 371                         break;
 372                 case INET_DIAG_BC_S_GE:
 373                         yes = entry->sport >= op[1].no;
 374                         break;
 375                 case INET_DIAG_BC_S_LE:
 376                         yes = entry->sport <= op[1].no;
 377                         break;
 378                 case INET_DIAG_BC_D_GE:
 379                         yes = entry->dport >= op[1].no;
 380                         break;
 381                 case INET_DIAG_BC_D_LE:
 382                         yes = entry->dport <= op[1].no;
 383                         break;
 384                 case INET_DIAG_BC_AUTO:
 385                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 386                         break;
 387                 case INET_DIAG_BC_S_COND:
 388                 case INET_DIAG_BC_D_COND: {
 389                         struct inet_diag_hostcond *cond;
 390                         __be32 *addr;
 391
 392                         cond = (struct inet_diag_hostcond *)(op + 1);
 393                         if (cond->port != -1 &&
 394                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 395                                              entry->sport : entry->dport)) {
 396                                 yes = 0;
 397                                 break;
 398                         }
 399
 400                         if (cond->prefix_len == 0)
 401                                 break;
 402
 403                         if (op->code == INET_DIAG_BC_S_COND)
 404                                 addr = entry->saddr;
 405                         else
 406                                 addr = entry->daddr;
 407
 408                         if (bitstring_match(addr, cond->addr,
 409                                             cond->prefix_len))
 410                                 break;
 411                         if (entry->family == AF_INET6 &&
 412                             cond->family == AF_INET) {
 413                                 if (addr[0] == 0 && addr[1] == 0 &&
 414                                     addr[2] == htonl(0xffff) &&
 415                                     bitstring_match(addr + 3, cond->addr,
 416                                                     cond->prefix_len))
 417                                         break;
 418                         }
 419                         yes = 0;
 420                         break;
 421                 }
 422                 }
 423
 424                 if (yes) {
 425                         len -= op->yes;
 426                         bc += op->yes;
 427                 } else {
 428                         len -= op->no;
 429                         bc += op->no;
 430                 }
 431         }
 432         return len == 0;
 433 }
 434
 435 static int valid_cc(const void *bc, int len, int cc)
 436 {
 437         while (len >= 0) {
 438                 const struct inet_diag_bc_op *op = bc;
 439
 440                 if (cc > len)
 441                         return 0;
 442                 if (cc == len)
 443                         return 1;
 444                 if (op->yes < 4 || op->yes & 3)
 445                         return 0;
 446                 len -= op->yes;
 447                 bc  += op->yes;
 448         }
 449         return 0;
 450 }
 451
 452 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 453 {
 454         const void *bc = bytecode;
 455         int  len = bytecode_len;
 456
 457         while (len > 0) {
 458                 const struct inet_diag_bc_op *op = bc;
 459
 460 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 461                 switch (op->code) {
 462                 case INET_DIAG_BC_AUTO:
 463                 case INET_DIAG_BC_S_COND:
 464                 case INET_DIAG_BC_D_COND:
 465                 case INET_DIAG_BC_S_GE:
 466                 case INET_DIAG_BC_S_LE:
 467                 case INET_DIAG_BC_D_GE:
 468                 case INET_DIAG_BC_D_LE:
 469                 case INET_DIAG_BC_JMP:
 470                         if (op->no < 4 || op->no > len + 4 || op->no & 3)
 471                                 return -EINVAL;
 472                         if (op->no < len &&
 473                             !valid_cc(bytecode, bytecode_len, len - op->no))
 474                                 return -EINVAL;
 475                         break;
 476                 case INET_DIAG_BC_NOP:
 477                         break;
 478                 default:
 479                         return -EINVAL;
 480                 }
 481                 if (op->yes < 4 || op->yes > len + 4 || op->yes & 3)
 482                         return -EINVAL;
 483                 bc  += op->yes;
 484                 len -= op->yes;
 485         }
 486         return len == 0 ? 0 : -EINVAL;
 487 }
 488
 489 static int inet_csk_diag_dump(struct sock *sk,
 490                               struct sk_buff *skb,
 491                               struct netlink_callback *cb)
 492 {
 493         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 494
 495         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 496                 struct inet_diag_entry entry;
 497                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 498                                                           sizeof(*r),
 499                                                           INET_DIAG_REQ_BYTECODE);
 500                 struct inet_sock *inet = inet_sk(sk);
 501
 502                 entry.family = sk->sk_family;
 503 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 504                 if (entry.family == AF_INET6) {
 505                         struct ipv6_pinfo *np = inet6_sk(sk);
 506
 507                         entry.saddr = np->rcv_saddr.s6_addr32;
 508                         entry.daddr = np->daddr.s6_addr32;
 509                 } else
 510 #endif
 511                 {
 512                         entry.saddr = &inet->inet_rcv_saddr;
 513                         entry.daddr = &inet->inet_daddr;
 514                 }
 515                 entry.sport = inet->inet_num;
 516                 entry.dport = ntohs(inet->inet_dport);
