14 files changed, 109 insertions(+), 30 deletions(-)

M ip/gre.c
M ip/icmp.c
M ip/ipaux.c
M ip/ipmux.c
M ip/tcp.c
M pc/etherga620.c
M pc/etherwavelan.c
M pc/fns.h
A pc/kpmemmove.c
M pc/mp.c
M port/devkprof.c
M port/portdat.h
M port/qio.c
M port/taslock.c

@@ 94,7 94,7 @@ grestate(Conv *c, char *state, int n)
 static void
 grecreate(Conv *c)
 {
-	c->rq = qopen(64*1024, 0, 0, c);
+	c->rq = qopen(64*1024, 1, 0, c);
 	c->wq = qopen(64*1024, 0, 0, 0);
 }
 

@@ 120,7 120,7 @@ icmpstate(Conv *c, char *state, int n)
 static void
 icmpcreate(Conv *c)
 {
-	c->rq = qopen(64*1024, 0, 0, c);
+	c->rq = qopen(64*1024, 1, 0, c);
 	c->wq = qopen(64*1024, 0, 0, 0);
 }
 

@@ 252,8 252,10 @@ isv4(uchar *ip)
 void
 v4tov6(uchar *v6, uchar *v4)
 {
-	memmove(v6, v4prefix, IPv4off);
-	memmove(v6 + IPv4off, v4, IPv4addrlen);
+	v6[0] = *(ulong)v4prefix;
+	v6[1] = *(ulong)(v4prefix+4);
+	v6[2] = *(ulong)(v4prefix+8);
+	v6[3] = *(ulong)v4;
 }
 
 int

@@ 586,7 586,7 @@ ipmuxcreate(Conv *c)
 {
 	Ipmuxrock *r;
 
-	c->rq = qopen(64*1024, 0, 0, c);
+	c->rq = qopen(64*1024, 1, 0, c);
 	c->wq = qopen(64*1024, 0, 0, 0);
 	r = (Ipmuxrock*)(c->ptcl);
 	r->chain = nil;

@@ 332,10 332,10 @@ tcpstate(Conv *c, char *state, int n)
 	s = (Tcpctl*)(c->ptcl);
 
 	return snprint(state, n,
-		"%s srtt %d mdev %d cwin %d swin %d rwin %d timer.start %d timer.count %d\n",
+		"%s srtt %d mdev %d cwin %d swin %d rwin %d timer.start %d timer.count %d rerecv %d\n",
 		tcpstates[s->state], s->srtt, s->mdev,
 		s->cwind, s->snd.wnd, s->rcv.wnd,
-		s->timer.start, s->timer.count);
+		s->timer.start, s->timer.count, s->rerecv);
 }
 
 static int


@@ 484,7 484,7 @@ tcpacktimer(void *v)
 static void
 tcpcreate(Conv *c)
 {
-	c->rq = qopen(QMAX, 0, tcpacktimer, c);
+	c->rq = qopen(QMAX, -1, tcpacktimer, c);
 	c->wq = qopen(2*QMAX, 0, 0, 0);
 }
 


@@ 770,7 770,7 @@ htontcp(Tcp *tcph, Block *data, Tcphdr *ph, Tcpctl *tcb)
 	}
 	else {
 		dlen = 0;
-		data = allocb(hdrlen + TCP_PKT);
+		data = allocb(hdrlen + TCP_PKT + 64);	/* the 64 pad is to meet mintu's */
 		if(data == nil)
 			return nil;
 		data->wp += hdrlen + TCP_PKT;


@@ 2137,7 2137,6 @@ int
 tcptrim(Tcpctl *tcb, Tcp *seg, Block **bp, ushort *length)
 {
 	ushort len;
-	Block *nbp;
 	uchar accept;
 	int dupcnt, excess;
 


@@ 2197,9 2196,9 @@ tcptrim(Tcpctl *tcb, Tcp *seg, Block **bp, ushort *length)
 	if(excess > 0) {
 		tcb->rerecv += excess;
 		*length -= excess;
-		nbp = copyblock(*bp, *length);
-		freeblist(*bp);
-		*bp = nbp;
+		*bp = trimblock(*bp, 0, *length);
+		if(*bp == nil)
+			panic("presotto is a boofhead");
 		seg->flags &= ~FIN;
 	}
 	return 0;

@@ 228,7 228,7 @@ enum {					/* Host/NIC Interface ring sizes */
 };
 
 enum {
-	NrsrHI		= 128,		/* Fill-level of Rsr (m.b. < Nrsr) */
+	NrsrHI		= 256,		/* (WAS 128) Fill-level of Rsr (m.b. < Nrsr) */
 	NrsrLO		= 64,		/* Level at which to top-up ring */
 	NrjrHI		= 0,		/* Fill-level of Rjr (m.b. < Nrjr) */
 	NrjrLO		= 0,		/* Level at which to top-up ring */


