Re: [PATCH 0/6] RfC: try improve native hotplug for pcie root ports

[Michael S. Tsirkin]

On Thu, Nov 11, 2021 at 06:08:11PM +0000, Daniel P. Berrangé wrote:
> On Thu, Nov 11, 2021 at 12:11:19PM -0500, Michael S. Tsirkin wrote:
> > On Thu, Nov 11, 2021 at 09:35:37AM +0000, Daniel P. Berrangé wrote:
> > > On Thu, Nov 11, 2021 at 03:20:07AM -0500, Michael S. Tsirkin wrote:
> > > > On Thu, Nov 11, 2021 at 08:53:06AM +0100, Gerd Hoffmann wrote:
> > > > >   Hi,
> > > > > 
> > > > > > Given it's a bugfix, and given that I hear through internal channels
> > > > > > that QE results so far have been encouraging, I am inclined to bite 
> > > > > > the
> > > > > > bullet and merge this for -rc1.
> > > > > 
> > > > > Fine with me.
> > > > > 
> > > > > > I don't think this conflicts with Julia's patches as users can still
> > > > > > disable ACPI hotplug into bridges.  Gerd, agree?  Worth the risk?
> > > > > 
> > > > > Combining this with Julia's patches looks a bit risky to me and surely
> > > > > needs testing.  I expect the problematic case is both native and acpi
> > > > > hotplug being enabled.
> > > > >  When the guest uses acpi hotplug nobody will
> > > > > turn on slot power on the pcie root port ...
> > > > 
> > > > I'm not sure I understand what the situation is, and how to trigger it.
> > > > Could you clarify pls?
> > > > 
> > > > > I'd suggest to just revert to pcie native hotplug for 6.2.
> > > > 
> > > > Hmm that kind of change seems even riskier to me. I think I'll try with
> > > > Igor's patches.
> > > 
> > > Why would it be risky ? PCIE native hotplug is what we've used in
> > > QEMU for years & years, until 6.1 enabled the buggy ACPI hotplug.
> > > The behaviour of the current PCIE native hotplug impl is a known
> > > quantity.
> > > 
> > > Regards,
> > > Daniel
> > 
> > For example we might regress some of the bugs that the switch to ACPI fixed 
> > back to
> > 6.0 state. There were a bunch of these. For example it should be
> > possible for guests to disable native and use ACPI then, but isn't.
> 
> Of course there were bugs fixed by switching to ACPI, but we'd
> lived with native hotplug in production and the majority of
> the time it worked as users needed. The bugs were edge cases
> essentially only affecting a small subset of users.
> 
> The switch to ACPI broke the out of the box configuration for
> used by OpenStack. That's not an edge case, that's a serious
> impact.

Right. It's pretty easy for downstreams to switch back if they want to,
though.

> > I'm very willing to consider the switch back to native by default
> > but given the timing doing big changes like that at the last
> > minute seems unusual.
> 
> I consider it to be fixing a serious regression by going back
> to a known working safe impl, that has been used in production
> successfully for a long time. We know there are bugs with
> native hotplug, but they're *known* problems.

Are you familiar with the issues or are just arguing generally?  From my
POV they made native too unreliable to be useful.  It's great that the
narrow usecase of openstack managed not to hit any of the races
involved, but I think it's more a question of luck.  This until these
specific patches, but they are only in v2 out of rfc status just today.

> The unsual thing about timing is having a major functional
> regression identified in the previous release and then not
> even having patches propposed to fix it until after soft
> freeze for the next release arrives :-(
> 
> It doesn't give a feeling of confidence, but makes me
> wonder what other *unknown* problems we're liable to hit
> still.
> 
> Regards,
> Daniel

Right. And this applies to both approaches.  So I thought hard about the
balance here, and am inclined to go with a rough consensus opinion.

I don't see any reasons or in fact anyone objecting strongly to Igor's
patches, so I'm taking these. These are just bugfixes.

But if we also reach a rough consensus on others and e.g. Igor acks
Gerd's patch to revert to native then I think I'll merge it, it's
a close thing.

> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|