OSDN Git Service

Merge android-4.4.155 (b3f777e) into msm-4.4
[sagit-ice-cold/kernel_xiaomi_msm8998.git] / net / ipv4 / tcp_timer.c
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 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
32 int sysctl_tcp_orphan_retries __read_mostly;
33 int sysctl_tcp_thin_linear_timeouts __read_mostly;
34
35 /*Function to reset tcp_ack related sysctl on resetting master control */
36 void set_tcp_default(void)
37 {
38         sysctl_tcp_delack_seg   = TCP_DELACK_SEG;
39 }
40
41 /*sysctl handler for tcp_ack realted master control */
42 int tcp_proc_delayed_ack_control(struct ctl_table *table, int write,
43                                  void __user *buffer, size_t *length,
44                                  loff_t *ppos)
45 {
46         int ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
47
48         /* The ret value will be 0 if the input validation is successful
49          * and the values are written to sysctl table. If not, the stack
50          * will continue to work with currently configured values
51          */
52         return ret;
53 }
54
55 /*sysctl handler for tcp_ack realted master control */
56 int tcp_use_userconfig_sysctl_handler(struct ctl_table *table, int write,
57                                       void __user *buffer, size_t *length,
58                                       loff_t *ppos)
59 {
60         int ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
61
62         if (write && ret == 0) {
63                 if (!sysctl_tcp_use_userconfig)
64                         set_tcp_default();
65         }
66         return ret;
67 }
68
69 static void tcp_write_err(struct sock *sk)
70 {
71         sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
72         sk->sk_error_report(sk);
73
74         tcp_done(sk);
75         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
76 }
77
78 /* Do not allow orphaned sockets to eat all our resources.
79  * This is direct violation of TCP specs, but it is required
80  * to prevent DoS attacks. It is called when a retransmission timeout
81  * or zero probe timeout occurs on orphaned socket.
82  *
83  * Also close if our net namespace is exiting; in that case there is no
84  * hope of ever communicating again since all netns interfaces are already
85  * down (or about to be down), and we need to release our dst references,
86  * which have been moved to the netns loopback interface, so the namespace
87  * can finish exiting.  This condition is only possible if we are a kernel
88  * socket, as those do not hold references to the namespace.
89  *
90  * Criteria is still not confirmed experimentally and may change.
91  * We kill the socket, if:
92  * 1. If number of orphaned sockets exceeds an administratively configured
93  *    limit.
94  * 2. If we have strong memory pressure.
95  * 3. If our net namespace is exiting.
96  */
97 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
98 {
99         struct tcp_sock *tp = tcp_sk(sk);
100         int shift = 0;
101
102         /* If peer does not open window for long time, or did not transmit
103          * anything for long time, penalize it. */
104         if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
105                 shift++;
106
107         /* If some dubious ICMP arrived, penalize even more. */
108         if (sk->sk_err_soft)
109                 shift++;
110
111         if (tcp_check_oom(sk, shift)) {
112                 /* Catch exceptional cases, when connection requires reset.
113                  *      1. Last segment was sent recently. */
114                 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
115                     /*  2. Window is closed. */
116                     (!tp->snd_wnd && !tp->packets_out))
117                         do_reset = true;
118                 if (do_reset)
119                         tcp_send_active_reset(sk, GFP_ATOMIC);
120                 tcp_done(sk);
121                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
122                 return 1;
123         }
124
125         if (!check_net(sock_net(sk))) {
126                 /* Not possible to send reset; just close */
127                 tcp_done(sk);
128                 return 1;
129         }
130
131         return 0;
132 }
133
134 /* Calculate maximal number or retries on an orphaned socket. */
135 static int tcp_orphan_retries(struct sock *sk, bool alive)
136 {
137         int retries = sysctl_tcp_orphan_retries; /* May be zero. */
138
139         /* We know from an ICMP that something is wrong. */
140         if (sk->sk_err_soft && !alive)
141                 retries = 0;
142
143         /* However, if socket sent something recently, select some safe
144          * number of retries. 8 corresponds to >100 seconds with minimal
145          * RTO of 200msec. */
146         if (retries == 0 && alive)
147                 retries = 8;
148         return retries;
149 }
150
151 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
152 {
153         struct net *net = sock_net(sk);
154
155         /* Black hole detection */
156         if (net->ipv4.sysctl_tcp_mtu_probing) {
157                 if (!icsk->icsk_mtup.enabled) {
158                         icsk->icsk_mtup.enabled = 1;
159                         icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
160                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
161                 } else {
162                         struct net *net = sock_net(sk);
163                         struct tcp_sock *tp = tcp_sk(sk);
164                         int mss;
165
166                         mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
167                         mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
168                         mss = max(mss, 68 - tp->tcp_header_len);
169                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
170                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
171                 }
172         }
173 }
174
175 /* This function calculates a "timeout" which is equivalent to the timeout of a
176  * TCP connection after "boundary" unsuccessful, exponentially backed-off
177  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
178  * syn_set flag is set.
