crypto: af_alg - Use bh_lock_sock in sk_destruct

Commit Message

Herbert Xu Dec. 5, 2019, 5:45 a.m. UTC

On Wed, Dec 04, 2019 at 08:59:11PM -0800, Eric Dumazet wrote:
>
> crypto layer (hash_sock_destruct()) is called from rcu callback (this in BH context) but tries to grab a socket lock.
> 
> A socket lock can schedule, which is illegal in BH context.

Fair enough.  Although I was rather intrigued as to how the RCU call
occured in the first place.  After some digging my theory is that
this is due to a SO_ATTACH_REUSEPORT_CBPF or SO_ATTACH_REUSEPORT_EBPF
setsockopt on the crypto socket.

What are these filters even suppposed to do on an af_alg socket?

Anyhow, this is a bug that could have been triggered even without
this, but it would have been almost impossible to do it through
syzbot as you need to have an outstanding async skcipher/aead request
that is freed in BH context.

---8<---
As af_alg_release_parent may be called from BH context (most notably
due to an async request that only completes after socket closure,
or as reported here because of an RCU-delayed sk_destruct call), we
must use bh_lock_sock instead of lock_sock.

Reported-by: [e-mail]
Reported-by: Eric Dumazet <[e-mail]>
Fixes: c840ac6af3f8 ("crypto: af_alg - Disallow bind/setkey/...")
Cc: <[e-mail]>
Signed-off-by: Herbert Xu <[e-mail]>

Patch

diff --git a/crypto/af_alg.c b/crypto/af_alg.c
index 0dceaabc6321..3d8e53010cda 100644
--- a/crypto/af_alg.c
+++ b/crypto/af_alg.c
@@ -134,11 +134,13 @@  void af_alg_release_parent(struct sock *sk)
 	sk = ask->parent;
 	ask = alg_sk(sk);
 
-	lock_sock(sk);
+	local_bh_disable();
+	bh_lock_sock(sk);
 	ask->nokey_refcnt -= nokey;
 	if (!last)
 		last = !--ask->refcnt;
-	release_sock(sk);
+	bh_unlock_sock(sk);
+	local_bh_enable();
 
 	if (last)
 		sock_put(sk);