eal: set name when creating a control thread
[dpdk.git] / lib / librte_vhost / socket.c
index aaa9c27..4a561ad 100644 (file)
@@ -1,39 +1,9 @@
-/*-
- *   BSD LICENSE
- *
- *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
- *   All rights reserved.
- *
- *   Redistribution and use in source and binary forms, with or without
- *   modification, are permitted provided that the following conditions
- *   are met:
- *
- *     * Redistributions of source code must retain the above copyright
- *       notice, this list of conditions and the following disclaimer.
- *     * Redistributions in binary form must reproduce the above copyright
- *       notice, this list of conditions and the following disclaimer in
- *       the documentation and/or other materials provided with the
- *       distribution.
- *     * Neither the name of Intel Corporation nor the names of its
- *       contributors may be used to endorse or promote products derived
- *       from this software without specific prior written permission.
- *
- *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
- *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
- *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
- *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
- *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
- *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
- *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
- *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2016 Intel Corporation
  */
 
 #include <stdint.h>
 #include <stdio.h>
-#include <stdbool.h>
 #include <limits.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include "vhost.h"
 #include "vhost_user.h"
 
+
+TAILQ_HEAD(vhost_user_connection_list, vhost_user_connection);
+
 /*
  * Every time rte_vhost_driver_register() is invoked, an associated
  * vhost_user_socket struct will be created.
  */
 struct vhost_user_socket {
+       struct vhost_user_connection_list conn_list;
+       pthread_mutex_t conn_mutex;
        char *path;
-       int listenfd;
-       int connfd;
+       int socket_fd;
+       struct sockaddr_un un;
        bool is_server;
        bool reconnect;
        bool dequeue_zero_copy;
+       bool iommu_support;
+       bool use_builtin_virtio_net;
+
+       /*
+        * The "supported_features" indicates the feature bits the
+        * vhost driver supports. The "features" indicates the feature
+        * bits after the rte_vhost_driver_features_disable/enable().
+        * It is also the final feature bits used for vhost-user
+        * features negotiation.
+        */
+       uint64_t supported_features;
+       uint64_t features;
+
+       /*
+        * Device id to identify a specific backend device.
+        * It's set to -1 for the default software implementation.
+        * If valid, one socket can have 1 connection only.
+        */
+       int vdpa_dev_id;
+
+       struct vhost_device_ops const *notify_ops;
 };
 
 struct vhost_user_connection {
        struct vhost_user_socket *vsocket;
+       int connfd;
        int vid;
+
+       TAILQ_ENTRY(vhost_user_connection) next;
 };
 
 #define MAX_VHOST_SOCKET 1024
@@ -82,7 +81,8 @@ struct vhost_user {
 
 static void vhost_user_server_new_connection(int fd, void *data, int *remove);
 static void vhost_user_read_cb(int fd, void *dat, int *remove);
-static int vhost_user_create_client(struct vhost_user_socket *vsocket);
+static int create_unix_socket(struct vhost_user_socket *vsocket);
+static int vhost_user_start_client(struct vhost_user_socket *vsocket);
 
 static struct vhost_user vhost_user = {
        .fdset = {
@@ -103,6 +103,7 @@ read_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
        size_t fdsize = fd_num * sizeof(int);
        char control[CMSG_SPACE(fdsize)];
        struct cmsghdr *cmsg;
+       int got_fds = 0;
        int ret;
 
        memset(&msgh, 0, sizeof(msgh));
@@ -129,11 +130,16 @@ read_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
                cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
                if ((cmsg->cmsg_level == SOL_SOCKET) &&
                        (cmsg->cmsg_type == SCM_RIGHTS)) {
-                       memcpy(fds, CMSG_DATA(cmsg), fdsize);
+                       got_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
+                       memcpy(fds, CMSG_DATA(cmsg), got_fds * sizeof(int));
                        break;
                }
        }
 
+       /* Clear out unused file descriptors */
+       while (got_fds < fd_num)
+               fds[got_fds++] = -1;
+
        return ret;
 }
 
