On Fri, Jun 16, 2017 at 03:36:00PM +0200, Andrea Arcangeli wrote:
> Hello Krill,
>
> On Thu, Jun 15, 2017 at 05:52:22PM +0300, Kirill A. Shutemov wrote:
> > +static inline pmd_t pmdp_establish(pmd_t *pmdp, pmd_t pmd)
> > +{
> > + pmd_t old;
> > +
> > + /*
> > + * We cannot assume what is value of pmd here, so there's no easy way
> > + * to set if half by half. We have to fall back to cmpxchg64.
> > + */
> > + {
> > + old = *pmdp;
> > + } while (cmpxchg64(&pmdp->pmd, old.pmd, pmd.pmd) != old.pmd);
> > +
> > + return old;
> > +}
>
> I see further margin for optimization here (although it's only for PAE
> x32..).
>
> pmd is stable so we could do:
>
> if (!(pmd & _PAGE_PRESENT)) {
> cast to split_pmd and use xchg on pmd_low like
> native_pmdp_get_and_clear and copy pmd_high non atomically
> } else {
> the above cmpxchg64 loop
> }
>
> Now thinking about the above I had a second thought if pmdp_establish
> is the right interface and if we shouldn't replace pmdp_establish with
> pmdp_mknotpresent instead to skip the pmd & _PAGE_PRESENT check that
> will always be true in practice, so pmdp_mknotpresent will call
> internally pmd_mknotpresent and it won't have to check for pmd &
> _PAGE_PRESENT and it would have no cons on x86-64.
With your proposed optimization, compiler is in good position to eliminate
cmpxchg loop for trivial cases as we have in pmdp_invalidate() case.
It can see that pmd is always has the present bit cleared.
I'll keep more flexible interface for now. Will see if anybody would see
more problems with it.
--
Kirill A. Shutemov