179  */
180 static bool retransmits_timed_out(struct sock *sk,
181                                   unsigned int boundary,
182                                   unsigned int timeout,
183                                   bool syn_set)
184 {
185         unsigned int linear_backoff_thresh, start_ts;
186         unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
187
188         if (!inet_csk(sk)->icsk_retransmits)
189                 return false;
190
191         start_ts = tcp_sk(sk)->retrans_stamp;
192         if (unlikely(!start_ts))
193                 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
194
195         if (likely(timeout == 0)) {
196                 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
197
198                 if (boundary <= linear_backoff_thresh)
199                         timeout = ((2 << boundary) - 1) * rto_base;
200                 else
201                         timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
202                                 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
203         }
204         return (tcp_time_stamp - start_ts) >= timeout;
205 }
206
207 /* A write timeout has occurred. Process the after effects. */
208 static int tcp_write_timeout(struct sock *sk)
209 {
210         struct inet_connection_sock *icsk = inet_csk(sk);
211         struct tcp_sock *tp = tcp_sk(sk);
212         int retry_until;
213         bool do_reset, syn_set = false;
214
215         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
216                 if (icsk->icsk_retransmits) {
217                         dst_negative_advice(sk);
218                         if (tp->syn_fastopen || tp->syn_data)
219                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
220                         if (tp->syn_data && icsk->icsk_retransmits == 1)
221                                 NET_INC_STATS_BH(sock_net(sk),
222                                                  LINUX_MIB_TCPFASTOPENACTIVEFAIL);
223                 }
224                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
225                 syn_set = true;
226         } else {
227                 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
228                         /* Some middle-boxes may black-hole Fast Open _after_
229                          * the handshake. Therefore we conservatively disable
230                          * Fast Open on this path on recurring timeouts with
231                          * few or zero bytes acked after Fast Open.
232                          */
233                         if (tp->syn_data_acked &&
234                             tp->bytes_acked <= tp->rx_opt.mss_clamp) {
235                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
236                                 if (icsk->icsk_retransmits == sysctl_tcp_retries1)
237                                         NET_INC_STATS_BH(sock_net(sk),
238                                                          LINUX_MIB_TCPFASTOPENACTIVEFAIL);
239                         }
240                         /* Black hole detection */
241                         tcp_mtu_probing(icsk, sk);
242
243                         dst_negative_advice(sk);
244                 }
245
246                 retry_until = sysctl_tcp_retries2;
247                 if (sock_flag(sk, SOCK_DEAD)) {
248                         const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
249
250                         retry_until = tcp_orphan_retries(sk, alive);
251                         do_reset = alive ||
252                                 !retransmits_timed_out(sk, retry_until, 0, 0);
253
254                         if (tcp_out_of_resources(sk, do_reset))
255                                 return 1;
256                 }
257         }
258
259         if (retransmits_timed_out(sk, retry_until,
260                                   syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
261                 /* Has it gone just too far? */
262                 tcp_write_err(sk);
263                 return 1;
264         }
265         return 0;
266 }
267
268 void tcp_delack_timer_handler(struct sock *sk)
269 {
270         struct tcp_sock *tp = tcp_sk(sk);
271         struct inet_connection_sock *icsk = inet_csk(sk);
272
273         sk_mem_reclaim_partial(sk);
274
275         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
276             !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
277                 goto out;
278
279         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
280                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
281                 goto out;
282         }
283         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
284
285         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
286                 struct sk_buff *skb;
287
288                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
289
290                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
291                         sk_backlog_rcv(sk, skb);
292
293                 tp->ucopy.memory = 0;
294         }
295
296         if (inet_csk_ack_scheduled(sk)) {
297                 if (!icsk->icsk_ack.pingpong) {
298                         /* Delayed ACK missed: inflate ATO. */
299                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
300                 } else {
301                         /* Delayed ACK missed: leave pingpong mode and
302                          * deflate ATO.
