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 #include <linux/sock_diag.h>
  37
  38 static const struct inet_diag_handler **inet_diag_table;
  39
  40 struct inet_diag_entry {
  41         __be32 *saddr;
  42         __be32 *daddr;
  43         u16 sport;
  44         u16 dport;
  45         u16 family;
  46         u16 userlocks;
  47 #if IS_ENABLED(CONFIG_IPV6)
  48         struct in6_addr saddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  49         struct in6_addr daddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  50 #endif
  51 };
  52
  53 static DEFINE_MUTEX(inet_diag_table_mutex);
  54
  55 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  56 {
  57         if (!inet_diag_table[proto])
  58                 request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
  59                                NETLINK_SOCK_DIAG, AF_INET, proto);
  60
  61         mutex_lock(&inet_diag_table_mutex);
  62         if (!inet_diag_table[proto])
  63                 return ERR_PTR(-ENOENT);
  64
  65         return inet_diag_table[proto];
  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 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
  75                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
  76                               struct user_namespace *user_ns,
  77                               u32 portid, u32 seq, u16 nlmsg_flags,
  78                               const struct nlmsghdr *unlh)
  79 {
  80         const struct inet_sock *inet = inet_sk(sk);
  81         struct inet_diag_msg *r;
  82         struct nlmsghdr  *nlh;
  83         struct nlattr *attr;
  84         void *info = NULL;
  85         const struct inet_diag_handler *handler;
  86         int ext = req->idiag_ext;
  87
  88         handler = inet_diag_table[req->sdiag_protocol];
  89         BUG_ON(handler == NULL);
  90
  91         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
  92                         nlmsg_flags);
  93         if (!nlh)
  94                 return -EMSGSIZE;
  95
  96         r = nlmsg_data(nlh);
  97         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  98
  99         r->idiag_family = sk->sk_family;
 100         r->idiag_state = sk->sk_state;
 101         r->idiag_timer = 0;
 102         r->idiag_retrans = 0;
 103
 104         r->id.idiag_if = sk->sk_bound_dev_if;
 105         sock_diag_save_cookie(sk, r->id.idiag_cookie);
 106
 107         r->id.idiag_sport = inet->inet_sport;
 108         r->id.idiag_dport = inet->inet_dport;
 109
 110         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 111         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 112
 113         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 114         r->id.idiag_dst[0] = inet->inet_daddr;
 115
 116         if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
 117                 goto errout;
 118
 119         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 120          * hence this needs to be included regardless of socket family.
 121          */
 122         if (ext & (1 << (INET_DIAG_TOS - 1)))
 123                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 124                         goto errout;
 125
 126 #if IS_ENABLED(CONFIG_IPV6)
 127         if (r->idiag_family == AF_INET6) {
 128                 const struct ipv6_pinfo *np = inet6_sk(sk);
 129
 130                 *(struct in6_addr *)r->id.idiag_src = np->rcv_saddr;
 131                 *(struct in6_addr *)r->id.idiag_dst = np->daddr;
 132
 133                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 134                         if (nla_put_u8(skb, INET_DIAG_TCLASS, np->tclass) < 0)
 135                                 goto errout;
 136         }
 137 #endif
 138
 139         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 140         r->idiag_inode = sock_i_ino(sk);
 141
 142         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 143                 struct inet_diag_meminfo minfo = {
 144                         .idiag_rmem = sk_rmem_alloc_get(sk),
 145                         .idiag_wmem = sk->sk_wmem_queued,
 146                         .idiag_fmem = sk->sk_forward_alloc,
 147                         .idiag_tmem = sk_wmem_alloc_get(sk),
 148                 };
 149
 150                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 151                         goto errout;
 152         }
 153
 154         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 155                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 156                         goto errout;
 157
 158         if (icsk == NULL) {
 159                 handler->idiag_get_info(sk, r, NULL);
 160                 goto out;
 161         }
 162
 163 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 164
 165         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
 166             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 167             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 168                 r->idiag_timer = 1;
 169                 r->idiag_retrans = icsk->icsk_retransmits;
 170                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 171         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 172                 r->idiag_timer = 4;
 173                 r->idiag_retrans = icsk->icsk_probes_out;
 174                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 175         } else if (timer_pending(&sk->sk_timer)) {
 176                 r->idiag_timer = 2;
 177                 r->idiag_retrans = icsk->icsk_probes_out;
 178                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 179         } else {
 180                 r->idiag_timer = 0;
 181                 r->idiag_expires = 0;
 182         }
