Re: [RFC PATCH] introduce sys_membarrier(): process-wide memory barrier (v5)

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Peter Zijlstra (peterz@infradead.org) wrote:
> On Thu, 2010-01-21 at 11:07 -0500, Mathieu Desnoyers wrote:
> > 
> > One efficient way to fit the requirement of sys_membarrier() would be to
> > create spin_lock_mb()/spin_unlock_mb(), which would have full memory
> > barriers rather than the acquire/release semantic. These could be used
> > within schedule() execution. On UP, they would turn into preempt off/on
> > and a compiler barrier, just like normal spin locks.
> > 
> > On architectures like x86, the atomic instructions already imply a full
> > memory barrier, so we have a direct mapping and no overhead. On
> > architecture where the spin lock only provides acquire semantic (e.g.
> > powerpc using lwsync and isync), then we would have to create an
> > alternate implementation with "sync". 
> 
> There's also clear_tsk_need_resched() which is an atomic op.

But clear_bit() only acts as a full memory barrier on x86 due to the
lock-prefix side-effect.

Ideally, if we add some kind of synchronization, it would be good to
piggy-back on spin lock/unlock, because these already identify
synchronization points (acquire/release semantic). It also surrounds the
scheduler execution. As we need memory barriers before and after the
data modification, this looks like a sane way to proceed: if data update
is protected by the spinlock, then we are sure that we have the matching
full memory barriers.

> 
> The thing I'm worrying about is not making schedule() more expensive for
> a relatively rare operation like sys_membarrier(), while at the same
> time trying to not make while (1) sys_membarrier() ruin your system.

Yep, I share your concern.

> 
> On x86 there is plenty that implies a full mb before rq->curr = next,
> the thing to figure out is what is generally the cheapest place to force
> one for other architectures.

Yep.

> 
> Not sure where that leaves us, since I'm not too familiar with !x86.
> 

As I proposed above, I think what we have to look for is: where do we
already have some weak memory barriers already required ? And then
upgrade these memory barriers to full memory barriers. The spinlock
approach is one possible solution.

The problem with piggy-backing on clear_flag/set_flag is that these
operations don't semantically imply memory barriers at all. So adding
an additional full mb() around these would be much more costly than
"upgrading" an already-existing barrier.

Thanks,

Mathieu


-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68