vhost: support interrupt mode
authorJunjie Chen <junjie.j.chen@intel.com>
Thu, 12 Apr 2018 16:28:34 +0000 (12:28 -0400)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 13 Apr 2018 22:43:30 +0000 (00:43 +0200)
In some cases we want vhost dequeue work in interrupt mode to
release cpus to others when no data to transmit. So we install
interrupt handler of vhost device and interrupt vectors for each
rx queue when creating new backend according to vhost interrupt
configuration. Thus, applications could register a epoll event fd
to associate rx queues with interrupt vectors.

Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
Reviewed-by: Jianfeng Tan <jianfeng.tan@intel.com>
drivers/net/vhost/rte_eth_vhost.c
lib/librte_vhost/vhost.c

index bd5e944..d7d44a0 100644 (file)
@@ -553,6 +553,136 @@ find_internal_resource(char *ifname)
        return list;
 }
 
+static int
+eth_rxq_intr_enable(struct rte_eth_dev *dev, uint16_t qid)
+{
+       struct rte_vhost_vring vring;
+       struct vhost_queue *vq;
+       int ret = 0;
+
+       vq = dev->data->rx_queues[qid];
+       if (!vq) {
+               RTE_LOG(ERR, PMD, "rxq%d is not setup yet\n", qid);
+               return -1;
+       }
+
+       ret = rte_vhost_get_vhost_vring(vq->vid, (qid << 1) + 1, &vring);
+       if (ret < 0) {
+               RTE_LOG(ERR, PMD, "Failed to get rxq%d's vring\n", qid);
+               return ret;
+       }
+       RTE_LOG(INFO, PMD, "Enable interrupt for rxq%d\n", qid);
+       rte_vhost_enable_guest_notification(vq->vid, (qid << 1) + 1, 1);
+       rte_wmb();
+
+       return ret;
+}
+
+static int
+eth_rxq_intr_disable(struct rte_eth_dev *dev, uint16_t qid)
+{
+       struct rte_vhost_vring vring;
+       struct vhost_queue *vq;
+       int ret = 0;
+
+       vq = dev->data->rx_queues[qid];
+       if (!vq) {
+               RTE_LOG(ERR, PMD, "rxq%d is not setup yet\n", qid);
+               return -1;
+       }
+
+       ret = rte_vhost_get_vhost_vring(vq->vid, (qid << 1) + 1, &vring);
+       if (ret < 0) {
+               RTE_LOG(ERR, PMD, "Failed to get rxq%d's vring", qid);
+               return ret;
+       }
+       RTE_LOG(INFO, PMD, "Disable interrupt for rxq%d\n", qid);
+       rte_vhost_enable_guest_notification(vq->vid, (qid << 1) + 1, 0);
+       rte_wmb();
+
+       return 0;
+}
+
+static void
+eth_vhost_uninstall_intr(struct rte_eth_dev *dev)
+{
+       struct rte_intr_handle *intr_handle = dev->intr_handle;
+
+       if (intr_handle) {
+               if (intr_handle->intr_vec)
+                       free(intr_handle->intr_vec);
+               free(intr_handle);
+       }
+
+       dev->intr_handle = NULL;
+}
+
+static int
+eth_vhost_install_intr(struct rte_eth_dev *dev)
+{
+       struct rte_vhost_vring vring;
+       struct vhost_queue *vq;
+       int count = 0;
+       int nb_rxq = dev->data->nb_rx_queues;
+       int i;
+       int ret;
+
+       /* uninstall firstly if we are reconnecting */
+       if (dev->intr_handle)
+               eth_vhost_uninstall_intr(dev);
+
+       dev->intr_handle = malloc(sizeof(*dev->intr_handle));
+       if (!dev->intr_handle) {
+               RTE_LOG(ERR, PMD, "Fail to allocate intr_handle\n");
+               return -ENOMEM;
+       }
+       memset(dev->intr_handle, 0, sizeof(*dev->intr_handle));
+
+       dev->intr_handle->efd_counter_size = sizeof(uint64_t);
+
+       dev->intr_handle->intr_vec =
+               malloc(nb_rxq * sizeof(dev->intr_handle->intr_vec[0]));
+
+       if (!dev->intr_handle->intr_vec) {
+               RTE_LOG(ERR, PMD,
+                       "Failed to allocate memory for interrupt vector\n");
+               free(dev->intr_handle);
+               return -ENOMEM;
+       }
+
+       RTE_LOG(INFO, PMD, "Prepare intr vec\n");
+       for (i = 0; i < nb_rxq; i++) {
+               vq = dev->data->rx_queues[i];
+               if (!vq) {
+                       RTE_LOG(INFO, PMD, "rxq-%d not setup yet, skip!\n", i);
+                       continue;
+               }
+
+               ret = rte_vhost_get_vhost_vring(vq->vid, (i << 1) + 1, &vring);
+               if (ret < 0) {
+                       RTE_LOG(INFO, PMD,
+                               "Failed to get rxq-%d's vring, skip!\n", i);
+                       continue;
+               }
+
+               if (vring.kickfd < 0) {
+                       RTE_LOG(INFO, PMD,
+                               "rxq-%d's kickfd is invalid, skip!\n", i);
+                       continue;
+               }
+               dev->intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + i;
+               dev->intr_handle->efds[i] = vring.kickfd;
+               count++;
+               RTE_LOG(INFO, PMD, "Installed intr vec for rxq-%d\n", i);
+       }
+
+       dev->intr_handle->nb_efd = count;
+       dev->intr_handle->max_intr = count + 1;
+       dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
+
+       return 0;
+}
+
 static void
 update_queuing_status(struct rte_eth_dev *dev)
 {
@@ -618,6 +748,7 @@ new_device(int vid)
        struct rte_eth_dev *eth_dev;
        struct internal_list *list;
        struct pmd_internal *internal;
+       struct rte_eth_conf *dev_conf;
        unsigned i;
        char ifname[PATH_MAX];
 #ifdef RTE_LIBRTE_VHOST_NUMA
@@ -633,6 +764,7 @@ new_device(int vid)
 
