Aw: Re: [PATCH 2/2] net: dsa: tag_ksz: dont let the hardware process the layer 4 checksum

[Lino Sanfilippo]

> Gesendet: Donnerstag, 15. Juli 2021 um 13:49 Uhr
> Von: "Vladimir Oltean" <olteanv@xxxxxxxxx>
> An: "Lino Sanfilippo" <LinoSanfilippo@xxxxxx>
> Cc: "Andrew Lunn" <andrew@xxxxxxx>, woojung.huh@xxxxxxxxxxxxx, UNGLinuxDriver@xxxxxxxxxxxxx, vivien.didelot@xxxxxxxxx, f.fainelli@xxxxxxxxx, davem@xxxxxxxxxxxxx, kuba@xxxxxxxxxx, netdev@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx
> Betreff: Re: [PATCH 2/2] net: dsa: tag_ksz: dont let the hardware process the layer 4 checksum
>
> On Thu, Jul 15, 2021 at 01:16:12PM +0200, Lino Sanfilippo wrote:
> > Sure, I will test this solution. But I think NETIF_F_FRAGLIST should also be
> > cleared in this case, right?
>
> Hmm, interesting question. I think only hns3 makes meaningful use of
> NETIF_F_FRAGLIST, right? I'm looking at hns3_fill_skb_to_desc().
> Other drivers seem to set it for ridiculous reasons - looking at commit
> 66aa0678efc2 ("ibmveth: Support to enable LSO/CSO for Trunk VEA.") -
> they set NETIF_F_FRAGLIST and then linearize the skb chain anyway. The
> claimed 4x throughput benefit probably has to do with less skbs
> traversing the stack? I don't know.
>
> Anyway, it is hard to imagine all the things that could go wrong with
> chains of IP fragments on a DSA interface, precisely because I have so
> few examples to look at. I would say, header taggers are probably fine,
> tail taggers not so much, so apply the same treatment as for NETIF_F_SG?
>

Please note that skb_put() asserts that the SKB is linearized. So I think we
should rather clear both NETIF_F_FRAGLIST and NETIF_F_SG unconditionally since also
header taggers use some form of skb_put() dont they?