@@ -159,6 +165,11 @@ send_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
                msgh.msg_control = control;
                msgh.msg_controllen = sizeof(control);
                cmsg = CMSG_FIRSTHDR(&msgh);
+               if (cmsg == NULL) {
+                       RTE_LOG(ERR, VHOST_CONFIG, "cmsg == NULL\n");
+                       errno = EINVAL;
+                       return -1;
+               }
                cmsg->cmsg_len = CMSG_LEN(fdsize);
                cmsg->cmsg_level = SOL_SOCKET;
                cmsg->cmsg_type = SCM_RIGHTS;
@@ -169,7 +180,7 @@ send_fd_message(int sockfd, char *buf, int buflen, int *fds, int fd_num)
        }
 
        do {
-               ret = sendmsg(sockfd, &msgh, 0);
+               ret = sendmsg(sockfd, &msgh, MSG_NOSIGNAL);
        } while (ret < 0 && errno == EINTR);
 
        if (ret < 0) {
@@ -196,32 +207,57 @@ vhost_user_add_connection(int fd, struct vhost_user_socket *vsocket)
 
        vid = vhost_new_device();
        if (vid == -1) {
-               close(fd);
-               free(conn);
-               return;
+               goto err;
        }
 
        size = strnlen(vsocket->path, PATH_MAX);
        vhost_set_ifname(vid, vsocket->path, size);
 
+       vhost_set_builtin_virtio_net(vid, vsocket->use_builtin_virtio_net);
+
+       vhost_attach_vdpa_device(vid, vsocket->vdpa_dev_id);
+
        if (vsocket->dequeue_zero_copy)
                vhost_enable_dequeue_zero_copy(vid);
 
        RTE_LOG(INFO, VHOST_CONFIG, "new device, handle is %d\n", vid);
 
-       vsocket->connfd = fd;
+       if (vsocket->notify_ops->new_connection) {
+               ret = vsocket->notify_ops->new_connection(vid);
+               if (ret < 0) {
+                       RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to add vhost user connection with fd %d\n",
+                               fd);
+                       goto err;
+               }
+       }
+
+       conn->connfd = fd;
        conn->vsocket = vsocket;
        conn->vid = vid;
        ret = fdset_add(&vhost_user.fdset, fd, vhost_user_read_cb,
                        NULL, conn);
        if (ret < 0) {
-               vsocket->connfd = -1;
-               free(conn);
-               close(fd);
                RTE_LOG(ERR, VHOST_CONFIG,
                        "failed to add fd %d into vhost server fdset\n",
                        fd);
+
+               if (vsocket->notify_ops->destroy_connection)
+                       vsocket->notify_ops->destroy_connection(conn->vid);
+
+               goto err;
        }
+
+       pthread_mutex_lock(&vsocket->conn_mutex);
+       TAILQ_INSERT_TAIL(&vsocket->conn_list, conn, next);
+       pthread_mutex_unlock(&vsocket->conn_mutex);
+
+       fdset_pipe_notify(&vhost_user.fdset);
+       return;
+
+err:
+       free(conn);
+       close(fd);
 }
 
 /* call back when there is new vhost-user connection from client  */
@@ -247,29 +283,39 @@ vhost_user_read_cb(int connfd, void *dat, int *remove)
 
        ret = vhost_user_msg_handler(conn->vid, connfd);
        if (ret < 0) {
-               vsocket->connfd = -1;
                close(connfd);
                *remove = 1;
                vhost_destroy_device(conn->vid);
+
+               if (vsocket->notify_ops->destroy_connection)
+                       vsocket->notify_ops->destroy_connection(conn->vid);
+
+               pthread_mutex_lock(&vsocket->conn_mutex);
+               TAILQ_REMOVE(&vsocket->conn_list, conn, next);
+               pthread_mutex_unlock(&vsocket->conn_mutex);
+
                free(conn);
 
-               if (vsocket->reconnect)
-                       vhost_user_create_client(vsocket);
+               if (vsocket->reconnect) {
+                       create_unix_socket(vsocket);
+                       vhost_user_start_client(vsocket);
+               }
        }
 }
 
 static int
-create_unix_socket(const char *path, struct sockaddr_un *un, bool is_server)
+create_unix_socket(struct vhost_user_socket *vsocket)
 {
        int fd;
+       struct sockaddr_un *un = &vsocket->un;
 
        fd = socket(AF_UNIX, SOCK_STREAM, 0);
        if (fd < 0)
                return -1;
        RTE_LOG(INFO, VHOST_CONFIG, "vhost-user %s: socket created, fd: %d\n",
-               is_server ? "server" : "client", fd);
+               vsocket->is_server ? "server" : "client", fd);
 
