Re: [PATCH v4 1/1] vhost: handle virtqueue locking for memory hotplug

[Maxime Coquelin <maxime.coquelin@redhat.com>]

Hello Danylo,

On 6/2/25 10:50 AM, Danylo Vodopianov wrote:
> For vDPA devices, virtqueues are not locked once the device has been
> configured. In the
> commit 5e8fcc60b59d ("vhost: enhance virtqueue access lock asserts"),
> the asserts were enhanced to trigger rte_panic functionality, preventing
> access to virtqueues without locking. However, this change introduced
> an issue where the memory hotplug mechanism, added in the
> commit 127f9c6f7b78 ("vhost: handle memory hotplug with vDPA devices"),
> no longer works. Since it expects for all queues are locked.
> 
> During the initialization of a vDPA device, the driver sets the
> VIRTIO_DEV_VDPA_CONFIGURED flag, which prevents the
> vhost_user_lock_all_queue_pairs function from locking the
> virtqueues. This leads to the error: the VIRTIO_DEV_VDPA_CONFIGURED
> flag allows the use of the hotplug mechanism, but it fails
> because the virtqueues are not locked, while it expects to be locked
> for VHOST_USER_SET_MEM_TABLE in the table VHOST_MESSAGE_HANDLERS.
> 
> This patch addresses the issue by enhancing the conditional statement
> to include a new condition. Specifically, when the device receives the
> VHOST_USER_SET_MEM_TABLE request, the virtqueues are locked to update
> the basic configurations and hotplug the guest memory.
> 
> This fix does not impact access lock when vDPA driver is configured
> for other unnecessary message handlers.
> 
> Manual memory configuring with "--socket-mem" option allows to avoid
> hotplug mechanism using.

s/using/use/

It needs a fixes tag, and stable@dpdk.org should be cc'ed, so that it
gets backported to LTS branches.

> 
> Signed-off-by: Danylo Vodopianov <dvo-plv@napatech.com>
> ---
>   lib/vhost/vhost_user.c | 8 +++++++-
>   1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
> index ec950acf97..16d03e1033 100644
> --- a/lib/vhost/vhost_user.c
> +++ b/lib/vhost/vhost_user.c
> @@ -3178,7 +3178,13 @@ vhost_user_msg_handler(int vid, int fd)
>   	 * would cause a dead lock.
>   	 */
>   	if (msg_handler != NULL && msg_handler->lock_all_qps) {
> -		if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED)) {
> +		/* Lock all queue pairs if the device is not configured for vDPA,
> +		 * or if it is configured for vDPA but the request is VHOST_USER_SET_MEM_TABLE.
> +		 * This ensures proper queue locking for memory table updates and guest
> +		 * memory hotplug.
> +		 */
> +		if (!(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) ||
> +			request == VHOST_USER_SET_MEM_TABLE) {

It looks like a workaround, and I'm afraid it could cause regression
with some vDPA devices, or that it would not be enough and we would have 
to add other requests as exception.


Wouldn't it better to modify VHOST_USER_ASSERT_LOCK() so that it takes
into account the VIRTIO_DEV_VDPA_CONFIGURED flag?

Thanks,
Maxime

>   			vhost_user_lock_all_queue_pairs(dev);
>   			unlock_required = 1;
>   		}