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 tcp_tw_bucket *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 tcp_tw_bucket *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 "tw_bucket %p refcnt=%d\n", tw,
89 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 tcp_tw_bucket *tw, struct sk_buff *skb,
125 struct tcphdr *th, unsigned len)
127 struct tcp_options_received tmp_opt;
130 tmp_opt.saw_tstamp = 0;
131 if (th->doff > (sizeof(struct tcphdr) >> 2) && tw->tw_ts_recent_stamp) {
132 tcp_parse_options(skb, &tmp_opt, 0);
134 if (tmp_opt.saw_tstamp) {
135 tmp_opt.ts_recent = tw->tw_ts_recent;
136 tmp_opt.ts_recent_stamp = tw->tw_ts_recent_stamp;
137 paws_reject = tcp_paws_check(&tmp_opt, th->rst);
141 if (tw->tw_substate == TCP_FIN_WAIT2) {
142 /* Just repeat all the checks of tcp_rcv_state_process() */
144 /* Out of window, send ACK */
146 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
148 tw->tw_rcv_nxt + tw->tw_rcv_wnd))
154 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt))
158 if (!after(TCP_SKB_CB(skb)->end_seq, tw->tw_rcv_nxt) ||
159 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
161 return TCP_TW_SUCCESS;
164 /* New data or FIN. If new data arrive after half-duplex close,
168 TCP_SKB_CB(skb)->end_seq != tw->tw_rcv_nxt + 1) {
170 tcp_tw_deschedule(tw);
175 /* FIN arrived, enter true time-wait state. */
176 tw->tw_substate = TCP_TIME_WAIT;
177 tw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
178 if (tmp_opt.saw_tstamp) {
179 tw->tw_ts_recent_stamp = xtime.tv_sec;
180 tw->tw_ts_recent = tmp_opt.rcv_tsval;
183 /* I am shamed, but failed to make it more elegant.
184 * Yes, it is direct reference to IP, which is impossible
185 * to generalize to IPv6. Taking into account that IPv6
186 * do not undertsnad recycling in any case, it not
187 * a big problem in practice. --ANK */
188 if (tw->tw_family == AF_INET &&
189 sysctl_tcp_tw_recycle && tw->tw_ts_recent_stamp &&
190 tcp_v4_tw_remember_stamp(tw))
191 tcp_tw_schedule(tw, tw->tw_timeout);
193 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
198 * Now real TIME-WAIT state.
201 * "When a connection is [...] on TIME-WAIT state [...]
202 * [a TCP] MAY accept a new SYN from the remote TCP to
203 * reopen the connection directly, if it:
205 * (1) assigns its initial sequence number for the new
206 * connection to be larger than the largest sequence
207 * number it used on the previous connection incarnation,
210 * (2) returns to TIME-WAIT state if the SYN turns out
211 * to be an old duplicate".
215 (TCP_SKB_CB(skb)->seq == tw->tw_rcv_nxt &&
216 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
217 /* In window segment, it may be only reset or bare ack. */
220 /* This is TIME_WAIT assasination, in two flavors.
221 * Oh well... nobody has a sufficient solution to this
224 if (sysctl_tcp_rfc1337 == 0) {
226 tcp_tw_deschedule(tw);
228 return TCP_TW_SUCCESS;
231 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
233 if (tmp_opt.saw_tstamp) {
234 tw->tw_ts_recent = tmp_opt.rcv_tsval;
235 tw->tw_ts_recent_stamp = xtime.tv_sec;
239 return TCP_TW_SUCCESS;
242 /* Out of window segment.
244 All the segments are ACKed immediately.
246 The only exception is new SYN. We accept it, if it is
247 not old duplicate and we are not in danger to be killed
248 by delayed old duplicates. RFC check is that it has
249 newer sequence number works at rates <40Mbit/sec.
250 However, if paws works, it is reliable AND even more,
251 we even may relax silly seq space cutoff.
