Hello Igor,

On Wed, Dec 18, 2019 at 01:40:39PM +0000, Igor Russkikh wrote:

quoted

@@ -2922,7 +3300,27 @@ static int macsec_changelink(struct net_device

*dev, struct nlattr *tb[],
 	    data[IFLA_MACSEC_PORT])
 		return -EINVAL;
 
-	return macsec_changelink_common(dev, data);
+	/* If h/w offloading is available, propagate to the device */
+	if (macsec_is_offloaded(macsec)) {
+		const struct macsec_ops *ops;
+		struct macsec_context ctx;
+		int ret;
+
+		ops = macsec_get_ops(netdev_priv(dev), &ctx);
+		if (!ops)
+			return -EOPNOTSUPP;
+
+		ctx.secy = &macsec->secy;
+		ret = macsec_offload(ops->mdo_upd_secy, &ctx);
+		if (ret)
+			return ret;
+	}
+
+	ret = macsec_changelink_common(dev, data);

In our mac driver verification we see that propagating upd_secy to
device before doing macsec_changelink_common is actually useless,
since in this case underlying device can't fetch any of the updated
parameters from the macsec structures.

Isn't it logical first doing `macsec_changelink_common` and then
propagate the event?

Doing the macsec_changelink_common after propagating the event to the
device driver was done to ease the fail case scenario (it's quite hard
to revert macsec_changelink_common). But then you're right that many
parameters are set by macsec_changelink_common, which means it must be
performed before the propagation of the upd_secy event.

I think the solution is to keep a copy of unmodified secy and tx_sc, and
in case of failure to revert the operation by copying the whole
structures back. That would allow to move macsec_changelink_common up.
Would that work for you?

Thanks for spotting this!
Antoine

-- 
Antoine Ténart, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com