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         if ((ext & (1 << (INET_DIAG_TOS - 1))) && (sk->sk_family != AF_INET6))
 112                 RTA_PUT_U8(skb, INET_DIAG_TOS, inet->tos);
 113
 114         r->idiag_family = sk->sk_family;
 115         r->idiag_state = sk->sk_state;
 116         r->idiag_timer = 0;
 117         r->idiag_retrans = 0;
 118
 119         r->id.idiag_if = sk->sk_bound_dev_if;
 120         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
 121         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 122
 123         r->id.idiag_sport = inet->inet_sport;
 124         r->id.idiag_dport = inet->inet_dport;
 125         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 126         r->id.idiag_dst[0] = inet->inet_daddr;
 127
 128 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 129         if (r->idiag_family == AF_INET6) {
 130                 const struct ipv6_pinfo *np = inet6_sk(sk);
 131
 132                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 133                                &np->rcv_saddr);
 134                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 135                                &np->daddr);
 136                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 137                         RTA_PUT_U8(skb, INET_DIAG_TCLASS, np->tclass);
 138         }
 139 #endif
 140
 141 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 142
 143         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 144                 r->idiag_timer = 1;
 145                 r->idiag_retrans = icsk->icsk_retransmits;
 146                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 147         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 148                 r->idiag_timer = 4;
 149                 r->idiag_retrans = icsk->icsk_probes_out;
 150                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 151         } else if (timer_pending(&sk->sk_timer)) {
 152                 r->idiag_timer = 2;
 153                 r->idiag_retrans = icsk->icsk_probes_out;
 154                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 155         } else {
 156                 r->idiag_timer = 0;
 157                 r->idiag_expires = 0;
 158         }
 159 #undef EXPIRES_IN_MS
 160
 161         r->idiag_uid = sock_i_uid(sk);
 162         r->idiag_inode = sock_i_ino(sk);
 163
 164         if (minfo) {
 165                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 166                 minfo->idiag_wmem = sk->sk_wmem_queued;
 167                 minfo->idiag_fmem = sk->sk_forward_alloc;
 168                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 169         }
 170
 171         handler->idiag_get_info(sk, r, info);
 172
 173         if (sk->sk_state < TCP_TIME_WAIT &&
 174             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 175                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 176
 177         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 178         return skb->len;
 179
 180 rtattr_failure:
 181 nlmsg_failure:
 182         nlmsg_trim(skb, b);
 183         return -EMSGSIZE;
 184 }
 185
 186 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 187                                struct sk_buff *skb, int ext, u32 pid,
 188                                u32 seq, u16 nlmsg_flags,
 189                                const struct nlmsghdr *unlh)
 190 {
 191         long tmo;
 192         struct inet_diag_msg *r;
 193         const unsigned char *previous_tail = skb_tail_pointer(skb);
 194         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 195                                          unlh->nlmsg_type, sizeof(*r));
 196
 197         r = NLMSG_DATA(nlh);
 198         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 199
 200         nlh->nlmsg_flags = nlmsg_flags;
 201
 202         tmo = tw->tw_ttd - jiffies;
 203         if (tmo < 0)
 204                 tmo = 0;
 205
 206         r->idiag_family       = tw->tw_family;
 207         r->idiag_retrans      = 0;
 208         r->id.idiag_if        = tw->tw_bound_dev_if;
 209         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 210         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 211         r->id.idiag_sport     = tw->tw_sport;
 212         r->id.idiag_dport     = tw->tw_dport;
 213         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 214         r->id.idiag_dst[0]    = tw->tw_daddr;
 215         r->idiag_state        = tw->tw_substate;
 216         r->idiag_timer        = 3;
 217         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 218         r->idiag_rqueue       = 0;
