• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <trace/hooks/net.h>
26 
tcp_clamp_rto_to_user_timeout(const struct sock * sk)27 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk)
28 {
29 	struct inet_connection_sock *icsk = inet_csk(sk);
30 	u32 elapsed, start_ts, user_timeout;
31 	s32 remaining;
32 
33 	start_ts = tcp_sk(sk)->retrans_stamp;
34 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
35 	if (!user_timeout)
36 		return icsk->icsk_rto;
37 	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
38 	remaining = user_timeout - elapsed;
39 	if (remaining <= 0)
40 		return 1; /* user timeout has passed; fire ASAP */
41 
42 	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
43 }
44 
tcp_clamp_probe0_to_user_timeout(const struct sock * sk,u32 when)45 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when)
46 {
47 	struct inet_connection_sock *icsk = inet_csk(sk);
48 	u32 remaining, user_timeout;
49 	s32 elapsed;
50 
51 	user_timeout = READ_ONCE(icsk->icsk_user_timeout);
52 	if (!user_timeout || !icsk->icsk_probes_tstamp)
53 		return when;
54 
55 	elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp;
56 	if (unlikely(elapsed < 0))
57 		elapsed = 0;
58 	remaining = msecs_to_jiffies(user_timeout) - elapsed;
59 	remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN);
60 
61 	return min_t(u32, remaining, when);
62 }
63 
64 /**
65  *  tcp_write_err() - close socket and save error info
66  *  @sk:  The socket the error has appeared on.
67  *
68  *  Returns: Nothing (void)
69  */
70 
tcp_write_err(struct sock * sk)71 static void tcp_write_err(struct sock *sk)
72 {
73 	tcp_done_with_error(sk, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT);
74 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
75 }
76 
77 /**
78  *  tcp_out_of_resources() - Close socket if out of resources
79  *  @sk:        pointer to current socket
80  *  @do_reset:  send a last packet with reset flag
81  *
82  *  Do not allow orphaned sockets to eat all our resources.
83  *  This is direct violation of TCP specs, but it is required
84  *  to prevent DoS attacks. It is called when a retransmission timeout
85  *  or zero probe timeout occurs on orphaned socket.
86  *
87  *  Also close if our net namespace is exiting; in that case there is no
88  *  hope of ever communicating again since all netns interfaces are already
89  *  down (or about to be down), and we need to release our dst references,
90  *  which have been moved to the netns loopback interface, so the namespace
91  *  can finish exiting.  This condition is only possible if we are a kernel
92  *  socket, as those do not hold references to the namespace.
93  *
94  *  Criteria is still not confirmed experimentally and may change.
95  *  We kill the socket, if:
96  *  1. If number of orphaned sockets exceeds an administratively configured
97  *     limit.
98  *  2. If we have strong memory pressure.
99  *  3. If our net namespace is exiting.
100  */
tcp_out_of_resources(struct sock * sk,bool do_reset)101 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
102 {
103 	struct tcp_sock *tp = tcp_sk(sk);
104 	int shift = 0;
105 
106 	/* If peer does not open window for long time, or did not transmit
107 	 * anything for long time, penalize it. */
108 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
109 		shift++;
110 
111 	/* If some dubious ICMP arrived, penalize even more. */
112 	if (READ_ONCE(sk->sk_err_soft))
113 		shift++;
114 
115 	if (tcp_check_oom(sk, shift)) {
116 		/* Catch exceptional cases, when connection requires reset.
117 		 *      1. Last segment was sent recently. */
118 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
119 		    /*  2. Window is closed. */
120 		    (!tp->snd_wnd && !tp->packets_out))
121 			do_reset = true;
122 		if (do_reset)
123 			tcp_send_active_reset(sk, GFP_ATOMIC);
124 		tcp_done(sk);
125 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
126 		return 1;
127 	}
128 
129 	if (!check_net(sock_net(sk))) {
130 		/* Not possible to send reset; just close */
131 		tcp_done(sk);
132 		return 1;
133 	}
134 
135 	return 0;
136 }
137 
138 /**
139  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
140  *  @sk:    Pointer to the current socket.
