net/vhost: fix interrupt mode
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Fri, 24 Jul 2020 14:45:30 +0000 (16:45 +0200)
committerFerruh Yigit <ferruh.yigit@intel.com>
Wed, 29 Jul 2020 22:41:24 +0000 (00:41 +0200)
At .new_device() time, only the first vring pair is
now ready, other vrings are configured later.

Problem is that when application will setup and enable
interrupts, only the first queue pair Rx interrupt will
be enabled.

This patches fixes the issue by setting the number of
max interrupts to the number of Rx queues that will be
later initialized. Then, as soon as a Rx vring is ready
and interrupt enabled by the application, it removes the
corresponding uninitialized epoll event, and installs a
new one with the valid FD.

Fixes: 604052ae5395 ("net/vhost: support queue update")

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
drivers/net/vhost/rte_eth_vhost.c

index 951929c..e55278a 100644 (file)
@@ -5,6 +5,7 @@
 #include <unistd.h>
 #include <pthread.h>
 #include <stdbool.h>
+#include <sys/epoll.h>
 
 #include <rte_mbuf.h>
 #include <rte_ethdev_driver.h>
@@ -95,6 +96,8 @@ struct vhost_queue {
        uint16_t port;
        uint16_t virtqueue_id;
        struct vhost_stats stats;
+       int intr_enable;
+       rte_spinlock_t intr_lock;
 };
 
 struct pmd_internal {
@@ -524,12 +527,58 @@ find_internal_resource(char *ifname)
        return list;
 }
 
+static int
+eth_vhost_update_intr(struct rte_eth_dev *eth_dev, uint16_t rxq_idx)
+{
+       struct rte_intr_handle *handle = eth_dev->intr_handle;
+       struct rte_epoll_event rev;
+       int epfd, ret;
+
+       if (!handle)
+               return 0;
+
+       if (handle->efds[rxq_idx] == handle->elist[rxq_idx].fd)
+               return 0;
+
+       VHOST_LOG(INFO, "kickfd for rxq-%d was changed, updating handler.\n",
+                       rxq_idx);
+
+       if (handle->elist[rxq_idx].fd != -1)
+               VHOST_LOG(ERR, "Unexpected previous kickfd value (Got %d, expected -1).\n",
+                               handle->elist[rxq_idx].fd);
+
+       /*
+        * First remove invalid epoll event, and then install
+        * the new one. May be solved with a proper API in the
+        * future.
+        */
+       epfd = handle->elist[rxq_idx].epfd;
+       rev = handle->elist[rxq_idx];
+       ret = rte_epoll_ctl(epfd, EPOLL_CTL_DEL, rev.fd,
+                       &handle->elist[rxq_idx]);
+       if (ret) {
+               VHOST_LOG(ERR, "Delete epoll event failed.\n");
+               return ret;
+       }
+
+       rev.fd = handle->efds[rxq_idx];
+       handle->elist[rxq_idx] = rev;
+       ret = rte_epoll_ctl(epfd, EPOLL_CTL_ADD, rev.fd,
+                       &handle->elist[rxq_idx]);
+       if (ret) {
+               VHOST_LOG(ERR, "Add epoll event failed.\n");
+               return ret;
+       }
+
+       return 0;
+}
+
 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;
+       int old_intr_enable, ret = 0;
 
        vq = dev->data->rx_queues[qid];
        if (!vq) {
@@ -537,6 +586,18 @@ eth_rxq_intr_enable(struct rte_eth_dev *dev, uint16_t qid)
                return -1;
        }
 
