Linux Audio

Check our new training course

Loading...
v3.1
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the  BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Implementation of the Transmission Control Protocol(TCP).
  7 *
  8 * Authors:	Ross Biro
  9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
 11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
 12 *		Florian La Roche, <flla@stud.uni-sb.de>
 13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
 15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
 16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
 17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18 *		Jorge Cwik, <jorge@laser.satlink.net>
 19 */
 20
 21#include <linux/mm.h>
 22#include <linux/module.h>
 23#include <linux/slab.h>
 24#include <linux/sysctl.h>
 25#include <linux/workqueue.h>
 26#include <net/tcp.h>
 27#include <net/inet_common.h>
 28#include <net/xfrm.h>
 29
 30int sysctl_tcp_syncookies __read_mostly = 1;
 31EXPORT_SYMBOL(sysctl_tcp_syncookies);
 32
 33int sysctl_tcp_abort_on_overflow __read_mostly;
 34
 35struct inet_timewait_death_row tcp_death_row = {
 36	.sysctl_max_tw_buckets = NR_FILE * 2,
 37	.period		= TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
 38	.death_lock	= __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
 39	.hashinfo	= &tcp_hashinfo,
 40	.tw_timer	= TIMER_INITIALIZER(inet_twdr_hangman, 0,
 41					    (unsigned long)&tcp_death_row),
 42	.twkill_work	= __WORK_INITIALIZER(tcp_death_row.twkill_work,
 43					     inet_twdr_twkill_work),
 44/* Short-time timewait calendar */
 45
 46	.twcal_hand	= -1,
 47	.twcal_timer	= TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
 48					    (unsigned long)&tcp_death_row),
 49};
 50EXPORT_SYMBOL_GPL(tcp_death_row);
 51
 52/* VJ's idea. Save last timestamp seen from this destination
 53 * and hold it at least for normal timewait interval to use for duplicate
 54 * segment detection in subsequent connections, before they enter synchronized
 55 * state.
 56 */
 57
 58static int tcp_remember_stamp(struct sock *sk)
 59{
 60	const struct inet_connection_sock *icsk = inet_csk(sk);
 61	struct tcp_sock *tp = tcp_sk(sk);
 62	struct inet_peer *peer;
 63	bool release_it;
 64
 65	peer = icsk->icsk_af_ops->get_peer(sk, &release_it);
 66	if (peer) {
 67		if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 ||
 68		    ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL &&
 69		     peer->tcp_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 70			peer->tcp_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 71			peer->tcp_ts = tp->rx_opt.ts_recent;
 72		}
 73		if (release_it)
 74			inet_putpeer(peer);
 75		return 1;
 76	}
 77
 78	return 0;
 79}
 80
 81static int tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 82{
 83	struct sock *sk = (struct sock *) tw;
 84	struct inet_peer *peer;
 85
 86	peer = twsk_getpeer(sk);
 87	if (peer) {
 88		const struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
 89
 90		if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 ||
 91		    ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL &&
 92		     peer->tcp_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 93			peer->tcp_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 94			peer->tcp_ts	   = tcptw->tw_ts_recent;
 95		}
 96		inet_putpeer(peer);
 97		return 1;
 98	}
 99	return 0;
100}
101
102static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
103{
104	if (seq == s_win)
105		return 1;
106	if (after(end_seq, s_win) && before(seq, e_win))
107		return 1;
108	return seq == e_win && seq == end_seq;
109}
110
111/*
112 * * Main purpose of TIME-WAIT state is to close connection gracefully,
113 *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
114 *   (and, probably, tail of data) and one or more our ACKs are lost.
115 * * What is TIME-WAIT timeout? It is associated with maximal packet
116 *   lifetime in the internet, which results in wrong conclusion, that
117 *   it is set to catch "old duplicate segments" wandering out of their path.
118 *   It is not quite correct. This timeout is calculated so that it exceeds
119 *   maximal retransmission timeout enough to allow to lose one (or more)
120 *   segments sent by peer and our ACKs. This time may be calculated from RTO.
121 * * When TIME-WAIT socket receives RST, it means that another end
122 *   finally closed and we are allowed to kill TIME-WAIT too.
123 * * Second purpose of TIME-WAIT is catching old duplicate segments.
124 *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
125 *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
126 * * If we invented some more clever way to catch duplicates
127 *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
128 *
129 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
130 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
131 * from the very beginning.
132 *
133 * NOTE. With recycling (and later with fin-wait-2) TW bucket
134 * is _not_ stateless. It means, that strictly speaking we must
135 * spinlock it. I do not want! Well, probability of misbehaviour
136 * is ridiculously low and, seems, we could use some mb() tricks
137 * to avoid misread sequence numbers, states etc.  --ANK
 
 
138 */
139enum tcp_tw_status
140tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
141			   const struct tcphdr *th)
142{
143	struct tcp_options_received tmp_opt;
144	u8 *hash_location;
145	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
146	int paws_reject = 0;
147
148	tmp_opt.saw_tstamp = 0;
149	if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
150		tcp_parse_options(skb, &tmp_opt, &hash_location, 0);
151
152		if (tmp_opt.saw_tstamp) {
 
