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.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_minisocks.c,v 1.15 2002/02/01 22:01:04 davem Exp $
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>
23 #include <linux/config.h>
25 #include <linux/module.h>
26 #include <linux/sysctl.h>
27 #include <linux/workqueue.h>
29 #include <net/inet_common.h>
33 #define SYNC_INIT 0 /* let the user enable it */
38 int sysctl_tcp_tw_recycle
;
39 int sysctl_tcp_max_tw_buckets
= NR_FILE
*2;
41 int sysctl_tcp_syncookies
= SYNC_INIT
;
42 int sysctl_tcp_abort_on_overflow
;
44 static void tcp_tw_schedule(struct inet_timewait_sock
*tw
, int timeo
);
46 static __inline__
int tcp_in_window(u32 seq
, u32 end_seq
, u32 s_win
, u32 e_win
)
50 if (after(end_seq
, s_win
) && before(seq
, e_win
))
52 return (seq
== e_win
&& seq
== end_seq
);
55 /* New-style handling of TIME_WAIT sockets. */
60 /* Must be called with locally disabled BHs. */
61 static void tcp_timewait_kill(struct inet_timewait_sock
*tw
)
63 struct inet_bind_hashbucket
*bhead
;
64 struct inet_bind_bucket
*tb
;
65 /* Unlink from established hashes. */
66 struct inet_ehash_bucket
*ehead
= &tcp_hashinfo
.ehash
[tw
->tw_hashent
];
68 write_lock(&ehead
->lock
);
69 if (hlist_unhashed(&tw
->tw_node
)) {
70 write_unlock(&ehead
->lock
);
73 __hlist_del(&tw
->tw_node
);
74 sk_node_init(&tw
->tw_node
);
75 write_unlock(&ehead
->lock
);
77 /* Disassociate with bind bucket. */
78 bhead
= &tcp_hashinfo
.bhash
[inet_bhashfn(tw
->tw_num
, tcp_hashinfo
.bhash_size
)];
79 spin_lock(&bhead
->lock
);
81 __hlist_del(&tw
->tw_bind_node
);
83 inet_bind_bucket_destroy(tcp_hashinfo
.bind_bucket_cachep
, tb
);
84 spin_unlock(&bhead
->lock
);
86 #ifdef SOCK_REFCNT_DEBUG
87 if (atomic_read(&tw
->tw_refcnt
) != 1) {
88 printk(KERN_DEBUG
"%s timewait_sock %p refcnt=%d\n",
89 tw
->tw_prot
->name
, tw
, atomic_read(&tw
->tw_refcnt
));
96 * * Main purpose of TIME-WAIT state is to close connection gracefully,
97 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
98 * (and, probably, tail of data) and one or more our ACKs are lost.
99 * * What is TIME-WAIT timeout? It is associated with maximal packet
100 * lifetime in the internet, which results in wrong conclusion, that
101 * it is set to catch "old duplicate segments" wandering out of their path.
102 * It is not quite correct. This timeout is calculated so that it exceeds
103 * maximal retransmission timeout enough to allow to lose one (or more)
104 * segments sent by peer and our ACKs. This time may be calculated from RTO.
105 * * When TIME-WAIT socket receives RST, it means that another end
106 * finally closed and we are allowed to kill TIME-WAIT too.
107 * * Second purpose of TIME-WAIT is catching old duplicate segments.
108 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
109 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
110 * * If we invented some more clever way to catch duplicates
111 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
113 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
114 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
115 * from the very beginning.
117 * NOTE. With recycling (and later with fin-wait-2) TW bucket
118 * is _not_ stateless. It means, that strictly speaking we must
119 * spinlock it. I do not want! Well, probability of misbehaviour
120 * is ridiculously low and, seems, we could use some mb() tricks
121 * to avoid misread sequence numbers, states etc. --ANK
124 tcp_timewait_state_process(struct inet_timewait_sock
*tw
, struct sk_buff
*skb
,
125 const struct tcphdr
*th
)
127 struct tcp_timewait_sock
*tcptw
= tcp_twsk((struct sock
*)tw
);
128 struct tcp_options_received tmp_opt
;
131 tmp_opt
.saw_tstamp
= 0;
132 if (th
->doff
> (sizeof(*th
) >> 2) && tcptw
->tw_ts_recent_stamp
) {
133 tcp_parse_options(skb
, &tmp_opt
, 0);
135 if (tmp_opt
.saw_tstamp
) {
136 tmp_opt
.ts_recent
= tcptw
->tw_ts_recent
;
137 tmp_opt
.ts_recent_stamp
= tcptw
->tw_ts_recent_stamp
;
138 paws_reject
= tcp_paws_check(&tmp_opt
, th
->rst
);
142 if (tw
->tw_substate
== TCP_FIN_WAIT2
) {
143 /* Just repeat all the checks of tcp_rcv_state_process() */
145 /* Out of window, send ACK */
147 !tcp_in_window(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
,
149 tcptw
->tw_rcv_nxt
+ tcptw
->tw_rcv_wnd
))
155 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tcptw
->tw_rcv_nxt
))
159 if (!after(TCP_SKB_CB(skb
)->end_seq
, tcptw
->tw_rcv_nxt
) ||
160 TCP_SKB_CB(skb
)->end_seq
== TCP_SKB_CB(skb
)->seq
) {
162 return TCP_TW_SUCCESS
;
165 /* New data or FIN. If new data arrive after half-duplex close,
169 TCP_SKB_CB(skb
)->end_seq
!= tcptw
->tw_rcv_nxt
+ 1) {
171 tcp_tw_deschedule(tw
);
176 /* FIN arrived, enter true time-wait state. */
177 tw
->tw_substate
= TCP_TIME_WAIT
;
178 tcptw
->tw_rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
179 if (tmp_opt
.saw_tstamp
) {
180 tcptw
->tw_ts_recent_stamp
= xtime
.tv_sec
;
181 tcptw
->tw_ts_recent
= tmp_opt
.rcv_tsval
;
184 /* I am shamed, but failed to make it more elegant.
