1 // SPDX-License-Identifier: GPL-2.0-only
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  * Authors:	Ross Biro
10  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
12  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
13  *		Florian La Roche, <flla@stud.uni-sb.de>
14  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
16  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
17  *		Matthew Dillon, <dillon@apollo.west.oic.com>
18  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19  *		Jorge Cwik, <jorge@laser.satlink.net>
20  */
21 
22 #include <linux/module.h>
23 #include <linux/gfp.h>
24 #include <net/tcp.h>
25 
tcp_clamp_rto_to_user_timeout(const struct sock * sk)26 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
27 {
28 	struct inet_connection_sock *icsk = inet_csk(sk);
29 	u32 elapsed, start_ts, user_timeout;
30 	s32 remaining;
31 
32 	start_ts = tcp_sk(sk)->retrans_stamp;
33 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
34 	if (!user_timeout)
35 		return icsk->icsk_rto;
36 	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
37 	remaining = user_timeout - elapsed;
38 	if (remaining <= 0)
39 		return 1; /* user timeout has passed; fire ASAP */
40 
41 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
42 }
43 
tcp_clamp_probe0_to_user_timeout(const struct sock * sk,u32 when)44 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
45 {
46 	struct inet_connection_sock *icsk = inet_csk(sk);
47 	u32 remaining, user_timeout;
48 	s32 elapsed;
49 
50 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
51 	if (!user_timeout || !icsk->icsk_probes_tstamp)
52 		return when;
53 
54 	elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
55 	if (unlikely(elapsed < 0))
56 		elapsed = 0;
57 	remaining = msecs_to_jiffies(user_timeout) - elapsed;
58 	remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
59 
60 	return min_t(u32, remaining, when);
61 }
62 
63 /**
64  *  tcp_write_err() - close socket and save error info
65  *  @sk:  The socket the error has appeared on.
66  *
67  *  Returns: Nothing (void)
68  */
69 
tcp_write_err(struct sock * sk)70 static void tcp_write_err(struct sock *sk)
71 {
72 	WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
73 	sk_error_report(sk);
74 
75 	tcp_write_queue_purge(sk);
76 	tcp_done(sk);
77 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
78 }
79 
80 /**
81  *  tcp_out_of_resources() - Close socket if out of resources
82  *  @sk:        pointer to current socket
83  *  @do_reset:  send a last packet with reset flag
84  *
85  *  Do not allow orphaned sockets to eat all our resources.
86  *  This is direct violation of TCP specs, but it is required
87  *  to prevent DoS attacks. It is called when a retransmission timeout
88  *  or zero probe timeout occurs on orphaned socket.
89  *
90  *  Also close if our net namespace is exiting; in that case there is no
91  *  hope of ever communicating again since all netns interfaces are already
92  *  down (or about to be down), and we need to release our dst references,
93  *  which have been moved to the netns loopback interface, so the namespace
94  *  can finish exiting.  This condition is only possible if we are a kernel
95  *  socket, as those do not hold references to the namespace.
96  *
97  *  Criteria is still not confirmed experimentally and may change.
98  *  We kill the socket, if:
99  *  1. If number of orphaned sockets exceeds an administratively configured
100  *     limit.
101  *  2. If we have strong memory pressure.
102  *  3. If our net namespace is exiting.
103  */
tcp_out_of_resources(struct sock * sk,bool do_reset)104 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
105 {
106 	struct tcp_sock *tp = tcp_sk(sk);
107 	int shift = 0;
108 
109 	/* If peer does not open window for long time, or did not transmit
110 	 * anything for long time, penalize it. */
111 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
112 		shift++;
113 
114 	/* If some dubious ICMP arrived, penalize even more. */
115 	if (READ_ONCE(sk->sk_err_soft))
116 		shift++;
117 
118 	if (tcp_check_oom(sk, shift)) {
119 		/* Catch exceptional cases, when connection requires reset.