-       if (!is_server && fcntl(fd, F_SETFL, O_NONBLOCK)) {
+       if (!vsocket->is_server && fcntl(fd, F_SETFL, O_NONBLOCK)) {
                RTE_LOG(ERR, VHOST_CONFIG,
                        "vhost-user: can't set nonblocking mode for socket, fd: "
                        "%d (%s)\n", fd, strerror(errno));
@@ -279,25 +325,31 @@ create_unix_socket(const char *path, struct sockaddr_un *un, bool is_server)
 
        memset(un, 0, sizeof(*un));
        un->sun_family = AF_UNIX;
-       strncpy(un->sun_path, path, sizeof(un->sun_path));
+       strncpy(un->sun_path, vsocket->path, sizeof(un->sun_path));
        un->sun_path[sizeof(un->sun_path) - 1] = '\0';
 
-       return fd;
+       vsocket->socket_fd = fd;
+       return 0;
 }
 
 static int
-vhost_user_create_server(struct vhost_user_socket *vsocket)
+vhost_user_start_server(struct vhost_user_socket *vsocket)
 {
-       int fd;
        int ret;
-       struct sockaddr_un un;
+       int fd = vsocket->socket_fd;
        const char *path = vsocket->path;
 
-       fd = create_unix_socket(path, &un, vsocket->is_server);
-       if (fd < 0)
-               return -1;
-
-       ret = bind(fd, (struct sockaddr *)&un, sizeof(un));
+       /*
+        * bind () may fail if the socket file with the same name already
+        * exists. But the library obviously should not delete the file
+        * provided by the user, since we can not be sure that it is not
+        * being used by other applications. Moreover, many applications form
+        * socket names based on user input, which is prone to errors.
+        *
+        * The user must ensure that the socket does not exist before
+        * registering the vhost driver in server mode.
+        */
+       ret = bind(fd, (struct sockaddr *)&vsocket->un, sizeof(vsocket->un));
        if (ret < 0) {
                RTE_LOG(ERR, VHOST_CONFIG,
                        "failed to bind to %s: %s; remove it and try again\n",
@@ -310,7 +362,6 @@ vhost_user_create_server(struct vhost_user_socket *vsocket)
        if (ret < 0)
                goto err;
 
-       vsocket->listenfd = fd;
        ret = fdset_add(&vhost_user.fdset, fd, vhost_user_server_new_connection,
                  NULL, vsocket);
        if (ret < 0) {
@@ -417,38 +468,42 @@ vhost_user_reconnect_init(void)
 {
        int ret;
 
-       pthread_mutex_init(&reconn_list.mutex, NULL);
+       ret = pthread_mutex_init(&reconn_list.mutex, NULL);
+       if (ret < 0) {
+               RTE_LOG(ERR, VHOST_CONFIG, "failed to initialize mutex");
+               return ret;
+       }
        TAILQ_INIT(&reconn_list.head);
 
-       ret = pthread_create(&reconn_tid, NULL,
+       ret = rte_ctrl_thread_create(&reconn_tid, "vhost_reconn", NULL,
                             vhost_user_client_reconnect, NULL);
-       if (ret < 0)
+       if (ret != 0) {
                RTE_LOG(ERR, VHOST_CONFIG, "failed to create reconnect thread");
+               if (pthread_mutex_destroy(&reconn_list.mutex)) {
+                       RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to destroy reconnect mutex");
+               }
+       }
 
        return ret;
 }
 
 static int
-vhost_user_create_client(struct vhost_user_socket *vsocket)
+vhost_user_start_client(struct vhost_user_socket *vsocket)
 {
-       int fd;
        int ret;
-       struct sockaddr_un un;
+       int fd = vsocket->socket_fd;
        const char *path = vsocket->path;
        struct vhost_user_reconnect *reconn;
 
-       fd = create_unix_socket(path, &un, vsocket->is_server);
-       if (fd < 0)
-               return -1;
-
-       ret = vhost_user_connect_nonblock(fd, (struct sockaddr *)&un,
-                                         sizeof(un));
+       ret = vhost_user_connect_nonblock(fd, (struct sockaddr *)&vsocket->un,
+                                         sizeof(vsocket->un));
        if (ret == 0) {
                vhost_user_add_connection(fd, vsocket);
                return 0;
        }
 
-       RTE_LOG(ERR, VHOST_CONFIG,
+       RTE_LOG(WARNING, VHOST_CONFIG,
                "failed to connect to %s: %s\n",
                path, strerror(errno));
 
@@ -457,7 +512,7 @@ vhost_user_create_client(struct vhost_user_socket *vsocket)
                return -1;
        }
 
