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;
MIME-Version: 1.0
References: <20220715115559.139691-1-shaozhengchao@huawei.com>
 <20220914111936.19881-1-oss@lmb.io> <CAKH8qBujKnFh8_g+npxHpo7RGFshus3N0iysmVBohTtG1X2yow@mail.gmail.com>
 <5a3c5ea9-d557-6070-d778-1092f3c51257@huawei.com> <aec8ef40-260c-4ded-b806-d381a3075ff0@www.fastmail.com>
 <c416473b-af8b-3bf6-7ede-e1198b3496f5@huawei.com>
In-Reply-To: <c416473b-af8b-3bf6-7ede-e1198b3496f5@huawei.com>
From:   Stanislav Fomichev <sdf@google.com>
Date:   Wed, 21 Sep 2022 13:59:39 -0700
Message-ID: <CAKH8qBtcRHU4UyUxSQe2tC0yopCaJ9jXugchuwDibAJc10FJww@mail.gmail.com>
Subject: Re: [PATCH v4,bpf-next] bpf: Don't redirect packets with invalid pkt_len
To:     shaozhengchao <shaozhengchao@huawei.com>
Cc:     Lorenz Bauer <oss@lmb.io>, ast@kernel.org, daniel@iogearbox.net,
        bpf@vger.kernel.org, linux-kernel@vger.kernel.org,
        yuehaibing@huawei.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Sep 21, 2022 at 1:48 AM shaozhengchao <shaozhengchao@huawei.com> wrote:
>
>
>
> On 2022/9/20 22:42, Lorenz Bauer wrote:
> > On Mon, 19 Sep 2022, at 11:55, shaozhengchao wrote:
> >> Sorry for the delay. I'm busy testing the TC module recently. I'm very
> >> sorry for the user-space breakage.
> >>
> >> The root cause of this problem is that eth_type_trans() is called when
> >> the protocol type of the SKB is parsed. The len value of the SKB is
> >> reduced to 0. If the user mode requires that the forwarding succeed, or
> >>    if the MAC header is added again after the MAC header is subtracted,
> >> is this appropriate?
> >
> > We don't require forwarding to succeed with a 14 byte input buffer. We also don't look at the MAC header.
> >
> > I think refusing to forward 0 length packets would be OK. Not 100% certain I understood you correctly, let me know if this helps.
> >
> > Best
> > Lorenz
> Hi Lorenz
>         Sorry. But how does the rejection of the 0 length affect the
> test case? Is the return value abnormal, send packet failure or some
> others?

Sorry for the late reply, I'm still traveling.
What we want to avoid is creating invalid skbs via prog_test_run.
We can reject <14 byte packets with ENIVAL, but that might break some
of the existing apps (as Lorenz reported).
So it seems like a safer alternative would be to accept those packets,
but just append missing bytes in the kernel to create the valid
packets? Padding with zeros seems fine?

> Zhengchao Shao