net/vhost: support queue update
authorMatan Azrad <matan@mellanox.com>
Tue, 21 Jul 2020 16:38:16 +0000 (16:38 +0000)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 21 Jul 2020 17:03:24 +0000 (19:03 +0200)
The commit below changed the readiness condition of vhost device to fix
multi-queues issues showed with QEMU versions.

Now, the vhost device is ready when the first queue-pair is ready.
When more queues are being ready, the queue state callback will be
triggered to notify the vhost manager.

In case of Rx interrupt configuration, the vhost driver set the
kickfd queue file descriptor in order to be notified on Rx traffic.

So, when queue becomes ready, the kickfd may be changed and should be
updated in the Rx interrupt structure.

Update kickfd when the queue state callback is invoked.
Also update event notification when it is enabled by the user.

Fixes: d0fcc38f5fa4 ("vhost: improve device readiness notifications")

Suggested-by: Marvin Liu <yong.liu@intel.com>
Signed-off-by: Matan Azrad <matan@mellanox.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
drivers/net/vhost/rte_eth_vhost.c

index 9fbf39f..bbf79b2 100644 (file)
@@ -94,6 +94,7 @@ struct vhost_queue {
        struct rte_mempool *mb_pool;
        uint16_t port;
        uint16_t virtqueue_id;
+       bool intr_en;
        struct vhost_stats stats;
 };
 
@@ -546,6 +547,8 @@ eth_rxq_intr_enable(struct rte_eth_dev *dev, uint16_t qid)
        rte_vhost_enable_guest_notification(vq->vid, (qid << 1) + 1, 1);
        rte_wmb();
 
+       vq->intr_en = true;
+
        return ret;
 }
 
@@ -571,6 +574,8 @@ 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_en = false;
+
        return 0;
 }
 
@@ -830,6 +835,45 @@ destroy_device(int vid)
        _rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
 }
 
+static int
+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 rte_vhost_vring vring;
+       struct vhost_queue *vq;
+       int rx_idx = vring_id % 2 ? (vring_id - 1) >> 1 : -1;
+       int ret = 0;
+
+       /*
+        * The vring kickfd may be changed after the new device notification.
+        * Update it when the vring state is updated.
+        */
+       if (rx_idx >= 0 && rx_idx < eth_dev->data->nb_rx_queues &&
+           rte_atomic32_read(&internal->dev_attached) &&
+           rte_atomic32_read(&internal->started) &&
+           dev_conf->intr_conf.rxq) {
+               vq = eth_dev->data->rx_queues[rx_idx];
+               ret = rte_vhost_get_vhost_vring(vid, vring_id, &vring);
+               if (!ret) {
+                       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;
+                       }
+
+                       rte_vhost_enable_guest_notification(vid, vring_id,
+                                                           vq->intr_en);
+                       rte_wmb();
+               }
+       }
+
+       return ret;
+}
+
 static int
 vring_state_changed(int vid, uint16_t vring, int enable)
 {
@@ -848,6 +892,11 @@ vring_state_changed(int vid, uint16_t vring, int enable)
        eth_dev = list->eth_dev;
        /* won't be NULL */
        state = vring_states[eth_dev->data->port_id];
+
+       if (enable && vring_conf_update(vid, eth_dev, vring))
+               VHOST_LOG(INFO, "Failed to update vring-%d configuration.\n",
+                         (int)vring);
+
        rte_spinlock_lock(&state->lock);
        if (state->cur[vring] == enable) {
                rte_spinlock_unlock(&state->lock);