tcp: md5: refine tcp_md5_do_add()/tcp_md5_hash_key() barriers

My prior fix went a bit too far, according to Herbert and Mathieu.

Since we accept that concurrent TCP MD5 lookups might see inconsistent
keys, we can use READ_ONCE()/WRITE_ONCE() instead of smp_rmb()/smp_wmb()

Clearing all key->key[] is needed to avoid possible KMSAN reports,
if key->keylen is increased. Since tcp_md5_do_add() is not fast path,
using __GFP_ZERO to clear all struct tcp_md5sig_key is simpler.

data_race() was added in linux-5.8 and will prevent KCSAN reports,
this can safely be removed in stable backports, if data_race() is
not yet backported.

v2: use data_race() both in tcp_md5_hash_key() and tcp_md5_do_add()

Fixes: 6a2febec ("tcp: md5: add missing memory barriers in tcp_md5_do_add()/tcp_md5_hash_key()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Marco Elver <elver@google.com>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv4/tcp.c

@@ -4033,14 +4033,14 @@ EXPORT_SYMBOL(tcp_md5_hash_skb_data);
 
 int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
 {
-	u8 keylen = key->keylen;
+	u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
 	struct scatterlist sg;
 
-	smp_rmb(); /* paired with smp_wmb() in tcp_md5_do_add() */
-
 	sg_init_one(&sg, key->key, keylen);
 	ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
-	return crypto_ahash_update(hp->md5_req);
+
+	/* We use data_race() because tcp_md5_do_add() might change key->key under us */
+	return data_race(crypto_ahash_update(hp->md5_req));
 }
 EXPORT_SYMBOL(tcp_md5_hash_key);
 

net/ipv4/tcp_ipv4.c

@@ -1111,12 +1111,21 @@ int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 
 	key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index);
 	if (key) {
-		/* Pre-existing entry - just update that one. */
-		memcpy(key->key, newkey, newkeylen);
+		/* Pre-existing entry - just update that one.
+		 * Note that the key might be used concurrently.
+		 * data_race() is telling kcsan that we do not care of
+		 * key mismatches, since changing MD5 key on live flows
+		 * can lead to packet drops.
+		 */
+		data_race(memcpy(key->key, newkey, newkeylen));
 
-		smp_wmb(); /* pairs with smp_rmb() in tcp_md5_hash_key() */
+		/* Pairs with READ_ONCE() in tcp_md5_hash_key().
+		 * Also note that a reader could catch new key->keylen value
+		 * but old key->key[], this is the reason we use __GFP_ZERO
+		 * at sock_kmalloc() time below these lines.
+		 */
+		WRITE_ONCE(key->keylen, newkeylen);
 
-		key->keylen = newkeylen;
 		return 0;
 	}
 
@@ -1132,7 +1141,7 @@ int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
 		rcu_assign_pointer(tp->md5sig_info, md5sig);
 	}
 
-	key = sock_kmalloc(sk, sizeof(*key), gfp);
+	key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
 	if (!key)
 		return -ENOMEM;
 	if (!tcp_alloc_md5sig_pool()) {