153			tmp_opt.ts_recent	= tcptw->tw_ts_recent;
154			tmp_opt.ts_recent_stamp	= tcptw->tw_ts_recent_stamp;
155			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
156		}
157	}
158
159	if (tw->tw_substate == TCP_FIN_WAIT2) {
160		/* Just repeat all the checks of tcp_rcv_state_process() */
161
162		/* Out of window, send ACK */
163		if (paws_reject ||
164		    !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
165				   tcptw->tw_rcv_nxt,
166				   tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
167			return TCP_TW_ACK;
168
169		if (th->rst)
170			goto kill;
171
172		if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
173			goto kill_with_rst;
174
175		/* Dup ACK? */
176		if (!th->ack ||
177		    !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
178		    TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
179			inet_twsk_put(tw);
180			return TCP_TW_SUCCESS;
181		}
182
183		/* New data or FIN. If new data arrive after half-duplex close,
184		 * reset.
185		 */
186		if (!th->fin ||
187		    TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
188kill_with_rst:
189			inet_twsk_deschedule(tw, &tcp_death_row);
190			inet_twsk_put(tw);
191			return TCP_TW_RST;
192		}
193
194		/* FIN arrived, enter true time-wait state. */
195		tw->tw_substate	  = TCP_TIME_WAIT;
196		tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
197		if (tmp_opt.saw_tstamp) {
198			tcptw->tw_ts_recent_stamp = get_seconds();
199			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
200		}
201
202		if (tcp_death_row.sysctl_tw_recycle &&
203		    tcptw->tw_ts_recent_stamp &&
204		    tcp_tw_remember_stamp(tw))
205			inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
206					   TCP_TIMEWAIT_LEN);
207		else
208			inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
209					   TCP_TIMEWAIT_LEN);
210		return TCP_TW_ACK;
211	}
212
213	/*
214	 *	Now real TIME-WAIT state.
215	 *
216	 *	RFC 1122:
217	 *	"When a connection is [...] on TIME-WAIT state [...]
218	 *	[a TCP] MAY accept a new SYN from the remote TCP to
219	 *	reopen the connection directly, if it:
220	 *
221	 *	(1)  assigns its initial sequence number for the new
222	 *	connection to be larger than the largest sequence
223	 *	number it used on the previous connection incarnation,
224	 *	and
225	 *
226	 *	(2)  returns to TIME-WAIT state if the SYN turns out
227	 *	to be an old duplicate".
228	 */
229
230	if (!paws_reject &&
231	    (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
232	     (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
233		/* In window segment, it may be only reset or bare ack. */
234
235		if (th->rst) {
236			/* This is TIME_WAIT assassination, in two flavors.
237			 * Oh well... nobody has a sufficient solution to this
238			 * protocol bug yet.
239			 */
240			if (sysctl_tcp_rfc1337 == 0) {
241kill:
242				inet_twsk_deschedule(tw, &tcp_death_row);
243				inet_twsk_put(tw);
244				return TCP_TW_SUCCESS;
245			}
246		}
247		inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
248				   TCP_TIMEWAIT_LEN);
249
250		if (tmp_opt.saw_tstamp) {
251			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
252			tcptw->tw_ts_recent_stamp = get_seconds();
253		}
254
255		inet_twsk_put(tw);
256		return TCP_TW_SUCCESS;
257	}
258
259	/* Out of window segment.
260
261	   All the segments are ACKed immediately.
262
263	   The only exception is new SYN. We accept it, if it is
264	   not old duplicate and we are not in danger to be killed
265	   by delayed old duplicates. RFC check is that it has
266	   newer sequence number works at rates <40Mbit/sec.
267	   However, if paws works, it is reliable AND even more,
268	   we even may relax silly seq space cutoff.
269
270	   RED-PEN: we violate main RFC requirement, if this SYN will appear
271	   old duplicate (i.e. we receive RST in reply to SYN-ACK),
272	   we must return socket to time-wait state. It is not good,
273	   but not fatal yet.
274	 */
275
276	if (th->syn && !th->rst && !th->ack && !paws_reject &&
277	    (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
278	     (tmp_opt.saw_tstamp &&
279	      (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
280		u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
281		if (isn == 0)
282			isn++;
283		TCP_SKB_CB(skb)->when = isn;
284		return TCP_TW_SYN;
285	}
286
287	if (paws_reject)
288		NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
289
290	if (!th->rst) {
291		/* In this case we must reset the TIMEWAIT timer.
292		 *
293		 * If it is ACKless SYN it may be both old duplicate
294		 * and new good SYN with random sequence number <rcv_nxt.
295		 * Do not reschedule in the last case.
296		 */
297		if (paws_reject || th->ack)
298			inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
299					   TCP_TIMEWAIT_LEN);
300
301		/* Send ACK. Note, we do not put the bucket,
302		 * it will be released by caller.
303		 */
304		return TCP_TW_ACK;
305	}
306	inet_twsk_put(tw);
307	return TCP_TW_SUCCESS;
308}
309EXPORT_SYMBOL(tcp_timewait_state_process);
310
311/*
312 * Move a socket to time-wait or dead fin-wait-2 state.
313 */
314void tcp_time_wait(struct sock *sk, int state, int timeo)
315{
316	struct inet_timewait_sock *tw = NULL;
317	const struct inet_connection_sock *icsk = inet_csk(sk);
318	const struct tcp_sock *tp = tcp_sk(sk);
319	int recycle_ok = 0;
320
321	if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
322		recycle_ok = tcp_remember_stamp(sk);
323
324	if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
325		tw = inet_twsk_alloc(sk, state);
326
327	if (tw != NULL) {
328		struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
329		const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
 
330
331		tw->tw_transparent	= inet_sk(sk)->transparent;
332		tw->tw_rcv_wscale	= tp->rx_opt.rcv_wscale;
333		tcptw->tw_rcv_nxt	= tp->rcv_nxt;
334		tcptw->tw_snd_nxt	= tp->snd_nxt;
335		tcptw->tw_rcv_wnd	= tcp_receive_window(tp);
336		tcptw->tw_ts_recent	= tp->rx_opt.ts_recent;
337		tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
 
338
339#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
340		if (tw->tw_family == PF_INET6) {
341			struct ipv6_pinfo *np = inet6_sk(sk);
342			struct inet6_timewait_sock *tw6;
343
344			tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
345			tw6 = inet6_twsk((struct sock *)tw);
346			ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
347			ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
348			tw->tw_ipv6only = np->ipv6only;
349		}
350#endif
351
352#ifdef CONFIG_TCP_MD5SIG
353		/*
354		 * The timewait bucket does not have the key DB from the
355		 * sock structure. We just make a quick copy of the
356		 * md5 key being used (if indeed we are using one)
357		 * so the timewait ack generating code has the key.
358		 */
359		do {
360			struct tcp_md5sig_key *key;
361			memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key));
362			tcptw->tw_md5_keylen = 0;
363			key = tp->af_specific->md5_lookup(sk, sk);
364			if (key != NULL) {
365				memcpy(&tcptw->tw_md5_key, key->key, key->keylen);
366				tcptw->tw_md5_keylen = key->keylen;
367				if (tcp_alloc_md5sig_pool(sk) == NULL)
368					BUG();
369			}
370		} while (0);
371#endif
372
373		/* Linkage updates. */
374		__inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
375
376		/* Get the TIME_WAIT timeout firing. */
377		if (timeo < rto)
378			timeo = rto;
379
380		if (recycle_ok) {
381			tw->tw_timeout = rto;
382		} else {
383			tw->tw_timeout = TCP_TIMEWAIT_LEN;
384			if (state == TCP_TIME_WAIT)
385				timeo = TCP_TIMEWAIT_LEN;
386		}
387
388		inet_twsk_schedule(tw, &tcp_death_row, timeo,
389				   TCP_TIMEWAIT_LEN);
390		inet_twsk_put(tw);
391	} else {
392		/* Sorry, if we're out of memory, just CLOSE this
393		 * socket up.  We've got bigger problems than
394		 * non-graceful socket closings.
395		 */
396		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
397	}
398
399	tcp_update_metrics(sk);
400	tcp_done(sk);
401}
402
403void tcp_twsk_destructor(struct sock *sk)
404{
405#ifdef CONFIG_TCP_MD5SIG
406	struct tcp_timewait_sock *twsk = tcp_twsk(sk);
407	if (twsk->tw_md5_keylen)
408		tcp_free_md5sig_pool();
 
