On Sun, May 19, 2019 at 10:37 PM Anshuman Khandual <anshuman.khandual@xxxxxxx> wrote: > > On 05/18/2019 03:20 AM, Andrew Morton wrote: > > On Fri, 17 May 2019 16:08:34 +0530 Anshuman Khandual <anshuman.khandual@xxxxxxx> wrote: > > > >> The presence of struct page does not guarantee linear mapping for the pfn > >> physical range. Device private memory which is non-coherent is excluded > >> from linear mapping during devm_memremap_pages() though they will still > >> have struct page coverage. Just check for device private memory before > >> giving out virtual address for a given pfn. > > > > I was going to give my standard "what are the user-visible runtime > > effects of this change?", but... > > > >> All these helper functions are all pfn_t related but could not figure out > >> another way of determining a private pfn without looking into it's struct > >> page. pfn_t_to_virt() is not getting used any where in mainline kernel.Is > >> it used by out of tree drivers ? Should we then drop it completely ? > > > > Yeah, let's kill it. +1 to killing it, since there has been a paucity of 'unsigned long pfn' code path conversions to 'pfn_t', and it continues to go unused. > > But first, let's fix it so that if someone brings it back, they bring > > back a non-buggy version. Not sure this can be solved without a rethink of who owns the virtual address space corresponding to MEMORY_DEVICE_PRIVATE, and clawing back some of the special-ness of HMM. > > Makes sense. > > > > > So... what (would be) the user-visible runtime effects of this change? > > I am not very well aware about the user interaction with the drivers which > hotplug and manage ZONE_DEVICE memory in general. Hence will not be able to > comment on it's user visible runtime impact. I just figured this out from > code audit while testing ZONE_DEVICE on arm64 platform. But the fix makes > the function bit more expensive as it now involve some additional memory > references. MEMORY_DEVICE_PRIVATE semantics were part of the package of the initial HMM submission that landed in the kernel without an upstream user. While pfn_t_to_virt() also does not have an upstream user it was at least modeled after the existing pfn_to_virt() api to allow for future 'unsigned long pfn' to 'pfn_t' conversions. As for what a fix might look like, it seems to me that we should try to unify 'pfn_t' and 'hmm_pfn's. I don't see why 'hmm_pfn's need to exist as their own concept vs trying consume more flag space out of pfn_t. That would at least allow the pfn_t_has_page() helper to detect the HMM case.
