Re: [PATCH] irqchip/riscv-imsic: Synchronize remote state with IRQ work

From: Anup Patel

Date: Tue Nov 10 2026 - 02:47:26 EST


On Tue, Nov 10, 2026 at 12:57 PM Troy Mitchell <troy.mitchell@xxxxxxxxx> wrote:
>
> On Wed Nov 4, 2026 at 9:59 AM +08, Troy Mitchell wrote:
> > On Tue Nov 3, 2026 at 11:52 PM +08, Anup Patel wrote:
> > > On Tue, Nov 3, 2026 at 1:34 PM Troy Mitchell <troy.mitchell@xxxxxxxxx> wrote:
> > > >
> > > > Remote IMSIC state updates are deferred to a timer on the target CPU.
> > > > If that CPU has stopped its local clockevent and is waiting for a
> > > > broadcast interrupt, enabling the broadcast IRQ can depend on delivery
> > > > of that same interrupt. This can stall timer delivery when the broadcast
> > > > IRQ is unmasked or moved to another CPU, including during CPU hotplug.
> > > >
> > > > Queue hard IRQ work to synchronize remote state independently of timer
> > > > delivery. Keep synchronous updates on the local CPU and the timer retry
> > > > for an unfinished vector move. Hard IRQ work also avoids depending on a
> > > > schedulable worker when PREEMPT_RT is enabled.
> > >
> > > Using timer retry for local CPU and IRQ work queue for remote CPUs looks
> > > strange. Why not use IRQ work queue even for local CPU ??
> > On the local CPU the sync runs synchronously, so there is nothing to
> > queue. The timer only retries an unfinished move, and by then the new
> > vector is already enabled, so it cannot stall on the broadcast IRQ.
> > Requeueing IRQ work from its own handler would just keep self-IPIing
> > until the old CPU catches up.
> >
> > I'll add more comments and modify the commit body.
> That was not the whole story. The retry cannot stall, but it can be
> late. The local CPU is usually the one entering idle and taking the
> broadcast IRQ, and the timer it arms on itself there is not programmed
> because its tick is already stopped. Nothing tells it when the old CPU
> is done, so the move only completes on its next interrupt. On K3 with
> deep idle I see move windows of up to 0.5s.
>
> So you are right that the timer should go, just not by requeueing on
> the local CPU. In v2 the old CPU will queue the IRQ work on the new CPU
> once it has released the old vector, and the retry timer will be
> removed. Move windows drop to about 100us. This will be a separate
> patch on top of this one.
>

Sounds good.

Regards,
Anup