On Tue, Oct 03, 2017 at 11:30:50AM -0400, Nicolas Pitre wrote: > Unless you have a better scheme altogether to suggest of course, given > the existing constraints. I still can't understand why this convoluted fault path that finds vma, attempts with all kinds of races and then tries to update things like vm_ops is even nessecary. We have direct mappings of physical address perfectly working in the DAX code (even with write support!) or in drivers using remap_pfn_range so a really good explanation why neither scheme can be used is needed first.