409#endif
410}
411EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
412
413static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
414					 struct request_sock *req)
415{
416	tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
417}
418
419/* This is not only more efficient than what we used to do, it eliminates
420 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
421 *
422 * Actually, we could lots of memory writes here. tp of listening
423 * socket contains all necessary default parameters.
424 */
425struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
426{
427	struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
428
429	if (newsk != NULL) {
430		const struct inet_request_sock *ireq = inet_rsk(req);
431		struct tcp_request_sock *treq = tcp_rsk(req);
432		struct inet_connection_sock *newicsk = inet_csk(newsk);
433		struct tcp_sock *newtp = tcp_sk(newsk);
434		struct tcp_sock *oldtp = tcp_sk(sk);
435		struct tcp_cookie_values *oldcvp = oldtp->cookie_values;
436
437		/* TCP Cookie Transactions require space for the cookie pair,
438		 * as it differs for each connection.  There is no need to
439		 * copy any s_data_payload stored at the original socket.
440		 * Failure will prevent resuming the connection.
441		 *
442		 * Presumed copied, in order of appearance:
443		 *	cookie_in_always, cookie_out_never
444		 */
445		if (oldcvp != NULL) {
446			struct tcp_cookie_values *newcvp =
447				kzalloc(sizeof(*newtp->cookie_values),
448					GFP_ATOMIC);
449
450			if (newcvp != NULL) {
451				kref_init(&newcvp->kref);
452				newcvp->cookie_desired =
453						oldcvp->cookie_desired;
454				newtp->cookie_values = newcvp;
455			} else {
456				/* Not Yet Implemented */
457				newtp->cookie_values = NULL;
458			}
459		}
460
461		/* Now setup tcp_sock */
462		newtp->pred_flags = 0;
463
464		newtp->rcv_wup = newtp->copied_seq =
465		newtp->rcv_nxt = treq->rcv_isn + 1;
466
467		newtp->snd_sml = newtp->snd_una =
468		newtp->snd_nxt = newtp->snd_up =
469			treq->snt_isn + 1 + tcp_s_data_size(oldtp);
470
471		tcp_prequeue_init(newtp);
 
472
473		tcp_init_wl(newtp, treq->rcv_isn);
474
475		newtp->srtt = 0;
476		newtp->mdev = TCP_TIMEOUT_INIT;
477		newicsk->icsk_rto = TCP_TIMEOUT_INIT;
478
479		newtp->packets_out = 0;
480		newtp->retrans_out = 0;
481		newtp->sacked_out = 0;
482		newtp->fackets_out = 0;
483		newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 
 
 
 
484
485		/* So many TCP implementations out there (incorrectly) count the
486		 * initial SYN frame in their delayed-ACK and congestion control
487		 * algorithms that we must have the following bandaid to talk
488		 * efficiently to them.  -DaveM
489		 */
490		newtp->snd_cwnd = TCP_INIT_CWND;
491		newtp->snd_cwnd_cnt = 0;
492		newtp->bytes_acked = 0;
493
494		newtp->frto_counter = 0;
495		newtp->frto_highmark = 0;
496
497		newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
498
499		tcp_set_ca_state(newsk, TCP_CA_Open);
500		tcp_init_xmit_timers(newsk);
501		skb_queue_head_init(&newtp->out_of_order_queue);
502		newtp->write_seq = newtp->pushed_seq =
503			treq->snt_isn + 1 + tcp_s_data_size(oldtp);
504
505		newtp->rx_opt.saw_tstamp = 0;
506
507		newtp->rx_opt.dsack = 0;
508		newtp->rx_opt.num_sacks = 0;
509
510		newtp->urg_data = 0;
511
512		if (sock_flag(newsk, SOCK_KEEPOPEN))
513			inet_csk_reset_keepalive_timer(newsk,
514						       keepalive_time_when(newtp));
515
516		newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
517		if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
518			if (sysctl_tcp_fack)
519				tcp_enable_fack(newtp);
520		}
521		newtp->window_clamp = req->window_clamp;
522		newtp->rcv_ssthresh = req->rcv_wnd;
523		newtp->rcv_wnd = req->rcv_wnd;
524		newtp->rx_opt.wscale_ok = ireq->wscale_ok;
525		if (newtp->rx_opt.wscale_ok) {
526			newtp->rx_opt.snd_wscale = ireq->snd_wscale;
527			newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
528		} else {
529			newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
530			newtp->window_clamp = min(newtp->window_clamp, 65535U);
531		}
532		newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
533				  newtp->rx_opt.snd_wscale);
534		newtp->max_window = newtp->snd_wnd;
535
536		if (newtp->rx_opt.tstamp_ok) {
537			newtp->rx_opt.ts_recent = req->ts_recent;
538			newtp->rx_opt.ts_recent_stamp = get_seconds();
539			newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
540		} else {
541			newtp->rx_opt.ts_recent_stamp = 0;
542			newtp->tcp_header_len = sizeof(struct tcphdr);
543		}
 
544#ifdef CONFIG_TCP_MD5SIG
545		newtp->md5sig_info = NULL;	/*XXX*/
546		if (newtp->af_specific->md5_lookup(sk, newsk))
547			newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
548#endif
549		if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
550			newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
551		newtp->rx_opt.mss_clamp = req->mss;
552		TCP_ECN_openreq_child(newtp, req);
 
 
553
554		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
555	}
556	return newsk;
557}
558EXPORT_SYMBOL(tcp_create_openreq_child);
559
560/*
561 *	Process an incoming packet for SYN_RECV sockets represented
562 *	as a request_sock.
 
 
 
 
 
 
563 */
564
565struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
566			   struct request_sock *req,
567			   struct request_sock **prev)
 
568{
569	struct tcp_options_received tmp_opt;
570	u8 *hash_location;
571	struct sock *child;
572	const struct tcphdr *th = tcp_hdr(skb);
573	__be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
574	int paws_reject = 0;
 
 
575
576	tmp_opt.saw_tstamp = 0;
577	if (th->doff > (sizeof(struct tcphdr)>>2)) {
578		tcp_parse_options(skb, &tmp_opt, &hash_location, 0);
579
580		if (tmp_opt.saw_tstamp) {
581			tmp_opt.ts_recent = req->ts_recent;
582			/* We do not store true stamp, but it is not required,
583			 * it can be estimated (approximately)
584			 * from another data.
585			 */
586			tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
587			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
588		}
589	}
590
591	/* Check for pure retransmitted SYN. */
592	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
593	    flg == TCP_FLAG_SYN &&
594	    !paws_reject) {
595		/*
596		 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
597		 * this case on figure 6 and figure 8, but formal
598		 * protocol description says NOTHING.
599		 * To be more exact, it says that we should send ACK,
600		 * because this segment (at least, if it has no data)
601		 * is out of window.
602		 *
603		 *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
604		 *  describe SYN-RECV state. All the description
605		 *  is wrong, we cannot believe to it and should
606		 *  rely only on common sense and implementation
607		 *  experience.
608		 *
609		 * Enforce "SYN-ACK" according to figure 8, figure 6
610		 * of RFC793, fixed by RFC1122.
 