 517                 entry.userlocks = sk->sk_userlocks;
 518
 519                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 520                         return 0;
 521         }
 522
 523         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 524                                   NETLINK_CB(cb->skb).pid,
 525                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 526 }
 527
 528 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 529                                struct sk_buff *skb,
 530                                struct netlink_callback *cb)
 531 {
 532         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 533
 534         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 535                 struct inet_diag_entry entry;
 536                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 537                                                           sizeof(*r),
 538                                                           INET_DIAG_REQ_BYTECODE);
 539
 540                 entry.family = tw->tw_family;
 541 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 542                 if (tw->tw_family == AF_INET6) {
 543                         struct inet6_timewait_sock *tw6 =
 544                                                 inet6_twsk((struct sock *)tw);
 545                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 546                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 547                 } else
 548 #endif
 549                 {
 550                         entry.saddr = &tw->tw_rcv_saddr;
 551                         entry.daddr = &tw->tw_daddr;
 552                 }
 553                 entry.sport = tw->tw_num;
 554                 entry.dport = ntohs(tw->tw_dport);
 555                 entry.userlocks = 0;
 556
 557                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 558                         return 0;
 559         }
 560
 561         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 562                                    NETLINK_CB(cb->skb).pid,
 563                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 564 }
 565
 566 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 567                               struct request_sock *req, u32 pid, u32 seq,
 568                               const struct nlmsghdr *unlh)
 569 {
 570         const struct inet_request_sock *ireq = inet_rsk(req);
 571         struct inet_sock *inet = inet_sk(sk);
 572         unsigned char *b = skb_tail_pointer(skb);
 573         struct inet_diag_msg *r;
 574         struct nlmsghdr *nlh;
 575         long tmo;
 576
 577         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 578         nlh->nlmsg_flags = NLM_F_MULTI;
 579         r = NLMSG_DATA(nlh);
 580
 581         r->idiag_family = sk->sk_family;
 582         r->idiag_state = TCP_SYN_RECV;
 583         r->idiag_timer = 1;
 584         r->idiag_retrans = req->retrans;
 585
 586         r->id.idiag_if = sk->sk_bound_dev_if;
 587         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 588         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 589
 590         tmo = req->expires - jiffies;
 591         if (tmo < 0)
 592                 tmo = 0;
 593
 594         r->id.idiag_sport = inet->inet_sport;
 595         r->id.idiag_dport = ireq->rmt_port;
 596         r->id.idiag_src[0] = ireq->loc_addr;
 597         r->id.idiag_dst[0] = ireq->rmt_addr;
 598         r->idiag_expires = jiffies_to_msecs(tmo);
 599         r->idiag_rqueue = 0;
 600         r->idiag_wqueue = 0;
 601         r->idiag_uid = sock_i_uid(sk);
 602         r->idiag_inode = 0;
 603 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 604         if (r->idiag_family == AF_INET6) {
 605                 *(struct in6_addr *)r->id.idiag_src = inet6_rsk(req)->loc_addr;
 606                 *(struct in6_addr *)r->id.idiag_dst = inet6_rsk(req)->rmt_addr;
 607         }
 608 #endif
 609         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 610
 611         return skb->len;
 612
 613 nlmsg_failure:
 614         nlmsg_trim(skb, b);
 615         return -1;
 616 }
 617
 618 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 619                                struct netlink_callback *cb)
 620 {
 621         struct inet_diag_entry entry;
 622         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 623         struct inet_connection_sock *icsk = inet_csk(sk);
 624         struct listen_sock *lopt;
 625         const struct nlattr *bc = NULL;
 626         struct inet_sock *inet = inet_sk(sk);
 627         int j, s_j;
 628         int reqnum, s_reqnum;
 629         int err = 0;
 630
 631         s_j = cb->args[3];
 632         s_reqnum = cb->args[4];
 633
 634         if (s_j > 0)
 635                 s_j--;
 636
 637         entry.family = sk->sk_family;
 638
 639         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 640
 641         lopt = icsk->icsk_accept_queue.listen_opt;
 642         if (!lopt || !lopt->qlen)
 643                 goto out;
 644
 645         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 646                 bc = nlmsg_find_attr(cb->nlh, sizeof(*r),
 647                                      INET_DIAG_REQ_BYTECODE);
 648                 entry.sport = inet->inet_num;
 649                 entry.userlocks = sk->sk_userlocks;
 650         }
 651
 652         for (j = s_j; j < lopt->nr_table_entries; j++) {
 653                 struct request_sock *req, *head = lopt->syn_table[j];
 654
 655                 reqnum = 0;
 656                 for (req = head; req; reqnum++, req = req->dl_next) {
 657                         struct inet_request_sock *ireq = inet_rsk(req);
 658
 659                         if (reqnum < s_reqnum)
 660                                 continue;
 661                         if (r->id.idiag_dport != ireq->rmt_port &&
 662                             r->id.idiag_dport)
 663                                 continue;
 664
 665                         if (bc) {
 666                                 entry.saddr =
 667 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 668                                         (entry.family == AF_INET6) ?
