Re: [PATCH 1/2] mm/debug_vm_pgtable/basic: Add validation for dirtiness after write protect

[Catalin Marinas <catalin.marinas@arm.com>]

On Fri, Nov 27, 2020 at 09:22:24AM +0100, Christophe Leroy wrote:
> Le 27/11/2020 à 06:06, Anshuman Khandual a écrit :
> > This adds validation tests for dirtiness after write protect conversion for
> > each page table level. This is important for platforms such as arm64 that
> > removes the hardware dirty bit while making it an write protected one. This
> > also fixes pxx_wrprotect() related typos in the documentation file.
> 
> > diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
> > index c05d9dcf7891..a5be11210597 100644
> > --- a/mm/debug_vm_pgtable.c
> > +++ b/mm/debug_vm_pgtable.c
> > @@ -70,6 +70,7 @@ static void __init pte_basic_tests(unsigned long pfn, pgprot_t prot)
> >   	WARN_ON(pte_young(pte_mkold(pte_mkyoung(pte))));
> >   	WARN_ON(pte_dirty(pte_mkclean(pte_mkdirty(pte))));
> >   	WARN_ON(pte_write(pte_wrprotect(pte_mkwrite(pte))));
> > +	WARN_ON(pte_dirty(pte_wrprotect(pte)));
> 
> Wondering what you are testing here exactly.
> 
> Do you expect that if PTE has the dirty bit, it gets cleared by pte_wrprotect() ?
> 
> Powerpc doesn't do that, it only clears the RW bit but the dirty bit remains
> if it is set, until you call pte_mkclean() explicitely.

Arm64 has an unusual way of setting a hardware dirty "bit", it actually
clears the PTE_RDONLY bit. The pte_wrprotect() sets the PTE_RDONLY bit
back and we can lose the dirty information. Will found this and posted
patches to fix the arm64 pte_wprotect() to set a software PTE_DIRTY if
!PTE_RDONLY (we do this for ptep_set_wrprotect() already). My concern
was that we may inadvertently make a fresh/clean pte dirty with such
change, hence the suggestion for the test.

That said, I think we also need a test in the other direction,
pte_wrprotect() should preserve any dirty information:

	WARN_ON(!pte_dirty(pte_wrprotect(pte_mkdirty(pte))));

If pte_mkwrite() makes a pte truly writable and potentially dirty, we
could also add a test as below. However, I think that's valid for arm64,
other architectures with a separate hardware dirty bit would fail this:

	WARN_ON(!pte_dirty(pte_wrprotect(pte_mkwrite(pte))));

-- 
Catalin