141  *  @alive: bool, socket alive state
142  */
tcp_orphan_retries(struct sock * sk,bool alive)143 static int tcp_orphan_retries(struct sock *sk, bool alive)
144 {
145 	int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */
146 
147 	/* We know from an ICMP that something is wrong. */
148 	if (READ_ONCE(sk->sk_err_soft) && !alive)
149 		retries = 0;
150 
151 	/* However, if socket sent something recently, select some safe
152 	 * number of retries. 8 corresponds to >100 seconds with minimal
153 	 * RTO of 200msec. */
154 	if (retries == 0 && alive)
155 		retries = 8;
156 	return retries;
157 }
158 
tcp_mtu_probing(struct inet_connection_sock * icsk,struct sock * sk)159 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
160 {
161 	const struct net *net = sock_net(sk);
162 	int mss;
163 
164 	/* Black hole detection */
165 	if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing))
166 		return;
167 
168 	if (!icsk->icsk_mtup.enabled) {
169 		icsk->icsk_mtup.enabled = 1;
170 		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
171 	} else {
172 		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
173 		mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss);
174 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor));
175 		mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss));
176 		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
177 	}
178 	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
179 }
180 
tcp_model_timeout(struct sock * sk,unsigned int boundary,unsigned int rto_base)181 static unsigned int tcp_model_timeout(struct sock *sk,
182 				      unsigned int boundary,
183 				      unsigned int rto_base)
184 {
185 	unsigned int linear_backoff_thresh, timeout;
186 
187 	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
188 	if (boundary <= linear_backoff_thresh)
189 		timeout = ((2 << boundary) - 1) * rto_base;
190 	else
191 		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
192 			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
193 	return jiffies_to_msecs(timeout);
194 }
195 /**
196  *  retransmits_timed_out() - returns true if this connection has timed out
197  *  @sk:       The current socket
198  *  @boundary: max number of retransmissions
199  *  @timeout:  A custom timeout value.
200  *             If set to 0 the default timeout is calculated and used.
201  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
202  *
203  * The default "timeout" value this function can calculate and use
204  * is equivalent to the timeout of a TCP Connection
205  * after "boundary" unsuccessful, exponentially backed-off
206  * retransmissions with an initial RTO of TCP_RTO_MIN.
207  */
retransmits_timed_out(struct sock * sk,unsigned int boundary,unsigned int timeout)208 static bool retransmits_timed_out(struct sock *sk,
209 				  unsigned int boundary,
210 				  unsigned int timeout)
211 {
212 	unsigned int start_ts;
213 
214 	if (!inet_csk(sk)->icsk_retransmits)
215 		return false;
216 
217 	start_ts = tcp_sk(sk)->retrans_stamp;
218 	if (likely(timeout == 0)) {
219 		unsigned int rto_base = TCP_RTO_MIN;
220 
221 		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
222 			rto_base = tcp_timeout_init(sk);
223 		timeout = tcp_model_timeout(sk, boundary, rto_base);
224 	}
225 
226 	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
227 }
228 
229 /* A write timeout has occurred. Process the after effects. */
tcp_write_timeout(struct sock * sk)230 static int tcp_write_timeout(struct sock *sk)
231 {
232 	struct inet_connection_sock *icsk = inet_csk(sk);
233 	struct tcp_sock *tp = tcp_sk(sk);
234 	struct net *net = sock_net(sk);
235 	bool expired = false, do_reset;
236 	int retry_until, max_retransmits;
237 
238 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
239 		if (icsk->icsk_retransmits)
240 			__dst_negative_advice(sk);
241 		/* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */
242 		retry_until = READ_ONCE(icsk->icsk_syn_retries) ? :
243 			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
244 
245 		max_retransmits = retry_until;
246 		if (sk->sk_state == TCP_SYN_SENT)
247 			max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts);
248 
249 		expired = icsk->icsk_retransmits >= max_retransmits;
250 	} else {
251 		if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) {
252 			/* Black hole detection */
253 			tcp_mtu_probing(icsk, sk);