120 		 *      1. Last segment was sent recently. */
121 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
122 		    /*  2. Window is closed. */
123 		    (!tp->snd_wnd && !tp->packets_out))
124 			do_reset = true;
125 		if (do_reset)
126 			tcp_send_active_reset(sk, GFP_ATOMIC);
127 		tcp_done(sk);
128 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
129 		return 1;
130 	}
131 
132 	if (!check_net(sock_net(sk))) {
133 		/* Not possible to send reset; just close */
134 		tcp_done(sk);
135 		return 1;
136 	}
137 
138 	return 0;
139 }
140 
141 /**
142  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
143  *  @sk:    Pointer to the current socket.
144  *  @alive: bool, socket alive state
145  */
tcp_orphan_retries(struct sock * sk,bool alive)146 static int tcp_orphan_retries(struct sock *sk, bool alive)
147 {
148 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
149 
150 	/* We know from an ICMP that something is wrong. */
151 	if (READ_ONCE(sk->sk_err_soft) && !alive)
152 		retries = 0;
153 
154 	/* However, if socket sent something recently, select some safe
155 	 * number of retries. 8 corresponds to >100 seconds with minimal
156 	 * RTO of 200msec. */
157 	if (retries == 0 && alive)
158 		retries = 8;
159 	return retries;
160 }
161 
tcp_mtu_probing(struct inet_connection_sock * icsk,struct sock * sk)162 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
163 {
164 	const struct net *net = sock_net(sk);
165 	int mss;
166 
167 	/* Black hole detection */
168 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
169 		return;
170 
171 	if (!icsk->icsk_mtup.enabled) {
172 		icsk->icsk_mtup.enabled = 1;
173 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
174 	} else {
175 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
176 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
177 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
178 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
179 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
180 	}
181 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
182 }
183 
tcp_model_timeout(struct sock * sk,unsigned int boundary,unsigned int rto_base)184 static unsigned int tcp_model_timeout(struct sock *sk,
185 				      unsigned int boundary,
186 				      unsigned int rto_base)
187 {
188 	unsigned int linear_backoff_thresh, timeout;
189 
190 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
191 	if (boundary <= linear_backoff_thresh)
192 		timeout = ((2 << boundary) - 1) * rto_base;
193 	else
194 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
195 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
196 	return jiffies_to_msecs(timeout);
197 }
198 /**
199  *  retransmits_timed_out() - returns true if this connection has timed out
200  *  @sk:       The current socket
201  *  @boundary: max number of retransmissions
202  *  @timeout:  A custom timeout value.
203  *             If set to 0 the default timeout is calculated and used.
204  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
205  *
206  * The default "timeout" value this function can calculate and use
207  * is equivalent to the timeout of a TCP Connection
208  * after "boundary" unsuccessful, exponentially backed-off
209  * retransmissions with an initial RTO of TCP_RTO_MIN.
