Re: [PATCH net-next 0/5] skbuff: introduce skbuff_heads bulking and reusing

[Alexander Lobakin]

From: Eric Dumazet <edumazet@xxxxxxxxxx>
Date: Tue, 12 Jan 2021 13:32:56 +0100

> On Tue, Jan 12, 2021 at 11:56 AM Alexander Lobakin <alobakin@xxxxx> wrote:
>>
>
>>
>> Ah, I should've mentioned that I use UDP GRO Fraglists, so these
>> numbers are for GRO.
>>
>
> Right, this suggests UDP GRO fraglist is a pathological case of GRO,
> not saving memory.
>
> Real GRO (TCP in most cases) will consume one skb, and have page
> fragments for each segment.
>
> Having skbs linked together is not cache friendly.

OK, so I rebased test setup a bit to clarify the things out.

I disabled fraglists and GRO/GSO fraglists support advertisement
in driver to exclude any "pathological" cases and switched it
from napi_get_frags() + napi_gro_frags() to napi_alloc_skb() +
napi_gro_receive() to disable local skb reusing (napi_reuse_skb()).
I also enabled GSO UDP L4 ("classic" one: one skbuff_head + frags)
for forwarding, not only local traffic, and disabled NF flow offload
to increase CPU loading and drop performance below link speed so I
could see the changes.

So, the traffic flows looked like:
 - TCP GRO (one head + frags) -> NAT -> hardware TSO;
 - UDP GRO (one head + frags) -> NAT -> driver-side GSO.

Baseline 5.11-rc3:
 - 865 Mbps TCP, 866 Mbps UDP.

This patch (both separate caches and Edward's unified cache):
 - 899 Mbps TCP, 893 Mbps UDP.

So that's cleary *not* only "pathological" UDP GRO Fraglists
"problem" as TCP also got ~35 Mbps from this, as well as
non-fraglisted UDP.

Regarding latencies: I remember there were talks about latencies when
Edward introduced batched GRO (using linked lists to pass skbs from
GRO layer to core stack instead of passing one by one), so I think
it's a perennial question when it comes to batching/caching.

Thanks for the feedback, will post v2 soon.
The question about if this caching is reasonable isn't closed anyway,
but I don't see significant "cons" for now.

> So I would try first to make this case better, instead of trying to
> work around the real issue.

Al