253 RED-PEN: we violate main RFC requirement, if this SYN will appear
254 old duplicate (i.e. we receive RST in reply to SYN-ACK),
255 we must return socket to time-wait state. It is not good,
259 if (th->syn && !th->rst && !th->ack && !paws_reject &&
260 (after(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt) ||
261 (tmp_opt.saw_tstamp && (s32)(tw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
262 u32 isn = tw->tw_snd_nxt + 65535 + 2;
265 TCP_SKB_CB(skb)->when = isn;
270 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
273 /* In this case we must reset the TIMEWAIT timer.
275 * If it is ACKless SYN it may be both old duplicate
276 * and new good SYN with random sequence number <rcv_nxt.
277 * Do not reschedule in the last case.
279 if (paws_reject || th->ack)
280 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
282 /* Send ACK. Note, we do not put the bucket,
283 * it will be released by caller.
288 return TCP_TW_SUCCESS;
291 /* Enter the time wait state. This is called with locally disabled BH.
292 * Essentially we whip up a timewait bucket, copy the
293 * relevant info into it from the SK, and mess with hash chains
296 static void __tcp_tw_hashdance(struct sock *sk, struct tcp_tw_bucket *tw)
298 const struct inet_sock *inet = inet_sk(sk);
299 struct inet_ehash_bucket *ehead = &tcp_hashinfo.ehash[sk->sk_hashent];
300 struct inet_bind_hashbucket *bhead;
301 /* Step 1: Put TW into bind hash. Original socket stays there too.
302 Note, that any socket with inet->num != 0 MUST be bound in
303 binding cache, even if it is closed.
305 bhead = &tcp_hashinfo.bhash[inet_bhashfn(inet->num, tcp_hashinfo.bhash_size)];
306 spin_lock(&bhead->lock);
307 tw->tw_tb = inet->bind_hash;
308 BUG_TRAP(inet->bind_hash);
309 tw_add_bind_node(tw, &tw->tw_tb->owners);
310 spin_unlock(&bhead->lock);
312 write_lock(&ehead->lock);
314 /* Step 2: Remove SK from established hash. */
315 if (__sk_del_node_init(sk))
316 sock_prot_dec_use(sk->sk_prot);
318 /* Step 3: Hash TW into TIMEWAIT half of established hash table. */
319 tw_add_node(tw, &(ehead + tcp_hashinfo.ehash_size)->chain);
320 atomic_inc(&tw->tw_refcnt);
322 write_unlock(&ehead->lock);
326 * Move a socket to time-wait or dead fin-wait-2 state.
328 void tcp_time_wait(struct sock *sk, int state, int timeo)
330 struct tcp_tw_bucket *tw = NULL;
331 struct tcp_sock *tp = tcp_sk(sk);
334 if (sysctl_tcp_tw_recycle && tp->rx_opt.ts_recent_stamp)
335 recycle_ok = tp->af_specific->remember_stamp(sk);
337 if (tcp_tw_count < sysctl_tcp_max_tw_buckets)
338 tw = kmem_cache_alloc(tcp_timewait_cachep, SLAB_ATOMIC);
341 struct inet_sock *inet = inet_sk(sk);
342 int rto = (tp->rto<<2) - (tp->rto>>1);
344 /* Give us an identity. */
345 tw->tw_daddr = inet->daddr;
346 tw->tw_rcv_saddr = inet->rcv_saddr;
347 tw->tw_bound_dev_if = sk->sk_bound_dev_if;
348 tw->tw_num = inet->num;
349 tw->tw_state = TCP_TIME_WAIT;
350 tw->tw_substate = state;
351 tw->tw_sport = inet->sport;
352 tw->tw_dport = inet->dport;
