Linux Audio

Check our new training course

Loading...
v5.4
  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
 26static 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;
 30	s32 remaining;
 31
 32	start_ts = tcp_sk(sk)->retrans_stamp;
 33	if (!icsk->icsk_user_timeout)
 34		return icsk->icsk_rto;
 35	elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts;
 36	remaining = icsk->icsk_user_timeout - elapsed;
 37	if (remaining <= 0)
 38		return 1; /* user timeout has passed; fire ASAP */
 39
 40	return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining));
 41}
 42
 43/**
 44 *  tcp_write_err() - close socket and save error info
 45 *  @sk:  The socket the error has appeared on.
 46 *
 47 *  Returns: Nothing (void)
 48 */
 49
 50static void tcp_write_err(struct sock *sk)
 51{
 52	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
 53	sk->sk_error_report(sk);
 54
 55	tcp_write_queue_purge(sk);
 56	tcp_done(sk);
 57	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
 58}
 59
 60/**
 61 *  tcp_out_of_resources() - Close socket if out of resources
 62 *  @sk:        pointer to current socket
 63 *  @do_reset:  send a last packet with reset flag
 64 *
 65 *  Do not allow orphaned sockets to eat all our resources.
 66 *  This is direct violation of TCP specs, but it is required
 67 *  to prevent DoS attacks. It is called when a retransmission timeout
 68 *  or zero probe timeout occurs on orphaned socket.
 69 *
 70 *  Also close if our net namespace is exiting; in that case there is no
 71 *  hope of ever communicating again since all netns interfaces are already
 72 *  down (or about to be down), and we need to release our dst references,
 73 *  which have been moved to the netns loopback interface, so the namespace
 74 *  can finish exiting.  This condition is only possible if we are a kernel
 75 *  socket, as those do not hold references to the namespace.
 76 *
 77 *  Criteria is still not confirmed experimentally and may change.
 78 *  We kill the socket, if:
 79 *  1. If number of orphaned sockets exceeds an administratively configured
 80 *     limit.
 81 *  2. If we have strong memory pressure.
 82 *  3. If our net namespace is exiting.
 83 */
 84static int tcp_out_of_resources(struct sock *sk, bool do_reset)
 85{
 86	struct tcp_sock *tp = tcp_sk(sk);
 87	int shift = 0;
 88
 89	/* If peer does not open window for long time, or did not transmit
 90	 * anything for long time, penalize it. */
 91	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
 92		shift++;
 93
 94	/* If some dubious ICMP arrived, penalize even more. */
 95	if (sk->sk_err_soft)
 96		shift++;
 97
 98	if (tcp_check_oom(sk, shift)) {
 99		/* Catch exceptional cases, when connection requires reset.
100		 *      1. Last segment was sent recently. */
101		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
102		    /*  2. Window is closed. */
103		    (!tp->snd_wnd && !tp->packets_out))
104			do_reset = true;
105		if (do_reset)
106			tcp_send_active_reset(sk, GFP_ATOMIC);
107		tcp_done(sk);
108		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
109		return 1;
110	}
111
112	if (!check_net(sock_net(sk))) {
113		/* Not possible to send reset; just close */
114		tcp_done(sk);
115		return 1;
116	}
117
118	return 0;
119}
120
121/**
122 *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
123 *  @sk:    Pointer to the current socket.
124 *  @alive: bool, socket alive state
125 */
126static int tcp_orphan_retries(struct sock *sk, bool alive)
127{
128	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
129
130	/* We know from an ICMP that something is wrong. */
131	if (sk->sk_err_soft && !alive)
132		retries = 0;
133
134	/* However, if socket sent something recently, select some safe
135	 * number of retries. 8 corresponds to >100 seconds with minimal
136	 * RTO of 200msec. */
137	if (retries == 0 && alive)
138		retries = 8;
139	return retries;
140}
141
142static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
143{
144	const struct net *net = sock_net(sk);
145	int mss;
146
147	/* Black hole detection */
148	if (!net->ipv4.sysctl_tcp_mtu_probing)
149		return;
 
 
 
 
 
 
 
150
151	if (!icsk->icsk_mtup.enabled) {
152		icsk->icsk_mtup.enabled = 1;
153		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
154	} else {
155		mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
156		mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
157		mss = max(mss, net->ipv4.sysctl_tcp_mtu_probe_floor);
158		mss = max(mss, net->ipv4.sysctl_tcp_min_snd_mss);
159		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
160	}
161	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
162}
163
164static unsigned int tcp_model_timeout(struct sock *sk,
165				      unsigned int boundary,
166				      unsigned int rto_base)
167{
168	unsigned int linear_backoff_thresh, timeout;
169
170	linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base);
171	if (boundary <= linear_backoff_thresh)
172		timeout = ((2 << boundary) - 1) * rto_base;
173	else
174		timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
175			(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
176	return jiffies_to_msecs(timeout);
177}
178/**
179 *  retransmits_timed_out() - returns true if this connection has timed out
180 *  @sk:       The current socket
181 *  @boundary: max number of retransmissions
182 *  @timeout:  A custom timeout value.
183 *             If set to 0 the default timeout is calculated and used.
184 *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
185 *
186 * The default "timeout" value this function can calculate and use
187 * is equivalent to the timeout of a TCP Connection
188 * after "boundary" unsuccessful, exponentially backed-off
189 * retransmissions with an initial RTO of TCP_RTO_MIN.
190 */
191static bool retransmits_timed_out(struct sock *sk,
192				  unsigned int boundary,
193				  unsigned int timeout)
 
