Differential D23342 Diff 67324 usr.sbin/bhyve/pci_virtio_net.c

Changeset View

Standalone View

usr.sbin/bhyve/pci_virtio_net.c

Show First 20 Lines • Show All 111 Lines • ▼ Show 20 Lines	struct pci_vtnet_softc {
pthread_mutex_t rx_mtx;		pthread_mutex_t rx_mtx;
int rx_merge; /* merged rx bufs in use */		int rx_merge; /* merged rx bufs in use */

pthread_t tx_tid;		pthread_t tx_tid;
pthread_mutex_t tx_mtx;		pthread_mutex_t tx_mtx;
pthread_cond_t tx_cond;		pthread_cond_t tx_cond;
int tx_in_progress;		int tx_in_progress;

		size_t vhdrlen;
		size_t be_vhdrlen;

struct virtio_net_config vsc_config;		struct virtio_net_config vsc_config;
struct virtio_consts vsc_consts;		struct virtio_consts vsc_consts;
};		};

static void pci_vtnet_reset(void *);		static void pci_vtnet_reset(void *);
/* static void pci_vtnet_notify(void , struct vqueue_info ); */		/* static void pci_vtnet_notify(void , struct vqueue_info ); */
static int pci_vtnet_cfgread(void , int, int, uint32_t );		static int pci_vtnet_cfgread(void , int, int, uint32_t );
static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);		static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
▲ Show 20 Lines • Show All 47 Lines • ▼ Show 20 Lines	pci_vtnet_reset(void *vsc)
*/		*/
vi_reset_dev(&sc->vsc_vs);		vi_reset_dev(&sc->vsc_vs);

sc->resetting = 0;		sc->resetting = 0;
pthread_mutex_unlock(&sc->tx_mtx);		pthread_mutex_unlock(&sc->tx_mtx);
pthread_mutex_unlock(&sc->rx_mtx);		pthread_mutex_unlock(&sc->rx_mtx);
}		}

		static __inline struct iovec *
		iov_trim_hdr(struct iovec iov, int iovcnt, unsigned int tlen)
		{
		struct iovec *riov;

		if (iov[0].iov_len < tlen) {
		/*
		* Not enough header space in the first fragment.
		* That's not ok for us.
		*/
		return NULL;
		}

		iov[0].iov_len -= tlen;
		if (iov[0].iov_len == 0) {
		*iovcnt -= 1;
		if (*iovcnt == 0) {
		/*
		* Only space for the header. That's not
		* enough for us.
		*/
		return NULL;
		}
		riov = &iov[1];
		} else {
		iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + tlen);
		riov = &iov[0];
		}

		return (riov);
		}

struct virtio_mrg_rxbuf_info {		struct virtio_mrg_rxbuf_info {
uint16_t idx;		uint16_t idx;
uint16_t pad;		uint16_t pad;
uint32_t len;		uint32_t len;
};		};

static void		static void
pci_vtnet_rx(struct pci_vtnet_softc *sc)		pci_vtnet_rx(struct pci_vtnet_softc *sc)
{		{
		int prepend_hdr_len = sc->vhdrlen - sc->be_vhdrlen;
struct virtio_mrg_rxbuf_info info[VTNET_MAXSEGS];		struct virtio_mrg_rxbuf_info info[VTNET_MAXSEGS];
struct iovec iov[VTNET_MAXSEGS + 1];		struct iovec iov[VTNET_MAXSEGS + 1];
struct vqueue_info *vq;		struct vqueue_info *vq;
uint32_t cur_iov_bytes;		uint32_t riov_bytes;
struct iovec *cur_iov;		struct iovec *riov;
uint16_t cur_iov_len;		int riov_len;
uint32_t ulen;		uint32_t ulen;
int n_chains;		int n_chains;
int len;		int len;

vq = &sc->vsc_queues[VTNET_RXQ];		vq = &sc->vsc_queues[VTNET_RXQ];
for (;;) {		for (;;) {
		struct virtio_net_rxhdr *hdr;

/*		/*
* Get a descriptor chain to store the next ingress		* Get a descriptor chain to store the next ingress
* packet. In case of mergeable rx buffers, get as		* packet. In case of mergeable rx buffers, get as
* many chains as necessary in order to make room		* many chains as necessary in order to make room
* for a maximum sized LRO packet.		* for a maximum sized LRO packet.
*/		*/
cur_iov_bytes = 0;		riov_bytes = 0;
cur_iov_len = 0;		riov_len = 0;
cur_iov = iov;		riov = iov;
n_chains = 0;		n_chains = 0;
do {		do {
int n = vq_getchain(vq, &info[n_chains].idx, cur_iov,		int n = vq_getchain(vq, &info[n_chains].idx, riov,
VTNET_MAXSEGS - cur_iov_len, NULL);		VTNET_MAXSEGS - riov_len, NULL);

if (n == 0) {		if (n == 0) {
/*		/*
* No rx buffers. Enable RX kicks and double		* No rx buffers. Enable RX kicks and double
* check.		* check.
*/		*/
vq_kick_enable(vq);		vq_kick_enable(vq);
if (!vq_has_descs(vq)) {		if (!vq_has_descs(vq)) {
Show All 9 Lines	do {
netbe_rx_disable(sc->vsc_be);		netbe_rx_disable(sc->vsc_be);
return;		return;
}		}

