Received-SPF: pass (google.com: domain of linux-kernel+bounces-197205-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Precedence: bulk
MIME-Version: 1.0
References: <cover.1717105215.git.yan@cloudflare.com> <9be3733eee16bb81a7e8e2e57ebcc008f95cae08.1717105215.git.yan@cloudflare.com>
 <CANn89iLo6A__U5HqeA65NuBnrg36jpt9EOUC7T0fLdNEpa6eRQ@mail.gmail.com>
In-Reply-To: <CANn89iLo6A__U5HqeA65NuBnrg36jpt9EOUC7T0fLdNEpa6eRQ@mail.gmail.com>
From: Yan Zhai <yan@cloudflare.com>
Date: Fri, 31 May 2024 11:58:13 -0500
Message-ID: <CAO3-PboQ68+xFe4Z10L-s-k3NCgciGXNWM00-3wgqbPmGaBB9A@mail.gmail.com>
Subject: Re: [RFC net-next 1/6] net: add kfree_skb_for_sk function
To: Eric Dumazet <edumazet@google.com>
Cc: netdev@vger.kernel.org, "David S. Miller" <davem@davemloft.net>, 
	Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>, Simon Horman <horms@kernel.org>, 
	David Ahern <dsahern@kernel.org>, Abhishek Chauhan <quic_abchauha@quicinc.com>, 
	Mina Almasry <almasrymina@google.com>, Florian Westphal <fw@strlen.de>, 
	Alexander Lobakin <aleksander.lobakin@intel.com>, David Howells <dhowells@redhat.com>, 
	Jiri Pirko <jiri@resnulli.us>, Daniel Borkmann <daniel@iogearbox.net>, 
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>, Lorenzo Bianconi <lorenzo@kernel.org>, 
	Pavel Begunkov <asml.silence@gmail.com>, linux-kernel@vger.kernel.org, 
	kernel-team@cloudflare.com, Jesper Dangaard Brouer <hawk@kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

Hi Eric,

 Thanks for the feedback.

On Fri, May 31, 2024 at 1:51=E2=80=AFAM Eric Dumazet <edumazet@google.com> =
wrote:
>
> On Thu, May 30, 2024 at 11:46=E2=80=AFPM Yan Zhai <yan@cloudflare.com> wr=
ote:
> >
> > Implement a new kfree_skb_for_sk to replace kfree_skb_reason on a few
> > local receive path. The function accepts an extra receiving socket
> > argument, which will be set in skb->cb for kfree_skb/consume_skb
> > tracepoint consumption. With this extra bit of information, it will be
> > easier to attribute dropped packets to netns/containers and
> > sockets/services for performance and error monitoring purposes.
>
> This is a lot of code churn...
>
> I have to ask : Why not simply adding an sk parameter to an existing
> trace point ?
>
Modifying a signature of the current tracepoint seems like a breaking
change, that's why I was saving the context inside skb->cb, hoping to
not impact any existing programs watching this tracepoint. But
thinking it twice, it might not cause a problem if the signature
becomes:

 trace_kfree_skb(const struct sk_buff *skb, void *location, enum
skb_drop_reason reason, const struct sock *sk)

As return values are usually not a thing for tracepoints, it is
probably still compatible. The cons is that the last "sk" still breaks
the integrity of naming. How about making a "kfree_skb_context"
internal struct and putting it as the last argument to "hide" the
naming confusion?

> If this not possible, I would rather add new tracepoints, adding new clas=
ses,
> because it will ease your debugging :
>
> When looking for TCP drops, simply use a tcp_event_sk_skb_reason instance=
,
> and voila, no distractions caused by RAW/ICMP/ICMPv6/af_packet drops.
>
> DECLARE_EVENT_CLASS(tcp_event_sk_skb_reason,
>
>      TP_PROTO(const struct sock *sk, const struct sk_buff *skb, enum
> skb_drop_reason reason),
> ...
> );

The alternative of adding another tracepoint could indeed work, we had
a few cases like that in the past, e.g.

https://lore.kernel.org/lkml/20230711043453.64095-1-ivan@cloudflare.com/
https://lore.kernel.org/netdev/20230707043923.35578-1-ivan@cloudflare.com/

But it does feel like a whack-a-mole thing. The problems are solvable
if we extend the kfree_skb tracepoint, so I would prefer to not add a
new tracepoint.

>
> Also, the name ( kfree_skb_for_sk) and order of parameters is confusing.
>
> I always prefer this kind of ordering/names :
>
> void sk_skb_reason_drop( [struct net *net ] // not relevant here, but
> to expand the rationale
>               struct sock *sk, struct sk_buff *skb, enum skb_drop_reason =
reason)
>
> Looking at the name, we immediately see the parameter order.
>
> The consume one (no @reason there) would be called
>
> void sk_skb_consume(struct sock *sk, struct sk_buff *skb);

I was intending to keep the "kfree_skb" prefix initially since it
would appear less surprising to kernel developers who used kfree_skb
and kfree_skb_reason. But your points do make good sense. How about
"kfree_sk_skb_reason" and "consume_sk_skb" here?

thanks
Yan