Re: [patch] sched: fix scheduling latencies for !PREEMPT kernels

[Nick Piggin <nickpiggin@yahoo.com.au>]

Ingo Molnar wrote:
> * Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> 
> 
>>OK.
>>
>>Alternatively, I'd say tell everyone who wants really low latency to
>>enable CONFIG_PREEMPT, which automatically gives the minimum possible
>>preempt latency, delimited (and defined) by critical sections, instead
>>of the more ad-hoc "sprinkling" ;)
> 
> 
> it's not ad-hoc. These are the 10 remaining points for which there is no
> natural might_sleep() point nearby (according to measurements). That's
> why i called them 'complementary'. They cause zero problems for the
> normal kernel (we already have another 70 cond_resched() points), but
> they _are_ the ones needed in addition if might_sleep() also does
> cond_resched().
> 

No, I mean putting cond_resched in might_sleep is ad hoc. But that
doesn't mean it doesn't work - it obviously does ;)

> the 'reliability' of latency break-up depends on the basic preemption
> model. Believe me, even with CONFIG_PREEMPT there were a boatload of
> critical sections that had insanely long latencies that nobody fixed
> until the VP patchset came along. Without CONFIG_PREEMPT the number of
> possibly latency-paths increases, but the situation is the same as with
> CONFIG_PREEMPT: you need tools, people that test stuff and lots of
> manual work to break them up reliably. You will never be 'done' but you
> can do a reasonably good job for workloads that people care about.
> 

All the other stuff in your patches are obviously very important with
or without "full preempt", and make up the bulk of the *hard* work.
You have no arguments from me about that.

> the 'final' preemption model [for hard-RT purposes] that i believe will
> make it into the Linux kernel one nice day is total preemptability of
> everything but the core preemption code (i.e. the scheduler and
> interrupt controllers). _That_ might be something that has provable
> latencies. Note that such a 'total preemption' model has prerequisites
> too, like the deterministic execution of hardirqs/softirqs.
> 
> note that the current lock-break-up activities still make alot of sense
> even under the total-preemption model: it decreases the latency of
> kernel-using hard-RT applications. (raw total preemption only guarantees
> quick scheduling of the hard-RT task - it doesnt guarantee that the task
> can complete any useful kernel/syscall work.)
> 
> since we already see at least 4 different viable preemption models
> placed on different points in the 'latency reliability' spectrum, it
> makes little sense to settle for any of them. So i'm aiming to keep the
> core code flexible to have them all without much fuss, and usage will
> decide which ones are needed. Maybe CONFIG_PREEMPT will merge into
> CONFIG_TOTAL_PREEMPT. Maybe CONFIG_NO_PREEMPT will merge into
> CONFIG_PREEMPT_VOLUNTARY. Maybe CONFIG_PREEMPT_VOLUNTARY will go away
> altogether. We cannot know at this point, it all depends on how usage
> (and consequently, hardware) evolves.
> 

OK I'll leave it at that. We'll see what happens.