 183 #undef EXPIRES_IN_MS
 184
 185         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 186                 attr = nla_reserve(skb, INET_DIAG_INFO,
 187                                    sizeof(struct tcp_info));
 188                 if (!attr)
 189                         goto errout;
 190
 191                 info = nla_data(attr);
 192         }
 193
 194         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 195                 if (nla_put_string(skb, INET_DIAG_CONG,
 196                                    icsk->icsk_ca_ops->name) < 0)
 197                         goto errout;
 198
 199         handler->idiag_get_info(sk, r, info);
 200
 201         if (sk->sk_state < TCP_TIME_WAIT &&
 202             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 203                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 204
 205 out:
 206         return nlmsg_end(skb, nlh);
 207
 208 errout:
 209         nlmsg_cancel(skb, nlh);
 210         return -EMSGSIZE;
 211 }
 212 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 213
 214 static int inet_csk_diag_fill(struct sock *sk,
 215                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
 216                               struct user_namespace *user_ns,
 217                               u32 portid, u32 seq, u16 nlmsg_flags,
 218                               const struct nlmsghdr *unlh)
 219 {
 220         return inet_sk_diag_fill(sk, inet_csk(sk),
 221                         skb, req, user_ns, portid, seq, nlmsg_flags, unlh);
 222 }
 223
 224 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 225                                struct sk_buff *skb, struct inet_diag_req_v2 *req,
 226                                u32 portid, u32 seq, u16 nlmsg_flags,
 227                                const struct nlmsghdr *unlh)
 228 {
 229         long tmo;
 230         struct inet_diag_msg *r;
 231         struct nlmsghdr *nlh;
 232
 233         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 234                         nlmsg_flags);
 235         if (!nlh)
 236                 return -EMSGSIZE;
 237
 238         r = nlmsg_data(nlh);
 239         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 240
 241         tmo = tw->tw_ttd - jiffies;
 242         if (tmo < 0)
 243                 tmo = 0;
 244
 245         r->idiag_family       = tw->tw_family;
 246         r->idiag_retrans      = 0;
 247
 248         r->id.idiag_if        = tw->tw_bound_dev_if;
 249         sock_diag_save_cookie(tw, r->id.idiag_cookie);
 250
 251         r->id.idiag_sport     = tw->tw_sport;
 252         r->id.idiag_dport     = tw->tw_dport;
 253
 254         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 255         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 256
 257         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 258         r->id.idiag_dst[0]    = tw->tw_daddr;
 259
 260         r->idiag_state        = tw->tw_substate;
 261         r->idiag_timer        = 3;
 262         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 263         r->idiag_rqueue       = 0;
 264         r->idiag_wqueue       = 0;
 265         r->idiag_uid          = 0;
 266         r->idiag_inode        = 0;
 267 #if IS_ENABLED(CONFIG_IPV6)
 268         if (tw->tw_family == AF_INET6) {
 269                 const struct inet6_timewait_sock *tw6 =
 270                                                 inet6_twsk((struct sock *)tw);
 271
 272                 *(struct in6_addr *)r->id.idiag_src = tw6->tw_v6_rcv_saddr;
 273                 *(struct in6_addr *)r->id.idiag_dst = tw6->tw_v6_daddr;
 274         }
 275 #endif
 276
 277         return nlmsg_end(skb, nlh);
 278 }
 279
 280 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 281                         struct inet_diag_req_v2 *r,
 282                         struct user_namespace *user_ns,
 283                         u32 portid, u32 seq, u16 nlmsg_flags,
 284                         const struct nlmsghdr *unlh)
 285 {
 286         if (sk->sk_state == TCP_TIME_WAIT)
 287                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 288                                            skb, r, portid, seq, nlmsg_flags,
 289                                            unlh);
 290         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq, nlmsg_flags, unlh);
 291 }
 292
 293 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *in_skb,
 294                 const struct nlmsghdr *nlh, struct inet_diag_req_v2 *req)
 295 {
 296         int err;
 297         struct sock *sk;
 298         struct sk_buff *rep;
 299         struct net *net = sock_net(in_skb->sk);
 300
 301         err = -EINVAL;
 302         if (req->sdiag_family == AF_INET) {
 303                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 304                                  req->id.idiag_dport, req->id.idiag_src[0],
 305                                  req->id.idiag_sport, req->id.idiag_if);
 306         }
 307 #if IS_ENABLED(CONFIG_IPV6)
 308         else if (req->sdiag_family == AF_INET6) {
 309                 sk = inet6_lookup(net, hashinfo,
 310                                   (struct in6_addr *)req->id.idiag_dst,
 311                                   req->id.idiag_dport,
 312                                   (struct in6_addr *)req->id.idiag_src,
 313                                   req->id.idiag_sport,
 314                                   req->id.idiag_if);
 315         }
 316 #endif
 317         else {
 318                 goto out_nosk;
 319         }
 320
 321         err = -ENOENT;
 322         if (sk == NULL)
 323                 goto out_nosk;
 324
 325         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 326         if (err)
 327                 goto out;
 328