 219         r->idiag_wqueue       = 0;
 220         r->idiag_uid          = 0;
 221         r->idiag_inode        = 0;
 222 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 223         if (tw->tw_family == AF_INET6) {
 224                 const struct inet6_timewait_sock *tw6 =
 225                                                 inet6_twsk((struct sock *)tw);
 226
 227                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 228                                &tw6->tw_v6_rcv_saddr);
 229                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 230                                &tw6->tw_v6_daddr);
 231         }
 232 #endif
 233         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 234         return skb->len;
 235 nlmsg_failure:
 236         nlmsg_trim(skb, previous_tail);
 237         return -EMSGSIZE;
 238 }
 239
 240 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 241                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 242                         const struct nlmsghdr *unlh)
 243 {
 244         if (sk->sk_state == TCP_TIME_WAIT)
 245                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 246                                            skb, ext, pid, seq, nlmsg_flags,
 247                                            unlh);
 248         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 249 }
 250
 251 static int inet_diag_get_exact(struct sk_buff *in_skb,
 252                                const struct nlmsghdr *nlh)
 253 {
 254         int err;
 255         struct sock *sk;
 256         struct inet_diag_req *req = NLMSG_DATA(nlh);
 257         struct sk_buff *rep;
 258         struct inet_hashinfo *hashinfo;
 259         const struct inet_diag_handler *handler;
 260
 261         handler = inet_diag_lock_handler(nlh->nlmsg_type);
 262         if (IS_ERR(handler)) {
 263                 err = PTR_ERR(handler);
 264                 goto unlock;
 265         }
 266
 267         hashinfo = handler->idiag_hashinfo;
 268         err = -EINVAL;
 269
 270         if (req->idiag_family == AF_INET) {
 271                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 272                                  req->id.idiag_dport, req->id.idiag_src[0],
 273                                  req->id.idiag_sport, req->id.idiag_if);
 274         }
 275 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 276         else if (req->idiag_family == AF_INET6) {
 277                 sk = inet6_lookup(&init_net, hashinfo,
 278                                   (struct in6_addr *)req->id.idiag_dst,
 279                                   req->id.idiag_dport,
 280                                   (struct in6_addr *)req->id.idiag_src,
 281                                   req->id.idiag_sport,
 282                                   req->id.idiag_if);
 283         }
 284 #endif
 285         else {
 286                 goto unlock;
 287         }
 288
 289         err = -ENOENT;
 290         if (sk == NULL)
 291                 goto unlock;
 292
 293         err = -ESTALE;
 294         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 295              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 296             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 297              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 298                 goto out;
 299
 300         err = -ENOMEM;
 301         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 302                                      sizeof(struct inet_diag_meminfo) +
 303                                      handler->idiag_info_size + 64)),
 304                         GFP_KERNEL);
 305         if (!rep)
 306                 goto out;
 307
 308         err = sk_diag_fill(sk, rep, req->idiag_ext,
 309                            NETLINK_CB(in_skb).pid,
 310                            nlh->nlmsg_seq, 0, nlh);
 311         if (err < 0) {
 312                 WARN_ON(err == -EMSGSIZE);
 313                 kfree_skb(rep);
 314                 goto out;
 315         }
 316         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 317                               MSG_DONTWAIT);
 318         if (err > 0)
 319                 err = 0;
 320
 321 out:
 322         if (sk) {
 323                 if (sk->sk_state == TCP_TIME_WAIT)
 324                         inet_twsk_put((struct inet_timewait_sock *)sk);
 325                 else
 326                         sock_put(sk);
 327         }
 328 unlock:
 329         inet_diag_unlock_handler(handler);
 330         return err;
 331 }
 332
