Re: [PATCH] PCI/ASPM: Enable ASPM for links under VMD domain
From: Dan Williams
Date: Thu Aug 27 2020 - 17:34:06 EST
On Thu, Aug 27, 2020 at 9:46 AM Derrick, Jonathan
<jonathan.derrick@xxxxxxxxx> wrote:
>
> On Thu, 2020-08-27 at 17:23 +0100, hch@xxxxxxxxxxxxx wrote:
> > On Thu, Aug 27, 2020 at 04:13:44PM +0000, Derrick, Jonathan wrote:
> > > On Thu, 2020-08-27 at 06:34 +0000, hch@xxxxxxxxxxxxx wrote:
> > > > On Wed, Aug 26, 2020 at 09:43:27PM +0000, Derrick, Jonathan wrote:
> > > > > Feel free to review my set to disable the MSI remapping which will
> > > > > make
> > > > > it perform as well as direct-attached:
> > > > >
> > > > > https://patchwork.kernel.org/project/linux-pci/list/?series=325681
> > > >
> > > > So that then we have to deal with your schemes to make individual
> > > > device direct assignment work in a convoluted way?
> > >
> > > That's not the intent of that patchset -at all-. It was to address the
> > > performance bottlenecks with VMD that you constantly complain about.
> >
> > I know. But once we fix that bottleneck we fix the next issue,
> > then to tackle the next. While at the same time VMD brings zero
> > actual benefits.
> >
>
> Just a few benefits and there are other users with unique use cases:
> 1. Passthrough of the endpoint to OSes which don't natively support
> hotplug can enable hotplug for that OS using the guest VMD driver
> 2. Some hypervisors have a limit on the number of devices that can be
> passed through. VMD endpoint is a single device that expands to many.
> 3. Expansion of possible bus numbers beyond 256 by using other
> segments.
> 4. Custom RAID LED patterns driven by ledctl
>
> I'm not trying to market this. Just pointing out that this isn't
> "bringing zero actual benefits" to many users.
>
The initial intent of the VMD driver was to allow Linux to find and
initialize devices behind a VMD configuration where VMD was required
for a non-Linux OS. For Linux, if full native PCI-E is an available
configuration option I think it makes sense to recommend Linux users
to flip that knob rather than continue to wrestle with the caveats of
the VMD driver. Where that knob isn't possible / available VMD can be
a fallback, but full native PCI-E is what Linux wants in the end.