Quoting Chris Wilson (2019-08-21 16:24:22)
> Quoting Christian König (2019-08-21 13:31:45)
> > @@ -117,17 +120,10 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data,
> > busy_check_writer(rcu_dereference(obj->base.resv->fence_excl));
> >
> > /* Translate shared fences to READ set of engines */
> > - list = rcu_dereference(obj->base.resv->fence);
> > - if (list) {
> > - unsigned int shared_count = list->shared_count, i;
> > -
> > - for (i = 0; i < shared_count; ++i) {
> > - struct dma_fence *fence =
> > - rcu_dereference(list->shared[i]);
> > -
> > - args->busy |= busy_check_reader(fence);
> > - }
> > - }
> > + readers = dma_resv_fences_get_rcu(&obj->base.resv->readers);
> > + dma_fence_array_for_each(fence, cursor, readers)
> > + args->busy |= busy_check_reader(fence);
> > + dma_fence_put(readers);
>
> That's underwhelming, the full-mb shows up in scaling tests (I'll test
> the impact of this series later). Something like,
To put some numbers to it, adding the full-mb adds 5ns to a single
thread on Kabylake and 20ns under contention.
-Chris
