Merge pull request #171 from libos-nuse/feature-virtio-rawsock

add raw socket (AF_PACKET) backend for virtio-net host

tools/lkl/include/lkl.h

@@ -301,6 +301,14 @@ struct lkl_netdev *lkl_netdev_dpdk_create(const char *ifname);
  */
 struct lkl_netdev *lkl_netdev_vde_create(const char *switch_path);
 
+/**
+ * lkl_netdev_raw_create - create raw socket net_device for the virtio net
+ *                         backend
+ *
+ * @ifname - interface name for the snoop device.
+ */
+struct lkl_netdev *lkl_netdev_raw_create(const char *ifname);
+
 /*
  * lkl_register_dbg_handler- register a signal handler that loads a debug lib.
  *

tools/lkl/lib/Build

@@ -12,6 +12,8 @@ lkl-y += virtio.o
 lkl-y += dbg.o
 lkl-y += dbg_handler.o
 lkl-$(CONFIG_AUTO_LKL_POSIX_HOST) += virtio_net.o
+lkl-$(CONFIG_AUTO_LKL_POSIX_HOST) += virtio_net_linux_fdnet.o
 lkl-$(CONFIG_AUTO_LKL_POSIX_HOST) += virtio_net_tap.o
+lkl-$(CONFIG_AUTO_LKL_POSIX_HOST) += virtio_net_raw.o
 lkl-$(CONFIG_AUTO_LKL_VIRTIO_NET_DPDK) += virtio_net_dpdk.o
 lkl-$(CONFIG_AUTO_LKL_VIRTIO_NET_VDE) += virtio_net_vde.o

tools/lkl/lib/hijack/hijack.c

@@ -158,7 +158,7 @@ HOST_CALL(socket);
 int socket(int domain, int type, int protocol)
 {
 	CHECK_HOST_CALL(socket);
-	if (domain == AF_UNIX)
+	if (domain == AF_UNIX || domain == PF_PACKET)
 		return host_socket(domain, type, protocol);
 
 	return lkl_call(__lkl__NR_socket, 3, domain, type, protocol);

tools/lkl/lib/hijack/init.c

@@ -25,7 +25,7 @@
 #include <lkl_host.h>
 
 #include "xlate.h"
-#include "../virtio_net_tap.h"
+#include "../virtio_net_linux_fdnet.h"
 
 #define __USE_GNU
 #include <dlfcn.h>
@@ -170,11 +170,11 @@ static void mount_cmds_exec(char *_cmds, int (*callback)(char*))
 	free(cmds);
 }
 
-void fixup_netdev_tap_ops(void)
+void fixup_netdev_linux_fdnet_ops(void)
 {
 	/* It's okay if this is NULL, because then netdev close will
 	 * fall back onto an uncloseable implementation. */
-	lkl_netdev_tap_ops.eventfd = dlsym(RTLD_NEXT, "eventfd");
+	lkl_netdev_linux_fdnet_ops.eventfd = dlsym(RTLD_NEXT, "eventfd");
 }
 
 void __attribute__((constructor(102)))
@@ -197,7 +197,7 @@ hijack_init(void)
 	char *arp_entries = getenv("LKL_HIJACK_NET_ARP");
 
 	/* Must be run before lkl_netdev_tap_create */
-	fixup_netdev_tap_ops();
+	fixup_netdev_linux_fdnet_ops();
 
 	if (tap) {
 		fprintf(stderr,
@@ -214,6 +214,8 @@ hijack_init(void)
 			nd = lkl_netdev_dpdk_create(ifparams);
 		else if (strcmp(iftype, "vde") == 0)
 			nd = lkl_netdev_vde_create(ifparams);
+		else if (strcmp(iftype, "raw") == 0)
+			nd = lkl_netdev_raw_create(ifparams);
 	}
 
