OSDN Git Service

Merge android-4.4.183 (94fd428) 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                         mss = max(mss, net->ipv4.sysctl_tcp_min_snd_mss);
170                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
171                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
172                 }
173         }
174 }
175
176 /* This function calculates a "timeout" which is equivalent to the timeout of a
177  * TCP connection after "boundary" unsuccessful, exponentially backed-off
178  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
179  * syn_set flag is set.
180  */
181 static bool retransmits_timed_out(struct sock *sk,
182                                   unsigned int boundary,
183                                   unsigned int timeout,
184                                   bool syn_set)
185 {
186         unsigned int linear_backoff_thresh, start_ts;
187         unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
188
189         if (!inet_csk(sk)->icsk_retransmits)
190                 return false;
191
192         start_ts = tcp_sk(sk)->retrans_stamp;
193         if (unlikely(!start_ts))
194                 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
195
196         if (likely(timeout == 0)) {
197                 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
198
199                 if (boundary <= linear_backoff_thresh)
200                         timeout = ((2 << boundary) - 1) * rto_base;
201                 else
202                         timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
203                                 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
204         }
205         return (tcp_time_stamp - start_ts) >= timeout;
206 }
207
208 /* A write timeout has occurred. Process the after effects. */
209 static int tcp_write_timeout(struct sock *sk)
210 {
211         struct inet_connection_sock *icsk = inet_csk(sk);
212         struct tcp_sock *tp = tcp_sk(sk);
213         int retry_until;
214         bool do_reset, syn_set = false;
215
216         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
217                 if (icsk->icsk_retransmits) {
218                         dst_negative_advice(sk);
219                         if (tp->syn_fastopen || tp->syn_data)
220                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
221                         if (tp->syn_data && icsk->icsk_retransmits == 1)
222                                 NET_INC_STATS_BH(sock_net(sk),
223                                                  LINUX_MIB_TCPFASTOPENACTIVEFAIL);
224                 }
225                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
226                 syn_set = true;
227         } else {
228                 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
229                         /* Some middle-boxes may black-hole Fast Open _after_
230                          * the handshake. Therefore we conservatively disable
231                          * Fast Open on this path on recurring timeouts with
232                          * few or zero bytes acked after Fast Open.
233                          */
234                         if (tp->syn_data_acked &&
235                             tp->bytes_acked <= tp->rx_opt.mss_clamp) {
236                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
237                                 if (icsk->icsk_retransmits == sysctl_tcp_retries1)
238                                         NET_INC_STATS_BH(sock_net(sk),
239                                                          LINUX_MIB_TCPFASTOPENACTIVEFAIL);
240                         }
241                         /* Black hole detection */
242                         tcp_mtu_probing(icsk, sk);
243
244                         dst_negative_advice(sk);
245                 }
246
247                 retry_until = sysctl_tcp_retries2;
248                 if (sock_flag(sk, SOCK_DEAD)) {
249                         const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
250
251                         retry_until = tcp_orphan_retries(sk, alive);
252                         do_reset = alive ||
253                                 !retransmits_timed_out(sk, retry_until, 0, 0);
254
255                         if (tcp_out_of_resources(sk, do_reset))
256                                 return 1;
257                 }
258         }
259
260         if (retransmits_timed_out(sk, retry_until,
261                                   syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
262                 /* Has it gone just too far? */
263                 tcp_write_err(sk);
264                 return 1;
265         }
266         return 0;
267 }
268
269 void tcp_delack_timer_handler(struct sock *sk)
270 {
271         struct tcp_sock *tp = tcp_sk(sk);
272         struct inet_connection_sock *icsk = inet_csk(sk);
273
274         sk_mem_reclaim_partial(sk);
275
276         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
277             !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
278                 goto out;
279
280         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
281                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
282                 goto out;
283         }
284         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
285
286         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
287                 struct sk_buff *skb;
288
289                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
290
291                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
292                         sk_backlog_rcv(sk, skb);
293
294                 tp->ucopy.memory = 0;
295         }
296
297         if (inet_csk_ack_scheduled(sk)) {
298                 if (!icsk->icsk_ack.pingpong) {
299                         /* Delayed ACK missed: inflate ATO. */
300                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
301                 } else {
302                         /* Delayed ACK missed: leave pingpong mode and
303                          * deflate ATO.