 329         rep = nlmsg_new(sizeof(struct inet_diag_msg) +
 330                         sizeof(struct inet_diag_meminfo) +
 331                         sizeof(struct tcp_info) + 64, GFP_KERNEL);
 332         if (!rep) {
 333                 err = -ENOMEM;
 334                 goto out;
 335         }
 336
 337         err = sk_diag_fill(sk, rep, req,
 338                            sk_user_ns(NETLINK_CB(in_skb).sk),
 339                            NETLINK_CB(in_skb).portid,
 340                            nlh->nlmsg_seq, 0, nlh);
 341         if (err < 0) {
 342                 WARN_ON(err == -EMSGSIZE);
 343                 nlmsg_free(rep);
 344                 goto out;
 345         }
 346         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 347                               MSG_DONTWAIT);
 348         if (err > 0)
 349                 err = 0;
 350
 351 out:
 352         if (sk) {
 353                 if (sk->sk_state == TCP_TIME_WAIT)
 354                         inet_twsk_put((struct inet_timewait_sock *)sk);
 355                 else
 356                         sock_put(sk);
 357         }
 358 out_nosk:
 359         return err;
 360 }
 361 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 362
 363 static int inet_diag_get_exact(struct sk_buff *in_skb,
 364                                const struct nlmsghdr *nlh,
 365                                struct inet_diag_req_v2 *req)
 366 {
 367         const struct inet_diag_handler *handler;
 368         int err;
 369
 370         handler = inet_diag_lock_handler(req->sdiag_protocol);
 371         if (IS_ERR(handler))
 372                 err = PTR_ERR(handler);
 373         else
 374                 err = handler->dump_one(in_skb, nlh, req);
 375         inet_diag_unlock_handler(handler);
 376
 377         return err;
 378 }
 379
 380 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 381 {
 382         int words = bits >> 5;
 383
 384         bits &= 0x1f;
 385
 386         if (words) {
 387                 if (memcmp(a1, a2, words << 2))
 388                         return 0;
 389         }
 390         if (bits) {
 391                 __be32 w1, w2;
 392                 __be32 mask;
 393
 394                 w1 = a1[words];
 395                 w2 = a2[words];
 396
 397                 mask = htonl((0xffffffff) << (32 - bits));
 398
 399                 if ((w1 ^ w2) & mask)
 400                         return 0;
 401         }
 402
 403         return 1;
 404 }
 405
 406
 407 static int inet_diag_bc_run(const struct nlattr *_bc,
 408                 const struct inet_diag_entry *entry)
 409 {
 410         const void *bc = nla_data(_bc);
 411         int len = nla_len(_bc);
 412
 413         while (len > 0) {
 414                 int yes = 1;
 415                 const struct inet_diag_bc_op *op = bc;
 416
 417                 switch (op->code) {
 418                 case INET_DIAG_BC_NOP:
 419                         break;
 420                 case INET_DIAG_BC_JMP:
 421                         yes = 0;
 422                         break;
 423                 case INET_DIAG_BC_S_GE:
 424                         yes = entry->sport >= op[1].no;
 425                         break;
 426                 case INET_DIAG_BC_S_LE:
 427                         yes = entry->sport <= op[1].no;
 428                         break;
 429                 case INET_DIAG_BC_D_GE:
 430                         yes = entry->dport >= op[1].no;
 431                         break;
 432                 case INET_DIAG_BC_D_LE:
 433                         yes = entry->dport <= op[1].no;
 434                         break;
 435                 case INET_DIAG_BC_AUTO:
 436                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 437                         break;
 438                 case INET_DIAG_BC_S_COND:
 439                 case INET_DIAG_BC_D_COND: {
 440                         struct inet_diag_hostcond *cond;
 441                         __be32 *addr;
 442
 443                         cond = (struct inet_diag_hostcond *)(op + 1);
 444                         if (cond->port != -1 &&
 445                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 446                                              entry->sport : entry->dport)) {
 447                                 yes = 0;
 448                                 break;
 449                         }
 450
 451                         if (op->code == INET_DIAG_BC_S_COND)
 452                                 addr = entry->saddr;
 453                         else
 454                                 addr = entry->daddr;
 455
 456                         if (cond->family != AF_UNSPEC &&
 457                             cond->family != entry->family) {
 458                                 if (entry->family == AF_INET6 &&
 459                                     cond->family == AF_INET) {
 460                                         if (addr[0] == 0 && addr[1] == 0 &&
 461                                             addr[2] == htonl(0xffff) &&
 462                                             bitstring_match(addr + 3,
 463                                                             cond->addr,
 464                                                             cond->prefix_len))
 465                                                 break;
 466                                 }
 467                                 yes = 0;
 468                                 break;
 469                         }
 470
 471                         if (cond->prefix_len == 0)
 472                                 break;
 473                         if (bitstring_match(addr, cond->addr,
 474                                             cond->prefix_len))
 475                                 break;
 476                         yes = 0;
