OS-6681 modernize viona driver

Reviewed by: Hans Rosenfeld <hans.rosenfeld@joyent.com>
Reviewed by: Mike Gerdts <mike.gerdts@joyent.com>
Reviewed by: Jerry Jelinek <jerry.jelinek@joyent.com>
Approved by: Jerry Jelinek <jerry.jelinek@joyent.com>

5 files changed, 312 insertions, 179 deletions

diff --git a/usr/src/cmd/bhyve/Makefile.com b/usr/src/cmd/bhyve/Makefile.com
index b80bf910de..b9ad3fde70 100644
--- a/usr/src/cmd/bhyve/Makefile.com
+++ b/usr/src/cmd/bhyve/Makefile.com
@@ -42,6 +42,7 @@ SRCS =	acpi.c			\
 	pci_virtio_block.c	\
 	pci_virtio_net.c	\
 	pci_virtio_rnd.c	\
+	pci_virtio_viona.c	\
 	pci_xhci.c		\
 	pm.c			\
 	post.c			\
diff --git a/usr/src/cmd/bhyve/pci_emul.c b/usr/src/cmd/bhyve/pci_emul.c
index 101773b4e5..d48af9521b 100644
--- a/usr/src/cmd/bhyve/pci_emul.c
+++ b/usr/src/cmd/bhyve/pci_emul.c
@@ -1580,6 +1580,11 @@ pci_lintr_update(struct pci_devinst *pi)
 		pci_irq_assert(pi);
 	}
 	pthread_mutex_unlock(&pi->pi_lintr.lock);
+#ifndef __FreeBSD__
+	if (pi->pi_d->pe_lintrupdate != NULL) {
+		pi->pi_d->pe_lintrupdate(pi);
+	}
+#endif /* __FreeBSD__ */
 }
 
 int
diff --git a/usr/src/cmd/bhyve/pci_emul.h b/usr/src/cmd/bhyve/pci_emul.h
index 0fffb19dee..a0aaa425be 100644
--- a/usr/src/cmd/bhyve/pci_emul.h
+++ b/usr/src/cmd/bhyve/pci_emul.h
@@ -25,6 +25,9 @@
  *
  * $FreeBSD$
  */
+/*
+ * Copyright 2018 Joyent, Inc.
+ */
 
 #ifndef _PCI_EMUL_H_
 #define _PCI_EMUL_H_
@@ -69,6 +72,10 @@ struct pci_devemu {
 	uint64_t  (*pe_barread)(struct vmctx *ctx, int vcpu,
 				struct pci_devinst *pi, int baridx,
 				uint64_t offset, int size);
+
+#ifndef __FreeBSD__
+	void	(*pe_lintrupdate)(struct pci_devinst *pi);
+#endif /* __FreeBSD__ */
 };
 #define PCI_EMUL_SET(x)   DATA_SET(pci_devemu_set, x);
 
diff --git a/usr/src/cmd/bhyve/pci_virtio_viona.c b/usr/src/cmd/bhyve/pci_virtio_viona.c
index e5a5cb584f..1c4b739711 100644
--- a/usr/src/cmd/bhyve/pci_virtio_viona.c
+++ b/usr/src/cmd/bhyve/pci_virtio_viona.c
@@ -34,7 +34,7 @@
  * http://www.illumos.org/license/CDDL.
  *
  * Copyright 2015 Pluribus Networks Inc.
- * Copyright 2017 Joyent, Inc.
+ * Copyright 2018 Joyent, Inc.
  */
 
 #include <sys/cdefs.h>
@@ -85,18 +85,6 @@
 #define	VIONA_REGSZ	VIONA_R_MAX+1
 
 /*
- * Host capabilities
- */
-#define	VIRTIO_NET_F_MAC	(1 <<  5) /* host supplies MAC */
-#define	VIRTIO_NET_F_MRG_RXBUF	(1 << 15) /* host can merge RX buffers */
-#define	VIRTIO_NET_F_STATUS	(1 << 16) /* config status field available */
-
-#define	VIONA_S_HOSTCAPS		\
-	(VIRTIO_NET_F_MAC |		\
-	VIRTIO_NET_F_MRG_RXBUF |	\
-	VIRTIO_NET_F_STATUS)
-
-/*
  * Queue definitions.
  */
 #define	VIONA_RXQ	0
@@ -108,7 +96,7 @@
 /*
  * Debug printf
  */
-static int pci_viona_debug;
+static volatile int pci_viona_debug;
 #define	DPRINTF(params) if (pci_viona_debug) printf params
 #define	WPRINTF(params) printf params
 
