Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * IPv4 specific functions
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
19/*
20 * Changes:
21 * David S. Miller : New socket lookup architecture.
22 * This code is dedicated to John Dyson.
23 * David S. Miller : Change semantics of established hash,
24 * half is devoted to TIME_WAIT sockets
25 * and the rest go in the other half.
26 * Andi Kleen : Add support for syncookies and fixed
27 * some bugs: ip options weren't passed to
28 * the TCP layer, missed a check for an
29 * ACK bit.
30 * Andi Kleen : Implemented fast path mtu discovery.
31 * Fixed many serious bugs in the
32 * request_sock handling and moved
33 * most of it into the af independent code.
34 * Added tail drop and some other bugfixes.
35 * Added new listen semantics.
36 * Mike McLagan : Routing by source
37 * Juan Jose Ciarlante: ip_dynaddr bits
38 * Andi Kleen: various fixes.
39 * Vitaly E. Lavrov : Transparent proxy revived after year
40 * coma.
41 * Andi Kleen : Fix new listen.
42 * Andi Kleen : Fix accept error reporting.
43 * YOSHIFUJI Hideaki @USAGI and: Support IPV6_V6ONLY socket option, which
44 * Alexey Kuznetsov allow both IPv4 and IPv6 sockets to bind
45 * a single port at the same time.
46 */
47
48#define pr_fmt(fmt) "TCP: " fmt
49
50#include <linux/bottom_half.h>
51#include <linux/types.h>
52#include <linux/fcntl.h>
53#include <linux/module.h>
54#include <linux/random.h>
55#include <linux/cache.h>
56#include <linux/jhash.h>
57#include <linux/init.h>
58#include <linux/times.h>
59#include <linux/slab.h>
60
61#include <net/net_namespace.h>
62#include <net/icmp.h>
63#include <net/inet_hashtables.h>
64#include <net/tcp.h>
65#include <net/transp_v6.h>
66#include <net/ipv6.h>
67#include <net/inet_common.h>
68#include <net/timewait_sock.h>
69#include <net/xfrm.h>
70#include <net/secure_seq.h>
71#include <net/busy_poll.h>
72
73#include <linux/inet.h>
74#include <linux/ipv6.h>
75#include <linux/stddef.h>
76#include <linux/proc_fs.h>
77#include <linux/seq_file.h>
78#include <linux/inetdevice.h>
79#include <linux/btf_ids.h>
80
81#include <crypto/hash.h>
82#include <linux/scatterlist.h>
83
84#include <trace/events/tcp.h>
85
86#ifdef CONFIG_TCP_MD5SIG
87static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
88 __be32 daddr, __be32 saddr, const struct tcphdr *th);
89#endif
90
91struct inet_hashinfo tcp_hashinfo;
92EXPORT_SYMBOL(tcp_hashinfo);
93
94static u32 tcp_v4_init_seq(const struct sk_buff *skb)
95{
96 return secure_tcp_seq(ip_hdr(skb)->daddr,
97 ip_hdr(skb)->saddr,
98 tcp_hdr(skb)->dest,
99 tcp_hdr(skb)->source);
100}
101
102static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
103{
104 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
105}
106
107int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
108{
109 const struct inet_timewait_sock *tw = inet_twsk(sktw);
110 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
111 struct tcp_sock *tp = tcp_sk(sk);
112 int reuse = sock_net(sk)->ipv4.sysctl_tcp_tw_reuse;
113
114 if (reuse == 2) {
115 /* Still does not detect *everything* that goes through
116 * lo, since we require a loopback src or dst address
117 * or direct binding to 'lo' interface.
118 */
119 bool loopback = false;
120 if (tw->tw_bound_dev_if == LOOPBACK_IFINDEX)
121 loopback = true;
122#if IS_ENABLED(CONFIG_IPV6)
123 if (tw->tw_family == AF_INET6) {
124 if (ipv6_addr_loopback(&tw->tw_v6_daddr) ||
125 ipv6_addr_v4mapped_loopback(&tw->tw_v6_daddr) ||
126 ipv6_addr_loopback(&tw->tw_v6_rcv_saddr) ||
127 ipv6_addr_v4mapped_loopback(&tw->tw_v6_rcv_saddr))
128 loopback = true;
129 } else
130#endif
131 {
132 if (ipv4_is_loopback(tw->tw_daddr) ||
133 ipv4_is_loopback(tw->tw_rcv_saddr))
134 loopback = true;
135 }
136 if (!loopback)
137 reuse = 0;
138 }
139
140 /* With PAWS, it is safe from the viewpoint
141 of data integrity. Even without PAWS it is safe provided sequence
142 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
143
144 Actually, the idea is close to VJ's one, only timestamp cache is
145 held not per host, but per port pair and TW bucket is used as state
146 holder.
147
148 If TW bucket has been already destroyed we fall back to VJ's scheme
149 and use initial timestamp retrieved from peer table.
150 */
151 if (tcptw->tw_ts_recent_stamp &&
152 (!twp || (reuse && time_after32(ktime_get_seconds(),
153 tcptw->tw_ts_recent_stamp)))) {
154 /* In case of repair and re-using TIME-WAIT sockets we still
155 * want to be sure that it is safe as above but honor the
156 * sequence numbers and time stamps set as part of the repair
157 * process.
158 *
159 * Without this check re-using a TIME-WAIT socket with TCP
160 * repair would accumulate a -1 on the repair assigned
161 * sequence number. The first time it is reused the sequence
162 * is -1, the second time -2, etc. This fixes that issue
163 * without appearing to create any others.
164 */
165 if (likely(!tp->repair)) {
166 u32 seq = tcptw->tw_snd_nxt + 65535 + 2;
167
168 if (!seq)
169 seq = 1;
170 WRITE_ONCE(tp->write_seq, seq);
171 tp->rx_opt.ts_recent = tcptw->tw_ts_recent;
172 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
173 }
174 sock_hold(sktw);
175 return 1;
176 }
177
178 return 0;
179}
180EXPORT_SYMBOL_GPL(tcp_twsk_unique);
181
182static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
183 int addr_len)
184{
185 /* This check is replicated from tcp_v4_connect() and intended to
186 * prevent BPF program called below from accessing bytes that are out
187 * of the bound specified by user in addr_len.
188 */
189 if (addr_len < sizeof(struct sockaddr_in))
190 return -EINVAL;
191
192 sock_owned_by_me(sk);
193
194 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr);
195}
196
197/* This will initiate an outgoing connection. */
198int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
199{
200 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
201 struct inet_sock *inet = inet_sk(sk);
202 struct tcp_sock *tp = tcp_sk(sk);
203 __be16 orig_sport, orig_dport;
204 __be32 daddr, nexthop;
205 struct flowi4 *fl4;
206 struct rtable *rt;
207 int err;
208 struct ip_options_rcu *inet_opt;
209 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
210
211 if (addr_len < sizeof(struct sockaddr_in))
212 return -EINVAL;
213
214 if (usin->sin_family != AF_INET)
215 return -EAFNOSUPPORT;
216
217 nexthop = daddr = usin->sin_addr.s_addr;
218 inet_opt = rcu_dereference_protected(inet->inet_opt,
219 lockdep_sock_is_held(sk));
220 if (inet_opt && inet_opt->opt.srr) {
221 if (!daddr)
222 return -EINVAL;
223 nexthop = inet_opt->opt.faddr;
224 }
225
226 orig_sport = inet->inet_sport;
227 orig_dport = usin->sin_port;
228 fl4 = &inet->cork.fl.u.ip4;
229 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
230 RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
231 IPPROTO_TCP,
232 orig_sport, orig_dport, sk);
233 if (IS_ERR(rt)) {
234 err = PTR_ERR(rt);
235 if (err == -ENETUNREACH)
236 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
237 return err;
238 }
239
240 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
241 ip_rt_put(rt);
242 return -ENETUNREACH;
243 }
244
245 if (!inet_opt || !inet_opt->opt.srr)
246 daddr = fl4->daddr;
247
248 if (!inet->inet_saddr)
249 inet->inet_saddr = fl4->saddr;
250 sk_rcv_saddr_set(sk, inet->inet_saddr);
251
252 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
253 /* Reset inherited state */
254 tp->rx_opt.ts_recent = 0;
255 tp->rx_opt.ts_recent_stamp = 0;
256 if (likely(!tp->repair))
257 WRITE_ONCE(tp->write_seq, 0);
258 }
259
260 inet->inet_dport = usin->sin_port;
261 sk_daddr_set(sk, daddr);
262
263 inet_csk(sk)->icsk_ext_hdr_len = 0;
264 if (inet_opt)
265 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
266
267 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
268
269 /* Socket identity is still unknown (sport may be zero).
270 * However we set state to SYN-SENT and not releasing socket
271 * lock select source port, enter ourselves into the hash tables and
272 * complete initialization after this.
273 */
274 tcp_set_state(sk, TCP_SYN_SENT);
275 err = inet_hash_connect(tcp_death_row, sk);
276 if (err)
277 goto failure;
278
279 sk_set_txhash(sk);
280
281 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
282 inet->inet_sport, inet->inet_dport, sk);
283 if (IS_ERR(rt)) {
284 err = PTR_ERR(rt);
285 rt = NULL;
286 goto failure;
287 }
288 /* OK, now commit destination to socket. */
289 sk->sk_gso_type = SKB_GSO_TCPV4;
290 sk_setup_caps(sk, &rt->dst);
291 rt = NULL;
292
293 if (likely(!tp->repair)) {
294 if (!tp->write_seq)
295 WRITE_ONCE(tp->write_seq,
296 secure_tcp_seq(inet->inet_saddr,
297 inet->inet_daddr,
298 inet->inet_sport,
299 usin->sin_port));
300 tp->tsoffset = secure_tcp_ts_off(sock_net(sk),
301 inet->inet_saddr,
302 inet->inet_daddr);
303 }
304
305 inet->inet_id = prandom_u32();
306
307 if (tcp_fastopen_defer_connect(sk, &err))
308 return err;
309 if (err)
310 goto failure;
311
312 err = tcp_connect(sk);
313
314 if (err)
315 goto failure;
316
317 return 0;
318
319failure:
320 /*
321 * This unhashes the socket and releases the local port,
322 * if necessary.
323 */
324 tcp_set_state(sk, TCP_CLOSE);
325 ip_rt_put(rt);
326 sk->sk_route_caps = 0;
327 inet->inet_dport = 0;
328 return err;
329}
330EXPORT_SYMBOL(tcp_v4_connect);
331
332/*
333 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
334 * It can be called through tcp_release_cb() if socket was owned by user
335 * at the time tcp_v4_err() was called to handle ICMP message.
336 */
337void tcp_v4_mtu_reduced(struct sock *sk)
338{
339 struct inet_sock *inet = inet_sk(sk);
340 struct dst_entry *dst;
341 u32 mtu;
342
343 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
344 return;
345 mtu = READ_ONCE(tcp_sk(sk)->mtu_info);
346 dst = inet_csk_update_pmtu(sk, mtu);
347 if (!dst)
348 return;
349
350 /* Something is about to be wrong... Remember soft error
351 * for the case, if this connection will not able to recover.
352 */
353 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
354 sk->sk_err_soft = EMSGSIZE;
355
356 mtu = dst_mtu(dst);
357
358 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
359 ip_sk_accept_pmtu(sk) &&
360 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
361 tcp_sync_mss(sk, mtu);
362
363 /* Resend the TCP packet because it's
364 * clear that the old packet has been
365 * dropped. This is the new "fast" path mtu
366 * discovery.
367 */
368 tcp_simple_retransmit(sk);
369 } /* else let the usual retransmit timer handle it */
370}
371EXPORT_SYMBOL(tcp_v4_mtu_reduced);
372
373static void do_redirect(struct sk_buff *skb, struct sock *sk)
374{
375 struct dst_entry *dst = __sk_dst_check(sk, 0);
376
377 if (dst)
378 dst->ops->redirect(dst, sk, skb);
379}
380
381
382/* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
383void tcp_req_err(struct sock *sk, u32 seq, bool abort)
384{
385 struct request_sock *req = inet_reqsk(sk);
386 struct net *net = sock_net(sk);
387
388 /* ICMPs are not backlogged, hence we cannot get
389 * an established socket here.
390 */
391 if (seq != tcp_rsk(req)->snt_isn) {
392 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
393 } else if (abort) {
394 /*
395 * Still in SYN_RECV, just remove it silently.
396 * There is no good way to pass the error to the newly
397 * created socket, and POSIX does not want network
398 * errors returned from accept().
399 */
400 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
401 tcp_listendrop(req->rsk_listener);
402 }
403 reqsk_put(req);
404}
405EXPORT_SYMBOL(tcp_req_err);
406
407/* TCP-LD (RFC 6069) logic */
408void tcp_ld_RTO_revert(struct sock *sk, u32 seq)
409{
410 struct inet_connection_sock *icsk = inet_csk(sk);
411 struct tcp_sock *tp = tcp_sk(sk);
412 struct sk_buff *skb;
413 s32 remaining;
414 u32 delta_us;
415
416 if (sock_owned_by_user(sk))
417 return;
418
419 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
420 !icsk->icsk_backoff)
421 return;
422
423 skb = tcp_rtx_queue_head(sk);
424 if (WARN_ON_ONCE(!skb))
425 return;
426
427 icsk->icsk_backoff--;
428 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) : TCP_TIMEOUT_INIT;
429 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
430
431 tcp_mstamp_refresh(tp);
432 delta_us = (u32)(tp->tcp_mstamp - tcp_skb_timestamp_us(skb));
433 remaining = icsk->icsk_rto - usecs_to_jiffies(delta_us);
434
435 if (remaining > 0) {
436 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
437 remaining, TCP_RTO_MAX);
438 } else {
439 /* RTO revert clocked out retransmission.
440 * Will retransmit now.
441 */
442 tcp_retransmit_timer(sk);
443 }
444}
445EXPORT_SYMBOL(tcp_ld_RTO_revert);
446
447/*
448 * This routine is called by the ICMP module when it gets some
449 * sort of error condition. If err < 0 then the socket should
450 * be closed and the error returned to the user. If err > 0
451 * it's just the icmp type << 8 | icmp code. After adjustment
452 * header points to the first 8 bytes of the tcp header. We need
453 * to find the appropriate port.
454 *
455 * The locking strategy used here is very "optimistic". When
456 * someone else accesses the socket the ICMP is just dropped
457 * and for some paths there is no check at all.
458 * A more general error queue to queue errors for later handling
459 * is probably better.
460 *
461 */
462
463int tcp_v4_err(struct sk_buff *skb, u32 info)
464{
465 const struct iphdr *iph = (const struct iphdr *)skb->data;
466 struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
467 struct tcp_sock *tp;
468 struct inet_sock *inet;
469 const int type = icmp_hdr(skb)->type;
470 const int code = icmp_hdr(skb)->code;
471 struct sock *sk;
472 struct request_sock *fastopen;
473 u32 seq, snd_una;
474 int err;
475 struct net *net = dev_net(skb->dev);
476
477 sk = __inet_lookup_established(net, &tcp_hashinfo, iph->daddr,
478 th->dest, iph->saddr, ntohs(th->source),
479 inet_iif(skb), 0);
480 if (!sk) {
481 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
482 return -ENOENT;
483 }
484 if (sk->sk_state == TCP_TIME_WAIT) {
485 inet_twsk_put(inet_twsk(sk));
486 return 0;
487 }
488 seq = ntohl(th->seq);
489 if (sk->sk_state == TCP_NEW_SYN_RECV) {
490 tcp_req_err(sk, seq, type == ICMP_PARAMETERPROB ||
491 type == ICMP_TIME_EXCEEDED ||
492 (type == ICMP_DEST_UNREACH &&
493 (code == ICMP_NET_UNREACH ||
494 code == ICMP_HOST_UNREACH)));
495 return 0;
496 }
497
498 bh_lock_sock(sk);
499 /* If too many ICMPs get dropped on busy
500 * servers this needs to be solved differently.
501 * We do take care of PMTU discovery (RFC1191) special case :
502 * we can receive locally generated ICMP messages while socket is held.
503 */
504 if (sock_owned_by_user(sk)) {
505 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
506 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
507 }
508 if (sk->sk_state == TCP_CLOSE)
509 goto out;
510
511 if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
512 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
513 goto out;
514 }
515
516 tp = tcp_sk(sk);
517 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
518 fastopen = rcu_dereference(tp->fastopen_rsk);
519 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
520 if (sk->sk_state != TCP_LISTEN &&
521 !between(seq, snd_una, tp->snd_nxt)) {
522 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
523 goto out;
524 }
525
526 switch (type) {
527 case ICMP_REDIRECT:
528 if (!sock_owned_by_user(sk))
529 do_redirect(skb, sk);
530 goto out;
531 case ICMP_SOURCE_QUENCH:
532 /* Just silently ignore these. */
533 goto out;
534 case ICMP_PARAMETERPROB:
535 err = EPROTO;
536 break;
537 case ICMP_DEST_UNREACH:
538 if (code > NR_ICMP_UNREACH)
539 goto out;
540
541 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
542 /* We are not interested in TCP_LISTEN and open_requests
543 * (SYN-ACKs send out by Linux are always <576bytes so
544 * they should go through unfragmented).
545 */
546 if (sk->sk_state == TCP_LISTEN)
547 goto out;
548
549 WRITE_ONCE(tp->mtu_info, info);
550 if (!sock_owned_by_user(sk)) {
551 tcp_v4_mtu_reduced(sk);
552 } else {
553 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
554 sock_hold(sk);
555 }
556 goto out;
557 }
558
559 err = icmp_err_convert[code].errno;
560 /* check if this ICMP message allows revert of backoff.
561 * (see RFC 6069)
562 */
563 if (!fastopen &&
564 (code == ICMP_NET_UNREACH || code == ICMP_HOST_UNREACH))
565 tcp_ld_RTO_revert(sk, seq);
566 break;
567 case ICMP_TIME_EXCEEDED:
568 err = EHOSTUNREACH;
569 break;
570 default:
571 goto out;
572 }
573
574 switch (sk->sk_state) {
575 case TCP_SYN_SENT:
576 case TCP_SYN_RECV:
577 /* Only in fast or simultaneous open. If a fast open socket is
578 * already accepted it is treated as a connected one below.
579 */
580 if (fastopen && !fastopen->sk)
581 break;
582
583 ip_icmp_error(sk, skb, err, th->dest, info, (u8 *)th);
584
585 if (!sock_owned_by_user(sk)) {
586 sk->sk_err = err;
587
588 sk_error_report(sk);
589
590 tcp_done(sk);
591 } else {
592 sk->sk_err_soft = err;
593 }
594 goto out;
595 }
596
597 /* If we've already connected we will keep trying
598 * until we time out, or the user gives up.
599 *
600 * rfc1122 4.2.3.9 allows to consider as hard errors
601 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
602 * but it is obsoleted by pmtu discovery).
603 *
604 * Note, that in modern internet, where routing is unreliable
605 * and in each dark corner broken firewalls sit, sending random
606 * errors ordered by their masters even this two messages finally lose
607 * their original sense (even Linux sends invalid PORT_UNREACHs)
608 *
609 * Now we are in compliance with RFCs.
610 * --ANK (980905)
611 */
612
613 inet = inet_sk(sk);
614 if (!sock_owned_by_user(sk) && inet->recverr) {
615 sk->sk_err = err;
616 sk_error_report(sk);
617 } else { /* Only an error on timeout */
618 sk->sk_err_soft = err;
619 }
620
621out:
622 bh_unlock_sock(sk);
623 sock_put(sk);
624 return 0;
625}
626
627void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
628{
629 struct tcphdr *th = tcp_hdr(skb);
630
631 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
632 skb->csum_start = skb_transport_header(skb) - skb->head;
633 skb->csum_offset = offsetof(struct tcphdr, check);
634}
635
636/* This routine computes an IPv4 TCP checksum. */
637void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
638{
639 const struct inet_sock *inet = inet_sk(sk);
640
641 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
642}
643EXPORT_SYMBOL(tcp_v4_send_check);
644
645/*
646 * This routine will send an RST to the other tcp.
647 *
648 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
649 * for reset.
650 * Answer: if a packet caused RST, it is not for a socket
651 * existing in our system, if it is matched to a socket,
652 * it is just duplicate segment or bug in other side's TCP.
653 * So that we build reply only basing on parameters
654 * arrived with segment.
655 * Exception: precedence violation. We do not implement it in any case.
656 */
657
658#ifdef CONFIG_TCP_MD5SIG
659#define OPTION_BYTES TCPOLEN_MD5SIG_ALIGNED
660#else
661#define OPTION_BYTES sizeof(__be32)
662#endif
663
664static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
665{
666 const struct tcphdr *th = tcp_hdr(skb);
667 struct {
668 struct tcphdr th;
669 __be32 opt[OPTION_BYTES / sizeof(__be32)];
670 } rep;
671 struct ip_reply_arg arg;
672#ifdef CONFIG_TCP_MD5SIG
673 struct tcp_md5sig_key *key = NULL;
674 const __u8 *hash_location = NULL;
675 unsigned char newhash[16];
676 int genhash;
677 struct sock *sk1 = NULL;
678#endif
679 u64 transmit_time = 0;
680 struct sock *ctl_sk;
681 struct net *net;
682
683 /* Never send a reset in response to a reset. */
684 if (th->rst)
685 return;
686
687 /* If sk not NULL, it means we did a successful lookup and incoming
688 * route had to be correct. prequeue might have dropped our dst.
689 */
690 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
691 return;
692
693 /* Swap the send and the receive. */
694 memset(&rep, 0, sizeof(rep));
695 rep.th.dest = th->source;
696 rep.th.source = th->dest;
697 rep.th.doff = sizeof(struct tcphdr) / 4;
698 rep.th.rst = 1;
699
700 if (th->ack) {
701 rep.th.seq = th->ack_seq;
702 } else {
703 rep.th.ack = 1;
704 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
705 skb->len - (th->doff << 2));
706 }
707
708 memset(&arg, 0, sizeof(arg));
709 arg.iov[0].iov_base = (unsigned char *)&rep;
710 arg.iov[0].iov_len = sizeof(rep.th);
711
712 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
713#ifdef CONFIG_TCP_MD5SIG
714 rcu_read_lock();
715 hash_location = tcp_parse_md5sig_option(th);
716 if (sk && sk_fullsock(sk)) {
717 const union tcp_md5_addr *addr;
718 int l3index;
719
720 /* sdif set, means packet ingressed via a device
721 * in an L3 domain and inet_iif is set to it.
722 */
723 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
724 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
725 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
726 } else if (hash_location) {
727 const union tcp_md5_addr *addr;
728 int sdif = tcp_v4_sdif(skb);
729 int dif = inet_iif(skb);
730 int l3index;
731
732 /*
733 * active side is lost. Try to find listening socket through
734 * source port, and then find md5 key through listening socket.
735 * we are not loose security here:
736 * Incoming packet is checked with md5 hash with finding key,
737 * no RST generated if md5 hash doesn't match.
738 */
739 sk1 = __inet_lookup_listener(net, &tcp_hashinfo, NULL, 0,
740 ip_hdr(skb)->saddr,
741 th->source, ip_hdr(skb)->daddr,
742 ntohs(th->source), dif, sdif);
743 /* don't send rst if it can't find key */
744 if (!sk1)
745 goto out;
746
747 /* sdif set, means packet ingressed via a device
748 * in an L3 domain and dif is set to it.
749 */
750 l3index = sdif ? dif : 0;
751 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
752 key = tcp_md5_do_lookup(sk1, l3index, addr, AF_INET);
753 if (!key)
754 goto out;
755
756
757 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
758 if (genhash || memcmp(hash_location, newhash, 16) != 0)
759 goto out;
760
761 }
762
763 if (key) {
764 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
765 (TCPOPT_NOP << 16) |
766 (TCPOPT_MD5SIG << 8) |
767 TCPOLEN_MD5SIG);
768 /* Update length and the length the header thinks exists */
769 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
770 rep.th.doff = arg.iov[0].iov_len / 4;
771
772 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
773 key, ip_hdr(skb)->saddr,
774 ip_hdr(skb)->daddr, &rep.th);
775 }
776#endif
777 /* Can't co-exist with TCPMD5, hence check rep.opt[0] */
778 if (rep.opt[0] == 0) {
779 __be32 mrst = mptcp_reset_option(skb);
780
781 if (mrst) {
782 rep.opt[0] = mrst;
783 arg.iov[0].iov_len += sizeof(mrst);
784 rep.th.doff = arg.iov[0].iov_len / 4;
785 }
786 }
787
788 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
789 ip_hdr(skb)->saddr, /* XXX */
790 arg.iov[0].iov_len, IPPROTO_TCP, 0);
791 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
792 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
793
794 /* When socket is gone, all binding information is lost.
795 * routing might fail in this case. No choice here, if we choose to force
796 * input interface, we will misroute in case of asymmetric route.
797 */
798 if (sk) {
799 arg.bound_dev_if = sk->sk_bound_dev_if;
800 if (sk_fullsock(sk))
801 trace_tcp_send_reset(sk, skb);
802 }
803
804 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
805 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
806
807 arg.tos = ip_hdr(skb)->tos;
808 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
809 local_bh_disable();
810 ctl_sk = this_cpu_read(*net->ipv4.tcp_sk);
811 if (sk) {
812 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
813 inet_twsk(sk)->tw_mark : sk->sk_mark;
814 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
815 inet_twsk(sk)->tw_priority : sk->sk_priority;
816 transmit_time = tcp_transmit_time(sk);
817 }
818 ip_send_unicast_reply(ctl_sk,
819 skb, &TCP_SKB_CB(skb)->header.h4.opt,
820 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
821 &arg, arg.iov[0].iov_len,
822 transmit_time);
823
824 ctl_sk->sk_mark = 0;
825 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
826 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
827 local_bh_enable();
828
829#ifdef CONFIG_TCP_MD5SIG
830out:
831 rcu_read_unlock();
832#endif
833}
834
835/* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
836 outside socket context is ugly, certainly. What can I do?