 477                         break;
 478                 }
 479                 }
 480
 481                 if (yes) {
 482                         len -= op->yes;
 483                         bc += op->yes;
 484                 } else {
 485                         len -= op->no;
 486                         bc += op->no;
 487                 }
 488         }
 489         return len == 0;
 490 }
 491
 492 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 493 {
 494         struct inet_diag_entry entry;
 495         struct inet_sock *inet = inet_sk(sk);
 496
 497         if (bc == NULL)
 498                 return 1;
 499
 500         entry.family = sk->sk_family;
 501 #if IS_ENABLED(CONFIG_IPV6)
 502         if (entry.family == AF_INET6) {
 503                 struct ipv6_pinfo *np = inet6_sk(sk);
 504
 505                 entry.saddr = np->rcv_saddr.s6_addr32;
 506                 entry.daddr = np->daddr.s6_addr32;
 507         } else
 508 #endif
 509         {
 510                 entry.saddr = &inet->inet_rcv_saddr;
 511                 entry.daddr = &inet->inet_daddr;
 512         }
 513         entry.sport = inet->inet_num;
 514         entry.dport = ntohs(inet->inet_dport);
 515         entry.userlocks = sk->sk_userlocks;
 516
 517         return inet_diag_bc_run(bc, &entry);
 518 }
 519 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 520
 521 static int valid_cc(const void *bc, int len, int cc)
 522 {
 523         while (len >= 0) {
 524                 const struct inet_diag_bc_op *op = bc;
 525
 526                 if (cc > len)
 527                         return 0;
 528                 if (cc == len)
 529                         return 1;
 530                 if (op->yes < 4 || op->yes & 3)
 531                         return 0;
 532                 len -= op->yes;
 533                 bc  += op->yes;
 534         }
 535         return 0;
 536 }
 537
 538 /* Validate an inet_diag_hostcond. */
 539 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 540                            int *min_len)
 541 {
 542         int addr_len;
 543         struct inet_diag_hostcond *cond;
 544
 545         /* Check hostcond space. */
 546         *min_len += sizeof(struct inet_diag_hostcond);
 547         if (len < *min_len)
 548                 return false;
 549         cond = (struct inet_diag_hostcond *)(op + 1);
 550
 551         /* Check address family and address length. */
 552         switch (cond->family) {
 553         case AF_UNSPEC:
 554                 addr_len = 0;
 555                 break;
 556         case AF_INET:
 557                 addr_len = sizeof(struct in_addr);
 558                 break;
 559         case AF_INET6:
 560                 addr_len = sizeof(struct in6_addr);
 561                 break;
 562         default:
 563                 return false;
 564         }
 565         *min_len += addr_len;
 566         if (len < *min_len)
 567                 return false;
 568
 569         /* Check prefix length (in bits) vs address length (in bytes). */
 570         if (cond->prefix_len > 8 * addr_len)
 571                 return false;
 572
 573         return true;
 574 }
 575
 576 /* Validate a port comparison operator. */
 577 static inline bool valid_port_comparison(const struct inet_diag_bc_op *op,
 578                                          int len, int *min_len)
 579 {
 580         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 581         *min_len += sizeof(struct inet_diag_bc_op);
 582         if (len < *min_len)
 583                 return false;
 584         return true;
 585 }
 586
 587 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 588 {
 589         const void *bc = bytecode;
 590         int  len = bytecode_len;
 591
 592         while (len > 0) {
 593                 const struct inet_diag_bc_op *op = bc;
 594                 int min_len = sizeof(struct inet_diag_bc_op);
 595
 596 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 597                 switch (op->code) {
 598                 case INET_DIAG_BC_S_COND:
 599                 case INET_DIAG_BC_D_COND:
 600                         if (!valid_hostcond(bc, len, &min_len))
 601                                 return -EINVAL;
 602                         break;
 603                 case INET_DIAG_BC_S_GE:
 604                 case INET_DIAG_BC_S_LE:
 605                 case INET_DIAG_BC_D_GE:
 606                 case INET_DIAG_BC_D_LE:
 607                         if (!valid_port_comparison(bc, len, &min_len))
 608                                 return -EINVAL;
 609                         break;
 610                 case INET_DIAG_BC_AUTO:
 611                 case INET_DIAG_BC_JMP:
 612                 case INET_DIAG_BC_NOP:
 613                         break;
 614                 default:
 615                         return -EINVAL;
 616                 }
 617
 618                 if (op->code != INET_DIAG_BC_NOP) {
 619                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 620                                 return -EINVAL;
 621                         if (op->no < len &&
 622                             !valid_cc(bytecode, bytecode_len, len - op->no))
 623                                 return -EINVAL;
 624                 }
 625
 626                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 627                         return -EINVAL;
 628                 bc  += op->yes;
 629                 len -= op->yes;
 630         }
 631         return len == 0 ? 0 : -EINVAL;
 632 }
 633
 634 static int inet_csk_diag_dump(struct sock *sk,
 635                               struct sk_buff *skb,
 636                               struct netlink_callback *cb,
 637                               struct inet_diag_req_v2 *r,
 638                               const struct nlattr *bc)