 669                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 670 #endif
 671                                         &ireq->loc_addr;
 672                                 entry.daddr =
 673 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 674                                         (entry.family == AF_INET6) ?
 675                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 676 #endif
 677                                         &ireq->rmt_addr;
 678                                 entry.dport = ntohs(ireq->rmt_port);
 679
 680                                 if (!inet_diag_bc_run(nla_data(bc),
 681                                                       nla_len(bc), &entry))
 682                                         continue;
 683                         }
 684
 685                         err = inet_diag_fill_req(skb, sk, req,
 686                                                NETLINK_CB(cb->skb).pid,
 687                                                cb->nlh->nlmsg_seq, cb->nlh);
 688                         if (err < 0) {
 689                                 cb->args[3] = j + 1;
 690                                 cb->args[4] = reqnum;
 691                                 goto out;
 692                         }
 693                 }
 694
 695                 s_reqnum = 0;
 696         }
 697
 698 out:
 699         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 700
 701         return err;
 702 }
 703
 704 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 705 {
 706         int i, num;
 707         int s_i, s_num;
 708         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 709         const struct inet_diag_handler *handler;
 710         struct inet_hashinfo *hashinfo;
 711
 712         handler = inet_diag_lock_handler(cb->nlh->nlmsg_type);
 713         if (IS_ERR(handler))
 714                 goto unlock;
 715
 716         hashinfo = handler->idiag_hashinfo;
 717
 718         s_i = cb->args[1];
 719         s_num = num = cb->args[2];
 720
 721         if (cb->args[0] == 0) {
 722                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 723                         goto skip_listen_ht;
 724
 725                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 726                         struct sock *sk;
 727                         struct hlist_nulls_node *node;
 728                         struct inet_listen_hashbucket *ilb;
 729
 730                         num = 0;
 731                         ilb = &hashinfo->listening_hash[i];
 732                         spin_lock_bh(&ilb->lock);
 733                         sk_nulls_for_each(sk, node, &ilb->head) {
 734                                 struct inet_sock *inet = inet_sk(sk);
 735
 736                                 if (num < s_num) {
 737                                         num++;
 738                                         continue;
 739                                 }
 740
 741                                 if (r->id.idiag_sport != inet->inet_sport &&
 742                                     r->id.idiag_sport)
 743                                         goto next_listen;
 744
 745                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 746                                     r->id.idiag_dport ||
 747                                     cb->args[3] > 0)
 748                                         goto syn_recv;
 749
 750                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 751                                         spin_unlock_bh(&ilb->lock);
 752                                         goto done;
 753                                 }
 754
 755 syn_recv:
 756                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 757                                         goto next_listen;
 758
 759                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 760                                         spin_unlock_bh(&ilb->lock);
 761                                         goto done;
 762                                 }
 763
 764 next_listen:
 765                                 cb->args[3] = 0;
 766                                 cb->args[4] = 0;
 767                                 ++num;
 768                         }
 769                         spin_unlock_bh(&ilb->lock);
 770
 771                         s_num = 0;
 772                         cb->args[3] = 0;
 773                         cb->args[4] = 0;
 774                 }
 775 skip_listen_ht:
 776                 cb->args[0] = 1;
 777                 s_i = num = s_num = 0;
 778         }
 779
 780         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 781                 goto unlock;
 782
 783         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 784                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 785                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 786                 struct sock *sk;
 787                 struct hlist_nulls_node *node;
 788
 789                 num = 0;
 790
 791                 if (hlist_nulls_empty(&head->chain) &&
 792                         hlist_nulls_empty(&head->twchain))
 793                         continue;
 794
 795                 if (i > s_i)
 796                         s_num = 0;
 797
 798                 spin_lock_bh(lock);
 799                 sk_nulls_for_each(sk, node, &head->chain) {
 800                         struct inet_sock *inet = inet_sk(sk);
 801
 802                         if (num < s_num)
 803                                 goto next_normal;
 804                         if (!(r->idiag_states & (1 << sk->sk_state)))
