Re: [PATCH v5 4/8] vfio: Wait for dma-buf invalidation to complete

[Jason Gunthorpe <jgg@ziepe.ca>]

On Mon, Feb 02, 2026 at 04:21:50PM +0100, Christian König wrote:
> > I admit I don't know a lot about VFIO PM support.. Though I thought in
> > the VFIO case PM was actually under userspace control as generally the
> > PM control is delegated to the VM.
> > 
> > Through that lens, what is happening here is correct. If the VM
> > requests to shut down VFIO PM (through a hypervisor vfio ioctl) then
> > we do want to revoke the DMABUF so that the VM can't trigger a AER/etc
> > by trying to access the sleeping PCI device.
> > 
> > I don't think VFIO uses automatic PM on a timer, that doesn't make
> > sense for it's programming model.
> 
> From your description I agree that this doesn't make sense, but from
> the code it looks like exactly that is done.
> 
> Grep for pm_runtime_* on drivers/vfio/pci, but could be that I
> misunderstood the functionality, e.g. didn't spend to much time on
> it.
> 
> Just keep it in the back of your mind and maybe double check if that
> is actually the desired behavior.

I had a small conversation with AlexW and we think VFIO is OK (bugs
excluded).

The use of the PM timer is still under userspace control, even though
a timer is still involved.

Basically there are a series of IOCTL defined in VFIO, like
LOW_POWER_ENTRY that all isolate the PCI device from userspace. The
mmap is blocked with SIBGUS and the DMABUFs are revoked.

The VFIO uAPI contract requries userspace to stop touching the device
immediately when using these IOCTLs. The PM timer may still be
involved, but is an implementation detail.

Effectively VFIO has a device state "isolated" meaning that userspace
cannot access the MMIO, and it enters this state based on various
IOCTLs from userspace. It ties mmap and DMABUF together so that if
mmap SIGBUS's the DMABUF is unmapped.

I understand your remarks, and this use of PM is certainly nothing
that any other driver should copy, but it does make sense for VFIO. If
there are bugs/issues we would continue to keep the overall property
that SGIBUS==DMABUF unmapped and only adjust when that happens.

TBH, I don't think people use the VFIO PM feature very much.

Jason