some support for the second interface.  it doesn't work, but at least
we see it.

Reference: /n/atom/patch/applied/duali40
Date: Thu Jun 12 04:53:08 CES 2014
Signed-off-by: quanstro@quanstro.net

--- /sys/src/nix/k10/etheri40.c	Thu Jun 12 04:52:06 2014
+++ /sys/src/nix/k10/etheri40.c	Thu Jun 12 04:52:09 2014
@@ -487,6 +487,8 @@
 	Mb1	= 1<<6,	/* must be 1 */
 	Rs	= 1<<5,
 	Teop	= 1<<4,
+	Txdesc	= 0<<0,
+	Nopdesc	= 1<<0,
 
 	/* Td status */
 	Tdd	= 0xf<<0,
@@ -512,7 +514,7 @@
 	Hmcpgsz		= 4096,
 	Hmcalign	= 4096,
 	Fpdeperpage	= Hmcpgsz/sizeof(FPDE),
-	Fpgshift		= 9,
+	Fpgshift		= 12,
 
 	Fpdevalid	= 1<<0,
 };
@@ -548,7 +550,7 @@
 	Block	**rb;
 	uint	rdt, rdfree;
 	Td	*tdba;
-	uint	tdh, tdt;
+	uint	tdh, tdt, tdfree;
 	Block	**tb;
 	uchar	ra[Eaddrlen];
 	uvlong	zstats[nelem(stattab)];
@@ -556,8 +558,10 @@
 	uchar	*rxbuf;
 	int	type;
 	uint	speeds[5];
-	uint	nobufs;
-/**/	uvlong irqcnt[4];
+	uint	phy;
+	uint	nobufs, txfull;
+	uvlong	irqcnt[4];
+	char	ename[16];
 };
 
 struct Rbpool {
@@ -575,6 +579,8 @@
 	union {
 		struct {
 			Block	*x;
+			uvlong	fasttick;
+			uvlong	slowtick;
 			uint	nfast;
 			uint	nslow;
 		};
@@ -585,7 +591,7 @@
 /* tweakable parameters */
 enum{
 	Nrd	= 256,		/* 2ⁿ */
-	Ntd	= 256,		/* 2ⁿ */
+	Ntd	= 2*256,		/* 2ⁿ */
 	Nrb	= 2048,
 	Nctlr	= 8,
 	Rbalign	= 8,
@@ -642,7 +648,24 @@
 	return cttab[c->type].name;
 }
 
-int
+static char*
+ename(Ether *e)
+{
+	Ctlr *c;
+
+	c = e->ctlr;
+	if(c->ename[0] == 0)
+		snprint(c->ename, sizeof c->ename, "#l%d: %s:pf %d",
+			e->ctlrno, cname(c), c->pf);
+	return c->ename;
+}
+
+static void
+readclear(u32int)
+{
+}
+
+static int
 seidtovsiid(int i)
 {
 	return i-512;
@@ -675,6 +698,7 @@
 	uint i;
 	uvlong v, *stats;
 	Ctlr *c;
+	Rbpool *r;
 
 	c = e->ctlr;
 	p = s = malloc(READSTR + nelem(stattab)*sizeof(uvlong));
@@ -692,12 +716,20 @@
 		if((v = stats[i] - c->zvstats[i])>0)
 			p = seprint(p, q, "%.10s	%llud\n", vstattab[i].name, v);
 	p = seprint(p, q, "type: %s\n", cttab[c->type].name);
+	r = c->pool;
+	p = seprint(p, q, "pool: fast %d %llud slow %d %lld starve %d wake %d on %d\n",
+		r->nfast, r->fasttick,
+		r->nslow, r->slowtick,
+		r->nstarve, r->nwakey, r->starve);
 	p = seprint(p, q, "speeds:");
 	for(i = 0; i < nelem(c->speeds); i++)
 		p = seprint(p, q, " %d:%d", speeds[i], c->speeds[i]);
 	p = seprint(p, q, "\n");
-	p = seprint(p, q, "irqcnt: a %lld r %lld t %lld s %lld\n", c->irqcnt[Cnta], c->irqcnt[Cntr], c->irqcnt[Cntt], c->irqcnt[Cnts]);
+	p = seprint(p, q, "phy: %.2ux\n", c->phy);
+	p = seprint(p, q, "irqcnt: a %lld r %lld t %lld s %lld\n",
+		c->irqcnt[Cnta], c->irqcnt[Cntr], c->irqcnt[Cntt], c->irqcnt[Cnts]);
 	seprint(p, q, "nobufs: %ud\n", c->nobufs);
+	seprint(p, q, "txfull: %ud\n", c->txfull);
 	n = readstr(offset, a, n, s);
 	free(s);
 
