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 *idiagnl;
  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_INET_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 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 126         if (r->idiag_family == AF_INET6) {
 127                 struct ipv6_pinfo *np = inet6_sk(sk);
 128
 129                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 130                                &np->rcv_saddr);
 131                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 132                                &np->daddr);
 133         }
 134 #endif
 135
 136 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 137
 138         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 139                 r->idiag_timer = 1;
 140                 r->idiag_retrans = icsk->icsk_retransmits;
 141                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 142         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 143                 r->idiag_timer = 4;
 144                 r->idiag_retrans = icsk->icsk_probes_out;
 145                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 146         } else if (timer_pending(&sk->sk_timer)) {
 147                 r->idiag_timer = 2;
 148                 r->idiag_retrans = icsk->icsk_probes_out;
 149                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 150         } else {
 151                 r->idiag_timer = 0;
 152                 r->idiag_expires = 0;
 153         }
 154 #undef EXPIRES_IN_MS
 155
 156         r->idiag_uid = sock_i_uid(sk);
 157         r->idiag_inode = sock_i_ino(sk);
 158
 159         if (minfo) {
 160                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 161                 minfo->idiag_wmem = sk->sk_wmem_queued;
 162                 minfo->idiag_fmem = sk->sk_forward_alloc;
 163                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 164         }
 165
 166         handler->idiag_get_info(sk, r, info);
 167
 168         if (sk->sk_state < TCP_TIME_WAIT &&
 169             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 170                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 171
 172         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 173         return skb->len;
 174
 175 rtattr_failure:
 176 nlmsg_failure:
 177         nlmsg_trim(skb, b);
 178         return -EMSGSIZE;
 179 }
 180
 181 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 182                                struct sk_buff *skb, int ext, u32 pid,
 183                                u32 seq, u16 nlmsg_flags,
 184                                const struct nlmsghdr *unlh)
 185 {
 186         long tmo;
 187         struct inet_diag_msg *r;
 188         const unsigned char *previous_tail = skb_tail_pointer(skb);
 189         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 190                                          unlh->nlmsg_type, sizeof(*r));
 191
 192         r = NLMSG_DATA(nlh);
 193         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 194
 195         nlh->nlmsg_flags = nlmsg_flags;
 196
 197         tmo = tw->tw_ttd - jiffies;
 198         if (tmo < 0)
 199                 tmo = 0;
 200
 201         r->idiag_family       = tw->tw_family;
 202         r->idiag_retrans      = 0;
 203         r->id.idiag_if        = tw->tw_bound_dev_if;
 204         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 205         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 206         r->id.idiag_sport     = tw->tw_sport;
 207         r->id.idiag_dport     = tw->tw_dport;
 208         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 209         r->id.idiag_dst[0]    = tw->tw_daddr;
 210         r->idiag_state        = tw->tw_substate;
 211         r->idiag_timer        = 3;
 212         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 213         r->idiag_rqueue       = 0;
 214         r->idiag_wqueue       = 0;
 215         r->idiag_uid          = 0;
 216         r->idiag_inode        = 0;
 217 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 218         if (tw->tw_family == AF_INET6) {
 219                 const struct inet6_timewait_sock *tw6 =
 220                                                 inet6_twsk((struct sock *)tw);
 221
 222                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 223                                &tw6->tw_v6_rcv_saddr);
 224                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 225                                &tw6->tw_v6_daddr);
 226         }
 227 #endif
 228         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 229         return skb->len;
 230 nlmsg_failure:
 231         nlmsg_trim(skb, previous_tail);
 232         return -EMSGSIZE;
 233 }
 234
 235 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 236                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 237                         const struct nlmsghdr *unlh)
 238 {
 239         if (sk->sk_state == TCP_TIME_WAIT)
 240                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 241                                            skb, ext, pid, seq, nlmsg_flags,
 242                                            unlh);
 243         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 244 }
 245
 246 static int inet_diag_get_exact(struct sk_buff *in_skb,
 247                                const struct nlmsghdr *nlh)
 248 {
