Re: [PATCH net-next 02/13] sk_buff: add skb extension infrastructure

On 12/12/2018 12:52 PM, Florian Westphal wrote:

Eric Dumazet [off-list ref] wrote:

quoted

quoted

If you prefer that, ok, but I don't see why we can't unify them behind
a single layer?

Well, for a start we do not use nf_brifge or secpath.

Then the extension framework isn't built and the result
is exactly the same as before these patches: helpers
turn into empty inline stubs.

If its built, sk_buff size is reduced by up to one pointer,
and the added one isn't initialised at allocation time.

sk_buff layout with the extension patches is almost same
as with XFRM=n BRIDGE_NF=n (because nf_bridge and sp pointers
get removed), there is only one additional pointer at the end,
not inited at alloc time.  The other new member fills a hole.

quoted

XDP is all about not unifying because unifying has a cost.

Do we really want to slow down the stack just because MPTCP is coming ?

I don't want to slow down the stack.  All places that gain a helper call
do so instead of the nf_bridge/secpath one.  They only expand into code
if at least one of secpath or nf_bridge are built in.
Both struct secpath and nf_bridge have their refcounts removed, so no
additional reference counts are added.

If you still think this is proposal is a bad idea, ok, let me know
and I will stop working on this.

Certainly not a bad idea.

MPTCP can then just add skb->cb2[], but neither nf_bridge nor secpath can
use it.

If not, I will send another iteration that just allocates the entire
extension space if first extension is added, it simplifies allocation
handling a little.

I am still unsure of the added extra costs, but for a start, TCP xmit clones
the skbs. 

Does it mean an additional pair of atomic operations if the skb comes from MPTCP enabled flow ?

The clone is used in the lower layers (IP , qdisc, device xmit), so I do not see why a clone
should share the extdata from the master.