tcp.c

这是一个同样来自贝尔实验室的和UNIX有着渊源的操作系统, 其简洁的设计和实现易于我们学习和理解

		seg->ack = 0;
	}
	else {
		rflags |= ACK;
		seg->ack = seg->seq;
		seg->seq = 0;
		if(seg->flags & SYN)
			seg->ack++;
		seg->ack += length;
		if(seg->flags & FIN)
			seg->ack++;
	}
	seg->flags = rflags;
	seg->wnd = 0;
	seg->urg = 0;
	seg->mss = 0;
	seg->ws = 0;
	switch(version) {
	case V4:
		hbp = htontcp4(seg, nil, &ph4, nil);
		if(hbp == nil)
			return;
		ipoput4(tcp->f, hbp, 0, MAXTTL, DFLTTOS, nil);
		break;
	case V6:
		hbp = htontcp6(seg, nil, &ph6, nil);
		if(hbp == nil)
			return;
		ipoput6(tcp->f, hbp, 0, MAXTTL, DFLTTOS, nil);
		break;
	default:
		panic("sndrst2: version %d", version);
	}
}

/*
 *  send a reset to the remote side and close the conversation
 *  called with s qlocked
 */
char*
tcphangup(Conv *s)
{
	Tcp seg;
	Tcpctl *tcb;
	Block *hbp;

	tcb = (Tcpctl*)s->ptcl;
	if(waserror())
		return commonerror();
	if(s->raddr != 0) {
		if(!waserror()){
			seg.flags = RST | ACK;
			seg.ack = tcb->rcv.nxt;
			tcb->rcv.una = 0;
			seg.seq = tcb->snd.ptr;
			seg.wnd = 0;
			seg.urg = 0;
			seg.mss = 0;
			seg.ws = 0;
			switch(s->ipversion) {
			case V4:
				tcb->protohdr.tcp4hdr.vihl = IP_VER4;
				hbp = htontcp4(&seg, nil, &tcb->protohdr.tcp4hdr, tcb);
				ipoput4(s->p->f, hbp, 0, s->ttl, s->tos, s);
				break;
			case V6:
				tcb->protohdr.tcp6hdr.vcf[0] = IP_VER6;
				hbp = htontcp6(&seg, nil, &tcb->protohdr.tcp6hdr, tcb);
				ipoput6(s->p->f, hbp, 0, s->ttl, s->tos, s);
				break;
			default:
				panic("tcphangup: version %d", s->ipversion);
			}
			poperror();
		}
	}
	localclose(s, nil);
	poperror();
	return nil;
}

/*
 *  (re)send a SYN ACK
 */
int
sndsynack(Proto *tcp, Limbo *lp)
{
	Block *hbp;
	Tcp4hdr ph4;
	Tcp6hdr ph6;
	Tcp seg;
	int scale;

	/* make pseudo header */
	switch(lp->version) {
	case V4:
		memset(&ph4, 0, sizeof(ph4));
		ph4.vihl = IP_VER4;
		v6tov4(ph4.tcpsrc, lp->laddr);
		v6tov4(ph4.tcpdst, lp->raddr);
		ph4.proto = IP_TCPPROTO;
		hnputs(ph4.tcplen, TCP4_HDRSIZE);
		hnputs(ph4.tcpsport, lp->lport);
		hnputs(ph4.tcpdport, lp->rport);
		break;
	case V6:
		memset(&ph6, 0, sizeof(ph6));
		ph6.vcf[0] = IP_VER6;
		ipmove(ph6.tcpsrc, lp->laddr);
		ipmove(ph6.tcpdst, lp->raddr);
		ph6.proto = IP_TCPPROTO;
		hnputs(ph6.ploadlen, TCP6_HDRSIZE);
		hnputs(ph6.tcpsport, lp->lport);
		hnputs(ph6.tcpdport, lp->rport);
		break;
	default:
		panic("sndrst: version %d", lp->version);
	}

	seg.seq = lp->iss;
	seg.ack = lp->irs+1;
	seg.flags = SYN|ACK;
	seg.urg = 0;
	seg.mss = tcpmtu(tcp, lp->laddr, lp->version, &scale);
	seg.wnd = QMAX;

