On Tue, 29 Nov 2011 18:32:18 +0100, Daniel Vetter <daniel.vetter at ffwll.ch> wrote: > Better be paranoid. The wmb should flush the wc writes, and > the chipset_flush hopefully flushes any mch buffers. There've been a > few overlay hangs I've never really diagnosed, unfortunately all the > reporters disappeared. One of the worries I've had in the past has been whether we need a wmb() inside intel_ring_begin() so that all the register writes (assuming we get WC registers one day) and the WC gtt writes are coherent with initiating work on the GPU. I attacked i915_gem_object_flush_gtt_write_domain() instead in the belief that would be sufficient. I think this demonstrates I was wrong. However, you could argue that the overlay code is circumventing the cache domain tracking of gem objects established for this very purpose. Though calling i915_gem_object_set_to_gtt_domain(obj,true) in map and i915_gem_object_flush_gtt_write_domain() in unmap might be overkill... -Chris -- Chris Wilson, Intel Open Source Technology Centre