 	if (nd) {

tools/lkl/lib/virtio_net_linux_fdnet.c

@@ -0,0 +1,251 @@
+/*
+ * Linux File descripter based virtual network interface feature for LKL
+ * Copyright (c) 2015,2016 Ryo Nakamura, Hajime Tazaki
+ *
+ * Author: Ryo Nakamura <upa@wide.ad.jp>
+ *         Hajime Tazaki <thehajime@gmail.com>
+ *         Octavian Purdila <octavian.purdila@intel.com>
+ *
+ * Current implementation is linux-specific.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <string.h>
+#include <sys/epoll.h>
+#include <sys/eventfd.h>
+
+#include "virtio.h"
+#include "virtio_net_linux_fdnet.h"
+
+struct lkl_netdev_linux_fdnet_ops lkl_netdev_linux_fdnet_ops = {
+	/*
+	 * /dev/net/tun is Linux specific so we know our host is some
+	 * flavor of Linux, but this allows graceful support if we're
+	 * on a kernel that's < 2.6.22.
+	 */
+	#ifdef __NR_eventfd
+	/* This sigature was recently (9/2014) changed in glibc. */
+	.eventfd = (int (*)(unsigned int, int))eventfd,
+	#endif /* __NR_eventfd */
+};
+
+static int linux_fdnet_net_tx(struct lkl_netdev *nd, void *data, int len)
+{
+	int ret;
+	struct lkl_netdev_linux_fdnet *nd_fdnet =
+		container_of(nd, struct lkl_netdev_linux_fdnet, dev);
+
+	do {
+		ret = write(nd_fdnet->fd, data, len);
+	} while (ret == -1 && errno == EINVAL);
+	if (ret > 0)
+		return 0;
+	if (ret < 0 && errno != EAGAIN)
+		perror("write to Linux fd netdev fails");
+
+	return -1;
+}
+
+static int linux_fdnet_net_rx(struct lkl_netdev *nd, void *data, int *len)
+{
+	int ret;
+	struct lkl_netdev_linux_fdnet *nd_fdnet =
+		container_of(nd, struct lkl_netdev_linux_fdnet, dev);
+
+	do {
+		ret = read(nd_fdnet->fd, data, *len);
+	} while (ret == -1 && errno == EINVAL);
+	if (ret > 0) {
+		*len = ret;
+		return 0;
+	}
+	if (ret < 0 && errno != EAGAIN)
+		perror("read from fdnet device fails");
+
+	return -1;
+}
+
+static int linux_fdnet_net_poll(struct lkl_netdev *nd, int events)
+{
+	struct lkl_netdev_linux_fdnet *nd_fdnet =
+		container_of(nd, struct lkl_netdev_linux_fdnet, dev);
+	int epoll_fd = -1;
+	struct epoll_event ev[2];
+	int ret;
+	const int is_rx = events & LKL_DEV_NET_POLL_RX;
+	const int is_tx = events & LKL_DEV_NET_POLL_TX;
+	int i;
+	int ret_ev = 0;
+	unsigned int event;
+
+	if (is_rx && is_tx) {
+		fprintf(stderr, "both LKL_DEV_NET_POLL_RX and "
+			"LKL_DEV_NET_POLL_TX are set\n");
+		lkl_host_ops.panic();
+		return -1;
+	}
+	if (!is_rx && !is_tx) {
+		fprintf(stderr, "Neither LKL_DEV_NET_POLL_RX nor"
+			" LKL_DEV_NET_POLL_TX are set.\n");
+		lkl_host_ops.panic();
+		return -1;
+	}
+
+	if (is_rx)
+		epoll_fd = nd_fdnet->epoll_rx_fd;
+	else if (is_tx)
+		epoll_fd = nd_fdnet->epoll_tx_fd;
+
+	do {
+		ret = epoll_wait(epoll_fd, ev, 2, -1);
+	} while (ret == -1 && errno == EINTR);
+	if (ret < 0) {
+		perror("epoll_wait");
+		return -1;
+	}
+
+	for (i = 0; i < ret; ++i) {
+		if (ev[i].data.fd == nd_fdnet->eventfd)
+			return -1;
+		if (ev[i].data.fd == nd_fdnet->fd) {
+			event = ev[i].events;
+			if (event & (EPOLLIN | EPOLLPRI))
+				ret_ev = LKL_DEV_NET_POLL_RX;
+			else if (event & EPOLLOUT)
+				ret_ev = LKL_DEV_NET_POLL_TX;
+			else
+				return -1;
+		}
+	}
+	return ret_ev;
+}
+
+static int linux_fdnet_net_close(struct lkl_netdev *nd)
+{
+	long buf = 1;
+	struct lkl_netdev_linux_fdnet *nd_fdnet =
+		container_of(nd, struct lkl_netdev_linux_fdnet, dev);
+
+	if (nd_fdnet->eventfd == -1) {
+		/* No eventfd support. */
+		return 0;
+	}
+
+	if (write(nd_fdnet->eventfd, &buf, sizeof(buf)) < 0) {
+		perror("linux-fdnet: failed to close fd");
+		/* This should never happen. */
+		return -1;
+	}
+
+	/* The order that we join in doesn't matter. */
+	if (lkl_host_ops.thread_join(nd->rx_tid) ||
+		lkl_host_ops.thread_join(nd->tx_tid))
+		return -1;
+
+	/* nor does the order that we close */
+	if (close(nd_fdnet->fd) || close(nd_fdnet->eventfd) ||
+		close(nd_fdnet->epoll_rx_fd) || close(nd_fdnet->epoll_tx_fd)) {
+		perror("linux-fdnet net_close fd");
+		return -1;
+	}
+
+	return 0;
+}
+
+struct lkl_dev_net_ops linux_fdnet_net_ops =  {
+	.tx = linux_fdnet_net_tx,
+	.rx = linux_fdnet_net_rx,
+	.poll = linux_fdnet_net_poll,
+	.close = linux_fdnet_net_close,
+};
+
+static int add_to_epoll(int epoll_fd, int fd, unsigned int events)
+{
+	struct epoll_event ev;
+	int ret;
+
+	memset(&ev, 0, sizeof(ev));
+	ev.events = events;
+	ev.data.fd = fd;
+	ret = epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fd, &ev);
+	if (ret) {
+		perror("EPOLL_CTL_ADD fails");
+		return -1;
+	}
+	return 0;
+}
+
+static int create_epoll_fd(int fd, unsigned int events)
+{
+	int ret = epoll_create1(0);
+
+	if (ret < 0) {
+		perror("epoll_create1");
+		return -1;
+	}
+	if (add_to_epoll(ret, fd, events)) {
+		close(ret);
+		return -1;
+	}
+	return ret;
+}
+
+
+struct lkl_netdev_linux_fdnet *lkl_register_netdev_linux_fdnet(int fd)
+{
+	struct lkl_netdev_linux_fdnet *nd;
+
+	nd = (struct lkl_netdev_linux_fdnet *)
+		malloc(sizeof(struct lkl_netdev_linux_fdnet));
+	if (!nd) {
+		fprintf(stderr, "fdnet: failed to allocate memory\n");
+		/* TODO: propagate the error state, maybe use errno for that? */
+		return NULL;
+	}
+
+	nd->fd = fd;
+	/* Making them edge-triggered to save CPU. */
+	nd->epoll_rx_fd = create_epoll_fd(nd->fd, EPOLLIN | EPOLLPRI | EPOLLET);
+	nd->epoll_tx_fd = create_epoll_fd(nd->fd, EPOLLOUT | EPOLLET);
+	if (nd->epoll_rx_fd < 0 || nd->epoll_tx_fd < 0) {
+		if (nd->epoll_rx_fd >= 0)
+			close(nd->epoll_rx_fd);
+		if (nd->epoll_tx_fd >= 0)
+			close(nd->epoll_tx_fd);
+		lkl_unregister_netdev_linux_fdnet(nd);
+		return NULL;
+	}
+
+	if (lkl_netdev_linux_fdnet_ops.eventfd) {
+		/* eventfd is supported by the host, all is well */
+		nd->eventfd = lkl_netdev_linux_fdnet_ops.eventfd(
+			0, EFD_NONBLOCK | EFD_SEMAPHORE);
+
+		if (nd->eventfd < 0) {
+			perror("fdnet: create eventfd");
+			lkl_unregister_netdev_linux_fdnet(nd);
+			return NULL;
+		}
+		if (add_to_epoll(nd->epoll_rx_fd, nd->eventfd, EPOLLIN) ||
+			add_to_epoll(nd->epoll_tx_fd, nd->eventfd, EPOLLIN)) {
+			lkl_unregister_netdev_linux_fdnet(nd);
+			return NULL;
+		}
+	} else {
+		/* no host eventfd support */
+		nd->eventfd = -1;
+	}
+
+	nd->dev.ops = &linux_fdnet_net_ops;
+	return nd;
+}
+
+void lkl_unregister_netdev_linux_fdnet(struct lkl_netdev_linux_fdnet *nd)
+{
+	close(nd->eventfd);
+	close(nd->epoll_rx_fd);
+	close(nd->epoll_tx_fd);
+	free(nd);
+}

