Re: [PATCH tip/core/rcu 02/40] rcu: Make arch select smp_mb__after_unlock_lock() strength

[Paul E. McKenney]

On Wed, Apr 19, 2017 at 11:38:22PM +1000, Michael Ellerman wrote:
> "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> writes:
> 
> > On Thu, Apr 13, 2017 at 06:37:57PM +0200, Peter Zijlstra wrote:
> >> On Thu, Apr 13, 2017 at 09:26:51AM -0700, Paul E. McKenney wrote:
> >> 
> >> > ARCH_WEAK_RELEASE_ACQUIRE actually works both ways.
> >> > 
> >> > To see this, imagine some strange alternate universe in which the Power
> >> > hardware guys actually did decide to switch PPC to doing RCsc as you
> >> > suggest.  There would still be a lot of Power hardware out there that
> >> > still does RCpc.  Therefore, powerpc builds that needed to run on old
> >> > Power hardware would select ARCH_WEAK_RELEASE_ACQUIRE, while kernels
> >> > built to run only on the shiny new (but mythical) alternate-universe
> >> > Power hardware would avoid selecting this Kconfig option.
> >> 
> >> Ah, but Power software guys could do it today by replacing an LWSYNC
> >> with a SYNC in say arch_spin_unlock().
> >> 
> >> And yes, I know this isn't a popular suggestion, but it would do the
> >> trick.
> >
> > Indeed, there is a fine line between motivating people to move to new
> > hardware on the one hand and terminally annoying existing users on
> > the other.  ;-)
> >
> >> Its just that since there's one (PPC) we can sort of pressure them with
> >> the pain of being the only ones to hit all the bugs. But the moment more
> >> appear (and I'm afraid it'll be MIPS, with the excuse that PPC already
> >> does this) it will be ever so much harder to get rid of it.
> >> 
> >> Then again, maybe I should just give up and accept the Linux kernel has
> >> RCpc locks..
> >
> > As usual, I must defer to the powerpc maintainers on this one.
> 
> I reworked my locking tests a bit, to run longer, disable ASLR and a few
> other things, and ran them again. They just bang repeatedly on an
> uncontended lock, so nothing fancy at all.
> 
> Switching the release barrier to sync (from lwsync) slows it down by
> about 18%.

Ouch!!!

> So I think that pretty much rules it out, at least on current CPUs.
> 
> I'll try and get some more time to make sure I didn't do something
> stupid in the test, and maybe do a version that includes some
> contention.

Looking forward to seeing what you come up with...

							Thanx, Paul