837 */
838
839static void tcp_v4_send_ack(const struct sock *sk,
840 struct sk_buff *skb, u32 seq, u32 ack,
841 u32 win, u32 tsval, u32 tsecr, int oif,
842 struct tcp_md5sig_key *key,
843 int reply_flags, u8 tos)
844{
845 const struct tcphdr *th = tcp_hdr(skb);
846 struct {
847 struct tcphdr th;
848 __be32 opt[(TCPOLEN_TSTAMP_ALIGNED >> 2)
849#ifdef CONFIG_TCP_MD5SIG
850 + (TCPOLEN_MD5SIG_ALIGNED >> 2)
851#endif
852 ];
853 } rep;
854 struct net *net = sock_net(sk);
855 struct ip_reply_arg arg;
856 struct sock *ctl_sk;
857 u64 transmit_time;
858
859 memset(&rep.th, 0, sizeof(struct tcphdr));
860 memset(&arg, 0, sizeof(arg));
861
862 arg.iov[0].iov_base = (unsigned char *)&rep;
863 arg.iov[0].iov_len = sizeof(rep.th);
864 if (tsecr) {
865 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
866 (TCPOPT_TIMESTAMP << 8) |
867 TCPOLEN_TIMESTAMP);
868 rep.opt[1] = htonl(tsval);
869 rep.opt[2] = htonl(tsecr);
870 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
871 }
872
873 /* Swap the send and the receive. */
874 rep.th.dest = th->source;
875 rep.th.source = th->dest;
876 rep.th.doff = arg.iov[0].iov_len / 4;
877 rep.th.seq = htonl(seq);
878 rep.th.ack_seq = htonl(ack);
879 rep.th.ack = 1;
880 rep.th.window = htons(win);
881
882#ifdef CONFIG_TCP_MD5SIG
883 if (key) {
884 int offset = (tsecr) ? 3 : 0;
885
886 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
887 (TCPOPT_NOP << 16) |
888 (TCPOPT_MD5SIG << 8) |
889 TCPOLEN_MD5SIG);
890 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
891 rep.th.doff = arg.iov[0].iov_len/4;
892
893 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
894 key, ip_hdr(skb)->saddr,
895 ip_hdr(skb)->daddr, &rep.th);
896 }
897#endif
898 arg.flags = reply_flags;
899 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
900 ip_hdr(skb)->saddr, /* XXX */
901 arg.iov[0].iov_len, IPPROTO_TCP, 0);
902 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
903 if (oif)
904 arg.bound_dev_if = oif;
905 arg.tos = tos;
906 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
907 local_bh_disable();
908 ctl_sk = this_cpu_read(*net->ipv4.tcp_sk);
909 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
910 inet_twsk(sk)->tw_mark : sk->sk_mark;
911 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
912 inet_twsk(sk)->tw_priority : sk->sk_priority;
913 transmit_time = tcp_transmit_time(sk);
914 ip_send_unicast_reply(ctl_sk,
915 skb, &TCP_SKB_CB(skb)->header.h4.opt,
916 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
917 &arg, arg.iov[0].iov_len,
918 transmit_time);
919
920 ctl_sk->sk_mark = 0;
921 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
922 local_bh_enable();
923}
924
925static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
926{
927 struct inet_timewait_sock *tw = inet_twsk(sk);
928 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
929
930 tcp_v4_send_ack(sk, skb,
931 tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
932 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
933 tcp_time_stamp_raw() + tcptw->tw_ts_offset,
934 tcptw->tw_ts_recent,
935 tw->tw_bound_dev_if,
936 tcp_twsk_md5_key(tcptw),
937 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
938 tw->tw_tos
939 );
940
941 inet_twsk_put(tw);
942}
943
944static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
945 struct request_sock *req)
946{
947 const union tcp_md5_addr *addr;
948 int l3index;
949
950 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
951 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
952 */
953 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
954 tcp_sk(sk)->snd_nxt;
955
956 /* RFC 7323 2.3
957 * The window field (SEG.WND) of every outgoing segment, with the
958 * exception of <SYN> segments, MUST be right-shifted by
959 * Rcv.Wind.Shift bits:
960 */
961 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
962 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
963 tcp_v4_send_ack(sk, skb, seq,
964 tcp_rsk(req)->rcv_nxt,
965 req->rsk_rcv_wnd >> inet_rsk(req)->rcv_wscale,
966 tcp_time_stamp_raw() + tcp_rsk(req)->ts_off,
967 req->ts_recent,
968 0,
969 tcp_md5_do_lookup(sk, l3index, addr, AF_INET),
970 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
971 ip_hdr(skb)->tos);
972}
973
974/*
975 * Send a SYN-ACK after having received a SYN.
976 * This still operates on a request_sock only, not on a big
977 * socket.
978 */
979static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
980 struct flowi *fl,
981 struct request_sock *req,
982 struct tcp_fastopen_cookie *foc,
983 enum tcp_synack_type synack_type,
984 struct sk_buff *syn_skb)
985{
986 const struct inet_request_sock *ireq = inet_rsk(req);
987 struct flowi4 fl4;
988 int err = -1;
989 struct sk_buff *skb;
990 u8 tos;
991
992 /* First, grab a route. */
993 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
994 return -1;
995
996 skb = tcp_make_synack(sk, dst, req, foc, synack_type, syn_skb);
997
998 if (skb) {
999 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
1000
1001 tos = sock_net(sk)->ipv4.sysctl_tcp_reflect_tos ?
1002 (tcp_rsk(req)->syn_tos & ~INET_ECN_MASK) |
1003 (inet_sk(sk)->tos & INET_ECN_MASK) :
1004 inet_sk(sk)->tos;
1005
1006 if (!INET_ECN_is_capable(tos) &&
1007 tcp_bpf_ca_needs_ecn((struct sock *)req))
1008 tos |= INET_ECN_ECT_0;
1009
1010 rcu_read_lock();
1011 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
1012 ireq->ir_rmt_addr,
1013 rcu_dereference(ireq->ireq_opt),
1014 tos);
1015 rcu_read_unlock();
1016 err = net_xmit_eval(err);
1017 }
1018
1019 return err;
1020}
1021
1022/*
1023 * IPv4 request_sock destructor.
1024 */
1025static void tcp_v4_reqsk_destructor(struct request_sock *req)
1026{
1027 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
1028}
1029
1030#ifdef CONFIG_TCP_MD5SIG
1031/*
1032 * RFC2385 MD5 checksumming requires a mapping of
1033 * IP address->MD5 Key.
1034 * We need to maintain these in the sk structure.
1035 */
1036
1037DEFINE_STATIC_KEY_FALSE(tcp_md5_needed);
1038EXPORT_SYMBOL(tcp_md5_needed);
1039
1040static bool better_md5_match(struct tcp_md5sig_key *old, struct tcp_md5sig_key *new)
1041{
1042 if (!old)
1043 return true;
1044
1045 /* l3index always overrides non-l3index */
1046 if (old->l3index && new->l3index == 0)
1047 return false;
1048 if (old->l3index == 0 && new->l3index)
1049 return true;
1050
1051 return old->prefixlen < new->prefixlen;
1052}
1053
1054/* Find the Key structure for an address. */
1055struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1056 const union tcp_md5_addr *addr,
1057 int family)
1058{
1059 const struct tcp_sock *tp = tcp_sk(sk);
1060 struct tcp_md5sig_key *key;
1061 const struct tcp_md5sig_info *md5sig;
1062 __be32 mask;
1063 struct tcp_md5sig_key *best_match = NULL;
1064 bool match;
1065
1066 /* caller either holds rcu_read_lock() or socket lock */
1067 md5sig = rcu_dereference_check(tp->md5sig_info,
1068 lockdep_sock_is_held(sk));
1069 if (!md5sig)
1070 return NULL;
1071
1072 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1073 lockdep_sock_is_held(sk)) {
1074 if (key->family != family)
1075 continue;
1076 if (key->l3index && key->l3index != l3index)
1077 continue;
1078 if (family == AF_INET) {
1079 mask = inet_make_mask(key->prefixlen);
1080 match = (key->addr.a4.s_addr & mask) ==
1081 (addr->a4.s_addr & mask);
1082#if IS_ENABLED(CONFIG_IPV6)
1083 } else if (family == AF_INET6) {
1084 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
1085 key->prefixlen);
1086#endif
1087 } else {
1088 match = false;
1089 }
1090
1091 if (match && better_md5_match(best_match, key))
1092 best_match = key;
1093 }
1094 return best_match;
1095}
1096EXPORT_SYMBOL(__tcp_md5_do_lookup);
1097
1098static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
1099 const union tcp_md5_addr *addr,
1100 int family, u8 prefixlen,
1101 int l3index)
1102{
1103 const struct tcp_sock *tp = tcp_sk(sk);
1104 struct tcp_md5sig_key *key;
1105 unsigned int size = sizeof(struct in_addr);
1106 const struct tcp_md5sig_info *md5sig;
1107
1108 /* caller either holds rcu_read_lock() or socket lock */
1109 md5sig = rcu_dereference_check(tp->md5sig_info,
1110 lockdep_sock_is_held(sk));
1111 if (!md5sig)
1112 return NULL;
1113#if IS_ENABLED(CONFIG_IPV6)
1114 if (family == AF_INET6)
1115 size = sizeof(struct in6_addr);
1116#endif
1117 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1118 lockdep_sock_is_held(sk)) {
1119 if (key->family != family)
1120 continue;
1121 if (key->l3index != l3index)
1122 continue;
1123 if (!memcmp(&key->addr, addr, size) &&
1124 key->prefixlen == prefixlen)
1125 return key;
1126 }
1127 return NULL;
1128}
1129
1130struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1131 const struct sock *addr_sk)
1132{
1133 const union tcp_md5_addr *addr;
1134 int l3index;
1135
1136 l3index = l3mdev_master_ifindex_by_index(sock_net(sk),
1137 addr_sk->sk_bound_dev_if);
1138 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
1139 return tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1140}
1141EXPORT_SYMBOL(tcp_v4_md5_lookup);
1142
1143/* This can be called on a newly created socket, from other files */
1144int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1145 int family, u8 prefixlen, int l3index,
1146 const u8 *newkey, u8 newkeylen, gfp_t gfp)
1147{
1148 /* Add Key to the list */
1149 struct tcp_md5sig_key *key;
1150 struct tcp_sock *tp = tcp_sk(sk);
1151 struct tcp_md5sig_info *md5sig;
1152
1153 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index);
1154 if (key) {
1155 /* Pre-existing entry - just update that one.
1156 * Note that the key might be used concurrently.
1157 * data_race() is telling kcsan that we do not care of
1158 * key mismatches, since changing MD5 key on live flows
1159 * can lead to packet drops.
1160 */
1161 data_race(memcpy(key->key, newkey, newkeylen));
1162
1163 /* Pairs with READ_ONCE() in tcp_md5_hash_key().
1164 * Also note that a reader could catch new key->keylen value
1165 * but old key->key[], this is the reason we use __GFP_ZERO
1166 * at sock_kmalloc() time below these lines.
1167 */
1168 WRITE_ONCE(key->keylen, newkeylen);
1169
1170 return 0;
1171 }
1172
1173 md5sig = rcu_dereference_protected(tp->md5sig_info,
1174 lockdep_sock_is_held(sk));
1175 if (!md5sig) {
1176 md5sig = kmalloc(sizeof(*md5sig), gfp);
1177 if (!md5sig)
1178 return -ENOMEM;
1179
1180 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1181 INIT_HLIST_HEAD(&md5sig->head);
1182 rcu_assign_pointer(tp->md5sig_info, md5sig);
1183 }
1184
1185 key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
1186 if (!key)
1187 return -ENOMEM;
1188 if (!tcp_alloc_md5sig_pool()) {
1189 sock_kfree_s(sk, key, sizeof(*key));
1190 return -ENOMEM;
1191 }
1192
1193 memcpy(key->key, newkey, newkeylen);
1194 key->keylen = newkeylen;
1195 key->family = family;
1196 key->prefixlen = prefixlen;
1197 key->l3index = l3index;
1198 memcpy(&key->addr, addr,
1199 (family == AF_INET6) ? sizeof(struct in6_addr) :
1200 sizeof(struct in_addr));
1201 hlist_add_head_rcu(&key->node, &md5sig->head);
1202 return 0;
1203}
1204EXPORT_SYMBOL(tcp_md5_do_add);
1205
1206int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1207 u8 prefixlen, int l3index)
1208{
1209 struct tcp_md5sig_key *key;
1210
1211 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index);
1212 if (!key)
1213 return -ENOENT;
1214 hlist_del_rcu(&key->node);
1215 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1216 kfree_rcu(key, rcu);
1217 return 0;
1218}
1219EXPORT_SYMBOL(tcp_md5_do_del);
1220
1221static void tcp_clear_md5_list(struct sock *sk)
1222{
1223 struct tcp_sock *tp = tcp_sk(sk);
1224 struct tcp_md5sig_key *key;
1225 struct hlist_node *n;
1226 struct tcp_md5sig_info *md5sig;
1227
1228 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1229
1230 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1231 hlist_del_rcu(&key->node);
1232 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1233 kfree_rcu(key, rcu);
1234 }
1235}
1236
1237static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1238 sockptr_t optval, int optlen)
1239{
1240 struct tcp_md5sig cmd;
1241 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1242 const union tcp_md5_addr *addr;
1243 u8 prefixlen = 32;
1244 int l3index = 0;
1245
1246 if (optlen < sizeof(cmd))
1247 return -EINVAL;
1248
1249 if (copy_from_sockptr(&cmd, optval, sizeof(cmd)))
1250 return -EFAULT;
1251
1252 if (sin->sin_family != AF_INET)
1253 return -EINVAL;
1254
1255 if (optname == TCP_MD5SIG_EXT &&
1256 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1257 prefixlen = cmd.tcpm_prefixlen;
1258 if (prefixlen > 32)
1259 return -EINVAL;
1260 }
1261
1262 if (optname == TCP_MD5SIG_EXT &&
1263 cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX) {
1264 struct net_device *dev;
1265
1266 rcu_read_lock();
1267 dev = dev_get_by_index_rcu(sock_net(sk), cmd.tcpm_ifindex);
1268 if (dev && netif_is_l3_master(dev))
1269 l3index = dev->ifindex;
1270
1271 rcu_read_unlock();
1272
1273 /* ok to reference set/not set outside of rcu;
1274 * right now device MUST be an L3 master
1275 */
1276 if (!dev || !l3index)
1277 return -EINVAL;
1278 }
1279
1280 addr = (union tcp_md5_addr *)&sin->sin_addr.s_addr;
1281
1282 if (!cmd.tcpm_keylen)
1283 return tcp_md5_do_del(sk, addr, AF_INET, prefixlen, l3index);
1284
1285 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1286 return -EINVAL;
1287
1288 return tcp_md5_do_add(sk, addr, AF_INET, prefixlen, l3index,
1289 cmd.tcpm_key, cmd.tcpm_keylen, GFP_KERNEL);
1290}
1291
1292static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
1293 __be32 daddr, __be32 saddr,
1294 const struct tcphdr *th, int nbytes)
1295{
1296 struct tcp4_pseudohdr *bp;
1297 struct scatterlist sg;
1298 struct tcphdr *_th;
1299
1300 bp = hp->scratch;
1301 bp->saddr = saddr;
1302 bp->daddr = daddr;
1303 bp->pad = 0;
1304 bp->protocol = IPPROTO_TCP;
1305 bp->len = cpu_to_be16(nbytes);
1306
1307 _th = (struct tcphdr *)(bp + 1);
1308 memcpy(_th, th, sizeof(*th));
1309 _th->check = 0;
1310
1311 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1312 ahash_request_set_crypt(hp->md5_req, &sg, NULL,
1313 sizeof(*bp) + sizeof(*th));
1314 return crypto_ahash_update(hp->md5_req);
1315}
1316
1317static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1318 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1319{
1320 struct tcp_md5sig_pool *hp;
1321 struct ahash_request *req;
1322
1323 hp = tcp_get_md5sig_pool();
1324 if (!hp)
1325 goto clear_hash_noput;
1326 req = hp->md5_req;
1327
1328 if (crypto_ahash_init(req))
1329 goto clear_hash;
1330 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
1331 goto clear_hash;
1332 if (tcp_md5_hash_key(hp, key))
1333 goto clear_hash;
1334 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1335 if (crypto_ahash_final(req))
1336 goto clear_hash;
1337
1338 tcp_put_md5sig_pool();
1339 return 0;
1340
1341clear_hash:
1342 tcp_put_md5sig_pool();
1343clear_hash_noput:
1344 memset(md5_hash, 0, 16);
1345 return 1;
1346}
1347
1348int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1349 const struct sock *sk,
1350 const struct sk_buff *skb)
1351{
1352 struct tcp_md5sig_pool *hp;
1353 struct ahash_request *req;
1354 const struct tcphdr *th = tcp_hdr(skb);
1355 __be32 saddr, daddr;
1356
1357 if (sk) { /* valid for establish/request sockets */
1358 saddr = sk->sk_rcv_saddr;
1359 daddr = sk->sk_daddr;
1360 } else {
1361 const struct iphdr *iph = ip_hdr(skb);
1362 saddr = iph->saddr;
1363 daddr = iph->daddr;
1364 }
1365
1366 hp = tcp_get_md5sig_pool();
1367 if (!hp)
1368 goto clear_hash_noput;
1369 req = hp->md5_req;
1370
1371 if (crypto_ahash_init(req))
1372 goto clear_hash;
1373
1374 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, skb->len))
1375 goto clear_hash;
1376 if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
1377 goto clear_hash;
1378 if (tcp_md5_hash_key(hp, key))
1379 goto clear_hash;
1380 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1381 if (crypto_ahash_final(req))
1382 goto clear_hash;
1383
1384 tcp_put_md5sig_pool();
1385 return 0;
1386
1387clear_hash:
1388 tcp_put_md5sig_pool();
1389clear_hash_noput:
1390 memset(md5_hash, 0, 16);
1391 return 1;
1392}
1393EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1394
1395#endif
1396
1397/* Called with rcu_read_lock() */
1398static bool tcp_v4_inbound_md5_hash(const struct sock *sk,
1399 const struct sk_buff *skb,
1400 int dif, int sdif)
1401{
1402#ifdef CONFIG_TCP_MD5SIG
1403 /*
1404 * This gets called for each TCP segment that arrives
1405 * so we want to be efficient.
1406 * We have 3 drop cases:
1407 * o No MD5 hash and one expected.
1408 * o MD5 hash and we're not expecting one.
1409 * o MD5 hash and its wrong.
1410 */
1411 const __u8 *hash_location = NULL;
1412 struct tcp_md5sig_key *hash_expected;
1413 const struct iphdr *iph = ip_hdr(skb);
1414 const struct tcphdr *th = tcp_hdr(skb);
1415 const union tcp_md5_addr *addr;
1416 unsigned char newhash[16];
1417 int genhash, l3index;
1418
1419 /* sdif set, means packet ingressed via a device
1420 * in an L3 domain and dif is set to the l3mdev
1421 */
1422 l3index = sdif ? dif : 0;
1423
1424 addr = (union tcp_md5_addr *)&iph->saddr;
1425 hash_expected = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1426 hash_location = tcp_parse_md5sig_option(th);
1427
1428 /* We've parsed the options - do we have a hash? */
1429 if (!hash_expected && !hash_location)
1430 return false;
1431
1432 if (hash_expected && !hash_location) {
1433 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
1434 return true;
1435 }
1436
1437 if (!hash_expected && hash_location) {
1438 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
1439 return true;
1440 }
1441
1442 /* Okay, so this is hash_expected and hash_location -
1443 * so we need to calculate the checksum.
1444 */
1445 genhash = tcp_v4_md5_hash_skb(newhash,
1446 hash_expected,
1447 NULL, skb);
1448
1449 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
1450 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
1451 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s L3 index %d\n",
1452 &iph->saddr, ntohs(th->source),
1453 &iph->daddr, ntohs(th->dest),
1454 genhash ? " tcp_v4_calc_md5_hash failed"
1455 : "", l3index);
1456 return true;
1457 }
1458 return false;
1459#endif
1460 return false;
1461}
1462
1463static void tcp_v4_init_req(struct request_sock *req,
1464 const struct sock *sk_listener,
1465 struct sk_buff *skb)
1466{
1467 struct inet_request_sock *ireq = inet_rsk(req);
1468 struct net *net = sock_net(sk_listener);
1469
1470 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1471 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1472 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1473}
1474
1475static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1476 struct sk_buff *skb,
1477 struct flowi *fl,
1478 struct request_sock *req)
1479{
1480 tcp_v4_init_req(req, sk, skb);
1481
1482 if (security_inet_conn_request(sk, skb, req))
1483 return NULL;
1484
1485 return inet_csk_route_req(sk, &fl->u.ip4, req);
1486}
1487
1488struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1489 .family = PF_INET,
1490 .obj_size = sizeof(struct tcp_request_sock),
1491 .rtx_syn_ack = tcp_rtx_synack,
1492 .send_ack = tcp_v4_reqsk_send_ack,
1493 .destructor = tcp_v4_reqsk_destructor,
1494 .send_reset = tcp_v4_send_reset,
1495 .syn_ack_timeout = tcp_syn_ack_timeout,
1496};
1497
1498const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1499 .mss_clamp = TCP_MSS_DEFAULT,
1500#ifdef CONFIG_TCP_MD5SIG
1501 .req_md5_lookup = tcp_v4_md5_lookup,
1502 .calc_md5_hash = tcp_v4_md5_hash_skb,
1503#endif
1504#ifdef CONFIG_SYN_COOKIES
1505 .cookie_init_seq = cookie_v4_init_sequence,
1506#endif
1507 .route_req = tcp_v4_route_req,
1508 .init_seq = tcp_v4_init_seq,
1509 .init_ts_off = tcp_v4_init_ts_off,
1510 .send_synack = tcp_v4_send_synack,
1511};
1512
1513int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1514{
1515 /* Never answer to SYNs send to broadcast or multicast */
1516 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1517 goto drop;
1518
1519 return tcp_conn_request(&tcp_request_sock_ops,
1520 &tcp_request_sock_ipv4_ops, sk, skb);
1521
1522drop:
1523 tcp_listendrop(sk);
1524 return 0;
1525}
1526EXPORT_SYMBOL(tcp_v4_conn_request);
1527
1528
1529/*
1530 * The three way handshake has completed - we got a valid synack -
1531 * now create the new socket.
1532 */
1533struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1534 struct request_sock *req,
1535 struct dst_entry *dst,
1536 struct request_sock *req_unhash,
1537 bool *own_req)
1538{
1539 struct inet_request_sock *ireq;
1540 bool found_dup_sk = false;
1541 struct inet_sock *newinet;
1542 struct tcp_sock *newtp;
1543 struct sock *newsk;
1544#ifdef CONFIG_TCP_MD5SIG
1545 const union tcp_md5_addr *addr;
1546 struct tcp_md5sig_key *key;
1547 int l3index;
1548#endif
1549 struct ip_options_rcu *inet_opt;
1550
1551 if (sk_acceptq_is_full(sk))
1552 goto exit_overflow;
1553
1554 newsk = tcp_create_openreq_child(sk, req, skb);
1555 if (!newsk)
1556 goto exit_nonewsk;
1557
1558 newsk->sk_gso_type = SKB_GSO_TCPV4;
1559 inet_sk_rx_dst_set(newsk, skb);
1560
1561 newtp = tcp_sk(newsk);
1562 newinet = inet_sk(newsk);
1563 ireq = inet_rsk(req);
1564 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1565 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1566 newsk->sk_bound_dev_if = ireq->ir_iif;
1567 newinet->inet_saddr = ireq->ir_loc_addr;
1568 inet_opt = rcu_dereference(ireq->ireq_opt);
1569 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1570 newinet->mc_index = inet_iif(skb);
1571 newinet->mc_ttl = ip_hdr(skb)->ttl;
1572 newinet->rcv_tos = ip_hdr(skb)->tos;
1573 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1574 if (inet_opt)
1575 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1576 newinet->inet_id = prandom_u32();
1577
1578 /* Set ToS of the new socket based upon the value of incoming SYN.
1579 * ECT bits are set later in tcp_init_transfer().
1580 */
1581 if (sock_net(sk)->ipv4.sysctl_tcp_reflect_tos)
1582 newinet->tos = tcp_rsk(req)->syn_tos & ~INET_ECN_MASK;
1583
1584 if (!dst) {
1585 dst = inet_csk_route_child_sock(sk, newsk, req);
1586 if (!dst)
1587 goto put_and_exit;
1588 } else {
1589 /* syncookie case : see end of cookie_v4_check() */
1590 }
1591 sk_setup_caps(newsk, dst);
1592
1593 tcp_ca_openreq_child(newsk, dst);
1594
1595 tcp_sync_mss(newsk, dst_mtu(dst));
1596 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1597
1598 tcp_initialize_rcv_mss(newsk);
1599
1600#ifdef CONFIG_TCP_MD5SIG
1601 l3index = l3mdev_master_ifindex_by_index(sock_net(sk), ireq->ir_iif);
1602 /* Copy over the MD5 key from the original socket */
1603 addr = (union tcp_md5_addr *)&newinet->inet_daddr;
1604 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1605 if (key) {
1606 /*
1607 * We're using one, so create a matching key
1608 * on the newsk structure. If we fail to get
1609 * memory, then we end up not copying the key
1610 * across. Shucks.
1611 */
1612 tcp_md5_do_add(newsk, addr, AF_INET, 32, l3index,
1613 key->key, key->keylen, GFP_ATOMIC);
1614 sk_nocaps_add(newsk, NETIF_F_GSO_MASK);
1615 }
1616#endif
1617
1618 if (__inet_inherit_port(sk, newsk) < 0)
1619 goto put_and_exit;
1620 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash),
1621 &found_dup_sk);
1622 if (likely(*own_req)) {
1623 tcp_move_syn(newtp, req);
1624 ireq->ireq_opt = NULL;
1625 } else {
1626 newinet->inet_opt = NULL;
1627
1628 if (!req_unhash && found_dup_sk) {
1629 /* This code path should only be executed in the
1630 * syncookie case only
1631 */
1632 bh_unlock_sock(newsk);
1633 sock_put(newsk);
1634 newsk = NULL;
1635 }
1636 }
1637 return newsk;
1638
1639exit_overflow:
1640 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1641exit_nonewsk:
1642 dst_release(dst);
1643exit:
1644 tcp_listendrop(sk);
1645 return NULL;
1646put_and_exit:
1647 newinet->inet_opt = NULL;
1648 inet_csk_prepare_forced_close(newsk);
1649 tcp_done(newsk);
1650 goto exit;
1651}
1652EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1653
1654static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1655{
1656#ifdef CONFIG_SYN_COOKIES
1657 const struct tcphdr *th = tcp_hdr(skb);
1658
1659 if (!th->syn)
1660 sk = cookie_v4_check(sk, skb);
1661#endif
1662 return sk;
1663}
1664
1665u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
1666 struct tcphdr *th, u32 *cookie)
1667{
1668 u16 mss = 0;
1669#ifdef CONFIG_SYN_COOKIES
1670 mss = tcp_get_syncookie_mss(&tcp_request_sock_ops,
1671 &tcp_request_sock_ipv4_ops, sk, th);
1672 if (mss) {
1673 *cookie = __cookie_v4_init_sequence(iph, th, &mss);
1674 tcp_synq_overflow(sk);
1675 }
1676#endif
1677 return mss;
1678}
1679
1680INDIRECT_CALLABLE_DECLARE(struct dst_entry *ipv4_dst_check(struct dst_entry *,
1681 u32));
1682/* The socket must have it's spinlock held when we get
1683 * here, unless it is a TCP_LISTEN socket.
1684 *
1685 * We have a potential double-lock case here, so even when
1686 * doing backlog processing we use the BH locking scheme.
1687 * This is because we cannot sleep with the original spinlock
1688 * held.
1689 */
1690int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1691{
1692 struct sock *rsk;
1693
1694 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1695 struct dst_entry *dst = sk->sk_rx_dst;
1696
1697 sock_rps_save_rxhash(sk, skb);
1698 sk_mark_napi_id(sk, skb);
1699 if (dst) {
1700 if (inet_sk(sk)->rx_dst_ifindex != skb->skb_iif ||
1701 !INDIRECT_CALL_1(dst->ops->check, ipv4_dst_check,
1702 dst, 0)) {
1703 dst_release(dst);
1704 sk->sk_rx_dst = NULL;
1705 }
1706 }
1707 tcp_rcv_established(sk, skb);
1708 return 0;
1709 }
1710
1711 if (tcp_checksum_complete(skb))
1712 goto csum_err;
1713
1714 if (sk->sk_state == TCP_LISTEN) {
1715 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1716
1717 if (!nsk)
1718 goto discard;
1719 if (nsk != sk) {
1720 if (tcp_child_process(sk, nsk, skb)) {
1721 rsk = nsk;
1722 goto reset;
1723 }
1724 return 0;
1725 }
1726 } else
1727 sock_rps_save_rxhash(sk, skb);
1728
1729 if (tcp_rcv_state_process(sk, skb)) {
1730 rsk = sk;
1731 goto reset;
1732 }
1733 return 0;
1734
1735reset:
1736 tcp_v4_send_reset(rsk, skb);
1737discard:
1738 kfree_skb(skb);
1739 /* Be careful here. If this function gets more complicated and
1740 * gcc suffers from register pressure on the x86, sk (in %ebx)
1741 * might be destroyed here. This current version compiles correctly,
1742 * but you have been warned.