	/* if the other side set scale, we should too */
	if(lp->rcvscale){
		seg.ws = scale;
		lp->sndscale = scale;
	} else {
		seg.ws = 0;
		lp->sndscale = 0;
	}

	switch(lp->version) {
	case V4:
		hbp = htontcp4(&seg, nil, &ph4, nil);
		if(hbp == nil)
			return -1;
		ipoput4(tcp->f, hbp, 0, MAXTTL, DFLTTOS, nil);
		break;
	case V6:
		hbp = htontcp6(&seg, nil, &ph6, nil);
		if(hbp == nil)
			return -1;
		ipoput6(tcp->f, hbp, 0, MAXTTL, DFLTTOS, nil);
		break;
	default:
		panic("sndsnack: version %d", lp->version);
	}
	lp->lastsend = NOW;
	return 0;
}

#define hashipa(a, p) ( ( (a)[IPaddrlen-2] + (a)[IPaddrlen-1] + p )&LHTMASK )

/*
 *  put a call into limbo and respond with a SYN ACK
 *
 *  called with proto locked
 */
static void
limbo(Conv *s, uchar *source, uchar *dest, Tcp *seg, int version)
{
	Limbo *lp, **l;
	Tcppriv *tpriv;
	int h;

	tpriv = s->p->priv;
	h = hashipa(source, seg->source);

	for(l = &tpriv->lht[h]; *l != nil; l = &lp->next){
		lp = *l;
		if(lp->lport != seg->dest || lp->rport != seg->source || lp->version != version)
			continue;
		if(ipcmp(lp->raddr, source) != 0)
			continue;
		if(ipcmp(lp->laddr, dest) != 0)
			continue;

		/* each new SYN restarts the retransmits */
		lp->irs = seg->seq;
		break;
	}
	lp = *l;
	if(lp == nil){
		if(tpriv->nlimbo >= Maxlimbo && tpriv->lht[h]){
			lp = tpriv->lht[h];
			tpriv->lht[h] = lp->next;
			lp->next = nil;
		} else {
			lp = malloc(sizeof(*lp));
			if(lp == nil)
				return;
			tpriv->nlimbo++;
		}
		*l = lp;
		lp->version = version;
		ipmove(lp->laddr, dest);
		ipmove(lp->raddr, source);
		lp->lport = seg->dest;
		lp->rport = seg->source;
		lp->mss = seg->mss;
		lp->rcvscale = seg->ws;
		lp->irs = seg->seq;
		lp->iss = (nrand(1<<16)<<16)|nrand(1<<16);
	}

	if(sndsynack(s->p, lp) < 0){
		*l = lp->next;
		tpriv->nlimbo--;
		free(lp);
	}
}

/*
 *  resend SYN ACK's once every SYNACK_RXTIMER ms.
 */
static void
limborexmit(Proto *tcp)
{
	Tcppriv *tpriv;
	Limbo **l, *lp;
	int h;
	int seen;
	ulong now;

	tpriv = tcp->priv;

	if(!canqlock(tcp))
		return;
	seen = 0;
	now = NOW;
	for(h = 0; h < NLHT && seen < tpriv->nlimbo; h++){
		for(l = &tpriv->lht[h]; *l != nil && seen < tpriv->nlimbo; ){
			lp = *l;
			seen++;
			if(now - lp->lastsend < (lp->rexmits+1)*SYNACK_RXTIMER)
				continue;

			/* time it out after 1 second */
			if(++(lp->rexmits) > 5){
				tpriv->nlimbo--;
				*l = lp->next;
				free(lp);
				continue;
			}

