Re: [PATCH] mm/thp: Correctly differentiate between mapped THP and PMD migration entry

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

On Tue, Oct 09, 2018 at 04:04:21PM +0300, Kirill A. Shutemov wrote:
> On Tue, Oct 09, 2018 at 09:28:58AM +0530, Anshuman Khandual wrote:
> > A normal mapped THP page at PMD level should be correctly differentiated
> > from a PMD migration entry while walking the page table. A mapped THP would
> > additionally check positive for pmd_present() along with pmd_trans_huge()
> > as compared to a PMD migration entry. This just adds a new conditional test
> > differentiating the two while walking the page table.
> > 
> > Fixes: 616b8371539a6 ("mm: thp: enable thp migration in generic path")
> > Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
> > ---
> > On X86, pmd_trans_huge() and is_pmd_migration_entry() are always mutually
> > exclusive which makes the current conditional block work for both mapped
> > and migration entries. This is not same with arm64 where pmd_trans_huge()
> > returns positive for both mapped and migration entries. Could some one
> > please explain why pmd_trans_huge() has to return false for migration
> > entries which just install swap bits and its still a PMD ?
> 
> I guess it's just a design choice. Any reason why arm64 cannot do the
> same?

Anshuman, would it work to:

#define pmd_trans_huge(pmd)     (pmd_present(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT))

?

> > Nonetheless pmd_present() seems to be a better check to distinguish
> > between mapped and (non-mapped non-present) migration entries without
> > any ambiguity.
> 
> Can we instead reverse order of check:
> 
> if (pmd_trans_huge(pmde) || is_pmd_migration_entry(pmde)) {
> 	pvmw->ptl = pmd_lock(mm, pvmw->pmd);
> 	if (!pmd_present(*pvmw->pmd)) {
> 		...
> 	} else if (likely(pmd_trans_huge(*pvmw->pmd))) {
> 		...
> 	} else {
> 		...
> 	}
> ...
> 
> This should cover both imeplementations of pmd_trans_huge().

I'd much rather have portable semantics for pmd_trans_huge(), if we can
achieve that efficiently. But that would be fast /and/ correct, so perhaps
I'm being too hopeful :)

Will