303                          */
304                         icsk->icsk_ack.pingpong = 0;
305                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
306                 }
307                 tcp_send_ack(sk);
308                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
309         }
310
311 out:
312         if (tcp_under_memory_pressure(sk))
313                 sk_mem_reclaim(sk);
314 }
315
316 static void tcp_delack_timer(unsigned long data)
317 {
318         struct sock *sk = (struct sock *)data;
319
320         bh_lock_sock(sk);
321         if (!sock_owned_by_user(sk)) {
322                 tcp_delack_timer_handler(sk);
323         } else {
324                 inet_csk(sk)->icsk_ack.blocked = 1;
325                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
326                 /* deleguate our work to tcp_release_cb() */
327                 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
328                         sock_hold(sk);
329         }
330         bh_unlock_sock(sk);
331         sock_put(sk);
332 }
333
334 static void tcp_probe_timer(struct sock *sk)
335 {
336         struct inet_connection_sock *icsk = inet_csk(sk);
337         struct tcp_sock *tp = tcp_sk(sk);
338         int max_probes;
339         u32 start_ts;
340
341         if (tp->packets_out || !tcp_send_head(sk)) {
342                 icsk->icsk_probes_out = 0;
343                 return;
344         }
345
346         /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
347          * long as the receiver continues to respond probes. We support this by
348          * default and reset icsk_probes_out with incoming ACKs. But if the
349          * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
350          * kill the socket when the retry count and the time exceeds the
351          * corresponding system limit. We also implement similar policy when
352          * we use RTO to probe window in tcp_retransmit_timer().
353          */
354         start_ts = tcp_skb_timestamp(tcp_send_head(sk));
355         if (!start_ts)
356                 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
357         else if (icsk->icsk_user_timeout &&
358                  (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
359                 goto abort;
360
361         max_probes = sysctl_tcp_retries2;
362         if (sock_flag(sk, SOCK_DEAD)) {
363                 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
364
365                 max_probes = tcp_orphan_retries(sk, alive);
366                 if (!alive && icsk->icsk_backoff >= max_probes)
367                         goto abort;
368                 if (tcp_out_of_resources(sk, true))
369                         return;
370         }
371
372         if (icsk->icsk_probes_out > max_probes) {
373 abort:          tcp_write_err(sk);
374         } else {
375                 /* Only send another probe if we didn't close things up. */
376                 tcp_send_probe0(sk);
377         }
378 }
379
380 /*
381  *      Timer for Fast Open socket to retransmit SYNACK. Note that the
382  *      sk here is the child socket, not the parent (listener) socket.
383  */
384 static void tcp_fastopen_synack_timer(struct sock *sk)
385 {
386         struct inet_connection_sock *icsk = inet_csk(sk);
387         int max_retries = icsk->icsk_syn_retries ? :
388             sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
389         struct request_sock *req;
390
391         req = tcp_sk(sk)->fastopen_rsk;
392         req->rsk_ops->syn_ack_timeout(req);
393
394         if (req->num_timeout >= max_retries) {
395                 tcp_write_err(sk);
396                 return;
397         }
398         /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
399          * returned from rtx_syn_ack() to make it more persistent like
400          * regular retransmit because if the child socket has been accepted
401          * it's not good to give up too easily.
402          */
403         inet_rtx_syn_ack(sk, req);
404         req->num_timeout++;
405         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
406                           TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
407 }
408
409 /*
410  *      The TCP retransmit timer.
411  */
412
413 void tcp_retransmit_timer(struct sock *sk)
414 {
415         struct tcp_sock *tp = tcp_sk(sk);
416         struct inet_connection_sock *icsk = inet_csk(sk);
417
418         if (tp->fastopen_rsk) {
419                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
420                              sk->sk_state != TCP_FIN_WAIT1);
421                 tcp_fastopen_synack_timer(sk);
422                 /* Before we receive ACK to our SYN-ACK don't retransmit
423                  * anything else (e.g., data or FIN segments).
