Re: [patch] sched: unlocked context-switches

[Nick Piggin]

Ingo Molnar wrote:
> * Nick Piggin <nickpiggin@xxxxxxxxxxxx> wrote:

> > I did propose doing unconditionally unlocked switches a while back 
> > when my patch first popped up - you were against it then, but I guess 
> > you've had second thoughts?
>
>
> the reordering of switch_to() and the switch_mm()-related logic was that 
> made it really worthwile and clean. I.e. we pick a task atomically, we 
> switch stacks, and then we switch the MM. Note that this setup still 
> leaves the possibility open to move the stack-switching back under the 
> irq-disabled section in a natural way.

Yeah true. I didn't come up with code for you to look at at
that point anyway so you were obviously just speculating!

> > It does add an extra couple of stores to on_cpu, and a wmb() for 
> > architectures that didn't previously need the unlocked switches. And 
> > ia64 needs the extra interrupt disable / enable. Probably worth it?
>
>
> it also removes extra stores to rq->prev_mm and other stores. I havent 
> measured any degradation on x86.

Yeah true, although that is just a single cacheline (which will be
hot for any context switch heavy workload).

On the other hand, I tried put oncpu near other fields that are
accessed during context switch, so maybe its not an issue.

> If the irq disable/enable becomes widespread i'll do another patch to 
> push the irq-enabling into switch_to() so the arch can do the 
> stack-switch first and then enable interrupts and do the rest - but i 
> didnt want to complicate things unnecessarily for now.
>
>
> > Minor style request: I like that you're accessing ->on_cpu through 
> > functions so the !SMP case doesn't clutter the code with ifdefs... but 
> > can you do set_task_on_cpu(p) and clear_task_on_cpu(p) ?
>
>
> yeah, i thought about these two variants and went for set_task_on_cpu() 
> so that it's less encapsulated (it's really just a conditional 
> assignment) and that it parallels set_task_cpu() use. But no strong 
> feelings either way. Anyway, lets try what we have now, i'll do the rest 
> in deltas.

Sounds good.

--
SUSE Labs, Novell Inc.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/