194{
195	unsigned int start_ts;
 
196
197	if (!inet_csk(sk)->icsk_retransmits)
198		return false;
199
200	start_ts = tcp_sk(sk)->retrans_stamp;
 
 
 
201	if (likely(timeout == 0)) {
202		unsigned int rto_base = TCP_RTO_MIN;
203
204		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
205			rto_base = tcp_timeout_init(sk);
206		timeout = tcp_model_timeout(sk, boundary, rto_base);
 
 
207	}
208
209	return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
210}
211
212/* A write timeout has occurred. Process the after effects. */
213static int tcp_write_timeout(struct sock *sk)
214{
215	struct inet_connection_sock *icsk = inet_csk(sk);
216	struct tcp_sock *tp = tcp_sk(sk);
217	struct net *net = sock_net(sk);
218	bool expired = false, do_reset;
219	int retry_until;
 
220
221	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
222		if (icsk->icsk_retransmits) {
223			dst_negative_advice(sk);
224		} else {
225			sk_rethink_txhash(sk);
 
 
 
226		}
227		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
228		expired = icsk->icsk_retransmits >= retry_until;
229	} else {
230		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
 
 
 
 
 
 
 
 
 
 
 
 
231			/* Black hole detection */
232			tcp_mtu_probing(icsk, sk);
233
234			dst_negative_advice(sk);
235		} else {
236			sk_rethink_txhash(sk);
237		}
238
239		retry_until = net->ipv4.sysctl_tcp_retries2;
240		if (sock_flag(sk, SOCK_DEAD)) {
241			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
242
243			retry_until = tcp_orphan_retries(sk, alive);
244			do_reset = alive ||
245				!retransmits_timed_out(sk, retry_until, 0);
246
247			if (tcp_out_of_resources(sk, do_reset))
248				return 1;
249		}
250	}
251	if (!expired)
252		expired = retransmits_timed_out(sk, retry_until,
253						icsk->icsk_user_timeout);
254	tcp_fastopen_active_detect_blackhole(sk, expired);
255
256	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG))
257		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB,
258				  icsk->icsk_retransmits,
259				  icsk->icsk_rto, (int)expired);
260
261	if (expired) {
 
262		/* Has it gone just too far? */
263		tcp_write_err(sk);
264		return 1;
265	}
266
267	return 0;
268}
269
270/* Called with BH disabled */
271void tcp_delack_timer_handler(struct sock *sk)
272{
 
273	struct inet_connection_sock *icsk = inet_csk(sk);
274
275	sk_mem_reclaim_partial(sk);
276
277	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
278	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
279		goto out;
280
281	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
282		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
283		goto out;
284	}
285	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
286
 
 
 
 
 
 
 
 
 
 
 
287	if (inet_csk_ack_scheduled(sk)) {
288		if (!inet_csk_in_pingpong_mode(sk)) {
289			/* Delayed ACK missed: inflate ATO. */
290			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
291		} else {
292			/* Delayed ACK missed: leave pingpong mode and
293			 * deflate ATO.
294			 */
295			inet_csk_exit_pingpong_mode(sk);
296			icsk->icsk_ack.ato      = TCP_ATO_MIN;
297		}
298		tcp_mstamp_refresh(tcp_sk(sk));
299		tcp_send_ack(sk);
300		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
301	}
302
303out:
304	if (tcp_under_memory_pressure(sk))
305		sk_mem_reclaim(sk);
306}
307
308
309/**
310 *  tcp_delack_timer() - The TCP delayed ACK timeout handler
311 *  @data:  Pointer to the current socket. (gets casted to struct sock *)
312 *
313 *  This function gets (indirectly) called when the kernel timer for a TCP packet
314 *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
315 *
316 *  Returns: Nothing (void)
317 */
318static void tcp_delack_timer(struct timer_list *t)
319{
320	struct inet_connection_sock *icsk =
321			from_timer(icsk, t, icsk_delack_timer);
322	struct sock *sk = &icsk->icsk_inet.sk;
323
324	bh_lock_sock(sk);
325	if (!sock_owned_by_user(sk)) {
326		tcp_delack_timer_handler(sk);
327	} else {
328		icsk->icsk_ack.blocked = 1;
329		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
330		/* deleguate our work to tcp_release_cb() */
331		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
332			sock_hold(sk);
333	}
334	bh_unlock_sock(sk);
335	sock_put(sk);
336}
337
338static void tcp_probe_timer(struct sock *sk)
339{
340	struct inet_connection_sock *icsk = inet_csk(sk);
341	struct sk_buff *skb = tcp_send_head(sk);
342	struct tcp_sock *tp = tcp_sk(sk);
343	int max_probes;
 