1743 */
1744 return 0;
1745
1746csum_err:
1747 trace_tcp_bad_csum(skb);
1748 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1749 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1750 goto discard;
1751}
1752EXPORT_SYMBOL(tcp_v4_do_rcv);
1753
1754int tcp_v4_early_demux(struct sk_buff *skb)
1755{
1756 const struct iphdr *iph;
1757 const struct tcphdr *th;
1758 struct sock *sk;
1759
1760 if (skb->pkt_type != PACKET_HOST)
1761 return 0;
1762
1763 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1764 return 0;
1765
1766 iph = ip_hdr(skb);
1767 th = tcp_hdr(skb);
1768
1769 if (th->doff < sizeof(struct tcphdr) / 4)
1770 return 0;
1771
1772 sk = __inet_lookup_established(dev_net(skb->dev), &tcp_hashinfo,
1773 iph->saddr, th->source,
1774 iph->daddr, ntohs(th->dest),
1775 skb->skb_iif, inet_sdif(skb));
1776 if (sk) {
1777 skb->sk = sk;
1778 skb->destructor = sock_edemux;
1779 if (sk_fullsock(sk)) {
1780 struct dst_entry *dst = READ_ONCE(sk->sk_rx_dst);
1781
1782 if (dst)
1783 dst = dst_check(dst, 0);
1784 if (dst &&
1785 inet_sk(sk)->rx_dst_ifindex == skb->skb_iif)
1786 skb_dst_set_noref(skb, dst);
1787 }
1788 }
1789 return 0;
1790}
1791
1792bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
1793{
1794 u32 limit = READ_ONCE(sk->sk_rcvbuf) + READ_ONCE(sk->sk_sndbuf);
1795 u32 tail_gso_size, tail_gso_segs;
1796 struct skb_shared_info *shinfo;
1797 const struct tcphdr *th;
1798 struct tcphdr *thtail;
1799 struct sk_buff *tail;
1800 unsigned int hdrlen;
1801 bool fragstolen;
1802 u32 gso_segs;
1803 u32 gso_size;
1804 int delta;
1805
1806 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
1807 * we can fix skb->truesize to its real value to avoid future drops.
1808 * This is valid because skb is not yet charged to the socket.
1809 * It has been noticed pure SACK packets were sometimes dropped
1810 * (if cooked by drivers without copybreak feature).
1811 */
1812 skb_condense(skb);
1813
1814 skb_dst_drop(skb);
1815
1816 if (unlikely(tcp_checksum_complete(skb))) {
1817 bh_unlock_sock(sk);
1818 trace_tcp_bad_csum(skb);
1819 __TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1820 __TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1821 return true;
1822 }
1823
1824 /* Attempt coalescing to last skb in backlog, even if we are
1825 * above the limits.
1826 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
1827 */
1828 th = (const struct tcphdr *)skb->data;
1829 hdrlen = th->doff * 4;
1830
1831 tail = sk->sk_backlog.tail;
1832 if (!tail)
1833 goto no_coalesce;
1834 thtail = (struct tcphdr *)tail->data;
1835
1836 if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
1837 TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
1838 ((TCP_SKB_CB(tail)->tcp_flags |
1839 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_SYN | TCPHDR_RST | TCPHDR_URG)) ||
1840 !((TCP_SKB_CB(tail)->tcp_flags &
1841 TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_ACK) ||
1842 ((TCP_SKB_CB(tail)->tcp_flags ^
1843 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
1844#ifdef CONFIG_TLS_DEVICE
1845 tail->decrypted != skb->decrypted ||
1846#endif
1847 thtail->doff != th->doff ||
1848 memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
1849 goto no_coalesce;
1850
1851 __skb_pull(skb, hdrlen);
1852
1853 shinfo = skb_shinfo(skb);
1854 gso_size = shinfo->gso_size ?: skb->len;
1855 gso_segs = shinfo->gso_segs ?: 1;
1856
1857 shinfo = skb_shinfo(tail);
1858 tail_gso_size = shinfo->gso_size ?: (tail->len - hdrlen);
1859 tail_gso_segs = shinfo->gso_segs ?: 1;
1860
1861 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
1862 TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
1863
1864 if (likely(!before(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))) {
1865 TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
1866 thtail->window = th->window;
1867 }
1868
1869 /* We have to update both TCP_SKB_CB(tail)->tcp_flags and
1870 * thtail->fin, so that the fast path in tcp_rcv_established()
1871 * is not entered if we append a packet with a FIN.
1872 * SYN, RST, URG are not present.
1873 * ACK is set on both packets.
1874 * PSH : we do not really care in TCP stack,
1875 * at least for 'GRO' packets.
1876 */
1877 thtail->fin |= th->fin;
1878 TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1879
1880 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1881 TCP_SKB_CB(tail)->has_rxtstamp = true;
1882 tail->tstamp = skb->tstamp;
1883 skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
1884 }
1885
1886 /* Not as strict as GRO. We only need to carry mss max value */
1887 shinfo->gso_size = max(gso_size, tail_gso_size);
1888 shinfo->gso_segs = min_t(u32, gso_segs + tail_gso_segs, 0xFFFF);
1889
1890 sk->sk_backlog.len += delta;
1891 __NET_INC_STATS(sock_net(sk),
1892 LINUX_MIB_TCPBACKLOGCOALESCE);
1893 kfree_skb_partial(skb, fragstolen);
1894 return false;
1895 }
1896 __skb_push(skb, hdrlen);
1897
1898no_coalesce:
1899 /* Only socket owner can try to collapse/prune rx queues
1900 * to reduce memory overhead, so add a little headroom here.
1901 * Few sockets backlog are possibly concurrently non empty.
1902 */
1903 limit += 64*1024;
1904
1905 if (unlikely(sk_add_backlog(sk, skb, limit))) {
1906 bh_unlock_sock(sk);
1907 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
1908 return true;
1909 }
1910 return false;
1911}
1912EXPORT_SYMBOL(tcp_add_backlog);
1913
1914int tcp_filter(struct sock *sk, struct sk_buff *skb)
1915{
1916 struct tcphdr *th = (struct tcphdr *)skb->data;
1917
1918 return sk_filter_trim_cap(sk, skb, th->doff * 4);
1919}
1920EXPORT_SYMBOL(tcp_filter);
1921
1922static void tcp_v4_restore_cb(struct sk_buff *skb)
1923{
1924 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
1925 sizeof(struct inet_skb_parm));
1926}
1927
1928static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
1929 const struct tcphdr *th)
1930{
1931 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
1932 * barrier() makes sure compiler wont play fool^Waliasing games.
1933 */
1934 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
1935 sizeof(struct inet_skb_parm));
1936 barrier();
1937
1938 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1939 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1940 skb->len - th->doff * 4);
1941 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1942 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
1943 TCP_SKB_CB(skb)->tcp_tw_isn = 0;
1944 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
1945 TCP_SKB_CB(skb)->sacked = 0;
1946 TCP_SKB_CB(skb)->has_rxtstamp =
1947 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
1948}
1949
1950/*
1951 * From tcp_input.c
1952 */
1953
1954int tcp_v4_rcv(struct sk_buff *skb)
1955{
1956 struct net *net = dev_net(skb->dev);
1957 struct sk_buff *skb_to_free;
1958 int sdif = inet_sdif(skb);
1959 int dif = inet_iif(skb);
1960 const struct iphdr *iph;
1961 const struct tcphdr *th;
1962 bool refcounted;
1963 struct sock *sk;
1964 int ret;
1965
1966 if (skb->pkt_type != PACKET_HOST)
1967 goto discard_it;
1968
1969 /* Count it even if it's bad */
1970 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
1971
1972 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1973 goto discard_it;
1974
1975 th = (const struct tcphdr *)skb->data;
1976
1977 if (unlikely(th->doff < sizeof(struct tcphdr) / 4))
1978 goto bad_packet;
1979 if (!pskb_may_pull(skb, th->doff * 4))
1980 goto discard_it;
1981
1982 /* An explanation is required here, I think.
1983 * Packet length and doff are validated by header prediction,
1984 * provided case of th->doff==0 is eliminated.
1985 * So, we defer the checks. */
1986
1987 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
1988 goto csum_error;
1989
1990 th = (const struct tcphdr *)skb->data;
1991 iph = ip_hdr(skb);
1992lookup:
1993 sk = __inet_lookup_skb(&tcp_hashinfo, skb, __tcp_hdrlen(th), th->source,
1994 th->dest, sdif, &refcounted);
1995 if (!sk)
1996 goto no_tcp_socket;
1997
1998process:
1999 if (sk->sk_state == TCP_TIME_WAIT)
2000 goto do_time_wait;
2001
2002 if (sk->sk_state == TCP_NEW_SYN_RECV) {
2003 struct request_sock *req = inet_reqsk(sk);
2004 bool req_stolen = false;
2005 struct sock *nsk;
2006
2007 sk = req->rsk_listener;
2008 if (unlikely(tcp_v4_inbound_md5_hash(sk, skb, dif, sdif))) {
2009 sk_drops_add(sk, skb);
2010 reqsk_put(req);
2011 goto discard_it;
2012 }
2013 if (tcp_checksum_complete(skb)) {
2014 reqsk_put(req);
2015 goto csum_error;
2016 }
2017 if (unlikely(sk->sk_state != TCP_LISTEN)) {
2018 nsk = reuseport_migrate_sock(sk, req_to_sk(req), skb);
2019 if (!nsk) {
2020 inet_csk_reqsk_queue_drop_and_put(sk, req);
2021 goto lookup;
2022 }
2023 sk = nsk;
2024 /* reuseport_migrate_sock() has already held one sk_refcnt
2025 * before returning.
2026 */
2027 } else {
2028 /* We own a reference on the listener, increase it again
2029 * as we might lose it too soon.
2030 */
2031 sock_hold(sk);
2032 }
2033 refcounted = true;
2034 nsk = NULL;
2035 if (!tcp_filter(sk, skb)) {
2036 th = (const struct tcphdr *)skb->data;
2037 iph = ip_hdr(skb);
2038 tcp_v4_fill_cb(skb, iph, th);
2039 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
2040 }
2041 if (!nsk) {
2042 reqsk_put(req);
2043 if (req_stolen) {
2044 /* Another cpu got exclusive access to req
2045 * and created a full blown socket.
2046 * Try to feed this packet to this socket
2047 * instead of discarding it.
2048 */
2049 tcp_v4_restore_cb(skb);
2050 sock_put(sk);
2051 goto lookup;
2052 }
2053 goto discard_and_relse;
2054 }
2055 if (nsk == sk) {
2056 reqsk_put(req);
2057 tcp_v4_restore_cb(skb);
2058 } else if (tcp_child_process(sk, nsk, skb)) {
2059 tcp_v4_send_reset(nsk, skb);
2060 goto discard_and_relse;
2061 } else {
2062 sock_put(sk);
2063 return 0;
2064 }
2065 }
2066 if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
2067 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
2068 goto discard_and_relse;
2069 }
2070
2071 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
2072 goto discard_and_relse;
2073
2074 if (tcp_v4_inbound_md5_hash(sk, skb, dif, sdif))
2075 goto discard_and_relse;
2076
2077 nf_reset_ct(skb);
2078
2079 if (tcp_filter(sk, skb))
2080 goto discard_and_relse;
2081 th = (const struct tcphdr *)skb->data;
2082 iph = ip_hdr(skb);
2083 tcp_v4_fill_cb(skb, iph, th);
2084
2085 skb->dev = NULL;
2086
2087 if (sk->sk_state == TCP_LISTEN) {
2088 ret = tcp_v4_do_rcv(sk, skb);
2089 goto put_and_return;
2090 }
2091
2092 sk_incoming_cpu_update(sk);
2093
2094 bh_lock_sock_nested(sk);
2095 tcp_segs_in(tcp_sk(sk), skb);
2096 ret = 0;
2097 if (!sock_owned_by_user(sk)) {
2098 skb_to_free = sk->sk_rx_skb_cache;
2099 sk->sk_rx_skb_cache = NULL;
2100 ret = tcp_v4_do_rcv(sk, skb);
2101 } else {
2102 if (tcp_add_backlog(sk, skb))
2103 goto discard_and_relse;
2104 skb_to_free = NULL;
2105 }
2106 bh_unlock_sock(sk);
2107 if (skb_to_free)
2108 __kfree_skb(skb_to_free);
2109
2110put_and_return:
2111 if (refcounted)
2112 sock_put(sk);
2113
2114 return ret;
2115
2116no_tcp_socket:
2117 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
2118 goto discard_it;
2119
2120 tcp_v4_fill_cb(skb, iph, th);
2121
2122 if (tcp_checksum_complete(skb)) {
2123csum_error:
2124 trace_tcp_bad_csum(skb);
2125 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
2126bad_packet:
2127 __TCP_INC_STATS(net, TCP_MIB_INERRS);
2128 } else {
2129 tcp_v4_send_reset(NULL, skb);
2130 }
2131
2132discard_it:
2133 /* Discard frame. */
2134 kfree_skb(skb);
2135 return 0;
2136
2137discard_and_relse:
2138 sk_drops_add(sk, skb);
2139 if (refcounted)
2140 sock_put(sk);
2141 goto discard_it;
2142
2143do_time_wait:
2144 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
2145 inet_twsk_put(inet_twsk(sk));
2146 goto discard_it;
2147 }
2148
2149 tcp_v4_fill_cb(skb, iph, th);
2150
2151 if (tcp_checksum_complete(skb)) {
2152 inet_twsk_put(inet_twsk(sk));
2153 goto csum_error;
2154 }
2155 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
2156 case TCP_TW_SYN: {
2157 struct sock *sk2 = inet_lookup_listener(dev_net(skb->dev),
2158 &tcp_hashinfo, skb,
2159 __tcp_hdrlen(th),
2160 iph->saddr, th->source,
2161 iph->daddr, th->dest,
2162 inet_iif(skb),
2163 sdif);
2164 if (sk2) {
2165 inet_twsk_deschedule_put(inet_twsk(sk));
2166 sk = sk2;
2167 tcp_v4_restore_cb(skb);
2168 refcounted = false;
2169 goto process;
2170 }
2171 }
2172 /* to ACK */
2173 fallthrough;
2174 case TCP_TW_ACK:
2175 tcp_v4_timewait_ack(sk, skb);
2176 break;
2177 case TCP_TW_RST:
2178 tcp_v4_send_reset(sk, skb);
2179 inet_twsk_deschedule_put(inet_twsk(sk));
2180 goto discard_it;
2181 case TCP_TW_SUCCESS:;
2182 }
2183 goto discard_it;
2184}
2185
2186static struct timewait_sock_ops tcp_timewait_sock_ops = {
2187 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
2188 .twsk_unique = tcp_twsk_unique,
2189 .twsk_destructor= tcp_twsk_destructor,
2190};
2191
2192void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
2193{
2194 struct dst_entry *dst = skb_dst(skb);
2195
2196 if (dst && dst_hold_safe(dst)) {
2197 sk->sk_rx_dst = dst;
2198 inet_sk(sk)->rx_dst_ifindex = skb->skb_iif;
2199 }
2200}
2201EXPORT_SYMBOL(inet_sk_rx_dst_set);
2202
2203const struct inet_connection_sock_af_ops ipv4_specific = {
2204 .queue_xmit = ip_queue_xmit,
2205 .send_check = tcp_v4_send_check,
2206 .rebuild_header = inet_sk_rebuild_header,
2207 .sk_rx_dst_set = inet_sk_rx_dst_set,
2208 .conn_request = tcp_v4_conn_request,
2209 .syn_recv_sock = tcp_v4_syn_recv_sock,
2210 .net_header_len = sizeof(struct iphdr),
2211 .setsockopt = ip_setsockopt,
2212 .getsockopt = ip_getsockopt,
2213 .addr2sockaddr = inet_csk_addr2sockaddr,
2214 .sockaddr_len = sizeof(struct sockaddr_in),
2215 .mtu_reduced = tcp_v4_mtu_reduced,
2216};
2217EXPORT_SYMBOL(ipv4_specific);
2218
2219#ifdef CONFIG_TCP_MD5SIG
2220static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
2221 .md5_lookup = tcp_v4_md5_lookup,
2222 .calc_md5_hash = tcp_v4_md5_hash_skb,
2223 .md5_parse = tcp_v4_parse_md5_keys,
2224};
2225#endif
2226
2227/* NOTE: A lot of things set to zero explicitly by call to
2228 * sk_alloc() so need not be done here.
2229 */
2230static int tcp_v4_init_sock(struct sock *sk)
2231{
2232 struct inet_connection_sock *icsk = inet_csk(sk);
2233
2234 tcp_init_sock(sk);
2235
2236 icsk->icsk_af_ops = &ipv4_specific;
2237
2238#ifdef CONFIG_TCP_MD5SIG
2239 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
2240#endif
2241
2242 return 0;
2243}
2244
2245void tcp_v4_destroy_sock(struct sock *sk)
2246{
2247 struct tcp_sock *tp = tcp_sk(sk);
2248
2249 trace_tcp_destroy_sock(sk);
2250
2251 tcp_clear_xmit_timers(sk);
2252
2253 tcp_cleanup_congestion_control(sk);
2254
2255 tcp_cleanup_ulp(sk);
2256
2257 /* Cleanup up the write buffer. */
2258 tcp_write_queue_purge(sk);
2259
2260 /* Check if we want to disable active TFO */
2261 tcp_fastopen_active_disable_ofo_check(sk);
2262
2263 /* Cleans up our, hopefully empty, out_of_order_queue. */
2264 skb_rbtree_purge(&tp->out_of_order_queue);
2265
2266#ifdef CONFIG_TCP_MD5SIG
2267 /* Clean up the MD5 key list, if any */
2268 if (tp->md5sig_info) {
2269 tcp_clear_md5_list(sk);
2270 kfree_rcu(rcu_dereference_protected(tp->md5sig_info, 1), rcu);
2271 tp->md5sig_info = NULL;
2272 }
2273#endif
2274
2275 /* Clean up a referenced TCP bind bucket. */
2276 if (inet_csk(sk)->icsk_bind_hash)
2277 inet_put_port(sk);
2278
2279 BUG_ON(rcu_access_pointer(tp->fastopen_rsk));
2280
2281 /* If socket is aborted during connect operation */
2282 tcp_free_fastopen_req(tp);
2283 tcp_fastopen_destroy_cipher(sk);
2284 tcp_saved_syn_free(tp);
2285
2286 sk_sockets_allocated_dec(sk);
2287}
2288EXPORT_SYMBOL(tcp_v4_destroy_sock);
2289
2290#ifdef CONFIG_PROC_FS
2291/* Proc filesystem TCP sock list dumping. */
2292
2293/*
2294 * Get next listener socket follow cur. If cur is NULL, get first socket
2295 * starting from bucket given in st->bucket; when st->bucket is zero the
2296 * very first socket in the hash table is returned.
2297 */
2298static void *listening_get_next(struct seq_file *seq, void *cur)
2299{
2300 struct tcp_seq_afinfo *afinfo;
2301 struct tcp_iter_state *st = seq->private;
2302 struct net *net = seq_file_net(seq);
2303 struct inet_listen_hashbucket *ilb;
2304 struct hlist_nulls_node *node;
2305 struct sock *sk = cur;
2306
2307 if (st->bpf_seq_afinfo)
2308 afinfo = st->bpf_seq_afinfo;
2309 else
2310 afinfo = PDE_DATA(file_inode(seq->file));
2311
2312 if (!sk) {
2313get_head:
2314 ilb = &tcp_hashinfo.listening_hash[st->bucket];
2315 spin_lock(&ilb->lock);
2316 sk = sk_nulls_head(&ilb->nulls_head);
2317 st->offset = 0;
2318 goto get_sk;
2319 }
2320 ilb = &tcp_hashinfo.listening_hash[st->bucket];
2321 ++st->num;
2322 ++st->offset;
2323
2324 sk = sk_nulls_next(sk);
2325get_sk:
2326 sk_nulls_for_each_from(sk, node) {
2327 if (!net_eq(sock_net(sk), net))
2328 continue;
2329 if (afinfo->family == AF_UNSPEC ||
2330 sk->sk_family == afinfo->family)
2331 return sk;
2332 }
2333 spin_unlock(&ilb->lock);
2334 st->offset = 0;
2335 if (++st->bucket < INET_LHTABLE_SIZE)
2336 goto get_head;
2337 return NULL;
2338}
2339
2340static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2341{
2342 struct tcp_iter_state *st = seq->private;
2343 void *rc;
2344
2345 st->bucket = 0;
2346 st->offset = 0;
2347 rc = listening_get_next(seq, NULL);
2348
2349 while (rc && *pos) {
2350 rc = listening_get_next(seq, rc);
2351 --*pos;
2352 }
2353 return rc;
2354}
2355
2356static inline bool empty_bucket(const struct tcp_iter_state *st)
2357{
2358 return hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].chain);
2359}
2360
2361/*
2362 * Get first established socket starting from bucket given in st->bucket.
2363 * If st->bucket is zero, the very first socket in the hash is returned.
2364 */
2365static void *established_get_first(struct seq_file *seq)
2366{
2367 struct tcp_seq_afinfo *afinfo;
2368 struct tcp_iter_state *st = seq->private;
2369 struct net *net = seq_file_net(seq);
2370 void *rc = NULL;
2371
2372 if (st->bpf_seq_afinfo)
2373 afinfo = st->bpf_seq_afinfo;
2374 else
2375 afinfo = PDE_DATA(file_inode(seq->file));
2376
2377 st->offset = 0;
2378 for (; st->bucket <= tcp_hashinfo.ehash_mask; ++st->bucket) {
2379 struct sock *sk;
2380 struct hlist_nulls_node *node;
2381 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, st->bucket);
2382
2383 /* Lockless fast path for the common case of empty buckets */
2384 if (empty_bucket(st))
2385 continue;
2386
2387 spin_lock_bh(lock);
2388 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
2389 if ((afinfo->family != AF_UNSPEC &&
2390 sk->sk_family != afinfo->family) ||
2391 !net_eq(sock_net(sk), net)) {
2392 continue;
2393 }
2394 rc = sk;
2395 goto out;
2396 }
2397 spin_unlock_bh(lock);
2398 }
2399out:
2400 return rc;
2401}
2402
2403static void *established_get_next(struct seq_file *seq, void *cur)
2404{
2405 struct tcp_seq_afinfo *afinfo;
2406 struct sock *sk = cur;
2407 struct hlist_nulls_node *node;
2408 struct tcp_iter_state *st = seq->private;
2409 struct net *net = seq_file_net(seq);
2410
2411 if (st->bpf_seq_afinfo)
2412 afinfo = st->bpf_seq_afinfo;
2413 else
2414 afinfo = PDE_DATA(file_inode(seq->file));
2415
2416 ++st->num;
2417 ++st->offset;
2418
2419 sk = sk_nulls_next(sk);
2420
2421 sk_nulls_for_each_from(sk, node) {
2422 if ((afinfo->family == AF_UNSPEC ||
2423 sk->sk_family == afinfo->family) &&
2424 net_eq(sock_net(sk), net))
2425 return sk;
2426 }
2427
2428 spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
2429 ++st->bucket;
2430 return established_get_first(seq);
2431}
2432
2433static void *established_get_idx(struct seq_file *seq, loff_t pos)
2434{
2435 struct tcp_iter_state *st = seq->private;
2436 void *rc;
2437
2438 st->bucket = 0;
2439 rc = established_get_first(seq);
2440
2441 while (rc && pos) {
2442 rc = established_get_next(seq, rc);
2443 --pos;
2444 }
2445 return rc;
2446}
2447
2448static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2449{
2450 void *rc;
2451 struct tcp_iter_state *st = seq->private;
2452
2453 st->state = TCP_SEQ_STATE_LISTENING;
2454 rc = listening_get_idx(seq, &pos);
2455
2456 if (!rc) {
2457 st->state = TCP_SEQ_STATE_ESTABLISHED;
2458 rc = established_get_idx(seq, pos);
2459 }
2460
2461 return rc;
2462}
2463
2464static void *tcp_seek_last_pos(struct seq_file *seq)
2465{
2466 struct tcp_iter_state *st = seq->private;
2467 int bucket = st->bucket;
2468 int offset = st->offset;
2469 int orig_num = st->num;
2470 void *rc = NULL;
2471
2472 switch (st->state) {
2473 case TCP_SEQ_STATE_LISTENING:
2474 if (st->bucket >= INET_LHTABLE_SIZE)
2475 break;
2476 st->state = TCP_SEQ_STATE_LISTENING;
2477 rc = listening_get_next(seq, NULL);
2478 while (offset-- && rc && bucket == st->bucket)
2479 rc = listening_get_next(seq, rc);
2480 if (rc)
2481 break;
2482 st->bucket = 0;
2483 st->state = TCP_SEQ_STATE_ESTABLISHED;
2484 fallthrough;
2485 case TCP_SEQ_STATE_ESTABLISHED:
2486 if (st->bucket > tcp_hashinfo.ehash_mask)
2487 break;
2488 rc = established_get_first(seq);
2489 while (offset-- && rc && bucket == st->bucket)
2490 rc = established_get_next(seq, rc);
2491 }
2492
2493 st->num = orig_num;
2494
2495 return rc;
2496}
2497
2498void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2499{
2500 struct tcp_iter_state *st = seq->private;
2501 void *rc;
2502
2503 if (*pos && *pos == st->last_pos) {
2504 rc = tcp_seek_last_pos(seq);
2505 if (rc)
2506 goto out;
2507 }
2508
2509 st->state = TCP_SEQ_STATE_LISTENING;
2510 st->num = 0;
2511 st->bucket = 0;
2512 st->offset = 0;
2513 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2514
2515out:
2516 st->last_pos = *pos;
2517 return rc;
2518}
2519EXPORT_SYMBOL(tcp_seq_start);
2520
2521void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2522{
2523 struct tcp_iter_state *st = seq->private;
2524 void *rc = NULL;
2525
2526 if (v == SEQ_START_TOKEN) {
2527 rc = tcp_get_idx(seq, 0);
2528 goto out;
2529 }
2530
2531 switch (st->state) {
2532 case TCP_SEQ_STATE_LISTENING:
2533 rc = listening_get_next(seq, v);
2534 if (!rc) {
2535 st->state = TCP_SEQ_STATE_ESTABLISHED;
2536 st->bucket = 0;
2537 st->offset = 0;
2538 rc = established_get_first(seq);
2539 }
2540 break;
2541 case TCP_SEQ_STATE_ESTABLISHED:
2542 rc = established_get_next(seq, v);
2543 break;
2544 }
2545out:
2546 ++*pos;
2547 st->last_pos = *pos;
2548 return rc;
2549}
2550EXPORT_SYMBOL(tcp_seq_next);
2551
2552void tcp_seq_stop(struct seq_file *seq, void *v)
2553{
2554 struct tcp_iter_state *st = seq->private;
2555
2556 switch (st->state) {
2557 case TCP_SEQ_STATE_LISTENING:
2558 if (v != SEQ_START_TOKEN)
2559 spin_unlock(&tcp_hashinfo.listening_hash[st->bucket].lock);
2560 break;
2561 case TCP_SEQ_STATE_ESTABLISHED:
2562 if (v)
2563 spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
2564 break;
2565 }
2566}
2567EXPORT_SYMBOL(tcp_seq_stop);
2568
2569static void get_openreq4(const struct request_sock *req,
2570 struct seq_file *f, int i)
2571{
2572 const struct inet_request_sock *ireq = inet_rsk(req);
2573 long delta = req->rsk_timer.expires - jiffies;
2574
2575 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2576 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2577 i,
2578 ireq->ir_loc_addr,
2579 ireq->ir_num,
2580 ireq->ir_rmt_addr,
2581 ntohs(ireq->ir_rmt_port),
2582 TCP_SYN_RECV,
2583 0, 0, /* could print option size, but that is af dependent. */
2584 1, /* timers active (only the expire timer) */
2585 jiffies_delta_to_clock_t(delta),
2586 req->num_timeout,
2587 from_kuid_munged(seq_user_ns(f),
2588 sock_i_uid(req->rsk_listener)),
2589 0, /* non standard timer */
2590 0, /* open_requests have no inode */
2591 0,
2592 req);
2593}
2594
2595static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2596{
2597 int timer_active;
2598 unsigned long timer_expires;
2599 const struct tcp_sock *tp = tcp_sk(sk);
2600 const struct inet_connection_sock *icsk = inet_csk(sk);
2601 const struct inet_sock *inet = inet_sk(sk);
2602 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2603 __be32 dest = inet->inet_daddr;
2604 __be32 src = inet->inet_rcv_saddr;
2605 __u16 destp = ntohs(inet->inet_dport);
2606 __u16 srcp = ntohs(inet->inet_sport);
2607 int rx_queue;
2608 int state;
2609
2610 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2611 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2612 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2613 timer_active = 1;
2614 timer_expires = icsk->icsk_timeout;
2615 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2616 timer_active = 4;
2617 timer_expires = icsk->icsk_timeout;
2618 } else if (timer_pending(&sk->sk_timer)) {
2619 timer_active = 2;
2620 timer_expires = sk->sk_timer.expires;
2621 } else {
2622 timer_active = 0;
2623 timer_expires = jiffies;
2624 }
2625
2626 state = inet_sk_state_load(sk);
2627 if (state == TCP_LISTEN)
2628 rx_queue = READ_ONCE(sk->sk_ack_backlog);
2629 else
2630 /* Because we don't lock the socket,
2631 * we might find a transient negative value.