+       rte_spinlock_lock(&vq->intr_lock);
+       old_intr_enable = vq->intr_enable;
+       vq->intr_enable = 1;
+       ret = eth_vhost_update_intr(dev, qid);
+       rte_spinlock_unlock(&vq->intr_lock);
+
+       if (ret < 0) {
+               VHOST_LOG(ERR, "Failed to update rxq%d's intr\n", qid);
+               vq->intr_enable = old_intr_enable;
+               return ret;
+       }
+
        ret = rte_vhost_get_vhost_vring(vq->vid, (qid << 1) + 1, &vring);
        if (ret < 0) {
                VHOST_LOG(ERR, "Failed to get rxq%d's vring\n", qid);
@@ -571,6 +632,8 @@ eth_rxq_intr_disable(struct rte_eth_dev *dev, uint16_t qid)
        rte_vhost_enable_guest_notification(vq->vid, (qid << 1) + 1, 0);
        rte_wmb();
 
+       vq->intr_enable = 0;
+
        return 0;
 }
 
@@ -593,7 +656,6 @@ 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;
@@ -623,6 +685,8 @@ eth_vhost_install_intr(struct rte_eth_dev *dev)
 
        VHOST_LOG(INFO, "Prepare intr vec\n");
        for (i = 0; i < nb_rxq; i++) {
+               dev->intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + i;
+               dev->intr_handle->efds[i] = -1;
                vq = dev->data->rx_queues[i];
                if (!vq) {
                        VHOST_LOG(INFO, "rxq-%d not setup yet, skip!\n", i);
@@ -641,14 +705,12 @@ eth_vhost_install_intr(struct rte_eth_dev *dev)
                                "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++;
                VHOST_LOG(INFO, "Installed intr vec for rxq-%d\n", i);
        }
 
-       dev->intr_handle->nb_efd = count;
-       dev->intr_handle->max_intr = count + 1;
+       dev->intr_handle->nb_efd = nb_rxq;
+       dev->intr_handle->max_intr = nb_rxq + 1;
        dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
 
        return 0;
@@ -835,6 +897,7 @@ vring_conf_update(int vid, struct rte_eth_dev *eth_dev, uint16_t vring_id)
 {
        struct rte_eth_conf *dev_conf = &eth_dev->data->dev_conf;
        struct pmd_internal *internal = eth_dev->data->dev_private;
+       struct vhost_queue *vq;
        struct rte_vhost_vring vring;
        int rx_idx = vring_id % 2 ? (vring_id - 1) >> 1 : -1;
        int ret = 0;
@@ -853,12 +916,18 @@ vring_conf_update(int vid, struct rte_eth_dev *eth_dev, uint16_t vring_id)
                                        vring_id);
                        return ret;
                }
+               eth_dev->intr_handle->efds[rx_idx] = vring.kickfd;
 
-               if (vring.kickfd != eth_dev->intr_handle->efds[rx_idx]) {
-                       VHOST_LOG(INFO, "kickfd for rxq-%d was changed.\n",
-                                         rx_idx);
-                       eth_dev->intr_handle->efds[rx_idx] = vring.kickfd;
+               vq = eth_dev->data->rx_queues[rx_idx];
+               if (!vq) {
+                       VHOST_LOG(ERR, "rxq%d is not setup yet\n", rx_idx);
+                       return -1;
                }
+
+               rte_spinlock_lock(&vq->intr_lock);
+               if (vq->intr_enable)
+                       ret = eth_vhost_update_intr(eth_dev, rx_idx);
+               rte_spinlock_unlock(&vq->intr_lock);
        }
 
        return ret;
@@ -1152,6 +1221,7 @@ eth_rx_queue_setup(struct rte_eth_dev *dev, uint16_t rx_queue_id,
 
        vq->mb_pool = mb_pool;
        vq->virtqueue_id = rx_queue_id * VIRTIO_QNUM + VIRTIO_TXQ;
+       rte_spinlock_init(&vq->intr_lock);
        dev->data->rx_queues[rx_queue_id] = vq;
 
        return 0;
@@ -1173,6 +1243,7 @@ eth_tx_queue_setup(struct rte_eth_dev *dev, uint16_t tx_queue_id,
        }
 
        vq->virtqueue_id = tx_queue_id * VIRTIO_QNUM + VIRTIO_RXQ;
+       rte_spinlock_init(&vq->intr_lock);
        dev->data->tx_queues[tx_queue_id] = vq;
 
        return 0;