from cinap, experimental support for virtio.

Reference: /n/atom/patch/applied/virtioscsik10
Date: Sat Jun  7 06:07:16 CES 2014
Signed-off-by: quanstro@quanstro.net

--- /sys/src/nix/k10/sdvirtio.c	Sat Jun  7 06:06:29 2014
+++ /sys/src/nix/k10/sdvirtio.c	Sat Jun  7 06:06:30 2014
@@ -15,6 +15,14 @@
 typedef struct Vqueue Vqueue;
 typedef struct Vdev Vdev;
 
+typedef struct ScsiCfg ScsiCfg;
+
+/* device types */
+enum {
+	TypBlk	= 2,
+	TypSCSI	= 8,
+};
+
 /* status flags */
 enum {
 	Acknowledge = 1,
@@ -72,6 +80,10 @@
 struct Vqueue
 {
 	Lock;
+
+	Vdev	*dev;
+	int	idx;
+
 	int	size;
 
 	int	free;
@@ -103,9 +115,30 @@
 	int	nqueue;
 	Vqueue	*queue[16];
 
+	void	*cfg;	/* device specific config (for scsi) */
+
 	Vdev	*next;
 };
 
+enum {
+	CDBSIZE		= 32,
+	SENSESIZE	= 96,
+};
+
+struct ScsiCfg
+{
+	u32int	num_queues;
+	u32int	seg_max;
+	u32int	max_sectors;
+	u32int	cmd_per_lun;
+	u32int	event_info_size;
+	u32int	sense_size;
+	u32int	cdb_size;
+	u16int	max_channel;
+	u16int	max_target;
+	u32int	max_lun;
+};
+
 static Vqueue*
 mkvqueue(int size)
 {
@@ -160,13 +193,14 @@
 viopnpdevs(int typ)
 {
 	Vdev *vd, *h, **hh;
+	Vqueue *q;
 	Pcidev *p;
 	int n, i;
 
 	h = nil;
 	hh = &h;
 	for(p = nil; p = pcimatch(p, 0x1af4, 0);){
-		if((p->did < 0x1000) || (p->did >= 0x1040))
+		if((p->did < 0x1000) || (p->did > 0x103F))
 			continue;
 		if(p->rid != 0)
 			continue;
@@ -178,7 +212,7 @@
 		}
 		vd->port = p->mem[0].bar & ~0x1;
 		if(ioalloc(vd->port, p->mem[0].size, 0, "virtio") < 0){
-			print("sdvirtio: port %lux in use\n", vd->port);
+			print("sdvirtio: port %ux in use\n", vd->port);
 			free(vd);
 			continue;
 		}
@@ -195,9 +229,12 @@
 			n = ins(vd->port+Qsize);
 			if(n == 0 || (n & (n-1)) != 0)
 				break;
-			if((vd->queue[i] = mkvqueue(n)) == nil)
+			if((q = mkvqueue(n)) == nil)
 				break;
-			coherence();	/* ? virtual machine can be uncoherent */
+			q->dev = vd;
+			q->idx = i;
+			vd->queue[i] = q;
+			coherence();
 			outl(vd->port+Qaddr, PADDR(vd->queue[i]->desc)/PGSZ);
 		}
 		vd->nqueue = i;
@@ -215,39 +252,42 @@
 };
 
 static void
-viointerrupt(Ureg *, void *arg)
+vqinterrupt(Vqueue *q)
 {
 	int id, free, m;
 	struct Rock *r;
 	Rendez *z;
-	Vqueue *q;
-	Vdev *vd;
 
-	vd = arg;
-	if(inb(vd->port+Isr) & 1){
-		q = vd->queue[0];
-		m = q->size-1;
+	m = q->size-1;
 
-		ilock(q);
-		while((q->lastused ^ q->used->idx) & m){
-			id = q->usedent[q->lastused++ & m].id;
-			if(r = q->rock[id]){
-				q->rock[id] = nil;
-				z = r->sleep;
-				r->done = 1;	/* hands off */
-				if(z != nil)
-					wakeup(z);
-			}
-			do {
-				free = id;
-				id = q->desc[free].next;
-				q->desc[free].next = q->free;
-				q->free = free;
-				q->nfree++;
-			} while(q->desc[free].flags & Next);
+	ilock(q);
+	while((q->lastused ^ q->used->idx) & m){
+		id = q->usedent[q->lastused++ & m].id;
+		if(r = q->rock[id]){
+			q->rock[id] = nil;
+			z = r->sleep;
+			r->done = 1;	/* hands off */
+			if(z != nil)
+				wakeup(z);
 		}
-		iunlock(q);
+		do {
+			free = id;
+			id = q->desc[free].next;
+			q->desc[free].next = q->free;
+			q->free = free;
+			q->nfree++;
+		} while(q->desc[free].flags & Next);
 	}
+	iunlock(q);
+}
+
+static void
+viointerrupt(Ureg *, void *arg)
+{
+	Vdev *vd = arg;
+
+	if(inb(vd->port+Isr) & 1)
+		vqinterrupt(vd->queue[vd->typ == TypSCSI ? 2 : 0]);
 }
 