@@ -124,26 +112,20 @@ struct pci_viona_softc {
 	int		vsc_isr;
 
 	datalink_id_t	vsc_linkid;
-	char		vsc_linkname[MAXLINKNAMELEN];
 	int		vsc_vnafd;
 
+	/* Configurable parameters */
+	char		vsc_linkname[MAXLINKNAMELEN];
+	uint32_t	vsc_feature_mask;
+	uint16_t	vsc_vq_size;
+
 	uint32_t	vsc_features;
 	uint8_t		vsc_macaddr[6];
 
 	uint64_t	vsc_pfn[VIONA_MAXQ];
 	uint16_t	vsc_msix_table_idx[VIONA_MAXQ];
-	/*
-	 * Flag to see if host is already sending data out.
-	 * If it is, no need to wait for lock and send interrupt to host
-	 * for new data.
-	 */
-	boolean_t	vsc_tx_kick_lock_held;
-
-	pthread_t	tx_tid;
-	pthread_mutex_t	tx_mtx;
-	pthread_cond_t	tx_cond;
+	boolean_t	vsc_msix_active;
 };
-#define	viona_ctx(sc)	((sc)->vsc_pi->pi_vmctx)
 
 /*
  * Return the size of IO BAR that maps virtio header and device specific
@@ -160,15 +142,14 @@ pci_viona_iosize(struct pci_devinst *pi)
 }
 
 static uint16_t
-pci_viona_qsize(int qnum)
+pci_viona_qsize(struct pci_viona_softc *sc, int qnum)
 {
 	/* XXX no ctl queue currently */
 	if (qnum == VIONA_CTLQ) {
 		return (0);
 	}
 
-	/* XXX fixed currently. Maybe different for tx/rx/ctl */
-	return (VIONA_RINGSZ);
+	return (sc->vsc_vq_size);
 }
 
 static void
@@ -180,21 +161,13 @@ pci_viona_ring_reset(struct pci_viona_softc *sc, int ring)
 
 	switch (ring) {
 	case VIONA_RXQ:
-		error = ioctl(sc->vsc_vnafd, VNA_IOC_RX_RING_RESET);
-		if (error != 0) {
-			WPRINTF(("ioctl viona rx ring reset failed %d\n",
-			    error));
-		} else {
-			sc->vsc_pfn[VIONA_RXQ] = 0;
-		}
-		break;
 	case VIONA_TXQ:
-		error = ioctl(sc->vsc_vnafd, VNA_IOC_TX_RING_RESET);
+		error = ioctl(sc->vsc_vnafd, VNA_IOC_RING_RESET, ring);
 		if (error != 0) {
-			WPRINTF(("ioctl viona tx ring reset failed %d\n",
-			    error));
+			WPRINTF(("ioctl viona ring %u reset failed %d\n",
+			    ring, errno));
 		} else {
-			sc->vsc_pfn[VIONA_TXQ] = 0;
+			sc->vsc_pfn[ring] = 0;
 		}
 		break;
 	case VIONA_CTLQ:
@@ -220,11 +193,11 @@ static void *
 pci_viona_poll_thread(void *param)
 {
 	struct pci_viona_softc *sc = param;
-	pollfd_t	pollset;
-	int			error;
+	pollfd_t pollset;
+	const int fd = sc->vsc_vnafd;
 
-	pollset.fd = sc->vsc_vnafd;
-	pollset.events = POLLIN | POLLOUT;
+	pollset.fd = fd;
+	pollset.events = POLLRDBAND;
 
