Re: [PATCH v4 1/9] introduce __pfn_t for scatterlists and pmem

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, Jun 5, 2015 at 2:19 PM, Dan Williams <dan.j.williams@intel.com> wrote:
> +enum {
> +#if BITS_PER_LONG == 64
> +       PFN_SHIFT = 3,
> +       /* device-pfn not covered by memmap */
> +       PFN_DEV = (1UL << 2),
> +#else
> +       PFN_SHIFT = 2,
> +#endif
> +       PFN_MASK = (1UL << PFN_SHIFT) - 1,
> +       PFN_SG_CHAIN = (1UL << 0),
> +       PFN_SG_LAST = (1UL << 1),
> +};

Ugh. Just make PFN_SHIFT unconditional. Make it 2, unconditionally.
Or, if you want to have more bits, make it three unconditionally, and
make 'struct page' just be at least 8-byte aligned even on 32-bit.

Even on 32-bit architectures, there's plenty of bits. There's no
reason to "pack" this optimally. Remember: it's a page frame number,
so there's the page size shifting going on in physical memory, and
even if you shift the PFN by 3 - or four  of five - bits
unconditionally (rather than try to shift it by some minimal number),
you're covering a *lot* of physical memory.

Say you're a 32-bit architecture with a 4k page size, and you lose
three bits to "type" bits. You still have 32+12-3=41 bits of physical
address space. Which is way more than realistic for a 32-bit
architecture anyway, even with PAE (or PXE or whatever ARM calls it).
Not that I see persistent memory being all that relevant on 32-bit
hardware anyway.

So I think if you actually do want that third bit, you're better off
just marking "struct page" as being __aligned__((8)) and getting the
three bits unconditionally. Just make the rule be that mem_map[] has
to be 8-byte aligned.

Even 16-byte alignment would probably be fine. No?

                Linus