Re: [PATCH PTI v3 03/10] x86/pti/64: Fix ESPFIX64 user mapping

[Andy Lutomirski]

On Thu, Dec 14, 2017 at 6:10 AM, Kirill A. Shutemov
<kirill@xxxxxxxxxxxxx> wrote:
> On Wed, Dec 13, 2017 at 09:01:50AM -0800, Andy Lutomirski wrote:
>> On Wed, Dec 13, 2017 at 5:12 AM, Kirill A. Shutemov
>> <kirill@xxxxxxxxxxxxx> wrote:
>> > On Tue, Dec 12, 2017 at 07:56:38AM -0800, Andy Lutomirski wrote:
>> >> The ESPFIX64 user mapping belongs in pti.c just like all the other
>> >> user mappings.  Move it there and make it work correctly while we're
>> >> at it.
>> >>
>> >> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
>> >
>> > Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
>> >
>> > BTW, why do we open-code p?d_alloc() in pti_user_pagetable_walk_*()?
>> > It seems unnecessary and potentially bogus: see smp_wmb() in __p?d_alloc()
>> > helpers.
>>
>> The helpers won't work -- we're allocating kernel-owned tables in the
>> usermode part of init_mm.  The p?d_alloc() helpers allocate
>> user-accounted tables in the kernelmode part of the mm.
>
> What's wrong to account them against init_mm?
>

Nothing.  But the other parts are showstoppers: the helpers won't set
_PAGE_USER in the higher level entries and they will put the pgds in
the wrong place.