353 tw->tw_family = sk->sk_family;
354 tw->tw_reuse = sk->sk_reuse;
355 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale;
356 atomic_set(&tw->tw_refcnt, 1);
358 tw->tw_hashent = sk->sk_hashent;
359 tw->tw_rcv_nxt = tp->rcv_nxt;
360 tw->tw_snd_nxt = tp->snd_nxt;
361 tw->tw_rcv_wnd = tcp_receive_window(tp);
362 tw->tw_ts_recent = tp->rx_opt.ts_recent;
363 tw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
364 tw_dead_node_init(tw);
366 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
367 if (tw->tw_family == PF_INET6) {
368 struct ipv6_pinfo *np = inet6_sk(sk);
370 ipv6_addr_copy(&tw->tw_v6_daddr, &np->daddr);
371 ipv6_addr_copy(&tw->tw_v6_rcv_saddr, &np->rcv_saddr);
372 tw->tw_v6_ipv6only = np->ipv6only;
374 memset(&tw->tw_v6_daddr, 0, sizeof(tw->tw_v6_daddr));
375 memset(&tw->tw_v6_rcv_saddr, 0, sizeof(tw->tw_v6_rcv_saddr));
376 tw->tw_v6_ipv6only = 0;
379 /* Linkage updates. */
380 __tcp_tw_hashdance(sk, tw);
382 /* Get the TIME_WAIT timeout firing. */
387 tw->tw_timeout = rto;
389 tw->tw_timeout = TCP_TIMEWAIT_LEN;
390 if (state == TCP_TIME_WAIT)
391 timeo = TCP_TIMEWAIT_LEN;
394 tcp_tw_schedule(tw, timeo);
397 /* Sorry, if we're out of memory, just CLOSE this
398 * socket up. We've got bigger problems than
399 * non-graceful socket closings.
402 printk(KERN_INFO "TCP: time wait bucket table overflow\n");
405 tcp_update_metrics(sk);
409 /* Kill off TIME_WAIT sockets once their lifetime has expired. */
410 static int tcp_tw_death_row_slot;
412 static void tcp_twkill(unsigned long);
414 /* TIME_WAIT reaping mechanism. */
415 #define TCP_TWKILL_SLOTS 8 /* Please keep this a power of 2. */
416 #define TCP_TWKILL_PERIOD (TCP_TIMEWAIT_LEN/TCP_TWKILL_SLOTS)
418 #define TCP_TWKILL_QUOTA 100
420 static struct hlist_head tcp_tw_death_row[TCP_TWKILL_SLOTS];
421 static DEFINE_SPINLOCK(tw_death_lock);
422 static struct timer_list tcp_tw_timer = TIMER_INITIALIZER(tcp_twkill, 0, 0);
423 static void twkill_work(void *);
424 static DECLARE_WORK(tcp_twkill_work, twkill_work, NULL);
425 static u32 twkill_thread_slots;
427 /* Returns non-zero if quota exceeded. */
428 static int tcp_do_twkill_work(int slot, unsigned int quota)
430 struct tcp_tw_bucket *tw;
431 struct hlist_node *node;
435 /* NOTE: compare this to previous version where lock
436 * was released after detaching chain. It was racy,
437 * because tw buckets are scheduled in not serialized context
438 * in 2.3 (with netfilter), and with softnet it is common, because
439 * soft irqs are not sequenced.
444 tw_for_each_inmate(tw, node, &tcp_tw_death_row[slot]) {
445 __tw_del_dead_node(tw);
446 spin_unlock(&tw_death_lock);
447 tcp_timewait_kill(tw);
450 spin_lock(&tw_death_lock);
451 if (killed > quota) {
456 /* While we dropped tw_death_lock, another cpu may have
457 * killed off the next TW bucket in the list, therefore
458 * do a fresh re-read of the hlist head node with the
459 * lock reacquired. We still use the hlist traversal
460 * macro in order to get the prefetches.
