Re: [PATCH net-next] net-core: add InMacErrors counter

[Cong Wang]

On Mon, Jan 24, 2022 at 09:29:24AM -0800, Jakub Kicinski wrote:
> On Mon, 24 Jan 2022 09:13:12 -0800 Brian Vazquez wrote:
> > On Fri, Jan 21, 2022 at 7:41 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > > On Sat, 22 Jan 2022 00:03:01 +0000 Jeffrey Ji wrote:  
> > > > From: jeffreyji <jeffreyji@xxxxxxxxxx>
> > > >
> > > > Increment InMacErrors counter when packet dropped due to incorrect dest
> > > > MAC addr.
> > > >
> > > > example output from nstat:
> > > > \~# nstat -z "*InMac*"
> > > > \#kernel
> > > > Ip6InMacErrors                  0                  0.0
> > > > IpExtInMacErrors                1                  0.0
> > > >
> > > > Tested: Created 2 netns, sent 1 packet using trafgen from 1 to the other
> > > > with "{eth(daddr=$INCORRECT_MAC...}", verified that nstat showed the
> > > > counter was incremented.
> > > >
> > > > Signed-off-by: jeffreyji <jeffreyji@xxxxxxxxxx>  
> > >
> > > How about we use the new kfree_skb_reason() instead to avoid allocating
> > > per-netns memory the stats?  
> > 
> > I'm not too familiar with the new kfree_skb_reason , but my
> > understanding is that it needs either the drop_monitor  or ebpf to get
> > the reason from the tracepoint, right? This is not too different from
> > using perf tool to find where the pkt is being dropped.
> > 
> > The idea here was to have a high level metric that is easier to find
> > for users that have less expertise on using more advance tools.
> 
> That much it's understood, but it's a trade off. This drop point
> existed for 20 years, why do we need to consume extra memory now?

kfree_skb_reason() is for tracing and tracing has overhead in
production, which is higher than just a percpu counter.

What memory overhead are you talking about? We have ~37 IP related
SNMP counters, this patch merely adds 1/37 memory overhead. So, what's the
point? :-/

Thanks.