Re: [PATCH v5 2/6] ksm: support unsharing zero pages placed by KSM

[David Hildenbrand]

On 11.03.23 06:37, xu xin wrote:
> [sorry to reply so late, on vacation too, and my mailing system has some kind of problem]
>
> > [sorry, was on vacation last week]
>
> > > Why use flags if they both conditions are mutually exclusive?
> >
> > Just to make the return value of break_ksm_pmd_entry() more expressive and
> > understandable. because break_ksm_pmd_entry have three types of returned
> > values (0, 1, 2).
>
> > It adds confusion. Just simplify it please.
>
> So I think it's good to add a enum value of 0 listed here as suggested
> by Claudio Imbrenda.

Please keep it simple.

> > > MADV_UNMERGEABLE -> unmerge_ksm_pages() will never unshare the shared
> > > zeropage? I thought the patch description mentions that that is one of
> > > the goals?
> >
> > No, MADV_UNMERGEABLE will trigger KSM to unshare the shared zeropages in the
> > context of "get_next_rmap_item() -> unshare_zero_pages(), but not directly in the
> > context of " madvise()-> unmerge_ksm_pages() ". The reason for this is to avoid
> > increasing long delays of madvise() calling on unsharing zero pages.
>
> > Why do we care and make this case special?
>
> Yeah, the code seems a bit special, but it is a helpless way and best choice, because the
> action of unsharing zero-pages is too complex and CPU consuming because checking whether the
> page we get is actually placed by KSM or not is not a easy thing in the context of
> unmerge_ksm_pages.
>
> In experiment, unsharing zero-pages in the context of unmerge_ksm_pages cause user' madvise()
> spend 5 times the time than the way of the current patch.

Who exactly cares  and why?

> So let's leave it as it is now. I will add a (short) explanation of when and why the new
> unshare_zero_page flag should be used.

I vote to keep it as simple as possible in the initial version.

--
Thanks,

David / dhildenb