Re: [RFC 2/8] mm/hugetlb: remove useless set_page_count()

[Mike Kravetz]

On 10/26/21 11:50 AM, Pasha Tatashin wrote:
> On Tue, Oct 26, 2021 at 2:45 PM Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
>>
>> On 10/26/21 10:38 AM, Pasha Tatashin wrote:
>>> prep_compound_gigantic_page() calls set_page_count(0, p), but it is not
>>> needed because page_ref_freeze(p, 1) already sets refcount to 0.
>>>
>>> Using, set_page_count() is dangerous, because it unconditionally resets
>>> refcount from the current value to unrestrained value, and therefore
>>> should be minimized.
>>>
>>> Signed-off-by: Pasha Tatashin <pasha.tatashin@xxxxxxxxxx>
>>
>> Thanks!
>>
>> My bad for not removing the set_page_count when adding the page_ref_freeze.
>>
>> FYI, there have been additional changes to this routine in Andrew's
>> tree.  Not really sure if we want/need the VM_BUG_ON_PAGE as that would
>> only check if there was a 'bug' in page_ref_freeze.
> 
> I would like to keep it. Part of the idea of this series is to reduce
> reliance on comments such as:
> 
> /* No worries, refcount is A therefore we can do B */
> 
> And instead enforce that via VM_BUG_ON().  It should be able to
> prevent existing and future _refcount related bugs from manifesting as
> memory corruptions.

Ok, but that seems a bit redundant to me. 

There is actually a VM_BUG_ON_PAGE(page_count(p), p) in the 'non-demote'
case in Andrew's tree.  This is in the path where we do not call
page_ref_freeze to zero page ref.  That seems sufficient to me.

Since you did point out the unnecessary set_page_count, I will submit a
code cleanup patch to remove it.  I think that is independent of your
efforts here, and adding VM_BUG_ON can be discussed in the context of
this series.
-- 
Mike Kravetz