Let's document why hugetlb still uses folio_mapcount() and is prone to leaking memory between processes, for example using vmsplice() that still uses FOLL_GET. More details can be found in [1], especially around how hugetlb pages cannot really be overcommitted, and why we don't particularly care about these vmsplice() leaks for hugetlb -- in contrast to ordinary memory. [1] https://lore.kernel.org/all/8b42a24d-caf0-46ef-9e15-0f88d47d2f21@xxxxxxxxxx/ Suggested-by: Peter Xu <peterx@xxxxxxxxxx> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx> --- mm/hugetlb.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/mm/hugetlb.c b/mm/hugetlb.c index 417fc5cdb6eeb..a7efb350f5d07 100644 --- a/mm/hugetlb.c +++ b/mm/hugetlb.c @@ -5963,6 +5963,13 @@ static vm_fault_t hugetlb_wp(struct folio *pagecache_folio, /* * If no-one else is actually using this page, we're the exclusive * owner and can reuse this page. + * + * Note that we don't rely on the (safer) folio refcount here, because + * copying the hugetlb folio when there are unexpected (temporary) + * folio references could harm simple fork()+exit() users when + * we run out of free hugetlb folios: we would have to kill processes + * in scenarios that used to work. As a side effect, there can still + * be leaks between processes, for example, with FOLL_GET users. */ if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { if (!PageAnonExclusive(&old_folio->page)) { -- 2.44.0
