[PATCH v2] userfaultfd: ensure mremap_userfaultfd_fail() releases mmap_changing

[Mike Rapoport]

From: "Mike Rapoport (Microsoft)" <rppt@xxxxxxxxxx>

Sashiko says:

  mremap_userfaultfd_prep() increments ctx->mmap_changing to stall
  concurrent operations, but mremap_userfaultfd_fail() does not
  decrement it before dropping the context reference.

If an mremap operation fails, ctx->mmap_changing remains elevated. This
will causes subsequent userfaultfd operations like a UFFDIO_COPY to fail
with -EAGAIN.

Decrement ctx->mmap_changing in mremap_userfaultfd_fail().

Link: https://sashiko.dev/#/patchset/20260430113512.115938-1-rppt@xxxxxxxxxx
Fixes: df2cc96e7701 ("userfaultfd: prevent non-cooperative events vs mcopy_atomic races")
Reviewed-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
Signed-off-by: Mike Rapoport (Microsoft) <rppt@xxxxxxxxxx>
---

I split the fix from the code movement series, will be easier to
everyone :)

v2 changes:
* VM_WARN() if mmap_changing is going negative

v1: https://lore.kernel.org/all/20260501145433.156211-1-rppt@xxxxxxxxxx
(patch 1/3)

 fs/userfaultfd.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 4b53dc4a3266..390e4b7d9cb9 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -786,6 +786,8 @@ void mremap_userfaultfd_fail(struct vm_userfaultfd_ctx *vm_ctx)
 	if (!ctx)
 		return;
 
+	atomic_dec(&ctx->mmap_changing);
+	VM_WARN_ON_ONCE(atomic_read(&ctx->mmap_changing) < 0);
 	userfaultfd_ctx_put(ctx);
 }
 

base-commit: 972c53e0ec3abfc6f5fe2cb503640710fb23cf95
-- 
2.53.0