On Fri, Dec 17, 2021 at 6:42 PM Linus Torvalds
[off-list ref] wrote:

IOW, I think that this stupid (AND UNTESTED) patch should likely just
fix David's test-case with the hugepage and splice thing..

Looking at that patch, the page lock is entirely pointless.

It existed because of that broken reuse_swap_page() that tried to
count page mappings etc, but once you get rid of that - like we got
rid of it for the regular pages - it's not even needed.

So as we hold the page table lock, and see a page_count() of 1, we
could be done without any page lock at all. So that whole
trylock/unlock is actually unnecessary.

That said, it's possibly woth re-using any swap cache pages at this
point, and that would want the page lock. So some complexity in this
area is likely worth it. Similar to how we did it in commit
f4c4a3f48480 ("mm: free idle swap cache page after COW") for regular
pages.

So that patch is not great, but I think it works as a guiding one.

And notice how *simple* it is. It doesn't require careful counting of
swap entries that depend on page locking.

                Linus