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 * * Main purpose of TIME-WAIT state is to close connection gracefully,
61 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
62 * (and, probably, tail of data) and one or more our ACKs are lost.
63 * * What is TIME-WAIT timeout? It is associated with maximal packet
64 * lifetime in the internet, which results in wrong conclusion, that
65 * it is set to catch "old duplicate segments" wandering out of their path.
66 * It is not quite correct. This timeout is calculated so that it exceeds
67 * maximal retransmission timeout enough to allow to lose one (or more)
68 * segments sent by peer and our ACKs. This time may be calculated from RTO.
69 * * When TIME-WAIT socket receives RST, it means that another end
70 * finally closed and we are allowed to kill TIME-WAIT too.
71 * * Second purpose of TIME-WAIT is catching old duplicate segments.
72 * Well, certainly it is pure paranoia, but if we load TIME-WAIT
73 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
74 * * If we invented some more clever way to catch duplicates
75 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
77 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
78 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
79 * from the very beginning.
81 * NOTE. With recycling (and later with fin-wait-2) TW bucket
82 * is _not_ stateless. It means, that strictly speaking we must
83 * spinlock it. I do not want! Well, probability of misbehaviour
84 * is ridiculously low and, seems, we could use some mb() tricks
85 * to avoid misread sequence numbers, states etc. --ANK
88 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
89 const struct tcphdr *th)
91 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
92 struct tcp_options_received tmp_opt;
95 tmp_opt.saw_tstamp = 0;
96 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
97 tcp_parse_options(skb, &tmp_opt, 0);
99 if (tmp_opt.saw_tstamp) {
100 tmp_opt.ts_recent = tcptw->tw_ts_recent;
101 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
102 paws_reject = tcp_paws_check(&tmp_opt, th->rst);
106 if (tw->tw_substate == TCP_FIN_WAIT2) {
107 /* Just repeat all the checks of tcp_rcv_state_process() */
109 /* Out of window, send ACK */
111 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
113 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
119 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
123 if (!after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
124 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
126 return TCP_TW_SUCCESS;
129 /* New data or FIN. If new data arrive after half-duplex close,
133 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
135 tcp_tw_deschedule(tw);
140 /* FIN arrived, enter true time-wait state. */
141 tw->tw_substate = TCP_TIME_WAIT;
142 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
143 if (tmp_opt.saw_tstamp) {
144 tcptw->tw_ts_recent_stamp = xtime.tv_sec;
145 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
148 /* I am shamed, but failed to make it more elegant.
149 * Yes, it is direct reference to IP, which is impossible
150 * to generalize to IPv6. Taking into account that IPv6
151 * do not undertsnad recycling in any case, it not
152 * a big problem in practice. --ANK */
153 if (tw->tw_family == AF_INET &&
154 sysctl_tcp_tw_recycle && tcptw->tw_ts_recent_stamp &&
155 tcp_v4_tw_remember_stamp(tw))
156 tcp_tw_schedule(tw, tw->tw_timeout);
158 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
163 * Now real TIME-WAIT state.
166 * "When a connection is [...] on TIME-WAIT state [...]
167 * [a TCP] MAY accept a new SYN from the remote TCP to
168 * reopen the connection directly, if it:
170 * (1) assigns its initial sequence number for the new
171 * connection to be larger than the largest sequence
172 * number it used on the previous connection incarnation,
175 * (2) returns to TIME-WAIT state if the SYN turns out
176 * to be an old duplicate".
180 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
181 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
182 /* In window segment, it may be only reset or bare ack. */
185 /* This is TIME_WAIT assasination, in two flavors.
186 * Oh well... nobody has a sufficient solution to this
189 if (sysctl_tcp_rfc1337 == 0) {
191 tcp_tw_deschedule(tw);
193 return TCP_TW_SUCCESS;
196 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
198 if (tmp_opt.saw_tstamp) {
199 tcptw->tw_ts_recent = tmp_opt.rcv_tsval;
200 tcptw->tw_ts_recent_stamp = xtime.tv_sec;
204 return TCP_TW_SUCCESS;
207 /* Out of window segment.
