[SK_BUFF]: Introduce skb_reset_network_header(skb)
[linux-2.6] / 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  * Version:     $Id: tcp_timer.c,v 1.88 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 #include <linux/module.h>
24 #include <net/tcp.h>
25
26 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
27 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
28 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
29 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
30 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
31 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
32 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
33 int sysctl_tcp_orphan_retries __read_mostly;
34
35 static void tcp_write_timer(unsigned long);
36 static void tcp_delack_timer(unsigned long);
37 static void tcp_keepalive_timer (unsigned long data);
38
39 void tcp_init_xmit_timers(struct sock *sk)
40 {
41         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
42                                   &tcp_keepalive_timer);
43 }
44
45 EXPORT_SYMBOL(tcp_init_xmit_timers);
46
47 static void tcp_write_err(struct sock *sk)
48 {
49         sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
50         sk->sk_error_report(sk);
51
52         tcp_done(sk);
53         NET_INC_STATS_BH(LINUX_MIB_TCPABORTONTIMEOUT);
54 }
55
56 /* Do not allow orphaned sockets to eat all our resources.
57  * This is direct violation of TCP specs, but it is required
58  * to prevent DoS attacks. It is called when a retransmission timeout
59  * or zero probe timeout occurs on orphaned socket.
60  *
61  * Criteria is still not confirmed experimentally and may change.
62  * We kill the socket, if:
63  * 1. If number of orphaned sockets exceeds an administratively configured
64  *    limit.
65  * 2. If we have strong memory pressure.
66  */
67 static int tcp_out_of_resources(struct sock *sk, int do_reset)
68 {
69         struct tcp_sock *tp = tcp_sk(sk);
70         int orphans = atomic_read(&tcp_orphan_count);
71
72         /* If peer does not open window for long time, or did not transmit
73          * anything for long time, penalize it. */
74         if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
75                 orphans <<= 1;
76
77         /* If some dubious ICMP arrived, penalize even more. */
78         if (sk->sk_err_soft)
79                 orphans <<= 1;
80
81         if (orphans >= sysctl_tcp_max_orphans ||
82             (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
83              atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])) {
84                 if (net_ratelimit())
85                         printk(KERN_INFO "Out of socket memory\n");
86
87                 /* Catch exceptional cases, when connection requires reset.
88                  *      1. Last segment was sent recently. */
89                 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
90                     /*  2. Window is closed. */
91                     (!tp->snd_wnd && !tp->packets_out))
92                         do_reset = 1;
93                 if (do_reset)
94                         tcp_send_active_reset(sk, GFP_ATOMIC);
95                 tcp_done(sk);
96                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY);
97                 return 1;
98         }
99         return 0;
100 }
101
102 /* Calculate maximal number or retries on an orphaned socket. */
103 static int tcp_orphan_retries(struct sock *sk, int alive)
104 {
105         int retries = sysctl_tcp_orphan_retries; /* May be zero. */
106
107         /* We know from an ICMP that something is wrong. */
108         if (sk->sk_err_soft && !alive)
109                 retries = 0;
110
111         /* However, if socket sent something recently, select some safe
112          * number of retries. 8 corresponds to >100 seconds with minimal
113          * RTO of 200msec. */
114         if (retries == 0 && alive)
115                 retries = 8;
116         return retries;
117 }
118
119 /* A write timeout has occurred. Process the after effects. */
120 static int tcp_write_timeout(struct sock *sk)
121 {
122         struct inet_connection_sock *icsk = inet_csk(sk);
123         struct tcp_sock *tp = tcp_sk(sk);
124         int retry_until;
125         int mss;
126
127         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
128                 if (icsk->icsk_retransmits)
129                         dst_negative_advice(&sk->sk_dst_cache);
130                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
131         } else {
132                 if (icsk->icsk_retransmits >= sysctl_tcp_retries1) {
133                         /* Black hole detection */
134                         if (sysctl_tcp_mtu_probing) {
135                                 if (!icsk->icsk_mtup.enabled) {
136                                         icsk->icsk_mtup.enabled = 1;
137                                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
138                                 } else {
139                                         mss = min(sysctl_tcp_base_mss,
140                                                   tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)/2);
