Re: [PATCH net-next v3 3/3] net: icmp: add reasons of the skb drops to icmp protocol

[David Ahern <dsahern@kernel.org>]

On 3/17/22 8:53 AM, David Laight wrote:
> From: David Ahern
>> Sent: 17 March 2022 14:49
>>
>> On 3/16/22 10:05 PM, Jakub Kicinski wrote:
>>> On Wed, 16 Mar 2022 21:35:47 -0600 David Ahern wrote:
>>>> On 3/16/22 9:18 PM, Jakub Kicinski wrote:
>>>>>
>>>>> I guess this set raises the follow up question to Dave if adding
>>>>> drop reasons to places with MIB exception stats means improving
>>>>> the granularity or one MIB stat == one reason?
>>>>
>>>> There are a few examples where multiple MIB stats are bumped on a drop,
>>>> but the reason code should always be set based on first failure. Did you
>>>> mean something else with your question?
>>>
>>> I meant whether we want to differentiate between TYPE, and BROADCAST or
>>> whatever other possible invalid protocol cases we can get here or just
>>> dump them all into a single protocol error code.
>>
>> I think a single one is a good starting point.
> 
> I remember looking at (I think) the packet drop stats a while back.
> Two machines on the same LAN were reporting rather different values.
> Basically 0 v quite a few.
> 
> It turned out that passing the packets to dhcp was deemed enough
> to stop them being reported as 'dropped'.
> And I think that version of dhcp fed every packed into its BPF? filter.
> (I never did decide whether that caused every skb to be duplicated.)
> 

I believe it depends on the type of socket. Packet sockets - e.g.,
running lldpd or tcpdump - do cause every packet to be cloned and kills
performance.