Dave Hansen <d...@sr71.net> writes:
> On 07/01/2016 07:25 AM, Eric W. Biederman wrote:
>> Linus Torvalds <torva...@linux-foundation.org> writes:
>>> > On Thu, Jun 30, 2016 at 9:39 PM, Dave Hansen <d...@sr71.net> wrote:
>>>> >>
>>>> >> I think what you suggest will work if we don't consider A/D in
>>>> >> pte_none(). I think there are a bunch of code path where assume that
>>>> >> !pte_present() && !pte_none() means swap.
>>> >
>>> > Yeah, we would need to change pte_none() to mask off D/A, but I think
>>> > that might be the only real change needed (other than making sure that
>>> > we don't use the bits in the swap entries, I didn't look at that part
>>> > at all)
>> It looks like __pte_to_swp_entry also needs to be changed to mask out
>> those bits when the swap code reads pte entries. For all of the same
>> reasons as pte_none.
>
> I guess that would be nice, but isn't it redundant?
>
> static inline swp_entry_t pte_to_swp_entry(pte_t pte)
> {
> ...
> arch_entry = __pte_to_swp_entry(pte);
> return swp_entry(__swp_type(arch_entry), __swp_offset(arch_entry));
> }
>
> As long as __swp_type() and __swp_offset() don't let A/D through, then
> we should be OK. This site is the only call to __pte_to_swp_entry()
> that I can find in the entire codebase.
>
> Or am I missing something?
Given that __pte_to_swp_entry on x86_64 is just __pte_val or pte.pte it
does no filtering. Similarly __swp_type(arch_entry) is a >> and
swp_entry(type, ...) is a << of what appears to be same amount
for the swap type.
So any corruption in the upper bits of the pte will be preserved as a
swap type.
In fact I strongly suspect that the compiler can optimize out all of the
work done by "swp_entry(__swp_type(arch_entry), _swp_offset(arch_entry))".
Eric