254 			trace_android_vh_tcp_write_timeout_estab_retrans(sk);
255 
256 			__dst_negative_advice(sk);
257 		}
258 
259 		retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2);
260 		if (sock_flag(sk, SOCK_DEAD)) {
261 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
262 
263 			retry_until = tcp_orphan_retries(sk, alive);
264 			do_reset = alive ||
265 				!retransmits_timed_out(sk, retry_until, 0);
266 
267 			if (tcp_out_of_resources(sk, do_reset))
268 				return 1;
269 		}
270 	}
271 	if (!expired)
272 		expired = retransmits_timed_out(sk, retry_until,
273 						READ_ONCE(icsk->icsk_user_timeout));
274 	tcp_fastopen_active_detect_blackhole(sk, expired);
275 
276 	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
277 		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
278 				  icsk->icsk_retransmits,
279 				  icsk->icsk_rto, (int)expired);
280 
281 	if (expired) {
282 		/* Has it gone just too far? */
283 
284 		trace_android_vh_tcp_state_change(sk, TCP_STATE_CHANGE_REASON_SYN_TIMEOUT, 0);
285 
286 		tcp_write_err(sk);
287 		return 1;
288 	}
289 
290 	trace_android_vh_tcp_state_change(sk, TCP_STATE_CHANGE_REASON_RETRANSMIT, 0);
291 
292 	if (sk_rethink_txhash(sk)) {
293 		tp->timeout_rehash++;
294 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH);
295 	}
296 
297 	return 0;
298 }
299 
300 /* Called with BH disabled */
tcp_delack_timer_handler(struct sock * sk)301 void tcp_delack_timer_handler(struct sock *sk)
302 {
303 	struct inet_connection_sock *icsk = inet_csk(sk);
304 	struct tcp_sock *tp = tcp_sk(sk);
305 
306 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
307 		return;
308 
309 	/* Handling the sack compression case */
310 	if (tp->compressed_ack) {
311 		tcp_mstamp_refresh(tp);
312 		tcp_sack_compress_send_ack(sk);
313 		return;
314 	}
315 
316 	if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
317 		return;
318 
319 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
320 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
321 		return;
322 	}
323 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
324 
325 	if (inet_csk_ack_scheduled(sk)) {
326 		if (!inet_csk_in_pingpong_mode(sk)) {
327 			/* Delayed ACK missed: inflate ATO. */
328 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
329 		} else {
330 			/* Delayed ACK missed: leave pingpong mode and
331 			 * deflate ATO.
332 			 */
333 			inet_csk_exit_pingpong_mode(sk);
334 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
335 		}
336 		tcp_mstamp_refresh(tp);
337 		tcp_send_ack(sk);
338 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
339 	}
340 }
341 
342 
343 /**
344  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
345  *  @t:  Pointer to the timer. (gets casted to struct sock *)
346  *
347  *  This function gets (indirectly) called when the kernel timer for a TCP packet
348  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
349  *
350  *  Returns: Nothing (void)
351  */
tcp_delack_timer(struct timer_list * t)352 static void tcp_delack_timer(struct timer_list *t)
353 {
354 	struct inet_connection_sock *icsk =
355 			from_timer(icsk, t, icsk_delack_timer);
356 	struct sock *sk = &icsk->icsk_inet.sk;
357 
358 	bh_lock_sock(sk);
359 	if (!sock_owned_by_user(sk)) {
360 		tcp_delack_timer_handler(sk);
361 	} else {
362 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
363 		/* deleguate our work to tcp_release_cb() */
364 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
365 			sock_hold(sk);
366 	}
367 	bh_unlock_sock(sk);
368 	sock_put(sk);
369 }
370 
tcp_probe_timer(struct sock * sk)371 static void tcp_probe_timer(struct sock *sk)
372 {
373 	struct inet_connection_sock *icsk = inet_csk(sk);
374 	struct sk_buff *skb = tcp_send_head(sk);
375 	struct tcp_sock *tp = tcp_sk(sk);
376 	int max_probes;
377 
378 	if (tp->packets_out || !skb) {
379 		icsk->icsk_probes_out = 0;
380 		icsk->icsk_probes_tstamp = 0;
381 		return;
382 	}
383 
384 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
385 	 * long as the receiver continues to respond probes. We support this by
386 	 * default and reset icsk_probes_out with incoming ACKs. But if the
387 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
388 	 * kill the socket when the retry count and the time exceeds the
389 	 * corresponding system limit. We also implement similar policy when
390 	 * we use RTO to probe window in tcp_retransmit_timer().