2632 */
2633 rx_queue = max_t(int, READ_ONCE(tp->rcv_nxt) -
2634 READ_ONCE(tp->copied_seq), 0);
2635
2636 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2637 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2638 i, src, srcp, dest, destp, state,
2639 READ_ONCE(tp->write_seq) - tp->snd_una,
2640 rx_queue,
2641 timer_active,
2642 jiffies_delta_to_clock_t(timer_expires - jiffies),
2643 icsk->icsk_retransmits,
2644 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2645 icsk->icsk_probes_out,
2646 sock_i_ino(sk),
2647 refcount_read(&sk->sk_refcnt), sk,
2648 jiffies_to_clock_t(icsk->icsk_rto),
2649 jiffies_to_clock_t(icsk->icsk_ack.ato),
2650 (icsk->icsk_ack.quick << 1) | inet_csk_in_pingpong_mode(sk),
2651 tp->snd_cwnd,
2652 state == TCP_LISTEN ?
2653 fastopenq->max_qlen :
2654 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2655}
2656
2657static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2658 struct seq_file *f, int i)
2659{
2660 long delta = tw->tw_timer.expires - jiffies;
2661 __be32 dest, src;
2662 __u16 destp, srcp;
2663
2664 dest = tw->tw_daddr;
2665 src = tw->tw_rcv_saddr;
2666 destp = ntohs(tw->tw_dport);
2667 srcp = ntohs(tw->tw_sport);
2668
2669 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2670 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2671 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2672 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2673 refcount_read(&tw->tw_refcnt), tw);
2674}
2675
2676#define TMPSZ 150
2677
2678static int tcp4_seq_show(struct seq_file *seq, void *v)
2679{
2680 struct tcp_iter_state *st;
2681 struct sock *sk = v;
2682
2683 seq_setwidth(seq, TMPSZ - 1);
2684 if (v == SEQ_START_TOKEN) {
2685 seq_puts(seq, " sl local_address rem_address st tx_queue "
2686 "rx_queue tr tm->when retrnsmt uid timeout "
2687 "inode");
2688 goto out;
2689 }
2690 st = seq->private;
2691
2692 if (sk->sk_state == TCP_TIME_WAIT)
2693 get_timewait4_sock(v, seq, st->num);
2694 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2695 get_openreq4(v, seq, st->num);
2696 else
2697 get_tcp4_sock(v, seq, st->num);
2698out:
2699 seq_pad(seq, '\n');
2700 return 0;
2701}
2702
2703#ifdef CONFIG_BPF_SYSCALL
2704struct bpf_iter__tcp {
2705 __bpf_md_ptr(struct bpf_iter_meta *, meta);
2706 __bpf_md_ptr(struct sock_common *, sk_common);
2707 uid_t uid __aligned(8);
2708};
2709
2710static int tcp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta,
2711 struct sock_common *sk_common, uid_t uid)
2712{
2713 struct bpf_iter__tcp ctx;
2714
2715 meta->seq_num--; /* skip SEQ_START_TOKEN */
2716 ctx.meta = meta;
2717 ctx.sk_common = sk_common;
2718 ctx.uid = uid;
2719 return bpf_iter_run_prog(prog, &ctx);
2720}
2721
2722static int bpf_iter_tcp_seq_show(struct seq_file *seq, void *v)
2723{
2724 struct bpf_iter_meta meta;
2725 struct bpf_prog *prog;
2726 struct sock *sk = v;
2727 uid_t uid;
2728
2729 if (v == SEQ_START_TOKEN)
2730 return 0;
2731
2732 if (sk->sk_state == TCP_TIME_WAIT) {
2733 uid = 0;
2734 } else if (sk->sk_state == TCP_NEW_SYN_RECV) {
2735 const struct request_sock *req = v;
2736
2737 uid = from_kuid_munged(seq_user_ns(seq),
2738 sock_i_uid(req->rsk_listener));
2739 } else {
2740 uid = from_kuid_munged(seq_user_ns(seq), sock_i_uid(sk));
2741 }
2742
2743 meta.seq = seq;
2744 prog = bpf_iter_get_info(&meta, false);
2745 return tcp_prog_seq_show(prog, &meta, v, uid);
2746}
2747
2748static void bpf_iter_tcp_seq_stop(struct seq_file *seq, void *v)
2749{
2750 struct bpf_iter_meta meta;
2751 struct bpf_prog *prog;
2752
2753 if (!v) {
2754 meta.seq = seq;
2755 prog = bpf_iter_get_info(&meta, true);
2756 if (prog)
2757 (void)tcp_prog_seq_show(prog, &meta, v, 0);
2758 }
2759
2760 tcp_seq_stop(seq, v);
2761}
2762
2763static const struct seq_operations bpf_iter_tcp_seq_ops = {
2764 .show = bpf_iter_tcp_seq_show,
2765 .start = tcp_seq_start,
2766 .next = tcp_seq_next,
2767 .stop = bpf_iter_tcp_seq_stop,
2768};
2769#endif
2770
2771static const struct seq_operations tcp4_seq_ops = {
2772 .show = tcp4_seq_show,
2773 .start = tcp_seq_start,
2774 .next = tcp_seq_next,
2775 .stop = tcp_seq_stop,
2776};
2777
2778static struct tcp_seq_afinfo tcp4_seq_afinfo = {
2779 .family = AF_INET,
2780};
2781
2782static int __net_init tcp4_proc_init_net(struct net *net)
2783{
2784 if (!proc_create_net_data("tcp", 0444, net->proc_net, &tcp4_seq_ops,
2785 sizeof(struct tcp_iter_state), &tcp4_seq_afinfo))
2786 return -ENOMEM;
2787 return 0;
2788}
2789
2790static void __net_exit tcp4_proc_exit_net(struct net *net)
2791{
2792 remove_proc_entry("tcp", net->proc_net);
2793}
2794
2795static struct pernet_operations tcp4_net_ops = {
2796 .init = tcp4_proc_init_net,
2797 .exit = tcp4_proc_exit_net,
2798};
2799
2800int __init tcp4_proc_init(void)
2801{
2802 return register_pernet_subsys(&tcp4_net_ops);
2803}
2804
2805void tcp4_proc_exit(void)
2806{
2807 unregister_pernet_subsys(&tcp4_net_ops);
2808}
2809#endif /* CONFIG_PROC_FS */
2810
2811/* @wake is one when sk_stream_write_space() calls us.
2812 * This sends EPOLLOUT only if notsent_bytes is half the limit.
2813 * This mimics the strategy used in sock_def_write_space().
2814 */
2815bool tcp_stream_memory_free(const struct sock *sk, int wake)
2816{
2817 const struct tcp_sock *tp = tcp_sk(sk);
2818 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
2819 READ_ONCE(tp->snd_nxt);
2820
2821 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
2822}
2823EXPORT_SYMBOL(tcp_stream_memory_free);
2824
2825struct proto tcp_prot = {
2826 .name = "TCP",
2827 .owner = THIS_MODULE,
2828 .close = tcp_close,
2829 .pre_connect = tcp_v4_pre_connect,
2830 .connect = tcp_v4_connect,
2831 .disconnect = tcp_disconnect,
2832 .accept = inet_csk_accept,
2833 .ioctl = tcp_ioctl,
2834 .init = tcp_v4_init_sock,
2835 .destroy = tcp_v4_destroy_sock,
2836 .shutdown = tcp_shutdown,
2837 .setsockopt = tcp_setsockopt,
2838 .getsockopt = tcp_getsockopt,
2839 .bpf_bypass_getsockopt = tcp_bpf_bypass_getsockopt,
2840 .keepalive = tcp_set_keepalive,
2841 .recvmsg = tcp_recvmsg,
2842 .sendmsg = tcp_sendmsg,
2843 .sendpage = tcp_sendpage,
2844 .backlog_rcv = tcp_v4_do_rcv,
2845 .release_cb = tcp_release_cb,
2846 .hash = inet_hash,
2847 .unhash = inet_unhash,
2848 .get_port = inet_csk_get_port,
2849#ifdef CONFIG_BPF_SYSCALL
2850 .psock_update_sk_prot = tcp_bpf_update_proto,
2851#endif
2852 .enter_memory_pressure = tcp_enter_memory_pressure,
2853 .leave_memory_pressure = tcp_leave_memory_pressure,
2854 .stream_memory_free = tcp_stream_memory_free,
2855 .sockets_allocated = &tcp_sockets_allocated,
2856 .orphan_count = &tcp_orphan_count,
2857 .memory_allocated = &tcp_memory_allocated,
2858 .memory_pressure = &tcp_memory_pressure,
2859 .sysctl_mem = sysctl_tcp_mem,
2860 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2861 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2862 .max_header = MAX_TCP_HEADER,
2863 .obj_size = sizeof(struct tcp_sock),
2864 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2865 .twsk_prot = &tcp_timewait_sock_ops,
2866 .rsk_prot = &tcp_request_sock_ops,
2867 .h.hashinfo = &tcp_hashinfo,
2868 .no_autobind = true,
2869 .diag_destroy = tcp_abort,
2870};
2871EXPORT_SYMBOL(tcp_prot);
2872
2873static void __net_exit tcp_sk_exit(struct net *net)
2874{
2875 int cpu;
2876
2877 if (net->ipv4.tcp_congestion_control)
2878 bpf_module_put(net->ipv4.tcp_congestion_control,
2879 net->ipv4.tcp_congestion_control->owner);
2880
2881 for_each_possible_cpu(cpu)
2882 inet_ctl_sock_destroy(*per_cpu_ptr(net->ipv4.tcp_sk, cpu));
2883 free_percpu(net->ipv4.tcp_sk);
2884}
2885
2886static int __net_init tcp_sk_init(struct net *net)
2887{
2888 int res, cpu, cnt;
2889
2890 net->ipv4.tcp_sk = alloc_percpu(struct sock *);
2891 if (!net->ipv4.tcp_sk)
2892 return -ENOMEM;
2893
2894 for_each_possible_cpu(cpu) {
2895 struct sock *sk;
2896
2897 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
2898 IPPROTO_TCP, net);
2899 if (res)
2900 goto fail;
2901 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
2902
2903 /* Please enforce IP_DF and IPID==0 for RST and
2904 * ACK sent in SYN-RECV and TIME-WAIT state.
2905 */
2906 inet_sk(sk)->pmtudisc = IP_PMTUDISC_DO;
2907
2908 *per_cpu_ptr(net->ipv4.tcp_sk, cpu) = sk;
2909 }
2910
2911 net->ipv4.sysctl_tcp_ecn = 2;
2912 net->ipv4.sysctl_tcp_ecn_fallback = 1;
2913
2914 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
2915 net->ipv4.sysctl_tcp_min_snd_mss = TCP_MIN_SND_MSS;
2916 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
2917 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
2918 net->ipv4.sysctl_tcp_mtu_probe_floor = TCP_MIN_SND_MSS;
2919
2920 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
2921 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
2922 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
2923
2924 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
2925 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
2926 net->ipv4.sysctl_tcp_syncookies = 1;
2927 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
2928 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
2929 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
2930 net->ipv4.sysctl_tcp_orphan_retries = 0;
2931 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
2932 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
2933 net->ipv4.sysctl_tcp_tw_reuse = 2;
2934 net->ipv4.sysctl_tcp_no_ssthresh_metrics_save = 1;
2935
2936 cnt = tcp_hashinfo.ehash_mask + 1;
2937 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
2938 net->ipv4.tcp_death_row.hashinfo = &tcp_hashinfo;
2939
2940 net->ipv4.sysctl_max_syn_backlog = max(128, cnt / 128);
2941 net->ipv4.sysctl_tcp_sack = 1;
2942 net->ipv4.sysctl_tcp_window_scaling = 1;
2943 net->ipv4.sysctl_tcp_timestamps = 1;
2944 net->ipv4.sysctl_tcp_early_retrans = 3;
2945 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
2946 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
2947 net->ipv4.sysctl_tcp_retrans_collapse = 1;
2948 net->ipv4.sysctl_tcp_max_reordering = 300;
2949 net->ipv4.sysctl_tcp_dsack = 1;
2950 net->ipv4.sysctl_tcp_app_win = 31;
2951 net->ipv4.sysctl_tcp_adv_win_scale = 1;
2952 net->ipv4.sysctl_tcp_frto = 2;
2953 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
2954 /* This limits the percentage of the congestion window which we
2955 * will allow a single TSO frame to consume. Building TSO frames
2956 * which are too large can cause TCP streams to be bursty.
2957 */
2958 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
2959 /* Default TSQ limit of 16 TSO segments */
2960 net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
2961 /* rfc5961 challenge ack rate limiting */
2962 net->ipv4.sysctl_tcp_challenge_ack_limit = 1000;
2963 net->ipv4.sysctl_tcp_min_tso_segs = 2;
2964 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
2965 net->ipv4.sysctl_tcp_autocorking = 1;
2966 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
2967 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
2968 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
2969 if (net != &init_net) {
2970 memcpy(net->ipv4.sysctl_tcp_rmem,
2971 init_net.ipv4.sysctl_tcp_rmem,
2972 sizeof(init_net.ipv4.sysctl_tcp_rmem));
2973 memcpy(net->ipv4.sysctl_tcp_wmem,
2974 init_net.ipv4.sysctl_tcp_wmem,
2975 sizeof(init_net.ipv4.sysctl_tcp_wmem));
2976 }
2977 net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
2978 net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
2979 net->ipv4.sysctl_tcp_comp_sack_nr = 44;
2980 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
2981 spin_lock_init(&net->ipv4.tcp_fastopen_ctx_lock);
2982 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
2983 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
2984
2985 /* Reno is always built in */
2986 if (!net_eq(net, &init_net) &&
2987 bpf_try_module_get(init_net.ipv4.tcp_congestion_control,
2988 init_net.ipv4.tcp_congestion_control->owner))
2989 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
2990 else
2991 net->ipv4.tcp_congestion_control = &tcp_reno;
2992
2993 return 0;
2994fail:
2995 tcp_sk_exit(net);
2996
2997 return res;
2998}
2999
3000static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
3001{
3002 struct net *net;
3003
3004 inet_twsk_purge(&tcp_hashinfo, AF_INET);
3005
3006 list_for_each_entry(net, net_exit_list, exit_list)
3007 tcp_fastopen_ctx_destroy(net);
3008}
3009
3010static struct pernet_operations __net_initdata tcp_sk_ops = {
3011 .init = tcp_sk_init,
3012 .exit = tcp_sk_exit,
3013 .exit_batch = tcp_sk_exit_batch,
3014};
3015
3016#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3017DEFINE_BPF_ITER_FUNC(tcp, struct bpf_iter_meta *meta,
3018 struct sock_common *sk_common, uid_t uid)
3019
3020static int bpf_iter_init_tcp(void *priv_data, struct bpf_iter_aux_info *aux)
3021{
3022 struct tcp_iter_state *st = priv_data;
3023 struct tcp_seq_afinfo *afinfo;
3024 int ret;
3025
3026 afinfo = kmalloc(sizeof(*afinfo), GFP_USER | __GFP_NOWARN);
3027 if (!afinfo)
3028 return -ENOMEM;
3029
3030 afinfo->family = AF_UNSPEC;
3031 st->bpf_seq_afinfo = afinfo;
3032 ret = bpf_iter_init_seq_net(priv_data, aux);
3033 if (ret)
3034 kfree(afinfo);
3035 return ret;
3036}
3037
3038static void bpf_iter_fini_tcp(void *priv_data)
3039{
3040 struct tcp_iter_state *st = priv_data;
3041
3042 kfree(st->bpf_seq_afinfo);
3043 bpf_iter_fini_seq_net(priv_data);
3044}
3045
3046static const struct bpf_iter_seq_info tcp_seq_info = {
3047 .seq_ops = &bpf_iter_tcp_seq_ops,
3048 .init_seq_private = bpf_iter_init_tcp,
3049 .fini_seq_private = bpf_iter_fini_tcp,
3050 .seq_priv_size = sizeof(struct tcp_iter_state),
3051};
3052
3053static struct bpf_iter_reg tcp_reg_info = {
3054 .target = "tcp",
3055 .ctx_arg_info_size = 1,
3056 .ctx_arg_info = {
3057 { offsetof(struct bpf_iter__tcp, sk_common),
3058 PTR_TO_BTF_ID_OR_NULL },
3059 },
3060 .seq_info = &tcp_seq_info,
3061};
3062
3063static void __init bpf_iter_register(void)
3064{
3065 tcp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON];
3066 if (bpf_iter_reg_target(&tcp_reg_info))
3067 pr_warn("Warning: could not register bpf iterator tcp\n");
3068}
3069
3070#endif
3071
3072void __init tcp_v4_init(void)
3073{
3074 if (register_pernet_subsys(&tcp_sk_ops))
3075 panic("Failed to create the TCP control socket.\n");
3076
3077#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3078 bpf_iter_register();
3079#endif
3080}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * IPv4 specific functions
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
19/*
20 * Changes:
21 * David S. Miller : New socket lookup architecture.
22 * This code is dedicated to John Dyson.
23 * David S. Miller : Change semantics of established hash,
24 * half is devoted to TIME_WAIT sockets
25 * and the rest go in the other half.
26 * Andi Kleen : Add support for syncookies and fixed
27 * some bugs: ip options weren't passed to
28 * the TCP layer, missed a check for an
29 * ACK bit.
30 * Andi Kleen : Implemented fast path mtu discovery.
31 * Fixed many serious bugs in the
32 * request_sock handling and moved
33 * most of it into the af independent code.
34 * Added tail drop and some other bugfixes.
35 * Added new listen semantics.
36 * Mike McLagan : Routing by source
37 * Juan Jose Ciarlante: ip_dynaddr bits
38 * Andi Kleen: various fixes.
39 * Vitaly E. Lavrov : Transparent proxy revived after year
40 * coma.
41 * Andi Kleen : Fix new listen.
42 * Andi Kleen : Fix accept error reporting.
43 * YOSHIFUJI Hideaki @USAGI and: Support IPV6_V6ONLY socket option, which
44 * Alexey Kuznetsov allow both IPv4 and IPv6 sockets to bind
45 * a single port at the same time.
46 */
47
48#define pr_fmt(fmt) "TCP: " fmt
49
50#include <linux/bottom_half.h>
51#include <linux/types.h>
52#include <linux/fcntl.h>
53#include <linux/module.h>
54#include <linux/random.h>
55#include <linux/cache.h>
56#include <linux/jhash.h>
57#include <linux/init.h>
58#include <linux/times.h>
59#include <linux/slab.h>
60#include <linux/sched.h>
61
62#include <net/net_namespace.h>
63#include <net/icmp.h>
64#include <net/inet_hashtables.h>
65#include <net/tcp.h>
66#include <net/transp_v6.h>
67#include <net/ipv6.h>
68#include <net/inet_common.h>
69#include <net/timewait_sock.h>
70#include <net/xfrm.h>
71#include <net/secure_seq.h>
72#include <net/busy_poll.h>
73
74#include <linux/inet.h>
75#include <linux/ipv6.h>
76#include <linux/stddef.h>
77#include <linux/proc_fs.h>
78#include <linux/seq_file.h>
79#include <linux/inetdevice.h>
80#include <linux/btf_ids.h>
81
82#include <crypto/hash.h>
83#include <linux/scatterlist.h>
84
85#include <trace/events/tcp.h>
86
87#ifdef CONFIG_TCP_MD5SIG
88static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
89 __be32 daddr, __be32 saddr, const struct tcphdr *th);
90#endif
91
92struct inet_hashinfo tcp_hashinfo;
93EXPORT_SYMBOL(tcp_hashinfo);
94
95static DEFINE_PER_CPU(struct sock *, ipv4_tcp_sk);
96
97static u32 tcp_v4_init_seq(const struct sk_buff *skb)
98{
99 return secure_tcp_seq(ip_hdr(skb)->daddr,
100 ip_hdr(skb)->saddr,
101 tcp_hdr(skb)->dest,
102 tcp_hdr(skb)->source);
103}
104
105static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
106{
107 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
108}
109
110int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
111{
112 int reuse = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tw_reuse);
113 const struct inet_timewait_sock *tw = inet_twsk(sktw);
114 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
115 struct tcp_sock *tp = tcp_sk(sk);
116
117 if (reuse == 2) {
118 /* Still does not detect *everything* that goes through
119 * lo, since we require a loopback src or dst address
120 * or direct binding to 'lo' interface.
121 */
122 bool loopback = false;
123 if (tw->tw_bound_dev_if == LOOPBACK_IFINDEX)
124 loopback = true;
125#if IS_ENABLED(CONFIG_IPV6)
126 if (tw->tw_family == AF_INET6) {
127 if (ipv6_addr_loopback(&tw->tw_v6_daddr) ||
128 ipv6_addr_v4mapped_loopback(&tw->tw_v6_daddr) ||
129 ipv6_addr_loopback(&tw->tw_v6_rcv_saddr) ||
130 ipv6_addr_v4mapped_loopback(&tw->tw_v6_rcv_saddr))
131 loopback = true;
132 } else
133#endif
134 {
135 if (ipv4_is_loopback(tw->tw_daddr) ||
136 ipv4_is_loopback(tw->tw_rcv_saddr))
137 loopback = true;
138 }
139 if (!loopback)
140 reuse = 0;
141 }
142
143 /* With PAWS, it is safe from the viewpoint
144 of data integrity. Even without PAWS it is safe provided sequence
145 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
146
147 Actually, the idea is close to VJ's one, only timestamp cache is
148 held not per host, but per port pair and TW bucket is used as state
149 holder.
150
151 If TW bucket has been already destroyed we fall back to VJ's scheme
152 and use initial timestamp retrieved from peer table.
153 */
154 if (tcptw->tw_ts_recent_stamp &&
155 (!twp || (reuse && time_after32(ktime_get_seconds(),
156 tcptw->tw_ts_recent_stamp)))) {
157 /* inet_twsk_hashdance() sets sk_refcnt after putting twsk
158 * and releasing the bucket lock.
159 */
160 if (unlikely(!refcount_inc_not_zero(&sktw->sk_refcnt)))
161 return 0;
162
163 /* In case of repair and re-using TIME-WAIT sockets we still
164 * want to be sure that it is safe as above but honor the
165 * sequence numbers and time stamps set as part of the repair
166 * process.
167 *
168 * Without this check re-using a TIME-WAIT socket with TCP
169 * repair would accumulate a -1 on the repair assigned
170 * sequence number. The first time it is reused the sequence
171 * is -1, the second time -2, etc. This fixes that issue
172 * without appearing to create any others.
173 */
174 if (likely(!tp->repair)) {
175 u32 seq = tcptw->tw_snd_nxt + 65535 + 2;
176
177 if (!seq)
178 seq = 1;
179 WRITE_ONCE(tp->write_seq, seq);
180 tp->rx_opt.ts_recent = tcptw->tw_ts_recent;
181 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
182 }
183
184 return 1;
185 }
186
187 return 0;
188}
189EXPORT_SYMBOL_GPL(tcp_twsk_unique);
190
191static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
192 int addr_len)
193{
194 /* This check is replicated from tcp_v4_connect() and intended to
195 * prevent BPF program called below from accessing bytes that are out
196 * of the bound specified by user in addr_len.
197 */
198 if (addr_len < sizeof(struct sockaddr_in))
199 return -EINVAL;
200
201 sock_owned_by_me(sk);
202
203 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr, &addr_len);
204}
205
206/* This will initiate an outgoing connection. */
207int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
208{
209 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
210 struct inet_timewait_death_row *tcp_death_row;
211 struct inet_sock *inet = inet_sk(sk);
212 struct tcp_sock *tp = tcp_sk(sk);
213 struct ip_options_rcu *inet_opt;
214 struct net *net = sock_net(sk);
215 __be16 orig_sport, orig_dport;
216 __be32 daddr, nexthop;
217 struct flowi4 *fl4;
218 struct rtable *rt;
219 int err;
220
221 if (addr_len < sizeof(struct sockaddr_in))
222 return -EINVAL;
223
224 if (usin->sin_family != AF_INET)
225 return -EAFNOSUPPORT;
226
227 nexthop = daddr = usin->sin_addr.s_addr;
228 inet_opt = rcu_dereference_protected(inet->inet_opt,
229 lockdep_sock_is_held(sk));
230 if (inet_opt && inet_opt->opt.srr) {
231 if (!daddr)
232 return -EINVAL;
233 nexthop = inet_opt->opt.faddr;
234 }
235
236 orig_sport = inet->inet_sport;
237 orig_dport = usin->sin_port;
238 fl4 = &inet->cork.fl.u.ip4;
239 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
240 sk->sk_bound_dev_if, IPPROTO_TCP, orig_sport,
241 orig_dport, sk);
242 if (IS_ERR(rt)) {
243 err = PTR_ERR(rt);
244 if (err == -ENETUNREACH)
245 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
246 return err;
247 }
248
249 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
250 ip_rt_put(rt);
251 return -ENETUNREACH;
252 }
253
254 if (!inet_opt || !inet_opt->opt.srr)
255 daddr = fl4->daddr;
256
257 tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
258
259 if (!inet->inet_saddr) {
260 err = inet_bhash2_update_saddr(sk, &fl4->saddr, AF_INET);
261 if (err) {
262 ip_rt_put(rt);
263 return err;
264 }
265 } else {
266 sk_rcv_saddr_set(sk, inet->inet_saddr);
267 }
268
269 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
270 /* Reset inherited state */
271 tp->rx_opt.ts_recent = 0;
272 tp->rx_opt.ts_recent_stamp = 0;
273 if (likely(!tp->repair))
274 WRITE_ONCE(tp->write_seq, 0);
275 }
276
277 inet->inet_dport = usin->sin_port;
278 sk_daddr_set(sk, daddr);
279
280 inet_csk(sk)->icsk_ext_hdr_len = 0;
281 if (inet_opt)
282 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
283
284 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
285
286 /* Socket identity is still unknown (sport may be zero).
287 * However we set state to SYN-SENT and not releasing socket
288 * lock select source port, enter ourselves into the hash tables and
289 * complete initialization after this.
290 */
291 tcp_set_state(sk, TCP_SYN_SENT);
292 err = inet_hash_connect(tcp_death_row, sk);
293 if (err)
294 goto failure;
295
296 sk_set_txhash(sk);
297
298 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
299 inet->inet_sport, inet->inet_dport, sk);
300 if (IS_ERR(rt)) {
301 err = PTR_ERR(rt);
302 rt = NULL;
303 goto failure;
304 }
305 tp->tcp_usec_ts = dst_tcp_usec_ts(&rt->dst);
306 /* OK, now commit destination to socket. */
307 sk->sk_gso_type = SKB_GSO_TCPV4;
308 sk_setup_caps(sk, &rt->dst);
309 rt = NULL;
310
311 if (likely(!tp->repair)) {
312 if (!tp->write_seq)
313 WRITE_ONCE(tp->write_seq,
314 secure_tcp_seq(inet->inet_saddr,
315 inet->inet_daddr,
316 inet->inet_sport,
317 usin->sin_port));
318 WRITE_ONCE(tp->tsoffset,
319 secure_tcp_ts_off(net, inet->inet_saddr,
320 inet->inet_daddr));
321 }
322
323 atomic_set(&inet->inet_id, get_random_u16());
324
325 if (tcp_fastopen_defer_connect(sk, &err))
326 return err;
327 if (err)
328 goto failure;
329
330 err = tcp_connect(sk);
331
332 if (err)
333 goto failure;
334
335 return 0;
336
337failure:
338 /*
339 * This unhashes the socket and releases the local port,
340 * if necessary.
