Re: [PATCH v4 3/4] mm: Support batched unmap for lazyfree large folios during reclamation

On 2025/6/25 20:25, David Hildenbrand wrote:

On 25.06.25 14:20, Lance Yang wrote:

quoted

On 2025/6/25 20:09, David Hildenbrand wrote:

quoted

quoted

Somehow, I feel we could combine your cleanup code—which handles a 
batch
size of "nr" between 1 and nr_pages—with the
"if (nr_pages == folio_nr_pages(folio)) goto walk_done" check.

Yeah, that's what I was suggesting. It would have to be part of the
cleanup I think.

I'm still wondering if there is a case where

if (nr_pages == folio_nr_pages(folio))
      goto walk_done;

would be wrong when dealing with small folios.

quoted

In practice, this would let us skip almost all unnecessary checks,
except for a few rare corner cases.

For those corner cases where "nr" truly falls between 1 and nr_pages,
we can just leave them as-is—performing the redundant check inside
page_vma_mapped_walk().

I mean, batching mapcount+refcount updates etc. is always a win. If we
end up doing some unnecessary pte_none() checks, that might be
suboptimal but mostly noise in contrast to the other stuff we will
optimize out 🙂

Agreed that if we can easily avoid these pte_none() checks, we should do
that. Optimizing that for "nr_pages == folio_nr_pages(folio)" makes 
sense.

Hmm... I have a question about the reference counting here ...

        if (vma->vm_flags & VM_LOCKED)
            mlock_drain_local();
        folio_put(folio);
        /* We have already batched the entire folio */

Does anyone else still hold a reference to this folio after folio_put()?

The caller of the unmap operation should better hold a reference :)

Ah, you're right. I should have realized that :(

Thanks,
Lance

Also, I am not sure why we don't perform a

folio_put_refs(folio, nr_pages);

... :)