391 	 */
392 	if (!icsk->icsk_probes_tstamp) {
393 		icsk->icsk_probes_tstamp = tcp_jiffies32;
394 	} else {
395 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
396 
397 		if (user_timeout &&
398 		    (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >=
399 		     msecs_to_jiffies(user_timeout))
400 		goto abort;
401 	}
402 	max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2);
403 	if (sock_flag(sk, SOCK_DEAD)) {
404 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
405 
406 		max_probes = tcp_orphan_retries(sk, alive);
407 		if (!alive && icsk->icsk_backoff >= max_probes)
408 			goto abort;
409 		if (tcp_out_of_resources(sk, true))
410 			return;
411 	}
412 
413 	if (icsk->icsk_probes_out >= max_probes) {
414 abort:		tcp_write_err(sk);
415 	} else {
416 		/* Only send another probe if we didn't close things up. */
417 		tcp_send_probe0(sk);
418 	}
419 }
420 
421 /*
422  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
423  *	sk here is the child socket, not the parent (listener) socket.
424  */
tcp_fastopen_synack_timer(struct sock * sk,struct request_sock * req)425 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
426 {
427 	struct inet_connection_sock *icsk = inet_csk(sk);
428 	struct tcp_sock *tp = tcp_sk(sk);
429 	int max_retries;
430 
431 	req->rsk_ops->syn_ack_timeout(req);
432 
433 	/* Add one more retry for fastopen.
434 	 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt()
435 	 */
436 	max_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
437 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1;
438 
439 	if (req->num_timeout >= max_retries) {
440 		tcp_write_err(sk);
441 		return;
442 	}
443 	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
444 	if (icsk->icsk_retransmits == 1)
445 		tcp_enter_loss(sk);
446 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
447 	 * returned from rtx_syn_ack() to make it more persistent like
448 	 * regular retransmit because if the child socket has been accepted
449 	 * it's not good to give up too easily.
450 	 */
451 	inet_rtx_syn_ack(sk, req);
452 	req->num_timeout++;
453 	icsk->icsk_retransmits++;
454 	if (!tp->retrans_stamp)
455 		tp->retrans_stamp = tcp_time_stamp(tp);
456 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
457 			  req->timeout << req->num_timeout, TCP_RTO_MAX);
458 }
459 
tcp_rtx_probe0_timed_out(const struct sock * sk,const struct sk_buff * skb)460 static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
461 				     const struct sk_buff *skb)
462 {
463 	const struct inet_connection_sock *icsk = inet_csk(sk);
464 	u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
465 	const struct tcp_sock *tp = tcp_sk(sk);
466 	int timeout = TCP_RTO_MAX * 2;
467 	u32 rtx_delta;
468 	s32 rcv_delta;
469 
470 	rtx_delta = (u32)msecs_to_jiffies(tcp_time_stamp(tp) -
471 			(tp->retrans_stamp ?: tcp_skb_timestamp(skb)));
472 
473 	if (user_timeout) {
474 		/* If user application specified a TCP_USER_TIMEOUT,
475 		 * it does not want win 0 packets to 'reset the timer'
476 		 * while retransmits are not making progress.
477 		 */
478 		if (rtx_delta > user_timeout)
479 			return true;
480 		timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
481 	}
482 
483 	/* Note: timer interrupt might have been delayed by at least one jiffy,
484 	 * and tp->rcv_tstamp might very well have been written recently.
485 	 * rcv_delta can thus be negative.
486 	 */
487 	rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
488 	if (rcv_delta <= timeout)
489 		return false;
490 
491 	return rtx_delta > timeout;
492 }
493 
494 /**
495  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
496  *  @sk:  Pointer to the current socket.
