On Fri, 18 May 2018, Christoph Hellwig wrote: > > This implementation of arch_setup_pdev_archdata() differs from the > > powerpc one, in that this one avoids clobbering a device dma mask > > which has already been initialized. > > I think your implementation should move into the generic implementation > in drivers/base/platform.c instead of being stuck in m68k. > > Also powerpc probably wants fixing, but that's something left to the > ppc folks..
On powerpc, all platform devices get a dma mask. But they don't do that in drivers/base/platform.c, they use arch_setup_pdev_archdata(). Why didn't they take the approach you suggest? How would I support the claim that replacing an empty platform device dma mask with 0xffffffff is safe on all architectures and platforms? Is there no code conditional upon dev.coherent_dma_mask or dev.dma_mask that could misbehave? (Didn't I cite an example in the other thread?*) If you can convince me that it is safe, I'd be happy to submit the patch you asked for. For now, I still think that patching the platform driver was the correct patch*. Maybe the real problem is your commit 205e1b7f51e4 ("dma-mapping: warn when there is no coherent_dma_mask"), because it assumes that all dma_ops implementations care about coherent_dma_mask. The dma_map_ops implementations that do use coherent_dma_mask should simply fail when it is unset, right? Would it not be better to revert your patch and fix the dma_map_ops failure paths, than to somehow audit all the platform drivers and patch drivers/base/platform.c? Thanks. * https://lkml.kernel.org/r/alpine.LNX.2.21.1805091804290.72%40nippy.intranet --