enable msi-x, and clean up interrupt locking.

since wakeup/register reads should be slow in comparison to locking,
take the imlock twice if necessary.  it will usually be necessary due to
the link interrupt, but measurements show latency to be about 5000
cycles lower (20%) for maximum bandwidth tcp transmit.

Reference: /n/atom/patch/applied/82563msix
Date: Sun Sep  6 20:41:35 CES 2015
Signed-off-by: quanstro@quanstro.net

--- /sys/src/nix/k10/ether82563.c	Sun Sep  6 20:37:59 2015
+++ /sys/src/nix/k10/ether82563.c	Sun Sep  6 20:38:01 2015
@@ -554,8 +554,6 @@
 	int	radv;			/* receive interrupt absolute delay timer */
 
 	Rendez	trendez;
-	QLock	tlock;
-	int	tbusy;
 	Td	*tdba;			/* transmit descriptor base address */
 	Block	**tb;			/* transmit buffers */
 	int	tdh;			/* transmit descriptor head */
@@ -1595,10 +1593,11 @@
 	ctlr = edev->ctlr;
 
 	ilock(&ctlr->imlock);
-	csr32w(ctlr, Imc, ~0);
 	im = ctlr->im;
+	ctlr->im = 0;
+	iunlock(&ctlr->imlock);
 
-	while(icr = csr32r(ctlr, Icr) & ctlr->im){
+	if(icr = csr32r(ctlr, Icr) & im){
 		if(icr & (Lsc | Omed)){
 			im &= ~(Lsc | Omed);
 			ctlr->lim = icr & (Lsc | Omed);
@@ -1613,14 +1612,17 @@
 		}
 		if(icr & Txdw){
 			im &= ~Txdw;
-			ctlr->tintr++;
 			wakeup(&ctlr->trendez);
+			ctlr->tintr++;
 		}
 	}
 
-	ctlr->im = im;
-	csr32w(ctlr, Ims, im);
-	iunlock(&ctlr->imlock);
+	if(im){
+		ilock(&ctlr->imlock);
+		ctlr->im |= im;
+		csr32w(ctlr, Ims, ctlr->im);
+		iunlock(&ctlr->imlock);
+	}
 }
 
 static int
@@ -1853,6 +1855,8 @@
 	csr32w(ctlr, Fcrth, ctlr->fcrth);
 	if(cttab[ctlr->type].flag & F75)
 		csr32w(ctlr, Eitr, 128<<2);		/* 128 ¼ microsecond intervals */
+	csr32w(ctlr, Ctrlext, csr32r(ctlr, Ctrlext) | Iame);
+	csr32w(ctlr, Iam, Lsc|Omed|Rxt0|Rxo|Rxdmt0|Rxseq|Ack|Txdw);
 	return 0;
 }
 
@@ -2079,6 +2083,7 @@
 	}
 	pcisetbme(p);
 	if(i82563reset(ctlr)){
+		pciclrbme(ctlr->pcidev);
 		vunmap(ctlr->nic, p->mem[0].size);
 		return -1;
 	}
@@ -2089,6 +2094,7 @@
 pnp(Ether *edev, int type)
 {
 	Ctlr *ctlr;
+	Vctl *v;
 	static int done;
 
 	if(!done) {
@@ -2114,6 +2120,16 @@
 				break;
 		}
 	}
+
+	v = malloc(sizeof(Vctl));
+	vctlinit(v);
+	v->irq = ctlr->pcidev->intl;
+	v->f = i82563interrupt;
+	v->tbdf = ctlr->pcidev->tbdf;
+	v->a = edev;
+	if(ctlr->type == i210)
+		v->flag |= Vmsix;
+	edev->vector = v;
 
 	edev->ctlr = ctlr;
 	edev->port = ctlr->port;