On Thu, 10 Sep 2020 10:02:33 -0300
Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> On Thu, Sep 10, 2020 at 11:39:25AM +0200, Alexander Gordeev wrote:
>
> > As Gerald mentioned, it is very difficult to explain in a clear way.
> > Hopefully, one could make sense ot of it.
>
> I would say the page table API requires this invariant:
>
> pud = pud_offset(p4d, addr);
> do {
> WARN_ON(pud != pud_offset(p4d, addr);
> next = pud_addr_end(addr, end);
> } while (pud++, addr = next, addr != end);
>
> ie pud++ is supposed to be a shortcut for
> pud_offset(p4d, next)
>
> While S390 does not follow this. Fixing addr_end brings it into
> alignment by preventing pud++ from happening.
>
> The only currently known side effect is that gup_fast crashes, but it
> sure is an unexpected thing.
It only is unexpected in a "top-level folding" world, see my other reply.
Consider it an optimization, which was possible because of how our dynamic
folding works, and e.g. because we can determine the correct pagetable
level from a pXd value in pXd_offset.
> This suggests another fix, which is to say that pud++ is undefined and
> pud_offset() must always be called, but I think that would cause worse
> codegen on all other archs.
There really is nothing to fix for s390 outside of gup_fast, or other
potential future READ_ONCE pagetable walkers. We do take the side-effect
of the generic change on all other pagetable walkers for s390, but it
really is rather a slight degradation than a fix.