 249         int err;
 250         struct sock *sk;
 251         struct inet_diag_req *req = NLMSG_DATA(nlh);
 252         struct sk_buff *rep;
 253         struct inet_hashinfo *hashinfo;
 254         const struct inet_diag_handler *handler;
 255
 256         handler = inet_diag_lock_handler(nlh->nlmsg_type);
 257         if (IS_ERR(handler)) {
 258                 err = PTR_ERR(handler);
 259                 goto unlock;
 260         }
 261
 262         hashinfo = handler->idiag_hashinfo;
 263         err = -EINVAL;
 264
 265         if (req->idiag_family == AF_INET) {
 266                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 267                                  req->id.idiag_dport, req->id.idiag_src[0],
 268                                  req->id.idiag_sport, req->id.idiag_if);
 269         }
 270 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 271         else if (req->idiag_family == AF_INET6) {
 272                 sk = inet6_lookup(&init_net, hashinfo,
 273                                   (struct in6_addr *)req->id.idiag_dst,
 274                                   req->id.idiag_dport,
 275                                   (struct in6_addr *)req->id.idiag_src,
 276                                   req->id.idiag_sport,
 277                                   req->id.idiag_if);
 278         }
 279 #endif
 280         else {
 281                 goto unlock;
 282         }
 283
 284         err = -ENOENT;
 285         if (sk == NULL)
 286                 goto unlock;
 287
 288         err = -ESTALE;
 289         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 290              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 291             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 292              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 293                 goto out;
 294
 295         err = -ENOMEM;
 296         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 297                                      sizeof(struct inet_diag_meminfo) +
 298                                      handler->idiag_info_size + 64)),
 299                         GFP_KERNEL);
 300         if (!rep)
 301                 goto out;
 302
 303         err = sk_diag_fill(sk, rep, req->idiag_ext,
 304                            NETLINK_CB(in_skb).pid,
 305                            nlh->nlmsg_seq, 0, nlh);
 306         if (err < 0) {
 307                 WARN_ON(err == -EMSGSIZE);
 308                 kfree_skb(rep);
 309                 goto out;
 310         }
 311         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 312                               MSG_DONTWAIT);
 313         if (err > 0)
 314                 err = 0;
 315
 316 out:
 317         if (sk) {
 318                 if (sk->sk_state == TCP_TIME_WAIT)
 319                         inet_twsk_put((struct inet_timewait_sock *)sk);
 320                 else
 321                         sock_put(sk);
 322         }
 323 unlock:
 324         inet_diag_unlock_handler(handler);
 325         return err;
 326 }
 327
 328 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 329 {
 330         int words = bits >> 5;
 331
 332         bits &= 0x1f;
 333
 334         if (words) {
 335                 if (memcmp(a1, a2, words << 2))
 336                         return 0;
 337         }
 338         if (bits) {
 339                 __be32 w1, w2;
 340                 __be32 mask;
 341
 342                 w1 = a1[words];
 343                 w2 = a2[words];
 344
 345                 mask = htonl((0xffffffff) << (32 - bits));
 346
 347                 if ((w1 ^ w2) & mask)
 348                         return 0;
 349         }
 350
 351         return 1;
 352 }
 353
 354
 355 static int inet_diag_bc_run(const void *bc, int len,
 356                             const struct inet_diag_entry *entry)
 357 {
 358         while (len > 0) {
 359                 int yes = 1;
 360                 const struct inet_diag_bc_op *op = bc;
 361
 362                 switch (op->code) {
 363                 case INET_DIAG_BC_NOP:
 364                         break;
 365                 case INET_DIAG_BC_JMP:
 366                         yes = 0;
 367                         break;
 368                 case INET_DIAG_BC_S_GE:
 369                         yes = entry->sport >= op[1].no;
 370                         break;
 371                 case INET_DIAG_BC_S_LE:
 372                         yes = entry->sport <= op[1].no;
 373                         break;
 374                 case INET_DIAG_BC_D_GE:
 375                         yes = entry->dport >= op[1].no;
 376                         break;
 377                 case INET_DIAG_BC_D_LE:
 378                         yes = entry->dport <= op[1].no;
 379                         break;
 380                 case INET_DIAG_BC_AUTO:
 381                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 382                         break;
 383                 case INET_DIAG_BC_S_COND:
 384                 case INET_DIAG_BC_D_COND: {
 385                         struct inet_diag_hostcond *cond;
 386                         __be32 *addr;
 387
 388                         cond = (struct inet_diag_hostcond *)(op + 1);
 389                         if (cond->port != -1 &&
 390                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 391                                              entry->sport : entry->dport)) {
 392                                 yes = 0;