341 */
342 tcp_set_state(sk, TCP_CLOSE);
343 inet_bhash2_reset_saddr(sk);
344 ip_rt_put(rt);
345 sk->sk_route_caps = 0;
346 inet->inet_dport = 0;
347 return err;
348}
349EXPORT_SYMBOL(tcp_v4_connect);
350
351/*
352 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
353 * It can be called through tcp_release_cb() if socket was owned by user
354 * at the time tcp_v4_err() was called to handle ICMP message.
355 */
356void tcp_v4_mtu_reduced(struct sock *sk)
357{
358 struct inet_sock *inet = inet_sk(sk);
359 struct dst_entry *dst;
360 u32 mtu;
361
362 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
363 return;
364 mtu = READ_ONCE(tcp_sk(sk)->mtu_info);
365 dst = inet_csk_update_pmtu(sk, mtu);
366 if (!dst)
367 return;
368
369 /* Something is about to be wrong... Remember soft error
370 * for the case, if this connection will not able to recover.
371 */
372 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
373 WRITE_ONCE(sk->sk_err_soft, EMSGSIZE);
374
375 mtu = dst_mtu(dst);
376
377 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
378 ip_sk_accept_pmtu(sk) &&
379 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
380 tcp_sync_mss(sk, mtu);
381
382 /* Resend the TCP packet because it's
383 * clear that the old packet has been
384 * dropped. This is the new "fast" path mtu
385 * discovery.
386 */
387 tcp_simple_retransmit(sk);
388 } /* else let the usual retransmit timer handle it */
389}
390EXPORT_SYMBOL(tcp_v4_mtu_reduced);
391
392static void do_redirect(struct sk_buff *skb, struct sock *sk)
393{
394 struct dst_entry *dst = __sk_dst_check(sk, 0);
395
396 if (dst)
397 dst->ops->redirect(dst, sk, skb);
398}
399
400
401/* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
402void tcp_req_err(struct sock *sk, u32 seq, bool abort)
403{
404 struct request_sock *req = inet_reqsk(sk);
405 struct net *net = sock_net(sk);
406
407 /* ICMPs are not backlogged, hence we cannot get
408 * an established socket here.
409 */
410 if (seq != tcp_rsk(req)->snt_isn) {
411 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
412 } else if (abort) {
413 /*
414 * Still in SYN_RECV, just remove it silently.
415 * There is no good way to pass the error to the newly
416 * created socket, and POSIX does not want network
417 * errors returned from accept().
418 */
419 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
420 tcp_listendrop(req->rsk_listener);
421 }
422 reqsk_put(req);
423}
424EXPORT_SYMBOL(tcp_req_err);
425
426/* TCP-LD (RFC 6069) logic */
427void tcp_ld_RTO_revert(struct sock *sk, u32 seq)
428{
429 struct inet_connection_sock *icsk = inet_csk(sk);
430 struct tcp_sock *tp = tcp_sk(sk);
431 struct sk_buff *skb;
432 s32 remaining;
433 u32 delta_us;
434
435 if (sock_owned_by_user(sk))
436 return;
437
438 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
439 !icsk->icsk_backoff)
440 return;
441
442 skb = tcp_rtx_queue_head(sk);
443 if (WARN_ON_ONCE(!skb))
444 return;
445
446 icsk->icsk_backoff--;
447 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) : TCP_TIMEOUT_INIT;
448 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
449
450 tcp_mstamp_refresh(tp);
451 delta_us = (u32)(tp->tcp_mstamp - tcp_skb_timestamp_us(skb));
452 remaining = icsk->icsk_rto - usecs_to_jiffies(delta_us);
453
454 if (remaining > 0) {
455 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
456 remaining, TCP_RTO_MAX);
457 } else {
458 /* RTO revert clocked out retransmission.
459 * Will retransmit now.
460 */
461 tcp_retransmit_timer(sk);
462 }
463}
464EXPORT_SYMBOL(tcp_ld_RTO_revert);
465
466/*
467 * This routine is called by the ICMP module when it gets some
468 * sort of error condition. If err < 0 then the socket should
469 * be closed and the error returned to the user. If err > 0
470 * it's just the icmp type << 8 | icmp code. After adjustment
471 * header points to the first 8 bytes of the tcp header. We need
472 * to find the appropriate port.
473 *
474 * The locking strategy used here is very "optimistic". When
475 * someone else accesses the socket the ICMP is just dropped
476 * and for some paths there is no check at all.
477 * A more general error queue to queue errors for later handling
478 * is probably better.
479 *
480 */
481
482int tcp_v4_err(struct sk_buff *skb, u32 info)
483{
484 const struct iphdr *iph = (const struct iphdr *)skb->data;
485 struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
486 struct tcp_sock *tp;
487 const int type = icmp_hdr(skb)->type;
488 const int code = icmp_hdr(skb)->code;
489 struct sock *sk;
490 struct request_sock *fastopen;
491 u32 seq, snd_una;
492 int err;
493 struct net *net = dev_net(skb->dev);
494
495 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
496 iph->daddr, th->dest, iph->saddr,
497 ntohs(th->source), inet_iif(skb), 0);
498 if (!sk) {
499 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
500 return -ENOENT;
501 }
502 if (sk->sk_state == TCP_TIME_WAIT) {
503 /* To increase the counter of ignored icmps for TCP-AO */
504 tcp_ao_ignore_icmp(sk, AF_INET, type, code);
505 inet_twsk_put(inet_twsk(sk));
506 return 0;
507 }
508 seq = ntohl(th->seq);
509 if (sk->sk_state == TCP_NEW_SYN_RECV) {
510 tcp_req_err(sk, seq, type == ICMP_PARAMETERPROB ||
511 type == ICMP_TIME_EXCEEDED ||
512 (type == ICMP_DEST_UNREACH &&
513 (code == ICMP_NET_UNREACH ||
514 code == ICMP_HOST_UNREACH)));
515 return 0;
516 }
517
518 if (tcp_ao_ignore_icmp(sk, AF_INET, type, code)) {
519 sock_put(sk);
520 return 0;
521 }
522
523 bh_lock_sock(sk);
524 /* If too many ICMPs get dropped on busy
525 * servers this needs to be solved differently.
526 * We do take care of PMTU discovery (RFC1191) special case :
527 * we can receive locally generated ICMP messages while socket is held.
528 */
529 if (sock_owned_by_user(sk)) {
530 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
531 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
532 }
533 if (sk->sk_state == TCP_CLOSE)
534 goto out;
535
536 if (static_branch_unlikely(&ip4_min_ttl)) {
537 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
538 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
539 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
540 goto out;
541 }
542 }
543
544 tp = tcp_sk(sk);
545 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
546 fastopen = rcu_dereference(tp->fastopen_rsk);
547 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
548 if (sk->sk_state != TCP_LISTEN &&
549 !between(seq, snd_una, tp->snd_nxt)) {
550 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
551 goto out;
552 }
553
554 switch (type) {
555 case ICMP_REDIRECT:
556 if (!sock_owned_by_user(sk))
557 do_redirect(skb, sk);
558 goto out;
559 case ICMP_SOURCE_QUENCH:
560 /* Just silently ignore these. */
561 goto out;
562 case ICMP_PARAMETERPROB:
563 err = EPROTO;
564 break;
565 case ICMP_DEST_UNREACH:
566 if (code > NR_ICMP_UNREACH)
567 goto out;
568
569 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
570 /* We are not interested in TCP_LISTEN and open_requests
571 * (SYN-ACKs send out by Linux are always <576bytes so
572 * they should go through unfragmented).
573 */
574 if (sk->sk_state == TCP_LISTEN)
575 goto out;
576
577 WRITE_ONCE(tp->mtu_info, info);
578 if (!sock_owned_by_user(sk)) {
579 tcp_v4_mtu_reduced(sk);
580 } else {
581 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
582 sock_hold(sk);
583 }
584 goto out;
585 }
586
587 err = icmp_err_convert[code].errno;
588 /* check if this ICMP message allows revert of backoff.
589 * (see RFC 6069)
590 */
591 if (!fastopen &&
592 (code == ICMP_NET_UNREACH || code == ICMP_HOST_UNREACH))
593 tcp_ld_RTO_revert(sk, seq);
594 break;
595 case ICMP_TIME_EXCEEDED:
596 err = EHOSTUNREACH;
597 break;
598 default:
599 goto out;
600 }
601
602 switch (sk->sk_state) {
603 case TCP_SYN_SENT:
604 case TCP_SYN_RECV:
605 /* Only in fast or simultaneous open. If a fast open socket is
606 * already accepted it is treated as a connected one below.
607 */
608 if (fastopen && !fastopen->sk)
609 break;
610
611 ip_icmp_error(sk, skb, err, th->dest, info, (u8 *)th);
612
613 if (!sock_owned_by_user(sk)) {
614 WRITE_ONCE(sk->sk_err, err);
615
616 sk_error_report(sk);
617
618 tcp_done(sk);
619 } else {
620 WRITE_ONCE(sk->sk_err_soft, err);
621 }
622 goto out;
623 }
624
625 /* If we've already connected we will keep trying
626 * until we time out, or the user gives up.
627 *
628 * rfc1122 4.2.3.9 allows to consider as hard errors
629 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
630 * but it is obsoleted by pmtu discovery).
631 *
632 * Note, that in modern internet, where routing is unreliable
633 * and in each dark corner broken firewalls sit, sending random
634 * errors ordered by their masters even this two messages finally lose
635 * their original sense (even Linux sends invalid PORT_UNREACHs)
636 *
637 * Now we are in compliance with RFCs.
638 * --ANK (980905)
639 */
640
641 if (!sock_owned_by_user(sk) &&
642 inet_test_bit(RECVERR, sk)) {
643 WRITE_ONCE(sk->sk_err, err);
644 sk_error_report(sk);
645 } else { /* Only an error on timeout */
646 WRITE_ONCE(sk->sk_err_soft, err);
647 }
648
649out:
650 bh_unlock_sock(sk);
651 sock_put(sk);
652 return 0;
653}
654
655void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
656{
657 struct tcphdr *th = tcp_hdr(skb);
658
659 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
660 skb->csum_start = skb_transport_header(skb) - skb->head;
661 skb->csum_offset = offsetof(struct tcphdr, check);
662}
663
664/* This routine computes an IPv4 TCP checksum. */
665void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
666{
667 const struct inet_sock *inet = inet_sk(sk);
668
669 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
670}
671EXPORT_SYMBOL(tcp_v4_send_check);
672
673#define REPLY_OPTIONS_LEN (MAX_TCP_OPTION_SPACE / sizeof(__be32))
674
675static bool tcp_v4_ao_sign_reset(const struct sock *sk, struct sk_buff *skb,
676 const struct tcp_ao_hdr *aoh,
677 struct ip_reply_arg *arg, struct tcphdr *reply,
678 __be32 reply_options[REPLY_OPTIONS_LEN])
679{
680#ifdef CONFIG_TCP_AO
681 int sdif = tcp_v4_sdif(skb);
682 int dif = inet_iif(skb);
683 int l3index = sdif ? dif : 0;
684 bool allocated_traffic_key;
685 struct tcp_ao_key *key;
686 char *traffic_key;
687 bool drop = true;
688 u32 ao_sne = 0;
689 u8 keyid;
690
691 rcu_read_lock();
692 if (tcp_ao_prepare_reset(sk, skb, aoh, l3index, ntohl(reply->seq),
693 &key, &traffic_key, &allocated_traffic_key,
694 &keyid, &ao_sne))
695 goto out;
696
697 reply_options[0] = htonl((TCPOPT_AO << 24) | (tcp_ao_len(key) << 16) |
698 (aoh->rnext_keyid << 8) | keyid);
699 arg->iov[0].iov_len += tcp_ao_len_aligned(key);
700 reply->doff = arg->iov[0].iov_len / 4;
701
702 if (tcp_ao_hash_hdr(AF_INET, (char *)&reply_options[1],
703 key, traffic_key,
704 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
705 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
706 reply, ao_sne))
707 goto out;
708 drop = false;
709out:
710 rcu_read_unlock();
711 if (allocated_traffic_key)
712 kfree(traffic_key);
713 return drop;
714#else
715 return true;
716#endif
717}
718
719/*
720 * This routine will send an RST to the other tcp.
721 *
722 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
723 * for reset.
724 * Answer: if a packet caused RST, it is not for a socket
725 * existing in our system, if it is matched to a socket,
726 * it is just duplicate segment or bug in other side's TCP.
727 * So that we build reply only basing on parameters
728 * arrived with segment.
729 * Exception: precedence violation. We do not implement it in any case.
730 */
731
732static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
733{
734 const struct tcphdr *th = tcp_hdr(skb);
735 struct {
736 struct tcphdr th;
737 __be32 opt[REPLY_OPTIONS_LEN];
738 } rep;
739 const __u8 *md5_hash_location = NULL;
740 const struct tcp_ao_hdr *aoh;
741 struct ip_reply_arg arg;
742#ifdef CONFIG_TCP_MD5SIG
743 struct tcp_md5sig_key *key = NULL;
744 unsigned char newhash[16];
745 struct sock *sk1 = NULL;
746 int genhash;
747#endif
748 u64 transmit_time = 0;
749 struct sock *ctl_sk;
750 struct net *net;
751 u32 txhash = 0;
752
753 /* Never send a reset in response to a reset. */
754 if (th->rst)
755 return;
756
757 /* If sk not NULL, it means we did a successful lookup and incoming
758 * route had to be correct. prequeue might have dropped our dst.
759 */
760 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
761 return;
762
763 /* Swap the send and the receive. */
764 memset(&rep, 0, sizeof(rep));
765 rep.th.dest = th->source;
766 rep.th.source = th->dest;
767 rep.th.doff = sizeof(struct tcphdr) / 4;
768 rep.th.rst = 1;
769
770 if (th->ack) {
771 rep.th.seq = th->ack_seq;
772 } else {
773 rep.th.ack = 1;
774 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
775 skb->len - (th->doff << 2));
776 }
777
778 memset(&arg, 0, sizeof(arg));
779 arg.iov[0].iov_base = (unsigned char *)&rep;
780 arg.iov[0].iov_len = sizeof(rep.th);
781
782 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
783
784 /* Invalid TCP option size or twice included auth */
785 if (tcp_parse_auth_options(tcp_hdr(skb), &md5_hash_location, &aoh))
786 return;
787
788 if (aoh && tcp_v4_ao_sign_reset(sk, skb, aoh, &arg, &rep.th, rep.opt))
789 return;
790
791#ifdef CONFIG_TCP_MD5SIG
792 rcu_read_lock();
793 if (sk && sk_fullsock(sk)) {
794 const union tcp_md5_addr *addr;
795 int l3index;
796
797 /* sdif set, means packet ingressed via a device
798 * in an L3 domain and inet_iif is set to it.
799 */
800 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
801 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
802 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
803 } else if (md5_hash_location) {
804 const union tcp_md5_addr *addr;
805 int sdif = tcp_v4_sdif(skb);
806 int dif = inet_iif(skb);
807 int l3index;
808
809 /*
810 * active side is lost. Try to find listening socket through
811 * source port, and then find md5 key through listening socket.
812 * we are not loose security here:
813 * Incoming packet is checked with md5 hash with finding key,
814 * no RST generated if md5 hash doesn't match.
815 */
816 sk1 = __inet_lookup_listener(net, net->ipv4.tcp_death_row.hashinfo,
817 NULL, 0, ip_hdr(skb)->saddr,
818 th->source, ip_hdr(skb)->daddr,
819 ntohs(th->source), dif, sdif);
820 /* don't send rst if it can't find key */
821 if (!sk1)
822 goto out;
823
824 /* sdif set, means packet ingressed via a device
825 * in an L3 domain and dif is set to it.
826 */
827 l3index = sdif ? dif : 0;
828 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
829 key = tcp_md5_do_lookup(sk1, l3index, addr, AF_INET);
830 if (!key)
831 goto out;
832
833
834 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
835 if (genhash || memcmp(md5_hash_location, newhash, 16) != 0)
836 goto out;
837
838 }
839
840 if (key) {
841 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
842 (TCPOPT_NOP << 16) |
843 (TCPOPT_MD5SIG << 8) |
844 TCPOLEN_MD5SIG);
845 /* Update length and the length the header thinks exists */
846 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
847 rep.th.doff = arg.iov[0].iov_len / 4;
848
849 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
850 key, ip_hdr(skb)->saddr,
851 ip_hdr(skb)->daddr, &rep.th);
852 }
853#endif
854 /* Can't co-exist with TCPMD5, hence check rep.opt[0] */
855 if (rep.opt[0] == 0) {
856 __be32 mrst = mptcp_reset_option(skb);
857
858 if (mrst) {
859 rep.opt[0] = mrst;
860 arg.iov[0].iov_len += sizeof(mrst);
861 rep.th.doff = arg.iov[0].iov_len / 4;
862 }
863 }
864
865 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
866 ip_hdr(skb)->saddr, /* XXX */
867 arg.iov[0].iov_len, IPPROTO_TCP, 0);
868 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
869 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
870
871 /* When socket is gone, all binding information is lost.
872 * routing might fail in this case. No choice here, if we choose to force
873 * input interface, we will misroute in case of asymmetric route.
874 */
875 if (sk) {
876 arg.bound_dev_if = sk->sk_bound_dev_if;
877 if (sk_fullsock(sk))
878 trace_tcp_send_reset(sk, skb);
879 }
880
881 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
882 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
883
884 arg.tos = ip_hdr(skb)->tos;
885 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
886 local_bh_disable();
887 ctl_sk = this_cpu_read(ipv4_tcp_sk);
888 sock_net_set(ctl_sk, net);
889 if (sk) {
890 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
891 inet_twsk(sk)->tw_mark : sk->sk_mark;
892 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
893 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
894 transmit_time = tcp_transmit_time(sk);
895 xfrm_sk_clone_policy(ctl_sk, sk);
896 txhash = (sk->sk_state == TCP_TIME_WAIT) ?
897 inet_twsk(sk)->tw_txhash : sk->sk_txhash;
898 } else {
899 ctl_sk->sk_mark = 0;
900 ctl_sk->sk_priority = 0;
901 }
902 ip_send_unicast_reply(ctl_sk,
903 skb, &TCP_SKB_CB(skb)->header.h4.opt,
904 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
905 &arg, arg.iov[0].iov_len,
906 transmit_time, txhash);
907
908 xfrm_sk_free_policy(ctl_sk);
909 sock_net_set(ctl_sk, &init_net);
910 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
911 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
912 local_bh_enable();
913
914#ifdef CONFIG_TCP_MD5SIG
915out:
916 rcu_read_unlock();
917#endif
918}
919
920/* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
921 outside socket context is ugly, certainly. What can I do?
922 */
923
924static void tcp_v4_send_ack(const struct sock *sk,
925 struct sk_buff *skb, u32 seq, u32 ack,
926 u32 win, u32 tsval, u32 tsecr, int oif,
927 struct tcp_key *key,
928 int reply_flags, u8 tos, u32 txhash)
929{
930 const struct tcphdr *th = tcp_hdr(skb);
931 struct {
932 struct tcphdr th;
933 __be32 opt[(MAX_TCP_OPTION_SPACE >> 2)];
934 } rep;
935 struct net *net = sock_net(sk);
936 struct ip_reply_arg arg;
937 struct sock *ctl_sk;
938 u64 transmit_time;
939
940 memset(&rep.th, 0, sizeof(struct tcphdr));
941 memset(&arg, 0, sizeof(arg));
942
943 arg.iov[0].iov_base = (unsigned char *)&rep;
944 arg.iov[0].iov_len = sizeof(rep.th);
945 if (tsecr) {
946 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
947 (TCPOPT_TIMESTAMP << 8) |
948 TCPOLEN_TIMESTAMP);
949 rep.opt[1] = htonl(tsval);
950 rep.opt[2] = htonl(tsecr);
951 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
952 }
953
954 /* Swap the send and the receive. */
955 rep.th.dest = th->source;
956 rep.th.source = th->dest;
957 rep.th.doff = arg.iov[0].iov_len / 4;
958 rep.th.seq = htonl(seq);
959 rep.th.ack_seq = htonl(ack);
960 rep.th.ack = 1;
961 rep.th.window = htons(win);
962
963#ifdef CONFIG_TCP_MD5SIG
964 if (tcp_key_is_md5(key)) {
965 int offset = (tsecr) ? 3 : 0;
966
967 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
968 (TCPOPT_NOP << 16) |
969 (TCPOPT_MD5SIG << 8) |
970 TCPOLEN_MD5SIG);
971 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
972 rep.th.doff = arg.iov[0].iov_len/4;
973
974 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
975 key->md5_key, ip_hdr(skb)->saddr,
976 ip_hdr(skb)->daddr, &rep.th);
977 }
978#endif
979#ifdef CONFIG_TCP_AO
980 if (tcp_key_is_ao(key)) {
981 int offset = (tsecr) ? 3 : 0;
982
983 rep.opt[offset++] = htonl((TCPOPT_AO << 24) |
984 (tcp_ao_len(key->ao_key) << 16) |
985 (key->ao_key->sndid << 8) |
986 key->rcv_next);
987 arg.iov[0].iov_len += tcp_ao_len_aligned(key->ao_key);
988 rep.th.doff = arg.iov[0].iov_len / 4;
989
990 tcp_ao_hash_hdr(AF_INET, (char *)&rep.opt[offset],
991 key->ao_key, key->traffic_key,
992 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
993 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
994 &rep.th, key->sne);
995 }
996#endif
997 arg.flags = reply_flags;
998 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
999 ip_hdr(skb)->saddr, /* XXX */
1000 arg.iov[0].iov_len, IPPROTO_TCP, 0);
1001 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1002 if (oif)
1003 arg.bound_dev_if = oif;
1004 arg.tos = tos;
1005 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
1006 local_bh_disable();
1007 ctl_sk = this_cpu_read(ipv4_tcp_sk);
1008 sock_net_set(ctl_sk, net);
1009 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
1010 inet_twsk(sk)->tw_mark : READ_ONCE(sk->sk_mark);
1011 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
1012 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
1013 transmit_time = tcp_transmit_time(sk);
1014 ip_send_unicast_reply(ctl_sk,
1015 skb, &TCP_SKB_CB(skb)->header.h4.opt,
1016 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
1017 &arg, arg.iov[0].iov_len,
1018 transmit_time, txhash);
1019
1020 sock_net_set(ctl_sk, &init_net);
1021 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
1022 local_bh_enable();
1023}
1024
1025static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
1026{
1027 struct inet_timewait_sock *tw = inet_twsk(sk);
1028 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
1029 struct tcp_key key = {};
1030#ifdef CONFIG_TCP_AO
1031 struct tcp_ao_info *ao_info;
1032
1033 if (static_branch_unlikely(&tcp_ao_needed.key)) {
1034 /* FIXME: the segment to-be-acked is not verified yet */
1035 ao_info = rcu_dereference(tcptw->ao_info);
1036 if (ao_info) {
1037 const struct tcp_ao_hdr *aoh;
1038
1039 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) {
1040 inet_twsk_put(tw);
1041 return;
1042 }
1043
1044 if (aoh)
1045 key.ao_key = tcp_ao_established_key(ao_info, aoh->rnext_keyid, -1);
1046 }
1047 }
1048 if (key.ao_key) {
1049 struct tcp_ao_key *rnext_key;
1050
1051 key.traffic_key = snd_other_key(key.ao_key);
1052 key.sne = READ_ONCE(ao_info->snd_sne);
1053 rnext_key = READ_ONCE(ao_info->rnext_key);
1054 key.rcv_next = rnext_key->rcvid;
1055 key.type = TCP_KEY_AO;
1056#else
1057 if (0) {
1058#endif
1059#ifdef CONFIG_TCP_MD5SIG
1060 } else if (static_branch_unlikely(&tcp_md5_needed.key)) {
1061 key.md5_key = tcp_twsk_md5_key(tcptw);
1062 if (key.md5_key)
1063 key.type = TCP_KEY_MD5;
1064#endif
1065 }
1066
1067 tcp_v4_send_ack(sk, skb,
1068 tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
1069 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
1070 tcp_tw_tsval(tcptw),
1071 tcptw->tw_ts_recent,
1072 tw->tw_bound_dev_if, &key,
1073 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
1074 tw->tw_tos,
1075 tw->tw_txhash);
1076
1077 inet_twsk_put(tw);
1078}
1079
1080static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
1081 struct request_sock *req)
1082{
1083 struct tcp_key key = {};
1084
1085 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
1086 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
1087 */
1088 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
1089 tcp_sk(sk)->snd_nxt;
1090
1091#ifdef CONFIG_TCP_AO
1092 if (static_branch_unlikely(&tcp_ao_needed.key) &&
1093 tcp_rsk_used_ao(req)) {
1094 const union tcp_md5_addr *addr;
1095 const struct tcp_ao_hdr *aoh;
1096 int l3index;
1097
1098 /* Invalid TCP option size or twice included auth */
1099 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
1100 return;
1101 if (!aoh)
1102 return;
1103
1104 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1105 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1106 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET,
1107 aoh->rnext_keyid, -1);
1108 if (unlikely(!key.ao_key)) {
1109 /* Send ACK with any matching MKT for the peer */
1110 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET, -1, -1);
1111 /* Matching key disappeared (user removed the key?)
1112 * let the handshake timeout.
1113 */
1114 if (!key.ao_key) {
1115 net_info_ratelimited("TCP-AO key for (%pI4, %d)->(%pI4, %d) suddenly disappeared, won't ACK new connection\n",
1116 addr,
1117 ntohs(tcp_hdr(skb)->source),
1118 &ip_hdr(skb)->daddr,
1119 ntohs(tcp_hdr(skb)->dest));
1120 return;
1121 }
1122 }
1123 key.traffic_key = kmalloc(tcp_ao_digest_size(key.ao_key), GFP_ATOMIC);
1124 if (!key.traffic_key)
1125 return;
1126
1127 key.type = TCP_KEY_AO;
1128 key.rcv_next = aoh->keyid;
1129 tcp_v4_ao_calc_key_rsk(key.ao_key, key.traffic_key, req);
1130#else
1131 if (0) {
1132#endif
1133#ifdef CONFIG_TCP_MD5SIG
1134 } else if (static_branch_unlikely(&tcp_md5_needed.key)) {
1135 const union tcp_md5_addr *addr;
1136 int l3index;
1137
1138 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1139 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1140 key.md5_key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1141 if (key.md5_key)
1142 key.type = TCP_KEY_MD5;
1143#endif
1144 }
1145
1146 /* RFC 7323 2.3
1147 * The window field (SEG.WND) of every outgoing segment, with the
1148 * exception of <SYN> segments, MUST be right-shifted by
1149 * Rcv.Wind.Shift bits:
1150 */
1151 tcp_v4_send_ack(sk, skb, seq,
1152 tcp_rsk(req)->rcv_nxt,
1153 req->rsk_rcv_wnd >> inet_rsk(req)->rcv_wscale,
1154 tcp_rsk_tsval(tcp_rsk(req)),
1155 READ_ONCE(req->ts_recent),
1156 0, &key,
1157 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
1158 ip_hdr(skb)->tos,
1159 READ_ONCE(tcp_rsk(req)->txhash));
1160 if (tcp_key_is_ao(&key))
1161 kfree(key.traffic_key);
1162}
1163
1164/*
1165 * Send a SYN-ACK after having received a SYN.
1166 * This still operates on a request_sock only, not on a big
1167 * socket.
