net/ipv4/tcp_ipv4.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  *              IPv4 specific functions
   9  *
  10  *
  11  *              code split from:
  12  *              linux/ipv4/tcp.c
  13  *              linux/ipv4/tcp_input.c
  14  *              linux/ipv4/tcp_output.c
  15  *
  16  *              See tcp.c for author information
  17  *
  18  *      This program is free software; you can redistribute it and/or
  19  *      modify it under the terms of the GNU General Public License
  20  *      as published by the Free Software Foundation; either version
  21  *      2 of the License, or (at your option) any later version.
  22  */
  23
  24 /*
  25  * Changes:
  26  *              David S. Miller :       New socket lookup architecture.
  27  *                                      This code is dedicated to John Dyson.
  28  *              David S. Miller :       Change semantics of established hash,
  29  *                                      half is devoted to TIME_WAIT sockets
  30  *                                      and the rest go in the other half.
  31  *              Andi Kleen :            Add support for syncookies and fixed
  32  *                                      some bugs: ip options weren't passed to
  33  *                                      the TCP layer, missed a check for an
  34  *                                      ACK bit.
  35  *              Andi Kleen :            Implemented fast path mtu discovery.
  36  *                                      Fixed many serious bugs in the
  37  *                                      request_sock handling and moved
  38  *                                      most of it into the af independent code.
  39  *                                      Added tail drop and some other bugfixes.
  40  *                                      Added new listen semantics.
  41  *              Mike McLagan    :       Routing by source
  42  *      Juan Jose Ciarlante:            ip_dynaddr bits
  43  *              Andi Kleen:             various fixes.
  44  *      Vitaly E. Lavrov        :       Transparent proxy revived after year
  45  *                                      coma.
  46  *      Andi Kleen              :       Fix new listen.
  47  *      Andi Kleen              :       Fix accept error reporting.
  48  *      YOSHIFUJI Hideaki @USAGI and:   Support IPV6_V6ONLY socket option, which
  49  *      Alexey Kuznetsov                allow both IPv4 and IPv6 sockets to bind
  50  *                                      a single port at the same time.
  51  */
  52
  53 #define pr_fmt(fmt) "TCP: " fmt
  54
  55 #include <linux/bottom_half.h>
  56 #include <linux/types.h>
  57 #include <linux/fcntl.h>
  58 #include <linux/module.h>
  59 #include <linux/random.h>
  60 #include <linux/cache.h>
  61 #include <linux/jhash.h>
  62 #include <linux/init.h>
  63 #include <linux/times.h>
  64 #include <linux/slab.h>
  65
  66 #include <net/net_namespace.h>
  67 #include <net/icmp.h>
  68 #include <net/inet_hashtables.h>
  69 #include <net/tcp.h>
  70 #include <net/transp_v6.h>
  71 #include <net/ipv6.h>
  72 #include <net/inet_common.h>
  73 #include <net/timewait_sock.h>
  74 #include <net/xfrm.h>
  75 #include <net/secure_seq.h>
  76 #include <net/tcp_memcontrol.h>
  77 #include <net/busy_poll.h>
  78
  79 #include <linux/inet.h>
  80 #include <linux/ipv6.h>
  81 #include <linux/stddef.h>
  82 #include <linux/proc_fs.h>
  83 #include <linux/seq_file.h>
  84
  85 #include <linux/crypto.h>
  86 #include <linux/scatterlist.h>
  87
  88 int sysctl_tcp_tw_reuse __read_mostly;
  89 int sysctl_tcp_low_latency __read_mostly;
  90 EXPORT_SYMBOL(sysctl_tcp_low_latency);
  91
  92 #ifdef CONFIG_TCP_MD5SIG
  93 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
  94                                __be32 daddr, __be32 saddr, const struct tcphdr *th);
  95 #endif
  96
  97 struct inet_hashinfo tcp_hashinfo;
  98 EXPORT_SYMBOL(tcp_hashinfo);
  99
 100 static  __u32 tcp_v4_init_sequence(const struct sk_buff *skb)
 101 {
 102         return secure_tcp_sequence_number(ip_hdr(skb)->daddr,
 103                                           ip_hdr(skb)->saddr,
 104                                           tcp_hdr(skb)->dest,
 105                                           tcp_hdr(skb)->source);
 106 }
 107
 108 int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
 109 {
 110         const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
 111         struct tcp_sock *tp = tcp_sk(sk);
 112
 113         /* With PAWS, it is safe from the viewpoint
 114            of data integrity. Even without PAWS it is safe provided sequence
 115            spaces do not overlap i.e. at data rates <= 80Mbit/sec.
 116
 117            Actually, the idea is close to VJ's one, only timestamp cache is
 118            held not per host, but per port pair and TW bucket is used as state
 119            holder.
 120
 121            If TW bucket has been already destroyed we fall back to VJ's scheme
 122            and use initial timestamp retrieved from peer table.
 123          */
 124         if (tcptw->tw_ts_recent_stamp &&
 125             (!twp || (sysctl_tcp_tw_reuse &&
 126                              get_seconds() - tcptw->tw_ts_recent_stamp > 1))) {
 127                 tp->write_seq = tcptw->tw_snd_nxt + 65535 + 2;
 128                 if (tp->write_seq == 0)
 129                         tp->write_seq = 1;
 130                 tp->rx_opt.ts_recent       = tcptw->tw_ts_recent;
 131                 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
 132                 sock_hold(sktw);
 133                 return 1;
 134         }
 135
 136         return 0;
 137 }
 138 EXPORT_SYMBOL_GPL(tcp_twsk_unique);
 139
 140 /* This will initiate an outgoing connection. */
 141 int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
 142 {
 143         struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
 144         struct inet_sock *inet = inet_sk(sk);
 145         struct tcp_sock *tp = tcp_sk(sk);
 146         __be16 orig_sport, orig_dport;
 147         __be32 daddr, nexthop;
 148         struct flowi4 *fl4;
 149         struct rtable *rt;
 150         int err;
 151         struct ip_options_rcu *inet_opt;
 152
 153         if (addr_len < sizeof(struct sockaddr_in))
 154                 return -EINVAL;
 155
 156         if (usin->sin_family != AF_INET)
 157                 return -EAFNOSUPPORT;
 158
 159         nexthop = daddr = usin->sin_addr.s_addr;
 160         inet_opt = rcu_dereference_protected(inet->inet_opt,
 161                                              sock_owned_by_user(sk));
 162         if (inet_opt && inet_opt->opt.srr) {
 163                 if (!daddr)
 164                         return -EINVAL;
 165                 nexthop = inet_opt->opt.faddr;
 166         }
 167
 168         orig_sport = inet->inet_sport;
 169         orig_dport = usin->sin_port;
 170         fl4 = &inet->cork.fl.u.ip4;
 171         rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
 172                               RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
 173                               IPPROTO_TCP,
 174                               orig_sport, orig_dport, sk);
 175         if (IS_ERR(rt)) {
 176                 err = PTR_ERR(rt);
 177                 if (err == -ENETUNREACH)
 178                         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 179                 return err;
 180         }
 181
 182         if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
 183                 ip_rt_put(rt);
 184                 return -ENETUNREACH;
 185         }
 186
 187         if (!inet_opt || !inet_opt->opt.srr)
 188                 daddr = fl4->daddr;
 189
 190         if (!inet->inet_saddr)
 191                 inet->inet_saddr = fl4->saddr;
 192         sk_rcv_saddr_set(sk, inet->inet_saddr);
 193
 194         if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
 195                 /* Reset inherited state */
 196                 tp->rx_opt.ts_recent       = 0;
 197                 tp->rx_opt.ts_recent_stamp = 0;
 198                 if (likely(!tp->repair))
 199                         tp->write_seq      = 0;
 200         }
 201
 202         if (tcp_death_row.sysctl_tw_recycle &&
 203             !tp->rx_opt.ts_recent_stamp && fl4->daddr == daddr)
 204                 tcp_fetch_timewait_stamp(sk, &rt->dst);
 205
 206         inet->inet_dport = usin->sin_port;
 207         sk_daddr_set(sk, daddr);
 208
 209         inet_csk(sk)->icsk_ext_hdr_len = 0;
 210         if (inet_opt)
 211                 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
 212
 213         tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
 214
 215         /* Socket identity is still unknown (sport may be zero).
 216          * However we set state to SYN-SENT and not releasing socket
 217          * lock select source port, enter ourselves into the hash tables and
 218          * complete initialization after this.
 219          */
 220         tcp_set_state(sk, TCP_SYN_SENT);
 221         err = inet_hash_connect(&tcp_death_row, sk);
 222         if (err)
 223                 goto failure;
 224
 225         sk_set_txhash(sk);
 226
 227         rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
 228                                inet->inet_sport, inet->inet_dport, sk);
 229         if (IS_ERR(rt)) {
 230                 err = PTR_ERR(rt);
 231                 rt = NULL;
 232                 goto failure;
 233         }
 234         /* OK, now commit destination to socket.  */
 235         sk->sk_gso_type = SKB_GSO_TCPV4;
 236         sk_setup_caps(sk, &rt->dst);
 237
 238         if (!tp->write_seq && likely(!tp->repair))
 239                 tp->write_seq = secure_tcp_sequence_number(inet->inet_saddr,
 240                                                            inet->inet_daddr,
 241                                                            inet->inet_sport,
 242                                                            usin->sin_port);
 243
 244         inet->inet_id = tp->write_seq ^ jiffies;
 245
 246         err = tcp_connect(sk);
 247
 248         rt = NULL;
 249         if (err)
 250                 goto failure;
 251
 252         return 0;
 253
 254 failure:
 255         /*
 256          * This unhashes the socket and releases the local port,
 257          * if necessary.
 258          */
 259         tcp_set_state(sk, TCP_CLOSE);
 260         ip_rt_put(rt);
 261         sk->sk_route_caps = 0;
 262         inet->inet_dport = 0;
 263         return err;
 264 }
 265 EXPORT_SYMBOL(tcp_v4_connect);
 266
 267 /*
 268  * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
 269  * It can be called through tcp_release_cb() if socket was owned by user
 270  * at the time tcp_v4_err() was called to handle ICMP message.
 271  */
 272 void tcp_v4_mtu_reduced(struct sock *sk)
 273 {
 274         struct dst_entry *dst;
 275         struct inet_sock *inet = inet_sk(sk);
 276         u32 mtu = tcp_sk(sk)->mtu_info;
 277
 278         dst = inet_csk_update_pmtu(sk, mtu);
 279         if (!dst)
 280                 return;
 281
 282         /* Something is about to be wrong... Remember soft error
 283          * for the case, if this connection will not able to recover.
 284          */
 285         if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
 286                 sk->sk_err_soft = EMSGSIZE;
 287
 288         mtu = dst_mtu(dst);
 289
 290         if (inet->pmtudisc != IP_PMTUDISC_DONT &&
 291             ip_sk_accept_pmtu(sk) &&
 292             inet_csk(sk)->icsk_pmtu_cookie > mtu) {
 293                 tcp_sync_mss(sk, mtu);
 294
 295                 /* Resend the TCP packet because it's
 296                  * clear that the old packet has been
 297                  * dropped. This is the new "fast" path mtu
 298                  * discovery.