 static int
@@ -257,7 +297,7 @@
 }
 
 static int
-vioreq(Vdev *vd, int typ, void *a, long count, long secsize, uvlong lba)
+vioblkreq(Vdev *vd, int typ, void *a, long count, long secsize, uvlong lba)
 {
 	struct Rock rock;
 	int free, head;
@@ -265,7 +305,7 @@
 	Vdesc *d;
 
 	u8int status;
-	struct Vioreqhdr {
+	struct Vioblkreqhdr {
 		u32int	typ;
 		u32int	prio;
 		u64int	lba;
@@ -316,7 +356,7 @@
 	coherence();
 	q->availent[q->avail->idx++ & (q->size-1)] = head;
 	coherence();
-	outs(vd->port+Qnotify, 0);
+	outs(vd->port+Qnotify, q->idx);
 	iunlock(q);
 
 	while(!rock.done){
@@ -326,33 +366,150 @@
 		poperror();
 
 		if(!rock.done)
-			viointerrupt(nil, vd);
+			vqinterrupt(q);
 	}
 
 	return status;
 }
 
+static int
+vioscsireq(SDreq *r)
+{
+	uchar resp[4+4+2+2+SENSESIZE], req[8+8+3+CDBSIZE];
+	struct Rock rock;
+	int free, head;
+	u32int len;
+	Vqueue *q;
+	Vdesc *d;
+	Vdev *vd;
+	SDunit *u;
+	ScsiCfg *cfg;
+
+	u = r->unit;
+	vd = u->dev->ctlr;
+	cfg = vd->cfg;
+
+	memset(resp, 0, sizeof(resp));
+	memset(req, 0, sizeof(req));
+	req[0] = 1;
+	req[1] = u->subno;
+	req[2] = r->lun>>8;
+	req[3] = r->lun&0xFF;
+	*(u64int*)(&req[8]) = (uintptr)r;
+
+	memmove(&req[8+8+3], r->cmd, r->clen);
+
+	rock.done = 0;
+	rock.sleep = &up->sleep;
+
+	q = vd->queue[2];
+	ilock(q);
+	while(q->nfree < 3){
+		iunlock(q);
+
+		if(!waserror())
+			tsleep(&up->sleep, return0, 0, 500);
+		poperror();
+
+		ilock(q);
+	}
+
+	head = free = q->free;
+
+	d = &q->desc[free]; free = d->next;
+	d->addr = PADDR(req);
+	d->len = 8+8+3+cfg->cdb_size;
+	d->flags = Next;
+
+	if(r->write && r->dlen > 0){
+		d = &q->desc[free]; free = d->next;
+		d->addr = PADDR(r->data);
+		d->len = r->dlen;
+		d->flags = Next;
+	}
+
+	d = &q->desc[free]; free = d->next;
+	d->addr = PADDR(resp);
+	d->len = 4+4+2+2+cfg->sense_size;
+	d->flags = Write;
+
+	if(!r->write && r->dlen > 0){
+		d->flags |= Next;
+
+		d = &q->desc[free]; free = d->next;
+		d->addr = PADDR(r->data);
+		d->len = r->dlen;
+		d->flags = Write;
+	}
+	
+	q->free = free;
+	q->nfree -= 2 + (r->dlen > 0);
+
+	q->rock[head] = &rock;
+
+//	coherence();
+	q->availent[q->avail->idx++ & (q->size-1)] = head;
+	coherence();
+	outs(vd->port+Qnotify, q->idx);
+	iunlock(q);
+
+	while(!rock.done){
+		while(waserror())
+			;
+		tsleep(rock.sleep, viodone, &rock, 1000);
+		poperror();
+
+		if(!rock.done)
+			vqinterrupt(q);
+	}
+
+	/* response+status */
+	r->status = resp[10];
+	if(resp[11] != 0)
+		r->status = SDcheck;
+
+	/* sense_len */
+	len = *((u32int*)&resp[0]);
+	if(len > 0){
+		if(len > sizeof(r->sense))
+			len = sizeof(r->sense);
+		memmove(r->sense, &resp[4+4+2+2], len);
+		r->flags |= SDvalidsense;
+	}
+
+	/* data residue */
+	len = *((u32int*)&resp[4]);
+	if(len > r->dlen)
+		r->rlen = 0;
+	else
+		r->rlen = r->dlen - len;
+
+	return r->status;
+
+}
+
 static long
-viobio(SDunit *u, int, int write, void *a, long count, uvlong lba)
+viobio(SDunit *u, int lun, int write, void *a, long count, uvlong lba)
 {
 	long ss, cc, max, ret;
 	Vdev *vd;
 
