On 12/05/2024 23:35, Sabrina Dubroca wrote:

2024-05-06, 03:16:22 +0200, Antonio Quartulli wrote:

quoted

+/* send skb to connected peer, if any */
+static void ovpn_queue_skb(struct ovpn_struct *ovpn, struct sk_buff *skb,
+			   struct ovpn_peer *peer)
+{
+	int ret;
+
+	if (likely(!peer))
+		/* retrieve peer serving the destination IP of this packet */
+		peer = ovpn_peer_get_by_dst(ovpn, skb);
+	if (unlikely(!peer)) {
+		net_dbg_ratelimited("%s: no peer to send data to\n",
+				    ovpn->dev->name);
+		goto drop;
+	}
+
+	ret = ptr_ring_produce_bh(&peer->tx_ring, skb);
+	if (unlikely(ret < 0)) {
+		net_err_ratelimited("%s: cannot queue packet to TX ring\n",
+				    peer->ovpn->dev->name);
+		goto drop;
+	}
+
+	if (!queue_work(ovpn->crypto_wq, &peer->encrypt_work))
+		ovpn_peer_put(peer);

I wanted to come back to this after going through the crypto patch,
because this felt like a strange construct when I first looked at this
patch.

Why are you using a workqueue here? Based on the kdoc for crypto_wq
("used to schedule crypto work that may sleep during TX/RX"), it's to
deal with async crypto.

If so, why not use the more standard way of dealing with async crypto
in contexts that cannot sleep, ie letting the crypto core call the
"done" callback asynchronously? You need to do all the proper refcount
handling, but IMO it's cleaner and simpler than this workqueue and
ptr_ring. You can see an example of that in macsec (macsec_encrypt_*
in drivers/net/macsec.c).

Aha! You don't know how happy I was when I found the doc describing how 
to convert the async code into sync-looking :-) With the detail that I 
had to move to a different context, as the code may want to sleep (hence 
the introduction of the workqueue).

It looks like I am little fan of WQs, while you are telling me to avoid 
them if possible.

I presume that using WQs comes with a non-negligible cost, therefore if 
we can just get things done without having to use them, then I should 
just don't.

I think I could go back to no-workqueue encrypt/decrypt.
Do you think this may have any impact on any future multi-core 
optimization? Back then I also thought that going through workers may 
make improvements in this area easier. But I could just be wrong.

Regards,


-- 
Antonio Quartulli
OpenVPN Inc.