185 * Yes, it is direct reference to IP, which is impossible
186 * to generalize to IPv6. Taking into account that IPv6
187 * do not undertsnad recycling in any case, it not
188 * a big problem in practice. --ANK */
189 if (tw
->tw_family
== AF_INET
&&
190 sysctl_tcp_tw_recycle
&& tcptw
->tw_ts_recent_stamp
&&
191 tcp_v4_tw_remember_stamp(tw
))
192 tcp_tw_schedule(tw
, tw
->tw_timeout
);
194 tcp_tw_schedule(tw
, TCP_TIMEWAIT_LEN
);
199 * Now real TIME-WAIT state.
202 * "When a connection is [...] on TIME-WAIT state [...]
203 * [a TCP] MAY accept a new SYN from the remote TCP to
204 * reopen the connection directly, if it:
206 * (1) assigns its initial sequence number for the new
207 * connection to be larger than the largest sequence
208 * number it used on the previous connection incarnation,
211 * (2) returns to TIME-WAIT state if the SYN turns out
212 * to be an old duplicate".
216 (TCP_SKB_CB(skb
)->seq
== tcptw
->tw_rcv_nxt
&&
217 (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
|| th
->rst
))) {
218 /* In window segment, it may be only reset or bare ack. */
221 /* This is TIME_WAIT assasination, in two flavors.
222 * Oh well... nobody has a sufficient solution to this
225 if (sysctl_tcp_rfc1337
== 0) {
227 tcp_tw_deschedule(tw
);
229 return TCP_TW_SUCCESS
;
232 tcp_tw_schedule(tw
, TCP_TIMEWAIT_LEN
);
234 if (tmp_opt
.saw_tstamp
) {
235 tcptw
->tw_ts_recent
= tmp_opt
.rcv_tsval
;
236 tcptw
->tw_ts_recent_stamp
= xtime
.tv_sec
;
240 return TCP_TW_SUCCESS
;
243 /* Out of window segment.
245 All the segments are ACKed immediately.
247 The only exception is new SYN. We accept it, if it is
248 not old duplicate and we are not in danger to be killed
249 by delayed old duplicates. RFC check is that it has
250 newer sequence number works at rates <40Mbit/sec.
251 However, if paws works, it is reliable AND even more,
252 we even may relax silly seq space cutoff.
254 RED-PEN: we violate main RFC requirement, if this SYN will appear
255 old duplicate (i.e. we receive RST in reply to SYN-ACK),
256 we must return socket to time-wait state. It is not good,
260 if (th
->syn
&& !th
->rst
&& !th
->ack
&& !paws_reject
&&
261 (after(TCP_SKB_CB(skb
)->seq
, tcptw
->tw_rcv_nxt
) ||
262 (tmp_opt
.saw_tstamp
&&
263 (s32
)(tcptw
->tw_ts_recent
- tmp_opt
.rcv_tsval
) < 0))) {
264 u32 isn
= tcptw
->tw_snd_nxt
+ 65535 + 2;
267 TCP_SKB_CB(skb
)->when
= isn
;
272 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
275 /* In this case we must reset the TIMEWAIT timer.
277 * If it is ACKless SYN it may be both old duplicate
278 * and new good SYN with random sequence number <rcv_nxt.
279 * Do not reschedule in the last case.
281 if (paws_reject
|| th
->ack
)
282 tcp_tw_schedule(tw
, TCP_TIMEWAIT_LEN
);
284 /* Send ACK. Note, we do not put the bucket,
285 * it will be released by caller.
290 return TCP_TW_SUCCESS
;
293 /* Enter the time wait state. This is called with locally disabled BH.
294 * Essentially we whip up a timewait bucket, copy the
295 * relevant info into it from the SK, and mess with hash chains
298 static void __tcp_tw_hashdance(struct sock
*sk
, struct inet_timewait_sock
*tw
)
300 const struct inet_sock
*inet
= inet_sk(sk
);
301 struct inet_ehash_bucket
*ehead
= &tcp_hashinfo
.ehash
[sk
->sk_hashent
];
302 struct inet_bind_hashbucket
*bhead
;
303 /* Step 1: Put TW into bind hash. Original socket stays there too.
304 Note, that any socket with inet->num != 0 MUST be bound in
305 binding cache, even if it is closed.
