Re: [RFC net-next 0/6] net: configurable extra TX headroom for local egress

[Dragos Tatulea <dtatulea@nvidia.com>]

On 16.10.26 10:02, David Woodhouse wrote:
> On Thu, 2026-10-15 at 23:53 +0000, David Keogh wrote:
>>
>> Measurements (net-next at commit 05b5019b0, veth, 500000 UDP datagrams, a
>> 20-byte ip-in-ip push attached at TC egress, only tx_extra_headroom varied):
>>
>>   tx_extra_headroom   pskb_expand_head reallocs / 500k
>>   0 (off)             500003   (~1 per packet)
>>   64                  4        (background)
>>   128                 2        (background)
> 
> I remember frowning at this years ago, in the context of PPPoE.
> 
> I have the same question I had back then: why, if *every* packet ends
> up having to call pskb_expand_head() because it didn't have enough
> headroom, does our kernel need to be *told* by userspace that it needs
> to expand the allocations?
> 
> We gained dev->needed_headroom and ip_tunnel_adj_headroom() since then,
> and the kernel doesn't actually need to be told.
> 
> You explained why you don't want to use dev->needed_headroom; I don't
> have much of an opinion on that front. But if you do want to introduce
> a new tx_extra_headroom, could we try really hard to have it auto-
> tuning?
For TC-BPF I suppose only the BPF program knows how much extra
headroom is needed.

Maybe this tx_extra_headroom could be declared as a BTF decl tag
in the program. If this is the only usecase there's no need for a
rtnetlink uAPI.

Thanks,
Dragos