210  */
retransmits_timed_out(struct sock * sk,unsigned int boundary,unsigned int timeout)211 static bool retransmits_timed_out(struct sock *sk,
212 				  unsigned int boundary,
213 				  unsigned int timeout)
214 {
215 	unsigned int start_ts;
216 
217 	if (!inet_csk(sk)->icsk_retransmits)
218 		return false;
219 
220 	start_ts = tcp_sk(sk)->retrans_stamp;
221 	if (likely(timeout == 0)) {
222 		unsigned int rto_base = TCP_RTO_MIN;
223 
224 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
225 			rto_base = tcp_timeout_init(sk);
226 		timeout = tcp_model_timeout(sk, boundary, rto_base);
227 	}
228 
229 	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
230 }
231 
232 /* A write timeout has occurred. Process the after effects. */
tcp_write_timeout(struct sock * sk)233 static int tcp_write_timeout(struct sock *sk)
234 {
235 	struct inet_connection_sock *icsk = inet_csk(sk);
236 	struct tcp_sock *tp = tcp_sk(sk);
237 	struct net *net = sock_net(sk);
238 	bool expired = false, do_reset;
239 	int retry_until, max_retransmits;
240 
241 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
242 		if (icsk->icsk_retransmits)
243 			__dst_negative_advice(sk);
244 		/* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
245 		retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
246 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
247 
248 		max_retransmits = retry_until;
249 		if (sk->sk_state == TCP_SYN_SENT)
250 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
251 
252 		expired = icsk->icsk_retransmits >= max_retransmits;
253 	} else {
254 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
255 			/* Black hole detection */
256 			tcp_mtu_probing(icsk, sk);
257 
258 			__dst_negative_advice(sk);
259 		}
260 
261 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
262 		if (sock_flag(sk, SOCK_DEAD)) {
263 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
264 
265 			retry_until = tcp_orphan_retries(sk, alive);
266 			do_reset = alive ||
267 				!retransmits_timed_out(sk, retry_until, 0);
268 
269 			if (tcp_out_of_resources(sk, do_reset))
270 				return 1;
271 		}
272 	}
273 	if (!expired)
274 		expired = retransmits_timed_out(sk, retry_until,
275 						READ_ONCE(icsk->icsk_user_timeout));
276 	tcp_fastopen_active_detect_blackhole(sk, expired);
277 
278 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
279 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
280 				  icsk->icsk_retransmits,
281 				  icsk->icsk_rto, (int)expired);
282 
283 	if (expired) {
284 		/* Has it gone just too far? */
285 		tcp_write_err(sk);
286 		return 1;
287 	}
288 
289 	if (sk_rethink_txhash(sk)) {
290 		tp->timeout_rehash++;
291 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
292 	}
293 
294 	return 0;
295 }
296 
297 /* Called with BH disabled */
tcp_delack_timer_handler(struct sock * sk)298 void tcp_delack_timer_handler(struct sock *sk)
299 {
300 	struct inet_connection_sock *icsk = inet_csk(sk);
301 	struct tcp_sock *tp = tcp_sk(sk);
302 
303 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
304 		return;
305 
306 	/* Handling the sack compression case */
307 	if (tp->compressed_ack) {
308 		tcp_mstamp_refresh(tp);
309 		tcp_sack_compress_send_ack(sk);
310 		return;
311 	}
312 
313 	if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
314 		return;
315 
316 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
317 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
318 		return;
319 	}
320 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
321 
322 	if (inet_csk_ack_scheduled(sk)) {
323 		if (!inet_csk_in_pingpong_mode(sk)) {
324 			/* Delayed ACK missed: inflate ATO. */
325 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
326 		} else {
327 			/* Delayed ACK missed: leave pingpong mode and
328 			 * deflate ATO.
329 			 */
330 			inet_csk_exit_pingpong_mode(sk);
331 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
332 		}
333 		tcp_mstamp_refresh(tp);
334 		tcp_send_ack(sk);
335 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
336 	}
337 }
338 
339 
340 /**
341  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
342  *  @t:  Pointer to the timer. (gets casted to struct sock *)
343  *
344  *  This function gets (indirectly) called when the kernel timer for a TCP packet
345  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
346  *
347  *  Returns: Nothing (void)
348  */
tcp_delack_timer(struct timer_list * t)349 static void tcp_delack_timer(struct timer_list *t)
350 {
351 	struct inet_connection_sock *icsk =
352 			from_timer(icsk, t, icsk_delack_timer);
353 	struct sock *sk = &icsk->icsk_inet.sk;
354 
355 	bh_lock_sock(sk);
356 	if (!sock_owned_by_user(sk)) {
357 		tcp_delack_timer_handler(sk);
358 	} else {
359 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
360 		/* deleguate our work to tcp_release_cb() */
361 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
362 			sock_hold(sk);
363 	}
364 	bh_unlock_sock(sk);
365 	sock_put(sk);
366 }
367 
tcp_probe_timer(struct sock * sk)368 static void tcp_probe_timer(struct sock *sk)
369 {
370 	struct inet_connection_sock *icsk = inet_csk(sk);
371 	struct sk_buff *skb = tcp_send_head(sk);
372 	struct tcp_sock *tp = tcp_sk(sk);
373 	int max_probes;
374 
375 	if (tp->packets_out || !skb) {
376 		icsk->icsk_probes_out = 0;
377 		icsk->icsk_probes_tstamp = 0;
378 		return;
379 	}
380 
381 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
382 	 * long as the receiver continues to respond probes. We support this by
383 	 * default and reset icsk_probes_out with incoming ACKs. But if the
384 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
385 	 * kill the socket when the retry count and the time exceeds the
386 	 * corresponding system limit. We also implement similar policy when
387 	 * we use RTO to probe window in tcp_retransmit_timer().