307 bhead
= &tcp_hashinfo
.bhash
[inet_bhashfn(inet
->num
, tcp_hashinfo
.bhash_size
)];
308 spin_lock(&bhead
->lock
);
309 tw
->tw_tb
= inet
->bind_hash
;
310 BUG_TRAP(inet
->bind_hash
);
311 inet_twsk_add_bind_node(tw
, &tw
->tw_tb
->owners
);
312 spin_unlock(&bhead
->lock
);
314 write_lock(&ehead
->lock
);
316 /* Step 2: Remove SK from established hash. */
317 if (__sk_del_node_init(sk
))
318 sock_prot_dec_use(sk
->sk_prot
);
320 /* Step 3: Hash TW into TIMEWAIT half of established hash table. */
321 inet_twsk_add_node(tw
, &(ehead
+ tcp_hashinfo
.ehash_size
)->chain
);
322 atomic_inc(&tw
->tw_refcnt
);
324 write_unlock(&ehead
->lock
);
328 * Move a socket to time-wait or dead fin-wait-2 state.
330 void tcp_time_wait(struct sock
*sk
, int state
, int timeo
)
332 struct inet_timewait_sock
*tw
= NULL
;
333 const struct tcp_sock
*tp
= tcp_sk(sk
);
336 if (sysctl_tcp_tw_recycle
&& tp
->rx_opt
.ts_recent_stamp
)
337 recycle_ok
= tp
->af_specific
->remember_stamp(sk
);
339 if (tcp_tw_count
< sysctl_tcp_max_tw_buckets
)
340 tw
= kmem_cache_alloc(sk
->sk_prot_creator
->twsk_slab
, SLAB_ATOMIC
);
343 struct tcp_timewait_sock
*tcptw
= tcp_twsk((struct sock
*)tw
);
344 const struct inet_sock
*inet
= inet_sk(sk
);
345 const int rto
= (tp
->rto
<< 2) - (tp
->rto
>> 1);
347 /* Remember our protocol */
348 tw
->tw_prot
= sk
->sk_prot_creator
;
350 /* Give us an identity. */
351 tw
->tw_daddr
= inet
->daddr
;
352 tw
->tw_rcv_saddr
= inet
->rcv_saddr
;
353 tw
->tw_bound_dev_if
= sk
->sk_bound_dev_if
;
354 tw
->tw_num
= inet
->num
;
355 tw
->tw_state
= TCP_TIME_WAIT
;
356 tw
->tw_substate
= state
;
357 tw
->tw_sport
= inet
->sport
;
358 tw
->tw_dport
= inet
->dport
;
359 tw
->tw_family
= sk
->sk_family
;
360 tw
->tw_reuse
= sk
->sk_reuse
;
361 tw
->tw_rcv_wscale
= tp
->rx_opt
.rcv_wscale
;
362 atomic_set(&tw
->tw_refcnt
, 1);
364 tw
->tw_hashent
= sk
->sk_hashent
;
365 tcptw
->tw_rcv_nxt
= tp
->rcv_nxt
;
366 tcptw
->tw_snd_nxt
= tp
->snd_nxt
;
367 tcptw
->tw_rcv_wnd
= tcp_receive_window(tp
);
368 tcptw
->tw_ts_recent
= tp
->rx_opt
.ts_recent
;
369 tcptw
->tw_ts_recent_stamp
= tp
->rx_opt
.ts_recent_stamp
;
370 inet_twsk_dead_node_init(tw
);
372 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
373 if (tw
->tw_family
== PF_INET6
) {
374 struct ipv6_pinfo
*np
= inet6_sk(sk
);
375 struct tcp6_timewait_sock
*tcp6tw
= tcp6_twsk((struct sock
*)tw
);
377 ipv6_addr_copy(&tcp6tw
->tw_v6_daddr
, &np
->daddr
);
378 ipv6_addr_copy(&tcp6tw
->tw_v6_rcv_saddr
, &np
->rcv_saddr
);
379 tw
->tw_ipv6only
= np
->ipv6only
;
383 /* Linkage updates. */
384 __tcp_tw_hashdance(sk
, tw
);
386 /* Get the TIME_WAIT timeout firing. */
391 tw
->tw_timeout
= rto
;
393 tw
->tw_timeout
= TCP_TIMEWAIT_LEN
;
394 if (state
== TCP_TIME_WAIT
)
395 timeo
= TCP_TIMEWAIT_LEN
;
398 tcp_tw_schedule(tw
, timeo
);
401 /* Sorry, if we're out of memory, just CLOSE this
402 * socket up. We've got bigger problems than
403 * non-graceful socket closings.