 	for (;;) {
 		if (poll(&pollset, 1, -1) < 0) {
@@ -236,23 +209,35 @@ pci_viona_poll_thread(void *param)
 				break;
 			}
 		}
-		if (pollset.revents & POLLIN) {
-			pci_generate_msix(sc->vsc_pi,
-			    sc->vsc_msix_table_idx[VIONA_RXQ]);
-			error = ioctl(sc->vsc_vnafd, VNA_IOC_RX_INTR_CLR);
-			if (error != 0) {
-				WPRINTF(("ioctl viona rx intr clear failed"
-				    " %d\n", error));
+		if (pollset.revents & POLLRDBAND) {
+			vioc_intr_poll_t vip;
+			uint_t i;
+			int res;
+			boolean_t assert_lintr = B_FALSE;
+			const boolean_t do_msix = pci_msix_enabled(sc->vsc_pi);
+
+			res = ioctl(fd, VNA_IOC_INTR_POLL, &vip);
+			for (i = 0; res > 0 && i < VIONA_VQ_MAX; i++) {
+				if (vip.vip_status[i] == 0) {
+					continue;
+				}
+				if (do_msix) {
+					pci_generate_msix(sc->vsc_pi,
+					    sc->vsc_msix_table_idx[i]);
+				} else {
+					assert_lintr = B_TRUE;
+				}
+				res = ioctl(fd, VNA_IOC_RING_INTR_CLR, i);
+				if (res != 0) {
+					WPRINTF(("ioctl viona vq %d intr "
+					    "clear failed %d\n", i, errno));
+				}
 			}
-		}
-
-		if (pollset.revents & POLLOUT) {
-			pci_generate_msix(sc->vsc_pi,
-			    sc->vsc_msix_table_idx[VIONA_TXQ]);
-			error = ioctl(sc->vsc_vnafd, VNA_IOC_TX_INTR_CLR);
-			if (error != 0) {
-				WPRINTF(("ioctl viona tx intr clear failed"
-				    " %d\n", error));
+			if (assert_lintr) {
+				pthread_mutex_lock(&sc->vsc_mtx);
+				sc->vsc_isr |= VTCFG_ISR_QUEUES;
+				pci_lintr_assert(sc->vsc_pi);
+				pthread_mutex_unlock(&sc->vsc_mtx);
 			}
 		}
 	}
@@ -261,57 +246,6 @@ pci_viona_poll_thread(void *param)
 }
 
 static void
-pci_viona_ping_rxq(struct pci_viona_softc *sc)
-{
-	int error;
-
-	error = ioctl(sc->vsc_vnafd, VNA_IOC_RX_RING_KICK);
-	if (error != 0) {
-		WPRINTF(("ioctl viona rx ring kick failed %d\n", error));
-	}
-}
-
-static void *
-pci_viona_tx_thread(void *param)
-{
-	struct pci_viona_softc *sc = (struct pci_viona_softc *)param;
-	int error;
-
-	pthread_mutex_lock(&sc->tx_mtx);
-	for (;;) {
-		error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
-		assert(error == 0);
-		sc->vsc_tx_kick_lock_held = B_TRUE;
-		error = ioctl(sc->vsc_vnafd, VNA_IOC_TX_RING_KICK);
-		if (error != 0) {
-			WPRINTF(("ioctl viona tx ring kick failed %d\n",
-			    error));
-		}
-		sc->vsc_tx_kick_lock_held = B_FALSE;
-	}
-	pthread_mutex_unlock(&sc->tx_mtx);
-
-	return (NULL);
-}
-
-static void
-pci_viona_ping_txq(struct pci_viona_softc *sc)
-{
-	/* Signal the tx thread for processing */
-	if (sc->vsc_tx_kick_lock_held)
-		return;
-	pthread_mutex_lock(&sc->tx_mtx);
-	pthread_cond_signal(&sc->tx_cond);
-	pthread_mutex_unlock(&sc->tx_mtx);
-}
-
-static void
-pci_viona_ping_ctlq(struct pci_viona_softc *sc)
-{
-	DPRINTF(("viona: control qnotify!\n\r"));
-}
-
-static void
 pci_viona_ring_init(struct pci_viona_softc *sc, uint64_t pfn)
 {
 	int			qnum = sc->vsc_curq;
@@ -320,29 +254,19 @@ pci_viona_ring_init(struct pci_viona_softc *sc, uint64_t pfn)
 
 	assert(qnum < VIONA_MAXQ);
 
+	if (qnum == VIONA_CTLQ) {
+		return;
+	}
+
 	sc->vsc_pfn[qnum] = (pfn << VRING_PFN);
 
-	vna_ri.ri_qsize = pci_viona_qsize(qnum);
+	vna_ri.ri_index = qnum;
+	vna_ri.ri_qsize = pci_viona_qsize(sc, qnum);
 	vna_ri.ri_qaddr = (pfn << VRING_PFN);
+	error = ioctl(sc->vsc_vnafd, VNA_IOC_RING_INIT, &vna_ri);
 
