skb_cow() vs skb_cow_head() - what exactly is "header"?

skb_cow() vs skb_cow_head() - what exactly is "header"?

[Radu Rendec]

Hello,

I'm working on some application-specific NIC driver. On the TX path, it
must remove a custom tag that sits between the Ethernet type field and
the actual Ethernet payload; then it must add a different tag in front
of the Ethernet header (the MAC DA field) before it hands over the frame
to the hardware for delivery.

I'm wondering if skb_cow_head() is enough to safely modify the skb as
described above, or I should use skb_cow() instead. I've been going
through the skbuff code and I still can't figure out the difference.
Comment blocks around the definitions of skb_cow(), skb_cow_head() and
SKB_DATAREF_SHIFT mention a "header" part and a "payload" part of the
skb, but it's still unclear to me which is what. As far as I understand,
there is only "data" and it sits between the headroom and tailroom. If
the skb is fragmented, additional data is stored in separate pages (in
frags[]) or skbs (in frag_list).

Last but not least, is there any *recent* documentation that explains
the subtleties of skbs? Most of the documentation I could find is very
old and/or covers only the basic concepts, such as head/data/tail/end.

Thank you in advance,
Radu Rendec

Re: skb_cow() vs skb_cow_head() - what exactly is "header"?

[David Miller]

From: Radu Rendec <radu.rendec@gmail.com>
Date: Sat, 29 Dec 2018 13:26:34 -0500

> I'm working on some application-specific NIC driver. On the TX path, it
> must remove a custom tag that sits between the Ethernet type field and
> the actual Ethernet payload; then it must add a different tag in front
> of the Ethernet header (the MAC DA field) before it hands over the frame
> to the hardware for delivery.
> 
> I'm wondering if skb_cow_head() is enough to safely modify the skb as
> described above, or I should use skb_cow() instead.

skb_cow_head() should be sufficient.  This is what the DSA layer tagging
code uses, and it is doing something similar if not exactly like what you
are doing.

See net/dsa/tag_dsa.c:dsa_xmit().

Re: skb_cow() vs skb_cow_head() - what exactly is "header"?

[Radu Rendec]

On Tue, 2019-01-01 at 09:53 -0800, David Miller wrote:
> From: Radu Rendec <radu.rendec@gmail.com
> Date: Sat, 29 Dec 2018 13:26:34 -0500
> 
> > I'm working on some application-specific NIC driver. On the TX path, it
> > must remove a custom tag that sits between the Ethernet type field and
> > the actual Ethernet payload; then it must add a different tag in front
> > of the Ethernet header (the MAC DA field) before it hands over the frame
> > to the hardware for delivery.
> 
> > I'm wondering if skb_cow_head() is enough to safely modify the skb as
> > described above, or I should use skb_cow() instead.
> 
> skb_cow_head() should be sufficient.  This is what the DSA layer tagging
> code uses, and it is doing something similar if not exactly like what you
> are doing.
> 
> See net/dsa/tag_dsa.c:dsa_xmit().

David, thank you - this is very helpful! Indeed, my use case is very
similar to DSA, but not identical, so unfortunately I can't use DSA.

However, my other question is still unanswered: what exactly is the
"header"? In the case of a fragmented skb, is it the entire "main" skb
(i.e. skb->data, without the fragments) or just a portion of it? Does
the "header" concept still apply to linear skb's?

Thanks,
Radu Rendec