Mailing List Archive: [Xen-merge] Re: [patch] SMP alternatives for i386

[Xen-merge] Re: [patch] SMP alternatives for i386

Dec 13, 2005, 8:58 AM

Post #1 of 5 (414 views)

The SMP problems with self modifying code are very real but the hotplug
architecture for CPU already deals with these synchronizations and you
can use the SMP cross calls to 'capture' the other processors. The x86
case of processors with differing feature sets is extremely common on
PII/PIII systems and can also occur on other platforms where features
like SSE3 were introduced in later steppings. Clean alternatives support
for such platforms automatically would be good.

Other than that it looks sane to me (well the CPU hotplug case is silly
but the rest makes sense). The various CPU errata come out ok that I can
see as well because they use rmb/wmb/etc and those don't change
behaviour so still maintain store order against the bus.

_______________________________________________
Xen-merge mailing list
Xen-merge@lists.xensource.com
http://lists.xensource.com/xen-merge

[Xen-merge] Re: [patch] SMP alternatives for i386 [ In reply to ]

zwane at arm

Dec 14, 2005, 12:35 PM

Post #2 of 5 (406 views)

Permalink

diff -urN -x 'build-*' -x '*~' -x Make -x scripts linux-2.6.15-rc5/arch/i386/kernel/smpboot.c work-2.6.15-rc5/arch/i386/kernel/smpboot.c
--- linux-2.6.15-rc5/arch/i386/kernel/smpboot.c 2005-12-06 17:00:36.000000000 +0100
+++ work-2.6.15-rc5/arch/i386/kernel/smpboot.c 2005-12-06 17:06:48.000000000 +0100
@@ -1351,6 +1352,9 @@
fixup_irqs(map);
/* It's now safe to remove this processor from the online map */
cpu_clear(cpu, cpu_online_map);
+
+ if (1 == num_online_cpus())
+ alternatives_smp_switch(0);
return 0;
}

Is that really safe there? Ideally you want to do the switch when the
processor going offline is no longer executing kernel code.

_______________________________________________
Xen-merge mailing list
Xen-merge@lists.xensource.com
http://lists.xensource.com/xen-merge

Re: [Xen-merge] Re: [patch] SMP alternatives for i386 [ In reply to ]

Keir.Fraser at cl

Dec 14, 2005, 1:15 PM

Post #3 of 5 (411 views)

Permalink

On 14 Dec 2005, at 19:35, Zwane Mwaikambo wrote:

> diff -urN -x 'build-*' -x '*~' -x Make -x scripts
> linux-2.6.15-rc5/arch/i386/kernel/smpboot.c
> work-2.6.15-rc5/arch/i386/kernel/smpboot.c
> --- linux-2.6.15-rc5/arch/i386/kernel/smpboot.c 2005-12-06
> 17:00:36.000000000 +0100
> +++ work-2.6.15-rc5/arch/i386/kernel/smpboot.c 2005-12-06
> 17:06:48.000000000 +0100
> @@ -1351,6 +1352,9 @@
> fixup_irqs(map);
> /* It's now safe to remove this processor from the online map */
> cpu_clear(cpu, cpu_online_map);
> +
> + if (1 == num_online_cpus())
> + alternatives_smp_switch(0);
> return 0;
> }
>
> Is that really safe there? Ideally you want to do the switch when the
> processor going offline is no longer executing kernel code.

Perhaps that check belongs at the end of __cpu_die()? That's where it
lives in xen's smpboot.c.

-- Keir

_______________________________________________
Xen-merge mailing list
Xen-merge@lists.xensource.com
http://lists.xensource.com/xen-merge

Re: [Xen-merge] Re: [patch] SMP alternatives for i386 [ In reply to ]

zwane at arm

Dec 14, 2005, 6:20 PM

Post #4 of 5 (405 views)

Permalink

Hello Keir,

On Wed, 14 Dec 2005, Keir Fraser wrote:

> > fixup_irqs(map);
> > /* It's now safe to remove this processor from the online map */
> > cpu_clear(cpu, cpu_online_map);
> > +
> > + if (1 == num_online_cpus())
> > + alternatives_smp_switch(0);
> > return 0;
> > }
> >
> > Is that really safe there? Ideally you want to do the switch when the
> > processor going offline is no longer executing kernel code.
>
> Perhaps that check belongs at the end of __cpu_die()? That's where it lives in
> xen's smpboot.c.

Yes indeed, end of __cpu_die would be perfect for x86 too as that's where
CPU_DEAD acknowledge is checked.

_______________________________________________
Xen-merge mailing list
Xen-merge@lists.xensource.com
http://lists.xensource.com/xen-merge

Re: [Xen-merge] Re: [patch] SMP alternatives for i386 [ In reply to ]

kraxel at suse

Dec 15, 2005, 6:44 AM

Post #5 of 5 (405 views)

Permalink

>>> + if (1 == num_online_cpus())
>>> + alternatives_smp_switch(0);

> Yes indeed, end of __cpu_die would be perfect for x86 too as that's where
> CPU_DEAD acknowledge is checked.

Good point, fixed patch attached.

cheers,

Gerd

Mailing List Archive

Mailing List Archive

Attached Files: