On Thu, Oct 10, 2019 at 01:05:56PM -0700, Vineet Gupta wrote:
>
> Hi Kirill,
>
> On 10/10/19 1:56 AM, Kirill A. Shutemov wrote:
> > On Wed, Oct 09, 2019 at 10:26:55PM +0000, Vineet Gupta wrote:
> >>
> >> This series elides extraneous generate code for folded p4d/pud.
> >> This came up when trying to remove __ARCH_USE_5LEVEL_HACK from ARC port.
> >> The code saving are not a while lot, but still worthwhile IMHO.
> >
> > Agreed.
>
> Thx.
>
> So given we are folding pmd too, it seemed we could do the following as well.
>
> +#ifndef __PAGETABLE_PMD_FOLDED
> void pmd_clear_bad(pmd_t *);
> +#else
> +#define pmd_clear_bad(pmd) do { } while (0)
> +#endif
>
> +#ifndef __PAGETABLE_PMD_FOLDED
> void pmd_clear_bad(pmd_t *pmd)
> {
> pmd_ERROR(*pmd);
> pmd_clear(pmd);
> }
> +#endif
>
> I stared at generated code and it seems a bit wrong.
> free_pgd_range() -> pgd_none_or_clear_bad() is no longer checking for
> unmapped pgd
> entries as pgd_none/pgd_bad are all stubs returning 0.
>
> This whole pmd folding is a bit confusing considering I only revisit it every
> few
> years :-) Abstraction wise, __PAGETABLE_PMD_FOLDED only has pgd, pte but even
> in
> this regime bunch of pmd macros are still valid
>
> pmd_set(pmdp, ptep) {
> *pmdp.pud.p4d.pgd = (unsigned long)ptep
> }
>
> Is there a better way to make a mental model of this code folding.
I don't have any. PMD folding predates me and have never looked at it
closely. Quick look brings more confusion than clarity. :P
> In an ideal world pmd folded would have meant pmd_* routines just vanish -
> poof.
> So in that sense I like your implementation under #[45]LEVEL_HACK where the
> level
> simply vanishes by code like #define p4d_t pgd_t. Perhaps there is lot of
> historic
> baggage, proliferated into arch code so hard to untangle.
In ideal world all these pgd/p4d/pud/pmd/pte should die and we have
something more flexible to begin with.
I played with this before:
https://lore.kernel.org/lkml/20180424154355.mfjgkf47kdp2b...@black.fi.intel.com/
--
Kirill A. Shutemov
