Re: [PATCH] Use BIOS reboot on Toshiba Portege 4000

[Eduardo Habkost]

On Mon, Nov 03, 2008 at 08:13:07AM -0800, Eric W. Biederman wrote:
> Avi Kivity <avi@xxxxxxxxxx> writes:
> 
> > Eduardo Habkost wrote:
> >> We probably need to disable vmx on all CPUs, but emergency reboot skips
> >> native_smp_send_stop() (where we could hook a virt_disable call in).
> >>
> >> As relying on IPIs defeats the whole point of emergency_restart, a proper
> >> fix will need to use NMIs like the kdump code does.
> >>
> >
> > They should use the same code; they have a similar environment at entry and
> > reliability requirements for e_r are not greater than kdump's.
> 
> Just a sec.
> 
> I think we are confusing two issues here.
> 
> - Ordinary machine_restart which happens to call emergency_restart.
>   And is proceeded by machine_shutdown.
> 
> - And emergency_restart itself.
> 

I am considering only emergency_restart() itself, that simply reboots
the machine. All other cases should be calling the KVM reboot notifier,
that disables virtualization on all CPUs.


> To some extent I would be a lot happier if Alt-sysrq-r did what
> was necessary to get into a context where it can call machine_restart
> or even better kernel_restart().
> emergency_restart() is a nice idea but is broken by design.

Eliminating emergency_restart() looks good in theory, but can we
eliminate it on all cases? We need something for cases when we are on
a possibly-broken state and we want to reboot the machine as reliably
as possible.


> 
> That said.  If we can turn off vmx on that one processor.
> That should be enough for the cpu to triple fault and let
> the BIOS do what it needs to do on that cpu i.e. outb(magic, 0x92)
> and toggle a motherboard level reset?
> 
> If I read the earlier comments correctly the deep issue is
> that going through ACPI to reset systems is less reliable than
> doing it the classic way.

That's what Ingo suggested: instead of defaulting to ACPI reboot,
disable VMX before rebooting if needed and get back to a safer default.

That leads us to the NMI stuff: we need to disable VMX on all CPUs,
and we can't use IPIs if we are on a possibly-broken state.

-- 
Eduardo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/