 299                  */
 300                 tcp_simple_retransmit(sk);
 301         } /* else let the usual retransmit timer handle it */
 302 }
 303 EXPORT_SYMBOL(tcp_v4_mtu_reduced);
 304
 305 static void do_redirect(struct sk_buff *skb, struct sock *sk)
 306 {
 307         struct dst_entry *dst = __sk_dst_check(sk, 0);
 308
 309         if (dst)
 310                 dst->ops->redirect(dst, sk, skb);
 311 }
 312
 313
 314 /* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
 315 void tcp_req_err(struct sock *sk, u32 seq)
 316 {
 317         struct request_sock *req = inet_reqsk(sk);
 318         struct net *net = sock_net(sk);
 319
 320         /* ICMPs are not backlogged, hence we cannot get
 321          * an established socket here.
 322          */
 323         WARN_ON(req->sk);
 324
 325         if (seq != tcp_rsk(req)->snt_isn) {
 326                 NET_INC_STATS_BH(net, LINUX_MIB_OUTOFWINDOWICMPS);
 327                 reqsk_put(req);
 328         } else {
 329                 /*
 330                  * Still in SYN_RECV, just remove it silently.
 331                  * There is no good way to pass the error to the newly
 332                  * created socket, and POSIX does not want network
 333                  * errors returned from accept().
 334                  */
 335                 NET_INC_STATS_BH(net, LINUX_MIB_LISTENDROPS);
 336                 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
 337         }
 338 }
 339 EXPORT_SYMBOL(tcp_req_err);
 340
 341 /*
 342  * This routine is called by the ICMP module when it gets some
 343  * sort of error condition.  If err < 0 then the socket should
 344  * be closed and the error returned to the user.  If err > 0
 345  * it's just the icmp type << 8 | icmp code.  After adjustment
 346  * header points to the first 8 bytes of the tcp header.  We need
 347  * to find the appropriate port.
 348  *
 349  * The locking strategy used here is very "optimistic". When
 350  * someone else accesses the socket the ICMP is just dropped
 351  * and for some paths there is no check at all.
 352  * A more general error queue to queue errors for later handling
 353  * is probably better.
 354  *
 355  */
 356
 357 void tcp_v4_err(struct sk_buff *icmp_skb, u32 info)
 358 {
 359         const struct iphdr *iph = (const struct iphdr *)icmp_skb->data;
 360         struct tcphdr *th = (struct tcphdr *)(icmp_skb->data + (iph->ihl << 2));
 361         struct inet_connection_sock *icsk;
 362         struct tcp_sock *tp;
 363         struct inet_sock *inet;
 364         const int type = icmp_hdr(icmp_skb)->type;
 365         const int code = icmp_hdr(icmp_skb)->code;
 366         struct sock *sk;
 367         struct sk_buff *skb;
 368         struct request_sock *fastopen;
 369         __u32 seq, snd_una;
 370         __u32 remaining;
 371         int err;
 372         struct net *net = dev_net(icmp_skb->dev);
 373
 374         sk = __inet_lookup_established(net, &tcp_hashinfo, iph->daddr,
 375                                        th->dest, iph->saddr, ntohs(th->source),
 376                                        inet_iif(icmp_skb));
 377         if (!sk) {
 378                 ICMP_INC_STATS_BH(net, ICMP_MIB_INERRORS);
 379                 return;
 380         }
 381         if (sk->sk_state == TCP_TIME_WAIT) {
 382                 inet_twsk_put(inet_twsk(sk));
 383                 return;
 384         }
 385         seq = ntohl(th->seq);
 386         if (sk->sk_state == TCP_NEW_SYN_RECV)
 387                 return tcp_req_err(sk, seq);
 388
 389         bh_lock_sock(sk);
 390         /* If too many ICMPs get dropped on busy
 391          * servers this needs to be solved differently.
 392          * We do take care of PMTU discovery (RFC1191) special case :
 393          * we can receive locally generated ICMP messages while socket is held.
 394          */
 395         if (sock_owned_by_user(sk)) {
 396                 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
 397                         NET_INC_STATS_BH(net, LINUX_MIB_LOCKDROPPEDICMPS);
 398         }
 399         if (sk->sk_state == TCP_CLOSE)
 400                 goto out;
 401
 402         if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
 403                 NET_INC_STATS_BH(net, LINUX_MIB_TCPMINTTLDROP);
 404                 goto out;
 405         }
 406
 407         icsk = inet_csk(sk);
 408         tp = tcp_sk(sk);
 409         /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
 410         fastopen = tp->fastopen_rsk;
 411         snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
 412         if (sk->sk_state != TCP_LISTEN &&
 413             !between(seq, snd_una, tp->snd_nxt)) {
 414                 NET_INC_STATS_BH(net, LINUX_MIB_OUTOFWINDOWICMPS);
 415                 goto out;
 416         }
 417
 418         switch (type) {
 419         case ICMP_REDIRECT:
 420                 do_redirect(icmp_skb, sk);
 421                 goto out;
 422         case ICMP_SOURCE_QUENCH:
 423                 /* Just silently ignore these. */
 424                 goto out;
 425         case ICMP_PARAMETERPROB:
 426                 err = EPROTO;
 427                 break;
 428         case ICMP_DEST_UNREACH:
 429                 if (code > NR_ICMP_UNREACH)
 430                         goto out;
 431
 432                 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
 433                         /* We are not interested in TCP_LISTEN and open_requests
 434                          * (SYN-ACKs send out by Linux are always <576bytes so
 435                          * they should go through unfragmented).
 436                          */
 437                         if (sk->sk_state == TCP_LISTEN)
 438                                 goto out;
 439
 440                         tp->mtu_info = info;
 441                         if (!sock_owned_by_user(sk)) {
 442                                 tcp_v4_mtu_reduced(sk);
 443                         } else {
 444                                 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &tp->tsq_flags))
 445                                         sock_hold(sk);
 446                         }
 447                         goto out;
 448                 }
 449
 450                 err = icmp_err_convert[code].errno;
 451                 /* check if icmp_skb allows revert of backoff
 452                  * (see draft-zimmermann-tcp-lcd) */
 453                 if (code != ICMP_NET_UNREACH && code != ICMP_HOST_UNREACH)
 454                         break;
 455                 if (seq != tp->snd_una  || !icsk->icsk_retransmits ||
 456                     !icsk->icsk_backoff || fastopen)
 457                         break;
 458
 459                 if (sock_owned_by_user(sk))
 460                         break;
 461
 462                 icsk->icsk_backoff--;
 463                 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) :
 464                                                TCP_TIMEOUT_INIT;
 465                 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
 466
 467                 skb = tcp_write_queue_head(sk);
 468                 BUG_ON(!skb);
 469
 470                 remaining = icsk->icsk_rto -
 471                             min(icsk->icsk_rto,
 472                                 tcp_time_stamp - tcp_skb_timestamp(skb));
 473
 474                 if (remaining) {
 475                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 476                                                   remaining, TCP_RTO_MAX);
 477                 } else {
 478                         /* RTO revert clocked out retransmission.
 479                          * Will retransmit now */
 480                         tcp_retransmit_timer(sk);
 481                 }
 482
 483                 break;
 484         case ICMP_TIME_EXCEEDED:
 485                 err = EHOSTUNREACH;
 486                 break;
 487         default:
 488                 goto out;
 489         }
 490
 491         switch (sk->sk_state) {
 492         case TCP_SYN_SENT:
 493         case TCP_SYN_RECV:
 494                 /* Only in fast or simultaneous open. If a fast open socket is
 495                  * is already accepted it is treated as a connected one below.
 496                  */
 497                 if (fastopen && !fastopen->sk)
 498                         break;
 499
 500                 if (!sock_owned_by_user(sk)) {
 501                         sk->sk_err = err;
 502
 503                         sk->sk_error_report(sk);
 504
 505                         tcp_done(sk);
 506                 } else {
 507                         sk->sk_err_soft = err;
 508                 }
 509                 goto out;
 510         }
 511
 512         /* If we've already connected we will keep trying
 513          * until we time out, or the user gives up.
 514          *
 515          * rfc1122 4.2.3.9 allows to consider as hard errors
 516          * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
 517          * but it is obsoleted by pmtu discovery).
 518          *
 519          * Note, that in modern internet, where routing is unreliable
 520          * and in each dark corner broken firewalls sit, sending random
 521          * errors ordered by their masters even this two messages finally lose
 522          * their original sense (even Linux sends invalid PORT_UNREACHs)
 523          *
 524          * Now we are in compliance with RFCs.
 525          *                                                      --ANK (980905)
 526          */
 527
 528         inet = inet_sk(sk);
 529         if (!sock_owned_by_user(sk) && inet->recverr) {
 530                 sk->sk_err = err;
 531                 sk->sk_error_report(sk);
 532         } else  { /* Only an error on timeout */
 533                 sk->sk_err_soft = err;
 534         }
 535
 536 out:
 537         bh_unlock_sock(sk);
 538         sock_put(sk);
 539 }
 540
 541 void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
 542 {
 543         struct tcphdr *th = tcp_hdr(skb);
 544
 545         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 546                 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
 547                 skb->csum_start = skb_transport_header(skb) - skb->head;
 548                 skb->csum_offset = offsetof(struct tcphdr, check);
 549         } else {
 550                 th->check = tcp_v4_check(skb->len, saddr, daddr,
 551                                          csum_partial(th,
 552                                                       th->doff << 2,
 553                                                       skb->csum));
 554         }
 555 }
 556
 557 /* This routine computes an IPv4 TCP checksum. */
 558 void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
 559 {
 560         const struct inet_sock *inet = inet_sk(sk);
 561
 562         __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
 563 }
 564 EXPORT_SYMBOL(tcp_v4_send_check);
 565
 566 /*
 567  *      This routine will send an RST to the other tcp.
 568  *
 569  *      Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
 570  *                    for reset.
 571  *      Answer: if a packet caused RST, it is not for a socket
 572  *              existing in our system, if it is matched to a socket,
 573  *              it is just duplicate segment or bug in other side's TCP.
 574  *              So that we build reply only basing on parameters
 575  *              arrived with segment.
 576  *      Exception: precedence violation. We do not implement it in any case.
 577  */
 578
 579 static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
 580 {
 581         const struct tcphdr *th = tcp_hdr(skb);
 582         struct {
 583                 struct tcphdr th;
 584 #ifdef CONFIG_TCP_MD5SIG
 585                 __be32 opt[(TCPOLEN_MD5SIG_ALIGNED >> 2)];
 586 #endif
 587         } rep;
 588         struct ip_reply_arg arg;
 589 #ifdef CONFIG_TCP_MD5SIG
 590         struct tcp_md5sig_key *key;
 591         const __u8 *hash_location = NULL;
 592         unsigned char newhash[16];
 593         int genhash;
 594         struct sock *sk1 = NULL;
 595 #endif
 596         struct net *net;
 597
 598         /* Never send a reset in response to a reset. */
 599         if (th->rst)
 600                 return;
 601
 602         /* If sk not NULL, it means we did a successful lookup and incoming
 603          * route had to be correct. prequeue might have dropped our dst.
