Huge page(contiguous bit) slow down

[will.deacon@arm.com (Will Deacon)]

On Tue, Sep 18, 2018 at 05:09:28PM +0100, Catalin Marinas wrote:
> On Tue, Sep 18, 2018 at 04:16:26PM +0100, Will Deacon wrote:
> > On Tue, Sep 18, 2018 at 03:58:32PM +0100, Catalin Marinas wrote:
> > > On Tue, Sep 18, 2018 at 12:33:01PM +0100, Will Deacon wrote:
> > > > On Tue, Sep 18, 2018 at 03:02:17AM +0000, Zhang, Lei wrote:
> > > > > --- a/arch/arm64/mm/hugetlbpage.c
> > > > > +++ b/arch/arm64/mm/hugetlbpage.c
> > > > > @@ -332,6 +332,9 @@ int huge_ptep_set_access_flags(struct vm_area_struct *vma,
> > > > >         if (!pte_cont(pte))
> > > > >                 return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
> > > > > 
> > > > > +       if(pte_same(pte, READ_ONCE(*ptep)))
> > > > > +               return 0;
> > > > > +
> > > > 
> > > > This broadly seems to follow the non-contiguous code, but I wonder if we
> > > > can then drop the subsequent pte_same() check on this path and always return
> > > > 1 when we actually update the entries?
> > > 
> > > I don't remember why we went for first clearing and then checking
> > > pte_same() (maybe Steve knows) but I think we can leave pte_same()
> > > outside the get_clear_flush()/set_pte_at() block. This code is executed
> > > with the mmap_sem taken, so there shouldn't be any race on the
> > > individual ptes.
> > 
> > I suspect it's just to avoid the additional load of the page-table entry,
> > since we still have to use get_clear_flush() even with this change.
> > 
> > One thing I don't really grok is the interaction between the contiguous
> > hint and HW_AFDBM. Is it possible for us to be e.g. halfway through the
> > set_pte_at() loop and then for the hardware to perform atomic PTE updates
> > for entries later in the loop? If so, we've got a race and need to use
> > cmpxchg() like we do for the non-contiguous code.
> 
> With the current code, no, since get_clear_flush() sets all of them to
> 0, so no hardware updates before set_pte_at().

The case I'm concerned about is when we've set_pte_at() half of the mapping,
though. At this point, a CPU can get a translation via one of the entries
that we've put down, and it's not clear to me whether this could establish
a contiguous TLB entry which could then result in access/dirty updates to
PTEs that we haven't yet written out.

Will