Re: [Qemu-devel] [PATCH] rtc: introduce nmi disable bit handler for cmos

[Gonglei (Arei)]

> From: Paolo Bonzini [mailto:address@hidden
> Sent: Monday, December 14, 2015 9:37 PM
> On 14/12/2015 14:27, Gonglei (Arei) wrote:
> >
> >>
> >> On 14/12/2015 13:49, Gonglei (Arei) wrote:
> >>>>>>> This patch introduce nmi disable bit handler to fix the problem
> >>>>>>> and make the emulated CMOS like the real hardware.
> >>>>>
> >>>>> I think that this only works with -machine kernel_irqchip=off, however.
> >>> IIRCC, the kernel_irqchip is disabled by default, and we used the
> >>> default
> >> value.
> >>
> >> No, it's enabled by default.
> >>
> >
> > Okay, yes, I saw the source code again. That means kmod finish the NMI
> > injection wrok, and the NMI will not pass Qemu side. So, you thought
> > this patch cannot block NMI injection when kernel_irqchip=on ?
> 
> I am not sure.  It depends on which NMIs are blocked by the bit.  For
> example, the IOAPIC can deliver NMIs, and they wouldn't be blocked.
> 
> Do you have any documentation, to see whether they can actually happen on
> emulated hardware?  I guess we support the TCO watchdog, so yes.
> 
Yes, watchdog is one case, and we have another case which need to use NMI to
tell guest do something when guest's cpu stuck or something like that.
And I can invoke qmp command "inject-nmi" when SeaBIOS try to close NMI by
invoking rtc_read() or rtc_write(). 

After the NMI injection, the guest will reboot:

[2015-12-14 16:41:57] In resume (status=0)
[2015-12-14 16:41:57] In 32bit resume
[2015-12-14 16:41:57] =====Attempting a hard reboot====
[2015-12-14 16:41:58] SeaBIOS (version 
rel-1.8.1-0-g4adadbd-20151214_135833-linux-jAPTBr)

[snip]

So, I think we should handle those scenarios, just like the real hardware.

> > Maybe we should pass the nmi_disable bit to Kmod when kernel_irqchip=on ,
> right?
> 
> Yes, that's the idea.
> 
That means I have much more work need to do.

> But first of all, I've read the thread you linked, and I couldn't find the 
> place
> where it says that the root cause is NMIs.
> 
That's complete true. I haven't direct proof, but I think I eliminated
all possible causes, except NMIs. Of course, if you find any other clues,
please let me know.

The most trouble thing is I couldn't reproduce this problem. :(

Thanks,
-Gonglei