Re: [RFC PATCH v2 0/4] vfio/pci: Atomic Ops completer support

[Alex Williamson <alex.williamson@redhat.com>]

On Wed, 31 May 2023 23:55:41 +0200
Robin Voetter <robin@streamhpc.com> wrote:

> Hi Alex,
> 
> Thanks for taking the time to implement support for Atomic Op completer
> support properly :). I have tested out these patches and the kernel
> patch, and apart from a minor issue with patch 2 everything works fine;

Yes, Cedric noted the extra semicolon as well, I'm about to post that
patch as non-RFC since it has no dependencies on anything else.

> ROCm programs that use device->host atomic operations work properly.

Great, thanks for testing!

> Something that I have been thinking about, are there any implications
> involved with enabling this feature automatically with no possibility of
> turning it off? I have no use case for that, though, and I cant really
> think of a reason other than preventing the guest from finding out
> hardware details about the host.

Not sure I follow, are you suggesting that Atomic Ops completion
support is proactively enabled in the VM to match the host, regardless
of attached assigned devices?  An obvious problem there is migration.
If I start a VM on a host with Atomic Ops support and want to migrate
to a host without Atomic Ops support, config space of the root ports is
now different and the migration fails.  QEMU would also require
elevated privileges to read config space to determine the host support,
and then what does it do if only some of the PCIe hierarchy supports
Atomic Ops.

Policy decisions like that are generally left to management tools, so
there would always be a means to enable or disable the feature.  In
fact, that's specifically why I test that the Atomic Op completer bits
are unset in the root port before changing them so that this automatic
enablement could live alongside a command line option to statically
enable some bits.

That does however remind me that it is often good with these sorts of
"clever" automatic features to have an opt-out, so I'll likely add an
x-no-rp-atomics device option in the next version to provide that.
Thanks,

Alex