344
345	if (tp->packets_out || !skb) {
346		icsk->icsk_probes_out = 0;
347		return;
348	}
349
350	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
351	 * long as the receiver continues to respond probes. We support this by
352	 * default and reset icsk_probes_out with incoming ACKs. But if the
353	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
354	 * kill the socket when the retry count and the time exceeds the
355	 * corresponding system limit. We also implement similar policy when
356	 * we use RTO to probe window in tcp_retransmit_timer().
357	 */
358	if (icsk->icsk_user_timeout) {
359		u32 elapsed = tcp_model_timeout(sk, icsk->icsk_probes_out,
360						tcp_probe0_base(sk));
361
362		if (elapsed >= icsk->icsk_user_timeout)
363			goto abort;
364	}
365
366	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
367	if (sock_flag(sk, SOCK_DEAD)) {
368		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
369
370		max_probes = tcp_orphan_retries(sk, alive);
371		if (!alive && icsk->icsk_backoff >= max_probes)
372			goto abort;
373		if (tcp_out_of_resources(sk, true))
374			return;
375	}
376
377	if (icsk->icsk_probes_out >= max_probes) {
378abort:		tcp_write_err(sk);
379	} else {
380		/* Only send another probe if we didn't close things up. */
381		tcp_send_probe0(sk);
382	}
383}
384
385/*
386 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
387 *	sk here is the child socket, not the parent (listener) socket.
388 */
389static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
390{
391	struct inet_connection_sock *icsk = inet_csk(sk);
392	int max_retries = icsk->icsk_syn_retries ? :
393	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
394	struct tcp_sock *tp = tcp_sk(sk);
395
 
396	req->rsk_ops->syn_ack_timeout(req);
397
398	if (req->num_timeout >= max_retries) {
399		tcp_write_err(sk);
400		return;
401	}
402	/* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */
403	if (icsk->icsk_retransmits == 1)
404		tcp_enter_loss(sk);
405	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
406	 * returned from rtx_syn_ack() to make it more persistent like
407	 * regular retransmit because if the child socket has been accepted
408	 * it's not good to give up too easily.
409	 */
410	inet_rtx_syn_ack(sk, req);
411	req->num_timeout++;
412	icsk->icsk_retransmits++;
413	if (!tp->retrans_stamp)
414		tp->retrans_stamp = tcp_time_stamp(tp);
415	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
416			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
417}
418
419
420/**
421 *  tcp_retransmit_timer() - The TCP retransmit timeout handler
422 *  @sk:  Pointer to the current socket.
423 *
424 *  This function gets called when the kernel timer for a TCP packet
425 *  of this socket expires.
426 *
427 *  It handles retransmission, timer adjustment and other necesarry measures.
428 *
429 *  Returns: Nothing (void)
430 */
 
431void tcp_retransmit_timer(struct sock *sk)
432{
433	struct tcp_sock *tp = tcp_sk(sk);
434	struct net *net = sock_net(sk);
435	struct inet_connection_sock *icsk = inet_csk(sk);
436	struct request_sock *req;
437
438	req = rcu_dereference_protected(tp->fastopen_rsk,
439					lockdep_sock_is_held(sk));
440	if (req) {
441		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
442			     sk->sk_state != TCP_FIN_WAIT1);
443		tcp_fastopen_synack_timer(sk, req);
444		/* Before we receive ACK to our SYN-ACK don't retransmit
445		 * anything else (e.g., data or FIN segments).
446		 */
447		return;
448	}
449	if (!tp->packets_out || WARN_ON_ONCE(tcp_rtx_queue_empty(sk)))
450		return;
 
 
451
452	tp->tlp_high_seq = 0;
453
454	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
455	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
456		/* Receiver dastardly shrinks window. Our retransmits
457		 * become zero probes, but we should not timeout this
458		 * connection. If the socket is an orphan, time it out,
459		 * we cannot allow such beasts to hang infinitely.
460		 */
461		struct inet_sock *inet = inet_sk(sk);
462		if (sk->sk_family == AF_INET) {
463			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
464					    &inet->inet_daddr,
465					    ntohs(inet->inet_dport),
466					    inet->inet_num,
467					    tp->snd_una, tp->snd_nxt);
468		}
469#if IS_ENABLED(CONFIG_IPV6)
470		else if (sk->sk_family == AF_INET6) {
471			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
472					    &sk->sk_v6_daddr,
473					    ntohs(inet->inet_dport),
474					    inet->inet_num,
475					    tp->snd_una, tp->snd_nxt);
476		}
477#endif
478		if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
479			tcp_write_err(sk);
480			goto out;
481		}
482		tcp_enter_loss(sk);
483		tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1);
484		__sk_dst_reset(sk);
485		goto out_reset_timer;
486	}
487
488	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
489	if (tcp_write_timeout(sk))
490		goto out;
491
492	if (icsk->icsk_retransmits == 0) {
493		int mib_idx = 0;
494
495		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
496			if (tcp_is_sack(tp))
497				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
498			else
499				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
500		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
501			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
502		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
503			   tp->sacked_out) {
504			if (tcp_is_sack(tp))
505				mib_idx = LINUX_MIB_TCPSACKFAILURES;
506			else
507				mib_idx = LINUX_MIB_TCPRENOFAILURES;
 
 
508		}
509		if (mib_idx)
510			__NET_INC_STATS(sock_net(sk), mib_idx);
511	}
512
513	tcp_enter_loss(sk);
514
515	icsk->icsk_retransmits++;
516	if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) {
517		/* Retransmission failed because of local congestion,
518		 * Let senders fight for local resources conservatively.
519		 */
 