-       RTE_LOG(ERR, VHOST_CONFIG, "%s: reconnecting...\n", path);
+       RTE_LOG(INFO, VHOST_CONFIG, "%s: reconnecting...\n", path);
        reconn = malloc(sizeof(*reconn));
        if (reconn == NULL) {
                RTE_LOG(ERR, VHOST_CONFIG,
@@ -465,7 +520,7 @@ vhost_user_create_client(struct vhost_user_socket *vsocket)
                close(fd);
                return -1;
        }
-       reconn->un = un;
+       reconn->un = vsocket->un;
        reconn->fd = fd;
        reconn->vsocket = vsocket;
        pthread_mutex_lock(&reconn_list.mutex);
@@ -475,6 +530,254 @@ vhost_user_create_client(struct vhost_user_socket *vsocket)
        return 0;
 }
 
+static struct vhost_user_socket *
+find_vhost_user_socket(const char *path)
+{
+       int i;
+
+       for (i = 0; i < vhost_user.vsocket_cnt; i++) {
+               struct vhost_user_socket *vsocket = vhost_user.vsockets[i];
+
+               if (!strcmp(vsocket->path, path))
+                       return vsocket;
+       }
+
+       return NULL;
+}
+
+int
+rte_vhost_driver_attach_vdpa_device(const char *path, int did)
+{
+       struct vhost_user_socket *vsocket;
+
+       if (rte_vdpa_get_device(did) == NULL)
+               return -1;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket)
+               vsocket->vdpa_dev_id = did;
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+int
+rte_vhost_driver_detach_vdpa_device(const char *path)
+{
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket)
+               vsocket->vdpa_dev_id = -1;
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+int
+rte_vhost_driver_get_vdpa_device_id(const char *path)
+{
+       struct vhost_user_socket *vsocket;
+       int did = -1;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket)
+               did = vsocket->vdpa_dev_id;
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return did;
+}
+
+int
+rte_vhost_driver_disable_features(const char *path, uint64_t features)
+{
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+
+       /* Note that use_builtin_virtio_net is not affected by this function
+        * since callers may want to selectively disable features of the
+        * built-in vhost net device backend.
+        */
+
+       if (vsocket)
+               vsocket->features &= ~features;
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+int
+rte_vhost_driver_enable_features(const char *path, uint64_t features)
+{
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket) {
+               if ((vsocket->supported_features & features) != features) {
+                       /*
+                        * trying to enable features the driver doesn't
+                        * support.
+                        */
+                       pthread_mutex_unlock(&vhost_user.mutex);
+                       return -1;
+               }
+               vsocket->features |= features;
+       }
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+int
+rte_vhost_driver_set_features(const char *path, uint64_t features)
+{
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket) {
+               vsocket->supported_features = features;
+               vsocket->features = features;
+
+               /* Anyone setting feature bits is implementing their own vhost
+                * device backend.
+                */
+               vsocket->use_builtin_virtio_net = false;
+       }
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+int
+rte_vhost_driver_get_features(const char *path, uint64_t *features)
+{
+       struct vhost_user_socket *vsocket;
+       uint64_t vdpa_features;
+       struct rte_vdpa_device *vdpa_dev;
+       int did = -1;
+       int ret = 0;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (!vsocket) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "socket file %s is not registered yet.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       did = vsocket->vdpa_dev_id;
+       vdpa_dev = rte_vdpa_get_device(did);
+       if (!vdpa_dev || !vdpa_dev->ops->get_features) {
+               *features = vsocket->features;
+               goto unlock_exit;
+       }
+
+       if (vdpa_dev->ops->get_features(did, &vdpa_features) < 0) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to get vdpa features "
+                               "for socket file %s.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       *features = vsocket->features & vdpa_features;
+
+unlock_exit:
+       pthread_mutex_unlock(&vhost_user.mutex);
+       return ret;
+}
+
+int
+rte_vhost_driver_get_protocol_features(const char *path,
+               uint64_t *protocol_features)
+{
+       struct vhost_user_socket *vsocket;
+       uint64_t vdpa_protocol_features;
+       struct rte_vdpa_device *vdpa_dev;
+       int did = -1;
+       int ret = 0;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (!vsocket) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "socket file %s is not registered yet.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       did = vsocket->vdpa_dev_id;
+       vdpa_dev = rte_vdpa_get_device(did);
+       if (!vdpa_dev || !vdpa_dev->ops->get_protocol_features) {
+               *protocol_features = VHOST_USER_PROTOCOL_FEATURES;
+               goto unlock_exit;
+       }
+
+       if (vdpa_dev->ops->get_protocol_features(did,
+                               &vdpa_protocol_features) < 0) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to get vdpa protocol features "
+                               "for socket file %s.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       *protocol_features = VHOST_USER_PROTOCOL_FEATURES
+               & vdpa_protocol_features;
+
+unlock_exit:
+       pthread_mutex_unlock(&vhost_user.mutex);
+       return ret;
+}
+
+int
+rte_vhost_driver_get_queue_num(const char *path, uint32_t *queue_num)
+{
+       struct vhost_user_socket *vsocket;
+       uint32_t vdpa_queue_num;
+       struct rte_vdpa_device *vdpa_dev;
+       int did = -1;
+       int ret = 0;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (!vsocket) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "socket file %s is not registered yet.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       did = vsocket->vdpa_dev_id;
+       vdpa_dev = rte_vdpa_get_device(did);
+       if (!vdpa_dev || !vdpa_dev->ops->get_queue_num) {
+               *queue_num = VHOST_MAX_QUEUE_PAIRS;
+               goto unlock_exit;
+       }
+
+       if (vdpa_dev->ops->get_queue_num(did, &vdpa_queue_num) < 0) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to get vdpa queue number "
+                               "for socket file %s.\n", path);
+               ret = -1;
+               goto unlock_exit;
+       }
+
+       *queue_num = RTE_MIN((uint32_t)VHOST_MAX_QUEUE_PAIRS, vdpa_queue_num);
+
+unlock_exit:
+       pthread_mutex_unlock(&vhost_user.mutex);
+       return ret;
+}
+
 /*
  * Register a new vhost-user socket; here we could act as server
  * (the default case), or client (when RTE_VHOST_USER_CLIENT) flag
@@ -502,31 +805,69 @@ rte_vhost_driver_register(const char *path, uint64_t flags)
                goto out;
        memset(vsocket, 0, sizeof(struct vhost_user_socket));
        vsocket->path = strdup(path);
-       vsocket->connfd = -1;
+       if (vsocket->path == NULL) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "error: failed to copy socket path string\n");
+               free(vsocket);
+               goto out;
+       }
+       TAILQ_INIT(&vsocket->conn_list);
+       ret = pthread_mutex_init(&vsocket->conn_mutex, NULL);
+       if (ret) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "error: failed to init connection mutex\n");
+               goto out_free;
+       }
        vsocket->dequeue_zero_copy = flags & RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
 
