Re: [PATCH v4 net-next] net: ftmac100: support mtu > 1500

[Vladimir Oltean <olteanv@gmail.com>]

On Mon, Oct 24, 2022 at 06:09:02PM +0200, Andrew Lunn wrote:
> On Mon, Oct 24, 2022 at 06:50:31PM +0300, Sergei Antonov wrote:
> > On Wed, 19 Oct 2022 at 21:42, Vladimir Oltean <olteanv@gmail.com> wrote:
> > >
> > > On Wed, Oct 19, 2022 at 09:36:21PM +0300, Sergei Antonov wrote:
> > > > On Wed, 19 Oct 2022 at 19:55, Vladimir Oltean <olteanv@gmail.com> wrote:
> > > > >
> > > > > On Wed, Oct 19, 2022 at 07:20:58PM +0300, Sergei Antonov wrote:
> > > > > > The ftmac100 controller considers some packets FTL (frame
> > > > > > too long) and drops them. An example of a dropped packet:
> > > > > > 6 bytes - dst MAC
> > > > > > 6 bytes - src MAC
> > > > > > 2 bytes - EtherType IPv4 (0800)
> > > > > > 1504 bytes - IPv4 packet
> > > > >
> > > > > Why do you insist writing these confusing messages?
> > 
> > Working on a better version of the patch. And here is another question.
> > Unless the flag is set, the controller drops packets bigger than 1518
> > (that is 1500 payload + 14 Ethernet header + 4 FCS). So if mtu is 1500
> > the driver can enable the controller's functionality (clear the
> > FTMAC100_MACCR_RX_FTL flag) and save CPU time. When mtu is less or
> > greater than 1500, should the driver do the following:
> > if (ftmac100_rxdes_frame_length(rxdes) > netdev->mtu + ETH_HLEN) {
> >     drop the packet
> > }
> > I mean, is it driver's duty to drop oversized packets?
> 
> It is not well defined what should happen here. Some drivers will
> deliver oversized packets to the stack, some will not.
> 
> The main purpose of the MTU is fragmentation on transmission. How do
> you break a UDP or TCP PDU up into L2 frames. MTU is not really used
> on reception. If the L2 frame is otherwise valid, i doubt the stack
> will drop it if it is longer than expected.
> 
>      Andrew

I think the general wisdom is for a driver to not go out of its way to
accept packets larger than the MTU, and not go out of its way to drop
them either. The only given guarantee is that packets with an L2 length
<= dev->mtu are accepted.

The sad reality is that some layers like DSA rely on this marginal
behavior in some cases (controllers accept packets slightly larger than
dev->mtu) and this is why we took the decisions that we took (discussion
which I linked to in the v2 of this patch:
https://patchwork.ozlabs.org/project/netdev/patch/20200421123110.13733-2-olteanv@gmail.com/)