Re: [RFC PATCHv2 2/2] Change khugepaged to respect MMF_THP_DISABLE flag

[Alex Thorlton <athorlton@sgi.com>]

On Fri, Jan 17, 2014 at 09:34:44PM +0100, Oleg Nesterov wrote:
> On 01/16, Alex Thorlton wrote:
> >
> >  static inline int khugepaged_test_exit(struct mm_struct *mm)
> >  {
> > -	return atomic_read(&mm->mm_users) == 0;
> > +	return atomic_read(&mm->mm_users) == 0 ||
> > +	       (mm->flags & MMF_THP_DISABLE_MASK);
>                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> 
> test_bit(MMF_THP_DISABLE) ?

Probably should just use the bitop here, good call.

> And I am not sure this and another check in transparent_hugepage_enabled
> is actually right...
> 
> I think that MMF_THP_DISABLE_MASK should not disable thp if this
> vma has VM_HUGEPAGE set, iow perhaps madvise() should work even
> after PR_SET_THP_DISABLE?
> 
> IOW, MMF_THP_DISABLE should act as khugepaged_req_madv().

I hadn't thought of this, but maybe that's a good idea.  That way we can
turn off THP in general for an mm, but the places in code that
*specifically* request THP will still get it.  I don't see why that
would be a problem, as long as we go with the assumption that, if
somebody is explicitly requesting THPs, they probably have a good reason
for doing so.

Thanks for the input!

- Alex