Re: [PATCH v2] x86, hyperv: fix kernel panic when kexec on HyperV

[Peter Zijlstra]

On Tue, Mar 05, 2019 at 09:34:13PM +0800, Kairui Song wrote:
> On Tue, Mar 5, 2019 at 8:33 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Tue, Mar 05, 2019 at 08:17:03PM +0800, Kairui Song wrote:
> > > diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
> > > index 7abb09e2eeb8..34aa1e953dfc 100644
> > > --- a/arch/x86/hyperv/hv_init.c
> > > +++ b/arch/x86/hyperv/hv_init.c
> > > @@ -406,6 +406,12 @@ void hyperv_cleanup(void)
> > >       /* Reset our OS id */
> > >       wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
> > >
> > > +     /* Cleanup hypercall page reference before reset the page */
> > > +     hv_hypercall_pg = NULL;
> > > +
> > > +     /* Make sure page reference is cleared before wrmsr */
> >
> > This comment forgets to tell us who cares about this. And why the wrmsr
> > itself isn't serializing enough.
> >
> > > +     wmb();
> > > +
> > >       /* Reset the hypercall page */
> > >       hypercall_msr.as_uint64 = 0;
> > >       wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
> >
> > That looks like a fake MSR; and you're telling me that VMEXIT doesn't
> > serialize?
> 
> Thanks for the review, seem I being a bit paranoid on this. Will drop
> it and send a v3 if no one has any other complaint.

Also; our wrmsr implementation has a "memory" clobber on, which means
you don't even need a compiler barrier to force emit that store before
the wrmsr.

You might want to retain the comment that explains this ordering though.

	hv_hypercall_pg = NULL;
	/*
	 * text that explains why hv_hypercall_pg must be set NULL
	 * before the wrmsr
	 */
	wrmsrl(...);