/* More rx buffers found, so keep going. */		/* More rx buffers found, so keep going. */
vq_kick_disable(vq);		vq_kick_disable(vq);
continue;		continue;
}		}
assert(n >= 1 && cur_iov_len + n <= VTNET_MAXSEGS);		assert(n >= 1 && riov_len + n <= VTNET_MAXSEGS);
cur_iov_len += n;		riov_len += n;
if (!sc->rx_merge) {		if (!sc->rx_merge) {
n_chains = 1;		n_chains = 1;
break;		break;
}		}
info[n_chains].len = (uint32_t)count_iov(cur_iov, n);		info[n_chains].len = (uint32_t)count_iov(riov, n);
cur_iov_bytes += info[n_chains].len;		riov_bytes += info[n_chains].len;
cur_iov += n;		riov += n;
n_chains++;		n_chains++;
} while (cur_iov_bytes < VTNET_MAX_PKT_LEN &&		} while (riov_bytes < VTNET_MAX_PKT_LEN &&
cur_iov_len < VTNET_MAXSEGS);		riov_len < VTNET_MAXSEGS);

len = netbe_recv(sc->vsc_be, iov, cur_iov_len);		riov = iov;
		hdr = riov[0].iov_base;
		if (prepend_hdr_len > 0) {
		/*
		* The frontend uses a virtio-net header, but the
		* backend does not. We need to prepend a zeroed
		* header.
		*/
		riov = iov_trim_hdr(riov, &riov_len, prepend_hdr_len);
		if (riov == NULL) {
		/*
		* The first collected chain is nonsensical,
		* as it is not even enough to store the
		* virtio-net header. Just drop it.
		*/
		vq_relchain(vq, info[0].idx, 0);
		vq_retchains(vq, n_chains - 1);
		continue;
		}
		memset(hdr, 0, prepend_hdr_len);
		}

		afedorovUnsubmitted Done Inline Actions We need to initialize 'struct virtio_net_rxhdr hdr = iov[0].iov_base;' before calling netbe_recv() which can change value of iov[0].iov_base. afedorov:* We need to initialize 'struct virtio_net_rxhdr *hdr = iov[0].iov_base;' before calling…
		vmaffioneAuthorUnsubmitted Done Inline Actions You're right, thank you! vmaffione: You're right, thank you!
		len = netbe_recv(sc->vsc_be, riov, riov_len);

if (len <= 0) {		if (len <= 0) {
/*		/*
* No more packets (len == 0), or backend errored		* No more packets (len == 0), or backend errored
* (err < 0). Return unused available buffers		* (err < 0). Return unused available buffers
* and stop.		* and stop.
*/		*/
vq_retchains(vq, n_chains);		vq_retchains(vq, n_chains);
/* Interrupt if needed/appropriate and stop. */		/* Interrupt if needed/appropriate and stop. */
vq_endchains(vq, /used_all_avail=/0);		vq_endchains(vq, /used_all_avail=/0);
return;		return;
}		}

ulen = (uint32_t)len; /* avoid too many casts below */		ulen = (uint32_t)(len + prepend_hdr_len);

/* Publish the used buffers to the guest. */		/*
		* Publish the used buffers to the guest, reporting the
		* number of bytes that we wrote.
		*/
if (!sc->rx_merge) {		if (!sc->rx_merge) {
vq_relchain(vq, info[0].idx, ulen);		vq_relchain(vq, info[0].idx, ulen);
} else {		} else {
struct virtio_net_rxhdr *hdr = iov[0].iov_base;
uint32_t iolen;		uint32_t iolen;
		afedorovUnsubmitted Done Inline Actions I tried to test this patch. And I found a bug. With a tap backend, the netbe_recv() function calls iov_trim() which changes the value of iov[0].iov_base (It increases the address by the size of the struct virtio_net_rxhdr). https://svnweb.freebsd.org/base/head/usr.sbin/bhyve/net_backends.c?view=markup#l819 https://svnweb.freebsd.org/base/head/usr.sbin/bhyve/net_backends.c?view=markup#l766 So, hdr points to an incorrect location. If I move the 'hdr' declaration before calling the netbe_recv () function, then everything works correctly: ping -f -s 8000 and iperf3 with MTU == 9000. afedorov: I tried to test this patch. And I found a bug. With a tap backend, the netbe_recv() function…
		vmaffioneAuthorUnsubmitted Done Inline Actions Thank you for the suggestion. I will try to take a different approach to avoid visible changes to the iov vmaffione: Thank you for the suggestion. I will try to take a different approach to avoid visible changes…
int i = 0;		int i = 0;