141                                         mss = max(mss, 68 - tp->tcp_header_len);
142                                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
143                                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
144                                 }
145                         }
146
147                         dst_negative_advice(&sk->sk_dst_cache);
148                 }
149
150                 retry_until = sysctl_tcp_retries2;
151                 if (sock_flag(sk, SOCK_DEAD)) {
152                         const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
153
154                         retry_until = tcp_orphan_retries(sk, alive);
155
156                         if (tcp_out_of_resources(sk, alive || icsk->icsk_retransmits < retry_until))
157                                 return 1;
158                 }
159         }
160
161         if (icsk->icsk_retransmits >= retry_until) {
162                 /* Has it gone just too far? */
163                 tcp_write_err(sk);
164                 return 1;
165         }
166         return 0;
167 }
168
169 static void tcp_delack_timer(unsigned long data)
170 {
171         struct sock *sk = (struct sock*)data;
172         struct tcp_sock *tp = tcp_sk(sk);
173         struct inet_connection_sock *icsk = inet_csk(sk);
174
175         bh_lock_sock(sk);
176         if (sock_owned_by_user(sk)) {
177                 /* Try again later. */
178                 icsk->icsk_ack.blocked = 1;
179                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOCKED);
180                 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN);
181                 goto out_unlock;
182         }
183
184         sk_stream_mem_reclaim(sk);
185
186         if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
187                 goto out;
188
189         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
190                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
191                 goto out;
192         }
193         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
194
195         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
196                 struct sk_buff *skb;
197
198                 NET_INC_STATS_BH(LINUX_MIB_TCPSCHEDULERFAILED);
199
200                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
201                         sk->sk_backlog_rcv(sk, skb);
202
203                 tp->ucopy.memory = 0;
204         }
205
206         if (inet_csk_ack_scheduled(sk)) {
207                 if (!icsk->icsk_ack.pingpong) {
208                         /* Delayed ACK missed: inflate ATO. */
209                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
210                 } else {
211                         /* Delayed ACK missed: leave pingpong mode and
212                          * deflate ATO.
213                          */
214                         icsk->icsk_ack.pingpong = 0;
215                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
216                 }
217                 tcp_send_ack(sk);
218                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKS);
219         }
220         TCP_CHECK_TIMER(sk);
221
222 out:
223         if (tcp_memory_pressure)
224                 sk_stream_mem_reclaim(sk);
225 out_unlock:
226         bh_unlock_sock(sk);
227         sock_put(sk);
228 }
229
230 static void tcp_probe_timer(struct sock *sk)
231 {
232         struct inet_connection_sock *icsk = inet_csk(sk);
233         struct tcp_sock *tp = tcp_sk(sk);
234         int max_probes;
235
236         if (tp->packets_out || !tcp_send_head(sk)) {
237                 icsk->icsk_probes_out = 0;
238                 return;
239         }
240
241         /* *WARNING* RFC 1122 forbids this
242          *
243          * It doesn't AFAIK, because we kill the retransmit timer -AK
244          *
245          * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
246          * this behaviour in Solaris down as a bug fix. [AC]
247          *
248          * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
249          * even if they advertise zero window. Hence, connection is killed only
250          * if we received no ACKs for normal connection timeout. It is not killed
251          * only because window stays zero for some time, window may be zero
252          * until armageddon and even later. We are in full accordance
253          * with RFCs, only probe timer combines both retransmission timeout
254          * and probe timeout in one bottle.                             --ANK
255          */
256         max_probes = sysctl_tcp_retries2;
257
258         if (sock_flag(sk, SOCK_DEAD)) {
259                 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
260
261                 max_probes = tcp_orphan_retries(sk, alive);
262
263                 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
264                         return;
265         }
266
267         if (icsk->icsk_probes_out > max_probes) {
268                 tcp_write_err(sk);
269         } else {
270                 /* Only send another probe if we didn't close things up. */
271                 tcp_send_probe0(sk);
272         }
273 }
274
275 /*
276  *      The TCP retransmit timer.