 604          */
 605         if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
 606                 return;
 607
 608         /* Swap the send and the receive. */
 609         memset(&rep, 0, sizeof(rep));
 610         rep.th.dest   = th->source;
 611         rep.th.source = th->dest;
 612         rep.th.doff   = sizeof(struct tcphdr) / 4;
 613         rep.th.rst    = 1;
 614
 615         if (th->ack) {
 616                 rep.th.seq = th->ack_seq;
 617         } else {
 618                 rep.th.ack = 1;
 619                 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
 620                                        skb->len - (th->doff << 2));
 621         }
 622
 623         memset(&arg, 0, sizeof(arg));
 624         arg.iov[0].iov_base = (unsigned char *)&rep;
 625         arg.iov[0].iov_len  = sizeof(rep.th);
 626
 627         net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
 628 #ifdef CONFIG_TCP_MD5SIG
 629         hash_location = tcp_parse_md5sig_option(th);
 630         if (!sk && hash_location) {
 631                 /*
 632                  * active side is lost. Try to find listening socket through
 633                  * source port, and then find md5 key through listening socket.
 634                  * we are not loose security here:
 635                  * Incoming packet is checked with md5 hash with finding key,
 636                  * no RST generated if md5 hash doesn't match.
 637                  */
 638                 sk1 = __inet_lookup_listener(net,
 639                                              &tcp_hashinfo, ip_hdr(skb)->saddr,
 640                                              th->source, ip_hdr(skb)->daddr,
 641                                              ntohs(th->source), inet_iif(skb));
 642                 /* don't send rst if it can't find key */
 643                 if (!sk1)
 644                         return;
 645                 rcu_read_lock();
 646                 key = tcp_md5_do_lookup(sk1, (union tcp_md5_addr *)
 647                                         &ip_hdr(skb)->saddr, AF_INET);
 648                 if (!key)
 649                         goto release_sk1;
 650
 651                 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
 652                 if (genhash || memcmp(hash_location, newhash, 16) != 0)
 653                         goto release_sk1;
 654         } else {
 655                 key = sk ? tcp_md5_do_lookup(sk, (union tcp_md5_addr *)
 656                                              &ip_hdr(skb)->saddr,
 657                                              AF_INET) : NULL;
 658         }
 659
 660         if (key) {
 661                 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
 662                                    (TCPOPT_NOP << 16) |
 663                                    (TCPOPT_MD5SIG << 8) |
 664                                    TCPOLEN_MD5SIG);
 665                 /* Update length and the length the header thinks exists */
 666                 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
 667                 rep.th.doff = arg.iov[0].iov_len / 4;
 668
 669                 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
 670                                      key, ip_hdr(skb)->saddr,
 671                                      ip_hdr(skb)->daddr, &rep.th);
 672         }
 673 #endif
 674         arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
 675                                       ip_hdr(skb)->saddr, /* XXX */
 676                                       arg.iov[0].iov_len, IPPROTO_TCP, 0);
 677         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
 678         arg.flags = (sk && inet_sk(sk)->transparent) ? IP_REPLY_ARG_NOSRCCHECK : 0;
 679         /* When socket is gone, all binding information is lost.
 680          * routing might fail in this case. No choice here, if we choose to force
 681          * input interface, we will misroute in case of asymmetric route.
 682          */
 683         if (sk)
 684                 arg.bound_dev_if = sk->sk_bound_dev_if;
 685
 686         arg.tos = ip_hdr(skb)->tos;
 687         ip_send_unicast_reply(*this_cpu_ptr(net->ipv4.tcp_sk),
 688                               skb, &TCP_SKB_CB(skb)->header.h4.opt,
 689                               ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
 690                               &arg, arg.iov[0].iov_len);
 691
 692         TCP_INC_STATS_BH(net, TCP_MIB_OUTSEGS);
 693         TCP_INC_STATS_BH(net, TCP_MIB_OUTRSTS);
 694
 695 #ifdef CONFIG_TCP_MD5SIG
 696 release_sk1:
 697         if (sk1) {
 698                 rcu_read_unlock();
 699                 sock_put(sk1);
 700         }
 701 #endif
 702 }
 703
 704 /* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
 705    outside socket context is ugly, certainly. What can I do?
 706  */
 707
 708 static void tcp_v4_send_ack(struct sk_buff *skb, u32 seq, u32 ack,
 709                             u32 win, u32 tsval, u32 tsecr, int oif,
 710                             struct tcp_md5sig_key *key,
 711                             int reply_flags, u8 tos)
 712 {
 713         const struct tcphdr *th = tcp_hdr(skb);
 714         struct {
 715                 struct tcphdr th;
 716                 __be32 opt[(TCPOLEN_TSTAMP_ALIGNED >> 2)
 717 #ifdef CONFIG_TCP_MD5SIG
 718                            + (TCPOLEN_MD5SIG_ALIGNED >> 2)
 719 #endif
 720                         ];
 721         } rep;
 722         struct ip_reply_arg arg;
 723         struct net *net = dev_net(skb_dst(skb)->dev);
 724
 725         memset(&rep.th, 0, sizeof(struct tcphdr));
 726         memset(&arg, 0, sizeof(arg));
 727
 728         arg.iov[0].iov_base = (unsigned char *)&rep;
 729         arg.iov[0].iov_len  = sizeof(rep.th);
 730         if (tsecr) {
 731                 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 732                                    (TCPOPT_TIMESTAMP << 8) |
 733                                    TCPOLEN_TIMESTAMP);
 734                 rep.opt[1] = htonl(tsval);
 735                 rep.opt[2] = htonl(tsecr);
 736                 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
 737         }
 738
 739         /* Swap the send and the receive. */
 740         rep.th.dest    = th->source;
 741         rep.th.source  = th->dest;
 742         rep.th.doff    = arg.iov[0].iov_len / 4;
 743         rep.th.seq     = htonl(seq);
 744         rep.th.ack_seq = htonl(ack);
 745         rep.th.ack     = 1;
 746         rep.th.window  = htons(win);
 747
 748 #ifdef CONFIG_TCP_MD5SIG
 749         if (key) {
 750                 int offset = (tsecr) ? 3 : 0;
 751
 752                 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
 753                                           (TCPOPT_NOP << 16) |
 754                                           (TCPOPT_MD5SIG << 8) |
 755                                           TCPOLEN_MD5SIG);
 756                 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
 757                 rep.th.doff = arg.iov[0].iov_len/4;
 758
 759                 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
 760                                     key, ip_hdr(skb)->saddr,
 761                                     ip_hdr(skb)->daddr, &rep.th);
 762         }
 763 #endif
 764         arg.flags = reply_flags;
 765         arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
 766                                       ip_hdr(skb)->saddr, /* XXX */
 767                                       arg.iov[0].iov_len, IPPROTO_TCP, 0);
 768         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
 769         if (oif)
 770                 arg.bound_dev_if = oif;
 771         arg.tos = tos;
 772         ip_send_unicast_reply(*this_cpu_ptr(net->ipv4.tcp_sk),
 773                               skb, &TCP_SKB_CB(skb)->header.h4.opt,
 774                               ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
 775                               &arg, arg.iov[0].iov_len);
 776
 777         TCP_INC_STATS_BH(net, TCP_MIB_OUTSEGS);
 778 }
 779
 780 static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
 781 {
 782         struct inet_timewait_sock *tw = inet_twsk(sk);
 783         struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
 784
 785         tcp_v4_send_ack(skb, tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
 786                         tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
 787                         tcp_time_stamp + tcptw->tw_ts_offset,
 788                         tcptw->tw_ts_recent,
 789                         tw->tw_bound_dev_if,
 790                         tcp_twsk_md5_key(tcptw),
 791                         tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
 792                         tw->tw_tos
 793                         );
 794
 795         inet_twsk_put(tw);
 796 }
 797
 798 static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
 799                                   struct request_sock *req)
 800 {
 801         /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
 802          * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
 803          */
 804         tcp_v4_send_ack(skb, (sk->sk_state == TCP_LISTEN) ?
 805                         tcp_rsk(req)->snt_isn + 1 : tcp_sk(sk)->snd_nxt,
 806                         tcp_rsk(req)->rcv_nxt, req->rcv_wnd,
 807                         tcp_time_stamp,
 808                         req->ts_recent,
 809                         0,
 810                         tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&ip_hdr(skb)->daddr,
 811                                           AF_INET),
 812                         inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
 813                         ip_hdr(skb)->tos);
 814 }
 815
 816 /*
 817  *      Send a SYN-ACK after having received a SYN.
 818  *      This still operates on a request_sock only, not on a big
 819  *      socket.
 820  */
 821 static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
 822                               struct flowi *fl,
 823                               struct request_sock *req,
 824                               u16 queue_mapping,
 825                               struct tcp_fastopen_cookie *foc)
 826 {
 827         const struct inet_request_sock *ireq = inet_rsk(req);
 828         struct flowi4 fl4;
 829         int err = -1;
 830         struct sk_buff *skb;
 831
 832         /* First, grab a route. */
 833         if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
 834                 return -1;
 835
 836         skb = tcp_make_synack(sk, dst, req, foc);
 837
 838         if (skb) {
 839                 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
 840
 841                 skb_set_queue_mapping(skb, queue_mapping);
 842                 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
 843                                             ireq->ir_rmt_addr,
 844                                             ireq->opt);
 845                 err = net_xmit_eval(err);
 846         }
 847
 848         return err;
 849 }
 850
 851 /*
 852  *      IPv4 request_sock destructor.
 853  */
 854 static void tcp_v4_reqsk_destructor(struct request_sock *req)
 855 {
 856         kfree(inet_rsk(req)->opt);
 857 }
 858
 859
 860 #ifdef CONFIG_TCP_MD5SIG
 861 /*
 862  * RFC2385 MD5 checksumming requires a mapping of
 863  * IP address->MD5 Key.
 864  * We need to maintain these in the sk structure.
