On Fri, Sep 22, 2017 at 02:44:51PM +0100, Robin Murphy wrote: > Not quite - I mean instead of adding an ops->device_is_coherent callback > (which cannot really have a safe fallback value either way) and trying > to enforce that dma_get_cache_alignment() should be the only valid > caller, just add an ops->get_cache_alignment callback directly. Exactly - and then fall back to ARCH_DMA_MINALIGN/1 if the ops vector is not provided, to keep the existing behavior.