Re: [PATCH v9 7/7] arm64: kgdb: Roundup cpus using the debug IPI

[Marc Zyngier]

On 2023-08-07 11:54, Mark Rutland wrote:
> On Mon, Aug 07, 2023 at 11:47:04AM +0100, Marc Zyngier wrote:
> > On 2023-08-07 11:28, Mark Rutland wrote:
> > > On Thu, Jun 01, 2023 at 02:31:51PM -0700, Douglas Anderson wrote:
> > > > From: Sumit Garg <sumit.garg@xxxxxxxxxx>
> > > >
> > > > Let's use the debug IPI for rounding up CPUs in kgdb. When the debug
> > > > IPI is backed by an NMI (or pseudo NMI) then this will let us debug
> > > > even hard locked CPUs. When the debug IPI isn't backed by an NMI then
> > > > this won't really have any huge benefit but it will still work.
> > > >
> > > > Signed-off-by: Sumit Garg <sumit.garg@xxxxxxxxxx>
> > > > Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> > > > ---
> > > >
> > > > Changes in v9:
> > > > - Remove fallback for when debug IPI isn't available.
> > > > - Renamed "NMI IPI" to "debug IPI" since it might not be backed by
> > > > NMI.
> > > >
> > > >  arch/arm64/kernel/ipi_debug.c |  5 +++++
> > > >  arch/arm64/kernel/kgdb.c      | 14 ++++++++++++++
> > > >  2 files changed, 19 insertions(+)
> > >
> > > This looks fine to me, but I'd feel a bit happier if we had separate
> > > SGIs for
> > > the backtrace and the KGDB callback as they're logically unrelated.
> >
> > Well, we're a bit stuck here.
> >
> > We have exactly *one* spare SGI with GICv3, as we lose 8 of them
> > to the secure side. One possibility would be to mux some of the
> > lesser used IPIs onto two SGIs (one with standard priority, and
> > one with NMI priority).
>
> Understood; Doug and I suggested two options for that:
>
> 1) Unify/mux the IPI_CPU_STOP and IPI_CPU_CRASH_STOP IPIs
>
>    The only *intended* difference between the two is that 
> IPI_CPU_CRASH_STOP
>    calls crash_save_cpu() before trying to stop the CPU, but the
>    implementations have diverged significantly for unrelated reasons.
>
> 2) Remove IPI_WAKEUP
>
>    We only use IPI_WAKEUP for the ACPI parking protocol, and we could 
> reuse
>    another IPI (e.g. IPI_RESCHEDULE) to achieve the same thing witout a
>    dedicated IPI.

Sure. My concern is that we're papering over the fundamental problem,
which is that IPIs are limited resource, and that we're bound to pile
more stuff on them.

I'm all for reclaiming the ones that can be merged, but we may 
ultimately
need a real fix for this.

        M.
--
Jazz is not dead. It just smells funny...