			/* if we're being attacked, don't bother resending SYN ACK's */
			if(tpriv->nlimbo > 100)
				continue;

			if(sndsynack(tcp, lp) < 0){
				tpriv->nlimbo--;
				*l = lp->next;
				free(lp);
				continue;
			}

			l = &lp->next;
		}
	}
	qunlock(tcp);
}

/*
 *  lookup call in limbo.  if found, throw it out.
 *
 *  called with proto locked
 */
static void
limborst(Conv *s, Tcp *segp, uchar *src, uchar *dst, uchar version)
{
	Limbo *lp, **l;
	int h;
	Tcppriv *tpriv;

	tpriv = s->p->priv;

	/* find a call in limbo */
	h = hashipa(src, segp->source);
	for(l = &tpriv->lht[h]; *l != nil; l = &lp->next){
		lp = *l;
		if(lp->lport != segp->dest || lp->rport != segp->source || lp->version != version)
			continue;
		if(ipcmp(lp->laddr, dst) != 0)
			continue;
		if(ipcmp(lp->raddr, src) != 0)
			continue;

		/* RST can only follow the SYN */
		if(segp->seq == lp->irs+1){
			tpriv->nlimbo--;
			*l = lp->next;
			free(lp);
		}
		break;
	}
}

/*
 *  come here when we finally get an ACK to our SYN-ACK.
 *  lookup call in limbo.  if found, create a new conversation
 *
 *  called with proto locked
 */
static Conv*
tcpincoming(Conv *s, Tcp *segp, uchar *src, uchar *dst, uchar version)
{
	Conv *new;
	Tcpctl *tcb;
	Tcppriv *tpriv;
	Tcp4hdr *h4;
	Tcp6hdr *h6;
	Limbo *lp, **l;
	int h;

	/* unless it's just an ack, it can't be someone coming out of limbo */
	if((segp->flags & SYN) || (segp->flags & ACK) == 0)
		return nil;

	tpriv = s->p->priv;

	/* find a call in limbo */
	h = hashipa(src, segp->source);
	for(l = &tpriv->lht[h]; (lp = *l) != nil; l = &lp->next){
		netlog(s->p->f, Logtcp, "tcpincoming s %I,%ux/%I,%ux d %I,%ux/%I,%ux v %d/%d",
			src, segp->source, lp->raddr, lp->rport,
			dst, segp->dest, lp->laddr, lp->lport,
			version, lp->version
 		);

		if(lp->lport != segp->dest || lp->rport != segp->source || lp->version != version)
			continue;
		if(ipcmp(lp->laddr, dst) != 0)
			continue;
		if(ipcmp(lp->raddr, src) != 0)
			continue;

		/* we're assuming no data with the initial SYN */
		if(segp->seq != lp->irs+1 || segp->ack != lp->iss+1){
			netlog(s->p->f, Logtcp, "tcpincoming s %lux/%lux a %lux %lux",
				segp->seq, lp->irs+1, segp->ack, lp->iss+1);
			lp = nil;
		} else {
			tpriv->nlimbo--;
			*l = lp->next;
		}
		break;
	}
	if(lp == nil)
		return nil;

	new = Fsnewcall(s, src, segp->source, dst, segp->dest, version);
	if(new == nil)
		return nil;

	memmove(new->ptcl, s->ptcl, sizeof(Tcpctl));
	tcb = (Tcpctl*)new->ptcl;
	tcb->flags &= ~CLONE;
	tcb->timer.arg = new;
	tcb->timer.state = TcptimerOFF;
	tcb->acktimer.arg = new;
	tcb->acktimer.state = TcptimerOFF;
	tcb->katimer.arg = new;
	tcb->katimer.state = TcptimerOFF;
	tcb->rtt_timer.arg = new;
	tcb->rtt_timer.state = TcptimerOFF;

	tcb->irs = lp->irs;
	tcb->rcv.nxt = tcb->irs+1;
	tcb->rcv.urg = tcb->rcv.nxt;

	tcb->iss = lp->iss;
	tcb->rttseq = tcb->iss;
	tcb->snd.wl2 = tcb->iss;
	tcb->snd.una = tcb->iss+1;
	tcb->snd.ptr = tcb->iss+1;
	tcb->snd.nxt = tcb->iss+1;
	tcb->flgcnt = 0;
	tcb->flags |= SYNACK;

