Re: [PATCH] Convert powerpc simple spinlocks into ticket locks

[Peter Zijlstra]

On Thu, Feb 06, 2014 at 06:37:27PM +0100, Torsten Duwe wrote:
> On Thu, Feb 06, 2014 at 05:38:37PM +0100, Peter Zijlstra wrote:
> > On Thu, Feb 06, 2014 at 11:37:37AM +0100, Torsten Duwe wrote:
> > > x86 has them, MIPS has them, ARM has them, even ia64 has them:
> > > ticket locks. They reduce memory bus and cache pressure especially
> > > for contended spinlocks, increasing performance.
> > > 
> > > This patch is a port of the x86 spin locks, mostly written in C,
> > > to the powerpc, introducing inline asm where needed. The pSeries
> > > directed yield for vCPUs is taken care of by an additional "holder"
> > > field in the lock.
> > > 
> > 
> > A few questions; what's with the ppc64 holder thing? Not having a 32bit
> > spinlock_t is sad.
> 
> I must admit that I haven't tested the patch on non-pseries ppc64 nor on
> ppc32. Only ppc64 has the ldarx and I tried to atomically replace the 
> holder along with the locks. That might prove unneccessary.

But what is the holder for? Can't we do away with that field?

> > Can you pair lwarx with sthcx ? I couldn't immediately find the answer
> > in the PowerISA doc. If so I think you can do better by being able to
> > atomically load both tickets but only storing the head without affecting
> > the tail.
> 
> V2.06b, Book II, Chapter 3, "sthcx" says:
> | If a reservation exists and the length associated [...] is not 2 bytes,
> | it is undefined whether (RS)_48:63 are stored [...]
> 
> That doesn't make me feel comfortable :(

That's on page 692, right? The way I read that is of the lharx/sthcx
don't have the exact same address, storage is undefined. But I can't
find mention of non-matching load and store size, although I can imagine
it being the same undefined.
--
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/