 
520		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
521					  TCP_RESOURCE_PROBE_INTERVAL,
522					  TCP_RTO_MAX);
523		goto out;
524	}
525
526	/* Increase the timeout each time we retransmit.  Note that
527	 * we do not increase the rtt estimate.  rto is initialized
528	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
529	 * that doubling rto each time is the least we can get away with.
530	 * In KA9Q, Karn uses this for the first few times, and then
531	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
532	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
533	 * defined in the protocol as the maximum possible RTT.  I guess
534	 * we'll have to use something other than TCP to talk to the
535	 * University of Mars.
536	 *
537	 * PAWS allows us longer timeouts and large windows, so once
538	 * implemented ftp to mars will work nicely. We will have to fix
539	 * the 120 second clamps though!
540	 */
541	icsk->icsk_backoff++;
 
542
543out_reset_timer:
544	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
545	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
546	 * might be increased if the stream oscillates between thin and thick,
547	 * thus the old value might already be too high compared to the value
548	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
549	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
550	 * exponential backoff behaviour to avoid continue hammering
551	 * linear-timeout retransmissions into a black hole
552	 */
553	if (sk->sk_state == TCP_ESTABLISHED &&
554	    (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) &&
555	    tcp_stream_is_thin(tp) &&
556	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
557		icsk->icsk_backoff = 0;
558		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
559	} else {
560		/* Use normal (exponential) backoff */
561		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
562	}
563	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
564				  tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX);
565	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
566		__sk_dst_reset(sk);
567
568out:;
569}
570
571/* Called with bottom-half processing disabled.
572   Called by tcp_write_timer() */
573void tcp_write_timer_handler(struct sock *sk)
574{
575	struct inet_connection_sock *icsk = inet_csk(sk);
576	int event;
577
578	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
579	    !icsk->icsk_pending)
580		goto out;
581
582	if (time_after(icsk->icsk_timeout, jiffies)) {
583		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
584		goto out;
585	}
586
587	tcp_mstamp_refresh(tcp_sk(sk));
588	event = icsk->icsk_pending;
589
590	switch (event) {
591	case ICSK_TIME_REO_TIMEOUT:
592		tcp_rack_reo_timeout(sk);
593		break;
594	case ICSK_TIME_LOSS_PROBE:
595		tcp_send_loss_probe(sk);
596		break;
597	case ICSK_TIME_RETRANS:
598		icsk->icsk_pending = 0;
599		tcp_retransmit_timer(sk);
600		break;
601	case ICSK_TIME_PROBE0:
602		icsk->icsk_pending = 0;
603		tcp_probe_timer(sk);
604		break;
605	}
606
607out:
608	sk_mem_reclaim(sk);
609}
610
611static void tcp_write_timer(struct timer_list *t)
612{
613	struct inet_connection_sock *icsk =
614			from_timer(icsk, t, icsk_retransmit_timer);
615	struct sock *sk = &icsk->icsk_inet.sk;
616
617	bh_lock_sock(sk);
618	if (!sock_owned_by_user(sk)) {
619		tcp_write_timer_handler(sk);
620	} else {
621		/* delegate our work to tcp_release_cb() */
622		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
623			sock_hold(sk);
624	}
625	bh_unlock_sock(sk);
626	sock_put(sk);
627}
628
629void tcp_syn_ack_timeout(const struct request_sock *req)
630{
631	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
632
633	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
634}
635EXPORT_SYMBOL(tcp_syn_ack_timeout);
636
637void tcp_set_keepalive(struct sock *sk, int val)
638{
639	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
640		return;
641
642	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
643		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
644	else if (!val)
645		inet_csk_delete_keepalive_timer(sk);
646}
647EXPORT_SYMBOL_GPL(tcp_set_keepalive);
648
649
650static void tcp_keepalive_timer (struct timer_list *t)
651{
652	struct sock *sk = from_timer(sk, t, sk_timer);
653	struct inet_connection_sock *icsk = inet_csk(sk);
654	struct tcp_sock *tp = tcp_sk(sk);
655	u32 elapsed;
656
657	/* Only process if socket is not in use. */
658	bh_lock_sock(sk);
659	if (sock_owned_by_user(sk)) {
660		/* Try again later. */
661		inet_csk_reset_keepalive_timer (sk, HZ/20);
662		goto out;
663	}
664
665	if (sk->sk_state == TCP_LISTEN) {
666		pr_err("Hmm... keepalive on a LISTEN ???\n");
667		goto out;
668	}
669
670	tcp_mstamp_refresh(tp);
671	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
672		if (tp->linger2 >= 0) {
673			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
674
675			if (tmo > 0) {
676				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
677				goto out;
678			}
679		}
680		tcp_send_active_reset(sk, GFP_ATOMIC);
681		goto death;
682	}
683
684	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
685	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
686		goto out;
687
688	elapsed = keepalive_time_when(tp);
689
690	/* It is alive without keepalive 8) */
691	if (tp->packets_out || !tcp_write_queue_empty(sk))
692		goto resched;
693
694	elapsed = keepalive_time_elapsed(tp);
695
696	if (elapsed >= keepalive_time_when(tp)) {
697		/* If the TCP_USER_TIMEOUT option is enabled, use that
698		 * to determine when to timeout instead.
699		 */
700		if ((icsk->icsk_user_timeout != 0 &&
701		    elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) &&
702		    icsk->icsk_probes_out > 0) ||
703		    (icsk->icsk_user_timeout == 0 &&
704		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
705			tcp_send_active_reset(sk, GFP_ATOMIC);
706			tcp_write_err(sk);
707			goto out;
708		}
709		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
710			icsk->icsk_probes_out++;
711			elapsed = keepalive_intvl_when(tp);
712		} else {
713			/* If keepalive was lost due to local congestion,
714			 * try harder.
715			 */
716			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
717		}
718	} else {
719		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
720		elapsed = keepalive_time_when(tp) - elapsed;
721	}
722
723	sk_mem_reclaim(sk);
724
725resched:
726	inet_csk_reset_keepalive_timer (sk, elapsed);
727	goto out;
728
729death:
730	tcp_done(sk);
731
732out:
733	bh_unlock_sock(sk);
734	sock_put(sk);
735}
736
737static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer)
738{
739	struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer);
740	struct sock *sk = (struct sock *)tp;
741
742	bh_lock_sock(sk);
743	if (!sock_owned_by_user(sk)) {
744		if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
745			tcp_send_ack(sk);
746	} else {
747		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
748				      &sk->sk_tsq_flags))
749			sock_hold(sk);
750	}
751	bh_unlock_sock(sk);
752
753	sock_put(sk);
754
755	return HRTIMER_NORESTART;
756}
757
758void tcp_init_xmit_timers(struct sock *sk)
759{
760	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
761				  &tcp_keepalive_timer);
762	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
763		     HRTIMER_MODE_ABS_PINNED_SOFT);
764	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
765
766	hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC,
767		     HRTIMER_MODE_REL_PINNED_SOFT);
768	tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick;
769}
v4.6
 
