Re: [RFC PATCH 00/86] Make the kernel preemptible

[Ankur Arora]

Mark Rutland <mark.rutland@xxxxxxx> writes:

> On Wed, Nov 08, 2023 at 04:34:41PM -0800, Ankur Arora wrote:
>> Mark Rutland <mark.rutland@xxxxxxx> writes:
>>
>> > On Tue, Nov 07, 2023 at 01:56:46PM -0800, Ankur Arora wrote:
>> >> What's broken:
>> >>  - ARCH_NO_PREEMPT (See patch-45 "preempt: ARCH_NO_PREEMPT only preempts
>> >>    lazily")
>> >>  - Non-x86 architectures. It's trivial to support other archs (only need
>> >>    to add TIF_NEED_RESCHED_LAZY) but wanted to hold off until I got some
>> >>    comments on the series.
>> >>    (From some testing on arm64, didn't find any surprises.)
>> >
>> > When you say "testing on arm64, didn't find any surprises", I assume you mean
>> > with an additional patch adding TIF_NEED_RESCHED_LAZY?
>>
>> Yeah. And, handling that in the user exit path.
>>
>> > Note that since arm64 doesn't use the generic entry code, that also requires
>> > changes to arm64_preempt_schedule_irq() in arch/arm64/kernel/entry-common.c, to
>> > handle TIF_NEED_RESCHED_LAZY.
>>
>> So, the intent (which got muddied due to this overly large series)
>> was to delay handling TIF_NEED_RESCHED_LAZY until we are about to
>> return to user.
>
> Ah, I missed that detail -- thanks for clarifying!
>
>> I think arm64_preempt_schedule_irq() should only handle TIF_NEED_RESCHED
>> and the _TIF_NEED_RESCHED_LAZY should be handled via _TIF_WORK_MASK
>> and do_notify_resume().
>
> Digging a bit more, I think that should still work.
>
> One slight clarification: arm64_preempt_schedule_irq() doesn't look at
> TIF_NEED_RESCHED today, as it relies on the scheduler IPI calling
> preempt_fold_need_resched() to propogate TIF_NEED_RESCHED into
> PREEMPT_NEED_RESCHED. That should still work since this series makes
> preempt_fold_need_resched() check tif_need_resched(RESCHED_eager).
>
> I was a bit cnofused because in the generic entry code,
> irqentry_exit_cond_resched() explicitly checks for TIF_NEED_RESCHED, and I'm
> not sure why it does that rather than relying on the scheduler IPI as above.

Yeah I found that confusing as well. I suspect the reason is that not
all archs do the folding and we need the explicit check for those that
don't.


--
ankur