Re: [PATCH 6/7] xen: Enable PV clocksource for HVM

[Sheng Yang <sheng@linux.intel.com>]

On Friday 05 March 2010 17:31:47 Ian Campbell wrote:
> On Fri, 2010-03-05 at 02:54 +0000, Sheng Yang wrote:
> > On Friday 05 March 2010 01:40:51 Jeremy Fitzhardinge wrote:
> > > On 03/04/2010 01:36 AM, Sheng Yang wrote:
> > > > And enable it by default in PV extended HVM guest.
> > > >
> > > > Signed-off-by: Sheng Yang<sheng@linux.intel.com>
> > > > ---
> > > >   arch/x86/xen/hvmpv.c |   20 ++++++++++++++++++++
> > > >   arch/x86/xen/time.c  |    4 +++-
> > > >   2 files changed, 23 insertions(+), 1 deletions(-)
> > > >
> > > > diff --git a/arch/x86/xen/hvmpv.c b/arch/x86/xen/hvmpv.c
> > > > index 7a741ed..284e021 100644
> > > > --- a/arch/x86/xen/hvmpv.c
> > > > +++ b/arch/x86/xen/hvmpv.c
> > > > @@ -43,6 +43,7 @@ static void __init xen_hvm_pv_banner(void)
> > > >   		pv_info.name);
> > > >   	printk(KERN_INFO "Xen version: %d.%d%s\n",
> > > >   		version>>  16, version&  0xffff, extra.extraversion);
> > > > +	printk(KERN_INFO "PV feature: PV clocksource enabled\n");
> > > >   }
> > > >
> > > >   static int __init xen_para_available(void)
> > > > @@ -117,6 +118,22 @@ static void __init init_shared_info(void)
> > > >   	per_cpu(xen_vcpu, 0) =&HYPERVISOR_shared_info->vcpu_info[0];
> > > >   }
> > > >
> > > > +extern struct clocksource xen_clocksource;
> > >
> > > No externs in .c files.  But aside from that, I'd prefer you export a
> > > function from xen/time.c to do the clocksource registration and call it
> > > from both places (even if its a one-liner).
> >
> > OK
> >
> > > > +
> > > > +static void __init init_pv_clocksource(void)
> > > > +{
> > > > +	if (enable_hvm_pv(HVM_PV_CLOCK))
> > > > +		BUG();
> > >
> > > BUG is a bit severe.  Will it really never, ever fail?  And if it does,
> > > the consequence is hardly serious; we just fall back to emulated
> > > devices.
> >
> > It shouldn't fail. If we got a hypervisor without these extension, this
> > one won't be called. Because the feature should be detected by CPUID, and
> > the hypervisor would mask the unsupported features. So the guest won't
> > see that bit if it's not supported, and this one wouldn't be involved.
> 
> It is still possible that a hypervisor might want to turn on the general
> extensions but not the PV clock extensions, and we can't predict what
> reasons we might have for doing that in the future. A BUG() is an
> awfully big hammer for a failure like this and destroys any hope of
> making changes in a forward/backward compatible manner in the future.
> Surely it is possible to simply continue with non-PV clock?

Evtchn can't work with it. 

If the potentially disable pv clock is your concern, I think the better way is 
to make PV clock as a separate feature, and let PV evtchn depends on it. I've 
updated the patches to reflect this change.

-- 
regards
Yang, Sheng