+       /*
+        * Set the supported features correctly for the builtin vhost-user
+        * net driver.
+        *
+        * Applications know nothing about features the builtin virtio net
+        * driver (virtio_net.c) supports, thus it's not possible for them
+        * to invoke rte_vhost_driver_set_features(). To workaround it, here
+        * we set it unconditionally. If the application want to implement
+        * another vhost-user driver (say SCSI), it should call the
+        * rte_vhost_driver_set_features(), which will overwrite following
+        * two values.
+        */
+       vsocket->use_builtin_virtio_net = true;
+       vsocket->supported_features = VIRTIO_NET_SUPPORTED_FEATURES;
+       vsocket->features           = VIRTIO_NET_SUPPORTED_FEATURES;
+
+       if (!(flags & RTE_VHOST_USER_IOMMU_SUPPORT)) {
+               vsocket->supported_features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
+               vsocket->features &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
+       }
+
        if ((flags & RTE_VHOST_USER_CLIENT) != 0) {
                vsocket->reconnect = !(flags & RTE_VHOST_USER_NO_RECONNECT);
                if (vsocket->reconnect && reconn_tid == 0) {
-                       if (vhost_user_reconnect_init() < 0) {
-                               free(vsocket->path);
-                               free(vsocket);
-                               goto out;
-                       }
+                       if (vhost_user_reconnect_init() != 0)
+                               goto out_mutex;
                }
-               ret = vhost_user_create_client(vsocket);
        } else {
                vsocket->is_server = true;
-               ret = vhost_user_create_server(vsocket);
        }
+       ret = create_unix_socket(vsocket);
        if (ret < 0) {
-               free(vsocket->path);
-               free(vsocket);
-               goto out;
+               goto out_mutex;
        }
 
        vhost_user.vsockets[vhost_user.vsocket_cnt++] = vsocket;
 