assert(iov[0].iov_len >= sizeof(*hdr));

do {		do {
iolen = info[i].len;		iolen = info[i].len;
if (iolen > ulen) {		if (iolen > ulen) {
iolen = ulen;		iolen = ulen;
}		}
vq_relchain_prepare(vq, info[i].idx, iolen);		vq_relchain_prepare(vq, info[i].idx, iolen);
ulen -= iolen;		ulen -= iolen;
i++;		i++;
Show All 39 Lines	pci_vtnet_ping_rxq(void vsc, struct vqueue_info vq)
pthread_mutex_unlock(&sc->rx_mtx);		pthread_mutex_unlock(&sc->rx_mtx);
}		}

/* TX virtqueue processing, called by the TX thread. */		/* TX virtqueue processing, called by the TX thread. */
static void		static void
pci_vtnet_proctx(struct pci_vtnet_softc sc, struct vqueue_info vq)		pci_vtnet_proctx(struct pci_vtnet_softc sc, struct vqueue_info vq)
{		{
struct iovec iov[VTNET_MAXSEGS + 1];		struct iovec iov[VTNET_MAXSEGS + 1];
		struct iovec *siov = iov;
uint16_t idx;		uint16_t idx;
ssize_t len;		ssize_t len;
int n;		int n;

/*		/*
* Obtain chain of descriptors. The first descriptor also		* Obtain chain of descriptors. The first descriptor also
* contains the virtio-net header.		* contains the virtio-net header.
*/		*/
n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);		n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
assert(n >= 1 && n <= VTNET_MAXSEGS);		assert(n >= 1 && n <= VTNET_MAXSEGS);

len = netbe_send(sc->vsc_be, iov, n);		if (sc->vhdrlen != sc->be_vhdrlen) {
		/*
		* The frontend uses a virtio-net header, but the backend
		* does not. We simply strip the header and ignore it, as
		* it should be zero-filled.
		*/
		siov = iov_trim_hdr(siov, &n, sc->vhdrlen);
		}

/* chain is processed, release it and set len */		if (siov == NULL) {
vq_relchain(vq, idx, len > 0 ? len : 0);		/* The chain is nonsensical. Just drop it. */
		len = 0;
		} else {
		len = netbe_send(sc->vsc_be, siov, n);
		if (len < 0) {
		/*
		* If send failed, report that 0 bytes
		* were written.
		*/
		len = 0;
}		}
		}

		/*
		* Return the processed chain to the guest, reporting
		* the number of bytes that we read.
		*/
		vq_relchain(vq, idx, len);
		}

/* Called on TX kick. */		/* Called on TX kick. */
static void		static void
pci_vtnet_ping_txq(void vsc, struct vqueue_info vq)		pci_vtnet_ping_txq(void vsc, struct vqueue_info vq)
{		{
struct pci_vtnet_softc *sc = vsc;		struct pci_vtnet_softc *sc = vsc;

/*		/*
* Any ring entries to process?		* Any ring entries to process?
▲ Show 20 Lines • Show All 98 Lines • ▼ Show 20 Lines
#endif		#endif

/*		/*
* Attempt to open the backend device and read the MAC address		* Attempt to open the backend device and read the MAC address
* if specified.		* if specified.
*/		*/
mac_provided = 0;		mac_provided = 0;
if (opts != NULL) {		if (opts != NULL) {
		int mrg_rxbuf_enabled = 0;
		int mrg_rxbuf_disabled = 0;
char *devname;		char *devname;
char *vtopts;		char *vtopts;
int err;		int err = 0;

		/* Get the device name. */
devname = vtopts = strdup(opts);		devname = vtopts = strdup(opts);
(void) strsep(&vtopts, ",");		(void) strsep(&vtopts, ",");

if (vtopts != NULL) {		/*
err = net_parsemac(vtopts, sc->vsc_config.mac);		* Parse the list of options in the form
if (err != 0) {		* key1=value1,...,keyN=valueN.
		*/
		while (vtopts != NULL) {
		char *value = vtopts;
		char *key;

		key = strsep(&value, "=");
		if (value == NULL)
		break;
		vtopts = value;
		(void) strsep(&vtopts, ",");