	/* our sending max segment size cannot be bigger than what he asked for */
	if(lp->mss != 0 && lp->mss < tcb->mss)
		tcb->mss = lp->mss;

	/* window scaling */
	tcpsetscale(new, tcb, lp->rcvscale, lp->sndscale);

	/* the congestion window always starts out as a single segment */
	tcb->snd.wnd = segp->wnd;
	tcb->cwind = tcb->mss;

	/* set initial round trip time */
	tcb->sndsyntime = lp->lastsend+lp->rexmits*SYNACK_RXTIMER;
	tcpsynackrtt(new);

	free(lp);

	/* set up proto header */
	switch(version){
	case V4:
		h4 = &tcb->protohdr.tcp4hdr;
		memset(h4, 0, sizeof(*h4));
		h4->proto = IP_TCPPROTO;
		hnputs(h4->tcpsport, new->lport);
		hnputs(h4->tcpdport, new->rport);
		v6tov4(h4->tcpsrc, dst);
		v6tov4(h4->tcpdst, src);
		break;
	case V6:
		h6 = &tcb->protohdr.tcp6hdr;
		memset(h6, 0, sizeof(*h6));
		h6->proto = IP_TCPPROTO;
		hnputs(h6->tcpsport, new->lport);
		hnputs(h6->tcpdport, new->rport);
		ipmove(h6->tcpsrc, dst);
		ipmove(h6->tcpdst, src);
		break;
	default:
		panic("tcpincoming: version %d", new->ipversion);
	}

	tcpsetstate(new, Established);

	iphtadd(&tpriv->ht, new);

	return new;
}

int
seq_within(ulong x, ulong low, ulong high)
{
	if(low <= high){
		if(low <= x && x <= high)
			return 1;
	}
	else {
		if(x >= low || x <= high)
			return 1;
	}
	return 0;
}

int
seq_lt(ulong x, ulong y)
{
	return (int)(x-y) < 0;
}

int
seq_le(ulong x, ulong y)
{
	return (int)(x-y) <= 0;
}

int
seq_gt(ulong x, ulong y)
{
	return (int)(x-y) > 0;
}

int
seq_ge(ulong x, ulong y)
{
	return (int)(x-y) >= 0;
}

/*
 *  use the time between the first SYN and it's ack as the
 *  initial round trip time
 */
void
tcpsynackrtt(Conv *s)
{
	Tcpctl *tcb;
	int delta;
	Tcppriv *tpriv;

	tcb = (Tcpctl*)s->ptcl;
	tpriv = s->p->priv;

	delta = NOW - tcb->sndsyntime;
	tcb->srtt = delta<<LOGAGAIN;
	tcb->mdev = delta<<LOGDGAIN;

	/* halt round trip timer */
	tcphalt(tpriv, &tcb->rtt_timer);
}

void
update(Conv *s, Tcp *seg)
{
	int rtt, delta;
	Tcpctl *tcb;
	ulong acked;
	ulong expand;
	Tcppriv *tpriv;

	tpriv = s->p->priv;
	tcb = (Tcpctl*)s->ptcl;

	/* if everything has been acked, force output(?) */
	if(seq_gt(seg->ack, tcb->snd.nxt)) {
		tcb->flags |= FORCE;
		return;
	}

	/* added by Dong Lin for fast retransmission */
	if(seg->ack == tcb->snd.una
	&& tcb->snd.una != tcb->snd.nxt
	&& seg->len == 0
	&& seg->wnd == tcb->snd.wnd) {

		/* this is a pure ack w/o window update */
		netlog(s->p->f, Logtcprxmt, "dupack %lud ack %lud sndwnd %d advwin %d\n",
			tcb->snd.dupacks, seg->ack, tcb->snd.wnd, seg->wnd);

		if(++tcb->snd.dupacks == TCPREXMTTHRESH) {
			/*
			 *  tahoe tcp rxt the packet, half sshthresh,
 			 *  and set cwnd to one packet
			 */
			tcb->snd.recovery = 1;
			tcb->snd.rxt = tcb->snd.nxt;
			netlog(s->p->f, Logtcprxmt, "fast rxt %lud, nxt %lud\n", tcb->snd.una, tcb->snd.nxt);
			tcprxmit(s);
		} else {
			/* do reno tcp here. */
		}
	}