209 All the segments are ACKed immediately.
211 The only exception is new SYN. We accept it, if it is
212 not old duplicate and we are not in danger to be killed
213 by delayed old duplicates. RFC check is that it has
214 newer sequence number works at rates <40Mbit/sec.
215 However, if paws works, it is reliable AND even more,
216 we even may relax silly seq space cutoff.
218 RED-PEN: we violate main RFC requirement, if this SYN will appear
219 old duplicate (i.e. we receive RST in reply to SYN-ACK),
220 we must return socket to time-wait state. It is not good,
224 if (th->syn && !th->rst && !th->ack && !paws_reject &&
225 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
226 (tmp_opt.saw_tstamp &&
227 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
228 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
231 TCP_SKB_CB(skb)->when = isn;
236 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
239 /* In this case we must reset the TIMEWAIT timer.
241 * If it is ACKless SYN it may be both old duplicate
242 * and new good SYN with random sequence number <rcv_nxt.
243 * Do not reschedule in the last case.
245 if (paws_reject || th->ack)
246 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
248 /* Send ACK. Note, we do not put the bucket,
249 * it will be released by caller.
254 return TCP_TW_SUCCESS;
258 * Move a socket to time-wait or dead fin-wait-2 state.
260 void tcp_time_wait(struct sock *sk, int state, int timeo)
262 struct inet_timewait_sock *tw = NULL;
263 const struct tcp_sock *tp = tcp_sk(sk);
266 if (sysctl_tcp_tw_recycle && tp->rx_opt.ts_recent_stamp)
267 recycle_ok = tp->af_specific->remember_stamp(sk);
269 if (tcp_tw_count < sysctl_tcp_max_tw_buckets)
270 tw = inet_twsk_alloc(sk, state);
273 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
274 const int rto = (tp->rto << 2) - (tp->rto >> 1);
276 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
277 tcptw->tw_rcv_nxt = tp->rcv_nxt;
278 tcptw->tw_snd_nxt = tp->snd_nxt;
279 tcptw->tw_rcv_wnd = tcp_receive_window(tp);
280 tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
281 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
283 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
284 if (tw->tw_family == PF_INET6) {
285 struct ipv6_pinfo *np = inet6_sk(sk);
286 struct tcp6_timewait_sock *tcp6tw = tcp6_twsk((struct sock *)tw);
288 ipv6_addr_copy(&tcp6tw->tw_v6_daddr, &np->daddr);
289 ipv6_addr_copy(&tcp6tw->tw_v6_rcv_saddr, &np->rcv_saddr);
290 tw->tw_ipv6only = np->ipv6only;
293 /* Linkage updates. */
294 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
296 /* Get the TIME_WAIT timeout firing. */
301 tw->tw_timeout = rto;
303 tw->tw_timeout = TCP_TIMEWAIT_LEN;
304 if (state == TCP_TIME_WAIT)
305 timeo = TCP_TIMEWAIT_LEN;
308 tcp_tw_schedule(tw, timeo);
311 /* Sorry, if we're out of memory, just CLOSE this
312 * socket up. We've got bigger problems than
313 * non-graceful socket closings.
316 printk(KERN_INFO "TCP: time wait bucket table overflow\n");
319 tcp_update_metrics(sk);
323 /* Kill off TIME_WAIT sockets once their lifetime has expired. */
324 static int tcp_tw_death_row_slot;
326 static void tcp_twkill(unsigned long);
328 /* TIME_WAIT reaping mechanism. */
329 #define TCP_TWKILL_SLOTS 8 /* Please keep this a power of 2. */
330 #define TCP_TWKILL_PERIOD (TCP_TIMEWAIT_LEN/TCP_TWKILL_SLOTS)
332 #define TCP_TWKILL_QUOTA 100
334 static struct hlist_head tcp_tw_death_row[TCP_TWKILL_SLOTS];
335 static DEFINE_SPINLOCK(tw_death_lock);
336 static struct timer_list tcp_tw_timer = TIMER_INITIALIZER(tcp_twkill, 0, 0);
337 static void twkill_work(void *);
338 static DECLARE_WORK(tcp_twkill_work, twkill_work, NULL);
339 static u32 twkill_thread_slots;
341 /* Returns non-zero if quota exceeded. */
342 static int tcp_do_twkill_work(int slot, unsigned int quota)
344 struct inet_timewait_sock *tw;
345 struct hlist_node *node;
349 /* NOTE: compare this to previous version where lock
350 * was released after detaching chain. It was racy,
351 * because tw buckets are scheduled in not serialized context
352 * in 2.3 (with netfilter), and with softnet it is common, because
353 * soft irqs are not sequenced.
