Re: [PATCH v2 net-next 0/3] net: skb: introduce kfree_skb_with_reason() and use it for tcp and udp

[Cong Wang <xiyou.wangcong@gmail.com>]

On Mon, Jan 03, 2022 at 07:01:30PM -0700, David Ahern wrote:
> On 1/3/22 6:47 PM, Cong Wang wrote:
> > On Thu, Dec 30, 2021 at 05:32:37PM +0800, menglong8.dong@gmail.com wrote:
> >> From: Menglong Dong <imagedong@tencent.com>
> >>
> >> In this series patch, the interface kfree_skb_with_reason() is
> >> introduced(), which is used to collect skb drop reason, and pass
> >> it to 'kfree_skb' tracepoint. Therefor, 'drop_monitor' or eBPF is
> >> able to monitor abnormal skb with detail reason.
> >>
> > 
> > We already something close, __dev_kfree_skb_any(). Can't we unify
> > all of these?
> 
> Specifically?
> 
> The 'reason' passed around by those is either SKB_REASON_CONSUMED or
> SKB_REASON_DROPPED and is used to call kfree_skb vs consume_skb. i.e.,
> this is unrelated to this patch set and goal.

What prevents you extending it?

> 
> > 
> > 
> >> In fact, this series patches are out of the intelligence of David
> >> and Steve, I'm just a truck man :/
> >>
> > 
> > I think there was another discussion before yours, which I got involved
> > as well.
> > 
> >> Previous discussion is here:
> >>
> >> https://lore.kernel.org/netdev/20211118105752.1d46e990@gandalf.local.home/
> >> https://lore.kernel.org/netdev/67b36bd8-2477-88ac-83a0-35a1eeaf40c9@gmail.com/
> >>
> >> In the first patch, kfree_skb_with_reason() is introduced and
> >> the 'reason' field is added to 'kfree_skb' tracepoint. In the
> >> second patch, 'kfree_skb()' in replaced with 'kfree_skb_with_reason()'
> >> in tcp_v4_rcv(). In the third patch, 'kfree_skb_with_reason()' is
> >> used in __udp4_lib_rcv().
> >>
> > 
> > I don't follow all the discussions here, but IIRC it would be nice
> > if we can provide the SNMP stat code (for instance, TCP_MIB_CSUMERRORS) to
> > user-space, because those stats are already exposed to user-space, so
> > you don't have to invent new ones.
> 
> Those SNMP macros are not unique and can not be fed into a generic
> kfree_skb_reason function.

Sure, you also have the skb itself, particularly skb protocol, with
these combined, it should be unique.

Thanks.