tools/lkl/lib/virtio_net_linux_fdnet.h

@@ -0,0 +1,41 @@
+#ifndef _VIRTIO_NET_LINUX_FDNET_H
+#define _VIRTIO_NET_LINUX_FDNET_H
+
+struct lkl_netdev_linux_fdnet {
+	struct lkl_netdev dev;
+	/* file-descriptor based device */
+	int fd;
+	/* Needed to initiate shutdown */
+	int eventfd;
+	/* epoll fds for rx and tx */
+	int epoll_rx_fd;
+	int epoll_tx_fd;
+};
+
+extern struct lkl_netdev_linux_fdnet_ops {
+	/*
+	 * We need this so that we can "unhijack" this function in
+	 * case we decided to hijack it.
+	 */
+	int (*eventfd)(unsigned int initval, int flags);
+} lkl_netdev_linux_fdnet_ops;
+
+/**
+ * lkl_register_netdev_linux_fdnet - register a file descriptor-based network
+ * device as a NIC
+ *
+ * @fd - a POSIX file descriptor number for input/output
+ * @returns a struct lkl_netdev_linux_fdnet entry for virtio-net
+ */
+struct lkl_netdev_linux_fdnet *lkl_register_netdev_linux_fdnet(int fd);
+
+
+/**
+ * lkl_unregister_netdev_linux_fdnet - unregister a file descriptor-based
+ * network device as a NIC
+ *
+ * @nd - a struct lkl_netdev_linux_fdnet entry to be unregistered
+ */
+void lkl_unregister_netdev_linux_fdnet(struct lkl_netdev_linux_fdnet *nd);
+
+#endif /* _VIRTIO_NET_LINUX_FDNET_H*/