Re: [PATCH v5 2/5] allow mapping page-less memremaped areas into KVA

[Dan Williams <dan.j.williams@intel.com>]

On Thu, Aug 13, 2015 at 10:35 AM, Matthew Wilcox <willy@linux.intel.com> wrote:
> On Wed, Aug 12, 2015 at 11:01:09PM -0400, Dan Williams wrote:
>> +static inline __pfn_t page_to_pfn_t(struct page *page)
>> +{
>> +     __pfn_t pfn = { .val = page_to_pfn(page) << PAGE_SHIFT, };
>> +
>> +     return pfn;
>> +}
>
> static inline __pfn_t page_to_pfn_t(struct page *page)
> {
>         __pfn_t __pfn;
>         unsigned long pfn = page_to_pfn(page);
>         BUG_ON(pfn > (-1UL >> PFN_SHIFT))
>         __pfn.val = pfn << PFN_SHIFT;
>
>         return __pfn;
> }
>
> I have a problem wih PFN_SHIFT being equal to PAGE_SHIFT.  Consider a
> 32-bit kernel; you're asserting that no memory represented by a struct
> page can have a physical address above 4GB.
>
> You only need three bits for flags so far ... how about making PFN_SHIFT
> be 6?  That supports physical addresses up to 2^38 (256GB).  That should
> be enough, but hardware designers have done some strange things in the
> past (I know that HP made PA-RISC hardware that can run 32-bit kernels
> with memory between 64GB and 68GB, and they can't be the only strange
> hardware people out there).

Sounds good, especially given we only use 4-bits today.