1168 */
1169static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
1170 struct flowi *fl,
1171 struct request_sock *req,
1172 struct tcp_fastopen_cookie *foc,
1173 enum tcp_synack_type synack_type,
1174 struct sk_buff *syn_skb)
1175{
1176 const struct inet_request_sock *ireq = inet_rsk(req);
1177 struct flowi4 fl4;
1178 int err = -1;
1179 struct sk_buff *skb;
1180 u8 tos;
1181
1182 /* First, grab a route. */
1183 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
1184 return -1;
1185
1186 skb = tcp_make_synack(sk, dst, req, foc, synack_type, syn_skb);
1187
1188 if (skb) {
1189 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
1190
1191 tos = READ_ONCE(inet_sk(sk)->tos);
1192
1193 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1194 tos = (tcp_rsk(req)->syn_tos & ~INET_ECN_MASK) |
1195 (tos & INET_ECN_MASK);
1196
1197 if (!INET_ECN_is_capable(tos) &&
1198 tcp_bpf_ca_needs_ecn((struct sock *)req))
1199 tos |= INET_ECN_ECT_0;
1200
1201 rcu_read_lock();
1202 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
1203 ireq->ir_rmt_addr,
1204 rcu_dereference(ireq->ireq_opt),
1205 tos);
1206 rcu_read_unlock();
1207 err = net_xmit_eval(err);
1208 }
1209
1210 return err;
1211}
1212
1213/*
1214 * IPv4 request_sock destructor.
1215 */
1216static void tcp_v4_reqsk_destructor(struct request_sock *req)
1217{
1218 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
1219}
1220
1221#ifdef CONFIG_TCP_MD5SIG
1222/*
1223 * RFC2385 MD5 checksumming requires a mapping of
1224 * IP address->MD5 Key.
1225 * We need to maintain these in the sk structure.
1226 */
1227
1228DEFINE_STATIC_KEY_DEFERRED_FALSE(tcp_md5_needed, HZ);
1229EXPORT_SYMBOL(tcp_md5_needed);
1230
1231static bool better_md5_match(struct tcp_md5sig_key *old, struct tcp_md5sig_key *new)
1232{
1233 if (!old)
1234 return true;
1235
1236 /* l3index always overrides non-l3index */
1237 if (old->l3index && new->l3index == 0)
1238 return false;
1239 if (old->l3index == 0 && new->l3index)
1240 return true;
1241
1242 return old->prefixlen < new->prefixlen;
1243}
1244
1245/* Find the Key structure for an address. */
1246struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1247 const union tcp_md5_addr *addr,
1248 int family, bool any_l3index)
1249{
1250 const struct tcp_sock *tp = tcp_sk(sk);
1251 struct tcp_md5sig_key *key;
1252 const struct tcp_md5sig_info *md5sig;
1253 __be32 mask;
1254 struct tcp_md5sig_key *best_match = NULL;
1255 bool match;
1256
1257 /* caller either holds rcu_read_lock() or socket lock */
1258 md5sig = rcu_dereference_check(tp->md5sig_info,
1259 lockdep_sock_is_held(sk));
1260 if (!md5sig)
1261 return NULL;
1262
1263 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1264 lockdep_sock_is_held(sk)) {
1265 if (key->family != family)
1266 continue;
1267 if (!any_l3index && key->flags & TCP_MD5SIG_FLAG_IFINDEX &&
1268 key->l3index != l3index)
1269 continue;
1270 if (family == AF_INET) {
1271 mask = inet_make_mask(key->prefixlen);
1272 match = (key->addr.a4.s_addr & mask) ==
1273 (addr->a4.s_addr & mask);
1274#if IS_ENABLED(CONFIG_IPV6)
1275 } else if (family == AF_INET6) {
1276 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
1277 key->prefixlen);
1278#endif
1279 } else {
1280 match = false;
1281 }
1282
1283 if (match && better_md5_match(best_match, key))
1284 best_match = key;
1285 }
1286 return best_match;
1287}
1288EXPORT_SYMBOL(__tcp_md5_do_lookup);
1289
1290static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
1291 const union tcp_md5_addr *addr,
1292 int family, u8 prefixlen,
1293 int l3index, u8 flags)
1294{
1295 const struct tcp_sock *tp = tcp_sk(sk);
1296 struct tcp_md5sig_key *key;
1297 unsigned int size = sizeof(struct in_addr);
1298 const struct tcp_md5sig_info *md5sig;
1299
1300 /* caller either holds rcu_read_lock() or socket lock */
1301 md5sig = rcu_dereference_check(tp->md5sig_info,
1302 lockdep_sock_is_held(sk));
1303 if (!md5sig)
1304 return NULL;
1305#if IS_ENABLED(CONFIG_IPV6)
1306 if (family == AF_INET6)
1307 size = sizeof(struct in6_addr);
1308#endif
1309 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1310 lockdep_sock_is_held(sk)) {
1311 if (key->family != family)
1312 continue;
1313 if ((key->flags & TCP_MD5SIG_FLAG_IFINDEX) != (flags & TCP_MD5SIG_FLAG_IFINDEX))
1314 continue;
1315 if (key->l3index != l3index)
1316 continue;
1317 if (!memcmp(&key->addr, addr, size) &&
1318 key->prefixlen == prefixlen)
1319 return key;
1320 }
1321 return NULL;
1322}
1323
1324struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1325 const struct sock *addr_sk)
1326{
1327 const union tcp_md5_addr *addr;
1328 int l3index;
1329
1330 l3index = l3mdev_master_ifindex_by_index(sock_net(sk),
1331 addr_sk->sk_bound_dev_if);
1332 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
1333 return tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1334}
1335EXPORT_SYMBOL(tcp_v4_md5_lookup);
1336
1337static int tcp_md5sig_info_add(struct sock *sk, gfp_t gfp)
1338{
1339 struct tcp_sock *tp = tcp_sk(sk);
1340 struct tcp_md5sig_info *md5sig;
1341
1342 md5sig = kmalloc(sizeof(*md5sig), gfp);
1343 if (!md5sig)
1344 return -ENOMEM;
1345
1346 sk_gso_disable(sk);
1347 INIT_HLIST_HEAD(&md5sig->head);
1348 rcu_assign_pointer(tp->md5sig_info, md5sig);
1349 return 0;
1350}
1351
1352/* This can be called on a newly created socket, from other files */
1353static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1354 int family, u8 prefixlen, int l3index, u8 flags,
1355 const u8 *newkey, u8 newkeylen, gfp_t gfp)
1356{
1357 /* Add Key to the list */
1358 struct tcp_md5sig_key *key;
1359 struct tcp_sock *tp = tcp_sk(sk);
1360 struct tcp_md5sig_info *md5sig;
1361
1362 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1363 if (key) {
1364 /* Pre-existing entry - just update that one.
1365 * Note that the key might be used concurrently.
1366 * data_race() is telling kcsan that we do not care of
1367 * key mismatches, since changing MD5 key on live flows
1368 * can lead to packet drops.
1369 */
1370 data_race(memcpy(key->key, newkey, newkeylen));
1371
1372 /* Pairs with READ_ONCE() in tcp_md5_hash_key().
1373 * Also note that a reader could catch new key->keylen value
1374 * but old key->key[], this is the reason we use __GFP_ZERO
1375 * at sock_kmalloc() time below these lines.
1376 */
1377 WRITE_ONCE(key->keylen, newkeylen);
1378
1379 return 0;
1380 }
1381
1382 md5sig = rcu_dereference_protected(tp->md5sig_info,
1383 lockdep_sock_is_held(sk));
1384
1385 key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
1386 if (!key)
1387 return -ENOMEM;
1388
1389 memcpy(key->key, newkey, newkeylen);
1390 key->keylen = newkeylen;
1391 key->family = family;
1392 key->prefixlen = prefixlen;
1393 key->l3index = l3index;
1394 key->flags = flags;
1395 memcpy(&key->addr, addr,
1396 (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) ? sizeof(struct in6_addr) :
1397 sizeof(struct in_addr));
1398 hlist_add_head_rcu(&key->node, &md5sig->head);
1399 return 0;
1400}
1401
1402int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1403 int family, u8 prefixlen, int l3index, u8 flags,
1404 const u8 *newkey, u8 newkeylen)
1405{
1406 struct tcp_sock *tp = tcp_sk(sk);
1407
1408 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1409 if (tcp_md5_alloc_sigpool())
1410 return -ENOMEM;
1411
1412 if (tcp_md5sig_info_add(sk, GFP_KERNEL)) {
1413 tcp_md5_release_sigpool();
1414 return -ENOMEM;
1415 }
1416
1417 if (!static_branch_inc(&tcp_md5_needed.key)) {
1418 struct tcp_md5sig_info *md5sig;
1419
1420 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1421 rcu_assign_pointer(tp->md5sig_info, NULL);
1422 kfree_rcu(md5sig, rcu);
1423 tcp_md5_release_sigpool();
1424 return -EUSERS;
1425 }
1426 }
1427
1428 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index, flags,
1429 newkey, newkeylen, GFP_KERNEL);
1430}
1431EXPORT_SYMBOL(tcp_md5_do_add);
1432
1433int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1434 int family, u8 prefixlen, int l3index,
1435 struct tcp_md5sig_key *key)
1436{
1437 struct tcp_sock *tp = tcp_sk(sk);
1438
1439 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1440 tcp_md5_add_sigpool();
1441
1442 if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC))) {
1443 tcp_md5_release_sigpool();
1444 return -ENOMEM;
1445 }
1446
1447 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
1448 struct tcp_md5sig_info *md5sig;
1449
1450 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1451 net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
1452 rcu_assign_pointer(tp->md5sig_info, NULL);
1453 kfree_rcu(md5sig, rcu);
1454 tcp_md5_release_sigpool();
1455 return -EUSERS;
1456 }
1457 }
1458
1459 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index,
1460 key->flags, key->key, key->keylen,
1461 sk_gfp_mask(sk, GFP_ATOMIC));
1462}
1463EXPORT_SYMBOL(tcp_md5_key_copy);
1464
1465int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1466 u8 prefixlen, int l3index, u8 flags)
1467{
1468 struct tcp_md5sig_key *key;
1469
1470 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1471 if (!key)
1472 return -ENOENT;
1473 hlist_del_rcu(&key->node);
1474 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1475 kfree_rcu(key, rcu);
1476 return 0;
1477}
1478EXPORT_SYMBOL(tcp_md5_do_del);
1479
1480void tcp_clear_md5_list(struct sock *sk)
1481{
1482 struct tcp_sock *tp = tcp_sk(sk);
1483 struct tcp_md5sig_key *key;
1484 struct hlist_node *n;
1485 struct tcp_md5sig_info *md5sig;
1486
1487 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1488
1489 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1490 hlist_del_rcu(&key->node);
1491 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1492 kfree_rcu(key, rcu);
1493 }
1494}
1495
1496static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1497 sockptr_t optval, int optlen)
1498{
1499 struct tcp_md5sig cmd;
1500 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1501 const union tcp_md5_addr *addr;
1502 u8 prefixlen = 32;
1503 int l3index = 0;
1504 bool l3flag;
1505 u8 flags;
1506
1507 if (optlen < sizeof(cmd))
1508 return -EINVAL;
1509
1510 if (copy_from_sockptr(&cmd, optval, sizeof(cmd)))
1511 return -EFAULT;
1512
1513 if (sin->sin_family != AF_INET)
1514 return -EINVAL;
1515
1516 flags = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1517 l3flag = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1518
1519 if (optname == TCP_MD5SIG_EXT &&
1520 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1521 prefixlen = cmd.tcpm_prefixlen;
1522 if (prefixlen > 32)
1523 return -EINVAL;
1524 }
1525
1526 if (optname == TCP_MD5SIG_EXT && cmd.tcpm_ifindex &&
1527 cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX) {
1528 struct net_device *dev;
1529
1530 rcu_read_lock();
1531 dev = dev_get_by_index_rcu(sock_net(sk), cmd.tcpm_ifindex);
1532 if (dev && netif_is_l3_master(dev))
1533 l3index = dev->ifindex;
1534
1535 rcu_read_unlock();
1536
1537 /* ok to reference set/not set outside of rcu;
1538 * right now device MUST be an L3 master
1539 */
1540 if (!dev || !l3index)
1541 return -EINVAL;
1542 }
1543
1544 addr = (union tcp_md5_addr *)&sin->sin_addr.s_addr;
1545
1546 if (!cmd.tcpm_keylen)
1547 return tcp_md5_do_del(sk, addr, AF_INET, prefixlen, l3index, flags);
1548
1549 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1550 return -EINVAL;
1551
1552 /* Don't allow keys for peers that have a matching TCP-AO key.
1553 * See the comment in tcp_ao_add_cmd()
1554 */
1555 if (tcp_ao_required(sk, addr, AF_INET, l3flag ? l3index : -1, false))
1556 return -EKEYREJECTED;
1557
1558 return tcp_md5_do_add(sk, addr, AF_INET, prefixlen, l3index, flags,
1559 cmd.tcpm_key, cmd.tcpm_keylen);
1560}
1561
1562static int tcp_v4_md5_hash_headers(struct tcp_sigpool *hp,
1563 __be32 daddr, __be32 saddr,
1564 const struct tcphdr *th, int nbytes)
1565{
1566 struct tcp4_pseudohdr *bp;
1567 struct scatterlist sg;
1568 struct tcphdr *_th;
1569
1570 bp = hp->scratch;
1571 bp->saddr = saddr;
1572 bp->daddr = daddr;
1573 bp->pad = 0;
1574 bp->protocol = IPPROTO_TCP;
1575 bp->len = cpu_to_be16(nbytes);
1576
1577 _th = (struct tcphdr *)(bp + 1);
1578 memcpy(_th, th, sizeof(*th));
1579 _th->check = 0;
1580
1581 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1582 ahash_request_set_crypt(hp->req, &sg, NULL,
1583 sizeof(*bp) + sizeof(*th));
1584 return crypto_ahash_update(hp->req);
1585}
1586
1587static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1588 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1589{
1590 struct tcp_sigpool hp;
1591
1592 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1593 goto clear_hash_nostart;
1594
1595 if (crypto_ahash_init(hp.req))
1596 goto clear_hash;
1597 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
1598 goto clear_hash;
1599 if (tcp_md5_hash_key(&hp, key))
1600 goto clear_hash;
1601 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1602 if (crypto_ahash_final(hp.req))
1603 goto clear_hash;
1604
1605 tcp_sigpool_end(&hp);
1606 return 0;
1607
1608clear_hash:
1609 tcp_sigpool_end(&hp);
1610clear_hash_nostart:
1611 memset(md5_hash, 0, 16);
1612 return 1;
1613}
1614
1615int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1616 const struct sock *sk,
1617 const struct sk_buff *skb)
1618{
1619 const struct tcphdr *th = tcp_hdr(skb);
1620 struct tcp_sigpool hp;
1621 __be32 saddr, daddr;
1622
1623 if (sk) { /* valid for establish/request sockets */
1624 saddr = sk->sk_rcv_saddr;
1625 daddr = sk->sk_daddr;
1626 } else {
1627 const struct iphdr *iph = ip_hdr(skb);
1628 saddr = iph->saddr;
1629 daddr = iph->daddr;
1630 }
1631
1632 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1633 goto clear_hash_nostart;
1634
1635 if (crypto_ahash_init(hp.req))
1636 goto clear_hash;
1637
1638 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
1639 goto clear_hash;
1640 if (tcp_sigpool_hash_skb_data(&hp, skb, th->doff << 2))
1641 goto clear_hash;
1642 if (tcp_md5_hash_key(&hp, key))
1643 goto clear_hash;
1644 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1645 if (crypto_ahash_final(hp.req))
1646 goto clear_hash;
1647
1648 tcp_sigpool_end(&hp);
1649 return 0;
1650
1651clear_hash:
1652 tcp_sigpool_end(&hp);
1653clear_hash_nostart:
1654 memset(md5_hash, 0, 16);
1655 return 1;
1656}
1657EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1658
1659#endif
1660
1661static void tcp_v4_init_req(struct request_sock *req,
1662 const struct sock *sk_listener,
1663 struct sk_buff *skb)
1664{
1665 struct inet_request_sock *ireq = inet_rsk(req);
1666 struct net *net = sock_net(sk_listener);
1667
1668 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1669 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1670 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1671}
1672
1673static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1674 struct sk_buff *skb,
1675 struct flowi *fl,
1676 struct request_sock *req)
1677{
1678 tcp_v4_init_req(req, sk, skb);
1679
1680 if (security_inet_conn_request(sk, skb, req))
1681 return NULL;
1682
1683 return inet_csk_route_req(sk, &fl->u.ip4, req);
1684}
1685
1686struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1687 .family = PF_INET,
1688 .obj_size = sizeof(struct tcp_request_sock),
1689 .rtx_syn_ack = tcp_rtx_synack,
1690 .send_ack = tcp_v4_reqsk_send_ack,
1691 .destructor = tcp_v4_reqsk_destructor,
1692 .send_reset = tcp_v4_send_reset,
1693 .syn_ack_timeout = tcp_syn_ack_timeout,
1694};
1695
1696const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1697 .mss_clamp = TCP_MSS_DEFAULT,
1698#ifdef CONFIG_TCP_MD5SIG
1699 .req_md5_lookup = tcp_v4_md5_lookup,
1700 .calc_md5_hash = tcp_v4_md5_hash_skb,
1701#endif
1702#ifdef CONFIG_TCP_AO
1703 .ao_lookup = tcp_v4_ao_lookup_rsk,
1704 .ao_calc_key = tcp_v4_ao_calc_key_rsk,
1705 .ao_synack_hash = tcp_v4_ao_synack_hash,
1706#endif
1707#ifdef CONFIG_SYN_COOKIES
1708 .cookie_init_seq = cookie_v4_init_sequence,
1709#endif
1710 .route_req = tcp_v4_route_req,
1711 .init_seq = tcp_v4_init_seq,
1712 .init_ts_off = tcp_v4_init_ts_off,
1713 .send_synack = tcp_v4_send_synack,
1714};
1715
1716int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1717{
1718 /* Never answer to SYNs send to broadcast or multicast */
1719 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1720 goto drop;
1721
1722 return tcp_conn_request(&tcp_request_sock_ops,
1723 &tcp_request_sock_ipv4_ops, sk, skb);
1724
1725drop:
1726 tcp_listendrop(sk);
1727 return 0;
1728}
1729EXPORT_SYMBOL(tcp_v4_conn_request);
1730
1731
1732/*
1733 * The three way handshake has completed - we got a valid synack -
1734 * now create the new socket.
1735 */
1736struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1737 struct request_sock *req,
1738 struct dst_entry *dst,
1739 struct request_sock *req_unhash,
1740 bool *own_req)
1741{
1742 struct inet_request_sock *ireq;
1743 bool found_dup_sk = false;
1744 struct inet_sock *newinet;
1745 struct tcp_sock *newtp;
1746 struct sock *newsk;
1747#ifdef CONFIG_TCP_MD5SIG
1748 const union tcp_md5_addr *addr;
1749 struct tcp_md5sig_key *key;
1750 int l3index;
1751#endif
1752 struct ip_options_rcu *inet_opt;
1753
1754 if (sk_acceptq_is_full(sk))
1755 goto exit_overflow;
1756
1757 newsk = tcp_create_openreq_child(sk, req, skb);
1758 if (!newsk)
1759 goto exit_nonewsk;
1760
1761 newsk->sk_gso_type = SKB_GSO_TCPV4;
1762 inet_sk_rx_dst_set(newsk, skb);
1763
1764 newtp = tcp_sk(newsk);
1765 newinet = inet_sk(newsk);
1766 ireq = inet_rsk(req);
1767 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1768 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1769 newsk->sk_bound_dev_if = ireq->ir_iif;
1770 newinet->inet_saddr = ireq->ir_loc_addr;
1771 inet_opt = rcu_dereference(ireq->ireq_opt);
1772 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1773 newinet->mc_index = inet_iif(skb);
1774 newinet->mc_ttl = ip_hdr(skb)->ttl;
1775 newinet->rcv_tos = ip_hdr(skb)->tos;
1776 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1777 if (inet_opt)
1778 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1779 atomic_set(&newinet->inet_id, get_random_u16());
1780
1781 /* Set ToS of the new socket based upon the value of incoming SYN.
1782 * ECT bits are set later in tcp_init_transfer().
1783 */
1784 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1785 newinet->tos = tcp_rsk(req)->syn_tos & ~INET_ECN_MASK;
1786
1787 if (!dst) {
1788 dst = inet_csk_route_child_sock(sk, newsk, req);
1789 if (!dst)
1790 goto put_and_exit;
1791 } else {
1792 /* syncookie case : see end of cookie_v4_check() */
1793 }
1794 sk_setup_caps(newsk, dst);
1795
1796 tcp_ca_openreq_child(newsk, dst);
1797
1798 tcp_sync_mss(newsk, dst_mtu(dst));
1799 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1800
1801 tcp_initialize_rcv_mss(newsk);
1802
1803#ifdef CONFIG_TCP_MD5SIG
1804 l3index = l3mdev_master_ifindex_by_index(sock_net(sk), ireq->ir_iif);
1805 /* Copy over the MD5 key from the original socket */
1806 addr = (union tcp_md5_addr *)&newinet->inet_daddr;
1807 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1808 if (key && !tcp_rsk_used_ao(req)) {
1809 if (tcp_md5_key_copy(newsk, addr, AF_INET, 32, l3index, key))
1810 goto put_and_exit;
1811 sk_gso_disable(newsk);
1812 }
1813#endif
1814#ifdef CONFIG_TCP_AO
1815 if (tcp_ao_copy_all_matching(sk, newsk, req, skb, AF_INET))
1816 goto put_and_exit; /* OOM, release back memory */
1817#endif
1818
1819 if (__inet_inherit_port(sk, newsk) < 0)
1820 goto put_and_exit;
1821 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash),
1822 &found_dup_sk);
1823 if (likely(*own_req)) {
1824 tcp_move_syn(newtp, req);
1825 ireq->ireq_opt = NULL;
1826 } else {
1827 newinet->inet_opt = NULL;
1828
1829 if (!req_unhash && found_dup_sk) {
1830 /* This code path should only be executed in the
1831 * syncookie case only
1832 */
1833 bh_unlock_sock(newsk);
1834 sock_put(newsk);
1835 newsk = NULL;
1836 }
1837 }
1838 return newsk;
1839
1840exit_overflow:
1841 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1842exit_nonewsk:
1843 dst_release(dst);
1844exit:
1845 tcp_listendrop(sk);
1846 return NULL;
1847put_and_exit:
1848 newinet->inet_opt = NULL;
1849 inet_csk_prepare_forced_close(newsk);
1850 tcp_done(newsk);
1851 goto exit;
1852}
1853EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1854
1855static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1856{
1857#ifdef CONFIG_SYN_COOKIES
1858 const struct tcphdr *th = tcp_hdr(skb);
1859
1860 if (!th->syn)
1861 sk = cookie_v4_check(sk, skb);
1862#endif
1863 return sk;
1864}
1865
1866u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
1867 struct tcphdr *th, u32 *cookie)
1868{
1869 u16 mss = 0;
1870#ifdef CONFIG_SYN_COOKIES
1871 mss = tcp_get_syncookie_mss(&tcp_request_sock_ops,
1872 &tcp_request_sock_ipv4_ops, sk, th);
1873 if (mss) {
1874 *cookie = __cookie_v4_init_sequence(iph, th, &mss);
1875 tcp_synq_overflow(sk);
1876 }
1877#endif
1878 return mss;
1879}
1880
1881INDIRECT_CALLABLE_DECLARE(struct dst_entry *ipv4_dst_check(struct dst_entry *,
1882 u32));
1883/* The socket must have it's spinlock held when we get
1884 * here, unless it is a TCP_LISTEN socket.
1885 *
1886 * We have a potential double-lock case here, so even when
1887 * doing backlog processing we use the BH locking scheme.
1888 * This is because we cannot sleep with the original spinlock
1889 * held.
1890 */
1891int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1892{
1893 enum skb_drop_reason reason;
1894 struct sock *rsk;
1895
1896 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1897 struct dst_entry *dst;
1898
1899 dst = rcu_dereference_protected(sk->sk_rx_dst,
1900 lockdep_sock_is_held(sk));
1901
1902 sock_rps_save_rxhash(sk, skb);
1903 sk_mark_napi_id(sk, skb);
1904 if (dst) {
1905 if (sk->sk_rx_dst_ifindex != skb->skb_iif ||
1906 !INDIRECT_CALL_1(dst->ops->check, ipv4_dst_check,
1907 dst, 0)) {
1908 RCU_INIT_POINTER(sk->sk_rx_dst, NULL);
1909 dst_release(dst);
1910 }
1911 }
1912 tcp_rcv_established(sk, skb);
1913 return 0;
1914 }
1915
1916 if (tcp_checksum_complete(skb))
1917 goto csum_err;
1918
1919 if (sk->sk_state == TCP_LISTEN) {
1920 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1921
1922 if (!nsk)
1923 return 0;
1924 if (nsk != sk) {
1925 reason = tcp_child_process(sk, nsk, skb);
1926 if (reason) {
1927 rsk = nsk;
1928 goto reset;
1929 }
1930 return 0;
1931 }
1932 } else
1933 sock_rps_save_rxhash(sk, skb);
1934
1935 reason = tcp_rcv_state_process(sk, skb);
1936 if (reason) {
1937 rsk = sk;
1938 goto reset;
1939 }
1940 return 0;
1941
1942reset:
1943 tcp_v4_send_reset(rsk, skb);
1944discard:
1945 kfree_skb_reason(skb, reason);
1946 /* Be careful here. If this function gets more complicated and
1947 * gcc suffers from register pressure on the x86, sk (in %ebx)
1948 * might be destroyed here. This current version compiles correctly,
1949 * but you have been warned.
1950 */
1951 return 0;
1952
1953csum_err:
1954 reason = SKB_DROP_REASON_TCP_CSUM;
1955 trace_tcp_bad_csum(skb);
1956 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1957 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1958 goto discard;
1959}
1960EXPORT_SYMBOL(tcp_v4_do_rcv);
1961
1962int tcp_v4_early_demux(struct sk_buff *skb)
1963{
1964 struct net *net = dev_net(skb->dev);
1965 const struct iphdr *iph;
1966 const struct tcphdr *th;
1967 struct sock *sk;
1968
1969 if (skb->pkt_type != PACKET_HOST)
1970 return 0;
1971
1972 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1973 return 0;
1974
1975 iph = ip_hdr(skb);
1976 th = tcp_hdr(skb);
1977
1978 if (th->doff < sizeof(struct tcphdr) / 4)
1979 return 0;
1980
1981 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
1982 iph->saddr, th->source,
1983 iph->daddr, ntohs(th->dest),
1984 skb->skb_iif, inet_sdif(skb));
1985 if (sk) {
1986 skb->sk = sk;
1987 skb->destructor = sock_edemux;
1988 if (sk_fullsock(sk)) {
1989 struct dst_entry *dst = rcu_dereference(sk->sk_rx_dst);
1990
1991 if (dst)
1992 dst = dst_check(dst, 0);
1993 if (dst &&
1994 sk->sk_rx_dst_ifindex == skb->skb_iif)
1995 skb_dst_set_noref(skb, dst);
1996 }
1997 }
1998 return 0;
1999}
2000
2001bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
2002 enum skb_drop_reason *reason)
2003{
2004 u32 tail_gso_size, tail_gso_segs;
2005 struct skb_shared_info *shinfo;
2006 const struct tcphdr *th;
2007 struct tcphdr *thtail;
2008 struct sk_buff *tail;
2009 unsigned int hdrlen;
2010 bool fragstolen;
2011 u32 gso_segs;
2012 u32 gso_size;
2013 u64 limit;
2014 int delta;
2015
2016 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
2017 * we can fix skb->truesize to its real value to avoid future drops.
2018 * This is valid because skb is not yet charged to the socket.
2019 * It has been noticed pure SACK packets were sometimes dropped
2020 * (if cooked by drivers without copybreak feature).
