Re: [PATCH net-next 3/4] net: Support GRO/GSO fraglist chaining.

[Steffen Klassert <steffen.klassert@secunet.com>]

On Thu, Dec 19, 2019 at 11:28:34AM -0500, Willem de Bruijn wrote:
> On Thu, Dec 19, 2019 at 3:22 AM Steffen Klassert
> <steffen.klassert@secunet.com> wrote:
> >
> > I tried to find the subset of __copy_skb_header() that is needed to copy.
> > Some fields of nskb are still valid, and some (csum related) fields
> > should not be copied from skb to nskb.
> 
> Duplicating that code is kind of fragile wrt new fields being added to
> skbs later (such as the recent skb_ext example).
> 
> Perhaps we can split __copy_skb_header further and call the
> inner part directly.

I thought already about that, but __copy_skb_header does a
memcpy over all the header fields. If we split this, we
would need change the memcpy to direct assignments.

Maybe we can be conservative here and do a full
__copy_skb_header for now. The initial version
does not necessarily need to be the most performant
version. We could try to identify the correct subset
of header fields later then.

> >
> > I had to set ip_summed to CHECKSUM_UNNECESSARY on GRO to
> > make sure the noone touches the checksum of the head
> > skb. Otherise netfilter etc. tries to touch the csum.
> >
> > Before chaining I make sure that ip_summed and csum_level is
> > the same for all chained skbs and here I restore the original
> > value from nskb.
> 
> This is safe because the skb_gro_checksum_validate will have validated
> already on CHECKSUM_PARTIAL? What happens if there is decap or encap
> in the path? We cannot revert to CHECKSUM_PARTIAL after that, I
> imagine.

Yes, the checksum is validated with skb_gro_checksum_validate. If the
packets are UDP encapsulated, they are segmented before decapsulation.
Original values are already restored. If an additional encapsulation
happens, the encap checksum will be calculated after segmentation.
Original values are restored before that.

> 
> Either way, would you mind briefly documenting the checksum behavior
> in the commit message? It's not trivial and I doubt I'll recall the
> details in six months.

Yes, can do this.

> Really about patch 4: that squashed in a lot of non-trivial scaffolding
> from previous patch 'UDP: enable GRO by default'. Does it make sense
> to keep that in a separate patch? That should be a noop, which we can
> verify. And it makes patch 4 easier to reason about on its own, too.

Patch 4 is not that big, so not sure it that makes really sense.
But I can split out a preparation patch if that is preferred.

Anyway, I likely do another RFC version because we are already
late in the development cycle.