 
 
 
 
 
611		 */
612		req->rsk_ops->rtx_syn_ack(sk, req, NULL);
 
 
613		return NULL;
614	}
615
616	/* Further reproduces section "SEGMENT ARRIVES"
617	   for state SYN-RECEIVED of RFC793.
618	   It is broken, however, it does not work only
619	   when SYNs are crossed.
620
621	   You would think that SYN crossing is impossible here, since
622	   we should have a SYN_SENT socket (from connect()) on our end,
623	   but this is not true if the crossed SYNs were sent to both
624	   ends by a malicious third party.  We must defend against this,
625	   and to do that we first verify the ACK (as per RFC793, page
626	   36) and reset if it is invalid.  Is this a true full defense?
627	   To convince ourselves, let us consider a way in which the ACK
628	   test can still pass in this 'malicious crossed SYNs' case.
629	   Malicious sender sends identical SYNs (and thus identical sequence
630	   numbers) to both A and B:
631
632		A: gets SYN, seq=7
633		B: gets SYN, seq=7
634
635	   By our good fortune, both A and B select the same initial
636	   send sequence number of seven :-)
637
638		A: sends SYN|ACK, seq=7, ack_seq=8
639		B: sends SYN|ACK, seq=7, ack_seq=8
640
641	   So we are now A eating this SYN|ACK, ACK test passes.  So
642	   does sequence test, SYN is truncated, and thus we consider
643	   it a bare ACK.
644
645	   If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
646	   bare ACK.  Otherwise, we create an established connection.  Both
647	   ends (listening sockets) accept the new incoming connection and try
648	   to talk to each other. 8-)
649
650	   Note: This case is both harmless, and rare.  Possibility is about the
651	   same as us discovering intelligent life on another plant tomorrow.
652
653	   But generally, we should (RFC lies!) to accept ACK
654	   from SYNACK both here and in tcp_rcv_state_process().
655	   tcp_rcv_state_process() does not, hence, we do not too.
656
657	   Note that the case is absolutely generic:
658	   we cannot optimize anything here without
659	   violating protocol. All the checks must be made
660	   before attempt to create socket.
661	 */
662
663	/* RFC793 page 36: "If the connection is in any non-synchronized state ...
664	 *                  and the incoming segment acknowledges something not yet
665	 *                  sent (the segment carries an unacceptable ACK) ...
666	 *                  a reset is sent."
667	 *
668	 * Invalid ACK: reset will be sent by listening socket
 
 
 
669	 */
670	if ((flg & TCP_FLAG_ACK) &&
671	    (TCP_SKB_CB(skb)->ack_seq !=
672	     tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk))))
673		return sk;
674
675	/* Also, it would be not so bad idea to check rcv_tsecr, which
676	 * is essentially ACK extension and too early or too late values
677	 * should cause reset in unsynchronized states.
678	 */
679
680	/* RFC793: "first check sequence number". */
681
682	if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
683					  tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
684		/* Out of window: send ACK and drop. */
685		if (!(flg & TCP_FLAG_RST))
686			req->rsk_ops->send_ack(sk, skb, req);
687		if (paws_reject)
688			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
689		return NULL;
690	}
691
692	/* In sequence, PAWS is OK. */
693
694	if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
695		req->ts_recent = tmp_opt.rcv_tsval;
696
697	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
698		/* Truncate SYN, it is out of window starting
699		   at tcp_rsk(req)->rcv_isn + 1. */
700		flg &= ~TCP_FLAG_SYN;
701	}
702
703	/* RFC793: "second check the RST bit" and
704	 *	   "fourth, check the SYN bit"
705	 */
706	if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
707		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
708		goto embryonic_reset;
709	}
710
711	/* ACK sequence verified above, just make sure ACK is
712	 * set.  If ACK not set, just silently drop the packet.
 
 
 
713	 */
714	if (!(flg & TCP_FLAG_ACK))
715		return NULL;
716
 
 
 
 
 
 
717	/* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
718	if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
719	    TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
720		inet_rsk(req)->acked = 1;
721		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
722		return NULL;
723	}
724	if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr)
725		tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr;
726	else if (req->retrans) /* don't take RTT sample if retrans && ~TS */
727		tcp_rsk(req)->snt_synack = 0;
728
729	/* OK, ACK is valid, create big socket and
730	 * feed this segment to it. It will repeat all
731	 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
732	 * ESTABLISHED STATE. If it will be dropped after
733	 * socket is created, wait for troubles.
734	 */
735	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
736	if (child == NULL)
737		goto listen_overflow;
738
739	inet_csk_reqsk_queue_unlink(sk, req, prev);
740	inet_csk_reqsk_queue_removed(sk, req);
741
742	inet_csk_reqsk_queue_add(sk, req, child);
743	return child;
744
745listen_overflow:
746	if (!sysctl_tcp_abort_on_overflow) {
747		inet_rsk(req)->acked = 1;
748		return NULL;
749	}
750
751embryonic_reset:
752	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
753	if (!(flg & TCP_FLAG_RST))
 
 
 
 
754		req->rsk_ops->send_reset(sk, skb);
755
756	inet_csk_reqsk_queue_drop(sk, req, prev);
 
 
 
 
 
 
757	return NULL;
758}
759EXPORT_SYMBOL(tcp_check_req);
760
761/*
762 * Queue segment on the new socket if the new socket is active,
763 * otherwise we just shortcircuit this and continue with
764 * the new socket.
 