@@ -764,8 +796,9 @@
 	e->link = link;
 	e->mbps = speeds[i];
 	c->speeds[i]++;
+	c->phy = b[0];
 
-	print("link %d; mbps %d; an %d; phy %.2ux\n", e->link, e->mbps, b[3] & 1<<0, b[0]);
+//	print("link %d; mbps %d; an %d; phy %.2ux\n", e->link, e->mbps, b[3] & 1<<0, b[0]);
 	return 0;
 }
 
@@ -779,6 +812,18 @@
 }
 
 static void
+txim(Ctlr *c, int on)
+{
+	c->reg[Itxqctl + 0] = Itxidx*Iqirqidx | Iqend | on*Iqenable;			/* tx: ics=0x02 */
+}
+
+static void
+rxim(Ctlr *c, int on)
+{
+	c->reg[Irxqctl + 0] = Irxidx*Iqirqidx | 0*Iqnxtqidx | Iqnxtt | on*Iqenable;	/* tx: ics=0x04 */
+}
+
+static void
 qrx(Ether *e, Ctlr *c, Ad *a)
 {
 	switch(a->op){
@@ -815,10 +860,28 @@
 	c->aim = 0;
 
 	if(oldaim & (Eccerr | Malice | Grst | Pcierr)){
-		print("%s: err %.8ux; we're doomed\n", cname(c), oldaim & Errint);
+		print("%s: err %.8ux; we're doomed\n", ename(e), oldaim & Errint);
 	}
 	if(oldaim & Errint){
-		print("hmerr %.8ux %.8ux\n", c->reg[Hmerr+c->pf], c->reg[Hmerrd+c->pf]);
+		/*
+		 * 0:4	pf
+		 * 8:11	error
+		 *	0	private memory fn invalid
+		 *	1	invalid lan queue idx
+		 *	2	object index larger than glhmc_*cnt
+		 *	3	address beyond pf's segment limits
+		 *	4	invalid segment descriptor
+		 *	5	sd too small
+		 *	6	page descriptor invalid
+		 *	7	unsupported pcie read
+		 *	8	pflan_qalloc_pmat or pf_vt_pfalloc_pmat unset
+		 *	9	invalid object type
+		 *	10	invalid fcoe idx
+		 * 16:20		object type	(tx=0, rx=1)
+		 * 31		error detected
+		 */
+		print("%s: hmcerr %.8ux %.8ux\n", ename(e),
+			c->reg[Hmerr+c->pf], c->reg[Hmerrd+c->pf]);
 	}
 	/*
 	 * fixme: it's not necessary to process link status changes both
@@ -830,7 +893,7 @@
 		end = c->reg[Arqh+ pf];
 		for(h = c->arh; h != end; h = NEXT(h, Narq)){
 			a = c->arq + h;
-			print("%s: cmd slot %d op %#ux\n", cname(c), h, a->op);
+			print("%s: cmd slot %d op %#ux\n", ename(e), h, a->op);
 			qrx(e, c, a);
 			adinitrx(c, h);
 			c->reg[Arqt + pf] = h;
@@ -876,12 +939,15 @@
 rballoc(Rbpool *p)
 {
 	Block *b;
+	uvlong t;
 
+	t = -rdtsc();
 	for(;;){
 		if((b = p->x) != nil){
 			p->nfast++;
 			p->x = b->next;
 			b->next = nil;
+			p->fasttick += t+rdtsc();
 			return b;
 		}
 
@@ -895,6 +961,7 @@
 			return nil;
 		}
 		p->nslow++;
+		p->slowtick += t+rdtsc();
 		iunlock(p);
 		p->x = b;
 	}
@@ -1046,19 +1113,66 @@
 	return ((Ctlr*)c)->tim != 0;
 }
 