 393                                 break;
 394                         }
 395
 396                         if (cond->prefix_len == 0)
 397                                 break;
 398
 399                         if (op->code == INET_DIAG_BC_S_COND)
 400                                 addr = entry->saddr;
 401                         else
 402                                 addr = entry->daddr;
 403
 404                         if (bitstring_match(addr, cond->addr,
 405                                             cond->prefix_len))
 406                                 break;
 407                         if (entry->family == AF_INET6 &&
 408                             cond->family == AF_INET) {
 409                                 if (addr[0] == 0 && addr[1] == 0 &&
 410                                     addr[2] == htonl(0xffff) &&
 411                                     bitstring_match(addr + 3, cond->addr,
 412                                                     cond->prefix_len))
 413                                         break;
 414                         }
 415                         yes = 0;
 416                         break;
 417                 }
 418                 }
 419
 420                 if (yes) {
 421                         len -= op->yes;
 422                         bc += op->yes;
 423                 } else {
 424                         len -= op->no;
 425                         bc += op->no;
 426                 }
 427         }
 428         return (len == 0);
 429 }
 430
 431 static int valid_cc(const void *bc, int len, int cc)
 432 {
 433         while (len >= 0) {
 434                 const struct inet_diag_bc_op *op = bc;
 435
 436                 if (cc > len)
 437                         return 0;
 438                 if (cc == len)
 439                         return 1;
 440                 if (op->yes < 4)
 441                         return 0;
 442                 len -= op->yes;
 443                 bc  += op->yes;
 444         }
 445         return 0;
 446 }
 447
 448 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 449 {
 450         const unsigned char *bc = bytecode;
 451         int  len = bytecode_len;
 452
 453         while (len > 0) {
 454                 struct inet_diag_bc_op *op = (struct inet_diag_bc_op *)bc;
 455
 456 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 457                 switch (op->code) {
 458                 case INET_DIAG_BC_AUTO:
 459                 case INET_DIAG_BC_S_COND:
 460                 case INET_DIAG_BC_D_COND:
 461                 case INET_DIAG_BC_S_GE:
 462                 case INET_DIAG_BC_S_LE:
 463                 case INET_DIAG_BC_D_GE:
 464                 case INET_DIAG_BC_D_LE:
 465                         if (op->yes < 4 || op->yes > len + 4)
 466                                 return -EINVAL;
 467                 case INET_DIAG_BC_JMP:
 468                         if (op->no < 4 || op->no > len + 4)
 469                                 return -EINVAL;
 470                         if (op->no < len &&
 471                             !valid_cc(bytecode, bytecode_len, len - op->no))
 472                                 return -EINVAL;
 473                         break;
 474                 case INET_DIAG_BC_NOP:
 475                         if (op->yes < 4 || op->yes > len + 4)
 476                                 return -EINVAL;
 477                         break;
 478                 default:
 479                         return -EINVAL;
 480                 }
 481                 bc  += op->yes;
 482                 len -= op->yes;
 483         }
 484         return len == 0 ? 0 : -EINVAL;
 485 }
 486
 487 static int inet_csk_diag_dump(struct sock *sk,
 488                               struct sk_buff *skb,
 489                               struct netlink_callback *cb)
 490 {
 491         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 492
 493         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 494                 struct inet_diag_entry entry;
 495                 struct rtattr *bc = (struct rtattr *)(r + 1);
 496                 struct inet_sock *inet = inet_sk(sk);
 497
 498                 entry.family = sk->sk_family;
 499 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 500                 if (entry.family == AF_INET6) {
 501                         struct ipv6_pinfo *np = inet6_sk(sk);
 502
 503                         entry.saddr = np->rcv_saddr.s6_addr32;
 504                         entry.daddr = np->daddr.s6_addr32;
 505                 } else
 506 #endif
 507                 {
 508                         entry.saddr = &inet->inet_rcv_saddr;
 509                         entry.daddr = &inet->inet_daddr;
 510                 }
 511                 entry.sport = inet->inet_num;
 512                 entry.dport = ntohs(inet->inet_dport);
 513                 entry.userlocks = sk->sk_userlocks;
 514
 515                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 516                         return 0;
 517         }
 518
 519         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 520                                   NETLINK_CB(cb->skb).pid,
 521                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 522 }
 523
 524 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 525                                struct sk_buff *skb,
 526                                struct netlink_callback *cb)
 527 {
