Re: [RFC PATCH 1/1] remove redundant compare, cmpxchg already does it

[Davidlohr Bueso]

On Thu, 2014-06-05 at 10:54 -0700, Davidlohr Bueso wrote:
> On Thu, 2014-06-05 at 09:22 +0200, Peter Zijlstra wrote:
> > On Wed, Jun 04, 2014 at 04:56:50PM -0400, Andev wrote:
> > > On Wed, Jun 4, 2014 at 4:38 PM, Pranith Kumar <pranith@xxxxxxxxxx> wrote:
> > > > remove a redundant comparision
> > > >
> > > > Signed-off-by: Pranith Kumar <bobby.prani@xxxxxxxxx>
> > > > ---
> > > >  kernel/locking/rwsem-xadd.c | 3 +--
> > > >  1 file changed, 1 insertion(+), 2 deletions(-)
> > > >
> > > > diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
> > > > index 1f99664b..6f8bd3c 100644
> > > > --- a/kernel/locking/rwsem-xadd.c
> > > > +++ b/kernel/locking/rwsem-xadd.c
> > > > @@ -249,8 +249,7 @@ static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
> > > >  {
> > > >      if (!(count & RWSEM_ACTIVE_MASK)) {
> > > >          /* try acquiring the write lock */
> > > > -        if (sem->count == RWSEM_WAITING_BIAS &&
> > > > -            cmpxchg(&sem->count, RWSEM_WAITING_BIAS,
> > > > +        if (cmpxchg(&sem->count, RWSEM_WAITING_BIAS,
> > > >                  RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_WAITING_BIAS) {
> > > 
> > > This was mainly done to avoid the cost of a cmpxchg in case where they
> > > are not equal. Not sure if it really makes a difference though.
> > 
> > It does, a cache hot cmpxchg instruction is 24 cycles (as is pretty much
> > any other LOCKed ins, as measured on my WSM-EP), not to mention that
> > cmpxchg is a RMW so it needs to grab the cacheline in exclusive mode.
> > 
> > A read, which allows the cacheline to remain in shared, and non LOCKed
> > ops are way faster.
> 
> Yep, and we also do it in mutexes. The numbers and benefits on larger
> systems speaks for themselves. It would, perhaps, be worth adding a
> comment as it does seem redundant if you're not thinking about the
> cacheline when reading the code.

I knew I had formally read this technique somewhere:
http://pdos.csail.mit.edu/6.828/2010/readings/mcs.pdf (part 2.1).

Peter, what do you think of adding a new cmp_cmpxchg() or dcmpxchg()
call for such scenarios?


--
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/