Re: [Patch bpf-next v3 4/5] skmsg: use skb ext instead of TCP_SKB_CB

From: Cong Wang <xiyou.wangcong@gmail.com>
To: John Fastabend <john.fastabend@gmail.com>
Cc: Linux Kernel Network Developers <netdev@vger.kernel.org>,
	bpf <bpf@vger.kernel.org>,
	duanxiongchun@bytedance.com,
	Dongdong Wang <wangdongdong.6@bytedance.com>,
	jiang.wang@bytedance.com, Cong Wang <cong.wang@bytedance.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Jakub Sitnicki <jakub@cloudflare.com>,
	Lorenz Bauer <lmb@cloudflare.com>
Subject: Re: [Patch bpf-next v3 4/5] skmsg: use skb ext instead of TCP_SKB_CB
Date: Mon, 15 Feb 2021 17:04:00 -0800	[thread overview]
Message-ID: <CAM_iQpWzRpfWwZHPK=+KWbu+nLxJ=GKRHNC+97NT2DoN0qRc2A@mail.gmail.com> (raw)
In-Reply-To: <602b17b0492a8_3ed41208f2@john-XPS-13-9370.notmuch>

On Mon, Feb 15, 2021 at 4:54 PM John Fastabend <john.fastabend@gmail.com> wrote:
>
> Cong Wang wrote:
> > On Mon, Feb 15, 2021 at 3:57 PM John Fastabend <john.fastabend@gmail.com> wrote:
> > >
> > > For TCP case we can continue to use CB and not pay the price. For UDP
> > > and AF_UNIX we can do the extra alloc.
> >
> > I see your point, but specializing TCP case does not give much benefit
> > here, the skmsg code would have to check skb->protocol etc. to decide
> > whether to use TCP_SKB_CB() or skb_ext:
> >
> > if (skb->protocol == ...)
> >   TCP_SKB_CB(skb) = ...;
> > else
> >   ext = skb_ext_find(skb);
> >
> > which looks ugly to me. And I doubt skb->protocol alone is sufficient to
> > distinguish TCP, so we may end up having more checks above.
> >
> > So do you really want to trade code readability with an extra alloc?
>
> Above is ugly. So I look at where the patch replaces things,
>
> sk_psock_tls_strp_read(), this is TLS specific read hook so can't really
> work in generic case anyways.
>
> sk_psock_strp_read(), will you have UDP, AF_UNIX stream parsers? Do these
> even work outside TCP cases.
>
> For these ones: sk_psock_verdict_apply(), sk_psock_verdict_recv(),
> sk_psock_backlog(), can't we just do some refactoring around their
> hook points so we know the context. For example sk_psock_tls_verdict_apply
> is calling sk_psock_skb_redirect(). Why not have a sk_psock_unix_redirect()
> and a sk_psock_udp_redirect(). There are likely some optimizations we can
> deploy this way. We've already don this for tls and sk_msg types for example.
>
> Then the helpers will know their types by program type, just use the right
> variants.
>
> So not suggestiong if/else the checks so much as having per type hooks.
>

Hmm, but sk_psock_backlog() is still the only one that handles all three
above cases, right? It uses TCP_SKB_CB() too and more importantly it
is also why we can't use a per-cpu struct here (see bpf_redirect_info).

Thanks.