-	switch (qnum) {
-	case VIONA_RXQ:
-		error = ioctl(sc->vsc_vnafd, VNA_IOC_RX_RING_INIT, &vna_ri);
-		if (error != 0) {
-			WPRINTF(("ioctl viona rx ring init failed %d\n",
-			    error));
-		}
-		break;
-	case VIONA_TXQ:
-		error = ioctl(sc->vsc_vnafd, VNA_IOC_TX_RING_INIT, &vna_ri);
-		if (error != 0) {
-			WPRINTF(("ioctl viona tx ring init failed %d\n",
-			    error));
-		}
-		break;
-	case VIONA_CTLQ:
-	default:
-		break;
+	if (error != 0) {
+		WPRINTF(("ioctl viona ring %u init failed %d\n", qnum, errno));
 	}
 }
 
@@ -350,30 +274,20 @@ static int
 pci_viona_viona_init(struct vmctx *ctx, struct pci_viona_softc *sc)
 {
 	vioc_create_t		vna_create;
-#if notyet
-	char			devname[MAXNAMELEN];
-	int			ctlfd;
-#endif
 	int			error;
 
-	sc->vsc_vnafd = open("/devices/pseudo/viona@0:ctl", O_RDWR | O_EXCL);
+	sc->vsc_vnafd = open("/dev/viona", O_RDWR | O_EXCL);
 	if (sc->vsc_vnafd == -1) {
-		WPRINTF(("open viona ctl failed\n"));
+		WPRINTF(("open viona ctl failed: %d\n", errno));
 		return (-1);
 	}
 
 	vna_create.c_linkid = sc->vsc_linkid;
-	strlcpy(vna_create.c_vmname, vmname,
-	    sizeof (vna_create.c_vmname));
-#if notyet
-	vm_get_memory_seg(ctx, 1 * (1024 * 1024UL), &vna_create.c_lomem_size,
-	    NULL);
-	vm_get_memory_seg(ctx, 4 * (1024 * 1024 * 1024UL),
-	    &vna_create.c_himem_size, NULL);
-#endif
+	vna_create.c_vmfd = vm_get_device_fd(ctx);
 	error = ioctl(sc->vsc_vnafd, VNA_IOC_CREATE, &vna_create);
 	if (error != 0) {
-		WPRINTF(("ioctl viona create failed %d\n", error));
+		(void) close(sc->vsc_vnafd);
+		WPRINTF(("ioctl viona create failed %d\n", errno));
 		return (-1);
 	}
 
@@ -381,15 +295,99 @@ pci_viona_viona_init(struct vmctx *ctx, struct pci_viona_softc *sc)
 }
 
 static int
+pci_viona_parse_opts(struct pci_viona_softc *sc, char *opts)
+{
+	char *next, *cp, *vnic = NULL;
+	int err = 0;
+
+	sc->vsc_vq_size = VIONA_RINGSZ;
+	sc->vsc_feature_mask = 0;
+
+	for (; opts != NULL && *opts != '\0'; opts = next) {
+		char *val;
+
+		if ((cp = strchr(opts, ',')) != NULL) {
+			*cp = '\0';
+			next = cp + 1;
+		} else {
+			next = NULL;
+		}
+
+		if ((cp = strchr(opts, '=')) == NULL) {
+			/* vnic chosen with bare name */
+			if (vnic != NULL) {
+				fprintf(stderr,
+				    "viona: unexpected vnic name '%s'", opts);
+				err = -1;
+			} else {
+				vnic = opts;
+			}
+			continue;
+		}
+
+		/* <param>=<value> handling */
+		val = cp + 1;
+		*cp = '\0';
+		if (strcmp(opts, "feature_mask") == 0) {
+			long num;
+
+			errno = 0;
+			num = strtol(val, NULL, 0);
+			if (errno != 0 || num < 0) {
+				fprintf(stderr,
+				    "viona: invalid mask '%s'", val);
+			} else {
+				sc->vsc_feature_mask = num;
+			}
+		} else if (strcmp(opts, "vqsize") == 0) {
+			long num;
+
+			errno = 0;
+			num = strtol(val, NULL, 0);
+			if (errno != 0) {
+				fprintf(stderr,
+				    "viona: invalid vsqize '%s'", val);
+				err = -1;
+			} else if (num <= 2 || num > 32768) {
+				fprintf(stderr,
+				    "viona: vqsize out of range", num);
+				err = -1;
+			} else if ((1 << (ffs(num) - 1)) != num) {
+				fprintf(stderr,
+				    "viona: vqsize must be power of 2", num);
+				err = -1;
+			} else {
+				sc->vsc_vq_size = num;
+			}
+		} else {
+			fprintf(stderr,
+			    "viona: unrecognized option '%s'", opts);
+			err = -1;
+		}
+	}
+	if (vnic == NULL) {
+		fprintf(stderr, "viona: vnic name required");
+		sc->vsc_linkname[0] = '\0';
+		err = -1;
+	} else {
+		(void) strlcpy(sc->vsc_linkname, vnic, MAXLINKNAMELEN);
+	}
+
+	DPRINTF(("viona=%p dev=%s vqsize=%x feature_mask=%x\n", sc,
+	    sc->vsc_linkname, sc->vsc_vq_size, sc->vsc_feature_mask));
+	return (err);
+}
+
+static int
 pci_viona_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
 {
 	dladm_handle_t		handle;
 	dladm_status_t		status;
 	dladm_vnic_attr_t	attr;
 	char			errmsg[DLADM_STRSIZE];
-	int error;
+	int error, i;
 	struct pci_viona_softc *sc;
-	int i;
+	uint64_t ioport;
 
