Re: [RFC][PATCH 5/5] powerpc: Remove SYNC from _switch

On Thu, 8 Jun 2017 14:45:40 +0200
Peter Zijlstra [off-list ref] wrote:

On Thu, Jun 08, 2017 at 08:00:15PM +1000, Nicholas Piggin wrote:

quoted

I have those, I just also would like one in the core scheduler's use
of smp_mb__after_spinlock(), because it would be easy for core scheduler
change to miss that quirk. Sure we can say that Peter and scheduler
maintainers know about powerpc oddities, but then why shouldn't it also
go into a comment there?  

So the core scheduler guarantees smp_mb() or equivalent full transitive
ordering happens at schedule() time.

It has for a fairly long time and I don't think we'll ever get rid of
that, its fairly fundamental.

PPC is special in that smp_mb() ends up being the strongest ordering
primitive there is. But note that PPC is not unique, afaict Alpha is in
the same boat. They rely on the MB from the scheduler core.

IA64 OTOH, while they have smp_mb() == mb() still needs SYNC.I in
__switch_to() to serialize against (instruction) cache flushes.

So while I'm all for adding comments explaining what the core provides,
I don't see immediate reasons to call out PPC.

I guess I see your point... okay, will constrain the comment to powerpc
context switch and primitives code. Any fundamental change to such
scheduler barriers I guess would require at least a glance over arch
switch code :)

My plan is to send the powerpc sync removal patch for hopefully 4.13
merge. I'm pretty sure it will be equally happy with your patches.
Unless you can see any problems with it? More eyes would be welcome.

Thanks,
Nick