+static int
+·cleanup(Ctlr *c, int tdh)
+{
+	Block *b;
+	uint m, n;
+
+	m = c->ntd-1;
+	while(c->nrd - c->tdfree > 16 && (c->tdba[n = Next(tdh, m)].cmd&Tdd) == Tdd){
+		tdh = n;
+		b = c->tb[tdh];
+		c->tb[tdh] = nil;
+		freeb(b);
+	//	c->tdba[tdh].cmd = 0;
+		c->tdfree++;
+	}
+	return tdh;
+}
+
 static void
 tproc(void *v)
 {
-	Ether *e;
+	uint m, w;
+	Block *b;
 	Ctlr *c;
+	Ether *e;
+	Td *t;
 
 	e = v;
 	c = e->ctlr;
-loop:
-	sleep(&c->trendez, tim, c);		/* transmit kicks us */
-	c->tim = 0;
-	transmit(e);
-	goto loop;
+	m = c->ntd-1;
+	w = 0;
+	for(;;){
+		c->tdh = ·cleanup(c, c->tdh);
+		if(Next(c->tdt, m) == c->tdh){
+			c->txfull++;
+/**/			print("%s: tproc full\n", ename(e));
+			txim(c, 1);
+			sleep(&c->trendez, tim, c);
+			txim(c, 0);
+/**/			print("%s: tproc tim %.8ux\n", ename(e), c->tim);
+			c->tim = 0;
+			continue;
+		}
+		if(w > 8 || qlen(e->oq) == 0){
+			c->reg[Txtail + c->tqno] = c->tdt;
+			w = 0;
+		}
+		b = qbread(e->oq, 100000);
+		t = c->tdba+c->tdt;
+		t->addr = Pciwaddr(b->rp);
+		t->length = BLEN(b)<<2;
+		t->offset = (6+6+2)/2;	// cksum offload # header words up to etype
+		t->cmd = Rs|Mb1|Teop|Txdesc;
+		sfence();
+		c->tb[c->tdt] = b;
+		c->tdt = Next(c->tdt, m);
+		c->reg[Txtail + c->tqno] = c->tdt;
+		c->tdfree--;
+		w++;
+	}
 }
 
 static int
@@ -1072,7 +1186,7 @@
 	m = c->nrd-1;
 	i = 0;
 	p = c->pool;
-	for(rdt = c->rdt; Next(rdt, m) != rdh; rdt = Next(rdt, m)){
+	for(rdt = c->rdt; c->nrd-c->rdfree >= 16 && Next(rdt, m) != rdh; rdt = Next(rdt, m)){
 		r = c->rdba+rdt;
 		while((b = rballoc(c->pool)) == nil){
 			c->nobufs++;
@@ -1117,26 +1231,28 @@
 	c = e->ctlr;
 	m = c->nrd-1;
 	rdh = 0;
+	im(c, Irx);
 loop:
 	replenish(c, rdh, 1);
-	im(c, Irx);
+	rxim(c, 1);
+im(c, Irx);	/* this isn't right; still big delays. */
 	sleep(&c->rrendez, rim, c);
+//	rxim(c, 0);
 loop1:
 	c->rim = 0;
-	if(c->nrd-c->rdfree >= 16)
+	if(c->nrd-c->rdfree >= 24)
 	if(replenish(c, rdh, 0) == -1)
 		goto loop;
 	r = c->rdba+rdh;
 	if(!(r->status&Rdd))
 		goto loop;
 	b = c->rb[rdh];
-	c->rb[rdh] = 0;
+	c->rb[rdh] = nil;
 	b->wp += r->length>>6;
 	b->lim = b->wp;		/* lie like a dog */
-	b->flag |= Bipck;	/* always */
 	if(r->status&l3l4p)
 		b->flag |= Bipck|Btcpck|Budpck;
-	r->status = 0;
+//	r->status = 0;
 	etheriq(e, b, 1);
 	c->rdfree--;
 	rdh = Next(rdh, m);
@@ -1208,9 +1324,9 @@
 	c = e->ctlr;
 	rv = setmc(c, on, c->vsiseid, ea);
 	if(rv != 0){
-		snprint(buf, sizeof buf, "%s: multicast: error %d", cname(c), rv);
+		snprint(buf, sizeof buf, "%s: multicast: error %d", ename(e), rv);
 		if(rv < nelem(errtab))
-			snprint(buf, sizeof buf, "%s: multicast: %s", cname(c), errtab[rv]);
+			snprint(buf, sizeof buf, "%s: multicast: %s", ename(e), errtab[rv]);
 		error(buf);
 	}
 }
@@ -1218,24 +1334,23 @@
 static int
 detach(Ctlr *c)
 {
-	int i, pf, wait;
+	int i, wait;
 