 528         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 529
 530         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 531                 struct inet_diag_entry entry;
 532                 struct rtattr *bc = (struct rtattr *)(r + 1);
 533
 534                 entry.family = tw->tw_family;
 535 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 536                 if (tw->tw_family == AF_INET6) {
 537                         struct inet6_timewait_sock *tw6 =
 538                                                 inet6_twsk((struct sock *)tw);
 539                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 540                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 541                 } else
 542 #endif
 543                 {
 544                         entry.saddr = &tw->tw_rcv_saddr;
 545                         entry.daddr = &tw->tw_daddr;
 546                 }
 547                 entry.sport = tw->tw_num;
 548                 entry.dport = ntohs(tw->tw_dport);
 549                 entry.userlocks = 0;
 550
 551                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 552                         return 0;
 553         }
 554
 555         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 556                                    NETLINK_CB(cb->skb).pid,
 557                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 558 }
 559
 560 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 561                               struct request_sock *req, u32 pid, u32 seq,
 562                               const struct nlmsghdr *unlh)
 563 {
 564         const struct inet_request_sock *ireq = inet_rsk(req);
 565         struct inet_sock *inet = inet_sk(sk);
 566         unsigned char *b = skb_tail_pointer(skb);
 567         struct inet_diag_msg *r;
 568         struct nlmsghdr *nlh;
 569         long tmo;
 570
 571         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 572         nlh->nlmsg_flags = NLM_F_MULTI;
 573         r = NLMSG_DATA(nlh);
 574
 575         r->idiag_family = sk->sk_family;
 576         r->idiag_state = TCP_SYN_RECV;
 577         r->idiag_timer = 1;
 578         r->idiag_retrans = req->retrans;
 579
 580         r->id.idiag_if = sk->sk_bound_dev_if;
 581         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 582         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 583
 584         tmo = req->expires - jiffies;
 585         if (tmo < 0)
 586                 tmo = 0;
 587
 588         r->id.idiag_sport = inet->inet_sport;
 589         r->id.idiag_dport = ireq->rmt_port;
 590         r->id.idiag_src[0] = ireq->loc_addr;
 591         r->id.idiag_dst[0] = ireq->rmt_addr;
 592         r->idiag_expires = jiffies_to_msecs(tmo);
 593         r->idiag_rqueue = 0;
 594         r->idiag_wqueue = 0;
 595         r->idiag_uid = sock_i_uid(sk);
 596         r->idiag_inode = 0;
 597 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 598         if (r->idiag_family == AF_INET6) {
 599                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 600                                &inet6_rsk(req)->loc_addr);
 601                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 602                                &inet6_rsk(req)->rmt_addr);
 603         }
 604 #endif
 605         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 606
 607         return skb->len;
 608
 609 nlmsg_failure:
 610         nlmsg_trim(skb, b);
 611         return -1;
 612 }
 613
 614 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 615                                struct netlink_callback *cb)
 616 {
 617         struct inet_diag_entry entry;
 618         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 619         struct inet_connection_sock *icsk = inet_csk(sk);
 620         struct listen_sock *lopt;
 621         struct rtattr *bc = NULL;
 622         struct inet_sock *inet = inet_sk(sk);
 623         int j, s_j;
 624         int reqnum, s_reqnum;
 625         int err = 0;
 626
 627         s_j = cb->args[3];
 628         s_reqnum = cb->args[4];
 629
 630         if (s_j > 0)
 631                 s_j--;
 632
 633         entry.family = sk->sk_family;
 634
 635         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 636
 637         lopt = icsk->icsk_accept_queue.listen_opt;
 638         if (!lopt || !lopt->qlen)
 639                 goto out;
 640
 641         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 642                 bc = (struct rtattr *)(r + 1);
 643                 entry.sport = inet->inet_num;
 644                 entry.userlocks = sk->sk_userlocks;
 645         }
 646
 647         for (j = s_j; j < lopt->nr_table_entries; j++) {
 648                 struct request_sock *req, *head = lopt->syn_table[j];
 649
 650                 reqnum = 0;
 651                 for (req = head; req; reqnum++, req = req->dl_next) {
 652                         struct inet_request_sock *ireq = inet_rsk(req);
 653
 654                         if (reqnum < s_reqnum)
 655                                 continue;
 656                         if (r->id.idiag_dport != ireq->rmt_port &&
 657                             r->id.idiag_dport)
 658                                 continue;
 659
 660                         if (bc) {
 661                                 entry.saddr =
 662 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 663                                         (entry.family == AF_INET6) ?
