On Mon, Nov 30, 2015 at 02:34:40PM +0000, Chris Wilson wrote:
> +bool intel_breadcrumbs_add_waiter(struct drm_i915_gem_request *request)
> +{
> + struct intel_breadcrumbs *b = &request->i915->breadcrumbs;
> + struct intel_breadcrumbs_engine *be = &b->engine[request->ring->id];
> + bool first = false;
> +
> + spin_lock(&be->lock);
> + if (request->waiters_count++ == 0) {
> + struct rb_node **p, *parent;
> +
> + if (be->first_waiter == NULL)
> + wake_up_process(b->task);
Oh, the irony.
Converting the other side to a lockless READ_ONCE() means that I do know
have to kick after setting be->first_waiter and not rely on the spinlock
serialising the reads.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