	/*
	 *  update window
	 */
	if(seq_gt(seg->ack, tcb->snd.wl2)
	|| (tcb->snd.wl2 == seg->ack && seg->wnd > tcb->snd.wnd)){
		tcb->snd.wnd = seg->wnd;
		tcb->snd.wl2 = seg->ack;
	}

	if(!seq_gt(seg->ack, tcb->snd.una)){
		/*
		 *  don't let us hangup if sending into a closed window and
		 *  we're still getting acks
		 */
		if((tcb->flags&RETRAN) && tcb->snd.wnd == 0){
			tcb->backedoff = MAXBACKMS/4;
		}
		return;
	}

	/*
	 *  any positive ack turns off fast rxt,
	 *  (should we do new-reno on partial acks?)
	 */
	if(!tcb->snd.recovery || seq_ge(seg->ack, tcb->snd.rxt)) {
		tcb->snd.dupacks = 0;
		tcb->snd.recovery = 0;
	} else
		netlog(s->p->f, Logtcp, "rxt next %lud, cwin %ud\n", seg->ack, tcb->cwind);

	/* Compute the new send window size */
	acked = seg->ack - tcb->snd.una;

	/* avoid slow start and timers for SYN acks */
	if((tcb->flags & SYNACK) == 0) {
		tcb->flags |= SYNACK;
		acked--;
		tcb->flgcnt--;
		goto done;
	}

	/* slow start as long as we're not recovering from lost packets */
	if(tcb->cwind < tcb->snd.wnd && !tcb->snd.recovery) {
		if(tcb->cwind < tcb->ssthresh) {
			expand = tcb->mss;
			if(acked < expand)
				expand = acked;
		}
		else
			expand = ((int)tcb->mss * tcb->mss) / tcb->cwind;

		if(tcb->cwind + expand < tcb->cwind)
			expand = tcb->snd.wnd - tcb->cwind;
		if(tcb->cwind + expand > tcb->snd.wnd)
			expand = tcb->snd.wnd - tcb->cwind;
		tcb->cwind += expand;
	}

	/* Adjust the timers according to the round trip time */
	if(tcb->rtt_timer.state == TcptimerON && seq_ge(seg->ack, tcb->rttseq)) {
		tcphalt(tpriv, &tcb->rtt_timer);
		if((tcb->flags&RETRAN) == 0) {
			tcb->backoff = 0;
			tcb->backedoff = 0;
			rtt = tcb->rtt_timer.start - tcb->rtt_timer.count;
			if(rtt == 0)
				rtt = 1;	/* otherwise all close systems will rexmit in 0 time */
			rtt *= MSPTICK;
			if(tcb->srtt == 0) {
				tcb->srtt = rtt << LOGAGAIN;
				tcb->mdev = rtt << LOGDGAIN;
			} else {
				delta = rtt - (tcb->srtt>>LOGAGAIN);
				tcb->srtt += delta;
				if(tcb->srtt <= 0)
					tcb->srtt = 1;

				delta = abs(delta) - (tcb->mdev>>LOGDGAIN);
				tcb->mdev += delta;
				if(tcb->mdev <= 0)
					tcb->mdev = 1;
			}
			tcpsettimer(tcb);
		}
	}

done:
	if(qdiscard(s->wq, acked) < acked)
		tcb->flgcnt--;

	tcb->snd.una = seg->ack;
	if(seq_gt(seg->ack, tcb->snd.urg))
		tcb->snd.urg = seg->ack;

	if(tcb->snd.una != tcb->snd.nxt)
		tcpgo(tpriv, &tcb->timer);
	else
		tcphalt(tpriv, &tcb->timer);

	if(seq_lt(tcb->snd.ptr, tcb->snd.una))
		tcb->snd.ptr = tcb->snd.una;