Re: [PATCH v20 00/16] vfio/pci: Add PCIe TPH support

[Alex Williamson <alex@shazbot.org>]

On Wed, 21 Oct 2026 09:43:05 +0800
fengchengwen <fengchengwen@huawei.com> wrote:

> Hi Alex,
> 
> Quick follow-up from our LPC discussion on the VFIO TPH patchset.
> 
> The P2P TPH path carries significant verification complexity, as it builds
> on the PCI P2PDMA stack which is still under active upstream revision.
> 
> Meanwhile there is clear real-world demand for the device-to-host scenario.
> AMD recently posted their debugfs series exposing CPU-to-ST mappings [1],
> and we also have strong use cases for D2H TPH enablement.

You might be working ahead of the industry.  The debugfs interface
provides evidence that the feature is coming, but is perhaps still
working through implementation pains.

The lack of community participation in the vfio interface actually
suggests to me that there isn't a great real-world demand for this
feature, yet. 

> I'd like to propose a phased upstream approach:
> land device-to-host support first, and defer P2P TPH to a follow-up series.
> 
> We can add P2P support cleanly on top once the interface and verification
> are solid. This way we can deliver usable TPH functionality sooner, while
> keeping the path open for full P2P support later.
> 
> Let me know your thoughts on going ahead with device-to-host support first.
> If you're okay with this, I'll post v21 accordingly.

As above, the demand is not apparent based on the lack of community
participation in these threads.  I'm reluctant to commit to a uAPI
that's largely been proposed by me, as the maintainer, trying to fill
gaps in the narrow scope of the original proposal.

This needs support from the community and validation across broader use
cases.  I think that's the actual barrier, rather than ordering versus
Zhiping's series.

Despite the LPC talk, there still hasn't been much engagement on the
question of whether device ownership grants the user the ability to
target specific cache levels and how we handle cases where the
driver:cpu association isn't static.  Thanks,

Alex