388 	 */
389 	if (!icsk->icsk_probes_tstamp) {
390 		icsk->icsk_probes_tstamp = tcp_jiffies32;
391 	} else {
392 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
393 
394 		if (user_timeout &&
395 		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
396 		     msecs_to_jiffies(user_timeout))
397 		goto abort;
398 	}
399 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
400 	if (sock_flag(sk, SOCK_DEAD)) {
401 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
402 
403 		max_probes = tcp_orphan_retries(sk, alive);
404 		if (!alive && icsk->icsk_backoff >= max_probes)
405 			goto abort;
406 		if (tcp_out_of_resources(sk, true))
407 			return;
408 	}
409 
410 	if (icsk->icsk_probes_out >= max_probes) {
411 abort:		tcp_write_err(sk);
412 	} else {
413 		/* Only send another probe if we didn't close things up. */
414 		tcp_send_probe0(sk);
415 	}
416 }
417 
418 /*
419  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
420  *	sk here is the child socket, not the parent (listener) socket.
421  */
tcp_fastopen_synack_timer(struct sock * sk,struct request_sock * req)422 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
423 {
424 	struct inet_connection_sock *icsk = inet_csk(sk);
425 	struct tcp_sock *tp = tcp_sk(sk);
426 	int max_retries;
427 
428 	req->rsk_ops->syn_ack_timeout(req);
429 
430 	/* Add one more retry for fastopen.
431 	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
432 	 */
433 	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
434 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
435 
436 	if (req->num_timeout >= max_retries) {
437 		tcp_write_err(sk);
438 		return;
439 	}
440 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
441 	if (icsk->icsk_retransmits == 1)
442 		tcp_enter_loss(sk);
443 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
444 	 * returned from rtx_syn_ack() to make it more persistent like
445 	 * regular retransmit because if the child socket has been accepted
446 	 * it's not good to give up too easily.
447 	 */
448 	inet_rtx_syn_ack(sk, req);
449 	req->num_timeout++;
450 	icsk->icsk_retransmits++;
451 	if (!tp->retrans_stamp)
452 		tp->retrans_stamp = tcp_time_stamp(tp);
453 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
454 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
455 }
456 
tcp_rtx_probe0_timed_out(const struct sock * sk,const struct sk_buff * skb)457 static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
458 				     const struct sk_buff *skb)
459 {
460 	const struct tcp_sock *tp = tcp_sk(sk);
461 	const int timeout = TCP_RTO_MAX * 2;
462 	u32 rcv_delta, rtx_delta;
463 
464 	rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
465 	if (rcv_delta <= timeout)
466 		return false;
467 
468 	rtx_delta = (u32)msecs_to_jiffies(tcp_time_stamp(tp) -
469 			(tp->retrans_stamp ?: tcp_skb_timestamp(skb)));
470 
471 	return rtx_delta > timeout;
472 }
473 
474 /**
475  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
476  *  @sk:  Pointer to the current socket.
477  *
478  *  This function gets called when the kernel timer for a TCP packet
479  *  of this socket expires.