 
 
 
 
 
765 */
766
767int tcp_child_process(struct sock *parent, struct sock *child,
768		      struct sk_buff *skb)
769{
770	int ret = 0;
771	int state = child->sk_state;
772
773	if (!sock_owned_by_user(child)) {
774		ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
775					    skb->len);
776		/* Wakeup parent, send SIGIO */
777		if (state == TCP_SYN_RECV && child->sk_state != state)
778			parent->sk_data_ready(parent, 0);
779	} else {
780		/* Alas, it is possible again, because we do lookup
781		 * in main socket hash table and lock on listening
782		 * socket does not protect us more.
783		 */
784		__sk_add_backlog(child, skb);
785	}
786
787	bh_unlock_sock(child);
788	sock_put(child);
789	return ret;
790}
791EXPORT_SYMBOL(tcp_child_process);
v3.15
  1/*
  2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
  3 *		operating system.  INET is implemented using the  BSD Socket
  4 *		interface as the means of communication with the user level.
  5 *
  6 *		Implementation of the Transmission Control Protocol(TCP).
  7 *
  8 * Authors:	Ross Biro
  9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
 11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
 12 *		Florian La Roche, <flla@stud.uni-sb.de>
 13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
 15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
 16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
 17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18 *		Jorge Cwik, <jorge@laser.satlink.net>
 19 */
 20
 21#include <linux/mm.h>
 22#include <linux/module.h>
 23#include <linux/slab.h>
 24#include <linux/sysctl.h>
 25#include <linux/workqueue.h>
 26#include <net/tcp.h>
 27#include <net/inet_common.h>
 28#include <net/xfrm.h>
 29
 30int sysctl_tcp_syncookies __read_mostly = 1;
 31EXPORT_SYMBOL(sysctl_tcp_syncookies);
 32
 33int sysctl_tcp_abort_on_overflow __read_mostly;
 34
 35struct inet_timewait_death_row tcp_death_row = {
 36	.sysctl_max_tw_buckets = NR_FILE * 2,
 37	.period		= TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
 38	.death_lock	= __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
 39	.hashinfo	= &tcp_hashinfo,
 40	.tw_timer	= TIMER_INITIALIZER(inet_twdr_hangman, 0,
 41					    (unsigned long)&tcp_death_row),
 42	.twkill_work	= __WORK_INITIALIZER(tcp_death_row.twkill_work,
 43					     inet_twdr_twkill_work),
 44/* Short-time timewait calendar */
 45
 46	.twcal_hand	= -1,
 47	.twcal_timer	= TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
 48					    (unsigned long)&tcp_death_row),
 49};
 50EXPORT_SYMBOL_GPL(tcp_death_row);
 51
 52static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 53{
 54	if (seq == s_win)
 55		return true;
 56	if (after(end_seq, s_win) && before(seq, e_win))
 57		return true;
 58	return seq == e_win && seq == end_seq;
 59}
 60
 61/*
 62 * * Main purpose of TIME-WAIT state is to close connection gracefully,
 63 *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
 64 *   (and, probably, tail of data) and one or more our ACKs are lost.
 65 * * What is TIME-WAIT timeout? It is associated with maximal packet
 66 *   lifetime in the internet, which results in wrong conclusion, that
 67 *   it is set to catch "old duplicate segments" wandering out of their path.
 68 *   It is not quite correct. This timeout is calculated so that it exceeds
 69 *   maximal retransmission timeout enough to allow to lose one (or more)
 70 *   segments sent by peer and our ACKs. This time may be calculated from RTO.
 71 * * When TIME-WAIT socket receives RST, it means that another end
 72 *   finally closed and we are allowed to kill TIME-WAIT too.
 73 * * Second purpose of TIME-WAIT is catching old duplicate segments.
 74 *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
 75 *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
 76 * * If we invented some more clever way to catch duplicates
 77 *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
 78 *
 79 * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
 80 * When you compare it to RFCs, please, read section SEGMENT ARRIVES
 81 * from the very beginning.
 82 *
 83 * NOTE. With recycling (and later with fin-wait-2) TW bucket
 84 * is _not_ stateless. It means, that strictly speaking we must
 85 * spinlock it. I do not want! Well, probability of misbehaviour
 86 * is ridiculously low and, seems, we could use some mb() tricks
 87 * to avoid misread sequence numbers, states etc.  --ANK
 88 *
 89 * We don't need to initialize tmp_out.sack_ok as we don't use the results
 90 */
 91enum tcp_tw_status
 92tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
 93			   const struct tcphdr *th)
 94{
 95	struct tcp_options_received tmp_opt;
 
