On Fri, Sep 14, 2018 at 6:14 AM, Christoph Hellwig <hch@xxxxxx> wrote: > On Wed, Sep 12, 2018 at 07:22:11PM -0700, Dan Williams wrote: >> Given the fact that devm_memremap_pages() requires a percpu_ref that is >> torn down by devm_memremap_pages_release() the current support for >> mapping RAM is broken. > > I agree. Do you remember why we even added it in the first place? It was initially a copy over from memremap() that catches these attempts and returns a direct-map pointer.