On Thu, May 12, 2022 at 10:07:09PM +0100, Matthew Wilcox wrote:
> On Tue, May 10, 2022 at 12:22:42PM +0000, cgel.zte@xxxxxxxxx wrote:
> > +++ b/Documentation/admin-guide/mm/ksm.rst
> > @@ -32,7 +32,7 @@ are swapped back in: ksmd must rediscover their identity and merge again).
> > Controlling KSM with madvise
> > ============================
> >
> > -KSM only operates on those areas of address space which an application
> > +KSM can operates on those areas of address space which an application
>
> "can operate on"
>
Thanks.
> > + * Force anonymous pages of this mm to be involved in KSM merging
> > + * without explicitly calling madvise.
> > + */
> > + if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
> > + err = __ksm_enter(mm);
> > + if (!err)
> > + mm->ksm_force = force;
> > + }
> > +
> > + mmap_write_unlock(mm);
> > + }
>
> There's a much simpler patch hiding inside this complicated one.
>
> if (force) {
> set_bit(MMF_VM_MERGEABLE, &mm->flags));
> for each VMA
> set VM_MERGEABLE;
> err = __ksm_enter(mm);
> } else {
> clear_bit(MMF_VM_MERGEABLE, &mm->flags));
> for each VMA
> clear VM_MERGEABLE;
> }
>
> ... and all the extra complications you added go away.
Sorry, but I don't think that is a better way of implementation, although it
is simpler. It overrides the intention of code ifself which is
unrecoverable. For example, if a program which madvise just a part of VMAs
(not all) as MERGEABLE then its ksm_force is turned on, and subsequently
it ksm_force is turned off again, the "madvised MERGEBLE" cannot be
recoverd.
===========================================================================
I have a idea: can we refer to the interface of THP? Substitute ksm_force
with ksm_enabled and give three values to it:
1) always: force all anonymous VMAs of this process to be scanned.
2) madvise: the default state, unless user code call madvise, don't
scan this process.
3) never: never be involed in KSM.
===========================================================================
How about this?
