Hi,

On 09/07/2024 10:51, Sabrina Dubroca wrote:

2024-06-27, 15:08:30 +0200, Antonio Quartulli wrote:

quoted

+/* removes the primary key from the crypto context */
+void ovpn_crypto_kill_primary(struct ovpn_crypto_state *cs)
+{
+	struct ovpn_crypto_key_slot *ks;
+
+	mutex_lock(&cs->mutex);
+	ks = rcu_replace_pointer(cs->primary, NULL,
+				 lockdep_is_held(&cs->mutex));

Should there be a check that we're killing the key that has expired
and not some other key?  I'm wondering if this could happen:

ovpn_encrypt_one
     ovpn_aead_encrypt
         ovpn_pktid_xmit_next
             seq_num reaches threshold
             returns -ERANGE
         returns -ERANGE

                                             ovpn_crypto_key_slots_swap
                                                 replaces cs->primary with cs->secondary

     ovpn_encrypt_post
         ret = -ERANGE
         ovpn_crypto_kill_primary
             kills the freshly installed primary key

quoted

+	ovpn_crypto_key_slot_put(ks);
+	mutex_unlock(&cs->mutex);
+}

hm I think you're right.

This is theoretically possible...
Userspace might be reneweing the key exactly when the primary key is 
expiring..

We need to specify the key_id and kill exactly that key.

quoted

+

[...]

quoted

+static void ovpn_aead_encrypt_done(void *data, int ret)
+{
+	struct sk_buff *skb = data;
+
+	aead_request_free(ovpn_skb_cb(skb)->req);
+	ovpn_encrypt_post(skb, ret);
+}
+
+int ovpn_aead_encrypt(struct ovpn_crypto_key_slot *ks, struct sk_buff *skb,
+		      u32 peer_id)
+{
+	const unsigned int tag_size = crypto_aead_authsize(ks->encrypt);
+	const unsigned int head_size = ovpn_aead_encap_overhead(ks);
+	struct scatterlist sg[MAX_SKB_FRAGS + 2];
+	DECLARE_CRYPTO_WAIT(wait);

unused? (also in _decrypt)

Right. Leftover from the sync crypto approach. Will drop it.

[...]

quoted

+
+	req = aead_request_alloc(ks->encrypt, GFP_ATOMIC);
+	if (unlikely(!req))
+		return -ENOMEM;
+
+	/* setup async crypto operation */
+	aead_request_set_tfm(req, ks->encrypt);
+	aead_request_set_callback(req, 0, ovpn_aead_encrypt_done, NULL);

NULL? That should be skb, ovpn_aead_encrypt_done needs it (same for
decrypt).

Ouch

I suspect you haven't triggered the async path in testing. For that,
you can use crconf:

Yeah, I doubt I did, although I tried pumping as much traffic as 
possible (but it may not have been enough to trigger the needed conditions).

git clone https://git.code.sf.net/p/crconf/code
cd code && make
./src/crconf add driver 'pcrypt(generic-gcm-aesni)' type 3  priority 10000

Then all packets encrypted with gcm(aes) should go through the async
code.

Amazing! Thanks! Will give it a go.

quoted

+	aead_request_set_crypt(req, sg, sg, skb->len - head_size, iv);
+	aead_request_set_ad(req, OVPN_OP_SIZE_V2 + NONCE_WIRE_SIZE);
+
+	ovpn_skb_cb(skb)->req = req;
+	ovpn_skb_cb(skb)->ks = ks;
+
+	/* encrypt it */
+	return crypto_aead_encrypt(req);
+}

[...]

quoted

@@ -77,14 +133,45 @@ static void ovpn_decrypt_post(struct sk_buff *skb, int ret)
  /* pick next packet from RX queue, decrypt and forward it to the device */
  void ovpn_recv(struct ovpn_peer *peer, struct sk_buff *skb)
  {
+	struct ovpn_crypto_key_slot *ks;
+	u8 key_id;
+
+	/* get the key slot matching the key ID in the received packet */
+	key_id = ovpn_key_id_from_skb(skb);
+	ks = ovpn_crypto_key_id_to_slot(&peer->crypto, key_id);

This takes a reference on the keyslot (ovpn_crypto_key_slot_hold), but
I don't see it getting released in ovpn_decrypt_post. In
ovpn_encrypt_post you're adding a ovpn_crypto_key_slot_put (to match
ovpn_crypto_key_slot_primary), but nothing equivalent in
ovpn_decrypt_post?

good catch! I don't think it triggered any critical side effect (Except 
from the leak) hence it went unnoticed.

Will add the ovpn_crypto_key_slot_put in the exit path.


Thanks!

quoted

+	if (unlikely(!ks)) {
+		net_info_ratelimited("%s: no available key for peer %u, key-id: %u\n",
+				     peer->ovpn->dev->name, peer->id, key_id);
+		dev_core_stats_rx_dropped_inc(peer->ovpn->dev);
+		kfree_skb(skb);
+		return;
+	}
+
  	ovpn_skb_cb(skb)->peer = peer;
-	ovpn_decrypt_post(skb, 0);
+	ovpn_decrypt_post(skb, ovpn_aead_decrypt(ks, skb));
  }

-- 
Antonio Quartulli
OpenVPN Inc.