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the  BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Implementation of the Transmission Control Protocol(TCP).
  7 *
  8 * Authors:	Ross Biro
  9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
 11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
 12 *		Florian La Roche, <flla@stud.uni-sb.de>
 13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
 15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
 16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
 17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18 *		Jorge Cwik, <jorge@laser.satlink.net>
 19 */
 20
 21#include <linux/module.h>
 22#include <linux/gfp.h>
 23#include <net/tcp.h>
 24
 25int sysctl_tcp_thin_linear_timeouts __read_mostly;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 26
 27static void tcp_write_err(struct sock *sk)
 28{
 29	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
 30	sk->sk_error_report(sk);
 31
 
 32	tcp_done(sk);
 33	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
 34}
 35
 36/* Do not allow orphaned sockets to eat all our resources.
 37 * This is direct violation of TCP specs, but it is required
 38 * to prevent DoS attacks. It is called when a retransmission timeout
 39 * or zero probe timeout occurs on orphaned socket.
 40 *
 41 * Criteria is still not confirmed experimentally and may change.
 42 * We kill the socket, if:
 43 * 1. If number of orphaned sockets exceeds an administratively configured
 44 *    limit.
 45 * 2. If we have strong memory pressure.
 
 
 
 
 
 
 
 
 
 
 
 
 
 46 */
 47static int tcp_out_of_resources(struct sock *sk, bool do_reset)
 48{
 49	struct tcp_sock *tp = tcp_sk(sk);
 50	int shift = 0;
 51
 52	/* If peer does not open window for long time, or did not transmit
 53	 * anything for long time, penalize it. */
 54	if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
 55		shift++;
 56
 57	/* If some dubious ICMP arrived, penalize even more. */
 58	if (sk->sk_err_soft)
 59		shift++;
 60
 61	if (tcp_check_oom(sk, shift)) {
 62		/* Catch exceptional cases, when connection requires reset.
 63		 *      1. Last segment was sent recently. */
 64		if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
 65		    /*  2. Window is closed. */
 66		    (!tp->snd_wnd && !tp->packets_out))
 67			do_reset = true;
 68		if (do_reset)
 69			tcp_send_active_reset(sk, GFP_ATOMIC);
 70		tcp_done(sk);
 71		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
 
 
 
 
 
 
 72		return 1;
 73	}
 
 74	return 0;
 75}
 76
 77/* Calculate maximal number or retries on an orphaned socket. */
 
 
 
 
 78static int tcp_orphan_retries(struct sock *sk, bool alive)
 79{
 80	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
 81
 82	/* We know from an ICMP that something is wrong. */
 83	if (sk->sk_err_soft && !alive)
 84		retries = 0;
 85
 86	/* However, if socket sent something recently, select some safe
 87	 * number of retries. 8 corresponds to >100 seconds with minimal
 88	 * RTO of 200msec. */
 89	if (retries == 0 && alive)
 90		retries = 8;
 91	return retries;
 92}
 93
 94static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
 95{
 96	struct net *net = sock_net(sk);
 
