Re: [PATCH] x86/smp: Factor out parts of native_smp_prepare_cpus()

[Peter Zijlstra]

On Mon, Nov 08, 2021 at 12:20:26PM -0500, Boris Ostrovsky wrote:
> 
> On 11/8/21 10:11 AM, Peter Zijlstra wrote:
> > On Tue, Nov 02, 2021 at 07:36:36PM -0400, Boris Ostrovsky wrote:
> > > Commit 66558b730f25 ("sched: Add cluster scheduler level for x86")
> > > introduced cpu_l2c_shared_map mask which is expected to be initialized
> > > by smp_op.smp_prepare_cpus(). That commit only updated
> > > native_smp_prepare_cpus() version but not xen_pv_smp_prepare_cpus().
> > > As result Xen PV guests crash in set_cpu_sibling_map().
> > > 
> > > While the new mask can be allocated in xen_pv_smp_prepare_cpus() one can
> > > see that both versions of smp_prepare_cpus ops share a number of common
> > > operations that can be factored out. So do that instead.
> > > 
> > > Fixes: 66558b730f25 ("sched: Add cluster scheduler level for x86")
> > > Signed-off-by: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
> > Thanks! I'll go stick that somewhere /urgent (I've had another report on
> > that here:
> > 
> >    https://lkml.kernel.org/r/20211105074139.GE174703@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> > )
> 
> 
> Thank you. (I don't see this message btw)

Urgh, that thread never went to lkml :/

> > But looking at those functions; there seems to be more spurious
> > differences. For example, the whole sched_topology thing.
> 
> 
> I did look at that and thought this should be benign given that Xen PV
> is not really topology-aware. I didn't see anything that would be a
> cause for concern but perhaps you can point me to things I missed.

And me not being Xen aware... What does Xen-PV guests see of the CPUID
topology fields? Does it fully sanitize the CPUID data, or is it a clean
pass-through from whatever CPU the vCPU happens to run on at the time?


> > Should we re-architect this whole smp_prepare_cpus() thing instead? Have
> > a common function and a guest function? HyperV for instance seems to
> > call native_smp_prepare_cpus() and then does something extra (as does
> > xen_hvm).
> 
> 
> Something like
> 
> 
> void smp_prepare_cpus()
> 
> {
> 
>     // Code that this patch moved to smp_prepare_cpus_common();
> 
> 
>    smp_ops.smp_prepare_cpus();  // Including baremetal
> 
> }
> 
> 
> ?
> 
> 
> XenHVM and hyperV will need to call native smp_op too. Not sure this
> will be prettier than what it is now?

Hurmph, yeah. I was thinking it would allow pre and post common code,
but yeah, doesn't seem to make sense for now.