480  *
481  *  It handles retransmission, timer adjustment and other necessary measures.
482  *
483  *  Returns: Nothing (void)
484  */
tcp_retransmit_timer(struct sock * sk)485 void tcp_retransmit_timer(struct sock *sk)
486 {
487 	struct tcp_sock *tp = tcp_sk(sk);
488 	struct net *net = sock_net(sk);
489 	struct inet_connection_sock *icsk = inet_csk(sk);
490 	struct request_sock *req;
491 	struct sk_buff *skb;
492 
493 	req = rcu_dereference_protected(tp->fastopen_rsk,
494 					lockdep_sock_is_held(sk));
495 	if (req) {
496 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
497 			     sk->sk_state != TCP_FIN_WAIT1);
498 		tcp_fastopen_synack_timer(sk, req);
499 		/* Before we receive ACK to our SYN-ACK don't retransmit
500 		 * anything else (e.g., data or FIN segments).
501 		 */
502 		return;
503 	}
504 
505 	if (!tp->packets_out)
506 		return;
507 
508 	skb = tcp_rtx_queue_head(sk);
509 	if (WARN_ON_ONCE(!skb))
510 		return;
511 
512 	tp->tlp_high_seq = 0;
513 
514 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
515 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
516 		/* Receiver dastardly shrinks window. Our retransmits
517 		 * become zero probes, but we should not timeout this
518 		 * connection. If the socket is an orphan, time it out,
519 		 * we cannot allow such beasts to hang infinitely.
520 		 */
521 		struct inet_sock *inet = inet_sk(sk);
522 		u32 rtx_delta;
523 
524 		rtx_delta = tcp_time_stamp(tp) - (tp->retrans_stamp ?: tcp_skb_timestamp(skb));
525 		if (sk->sk_family == AF_INET) {
526 			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
527 				&inet->inet_daddr, ntohs(inet->inet_dport),
528 				inet->inet_num, tp->snd_una, tp->snd_nxt,
529 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
530 				rtx_delta);
531 		}
532 #if IS_ENABLED(CONFIG_IPV6)
533 		else if (sk->sk_family == AF_INET6) {
534 			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
535 				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
536 				inet->inet_num, tp->snd_una, tp->snd_nxt,
537 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
538 				rtx_delta);
539 		}
540 #endif
541 		if (tcp_rtx_probe0_timed_out(sk, skb)) {
542 			tcp_write_err(sk);
543 			goto out;
544 		}
545 		tcp_enter_loss(sk);
546 		tcp_retransmit_skb(sk, skb, 1);
547 		__sk_dst_reset(sk);
548 		goto out_reset_timer;
549 	}
550 
551 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
552 	if (tcp_write_timeout(sk))
553 		goto out;
554 
555 	if (icsk->icsk_retransmits == 0) {
556 		int mib_idx = 0;
557 
558 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
559 			if (tcp_is_sack(tp))
560 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
561 			else
562 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
563 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
564 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
565 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
566 			   tp->sacked_out) {
567 			if (tcp_is_sack(tp))
568 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
569 			else
570 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
571 		}
572 		if (mib_idx)
573 			__NET_INC_STATS(sock_net(sk), mib_idx);
574 	}
575 
576 	tcp_enter_loss(sk);
577 
578 	icsk->icsk_retransmits++;
579 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
580 		/* Retransmission failed because of local congestion,
581 		 * Let senders fight for local resources conservatively.
582 		 */
583 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
584 					  TCP_RESOURCE_PROBE_INTERVAL,
585 					  TCP_RTO_MAX);
586 		goto out;
587 	}
588 
589 	/* Increase the timeout each time we retransmit.  Note that
590 	 * we do not increase the rtt estimate.  rto is initialized
591 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
592 	 * that doubling rto each time is the least we can get away with.
