Following the previous patch, RXSCs are held and properly refcounted in
the RX path (instead of being implicitly held by their SA), so the SA
doesn't need to hold a reference on its parent RXSC.
This also avoids panics on module unload caused by the double layer of
RCU callbacks (call_rcu frees the RXSA, which puts the final reference
on the RXSC and allows to free it in its own call_rcu) that commit
b196c22af5c3 ("macsec: add rcu_barrier() on module exit") didn't
protect against.
There were also some refcounting bugs in macsec_add_rxsa where I didn't
put the reference on the RXSC on the error paths, which would lead to
memory leaks.
Fixes:
c09440f7dcb3 ("macsec: introduce IEEE 802.1AE driver")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
crypto_free_aead(sa->key.tfm);
free_percpu(sa->stats);
- macsec_rxsc_put(sa->sc);
kfree(sa);
}
rtnl_lock();
rx_sc = get_rxsc_from_nl(genl_info_net(info), attrs, tb_rxsc, &dev, &secy);
- if (IS_ERR(rx_sc) || !macsec_rxsc_get(rx_sc)) {
+ if (IS_ERR(rx_sc)) {
rtnl_unlock();
return PTR_ERR(rx_sc);
}