304                          */
305                         icsk->icsk_ack.pingpong = 0;
306                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
307                 }
308                 tcp_send_ack(sk);
309                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
310         }
311
312 out:
313         if (tcp_under_memory_pressure(sk))
314                 sk_mem_reclaim(sk);
315 }
316
317 static void tcp_delack_timer(unsigned long data)
318 {
319         struct sock *sk = (struct sock *)data;
320
321         bh_lock_sock(sk);
322         if (!sock_owned_by_user(sk)) {
323                 tcp_delack_timer_handler(sk);
324         } else {
325                 inet_csk(sk)->icsk_ack.blocked = 1;
326                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
327                 /* deleguate our work to tcp_release_cb() */
328                 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
329                         sock_hold(sk);
330         }
331         bh_unlock_sock(sk);
332         sock_put(sk);
333 }
334
335 static void tcp_probe_timer(struct sock *sk)
336 {
337         struct inet_connection_sock *icsk = inet_csk(sk);
338         struct tcp_sock *tp = tcp_sk(sk);
339         int max_probes;
340         u32 start_ts;
341
342         if (tp->packets_out || !tcp_send_head(sk)) {
343                 icsk->icsk_probes_out = 0;
344                 return;
345         }
346
347         /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
348          * long as the receiver continues to respond probes. We support this by
349          * default and reset icsk_probes_out with incoming ACKs. But if the
350          * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
351          * kill the socket when the retry count and the time exceeds the
352          * corresponding system limit. We also implement similar policy when
353          * we use RTO to probe window in tcp_retransmit_timer().
354          */
355         start_ts = tcp_skb_timestamp(tcp_send_head(sk));
356         if (!start_ts)
357                 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
358         else if (icsk->icsk_user_timeout &&
359                  (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
360                 goto abort;
361
362         max_probes = sysctl_tcp_retries2;
363         if (sock_flag(sk, SOCK_DEAD)) {
364                 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
365
366                 max_probes = tcp_orphan_retries(sk, alive);
367                 if (!alive && icsk->icsk_backoff >= max_probes)
368                         goto abort;
369                 if (tcp_out_of_resources(sk, true))
370                         return;
371         }
372
373         if (icsk->icsk_probes_out > max_probes) {
374 abort:          tcp_write_err(sk);
375         } else {
376                 /* Only send another probe if we didn't close things up. */
377                 tcp_send_probe0(sk);
378         }
379 }
380
381 /*
382  *      Timer for Fast Open socket to retransmit SYNACK. Note that the
383  *      sk here is the child socket, not the parent (listener) socket.
384  */
385 static void tcp_fastopen_synack_timer(struct sock *sk)
386 {
387         struct inet_connection_sock *icsk = inet_csk(sk);
388         int max_retries = icsk->icsk_syn_retries ? :
389             sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
390         struct request_sock *req;
391
392         req = tcp_sk(sk)->fastopen_rsk;
393         req->rsk_ops->syn_ack_timeout(req);
394
395         if (req->num_timeout >= max_retries) {
396                 tcp_write_err(sk);
397                 return;
398         }
399         /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
400          * returned from rtx_syn_ack() to make it more persistent like
401          * regular retransmit because if the child socket has been accepted
402          * it's not good to give up too easily.
403          */
404         inet_rtx_syn_ack(sk, req);
405         req->num_timeout++;
406         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
407                           TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
408 }
409
410 /*
411  *      The TCP retransmit timer.
412  */
413
414 void tcp_retransmit_timer(struct sock *sk)
415 {
416         struct tcp_sock *tp = tcp_sk(sk);
417         struct inet_connection_sock *icsk = inet_csk(sk);
418
419         if (tp->fastopen_rsk) {
420                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
421                              sk->sk_state != TCP_FIN_WAIT1);
422                 tcp_fastopen_synack_timer(sk);
423                 /* Before we receive ACK to our SYN-ACK don't retransmit
424                  * anything else (e.g., data or FIN segments).
425                  */
426                 return;
427         }
428         if (!tp->packets_out)
429                 goto out;
430
431         WARN_ON(tcp_write_queue_empty(sk));
432
433         tp->tlp_high_seq = 0;
434
435         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
436             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
437                 /* Receiver dastardly shrinks window. Our retransmits
438                  * become zero probes, but we should not timeout this
439                  * connection. If the socket is an orphan, time it out,
440                  * we cannot allow such beasts to hang infinitely.