593 	 * In KA9Q, Karn uses this for the first few times, and then
594 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
595 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
596 	 * defined in the protocol as the maximum possible RTT.  I guess
597 	 * we'll have to use something other than TCP to talk to the
598 	 * University of Mars.
599 	 *
600 	 * PAWS allows us longer timeouts and large windows, so once
601 	 * implemented ftp to mars will work nicely. We will have to fix
602 	 * the 120 second clamps though!
603 	 */
604 	icsk->icsk_backoff++;
605 
606 out_reset_timer:
607 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
608 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
609 	 * might be increased if the stream oscillates between thin and thick,
610 	 * thus the old value might already be too high compared to the value
611 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
612 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
613 	 * exponential backoff behaviour to avoid continue hammering
614 	 * linear-timeout retransmissions into a black hole
615 	 */
616 	if (sk->sk_state == TCP_ESTABLISHED &&
617 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
618 	    tcp_stream_is_thin(tp) &&
619 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
620 		icsk->icsk_backoff = 0;
621 		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
622 				       tcp_rto_min(sk),
623 				       TCP_RTO_MAX);
624 	} else if (sk->sk_state != TCP_SYN_SENT ||
625 		   icsk->icsk_backoff >
626 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
627 		/* Use normal (exponential) backoff unless linear timeouts are
628 		 * activated.
629 		 */
630 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
631 	}
632 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
633 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
634 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
635 		__sk_dst_reset(sk);
636 
637 out:;
638 }
639 
640 /* Called with bottom-half processing disabled.
641    Called by tcp_write_timer() */
tcp_write_timer_handler(struct sock * sk)642 void tcp_write_timer_handler(struct sock *sk)
643 {
644 	struct inet_connection_sock *icsk = inet_csk(sk);
645 	int event;
646 
647 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
648 	    !icsk->icsk_pending)
649 		return;
650 
651 	if (time_after(icsk->icsk_timeout, jiffies)) {
652 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
653 		return;
654 	}
655 
656 	tcp_mstamp_refresh(tcp_sk(sk));
657 	event = icsk->icsk_pending;
658 
659 	switch (event) {
660 	case ICSK_TIME_REO_TIMEOUT:
661 		tcp_rack_reo_timeout(sk);
662 		break;
663 	case ICSK_TIME_LOSS_PROBE:
664 		tcp_send_loss_probe(sk);
665 		break;
666 	case ICSK_TIME_RETRANS:
667 		icsk->icsk_pending = 0;
668 		tcp_retransmit_timer(sk);
669 		break;
670 	case ICSK_TIME_PROBE0:
671 		icsk->icsk_pending = 0;
672 		tcp_probe_timer(sk);
673 		break;
674 	}
675 }
676 
tcp_write_timer(struct timer_list * t)677 static void tcp_write_timer(struct timer_list *t)
678 {
679 	struct inet_connection_sock *icsk =
680 			from_timer(icsk, t, icsk_retransmit_timer);
681 	struct sock *sk = &icsk->icsk_inet.sk;
682 
683 	bh_lock_sock(sk);
684 	if (!sock_owned_by_user(sk)) {
685 		tcp_write_timer_handler(sk);
686 	} else {
687 		/* delegate our work to tcp_release_cb() */
688 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
689 			sock_hold(sk);
690 	}
691 	bh_unlock_sock(sk);
692 	sock_put(sk);
693 }
694 
tcp_syn_ack_timeout(const struct request_sock * req)695 void tcp_syn_ack_timeout(const struct request_sock *req)
696 {
697 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
698 
699 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
700 }
701 EXPORT_SYMBOL(tcp_syn_ack_timeout);
702 
tcp_set_keepalive(struct sock * sk,int val)703 void tcp_set_keepalive(struct sock *sk, int val)
704 {
705 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
706 		return;
707 
708 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