358 inet_twsk_for_each_inmate(tw, node, &tcp_tw_death_row[slot]) {
359 __inet_twsk_del_dead_node(tw);
360 spin_unlock(&tw_death_lock);
361 __inet_twsk_kill(tw, &tcp_hashinfo);
364 spin_lock(&tw_death_lock);
365 if (killed > quota) {
370 /* While we dropped tw_death_lock, another cpu may have
371 * killed off the next TW bucket in the list, therefore
372 * do a fresh re-read of the hlist head node with the
373 * lock reacquired. We still use the hlist traversal
374 * macro in order to get the prefetches.
379 tcp_tw_count -= killed;
380 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
385 static void tcp_twkill(unsigned long dummy)
389 spin_lock(&tw_death_lock);
391 if (tcp_tw_count == 0)
395 ret = tcp_do_twkill_work(tcp_tw_death_row_slot, TCP_TWKILL_QUOTA);
397 twkill_thread_slots |= (1 << tcp_tw_death_row_slot);
399 schedule_work(&tcp_twkill_work);
402 /* We purged the entire slot, anything left? */
406 tcp_tw_death_row_slot =
407 ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
409 mod_timer(&tcp_tw_timer, jiffies + TCP_TWKILL_PERIOD);
411 spin_unlock(&tw_death_lock);
414 extern void twkill_slots_invalid(void);
416 static void twkill_work(void *dummy)
420 if ((TCP_TWKILL_SLOTS - 1) > (sizeof(twkill_thread_slots) * 8))
421 twkill_slots_invalid();
423 while (twkill_thread_slots) {
424 spin_lock_bh(&tw_death_lock);
425 for (i = 0; i < TCP_TWKILL_SLOTS; i++) {
426 if (!(twkill_thread_slots & (1 << i)))
429 while (tcp_do_twkill_work(i, TCP_TWKILL_QUOTA) != 0) {
430 if (need_resched()) {
431 spin_unlock_bh(&tw_death_lock);
433 spin_lock_bh(&tw_death_lock);
437 twkill_thread_slots &= ~(1 << i);
439 spin_unlock_bh(&tw_death_lock);
443 /* These are always called from BH context. See callers in
444 * tcp_input.c to verify this.
447 /* This is for handling early-kills of TIME_WAIT sockets. */
448 void tcp_tw_deschedule(struct inet_timewait_sock *tw)
450 spin_lock(&tw_death_lock);
451 if (inet_twsk_del_dead_node(tw)) {
453 if (--tcp_tw_count == 0)
454 del_timer(&tcp_tw_timer);
456 spin_unlock(&tw_death_lock);
457 __inet_twsk_kill(tw, &tcp_hashinfo);
460 /* Short-time timewait calendar */
462 static int tcp_twcal_hand = -1;
463 static int tcp_twcal_jiffie;
464 static void tcp_twcal_tick(unsigned long);
465 static struct timer_list tcp_twcal_timer =
466 TIMER_INITIALIZER(tcp_twcal_tick, 0, 0);
467 static struct hlist_head tcp_twcal_row[TCP_TW_RECYCLE_SLOTS];
469 static void tcp_tw_schedule(struct inet_timewait_sock *tw, const int timeo)
471 struct hlist_head *list;
474 /* timeout := RTO * 3.5
476 * 3.5 = 1+2+0.5 to wait for two retransmits.
478 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
479 * our ACK acking that FIN can be lost. If N subsequent retransmitted
480 * FINs (or previous seqments) are lost (probability of such event
481 * is p^(N+1), where p is probability to lose single packet and
482 * time to detect the loss is about RTO*(2^N - 1) with exponential
483 * backoff). Normal timewait length is calculated so, that we
484 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
485 * [ BTW Linux. following BSD, violates this requirement waiting
486 * only for 60sec, we should wait at least for 240 secs.