 	if (opts == NULL) {
 		printf("virtio-viona: vnic required\n");
@@ -404,7 +402,10 @@ pci_viona_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
 
 	pthread_mutex_init(&sc->vsc_mtx, NULL);
 
-	strlcpy(sc->vsc_linkname, opts, MAXLINKNAMELEN);
+	if (pci_viona_parse_opts(sc, opts) != 0) {
+		free(sc);
+		return (1);
+	}
 
 	if ((status = dladm_open(&handle)) != DLADM_STATUS_OK) {
 		WPRINTF(("could not open /dev/dld"));
@@ -430,7 +431,6 @@ pci_viona_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
 		return (1);
 	}
 
-	sc->vsc_tx_kick_lock_held = B_FALSE;
 	memcpy(sc->vsc_macaddr, attr.va_mac_addr, ETHERADDRL);
 
 	dladm_close(handle);
@@ -449,42 +449,44 @@ pci_viona_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
 	pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
 	pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET);
+	pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
 
 	/* MSI-X support */
 	for (i = 0; i < VIONA_MAXQ; i++)
 		sc->vsc_msix_table_idx[i] = VIRTIO_MSI_NO_VECTOR;
 
-	/*
-	 * BAR 1 used to map MSI-X table and PBA
-	 */
+	/* BAR 1 used to map MSI-X table and PBA */
 	if (pci_emul_add_msixcap(pi, VIONA_MAXQ, 1)) {
 		free(sc);
 		return (1);
 	}
 
-	pci_emul_alloc_bar(pi, 0, PCIBAR_IO, VIONA_REGSZ);
+	/* BAR 0 for legacy-style virtio register access. */
+	error = pci_emul_alloc_bar(pi, 0, PCIBAR_IO, VIONA_REGSZ);
+	if (error != NULL) {
+		WPRINTF(("could not allocate virtio BAR\n"));
+		free(sc);
+		return (1);
+	}
+
+	/* Install ioport hook for virtqueue notification */
+	ioport = pi->pi_bar[0].addr + VTCFG_R_QNOTIFY;
+	error = ioctl(sc->vsc_vnafd, VNA_IOC_SET_NOTIFY_IOP, ioport);
+	if (error != 0) {
+		WPRINTF(("could not install ioport hook at %x\n", ioport));
+		free(sc);
+		return (1);
+	}
 
 	/*
-	 * Initialize tx semaphore & spawn TX processing thread
-	 * As of now, only one thread for TX desc processing is
-	 * spawned.
+	 * Need a legacy interrupt for virtio compliance, even though MSI-X
+	 * operation is _strongly_ suggested for adequate performance.
 	 */
-	pthread_mutex_init(&sc->tx_mtx, NULL);
-	pthread_cond_init(&sc->tx_cond, NULL);
-	pthread_create(&sc->tx_tid, NULL, pci_viona_tx_thread, (void *)sc);
+	pci_lintr_request(pi);
 