465 tcp_tw_count -= killed;
466 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITED, killed);
471 static void tcp_twkill(unsigned long dummy)
475 spin_lock(&tw_death_lock);
477 if (tcp_tw_count == 0)
481 ret = tcp_do_twkill_work(tcp_tw_death_row_slot, TCP_TWKILL_QUOTA);
483 twkill_thread_slots |= (1 << tcp_tw_death_row_slot);
485 schedule_work(&tcp_twkill_work);
488 /* We purged the entire slot, anything left? */
492 tcp_tw_death_row_slot =
493 ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
495 mod_timer(&tcp_tw_timer, jiffies + TCP_TWKILL_PERIOD);
497 spin_unlock(&tw_death_lock);
500 extern void twkill_slots_invalid(void);
502 static void twkill_work(void *dummy)
506 if ((TCP_TWKILL_SLOTS - 1) > (sizeof(twkill_thread_slots) * 8))
507 twkill_slots_invalid();
509 while (twkill_thread_slots) {
510 spin_lock_bh(&tw_death_lock);
511 for (i = 0; i < TCP_TWKILL_SLOTS; i++) {
512 if (!(twkill_thread_slots & (1 << i)))
515 while (tcp_do_twkill_work(i, TCP_TWKILL_QUOTA) != 0) {
516 if (need_resched()) {
517 spin_unlock_bh(&tw_death_lock);
519 spin_lock_bh(&tw_death_lock);
523 twkill_thread_slots &= ~(1 << i);
525 spin_unlock_bh(&tw_death_lock);
529 /* These are always called from BH context. See callers in
530 * tcp_input.c to verify this.
533 /* This is for handling early-kills of TIME_WAIT sockets. */
534 void tcp_tw_deschedule(struct tcp_tw_bucket *tw)
536 spin_lock(&tw_death_lock);
537 if (tw_del_dead_node(tw)) {
539 if (--tcp_tw_count == 0)
540 del_timer(&tcp_tw_timer);
542 spin_unlock(&tw_death_lock);
543 tcp_timewait_kill(tw);
546 /* Short-time timewait calendar */
548 static int tcp_twcal_hand = -1;
549 static int tcp_twcal_jiffie;
550 static void tcp_twcal_tick(unsigned long);
551 static struct timer_list tcp_twcal_timer =
552 TIMER_INITIALIZER(tcp_twcal_tick, 0, 0);
553 static struct hlist_head tcp_twcal_row[TCP_TW_RECYCLE_SLOTS];
555 static void tcp_tw_schedule(struct tcp_tw_bucket *tw, int timeo)
557 struct hlist_head *list;
560 /* timeout := RTO * 3.5
562 * 3.5 = 1+2+0.5 to wait for two retransmits.
564 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
565 * our ACK acking that FIN can be lost. If N subsequent retransmitted
566 * FINs (or previous seqments) are lost (probability of such event
567 * is p^(N+1), where p is probability to lose single packet and
568 * time to detect the loss is about RTO*(2^N - 1) with exponential
569 * backoff). Normal timewait length is calculated so, that we
570 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
571 * [ BTW Linux. following BSD, violates this requirement waiting
572 * only for 60sec, we should wait at least for 240 secs.
573 * Well, 240 consumes too much of resources 8)
575 * This interval is not reduced to catch old duplicate and
576 * responces to our wandering segments living for two MSLs.
