mlx5_vdpa_dev_add() doesn’t initialize mvdev->actual_features. It’s initialized later by mlx5_vdpa_set_driver_features(). However, mlx5_vdpa_get_config() depends on the VIRTIO_F_VERSION_1 flag in actual_features, to return data with correct endianness. When it’s called before mlx5_vdpa_set_driver_features(), the data are incorrectly returned as big-endian on big-endian machines, while QEMU then interprets them as little-endian. The fix is to initialize this VIRTIO_F_VERSION_1 as early as possible, especially considering that mlx5_vdpa_dev_add() insists on this flag to always be set anyway. Signed-off-by: Konstantin Shkolnyy <kshk@xxxxxxxxxxxxx> --- Changes in V2: Initialize only VIRTIO_F_VERSION_1 in actual_features, rather than all flags. drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c index 36099047560d..cccc49a08a1a 100644 --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c @@ -3884,6 +3884,9 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name, ndev->mvdev.max_vqs = max_vqs; mvdev = &ndev->mvdev; mvdev->mdev = mdev; + /* cpu_to_mlx5vdpa16() below depends on this flag */ + mvdev->actual_features = + (device_features & BIT_ULL(VIRTIO_F_VERSION_1)); ndev->vqs = kcalloc(max_vqs, sizeof(*ndev->vqs), GFP_KERNEL); ndev->event_cbs = kcalloc(max_vqs + 1, sizeof(*ndev->event_cbs), GFP_KERNEL); -- 2.34.1