406 printk(KERN_INFO
"TCP: time wait bucket table overflow\n");
409 tcp_update_metrics(sk
);
413 /* Kill off TIME_WAIT sockets once their lifetime has expired. */
414 static int tcp_tw_death_row_slot
;
416 static void tcp_twkill(unsigned long);
418 /* TIME_WAIT reaping mechanism. */
419 #define TCP_TWKILL_SLOTS 8 /* Please keep this a power of 2. */
420 #define TCP_TWKILL_PERIOD (TCP_TIMEWAIT_LEN/TCP_TWKILL_SLOTS)
422 #define TCP_TWKILL_QUOTA 100
424 static struct hlist_head tcp_tw_death_row
[TCP_TWKILL_SLOTS
];
425 static DEFINE_SPINLOCK(tw_death_lock
);
426 static struct timer_list tcp_tw_timer
= TIMER_INITIALIZER(tcp_twkill
, 0, 0);
427 static void twkill_work(void *);
428 static DECLARE_WORK(tcp_twkill_work
, twkill_work
, NULL
);
429 static u32 twkill_thread_slots
;
431 /* Returns non-zero if quota exceeded. */
432 static int tcp_do_twkill_work(int slot
, unsigned int quota
)
434 struct inet_timewait_sock
*tw
;
435 struct hlist_node
*node
;
439 /* NOTE: compare this to previous version where lock
440 * was released after detaching chain. It was racy,
441 * because tw buckets are scheduled in not serialized context
442 * in 2.3 (with netfilter), and with softnet it is common, because
443 * soft irqs are not sequenced.
448 inet_twsk_for_each_inmate(tw
, node
, &tcp_tw_death_row
[slot
]) {
449 __inet_twsk_del_dead_node(tw
);
450 spin_unlock(&tw_death_lock
);
451 tcp_timewait_kill(tw
);
454 spin_lock(&tw_death_lock
);
455 if (killed
> quota
) {
460 /* While we dropped tw_death_lock, another cpu may have
461 * killed off the next TW bucket in the list, therefore
462 * do a fresh re-read of the hlist head node with the
463 * lock reacquired. We still use the hlist traversal
464 * macro in order to get the prefetches.
469 tcp_tw_count
-= killed
;
470 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED
, killed
);
475 static void tcp_twkill(unsigned long dummy
)
479 spin_lock(&tw_death_lock
);
481 if (tcp_tw_count
== 0)
485 ret
= tcp_do_twkill_work(tcp_tw_death_row_slot
, TCP_TWKILL_QUOTA
);
487 twkill_thread_slots
|= (1 << tcp_tw_death_row_slot
);
489 schedule_work(&tcp_twkill_work
);
492 /* We purged the entire slot, anything left? */
496 tcp_tw_death_row_slot
=
497 ((tcp_tw_death_row_slot
+ 1) & (TCP_TWKILL_SLOTS
- 1));
499 mod_timer(&tcp_tw_timer
, jiffies
+ TCP_TWKILL_PERIOD
);
501 spin_unlock(&tw_death_lock
);
504 extern void twkill_slots_invalid(void);
506 static void twkill_work(void *dummy
)
510 if ((TCP_TWKILL_SLOTS
- 1) > (sizeof(twkill_thread_slots
) * 8))
511 twkill_slots_invalid();
513 while (twkill_thread_slots
) {
514 spin_lock_bh(&tw_death_lock
);
515 for (i
= 0; i
< TCP_TWKILL_SLOTS
; i
++) {
516 if (!(twkill_thread_slots
& (1 << i
)))
519 while (tcp_do_twkill_work(i
, TCP_TWKILL_QUOTA
) != 0) {
520 if (need_resched()) {
521 spin_unlock_bh(&tw_death_lock
);
523 spin_lock_bh(&tw_death_lock
);
527 twkill_thread_slots
&= ~(1 << i
);
529 spin_unlock_bh(&tw_death_lock
);
533 /* These are always called from BH context. See callers in
534 * tcp_input.c to verify this.
537 /* This is for handling early-kills of TIME_WAIT sockets. */
538 void tcp_tw_deschedule(struct inet_timewait_sock
*tw
)
540 spin_lock(&tw_death_lock
);
541 if (inet_twsk_del_dead_node(tw
)) {
543 if (--tcp_tw_count
== 0)
544 del_timer(&tcp_tw_timer
);
546 spin_unlock(&tw_death_lock
);
547 tcp_timewait_kill(tw
);
550 /* Short-time timewait calendar */
552 static int tcp_twcal_hand
= -1;
553 static int tcp_twcal_jiffie
;
554 static void tcp_twcal_tick(unsigned long);
555 static struct timer_list tcp_twcal_timer
=
556 TIMER_INITIALIZER(tcp_twcal_tick
, 0, 0);
557 static struct hlist_head tcp_twcal_row
[TCP_TW_RECYCLE_SLOTS
];
559 static void tcp_tw_schedule(struct inet_timewait_sock
*tw
, const int timeo
)
561 struct hlist_head
*list
;
564 /* timeout := RTO * 3.5
566 * 3.5 = 1+2+0.5 to wait for two retransmits.
568 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
569 * our ACK acking that FIN can be lost. If N subsequent retransmitted
570 * FINs (or previous seqments) are lost (probability of such event
571 * is p^(N+1), where p is probability to lose single packet and
572 * time to detect the loss is about RTO*(2^N - 1) with exponential
573 * backoff). Normal timewait length is calculated so, that we
574 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
575 * [ BTW Linux. following BSD, violates this requirement waiting
576 * only for 60sec, we should wait at least for 240 secs.
577 * Well, 240 consumes too much of resources 8)
579 * This interval is not reduced to catch old duplicate and
580 * responces to our wandering segments living for two MSLs.