577 * However, if we use PAWS to detect
578 * old duplicates, we can reduce the interval to bounds required
579 * by RTO, rather than MSL. So, if peer understands PAWS, we
580 * kill tw bucket after 3.5*RTO (it is important that this number
581 * is greater than TS tick!) and detect old duplicates with help
584 slot = (timeo + (1<<TCP_TW_RECYCLE_TICK) - 1) >> TCP_TW_RECYCLE_TICK;
586 spin_lock(&tw_death_lock);
588 /* Unlink it, if it was scheduled */
589 if (tw_del_dead_node(tw))
592 atomic_inc(&tw->tw_refcnt);
594 if (slot >= TCP_TW_RECYCLE_SLOTS) {
595 /* Schedule to slow timer */
596 if (timeo >= TCP_TIMEWAIT_LEN) {
597 slot = TCP_TWKILL_SLOTS-1;
599 slot = (timeo + TCP_TWKILL_PERIOD-1) / TCP_TWKILL_PERIOD;
600 if (slot >= TCP_TWKILL_SLOTS)
601 slot = TCP_TWKILL_SLOTS-1;
603 tw->tw_ttd = jiffies + timeo;
604 slot = (tcp_tw_death_row_slot + slot) & (TCP_TWKILL_SLOTS - 1);
605 list = &tcp_tw_death_row[slot];
607 tw->tw_ttd = jiffies + (slot << TCP_TW_RECYCLE_TICK);
609 if (tcp_twcal_hand < 0) {
611 tcp_twcal_jiffie = jiffies;
612 tcp_twcal_timer.expires = tcp_twcal_jiffie + (slot<<TCP_TW_RECYCLE_TICK);
613 add_timer(&tcp_twcal_timer);
615 if (time_after(tcp_twcal_timer.expires, jiffies + (slot<<TCP_TW_RECYCLE_TICK)))
616 mod_timer(&tcp_twcal_timer, jiffies + (slot<<TCP_TW_RECYCLE_TICK));
617 slot = (tcp_twcal_hand + slot)&(TCP_TW_RECYCLE_SLOTS-1);
619 list = &tcp_twcal_row[slot];
622 hlist_add_head(&tw->tw_death_node, list);
624 if (tcp_tw_count++ == 0)
625 mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
626 spin_unlock(&tw_death_lock);
629 void tcp_twcal_tick(unsigned long dummy)
633 unsigned long now = jiffies;
637 spin_lock(&tw_death_lock);
638 if (tcp_twcal_hand < 0)
641 slot = tcp_twcal_hand;
642 j = tcp_twcal_jiffie;
644 for (n=0; n<TCP_TW_RECYCLE_SLOTS; n++) {
645 if (time_before_eq(j, now)) {
646 struct hlist_node *node, *safe;
647 struct tcp_tw_bucket *tw;
649 tw_for_each_inmate_safe(tw, node, safe,
650 &tcp_twcal_row[slot]) {
651 __tw_del_dead_node(tw);
652 tcp_timewait_kill(tw);
659 tcp_twcal_jiffie = j;
660 tcp_twcal_hand = slot;
663 if (!hlist_empty(&tcp_twcal_row[slot])) {
664 mod_timer(&tcp_twcal_timer, j);
668 j += (1<<TCP_TW_RECYCLE_TICK);
669 slot = (slot+1)&(TCP_TW_RECYCLE_SLOTS-1);
674 if ((tcp_tw_count -= killed) == 0)
675 del_timer(&tcp_tw_timer);
676 NET_ADD_STATS_BH(LINUX_MIB_TIMEWAITKILLED, killed);
677 spin_unlock(&tw_death_lock);
680 /* This is not only more efficient than what we used to do, it eliminates
681 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
683 * Actually, we could lots of memory writes here. tp of listening
684 * socket contains all necessary default parameters.
686 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
688 /* allocate the newsk from the same slab of the master sock,
689 * if not, at sk_free time we'll try to free it from the wrong
690 * slabcache (i.e. is it TCPv4 or v6?), this is handled thru sk->sk_prot -acme */
691 struct sock *newsk = sk_alloc(PF_INET, GFP_ATOMIC, sk->sk_prot, 0);
694 struct inet_request_sock *ireq = inet_rsk(req);
695 struct tcp_request_sock *treq = tcp_rsk(req);
696 struct inet_sock *newinet = inet_sk(newsk);
697 struct tcp_sock *newtp;
698 struct sk_filter *filter;
700 memcpy(newsk, sk, sizeof(struct tcp_sock));
701 newsk->sk_state = TCP_SYN_RECV;
704 sk_node_init(&newsk->sk_node);
705 newinet->bind_hash = NULL;
707 /* Clone the TCP header template */
708 newinet->dport = ireq->rmt_port;
710 sock_lock_init(newsk);
713 rwlock_init(&newsk->sk_dst_lock);
714 newsk->sk_dst_cache = NULL;
715 atomic_set(&newsk->sk_rmem_alloc, 0);
716 skb_queue_head_init(&newsk->sk_receive_queue);
717 atomic_set(&newsk->sk_wmem_alloc, 0);
