net/vhost: fix crash when creating vdev dynamically
authorJunjie Chen <junjie.j.chen@intel.com>
Fri, 30 Mar 2018 06:58:31 +0000 (14:58 +0800)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 30 Mar 2018 12:08:44 +0000 (14:08 +0200)
When creating vdev dynamically, vhost pmd driver starts directly without
checking TX/RX queues are ready or not, and thus causes segmentation fault
when vhost library accesses queues. This patch adds a flag to check whether
queues are setup or not, and adds queues setup into dev_start function to
allow user to start them after setting up.

Fixes: aed0b12930b3 ("net/vhost: fix socket file deleted on stop")
Cc: stable@dpdk.org
Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
Tested-by: Zhiyong Yang <zhiyong.yang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
drivers/net/vhost/rte_eth_vhost.c

index 3aae01c..11b6076 100644 (file)
@@ -117,6 +117,7 @@ struct pmd_internal {
        char *dev_name;
        char *iface_name;
        uint16_t max_queues;
+       uint16_t vid;
        rte_atomic32_t started;
 };
 
@@ -527,8 +528,10 @@ update_queuing_status(struct rte_eth_dev *dev)
        unsigned int i;
        int allow_queuing = 1;
 
-       if (rte_atomic32_read(&internal->started) == 0 ||
-           rte_atomic32_read(&internal->dev_attached) == 0)
+       if (rte_atomic32_read(&internal->dev_attached) == 0)
+               return;
+
+       if (rte_atomic32_read(&internal->started) == 0)
                allow_queuing = 0;
 
        /* Wait until rx/tx_pkt_burst stops accessing vhost device */
@@ -551,13 +554,36 @@ update_queuing_status(struct rte_eth_dev *dev)
        }
 }
 
+static void
+queue_setup(struct rte_eth_dev *eth_dev, struct pmd_internal *internal)
+{
+       struct vhost_queue *vq;
+       int i;
+
+       for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
+               vq = eth_dev->data->rx_queues[i];
+               if (!vq)
+                       continue;
+               vq->vid = internal->vid;
+               vq->internal = internal;
+               vq->port = eth_dev->data->port_id;
+       }
+       for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
+               vq = eth_dev->data->tx_queues[i];
+               if (!vq)
+                       continue;
+               vq->vid = internal->vid;
+               vq->internal = internal;
+               vq->port = eth_dev->data->port_id;
+       }
+}
+
 static int
 new_device(int vid)
 {
        struct rte_eth_dev *eth_dev;
        struct internal_list *list;
        struct pmd_internal *internal;
-       struct vhost_queue *vq;
        unsigned i;
        char ifname[PATH_MAX];
 #ifdef RTE_LIBRTE_VHOST_NUMA
@@ -580,21 +606,13 @@ new_device(int vid)
                eth_dev->data->numa_node = newnode;
 #endif
 
-       for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
-               vq = eth_dev->data->rx_queues[i];
-               if (vq == NULL)
-                       continue;
-               vq->vid = vid;
-               vq->internal = internal;
-               vq->port = eth_dev->data->port_id;
-       }
-       for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
-               vq = eth_dev->data->tx_queues[i];
-               if (vq == NULL)
-                       continue;
-               vq->vid = vid;
-               vq->internal = internal;
-               vq->port = eth_dev->data->port_id;
+       internal->vid = vid;
+       if (eth_dev->data->rx_queues && eth_dev->data->tx_queues) {
+               queue_setup(eth_dev, internal);
+               rte_atomic32_set(&internal->dev_attached, 1);
+       } else {
+               RTE_LOG(INFO, PMD, "RX/TX queues have not setup yet\n");
+               rte_atomic32_set(&internal->dev_attached, 0);
        }
 
        for (i = 0; i < rte_vhost_get_vring_num(vid); i++)
@@ -604,7 +622,6 @@ new_device(int vid)
 
        eth_dev->data->dev_link.link_status = ETH_LINK_UP;
 
-       rte_atomic32_set(&internal->dev_attached, 1);
        update_queuing_status(eth_dev);
 
        RTE_LOG(INFO, PMD, "Vhost device %d created\n", vid);
@@ -634,8 +651,9 @@ destroy_device(int vid)
        eth_dev = list->eth_dev;
        internal = eth_dev->data->dev_private;
 
-       rte_atomic32_set(&internal->dev_attached, 0);
+       rte_atomic32_set(&internal->started, 0);
        update_queuing_status(eth_dev);
+       rte_atomic32_set(&internal->dev_attached, 0);
 
        eth_dev->data->dev_link.link_status = ETH_LINK_DOWN;
 
@@ -770,12 +788,17 @@ rte_eth_vhost_get_vid_from_port_id(uint16_t port_id)
 }
 
 static int
-eth_dev_start(struct rte_eth_dev *dev)
+eth_dev_start(struct rte_eth_dev *eth_dev)
 {
-       struct pmd_internal *internal = dev->data->dev_private;
+       struct pmd_internal *internal = eth_dev->data->dev_private;
+
+       if (unlikely(rte_atomic32_read(&internal->dev_attached) == 0)) {
+               queue_setup(eth_dev, internal);
+               rte_atomic32_set(&internal->dev_attached, 1);
+       }
 
        rte_atomic32_set(&internal->started, 1);
-       update_queuing_status(dev);
+       update_queuing_status(eth_dev);
 
        return 0;
 }