[PATCH 1/6] mm/mremap: Account memory on do_munmap() failure

From: Dmitry Safonov
Date: Mon Oct 12 2020 - 22:42:05 EST


move_vma() copies VMA without adding it to account, then unmaps old part
of VMA. On failure it unmaps the new VMA. With hacks accounting in
munmap is disabled as it's a copy of existing VMA.

Account the memory on munmap() failure which was previously copied into
a new VMA.

Fixes: commit e2ea83742133 ("[PATCH] mremap: move_vma fixes and cleanup")
Signed-off-by: Dmitry Safonov <dima@xxxxxxxxxx>
---
mm/mremap.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/mremap.c b/mm/mremap.c
index 138abbae4f75..03d31a0d4c67 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -450,7 +450,8 @@ static unsigned long move_vma(struct vm_area_struct *vma,

if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
/* OOM: unable to split vma, just get accounts right */
- vm_unacct_memory(excess >> PAGE_SHIFT);
+ if (vm_flags & VM_ACCOUNT)
+ vm_acct_memory(new_len >> PAGE_SHIFT);
excess = 0;
}

--
2.28.0