718 skb_queue_head_init(&newsk->sk_write_queue);
719 atomic_set(&newsk->sk_omem_alloc, 0);
720 newsk->sk_wmem_queued = 0;
721 newsk->sk_forward_alloc = 0;
723 sock_reset_flag(newsk, SOCK_DONE);
724 newsk->sk_userlocks = sk->sk_userlocks & ~SOCK_BINDPORT_LOCK;
725 newsk->sk_backlog.head = newsk->sk_backlog.tail = NULL;
726 newsk->sk_send_head = NULL;
727 rwlock_init(&newsk->sk_callback_lock);
728 skb_queue_head_init(&newsk->sk_error_queue);
729 newsk->sk_write_space = sk_stream_write_space;
731 if ((filter = newsk->sk_filter) != NULL)
732 sk_filter_charge(newsk, filter);
734 if (unlikely(xfrm_sk_clone_policy(newsk))) {
735 /* It is still raw copy of parent, so invalidate
736 * destructor and make plain sk_free() */
737 newsk->sk_destruct = NULL;
742 /* Now setup tcp_sock */
743 newtp = tcp_sk(newsk);
744 newtp->pred_flags = 0;
745 newtp->rcv_nxt = treq->rcv_isn + 1;
746 newtp->snd_nxt = treq->snt_isn + 1;
747 newtp->snd_una = treq->snt_isn + 1;
748 newtp->snd_sml = treq->snt_isn + 1;
750 tcp_prequeue_init(newtp);
752 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
754 newtp->retransmits = 0;
757 newtp->mdev = TCP_TIMEOUT_INIT;
758 newtp->rto = TCP_TIMEOUT_INIT;
760 newtp->packets_out = 0;
762 newtp->retrans_out = 0;
763 newtp->sacked_out = 0;
764 newtp->fackets_out = 0;
765 newtp->snd_ssthresh = 0x7fffffff;
767 /* So many TCP implementations out there (incorrectly) count the
768 * initial SYN frame in their delayed-ACK and congestion control
769 * algorithms that we must have the following bandaid to talk
770 * efficiently to them. -DaveM
773 newtp->snd_cwnd_cnt = 0;
775 newtp->frto_counter = 0;
776 newtp->frto_highmark = 0;
778 newtp->ca_ops = &tcp_reno;
780 tcp_set_ca_state(newtp, TCP_CA_Open);
781 tcp_init_xmit_timers(newsk);
782 skb_queue_head_init(&newtp->out_of_order_queue);
783 newtp->rcv_wup = treq->rcv_isn + 1;
784 newtp->write_seq = treq->snt_isn + 1;
785 newtp->pushed_seq = newtp->write_seq;
786 newtp->copied_seq = treq->rcv_isn + 1;
788 newtp->rx_opt.saw_tstamp = 0;
790 newtp->rx_opt.dsack = 0;
791 newtp->rx_opt.eff_sacks = 0;
793 newtp->probes_out = 0;
794 newtp->rx_opt.num_sacks = 0;
796 /* Deinitialize accept_queue to trap illegal accesses. */
797 memset(&newtp->accept_queue, 0, sizeof(newtp->accept_queue));
799 /* Back to base struct sock members. */
801 newsk->sk_priority = 0;
802 atomic_set(&newsk->sk_refcnt, 2);
805 * Increment the counter in the same struct proto as the master
806 * sock (sk_refcnt_debug_inc uses newsk->sk_prot->socks, that
807 * is the same as sk->sk_prot->socks, as this field was copied
808 * with memcpy), same rationale as the first comment in this
811 * This _changes_ the previous behaviour, where
812 * tcp_create_openreq_child always was incrementing the
813 * equivalent to tcp_prot->socks (inet_sock_nr), so this have
814 * to be taken into account in all callers. -acme
816 sk_refcnt_debug_inc(newsk);
818 atomic_inc(&tcp_sockets_allocated);
820 if (sock_flag(newsk, SOCK_KEEPOPEN))
821 tcp_reset_keepalive_timer(newsk,
822 keepalive_time_when(newtp));
823 newsk->sk_socket = NULL;
824 newsk->sk_sleep = NULL;
826 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
827 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
829 newtp->rx_opt.sack_ok |= 2;
831 newtp->window_clamp = req->window_clamp;
832 newtp->rcv_ssthresh = req->rcv_wnd;
833 newtp->rcv_wnd = req->rcv_wnd;
834 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
835 if (newtp->rx_opt.wscale_ok) {
836 newtp->rx_opt.snd_wscale = ireq->snd_wscale;
837 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
839 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
840 newtp->window_clamp = min(newtp->window_clamp, 65535U);