 333 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 334 {
 335         int words = bits >> 5;
 336
 337         bits &= 0x1f;
 338
 339         if (words) {
 340                 if (memcmp(a1, a2, words << 2))
 341                         return 0;
 342         }
 343         if (bits) {
 344                 __be32 w1, w2;
 345                 __be32 mask;
 346
 347                 w1 = a1[words];
 348                 w2 = a2[words];
 349
 350                 mask = htonl((0xffffffff) << (32 - bits));
 351
 352                 if ((w1 ^ w2) & mask)
 353                         return 0;
 354         }
 355
 356         return 1;
 357 }
 358
 359
 360 static int inet_diag_bc_run(const void *bc, int len,
 361                             const struct inet_diag_entry *entry)
 362 {
 363         while (len > 0) {
 364                 int yes = 1;
 365                 const struct inet_diag_bc_op *op = bc;
 366
 367                 switch (op->code) {
 368                 case INET_DIAG_BC_NOP:
 369                         break;
 370                 case INET_DIAG_BC_JMP:
 371                         yes = 0;
 372                         break;
 373                 case INET_DIAG_BC_S_GE:
 374                         yes = entry->sport >= op[1].no;
 375                         break;
 376                 case INET_DIAG_BC_S_LE:
 377                         yes = entry->sport <= op[1].no;
 378                         break;
 379                 case INET_DIAG_BC_D_GE:
 380                         yes = entry->dport >= op[1].no;
 381                         break;
 382                 case INET_DIAG_BC_D_LE:
 383                         yes = entry->dport <= op[1].no;
 384                         break;
 385                 case INET_DIAG_BC_AUTO:
 386                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 387                         break;
 388                 case INET_DIAG_BC_S_COND:
 389                 case INET_DIAG_BC_D_COND: {
 390                         struct inet_diag_hostcond *cond;
 391                         __be32 *addr;
 392
 393                         cond = (struct inet_diag_hostcond *)(op + 1);
 394                         if (cond->port != -1 &&
 395                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 396                                              entry->sport : entry->dport)) {
 397                                 yes = 0;
 398                                 break;
 399                         }
 400
 401                         if (cond->prefix_len == 0)
 402                                 break;
 403
 404                         if (op->code == INET_DIAG_BC_S_COND)
 405                                 addr = entry->saddr;
 406                         else
 407                                 addr = entry->daddr;
 408
 409                         if (bitstring_match(addr, cond->addr,
 410                                             cond->prefix_len))
 411                                 break;
 412                         if (entry->family == AF_INET6 &&
 413                             cond->family == AF_INET) {
 414                                 if (addr[0] == 0 && addr[1] == 0 &&
 415                                     addr[2] == htonl(0xffff) &&
 416                                     bitstring_match(addr + 3, cond->addr,
 417                                                     cond->prefix_len))
 418                                         break;
 419                         }
 420                         yes = 0;
 421                         break;
 422                 }
 423                 }
 424
 425                 if (yes) {
 426                         len -= op->yes;
 427                         bc += op->yes;
 428                 } else {
 429                         len -= op->no;
 430                         bc += op->no;
 431                 }
 432         }
 433         return len == 0;
 434 }
 435
 436 static int valid_cc(const void *bc, int len, int cc)
 437 {
 438         while (len >= 0) {
 439                 const struct inet_diag_bc_op *op = bc;
 440
 441                 if (cc > len)
 442                         return 0;
 443                 if (cc == len)
 444                         return 1;
 445                 if (op->yes < 4 || op->yes & 3)
 446                         return 0;
 447                 len -= op->yes;
 448                 bc  += op->yes;
 449         }
 450         return 0;
 451 }
 452
 453 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 454 {
 455         const void *bc = bytecode;
 456         int  len = bytecode_len;
 457
 458         while (len > 0) {
