> On May 20, 2020, at 1:10 PM, Thomas Gleixner <[email protected]> wrote:
>
> Peter Zijlstra <[email protected]> writes:
>>> On Mon, May 18, 2020 at 10:05:56AM +0200, Thomas Gleixner wrote:
>>> Peter Zijlstra <[email protected]> writes:
>>>> On Sat, May 16, 2020 at 01:45:51AM +0200, Thomas Gleixner wrote:
>>>>> --- a/arch/x86/kernel/nmi.c
>>>>> +++ b/arch/x86/kernel/nmi.c
>>>>> @@ -334,6 +334,7 @@ static noinstr void default_do_nmi(struc
>>>>> __this_cpu_write(last_nmi_rip, regs->ip);
>>>>>
>>>>> instrumentation_begin();
>>>>> + ftrace_nmi_handler_enter();
>>>>>
>>>>> handled = nmi_handle(NMI_LOCAL, regs);
>>>>> __this_cpu_add(nmi_stats.normal, handled);
>>>>> @@ -420,6 +421,7 @@ static noinstr void default_do_nmi(struc
>>>>> unknown_nmi_error(reason, regs);
>>>>>
>>>>> out:
>>>>> + ftrace_nmi_handler_exit();
>>>>> instrumentation_end();
>>>>> }
>>>>
>>>> Yeah, so I'm confused about this and the previous patch too. Why not
>>>> do just this? Remove that ftrace_nmi_handler.* crud from
>>>> nmi_{enter,exit}() and stick it here? Why do we needs the
>>>> nmi_{enter,exit}_notrace() thing?
>>>
>>> Because you then have to fixup _all_ architectures which use
>>> nmi_enter/exit().
>>
>> We probably have to anyway. But I can do that later I suppose.
>
> Second thoughts. For #DB and #INT3 we can just keep nmi_enter(), needs
> just annotation in nmi_enter() around that trace muck.
>
> For #NMI and #MCE I rather avoid the early trace call and do it once we
> have reached "stable" state, i.e. avoid it in the whole nested NMI mess.
>
>
What’s the issue? The actual meat is mostly in the asm for NMI, and for MCE
it’s just the sync-all-the-cores thing. The actual simultaneous NMI-and-MCE
case is utterly busted regardless, and I’ve been thinking about how to fix it.
It won’t be pretty, but nmi_enter() will have nothing to do with it.