842 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
843 newtp->max_window = newtp->snd_wnd;
845 if (newtp->rx_opt.tstamp_ok) {
846 newtp->rx_opt.ts_recent = req->ts_recent;
847 newtp->rx_opt.ts_recent_stamp = xtime.tv_sec;
848 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
850 newtp->rx_opt.ts_recent_stamp = 0;
851 newtp->tcp_header_len = sizeof(struct tcphdr);
853 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
854 newtp->ack.last_seg_size = skb->len-newtp->tcp_header_len;
855 newtp->rx_opt.mss_clamp = req->mss;
856 TCP_ECN_openreq_child(newtp, req);
857 if (newtp->ecn_flags&TCP_ECN_OK)
858 sock_set_flag(newsk, SOCK_NO_LARGESEND);
860 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
866 * Process an incoming packet for SYN_RECV sockets represented
870 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
871 struct request_sock *req,
872 struct request_sock **prev)
874 struct tcphdr *th = skb->h.th;
875 struct tcp_sock *tp = tcp_sk(sk);
876 u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
878 struct tcp_options_received tmp_opt;
881 tmp_opt.saw_tstamp = 0;
882 if (th->doff > (sizeof(struct tcphdr)>>2)) {
883 tcp_parse_options(skb, &tmp_opt, 0);
885 if (tmp_opt.saw_tstamp) {
886 tmp_opt.ts_recent = req->ts_recent;
887 /* We do not store true stamp, but it is not required,
888 * it can be estimated (approximately)
891 tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
892 paws_reject = tcp_paws_check(&tmp_opt, th->rst);
896 /* Check for pure retransmitted SYN. */
897 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
898 flg == TCP_FLAG_SYN &&
901 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
902 * this case on figure 6 and figure 8, but formal
903 * protocol description says NOTHING.
904 * To be more exact, it says that we should send ACK,
905 * because this segment (at least, if it has no data)
908 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT
909 * describe SYN-RECV state. All the description
910 * is wrong, we cannot believe to it and should
911 * rely only on common sense and implementation
914 * Enforce "SYN-ACK" according to figure 8, figure 6
915 * of RFC793, fixed by RFC1122.
917 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
921 /* Further reproduces section "SEGMENT ARRIVES"
922 for state SYN-RECEIVED of RFC793.
923 It is broken, however, it does not work only
924 when SYNs are crossed.
926 You would think that SYN crossing is impossible here, since
927 we should have a SYN_SENT socket (from connect()) on our end,
928 but this is not true if the crossed SYNs were sent to both
929 ends by a malicious third party. We must defend against this,
930 and to do that we first verify the ACK (as per RFC793, page
931 36) and reset if it is invalid. Is this a true full defense?
932 To convince ourselves, let us consider a way in which the ACK
933 test can still pass in this 'malicious crossed SYNs' case.
934 Malicious sender sends identical SYNs (and thus identical sequence
935 numbers) to both A and B:
940 By our good fortune, both A and B select the same initial
941 send sequence number of seven :-)
943 A: sends SYN|ACK, seq=7, ack_seq=8
944 B: sends SYN|ACK, seq=7, ack_seq=8
946 So we are now A eating this SYN|ACK, ACK test passes. So
947 does sequence test, SYN is truncated, and thus we consider
950 If tp->defer_accept, we silently drop this bare ACK. Otherwise,
951 we create an established connection. Both ends (listening sockets)
952 accept the new incoming connection and try to talk to each other. 8-)
954 Note: This case is both harmless, and rare. Possibility is about the
955 same as us discovering intelligent life on another plant tomorrow.