 97
 98	/* Black hole detection */
 99	if (net->ipv4.sysctl_tcp_mtu_probing) {
100		if (!icsk->icsk_mtup.enabled) {
101			icsk->icsk_mtup.enabled = 1;
102			icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
103			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
104		} else {
105			struct net *net = sock_net(sk);
106			struct tcp_sock *tp = tcp_sk(sk);
107			int mss;
108
109			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
110			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
111			mss = max(mss, 68 - tp->tcp_header_len);
112			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
113			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
114		}
115	}
116}
117
118/* This function calculates a "timeout" which is equivalent to the timeout of a
119 * TCP connection after "boundary" unsuccessful, exponentially backed-off
120 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
121 * syn_set flag is set.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
122 */
123static bool retransmits_timed_out(struct sock *sk,
124				  unsigned int boundary,
125				  unsigned int timeout,
126				  bool syn_set)
127{
128	unsigned int linear_backoff_thresh, start_ts;
129	unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
130
131	if (!inet_csk(sk)->icsk_retransmits)
132		return false;
133
134	start_ts = tcp_sk(sk)->retrans_stamp;
135	if (unlikely(!start_ts))
136		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
137
138	if (likely(timeout == 0)) {
139		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
140
141		if (boundary <= linear_backoff_thresh)
142			timeout = ((2 << boundary) - 1) * rto_base;
143		else
144			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
145				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
146	}
147	return (tcp_time_stamp - start_ts) >= timeout;
 
148}
149
150/* A write timeout has occurred. Process the after effects. */
151static int tcp_write_timeout(struct sock *sk)
152{
153	struct inet_connection_sock *icsk = inet_csk(sk);
154	struct tcp_sock *tp = tcp_sk(sk);
155	struct net *net = sock_net(sk);
 
156	int retry_until;
157	bool do_reset, syn_set = false;
158
159	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
160		if (icsk->icsk_retransmits) {
161			dst_negative_advice(sk);
162			if (tp->syn_fastopen || tp->syn_data)
163				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
164			if (tp->syn_data && icsk->icsk_retransmits == 1)
165				NET_INC_STATS_BH(sock_net(sk),
166						 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
167		}
168		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
169		syn_set = true;
170	} else {
171		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
172			/* Some middle-boxes may black-hole Fast Open _after_
173			 * the handshake. Therefore we conservatively disable
174			 * Fast Open on this path on recurring timeouts with
175			 * few or zero bytes acked after Fast Open.
176			 */
177			if (tp->syn_data_acked &&
178			    tp->bytes_acked <= tp->rx_opt.mss_clamp) {
179				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
180				if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
181					NET_INC_STATS_BH(sock_net(sk),
182							 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
183			}
184			/* Black hole detection */
185			tcp_mtu_probing(icsk, sk);
186
187			dst_negative_advice(sk);
 
 
188		}
189
190		retry_until = net->ipv4.sysctl_tcp_retries2;
191		if (sock_flag(sk, SOCK_DEAD)) {
192			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
193
194			retry_until = tcp_orphan_retries(sk, alive);
195			do_reset = alive ||
196				!retransmits_timed_out(sk, retry_until, 0, 0);
197
198			if (tcp_out_of_resources(sk, do_reset))
199				return 1;
200		}
201	}
 
 
 
 
 
 
 
 
 
202
203	if (retransmits_timed_out(sk, retry_until,
204				  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
205		/* Has it gone just too far? */
206		tcp_write_err(sk);
207		return 1;
208	}
 
209	return 0;
210}
211
 
212void tcp_delack_timer_handler(struct sock *sk)
213{
214	struct tcp_sock *tp = tcp_sk(sk);
215	struct inet_connection_sock *icsk = inet_csk(sk);
216
217	sk_mem_reclaim_partial(sk);
218
219	if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 
220		goto out;
221
222	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
223		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
224		goto out;
225	}
226	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
227
228	if (!skb_queue_empty(&tp->ucopy.prequeue)) {
229		struct sk_buff *skb;
230
231		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
232
233		while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
234			sk_backlog_rcv(sk, skb);
235
236		tp->ucopy.memory = 0;
237	}
238
239	if (inet_csk_ack_scheduled(sk)) {
240		if (!icsk->icsk_ack.pingpong) {
241			/* Delayed ACK missed: inflate ATO. */
242			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
243		} else {
244			/* Delayed ACK missed: leave pingpong mode and
245			 * deflate ATO.
246			 */
247			icsk->icsk_ack.pingpong = 0;
248			icsk->icsk_ack.ato      = TCP_ATO_MIN;
249		}
 
250		tcp_send_ack(sk);
251		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
252	}
253
254out:
255	if (tcp_under_memory_pressure(sk))
256		sk_mem_reclaim(sk);
257}
258
259static void tcp_delack_timer(unsigned long data)
 
 
 
 
 
 
 
 
 
 
260{
261	struct sock *sk = (struct sock *)data;
 
 
262
263	bh_lock_sock(sk);
264	if (!sock_owned_by_user(sk)) {
265		tcp_delack_timer_handler(sk);
266	} else {
267		inet_csk(sk)->icsk_ack.blocked = 1;
268		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
269		/* deleguate our work to tcp_release_cb() */
270		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
271			sock_hold(sk);
272	}
273	bh_unlock_sock(sk);
274	sock_put(sk);
275}
276
277static void tcp_probe_timer(struct sock *sk)
278{
279	struct inet_connection_sock *icsk = inet_csk(sk);
 
280	struct tcp_sock *tp = tcp_sk(sk);
281	int max_probes;
282	u32 start_ts;
283
284	if (tp->packets_out || !tcp_send_head(sk)) {
285		icsk->icsk_probes_out = 0;
286		return;
287	}
288
289	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
290	 * long as the receiver continues to respond probes. We support this by
291	 * default and reset icsk_probes_out with incoming ACKs. But if the
292	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
293	 * kill the socket when the retry count and the time exceeds the
294	 * corresponding system limit. We also implement similar policy when
295	 * we use RTO to probe window in tcp_retransmit_timer().
296	 */
297	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
298	if (!start_ts)
299		skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
300	else if (icsk->icsk_user_timeout &&
301		 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
302		goto abort;
 