-	max = 32;
-	ss = u->secsize;
 	vd = u->dev->ctlr;
+	if(vd->typ == TypSCSI)
+		return scsibio(u, lun, write, a, count, lba);
 
+	max = 32;
+	ss = u->secsize;
 	ret = 0;
 	while(count > 0){
 		if((cc = count) > max)
 			cc = max;
-		if(vioreq(vd, write != 0, (uchar*)a + ret, cc, ss, lba) != 0)
+		if(vioblkreq(vd, write != 0, (uchar*)a + ret, cc, ss, lba) != 0)
 			error(Eio);
 		ret += cc*ss;
 		count -= cc;
 		lba += cc;
 	}
-
 	return ret;
 }
 
@@ -362,10 +519,14 @@
 	int i, count, rw;
 	uvlong lba;
 	SDunit *u;
+	Vdev *vd;
 
 	u = r->unit;
+	vd = u->dev->ctlr;
+	if(vd->typ == TypSCSI)
+		return vioscsireq(r);
 	if(r->cmd[0] == 0x35 || r->cmd[0] == 0x91){
-		if(vioreq(u->dev->ctlr, 4, nil, 0, 0, 0) != 0)
+		if(vioblkreq(vd, 4, nil, 0, 0, 0) != 0)
 			return sdsetsense(r, SDcheck, 3, 0xc, 2);
 		return sdsetsense(r, SDok, 0, 0, 0);
 	}
@@ -384,6 +545,9 @@
 	Vdev *vd;
 
 	vd = u->dev->ctlr;
+	if(vd->typ == TypSCSI)
+		return scsionline(u);
+
 	cap = inl(vd->port+Devspec+4);
 	cap <<= 32;
 	cap |= inl(vd->port+Devspec);
@@ -396,13 +560,26 @@
 }
 
 static int
-vioverify(SDunit *)
+vioverify(SDunit *u)
 {
+	Vdev *vd;
+
+	vd = u->dev->ctlr;
+	if(vd->typ == TypSCSI)
+		return scsiverify(u);
+
 	return 1;
 }
 
 SDifc sdvirtioifc;
 
+static void
+vdevenable(Vdev *vd)
+{
+	intrenable(vd->pci->intl, viointerrupt, vd, vd->pci->tbdf, "virtio");
+	outb(vd->port+Status, inb(vd->port+Status) | DriverOk);
+}
+
 static SDev*
 viopnp(void)
 {
@@ -413,12 +590,11 @@
 	id = 'F';
 	h = nil;
 	hh = &h;
-	for(vd =  viopnpdevs(2); vd != nil; vd = vd->next){
+	for(vd =  viopnpdevs(TypBlk); vd != nil; vd = vd->next){
 		if(vd->nqueue != 1)
 			continue;
 
-		intrenable(vd->pci->intl, viointerrupt, vd, vd->pci->tbdf, "virtio");
-		outb(vd->port+Status, inb(vd->port+Status) | DriverOk);
+		vdevenable(vd);
 
 		if((s = malloc(sizeof(*s))) == nil)
 			break;
@@ -426,6 +602,50 @@
 		s->idno = id++;
 		s->ifc = &sdvirtioifc;
 		s->nunit = 1;
+		*hh = s;
+		hh = &s->next;
+	}
+
+	id = '0';
+	for(vd = viopnpdevs(TypSCSI); vd; vd = vd->next){
+		ScsiCfg *cfg;
+
+		if(vd->nqueue < 3)
+			continue;
+
+		if((cfg = malloc(sizeof(*cfg))) == nil)
+			break;
+		cfg->num_queues = inl(vd->port+Devspec+4*0);
+		cfg->seg_max = inl(vd->port+Devspec+4*1);
+		cfg->max_sectors = inl(vd->port+Devspec+4*2);
+		cfg->cmd_per_lun = inl(vd->port+Devspec+4*3);
+		cfg->event_info_size = inl(vd->port+Devspec+4*4);
+		cfg->sense_size = inl(vd->port+Devspec+4*5);
+		cfg->cdb_size = inl(vd->port+Devspec+4*6);
+		cfg->max_channel = ins(vd->port+Devspec+4*7);
+		cfg->max_target = ins(vd->port+Devspec+4*7+2);
+		cfg->max_lun = inl(vd->port+Devspec+4*8);
+
+		if(cfg->max_target == 0){
+			free(cfg);
+			continue;
+		}
+		if((cfg->cdb_size > CDBSIZE) || (cfg->sense_size > SENSESIZE)){
+			print("sdvirtio: cdb %ud or sense size %ud too big\n",
+				cfg->cdb_size, cfg->sense_size);
+			free(cfg);
+			continue;
+		}
+		vd->cfg = cfg;
+			
+		vdevenable(vd);
+
+		if((s = malloc(sizeof(*s))) == nil)
+			break;
+		s->ctlr = vd;
+		s->idno = id++;
+		s->ifc = &sdvirtioifc;
+		s->nunit = cfg->max_target;
 		*hh = s;
 		hh = &s->next;
 	}
--- /sys/src/nix/k10/cpu	Sat Jun  7 06:06:32 2014
+++ /sys/src/nix/k10/cpu	Sat Jun  7 06:06:33 2014
@@ -16,7 +16,7 @@
 	kprof
 #	segment
 	acpi
-	ws
+	ws2
 