441                  */
442                 struct inet_sock *inet = inet_sk(sk);
443                 if (sk->sk_family == AF_INET) {
444                         net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
445                                             &inet->inet_daddr,
446                                             ntohs(inet->inet_dport),
447                                             inet->inet_num,
448                                             tp->snd_una, tp->snd_nxt);
449                 }
450 #if IS_ENABLED(CONFIG_IPV6)
451                 else if (sk->sk_family == AF_INET6) {
452                         net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
453                                             &sk->sk_v6_daddr,
454                                             ntohs(inet->inet_dport),
455                                             inet->inet_num,
456                                             tp->snd_una, tp->snd_nxt);
457                 }
458 #endif
459                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
460                         tcp_write_err(sk);
461                         goto out;
462                 }
463                 tcp_enter_loss(sk);
464                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
465                 __sk_dst_reset(sk);
466                 goto out_reset_timer;
467         }
468
469         if (tcp_write_timeout(sk))
470                 goto out;
471
472         if (icsk->icsk_retransmits == 0) {
473                 int mib_idx;
474
475                 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
476                         if (tcp_is_sack(tp))
477                                 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
478                         else
479                                 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
480                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
481                         mib_idx = LINUX_MIB_TCPLOSSFAILURES;
482                 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
483                            tp->sacked_out) {
484                         if (tcp_is_sack(tp))
485                                 mib_idx = LINUX_MIB_TCPSACKFAILURES;
486                         else
487                                 mib_idx = LINUX_MIB_TCPRENOFAILURES;
488                 } else {
489                         mib_idx = LINUX_MIB_TCPTIMEOUTS;
490                 }
491                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
492         }
493
494         tcp_enter_loss(sk);
495
496         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
497                 /* Retransmission failed because of local congestion,
498                  * do not backoff.
499                  */
500                 if (!icsk->icsk_retransmits)
501                         icsk->icsk_retransmits = 1;
502                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
503                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
504                                           TCP_RTO_MAX);
505                 goto out;
506         }
507
508         /* Increase the timeout each time we retransmit.  Note that
509          * we do not increase the rtt estimate.  rto is initialized
510          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
511          * that doubling rto each time is the least we can get away with.
512          * In KA9Q, Karn uses this for the first few times, and then
513          * goes to quadratic.  netBSD doubles, but only goes up to *64,
514          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
515          * defined in the protocol as the maximum possible RTT.  I guess
516          * we'll have to use something other than TCP to talk to the
517          * University of Mars.
518          *
519          * PAWS allows us longer timeouts and large windows, so once
520          * implemented ftp to mars will work nicely. We will have to fix
521          * the 120 second clamps though!
522          */
523         icsk->icsk_backoff++;
524         icsk->icsk_retransmits++;
525
526 out_reset_timer:
527         /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
528          * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
529          * might be increased if the stream oscillates between thin and thick,
530          * thus the old value might already be too high compared to the value
531          * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
532          * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
533          * exponential backoff behaviour to avoid continue hammering
534          * linear-timeout retransmissions into a black hole
535          */
536         if (sk->sk_state == TCP_ESTABLISHED &&
537             (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
538             tcp_stream_is_thin(tp) &&
539             icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
540                 icsk->icsk_backoff = 0;
541                 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
542         } else {
543                 /* Use normal (exponential) backoff */
544                 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
545         }
546         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
547         if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
548                 __sk_dst_reset(sk);
549
550 out:;
551 }
552
553 void tcp_write_timer_handler(struct sock *sk)
554 {
555         struct inet_connection_sock *icsk = inet_csk(sk);