497  *
498  *  This function gets called when the kernel timer for a TCP packet
499  *  of this socket expires.
500  *
501  *  It handles retransmission, timer adjustment and other necessary measures.
502  *
503  *  Returns: Nothing (void)
504  */
tcp_retransmit_timer(struct sock * sk)505 void tcp_retransmit_timer(struct sock *sk)
506 {
507 	struct tcp_sock *tp = tcp_sk(sk);
508 	struct net *net = sock_net(sk);
509 	struct inet_connection_sock *icsk = inet_csk(sk);
510 	struct request_sock *req;
511 	struct sk_buff *skb;
512 
513 	req = rcu_dereference_protected(tp->fastopen_rsk,
514 					lockdep_sock_is_held(sk));
515 	if (req) {
516 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
517 			     sk->sk_state != TCP_FIN_WAIT1);
518 		tcp_fastopen_synack_timer(sk, req);
519 		/* Before we receive ACK to our SYN-ACK don't retransmit
520 		 * anything else (e.g., data or FIN segments).
521 		 */
522 		return;
523 	}
524 
525 	if (!tp->packets_out)
526 		return;
527 
528 	skb = tcp_rtx_queue_head(sk);
529 	if (WARN_ON_ONCE(!skb))
530 		return;
531 
532 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
533 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
534 		/* Receiver dastardly shrinks window. Our retransmits
535 		 * become zero probes, but we should not timeout this
536 		 * connection. If the socket is an orphan, time it out,
537 		 * we cannot allow such beasts to hang infinitely.
538 		 */
539 		struct inet_sock *inet = inet_sk(sk);
540 		u32 rtx_delta;
541 
542 		rtx_delta = tcp_time_stamp(tp) - (tp->retrans_stamp ?: tcp_skb_timestamp(skb));
543 		if (sk->sk_family == AF_INET) {
544 			net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
545 				&inet->inet_daddr, ntohs(inet->inet_dport),
546 				inet->inet_num, tp->snd_una, tp->snd_nxt,
547 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
548 				rtx_delta);
549 		}
550 #if IS_ENABLED(CONFIG_IPV6)
551 		else if (sk->sk_family == AF_INET6) {
552 			net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n",
553 				&sk->sk_v6_daddr, ntohs(inet->inet_dport),
554 				inet->inet_num, tp->snd_una, tp->snd_nxt,
555 				jiffies_to_msecs(jiffies - tp->rcv_tstamp),
556 				rtx_delta);
557 		}
558 #endif
559 		if (tcp_rtx_probe0_timed_out(sk, skb)) {
560 			tcp_write_err(sk);
561 			goto out;
562 		}
563 		tcp_enter_loss(sk);
564 		tcp_retransmit_skb(sk, skb, 1);
565 		__sk_dst_reset(sk);
566 		goto out_reset_timer;
567 	}
568 
569 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
570 	if (tcp_write_timeout(sk))
571 		goto out;
572 
573 	if (icsk->icsk_retransmits == 0) {
574 		int mib_idx = 0;
575 
576 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
577 			if (tcp_is_sack(tp))
578 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
579 			else
580 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
581 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
582 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
583 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
584 			   tp->sacked_out) {
585 			if (tcp_is_sack(tp))
586 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
587 			else
588 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
589 		}
590 		if (mib_idx)
591 			__NET_INC_STATS(sock_net(sk), mib_idx);
592 	}
593 
594 	tcp_enter_loss(sk);
595 
596 	icsk->icsk_retransmits++;
597 	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
598 		/* Retransmission failed because of local congestion,
599 		 * Let senders fight for local resources conservatively.
600 		 */
601 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
602 					  TCP_RESOURCE_PROBE_INTERVAL,
603 					  TCP_RTO_MAX);
604 		goto out;
605 	}
606 
607 	/* Increase the timeout each time we retransmit.  Note that
608 	 * we do not increase the rtt estimate.  rto is initialized
609 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
610 	 * that doubling rto each time is the least we can get away with.
