Re: [Qemu-devel] [PATCH v3 1/4] cpus: Define NMI callback

[Cornelia Huck]

On Wed, 11 Jun 2014 16:50:35 +1000
Alexey Kardashevskiy <address@hidden> wrote:

> On 06/11/2014 12:41 AM, Cornelia Huck wrote:
> > On Tue, 10 Jun 2014 09:39:51 -0400
> > Luiz Capitulino <address@hidden> wrote:
> > 
> >> On Wed,  4 Jun 2014 18:08:47 +1000
> >> Alexey Kardashevskiy <address@hidden> wrote:
> >>
> >>> This introduces an NMI (non maskable interrupt) callback per CPU class
> >>> which QMP's "nmi" command may use to issue NMI on a CPU.
> >>>
> >>> This adds support for it in qmp_inject_nmi(). Since no architecture
> >>> supports it at the moment, there is no change in behaviour.
> >>>
> >>> This changes inject-nmi command description for HMP and QMP.
> >>>
> >>> Signed-off-by: Alexey Kardashevskiy <address@hidden>
> >>> ---
> >>> Changes:
> >>> v3:
> >>> * actual nmi() enablement moved from last patch to first patch
> >>> * changed description for QMP command too
> >>> ---
> >>>  cpus.c            | 11 ++++++++++-
> >>>  hmp-commands.hx   |  6 ++----
> >>>  include/qom/cpu.h |  1 +
> >>>  qapi-schema.json  |  4 +---
> >>>  qmp-commands.hx   |  3 +--
> >>>  5 files changed, 15 insertions(+), 10 deletions(-)
> >>>

> >>> diff --git a/qapi-schema.json b/qapi-schema.json
> >>> index 7bc33ea..dcf6642 100644
> >>> --- a/qapi-schema.json
> >>> +++ b/qapi-schema.json
> >>> @@ -1748,13 +1748,11 @@
> >>>  ##
> >>>  # @inject-nmi:
> >>>  #
> >>> -# Injects an Non-Maskable Interrupt into all guest's VCPUs.
> >>> +# Injects an Non-Maskable Interrupt into the given guest's VCPU.
> >>
> >> QMP doesn't have the concept of "current monitored CPU" you talk in the
> >> intro email. In QMP you have to specify the CPU. You have to choices:
> >>
> >>  - Add a new command that takes a CPU parameter (seems the best to me, as
> >>    people were asking for a different command anyways)
> >>
> >>  - Add an optional parameter to inject-nmi. When the CPU parameter is
> >>    not specified, the command sends the NMI to all CPUs
> > 
> > The s390 restart interrupt is a per-vcpu interrupt, which we really
> > don't want to inject on _all_ vcpus. OTOH, we want to inject that
> > interrupt on any vcpu - we don't care which one it is. So I'd really
> > like an "inject nmi on default cpu" option.
> 
> 
> Interesting thing. I was pushing "nmi" for ppc which did nmi only on the
> current cpu but according to our architect Ben, I should inject NMI on all
> CPUs because although XMON (inkernel debugger) tries to stop other CPUs, it
> may fail to do that because of some interrupt deadlock. And - this is
> important for us - this is what pHyp (native IBM hypervisor) does.
> 
> Now I wonder how you are getting away with injecting NMI on one CPU only
> and what will happen if we inject NMI on all?

The Linux restart handler does stop the other cpus before doing any
other work - see restart_int_handler in arch/s390/kernel/entry64.S and
do_restart in arch/s390/kernel/ipl.c for the gory details ;)

Other operating systems need to have something similar in place, as the
restart interrupt is always directed to a single cpu (for example, z/VM
has SYSTEM RESTART, which is always directed either to the default cpu
or to a specified cpu).