Re: [Xen-devel] [PATCH for-4.12] arm/p2m: do not invalidate p2m root if it is shared with IOMMU

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Tue, 29 Jan 2019, Andrii Anisov wrote:
> On 29.01.19 17:17, Julien Grall wrote:
> > No we don't have to.
> I mean, EPAM systems as a XEN based virtualization solution provider have to
> handle that issue.
> 
> > They have been lucky to see this working even on baremetal.
> > Set/Way operations have been dropped from Linux for a long time, so I really
> > can't see why a proprietary driver is still using them.
> I guess there were some performance concerns. And they are really lucky, so
> they did not face related issues on baremetal.
> 
> > I think the policy for Set/Way operations is correct in Xen. This works to
> > avoid breaking basic case but the most complex one are going to break.
> Maybe it worth to get a notification from XEN, i.e. WARN_ONCE, saying that
> Set/Way cache operations are undesirable in VM and will lead to the system
> performance drop, or even instability in case of IOMMU sharing p2m with CPU.
> This might save days of debugging to XEN on ARM users.

A WARN_ONCE would be nice. If you send a patch I would be happy to take
in 4.12 (if the release manager agrees).


> > You solution is only delaying the real fix
> I would not say it is delaying the fix, but allows us to have a functional
> system until we get that fix.
> 
> > (i.e removing Set/Way operation from the software). So here the best
> > solution is to go to the vendor and ask them to fix their software.
> Actually escalation the issue to the vendor is one of our next steps here.

You really want to fix this in the Linux driver if you can, because
you'll get better performance and a more stable system. In the short
term, I would suggest you keep the work-around in your private tree.
However, do let us know how the escalation with the vendor proceeds and
if you have any troubles with it. Julien and I would be happy to provide
help in terms of information and docs on the reasons why this should be
fixed in the driver. Julien might even be able to help further via his
employer.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel