Re: [PATCH v9 1/9] hw/vfio/common: Remove error print on mmio region translation by viommu

[Auger Eric]

Hi Alex,

On 3/24/20 12:08 AM, Alex Williamson wrote:
> [Cc +dwg who originated this warning]
> 
> On Mon, 23 Mar 2020 14:16:09 +0530
> Bharat Bhushan <address@hidden> wrote:
> 
>> On ARM, the MSI doorbell is translated by the virtual IOMMU.
>> As such address_space_translate() returns the MSI controller
>> MMIO region and we get an "iommu map to non memory area"
>> message. Let's remove this latter.
>>
>> Signed-off-by: Eric Auger <address@hidden>
>> Signed-off-by: Bharat Bhushan <address@hidden>
>> ---
>>  hw/vfio/common.c | 2 --
>>  1 file changed, 2 deletions(-)
>>
>> diff --git a/hw/vfio/common.c b/hw/vfio/common.c
>> index 5ca11488d6..c586edf47a 100644
>> --- a/hw/vfio/common.c
>> +++ b/hw/vfio/common.c
>> @@ -426,8 +426,6 @@ static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void 
>> **vaddr,
>>                                   &xlat, &len, writable,
>>                                   MEMTXATTRS_UNSPECIFIED);
>>      if (!memory_region_is_ram(mr)) {
>> -        error_report("iommu map to non memory area %"HWADDR_PRIx"",
>> -                     xlat);
>>          return false;
>>      }
>>  
> 
> I'm a bit confused here, I think we need more justification beyond "we
> hit this warning and we don't want to because it's ok in this one
> special case, therefore remove it".  I assume the special case is that
> the device MSI address is managed via the SET_IRQS ioctl and therefore
> we won't actually get DMAs to this range.
Yes exactly. The guest creates a mapping between one giova and this gpa
(corresponding to the MSI controller doorbell) because MSIs are mapped
on ARM. But practically the physical device is programmed with an host
chosen iova that maps onto the physical MSI controller's doorbell. so
the device never performs DMA accesses to this range.

  But I imagine the case that
> was in mind when adding this warning was general peer-to-peer between
> and assigned and emulated device.
yes makes sense.

  Maybe there's an argument to be made
> that such a p2p mapping might also be used in a non-vIOMMU case.  We
> skip creating those mappings and drivers continue to work, maybe
> because nobody attempts to do p2p DMA with the types of devices we
> emulate, maybe because p2p DMA is not absolutely reliable on bare metal
> and drivers test it before using it.
MSI doorbells are mapped using the IOMMU_MMIO flag (dma-iommu.c
iommu_dma_get_msi_page).
One idea could be to pass that flag through the IOMMU Notifier mechanism
into the iotlb->perm. Eventually when we get this in vfio_get_vaddr() we
would not print the warning. Could that make sense?

Thanks

Eric



  Anyway, I need a better argument
> why this is ok.  Thanks,
> 
> Alex
>