487 * Well, 240 consumes too much of resources 8)
489 * This interval is not reduced to catch old duplicate and
490 * responces to our wandering segments living for two MSLs.
491 * However, if we use PAWS to detect
492 * old duplicates, we can reduce the interval to bounds required
493 * by RTO, rather than MSL. So, if peer understands PAWS, we
494 * kill tw bucket after 3.5*RTO (it is important that this number
495 * is greater than TS tick!) and detect old duplicates with help
498 slot = (timeo + (1<<TCP_TW_RECYCLE_TICK) - 1) >> TCP_TW_RECYCLE_TICK;
500 spin_lock(&tw_death_lock);
502 /* Unlink it, if it was scheduled */
503 if (inet_twsk_del_dead_node(tw))
506 atomic_inc(&tw->tw_refcnt);
508 if (slot >= TCP_TW_RECYCLE_SLOTS) {
509 /* Schedule to slow timer */
510 if (timeo >= TCP_TIMEWAIT_LEN) {
511 slot = TCP_TWKILL_SLOTS-1;
513 slot = (timeo + TCP_TWKILL_PERIOD-1) / TCP_TWKILL_PERIOD;
514 if (slot >= TCP_TWKILL_SLOTS)
515 slot = TCP_TWKILL_SLOTS-1;
517 tw->tw_ttd = jiffies + timeo;
518 slot = (tcp_tw_death_row_slot + slot) & (TCP_TWKILL_SLOTS - 1);
519 list = &tcp_tw_death_row[slot];
521 tw->tw_ttd = jiffies + (slot << TCP_TW_RECYCLE_TICK);
523 if (tcp_twcal_hand < 0) {
525 tcp_twcal_jiffie = jiffies;
526 tcp_twcal_timer.expires = tcp_twcal_jiffie + (slot<<TCP_TW_RECYCLE_TICK);
527 add_timer(&tcp_twcal_timer);
529 if (time_after(tcp_twcal_timer.expires, jiffies + (slot<<TCP_TW_RECYCLE_TICK)))
530 mod_timer(&tcp_twcal_timer, jiffies + (slot<<TCP_TW_RECYCLE_TICK));
531 slot = (tcp_twcal_hand + slot)&(TCP_TW_RECYCLE_SLOTS-1);
533 list = &tcp_twcal_row[slot];
536 hlist_add_head(&tw->tw_death_node, list);
538 if (tcp_tw_count++ == 0)
539 mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
540 spin_unlock(&tw_death_lock);
543 void tcp_twcal_tick(unsigned long dummy)
547 unsigned long now = jiffies;
551 spin_lock(&tw_death_lock);
552 if (tcp_twcal_hand < 0)
555 slot = tcp_twcal_hand;
556 j = tcp_twcal_jiffie;
558 for (n=0; n<TCP_TW_RECYCLE_SLOTS; n++) {
559 if (time_before_eq(j, now)) {
560 struct hlist_node *node, *safe;
561 struct inet_timewait_sock *tw;
563 inet_twsk_for_each_inmate_safe(tw, node, safe,
564 &tcp_twcal_row[slot]) {
565 __inet_twsk_del_dead_node(tw);
566 __inet_twsk_kill(tw, &tcp_hashinfo);
573 tcp_twcal_jiffie = j;
574 tcp_twcal_hand = slot;
577 if (!hlist_empty(&tcp_twcal_row[slot])) {
578 mod_timer(&tcp_twcal_timer, j);
582 j += (1<<TCP_TW_RECYCLE_TICK);
583 slot = (slot+1)&(TCP_TW_RECYCLE_SLOTS-1);
588 if ((tcp_tw_count -= killed) == 0)
589 del_timer(&tcp_tw_timer);
590 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
591 spin_unlock(&tw_death_lock);
594 /* This is not only more efficient than what we used to do, it eliminates
595 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
597 * Actually, we could lots of memory writes here. tp of listening
598 * socket contains all necessary default parameters.
