Re: [PATCH 2/2] mutex: Apply adaptive spinning on mutex_trylock()

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 2011-03-25 at 07:53 +0100, Tejun Heo wrote:
> Hello, Steven, Linus.
> 
> On Thu, Mar 24, 2011 at 09:38:58PM -0700, Linus Torvalds wrote:
> > On Thu, Mar 24, 2011 at 8:39 PM, Steven Rostedt <rostedt@goodmis.org> wrote:
> > >
> > > But now, mutex_trylock(B) becomes a spinner too, and since the B's owner
> > > is running (spinning on A) it will spin as well waiting for A's owner to
> > > release it. Unfortunately, A's owner is also spinning waiting for B to
> > > release it.
> > >
> > > If both A and B's owners are real time tasks, then boom! deadlock.
> > 
> > Hmm. I think you're right. And it looks pretty fundamental - I don't
> > see any reasonable approach to avoid it.
> 
> Hmmm... I have an idea.  Will play with it a bit and post if it works
> out okay.

One solution is to have this be only done on explicit trylocks. Perhaps
introduce a mutex_trylock_spin()? Then when the developer knows that
this scenario does not exist, they can convert mutex_trylocks() into
this spinning version.

> 
> > I think the RT issue is a red herring too - afaik, you can get a
> > deadlock with two perfectly normal processes too. Of course, for
> > non-RT tasks, any other process will eventually disturb the situation
> > and you'd get kicked out due to need_resched(), but even that might be
> > avoided for a long time if there are other CPU's - leading to tons of
> > wasted CPU time.
> 
> Yeap, need_resched() currently is the only thing which limits the
> duration of spinning when the owner continues to run.

Yeah, I was about to complain about the long latencies that this could
cause, then I realized that RT tasks would in fact deadlock the system
here, which I thought was a bigger problem, and focused on that issue.

-- Steve