On Fri, May 29, 2020 at 11:14 PM Willem de Bruijn
<willemdebruijn.ker...@gmail.com> wrote:
>
> On Fri, May 29, 2020 at 8:27 PM David Miller <da...@davemloft.net> wrote:
> >
> > From: Willem de Bruijn <willemdebruijn.ker...@gmail.com>
> > Date: Thu, 28 May 2020 13:05:32 -0400
> >
> > > Temporarily pull ETH_HLEN to make control flow the same for frags and
> > > not frags. Then push the header just before calling napi_gro_frags.
> > ...
> > > case IFF_TAP:
> > > - if (!frags)
> > > - skb->protocol = eth_type_trans(skb, tun->dev);
> > > + if (frags && !pskb_may_pull(skb, ETH_HLEN)) {
> > > + err = -ENOMEM;
> > > + goto drop;
> > > + }
> > > + skb->protocol = eth_type_trans(skb, tun->dev);
> > ...
> > > /* Exercise flow dissector code path. */
> > > - u32 headlen = eth_get_headlen(tun->dev, skb->data,
> > > - skb_headlen(skb));
> > > + skb_push(skb, ETH_HLEN);
> > > + headlen = eth_get_headlen(tun->dev, skb->data,
> > > + skb_headlen(skb));
> >
> > I hate to be a stickler on wording in the commit message, but the
> > change is not really "pulling" the ethernet header from the SKB.
> >
> > Instead it is invoking pskb_may_pull() which just makes sure the
> > header is there in the linear SKB data area.
> >
> > Can you please refine this description and resubmit?
>
> Of course. How is this
>
> "
> Ensure the link layer header lies in linear as eth_type_trans pulls
> ETH_HLEN. Then take the same code paths for frags as for not frags.
> Push the link layer header back just before calling napi_gro_frags.
>
> By pulling up to ETH_HLEN from frag0 into linear, this disables the
> frag0 optimization in the special case when IFF_NAPI_FRAGS is
> called with zero length iov[0] (and thus empty skb->linear).
> "
>
> Seemed good to add the extra clarification. I don't see a reasonable way
> to avoid that consequence, especially as I cannot restore the first skb frag
> (iov[1]) if it was exactly ETH_HLEN bytes and thus freed by __skb_pull_tail.
Sent. Probably faster that way. Do let me know if still too fast and
loose with wording. I can always do a v3.
Or to add some frags gymnastics to try to maintain the frag0 optimization
when iov[1] > ETH_LEN and frag0 thus can be restored. That just makes
for a more complicated fix.
