whiterose

linux unikernel
Log | Files | Refs | README | LICENSE | git clone https://git.ne02ptzero.me/git/whiterose

tcp_timer.c (22230B)


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