Re: [Qemu-devel] [PATCH 1/1 V6 resent ] qemu-kvm: fix improper nmi emulation

[Marcelo Tosatti]

On Tue, Oct 25, 2011 at 05:55:28PM +0800, Lai Jiangshan wrote:
> Previous discussions------------:
> >> >>
> >> >> Which approach you prefer to?
> >> >> I need to know the result before wasting too much time to respin
> >> >> the approach.
> > > 
> > > Yes, sorry about the slow and sometimes conflicting feedback.
> > > 
> >> >> 1) Fix KVM_NMI emulation approach  (which is v3 patchset)
> >> >>         - It directly fixes the problem and matches the
> >> >>           real hard ware more, but it changes KVM_NMI bahavior.
> >> >>         - Require both kernel-site and userspace-site fix.
> >> >>
> >> >> 2) Get the LAPIC state from kernel irqchip, and inject NMI if it is 
> >> >> allowed
> >> >>    (which is v4 patchset)
> >> >>         - Simple, don't changes any kernel behavior.
> >> >>         - Only need the userspace-site fix
> >> >>
> >> >> 3) Add KVM_SET_LINT1 approach (which is v5 patchset)
> >> >>         - don't changes the kernel's KVM_NMI behavior.
> >> >>         - much complex
> >> >>         - Require both kernel-site and userspace-site fix.
> >> >>         - userspace-site should also handle the !KVM_SET_LINT1
> >> >>           condition, it uses all the 2) approach' code. it means
> >> >>           this approach equals the 2) approach + KVM_SET_LINT1 ioctl.
> >> >>
> >> >> This is an urgent bug of us, we need to settle it down soo
> > > 
> > > While (1) is simple, it overloads a single ioctl with two meanings,
> > > that's not so good.
> > > 
> > > Whether we do (1) or (3), we need (2) as well, for older kernels.
> > > 
> > > So I recommend first focusing on (2) and merging it, then doing (3).
> > > 
> > > (note an additional issue with 3 is whether to make it a vm or vcpu
> > > ioctl - we've been assuming vcpu ioctl but it's not necessarily the best
> > > choice).
> > > 
> It is the 2) approach.
> It only changes the user space site, the kernel site is not touched.
> It is changed from previous v4 patch, fixed problems found by Jan.
> ----------------------------end previous discussions
> 
> 
> From: Lai Jiangshan <address@hidden>
> 
> 
> Currently, NMI interrupt is blindly sent to all the vCPUs when NMI
> button event happens. This doesn't properly emulate real hardware on
> which NMI button event triggers LINT1. Because of this, NMI is sent to
> the processor even when LINT1 is maskied in LVT. For example, this
> causes the problem that kdump initiated by NMI sometimes doesn't work
> on KVM, because kdump assumes NMI is masked on CPUs other than CPU0.
> 
> With this patch, inject-nmi request is handled as follows.
> 
> - When in-kernel irqchip is disabled, deliver LINT1 instead of NMI
>   interrupt.
> - When in-kernel irqchip is enabled, get the in-kernel LAPIC states
>   and test the APIC_LVT_MASKED, if LINT1 is unmasked, and then
>   delivering the NMI directly. (Suggested by Jan Kiszka)
> 
> Changed from old version:
>   re-implement it by the Jan's suggestion.
>   fix the race found by Jan.
> 
> Signed-off-by: Lai Jiangshan <address@hidden>
> Reported-by: Kenji Kaneshige <address@hidden>
> Acked-by: Avi Kivity <address@hidden>
> Acked-by: Jan Kiszka <address@hidden>

Please rebase.