581 * However, if we use PAWS to detect
582 * old duplicates, we can reduce the interval to bounds required
583 * by RTO, rather than MSL. So, if peer understands PAWS, we
584 * kill tw bucket after 3.5*RTO (it is important that this number
585 * is greater than TS tick!) and detect old duplicates with help
588 slot
= (timeo
+ (1<<TCP_TW_RECYCLE_TICK
) - 1) >> TCP_TW_RECYCLE_TICK
;
590 spin_lock(&tw_death_lock
);
592 /* Unlink it, if it was scheduled */
593 if (inet_twsk_del_dead_node(tw
))
596 atomic_inc(&tw
->tw_refcnt
);
598 if (slot
>= TCP_TW_RECYCLE_SLOTS
) {
599 /* Schedule to slow timer */
600 if (timeo
>= TCP_TIMEWAIT_LEN
) {
601 slot
= TCP_TWKILL_SLOTS
-1;
603 slot
= (timeo
+ TCP_TWKILL_PERIOD
-1) / TCP_TWKILL_PERIOD
;
604 if (slot
>= TCP_TWKILL_SLOTS
)
605 slot
= TCP_TWKILL_SLOTS
-1;
607 tw
->tw_ttd
= jiffies
+ timeo
;
608 slot
= (tcp_tw_death_row_slot
+ slot
) & (TCP_TWKILL_SLOTS
- 1);
609 list
= &tcp_tw_death_row
[slot
];
611 tw
->tw_ttd
= jiffies
+ (slot
<< TCP_TW_RECYCLE_TICK
);
613 if (tcp_twcal_hand
< 0) {
615 tcp_twcal_jiffie
= jiffies
;
616 tcp_twcal_timer
.expires
= tcp_twcal_jiffie
+ (slot
<<TCP_TW_RECYCLE_TICK
);
617 add_timer(&tcp_twcal_timer
);
619 if (time_after(tcp_twcal_timer
.expires
, jiffies
+ (slot
<<TCP_TW_RECYCLE_TICK
)))
620 mod_timer(&tcp_twcal_timer
, jiffies
+ (slot
<<TCP_TW_RECYCLE_TICK
));
621 slot
= (tcp_twcal_hand
+ slot
)&(TCP_TW_RECYCLE_SLOTS
-1);
623 list
= &tcp_twcal_row
[slot
];
626 hlist_add_head(&tw
->tw_death_node
, list
);
628 if (tcp_tw_count
++ == 0)
629 mod_timer(&tcp_tw_timer
, jiffies
+TCP_TWKILL_PERIOD
);
630 spin_unlock(&tw_death_lock
);
633 void tcp_twcal_tick(unsigned long dummy
)
637 unsigned long now
= jiffies
;
641 spin_lock(&tw_death_lock
);
642 if (tcp_twcal_hand
< 0)
645 slot
= tcp_twcal_hand
;
646 j
= tcp_twcal_jiffie
;
648 for (n
=0; n
<TCP_TW_RECYCLE_SLOTS
; n
++) {
649 if (time_before_eq(j
, now
)) {
650 struct hlist_node
*node
, *safe
;
651 struct inet_timewait_sock
*tw
;
653 inet_twsk_for_each_inmate_safe(tw
, node
, safe
,
654 &tcp_twcal_row
[slot
]) {
655 __inet_twsk_del_dead_node(tw
);
656 tcp_timewait_kill(tw
);
663 tcp_twcal_jiffie
= j
;
664 tcp_twcal_hand
= slot
;
667 if (!hlist_empty(&tcp_twcal_row
[slot
])) {
668 mod_timer(&tcp_twcal_timer
, j
);
672 j
+= (1<<TCP_TW_RECYCLE_TICK
);
673 slot
= (slot
+1)&(TCP_TW_RECYCLE_SLOTS
-1);
678 if ((tcp_tw_count
-= killed
) == 0)
679 del_timer(&tcp_tw_timer
);
680 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED
, killed
);
681 spin_unlock(&tw_death_lock
);
684 /* This is not only more efficient than what we used to do, it eliminates
685 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
687 * Actually, we could lots of memory writes here. tp of listening
688 * socket contains all necessary default parameters.