709 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
710 	else if (!val)
711 		inet_csk_delete_keepalive_timer(sk);
712 }
713 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
714 
715 
tcp_keepalive_timer(struct timer_list * t)716 static void tcp_keepalive_timer (struct timer_list *t)
717 {
718 	struct sock *sk = from_timer(sk, t, sk_timer);
719 	struct inet_connection_sock *icsk = inet_csk(sk);
720 	struct tcp_sock *tp = tcp_sk(sk);
721 	u32 elapsed;
722 
723 	/* Only process if socket is not in use. */
724 	bh_lock_sock(sk);
725 	if (sock_owned_by_user(sk)) {
726 		/* Try again later. */
727 		inet_csk_reset_keepalive_timer (sk, HZ/20);
728 		goto out;
729 	}
730 
731 	if (sk->sk_state == TCP_LISTEN) {
732 		pr_err("Hmm... keepalive on a LISTEN ???\n");
733 		goto out;
734 	}
735 
736 	tcp_mstamp_refresh(tp);
737 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
738 		if (READ_ONCE(tp->linger2) >= 0) {
739 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
740 
741 			if (tmo > 0) {
742 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
743 				goto out;
744 			}
745 		}
746 		tcp_send_active_reset(sk, GFP_ATOMIC);
747 		goto death;
748 	}
749 
750 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
751 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
752 		goto out;
753 
754 	elapsed = keepalive_time_when(tp);
755 
756 	/* It is alive without keepalive 8) */
757 	if (tp->packets_out || !tcp_write_queue_empty(sk))
758 		goto resched;
759 
760 	elapsed = keepalive_time_elapsed(tp);
761 
762 	if (elapsed >= keepalive_time_when(tp)) {
763 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
764 
765 		/* If the TCP_USER_TIMEOUT option is enabled, use that
766 		 * to determine when to timeout instead.
767 		 */
768 		if ((user_timeout != 0 &&
769 		    elapsed >= msecs_to_jiffies(user_timeout) &&
770 		    icsk->icsk_probes_out > 0) ||
771 		    (user_timeout == 0 &&
772 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
773 			tcp_send_active_reset(sk, GFP_ATOMIC);
774 			tcp_write_err(sk);
775 			goto out;
776 		}
777 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
778 			icsk->icsk_probes_out++;
779 			elapsed = keepalive_intvl_when(tp);
780 		} else {
781 			/* If keepalive was lost due to local congestion,
782 			 * try harder.
783 			 */
784 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
785 		}
786 	} else {
787 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
788 		elapsed = keepalive_time_when(tp) - elapsed;
789 	}
790 
791 resched:
792 	inet_csk_reset_keepalive_timer (sk, elapsed);
793 	goto out;
794 
795 death:
796 	tcp_done(sk);
797 
798 out:
799 	bh_unlock_sock(sk);
800 	sock_put(sk);
801 }
802 
tcp_compressed_ack_kick(struct hrtimer * timer)803 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
804 {
805 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
806 	struct sock *sk = (struct sock *)tp;
807 
808 	bh_lock_sock(sk);
809 	if (!sock_owned_by_user(sk)) {
810 		if (tp->compressed_ack) {
811 			/* Since we have to send one ack finally,
812 			 * subtract one from tp->compressed_ack to keep
813 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
814 			 */
815 			tp->compressed_ack--;
816 			tcp_send_ack(sk);
817 		}
818 	} else {
819 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
820 				      &sk->sk_tsq_flags))
821 			sock_hold(sk);
822 	}
823 	bh_unlock_sock(sk);
824 
825 	sock_put(sk);
826 
827 	return HRTIMER_NORESTART;
828 }
829 
tcp_init_xmit_timers(struct sock * sk)830 void tcp_init_xmit_timers(struct sock *sk)
831 {
832 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
833 				  &tcp_keepalive_timer);
834 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
835 		     HRTIMER_MODE_ABS_PINNED_SOFT);
836 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
837 
838 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
839 		     HRTIMER_MODE_REL_PINNED_SOFT);
840 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
841 }
842