Re: [PATCH v2 4/5] clocksource: mips-gic-timer: Set next GIC event on the correct VP

[Thomas Gleixner <tglx@kernel.org>]

On Tue, Sep 01 2026 at 17:53, Benoît Monin wrote:
> On Thursday, 20 August 2026 at 10:41:29 CEST, Thomas Gleixner wrote:
>> > On a multi-cluster system each cluster has its own GIC. The VO window of
>> > the local GIC is what write_gic_vl_other() redirects, so when the target
>> > CPU belongs to another cluster, write_gic_vo_compare() lands in the
>> > wrong GIC, overwriting the COMPARE register of a VP in the local cluster.
>> 
>> The MIPS GIC clockevent device is strictly registered per CPU with the
>> cpumask of that CPU. See gic_clockevent_cpu_init().
>> 
>> The set_next_event() callback of a per CPU clockevent device is _always_
>> invoked on that CPU and cannot be invoked on a different CPU.
>> 
>> So I completely fail to understand what this cluster and non-current CPU
>> voodoo is about.
>> 
> Going through the log, here is what I found:
>
> The commit 414408d0eedd ("MIPS: allow GIC clockevent device config
> from other CPUs") was merged in v3.16, its intent was to make the GIC
> clockevent usable as a tick broadcast device.

Ah. Nice detective work.

> Also in v3.16, the commit 0467811e9bdb ("MIPS: mark GIC clockevent device
> with CLOCK_EVT_FEAT_C3STOP") sets the C3STOP flag. With this flag, as
> I understand it, the GIC clockevent cannot be used as a tick broadcast
> device. So the set_next_event() callback is always invoked on the CPU
> of the event.

Correct. A broadcast device cannot stop. The whole point of a broadcast
decive is to act as a global backup for per CPU devices which are
affected by C3STOP. The PPC people added a hrtimer broadcast device to
handle the case where no other devices are available

> The `if/else` was added later in v4.15 by commit f16ff2bdb135
> ("clocksource/drivers/mips-gic-timer: Add fastpath for local timer
> updates"). But the `else` path was never called.

So my assumption about voodoo was not too far off :)

> I am going to drop the `else` path, should I also add the
> CLOCK_EVT_FEAT_PERCPU flag, or is C3STOP and the cpumask enough?

You can, though the device i already excluded from being qualified for
broadcast by C3STOP. But for clarity sake, yes.

Thanks,

        tglx