303
304	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
305	if (sock_flag(sk, SOCK_DEAD)) {
306		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
307
308		max_probes = tcp_orphan_retries(sk, alive);
309		if (!alive && icsk->icsk_backoff >= max_probes)
310			goto abort;
311		if (tcp_out_of_resources(sk, true))
312			return;
313	}
314
315	if (icsk->icsk_probes_out > max_probes) {
316abort:		tcp_write_err(sk);
317	} else {
318		/* Only send another probe if we didn't close things up. */
319		tcp_send_probe0(sk);
320	}
321}
322
323/*
324 *	Timer for Fast Open socket to retransmit SYNACK. Note that the
325 *	sk here is the child socket, not the parent (listener) socket.
326 */
327static void tcp_fastopen_synack_timer(struct sock *sk)
328{
329	struct inet_connection_sock *icsk = inet_csk(sk);
330	int max_retries = icsk->icsk_syn_retries ? :
331	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
332	struct request_sock *req;
333
334	req = tcp_sk(sk)->fastopen_rsk;
335	req->rsk_ops->syn_ack_timeout(req);
336
337	if (req->num_timeout >= max_retries) {
338		tcp_write_err(sk);
339		return;
340	}
 
 
 
341	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
342	 * returned from rtx_syn_ack() to make it more persistent like
343	 * regular retransmit because if the child socket has been accepted
344	 * it's not good to give up too easily.
345	 */
346	inet_rtx_syn_ack(sk, req);
347	req->num_timeout++;
 
 
 
348	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
349			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
350}
351
352/*
353 *	The TCP retransmit timer.
 
 
 
 
 
 
 
 
 
354 */
355
356void tcp_retransmit_timer(struct sock *sk)
357{
358	struct tcp_sock *tp = tcp_sk(sk);
359	struct net *net = sock_net(sk);
360	struct inet_connection_sock *icsk = inet_csk(sk);
 
361
362	if (tp->fastopen_rsk) {
 
 
363		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
364			     sk->sk_state != TCP_FIN_WAIT1);
365		tcp_fastopen_synack_timer(sk);
366		/* Before we receive ACK to our SYN-ACK don't retransmit
367		 * anything else (e.g., data or FIN segments).
368		 */
369		return;
370	}
371	if (!tp->packets_out)
372		goto out;
373
374	WARN_ON(tcp_write_queue_empty(sk));
375
376	tp->tlp_high_seq = 0;
377
378	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
379	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
380		/* Receiver dastardly shrinks window. Our retransmits
381		 * become zero probes, but we should not timeout this
382		 * connection. If the socket is an orphan, time it out,
383		 * we cannot allow such beasts to hang infinitely.
384		 */
385		struct inet_sock *inet = inet_sk(sk);
386		if (sk->sk_family == AF_INET) {
387			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
388					    &inet->inet_daddr,
389					    ntohs(inet->inet_dport),
390					    inet->inet_num,
391					    tp->snd_una, tp->snd_nxt);
392		}
393#if IS_ENABLED(CONFIG_IPV6)
394		else if (sk->sk_family == AF_INET6) {
395			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
396					    &sk->sk_v6_daddr,
397					    ntohs(inet->inet_dport),
398					    inet->inet_num,
399					    tp->snd_una, tp->snd_nxt);
400		}
401#endif
402		if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
403			tcp_write_err(sk);
404			goto out;
405		}
406		tcp_enter_loss(sk);
407		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
408		__sk_dst_reset(sk);
409		goto out_reset_timer;
410	}
411
 
412	if (tcp_write_timeout(sk))
413		goto out;
414
415	if (icsk->icsk_retransmits == 0) {
416		int mib_idx;
417
418		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
419			if (tcp_is_sack(tp))
420				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
421			else
422				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
423		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
424			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
425		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
426			   tp->sacked_out) {
427			if (tcp_is_sack(tp))
428				mib_idx = LINUX_MIB_TCPSACKFAILURES;
429			else
430				mib_idx = LINUX_MIB_TCPRENOFAILURES;
431		} else {
432			mib_idx = LINUX_MIB_TCPTIMEOUTS;
433		}
434		NET_INC_STATS_BH(sock_net(sk), mib_idx);
 
435	}
436
437	tcp_enter_loss(sk);
438
439	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
 
440		/* Retransmission failed because of local congestion,
441		 * do not backoff.
442		 */
443		if (!icsk->icsk_retransmits)
444			icsk->icsk_retransmits = 1;
445		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
446					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
447					  TCP_RTO_MAX);
448		goto out;
449	}
450
451	/* Increase the timeout each time we retransmit.  Note that
452	 * we do not increase the rtt estimate.  rto is initialized
453	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
454	 * that doubling rto each time is the least we can get away with.
455	 * In KA9Q, Karn uses this for the first few times, and then
456	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
457	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
458	 * defined in the protocol as the maximum possible RTT.  I guess
459	 * we'll have to use something other than TCP to talk to the
460	 * University of Mars.
461	 *
462	 * PAWS allows us longer timeouts and large windows, so once
463	 * implemented ftp to mars will work nicely. We will have to fix
464	 * the 120 second clamps though!
465	 */
466	icsk->icsk_backoff++;
467	icsk->icsk_retransmits++;
468
469out_reset_timer:
470	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
471	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
472	 * might be increased if the stream oscillates between thin and thick,
473	 * thus the old value might already be too high compared to the value
474	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
475	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
476	 * exponential backoff behaviour to avoid continue hammering
477	 * linear-timeout retransmissions into a black hole
478	 */
479	if (sk->sk_state == TCP_ESTABLISHED &&
480	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
481	    tcp_stream_is_thin(tp) &&
482	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
483		icsk->icsk_backoff = 0;
484		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
485	} else {
486		/* Use normal (exponential) backoff */
487		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
488	}
489	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
490	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
 