 664                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 665 #endif
 666                                         &ireq->loc_addr;
 667                                 entry.daddr =
 668 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 669                                         (entry.family == AF_INET6) ?
 670                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 671 #endif
 672                                         &ireq->rmt_addr;
 673                                 entry.dport = ntohs(ireq->rmt_port);
 674
 675                                 if (!inet_diag_bc_run(RTA_DATA(bc),
 676                                                     RTA_PAYLOAD(bc), &entry))
 677                                         continue;
 678                         }
 679
 680                         err = inet_diag_fill_req(skb, sk, req,
 681                                                NETLINK_CB(cb->skb).pid,
 682                                                cb->nlh->nlmsg_seq, cb->nlh);
 683                         if (err < 0) {
 684                                 cb->args[3] = j + 1;
 685                                 cb->args[4] = reqnum;
 686                                 goto out;
 687                         }
 688                 }
 689
 690                 s_reqnum = 0;
 691         }
 692
 693 out:
 694         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 695
 696         return err;
 697 }
 698
 699 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 700 {
 701         int i, num;
 702         int s_i, s_num;
 703         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 704         const struct inet_diag_handler *handler;
 705         struct inet_hashinfo *hashinfo;
 706
 707         handler = inet_diag_lock_handler(cb->nlh->nlmsg_type);
 708         if (IS_ERR(handler))
 709                 goto unlock;
 710
 711         hashinfo = handler->idiag_hashinfo;
 712
 713         s_i = cb->args[1];
 714         s_num = num = cb->args[2];
 715
 716         if (cb->args[0] == 0) {
 717                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 718                         goto skip_listen_ht;
 719
 720                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 721                         struct sock *sk;
 722                         struct hlist_nulls_node *node;
 723                         struct inet_listen_hashbucket *ilb;
 724
 725                         num = 0;
 726                         ilb = &hashinfo->listening_hash[i];
 727                         spin_lock_bh(&ilb->lock);
 728                         sk_nulls_for_each(sk, node, &ilb->head) {
 729                                 struct inet_sock *inet = inet_sk(sk);
 730
 731                                 if (num < s_num) {
 732                                         num++;
 733                                         continue;
 734                                 }
 735
 736                                 if (r->id.idiag_sport != inet->inet_sport &&
 737                                     r->id.idiag_sport)
 738                                         goto next_listen;
 739
 740                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 741                                     r->id.idiag_dport ||
 742                                     cb->args[3] > 0)
 743                                         goto syn_recv;
 744
 745                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 746                                         spin_unlock_bh(&ilb->lock);
 747                                         goto done;
 748                                 }
 749
 750 syn_recv:
 751                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 752                                         goto next_listen;
 753
 754                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 755                                         spin_unlock_bh(&ilb->lock);
 756                                         goto done;
 757                                 }
 758
 759 next_listen:
 760                                 cb->args[3] = 0;
 761                                 cb->args[4] = 0;
 762                                 ++num;
 763                         }
 764                         spin_unlock_bh(&ilb->lock);
 765
 766                         s_num = 0;
 767                         cb->args[3] = 0;
 768                         cb->args[4] = 0;
 769                 }
 770 skip_listen_ht:
 771                 cb->args[0] = 1;
 772                 s_i = num = s_num = 0;
 773         }
 774
 775         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 776                 goto unlock;
 777
 778         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 779                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 780                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 781                 struct sock *sk;
 782                 struct hlist_nulls_node *node;
 783
 784                 num = 0;
 785
 786                 if (hlist_nulls_empty(&head->chain) &&
 787                         hlist_nulls_empty(&head->twchain))
 788                         continue;
 789
 790                 if (i > s_i)
 791                         s_num = 0;
 792
 793                 spin_lock_bh(lock);
 794                 sk_nulls_for_each(sk, node, &head->chain) {
 795                         struct inet_sock *inet = inet_sk(sk);
 796
 797                         if (num < s_num)
 798                                 goto next_normal;
