From 34cde75c1b3bf345b7af668db3f8c1ce4f89e0fa Mon Sep 17 00:00:00 2001 From: Yunseong Kim Date: Mon, 25 Nov 2024 16:45:55 +0900 Subject: [PATCH] ksmbd: fix use-after-free in SMB request handling BugLink: https://bugs.launchpad.net/bugs/2089884 A race condition exists between SMB request handling in `ksmbd_conn_handler_loop()` and the freeing of `ksmbd_conn` in the workqueue handler `handle_ksmbd_work()`. This leads to a UAF. - KASAN: slab-use-after-free Read in handle_ksmbd_work - KASAN: slab-use-after-free in rtlock_slowlock_locked This race condition arises as follows: - `ksmbd_conn_handler_loop()` waits for `conn->r_count` to reach zero: `wait_event(conn->r_count_q, atomic_read(&conn->r_count) == 0);` - Meanwhile, `handle_ksmbd_work()` decrements `conn->r_count` using `atomic_dec_return(&conn->r_count)`, and if it reaches zero, calls `ksmbd_conn_free()`, which frees `conn`. - However, after `handle_ksmbd_work()` decrements `conn->r_count`, it may still access `conn->r_count_q` in the following line: `waitqueue_active(&conn->r_count_q)` or `wake_up(&conn->r_count_q)` This results in a UAF, as `conn` has already been freed. The discovery of this UAF can be referenced in the following PR for syzkaller's support for SMB requests. Link: https://github.com/google/syzkaller/pull/5524 Fixes: ee426bfb9d09 ("ksmbd: add refcnt to ksmbd_conn struct") Cc: linux-cifs@vger.kernel.org Cc: stable@vger.kernel.org # v6.6.55+, v6.10.14+, v6.11.3+ Cc: syzkaller@googlegroups.com Signed-off-by: Yunseong Kim Acked-by: Namjae Jeon Signed-off-by: Steve French (cherry picked from commit 9a8c5d89d327ff58e9b2517f8a6afb4181d32c6e linux-next) [koichiroden: pulled follow-up fix from linux-next] Signed-off-by: Koichiro Den Signed-off-by: Roxana Nicolescu --- fs/smb/server/server.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/fs/smb/server/server.c b/fs/smb/server/server.c index 2bbc3c3316f0..d70a68abc219 100644 --- a/fs/smb/server/server.c +++ b/fs/smb/server/server.c @@ -273,8 +273,12 @@ static void handle_ksmbd_work(struct work_struct *wk) * disconnection. waitqueue_active is safe because it * uses atomic operation for condition. */ + atomic_inc(&conn->refcnt); if (!atomic_dec_return(&conn->r_count) && waitqueue_active(&conn->r_count_q)) wake_up(&conn->r_count_q); + + if (atomic_dec_and_test(&conn->refcnt)) + kfree(conn); } /**