-	pf = c->pf;
-	wait = (c->reg[Rstctl + pf] & 0x3f)*100;
+	wait = (c->reg[Rstctl] & 0x3f)*100;
 	for(i = 0;; i++){
 		if(i == wait)
 			return -1;
-		if((c->reg[Rststat + pf] & Devstate) == 0)
+		if((c->reg[Rststat] & Devstate) == 0)
 			break;
 		delay(1);
 	}
 	if(i == 0){
 		/* need to use pf here not the port# */
-		c->reg[Ctl + pf] |= Sreset;
+		c->reg[Ctl] |= Sreset;
 		for (i = 0; i < 10; i++) {
 			if(i == 10)
 				return -1;
-			if((c->reg[Ctl + pf] & Sreset) == 0)
+			if((c->reg[Ctl] & Sreset) == 0)
 				break;
 			delay(1);
 		}
@@ -1390,39 +1505,42 @@
 	/* disable pause here */
 }
 
+#define Pageno(fpa)	((fpa)>>Fpgshift)
+#define Pfpageidx(pf, fpa)	(Pageno(fpa) - Pageno(2*MiB*(pf)))
+#define Hmcblock(fpa)	((fpa)>>9)
+
 static void
 hmcmap(Ctlr *c, u64int fpa)
 {
-	uint pgno;
+	uint pfpg;
 
 	if(fpa > Fpdeperpage * Hmcpgsz)
 		panic("%s: hmcmap:  %#llux\n", cname(c), fpa);
 	if(fpa & (1<<Fpgshift)-1)
 		panic("%s: hmcmap: misaligned fpa %llux", cname(c), fpa);
-	pgno = fpa>>Fpgshift;
-	if(pgno >= nelem(c->pagetab))
-		panic("%s: hmcmap: pgno %d > %d", cname(c), pgno, nelem(c->pagetab));
-	if(c->pagetab[pgno] != nil)
+	pfpg = Pfpageidx(c->pf, fpa);
+	if(pfpg >= nelem(c->pagetab))
+		panic("%s: hmcmap", cname(c));
+if(c->pf==1)print("PFPG	%d\n", pfpg);
+	if(c->pagetab[pfpg] != nil)
 		return;
-	c->pagetab[pgno] = mallocalign(Hmcpgsz, Hmcalign, 0, 0);
+	c->pagetab[pfpg] = mallocalign(Hmcpgsz, Hmcalign, 0, 0);
 }
 
 static void
 hmcfault(Ctlr *c, u64int fpa)
 {
-	uint pgno;
+	uint pfpg;
 
 	hmcmap(c, fpa);
-	pgno = fpa>>Fpgshift;
-	c->pdpage[pgno] = Pciwaddr(c->pagetab[pgno]) | Fpdevalid;
+	pfpg = Pfpageidx(c->pf, fpa);
+	c->pdpage[pfpg] = Pciwaddr(c->pagetab[pfpg]) | Fpdevalid;
 }
 
-#define Pageno(x)	((x)>>Fpgshift)
-
 static void
 hmcinit(Ctlr *c)
 {
-	int pf;
+	uint pf;
 
 	pf = c->pf;
 	c->pdpage = mallocalign(Hmcpgsz, Hmcalign, 0, 0);
@@ -1432,15 +1550,15 @@
 	 * to each pf.   populate each card's addres space with
 	 * 1 FPDE at [2M*pf, 2M*pf+(1<<Ppgshift)).
 	 */
-	c->reg[Hmsdpart + pf] = pf*(2*MiB>>Fpgshift) | 1*Segentries;
+	c->reg[Hmsdpart + pf] = Pageno(pf*2*MiB) | 1*Segentries;
 
