Re: [PATCH] mm/khugepaged: Detecting uffd-wp vma more efficiently

On Wed, Sep 22, 2021 at 7:18 PM Peter Xu [off-list ref] wrote:

On Wed, Sep 22, 2021 at 06:22:45PM -0700, Hugh Dickins wrote:

quoted

No, I think I misunderstood you before: thanks for re-explaining.
(And Axel's !userfaultfd_minor() check before calling do_fault_around()
plays an important part in making sure that it does reach shmem_fault().)

Still thanks for confirming this, Hugh.

Said that, Axel, I didn't mean I'm against doing something similar like
uffd-wp; it's just a heads-up that maybe you won't find a reproducer with real
issues with minor mode.

Even if I think minor mode should be fine with current code, we could still
choose to disable khugepaged from removing the pmd for VM_UFFD_MINOR vmas, just
like what we'll do with VM_UFFD_WP.  At least it can still reduce false
positives.

So far in my local branch I queued the patch which I attached, that's required
for uffd-wp shmem afaict.  If you think minor mode would like that too, I can
post it separately with minor mode added in.

No worries, you can leave the minor fault case to me.

My thinking there was a THP collapse bug was really just based on
speculation, not a real reproducer, so it's very possible my
speculation was wrong. It will take some more thinking and reading to
convince myself one way or the other. :) Thanks to you and Hugh for
all the details.

I'd prefer not to add this fix "just in case", if it isn't a real
problem, as it seems like it may confuse future readers of the code.

I'll send out a patch for it if / when I manage to build a real
reproducer. Or, in the meantime, some of my Google colleagues are
testing this code via their live migration implementation, so if there
is a bug here there's a good chance we'll find it that way too.

Note that it's slightly different from what I pasted in reply to Yang Shi - I
made it slightly more complicated just to make sure there's no race.  I
mentioned the possible race (I think) in the commit log.

Let me know your preference.

Thanks,

--
Peter Xu