On Wed, May 31, 2017 at 07:51:59PM +0100, Matthew Auld wrote: > err = mutex_lock_interruptible(&obj->mm.lock); > @@ -2533,7 +2543,33 @@ int __i915_gem_object_get_pages(struct > drm_i915_gem_object *obj) > > unlock: > mutex_unlock(&obj->mm.lock); > - return err; > + > + if (err) > + return err; > + > + for_each_sg(obj->mm.pages->sgl, sg, obj->mm.pages->nents, i) > + sg_mask |= sg->length;
This is the worst place to put a loop over sg. Not only is this synchronous, but pages may not have been allocated yet. Using set_pages was at least correct in that regard! The interface I will keep nagging for is for each task to compute sg_mask as they populate the scatterlist, and then they pass sg_mask to __i915_gem_object_set_pages() alongside the pages. Since this didn't fail BAT, that only means we do not have any userptr tests there... -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