690 struct sock
*tcp_create_openreq_child(struct sock
*sk
, struct request_sock
*req
, struct sk_buff
*skb
)
692 /* allocate the newsk from the same slab of the master sock,
693 * if not, at sk_free time we'll try to free it from the wrong
694 * slabcache (i.e. is it TCPv4 or v6?), this is handled thru sk->sk_prot -acme */
695 struct sock
*newsk
= sk_alloc(PF_INET
, GFP_ATOMIC
, sk
->sk_prot
, 0);
698 struct inet_request_sock
*ireq
= inet_rsk(req
);
699 struct tcp_request_sock
*treq
= tcp_rsk(req
);
700 struct inet_sock
*newinet
= inet_sk(newsk
);
701 struct tcp_sock
*newtp
;
702 struct sk_filter
*filter
;
704 memcpy(newsk
, sk
, sizeof(struct tcp_sock
));
705 newsk
->sk_state
= TCP_SYN_RECV
;
708 sk_node_init(&newsk
->sk_node
);
709 newinet
->bind_hash
= NULL
;
711 /* Clone the TCP header template */
712 newinet
->dport
= ireq
->rmt_port
;
714 sock_lock_init(newsk
);
717 rwlock_init(&newsk
->sk_dst_lock
);
718 newsk
->sk_dst_cache
= NULL
;
719 atomic_set(&newsk
->sk_rmem_alloc
, 0);
720 skb_queue_head_init(&newsk
->sk_receive_queue
);
721 atomic_set(&newsk
->sk_wmem_alloc
, 0);
722 skb_queue_head_init(&newsk
->sk_write_queue
);
723 atomic_set(&newsk
->sk_omem_alloc
, 0);
724 newsk
->sk_wmem_queued
= 0;
725 newsk
->sk_forward_alloc
= 0;
727 sock_reset_flag(newsk
, SOCK_DONE
);
728 newsk
->sk_userlocks
= sk
->sk_userlocks
& ~SOCK_BINDPORT_LOCK
;
729 newsk
->sk_backlog
.head
= newsk
->sk_backlog
.tail
= NULL
;
730 newsk
->sk_send_head
= NULL
;
731 rwlock_init(&newsk
->sk_callback_lock
);
732 skb_queue_head_init(&newsk
->sk_error_queue
);
733 newsk
->sk_write_space
= sk_stream_write_space
;
735 if ((filter
= newsk
->sk_filter
) != NULL
)
736 sk_filter_charge(newsk
, filter
);
738 if (unlikely(xfrm_sk_clone_policy(newsk
))) {
739 /* It is still raw copy of parent, so invalidate
740 * destructor and make plain sk_free() */
741 newsk
->sk_destruct
= NULL
;
746 /* Now setup tcp_sock */
747 newtp
= tcp_sk(newsk
);
748 newtp
->pred_flags
= 0;
749 newtp
->rcv_nxt
= treq
->rcv_isn
+ 1;
750 newtp
->snd_nxt
= treq
->snt_isn
+ 1;
751 newtp
->snd_una
= treq
->snt_isn
+ 1;
752 newtp
->snd_sml
= treq
->snt_isn
+ 1;
754 tcp_prequeue_init(newtp
);
756 tcp_init_wl(newtp
, treq
->snt_isn
, treq
->rcv_isn
);
758 newtp
->retransmits
= 0;
761 newtp
->mdev
= TCP_TIMEOUT_INIT
;
762 newtp
->rto
= TCP_TIMEOUT_INIT
;
764 newtp
->packets_out
= 0;
766 newtp
->retrans_out
= 0;
767 newtp
->sacked_out
= 0;
768 newtp
->fackets_out
= 0;
769 newtp
->snd_ssthresh
= 0x7fffffff;
771 /* So many TCP implementations out there (incorrectly) count the
772 * initial SYN frame in their delayed-ACK and congestion control
773 * algorithms that we must have the following bandaid to talk
774 * efficiently to them. -DaveM
777 newtp
->snd_cwnd_cnt
= 0;
779 newtp
->frto_counter
= 0;
780 newtp
->frto_highmark
= 0;
782 newtp
->ca_ops
= &tcp_reno
;
784 tcp_set_ca_state(newtp
, TCP_CA_Open
);
785 tcp_init_xmit_timers(newsk
);
786 skb_queue_head_init(&newtp
->out_of_order_queue
);
787 newtp
->rcv_wup
= treq
->rcv_isn
+ 1;
788 newtp
->write_seq
= treq
->snt_isn
+ 1;
789 newtp
->pushed_seq
= newtp
->write_seq
;
790 newtp
->copied_seq
= treq
->rcv_isn
+ 1;
792 newtp
->rx_opt
.saw_tstamp
= 0;
794 newtp
->rx_opt
.dsack
= 0;
795 newtp
->rx_opt
.eff_sacks
= 0;
797 newtp
->probes_out
= 0;
798 newtp
->rx_opt
.num_sacks
= 0;
800 /* Deinitialize accept_queue to trap illegal accesses. */
801 memset(&newtp
->accept_queue
, 0, sizeof(newtp
->accept_queue
));
803 /* Back to base struct sock members. */
805 newsk
->sk_priority
= 0;
806 atomic_set(&newsk
->sk_refcnt
, 2);
809 * Increment the counter in the same struct proto as the master
810 * sock (sk_refcnt_debug_inc uses newsk->sk_prot->socks, that
811 * is the same as sk->sk_prot->socks, as this field was copied
812 * with memcpy), same rationale as the first comment in this
815 * This _changes_ the previous behaviour, where
816 * tcp_create_openreq_child always was incrementing the
817 * equivalent to tcp_prot->socks (inet_sock_nr), so this have
