Re: [PATCH net-next v2 6/9] net: macsec: hardware offloading infrastructure

From: Sabrina Dubroca
Date: Fri Aug 16 2019 - 09:25:10 EST


2019-08-13, 10:58:17 +0200, Antoine Tenart wrote:
> Hi Igor,
>
> On Sat, Aug 10, 2019 at 01:20:32PM +0000, Igor Russkikh wrote:
> > On 08.08.2019 17:05, Antoine Tenart wrote:
> >
> > > The Rx and TX handlers are modified to take in account the special case
> > > were the MACsec transformation happens in the hardware, whether in a PHY
> > > or in a MAC, as the packets seen by the networking stack on both the
> >
> > Don't you think we may eventually may need xmit / handle_frame ops to be
> > a part of macsec_ops?
> >
> > That way software macsec could be extract to just another type of offload.
> > The drawback of current code is it doesn't show explicitly the path of
> > offloaded packets. It is hidden in `handle_not_macsec` and in
> > `macsec_start_xmit` branch. This makes incorrect counters to tick (see my below
> > comment)
> >
> > Another thing is that both xmit / macsec_handle_frame can't now be customized
> > by device driver. But this may be required.
> > We for example have usecases and HW features to allow specific flows to bypass
> > macsec encryption. This is normally used for macsec key control protocols,
> > identified by ethertype. Your phy is also capable on that as I see.
>
> I think this question is linked to the use of a MACsec virtual interface
> when using h/w offloading. The starting point for me was that I wanted
> to reuse the data structures and the API exposed to the userspace by the
> s/w implementation of MACsec. I then had two choices: keeping the exact
> same interface for the user (having a virtual MACsec interface), or

Unless it's really infeasible, yes, that's how things should be done IMO.

> registering the MACsec genl ops onto the real net devices (and making
> the s/w implementation a virtual net dev and a provider of the MACsec
> "offloading" ops).

Please, no :( Let's keep it as close as possible to the software
implementation, unless there's a really good reason not to. It's not
just "ip macsec" btw, wpa_supplicant can also configure MACsec and
would also need some logic to pick the device on which to do the genl
operations in that case.

> The advantages of the first option were that nearly all the logic of the
> s/w implementation could be kept and especially that it would be
> transparent for the user to use both implementations of MACsec. But this
> raised an issue as I had to modify the xmit / handle_frame ops to let
> all the traffic pass. This is because we have no way of knowing if a
> frame was handled by the MACsec h/w or not in ingress. So the virtual
> interface here only serve as the entrypoint for the API...

It's also the interface on which you'll run DHCP or install IP addresses.

> The second option would have the advantage to better represent the actual
> flow, but the way of configuring MACsec would be a bit different for the
> user, whether he wants to use s/w or h/w MACsec. If we were to do this I
> think we could extract the genl functions from the MACsec s/w
> implementation, and let it implement the MACsec ops (exactly as the
> offloading drivers).
>
> I'm open to discussing this :)
>
> As for the need for xmit / handle_frame ops (for a MAC w/ MACsec
> offloading), I'd say the xmit / handle_frame ops of the real net device
> driver could be used as the one of the MACsec virtual interface do not
> do much (regardless of the implementation choice discussed above).

There's no "handle_frame" op on a real device. macsec_handle_frame is
an rx_handler specificity that grabs packets from a real device and
sends them into a virtual device stacked on top of it. A real device
just hands packets over to the stack via NAPI.


> > > @@ -2546,11 +2814,15 @@ static netdev_tx_t macsec_start_xmit(struct sk_buff *skb,
> > > {
> > > struct macsec_dev *macsec = netdev_priv(dev);
> > > struct macsec_secy *secy = &macsec->secy;
> > > + struct macsec_tx_sc *tx_sc = &secy->tx_sc;
> > > struct pcpu_secy_stats *secy_stats;
> > > + struct macsec_tx_sa *tx_sa;
> > > int ret, len;
> > >
> > > + tx_sa = macsec_txsa_get(tx_sc->sa[tx_sc->encoding_sa]);
> >
> > Declared, but not used?
>
> I'll remove it then.

That's also a refcount leak, so, yes, please get rid of it.

[I'll answer the rest of the patch separately]

--
Sabrina