 805                                 goto next_normal;
 806                         if (r->id.idiag_sport != inet->inet_sport &&
 807                             r->id.idiag_sport)
 808                                 goto next_normal;
 809                         if (r->id.idiag_dport != inet->inet_dport &&
 810                             r->id.idiag_dport)
 811                                 goto next_normal;
 812                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 813                                 spin_unlock_bh(lock);
 814                                 goto done;
 815                         }
 816 next_normal:
 817                         ++num;
 818                 }
 819
 820                 if (r->idiag_states & TCPF_TIME_WAIT) {
 821                         struct inet_timewait_sock *tw;
 822
 823                         inet_twsk_for_each(tw, node,
 824                                     &head->twchain) {
 825
 826                                 if (num < s_num)
 827                                         goto next_dying;
 828                                 if (r->id.idiag_sport != tw->tw_sport &&
 829                                     r->id.idiag_sport)
 830                                         goto next_dying;
 831                                 if (r->id.idiag_dport != tw->tw_dport &&
 832                                     r->id.idiag_dport)
 833                                         goto next_dying;
 834                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 835                                         spin_unlock_bh(lock);
 836                                         goto done;
 837                                 }
 838 next_dying:
 839                                 ++num;
 840                         }
 841                 }
 842                 spin_unlock_bh(lock);
 843         }
 844
 845 done:
 846         cb->args[1] = i;
 847         cb->args[2] = num;
 848 unlock:
 849         inet_diag_unlock_handler(handler);
 850         return skb->len;
 851 }
 852
 853 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 854 {
 855         int hdrlen = sizeof(struct inet_diag_req);
 856
 857         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 858             nlmsg_len(nlh) < hdrlen)
 859                 return -EINVAL;
 860
 861         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 862                 if (nlmsg_attrlen(nlh, hdrlen)) {
 863                         struct nlattr *attr;
 864
 865                         attr = nlmsg_find_attr(nlh, hdrlen,
 866                                                INET_DIAG_REQ_BYTECODE);
 867                         if (attr == NULL ||
 868                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 869                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 870                                 return -EINVAL;
 871                 }
 872
 873                 return netlink_dump_start(sdiagnl, skb, nlh,
 874                                           inet_diag_dump, NULL, 0);
 875         }
 876
 877         return inet_diag_get_exact(skb, nlh);
 878 }
 879
 880 static int sock_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 881 {
 882         return inet_diag_rcv_msg(skb, nlh);
 883 }
 884
 885 static DEFINE_MUTEX(sock_diag_mutex);
 886
 887 static void sock_diag_rcv(struct sk_buff *skb)
 888 {
 889         mutex_lock(&sock_diag_mutex);
 890         netlink_rcv_skb(skb, &sock_diag_rcv_msg);
 891         mutex_unlock(&sock_diag_mutex);
 892 }
 893
 894 int inet_diag_register(const struct inet_diag_handler *h)
 895 {
 896         const __u16 type = h->idiag_type;
 897         int err = -EINVAL;
 898
 899         if (type >= INET_DIAG_GETSOCK_MAX)
 900                 goto out;
 901
 902         mutex_lock(&inet_diag_table_mutex);
 903         err = -EEXIST;
 904         if (inet_diag_table[type] == NULL) {
 905                 inet_diag_table[type] = h;
 906                 err = 0;
 907         }
 908         mutex_unlock(&inet_diag_table_mutex);
 909 out:
 910         return err;
 911 }
 912 EXPORT_SYMBOL_GPL(inet_diag_register);
 913
 914 void inet_diag_unregister(const struct inet_diag_handler *h)
 915 {
 916         const __u16 type = h->idiag_type;
 917
 918         if (type >= INET_DIAG_GETSOCK_MAX)
 919                 return;
 920
 921         mutex_lock(&inet_diag_table_mutex);
 922         inet_diag_table[type] = NULL;
 923         mutex_unlock(&inet_diag_table_mutex);
 924 }
 925 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 926
 927 static int __init inet_diag_init(void)
 928 {
 929         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 930                                           sizeof(struct inet_diag_handler *));
 931         int err = -ENOMEM;
 932
 933         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 934         if (!inet_diag_table)
 935                 goto out;
 936
 937         sdiagnl = netlink_kernel_create(&init_net, NETLINK_SOCK_DIAG, 0,
 938                                         sock_diag_rcv, NULL, THIS_MODULE);
 939         if (sdiagnl == NULL)
 940                 goto out_free_table;
 941         err = 0;
 942 out:
 943         return err;
 944 out_free_table:
 945         kfree(inet_diag_table);
 946         goto out;
 947 }
 948
 949 static void __exit inet_diag_exit(void)
 950 {
 951         netlink_kernel_release(sdiagnl);
 952         kfree(inet_diag_table);
 953 }
 954
 955 module_init(inet_diag_init);
 956 module_exit(inet_diag_exit);
 957 MODULE_LICENSE("GPL");
 958 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_SOCK_DIAG);