818 * to be taken into account in all callers. -acme
820 sk_refcnt_debug_inc(newsk
);
822 atomic_inc(&tcp_sockets_allocated
);
824 if (sock_flag(newsk
, SOCK_KEEPOPEN
))
825 tcp_reset_keepalive_timer(newsk
,
826 keepalive_time_when(newtp
));
827 newsk
->sk_socket
= NULL
;
828 newsk
->sk_sleep
= NULL
;
830 newtp
->rx_opt
.tstamp_ok
= ireq
->tstamp_ok
;
831 if((newtp
->rx_opt
.sack_ok
= ireq
->sack_ok
) != 0) {
833 newtp
->rx_opt
.sack_ok
|= 2;
835 newtp
->window_clamp
= req
->window_clamp
;
836 newtp
->rcv_ssthresh
= req
->rcv_wnd
;
837 newtp
->rcv_wnd
= req
->rcv_wnd
;
838 newtp
->rx_opt
.wscale_ok
= ireq
->wscale_ok
;
839 if (newtp
->rx_opt
.wscale_ok
) {
840 newtp
->rx_opt
.snd_wscale
= ireq
->snd_wscale
;
841 newtp
->rx_opt
.rcv_wscale
= ireq
->rcv_wscale
;
843 newtp
->rx_opt
.snd_wscale
= newtp
->rx_opt
.rcv_wscale
= 0;
844 newtp
->window_clamp
= min(newtp
->window_clamp
, 65535U);
846 newtp
->snd_wnd
= ntohs(skb
->h
.th
->window
) << newtp
->rx_opt
.snd_wscale
;
847 newtp
->max_window
= newtp
->snd_wnd
;
849 if (newtp
->rx_opt
.tstamp_ok
) {
850 newtp
->rx_opt
.ts_recent
= req
->ts_recent
;
851 newtp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
852 newtp
->tcp_header_len
= sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
854 newtp
->rx_opt
.ts_recent_stamp
= 0;
855 newtp
->tcp_header_len
= sizeof(struct tcphdr
);
857 if (skb
->len
>= TCP_MIN_RCVMSS
+newtp
->tcp_header_len
)
858 newtp
->ack
.last_seg_size
= skb
->len
-newtp
->tcp_header_len
;
859 newtp
->rx_opt
.mss_clamp
= req
->mss
;
860 TCP_ECN_openreq_child(newtp
, req
);
861 if (newtp
->ecn_flags
&TCP_ECN_OK
)
862 sock_set_flag(newsk
, SOCK_NO_LARGESEND
);
864 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS
);
870 * Process an incoming packet for SYN_RECV sockets represented
874 struct sock
*tcp_check_req(struct sock
*sk
,struct sk_buff
*skb
,
875 struct request_sock
*req
,
876 struct request_sock
**prev
)
878 struct tcphdr
*th
= skb
->h
.th
;
879 struct tcp_sock
*tp
= tcp_sk(sk
);
880 u32 flg
= tcp_flag_word(th
) & (TCP_FLAG_RST
|TCP_FLAG_SYN
|TCP_FLAG_ACK
);
882 struct tcp_options_received tmp_opt
;
885 tmp_opt
.saw_tstamp
= 0;
886 if (th
->doff
> (sizeof(struct tcphdr
)>>2)) {
887 tcp_parse_options(skb
, &tmp_opt
, 0);
889 if (tmp_opt
.saw_tstamp
) {
890 tmp_opt
.ts_recent
= req
->ts_recent
;
891 /* We do not store true stamp, but it is not required,
892 * it can be estimated (approximately)
895 tmp_opt
.ts_recent_stamp
= xtime
.tv_sec
- ((TCP_TIMEOUT_INIT
/HZ
)<<req
->retrans
);
896 paws_reject
= tcp_paws_check(&tmp_opt
, th
->rst
);
900 /* Check for pure retransmitted SYN. */
901 if (TCP_SKB_CB(skb
)->seq
== tcp_rsk(req
)->rcv_isn
&&
902 flg
== TCP_FLAG_SYN
&&
905 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
906 * this case on figure 6 and figure 8, but formal
907 * protocol description says NOTHING.
908 * To be more exact, it says that we should send ACK,
909 * because this segment (at least, if it has no data)
912 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
913 * describe SYN-RECV state. All the description
914 * is wrong, we cannot believe to it and should
915 * rely only on common sense and implementation
918 * Enforce "SYN-ACK" according to figure 8, figure 6
919 * of RFC793, fixed by RFC1122.
921 req
->rsk_ops
->rtx_syn_ack(sk
, req
, NULL
);
925 /* Further reproduces section "SEGMENT ARRIVES"
926 for state SYN-RECEIVED of RFC793.
927 It is broken, however, it does not work only
928 when SYNs are crossed.
930 You would think that SYN crossing is impossible here, since
931 we should have a SYN_SENT socket (from connect()) on our end,
932 but this is not true if the crossed SYNs were sent to both
933 ends by a malicious third party. We must defend against this,
934 and to do that we first verify the ACK (as per RFC793, page
935 36) and reset if it is invalid. Is this a true full defense?
936 To convince ourselves, let us consider a way in which the ACK
937 test can still pass in this 'malicious crossed SYNs' case.