+       pthread_mutex_unlock(&vhost_user.mutex);
+       return ret;
+
+out_mutex:
+       if (pthread_mutex_destroy(&vsocket->conn_mutex)) {
+               RTE_LOG(ERR, VHOST_CONFIG,
+                       "error: failed to destroy connection mutex\n");
+       }
+out_free:
+       free(vsocket->path);
+       free(vsocket);
 out:
        pthread_mutex_unlock(&vhost_user.mutex);
 
@@ -565,7 +906,7 @@ rte_vhost_driver_unregister(const char *path)
 {
        int i;
        int count;
-       struct vhost_user_connection *conn;
+       struct vhost_user_connection *conn, *next;
 
        pthread_mutex_lock(&vhost_user.mutex);
 
@@ -574,23 +915,31 @@ rte_vhost_driver_unregister(const char *path)
 
                if (!strcmp(vsocket->path, path)) {
                        if (vsocket->is_server) {
-                               fdset_del(&vhost_user.fdset, vsocket->listenfd);
-                               close(vsocket->listenfd);
+                               fdset_del(&vhost_user.fdset, vsocket->socket_fd);
+                               close(vsocket->socket_fd);
                                unlink(path);
                        } else if (vsocket->reconnect) {
                                vhost_user_remove_reconnect(vsocket);
                        }
 
-                       conn = fdset_del(&vhost_user.fdset, vsocket->connfd);
-                       if (conn) {
+                       pthread_mutex_lock(&vsocket->conn_mutex);
+                       for (conn = TAILQ_FIRST(&vsocket->conn_list);
+                            conn != NULL;
+                            conn = next) {
+                               next = TAILQ_NEXT(conn, next);
+
+                               fdset_del(&vhost_user.fdset, conn->connfd);
                                RTE_LOG(INFO, VHOST_CONFIG,
                                        "free connfd = %d for device '%s'\n",
-                                       vsocket->connfd, path);
-                               close(vsocket->connfd);
+                                       conn->connfd, path);
+                               close(conn->connfd);
                                vhost_destroy_device(conn->vid);
+                               TAILQ_REMOVE(&vsocket->conn_list, conn, next);
                                free(conn);
                        }
+                       pthread_mutex_unlock(&vsocket->conn_mutex);
 
+                       pthread_mutex_destroy(&vsocket->conn_mutex);
                        free(vsocket->path);
                        free(vsocket);
 
@@ -607,9 +956,74 @@ rte_vhost_driver_unregister(const char *path)
        return -1;
 }
 
+/*
+ * Register ops so that we can add/remove device to data core.
+ */
 int
-rte_vhost_driver_session_start(void)
+rte_vhost_driver_callback_register(const char *path,
+       struct vhost_device_ops const * const ops)
 {
-       fdset_event_dispatch(&vhost_user.fdset);
-       return 0;
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       if (vsocket)
+               vsocket->notify_ops = ops;
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? 0 : -1;
+}
+
+struct vhost_device_ops const *
+vhost_driver_callback_get(const char *path)
+{
+       struct vhost_user_socket *vsocket;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       return vsocket ? vsocket->notify_ops : NULL;
+}
+
+int
+rte_vhost_driver_start(const char *path)
+{
+       struct vhost_user_socket *vsocket;
+       static pthread_t fdset_tid;
+
+       pthread_mutex_lock(&vhost_user.mutex);
+       vsocket = find_vhost_user_socket(path);
+       pthread_mutex_unlock(&vhost_user.mutex);
+
+       if (!vsocket)
+               return -1;
+
+       if (fdset_tid == 0) {
+               /**
+                * create a pipe which will be waited by poll and notified to
+                * rebuild the wait list of poll.
+                */
+               if (fdset_pipe_init(&vhost_user.fdset) < 0) {
+                       RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to create pipe for vhost fdset\n");
+                       return -1;
+               }
+
+               int ret = rte_ctrl_thread_create(&fdset_tid,
+                       "vhost-events", NULL, fdset_event_dispatch,
+                       &vhost_user.fdset);
+               if (ret != 0) {
+                       RTE_LOG(ERR, VHOST_CONFIG,
+                               "failed to create fdset handling thread");
+
+                       fdset_pipe_uninit(&vhost_user.fdset);
+                       return -1;
+               }
+       }
+
+       if (vsocket->is_server)
+               return vhost_user_start_server(vsocket);
+       else
+               return vhost_user_start_client(vsocket);
 }