 459                 const struct inet_diag_bc_op *op = bc;
 460
 461 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 462                 switch (op->code) {
 463                 case INET_DIAG_BC_AUTO:
 464                 case INET_DIAG_BC_S_COND:
 465                 case INET_DIAG_BC_D_COND:
 466                 case INET_DIAG_BC_S_GE:
 467                 case INET_DIAG_BC_S_LE:
 468                 case INET_DIAG_BC_D_GE:
 469                 case INET_DIAG_BC_D_LE:
 470                 case INET_DIAG_BC_JMP:
 471                         if (op->no < 4 || op->no > len + 4 || op->no & 3)
 472                                 return -EINVAL;
 473                         if (op->no < len &&
 474                             !valid_cc(bytecode, bytecode_len, len - op->no))
 475                                 return -EINVAL;
 476                         break;
 477                 case INET_DIAG_BC_NOP:
 478                         break;
 479                 default:
 480                         return -EINVAL;
 481                 }
 482                 if (op->yes < 4 || op->yes > len + 4 || op->yes & 3)
 483                         return -EINVAL;
 484                 bc  += op->yes;
 485                 len -= op->yes;
 486         }
 487         return len == 0 ? 0 : -EINVAL;
 488 }
 489
 490 static int inet_csk_diag_dump(struct sock *sk,
 491                               struct sk_buff *skb,
 492                               struct netlink_callback *cb)
 493 {
 494         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 495
 496         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 497                 struct inet_diag_entry entry;
 498                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 499                                                           sizeof(*r),
 500                                                           INET_DIAG_REQ_BYTECODE);
 501                 struct inet_sock *inet = inet_sk(sk);
 502
 503                 entry.family = sk->sk_family;
 504 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 505                 if (entry.family == AF_INET6) {
 506                         struct ipv6_pinfo *np = inet6_sk(sk);
 507
 508                         entry.saddr = np->rcv_saddr.s6_addr32;
 509                         entry.daddr = np->daddr.s6_addr32;
 510                 } else
 511 #endif
 512                 {
 513                         entry.saddr = &inet->inet_rcv_saddr;
 514                         entry.daddr = &inet->inet_daddr;
 515                 }
 516                 entry.sport = inet->inet_num;
 517                 entry.dport = ntohs(inet->inet_dport);
 518                 entry.userlocks = sk->sk_userlocks;
 519
 520                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 521                         return 0;
 522         }
 523
 524         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 525                                   NETLINK_CB(cb->skb).pid,
 526                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 527 }
 528
 529 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 530                                struct sk_buff *skb,
 531                                struct netlink_callback *cb)
 532 {
 533         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 534
 535         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 536                 struct inet_diag_entry entry;
 537                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 538                                                           sizeof(*r),
 539                                                           INET_DIAG_REQ_BYTECODE);
 540
 541                 entry.family = tw->tw_family;
 542 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 543                 if (tw->tw_family == AF_INET6) {
 544                         struct inet6_timewait_sock *tw6 =
 545                                                 inet6_twsk((struct sock *)tw);
 546                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 547                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 548                 } else
 549 #endif
 550                 {
 551                         entry.saddr = &tw->tw_rcv_saddr;
 552                         entry.daddr = &tw->tw_daddr;
 553                 }
 554                 entry.sport = tw->tw_num;
 555                 entry.dport = ntohs(tw->tw_dport);
 556                 entry.userlocks = 0;
 557
 558                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 559                         return 0;
 560         }
 561