611 	 * In KA9Q, Karn uses this for the first few times, and then
612 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
613 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
614 	 * defined in the protocol as the maximum possible RTT.  I guess
615 	 * we'll have to use something other than TCP to talk to the
616 	 * University of Mars.
617 	 *
618 	 * PAWS allows us longer timeouts and large windows, so once
619 	 * implemented ftp to mars will work nicely. We will have to fix
620 	 * the 120 second clamps though!
621 	 */
622 	icsk->icsk_backoff++;
623 
624 out_reset_timer:
625 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
626 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
627 	 * might be increased if the stream oscillates between thin and thick,
628 	 * thus the old value might already be too high compared to the value
629 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
630 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
631 	 * exponential backoff behaviour to avoid continue hammering
632 	 * linear-timeout retransmissions into a black hole
633 	 */
634 	if (sk->sk_state == TCP_ESTABLISHED &&
635 	    (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) &&
636 	    tcp_stream_is_thin(tp) &&
637 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
638 		icsk->icsk_backoff = 0;
639 		icsk->icsk_rto = clamp(__tcp_set_rto(tp),
640 				       tcp_rto_min(sk),
641 				       TCP_RTO_MAX);
642 	} else if (sk->sk_state != TCP_SYN_SENT ||
643 		   icsk->icsk_backoff >
644 		   READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) {
645 		/* Use normal (exponential) backoff unless linear timeouts are
646 		 * activated.
647 		 */
648 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
649 
650 		trace_android_vh_tcp_fastsyn(sk);
651 	}
652 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
653 				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
654 	if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0))
655 		__sk_dst_reset(sk);
656 
657 out:;
658 }
659 
660 /* Called with bottom-half processing disabled.
661    Called by tcp_write_timer() */
tcp_write_timer_handler(struct sock * sk)662 void tcp_write_timer_handler(struct sock *sk)
663 {
664 	struct inet_connection_sock *icsk = inet_csk(sk);
665 	int event;
666 
667 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
668 	    !icsk->icsk_pending)
669 		return;
670 
671 	if (time_after(icsk->icsk_timeout, jiffies)) {
672 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
673 		return;
674 	}
675 
676 	tcp_mstamp_refresh(tcp_sk(sk));
677 	event = icsk->icsk_pending;
678 
679 	switch (event) {
680 	case ICSK_TIME_REO_TIMEOUT:
681 		tcp_rack_reo_timeout(sk);
682 		break;
683 	case ICSK_TIME_LOSS_PROBE:
684 		tcp_send_loss_probe(sk);
685 		break;
686 	case ICSK_TIME_RETRANS:
687 		icsk->icsk_pending = 0;
688 		tcp_retransmit_timer(sk);
689 		break;
690 	case ICSK_TIME_PROBE0:
691 		icsk->icsk_pending = 0;
692 		tcp_probe_timer(sk);
693 		break;
694 	}
695 }
696 
tcp_write_timer(struct timer_list * t)697 static void tcp_write_timer(struct timer_list *t)
698 {
699 	struct inet_connection_sock *icsk =
700 			from_timer(icsk, t, icsk_retransmit_timer);
701 	struct sock *sk = &icsk->icsk_inet.sk;
702 
703 	bh_lock_sock(sk);
704 	if (!sock_owned_by_user(sk)) {
705 		tcp_write_timer_handler(sk);
706 	} else {
707 		/* delegate our work to tcp_release_cb() */
708 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
709 			sock_hold(sk);
710 	}
711 	bh_unlock_sock(sk);
712 	sock_put(sk);
713 }
714 
tcp_syn_ack_timeout(const struct request_sock * req)715 void tcp_syn_ack_timeout(const struct request_sock *req)
716 {
717 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
718 
719 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
720 }
721 EXPORT_SYMBOL(tcp_syn_ack_timeout);
722 
tcp_set_keepalive(struct sock * sk,int val)723 void tcp_set_keepalive(struct sock *sk, int val)
724 {
725 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
726 		return;
727 