 639 {
 640         if (!inet_diag_bc_sk(bc, sk))
 641                 return 0;
 642
 643         return inet_csk_diag_fill(sk, skb, r,
 644                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 645                                   NETLINK_CB(cb->skb).portid,
 646                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 647 }
 648
 649 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 650                                struct sk_buff *skb,
 651                                struct netlink_callback *cb,
 652                                struct inet_diag_req_v2 *r,
 653                                const struct nlattr *bc)
 654 {
 655         if (bc != NULL) {
 656                 struct inet_diag_entry entry;
 657
 658                 entry.family = tw->tw_family;
 659 #if IS_ENABLED(CONFIG_IPV6)
 660                 if (tw->tw_family == AF_INET6) {
 661                         struct inet6_timewait_sock *tw6 =
 662                                                 inet6_twsk((struct sock *)tw);
 663                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 664                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 665                 } else
 666 #endif
 667                 {
 668                         entry.saddr = &tw->tw_rcv_saddr;
 669                         entry.daddr = &tw->tw_daddr;
 670                 }
 671                 entry.sport = tw->tw_num;
 672                 entry.dport = ntohs(tw->tw_dport);
 673                 entry.userlocks = 0;
 674
 675                 if (!inet_diag_bc_run(bc, &entry))
 676                         return 0;
 677         }
 678
 679         return inet_twsk_diag_fill(tw, skb, r,
 680                                    NETLINK_CB(cb->skb).portid,
 681                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 682 }
 683
 684 /* Get the IPv4, IPv6, or IPv4-mapped-IPv6 local and remote addresses
 685  * from a request_sock. For IPv4-mapped-IPv6 we must map IPv4 to IPv6.
 686  */
 687 static inline void inet_diag_req_addrs(const struct sock *sk,
 688                                        const struct request_sock *req,
 689                                        struct inet_diag_entry *entry)
 690 {
 691         struct inet_request_sock *ireq = inet_rsk(req);
 692
 693 #if IS_ENABLED(CONFIG_IPV6)
 694         if (sk->sk_family == AF_INET6) {
 695                 if (req->rsk_ops->family == AF_INET6) {
 696                         entry->saddr = inet6_rsk(req)->loc_addr.s6_addr32;
 697                         entry->daddr = inet6_rsk(req)->rmt_addr.s6_addr32;
 698                 } else if (req->rsk_ops->family == AF_INET) {
 699                         ipv6_addr_set_v4mapped(ireq->loc_addr,
 700                                                &entry->saddr_storage);
 701                         ipv6_addr_set_v4mapped(ireq->rmt_addr,
 702                                                &entry->daddr_storage);
 703                         entry->saddr = entry->saddr_storage.s6_addr32;
 704                         entry->daddr = entry->daddr_storage.s6_addr32;
 705                 }
 706         } else
 707 #endif
 708         {
 709                 entry->saddr = &ireq->loc_addr;
 710                 entry->daddr = &ireq->rmt_addr;
 711         }
 712 }
 713
 714 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 715                               struct request_sock *req,
 716                               struct user_namespace *user_ns,
 717                               u32 portid, u32 seq,
 718                               const struct nlmsghdr *unlh)
 719 {
 720         const struct inet_request_sock *ireq = inet_rsk(req);
 721         struct inet_sock *inet = inet_sk(sk);
 722         struct inet_diag_msg *r;
 723         struct nlmsghdr *nlh;
 724         long tmo;
 725
 726         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 727                         NLM_F_MULTI);
 728         if (!nlh)
 729                 return -EMSGSIZE;
 730
 731         r = nlmsg_data(nlh);
 732         r->idiag_family = sk->sk_family;
 733         r->idiag_state = TCP_SYN_RECV;
 734         r->idiag_timer = 1;
 735         r->idiag_retrans = req->num_retrans;
 736
 737         r->id.idiag_if = sk->sk_bound_dev_if;
 738         sock_diag_save_cookie(req, r->id.idiag_cookie);
 739
 740         tmo = req->expires - jiffies;
 741         if (tmo < 0)
 742                 tmo = 0;
 743
 744         r->id.idiag_sport = inet->inet_sport;
 745         r->id.idiag_dport = ireq->rmt_port;
 746
 747         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 748         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 749
 750         r->id.idiag_src[0] = ireq->loc_addr;
 751         r->id.idiag_dst[0] = ireq->rmt_addr;
 752
 753         r->idiag_expires = jiffies_to_msecs(tmo);
 754         r->idiag_rqueue = 0;
 755         r->idiag_wqueue = 0;
 756         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 757         r->idiag_inode = 0;
 758 #if IS_ENABLED(CONFIG_IPV6)
 759         if (r->idiag_family == AF_INET6) {
 760                 struct inet_diag_entry entry;
 761                 inet_diag_req_addrs(sk, req, &entry);
 762                 memcpy(r->id.idiag_src, entry.saddr, sizeof(struct in6_addr));
 763                 memcpy(r->id.idiag_dst, entry.daddr, sizeof(struct in6_addr));
 764         }
 765 #endif
 766
 767         return nlmsg_end(skb, nlh);
 768 }
 769
 770 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 771                                struct netlink_callback *cb,
 772                                struct inet_diag_req_v2 *r,
 773                                const struct nlattr *bc)
 774 {
 775         struct inet_diag_entry entry;