 562         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 563                                    NETLINK_CB(cb->skb).pid,
 564                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 565 }
 566
 567 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 568                               struct request_sock *req, u32 pid, u32 seq,
 569                               const struct nlmsghdr *unlh)
 570 {
 571         const struct inet_request_sock *ireq = inet_rsk(req);
 572         struct inet_sock *inet = inet_sk(sk);
 573         unsigned char *b = skb_tail_pointer(skb);
 574         struct inet_diag_msg *r;
 575         struct nlmsghdr *nlh;
 576         long tmo;
 577
 578         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 579         nlh->nlmsg_flags = NLM_F_MULTI;
 580         r = NLMSG_DATA(nlh);
 581
 582         r->idiag_family = sk->sk_family;
 583         r->idiag_state = TCP_SYN_RECV;
 584         r->idiag_timer = 1;
 585         r->idiag_retrans = req->retrans;
 586
 587         r->id.idiag_if = sk->sk_bound_dev_if;
 588         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 589         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 590
 591         tmo = req->expires - jiffies;
 592         if (tmo < 0)
 593                 tmo = 0;
 594
 595         r->id.idiag_sport = inet->inet_sport;
 596         r->id.idiag_dport = ireq->rmt_port;
 597         r->id.idiag_src[0] = ireq->loc_addr;
 598         r->id.idiag_dst[0] = ireq->rmt_addr;
 599         r->idiag_expires = jiffies_to_msecs(tmo);
 600         r->idiag_rqueue = 0;
 601         r->idiag_wqueue = 0;
 602         r->idiag_uid = sock_i_uid(sk);
 603         r->idiag_inode = 0;
 604 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 605         if (r->idiag_family == AF_INET6) {
 606                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 607                                &inet6_rsk(req)->loc_addr);
 608                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 609                                &inet6_rsk(req)->rmt_addr);
 610         }
 611 #endif
 612         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 613
 614         return skb->len;
 615
 616 nlmsg_failure:
 617         nlmsg_trim(skb, b);
 618         return -1;
 619 }
 620
 621 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 622                                struct netlink_callback *cb)
 623 {
 624         struct inet_diag_entry entry;
 625         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 626         struct inet_connection_sock *icsk = inet_csk(sk);
 627         struct listen_sock *lopt;
 628         const struct nlattr *bc = NULL;
 629         struct inet_sock *inet = inet_sk(sk);
 630         int j, s_j;
 631         int reqnum, s_reqnum;
 632         int err = 0;
 633
 634         s_j = cb->args[3];
 635         s_reqnum = cb->args[4];
 636
 637         if (s_j > 0)
 638                 s_j--;
 639
 640         entry.family = sk->sk_family;
 641
 642         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 643
 644         lopt = icsk->icsk_accept_queue.listen_opt;
 645         if (!lopt || !lopt->qlen)
 646                 goto out;
 647
 648         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 649                 bc = nlmsg_find_attr(cb->nlh, sizeof(*r),
 650                                      INET_DIAG_REQ_BYTECODE);
 651                 entry.sport = inet->inet_num;
 652                 entry.userlocks = sk->sk_userlocks;
 653         }
 654
 655         for (j = s_j; j < lopt->nr_table_entries; j++) {
 656                 struct request_sock *req, *head = lopt->syn_table[j];
 657
 658                 reqnum = 0;
 659                 for (req = head; req; reqnum++, req = req->dl_next) {
 660                         struct inet_request_sock *ireq = inet_rsk(req);
 661
 662                         if (reqnum < s_reqnum)
 663                                 continue;
 664                         if (r->id.idiag_dport != ireq->rmt_port &&
 665                             r->id.idiag_dport)
 666                                 continue;
 667
 668                         if (bc) {
 669                                 entry.saddr =
 670 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 671                                         (entry.family == AF_INET6) ?
 672                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 673 #endif
 674                                         &ireq->loc_addr;
 675                                 entry.daddr =
 676 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 677                                         (entry.family == AF_INET6) ?