 	/*
 	 * allocate 4 rx and 4 tx queues.  allocation is in 512b blocks.
 	 * divide our single allocated page in half between tx/rx.
 	 */
-	c->reg[Hmtxbase + pf] = 0/512;
+	c->reg[Hmtxbase + pf] = (0+2*MiB*pf)/512;
 	c->reg[Hmtxcnt + pf] = 4;
-	c->reg[Hmrxbase + pf] = (Hmcpgsz/2)/512;
+	c->reg[Hmrxbase + pf] = (0+2*MiB*pf + Hmcpgsz/2)/512;
 	c->reg[Hmrxcnt + pf] = 4;
 
 	/*
@@ -1451,9 +1569,9 @@
 	/*
 	 * program SD with our pdpage.
 	 */
-	c->reg[Hmsdaddrhi] = Pciwaddrh(c->pdpage);
-	c->reg[Hmsdaddrlo] = Pciwaddrl(c->pdpage) | 1*Sdcount | Sdvalid | Sdpaged;
-	c->reg[Hmsdcmd] = Hmcmdwrite | Pageno(0);
+	c->reg[Hmsdaddrhi + pf] = Pciwaddrh(c->pdpage);
+	c->reg[Hmsdaddrlo + pf] = Pciwaddrl(c->pdpage) | 1*Sdcount | Sdvalid | Sdpaged;
+	c->reg[Hmsdcmd + pf] = Hmcmdwrite | Pageno(pf*2*MiB);
 }
 
 static void
@@ -1629,11 +1747,6 @@
 //	print("tc0 on qset %#.4ux\n", c->qsh);
 }
 
-void
-readclear(u32int)
-{
-}
-
 static void
 irqenable(Ctlr *c)
 {
@@ -1641,12 +1754,11 @@
 
 	readclear(c->reg[Icr + c->pf]);
 	for(pf = 0; pf < c->npf; pf++){
-		c->reg[Irxqctl + 0] = Irxidx*Iqirqidx | 0*Iqnxtqidx | Iqnxtt | Iqenable;	/* rx: ics=0x04 */
-		c->reg[Itxqctl + 0] = Itxidx*Iqirqidx | Iqend | Iqenable;		/* tx: ics=0x02 */
+		rxim(c, 0);
+		txim(c, 0);
 		c->reg[Ilnklst + pf] = 0*Firstqidx | Firstrx;				/* 1st: rxq idx 0 */
-
 		c->reg[Icren + pf] = c->im;
-		c->reg[Idynctl + pf] = Intena | Clearpba;
+		c->reg[Idynctl + pf] = 0<<5 | 0<<3 | Intena | Clearpba;		/* 0*4µs coalescing */
 	}
 	im(c, Errint);
 }
@@ -1723,18 +1835,19 @@
 	u[25] |= p>>8;
 }
 
+//	c->reg[Vsiqbase + seidtovsiid(c->vsiseid)] = Vsicontig | 0;
+
 static void
 txrxinit(Ctlr *c)
 {
 	uchar *t, *r;
 
-//	c->reg[Vsiqbase + seidtovsiid(c->vsiseid)] = Vsicontig | 0;
-
 	c->tqno = c->pf*4 + 0;
 	c->rqno = c->pf*4 + 0;		/* skip 1 to ignore flow director (doesn't work) */
 
 	t = (uchar*)c->pagetab[0] + 0 + c->tqno*128;
 	r = (uchar*)c->pagetab[0] + Hmcpgsz/2 + c->rqno*32;
+
 	txcontext(c, t);
 	rxcontext(c, r);
 	sfence();
@@ -1760,9 +1873,17 @@
 }
 
 static void
-attach(Ether *e)
+vkproc(char c, Ether *e, void(*f)(void*))
 {
 	char buf[KNAMELEN];
+
+	snprint(buf, sizeof buf, "#l%d%c", e->ctlrno, c);
+	kproc(buf, f, e);
+}
+
+static void
+attach(Ether *e)
+{
 	int t;
 	Ctlr *c;
 