        eth_dev = list->eth_dev;
        internal = eth_dev->data->dev_private;
+       dev_conf = &eth_dev->data->dev_conf;
 
 #ifdef RTE_LIBRTE_VHOST_NUMA
        newnode = rte_vhost_get_numa_node(vid);
@@ -641,10 +773,19 @@ new_device(int vid)
 #endif
 
        internal->vid = vid;
-       if (rte_atomic32_read(&internal->started) == 1)
+       if (rte_atomic32_read(&internal->started) == 1) {
                queue_setup(eth_dev, internal);
-       else
+
+               if (dev_conf->intr_conf.rxq) {
+                       if (eth_vhost_install_intr(eth_dev) < 0) {
+                               RTE_LOG(INFO, PMD,
+                                       "Failed to install interrupt handler.");
+                                       return -1;
+                       }
+               }
+       } else {
                RTE_LOG(INFO, PMD, "RX/TX queues not exist yet\n");
+       }
 
        for (i = 0; i < rte_vhost_get_vring_num(vid); i++)
                rte_vhost_enable_guest_notification(vid, i, 0);
@@ -713,6 +854,7 @@ destroy_device(int vid)
        rte_spinlock_unlock(&state->lock);
 
        RTE_LOG(INFO, PMD, "Vhost device %d destroyed\n", vid);
+       eth_vhost_uninstall_intr(eth_dev);
 
        _rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
 }
@@ -824,8 +966,20 @@ static int
 eth_dev_start(struct rte_eth_dev *eth_dev)
 {
        struct pmd_internal *internal = eth_dev->data->dev_private;
+       struct rte_eth_conf *dev_conf = &eth_dev->data->dev_conf;
 
        queue_setup(eth_dev, internal);
+
+       if (rte_atomic32_read(&internal->dev_attached) == 1) {
+               if (dev_conf->intr_conf.rxq) {
+                       if (eth_vhost_install_intr(eth_dev) < 0) {
+                               RTE_LOG(INFO, PMD,
+                                       "Failed to install interrupt handler.");
+                                       return -1;
+                       }
+               }
+       }
+
        rte_atomic32_set(&internal->started, 1);
        update_queuing_status(eth_dev);
 
@@ -1062,6 +1216,8 @@ static const struct eth_dev_ops ops = {
        .xstats_reset = vhost_dev_xstats_reset,
        .xstats_get = vhost_dev_xstats_get,
        .xstats_get_names = vhost_dev_xstats_get_names,
+       .rx_queue_intr_enable = eth_rxq_intr_enable,
+       .rx_queue_intr_disable = eth_rxq_intr_disable,
 };
 
 static struct rte_vdev_driver pmd_vhost_drv;
index 8ad9ea8..5ddf55e 100644 (file)
@@ -577,16 +577,14 @@ rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
 {
        struct virtio_net *dev = get_device(vid);
 
-       if (dev == NULL)
-               return -1;
-
-       if (enable) {
-               RTE_LOG(ERR, VHOST_CONFIG,
-                       "guest notification isn't supported.\n");
+       if (!dev)
                return -1;
-       }
 
-       dev->virtqueue[queue_id]->used->flags = VRING_USED_F_NO_NOTIFY;
+       if (enable)
+               dev->virtqueue[queue_id]->used->flags &=
+                       ~VRING_USED_F_NO_NOTIFY;
+       else
+               dev->virtqueue[queue_id]->used->flags |= VRING_USED_F_NO_NOTIFY;
        return 0;
 }