600 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
602 /* allocate the newsk from the same slab of the master sock,
603 * if not, at sk_free time we'll try to free it from the wrong
604 * slabcache (i.e. is it TCPv4 or v6?), this is handled thru sk->sk_prot -acme */
605 struct sock *newsk = sk_alloc(PF_INET, GFP_ATOMIC, sk->sk_prot, 0);
608 struct inet_request_sock *ireq = inet_rsk(req);
609 struct tcp_request_sock *treq = tcp_rsk(req);
610 struct inet_sock *newinet = inet_sk(newsk);
611 struct tcp_sock *newtp;
612 struct sk_filter *filter;
614 memcpy(newsk, sk, sizeof(struct tcp_sock));
615 newsk->sk_state = TCP_SYN_RECV;
618 sk_node_init(&newsk->sk_node);
619 newinet->bind_hash = NULL;
621 /* Clone the TCP header template */
622 newinet->dport = ireq->rmt_port;
624 sock_lock_init(newsk);
627 rwlock_init(&newsk->sk_dst_lock);
628 newsk->sk_dst_cache = NULL;
629 atomic_set(&newsk->sk_rmem_alloc, 0);
630 skb_queue_head_init(&newsk->sk_receive_queue);
631 atomic_set(&newsk->sk_wmem_alloc, 0);
632 skb_queue_head_init(&newsk->sk_write_queue);
633 atomic_set(&newsk->sk_omem_alloc, 0);
634 newsk->sk_wmem_queued = 0;
635 newsk->sk_forward_alloc = 0;
637 sock_reset_flag(newsk, SOCK_DONE);
638 newsk->sk_userlocks = sk->sk_userlocks & ~SOCK_BINDPORT_LOCK;
639 newsk->sk_backlog.head = newsk->sk_backlog.tail = NULL;
640 newsk->sk_send_head = NULL;
641 rwlock_init(&newsk->sk_callback_lock);
642 skb_queue_head_init(&newsk->sk_error_queue);
643 newsk->sk_write_space = sk_stream_write_space;
645 if ((filter = newsk->sk_filter) != NULL)
646 sk_filter_charge(newsk, filter);
648 if (unlikely(xfrm_sk_clone_policy(newsk))) {
649 /* It is still raw copy of parent, so invalidate
650 * destructor and make plain sk_free() */
651 newsk->sk_destruct = NULL;
656 /* Now setup tcp_sock */
657 newtp = tcp_sk(newsk);
658 newtp->pred_flags = 0;
659 newtp->rcv_nxt = treq->rcv_isn + 1;
660 newtp->snd_nxt = treq->snt_isn + 1;
661 newtp->snd_una = treq->snt_isn + 1;
662 newtp->snd_sml = treq->snt_isn + 1;
664 tcp_prequeue_init(newtp);
666 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
668 newtp->retransmits = 0;
671 newtp->mdev = TCP_TIMEOUT_INIT;
672 newtp->rto = TCP_TIMEOUT_INIT;
674 newtp->packets_out = 0;
676 newtp->retrans_out = 0;
677 newtp->sacked_out = 0;
678 newtp->fackets_out = 0;
679 newtp->snd_ssthresh = 0x7fffffff;
681 /* So many TCP implementations out there (incorrectly) count the
682 * initial SYN frame in their delayed-ACK and congestion control
683 * algorithms that we must have the following bandaid to talk
684 * efficiently to them. -DaveM
687 newtp->snd_cwnd_cnt = 0;
689 newtp->frto_counter = 0;
690 newtp->frto_highmark = 0;
692 newtp->ca_ops = &tcp_reno;
694 tcp_set_ca_state(newtp, TCP_CA_Open);
695 tcp_init_xmit_timers(newsk);
696 skb_queue_head_init(&newtp->out_of_order_queue);
697 newtp->rcv_wup = treq->rcv_isn + 1;
698 newtp->write_seq = treq->snt_isn + 1;
699 newtp->pushed_seq = newtp->write_seq;
700 newtp->copied_seq = treq->rcv_isn + 1;
702 newtp->rx_opt.saw_tstamp = 0;
704 newtp->rx_opt.dsack = 0;
705 newtp->rx_opt.eff_sacks = 0;
707 newtp->probes_out = 0;
708 newtp->rx_opt.num_sacks = 0;
710 /* Deinitialize accept_queue to trap illegal accesses. */
711 memset(&newtp->accept_queue, 0, sizeof(newtp->accept_queue));
713 /* Back to base struct sock members. */
715 newsk->sk_priority = 0;
716 atomic_set(&newsk->sk_refcnt, 2);
719 * Increment the counter in the same struct proto as the master
720 * sock (sk_refcnt_debug_inc uses newsk->sk_prot->socks, that
721 * is the same as sk->sk_prot->socks, as this field was copied
722 * with memcpy), same rationale as the first comment in this
725 * This _changes_ the previous behaviour, where
726 * tcp_create_openreq_child always was incrementing the
727 * equivalent to tcp_prot->socks (inet_sock_nr), so this have