 96	struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
 97	bool paws_reject = false;
 98
 99	tmp_opt.saw_tstamp = 0;
100	if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
101		tcp_parse_options(skb, &tmp_opt, 0, NULL);
102
103		if (tmp_opt.saw_tstamp) {
104			tmp_opt.rcv_tsecr	-= tcptw->tw_ts_offset;
105			tmp_opt.ts_recent	= tcptw->tw_ts_recent;
106			tmp_opt.ts_recent_stamp	= tcptw->tw_ts_recent_stamp;
107			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
108		}
109	}
110
111	if (tw->tw_substate == TCP_FIN_WAIT2) {
112		/* Just repeat all the checks of tcp_rcv_state_process() */
113
114		/* Out of window, send ACK */
115		if (paws_reject ||
116		    !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
117				   tcptw->tw_rcv_nxt,
118				   tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
119			return TCP_TW_ACK;
120
121		if (th->rst)
122			goto kill;
123
124		if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
125			goto kill_with_rst;
126
127		/* Dup ACK? */
128		if (!th->ack ||
129		    !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
130		    TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
131			inet_twsk_put(tw);
132			return TCP_TW_SUCCESS;
133		}
134
135		/* New data or FIN. If new data arrive after half-duplex close,
136		 * reset.
137		 */
138		if (!th->fin ||
139		    TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
140kill_with_rst:
141			inet_twsk_deschedule(tw, &tcp_death_row);
142			inet_twsk_put(tw);
143			return TCP_TW_RST;
144		}
145
146		/* FIN arrived, enter true time-wait state. */
147		tw->tw_substate	  = TCP_TIME_WAIT;
148		tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
149		if (tmp_opt.saw_tstamp) {
150			tcptw->tw_ts_recent_stamp = get_seconds();
151			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
152		}
153
154		if (tcp_death_row.sysctl_tw_recycle &&
155		    tcptw->tw_ts_recent_stamp &&
156		    tcp_tw_remember_stamp(tw))
157			inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
158					   TCP_TIMEWAIT_LEN);
159		else
160			inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
161					   TCP_TIMEWAIT_LEN);
162		return TCP_TW_ACK;
163	}
164
165	/*
166	 *	Now real TIME-WAIT state.
167	 *
168	 *	RFC 1122:
169	 *	"When a connection is [...] on TIME-WAIT state [...]
170	 *	[a TCP] MAY accept a new SYN from the remote TCP to
171	 *	reopen the connection directly, if it:
172	 *
173	 *	(1)  assigns its initial sequence number for the new
174	 *	connection to be larger than the largest sequence
175	 *	number it used on the previous connection incarnation,
176	 *	and
177	 *
178	 *	(2)  returns to TIME-WAIT state if the SYN turns out
179	 *	to be an old duplicate".
180	 */
181
182	if (!paws_reject &&
183	    (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
184	     (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
185		/* In window segment, it may be only reset or bare ack. */
186
187		if (th->rst) {
188			/* This is TIME_WAIT assassination, in two flavors.
189			 * Oh well... nobody has a sufficient solution to this
190			 * protocol bug yet.
191			 */
192			if (sysctl_tcp_rfc1337 == 0) {
193kill:
194				inet_twsk_deschedule(tw, &tcp_death_row);
195				inet_twsk_put(tw);
196				return TCP_TW_SUCCESS;
197			}
198		}
199		inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
200				   TCP_TIMEWAIT_LEN);
201
202		if (tmp_opt.saw_tstamp) {
203			tcptw->tw_ts_recent	  = tmp_opt.rcv_tsval;
204			tcptw->tw_ts_recent_stamp = get_seconds();
205		}
206
207		inet_twsk_put(tw);
208		return TCP_TW_SUCCESS;
209	}
210
211	/* Out of window segment.
212
213	   All the segments are ACKed immediately.
214
215	   The only exception is new SYN. We accept it, if it is
216	   not old duplicate and we are not in danger to be killed
217	   by delayed old duplicates. RFC check is that it has
218	   newer sequence number works at rates <40Mbit/sec.
219	   However, if paws works, it is reliable AND even more,
220	   we even may relax silly seq space cutoff.
221
222	   RED-PEN: we violate main RFC requirement, if this SYN will appear
223	   old duplicate (i.e. we receive RST in reply to SYN-ACK),
224	   we must return socket to time-wait state. It is not good,
225	   but not fatal yet.
226	 */
227
228	if (th->syn && !th->rst && !th->ack && !paws_reject &&
229	    (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
230	     (tmp_opt.saw_tstamp &&
231	      (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
232		u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
233		if (isn == 0)
234			isn++;
235		TCP_SKB_CB(skb)->when = isn;
236		return TCP_TW_SYN;
237	}
238
239	if (paws_reject)
240		NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED);
241
242	if (!th->rst) {
243		/* In this case we must reset the TIMEWAIT timer.
244		 *
245		 * If it is ACKless SYN it may be both old duplicate
246		 * and new good SYN with random sequence number <rcv_nxt.
247		 * Do not reschedule in the last case.
248		 */
249		if (paws_reject || th->ack)
250			inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
251					   TCP_TIMEWAIT_LEN);
252
253		/* Send ACK. Note, we do not put the bucket,
254		 * it will be released by caller.
255		 */
256		return TCP_TW_ACK;
257	}
258	inet_twsk_put(tw);
259	return TCP_TW_SUCCESS;
260}
261EXPORT_SYMBOL(tcp_timewait_state_process);
262
263/*
264 * Move a socket to time-wait or dead fin-wait-2 state.
265 */
266void tcp_time_wait(struct sock *sk, int state, int timeo)
267{
268	struct inet_timewait_sock *tw = NULL;
269	const struct inet_connection_sock *icsk = inet_csk(sk);
270	const struct tcp_sock *tp = tcp_sk(sk);
271	bool recycle_ok = false;
272
273	if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
274		recycle_ok = tcp_remember_stamp(sk);
275
276	if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
277		tw = inet_twsk_alloc(sk, state);
278
279	if (tw != NULL) {
280		struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
281		const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
282		struct inet_sock *inet = inet_sk(sk);
283
284		tw->tw_transparent	= inet->transparent;
285		tw->tw_rcv_wscale	= tp->rx_opt.rcv_wscale;
286		tcptw->tw_rcv_nxt	= tp->rcv_nxt;
287		tcptw->tw_snd_nxt	= tp->snd_nxt;
288		tcptw->tw_rcv_wnd	= tcp_receive_window(tp);
289		tcptw->tw_ts_recent	= tp->rx_opt.ts_recent;
290		tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
291		tcptw->tw_ts_offset	= tp->tsoffset;
292
293#if IS_ENABLED(CONFIG_IPV6)
294		if (tw->tw_family == PF_INET6) {
295			struct ipv6_pinfo *np = inet6_sk(sk);
 
296
297			tw->tw_v6_daddr = sk->sk_v6_daddr;
298			tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
299			tw->tw_tclass = np->tclass;
300			tw->tw_flowlabel = np->flow_label >> 12;
301			tw->tw_ipv6only = np->ipv6only;
302		}
303#endif
304
305#ifdef CONFIG_TCP_MD5SIG
306		/*
307		 * The timewait bucket does not have the key DB from the
308		 * sock structure. We just make a quick copy of the
309		 * md5 key being used (if indeed we are using one)
310		 * so the timewait ack generating code has the key.
311		 */
312		do {
313			struct tcp_md5sig_key *key;
314			tcptw->tw_md5_key = NULL;
 
315			key = tp->af_specific->md5_lookup(sk, sk);
316			if (key != NULL) {
317				tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
318				if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool())
 
319					BUG();
320			}
321		} while (0);
322#endif
323
324		/* Linkage updates. */
325		__inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
326
327		/* Get the TIME_WAIT timeout firing. */
328		if (timeo < rto)
329			timeo = rto;
330
331		if (recycle_ok) {
332			tw->tw_timeout = rto;
333		} else {
334			tw->tw_timeout = TCP_TIMEWAIT_LEN;
335			if (state == TCP_TIME_WAIT)
336				timeo = TCP_TIMEWAIT_LEN;
337		}
338
339		inet_twsk_schedule(tw, &tcp_death_row, timeo,
340				   TCP_TIMEWAIT_LEN);
341		inet_twsk_put(tw);
342	} else {
343		/* Sorry, if we're out of memory, just CLOSE this
344		 * socket up.  We've got bigger problems than
345		 * non-graceful socket closings.
346		 */
347		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW);
348	}
349
350	tcp_update_metrics(sk);
351	tcp_done(sk);
352}
353
354void tcp_twsk_destructor(struct sock *sk)
355{
356#ifdef CONFIG_TCP_MD5SIG
357	struct tcp_timewait_sock *twsk = tcp_twsk(sk);
358
359	if (twsk->tw_md5_key)
360		kfree_rcu(twsk->tw_md5_key, rcu);
361#endif
362}
363EXPORT_SYMBOL_GPL(tcp_twsk_destructor);
364
365static inline void TCP_ECN_openreq_child(struct tcp_sock *tp,
366					 struct request_sock *req)
367{
368	tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0;
369}
370
371/* This is not only more efficient than what we used to do, it eliminates
372 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
373 *
374 * Actually, we could lots of memory writes here. tp of listening
375 * socket contains all necessary default parameters.
376 */
377struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
378{
379	struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC);
380
381	if (newsk != NULL) {
382		const struct inet_request_sock *ireq = inet_rsk(req);
383		struct tcp_request_sock *treq = tcp_rsk(req);
384		struct inet_connection_sock *newicsk = inet_csk(newsk);
385		struct tcp_sock *newtp = tcp_sk(newsk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
386
387		/* Now setup tcp_sock */
388		newtp->pred_flags = 0;
389
390		newtp->rcv_wup = newtp->copied_seq =
391		newtp->rcv_nxt = treq->rcv_isn + 1;
392
393		newtp->snd_sml = newtp->snd_una =
394		newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1;
 