		if (strcmp(key, "mac") == 0) {
		err = net_parsemac(value, sc->vsc_config.mac);
		if (err)
		break;
		mac_provided = 1;
		} else if (strcmp(key, "mrgrxbuf") == 0) {
		if (strcmp(value, "on") == 0) {
		mrg_rxbuf_enabled = 1;
		mrg_rxbuf_disabled = 0;
		} else {
		mrg_rxbuf_disabled = 1;
		mrg_rxbuf_enabled = 0;
		}
		}
		}

		if (err) {
free(devname);		free(devname);
free(sc);		free(sc);
return (err);		return (err);
}		}
mac_provided = 1;
}

err = netbe_init(&sc->vsc_be, devname, pci_vtnet_rx_callback,		err = netbe_init(&sc->vsc_be, devname, pci_vtnet_rx_callback,
sc);		sc);
free(devname);		free(devname);
if (err) {		if (err) {
free(sc);		free(sc);
return (err);		return (err);
}		}
sc->vsc_consts.vc_hv_caps \|= netbe_get_cap(sc->vsc_be);		sc->vsc_consts.vc_hv_caps \|= netbe_get_cap(sc->vsc_be);
		if (mrg_rxbuf_enabled)
		sc->vsc_consts.vc_hv_caps \|= VIRTIO_NET_F_MRG_RXBUF;
		else if (mrg_rxbuf_disabled)
		sc->vsc_consts.vc_hv_caps &= ~VIRTIO_NET_F_MRG_RXBUF;

}		}

if (!mac_provided) {		if (!mac_provided) {
net_genmac(pi, sc->vsc_config.mac);		net_genmac(pi, sc->vsc_config.mac);
}		}

/* initialize config space */		/* initialize config space */
pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);		pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
Show All 15 Lines	#endif
}		}

/* use BAR 0 to map config regs in IO space */		/* use BAR 0 to map config regs in IO space */
vi_set_io_bar(&sc->vsc_vs, 0);		vi_set_io_bar(&sc->vsc_vs, 0);

sc->resetting = 0;		sc->resetting = 0;

sc->rx_merge = 0;		sc->rx_merge = 0;
		sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
pthread_mutex_init(&sc->rx_mtx, NULL);		pthread_mutex_init(&sc->rx_mtx, NULL);

/*		/*
* Initialize tx semaphore & spawn TX processing thread.		* Initialize tx semaphore & spawn TX processing thread.
* As of now, only one thread for TX desc processing is		* As of now, only one thread for TX desc processing is
* spawned.		* spawned.
*/		*/
sc->tx_in_progress = 0;		sc->tx_in_progress = 0;
Show All 38 Lines	pci_vtnet_cfgread(void vsc, int offset, int size, uint32_t retval)
memcpy(retval, ptr, size);		memcpy(retval, ptr, size);
return (0);		return (0);
}		}

static void		static void
pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)		pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
{		{
struct pci_vtnet_softc *sc = vsc;		struct pci_vtnet_softc *sc = vsc;
unsigned int rx_vhdrlen;

sc->vsc_features = negotiated_features;		sc->vsc_features = negotiated_features;

if (negotiated_features & VIRTIO_NET_F_MRG_RXBUF) {		if (negotiated_features & VIRTIO_NET_F_MRG_RXBUF) {
rx_vhdrlen = sizeof(struct virtio_net_rxhdr);		sc->vhdrlen = sizeof(struct virtio_net_rxhdr);
sc->rx_merge = 1;		sc->rx_merge = 1;
} else {		} else {
/*		/*
* Without mergeable rx buffers, virtio-net header is 2		* Without mergeable rx buffers, virtio-net header is 2
* bytes shorter than sizeof(struct virtio_net_rxhdr).		* bytes shorter than sizeof(struct virtio_net_rxhdr).
*/		*/
rx_vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;		sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
sc->rx_merge = 0;		sc->rx_merge = 0;
}		}

/* Tell the backend to enable some capabilities it has advertised. */		/* Tell the backend to enable some capabilities it has advertised. */
netbe_set_cap(sc->vsc_be, negotiated_features, rx_vhdrlen);		netbe_set_cap(sc->vsc_be, negotiated_features, sc->vhdrlen);
		sc->be_vhdrlen = netbe_get_vnet_hdr_len(sc->vsc_be);
		assert(sc->be_vhdrlen == 0 \|\| sc->be_vhdrlen == sc->vhdrlen);
}		}

static struct pci_devemu pci_de_vnet = {		static struct pci_devemu pci_de_vnet = {
.pe_emu = "virtio-net",		.pe_emu = "virtio-net",
.pe_init = pci_vtnet_init,		.pe_init = pci_vtnet_init,
.pe_barwrite = vi_pci_write,		.pe_barwrite = vi_pci_write,
.pe_barread = vi_pci_read		.pe_barread = vi_pci_read
};		};
PCI_EMUL_SET(pci_de_vnet);		PCI_EMUL_SET(pci_de_vnet);