Re: [PATCH v4 00/41] vfio: Adopt iommufd

["Cédric Le Goater" <clg@redhat.com>]

On 11/8/23 04:26, Matthew Rosato wrote:
> On 11/7/23 1:28 PM, Cédric Le Goater wrote:
>> On 11/2/23 08:12, Zhenzhong Duan wrote:
>>> Hi,
>>>
>>> Thanks all for giving guides and comments on previous series, here is
>>> the v4 of pure iommufd support part.
>>>
>>> Based on Cédric's suggestion, this series includes an effort to remove
>>> spapr code from container.c, now all spapr functions are moved to spapr.c
>>> or spapr_pci_vfio.c, but there are still a few trival check on
>>> VFIO_SPAPR_TCE_*_IOMMU which I am not sure if deserved to introduce many
>>> callbacks and duplicate code just to remove them. Some functions are moved
>>> to spapr.c instead of spapr_pci_vfio.c to avoid compile issue because
>>> spapr_pci_vfio.c is arch specific, or else we need to introduce stub
>>> functions to those spapr functions moved.
>>>
>>>
>>> PATCH 1-5: Move spapr functions to spapr*.c
>>> PATCH 6-20: Abstract out base container
>>> PATCH 21-24: Introduce sparpr container and its specific interface
>>
>> PATCH 6-24 applied to vfio-next :
>>
>>    https://github.com/legoater/qemu/commits/vfio-next
>>
>> (with a global s/fucntional/functional/)
>>
>>
>> I also pushed the remaining patches on :
>>
>>    https://github.com/legoater/qemu/commits/vfio-8.2
>>
>> with a slight rework of the IOMMUFD configuration, now done per platform.
>> The VFIO frontend and the 'iommufd' object are only available on x86_64,
>> arm, s390x.
> 
> FYI, I first tried this vfio-8.2 branch on s390x but wasn't actually able to use the iommufd backend (was getting errors like Property 'vfio-pci.iommufd' not found) so I think something isn't actually enabling IOMMUFD as expected with your change...

yes. The previous method used to enable the IOMMUFD device with
a ./configure script option was exposing the CONFIG_IOMMUFD define
globally.

The current method using the Kconfig files requires an extra :

#include CONFIG_DEVICES

in each file using CONFIG_IOMMUFD.

I didn't see it because when compiled natively on x86_64 the
CONFIG_IOMMUFD define is included for some (magic) reason.
It is not the case on other arches, ppc64, aarch64, s390x.

I did the update and repushed vfio-8.2. Should work now.

> 
> Instead I tested on s390x using vfio-next + patches 25-41 of this series on top.
> 
> Legacy backend regression testing worked fine for vfio-pci, vfio-ap and vfio-ccw.

ok. Good. This means that vfio-next is in good shape.

> Using iommufd backend for vfio-pci on s390 exposes an s390-only issue related to accounting of vfio DMA limit (code in hw/s390x/s390-pci-vfio.c assumes VFIODevice.group is never null, but that's no longer true when we use the iommufd backend with cdev).  We don't even need to track this when using the iommufd backend -- With that issue bypassed, vfio-pci testing on s390x looks good so far.  I'll send a separate fix for that.

Thanks,

C.

  
> Using the iommufd backend for vfio-ccw and vfio-ap did not work, see response to patch 28.
> 
> Thanks,
> Matt
>