Linux Virtualization - May 2020 - [PATCH] vdpa_sim: do not reset IOTLB during device reset

We reset IOTLB during device reset this breaks the assumption that the
mapping needs to be controlled via vDPA DMA ops explicitly in a
incremental way. So the networking will be broken after e.g a guest
reset.

Fix this by not resetting the IOTLB during device reset.

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/vdpa/vdpa_sim/vdpa_sim.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.c b/drivers/vdpa/vdpa_sim/vdpa_sim.c
index 7957d2d41fc4..cc5525743a25 100644
--- a/drivers/vdpa/vdpa_sim/vdpa_sim.c
+++ b/drivers/vdpa/vdpa_sim/vdpa_sim.c
@@ -119,8 +119,6 @@ static void vdpasim_reset(struct vdpasim *vdpasim)
 	for (i = 0; i < VDPASIM_VQ_NUM; i++)
 		vdpasim_vq_reset(&vdpasim->vqs[i]);
 
-	vhost_iotlb_reset(vdpasim->iommu);
-
 	vdpasim->features = 0;
 	vdpasim->status = 0;
 	++vdpasim->generation;
-- 
2.20.1

On Thu, May 14, 2020 at 03:25:49PM +0800, Jason Wang
wrote:
> We reset IOTLB during device reset this breaks the assumption that the
> mapping needs to be controlled via vDPA DMA ops explicitly in a
> incremental way. So the networking will be broken after e.g a guest
> reset.
> 
> Fix this by not resetting the IOTLB during device reset.
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>

That's a bit weird, and can be a security risk if state
leaks between security domains through this.
And there's 0 chance any hardware implementation can
keep the translations around across resets - there
is simply nowhere to keep them.

IMHO we need a different way to make this work, simulator
needs to look like a hardware device as much as possible.

> ---
>  drivers/vdpa/vdpa_sim/vdpa_sim.c | 2 --
>  1 file changed, 2 deletions(-)
> 
> diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.c
b/drivers/vdpa/vdpa_sim/vdpa_sim.c
> index 7957d2d41fc4..cc5525743a25 100644
> --- a/drivers/vdpa/vdpa_sim/vdpa_sim.c
> +++ b/drivers/vdpa/vdpa_sim/vdpa_sim.c
> @@ -119,8 +119,6 @@ static void vdpasim_reset(struct vdpasim *vdpasim)
>  	for (i = 0; i < VDPASIM_VQ_NUM; i++)
>  		vdpasim_vq_reset(&vdpasim->vqs[i]);
>  
> -	vhost_iotlb_reset(vdpasim->iommu);
> -
>  	vdpasim->features = 0;
>  	vdpasim->status = 0;
>  	++vdpasim->generation;
> -- 
> 2.20.1

On 2020/5/14 ??5:35, Michael S. Tsirkin wrote:
> On Thu, May 14, 2020 at 03:25:49PM +0800, Jason Wang wrote:
>> We reset IOTLB during device reset this breaks the assumption that the
>> mapping needs to be controlled via vDPA DMA ops explicitly in a
>> incremental way. So the networking will be broken after e.g a guest
>> reset.
>>
>> Fix this by not resetting the IOTLB during device reset.
>>
>> Signed-off-by: Jason Wang <jasowang at redhat.com>
>
> That's a bit weird, and can be a security risk if state
> leaks between security domains through this.

I'm not sure I get this. Note that:

1) For devices that depend on platform IOMMU, the mappings are valid 
across device reset
2) vhost_vdpa will reset IOTLB during release, so I think there's no 
security leak in this case

If we reset IOTLB during device reset, there will be an inconsistency 
between on-chip IOMMU devices and platform IOMMU devices. We can fix 
this inconsistency in another way, e.g unmap during vhost_vdpa_reset. 
This means userspace need to replay the mapping before DRIVER_OK, which 
seems a burden to userspace.

> And there's 0 chance any hardware implementation can
> keep the translations around across resets - there
> is simply nowhere to keep them.

It depends on the hardware implementation, e.g the IOMMU does not belong 
to VF but PF.

Thanks

>
> IMHO we need a different way to make this work, simulator
> needs to look like a hardware device as much as possible.
>
>
>> ---
>>   drivers/vdpa/vdpa_sim/vdpa_sim.c | 2 --
>>   1 file changed, 2 deletions(-)
>>
>> diff --git a/drivers/vdpa/vdpa_sim/vdpa_sim.c
b/drivers/vdpa/vdpa_sim/vdpa_sim.c
>> index 7957d2d41fc4..cc5525743a25 100644
>> --- a/drivers/vdpa/vdpa_sim/vdpa_sim.c
>> +++ b/drivers/vdpa/vdpa_sim/vdpa_sim.c
>> @@ -119,8 +119,6 @@ static void vdpasim_reset(struct vdpasim *vdpasim)
>>   	for (i = 0; i < VDPASIM_VQ_NUM; i++)
>>   		vdpasim_vq_reset(&vdpasim->vqs[i]);
>>   
>> -	vhost_iotlb_reset(vdpasim->iommu);
>> -
>>   	vdpasim->features = 0;
>>   	vdpasim->status = 0;
>>   	++vdpasim->generation;
>> -- 
>> 2.20.1