556         int event;
557
558         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
559             !icsk->icsk_pending)
560                 goto out;
561
562         if (time_after(icsk->icsk_timeout, jiffies)) {
563                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
564                 goto out;
565         }
566
567         event = icsk->icsk_pending;
568
569         switch (event) {
570         case ICSK_TIME_EARLY_RETRANS:
571                 tcp_resume_early_retransmit(sk);
572                 break;
573         case ICSK_TIME_LOSS_PROBE:
574                 tcp_send_loss_probe(sk);
575                 break;
576         case ICSK_TIME_RETRANS:
577                 icsk->icsk_pending = 0;
578                 tcp_retransmit_timer(sk);
579                 break;
580         case ICSK_TIME_PROBE0:
581                 icsk->icsk_pending = 0;
582                 tcp_probe_timer(sk);
583                 break;
584         }
585
586 out:
587         sk_mem_reclaim(sk);
588 }
589
590 static void tcp_write_timer(unsigned long data)
591 {
592         struct sock *sk = (struct sock *)data;
593
594         bh_lock_sock(sk);
595         if (!sock_owned_by_user(sk)) {
596                 tcp_write_timer_handler(sk);
597         } else {
598                 /* deleguate our work to tcp_release_cb() */
599                 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
600                         sock_hold(sk);
601         }
602         bh_unlock_sock(sk);
603         sock_put(sk);
604 }
605
606 void tcp_syn_ack_timeout(const struct request_sock *req)
607 {
608         struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
609
610         NET_INC_STATS_BH(net, LINUX_MIB_TCPTIMEOUTS);
611 }
612 EXPORT_SYMBOL(tcp_syn_ack_timeout);
613
614 void tcp_set_keepalive(struct sock *sk, int val)
615 {
616         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
617                 return;
618
619         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
620                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
621         else if (!val)
622                 inet_csk_delete_keepalive_timer(sk);
623 }
624
625
626 static void tcp_keepalive_timer (unsigned long data)
627 {
628         struct sock *sk = (struct sock *) data;
629         struct inet_connection_sock *icsk = inet_csk(sk);
630         struct tcp_sock *tp = tcp_sk(sk);
631         u32 elapsed;
632
633         /* Only process if socket is not in use. */
634         bh_lock_sock(sk);
635         if (sock_owned_by_user(sk)) {
636                 /* Try again later. */
637                 inet_csk_reset_keepalive_timer (sk, HZ/20);
638                 goto out;
639         }
640
641         if (sk->sk_state == TCP_LISTEN) {
642                 pr_err("Hmm... keepalive on a LISTEN ???\n");
643                 goto out;
644         }
645
646         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
647                 if (tp->linger2 >= 0) {
648                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
649
650                         if (tmo > 0) {
651                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
652                                 goto out;
653                         }
654                 }
655                 tcp_send_active_reset(sk, GFP_ATOMIC);
656                 goto death;
657         }
658
659         if (!sock_flag(sk, SOCK_KEEPOPEN) ||
660             ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
661                 goto out;
662
663         elapsed = keepalive_time_when(tp);
664
665         /* It is alive without keepalive 8) */
666         if (tp->packets_out || tcp_send_head(sk))
667                 goto resched;
668
669         elapsed = keepalive_time_elapsed(tp);
670
671         if (elapsed >= keepalive_time_when(tp)) {
672                 /* If the TCP_USER_TIMEOUT option is enabled, use that
673                  * to determine when to timeout instead.
674                  */
675                 if ((icsk->icsk_user_timeout != 0 &&
676                     elapsed >= icsk->icsk_user_timeout &&
677                     icsk->icsk_probes_out > 0) ||
678                     (icsk->icsk_user_timeout == 0 &&
679                     icsk->icsk_probes_out >= keepalive_probes(tp))) {
680                         tcp_send_active_reset(sk, GFP_ATOMIC);
681                         tcp_write_err(sk);
682                         goto out;
683                 }
684                 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
685                         icsk->icsk_probes_out++;
686                         elapsed = keepalive_intvl_when(tp);
687                 } else {
688                         /* If keepalive was lost due to local congestion,
689                          * try harder.
690                          */
691                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
692                 }
693         } else {
694                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
695                 elapsed = keepalive_time_when(tp) - elapsed;
696         }
697
698         sk_mem_reclaim(sk);
699
700 resched:
701         inet_csk_reset_keepalive_timer (sk, elapsed);
702         goto out;
703
704 death:
705         tcp_done(sk);
706
707 out:
708         bh_unlock_sock(sk);
709         sock_put(sk);
710 }
711
712 void tcp_init_xmit_timers(struct sock *sk)
713 {
714         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
715                                   &tcp_keepalive_timer);
716 }