2021 */
2022 skb_condense(skb);
2023
2024 skb_dst_drop(skb);
2025
2026 if (unlikely(tcp_checksum_complete(skb))) {
2027 bh_unlock_sock(sk);
2028 trace_tcp_bad_csum(skb);
2029 *reason = SKB_DROP_REASON_TCP_CSUM;
2030 __TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
2031 __TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
2032 return true;
2033 }
2034
2035 /* Attempt coalescing to last skb in backlog, even if we are
2036 * above the limits.
2037 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
2038 */
2039 th = (const struct tcphdr *)skb->data;
2040 hdrlen = th->doff * 4;
2041
2042 tail = sk->sk_backlog.tail;
2043 if (!tail)
2044 goto no_coalesce;
2045 thtail = (struct tcphdr *)tail->data;
2046
2047 if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
2048 TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
2049 ((TCP_SKB_CB(tail)->tcp_flags |
2050 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_SYN | TCPHDR_RST | TCPHDR_URG)) ||
2051 !((TCP_SKB_CB(tail)->tcp_flags &
2052 TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_ACK) ||
2053 ((TCP_SKB_CB(tail)->tcp_flags ^
2054 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
2055#ifdef CONFIG_TLS_DEVICE
2056 tail->decrypted != skb->decrypted ||
2057#endif
2058 !mptcp_skb_can_collapse(tail, skb) ||
2059 thtail->doff != th->doff ||
2060 memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
2061 goto no_coalesce;
2062
2063 __skb_pull(skb, hdrlen);
2064
2065 shinfo = skb_shinfo(skb);
2066 gso_size = shinfo->gso_size ?: skb->len;
2067 gso_segs = shinfo->gso_segs ?: 1;
2068
2069 shinfo = skb_shinfo(tail);
2070 tail_gso_size = shinfo->gso_size ?: (tail->len - hdrlen);
2071 tail_gso_segs = shinfo->gso_segs ?: 1;
2072
2073 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
2074 TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
2075
2076 if (likely(!before(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))) {
2077 TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
2078 thtail->window = th->window;
2079 }
2080
2081 /* We have to update both TCP_SKB_CB(tail)->tcp_flags and
2082 * thtail->fin, so that the fast path in tcp_rcv_established()
2083 * is not entered if we append a packet with a FIN.
2084 * SYN, RST, URG are not present.
2085 * ACK is set on both packets.
2086 * PSH : we do not really care in TCP stack,
2087 * at least for 'GRO' packets.
2088 */
2089 thtail->fin |= th->fin;
2090 TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2091
2092 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2093 TCP_SKB_CB(tail)->has_rxtstamp = true;
2094 tail->tstamp = skb->tstamp;
2095 skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
2096 }
2097
2098 /* Not as strict as GRO. We only need to carry mss max value */
2099 shinfo->gso_size = max(gso_size, tail_gso_size);
2100 shinfo->gso_segs = min_t(u32, gso_segs + tail_gso_segs, 0xFFFF);
2101
2102 sk->sk_backlog.len += delta;
2103 __NET_INC_STATS(sock_net(sk),
2104 LINUX_MIB_TCPBACKLOGCOALESCE);
2105 kfree_skb_partial(skb, fragstolen);
2106 return false;
2107 }
2108 __skb_push(skb, hdrlen);
2109
2110no_coalesce:
2111 /* sk->sk_backlog.len is reset only at the end of __release_sock().
2112 * Both sk->sk_backlog.len and sk->sk_rmem_alloc could reach
2113 * sk_rcvbuf in normal conditions.
2114 */
2115 limit = ((u64)READ_ONCE(sk->sk_rcvbuf)) << 1;
2116
2117 limit += ((u32)READ_ONCE(sk->sk_sndbuf)) >> 1;
2118
2119 /* Only socket owner can try to collapse/prune rx queues
2120 * to reduce memory overhead, so add a little headroom here.
2121 * Few sockets backlog are possibly concurrently non empty.
2122 */
2123 limit += 64 * 1024;
2124
2125 limit = min_t(u64, limit, UINT_MAX);
2126
2127 if (unlikely(sk_add_backlog(sk, skb, limit))) {
2128 bh_unlock_sock(sk);
2129 *reason = SKB_DROP_REASON_SOCKET_BACKLOG;
2130 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
2131 return true;
2132 }
2133 return false;
2134}
2135EXPORT_SYMBOL(tcp_add_backlog);
2136
2137int tcp_filter(struct sock *sk, struct sk_buff *skb)
2138{
2139 struct tcphdr *th = (struct tcphdr *)skb->data;
2140
2141 return sk_filter_trim_cap(sk, skb, th->doff * 4);
2142}
2143EXPORT_SYMBOL(tcp_filter);
2144
2145static void tcp_v4_restore_cb(struct sk_buff *skb)
2146{
2147 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
2148 sizeof(struct inet_skb_parm));
2149}
2150
2151static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
2152 const struct tcphdr *th)
2153{
2154 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
2155 * barrier() makes sure compiler wont play fool^Waliasing games.
2156 */
2157 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
2158 sizeof(struct inet_skb_parm));
2159 barrier();
2160
2161 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
2162 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
2163 skb->len - th->doff * 4);
2164 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
2165 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
2166 TCP_SKB_CB(skb)->tcp_tw_isn = 0;
2167 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
2168 TCP_SKB_CB(skb)->sacked = 0;
2169 TCP_SKB_CB(skb)->has_rxtstamp =
2170 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
2171}
2172
2173/*
2174 * From tcp_input.c
2175 */
2176
2177int tcp_v4_rcv(struct sk_buff *skb)
2178{
2179 struct net *net = dev_net(skb->dev);
2180 enum skb_drop_reason drop_reason;
2181 int sdif = inet_sdif(skb);
2182 int dif = inet_iif(skb);
2183 const struct iphdr *iph;
2184 const struct tcphdr *th;
2185 bool refcounted;
2186 struct sock *sk;
2187 int ret;
2188
2189 drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
2190 if (skb->pkt_type != PACKET_HOST)
2191 goto discard_it;
2192
2193 /* Count it even if it's bad */
2194 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
2195
2196 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
2197 goto discard_it;
2198
2199 th = (const struct tcphdr *)skb->data;
2200
2201 if (unlikely(th->doff < sizeof(struct tcphdr) / 4)) {
2202 drop_reason = SKB_DROP_REASON_PKT_TOO_SMALL;
2203 goto bad_packet;
2204 }
2205 if (!pskb_may_pull(skb, th->doff * 4))
2206 goto discard_it;
2207
2208 /* An explanation is required here, I think.
2209 * Packet length and doff are validated by header prediction,
2210 * provided case of th->doff==0 is eliminated.
2211 * So, we defer the checks. */
2212
2213 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
2214 goto csum_error;
2215
2216 th = (const struct tcphdr *)skb->data;
2217 iph = ip_hdr(skb);
2218lookup:
2219 sk = __inet_lookup_skb(net->ipv4.tcp_death_row.hashinfo,
2220 skb, __tcp_hdrlen(th), th->source,
2221 th->dest, sdif, &refcounted);
2222 if (!sk)
2223 goto no_tcp_socket;
2224
2225process:
2226 if (sk->sk_state == TCP_TIME_WAIT)
2227 goto do_time_wait;
2228
2229 if (sk->sk_state == TCP_NEW_SYN_RECV) {
2230 struct request_sock *req = inet_reqsk(sk);
2231 bool req_stolen = false;
2232 struct sock *nsk;
2233
2234 sk = req->rsk_listener;
2235 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
2236 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2237 else
2238 drop_reason = tcp_inbound_hash(sk, req, skb,
2239 &iph->saddr, &iph->daddr,
2240 AF_INET, dif, sdif);
2241 if (unlikely(drop_reason)) {
2242 sk_drops_add(sk, skb);
2243 reqsk_put(req);
2244 goto discard_it;
2245 }
2246 if (tcp_checksum_complete(skb)) {
2247 reqsk_put(req);
2248 goto csum_error;
2249 }
2250 if (unlikely(sk->sk_state != TCP_LISTEN)) {
2251 nsk = reuseport_migrate_sock(sk, req_to_sk(req), skb);
2252 if (!nsk) {
2253 inet_csk_reqsk_queue_drop_and_put(sk, req);
2254 goto lookup;
2255 }
2256 sk = nsk;
2257 /* reuseport_migrate_sock() has already held one sk_refcnt
2258 * before returning.
2259 */
2260 } else {
2261 /* We own a reference on the listener, increase it again
2262 * as we might lose it too soon.
2263 */
2264 sock_hold(sk);
2265 }
2266 refcounted = true;
2267 nsk = NULL;
2268 if (!tcp_filter(sk, skb)) {
2269 th = (const struct tcphdr *)skb->data;
2270 iph = ip_hdr(skb);
2271 tcp_v4_fill_cb(skb, iph, th);
2272 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
2273 } else {
2274 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2275 }
2276 if (!nsk) {
2277 reqsk_put(req);
2278 if (req_stolen) {
2279 /* Another cpu got exclusive access to req
2280 * and created a full blown socket.
2281 * Try to feed this packet to this socket
2282 * instead of discarding it.
2283 */
2284 tcp_v4_restore_cb(skb);
2285 sock_put(sk);
2286 goto lookup;
2287 }
2288 goto discard_and_relse;
2289 }
2290 nf_reset_ct(skb);
2291 if (nsk == sk) {
2292 reqsk_put(req);
2293 tcp_v4_restore_cb(skb);
2294 } else {
2295 drop_reason = tcp_child_process(sk, nsk, skb);
2296 if (drop_reason) {
2297 tcp_v4_send_reset(nsk, skb);
2298 goto discard_and_relse;
2299 }
2300 sock_put(sk);
2301 return 0;
2302 }
2303 }
2304
2305 if (static_branch_unlikely(&ip4_min_ttl)) {
2306 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
2307 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
2308 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
2309 drop_reason = SKB_DROP_REASON_TCP_MINTTL;
2310 goto discard_and_relse;
2311 }
2312 }
2313
2314 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) {
2315 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2316 goto discard_and_relse;
2317 }
2318
2319 drop_reason = tcp_inbound_hash(sk, NULL, skb, &iph->saddr, &iph->daddr,
2320 AF_INET, dif, sdif);
2321 if (drop_reason)
2322 goto discard_and_relse;
2323
2324 nf_reset_ct(skb);
2325
2326 if (tcp_filter(sk, skb)) {
2327 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2328 goto discard_and_relse;
2329 }
2330 th = (const struct tcphdr *)skb->data;
2331 iph = ip_hdr(skb);
2332 tcp_v4_fill_cb(skb, iph, th);
2333
2334 skb->dev = NULL;
2335
2336 if (sk->sk_state == TCP_LISTEN) {
2337 ret = tcp_v4_do_rcv(sk, skb);
2338 goto put_and_return;
2339 }
2340
2341 sk_incoming_cpu_update(sk);
2342
2343 bh_lock_sock_nested(sk);
2344 tcp_segs_in(tcp_sk(sk), skb);
2345 ret = 0;
2346 if (!sock_owned_by_user(sk)) {
2347 ret = tcp_v4_do_rcv(sk, skb);
2348 } else {
2349 if (tcp_add_backlog(sk, skb, &drop_reason))
2350 goto discard_and_relse;
2351 }
2352 bh_unlock_sock(sk);
2353
2354put_and_return:
2355 if (refcounted)
2356 sock_put(sk);
2357
2358 return ret;
2359
2360no_tcp_socket:
2361 drop_reason = SKB_DROP_REASON_NO_SOCKET;
2362 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
2363 goto discard_it;
2364
2365 tcp_v4_fill_cb(skb, iph, th);
2366
2367 if (tcp_checksum_complete(skb)) {
2368csum_error:
2369 drop_reason = SKB_DROP_REASON_TCP_CSUM;
2370 trace_tcp_bad_csum(skb);
2371 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
2372bad_packet:
2373 __TCP_INC_STATS(net, TCP_MIB_INERRS);
2374 } else {
2375 tcp_v4_send_reset(NULL, skb);
2376 }
2377
2378discard_it:
2379 SKB_DR_OR(drop_reason, NOT_SPECIFIED);
2380 /* Discard frame. */
2381 kfree_skb_reason(skb, drop_reason);
2382 return 0;
2383
2384discard_and_relse:
2385 sk_drops_add(sk, skb);
2386 if (refcounted)
2387 sock_put(sk);
2388 goto discard_it;
2389
2390do_time_wait:
2391 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
2392 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2393 inet_twsk_put(inet_twsk(sk));
2394 goto discard_it;
2395 }
2396
2397 tcp_v4_fill_cb(skb, iph, th);
2398
2399 if (tcp_checksum_complete(skb)) {
2400 inet_twsk_put(inet_twsk(sk));
2401 goto csum_error;
2402 }
2403 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
2404 case TCP_TW_SYN: {
2405 struct sock *sk2 = inet_lookup_listener(net,
2406 net->ipv4.tcp_death_row.hashinfo,
2407 skb, __tcp_hdrlen(th),
2408 iph->saddr, th->source,
2409 iph->daddr, th->dest,
2410 inet_iif(skb),
2411 sdif);
2412 if (sk2) {
2413 inet_twsk_deschedule_put(inet_twsk(sk));
2414 sk = sk2;
2415 tcp_v4_restore_cb(skb);
2416 refcounted = false;
2417 goto process;
2418 }
2419 }
2420 /* to ACK */
2421 fallthrough;
2422 case TCP_TW_ACK:
2423 tcp_v4_timewait_ack(sk, skb);
2424 break;
2425 case TCP_TW_RST:
2426 tcp_v4_send_reset(sk, skb);
2427 inet_twsk_deschedule_put(inet_twsk(sk));
2428 goto discard_it;
2429 case TCP_TW_SUCCESS:;
2430 }
2431 goto discard_it;
2432}
2433
2434static struct timewait_sock_ops tcp_timewait_sock_ops = {
2435 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
2436 .twsk_unique = tcp_twsk_unique,
2437 .twsk_destructor= tcp_twsk_destructor,
2438};
2439
2440void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
2441{
2442 struct dst_entry *dst = skb_dst(skb);
2443
2444 if (dst && dst_hold_safe(dst)) {
2445 rcu_assign_pointer(sk->sk_rx_dst, dst);
2446 sk->sk_rx_dst_ifindex = skb->skb_iif;
2447 }
2448}
2449EXPORT_SYMBOL(inet_sk_rx_dst_set);
2450
2451const struct inet_connection_sock_af_ops ipv4_specific = {
2452 .queue_xmit = ip_queue_xmit,
2453 .send_check = tcp_v4_send_check,
2454 .rebuild_header = inet_sk_rebuild_header,
2455 .sk_rx_dst_set = inet_sk_rx_dst_set,
2456 .conn_request = tcp_v4_conn_request,
2457 .syn_recv_sock = tcp_v4_syn_recv_sock,
2458 .net_header_len = sizeof(struct iphdr),
2459 .setsockopt = ip_setsockopt,
2460 .getsockopt = ip_getsockopt,
2461 .addr2sockaddr = inet_csk_addr2sockaddr,
2462 .sockaddr_len = sizeof(struct sockaddr_in),
2463 .mtu_reduced = tcp_v4_mtu_reduced,
2464};
2465EXPORT_SYMBOL(ipv4_specific);
2466
2467#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2468static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
2469#ifdef CONFIG_TCP_MD5SIG
2470 .md5_lookup = tcp_v4_md5_lookup,
2471 .calc_md5_hash = tcp_v4_md5_hash_skb,
2472 .md5_parse = tcp_v4_parse_md5_keys,
2473#endif
2474#ifdef CONFIG_TCP_AO
2475 .ao_lookup = tcp_v4_ao_lookup,
2476 .calc_ao_hash = tcp_v4_ao_hash_skb,
2477 .ao_parse = tcp_v4_parse_ao,
2478 .ao_calc_key_sk = tcp_v4_ao_calc_key_sk,
2479#endif
2480};
2481#endif
2482
2483/* NOTE: A lot of things set to zero explicitly by call to
2484 * sk_alloc() so need not be done here.
2485 */
2486static int tcp_v4_init_sock(struct sock *sk)
2487{
2488 struct inet_connection_sock *icsk = inet_csk(sk);
2489
2490 tcp_init_sock(sk);
2491
2492 icsk->icsk_af_ops = &ipv4_specific;
2493
2494#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2495 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
2496#endif
2497
2498 return 0;
2499}
2500
2501#ifdef CONFIG_TCP_MD5SIG
2502static void tcp_md5sig_info_free_rcu(struct rcu_head *head)
2503{
2504 struct tcp_md5sig_info *md5sig;
2505
2506 md5sig = container_of(head, struct tcp_md5sig_info, rcu);
2507 kfree(md5sig);
2508 static_branch_slow_dec_deferred(&tcp_md5_needed);
2509 tcp_md5_release_sigpool();
2510}
2511#endif
2512
2513void tcp_v4_destroy_sock(struct sock *sk)
2514{
2515 struct tcp_sock *tp = tcp_sk(sk);
2516
2517 trace_tcp_destroy_sock(sk);
2518
2519 tcp_clear_xmit_timers(sk);
2520
2521 tcp_cleanup_congestion_control(sk);
2522
2523 tcp_cleanup_ulp(sk);
2524
2525 /* Cleanup up the write buffer. */
2526 tcp_write_queue_purge(sk);
2527
2528 /* Check if we want to disable active TFO */
2529 tcp_fastopen_active_disable_ofo_check(sk);
2530
2531 /* Cleans up our, hopefully empty, out_of_order_queue. */
2532 skb_rbtree_purge(&tp->out_of_order_queue);
2533
2534#ifdef CONFIG_TCP_MD5SIG
2535 /* Clean up the MD5 key list, if any */
2536 if (tp->md5sig_info) {
2537 struct tcp_md5sig_info *md5sig;
2538
2539 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
2540 tcp_clear_md5_list(sk);
2541 call_rcu(&md5sig->rcu, tcp_md5sig_info_free_rcu);
2542 rcu_assign_pointer(tp->md5sig_info, NULL);
2543 }
2544#endif
2545 tcp_ao_destroy_sock(sk, false);
2546
2547 /* Clean up a referenced TCP bind bucket. */
2548 if (inet_csk(sk)->icsk_bind_hash)
2549 inet_put_port(sk);
2550
2551 BUG_ON(rcu_access_pointer(tp->fastopen_rsk));
2552
2553 /* If socket is aborted during connect operation */
2554 tcp_free_fastopen_req(tp);
2555 tcp_fastopen_destroy_cipher(sk);
2556 tcp_saved_syn_free(tp);
2557
2558 sk_sockets_allocated_dec(sk);
2559}
2560EXPORT_SYMBOL(tcp_v4_destroy_sock);
2561
2562#ifdef CONFIG_PROC_FS
2563/* Proc filesystem TCP sock list dumping. */
2564
2565static unsigned short seq_file_family(const struct seq_file *seq);
2566
2567static bool seq_sk_match(struct seq_file *seq, const struct sock *sk)
2568{
2569 unsigned short family = seq_file_family(seq);
2570
2571 /* AF_UNSPEC is used as a match all */
2572 return ((family == AF_UNSPEC || family == sk->sk_family) &&
2573 net_eq(sock_net(sk), seq_file_net(seq)));
2574}
2575
2576/* Find a non empty bucket (starting from st->bucket)
2577 * and return the first sk from it.
2578 */
2579static void *listening_get_first(struct seq_file *seq)
2580{
2581 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2582 struct tcp_iter_state *st = seq->private;
2583
2584 st->offset = 0;
2585 for (; st->bucket <= hinfo->lhash2_mask; st->bucket++) {
2586 struct inet_listen_hashbucket *ilb2;
2587 struct hlist_nulls_node *node;
2588 struct sock *sk;
2589
2590 ilb2 = &hinfo->lhash2[st->bucket];
2591 if (hlist_nulls_empty(&ilb2->nulls_head))
2592 continue;
2593
2594 spin_lock(&ilb2->lock);
2595 sk_nulls_for_each(sk, node, &ilb2->nulls_head) {
2596 if (seq_sk_match(seq, sk))
2597 return sk;
2598 }
2599 spin_unlock(&ilb2->lock);
2600 }
2601
2602 return NULL;
2603}
2604
2605/* Find the next sk of "cur" within the same bucket (i.e. st->bucket).
2606 * If "cur" is the last one in the st->bucket,
2607 * call listening_get_first() to return the first sk of the next
2608 * non empty bucket.
2609 */
2610static void *listening_get_next(struct seq_file *seq, void *cur)
2611{
2612 struct tcp_iter_state *st = seq->private;
2613 struct inet_listen_hashbucket *ilb2;
2614 struct hlist_nulls_node *node;
2615 struct inet_hashinfo *hinfo;
2616 struct sock *sk = cur;
2617
2618 ++st->num;
2619 ++st->offset;
2620
2621 sk = sk_nulls_next(sk);
2622 sk_nulls_for_each_from(sk, node) {
2623 if (seq_sk_match(seq, sk))
2624 return sk;
2625 }
2626
2627 hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2628 ilb2 = &hinfo->lhash2[st->bucket];
2629 spin_unlock(&ilb2->lock);
2630 ++st->bucket;
2631 return listening_get_first(seq);
2632}
2633
2634static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2635{
2636 struct tcp_iter_state *st = seq->private;
2637 void *rc;
2638
2639 st->bucket = 0;
2640 st->offset = 0;
2641 rc = listening_get_first(seq);
2642
2643 while (rc && *pos) {
2644 rc = listening_get_next(seq, rc);
2645 --*pos;
2646 }
2647 return rc;
2648}
2649
2650static inline bool empty_bucket(struct inet_hashinfo *hinfo,
2651 const struct tcp_iter_state *st)
2652{
2653 return hlist_nulls_empty(&hinfo->ehash[st->bucket].chain);
2654}
2655
2656/*
2657 * Get first established socket starting from bucket given in st->bucket.
2658 * If st->bucket is zero, the very first socket in the hash is returned.
2659 */
2660static void *established_get_first(struct seq_file *seq)
2661{
2662 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2663 struct tcp_iter_state *st = seq->private;
2664
2665 st->offset = 0;
2666 for (; st->bucket <= hinfo->ehash_mask; ++st->bucket) {
2667 struct sock *sk;
2668 struct hlist_nulls_node *node;
2669 spinlock_t *lock = inet_ehash_lockp(hinfo, st->bucket);
2670
2671 cond_resched();
2672
2673 /* Lockless fast path for the common case of empty buckets */
2674 if (empty_bucket(hinfo, st))
2675 continue;
2676
2677 spin_lock_bh(lock);
2678 sk_nulls_for_each(sk, node, &hinfo->ehash[st->bucket].chain) {
2679 if (seq_sk_match(seq, sk))
2680 return sk;
2681 }
2682 spin_unlock_bh(lock);
2683 }
2684
2685 return NULL;
2686}
2687
2688static void *established_get_next(struct seq_file *seq, void *cur)
2689{
2690 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2691 struct tcp_iter_state *st = seq->private;
2692 struct hlist_nulls_node *node;
2693 struct sock *sk = cur;
2694
2695 ++st->num;
2696 ++st->offset;
2697
2698 sk = sk_nulls_next(sk);
2699
2700 sk_nulls_for_each_from(sk, node) {
2701 if (seq_sk_match(seq, sk))
2702 return sk;
2703 }
2704
2705 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2706 ++st->bucket;
2707 return established_get_first(seq);
2708}
2709
2710static void *established_get_idx(struct seq_file *seq, loff_t pos)
2711{
2712 struct tcp_iter_state *st = seq->private;
2713 void *rc;
2714
2715 st->bucket = 0;
2716 rc = established_get_first(seq);
2717
2718 while (rc && pos) {
2719 rc = established_get_next(seq, rc);
2720 --pos;
2721 }
2722 return rc;
2723}
2724
2725static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2726{
2727 void *rc;
2728 struct tcp_iter_state *st = seq->private;
2729
2730 st->state = TCP_SEQ_STATE_LISTENING;
2731 rc = listening_get_idx(seq, &pos);
2732
2733 if (!rc) {
2734 st->state = TCP_SEQ_STATE_ESTABLISHED;
2735 rc = established_get_idx(seq, pos);
2736 }
2737
2738 return rc;
2739}
2740
2741static void *tcp_seek_last_pos(struct seq_file *seq)
2742{
2743 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2744 struct tcp_iter_state *st = seq->private;
2745 int bucket = st->bucket;
2746 int offset = st->offset;
2747 int orig_num = st->num;
2748 void *rc = NULL;
2749
2750 switch (st->state) {
2751 case TCP_SEQ_STATE_LISTENING:
2752 if (st->bucket > hinfo->lhash2_mask)
2753 break;
2754 rc = listening_get_first(seq);
2755 while (offset-- && rc && bucket == st->bucket)
2756 rc = listening_get_next(seq, rc);
2757 if (rc)
2758 break;
2759 st->bucket = 0;
2760 st->state = TCP_SEQ_STATE_ESTABLISHED;
2761 fallthrough;
2762 case TCP_SEQ_STATE_ESTABLISHED:
2763 if (st->bucket > hinfo->ehash_mask)
2764 break;
2765 rc = established_get_first(seq);
2766 while (offset-- && rc && bucket == st->bucket)
2767 rc = established_get_next(seq, rc);
2768 }
2769
2770 st->num = orig_num;
2771
2772 return rc;
2773}
2774
2775void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2776{
2777 struct tcp_iter_state *st = seq->private;
2778 void *rc;
2779
2780 if (*pos && *pos == st->last_pos) {
2781 rc = tcp_seek_last_pos(seq);
2782 if (rc)
2783 goto out;
2784 }
2785
2786 st->state = TCP_SEQ_STATE_LISTENING;
2787 st->num = 0;
2788 st->bucket = 0;
2789 st->offset = 0;
2790 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2791
2792out:
2793 st->last_pos = *pos;
2794 return rc;
2795}
2796EXPORT_SYMBOL(tcp_seq_start);
2797
2798void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2799{
2800 struct tcp_iter_state *st = seq->private;
2801 void *rc = NULL;
2802
2803 if (v == SEQ_START_TOKEN) {
2804 rc = tcp_get_idx(seq, 0);
2805 goto out;
2806 }
2807
2808 switch (st->state) {
2809 case TCP_SEQ_STATE_LISTENING:
2810 rc = listening_get_next(seq, v);
2811 if (!rc) {
2812 st->state = TCP_SEQ_STATE_ESTABLISHED;
2813 st->bucket = 0;
2814 st->offset = 0;
2815 rc = established_get_first(seq);
2816 }
2817 break;
2818 case TCP_SEQ_STATE_ESTABLISHED:
2819 rc = established_get_next(seq, v);
2820 break;
2821 }
2822out:
2823 ++*pos;
2824 st->last_pos = *pos;
2825 return rc;
2826}
2827EXPORT_SYMBOL(tcp_seq_next);
2828
2829void tcp_seq_stop(struct seq_file *seq, void *v)
2830{
2831 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2832 struct tcp_iter_state *st = seq->private;
2833
2834 switch (st->state) {
2835 case TCP_SEQ_STATE_LISTENING:
2836 if (v != SEQ_START_TOKEN)
2837 spin_unlock(&hinfo->lhash2[st->bucket].lock);
2838 break;
2839 case TCP_SEQ_STATE_ESTABLISHED:
2840 if (v)
2841 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2842 break;
2843 }
2844}
2845EXPORT_SYMBOL(tcp_seq_stop);
2846
2847static void get_openreq4(const struct request_sock *req,
2848 struct seq_file *f, int i)
2849{
2850 const struct inet_request_sock *ireq = inet_rsk(req);
2851 long delta = req->rsk_timer.expires - jiffies;
2852
2853 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2854 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2855 i,
2856 ireq->ir_loc_addr,
2857 ireq->ir_num,
2858 ireq->ir_rmt_addr,
2859 ntohs(ireq->ir_rmt_port),
2860 TCP_SYN_RECV,
2861 0, 0, /* could print option size, but that is af dependent. */
2862 1, /* timers active (only the expire timer) */
2863 jiffies_delta_to_clock_t(delta),
2864 req->num_timeout,
2865 from_kuid_munged(seq_user_ns(f),
2866 sock_i_uid(req->rsk_listener)),
2867 0, /* non standard timer */
2868 0, /* open_requests have no inode */
2869 0,
2870 req);
2871}
2872
2873static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2874{
2875 int timer_active;
2876 unsigned long timer_expires;
2877 const struct tcp_sock *tp = tcp_sk(sk);
2878 const struct inet_connection_sock *icsk = inet_csk(sk);
2879 const struct inet_sock *inet = inet_sk(sk);
2880 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2881 __be32 dest = inet->inet_daddr;
2882 __be32 src = inet->inet_rcv_saddr;
2883 __u16 destp = ntohs(inet->inet_dport);
2884 __u16 srcp = ntohs(inet->inet_sport);
2885 int rx_queue;
2886 int state;
2887
2888 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2889 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2890 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2891 timer_active = 1;
2892 timer_expires = icsk->icsk_timeout;
2893 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2894 timer_active = 4;
2895 timer_expires = icsk->icsk_timeout;
2896 } else if (timer_pending(&sk->sk_timer)) {
2897 timer_active = 2;
2898 timer_expires = sk->sk_timer.expires;
2899 } else {
2900 timer_active = 0;
2901 timer_expires = jiffies;
2902 }
2903
2904 state = inet_sk_state_load(sk);
2905 if (state == TCP_LISTEN)
2906 rx_queue = READ_ONCE(sk->sk_ack_backlog);
2907 else
2908 /* Because we don't lock the socket,
2909 * we might find a transient negative value.
