Re: [PATCH v3 1/2] hw/i386/amd_iommu: Fix device setup failure when PT is on.

[Sairaj Kodilkar <sarunkod@amd.com>]

On 5/16/2025 8:13 PM, Philippe Mathieu-Daudé wrote:
> On 16/5/25 12:05, Sairaj Kodilkar wrote:
>> Commit c1f46999ef506 ("amd_iommu: Add support for pass though mode")
>> introduces the support for "pt" flag by enabling nodma memory when
>> "pt=off". This allowed VFIO devices to successfully register notifiers
>> by using nodma region.
>>
>> But, This also broke things when guest is booted with the iommu=nopt
>> because, devices bypass the IOMMU and use untranslated addresses 
>> (IOVA) to
>> perform DMA reads/writes to the nodma memory region, ultimately 
>> resulting in
>> a failure to setup the devices in the guest.
>>
>> Fix the above issue by always enabling the amdvi_dev_as->iommu memory 
>> region.
>> But this will once again cause VFIO devices to fail while registering the
>> notifiers with AMD IOMMU memory region.
>>
>> Fixes: c1f46999ef506 ("amd_iommu: Add support for pass though mode")
>> Signed-off-by: Sairaj Kodilkar <sarunkod@amd.com>
>> Reviewed-by: Vasant Hegde <vasant.hegde@amd.com>
>> ---
>>   hw/i386/amd_iommu.c | 12 ++----------
>>   1 file changed, 2 insertions(+), 10 deletions(-)
>>
>> diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
>> index 5f9b95279997..df8ba5d39ada 100644
>> --- a/hw/i386/amd_iommu.c
>> +++ b/hw/i386/amd_iommu.c
>> @@ -1426,7 +1426,6 @@ static AddressSpace *amdvi_host_dma_iommu(PCIBus 
>> *bus, void *opaque, int devfn)
>>       AMDVIState *s = opaque;
>>       AMDVIAddressSpace **iommu_as, *amdvi_dev_as;
>>       int bus_num = pci_bus_num(bus);
>> -    X86IOMMUState *x86_iommu = X86_IOMMU_DEVICE(s);
>>       iommu_as = s->address_spaces[bus_num];
>> @@ -1486,15 +1485,8 @@ static AddressSpace 
>> *amdvi_host_dma_iommu(PCIBus *bus, void *opaque, int devfn)
>>                                               AMDVI_INT_ADDR_FIRST,
>>                                               &amdvi_dev_as->iommu_ir, 
>> 1);
>> -        if (!x86_iommu->pt_supported) {
>> -            memory_region_set_enabled(&amdvi_dev_as->iommu_nodma, 
>> false);
>> -            memory_region_set_enabled(MEMORY_REGION(&amdvi_dev_as- 
>> >iommu),
>> -                                      true);
>> -        } else {
>> -            memory_region_set_enabled(MEMORY_REGION(&amdvi_dev_as- 
>> >iommu),
>> -                                      false);
>> -            memory_region_set_enabled(&amdvi_dev_as->iommu_nodma, true);
>> -        }
>> +        memory_region_set_enabled(&amdvi_dev_as->iommu_nodma, false);
> 
> I have no clue about this device but wonder what is the usefulness of
> iommu_nodma now, isn't it dead code?
> 

Hi Philippe,

Indeed the iommu_nodma is dead. The reason I did not remove the
iommu_nodma region completely is that, Alejandro's DMA remapping patches
[1] uses this region to dynamically switch the address space.

[1] 
https://lore.kernel.org/qemu-devel/20250502021605.1795985-1-alejandro.j.jimenez@oracle.com/

Thanks
Sairaj

>> +        memory_region_set_enabled(MEMORY_REGION(&amdvi_dev_as- 
>> >iommu), true);
>>       }
>>       return &iommu_as[devfn]->as;
>>   }
>