957 But generally, we should (RFC lies!) to accept ACK
958 from SYNACK both here and in tcp_rcv_state_process().
959 tcp_rcv_state_process() does not, hence, we do not too.
961 Note that the case is absolutely generic:
962 we cannot optimize anything here without
963 violating protocol. All the checks must be made
964 before attempt to create socket.
967 /* RFC793 page 36: "If the connection is in any non-synchronized state ...
968 * and the incoming segment acknowledges something not yet
969 * sent (the segment carries an unaccaptable ACK) ...
972 * Invalid ACK: reset will be sent by listening socket
974 if ((flg & TCP_FLAG_ACK) &&
975 (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
978 /* Also, it would be not so bad idea to check rcv_tsecr, which
979 * is essentially ACK extension and too early or too late values
980 * should cause reset in unsynchronized states.
983 /* RFC793: "first check sequence number". */
985 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
986 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
987 /* Out of window: send ACK and drop. */
988 if (!(flg & TCP_FLAG_RST))
989 req->rsk_ops->send_ack(skb, req);
991 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
995 /* In sequence, PAWS is OK. */
997 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
998 req->ts_recent = tmp_opt.rcv_tsval;
1000 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
1001 /* Truncate SYN, it is out of window starting
1002 at tcp_rsk(req)->rcv_isn + 1. */
1003 flg &= ~TCP_FLAG_SYN;
1006 /* RFC793: "second check the RST bit" and
1007 * "fourth, check the SYN bit"
1009 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN))
1010 goto embryonic_reset;
1012 /* ACK sequence verified above, just make sure ACK is
1013 * set. If ACK not set, just silently drop the packet.
1015 if (!(flg & TCP_FLAG_ACK))
1018 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
1019 if (tp->defer_accept && TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
1020 inet_rsk(req)->acked = 1;
1024 /* OK, ACK is valid, create big socket and
1025 * feed this segment to it. It will repeat all
1026 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
1027 * ESTABLISHED STATE. If it will be dropped after
1028 * socket is created, wait for troubles.
1030 child = tp->af_specific->syn_recv_sock(sk, skb, req, NULL);
1032 goto listen_overflow;
1034 tcp_synq_unlink(tp, req, prev);
1035 tcp_synq_removed(sk, req);
1037 tcp_acceptq_queue(sk, req, child);
1041 if (!sysctl_tcp_abort_on_overflow) {
1042 inet_rsk(req)->acked = 1;
1047 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
1048 if (!(flg & TCP_FLAG_RST))
1049 req->rsk_ops->send_reset(skb);
1051 tcp_synq_drop(sk, req, prev);
1056 * Queue segment on the new socket if the new socket is active,
1057 * otherwise we just shortcircuit this and continue with
1061 int tcp_child_process(struct sock *parent, struct sock *child,
1062 struct sk_buff *skb)
1065 int state = child->sk_state;
1067 if (!sock_owned_by_user(child)) {
1068 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
1070 /* Wakeup parent, send SIGIO */
1071 if (state == TCP_SYN_RECV && child->sk_state != state)
1072 parent->sk_data_ready(parent, 0);
1074 /* Alas, it is possible again, because we do lookup
1075 * in main socket hash table and lock on listening
1076 * socket does not protect us more.
1078 sk_add_backlog(child, skb);
1081 bh_unlock_sock(child);
1086 EXPORT_SYMBOL(tcp_check_req);
1087 EXPORT_SYMBOL(tcp_child_process);
1088 EXPORT_SYMBOL(tcp_create_openreq_child);
1089 EXPORT_SYMBOL(tcp_timewait_state_process);
1090 EXPORT_SYMBOL(tcp_tw_deschedule);