 865  */
 866
 867 /* Find the Key structure for an address.  */
 868 struct tcp_md5sig_key *tcp_md5_do_lookup(const struct sock *sk,
 869                                          const union tcp_md5_addr *addr,
 870                                          int family)
 871 {
 872         const struct tcp_sock *tp = tcp_sk(sk);
 873         struct tcp_md5sig_key *key;
 874         unsigned int size = sizeof(struct in_addr);
 875         const struct tcp_md5sig_info *md5sig;
 876
 877         /* caller either holds rcu_read_lock() or socket lock */
 878         md5sig = rcu_dereference_check(tp->md5sig_info,
 879                                        sock_owned_by_user(sk) ||
 880                                        lockdep_is_held((spinlock_t *)&sk->sk_lock.slock));
 881         if (!md5sig)
 882                 return NULL;
 883 #if IS_ENABLED(CONFIG_IPV6)
 884         if (family == AF_INET6)
 885                 size = sizeof(struct in6_addr);
 886 #endif
 887         hlist_for_each_entry_rcu(key, &md5sig->head, node) {
 888                 if (key->family != family)
 889                         continue;
 890                 if (!memcmp(&key->addr, addr, size))
 891                         return key;
 892         }
 893         return NULL;
 894 }
 895 EXPORT_SYMBOL(tcp_md5_do_lookup);
 896
 897 struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
 898                                          const struct sock *addr_sk)
 899 {
 900         const union tcp_md5_addr *addr;
 901
 902         addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
 903         return tcp_md5_do_lookup(sk, addr, AF_INET);
 904 }
 905 EXPORT_SYMBOL(tcp_v4_md5_lookup);
 906
 907 /* This can be called on a newly created socket, from other files */
 908 int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 909                    int family, const u8 *newkey, u8 newkeylen, gfp_t gfp)
 910 {
 911         /* Add Key to the list */
 912         struct tcp_md5sig_key *key;
 913         struct tcp_sock *tp = tcp_sk(sk);
 914         struct tcp_md5sig_info *md5sig;
 915
 916         key = tcp_md5_do_lookup(sk, addr, family);
 917         if (key) {
 918                 /* Pre-existing entry - just update that one. */
 919                 memcpy(key->key, newkey, newkeylen);
 920                 key->keylen = newkeylen;
 921                 return 0;
 922         }
 923
 924         md5sig = rcu_dereference_protected(tp->md5sig_info,
 925                                            sock_owned_by_user(sk));
 926         if (!md5sig) {
 927                 md5sig = kmalloc(sizeof(*md5sig), gfp);
 928                 if (!md5sig)
 929                         return -ENOMEM;
 930
 931                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
 932                 INIT_HLIST_HEAD(&md5sig->head);
 933                 rcu_assign_pointer(tp->md5sig_info, md5sig);
 934         }
 935
 936         key = sock_kmalloc(sk, sizeof(*key), gfp);
 937         if (!key)
 938                 return -ENOMEM;
 939         if (!tcp_alloc_md5sig_pool()) {
 940                 sock_kfree_s(sk, key, sizeof(*key));
 941                 return -ENOMEM;
 942         }
 943
 944         memcpy(key->key, newkey, newkeylen);
 945         key->keylen = newkeylen;
 946         key->family = family;
 947         memcpy(&key->addr, addr,
 948                (family == AF_INET6) ? sizeof(struct in6_addr) :
 949                                       sizeof(struct in_addr));
 950         hlist_add_head_rcu(&key->node, &md5sig->head);
 951         return 0;
 952 }
 953 EXPORT_SYMBOL(tcp_md5_do_add);
 954
 955 int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family)
 956 {
 957         struct tcp_md5sig_key *key;
 958
 959         key = tcp_md5_do_lookup(sk, addr, family);
 960         if (!key)
 961                 return -ENOENT;
 962         hlist_del_rcu(&key->node);
 963         atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
 964         kfree_rcu(key, rcu);
 965         return 0;
 966 }
 967 EXPORT_SYMBOL(tcp_md5_do_del);
 968
 969 static void tcp_clear_md5_list(struct sock *sk)
 970 {
 971         struct tcp_sock *tp = tcp_sk(sk);
 972         struct tcp_md5sig_key *key;
 973         struct hlist_node *n;
 974         struct tcp_md5sig_info *md5sig;
 975
 976         md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
 977
 978         hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
 979                 hlist_del_rcu(&key->node);
 980                 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
 981                 kfree_rcu(key, rcu);
 982         }
 983 }
 984
 985 static int tcp_v4_parse_md5_keys(struct sock *sk, char __user *optval,
 986                                  int optlen)
 987 {
 988         struct tcp_md5sig cmd;
 989         struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
 990
 991         if (optlen < sizeof(cmd))
 992                 return -EINVAL;
 993
 994         if (copy_from_user(&cmd, optval, sizeof(cmd)))
 995                 return -EFAULT;
 996
 997         if (sin->sin_family != AF_INET)
 998                 return -EINVAL;
 999
1000         if (!cmd.tcpm_keylen)
1001                 return tcp_md5_do_del(sk, (union tcp_md5_addr *)&sin->sin_addr.s_addr,
1002                                       AF_INET);
1003
1004         if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1005                 return -EINVAL;
1006
1007         return tcp_md5_do_add(sk, (union tcp_md5_addr *)&sin->sin_addr.s_addr,
1008                               AF_INET, cmd.tcpm_key, cmd.tcpm_keylen,
1009                               GFP_KERNEL);
1010 }
1011
1012 static int tcp_v4_md5_hash_pseudoheader(struct tcp_md5sig_pool *hp,
1013                                         __be32 daddr, __be32 saddr, int nbytes)
1014 {
1015         struct tcp4_pseudohdr *bp;
1016         struct scatterlist sg;
1017
1018         bp = &hp->md5_blk.ip4;
1019
1020         /*
1021          * 1. the TCP pseudo-header (in the order: source IP address,
1022          * destination IP address, zero-padded protocol number, and
1023          * segment length)
1024          */
1025         bp->saddr = saddr;
1026         bp->daddr = daddr;
1027         bp->pad = 0;
1028         bp->protocol = IPPROTO_TCP;
1029         bp->len = cpu_to_be16(nbytes);
1030
1031         sg_init_one(&sg, bp, sizeof(*bp));
1032         return crypto_hash_update(&hp->md5_desc, &sg, sizeof(*bp));
1033 }
1034
1035 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1036                                __be32 daddr, __be32 saddr, const struct tcphdr *th)
1037 {
1038         struct tcp_md5sig_pool *hp;
1039         struct hash_desc *desc;
1040
1041         hp = tcp_get_md5sig_pool();
1042         if (!hp)
1043                 goto clear_hash_noput;
1044         desc = &hp->md5_desc;
1045
1046         if (crypto_hash_init(desc))
1047                 goto clear_hash;
1048         if (tcp_v4_md5_hash_pseudoheader(hp, daddr, saddr, th->doff << 2))
1049                 goto clear_hash;
1050         if (tcp_md5_hash_header(hp, th))
1051                 goto clear_hash;
1052         if (tcp_md5_hash_key(hp, key))
1053                 goto clear_hash;
1054         if (crypto_hash_final(desc, md5_hash))
1055                 goto clear_hash;
1056
1057         tcp_put_md5sig_pool();
1058         return 0;
1059
1060 clear_hash:
1061         tcp_put_md5sig_pool();
1062 clear_hash_noput:
1063         memset(md5_hash, 0, 16);
1064         return 1;
1065 }
1066
1067 int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1068                         const struct sock *sk,
1069                         const struct sk_buff *skb)
1070 {
1071         struct tcp_md5sig_pool *hp;
1072         struct hash_desc *desc;
1073         const struct tcphdr *th = tcp_hdr(skb);
1074         __be32 saddr, daddr;
1075
1076         if (sk) { /* valid for establish/request sockets */
1077                 saddr = sk->sk_rcv_saddr;
1078                 daddr = sk->sk_daddr;
1079         } else {
1080                 const struct iphdr *iph = ip_hdr(skb);
1081                 saddr = iph->saddr;
1082                 daddr = iph->daddr;
1083         }
1084
1085         hp = tcp_get_md5sig_pool();
1086         if (!hp)
1087                 goto clear_hash_noput;
1088         desc = &hp->md5_desc;
1089
1090         if (crypto_hash_init(desc))
1091                 goto clear_hash;
1092
1093         if (tcp_v4_md5_hash_pseudoheader(hp, daddr, saddr, skb->len))
1094                 goto clear_hash;
1095         if (tcp_md5_hash_header(hp, th))
1096                 goto clear_hash;
1097         if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
1098                 goto clear_hash;
1099         if (tcp_md5_hash_key(hp, key))
1100                 goto clear_hash;
1101         if (crypto_hash_final(desc, md5_hash))
1102                 goto clear_hash;
1103
1104         tcp_put_md5sig_pool();
1105         return 0;
1106
1107 clear_hash:
1108         tcp_put_md5sig_pool();
1109 clear_hash_noput:
1110         memset(md5_hash, 0, 16);
1111         return 1;
1112 }
1113 EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1114
1115 #endif
1116
1117 /* Called with rcu_read_lock() */
1118 static bool tcp_v4_inbound_md5_hash(const struct sock *sk,
1119                                     const struct sk_buff *skb)
1120 {
1121 #ifdef CONFIG_TCP_MD5SIG
1122         /*
1123          * This gets called for each TCP segment that arrives
1124          * so we want to be efficient.
1125          * We have 3 drop cases:
1126          * o No MD5 hash and one expected.
1127          * o MD5 hash and we're not expecting one.
1128          * o MD5 hash and its wrong.
1129          */
1130         const __u8 *hash_location = NULL;
1131         struct tcp_md5sig_key *hash_expected;
1132         const struct iphdr *iph = ip_hdr(skb);
1133         const struct tcphdr *th = tcp_hdr(skb);
1134         int genhash;
1135         unsigned char newhash[16];
1136
1137         hash_expected = tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&iph->saddr,
1138                                           AF_INET);
1139         hash_location = tcp_parse_md5sig_option(th);
1140
1141         /* We've parsed the options - do we have a hash? */
1142         if (!hash_expected && !hash_location)
1143                 return false;
1144
1145         if (hash_expected && !hash_location) {
1146                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
1147                 return true;
1148         }
1149
1150         if (!hash_expected && hash_location) {
1151                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
1152                 return true;
1153         }
1154
1155         /* Okay, so this is hash_expected and hash_location -
1156          * so we need to calculate the checksum.
1157          */
1158         genhash = tcp_v4_md5_hash_skb(newhash,
1159                                       hash_expected,
1160                                       NULL, skb);
1161
1162         if (genhash || memcmp(hash_location, newhash, 16) != 0) {
1163                 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s\n",
1164                                      &iph->saddr, ntohs(th->source),
1165                                      &iph->daddr, ntohs(th->dest),
1166                                      genhash ? " tcp_v4_calc_md5_hash failed"
1167                                      : "");
1168                 return true;
1169         }
1170         return false;
1171 #endif
1172         return false;
1173 }
1174
1175 static void tcp_v4_init_req(struct request_sock *req,
1176                             const struct sock *sk_listener,
1177                             struct sk_buff *skb)
1178 {
1179         struct inet_request_sock *ireq = inet_rsk(req);
1180
1181         sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1182         sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1183         ireq->no_srccheck = inet_sk(sk_listener)->transparent;
1184         ireq->opt = tcp_v4_save_options(skb);
1185 }
1186
1187 static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1188                                           struct flowi *fl,
1189                                           const struct request_sock *req,
1190                                           bool *strict)
1191 {
1192         struct dst_entry *dst = inet_csk_route_req(sk, &fl->u.ip4, req);
1193
1194         if (strict) {
1195                 if (fl->u.ip4.daddr == inet_rsk(req)->ir_rmt_addr)
1196                         *strict = true;
1197                 else
1198                         *strict = false;
1199         }
1200
1201         return dst;
1202 }
1203
1204 struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1205         .family         =       PF_INET,
1206         .obj_size       =       sizeof(struct tcp_request_sock),
1207         .rtx_syn_ack    =       tcp_rtx_synack,
1208         .send_ack       =       tcp_v4_reqsk_send_ack,
1209         .destructor     =       tcp_v4_reqsk_destructor,
1210         .send_reset     =       tcp_v4_send_reset,
1211         .syn_ack_timeout =      tcp_syn_ack_timeout,
1212 };
1213
1214 static const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1215         .mss_clamp      =       TCP_MSS_DEFAULT,
1216 #ifdef CONFIG_TCP_MD5SIG
1217         .req_md5_lookup =       tcp_v4_md5_lookup,
1218         .calc_md5_hash  =       tcp_v4_md5_hash_skb,
1219 #endif
1220         .init_req       =       tcp_v4_init_req,
1221 #ifdef CONFIG_SYN_COOKIES
1222         .cookie_init_seq =      cookie_v4_init_sequence,
1223 #endif
1224         .route_req      =       tcp_v4_route_req,
1225         .init_seq       =       tcp_v4_init_sequence,
1226         .send_synack    =       tcp_v4_send_synack,
1227         .queue_hash_add =       inet_csk_reqsk_queue_hash_add,
1228 };
1229
1230 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1231 {
1232         /* Never answer to SYNs send to broadcast or multicast */
1233         if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1234                 goto drop;
1235
1236         return tcp_conn_request(&tcp_request_sock_ops,
1237                                 &tcp_request_sock_ipv4_ops, sk, skb);
1238
1239 drop:
1240         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
1241         return 0;
1242 }
1243 EXPORT_SYMBOL(tcp_v4_conn_request);
1244
1245
1246 /*
1247  * The three way handshake has completed - we got a valid synack -
1248  * now create the new socket.