 678                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 679 #endif
 680                                         &ireq->rmt_addr;
 681                                 entry.dport = ntohs(ireq->rmt_port);
 682
 683                                 if (!inet_diag_bc_run(nla_data(bc),
 684                                                       nla_len(bc), &entry))
 685                                         continue;
 686                         }
 687
 688                         err = inet_diag_fill_req(skb, sk, req,
 689                                                NETLINK_CB(cb->skb).pid,
 690                                                cb->nlh->nlmsg_seq, cb->nlh);
 691                         if (err < 0) {
 692                                 cb->args[3] = j + 1;
 693                                 cb->args[4] = reqnum;
 694                                 goto out;
 695                         }
 696                 }
 697
 698                 s_reqnum = 0;
 699         }
 700
 701 out:
 702         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 703
 704         return err;
 705 }
 706
 707 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 708 {
 709         int i, num;
 710         int s_i, s_num;
 711         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 712         const struct inet_diag_handler *handler;
 713         struct inet_hashinfo *hashinfo;
 714
 715         handler = inet_diag_lock_handler(cb->nlh->nlmsg_type);
 716         if (IS_ERR(handler))
 717                 goto unlock;
 718
 719         hashinfo = handler->idiag_hashinfo;
 720
 721         s_i = cb->args[1];
 722         s_num = num = cb->args[2];
 723
 724         if (cb->args[0] == 0) {
 725                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 726                         goto skip_listen_ht;
 727
 728                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 729                         struct sock *sk;
 730                         struct hlist_nulls_node *node;
 731                         struct inet_listen_hashbucket *ilb;
 732
 733                         num = 0;
 734                         ilb = &hashinfo->listening_hash[i];
 735                         spin_lock_bh(&ilb->lock);
 736                         sk_nulls_for_each(sk, node, &ilb->head) {
 737                                 struct inet_sock *inet = inet_sk(sk);
 738
 739                                 if (num < s_num) {
 740                                         num++;
 741                                         continue;
 742                                 }
 743
 744                                 if (r->id.idiag_sport != inet->inet_sport &&
 745                                     r->id.idiag_sport)
 746                                         goto next_listen;
 747
 748                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 749                                     r->id.idiag_dport ||
 750                                     cb->args[3] > 0)
 751                                         goto syn_recv;
 752
 753                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 754                                         spin_unlock_bh(&ilb->lock);
 755                                         goto done;
 756                                 }
 757
 758 syn_recv:
 759                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 760                                         goto next_listen;
 761
 762                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 763                                         spin_unlock_bh(&ilb->lock);
 764                                         goto done;
 765                                 }
 766
 767 next_listen:
 768                                 cb->args[3] = 0;
 769                                 cb->args[4] = 0;
 770                                 ++num;
 771                         }
 772                         spin_unlock_bh(&ilb->lock);
 773
 774                         s_num = 0;
 775                         cb->args[3] = 0;
 776                         cb->args[4] = 0;
 777                 }
 778 skip_listen_ht:
 779                 cb->args[0] = 1;
 780                 s_i = num = s_num = 0;
 781         }
 782
 783         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 784                 goto unlock;
 785
 786         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 787                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 788                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 789                 struct sock *sk;
 790                 struct hlist_nulls_node *node;
 791
 792                 num = 0;
 793
 794                 if (hlist_nulls_empty(&head->chain) &&
 795                         hlist_nulls_empty(&head->twchain))
 796                         continue;
 797
 798                 if (i > s_i)
 799                         s_num = 0;
 800
 801                 spin_lock_bh(lock);
 802                 sk_nulls_for_each(sk, node, &head->chain) {
 803                         struct inet_sock *inet = inet_sk(sk);
 804
 805                         if (num < s_num)
 806                                 goto next_normal;
 807                         if (!(r->idiag_states & (1 << sk->sk_state)))
 808                                 goto next_normal;
