Re: [PATCH 4/5] iommu/vt-d: Disallow SVA if devices don't support 64-bit address

[Robin Murphy <robin.murphy@arm.com>]

On 2021-07-21 02:50, Lu Baolu wrote:
> Hi Robin,
> 
> Thanks a lot for reviewing my patch!
> 
> On 7/20/21 5:27 PM, Robin Murphy wrote:
>> On 2021-07-20 02:38, Lu Baolu wrote:
>>> When the device and CPU share an address space (such as SVA), the device
>>> must support the same addressing capability as the CPU. The CPU does not
>>> consider the addressing ability of any device when managing the page 
>>> table
>>> of a process, so the device must have enough addressing ability to bind
>>> the page table of the process.
>>>
>>> Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
>>> ---
>>>   drivers/iommu/intel/iommu.c | 3 +++
>>>   1 file changed, 3 insertions(+)
>>>
>>> diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c
>>> index f45c80ce2381..f3cca1dd384d 100644
>>> --- a/drivers/iommu/intel/iommu.c
>>> +++ b/drivers/iommu/intel/iommu.c
>>> @@ -5372,6 +5372,9 @@ static int intel_iommu_enable_sva(struct device 
>>> *dev)
>>>       if (!(iommu->flags & VTD_FLAG_SVM_CAPABLE))
>>>           return -ENODEV;
>>> +    if (!dev->dma_mask || *dev->dma_mask != DMA_BIT_MASK(64))
>>
>> Careful - VFIO doesn't set DMA masks (since it doesn't use the DMA API),
> 
> SVA doesn't work through the VFIO framework.

Did anyone say it does? My point is that, as far as I understand, the 
SVA UAPI is very much intended to work *with* VFIO, and even if the 
finer details are still mired in the /dev/ioasid discussion today we 
should definitely expect to see VFIO-like use-cases at some point. I 
certainly don't see why any of the guest SVA stuff exists already if not 
for VFIO-assigned devices?

>> so this appears to be relying on another driver having bound previously,
> 
> Yes. You are right.
> 
>> otherwise the mask would still be the default 32-bit one from 
>> pci_setup_device(). I'm not sure that's an entirely robust assumption.
> 
> Currently SVA implementation always requires a native kernel driver. The
> assumption is that the drivers should check and set 64-bit addressing
> capability before calling iommu_sva_xxx() APIs.

...and given that that is not a documented requirement, and certainly 
not a technical one (even a self-contained kernel driver could choose to 
only use SVA contexts and not touch the DMA API), it's an inherently 
fragile assumption which I'm confident *will* be broken eventually :)

Robin.

>>> +        return -ENODEV;
>>> +
>>>       if (intel_iommu_enable_pasid(iommu, dev))
>>>           return -ENODEV;
>>>
> 
> Best regards,
> baolu