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

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Feb 07, 2014 at 06:08:45PM +0100, Torsten Duwe wrote:
> > static inline unsigned int xadd(unsigned int *v, unsigned int i)
> > {
> > 	int t, ret;
> > 	
> > 	__asm__ __volatile__ (
> > "1:	lwarx	%0, 0, %4\n"
> > "	mr	%1, %0\n"
> > "	add	%0, %3, %0\n"
> > "	stwcx.	%0, %0, %4\n"
> > "	bne-	1b\n"
> > 	: "=&r" (t), "=&r" (ret), "+m" (*v)
> > 	: "r" (i), "r" (v)
> > 	: "cc");
> > 
> > 	return ret;
> > }
> > 
> I don't like this xadd thing -- it's so x86 ;)
> x86 has its LOCK prefix, ppc has ll/sc.
> That should be reflected somehow IMHO.

Its the operational semantics I care about; this version is actually
nicer in that it doesn't actually imply all sorts of barriers :-)

> Maybe if xadd became mandatory for some kernel library.

call it fetch_add() its not an uncommon operation and many people
understand the semantics.

But you can simply include the asm bits in ticket_lock() and be done
with it. In that case you can also replace the add with an addi which
might be a little more efficient.

> > void ticket_unlock(tickets_t *lock)
> > {
> > 	ticket_t tail = lock->tail + 1;
> > 
> > 	/*
> > 	 * The store is save against the xadd for it will make the ll/sc fail
> > 	 * and try again. Aside from that PowerISA guarantees single-copy
> > 	 * atomicy for half-word writes.
> > 	 *
> > 	 * And since only the lock owner will ever write the tail, we're good.
> > 	 */
> > 	smp_store_release(&lock->tail, tail);
> > }
> 
> Yeah, let's try that on top of v2 (just posted).
> First, I want to see v2 work as nicely as v1 --
> compiling a debug kernel takes a while...

Use a faster machine... it can be done < 1 minute :-)