938 Malicious sender sends identical SYNs (and thus identical sequence
939 numbers) to both A and B:
944 By our good fortune, both A and B select the same initial
945 send sequence number of seven :-)
947 A: sends SYN|ACK, seq=7, ack_seq=8
948 B: sends SYN|ACK, seq=7, ack_seq=8
950 So we are now A eating this SYN|ACK, ACK test passes. So
951 does sequence test, SYN is truncated, and thus we consider
954 If tp->defer_accept, we silently drop this bare ACK. Otherwise,
955 we create an established connection. Both ends (listening sockets)
956 accept the new incoming connection and try to talk to each other. 8-)
958 Note: This case is both harmless, and rare. Possibility is about the
959 same as us discovering intelligent life on another plant tomorrow.
961 But generally, we should (RFC lies!) to accept ACK
962 from SYNACK both here and in tcp_rcv_state_process().
963 tcp_rcv_state_process() does not, hence, we do not too.
965 Note that the case is absolutely generic:
966 we cannot optimize anything here without
967 violating protocol. All the checks must be made
968 before attempt to create socket.
971 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
972 * and the incoming segment acknowledges something not yet
973 * sent (the segment carries an unaccaptable ACK) ...
976 * Invalid ACK: reset will be sent by listening socket
978 if ((flg
& TCP_FLAG_ACK
) &&
979 (TCP_SKB_CB(skb
)->ack_seq
!= tcp_rsk(req
)->snt_isn
+ 1))
982 /* Also, it would be not so bad idea to check rcv_tsecr, which
983 * is essentially ACK extension and too early or too late values
984 * should cause reset in unsynchronized states.
987 /* RFC793: "first check sequence number". */
989 if (paws_reject
|| !tcp_in_window(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
,
990 tcp_rsk(req
)->rcv_isn
+ 1, tcp_rsk(req
)->rcv_isn
+ 1 + req
->rcv_wnd
)) {
991 /* Out of window: send ACK and drop. */
992 if (!(flg
& TCP_FLAG_RST
))
993 req
->rsk_ops
->send_ack(skb
, req
);
995 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
999 /* In sequence, PAWS is OK. */
1001 if (tmp_opt
.saw_tstamp
&& !after(TCP_SKB_CB(skb
)->seq
, tcp_rsk(req
)->rcv_isn
+ 1))
1002 req
->ts_recent
= tmp_opt
.rcv_tsval
;
1004 if (TCP_SKB_CB(skb
)->seq
== tcp_rsk(req
)->rcv_isn
) {
1005 /* Truncate SYN, it is out of window starting
1006 at tcp_rsk(req)->rcv_isn + 1. */
1007 flg
&= ~TCP_FLAG_SYN
;
1010 /* RFC793: "second check the RST bit" and
1011 * "fourth, check the SYN bit"
1013 if (flg
& (TCP_FLAG_RST
|TCP_FLAG_SYN
))
1014 goto embryonic_reset
;
1016 /* ACK sequence verified above, just make sure ACK is
1017 * set. If ACK not set, just silently drop the packet.
1019 if (!(flg
& TCP_FLAG_ACK
))
1022 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
1023 if (tp
->defer_accept
&& TCP_SKB_CB(skb
)->end_seq
== tcp_rsk(req
)->rcv_isn
+ 1) {
1024 inet_rsk(req
)->acked
= 1;
1028 /* OK, ACK is valid, create big socket and
1029 * feed this segment to it. It will repeat all
1030 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
1031 * ESTABLISHED STATE. If it will be dropped after
1032 * socket is created, wait for troubles.
1034 child
= tp
->af_specific
->syn_recv_sock(sk
, skb
, req
, NULL
);
1036 goto listen_overflow
;
1038 tcp_synq_unlink(tp
, req
, prev
);
1039 tcp_synq_removed(sk
, req
);
1041 tcp_acceptq_queue(sk
, req
, child
);
1045 if (!sysctl_tcp_abort_on_overflow
) {
1046 inet_rsk(req
)->acked
= 1;
1051 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS
);
1052 if (!(flg
& TCP_FLAG_RST
))
1053 req
->rsk_ops
->send_reset(skb
);
1055 tcp_synq_drop(sk
, req
, prev
);
1060 * Queue segment on the new socket if the new socket is active,
1061 * otherwise we just shortcircuit this and continue with
1065 int tcp_child_process(struct sock
*parent
, struct sock
*child
,
1066 struct sk_buff
*skb
)
1069 int state
= child
->sk_state
;
1071 if (!sock_owned_by_user(child
)) {
1072 ret
= tcp_rcv_state_process(child
, skb
, skb
->h
.th
, skb
->len
);
1074 /* Wakeup parent, send SIGIO */
1075 if (state
== TCP_SYN_RECV
&& child
->sk_state
!= state
)
1076 parent
->sk_data_ready(parent
, 0);
1078 /* Alas, it is possible again, because we do lookup
1079 * in main socket hash table and lock on listening
1080 * socket does not protect us more.
1082 sk_add_backlog(child
, skb
);
1085 bh_unlock_sock(child
);
1090 EXPORT_SYMBOL(tcp_check_req
);
1091 EXPORT_SYMBOL(tcp_child_process
);
1092 EXPORT_SYMBOL(tcp_create_openreq_child
);
1093 EXPORT_SYMBOL(tcp_timewait_state_process
);
1094 EXPORT_SYMBOL(tcp_tw_deschedule
);