Date:   Tue, 30 Jul 2019 18:21:44 -0700
From:   Jakub Kicinski <jakub.kicinski@...ronome.com>
To:     Alexei Starovoitov <alexei.starovoitov@...il.com>
Cc:     "Daniel T. Lee" <danieltimlee@...il.com>,
        Stephen Hemminger <stephen@...workplumber.org>,
        David Ahern <dsahern@...il.com>,
        Jesper Dangaard Brouer <brouer@...hat.com>,
        John Fastabend <john.fastabend@...il.com>,
        Daniel Borkmann <daniel@...earbox.net>,
        Alexei Starovoitov <ast@...nel.org>,
        David Miller <davem@...emloft.net>, netdev@...r.kernel.org
Subject: Re: [PATCH 0/2] tools: bpftool: add net (un)load command to load
 XDP

On Tue, 30 Jul 2019 17:23:39 -0700, Alexei Starovoitov wrote:
> On Tue, Jul 30, 2019 at 05:07:25PM -0700, Jakub Kicinski wrote:
> > Nothing meaning you disagree it's duplicated effort and unnecessary 
> > LoC the community has to maintain, review, test..?  
> 
> I don't see duplicated effort.

Could you walk me through a scenario where, say, a new xdp attachment
flag is added? How can there be support in both bpftool and iproute2 
for specifying it without modifying both?

How can we say we want to make iproute2 use libbpf instead of it's own
library to avoid duplicated effort on the back end, and at the same time
claim having two tools do the same thing is no duplication 🤯

> > Duplicating the same features in bpftool will only diminish the
> > incentive for moving iproute2 to libbpf.   
> 
> not at all. why do you think so?

Because iproute2's BPF has fallen behind so the simplest thing is to
just contribute to bpftool. But iproute2 is the tool set for Linux
networking, we can't let it bit rot :(

Admittedly that's just me reading the tea leaves.

> > And for folks who deal
> > with a wide variety of customers, often novices maintaining two
> > ways of doing the same thing is a hassle :(  
> 
> It's not the same thing.
> I'm talking about my experience dealing with 'wide variety of bpf customers'.
> They only have a fraction of their time to learn one tool.
> Making every bpf customer learn ten tools is not an option.

IMHO vaguely competent users of Linux networking will know ip link.
If they are not vaguely competent, they should not attach XDP programs
to interfaces by hand...

> > Let's just put the two commands next to each other:
> > 
> >        ip link set xdp $PROG dev $DEV
> > 
> > bpftool net attach xdp $PROG dev $DEV
> > 
> > Are they that different?  
> 
> yes.
> they're different tools with they own upgrade/rollout cycle

I think this came up when bpftool net was added and I never really
understood it.

> > > If bpftool was taught to do equivalent of 'ip link' that would be
> > > very different story and I would be opposed to that.  
> > 
> > Yes, that'd be pretty clear cut, only the XDP stuff is a bit more 
> > of a judgement call.  
> 
> bpftool must be able to introspect every aspect of bpf programming.
> That includes detaching and attaching anywhere.
> Anyone doing 'bpftool p s' should be able to switch off particular
> prog id without learning ten different other tools.

I think the fact that we already have an implementation in iproute2,
which is at the risk of bit rot is more important to me that the
hypothetical scenario where everyone knows to just use bpftool (for
XDP, for TC it's still iproute2 unless there's someone crazy enough 
to reimplement the TC functionality :))

I'm not sure we can settle our differences over email :)
I have tremendous respect for all the maintainers I CCed here, 
if nobody steps up to agree with me I'll concede the bpftool net
battle entirely :)

Powered by blists - more mailing lists