 809                         if (r->id.idiag_sport != inet->inet_sport &&
 810                             r->id.idiag_sport)
 811                                 goto next_normal;
 812                         if (r->id.idiag_dport != inet->inet_dport &&
 813                             r->id.idiag_dport)
 814                                 goto next_normal;
 815                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 816                                 spin_unlock_bh(lock);
 817                                 goto done;
 818                         }
 819 next_normal:
 820                         ++num;
 821                 }
 822
 823                 if (r->idiag_states & TCPF_TIME_WAIT) {
 824                         struct inet_timewait_sock *tw;
 825
 826                         inet_twsk_for_each(tw, node,
 827                                     &head->twchain) {
 828
 829                                 if (num < s_num)
 830                                         goto next_dying;
 831                                 if (r->id.idiag_sport != tw->tw_sport &&
 832                                     r->id.idiag_sport)
 833                                         goto next_dying;
 834                                 if (r->id.idiag_dport != tw->tw_dport &&
 835                                     r->id.idiag_dport)
 836                                         goto next_dying;
 837                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 838                                         spin_unlock_bh(lock);
 839                                         goto done;
 840                                 }
 841 next_dying:
 842                                 ++num;
 843                         }
 844                 }
 845                 spin_unlock_bh(lock);
 846         }
 847
 848 done:
 849         cb->args[1] = i;
 850         cb->args[2] = num;
 851 unlock:
 852         inet_diag_unlock_handler(handler);
 853         return skb->len;
 854 }
 855
 856 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 857 {
 858         int hdrlen = sizeof(struct inet_diag_req);
 859
 860         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 861             nlmsg_len(nlh) < hdrlen)
 862                 return -EINVAL;
 863
 864         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 865                 if (nlmsg_attrlen(nlh, hdrlen)) {
 866                         struct nlattr *attr;
 867
 868                         attr = nlmsg_find_attr(nlh, hdrlen,
 869                                                INET_DIAG_REQ_BYTECODE);
 870                         if (attr == NULL ||
 871                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 872                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 873                                 return -EINVAL;
 874                 }
 875
 876                 return netlink_dump_start(idiagnl, skb, nlh,
 877                                           inet_diag_dump, NULL, 0);
 878         }
 879
 880         return inet_diag_get_exact(skb, nlh);
 881 }
 882
 883 static DEFINE_MUTEX(inet_diag_mutex);
 884
 885 static void inet_diag_rcv(struct sk_buff *skb)
 886 {
 887         mutex_lock(&inet_diag_mutex);
 888         netlink_rcv_skb(skb, &inet_diag_rcv_msg);
 889         mutex_unlock(&inet_diag_mutex);
 890 }
 891
 892 int inet_diag_register(const struct inet_diag_handler *h)
 893 {
 894         const __u16 type = h->idiag_type;
 895         int err = -EINVAL;
 896
 897         if (type >= INET_DIAG_GETSOCK_MAX)
 898                 goto out;
 899
 900         mutex_lock(&inet_diag_table_mutex);
 901         err = -EEXIST;
 902         if (inet_diag_table[type] == NULL) {
 903                 inet_diag_table[type] = h;
 904                 err = 0;
 905         }
 906         mutex_unlock(&inet_diag_table_mutex);
 907 out:
 908         return err;
 909 }
 910 EXPORT_SYMBOL_GPL(inet_diag_register);
 911
 912 void inet_diag_unregister(const struct inet_diag_handler *h)
 913 {
 914         const __u16 type = h->idiag_type;
 915
 916         if (type >= INET_DIAG_GETSOCK_MAX)
 917                 return;
 918
 919         mutex_lock(&inet_diag_table_mutex);
 920         inet_diag_table[type] = NULL;
 921         mutex_unlock(&inet_diag_table_mutex);
 922 }
 923 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 924
 925 static int __init inet_diag_init(void)
 926 {
 927         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 928                                           sizeof(struct inet_diag_handler *));
 929         int err = -ENOMEM;
 930
 931         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 932         if (!inet_diag_table)
 933                 goto out;
 934
 935         idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
 936                                         inet_diag_rcv, NULL, THIS_MODULE);
 937         if (idiagnl == NULL)
 938                 goto out_free_table;
 939         err = 0;
 940 out:
 941         return err;
 942 out_free_table:
 943         kfree(inet_diag_table);
 944         goto out;
 945 }
 946
 947 static void __exit inet_diag_exit(void)
 948 {
 949         netlink_kernel_release(idiagnl);
 950         kfree(inet_diag_table);
 951 }
 952
 953 module_init(inet_diag_init);
 954 module_exit(inet_diag_exit);
 955 MODULE_LICENSE("GPL");
 956 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_INET_DIAG);