 	ether		netif
 	cec
@@ -46,7 +46,8 @@
 
 	ether8169	pci ethermii
 #	ether82557	pci
-	ether82563	pci
+#	ether82563	pci
+	ether82563x	pci
 	etherbcm	pci ethermii
 	etherigbe	pci ethermii
 
@@ -56,8 +57,11 @@
 
 	etheri40		pci
 
+	etherusb
+
 	ethermedium
-#	loopbackmedium
+#	must have for venti (please fix venti)
+	loopbackmedium
 #	netdevmedium
 	usbuhci
 	usbohci
@@ -70,11 +74,12 @@
 	sdodin		sdscsifis sdatafis
 	sdvanir		sdscsifis sdatafis
 	sdmpt2		sdscsifis sdatafis
+	sdvirtio		sdscsi
 	sdloop
 
 misc +dev
-#	mp		mpacpi lapic ioapic msi pci sipi
-	acpi		acpiio lapic ioapic msi pci sipi
+#	mp		mpacpi lapic ioapic msi msix pci sipi
+	acpi		acpiio lapic ioapic msi msix pci sipi
 
 boot cpu
 	il
--- /sys/src/nix/k10/term	Sat Jun  7 06:06:34 2014
+++ /sys/src/nix/k10/term	Sat Jun  7 06:06:35 2014
@@ -17,7 +17,7 @@
 #	segment
 	acpi
 	audio
-	ws
+	ws2
 
 	ether		netif
 	cec
@@ -82,11 +82,12 @@
 	sdodin		sdscsifis sdatafis
 	sdvanir		sdscsifis sdatafis
 	sdmpt2		sdscsifis sdatafis
+	sdvirtio		sdscsi
 	sdloop
 
 misc +dev
 #	mp		mpacpi lapic ioapic msi pci sipi
-	acpi		acpiio lapic ioapic msi pci sipi
+	acpi		acpiio lapic ioapic msi msix pci sipi
 
 boot terminal
 	il
--- /sys/src/nix/port/sdscsi.c	Sat Jun  7 06:06:37 2014
+++ /sys/src/nix/port/sdscsi.c	Sat Jun  7 06:06:38 2014
@@ -303,56 +303,6 @@
 		return 0;
 }
 
-int
-scsiexec(SDunit* unit, int write, uchar* cmd, int clen, void* data, int* dlen)
-{
-	SDreq *r;
-	int status;
-
-	if((r = malloc(sizeof(SDreq))) == nil)
-		return SDmalloc;
-	r->unit = unit;
-	r->lun = cmd[1]>>5;		/* ??? */
-	r->write = write;
-	memmove(r->cmd, cmd, clen);
-	r->clen = clen;
-	r->data = data;
-	if(dlen)
-		r->dlen = *dlen;
-	r->flags = 0;
-
-	r->status = ~0;
-
-	/*
-	 * Call the device-specific I/O routine.
-	 * There should be no calls to 'error()' below this
-	 * which percolate back up.
-	 */
-	switch(status = unit->dev->ifc->rio(r)){
-	case SDok:
-		if(dlen)
-			*dlen = r->rlen;
-		/*FALLTHROUGH*/
-	case SDcheck:
-		/*FALLTHROUGH*/
-	default:
-		/*
-		 * It's more complicated than this. There are conditions
-		 * which are 'ok' but for which the returned status code
-		 * is not 'SDok'.
-		 * Also, not all conditions require a reqsense, might
-		 * need to do a reqsense here and make it available to the
-		 * caller somehow.
-		 *
-		 * Mañana.
-		 */
-		break;
-	}
-	sdfree(r);
-
-	return status;
-}
-
 static void
 scsifmt10(SDreq *r, int write, int lun, ulong nb, uvlong bno)
 {