395
396		tcp_prequeue_init(newtp);
397		INIT_LIST_HEAD(&newtp->tsq_node);
398
399		tcp_init_wl(newtp, treq->rcv_isn);
400
401		newtp->srtt_us = 0;
402		newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
403		newicsk->icsk_rto = TCP_TIMEOUT_INIT;
404
405		newtp->packets_out = 0;
406		newtp->retrans_out = 0;
407		newtp->sacked_out = 0;
408		newtp->fackets_out = 0;
409		newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
410		tcp_enable_early_retrans(newtp);
411		newtp->tlp_high_seq = 0;
412		newtp->lsndtime = treq->snt_synack;
413		newtp->total_retrans = req->num_retrans;
414
415		/* So many TCP implementations out there (incorrectly) count the
416		 * initial SYN frame in their delayed-ACK and congestion control
417		 * algorithms that we must have the following bandaid to talk
418		 * efficiently to them.  -DaveM
419		 */
420		newtp->snd_cwnd = TCP_INIT_CWND;
421		newtp->snd_cwnd_cnt = 0;
 
422
423		if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops &&
424		    !try_module_get(newicsk->icsk_ca_ops->owner))
425			newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
 
426
427		tcp_set_ca_state(newsk, TCP_CA_Open);
428		tcp_init_xmit_timers(newsk);
429		__skb_queue_head_init(&newtp->out_of_order_queue);
430		newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1;
 
431
432		newtp->rx_opt.saw_tstamp = 0;
433
434		newtp->rx_opt.dsack = 0;
435		newtp->rx_opt.num_sacks = 0;
436
437		newtp->urg_data = 0;
438
439		if (sock_flag(newsk, SOCK_KEEPOPEN))
440			inet_csk_reset_keepalive_timer(newsk,
441						       keepalive_time_when(newtp));
442
443		newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
444		if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
445			if (sysctl_tcp_fack)
446				tcp_enable_fack(newtp);
447		}
448		newtp->window_clamp = req->window_clamp;
449		newtp->rcv_ssthresh = req->rcv_wnd;
450		newtp->rcv_wnd = req->rcv_wnd;
451		newtp->rx_opt.wscale_ok = ireq->wscale_ok;
452		if (newtp->rx_opt.wscale_ok) {
453			newtp->rx_opt.snd_wscale = ireq->snd_wscale;
454			newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
455		} else {
456			newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
457			newtp->window_clamp = min(newtp->window_clamp, 65535U);
458		}
459		newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) <<
460				  newtp->rx_opt.snd_wscale);
461		newtp->max_window = newtp->snd_wnd;
462
463		if (newtp->rx_opt.tstamp_ok) {
464			newtp->rx_opt.ts_recent = req->ts_recent;
465			newtp->rx_opt.ts_recent_stamp = get_seconds();
466			newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
467		} else {
468			newtp->rx_opt.ts_recent_stamp = 0;
469			newtp->tcp_header_len = sizeof(struct tcphdr);
470		}
471		newtp->tsoffset = 0;
472#ifdef CONFIG_TCP_MD5SIG
473		newtp->md5sig_info = NULL;	/*XXX*/
474		if (newtp->af_specific->md5_lookup(sk, newsk))
475			newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
476#endif
477		if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len)
478			newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
479		newtp->rx_opt.mss_clamp = req->mss;
480		TCP_ECN_openreq_child(newtp, req);
481		newtp->fastopen_rsk = NULL;
482		newtp->syn_data_acked = 0;
483
484		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS);
485	}
486	return newsk;
487}
488EXPORT_SYMBOL(tcp_create_openreq_child);
489
490/*
491 * Process an incoming packet for SYN_RECV sockets represented as a
492 * request_sock. Normally sk is the listener socket but for TFO it
493 * points to the child socket.
494 *
495 * XXX (TFO) - The current impl contains a special check for ack
496 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better?
497 *
498 * We don't need to initialize tmp_opt.sack_ok as we don't use the results
499 */
500
501struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb,
502			   struct request_sock *req,
503			   struct request_sock **prev,
504			   bool fastopen)
505{
506	struct tcp_options_received tmp_opt;
 