1249  */
1250 struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1251                                   struct request_sock *req,
1252                                   struct dst_entry *dst)
1253 {
1254         struct inet_request_sock *ireq;
1255         struct inet_sock *newinet;
1256         struct tcp_sock *newtp;
1257         struct sock *newsk;
1258 #ifdef CONFIG_TCP_MD5SIG
1259         struct tcp_md5sig_key *key;
1260 #endif
1261         struct ip_options_rcu *inet_opt;
1262
1263         if (sk_acceptq_is_full(sk))
1264                 goto exit_overflow;
1265
1266         newsk = tcp_create_openreq_child(sk, req, skb);
1267         if (!newsk)
1268                 goto exit_nonewsk;
1269
1270         newsk->sk_gso_type = SKB_GSO_TCPV4;
1271         inet_sk_rx_dst_set(newsk, skb);
1272
1273         newtp                 = tcp_sk(newsk);
1274         newinet               = inet_sk(newsk);
1275         ireq                  = inet_rsk(req);
1276         sk_daddr_set(newsk, ireq->ir_rmt_addr);
1277         sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1278         newinet->inet_saddr           = ireq->ir_loc_addr;
1279         inet_opt              = ireq->opt;
1280         rcu_assign_pointer(newinet->inet_opt, inet_opt);
1281         ireq->opt             = NULL;
1282         newinet->mc_index     = inet_iif(skb);
1283         newinet->mc_ttl       = ip_hdr(skb)->ttl;
1284         newinet->rcv_tos      = ip_hdr(skb)->tos;
1285         inet_csk(newsk)->icsk_ext_hdr_len = 0;
1286         if (inet_opt)
1287                 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1288         newinet->inet_id = newtp->write_seq ^ jiffies;
1289
1290         if (!dst) {
1291                 dst = inet_csk_route_child_sock(sk, newsk, req);
1292                 if (!dst)
1293                         goto put_and_exit;
1294         } else {
1295                 /* syncookie case : see end of cookie_v4_check() */
1296         }
1297         sk_setup_caps(newsk, dst);
1298
1299         tcp_ca_openreq_child(newsk, dst);
1300
1301         tcp_sync_mss(newsk, dst_mtu(dst));
1302         newtp->advmss = dst_metric_advmss(dst);
1303         if (tcp_sk(sk)->rx_opt.user_mss &&
1304             tcp_sk(sk)->rx_opt.user_mss < newtp->advmss)
1305                 newtp->advmss = tcp_sk(sk)->rx_opt.user_mss;
1306
1307         tcp_initialize_rcv_mss(newsk);
1308
1309 #ifdef CONFIG_TCP_MD5SIG
1310         /* Copy over the MD5 key from the original socket */
1311         key = tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&newinet->inet_daddr,
1312                                 AF_INET);
1313         if (key) {
1314                 /*
1315                  * We're using one, so create a matching key
1316                  * on the newsk structure. If we fail to get
1317                  * memory, then we end up not copying the key
1318                  * across. Shucks.
1319                  */
1320                 tcp_md5_do_add(newsk, (union tcp_md5_addr *)&newinet->inet_daddr,
1321                                AF_INET, key->key, key->keylen, GFP_ATOMIC);
1322                 sk_nocaps_add(newsk, NETIF_F_GSO_MASK);
1323         }
1324 #endif
1325
1326         if (__inet_inherit_port(sk, newsk) < 0)
1327                 goto put_and_exit;
1328         __inet_hash_nolisten(newsk, NULL);
1329
1330         return newsk;
1331
1332 exit_overflow:
1333         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1334 exit_nonewsk:
1335         dst_release(dst);
1336 exit:
1337         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
1338         return NULL;
1339 put_and_exit:
1340         inet_csk_prepare_forced_close(newsk);
1341         tcp_done(newsk);
1342         goto exit;
1343 }
1344 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1345
1346 static struct sock *tcp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
1347 {
1348         const struct tcphdr *th = tcp_hdr(skb);
1349         const struct iphdr *iph = ip_hdr(skb);
1350         struct request_sock *req;
1351         struct sock *nsk;
1352
1353         req = inet_csk_search_req(sk, th->source, iph->saddr, iph->daddr);
1354         if (req) {
1355                 nsk = tcp_check_req(sk, skb, req, false);
1356                 if (!nsk || nsk == sk)
1357                         reqsk_put(req);
1358                 return nsk;
1359         }
1360
1361         nsk = inet_lookup_established(sock_net(sk), &tcp_hashinfo, iph->saddr,
1362                         th->source, iph->daddr, th->dest, inet_iif(skb));
1363
1364         if (nsk) {
1365                 if (nsk->sk_state != TCP_TIME_WAIT) {
1366                         bh_lock_sock(nsk);
1367                         return nsk;
1368                 }
1369                 inet_twsk_put(inet_twsk(nsk));
1370                 return NULL;
1371         }
1372
1373 #ifdef CONFIG_SYN_COOKIES
1374         if (!th->syn)
1375                 sk = cookie_v4_check(sk, skb);
1376 #endif
1377         return sk;
1378 }
1379
1380 /* The socket must have it's spinlock held when we get
1381  * here.
1382  *
1383  * We have a potential double-lock case here, so even when
1384  * doing backlog processing we use the BH locking scheme.
1385  * This is because we cannot sleep with the original spinlock
1386  * held.
1387  */
1388 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1389 {
1390         struct sock *rsk;
1391
1392         if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1393                 struct dst_entry *dst = sk->sk_rx_dst;
1394
1395                 sock_rps_save_rxhash(sk, skb);
1396                 sk_mark_napi_id(sk, skb);
1397                 if (dst) {
1398                         if (inet_sk(sk)->rx_dst_ifindex != skb->skb_iif ||
1399                             !dst->ops->check(dst, 0)) {
1400                                 dst_release(dst);
1401                                 sk->sk_rx_dst = NULL;
1402                         }
1403                 }
1404                 tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len);
1405                 return 0;
1406         }
1407
1408         if (tcp_checksum_complete(skb))
1409                 goto csum_err;
1410
1411         if (sk->sk_state == TCP_LISTEN) {
1412                 struct sock *nsk = tcp_v4_hnd_req(sk, skb);
1413                 if (!nsk)
1414                         goto discard;
1415
1416                 if (nsk != sk) {
1417                         sock_rps_save_rxhash(nsk, skb);
1418                         sk_mark_napi_id(nsk, skb);
1419                         if (tcp_child_process(sk, nsk, skb)) {
1420                                 rsk = nsk;
1421                                 goto reset;
1422                         }
1423                         return 0;
1424                 }
1425         } else
1426                 sock_rps_save_rxhash(sk, skb);
1427
1428         if (tcp_rcv_state_process(sk, skb)) {
1429                 rsk = sk;
1430                 goto reset;
1431         }
1432         return 0;
1433
1434 reset:
1435         tcp_v4_send_reset(rsk, skb);
1436 discard:
1437         kfree_skb(skb);
1438         /* Be careful here. If this function gets more complicated and
1439          * gcc suffers from register pressure on the x86, sk (in %ebx)
1440          * might be destroyed here. This current version compiles correctly,
1441          * but you have been warned.
1442          */
1443         return 0;
1444
1445 csum_err:
1446         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
1447         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
1448         goto discard;
1449 }
1450 EXPORT_SYMBOL(tcp_v4_do_rcv);
1451
1452 void tcp_v4_early_demux(struct sk_buff *skb)
1453 {
1454         const struct iphdr *iph;
1455         const struct tcphdr *th;
1456         struct sock *sk;
1457
1458         if (skb->pkt_type != PACKET_HOST)
1459                 return;
1460
1461         if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1462                 return;
1463
1464         iph = ip_hdr(skb);
1465         th = tcp_hdr(skb);
1466
1467         if (th->doff < sizeof(struct tcphdr) / 4)
1468                 return;
1469
1470         sk = __inet_lookup_established(dev_net(skb->dev), &tcp_hashinfo,
1471                                        iph->saddr, th->source,
1472                                        iph->daddr, ntohs(th->dest),
1473                                        skb->skb_iif);
1474         if (sk) {
1475                 skb->sk = sk;
1476                 skb->destructor = sock_edemux;
1477                 if (sk_fullsock(sk)) {
1478                         struct dst_entry *dst = READ_ONCE(sk->sk_rx_dst);
1479
1480                         if (dst)
1481                                 dst = dst_check(dst, 0);
1482                         if (dst &&
1483                             inet_sk(sk)->rx_dst_ifindex == skb->skb_iif)
1484                                 skb_dst_set_noref(skb, dst);
1485                 }
1486         }
1487 }
1488
1489 /* Packet is added to VJ-style prequeue for processing in process
1490  * context, if a reader task is waiting. Apparently, this exciting
1491  * idea (VJ's mail "Re: query about TCP header on tcp-ip" of 07 Sep 93)
1492  * failed somewhere. Latency? Burstiness? Well, at least now we will
1493  * see, why it failed. 8)8)                               --ANK
1494  *
1495  */
1496 bool tcp_prequeue(struct sock *sk, struct sk_buff *skb)
1497 {
1498         struct tcp_sock *tp = tcp_sk(sk);
1499
1500         if (sysctl_tcp_low_latency || !tp->ucopy.task)
1501                 return false;
1502
1503         if (skb->len <= tcp_hdrlen(skb) &&
1504             skb_queue_len(&tp->ucopy.prequeue) == 0)
1505                 return false;
1506
1507         /* Before escaping RCU protected region, we need to take care of skb
1508          * dst. Prequeue is only enabled for established sockets.
1509          * For such sockets, we might need the skb dst only to set sk->sk_rx_dst
1510          * Instead of doing full sk_rx_dst validity here, let's perform
1511          * an optimistic check.
