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

[Ankur Arora]

Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Wed, Nov 08, 2023 at 02:04:02AM -0800, Ankur Arora wrote:
>
>> >>   Revert "riscv: support PREEMPT_DYNAMIC with static keys"
>> >>   Revert "livepatch,sched: Add livepatch task switching to
>> >>     cond_resched()"
>> >>   Revert "arm64: Support PREEMPT_DYNAMIC"
>> >>   Revert "sched/preempt: Add PREEMPT_DYNAMIC using static keys"
>>
>> What's the best way to handle these? With the lazy bit, cond_resched()
>> and might_resched() are gone. So we don't need all of the static
>> key inftrastructure for toggling etc.
>>
>> The part of preempt_dynamic that makes sense to me is the one that
>> switches dynamically between none/voluntary/full. Here it would need
>> to be wired onto controls of the lazy bit.
>> (Right now the preemption policy is controlled by sched_feat in
>> patches 43, and 44 but sched/preempt is a much better interface.)
>
> I'm not understanding, those should stay obviously.
>
> The current preempt_dynamic stuff has 5 toggles:
>
> /*
>  * SC:cond_resched
>  * SC:might_resched
>  * SC:preempt_schedule
>  * SC:preempt_schedule_notrace
>  * SC:irqentry_exit_cond_resched
>  *
>  *
>  * NONE:
>  *   cond_resched               <- __cond_resched
>  *   might_resched              <- RET0
>  *   preempt_schedule           <- NOP
>  *   preempt_schedule_notrace   <- NOP
>  *   irqentry_exit_cond_resched <- NOP
>  *
>  * VOLUNTARY:
>  *   cond_resched               <- __cond_resched
>  *   might_resched              <- __cond_resched
>  *   preempt_schedule           <- NOP
>  *   preempt_schedule_notrace   <- NOP
>  *   irqentry_exit_cond_resched <- NOP
>  *
>  * FULL:
>  *   cond_resched               <- RET0
>  *   might_resched              <- RET0
>  *   preempt_schedule           <- preempt_schedule
>  *   preempt_schedule_notrace   <- preempt_schedule_notrace
>  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
>  */
>
> If you kill voluntary as we know it today, you can remove cond_resched
> and might_resched, but the remaining 3 are still needed to switch
> between NONE and FULL.

Ah now I see what you are saying.

Quick thought: even if we were running under NONE, eventually you'll
want to forcibly preempt out a CPU hog. So we will need to have
at least this one enabled always:

>  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched

These two, it might make sense to toggle them based on model.

>  *   preempt_schedule           <- preempt_schedule
>  *   preempt_schedule_notrace   <- preempt_schedule_notrace

Anyway let me think about this more and respond tomorrow.

For now, time for bed.

Thanks for clarifying btw.

Ankur

> Additionally, you'll get one new state to enable/disable the LAZY stuff.
> Neither NONE nor FULL want the LAZY thing on.
>
> You'll then end up with something like:
>
> /*
>  * SK:preempt_lazy
>  * SC:preempt_schedule
>  * SC:preempt_schedule_notrace
>  * SC:irqentry_exit_cond_resched
>  *
>  *
>  * NONE:
>  *   preempt_lazy		<- OFF
>  *   preempt_schedule           <- NOP
>  *   preempt_schedule_notrace   <- NOP
>  *   irqentry_exit_cond_resched <- NOP
>  *
>  * VOLUNTARY:
>  *   preempt_lazy		<- ON
>  *   preempt_schedule           <- preempt_schedule
>  *   preempt_schedule_notrace   <- preempt_schedule_notrace
>  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
>  *
>  * FULL:
>  *   preempt_lazy		<- OFF
>  *   preempt_schedule           <- preempt_schedule
>  *   preempt_schedule_notrace   <- preempt_schedule_notrace
>  *   irqentry_exit_cond_resched <- irqentry_exit_cond_resched
>  */
>
> For the architectures that do not have static_call but instead use
> static_key for everything, the SC's are obviously static_key based
> wrappers around the function calls -- like now.