507	struct sock *child;
508	const struct tcphdr *th = tcp_hdr(skb);
509	__be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
510	bool paws_reject = false;
511
512	BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN));
513
514	tmp_opt.saw_tstamp = 0;
515	if (th->doff > (sizeof(struct tcphdr)>>2)) {
516		tcp_parse_options(skb, &tmp_opt, 0, NULL);
517
518		if (tmp_opt.saw_tstamp) {
519			tmp_opt.ts_recent = req->ts_recent;
520			/* We do not store true stamp, but it is not required,
521			 * it can be estimated (approximately)
522			 * from another data.
523			 */
524			tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout);
525			paws_reject = tcp_paws_reject(&tmp_opt, th->rst);
526		}
527	}
528
529	/* Check for pure retransmitted SYN. */
530	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
531	    flg == TCP_FLAG_SYN &&
532	    !paws_reject) {
533		/*
534		 * RFC793 draws (Incorrectly! It was fixed in RFC1122)
535		 * this case on figure 6 and figure 8, but formal
536		 * protocol description says NOTHING.
537		 * To be more exact, it says that we should send ACK,
538		 * because this segment (at least, if it has no data)
539		 * is out of window.
540		 *
541		 *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
542		 *  describe SYN-RECV state. All the description
543		 *  is wrong, we cannot believe to it and should
544		 *  rely only on common sense and implementation
545		 *  experience.
546		 *
547		 * Enforce "SYN-ACK" according to figure 8, figure 6
548		 * of RFC793, fixed by RFC1122.
549		 *
550		 * Note that even if there is new data in the SYN packet
551		 * they will be thrown away too.
552		 *
553		 * Reset timer after retransmitting SYNACK, similar to
554		 * the idea of fast retransmit in recovery.
555		 */
556		if (!inet_rtx_syn_ack(sk, req))
557			req->expires = min(TCP_TIMEOUT_INIT << req->num_timeout,
558					   TCP_RTO_MAX) + jiffies;
559		return NULL;
560	}
561
562	/* Further reproduces section "SEGMENT ARRIVES"
563	   for state SYN-RECEIVED of RFC793.
564	   It is broken, however, it does not work only
565	   when SYNs are crossed.
566
567	   You would think that SYN crossing is impossible here, since
568	   we should have a SYN_SENT socket (from connect()) on our end,
569	   but this is not true if the crossed SYNs were sent to both
570	   ends by a malicious third party.  We must defend against this,
571	   and to do that we first verify the ACK (as per RFC793, page
572	   36) and reset if it is invalid.  Is this a true full defense?
573	   To convince ourselves, let us consider a way in which the ACK
574	   test can still pass in this 'malicious crossed SYNs' case.
575	   Malicious sender sends identical SYNs (and thus identical sequence
576	   numbers) to both A and B:
577
578		A: gets SYN, seq=7
579		B: gets SYN, seq=7
580
581	   By our good fortune, both A and B select the same initial
582	   send sequence number of seven :-)
583
584		A: sends SYN|ACK, seq=7, ack_seq=8
585		B: sends SYN|ACK, seq=7, ack_seq=8
586
587	   So we are now A eating this SYN|ACK, ACK test passes.  So
588	   does sequence test, SYN is truncated, and thus we consider
589	   it a bare ACK.
590
591	   If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
592	   bare ACK.  Otherwise, we create an established connection.  Both
593	   ends (listening sockets) accept the new incoming connection and try
594	   to talk to each other. 8-)
595
596	   Note: This case is both harmless, and rare.  Possibility is about the
597	   same as us discovering intelligent life on another plant tomorrow.
598
599	   But generally, we should (RFC lies!) to accept ACK
600	   from SYNACK both here and in tcp_rcv_state_process().
601	   tcp_rcv_state_process() does not, hence, we do not too.
602
603	   Note that the case is absolutely generic:
604	   we cannot optimize anything here without
605	   violating protocol. All the checks must be made
606	   before attempt to create socket.
607	 */
608
609	/* RFC793 page 36: "If the connection is in any non-synchronized state ...
610	 *                  and the incoming segment acknowledges something not yet
611	 *                  sent (the segment carries an unacceptable ACK) ...
612	 *                  a reset is sent."
613	 *
614	 * Invalid ACK: reset will be sent by listening socket.
615	 * Note that the ACK validity check for a Fast Open socket is done
616	 * elsewhere and is checked directly against the child socket rather
617	 * than req because user data may have been sent out.
618	 */
619	if ((flg & TCP_FLAG_ACK) && !fastopen &&
620	    (TCP_SKB_CB(skb)->ack_seq !=
621	     tcp_rsk(req)->snt_isn + 1))
622		return sk;
623
624	/* Also, it would be not so bad idea to check rcv_tsecr, which
625	 * is essentially ACK extension and too early or too late values
626	 * should cause reset in unsynchronized states.
627	 */
628
629	/* RFC793: "first check sequence number". */
630
631	if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
632					  tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) {
633		/* Out of window: send ACK and drop. */
634		if (!(flg & TCP_FLAG_RST))
635			req->rsk_ops->send_ack(sk, skb, req);
636		if (paws_reject)
637			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
638		return NULL;
639	}
640
641	/* In sequence, PAWS is OK. */
642
643	if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt))
644		req->ts_recent = tmp_opt.rcv_tsval;
645
646	if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
647		/* Truncate SYN, it is out of window starting
648		   at tcp_rsk(req)->rcv_isn + 1. */
649		flg &= ~TCP_FLAG_SYN;
650	}
651
652	/* RFC793: "second check the RST bit" and
653	 *	   "fourth, check the SYN bit"
654	 */
655	if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
656		TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
657		goto embryonic_reset;
658	}
659
660	/* ACK sequence verified above, just make sure ACK is
661	 * set.  If ACK not set, just silently drop the packet.
662	 *
663	 * XXX (TFO) - if we ever allow "data after SYN", the
664	 * following check needs to be removed.
665	 */
666	if (!(flg & TCP_FLAG_ACK))
667		return NULL;
668
669	/* For Fast Open no more processing is needed (sk is the
670	 * child socket).
671	 */
672	if (fastopen)
673		return sk;
674
675	/* While TCP_DEFER_ACCEPT is active, drop bare ACK. */
676	if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
677	    TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
678		inet_rsk(req)->acked = 1;
679		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP);
680		return NULL;
681	}
 
 
 
 
682
683	/* OK, ACK is valid, create big socket and
684	 * feed this segment to it. It will repeat all
685	 * the tests. THIS SEGMENT MUST MOVE SOCKET TO
686	 * ESTABLISHED STATE. If it will be dropped after
687	 * socket is created, wait for troubles.
688	 */
689	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL);
690	if (child == NULL)
691		goto listen_overflow;
692
693	inet_csk_reqsk_queue_unlink(sk, req, prev);
694	inet_csk_reqsk_queue_removed(sk, req);
695
696	inet_csk_reqsk_queue_add(sk, req, child);
697	return child;
698
699listen_overflow:
700	if (!sysctl_tcp_abort_on_overflow) {
701		inet_rsk(req)->acked = 1;
702		return NULL;
703	}
704
705embryonic_reset:
706	if (!(flg & TCP_FLAG_RST)) {
707		/* Received a bad SYN pkt - for TFO We try not to reset
708		 * the local connection unless it's really necessary to
709		 * avoid becoming vulnerable to outside attack aiming at
710		 * resetting legit local connections.
711		 */
712		req->rsk_ops->send_reset(sk, skb);
713	} else if (fastopen) { /* received a valid RST pkt */
714		reqsk_fastopen_remove(sk, req, true);
715		tcp_reset(sk);
716	}
717	if (!fastopen) {
718		inet_csk_reqsk_queue_drop(sk, req, prev);
719		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS);
720	}
721	return NULL;
722}
723EXPORT_SYMBOL(tcp_check_req);
724
725/*
726 * Queue segment on the new socket if the new socket is active,
727 * otherwise we just shortcircuit this and continue with
728 * the new socket.
729 *
730 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV
731 * when entering. But other states are possible due to a race condition
732 * where after __inet_lookup_established() fails but before the listener
733 * locked is obtained, other packets cause the same connection to
734 * be created.
735 */
736
737int tcp_child_process(struct sock *parent, struct sock *child,
738		      struct sk_buff *skb)
739{
740	int ret = 0;
741	int state = child->sk_state;
742
743	if (!sock_owned_by_user(child)) {
744		ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb),
745					    skb->len);
746		/* Wakeup parent, send SIGIO */
747		if (state == TCP_SYN_RECV && child->sk_state != state)
748			parent->sk_data_ready(parent);
749	} else {
750		/* Alas, it is possible again, because we do lookup
751		 * in main socket hash table and lock on listening
752		 * socket does not protect us more.
753		 */
754		__sk_add_backlog(child, skb);
755	}
756
757	bh_unlock_sock(child);
758	sock_put(child);
759	return ret;
760}
761EXPORT_SYMBOL(tcp_child_process);