Re: [PATCH] armpmu: broadcast overflow irq on multi-core system having one muxed SPI for PMU.

From: Mark Rutland
Date: Thu May 10 2018 - 06:21:32 EST


Hi,
On Thu, May 10, 2018 at 05:36:17PM +0900, Hoeun Ryu wrote:
> From: Hoeun Ryu <hoeun.ryu@xxxxxxx>
>
> On some SoCs like i.MX6DL/QL have only one muxed SPI for multi-core system.
> On the systems, a CPU can be interrupted by overflow irq but it is possible that
> the overflow actually occurs on another CPU.

Muxing the PMU IRQs is a really broken system design, and there's no
good way of supporting it.

What we should do for such systems is:

* Add a flag to the DT to describe that the IRQs are muxed, as this
cannot be probed.

* Add hrtimer code to periodically update the counters, to avoid
overflow (e.g. as we do in the l2x0 PMU).

* Reject sampling for such systems, as this cannot be done reliably or
efficiently.

NAK to broadcasting the IRQ -- there are a number of issues with the
general approach.

We should update the PMU probing code to warn when we have fewer IRQs
than CPUs, and fail gracefully to the above.

[...]

> static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
> {

> + /* smp_call_function cannot be called with irq disabled */
> + local_irq_enable();
> + preempt_disable();
> + smp_call_function_many(&mask, __armpmu_handle_irq, dev, 0);
> + preempt_enable();
> + local_irq_disable();

For many reasons, this sequence is not safe.

It is not safe to enable IRQs in irq handlers. Please never do this.

Thus it's also never safe to call smp_call_function*() in IRQ handlers.

Futher, If you ever encounter a case where you need to avoid preemption
across enabling IRQs, preemption must be disabled *before* enabling
IRQs.

Thanks,
Mark.