 776         struct inet_connection_sock *icsk = inet_csk(sk);
 777         struct listen_sock *lopt;
 778         struct inet_sock *inet = inet_sk(sk);
 779         int j, s_j;
 780         int reqnum, s_reqnum;
 781         int err = 0;
 782
 783         s_j = cb->args[3];
 784         s_reqnum = cb->args[4];
 785
 786         if (s_j > 0)
 787                 s_j--;
 788
 789         entry.family = sk->sk_family;
 790
 791         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 792
 793         lopt = icsk->icsk_accept_queue.listen_opt;
 794         if (!lopt || !lopt->qlen)
 795                 goto out;
 796
 797         if (bc != NULL) {
 798                 entry.sport = inet->inet_num;
 799                 entry.userlocks = sk->sk_userlocks;
 800         }
 801
 802         for (j = s_j; j < lopt->nr_table_entries; j++) {
 803                 struct request_sock *req, *head = lopt->syn_table[j];
 804
 805                 reqnum = 0;
 806                 for (req = head; req; reqnum++, req = req->dl_next) {
 807                         struct inet_request_sock *ireq = inet_rsk(req);
 808
 809                         if (reqnum < s_reqnum)
 810                                 continue;
 811                         if (r->id.idiag_dport != ireq->rmt_port &&
 812                             r->id.idiag_dport)
 813                                 continue;
 814
 815                         if (bc) {
 816                                 inet_diag_req_addrs(sk, req, &entry);
 817                                 entry.dport = ntohs(ireq->rmt_port);
 818
 819                                 if (!inet_diag_bc_run(bc, &entry))
 820                                         continue;
 821                         }
 822
 823                         err = inet_diag_fill_req(skb, sk, req,
 824                                                sk_user_ns(NETLINK_CB(cb->skb).sk),
 825                                                NETLINK_CB(cb->skb).portid,
 826                                                cb->nlh->nlmsg_seq, cb->nlh);
 827                         if (err < 0) {
 828                                 cb->args[3] = j + 1;
 829                                 cb->args[4] = reqnum;
 830                                 goto out;
 831                         }
 832                 }
 833
 834                 s_reqnum = 0;
 835         }
 836
 837 out:
 838         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 839
 840         return err;
 841 }
 842
 843 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 844                 struct netlink_callback *cb, struct inet_diag_req_v2 *r, struct nlattr *bc)
 845 {
 846         int i, num;
 847         int s_i, s_num;
 848         struct net *net = sock_net(skb->sk);
 849
 850         s_i = cb->args[1];
 851         s_num = num = cb->args[2];
 852
 853         if (cb->args[0] == 0) {
 854                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 855                         goto skip_listen_ht;
 856
 857                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 858                         struct sock *sk;
 859                         struct hlist_nulls_node *node;
 860                         struct inet_listen_hashbucket *ilb;
 861
 862                         num = 0;
 863                         ilb = &hashinfo->listening_hash[i];
 864                         spin_lock_bh(&ilb->lock);
 865                         sk_nulls_for_each(sk, node, &ilb->head) {
 866                                 struct inet_sock *inet = inet_sk(sk);
 867
 868                                 if (!net_eq(sock_net(sk), net))
 869                                         continue;
 870
 871                                 if (num < s_num) {
 872                                         num++;
 873                                         continue;
 874                                 }
 875
 876                                 if (r->sdiag_family != AF_UNSPEC &&
 877                                                 sk->sk_family != r->sdiag_family)
 878                                         goto next_listen;
 879
 880                                 if (r->id.idiag_sport != inet->inet_sport &&
 881                                     r->id.idiag_sport)
 882                                         goto next_listen;
 883
 884                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 885                                     r->id.idiag_dport ||
 886                                     cb->args[3] > 0)
 887                                         goto syn_recv;
 888
 889                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 890                                         spin_unlock_bh(&ilb->lock);
 891                                         goto done;
 892                                 }
 893
 894 syn_recv:
 895                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 896                                         goto next_listen;
 897
 898                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 899                                         spin_unlock_bh(&ilb->lock);
 900                                         goto done;
 901                                 }
 902
 903 next_listen:
 904                                 cb->args[3] = 0;
 905                                 cb->args[4] = 0;
 906                                 ++num;
 907                         }
 908                         spin_unlock_bh(&ilb->lock);
 909
 910                         s_num = 0;
 911                         cb->args[3] = 0;
 912                         cb->args[4] = 0;
 913                 }
 914 skip_listen_ht:
 915                 cb->args[0] = 1;
 916                 s_i = num = s_num = 0;
 917         }
 918