1512          */
1513         if (likely(sk->sk_rx_dst))
1514                 skb_dst_drop(skb);
1515         else
1516                 skb_dst_force(skb);
1517
1518         __skb_queue_tail(&tp->ucopy.prequeue, skb);
1519         tp->ucopy.memory += skb->truesize;
1520         if (tp->ucopy.memory > sk->sk_rcvbuf) {
1521                 struct sk_buff *skb1;
1522
1523                 BUG_ON(sock_owned_by_user(sk));
1524
1525                 while ((skb1 = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) {
1526                         sk_backlog_rcv(sk, skb1);
1527                         NET_INC_STATS_BH(sock_net(sk),
1528                                          LINUX_MIB_TCPPREQUEUEDROPPED);
1529                 }
1530
1531                 tp->ucopy.memory = 0;
1532         } else if (skb_queue_len(&tp->ucopy.prequeue) == 1) {
1533                 wake_up_interruptible_sync_poll(sk_sleep(sk),
1534                                            POLLIN | POLLRDNORM | POLLRDBAND);
1535                 if (!inet_csk_ack_scheduled(sk))
1536                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
1537                                                   (3 * tcp_rto_min(sk)) / 4,
1538                                                   TCP_RTO_MAX);
1539         }
1540         return true;
1541 }
1542 EXPORT_SYMBOL(tcp_prequeue);
1543
1544 /*
1545  *      From tcp_input.c
1546  */
1547
1548 int tcp_v4_rcv(struct sk_buff *skb)
1549 {
1550         const struct iphdr *iph;
1551         const struct tcphdr *th;
1552         struct sock *sk;
1553         int ret;
1554         struct net *net = dev_net(skb->dev);
1555
1556         if (skb->pkt_type != PACKET_HOST)
1557                 goto discard_it;
1558
1559         /* Count it even if it's bad */
1560         TCP_INC_STATS_BH(net, TCP_MIB_INSEGS);
1561
1562         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1563                 goto discard_it;
1564
1565         th = tcp_hdr(skb);
1566
1567         if (th->doff < sizeof(struct tcphdr) / 4)
1568                 goto bad_packet;
1569         if (!pskb_may_pull(skb, th->doff * 4))
1570                 goto discard_it;
1571
1572         /* An explanation is required here, I think.
1573          * Packet length and doff are validated by header prediction,
1574          * provided case of th->doff==0 is eliminated.
1575          * So, we defer the checks. */
1576
1577         if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
1578                 goto csum_error;
1579
1580         th = tcp_hdr(skb);
1581         iph = ip_hdr(skb);
1582         /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
1583          * barrier() makes sure compiler wont play fool^Waliasing games.
1584          */
1585         memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
1586                 sizeof(struct inet_skb_parm));
1587         barrier();
1588
1589         TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1590         TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1591                                     skb->len - th->doff * 4);
1592         TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1593         TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
1594         TCP_SKB_CB(skb)->tcp_tw_isn = 0;
1595         TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
1596         TCP_SKB_CB(skb)->sacked  = 0;
1597
1598         sk = __inet_lookup_skb(&tcp_hashinfo, skb, th->source, th->dest);
1599         if (!sk)
1600                 goto no_tcp_socket;
1601
1602 process:
1603         if (sk->sk_state == TCP_TIME_WAIT)
1604                 goto do_time_wait;
1605
1606         if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
1607                 NET_INC_STATS_BH(net, LINUX_MIB_TCPMINTTLDROP);
1608                 goto discard_and_relse;
1609         }
1610
1611         if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1612                 goto discard_and_relse;
1613
1614         if (tcp_v4_inbound_md5_hash(sk, skb))
1615                 goto discard_and_relse;
1616
1617         nf_reset(skb);
1618
1619         if (sk_filter(sk, skb))
1620                 goto discard_and_relse;
1621
1622         sk_incoming_cpu_update(sk);
1623         skb->dev = NULL;
1624
1625         bh_lock_sock_nested(sk);
1626         tcp_sk(sk)->segs_in += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
1627         ret = 0;
1628         if (!sock_owned_by_user(sk)) {
1629                 if (!tcp_prequeue(sk, skb))
1630                         ret = tcp_v4_do_rcv(sk, skb);
1631         } else if (unlikely(sk_add_backlog(sk, skb,
1632                                            sk->sk_rcvbuf + sk->sk_sndbuf))) {
1633                 bh_unlock_sock(sk);
1634                 NET_INC_STATS_BH(net, LINUX_MIB_TCPBACKLOGDROP);
1635                 goto discard_and_relse;
1636         }
1637         bh_unlock_sock(sk);
1638
1639         sock_put(sk);
1640
1641         return ret;
1642
1643 no_tcp_socket:
1644         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1645                 goto discard_it;
1646
1647         if (tcp_checksum_complete(skb)) {
1648 csum_error:
1649                 TCP_INC_STATS_BH(net, TCP_MIB_CSUMERRORS);
1650 bad_packet:
1651                 TCP_INC_STATS_BH(net, TCP_MIB_INERRS);
1652         } else {
1653                 tcp_v4_send_reset(NULL, skb);
1654         }
1655
1656 discard_it:
1657         /* Discard frame. */
1658         kfree_skb(skb);
1659         return 0;
1660
1661 discard_and_relse:
1662         sock_put(sk);
1663         goto discard_it;
1664
1665 do_time_wait:
1666         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1667                 inet_twsk_put(inet_twsk(sk));
1668                 goto discard_it;
1669         }
1670
1671         if (tcp_checksum_complete(skb)) {
1672                 inet_twsk_put(inet_twsk(sk));
1673                 goto csum_error;
1674         }
1675         switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
1676         case TCP_TW_SYN: {
1677                 struct sock *sk2 = inet_lookup_listener(dev_net(skb->dev),
1678                                                         &tcp_hashinfo,
1679                                                         iph->saddr, th->source,
1680                                                         iph->daddr, th->dest,
1681                                                         inet_iif(skb));
1682                 if (sk2) {
1683                         inet_twsk_deschedule_put(inet_twsk(sk));
1684                         sk = sk2;
1685                         goto process;
1686                 }
1687                 /* Fall through to ACK */
1688         }
1689         case TCP_TW_ACK:
1690                 tcp_v4_timewait_ack(sk, skb);
1691                 break;
1692         case TCP_TW_RST:
1693                 goto no_tcp_socket;
1694         case TCP_TW_SUCCESS:;
1695         }
1696         goto discard_it;
1697 }
1698
1699 static struct timewait_sock_ops tcp_timewait_sock_ops = {
1700         .twsk_obj_size  = sizeof(struct tcp_timewait_sock),
1701         .twsk_unique    = tcp_twsk_unique,
1702         .twsk_destructor= tcp_twsk_destructor,
1703 };
1704
1705 void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
1706 {
1707         struct dst_entry *dst = skb_dst(skb);
1708
1709         if (dst) {
1710                 dst_hold(dst);
1711                 sk->sk_rx_dst = dst;
1712                 inet_sk(sk)->rx_dst_ifindex = skb->skb_iif;
1713         }
1714 }
1715 EXPORT_SYMBOL(inet_sk_rx_dst_set);
1716
1717 const struct inet_connection_sock_af_ops ipv4_specific = {
1718         .queue_xmit        = ip_queue_xmit,
1719         .send_check        = tcp_v4_send_check,
1720         .rebuild_header    = inet_sk_rebuild_header,
1721         .sk_rx_dst_set     = inet_sk_rx_dst_set,
1722         .conn_request      = tcp_v4_conn_request,
1723         .syn_recv_sock     = tcp_v4_syn_recv_sock,
1724         .net_header_len    = sizeof(struct iphdr),
1725         .setsockopt        = ip_setsockopt,
1726         .getsockopt        = ip_getsockopt,
1727         .addr2sockaddr     = inet_csk_addr2sockaddr,
1728         .sockaddr_len      = sizeof(struct sockaddr_in),
1729         .bind_conflict     = inet_csk_bind_conflict,
1730 #ifdef CONFIG_COMPAT
1731         .compat_setsockopt = compat_ip_setsockopt,
1732         .compat_getsockopt = compat_ip_getsockopt,
1733 #endif
1734         .mtu_reduced       = tcp_v4_mtu_reduced,
1735 };
1736 EXPORT_SYMBOL(ipv4_specific);
1737
1738 #ifdef CONFIG_TCP_MD5SIG
1739 static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
1740         .md5_lookup             = tcp_v4_md5_lookup,
1741         .calc_md5_hash          = tcp_v4_md5_hash_skb,
1742         .md5_parse              = tcp_v4_parse_md5_keys,
1743 };
1744 #endif
1745
1746 /* NOTE: A lot of things set to zero explicitly by call to
1747  *       sk_alloc() so need not be done here.
1748  */
1749 static int tcp_v4_init_sock(struct sock *sk)
1750 {
1751         struct inet_connection_sock *icsk = inet_csk(sk);
1752
1753         tcp_init_sock(sk);
1754
1755         icsk->icsk_af_ops = &ipv4_specific;
1756
1757 #ifdef CONFIG_TCP_MD5SIG
1758         tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
1759 #endif
1760
1761         return 0;
1762 }
1763
1764 void tcp_v4_destroy_sock(struct sock *sk)
1765 {
1766         struct tcp_sock *tp = tcp_sk(sk);
1767
1768         tcp_clear_xmit_timers(sk);
1769
1770         tcp_cleanup_congestion_control(sk);
1771
1772         /* Cleanup up the write buffer. */
1773         tcp_write_queue_purge(sk);
1774
1775         /* Cleans up our, hopefully empty, out_of_order_queue. */
1776         __skb_queue_purge(&tp->out_of_order_queue);
1777
1778 #ifdef CONFIG_TCP_MD5SIG
1779         /* Clean up the MD5 key list, if any */
1780         if (tp->md5sig_info) {
1781                 tcp_clear_md5_list(sk);
1782                 kfree_rcu(tp->md5sig_info, rcu);
1783                 tp->md5sig_info = NULL;
1784         }
1785 #endif
1786
1787         /* Clean prequeue, it must be empty really */
1788         __skb_queue_purge(&tp->ucopy.prequeue);
1789
1790         /* Clean up a referenced TCP bind bucket. */
1791         if (inet_csk(sk)->icsk_bind_hash)
1792                 inet_put_port(sk);
1793
1794         BUG_ON(tp->fastopen_rsk);
1795
1796         /* If socket is aborted during connect operation */
1797         tcp_free_fastopen_req(tp);
1798         tcp_saved_syn_free(tp);
1799
1800         sk_sockets_allocated_dec(sk);
1801         sock_release_memcg(sk);
1802 }
1803 EXPORT_SYMBOL(tcp_v4_destroy_sock);
1804
1805 #ifdef CONFIG_PROC_FS
1806 /* Proc filesystem TCP sock list dumping. */
1807
1808 /*
1809  * Get next listener socket follow cur.  If cur is NULL, get first socket
1810  * starting from bucket given in st->bucket; when st->bucket is zero the
1811  * very first socket in the hash table is returned.