277  */
278
279 static void tcp_retransmit_timer(struct sock *sk)
280 {
281         struct tcp_sock *tp = tcp_sk(sk);
282         struct inet_connection_sock *icsk = inet_csk(sk);
283
284         if (!tp->packets_out)
285                 goto out;
286
287         BUG_TRAP(!tcp_write_queue_empty(sk));
288
289         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
290             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
291                 /* Receiver dastardly shrinks window. Our retransmits
292                  * become zero probes, but we should not timeout this
293                  * connection. If the socket is an orphan, time it out,
294                  * we cannot allow such beasts to hang infinitely.
295                  */
296 #ifdef TCP_DEBUG
297                 if (net_ratelimit()) {
298                         struct inet_sock *inet = inet_sk(sk);
299                         printk(KERN_DEBUG "TCP: Treason uncloaked! Peer %u.%u.%u.%u:%u/%u shrinks window %u:%u. Repaired.\n",
300                                NIPQUAD(inet->daddr), ntohs(inet->dport),
301                                inet->num, tp->snd_una, tp->snd_nxt);
302                 }
303 #endif
304                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
305                         tcp_write_err(sk);
306                         goto out;
307                 }
308                 tcp_enter_loss(sk, 0);
309                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
310                 __sk_dst_reset(sk);
311                 goto out_reset_timer;
312         }
313
314         if (tcp_write_timeout(sk))
315                 goto out;
316
317         if (icsk->icsk_retransmits == 0) {
318                 if (icsk->icsk_ca_state == TCP_CA_Disorder ||
319                     icsk->icsk_ca_state == TCP_CA_Recovery) {
320                         if (tp->rx_opt.sack_ok) {
321                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
322                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERYFAIL);
323                                 else
324                                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKFAILURES);
325                         } else {
326                                 if (icsk->icsk_ca_state == TCP_CA_Recovery)
327                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERYFAIL);
328                                 else
329                                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOFAILURES);
330                         }
331                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
332                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSFAILURES);
333                 } else {
334                         NET_INC_STATS_BH(LINUX_MIB_TCPTIMEOUTS);
335                 }
336         }
337
338         if (tcp_use_frto(sk)) {
339                 tcp_enter_frto(sk);
340         } else {
341                 tcp_enter_loss(sk, 0);
342         }
343
344         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
345                 /* Retransmission failed because of local congestion,
346                  * do not backoff.
347                  */
348                 if (!icsk->icsk_retransmits)
349                         icsk->icsk_retransmits = 1;
350                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
351                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
352                                           TCP_RTO_MAX);
353                 goto out;
354         }
355
356         /* Increase the timeout each time we retransmit.  Note that
357          * we do not increase the rtt estimate.  rto is initialized
358          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
359          * that doubling rto each time is the least we can get away with.
360          * In KA9Q, Karn uses this for the first few times, and then
361          * goes to quadratic.  netBSD doubles, but only goes up to *64,
362          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
363          * defined in the protocol as the maximum possible RTT.  I guess
364          * we'll have to use something other than TCP to talk to the
365          * University of Mars.
366          *
367          * PAWS allows us longer timeouts and large windows, so once
368          * implemented ftp to mars will work nicely. We will have to fix
369          * the 120 second clamps though!