@@ -1773,7 +1894,6 @@
 	unlock(&c->alock);
 	if(t & Fstarted)
 		return;
-
 	hmcinit(c);
 	linkinit(c);
 
@@ -1785,12 +1905,9 @@
 	allocblocks(c);		/* wrong place? move to rproc */
 	txrxinit(c);
 
-	snprint(buf, sizeof buf, "#l%da", e->ctlrno);
-	kproc(buf, aproc, e);
-	snprint(buf, sizeof buf, "#l%dr", e->ctlrno);
-	kproc(buf, rproc, e);
-	snprint(buf, sizeof buf, "#l%dt", e->ctlrno);
-	kproc(buf, tproc, e);
+	vkproc('a', e, aproc);
+	vkproc('r', e, rproc);
+	vkproc('t', e, tproc);
 }
 
 static void
@@ -1806,32 +1923,33 @@
 	ilock(&c->imlock);
 	im = c->im;
 	pf = c->pf;
+int debug = pf==1;
 
-	icr = c->reg[Icr + pf] & c->im;
-	c->reg[Idynctl + pf] = Intena | Clearpba;
+	icr = c->reg[Icr + pf] & im;
+//	c->reg[Idynctl + pf] = Clearpba;
 
-//	iprint("%s: irq %.8ux im %.8ux\n", cname(c), icr, c->im);
-/**/	if((icr & (Aprocint|Irx|Itx)) == 0)
-/**/		c->irqcnt[Cnts]++;
+if(debug)iprint("icr %.8ux %.8ux %.8ux\n", icr, c->reg[Hmerr+c->pf], c->reg[Hmerrd+c->pf]);
+	if((icr & (Aprocint|Irx|Itx)) == 0)
+		c->irqcnt[Cnts]++;
 	if(icr&Aprocint){
-/**/		c->irqcnt[Cnta]++;
+		c->irqcnt[Cnta]++;
 		im &= ~Aprocint;
 		c->aim = icr&Aprocint;
 		wakeup(&c->arendez);
 	}
 	if(icr&Irx){
-/**/		c->irqcnt[Cntr]++;
-		im &= ~Irx;
+		c->irqcnt[Cntr]++;
 		c->rim = icr&Irx;
 		wakeup(&c->rrendez);
 	}
 	if(icr&Itx){
-/**/		c->irqcnt[Cntt]++;
-		im &= ~Itx;
+		c->irqcnt[Cntt]++;
 		c->tim = icr&Itx;
 		wakeup(&c->trendez);
 	}
 	c->reg[Icren + pf] = im;
+	c->reg[Idynctl + pf] = 3<<3 | Intena;		/* no change to coalescing; enable */
+	c->im = im;
 	iunlock(&c->imlock);
 }
 
@@ -1874,8 +1992,7 @@
 	Ctlr *c;
 	Pcidev *p;
 
-	p = 0;
-	while(p = pcimatch(p, 0x8086, 0)){
+	for(p = nil; p = pcimatch(p, 0x8086, 0);){
 		hbafixup(p);
 		if((type = didmatch(p->did)) == -1)
 			continue;
@@ -1891,11 +2008,11 @@
 			continue;
 		}
 		c = malloc(sizeof *c);
+		c->type = type;
 		c->p = p;
 		c->port = io;
 		c->reg = (u32int*)mem;
 		c->rbsz = cttab[type].mtu;
-		c->type = type;
 		pcisetbme(p);
 		c->pf = c->reg[Portnum]&3;
 		c->npf = 1;				/* wrong */
@@ -1916,9 +2033,12 @@
 {
 	int i;
 	Ctlr *c;
+	static int once;
 
-	if(nctlr == 0)
+	if(once == 0){
 		scan();
+		once++;
+	}
 	for(i = 0; i<nctlr; i++){
 		c = ctlrtab[i];
 		if(c == nil || c->flag&Factive)
@@ -1944,7 +2064,7 @@
 	e->multicast = multicast;
 	e->promiscuous = promiscuous;
 	e->shutdown = shutdown;
-	e->transmit = transmit;
+//	e->transmit = transmit;
 	return 0;
 }