1812  */
1813 static void *listening_get_next(struct seq_file *seq, void *cur)
1814 {
1815         struct inet_connection_sock *icsk;
1816         struct hlist_nulls_node *node;
1817         struct sock *sk = cur;
1818         struct inet_listen_hashbucket *ilb;
1819         struct tcp_iter_state *st = seq->private;
1820         struct net *net = seq_file_net(seq);
1821
1822         if (!sk) {
1823                 ilb = &tcp_hashinfo.listening_hash[st->bucket];
1824                 spin_lock_bh(&ilb->lock);
1825                 sk = sk_nulls_head(&ilb->head);
1826                 st->offset = 0;
1827                 goto get_sk;
1828         }
1829         ilb = &tcp_hashinfo.listening_hash[st->bucket];
1830         ++st->num;
1831         ++st->offset;
1832
1833         if (st->state == TCP_SEQ_STATE_OPENREQ) {
1834                 struct request_sock *req = cur;
1835
1836                 icsk = inet_csk(st->syn_wait_sk);
1837                 req = req->dl_next;
1838                 while (1) {
1839                         while (req) {
1840                                 if (req->rsk_ops->family == st->family) {
1841                                         cur = req;
1842                                         goto out;
1843                                 }
1844                                 req = req->dl_next;
1845                         }
1846                         if (++st->sbucket >= icsk->icsk_accept_queue.listen_opt->nr_table_entries)
1847                                 break;
1848 get_req:
1849                         req = icsk->icsk_accept_queue.listen_opt->syn_table[st->sbucket];
1850                 }
1851                 sk        = sk_nulls_next(st->syn_wait_sk);
1852                 st->state = TCP_SEQ_STATE_LISTENING;
1853                 spin_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1854         } else {
1855                 icsk = inet_csk(sk);
1856                 spin_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1857                 if (reqsk_queue_len(&icsk->icsk_accept_queue))
1858                         goto start_req;
1859                 spin_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1860                 sk = sk_nulls_next(sk);
1861         }
1862 get_sk:
1863         sk_nulls_for_each_from(sk, node) {
1864                 if (!net_eq(sock_net(sk), net))
1865                         continue;
1866                 if (sk->sk_family == st->family) {
1867                         cur = sk;
1868                         goto out;
1869                 }
1870                 icsk = inet_csk(sk);
1871                 spin_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1872                 if (reqsk_queue_len(&icsk->icsk_accept_queue)) {
1873 start_req:
1874                         st->syn_wait_sk = sk;
1875                         st->state       = TCP_SEQ_STATE_OPENREQ;
1876                         st->sbucket     = 0;
1877                         goto get_req;
1878                 }
1879                 spin_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
1880         }
1881         spin_unlock_bh(&ilb->lock);
1882         st->offset = 0;
1883         if (++st->bucket < INET_LHTABLE_SIZE) {
1884                 ilb = &tcp_hashinfo.listening_hash[st->bucket];
1885                 spin_lock_bh(&ilb->lock);
1886                 sk = sk_nulls_head(&ilb->head);
1887                 goto get_sk;
1888         }
1889         cur = NULL;
1890 out:
1891         return cur;
1892 }
1893
1894 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
1895 {
1896         struct tcp_iter_state *st = seq->private;
1897         void *rc;
1898
1899         st->bucket = 0;
1900         st->offset = 0;
1901         rc = listening_get_next(seq, NULL);
1902
1903         while (rc && *pos) {
1904                 rc = listening_get_next(seq, rc);
1905                 --*pos;
1906         }
1907         return rc;
1908 }
1909
1910 static inline bool empty_bucket(const struct tcp_iter_state *st)
1911 {
1912         return hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].chain);
1913 }
1914
1915 /*
1916  * Get first established socket starting from bucket given in st->bucket.
1917  * If st->bucket is zero, the very first socket in the hash is returned.
1918  */
1919 static void *established_get_first(struct seq_file *seq)
1920 {
1921         struct tcp_iter_state *st = seq->private;
1922         struct net *net = seq_file_net(seq);
1923         void *rc = NULL;
1924
1925         st->offset = 0;
1926         for (; st->bucket <= tcp_hashinfo.ehash_mask; ++st->bucket) {
1927                 struct sock *sk;
1928                 struct hlist_nulls_node *node;
1929                 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, st->bucket);
1930
1931                 /* Lockless fast path for the common case of empty buckets */
1932                 if (empty_bucket(st))
1933                         continue;
1934
1935                 spin_lock_bh(lock);
1936                 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
1937                         if (sk->sk_family != st->family ||
1938                             !net_eq(sock_net(sk), net)) {
1939                                 continue;
1940                         }
1941                         rc = sk;
1942                         goto out;
1943                 }
1944                 spin_unlock_bh(lock);
1945         }
1946 out:
1947         return rc;
1948 }
1949
1950 static void *established_get_next(struct seq_file *seq, void *cur)
1951 {
1952         struct sock *sk = cur;
1953         struct hlist_nulls_node *node;
1954         struct tcp_iter_state *st = seq->private;
1955         struct net *net = seq_file_net(seq);
1956
1957         ++st->num;
1958         ++st->offset;
1959
1960         sk = sk_nulls_next(sk);
1961
1962         sk_nulls_for_each_from(sk, node) {
1963                 if (sk->sk_family == st->family && net_eq(sock_net(sk), net))
1964                         return sk;
1965         }
1966
1967         spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
1968         ++st->bucket;
1969         return established_get_first(seq);
1970 }
1971
1972 static void *established_get_idx(struct seq_file *seq, loff_t pos)
1973 {
1974         struct tcp_iter_state *st = seq->private;
1975         void *rc;
1976
1977         st->bucket = 0;
1978         rc = established_get_first(seq);
1979
1980         while (rc && pos) {
1981                 rc = established_get_next(seq, rc);
1982                 --pos;
1983         }
1984         return rc;
1985 }
1986
1987 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
1988 {
1989         void *rc;
1990         struct tcp_iter_state *st = seq->private;
1991
1992         st->state = TCP_SEQ_STATE_LISTENING;
1993         rc        = listening_get_idx(seq, &pos);
1994
1995         if (!rc) {
1996                 st->state = TCP_SEQ_STATE_ESTABLISHED;
1997                 rc        = established_get_idx(seq, pos);
1998         }
1999
2000         return rc;
2001 }
2002
2003 static void *tcp_seek_last_pos(struct seq_file *seq)
2004 {
2005         struct tcp_iter_state *st = seq->private;
2006         int offset = st->offset;
2007         int orig_num = st->num;
2008         void *rc = NULL;
2009
2010         switch (st->state) {
2011         case TCP_SEQ_STATE_OPENREQ:
2012         case TCP_SEQ_STATE_LISTENING:
2013                 if (st->bucket >= INET_LHTABLE_SIZE)
2014                         break;
2015                 st->state = TCP_SEQ_STATE_LISTENING;
2016                 rc = listening_get_next(seq, NULL);
2017                 while (offset-- && rc)
2018                         rc = listening_get_next(seq, rc);
2019                 if (rc)
2020                         break;
2021                 st->bucket = 0;
2022                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2023                 /* Fallthrough */
2024         case TCP_SEQ_STATE_ESTABLISHED:
2025                 if (st->bucket > tcp_hashinfo.ehash_mask)
2026                         break;
2027                 rc = established_get_first(seq);
2028                 while (offset-- && rc)
2029                         rc = established_get_next(seq, rc);
2030         }
2031
2032         st->num = orig_num;
2033
2034         return rc;
2035 }
2036
2037 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2038 {
2039         struct tcp_iter_state *st = seq->private;
2040         void *rc;
2041
2042         if (*pos && *pos == st->last_pos) {
2043                 rc = tcp_seek_last_pos(seq);
2044                 if (rc)
2045                         goto out;
2046         }
2047
2048         st->state = TCP_SEQ_STATE_LISTENING;
2049         st->num = 0;
2050         st->bucket = 0;
2051         st->offset = 0;
2052         rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2053
2054 out:
2055         st->last_pos = *pos;
2056         return rc;
2057 }
2058
2059 static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2060 {
2061         struct tcp_iter_state *st = seq->private;
2062         void *rc = NULL;
2063
2064         if (v == SEQ_START_TOKEN) {
2065                 rc = tcp_get_idx(seq, 0);
2066                 goto out;
2067         }
2068
2069         switch (st->state) {
2070         case TCP_SEQ_STATE_OPENREQ:
2071         case TCP_SEQ_STATE_LISTENING:
2072                 rc = listening_get_next(seq, v);
2073                 if (!rc) {
2074                         st->state = TCP_SEQ_STATE_ESTABLISHED;
2075                         st->bucket = 0;
2076                         st->offset = 0;
2077                         rc        = established_get_first(seq);
2078                 }
2079                 break;
2080         case TCP_SEQ_STATE_ESTABLISHED:
2081                 rc = established_get_next(seq, v);
2082                 break;
2083         }
2084 out:
2085         ++*pos;
2086         st->last_pos = *pos;
2087         return rc;
2088 }
2089
2090 static void tcp_seq_stop(struct seq_file *seq, void *v)
2091 {
2092         struct tcp_iter_state *st = seq->private;
2093
2094         switch (st->state) {
2095         case TCP_SEQ_STATE_OPENREQ:
2096                 if (v) {
2097                         struct inet_connection_sock *icsk = inet_csk(st->syn_wait_sk);
2098                         spin_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
2099                 }
2100         case TCP_SEQ_STATE_LISTENING:
2101                 if (v != SEQ_START_TOKEN)
2102                         spin_unlock_bh(&tcp_hashinfo.listening_hash[st->bucket].lock);
2103                 break;
2104         case TCP_SEQ_STATE_ESTABLISHED:
2105                 if (v)
2106                         spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
2107                 break;
2108         }
2109 }
2110
2111 int tcp_seq_open(struct inode *inode, struct file *file)
2112 {
2113         struct tcp_seq_afinfo *afinfo = PDE_DATA(inode);
2114         struct tcp_iter_state *s;
2115         int err;
2116
2117         err = seq_open_net(inode, file, &afinfo->seq_ops,
2118                           sizeof(struct tcp_iter_state));
2119         if (err < 0)
2120                 return err;
2121
2122         s = ((struct seq_file *)file->private_data)->private;
2123         s->family               = afinfo->family;
2124         s->last_pos             = 0;
2125         return 0;
2126 }
2127 EXPORT_SYMBOL(tcp_seq_open);
2128
2129 int tcp_proc_register(struct net *net, struct tcp_seq_afinfo *afinfo)
2130 {
2131         int rc = 0;
2132         struct proc_dir_entry *p;
2133
2134         afinfo->seq_ops.start           = tcp_seq_start;
2135         afinfo->seq_ops.next            = tcp_seq_next;
2136         afinfo->seq_ops.stop            = tcp_seq_stop;
2137
2138         p = proc_create_data(afinfo->name, S_IRUGO, net->proc_net,