@@ 665,11 665,11 @@ ga620init(Ether* edev)
 	 * These defaults are based on the tuning hints in the Alteon
 	 * Host/NIC Software Interface Definition and example software.
 	 */
-	csr32w(ctlr, Rct, 100);
+	csr32w(ctlr, Rct, 1000);	/* was 100 */
 	csr32w(ctlr, Sct, 0);
 	csr32w(ctlr, St, 100000);
 	csr32w(ctlr, SmcBD, Nsr/4);
-	csr32w(ctlr, RmcBD, 6);
+	csr32w(ctlr, RmcBD, 100);	/* was 6 */
 
 	/*
 	 * Enable DMA Assist Logic.


@@ 993,6 993,7 @@ ga620pnp(Ether* edev)
 	edev->irq = ctlr->pcidev->intl;
 	edev->tbdf = ctlr->pcidev->tbdf;
 	edev->mbps = 1000;
+	edev->oq = qopen(512*1024, 1, 0, 0);
 
 	/*
 	 * Check if the adapter's station address is to be overridden.

@@ 152,6 152,7 @@ enum
 		WCmdMsk		= 0x003f,
 		WCmdAccRd	= 0x0021,
 		WCmdAccWr	= 0x0121,
+		WCmdBusy	= 0x8000,
 		WCmdTxFree	= 0x000b|0x0100,
 	WR_Parm0	= 0x02,
 	WR_Parm1	= 0x04,


@@ 313,7 314,11 @@ w_cmd(Ctlr *ctlr, ushort cmd, ushort arg)
 	int i;
 	int rc;
 
+if(csr_ins(ctlr, WR_Cmd) & WCmdBusy)
+    print("busy\n");
 	csr_outs(ctlr, WR_Parm0, arg);
+	csr_outs(ctlr, WR_Parm1, 0);
+	csr_outs(ctlr, WR_Parm2, 0);
 	csr_outs(ctlr, WR_Cmd, cmd);
 	for (i = 0; i<WTmOut; i++){
 		rc = csr_ins(ctlr, WR_EvSts);


@@ 384,6 389,7 @@ w_outltv(Ctlr* ctlr, Wltv* l)
 {
 	int i,len;
 	ushort *p;
+delay(100);
 
 	if (w_seek(ctlr,l->type,0,1))
 		return;


@@ 400,6 406,7 @@ static void
 ltv_outs(Ctlr* ctlr, int type, ushort val)
 {
 	Wltv l;
+delay(100);
 
 	l.type = type;
 	l.val = val;


@@ 424,6 431,7 @@ ltv_outstr(Ctlr* ctlr, int type, char *val)
 {
 	Wltv l;
 	int len;
+delay(100);
 
 	len = strlen(val);
 	if(len > sizeof(l.s))


@@ 589,10 597,11 @@ w_enable(Ether* ether)
 	w_intdis(ctlr);
 	w_cmd(ctlr, WCmdDis, 0);
 	w_intdis(ctlr);
+delay(100);
 	if(w_cmd(ctlr, WCmdIni, 0))
 		return -1;
-
 	w_intdis(ctlr);
+
 	ltv_outs(ctlr, WTyp_Tick, 8);
 	ltv_outs(ctlr, WTyp_MaxLen, ctlr->maxlen);
 	ltv_outs(ctlr, WTyp_Ptype, ctlr->ptype);


@@ 1159,7 1168,7 @@ ctl(Ether* ether, void* buf, long n)
 			ctlr->keyid = key;
 			kp = &ctlr->key[key-1];
 			kp->len = strlen(cb->f[1]);
-			if (kp->len > WKeyLen)
+//			if (kp->len > WKeyLen)
 				kp->len = WKeyLen;
 			memset(kp->dat, 0, sizeof(kp->dat));
 			strncpy(kp->dat, cb->f[1], kp->len);


@@ 1361,6 1370,7 @@ reset(Ether* ether)
 	}
 
 	w_intdis(ctlr);
+delay(500);
 	if (w_cmd(ctlr,WCmdIni,0)){
 		print("#l%d: init failed\n", ether->ctlrno);
 		goto abort;

@@ 134,3 134,9 @@ ulong	TK2MS(ulong);				/* ticks to milliseconds */
 #define PADDR(a)	((ulong)(a)&~KZERO)
 
 #define	dcflush(a, b)
+
+/*
+ * wrapper around memmove that causes kprof to credit caller for time
+ */
+// extern void	*kpmemmove(void*, void*, long);
+// #define memmove kpmemmove