370          */
371         icsk->icsk_backoff++;
372         icsk->icsk_retransmits++;
373
374 out_reset_timer:
375         icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
376         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
377         if (icsk->icsk_retransmits > sysctl_tcp_retries1)
378                 __sk_dst_reset(sk);
379
380 out:;
381 }
382
383 static void tcp_write_timer(unsigned long data)
384 {
385         struct sock *sk = (struct sock*)data;
386         struct inet_connection_sock *icsk = inet_csk(sk);
387         int event;
388
389         bh_lock_sock(sk);
390         if (sock_owned_by_user(sk)) {
391                 /* Try again later */
392                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20));
393                 goto out_unlock;
394         }
395
396         if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
397                 goto out;
398
399         if (time_after(icsk->icsk_timeout, jiffies)) {
400                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
401                 goto out;
402         }
403
404         event = icsk->icsk_pending;
405         icsk->icsk_pending = 0;
406
407         switch (event) {
408         case ICSK_TIME_RETRANS:
409                 tcp_retransmit_timer(sk);
410                 break;
411         case ICSK_TIME_PROBE0:
412                 tcp_probe_timer(sk);
413                 break;
414         }
415         TCP_CHECK_TIMER(sk);
416
417 out:
418         sk_stream_mem_reclaim(sk);
419 out_unlock:
420         bh_unlock_sock(sk);
421         sock_put(sk);
422 }
423
424 /*
425  *      Timer for listening sockets
426  */
427
428 static void tcp_synack_timer(struct sock *sk)
429 {
430         inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
431                                    TCP_TIMEOUT_INIT, TCP_RTO_MAX);
432 }
433
434 void tcp_set_keepalive(struct sock *sk, int val)
435 {
436         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
437                 return;
438
439         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
440                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
441         else if (!val)
442                 inet_csk_delete_keepalive_timer(sk);
443 }
444
445
446 static void tcp_keepalive_timer (unsigned long data)
447 {
448         struct sock *sk = (struct sock *) data;
449         struct inet_connection_sock *icsk = inet_csk(sk);
450         struct tcp_sock *tp = tcp_sk(sk);
451         __u32 elapsed;
452
453         /* Only process if socket is not in use. */
454         bh_lock_sock(sk);
455         if (sock_owned_by_user(sk)) {
456                 /* Try again later. */
457                 inet_csk_reset_keepalive_timer (sk, HZ/20);
458                 goto out;
459         }
460
461         if (sk->sk_state == TCP_LISTEN) {
462                 tcp_synack_timer(sk);
463                 goto out;
464         }
465
466         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
467                 if (tp->linger2 >= 0) {
468                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
469
470                         if (tmo > 0) {
471                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
472                                 goto out;
473                         }
474                 }
475                 tcp_send_active_reset(sk, GFP_ATOMIC);
476                 goto death;
477         }
478
479         if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
480                 goto out;
481
482         elapsed = keepalive_time_when(tp);
483
484         /* It is alive without keepalive 8) */
485         if (tp->packets_out || tcp_send_head(sk))
486                 goto resched;
487
488         elapsed = tcp_time_stamp - tp->rcv_tstamp;
489
490         if (elapsed >= keepalive_time_when(tp)) {
491                 if ((!tp->keepalive_probes && icsk->icsk_probes_out >= sysctl_tcp_keepalive_probes) ||
492                      (tp->keepalive_probes && icsk->icsk_probes_out >= tp->keepalive_probes)) {
493                         tcp_send_active_reset(sk, GFP_ATOMIC);
494                         tcp_write_err(sk);
495                         goto out;
496                 }
497                 if (tcp_write_wakeup(sk) <= 0) {
498                         icsk->icsk_probes_out++;
499                         elapsed = keepalive_intvl_when(tp);
500                 } else {
501                         /* If keepalive was lost due to local congestion,
502                          * try harder.
503                          */
504                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
505                 }
506         } else {
507                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
508                 elapsed = keepalive_time_when(tp) - elapsed;
509         }
510
511         TCP_CHECK_TIMER(sk);
512         sk_stream_mem_reclaim(sk);
513
514 resched:
515         inet_csk_reset_keepalive_timer (sk, elapsed);
516         goto out;
517
518 death:
519         tcp_done(sk);
520
521 out:
522         bh_unlock_sock(sk);
523         sock_put(sk);
524 }