On Wed, Nov 30, 2016 at 01:29:39PM +0100, Maarten Lankhorst wrote:
> > +static void stress_work(struct work_struct *work)
> > +{
> > + struct stress *stress = container_of(work, typeof(*stress), work);
> > + const int nlocks = stress->nlocks;
> > + struct ww_mutex *locks = stress->locks;
> > + struct ww_acquire_ctx ctx;
> > + int contended = -1;
> > + int *order;
> > + int n, ret;
> > +
> > + order = get_random_order(nlocks);
> > + if (!order)
> > + return;
> > +
> > + ww_acquire_init(&ctx, &ww_class);
> > +
> > +retry:
> > + ret = 0;
> > + for (n = 0; n < nlocks; n++) {
> > + if (n == contended)
> > + continue;
> > +
> > + ret = ww_mutex_lock(&locks[order[n]], &ctx);
> > + if (ret < 0)
> > + break;
> > + }
> What's wrong with attempting to lock the contended lock here?
> Who knows, this might find some more bugs than the functional tests already
> do.
I was trying to follow the guide, which was lock, backoff by unlocking
everything, slowlock the contended lock, then lock everything else.
I have now a second worker that follows the reordering method as well.
(As well as a test that slowlock after the ABBA deadlock detection
resolves the locking order.)
If you have a sketch of something else to try, I'll add it.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
