On Tue, Aug 06, 2024 at 03:02:00PM +0200, David Hildenbrand wrote:

quoted

Right.

I don't have a reason to change numa_pte_updates semantics yet so far, but
here there's the problem where numa_huge_pte_updates can be ambiguous when
there is even PUD involved.

In general, I don't know how I should treat this counter in PUD path even
if NUMA isn't involved in dax yet; it can be soon involved if we move on
with using this same path for hugetlb, or when 1G thp can be possible (with
Yu Zhao's TAO?).

We shouldn't bother about it in the PUD path at all I think. Especially as
long as NUMA hinting doesn't apply to any of what we would handle on the PUD
path :)

Hmm, I just noticed that hugetlb was never involved.. but then how about a
potential 1G THP?  Do you mean 1G THP will not be accounted in numa
balancing too even in the future?

The motivation I had this patch in this series is I want to be clear on how
I should treat this counter in pud path if it won't go.  And when people
compare the two paths we'll need to be clear why there's such difference if
I ignore it in pud path.

Per my current read on this counter, it might be an overkill to do that at
all, and it might be simpler we drop it now.

quoted

One other thing I can do is I drop this patch, ignore NUMA_HUGE_PTE_UPDATES
in PUD dax processing for now.  It'll work for this series, but it'll still
be a problem later.  I figured maybe we should simply drop it from now.

It probably shouldn't block your other fixes and we should likely discuss
that separately.

I agree that we should look into dropping that PMD counter completely.

No strong opinion here.  If we prefer keeping that as separate topic, I'll
drop this patch.  You're right, it's not yet relevant to the fix.

Thanks,

-- 
Peter Xu