Re: [PATCH 2/3] mm: thp: Fix the update_mmu_cache() last argument passing in mm/huge_memory.c

[Will Deacon <will.deacon@arm.com>]

Hi Ralf,

On Wed, Sep 19, 2012 at 04:53:46PM +0100, Ralf Baechle wrote:
> On Wed, Sep 19, 2012 at 10:12:28AM +0100, Catalin Marinas wrote:
> 
> > >> 5) void update_mmu_cache(struct vm_area_struct *vma,
> > >>                          unsigned long address, pte_t *ptep)
> > >
> > > Yes please.
> > 
> > Should we just use a generic (void *) for the last argument or force a
> > cast in mm/huge_memory.c?
> > 
> > Ralf's point is that transparent huge page code calls update_mmu_cache
> > with a (pmd_t *) as the last argument. This could make sense for THP
> > as it assumes that huge pages can only be created at the pmd level.
> > But that's unlike mm/hugetlb.c which casts huge page types to pte_t,
> > even though on ARM they are implemented at the pmd level.
> > 
> > On ARM (with VIPT caches) update_mmu_cache() is empty like on x86,
> > though a static inline rather than macro.
> 
> It's even worse - mm/huge_memory.c is passing a pmd_t, not a pointer so
> changing the type of update_mmu_cache's 3rd argument alone won't cut it.
> 
> This went unnoticed so far because all existing architectures supporting
> transparent huge pages implement update_mmu_cache() as do { } while (0).
> 
> MIPS uses update_mmu_cache() as the hook to deal with cache aliases and
> pre-faulting a TLB entry.  But aliases don't affect small pages and having
> a separate variant of update_mmu_cache for huge pages will allow some other
> optimizations.  That's minor but it's the argument types that really need
> to be fixed and because MIPS also implements huge pages at PMD level I'd
> be happy if we settle on pte_t * for mm/huge_memory.c.

Just to clarify: do you want a cast from pmd_t * to pte_t * or instead copy
the hugetlb code and pass ptes around instead of pmds? The latter is pretty
invasive...

Will