2139                              afinfo->seq_fops, afinfo);
2140         if (!p)
2141                 rc = -ENOMEM;
2142         return rc;
2143 }
2144 EXPORT_SYMBOL(tcp_proc_register);
2145
2146 void tcp_proc_unregister(struct net *net, struct tcp_seq_afinfo *afinfo)
2147 {
2148         remove_proc_entry(afinfo->name, net->proc_net);
2149 }
2150 EXPORT_SYMBOL(tcp_proc_unregister);
2151
2152 static void get_openreq4(const struct request_sock *req,
2153                          struct seq_file *f, int i)
2154 {
2155         const struct inet_request_sock *ireq = inet_rsk(req);
2156         long delta = req->rsk_timer.expires - jiffies;
2157
2158         seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2159                 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2160                 i,
2161                 ireq->ir_loc_addr,
2162                 ireq->ir_num,
2163                 ireq->ir_rmt_addr,
2164                 ntohs(ireq->ir_rmt_port),
2165                 TCP_SYN_RECV,
2166                 0, 0, /* could print option size, but that is af dependent. */
2167                 1,    /* timers active (only the expire timer) */
2168                 jiffies_delta_to_clock_t(delta),
2169                 req->num_timeout,
2170                 from_kuid_munged(seq_user_ns(f),
2171                                  sock_i_uid(req->rsk_listener)),
2172                 0,  /* non standard timer */
2173                 0, /* open_requests have no inode */
2174                 0,
2175                 req);
2176 }
2177
2178 static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2179 {
2180         int timer_active;
2181         unsigned long timer_expires;
2182         const struct tcp_sock *tp = tcp_sk(sk);
2183         const struct inet_connection_sock *icsk = inet_csk(sk);
2184         const struct inet_sock *inet = inet_sk(sk);
2185         const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2186         __be32 dest = inet->inet_daddr;
2187         __be32 src = inet->inet_rcv_saddr;
2188         __u16 destp = ntohs(inet->inet_dport);
2189         __u16 srcp = ntohs(inet->inet_sport);
2190         int rx_queue;
2191
2192         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2193             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2194             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2195                 timer_active    = 1;
2196                 timer_expires   = icsk->icsk_timeout;
2197         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2198                 timer_active    = 4;
2199                 timer_expires   = icsk->icsk_timeout;
2200         } else if (timer_pending(&sk->sk_timer)) {
2201                 timer_active    = 2;
2202                 timer_expires   = sk->sk_timer.expires;
2203         } else {
2204                 timer_active    = 0;
2205                 timer_expires = jiffies;
2206         }
2207
2208         if (sk->sk_state == TCP_LISTEN)
2209                 rx_queue = sk->sk_ack_backlog;
2210         else
2211                 /*
2212                  * because we dont lock socket, we might find a transient negative value
2213                  */
2214                 rx_queue = max_t(int, tp->rcv_nxt - tp->copied_seq, 0);
2215
2216         seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2217                         "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2218                 i, src, srcp, dest, destp, sk->sk_state,
2219                 tp->write_seq - tp->snd_una,
2220                 rx_queue,
2221                 timer_active,
2222                 jiffies_delta_to_clock_t(timer_expires - jiffies),
2223                 icsk->icsk_retransmits,
2224                 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2225                 icsk->icsk_probes_out,
2226                 sock_i_ino(sk),
2227                 atomic_read(&sk->sk_refcnt), sk,
2228                 jiffies_to_clock_t(icsk->icsk_rto),
2229                 jiffies_to_clock_t(icsk->icsk_ack.ato),
2230                 (icsk->icsk_ack.quick << 1) | icsk->icsk_ack.pingpong,
2231                 tp->snd_cwnd,
2232                 sk->sk_state == TCP_LISTEN ?
2233                     (fastopenq ? fastopenq->max_qlen : 0) :
2234                     (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2235 }
2236
2237 static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2238                                struct seq_file *f, int i)
2239 {
2240         long delta = tw->tw_timer.expires - jiffies;
2241         __be32 dest, src;
2242         __u16 destp, srcp;
2243
2244         dest  = tw->tw_daddr;
2245         src   = tw->tw_rcv_saddr;
2246         destp = ntohs(tw->tw_dport);
2247         srcp  = ntohs(tw->tw_sport);
2248
2249         seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2250                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2251                 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2252                 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2253                 atomic_read(&tw->tw_refcnt), tw);
2254 }
2255
2256 #define TMPSZ 150
2257
2258 static int tcp4_seq_show(struct seq_file *seq, void *v)
2259 {
2260         struct tcp_iter_state *st;
2261         struct sock *sk = v;
2262
2263         seq_setwidth(seq, TMPSZ - 1);
2264         if (v == SEQ_START_TOKEN) {
2265                 seq_puts(seq, "  sl  local_address rem_address   st tx_queue "
2266                            "rx_queue tr tm->when retrnsmt   uid  timeout "
2267                            "inode");
2268                 goto out;
2269         }
2270         st = seq->private;
2271
2272         switch (st->state) {
2273         case TCP_SEQ_STATE_LISTENING:
2274         case TCP_SEQ_STATE_ESTABLISHED:
2275                 if (sk->sk_state == TCP_TIME_WAIT)
2276                         get_timewait4_sock(v, seq, st->num);
2277                 else
2278                         get_tcp4_sock(v, seq, st->num);
2279                 break;
2280         case TCP_SEQ_STATE_OPENREQ:
2281                 get_openreq4(v, seq, st->num);
2282                 break;
2283         }
2284 out:
2285         seq_pad(seq, '\n');
2286         return 0;
2287 }
2288
2289 static const struct file_operations tcp_afinfo_seq_fops = {
2290         .owner   = THIS_MODULE,
2291         .open    = tcp_seq_open,
2292         .read    = seq_read,
2293         .llseek  = seq_lseek,
2294         .release = seq_release_net
2295 };
2296
2297 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
2298         .name           = "tcp",
2299         .family         = AF_INET,
2300         .seq_fops       = &tcp_afinfo_seq_fops,
2301         .seq_ops        = {
2302                 .show           = tcp4_seq_show,
2303         },
2304 };
2305
2306 static int __net_init tcp4_proc_init_net(struct net *net)
2307 {
2308         return tcp_proc_register(net, &tcp4_seq_afinfo);
2309 }
2310
2311 static void __net_exit tcp4_proc_exit_net(struct net *net)
2312 {
2313         tcp_proc_unregister(net, &tcp4_seq_afinfo);
2314 }
2315
2316 static struct pernet_operations tcp4_net_ops = {
2317         .init = tcp4_proc_init_net,
2318         .exit = tcp4_proc_exit_net,
2319 };
2320
2321 int __init tcp4_proc_init(void)
2322 {
2323         return register_pernet_subsys(&tcp4_net_ops);
2324 }
2325
2326 void tcp4_proc_exit(void)
2327 {
2328         unregister_pernet_subsys(&tcp4_net_ops);
2329 }
2330 #endif /* CONFIG_PROC_FS */
2331
2332 struct proto tcp_prot = {
2333         .name                   = "TCP",
2334         .owner                  = THIS_MODULE,
2335         .close                  = tcp_close,
2336         .connect                = tcp_v4_connect,
2337         .disconnect             = tcp_disconnect,
2338         .accept                 = inet_csk_accept,
2339         .ioctl                  = tcp_ioctl,
2340         .init                   = tcp_v4_init_sock,
2341         .destroy                = tcp_v4_destroy_sock,
2342         .shutdown               = tcp_shutdown,
2343         .setsockopt             = tcp_setsockopt,
2344         .getsockopt             = tcp_getsockopt,
2345         .recvmsg                = tcp_recvmsg,
2346         .sendmsg                = tcp_sendmsg,
2347         .sendpage               = tcp_sendpage,
2348         .backlog_rcv            = tcp_v4_do_rcv,
2349         .release_cb             = tcp_release_cb,
2350         .hash                   = inet_hash,
2351         .unhash                 = inet_unhash,
2352         .get_port               = inet_csk_get_port,
2353         .enter_memory_pressure  = tcp_enter_memory_pressure,
2354         .stream_memory_free     = tcp_stream_memory_free,
2355         .sockets_allocated      = &tcp_sockets_allocated,
2356         .orphan_count           = &tcp_orphan_count,
2357         .memory_allocated       = &tcp_memory_allocated,
2358         .memory_pressure        = &tcp_memory_pressure,
2359         .sysctl_mem             = sysctl_tcp_mem,
2360         .sysctl_wmem            = sysctl_tcp_wmem,
2361         .sysctl_rmem            = sysctl_tcp_rmem,
2362         .max_header             = MAX_TCP_HEADER,
2363         .obj_size               = sizeof(struct tcp_sock),
2364         .slab_flags             = SLAB_DESTROY_BY_RCU,
2365         .twsk_prot              = &tcp_timewait_sock_ops,
2366         .rsk_prot               = &tcp_request_sock_ops,
2367         .h.hashinfo             = &tcp_hashinfo,
2368         .no_autobind            = true,
2369 #ifdef CONFIG_COMPAT
2370         .compat_setsockopt      = compat_tcp_setsockopt,
2371         .compat_getsockopt      = compat_tcp_getsockopt,
2372 #endif
2373 #ifdef CONFIG_MEMCG_KMEM
2374         .init_cgroup            = tcp_init_cgroup,
2375         .destroy_cgroup         = tcp_destroy_cgroup,
2376         .proto_cgroup           = tcp_proto_cgroup,
2377 #endif
2378 };
2379 EXPORT_SYMBOL(tcp_prot);
2380
2381 static void __net_exit tcp_sk_exit(struct net *net)
2382 {
2383         int cpu;
2384
2385         for_each_possible_cpu(cpu)
2386                 inet_ctl_sock_destroy(*per_cpu_ptr(net->ipv4.tcp_sk, cpu));
2387         free_percpu(net->ipv4.tcp_sk);
2388 }
2389
2390 static int __net_init tcp_sk_init(struct net *net)
2391 {
2392         int res, cpu;
2393
2394         net->ipv4.tcp_sk = alloc_percpu(struct sock *);
2395         if (!net->ipv4.tcp_sk)
2396                 return -ENOMEM;
2397
2398         for_each_possible_cpu(cpu) {
2399                 struct sock *sk;
2400
2401                 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
2402                                            IPPROTO_TCP, net);
2403                 if (res)
2404                         goto fail;
2405                 *per_cpu_ptr(net->ipv4.tcp_sk, cpu) = sk;
2406         }
2407
2408         net->ipv4.sysctl_tcp_ecn = 2;
2409         net->ipv4.sysctl_tcp_ecn_fallback = 1;
2410
2411         net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
2412         net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
2413         net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
2414
2415         return 0;
2416 fail:
2417         tcp_sk_exit(net);
2418
2419         return res;
2420 }
2421
2422 static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
2423 {
2424         inet_twsk_purge(&tcp_hashinfo, &tcp_death_row, AF_INET);
2425 }
2426
2427 static struct pernet_operations __net_initdata tcp_sk_ops = {
2428        .init       = tcp_sk_init,
2429        .exit       = tcp_sk_exit,
2430        .exit_batch = tcp_sk_exit_batch,
2431 };
2432
2433 void __init tcp_v4_init(void)
2434 {
2435         inet_hashinfo_init(&tcp_hashinfo);
2436         if (register_pernet_subsys(&tcp_sk_ops))
2437                 panic("Failed to create the TCP control socket.\n");
2438 }