728 * to be taken into account in all callers. -acme
730 sk_refcnt_debug_inc(newsk);
732 atomic_inc(&tcp_sockets_allocated);
734 if (sock_flag(newsk, SOCK_KEEPOPEN))
735 tcp_reset_keepalive_timer(newsk,
736 keepalive_time_when(newtp));
737 newsk->sk_socket = NULL;
738 newsk->sk_sleep = NULL;
740 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
741 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
743 newtp->rx_opt.sack_ok |= 2;
745 newtp->window_clamp = req->window_clamp;
746 newtp->rcv_ssthresh = req->rcv_wnd;
747 newtp->rcv_wnd = req->rcv_wnd;
748 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
749 if (newtp->rx_opt.wscale_ok) {
750 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
751 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
753 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
754 newtp->window_clamp = min(newtp->window_clamp, 65535U);
756 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
757 newtp->max_window = newtp->snd_wnd;
759 if (newtp->rx_opt.tstamp_ok) {
760 newtp->rx_opt.ts_recent = req->ts_recent;
761 newtp->rx_opt.ts_recent_stamp = xtime.tv_sec;
762 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
764 newtp->rx_opt.ts_recent_stamp = 0;
765 newtp->tcp_header_len = sizeof(struct tcphdr);
767 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
768 newtp->ack.last_seg_size = skb->len-newtp->tcp_header_len;
769 newtp->rx_opt.mss_clamp = req->mss;
770 TCP_ECN_openreq_child(newtp, req);
771 if (newtp->ecn_flags&TCP_ECN_OK)
772 sock_set_flag(newsk, SOCK_NO_LARGESEND);
774 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
780 * Process an incoming packet for SYN_RECV sockets represented
784 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
785 struct request_sock *req,
786 struct request_sock **prev)
788 struct tcphdr *th = skb->h.th;
789 struct tcp_sock *tp = tcp_sk(sk);
790 u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
792 struct tcp_options_received tmp_opt;
795 tmp_opt.saw_tstamp = 0;
796 if (th->doff > (sizeof(struct tcphdr)>>2)) {
797 tcp_parse_options(skb, &tmp_opt, 0);
799 if (tmp_opt.saw_tstamp) {
800 tmp_opt.ts_recent = req->ts_recent;
801 /* We do not store true stamp, but it is not required,
802 * it can be estimated (approximately)
805 tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
806 paws_reject = tcp_paws_check(&tmp_opt, th->rst);
810 /* Check for pure retransmitted SYN. */
811 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
812 flg == TCP_FLAG_SYN &&
815 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
816 * this case on figure 6 and figure 8, but formal
817 * protocol description says NOTHING.
818 * To be more exact, it says that we should send ACK,
819 * because this segment (at least, if it has no data)
822 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
823 * describe SYN-RECV state. All the description
824 * is wrong, we cannot believe to it and should
825 * rely only on common sense and implementation
828 * Enforce "SYN-ACK" according to figure 8, figure 6
829 * of RFC793, fixed by RFC1122.
831 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
835 /* Further reproduces section "SEGMENT ARRIVES"
836 for state SYN-RECEIVED of RFC793.
837 It is broken, however, it does not work only
838 when SYNs are crossed.
840 You would think that SYN crossing is impossible here, since
841 we should have a SYN_SENT socket (from connect()) on our end,
842 but this is not true if the crossed SYNs were sent to both
843 ends by a malicious third party. We must defend against this,
844 and to do that we first verify the ACK (as per RFC793, page
845 36) and reset if it is invalid. Is this a true full defense?
846 To convince ourselves, let us consider a way in which the ACK
847 test can still pass in this 'malicious crossed SYNs' case.
