Re: [PATCH] x86/nmi: Make external NMI injection reliably crash the host

[Ross Lagerwall <ross.lagerwall@citrix.com>]

On 08/26/2014 04:38 PM, Jan Beulich wrote:
>>>> On 26.08.14 at 17:26, <ross.lagerwall@citrix.com> wrote:
>> On 08/26/2014 01:59 PM, Jan Beulich wrote:
>>>>>> On 26.08.14 at 12:10, <ross.lagerwall@citrix.com> wrote:
>>>> @@ -3323,7 +3323,7 @@ void do_nmi(const struct cpu_user_regs *regs)
>>>>                pci_serr_error(regs);
>>>>            if ( reason & 0x40 )
>>>>                io_check_error(regs);
>>>> -        if ( !(reason & 0xc0) && !nmi_watchdog )
>>>> +        if ( !(reason & 0xc0) )
>>>>                unknown_nmi_error(regs, reason);
>>>
>>> As much as I like the original idea, I'm afraid this won't fly: I do
>>> know of systems where bad motherboard design leads to neither
>>> of these two bits ever getting set. I.e. at the very minimum we'd
>>> need a command line option to restore old behavior. Personally I
>>> think it should in fact remain default behavior, and new behavior
>>> should only be enabled via command line option.
>>
>> Well the old behavior was different depending on whether the watchdog
>> was enabled or not. Since the watchdog was disabled by default, that's
>> no different from the behavior here.
>>
>> So are you thinking something like an ignore_unknown_nmi boolean
>> parameter that defaults to true?
>
> More like a "watchdog=force" one, but right, since the watchdog
> isn't being enabled by default, maybe making it an opt-out instead
> of opt-in would indeed be acceptable.
>

If bad motherboard design leads to neither of these bits being set (thus 
always giving an unknown nmi error), can't the user set nmi=ignore on 
the xen command-line to get the previous behavior?

We already have an tristate nmi parameter, a boolean watchdog parameter, 
and a watchdog timeout parameter. I'm loathe to introduce even more 
possible states.

Regards
-- 
Ross Lagerwall