491		__sk_dst_reset(sk);
492
493out:;
494}
495
 
 
496void tcp_write_timer_handler(struct sock *sk)
497{
498	struct inet_connection_sock *icsk = inet_csk(sk);
499	int event;
500
501	if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
 
502		goto out;
503
504	if (time_after(icsk->icsk_timeout, jiffies)) {
505		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
506		goto out;
507	}
508
 
509	event = icsk->icsk_pending;
510
511	switch (event) {
512	case ICSK_TIME_EARLY_RETRANS:
513		tcp_resume_early_retransmit(sk);
514		break;
515	case ICSK_TIME_LOSS_PROBE:
516		tcp_send_loss_probe(sk);
517		break;
518	case ICSK_TIME_RETRANS:
519		icsk->icsk_pending = 0;
520		tcp_retransmit_timer(sk);
521		break;
522	case ICSK_TIME_PROBE0:
523		icsk->icsk_pending = 0;
524		tcp_probe_timer(sk);
525		break;
526	}
527
528out:
529	sk_mem_reclaim(sk);
530}
531
532static void tcp_write_timer(unsigned long data)
533{
534	struct sock *sk = (struct sock *)data;
 
 
535
536	bh_lock_sock(sk);
537	if (!sock_owned_by_user(sk)) {
538		tcp_write_timer_handler(sk);
539	} else {
540		/* deleguate our work to tcp_release_cb() */
541		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
542			sock_hold(sk);
543	}
544	bh_unlock_sock(sk);
545	sock_put(sk);
546}
547
548void tcp_syn_ack_timeout(const struct request_sock *req)
549{
550	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
551
552	NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
553}
554EXPORT_SYMBOL(tcp_syn_ack_timeout);
555
556void tcp_set_keepalive(struct sock *sk, int val)
557{
558	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
559		return;
560
561	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
562		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
563	else if (!val)
564		inet_csk_delete_keepalive_timer(sk);
565}
 
566
567
568static void tcp_keepalive_timer (unsigned long data)
569{
570	struct sock *sk = (struct sock *) data;
571	struct inet_connection_sock *icsk = inet_csk(sk);
572	struct tcp_sock *tp = tcp_sk(sk);
573	u32 elapsed;
574
575	/* Only process if socket is not in use. */
576	bh_lock_sock(sk);
577	if (sock_owned_by_user(sk)) {
578		/* Try again later. */
579		inet_csk_reset_keepalive_timer (sk, HZ/20);
580		goto out;
581	}
582
583	if (sk->sk_state == TCP_LISTEN) {
584		pr_err("Hmm... keepalive on a LISTEN ???\n");
585		goto out;
586	}
587
 
588	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
589		if (tp->linger2 >= 0) {
590			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
591
592			if (tmo > 0) {
593				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
594				goto out;
595			}
596		}
597		tcp_send_active_reset(sk, GFP_ATOMIC);
598		goto death;
599	}
600
601	if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
 
602		goto out;
603
604	elapsed = keepalive_time_when(tp);
605
606	/* It is alive without keepalive 8) */
607	if (tp->packets_out || tcp_send_head(sk))
608		goto resched;
609
610	elapsed = keepalive_time_elapsed(tp);
611
612	if (elapsed >= keepalive_time_when(tp)) {
613		/* If the TCP_USER_TIMEOUT option is enabled, use that
614		 * to determine when to timeout instead.
615		 */
616		if ((icsk->icsk_user_timeout != 0 &&
617		    elapsed >= icsk->icsk_user_timeout &&
618		    icsk->icsk_probes_out > 0) ||
619		    (icsk->icsk_user_timeout == 0 &&
620		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
621			tcp_send_active_reset(sk, GFP_ATOMIC);
622			tcp_write_err(sk);
623			goto out;
624		}
625		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
626			icsk->icsk_probes_out++;
627			elapsed = keepalive_intvl_when(tp);
628		} else {
629			/* If keepalive was lost due to local congestion,
630			 * try harder.
631			 */
632			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
633		}
634	} else {
635		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
636		elapsed = keepalive_time_when(tp) - elapsed;
637	}
638
639	sk_mem_reclaim(sk);
640
641resched:
642	inet_csk_reset_keepalive_timer (sk, elapsed);
643	goto out;
644
645death:
646	tcp_done(sk);
647
648out:
649	bh_unlock_sock(sk);
650	sock_put(sk);
651}
652
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
653void tcp_init_xmit_timers(struct sock *sk)
654{
655	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
656				  &tcp_keepalive_timer);
 
 
 
 
 
 
 
657}