2910 */
2911 rx_queue = max_t(int, READ_ONCE(tp->rcv_nxt) -
2912 READ_ONCE(tp->copied_seq), 0);
2913
2914 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2915 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2916 i, src, srcp, dest, destp, state,
2917 READ_ONCE(tp->write_seq) - tp->snd_una,
2918 rx_queue,
2919 timer_active,
2920 jiffies_delta_to_clock_t(timer_expires - jiffies),
2921 icsk->icsk_retransmits,
2922 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2923 icsk->icsk_probes_out,
2924 sock_i_ino(sk),
2925 refcount_read(&sk->sk_refcnt), sk,
2926 jiffies_to_clock_t(icsk->icsk_rto),
2927 jiffies_to_clock_t(icsk->icsk_ack.ato),
2928 (icsk->icsk_ack.quick << 1) | inet_csk_in_pingpong_mode(sk),
2929 tcp_snd_cwnd(tp),
2930 state == TCP_LISTEN ?
2931 fastopenq->max_qlen :
2932 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2933}
2934
2935static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2936 struct seq_file *f, int i)
2937{
2938 long delta = tw->tw_timer.expires - jiffies;
2939 __be32 dest, src;
2940 __u16 destp, srcp;
2941
2942 dest = tw->tw_daddr;
2943 src = tw->tw_rcv_saddr;
2944 destp = ntohs(tw->tw_dport);
2945 srcp = ntohs(tw->tw_sport);
2946
2947 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2948 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2949 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2950 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2951 refcount_read(&tw->tw_refcnt), tw);
2952}
2953
2954#define TMPSZ 150
2955
2956static int tcp4_seq_show(struct seq_file *seq, void *v)
2957{
2958 struct tcp_iter_state *st;
2959 struct sock *sk = v;
2960
2961 seq_setwidth(seq, TMPSZ - 1);
2962 if (v == SEQ_START_TOKEN) {
2963 seq_puts(seq, " sl local_address rem_address st tx_queue "
2964 "rx_queue tr tm->when retrnsmt uid timeout "
2965 "inode");
2966 goto out;
2967 }
2968 st = seq->private;
2969
2970 if (sk->sk_state == TCP_TIME_WAIT)
2971 get_timewait4_sock(v, seq, st->num);
2972 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2973 get_openreq4(v, seq, st->num);
2974 else
2975 get_tcp4_sock(v, seq, st->num);
2976out:
2977 seq_pad(seq, '\n');
2978 return 0;
2979}
2980
2981#ifdef CONFIG_BPF_SYSCALL
2982struct bpf_tcp_iter_state {
2983 struct tcp_iter_state state;
2984 unsigned int cur_sk;
2985 unsigned int end_sk;
2986 unsigned int max_sk;
2987 struct sock **batch;
2988 bool st_bucket_done;
2989};
2990
2991struct bpf_iter__tcp {
2992 __bpf_md_ptr(struct bpf_iter_meta *, meta);
2993 __bpf_md_ptr(struct sock_common *, sk_common);
2994 uid_t uid __aligned(8);
2995};
2996
2997static int tcp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta,
2998 struct sock_common *sk_common, uid_t uid)
2999{
3000 struct bpf_iter__tcp ctx;
3001
3002 meta->seq_num--; /* skip SEQ_START_TOKEN */
3003 ctx.meta = meta;
3004 ctx.sk_common = sk_common;
3005 ctx.uid = uid;
3006 return bpf_iter_run_prog(prog, &ctx);
3007}
3008
3009static void bpf_iter_tcp_put_batch(struct bpf_tcp_iter_state *iter)
3010{
3011 while (iter->cur_sk < iter->end_sk)
3012 sock_gen_put(iter->batch[iter->cur_sk++]);
3013}
3014
3015static int bpf_iter_tcp_realloc_batch(struct bpf_tcp_iter_state *iter,
3016 unsigned int new_batch_sz)
3017{
3018 struct sock **new_batch;
3019
3020 new_batch = kvmalloc(sizeof(*new_batch) * new_batch_sz,
3021 GFP_USER | __GFP_NOWARN);
3022 if (!new_batch)
3023 return -ENOMEM;
3024
3025 bpf_iter_tcp_put_batch(iter);
3026 kvfree(iter->batch);
3027 iter->batch = new_batch;
3028 iter->max_sk = new_batch_sz;
3029
3030 return 0;
3031}
3032
3033static unsigned int bpf_iter_tcp_listening_batch(struct seq_file *seq,
3034 struct sock *start_sk)
3035{
3036 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3037 struct bpf_tcp_iter_state *iter = seq->private;
3038 struct tcp_iter_state *st = &iter->state;
3039 struct hlist_nulls_node *node;
3040 unsigned int expected = 1;
3041 struct sock *sk;
3042
3043 sock_hold(start_sk);
3044 iter->batch[iter->end_sk++] = start_sk;
3045
3046 sk = sk_nulls_next(start_sk);
3047 sk_nulls_for_each_from(sk, node) {
3048 if (seq_sk_match(seq, sk)) {
3049 if (iter->end_sk < iter->max_sk) {
3050 sock_hold(sk);
3051 iter->batch[iter->end_sk++] = sk;
3052 }
3053 expected++;
3054 }
3055 }
3056 spin_unlock(&hinfo->lhash2[st->bucket].lock);
3057
3058 return expected;
3059}
3060
3061static unsigned int bpf_iter_tcp_established_batch(struct seq_file *seq,
3062 struct sock *start_sk)
3063{
3064 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3065 struct bpf_tcp_iter_state *iter = seq->private;
3066 struct tcp_iter_state *st = &iter->state;
3067 struct hlist_nulls_node *node;
3068 unsigned int expected = 1;
3069 struct sock *sk;
3070
3071 sock_hold(start_sk);
3072 iter->batch[iter->end_sk++] = start_sk;
3073
3074 sk = sk_nulls_next(start_sk);
3075 sk_nulls_for_each_from(sk, node) {
3076 if (seq_sk_match(seq, sk)) {
3077 if (iter->end_sk < iter->max_sk) {
3078 sock_hold(sk);
3079 iter->batch[iter->end_sk++] = sk;
3080 }
3081 expected++;
3082 }
3083 }
3084 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
3085
3086 return expected;
3087}
3088
3089static struct sock *bpf_iter_tcp_batch(struct seq_file *seq)
3090{
3091 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3092 struct bpf_tcp_iter_state *iter = seq->private;
3093 struct tcp_iter_state *st = &iter->state;
3094 unsigned int expected;
3095 bool resized = false;
3096 struct sock *sk;
3097
3098 /* The st->bucket is done. Directly advance to the next
3099 * bucket instead of having the tcp_seek_last_pos() to skip
3100 * one by one in the current bucket and eventually find out
3101 * it has to advance to the next bucket.
3102 */
3103 if (iter->st_bucket_done) {
3104 st->offset = 0;
3105 st->bucket++;
3106 if (st->state == TCP_SEQ_STATE_LISTENING &&
3107 st->bucket > hinfo->lhash2_mask) {
3108 st->state = TCP_SEQ_STATE_ESTABLISHED;
3109 st->bucket = 0;
3110 }
3111 }
3112
3113again:
3114 /* Get a new batch */
3115 iter->cur_sk = 0;
3116 iter->end_sk = 0;
3117 iter->st_bucket_done = false;
3118
3119 sk = tcp_seek_last_pos(seq);
3120 if (!sk)
3121 return NULL; /* Done */
3122
3123 if (st->state == TCP_SEQ_STATE_LISTENING)
3124 expected = bpf_iter_tcp_listening_batch(seq, sk);
3125 else
3126 expected = bpf_iter_tcp_established_batch(seq, sk);
3127
3128 if (iter->end_sk == expected) {
3129 iter->st_bucket_done = true;
3130 return sk;
3131 }
3132
3133 if (!resized && !bpf_iter_tcp_realloc_batch(iter, expected * 3 / 2)) {
3134 resized = true;
3135 goto again;
3136 }
3137
3138 return sk;
3139}
3140
3141static void *bpf_iter_tcp_seq_start(struct seq_file *seq, loff_t *pos)
3142{
3143 /* bpf iter does not support lseek, so it always
3144 * continue from where it was stop()-ped.
3145 */
3146 if (*pos)
3147 return bpf_iter_tcp_batch(seq);
3148
3149 return SEQ_START_TOKEN;
3150}
3151
3152static void *bpf_iter_tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
3153{
3154 struct bpf_tcp_iter_state *iter = seq->private;
3155 struct tcp_iter_state *st = &iter->state;
3156 struct sock *sk;
3157
3158 /* Whenever seq_next() is called, the iter->cur_sk is
3159 * done with seq_show(), so advance to the next sk in
3160 * the batch.
3161 */
3162 if (iter->cur_sk < iter->end_sk) {
3163 /* Keeping st->num consistent in tcp_iter_state.
3164 * bpf_iter_tcp does not use st->num.
3165 * meta.seq_num is used instead.
3166 */
3167 st->num++;
3168 /* Move st->offset to the next sk in the bucket such that
3169 * the future start() will resume at st->offset in
3170 * st->bucket. See tcp_seek_last_pos().
3171 */
3172 st->offset++;
3173 sock_gen_put(iter->batch[iter->cur_sk++]);
3174 }
3175
3176 if (iter->cur_sk < iter->end_sk)
3177 sk = iter->batch[iter->cur_sk];
3178 else
3179 sk = bpf_iter_tcp_batch(seq);
3180
3181 ++*pos;
3182 /* Keeping st->last_pos consistent in tcp_iter_state.
3183 * bpf iter does not do lseek, so st->last_pos always equals to *pos.
3184 */
3185 st->last_pos = *pos;
3186 return sk;
3187}
3188
3189static int bpf_iter_tcp_seq_show(struct seq_file *seq, void *v)
3190{
3191 struct bpf_iter_meta meta;
3192 struct bpf_prog *prog;
3193 struct sock *sk = v;
3194 uid_t uid;
3195 int ret;
3196
3197 if (v == SEQ_START_TOKEN)
3198 return 0;
3199
3200 if (sk_fullsock(sk))
3201 lock_sock(sk);
3202
3203 if (unlikely(sk_unhashed(sk))) {
3204 ret = SEQ_SKIP;
3205 goto unlock;
3206 }
3207
3208 if (sk->sk_state == TCP_TIME_WAIT) {
3209 uid = 0;
3210 } else if (sk->sk_state == TCP_NEW_SYN_RECV) {
3211 const struct request_sock *req = v;
3212
3213 uid = from_kuid_munged(seq_user_ns(seq),
3214 sock_i_uid(req->rsk_listener));
3215 } else {
3216 uid = from_kuid_munged(seq_user_ns(seq), sock_i_uid(sk));
3217 }
3218
3219 meta.seq = seq;
3220 prog = bpf_iter_get_info(&meta, false);
3221 ret = tcp_prog_seq_show(prog, &meta, v, uid);
3222
3223unlock:
3224 if (sk_fullsock(sk))
3225 release_sock(sk);
3226 return ret;
3227
3228}
3229
3230static void bpf_iter_tcp_seq_stop(struct seq_file *seq, void *v)
3231{
3232 struct bpf_tcp_iter_state *iter = seq->private;
3233 struct bpf_iter_meta meta;
3234 struct bpf_prog *prog;
3235
3236 if (!v) {
3237 meta.seq = seq;
3238 prog = bpf_iter_get_info(&meta, true);
3239 if (prog)
3240 (void)tcp_prog_seq_show(prog, &meta, v, 0);
3241 }
3242
3243 if (iter->cur_sk < iter->end_sk) {
3244 bpf_iter_tcp_put_batch(iter);
3245 iter->st_bucket_done = false;
3246 }
3247}
3248
3249static const struct seq_operations bpf_iter_tcp_seq_ops = {
3250 .show = bpf_iter_tcp_seq_show,
3251 .start = bpf_iter_tcp_seq_start,
3252 .next = bpf_iter_tcp_seq_next,
3253 .stop = bpf_iter_tcp_seq_stop,
3254};
3255#endif
3256static unsigned short seq_file_family(const struct seq_file *seq)
3257{
3258 const struct tcp_seq_afinfo *afinfo;
3259
3260#ifdef CONFIG_BPF_SYSCALL
3261 /* Iterated from bpf_iter. Let the bpf prog to filter instead. */
3262 if (seq->op == &bpf_iter_tcp_seq_ops)
3263 return AF_UNSPEC;
3264#endif
3265
3266 /* Iterated from proc fs */
3267 afinfo = pde_data(file_inode(seq->file));
3268 return afinfo->family;
3269}
3270
3271static const struct seq_operations tcp4_seq_ops = {
3272 .show = tcp4_seq_show,
3273 .start = tcp_seq_start,
3274 .next = tcp_seq_next,
3275 .stop = tcp_seq_stop,
3276};
3277
3278static struct tcp_seq_afinfo tcp4_seq_afinfo = {
3279 .family = AF_INET,
3280};
3281
3282static int __net_init tcp4_proc_init_net(struct net *net)
3283{
3284 if (!proc_create_net_data("tcp", 0444, net->proc_net, &tcp4_seq_ops,
3285 sizeof(struct tcp_iter_state), &tcp4_seq_afinfo))
3286 return -ENOMEM;
3287 return 0;
3288}
3289
3290static void __net_exit tcp4_proc_exit_net(struct net *net)
3291{
3292 remove_proc_entry("tcp", net->proc_net);
3293}
3294
3295static struct pernet_operations tcp4_net_ops = {
3296 .init = tcp4_proc_init_net,
3297 .exit = tcp4_proc_exit_net,
3298};
3299
3300int __init tcp4_proc_init(void)
3301{
3302 return register_pernet_subsys(&tcp4_net_ops);
3303}
3304
3305void tcp4_proc_exit(void)
3306{
3307 unregister_pernet_subsys(&tcp4_net_ops);
3308}
3309#endif /* CONFIG_PROC_FS */
3310
3311/* @wake is one when sk_stream_write_space() calls us.
3312 * This sends EPOLLOUT only if notsent_bytes is half the limit.
3313 * This mimics the strategy used in sock_def_write_space().
3314 */
3315bool tcp_stream_memory_free(const struct sock *sk, int wake)
3316{
3317 const struct tcp_sock *tp = tcp_sk(sk);
3318 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
3319 READ_ONCE(tp->snd_nxt);
3320
3321 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
3322}
3323EXPORT_SYMBOL(tcp_stream_memory_free);
3324
3325struct proto tcp_prot = {
3326 .name = "TCP",
3327 .owner = THIS_MODULE,
3328 .close = tcp_close,
3329 .pre_connect = tcp_v4_pre_connect,
3330 .connect = tcp_v4_connect,
3331 .disconnect = tcp_disconnect,
3332 .accept = inet_csk_accept,
3333 .ioctl = tcp_ioctl,
3334 .init = tcp_v4_init_sock,
3335 .destroy = tcp_v4_destroy_sock,
3336 .shutdown = tcp_shutdown,
3337 .setsockopt = tcp_setsockopt,
3338 .getsockopt = tcp_getsockopt,
3339 .bpf_bypass_getsockopt = tcp_bpf_bypass_getsockopt,
3340 .keepalive = tcp_set_keepalive,
3341 .recvmsg = tcp_recvmsg,
3342 .sendmsg = tcp_sendmsg,
3343 .splice_eof = tcp_splice_eof,
3344 .backlog_rcv = tcp_v4_do_rcv,
3345 .release_cb = tcp_release_cb,
3346 .hash = inet_hash,
3347 .unhash = inet_unhash,
3348 .get_port = inet_csk_get_port,
3349 .put_port = inet_put_port,
3350#ifdef CONFIG_BPF_SYSCALL
3351 .psock_update_sk_prot = tcp_bpf_update_proto,
3352#endif
3353 .enter_memory_pressure = tcp_enter_memory_pressure,
3354 .leave_memory_pressure = tcp_leave_memory_pressure,
3355 .stream_memory_free = tcp_stream_memory_free,
3356 .sockets_allocated = &tcp_sockets_allocated,
3357 .orphan_count = &tcp_orphan_count,
3358
3359 .memory_allocated = &tcp_memory_allocated,
3360 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3361
3362 .memory_pressure = &tcp_memory_pressure,
3363 .sysctl_mem = sysctl_tcp_mem,
3364 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3365 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3366 .max_header = MAX_TCP_HEADER,
3367 .obj_size = sizeof(struct tcp_sock),
3368 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3369 .twsk_prot = &tcp_timewait_sock_ops,
3370 .rsk_prot = &tcp_request_sock_ops,
3371 .h.hashinfo = NULL,
3372 .no_autobind = true,
3373 .diag_destroy = tcp_abort,
3374};
3375EXPORT_SYMBOL(tcp_prot);
3376
3377static void __net_exit tcp_sk_exit(struct net *net)
3378{
3379 if (net->ipv4.tcp_congestion_control)
3380 bpf_module_put(net->ipv4.tcp_congestion_control,
3381 net->ipv4.tcp_congestion_control->owner);
3382}
3383
3384static void __net_init tcp_set_hashinfo(struct net *net)
3385{
3386 struct inet_hashinfo *hinfo;
3387 unsigned int ehash_entries;
3388 struct net *old_net;
3389
3390 if (net_eq(net, &init_net))
3391 goto fallback;
3392
3393 old_net = current->nsproxy->net_ns;
3394 ehash_entries = READ_ONCE(old_net->ipv4.sysctl_tcp_child_ehash_entries);
3395 if (!ehash_entries)
3396 goto fallback;
3397
3398 ehash_entries = roundup_pow_of_two(ehash_entries);
3399 hinfo = inet_pernet_hashinfo_alloc(&tcp_hashinfo, ehash_entries);
3400 if (!hinfo) {
3401 pr_warn("Failed to allocate TCP ehash (entries: %u) "
3402 "for a netns, fallback to the global one\n",
3403 ehash_entries);
3404fallback:
3405 hinfo = &tcp_hashinfo;
3406 ehash_entries = tcp_hashinfo.ehash_mask + 1;
3407 }
3408
3409 net->ipv4.tcp_death_row.hashinfo = hinfo;
3410 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = ehash_entries / 2;
3411 net->ipv4.sysctl_max_syn_backlog = max(128U, ehash_entries / 128);
3412}
3413
3414static int __net_init tcp_sk_init(struct net *net)
3415{
3416 net->ipv4.sysctl_tcp_ecn = 2;
3417 net->ipv4.sysctl_tcp_ecn_fallback = 1;
3418
3419 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
3420 net->ipv4.sysctl_tcp_min_snd_mss = TCP_MIN_SND_MSS;
3421 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
3422 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
3423 net->ipv4.sysctl_tcp_mtu_probe_floor = TCP_MIN_SND_MSS;
3424
3425 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
3426 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
3427 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
3428
3429 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
3430 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
3431 net->ipv4.sysctl_tcp_syncookies = 1;
3432 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
3433 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
3434 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
3435 net->ipv4.sysctl_tcp_orphan_retries = 0;
3436 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
3437 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
3438 net->ipv4.sysctl_tcp_tw_reuse = 2;
3439 net->ipv4.sysctl_tcp_no_ssthresh_metrics_save = 1;
3440
3441 refcount_set(&net->ipv4.tcp_death_row.tw_refcount, 1);
3442 tcp_set_hashinfo(net);
3443
3444 net->ipv4.sysctl_tcp_sack = 1;
3445 net->ipv4.sysctl_tcp_window_scaling = 1;
3446 net->ipv4.sysctl_tcp_timestamps = 1;
3447 net->ipv4.sysctl_tcp_early_retrans = 3;
3448 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
3449 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
3450 net->ipv4.sysctl_tcp_retrans_collapse = 1;
3451 net->ipv4.sysctl_tcp_max_reordering = 300;
3452 net->ipv4.sysctl_tcp_dsack = 1;
3453 net->ipv4.sysctl_tcp_app_win = 31;
3454 net->ipv4.sysctl_tcp_adv_win_scale = 1;
3455 net->ipv4.sysctl_tcp_frto = 2;
3456 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
3457 /* This limits the percentage of the congestion window which we
3458 * will allow a single TSO frame to consume. Building TSO frames
3459 * which are too large can cause TCP streams to be bursty.
3460 */
3461 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
3462 /* Default TSQ limit of 16 TSO segments */
3463 net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
3464
3465 /* rfc5961 challenge ack rate limiting, per net-ns, disabled by default. */
3466 net->ipv4.sysctl_tcp_challenge_ack_limit = INT_MAX;
3467
3468 net->ipv4.sysctl_tcp_min_tso_segs = 2;
3469 net->ipv4.sysctl_tcp_tso_rtt_log = 9; /* 2^9 = 512 usec */
3470 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
3471 net->ipv4.sysctl_tcp_autocorking = 1;
3472 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
3473 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
3474 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
3475 if (net != &init_net) {
3476 memcpy(net->ipv4.sysctl_tcp_rmem,
3477 init_net.ipv4.sysctl_tcp_rmem,
3478 sizeof(init_net.ipv4.sysctl_tcp_rmem));
3479 memcpy(net->ipv4.sysctl_tcp_wmem,
3480 init_net.ipv4.sysctl_tcp_wmem,
3481 sizeof(init_net.ipv4.sysctl_tcp_wmem));
3482 }
3483 net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
3484 net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
3485 net->ipv4.sysctl_tcp_comp_sack_nr = 44;
3486 net->ipv4.sysctl_tcp_backlog_ack_defer = 1;
3487 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
3488 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
3489 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
3490
3491 /* Set default values for PLB */
3492 net->ipv4.sysctl_tcp_plb_enabled = 0; /* Disabled by default */
3493 net->ipv4.sysctl_tcp_plb_idle_rehash_rounds = 3;
3494 net->ipv4.sysctl_tcp_plb_rehash_rounds = 12;
3495 net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60;
3496 /* Default congestion threshold for PLB to mark a round is 50% */
3497 net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2;
3498
3499 /* Reno is always built in */
3500 if (!net_eq(net, &init_net) &&
3501 bpf_try_module_get(init_net.ipv4.tcp_congestion_control,
3502 init_net.ipv4.tcp_congestion_control->owner))
3503 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
3504 else
3505 net->ipv4.tcp_congestion_control = &tcp_reno;
3506
3507 net->ipv4.sysctl_tcp_syn_linear_timeouts = 4;
3508 net->ipv4.sysctl_tcp_shrink_window = 0;
3509
3510 net->ipv4.sysctl_tcp_pingpong_thresh = 1;
3511
3512 return 0;
3513}
3514
3515static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
3516{
3517 struct net *net;
3518
3519 tcp_twsk_purge(net_exit_list, AF_INET);
3520
3521 list_for_each_entry(net, net_exit_list, exit_list) {
3522 inet_pernet_hashinfo_free(net->ipv4.tcp_death_row.hashinfo);
3523 WARN_ON_ONCE(!refcount_dec_and_test(&net->ipv4.tcp_death_row.tw_refcount));
3524 tcp_fastopen_ctx_destroy(net);
3525 }
3526}
3527
3528static struct pernet_operations __net_initdata tcp_sk_ops = {
3529 .init = tcp_sk_init,
3530 .exit = tcp_sk_exit,
3531 .exit_batch = tcp_sk_exit_batch,
3532};
3533
3534#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3535DEFINE_BPF_ITER_FUNC(tcp, struct bpf_iter_meta *meta,
3536 struct sock_common *sk_common, uid_t uid)
3537
3538#define INIT_BATCH_SZ 16
3539
3540static int bpf_iter_init_tcp(void *priv_data, struct bpf_iter_aux_info *aux)
3541{
3542 struct bpf_tcp_iter_state *iter = priv_data;
3543 int err;
3544
3545 err = bpf_iter_init_seq_net(priv_data, aux);
3546 if (err)
3547 return err;
3548
3549 err = bpf_iter_tcp_realloc_batch(iter, INIT_BATCH_SZ);
3550 if (err) {
3551 bpf_iter_fini_seq_net(priv_data);
3552 return err;
3553 }
3554
3555 return 0;
3556}
3557
3558static void bpf_iter_fini_tcp(void *priv_data)
3559{
3560 struct bpf_tcp_iter_state *iter = priv_data;
3561
3562 bpf_iter_fini_seq_net(priv_data);
3563 kvfree(iter->batch);
3564}
3565
3566static const struct bpf_iter_seq_info tcp_seq_info = {
3567 .seq_ops = &bpf_iter_tcp_seq_ops,
3568 .init_seq_private = bpf_iter_init_tcp,
3569 .fini_seq_private = bpf_iter_fini_tcp,
3570 .seq_priv_size = sizeof(struct bpf_tcp_iter_state),
3571};
3572
3573static const struct bpf_func_proto *
3574bpf_iter_tcp_get_func_proto(enum bpf_func_id func_id,
3575 const struct bpf_prog *prog)
3576{
3577 switch (func_id) {
3578 case BPF_FUNC_setsockopt:
3579 return &bpf_sk_setsockopt_proto;
3580 case BPF_FUNC_getsockopt:
3581 return &bpf_sk_getsockopt_proto;
3582 default:
3583 return NULL;
3584 }
3585}
3586
3587static struct bpf_iter_reg tcp_reg_info = {
3588 .target = "tcp",
3589 .ctx_arg_info_size = 1,
3590 .ctx_arg_info = {
3591 { offsetof(struct bpf_iter__tcp, sk_common),
3592 PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
3593 },
3594 .get_func_proto = bpf_iter_tcp_get_func_proto,
3595 .seq_info = &tcp_seq_info,
3596};
3597
3598static void __init bpf_iter_register(void)
3599{
3600 tcp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON];
3601 if (bpf_iter_reg_target(&tcp_reg_info))
3602 pr_warn("Warning: could not register bpf iterator tcp\n");
3603}
3604
3605#endif
3606
3607void __init tcp_v4_init(void)
3608{
3609 int cpu, res;
3610
3611 for_each_possible_cpu(cpu) {
3612 struct sock *sk;
3613
3614 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
3615 IPPROTO_TCP, &init_net);
3616 if (res)
3617 panic("Failed to create the TCP control socket.\n");
3618 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
3619
3620 /* Please enforce IP_DF and IPID==0 for RST and
3621 * ACK sent in SYN-RECV and TIME-WAIT state.
3622 */
3623 inet_sk(sk)->pmtudisc = IP_PMTUDISC_DO;
3624
3625 per_cpu(ipv4_tcp_sk, cpu) = sk;
3626 }
3627 if (register_pernet_subsys(&tcp_sk_ops))
3628 panic("Failed to create the TCP control socket.\n");
3629
3630#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3631 bpf_iter_register();
3632#endif
3633}