Re: [PATCH v2] introduce atomic_pointer to fix a race condition in cancelable mcs spinlocks

From: James Bottomley
Date: Mon Jun 02 2014 - 16:25:23 EST

On Mon, 2014-06-02 at 22:05 +0200, Peter Zijlstra wrote:
> On Mon, Jun 02, 2014 at 01:33:34PM -0400, Waiman Long wrote:
> > On 06/02/2014 12:30 PM, Peter Zijlstra wrote:
> > >On Mon, Jun 02, 2014 at 06:25:25PM +0200, Peter Zijlstra wrote:
> > >>I'm almost inclined to just exclude parisc from using opt spinning.
> > >>
> > >>That said, this patch still doesn't address the far more interesting
> > >>problem of actually finding these issues for these few weird archs.
> > >So why do these archs provide xchg() and cmpxchg() at all? Wouldn't it
> > >be much simpler if archs that cannot sanely do this, not provide these
> > >primitives at all?
> >
> > I believe xchg() and cmpxchg() are used in quite a number of places within
> > the generic kernel code. So kernel compilation will fail if those APIs
> > aren't provided by an architecture.
> Yep.. so this is going to be painful for a while. But given their
> (parisc, sparc32, metag-lock1) constraints, who knows how many of those
> uses are actually broken.
> So the question is, do you prefer subtly broken code or hard compile
> fails? Me, I go for the compile fail.

The failure is only when a variable that will have an atomic exchange
done on it is updated by a simple operation. To do this properly, we'd
probably need an update macro we could supply the locking to, and a way
of marking the variable to get the compiler to cause a build error if it
was ever updated improperly, but that's starting to look very similar to
Mikulas' proposal.


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at