 	return (0);
 }
 
-/*
- * Function pointer array to handle queue notifications
- */
-static void (*pci_viona_qnotify[VIONA_MAXQ])(struct pci_viona_softc *) = {
-	pci_viona_ping_rxq,
-	pci_viona_ping_txq,
-	pci_viona_ping_ctlq
-};
-
 static uint64_t
 viona_adjust_offset(struct pci_devinst *pi, uint64_t offset)
 {
@@ -501,6 +503,109 @@ viona_adjust_offset(struct pci_devinst *pi, uint64_t offset)
 }
 
 static void
+pci_viona_ring_set_msix(struct pci_devinst *pi, uint_t ring)
+{
+	struct pci_viona_softc *sc = pi->pi_arg;
+	struct msix_table_entry mte;
+	uint16_t tab_index;
+	vioc_ring_msi_t vrm;
+	int res;
+
+	assert(ring <= VIONA_VQ_TX);
+
+	vrm.rm_index = ring;
+	vrm.rm_addr = 0;
+	vrm.rm_msg = 0;
+	tab_index = sc->vsc_msix_table_idx[ring];
+
+	if (tab_index != VIRTIO_MSI_NO_VECTOR && sc->vsc_msix_active) {
+		mte = pi->pi_msix.table[tab_index];
+		if ((mte.vector_control & PCIM_MSIX_VCTRL_MASK) == 0) {
+			vrm.rm_addr = mte.addr;
+			vrm.rm_msg = mte.msg_data;
+		}
+	}
+
+	res = ioctl(sc->vsc_vnafd, VNA_IOC_RING_SET_MSI, &vrm);
+	if (res != 0) {
+		WPRINTF(("ioctl viona set_msi %d failed %d\n", ring, errno));
+	}
+}
+
+static void
+pci_viona_lintrupdate(struct pci_devinst *pi)
+{
+	struct pci_viona_softc *sc = pi->pi_arg;
+	boolean_t msix_on = B_FALSE;
+
+	pthread_mutex_lock(&sc->vsc_mtx);
+	msix_on = pci_msix_enabled(pi) && (pi->pi_msix.function_mask == 0);
+	if ((sc->vsc_msix_active && !msix_on) ||
+	    (msix_on && !sc->vsc_msix_active)) {
+		uint_t i;
+
+		sc->vsc_msix_active = msix_on;
+		/* Update in-kernel ring configs */
+		for (i = 0; i <= VIONA_VQ_TX; i++) {
+			pci_viona_ring_set_msix(pi, i);
+		}
+	}
+	pthread_mutex_unlock(&sc->vsc_mtx);
+}
+
+static void
+pci_viona_msix_update(struct pci_devinst *pi, uint64_t offset)
+{
+	struct pci_viona_softc *sc = pi->pi_arg;
+	uint_t tab_index, i;
+
+	pthread_mutex_lock(&sc->vsc_mtx);
+	if (!sc->vsc_msix_active) {
+		pthread_mutex_unlock(&sc->vsc_mtx);
+		return;
+	}
+
+	/*
+	 * Rather than update every possible MSI-X vector, cheat and use the
+	 * offset to calculate the entry within the table.  Since this should
+	 * only be called when a write to the table succeeds, the index should
+	 * be valid.
+	 */
+	tab_index = offset / MSIX_TABLE_ENTRY_SIZE;
+
+	for (i = 0; i <= VIONA_VQ_TX; i++) {
+		if (sc->vsc_msix_table_idx[i] != tab_index) {
+			continue;
+		}
+		pci_viona_ring_set_msix(pi, i);
+	}
+
+	pthread_mutex_unlock(&sc->vsc_mtx);
+}
+
+static void
+pci_viona_qnotify(struct pci_viona_softc *sc, int ring)
+{
+	int error;
+
+	switch (ring) {
+	case VIONA_TXQ:
+	case VIONA_RXQ:
+		error = ioctl(sc->vsc_vnafd, VNA_IOC_RING_KICK, ring);
+		if (error != 0) {
+			WPRINTF(("ioctl viona ring %d kick failed %d\n",
+			    ring, errno));
+		}
+		break;
+	case VIONA_CTLQ:
+		DPRINTF(("viona: control qnotify!\n"));
+		break;
+	default:
+		break;
+	}
+}
+
+static void
 pci_viona_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
     int baridx, uint64_t offset, int size, uint64_t value)
 {
@@ -510,7 +615,9 @@ pci_viona_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 
 	if (baridx == pci_msix_table_bar(pi) ||
 	    baridx == pci_msix_pba_bar(pi)) {
-		pci_emul_msix_twrite(pi, offset, size, value);
+		if (pci_emul_msix_twrite(pi, offset, size, value) == 0) {
+			pci_viona_msix_update(pi, offset);
+		}
 		return;
 	}
 
