Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 3 Feb 2022 11:34:26 -0800
From:   Jakub Kicinski <kuba@kernel.org>
To:     Eric Dumazet <edumazet@google.com>
Cc:     Jeffrey Ji <jeffreyjilinux@gmail.com>,
        "David S . Miller" <davem@davemloft.net>,
        Brian Vazquez <brianvv@google.com>,
        LKML <linux-kernel@vger.kernel.org>,
        netdev <netdev@vger.kernel.org>, jeffreyji <jeffreyji@google.com>
Subject: Re: [PATCH v6 net-next] net-core: add InMacErrors counter
Message-ID: <20220203113426.49c70285@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
In-Reply-To: <CANn89iJqQwUVnh3SPZ7j4RGMhEZsBk3uT3wosAbb1aFSzoyS+A@mail.gmail.com>
References: <20220201222845.3640041-1-jeffreyji@google.com>
        <20220202205916.58f4a592@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com>
        <CANn89iJqQwUVnh3SPZ7j4RGMhEZsBk3uT3wosAbb1aFSzoyS+A@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Thu, 3 Feb 2022 10:13:59 -0800 Eric Dumazet wrote:
> > I had another thing and this still doesn't sit completely well
> > with me :(
> >
> > Shouldn't we count those drops as skb->dev->rx_dropped?
> > Commonly NICs will do such filtering and if I got it right
> > in struct rtnl_link_stats64 kdoc - report them as rx_dropped.
> > It'd be inconsistent if on a physical interface we count
> > these as rx_dropped and on SW interface (or with promisc enabled
> > etc.) in the SNMP counters.  
> 
> I like to see skb->dev->rx_dropped as a fallback-catch-all bucket
> for all cases we do not already have a more specific counter.

Indeed, it's a fallback so counting relatively common events like
unicast filtering into generic "drops" feels wrong. I heard complaints
that this is non-intuitive and makes debug harder in the past.

> > Or we can add a new link stat that NICs can use as well.  
> 
> Yes, this could be done, but we have to be careful about not hitting
> a single cache line, for the cases we receive floods of such messages
> on multiqueue NIC.
> (The single atomic in dev->rx_dropped) is suffering from this issue btw)

Even more of a reason to bite the bullet and move from the atomic
counters to pcpu stats?

> > In fact I'm not sure this is really a IP AKA L3 statistic,
> > it's the L2 address that doesn't match.
> >
> >
> > If everyone disagrees - should we at least rename the MIB counter
> > similarly to the drop reason? Experience shows that users call for
> > help when they see counters with Error in their name, I'd vote for
> > IpExtInDropOtherhost or some such. The statistic should also be
> > documented in Documentation/networking/snmp_counter.rst