RE: [PATCH v2 net-next 2/5] net: enetc: add Tx checksum offload for i.MX95 ENETC

[Claudiu Manoil]

> -----Original Message-----
> From: Wei Fang <wei.fang@xxxxxxx>
> Sent: Monday, November 11, 2024 2:16 PM
> To: Claudiu Manoil <claudiu.manoil@xxxxxxx>
> Cc: netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; imx@xxxxxxxxxxxxxxx;
> Vladimir Oltean <vladimir.oltean@xxxxxxx>; Clark Wang
> <xiaoning.wang@xxxxxxx>; andrew+netdev@xxxxxxx;
> davem@xxxxxxxxxxxxx; edumazet@xxxxxxxxxx; kuba@xxxxxxxxxx;
> pabeni@xxxxxxxxxx; Frank Li <frank.li@xxxxxxx>
> Subject: RE: [PATCH v2 net-next 2/5] net: enetc: add Tx checksum offload for
> i.MX95 ENETC
> 
> Hi Claudiu,
> 
> > -----Original Message-----
> > From: Wei Fang
> > Sent: 2024年11月11日 17:26
> > To: Claudiu Manoil <claudiu.manoil@xxxxxxx>
> > Cc: netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > imx@xxxxxxxxxxxxxxx; Vladimir Oltean <vladimir.oltean@xxxxxxx>; Clark
> > Wang <xiaoning.wang@xxxxxxx>; andrew+netdev@xxxxxxx;
> > davem@xxxxxxxxxxxxx; edumazet@xxxxxxxxxx; kuba@xxxxxxxxxx;
> > pabeni@xxxxxxxxxx; Frank Li <frank.li@xxxxxxx>
> > Subject: RE: [PATCH v2 net-next 2/5] net: enetc: add Tx checksum
> > offload for
> > i.MX95 ENETC
> >
> > > > --- a/drivers/net/ethernet/freescale/enetc/enetc.c
> > > > +++ b/drivers/net/ethernet/freescale/enetc/enetc.c
> > > > @@ -143,6 +143,27 @@ static int enetc_ptp_parse(struct sk_buff
> > > > *skb, u8 *udp,
> > > >  	return 0;
> > > >  }
> > > >
> > > > +static bool enetc_tx_csum_offload_check(struct sk_buff *skb) {
> > > > +	if (ip_hdr(skb)->version == 4)
> > >
> > > I would avoid using ip_hdr(), or any form of touching packed data
> > > and try extract this kind of info directly from the skb metadata
> > > instead, see also comment below.
> > >
> > > i.e., why not:
> > > if (skb->protocol == htons(ETH_P_IPV6)) ..  etc. ?
> >
> > skb->protocol may be VLAN protocol, such as ETH_P_8021Q,
> ETH_P_8021AD.
> > If so, it is impossible to determine whether it is an IPv4 or IPv6
> > frames through protocol.
> >
> > > or
> > > switch (skb->csum_offset) {
> > > case offsetof(struct tcphdr, check):
> > > [...]
> > > case offsetof(struct udphdr, check):
> > > [...]
> >
> > This seems to be able to be used to determine whether it is a UDP or
> > TCP frame.
> > Thanks.
> >
> > >
> > > > +		return ip_hdr(skb)->protocol == IPPROTO_TCP ||
> > > > +		       ip_hdr(skb)->protocol == IPPROTO_UDP;
> > > > +
> > > > +	if (ip_hdr(skb)->version == 6)
> > > > +		return ipv6_hdr(skb)->nexthdr == NEXTHDR_TCP ||
> > > > +		       ipv6_hdr(skb)->nexthdr == NEXTHDR_UDP;
> > > > +
> > > > +	return false;
> > > > +}
> > > > +
> > > > +static bool enetc_skb_is_tcp(struct sk_buff *skb) {
> > >
> > > There is a more efficient way of checking if L4 is TCP, without
> > > touching packet data, i.e. through the 'csum_offset' skb field:
> > > return skb->csum_offset == offsetof(struct tcphdr, check);
> > >
> > > Pls. have a look at these optimizations, I would expect visible
> > > improvements in throughput. Thanks.
> >
> > For small packets this might improve performance, but I'm not sure if
> > it would be a significant improvement. :)
> >
> 
> I didn't see any visible improvements in performance after using csum_offset.
> For example, when using pktgen to send 10,000,000 packets, the time taken is
> almost the same regardless of whether they are large or small packets, and the
> CPU idle ratio seen through the top command is also basically the same. Also,
> the UDP performance tested by iperf3 is basically the same.
> 

Maybe there's a bigger bottleneck somewhere else. I would change enetc_skb_is_tcp()
regardless, instead of 'if (ip_hdr() == 4) ... else ...', you can have the one line return statement
above.

As commented before, I would try to get rid of any access to packet content if skb metadata
fields could be used instead, but I don have a solution now on how to retrieve the IP protocol
version this way.

Thanks for testing anyway.