Re: virtio-iommu hotplug issue

[Akihiko Odaki]

On 2023/04/13 19:40, Jean-Philippe Brucker wrote:
> Hello,
>
> On Thu, Apr 13, 2023 at 01:49:43PM +0900, Akihiko Odaki wrote:
> > Hi,
> >
> > Recently I encountered a problem with the combination of Linux's
> > virtio-iommu driver and QEMU when a SR-IOV virtual function gets disabled.
> > I'd like to ask you what kind of solution is appropriate here and implement
> > the solution if possible.
> >
> > A PCIe device implementing the SR-IOV specification exports a virtual
> > function, and the guest can enable or disable it at runtime by writing to a
> > configuration register. This effectively looks like a PCI device is
> > hotplugged for the guest.
>
> Just so I understand this better: the guest gets a whole PCIe device PF
> that implements SR-IOV, and so the guest can dynamically create VFs?  Out
> of curiosity, is that a hardware device assigned to the guest with VFIO,
> or a device emulated by QEMU?

Yes, that's right. The guest can dynamically create and delete VFs. The 
device is emulated by QEMU: igb, an Intel NIC recently added to QEMU and 
projected to be released as part of QEMU 8.0.

> > In such a case, the kernel assumes the endpoint is
> > detached from the virtio-iommu domain, but QEMU actually does not detach it.
> >
> > This inconsistent view of the removed device sometimes prevents the VM from
> > correctly performing the following procedure, for example:
> > 1. Enable a VF.
> > 2. Disable the VF.
> > 3. Open a vfio container.
> > 4. Open the group which the PF belongs to.
> > 5. Add the group to the vfio container.
> > 6. Map some memory region.
> > 7. Close the group.
> > 8. Close the vfio container.
> > 9. Repeat 3-8
> >
> > When the VF gets disabled, the kernel assumes the endpoint is detached from
> > the IOMMU domain, but QEMU actually doesn't detach it. Later, the domain
> > will be reused in step 3-8.
> >
> > In step 7, the PF will be detached, and the kernel thinks there is no
> > endpoint attached and the mapping the domain holds is cleared, but the VF
> > endpoint is still attached and the mapping is kept intact.
> >
> > In step 9, the same domain will be reused again, and the kernel requests to
> > create a new mapping, but it will conflict with the existing mapping and
> > result in -EINVAL.
> >
> > This problem can be fixed by either of:
> > - requesting the detachment of the endpoint from the guest when the PCI
> > device is unplugged (the VF is disabled)
>
> Yes, I think this is an issue in the virtio-iommu driver, which should be
> sending a DETACH request when the VF is disabled, likely from
> viommu_release_device(). I'll work on a fix unless you would like to do it

It will be nice if you prepare a fix. I will test your patch with my 
workload if you share it with me.

Regards,
Akihiko Odaki

> > - detecting that the PCI device is gone and automatically detach it on
> > QEMU-side.
> >
> > It is not completely clear for me which solution is more appropriate as the
> > virtio-iommu specification is written in a way independent of the endpoint
> > mechanism and does not say what should be done when a PCI device is
> > unplugged.
>
> Yes, I'm not sure it's in scope for the specification, it's more about
> software guidance
>
> Thanks,
> Jean