 919         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 920                 goto out;
 921
 922         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 923                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 924                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 925                 struct sock *sk;
 926                 struct hlist_nulls_node *node;
 927
 928                 num = 0;
 929
 930                 if (hlist_nulls_empty(&head->chain) &&
 931                         hlist_nulls_empty(&head->twchain))
 932                         continue;
 933
 934                 if (i > s_i)
 935                         s_num = 0;
 936
 937                 spin_lock_bh(lock);
 938                 sk_nulls_for_each(sk, node, &head->chain) {
 939                         struct inet_sock *inet = inet_sk(sk);
 940
 941                         if (!net_eq(sock_net(sk), net))
 942                                 continue;
 943                         if (num < s_num)
 944                                 goto next_normal;
 945                         if (!(r->idiag_states & (1 << sk->sk_state)))
 946                                 goto next_normal;
 947                         if (r->sdiag_family != AF_UNSPEC &&
 948                                         sk->sk_family != r->sdiag_family)
 949                                 goto next_normal;
 950                         if (r->id.idiag_sport != inet->inet_sport &&
 951                             r->id.idiag_sport)
 952                                 goto next_normal;
 953                         if (r->id.idiag_dport != inet->inet_dport &&
 954                             r->id.idiag_dport)
 955                                 goto next_normal;
 956                         if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 957                                 spin_unlock_bh(lock);
 958                                 goto done;
 959                         }
 960 next_normal:
 961                         ++num;
 962                 }
 963
 964                 if (r->idiag_states & (TCPF_TIME_WAIT | TCPF_FIN_WAIT2)) {
 965                         struct inet_timewait_sock *tw;
 966
 967                         inet_twsk_for_each(tw, node,
 968                                     &head->twchain) {
 969                                 if (!net_eq(twsk_net(tw), net))
 970                                         continue;
 971
 972                                 if (num < s_num)
 973                                         goto next_dying;
 974                                 if (!(r->idiag_states & (1 << tw->tw_substate)))
 975                                         goto next_dying;
 976                                 if (r->sdiag_family != AF_UNSPEC &&
 977                                                 tw->tw_family != r->sdiag_family)
 978                                         goto next_dying;
 979                                 if (r->id.idiag_sport != tw->tw_sport &&
 980                                     r->id.idiag_sport)
 981                                         goto next_dying;
 982                                 if (r->id.idiag_dport != tw->tw_dport &&
 983                                     r->id.idiag_dport)
 984                                         goto next_dying;
 985                                 if (inet_twsk_diag_dump(tw, skb, cb, r, bc) < 0) {
 986                                         spin_unlock_bh(lock);
 987                                         goto done;
 988                                 }
 989 next_dying:
 990                                 ++num;
 991                         }
 992                 }
 993                 spin_unlock_bh(lock);
 994         }
 995
 996 done:
 997         cb->args[1] = i;
 998         cb->args[2] = num;
 999 out:
1000         ;
1001 }
1002 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
1003
1004 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
1005                 struct inet_diag_req_v2 *r, struct nlattr *bc)
1006 {
1007         const struct inet_diag_handler *handler;
1008         int err = 0;
1009
1010         handler = inet_diag_lock_handler(r->sdiag_protocol);
1011         if (!IS_ERR(handler))
1012                 handler->dump(skb, cb, r, bc);
1013         else
1014                 err = PTR_ERR(handler);
1015         inet_diag_unlock_handler(handler);
1016
1017         return err ? : skb->len;
1018 }
1019
1020 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
1021 {
1022         struct nlattr *bc = NULL;
1023         int hdrlen = sizeof(struct inet_diag_req_v2);
1024
1025         if (nlmsg_attrlen(cb->nlh, hdrlen))
1026                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1027
1028         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
1029 }
1030
1031 static inline int inet_diag_type2proto(int type)
1032 {
1033         switch (type) {
1034         case TCPDIAG_GETSOCK:
1035                 return IPPROTO_TCP;
1036         case DCCPDIAG_GETSOCK:
1037                 return IPPROTO_DCCP;
1038         default:
1039                 return 0;
1040         }
1041 }
1042
1043 static int inet_diag_dump_compat(struct sk_buff *skb, struct netlink_callback *cb)
1044 {
1045         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
1046         struct inet_diag_req_v2 req;
1047         struct nlattr *bc = NULL;
1048         int hdrlen = sizeof(struct inet_diag_req);
1049
1050         req.sdiag_family = AF_UNSPEC; /* compatibility */
1051         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
1052         req.idiag_ext = rc->idiag_ext;
1053         req.idiag_states = rc->idiag_states;
1054         req.id = rc->id;
1055
1056         if (nlmsg_attrlen(cb->nlh, hdrlen))
1057                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1058