424                  */
425                 return;
426         }
427         if (!tp->packets_out)
428                 goto out;
429
430         WARN_ON(tcp_write_queue_empty(sk));
431
432         tp->tlp_high_seq = 0;
433
434         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
435             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
436                 /* Receiver dastardly shrinks window. Our retransmits
437                  * become zero probes, but we should not timeout this
438                  * connection. If the socket is an orphan, time it out,
439                  * we cannot allow such beasts to hang infinitely.
440                  */
441                 struct inet_sock *inet = inet_sk(sk);
442                 if (sk->sk_family == AF_INET) {
443                         net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
444                                             &inet->inet_daddr,
445                                             ntohs(inet->inet_dport),
446                                             inet->inet_num,
447                                             tp->snd_una, tp->snd_nxt);
448                 }
449 #if IS_ENABLED(CONFIG_IPV6)
450                 else if (sk->sk_family == AF_INET6) {
451                         net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
452                                             &sk->sk_v6_daddr,
453                                             ntohs(inet->inet_dport),
454                                             inet->inet_num,
455                                             tp->snd_una, tp->snd_nxt);
456                 }
457 #endif
458                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
459                         tcp_write_err(sk);
460                         goto out;
461                 }
462                 tcp_enter_loss(sk);
463                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
464                 __sk_dst_reset(sk);
465                 goto out_reset_timer;
466         }
467
468         if (tcp_write_timeout(sk))
469                 goto out;
470
471         if (icsk->icsk_retransmits == 0) {
472                 int mib_idx;
473
474                 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
475                         if (tcp_is_sack(tp))
476                                 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
477                         else
478                                 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
479                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
480                         mib_idx = LINUX_MIB_TCPLOSSFAILURES;
481                 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
482                            tp->sacked_out) {
483                         if (tcp_is_sack(tp))
484                                 mib_idx = LINUX_MIB_TCPSACKFAILURES;
485                         else
486                                 mib_idx = LINUX_MIB_TCPRENOFAILURES;
487                 } else {
488                         mib_idx = LINUX_MIB_TCPTIMEOUTS;
489                 }
490                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
491         }
492
493         tcp_enter_loss(sk);
494
495         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
496                 /* Retransmission failed because of local congestion,
497                  * do not backoff.
498                  */
499                 if (!icsk->icsk_retransmits)
500                         icsk->icsk_retransmits = 1;
501                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
502                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
503                                           TCP_RTO_MAX);
504                 goto out;
505         }
506
507         /* Increase the timeout each time we retransmit.  Note that
508          * we do not increase the rtt estimate.  rto is initialized
509          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
510          * that doubling rto each time is the least we can get away with.
511          * In KA9Q, Karn uses this for the first few times, and then
512          * goes to quadratic.  netBSD doubles, but only goes up to *64,
513          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
514          * defined in the protocol as the maximum possible RTT.  I guess
515          * we'll have to use something other than TCP to talk to the
516          * University of Mars.
517          *
518          * PAWS allows us longer timeouts and large windows, so once
519          * implemented ftp to mars will work nicely. We will have to fix
520          * the 120 second clamps though!