728 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
729 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
730 	else if (!val)
731 		inet_csk_delete_keepalive_timer(sk);
732 }
733 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
734 
735 
tcp_keepalive_timer(struct timer_list * t)736 static void tcp_keepalive_timer (struct timer_list *t)
737 {
738 	struct sock *sk = from_timer(sk, t, sk_timer);
739 	struct inet_connection_sock *icsk = inet_csk(sk);
740 	struct tcp_sock *tp = tcp_sk(sk);
741 	u32 elapsed;
742 
743 	/* Only process if socket is not in use. */
744 	bh_lock_sock(sk);
745 	if (sock_owned_by_user(sk)) {
746 		/* Try again later. */
747 		inet_csk_reset_keepalive_timer (sk, HZ/20);
748 		goto out;
749 	}
750 
751 	if (sk->sk_state == TCP_LISTEN) {
752 		pr_err("Hmm... keepalive on a LISTEN ???\n");
753 		goto out;
754 	}
755 
756 	tcp_mstamp_refresh(tp);
757 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
758 		if (READ_ONCE(tp->linger2) >= 0) {
759 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
760 
761 			if (tmo > 0) {
762 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
763 				goto out;
764 			}
765 		}
766 		tcp_send_active_reset(sk, GFP_ATOMIC);
767 		goto death;
768 	}
769 
770 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
771 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
772 		goto out;
773 
774 	elapsed = keepalive_time_when(tp);
775 
776 	/* It is alive without keepalive 8) */
777 	if (tp->packets_out || !tcp_write_queue_empty(sk))
778 		goto resched;
779 
780 	elapsed = keepalive_time_elapsed(tp);
781 
782 	if (elapsed >= keepalive_time_when(tp)) {
783 		u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
784 
785 		/* If the TCP_USER_TIMEOUT option is enabled, use that
786 		 * to determine when to timeout instead.
787 		 */
788 		if ((user_timeout != 0 &&
789 		    elapsed >= msecs_to_jiffies(user_timeout) &&
790 		    icsk->icsk_probes_out > 0) ||
791 		    (user_timeout == 0 &&
792 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
793 			tcp_send_active_reset(sk, GFP_ATOMIC);
794 			tcp_write_err(sk);
795 			goto out;
796 		}
797 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
798 			icsk->icsk_probes_out++;
799 			elapsed = keepalive_intvl_when(tp);
800 		} else {
801 			/* If keepalive was lost due to local congestion,
802 			 * try harder.
803 			 */
804 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
805 		}
806 	} else {
807 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
808 		elapsed = keepalive_time_when(tp) - elapsed;
809 	}
810 
811 resched:
812 	inet_csk_reset_keepalive_timer (sk, elapsed);
813 	goto out;
814 
815 death:
816 	tcp_done(sk);
817 
818 out:
819 	bh_unlock_sock(sk);
820 	sock_put(sk);
821 }
822 
tcp_compressed_ack_kick(struct hrtimer * timer)823 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
824 {
825 	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
826 	struct sock *sk = (struct sock *)tp;
827 
828 	bh_lock_sock(sk);
829 	if (!sock_owned_by_user(sk)) {
830 		if (tp->compressed_ack) {
831 			/* Since we have to send one ack finally,
832 			 * subtract one from tp->compressed_ack to keep
833 			 * LINUX_MIB_TCPACKCOMPRESSED accurate.
834 			 */
835 			tp->compressed_ack--;
836 			tcp_send_ack(sk);
837 		}
838 	} else {
839 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
840 				      &sk->sk_tsq_flags))
841 			sock_hold(sk);
842 	}
843 	bh_unlock_sock(sk);
844 
845 	sock_put(sk);
846 
847 	return HRTIMER_NORESTART;
848 }
849 
tcp_init_xmit_timers(struct sock * sk)850 void tcp_init_xmit_timers(struct sock *sk)
851 {
852 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
853 				  &tcp_keepalive_timer);
854 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
855 		     HRTIMER_MODE_ABS_PINNED_SOFT);
856 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
857 
858 	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
859 		     HRTIMER_MODE_REL_PINNED_SOFT);
860 	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
861 }
862