1059         return __inet_diag_dump(skb, cb, &req, bc);
1060 }
1061
1062 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
1063                                const struct nlmsghdr *nlh)
1064 {
1065         struct inet_diag_req *rc = nlmsg_data(nlh);
1066         struct inet_diag_req_v2 req;
1067
1068         req.sdiag_family = rc->idiag_family;
1069         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
1070         req.idiag_ext = rc->idiag_ext;
1071         req.idiag_states = rc->idiag_states;
1072         req.id = rc->id;
1073
1074         return inet_diag_get_exact(in_skb, nlh, &req);
1075 }
1076
1077 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1078 {
1079         int hdrlen = sizeof(struct inet_diag_req);
1080         struct net *net = sock_net(skb->sk);
1081
1082         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1083             nlmsg_len(nlh) < hdrlen)
1084                 return -EINVAL;
1085
1086         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1087                 if (nlmsg_attrlen(nlh, hdrlen)) {
1088                         struct nlattr *attr;
1089
1090                         attr = nlmsg_find_attr(nlh, hdrlen,
1091                                                INET_DIAG_REQ_BYTECODE);
1092                         if (attr == NULL ||
1093                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1094                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1095                                 return -EINVAL;
1096                 }
1097                 {
1098                         struct netlink_dump_control c = {
1099                                 .dump = inet_diag_dump_compat,
1100                         };
1101                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1102                 }
1103         }
1104
1105         return inet_diag_get_exact_compat(skb, nlh);
1106 }
1107
1108 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1109 {
1110         int hdrlen = sizeof(struct inet_diag_req_v2);
1111         struct net *net = sock_net(skb->sk);
1112
1113         if (nlmsg_len(h) < hdrlen)
1114                 return -EINVAL;
1115
1116         if (h->nlmsg_flags & NLM_F_DUMP) {
1117                 if (nlmsg_attrlen(h, hdrlen)) {
1118                         struct nlattr *attr;
1119                         attr = nlmsg_find_attr(h, hdrlen,
1120                                                INET_DIAG_REQ_BYTECODE);
1121                         if (attr == NULL ||
1122                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1123                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1124                                 return -EINVAL;
1125                 }
1126                 {
1127                         struct netlink_dump_control c = {
1128                                 .dump = inet_diag_dump,
1129                         };
1130                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1131                 }
1132         }
1133
1134         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1135 }
1136
1137 static const struct sock_diag_handler inet_diag_handler = {
1138         .family = AF_INET,
1139         .dump = inet_diag_handler_dump,
1140 };
1141
1142 static const struct sock_diag_handler inet6_diag_handler = {
1143         .family = AF_INET6,
1144         .dump = inet_diag_handler_dump,
1145 };
1146
1147 int inet_diag_register(const struct inet_diag_handler *h)
1148 {
1149         const __u16 type = h->idiag_type;
1150         int err = -EINVAL;
1151
1152         if (type >= IPPROTO_MAX)
1153                 goto out;
1154
1155         mutex_lock(&inet_diag_table_mutex);
1156         err = -EEXIST;
1157         if (inet_diag_table[type] == NULL) {
1158                 inet_diag_table[type] = h;
1159                 err = 0;
1160         }
1161         mutex_unlock(&inet_diag_table_mutex);
1162 out:
1163         return err;
1164 }
1165 EXPORT_SYMBOL_GPL(inet_diag_register);
1166
1167 void inet_diag_unregister(const struct inet_diag_handler *h)
1168 {
1169         const __u16 type = h->idiag_type;
1170
1171         if (type >= IPPROTO_MAX)
1172                 return;
1173
1174         mutex_lock(&inet_diag_table_mutex);
1175         inet_diag_table[type] = NULL;
1176         mutex_unlock(&inet_diag_table_mutex);
1177 }
1178 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1179
1180 static int __init inet_diag_init(void)
1181 {
1182         const int inet_diag_table_size = (IPPROTO_MAX *
1183                                           sizeof(struct inet_diag_handler *));
1184         int err = -ENOMEM;
1185
1186         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1187         if (!inet_diag_table)
1188                 goto out;
1189
1190         err = sock_diag_register(&inet_diag_handler);
1191         if (err)
1192                 goto out_free_nl;
1193
1194         err = sock_diag_register(&inet6_diag_handler);
1195         if (err)
1196                 goto out_free_inet;
1197
1198         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1199 out:
1200         return err;
1201
1202 out_free_inet:
1203         sock_diag_unregister(&inet_diag_handler);
1204 out_free_nl:
1205         kfree(inet_diag_table);
1206         goto out;
1207 }
1208
1209 static void __exit inet_diag_exit(void)
1210 {
1211         sock_diag_unregister(&inet6_diag_handler);
1212         sock_diag_unregister(&inet_diag_handler);
1213         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1214         kfree(inet_diag_table);
1215 }
1216
1217 module_init(inet_diag_init);
1218 module_exit(inet_diag_exit);
1219 MODULE_LICENSE("GPL");
1220 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1221 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);