521          */
522         icsk->icsk_backoff++;
523         icsk->icsk_retransmits++;
524
525 out_reset_timer:
526         /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
527          * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
528          * might be increased if the stream oscillates between thin and thick,
529          * thus the old value might already be too high compared to the value
530          * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
531          * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
532          * exponential backoff behaviour to avoid continue hammering
533          * linear-timeout retransmissions into a black hole
534          */
535         if (sk->sk_state == TCP_ESTABLISHED &&
536             (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
537             tcp_stream_is_thin(tp) &&
538             icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
539                 icsk->icsk_backoff = 0;
540                 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
541         } else {
542                 /* Use normal (exponential) backoff */
543                 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
544         }
545         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
546         if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
547                 __sk_dst_reset(sk);
548
549 out:;
550 }
551
552 void tcp_write_timer_handler(struct sock *sk)
553 {
554         struct inet_connection_sock *icsk = inet_csk(sk);
555         int event;
556
557         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
558             !icsk->icsk_pending)
559                 goto out;
560
561         if (time_after(icsk->icsk_timeout, jiffies)) {
562                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
563                 goto out;
564         }
565
566         event = icsk->icsk_pending;
567
568         switch (event) {
569         case ICSK_TIME_EARLY_RETRANS:
570                 tcp_resume_early_retransmit(sk);
571                 break;
572         case ICSK_TIME_LOSS_PROBE:
573                 tcp_send_loss_probe(sk);
574                 break;
575         case ICSK_TIME_RETRANS:
576                 icsk->icsk_pending = 0;
577                 tcp_retransmit_timer(sk);
578                 break;
579         case ICSK_TIME_PROBE0:
580                 icsk->icsk_pending = 0;
581                 tcp_probe_timer(sk);
582                 break;
583         }
584
585 out:
586         sk_mem_reclaim(sk);
587 }
588
589 static void tcp_write_timer(unsigned long data)
590 {
591         struct sock *sk = (struct sock *)data;
592
593         bh_lock_sock(sk);
594         if (!sock_owned_by_user(sk)) {
595                 tcp_write_timer_handler(sk);
596         } else {
597                 /* deleguate our work to tcp_release_cb() */
598                 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
599                         sock_hold(sk);
600         }
601         bh_unlock_sock(sk);
602         sock_put(sk);
603 }
604
605 void tcp_syn_ack_timeout(const struct request_sock *req)
606 {
607         struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
608
609         NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
610 }
611 EXPORT_SYMBOL(tcp_syn_ack_timeout);
612
613 void tcp_set_keepalive(struct sock *sk, int val)
614 {
615         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
616                 return;
617
618         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
619                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
620         else if (!val)
621                 inet_csk_delete_keepalive_timer(sk);
622 }
623
624
625 static void tcp_keepalive_timer (unsigned long data)
626 {
627         struct sock *sk = (struct sock *) data;
628         struct inet_connection_sock *icsk = inet_csk(sk);
629         struct tcp_sock *tp = tcp_sk(sk);
630         u32 elapsed;
631
632         /* Only process if socket is not in use. */
633         bh_lock_sock(sk);
634         if (sock_owned_by_user(sk)) {
635                 /* Try again later. */
636                 inet_csk_reset_keepalive_timer (sk, HZ/20);
637                 goto out;
638         }
639
640         if (sk->sk_state == TCP_LISTEN) {
641                 pr_err("Hmm... keepalive on a LISTEN ???\n");
642                 goto out;
643         }
644
645         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
646                 if (tp->linger2 >= 0) {
647                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
648
649                         if (tmo > 0) {
650                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
651                                 goto out;
652                         }
653                 }
654                 tcp_send_active_reset(sk, GFP_ATOMIC);
655                 goto death;
656         }
657
658         if (!sock_flag(sk, SOCK_KEEPOPEN) ||
659             ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
660                 goto out;
661
662         elapsed = keepalive_time_when(tp);
663
664         /* It is alive without keepalive 8) */
665         if (tp->packets_out || tcp_send_head(sk))
666                 goto resched;
667
668         elapsed = keepalive_time_elapsed(tp);
669
670         if (elapsed >= keepalive_time_when(tp)) {
671                 /* If the TCP_USER_TIMEOUT option is enabled, use that
672                  * to determine when to timeout instead.
673                  */
674                 if ((icsk->icsk_user_timeout != 0 &&
675                     elapsed >= icsk->icsk_user_timeout &&
676                     icsk->icsk_probes_out > 0) ||
677                     (icsk->icsk_user_timeout == 0 &&
678                     icsk->icsk_probes_out >= keepalive_probes(tp))) {
679                         tcp_send_active_reset(sk, GFP_ATOMIC);
680                         tcp_write_err(sk);
681                         goto out;
682                 }
683                 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
684                         icsk->icsk_probes_out++;
685                         elapsed = keepalive_intvl_when(tp);
686                 } else {
687                         /* If keepalive was lost due to local congestion,
688                          * try harder.
689                          */
690                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
691                 }
692         } else {
693                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
694                 elapsed = keepalive_time_when(tp) - elapsed;
695         }
696
697         sk_mem_reclaim(sk);
698
699 resched:
700         inet_csk_reset_keepalive_timer (sk, elapsed);
701         goto out;
702
703 death:
704         tcp_done(sk);
705
706 out:
707         bh_unlock_sock(sk);
708         sock_put(sk);
709 }
710
711 void tcp_init_xmit_timers(struct sock *sk)
712 {
713         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
714                                   &tcp_keepalive_timer);
715 }