Re: [Qemu-ppc] [Qemu-devel] [PATCH 0/5] spapr: DRC cleanups (part VI)

[Daniel Henrique Barboza]

On 07/07/2017 04:14 AM, David Gibson wrote:
> On Thu, Jul 06, 2017 at 11:31:35AM -0300, Daniel Henrique Barboza wrote:
> > On 07/05/2017 06:53 PM, Daniel Henrique Barboza wrote:
> > > Worth a try. I'll see if I can make a POC of this DRC reset at incoming
> > > migration time.
> >
> > Just played a little with the idea of manual reset during migration. I've
> > created a
> > POC that resets the CPU DRCs in a pre_load hook.
> Ok.  I'm not sure does the pre_load hook get called for objects where
> nothing ends up being sent in the migration stream?

In this POC I've added a pre_load hook in the machine state to ensure
that it will be called.  I saw that we already had a post_load hook in there
that does adjustments in the CPU and RADIX state, so adding a pre_load hook
sounded viable too. I am not sure if we can add some form of hook for
the DRCs that won't be migrated in the DRC state or if there is a better
VMSD to do the reset of the DRCs.

Anyway, this was more of an experiment than a serious fix proposal.
I believe the general idea is solid but it'll get a couple of spins to get a
good patch for pushing. I'll send the POC as a RFC to get things started and
we'll see what happens.


Daniel

> > Then I've done the same
> > experiment - device_add on both source and target before the migration,
> > hot unplug the device after migration is completed. The hot unplug worked as
> > expected in both QEMU and guest.
> That's encouraging.
>
> > To minimize the impact I am resetting only the DRCs of the CPUs that were
> > hotplugged
> > in the target instead of resetting everybody.
> That sounds like more trouble than it's worth, I think it'll be easier
> and safer to reset all DRCs.
>
> > I'll see if this solution works for LMBs and PCI devices. In case
> > affirmative, and
> > if we are fine with this solution of resetting the DRCs in pre_load (not
> > sure if pre_load is
> > the right place for doing it - suggestions welcome), I can send a patch to
> > be applied on top
> > of this series.
> >
> >
> >
> > Thanks,
> >
> > Daniel