@@ 0,0 1,21 @@
+#include	"u.h"
+#include	"../port/lib.h"
+#include	"mem.h"
+#include	"dat.h"
+#include	"fns.h"
+#include	"../port/error.h"
+
+#undef memmove
+
+void	*
+kpmemmove(void *s, void *d, long n)
+{
+	void *p;
+	if(up == nil)
+		return memmove(s, d, n);
+
+	up->kppc = getcallerpc(&s);
+	p = memmove(s, d, n);
+	up->kppc = 0;
+	return p;
+}

@@ 462,6 462,8 @@ mpstartap(Apic* apic)
 void
 mpinit(void)
 {
+	int ncpu;
+	char *cp;
 	PCMP *pcmp;
 	uchar *e, *p;
 	Apic *apic, *bpapic;


@@ 562,11 564,22 @@ mpinit(void)
 	/*
 	 * Initialise the application processors.
 	 */
+	if(cp = getconf("*ncpu")){
+		ncpu = strtol(cp, 0, 0);
+		if(ncpu < 1)
+			ncpu = 1;
+	}
+	else
+		ncpu = MaxAPICNO;
 	memmove((void*)APBOOTSTRAP, apbootstrap, sizeof(apbootstrap));
 	for(apic = mpapic; apic <= &mpapic[MaxAPICNO]; apic++){
+		if(ncpu <= 1)
+			break;
 		if((apic->flags & (PcmpBP|PcmpEN)) == PcmpEN
-		&& apic->type == PcmpPROCESSOR)
+		&& apic->type == PcmpPROCESSOR){
 			mpstartap(apic);
+			ncpu--;
+		}
 	}
 
 	/*

@@ 41,7 41,8 @@ _kproftimer(ulong pc)
 	 */
 	if(pc>=(ulong)spllo && pc<=(ulong)spldone)
 		pc = m->splpc;
-
+	if(up && up->kppc)
+		pc = up->kppc;
 	kprof.buf[0] += TK2MS(1);
 	if(kprof.minpc<=pc && pc<kprof.maxpc){
 		pc -= kprof.minpc;

@@ 621,6 621,7 @@ struct Proc
 				 *  that last preempted it
 				 */
 	ulong	qpc;		/* pc calling last blocking qlock */
+	ulong	kppc;		/* kprof calling pc */
 
 	void	*ureg;		/* User registers for notes */
 	void	*dbgreg;	/* User registers for devproc */

@@ 55,6 55,7 @@ enum
 	Qmsg		= (1<<1),	/* message stream */
 	Qclosed		= (1<<2),
 	Qflow		= (1<<3),
+	Qcoalesce	= (1<<4),	/* coallesce packets on read */
 
 	Maxatomic	= 32*1024,
 };


@@ 749,7 750,12 @@ qopen(int limit, int msg, void (*kick)(void*), void *arg)
 	q->limit = q->inilim = limit;
 	q->kick = kick;
 	q->arg = arg;
-	q->state = msg ? Qmsg : 0;
+	q->state = 0;
+	if(msg > 0)
+		q->state |= Qmsg;
+	else if(msg < 0)
+		q->state |= Qcoalesce;
+	
 	q->state |= Qstarve;
 	q->eof = 0;
 	q->noblock = 0;


@@ 865,15 871,34 @@ long
 qread(Queue *q, void *vp, int len)
 {
 	Block *b;
+	int m;
+	uchar *p;
+	uchar *e;
 
-	b = qbread(q, len);
-	if(b == 0)
-		return 0;
+	p = vp;
 
-	len = BLEN(b);
-	memmove(vp, b->rp, len);
-	freeb(b);
-	return len;
+	if((q->state & Qcoalesce) == 0){
+		b = qbread(q, len);
+		if(b == 0)
+			return 0;
+
+		m = BLEN(b);
+		memmove(p, b->rp, m);
+		freeb(b);
+		return m;
+	}
+
+	for(e = p + len; p < e; p += m){
+		b = qbread(q, e-p);
+		if(b == 0)
+			return 0;
+
+		m = BLEN(b);
+		memmove(p, b->rp, m);
+		freeb(b);
+	}
+
+	return p-(uchar*)vp;
 }
 
 static int

@@ 170,7 170,7 @@ unlock(Lock *l)
 		print("unlock of ilock: pc %lux, held by %lux\n", getcallerpc(&l), l->pc);
 	l->pc = 0;
 	l->key = 0;
-	coherence();
+	//coherence();
 }
 
 void


@@ 186,7 186,7 @@ iunlock(Lock *l)
 	sr = l->sr;
 	l->pc = 0;
 	l->key = 0;
-	coherence();
+	//coherence();
 
 	m->splpc = getcallerpc(&l);
 	splxpc(sr);