On Mon, Apr 20, 2015 at 03:56:09PM +0100, Tvrtko Ursulin wrote: > >Will take a bit of judicious planning (like doing a self-copy on the dst > >first to make sure it bound ahead of the leak and not reallocating the > >batch handle) > > Hm.. not bad! > > >Or alternatively, we know the name of the flinked object, so we can do > >an explicit search! Ok, that seems more robust, but you do some sync > >point in there (either gem_sync() every child before exit or > >gem_quiescent_gpu()). > > Anything to avoid debugfs parsing I think. Well not anything, if > exec[].offset is not retrievable from intel_batchbuffer helpers I'll > reconsider it. Definitely need to at least count ppgtt sections > since the IGT already failed (to fail) on one older kernel with the > poor character counting approach. With the libdrm interface, drm_intel_bo->offset will be kept up to date with the post-execbuffer value. I'd suggest just using the rawer ioctl wrappers for fine control, especially when allocating objects and offsets. -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