@@ -529,10 +636,14 @@ pci_viona_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 	switch (offset) {
 	case VTCFG_R_GUESTCAP:
 		assert(size == 4);
+		value &= ~(sc->vsc_feature_mask);
 		err = ioctl(sc->vsc_vnafd, VNA_IOC_SET_FEATURES, &value);
-		if (err != 0)
+		if (err != 0) {
 			WPRINTF(("ioctl feature negotiation returned"
-			    " err = %d\n", err));
+			    " err = %d\n", errno));
+		} else {
+			sc->vsc_features = value;
+		}
 		break;
 	case VTCFG_R_PFN:
 		assert(size == 4);
@@ -546,7 +657,7 @@ pci_viona_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 	case VTCFG_R_QNOTIFY:
 		assert(size == 2);
 		assert(value < VIONA_MAXQ);
-		(*pci_viona_qnotify[value])(sc);
+		pci_viona_qnotify(sc, value);
 		break;
 	case VTCFG_R_STATUS:
 		assert(size == 1);
@@ -597,7 +708,7 @@ pci_viona_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 	pthread_mutex_unlock(&sc->vsc_mtx);
 }
 
-uint64_t
+static uint64_t
 pci_viona_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
     int baridx, uint64_t offset, int size)
 {
@@ -627,9 +738,11 @@ pci_viona_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 	case VTCFG_R_HOSTCAP:
 		assert(size == 4);
 		err = ioctl(sc->vsc_vnafd, VNA_IOC_GET_FEATURES, &value);
-		if (err != 0)
+		if (err != 0) {
 			WPRINTF(("ioctl get host features returned"
-			    " err = %d\n", err));
+			    " err = %d\n", errno));
+		}
+		value &= ~sc->vsc_feature_mask;
 		break;
 	case VTCFG_R_GUESTCAP:
 		assert(size == 4);
@@ -641,7 +754,7 @@ pci_viona_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 		break;
 	case VTCFG_R_QNUM:
 		assert(size == 2);
-		value = pci_viona_qsize(sc->vsc_curq);
+		value = pci_viona_qsize(sc, sc->vsc_curq);
 		break;
 	case VTCFG_R_QSEL:
 		assert(size == 2);
@@ -705,9 +818,10 @@ pci_viona_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
 }
 
 struct pci_devemu pci_de_viona = {
-	.pe_emu = 	"virtio-net-viona",
+	.pe_emu =	"virtio-net-viona",
 	.pe_init =	pci_viona_init,
 	.pe_barwrite =	pci_viona_write,
-	.pe_barread =	pci_viona_read
+	.pe_barread =	pci_viona_read,
+	.pe_lintrupdate = pci_viona_lintrupdate
 };
 PCI_EMUL_SET(pci_de_viona);
diff --git a/usr/src/cmd/devfsadm/i386/misc_link_i386.c b/usr/src/cmd/devfsadm/i386/misc_link_i386.c
index f1a6ef1172..99085a3d4f 100644
--- a/usr/src/cmd/devfsadm/i386/misc_link_i386.c
+++ b/usr/src/cmd/devfsadm/i386/misc_link_i386.c
@@ -84,6 +84,9 @@ static devfsadm_create_t misc_cbt[] = {
 	{ "pseudo", "ddi_pseudo", "ucode",
 	    TYPE_EXACT | DRV_EXACT, ILEVEL_0, ln_minor_name,
 	},
+	{ "pseudo", "ddi_pseudo", "viona",
+	    TYPE_EXACT | DRV_EXACT, ILEVEL_0, ln_minor_name,
+	}
 };
 
 DEVFSADM_CREATE_INIT_V0(misc_cbt);
@@ -109,6 +112,9 @@ static devfsadm_remove_t misc_remove_cbt[] = {
 	},
 	{ "serial", "^tty[a-z]$", RM_ALWAYS | RM_PRE,
 		ILEVEL_1, devfsadm_rm_all
+	},
+	{ "pseudo", "^viona$", RM_ALWAYS | RM_PRE | RM_HOT,
+		ILEVEL_0, devfsadm_rm_all
 	}
 };