On Thu, 20 Jun 2024, David Hildenbrand wrote:
(I do have doubts about Barry's: the "_new" in folio_add_new_anon_rmap()
was all about optimizing a known-exclusive case, so it surprises me
to see it being extended to non-exclusive; and I worry over how its
atomic_set(&page->_mapcount, 0)s can be safe when non-exclusive (but
I've never caught up with David's exclusive changes, I'm out of date).
We discussed that a while ago: if we wouldn't be holding the folio lock in the
"folio == swapcache" at that point (which we do for both do_swap_page() and
unuse_pte()) things would already be pretty broken.
You're thinking of the non-atomic-ness of atomic_set(): I agree that
the folio lock makes that safe (against other adds; but not against
concurrent removes, which could never occur with the old "_new" usage).
But what I'm worried about is the 0 in atomic_set(&page->_mapcount, 0):
once the folio lock has been dropped, another non-exclusive add could
come in and set _mapcount again to 0 instead of to 1 (mapcount 1 when
it should be 2)?
That's I added a while ago:
if (unlikely(!folio_test_anon(folio))) {
VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
/*
* For a PTE-mapped large folio, we only know that the single
* PTE is exclusive. Further, __folio_set_anon() might not get
* folio->index right when not given the address of the head
* page.
*/
...
We should probably move that VM_WARN_ON_FOLIO() to folio_add_new_anon_rmap()
and document that it's required in the non-exclusive case.
But even if those are wrong, I'd expect them to tend towards a mapped
page becoming unreclaimable, then "Bad page map" when munmapped,
not to any of the double-free symptoms I've actually seen.)
What's the first known-good commit?
I cannot answer that with any certainty: we're on the shifting sands
of next and mm-unstable, and the bug itself is looking rather like
something which gets amplified or masked by other changes - witness
my confident arrival at Barry's series as introducing the badness,
only for a longer run then to contradict that conclusion.
There was no sign of a problem in a 20-hour run of the same load on
rc3-based next-2024-06-13 (plus my posted fixes); there has been no
sign of this problem on 6.10-rc1, rc2, rc3 (but I've not tried rc4
itself, mm.git needing the more urgent attention). mm-everything-
2024-06-15 (minus Chris's mTHP swap) did not show this problem, but
did show filemap_get_folio() hang. mm-everything-2024-06-18 (minus
Baolin's mTHP shmem swap) is where I'm hunting it.