Re: [dpdk-dev] [PATCH] net/vhost: fix vhost invalid state

["Tan, Jianfeng" <jianfeng.tan@intel.com>]

Hi Junjie,

I think the code is still buggy. As vhost thread and master thread are 
separately invoking new_device() and dev_start().

On 4/10/2018 10:18 PM, Junjie Chen wrote:
> dev_start sets *dev_attached* after setup queues, this sets device to
> invalid state since no frontend is attached. Also destroy_device set
> *started* to zero which makes *allow_queuing* always zero until dev_start
> get called again. Actually, we should not determine queues existence by
> *dev_attached* but by queues pointers or other separated variable(s).
>
> Fixes: 30a701a53737 ("net/vhost: fix crash when creating vdev
> dynamically")
>
> Signed-off-by: Junjie Chen <junjie.j.chen@intel.com>
> ---
>   drivers/net/vhost/rte_eth_vhost.c | 64 +++++++++++++++++++++++----------------
>   1 file changed, 38 insertions(+), 26 deletions(-)
>
> diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
> index 11b6076..6a2ff76 100644
> --- a/drivers/net/vhost/rte_eth_vhost.c
> +++ b/drivers/net/vhost/rte_eth_vhost.c
> @@ -118,6 +118,7 @@ struct pmd_internal {
>   	char *iface_name;
>   	uint16_t max_queues;
>   	uint16_t vid;
> +	uint16_t queue_ready;

If we can reuse below *started* field, instead of introducing a new 
variable?

>   	rte_atomic32_t started;
>   };
>   
> @@ -528,10 +529,13 @@ update_queuing_status(struct rte_eth_dev *dev)
>   	unsigned int i;
>   	int allow_queuing = 1;
>   
> -	if (rte_atomic32_read(&internal->dev_attached) == 0)
> +	if (!dev->data->rx_queues || !dev->data->tx_queues) {
> +		RTE_LOG(ERR, PMD, "RX/TX queues not setup yet\n");
>   		return;
> +	}
>   
> -	if (rte_atomic32_read(&internal->started) == 0)
> +	if (rte_atomic32_read(&internal->started) == 0 ||
> +	    rte_atomic32_read(&internal->dev_attached) == 0)
>   		allow_queuing = 0;
>   
>   	/* Wait until rx/tx_pkt_burst stops accessing vhost device */
> @@ -576,6 +580,8 @@ queue_setup(struct rte_eth_dev *eth_dev, struct pmd_internal *internal)
>   		vq->internal = internal;
>   		vq->port = eth_dev->data->port_id;
>   	}
> +
> +	internal->queue_ready = 1;
>   }
>   
>   static int
> @@ -607,13 +613,10 @@ new_device(int vid)
>   #endif
>   
>   	internal->vid = vid;
> -	if (eth_dev->data->rx_queues && eth_dev->data->tx_queues) {
> +	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);
> -	}
> +	else

vhost thread (t1): goes here and before setting dev_attached ...

> +		RTE_LOG(INFO, PMD, "RX/TX queues not setup yet\n");
>   
>   	for (i = 0; i < rte_vhost_get_vring_num(vid); i++)
>   		rte_vhost_enable_guest_notification(vid, i, 0);
> @@ -622,6 +625,7 @@ 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);
> @@ -657,17 +661,19 @@ destroy_device(int vid)
>   
>   	eth_dev->data->dev_link.link_status = ETH_LINK_DOWN;
>   
> -	for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
> -		vq = eth_dev->data->rx_queues[i];
> -		if (vq == NULL)
> -			continue;
> -		vq->vid = -1;
> -	}
> -	for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
> -		vq = eth_dev->data->tx_queues[i];
> -		if (vq == NULL)
> -			continue;
> -		vq->vid = -1;
> +	if (eth_dev->data->rx_queues && eth_dev->data->tx_queues) {
> +		for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
> +			vq = eth_dev->data->rx_queues[i];
> +			if (!vq)
> +				continue;
> +			vq->vid = -1;
> +		}
> +		for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
> +			vq = eth_dev->data->tx_queues[i];
> +			if (!vq)
> +				continue;
> +			vq->vid = -1;
> +		}
>   	}
>   
>   	state = vring_states[eth_dev->data->port_id];
> @@ -792,11 +798,14 @@ eth_dev_start(struct rte_eth_dev *eth_dev)
>   {
>   	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);
> +	if (!eth_dev->data->rx_queues || !eth_dev->data->tx_queues) {
> +		RTE_LOG(ERR, PMD, "RX/TX queues not exist yet\n");
> +		return -1;
>   	}

I think in DPDK, we can make the assumption that dev_start() comes later 
than queue setup. So we don't need this check.

>   
> +	if (!internal->queue_ready)
> +		queue_setup(eth_dev, internal);

Master thread (t2): setting invalid values.

> +
>   	rte_atomic32_set(&internal->started, 1);
>   	update_queuing_status(eth_dev);
>   
> @@ -836,10 +845,13 @@ eth_dev_close(struct rte_eth_dev *dev)
>   	pthread_mutex_unlock(&internal_list_lock);
>   	rte_free(list);
>   
> -	for (i = 0; i < dev->data->nb_rx_queues; i++)
> -		rte_free(dev->data->rx_queues[i]);
> -	for (i = 0; i < dev->data->nb_tx_queues; i++)
> -		rte_free(dev->data->tx_queues[i]);
> +	if (dev->data->rx_queues)
> +		for (i = 0; i < dev->data->nb_rx_queues; i++)
> +			rte_free(dev->data->rx_queues[i]);
> +
> +	if (dev->data->tx_queues)
> +		for (i = 0; i < dev->data->nb_tx_queues; i++)
> +			rte_free(dev->data->tx_queues[i]);
>   
>   	rte_free(dev->data->mac_addrs);
>   	free(internal->dev_name);