 799                         if (!(r->idiag_states & (1 << sk->sk_state)))
 800                                 goto next_normal;
 801                         if (r->id.idiag_sport != inet->inet_sport &&
 802                             r->id.idiag_sport)
 803                                 goto next_normal;
 804                         if (r->id.idiag_dport != inet->inet_dport &&
 805                             r->id.idiag_dport)
 806                                 goto next_normal;
 807                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 808                                 spin_unlock_bh(lock);
 809                                 goto done;
 810                         }
 811 next_normal:
 812                         ++num;
 813                 }
 814
 815                 if (r->idiag_states & TCPF_TIME_WAIT) {
 816                         struct inet_timewait_sock *tw;
 817
 818                         inet_twsk_for_each(tw, node,
 819                                     &head->twchain) {
 820
 821                                 if (num < s_num)
 822                                         goto next_dying;
 823                                 if (r->id.idiag_sport != tw->tw_sport &&
 824                                     r->id.idiag_sport)
 825                                         goto next_dying;
 826                                 if (r->id.idiag_dport != tw->tw_dport &&
 827                                     r->id.idiag_dport)
 828                                         goto next_dying;
 829                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 830                                         spin_unlock_bh(lock);
 831                                         goto done;
 832                                 }
 833 next_dying:
 834                                 ++num;
 835                         }
 836                 }
 837                 spin_unlock_bh(lock);
 838         }
 839
 840 done:
 841         cb->args[1] = i;
 842         cb->args[2] = num;
 843 unlock:
 844         inet_diag_unlock_handler(handler);
 845         return skb->len;
 846 }
 847
 848 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 849 {
 850         int hdrlen = sizeof(struct inet_diag_req);
 851
 852         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 853             nlmsg_len(nlh) < hdrlen)
 854                 return -EINVAL;
 855
 856         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 857                 if (nlmsg_attrlen(nlh, hdrlen)) {
 858                         struct nlattr *attr;
 859
 860                         attr = nlmsg_find_attr(nlh, hdrlen,
 861                                                INET_DIAG_REQ_BYTECODE);
 862                         if (attr == NULL ||
 863                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 864                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 865                                 return -EINVAL;
 866                 }
 867
 868                 return netlink_dump_start(idiagnl, skb, nlh,
 869                                           inet_diag_dump, NULL);
 870         }
 871
 872         return inet_diag_get_exact(skb, nlh);
 873 }
 874
 875 static DEFINE_MUTEX(inet_diag_mutex);
 876
 877 static void inet_diag_rcv(struct sk_buff *skb)
 878 {
 879         mutex_lock(&inet_diag_mutex);
 880         netlink_rcv_skb(skb, &inet_diag_rcv_msg);
 881         mutex_unlock(&inet_diag_mutex);
 882 }
 883
 884 int inet_diag_register(const struct inet_diag_handler *h)
 885 {
 886         const __u16 type = h->idiag_type;
 887         int err = -EINVAL;
 888
 889         if (type >= INET_DIAG_GETSOCK_MAX)
 890                 goto out;
 891
 892         mutex_lock(&inet_diag_table_mutex);
 893         err = -EEXIST;
 894         if (inet_diag_table[type] == NULL) {
 895                 inet_diag_table[type] = h;
 896                 err = 0;
 897         }
 898         mutex_unlock(&inet_diag_table_mutex);
 899 out:
 900         return err;
 901 }
 902 EXPORT_SYMBOL_GPL(inet_diag_register);
 903
 904 void inet_diag_unregister(const struct inet_diag_handler *h)
 905 {
 906         const __u16 type = h->idiag_type;
 907
 908         if (type >= INET_DIAG_GETSOCK_MAX)
 909                 return;
 910
 911         mutex_lock(&inet_diag_table_mutex);
 912         inet_diag_table[type] = NULL;
 913         mutex_unlock(&inet_diag_table_mutex);
 914 }
 915 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 916
 917 static int __init inet_diag_init(void)
 918 {
 919         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 920                                           sizeof(struct inet_diag_handler *));
 921         int err = -ENOMEM;
 922
 923         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 924         if (!inet_diag_table)
 925                 goto out;
 926
 927         idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
 928                                         inet_diag_rcv, NULL, THIS_MODULE);
 929         if (idiagnl == NULL)
 930                 goto out_free_table;
 931         err = 0;
 932 out:
 933         return err;
 934 out_free_table:
 935         kfree(inet_diag_table);
 936         goto out;
 937 }
 938
 939 static void __exit inet_diag_exit(void)
 940 {
 941         netlink_kernel_release(idiagnl);
 942         kfree(inet_diag_table);
 943 }
 944
 945 module_init(inet_diag_init);
 946 module_exit(inet_diag_exit);
 947 MODULE_LICENSE("GPL");
 948 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_INET_DIAG);