848 Malicious sender sends identical SYNs (and thus identical sequence
849 numbers) to both A and B:
854 By our good fortune, both A and B select the same initial
855 send sequence number of seven :-)
857 A: sends SYN|ACK, seq=7, ack_seq=8
858 B: sends SYN|ACK, seq=7, ack_seq=8
860 So we are now A eating this SYN|ACK, ACK test passes. So
861 does sequence test, SYN is truncated, and thus we consider
864 If tp->defer_accept, we silently drop this bare ACK. Otherwise,
865 we create an established connection. Both ends (listening sockets)
866 accept the new incoming connection and try to talk to each other. 8-)
868 Note: This case is both harmless, and rare. Possibility is about the
869 same as us discovering intelligent life on another plant tomorrow.
871 But generally, we should (RFC lies!) to accept ACK
872 from SYNACK both here and in tcp_rcv_state_process().
873 tcp_rcv_state_process() does not, hence, we do not too.
875 Note that the case is absolutely generic:
876 we cannot optimize anything here without
877 violating protocol. All the checks must be made
878 before attempt to create socket.
881 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
882 * and the incoming segment acknowledges something not yet
883 * sent (the segment carries an unaccaptable ACK) ...
886 * Invalid ACK: reset will be sent by listening socket
888 if ((flg & TCP_FLAG_ACK) &&
889 (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
892 /* Also, it would be not so bad idea to check rcv_tsecr, which
893 * is essentially ACK extension and too early or too late values
894 * should cause reset in unsynchronized states.
897 /* RFC793: "first check sequence number". */
899 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
900 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
901 /* Out of window: send ACK and drop. */
902 if (!(flg & TCP_FLAG_RST))
903 req->rsk_ops->send_ack(skb, req);
905 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
909 /* In sequence, PAWS is OK. */
911 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
912 req->ts_recent = tmp_opt.rcv_tsval;
914 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
915 /* Truncate SYN, it is out of window starting
916 at tcp_rsk(req)->rcv_isn + 1. */
917 flg &= ~TCP_FLAG_SYN;
920 /* RFC793: "second check the RST bit" and
921 * "fourth, check the SYN bit"
923 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN))
924 goto embryonic_reset;
926 /* ACK sequence verified above, just make sure ACK is
927 * set. If ACK not set, just silently drop the packet.
929 if (!(flg & TCP_FLAG_ACK))
932 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
933 if (tp->defer_accept && TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
934 inet_rsk(req)->acked = 1;
938 /* OK, ACK is valid, create big socket and
939 * feed this segment to it. It will repeat all
940 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
941 * ESTABLISHED STATE. If it will be dropped after
942 * socket is created, wait for troubles.
944 child = tp->af_specific->syn_recv_sock(sk, skb, req, NULL);
946 goto listen_overflow;
948 tcp_synq_unlink(tp, req, prev);
949 tcp_synq_removed(sk, req);
951 tcp_acceptq_queue(sk, req, child);
955 if (!sysctl_tcp_abort_on_overflow) {
956 inet_rsk(req)->acked = 1;
961 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
962 if (!(flg & TCP_FLAG_RST))
963 req->rsk_ops->send_reset(skb);
965 tcp_synq_drop(sk, req, prev);
970 * Queue segment on the new socket if the new socket is active,
971 * otherwise we just shortcircuit this and continue with
975 int tcp_child_process(struct sock *parent, struct sock *child,
979 int state = child->sk_state;
981 if (!sock_owned_by_user(child)) {
982 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
984 /* Wakeup parent, send SIGIO */
985 if (state == TCP_SYN_RECV && child->sk_state != state)
986 parent->sk_data_ready(parent, 0);
988 /* Alas, it is possible again, because we do lookup
989 * in main socket hash table and lock on listening
990 * socket does not protect us more.
992 sk_add_backlog(child, skb);
995 bh_unlock_sock(child);
1000 EXPORT_SYMBOL(tcp_check_req);
1001 EXPORT_SYMBOL(tcp_child_process);
1002 EXPORT_SYMBOL(tcp_create_openreq_child);
1003 EXPORT_SYMBOL(tcp_timewait_state_process);
1004 EXPORT_SYMBOL(tcp_tw_deschedule);