Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.35 as permitted sender) client-ip=23.128.96.35;
From:   =?utf-8?Q?Bj=C3=B8rn_Mork?= <bjorn@mork.no>
To:     Oliver Neukum <oneukum@suse.com>
Cc:     Ren Mingshuai <renmingshuai@huawei.com>, kuba@kernel.org,
        caowangbao@huawei.com, davem@davemloft.net, khlebnikov@openvz.org,
        liaichun@huawei.com, linux-kernel@vger.kernel.org,
        netdev@vger.kernel.org, yanan@huawei.com
Subject: Re: [PATCH] net: usbnet: Fix potential NULL pointer dereference
Organization: m
References: <20231101213832.77bd657b@kernel.org>
        <20231102090630.938759-1-renmingshuai@huawei.com>
        <80af8b7a-c543-4386-bb0c-a356189581a0@suse.com>
Date:   Mon, 06 Nov 2023 11:55:49 +0100
In-Reply-To: <80af8b7a-c543-4386-bb0c-a356189581a0@suse.com> (Oliver Neukum's
        message of "Mon, 6 Nov 2023 11:18:39 +0100")
Message-ID: <871qd3up56.fsf@miraculix.mork.no>
User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/28.2 (gnu/linux)
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Oliver Neukum <oneukum@suse.com> writes:

> yes it looks like NCM does funky things, but what does that mean?
>
> ndp_to_end_store()
>
>         /* flush pending data before changing flag */
>         netif_tx_lock_bh(dev->net);
>         usbnet_start_xmit(NULL, dev->net);
>         spin_lock_bh(&ctx->mtx);
>         if (enable)
>
> expects some odd semantics from it. The proposed patch simply
> increases the drop counter, which is by itself questionable, as
> we drop nothing.
>
> But it definitely does no IO, so we flush nothing.
> That is, we clearly have bug(s) but the patch only papers over
> them.
> And frankly, the basic question needs to be answered:
> Are you allowed to call ndo_start_xmit() with a NULL skb?
>
> My understanding until now was that you must not.

Yuck.  I see that I'm to blame for that code, so I've tried to figure
out what the idea behind it could possibly have been.

I believe that code is based on the (safe?) assumption that the struct
usbnet driver_info->tx_fixup points to cdc_ncm_tx_fixup().  And
cdc_ncm_tx_fixup does lots of weird stuff, including special handling of
NULL skb. It might return a valid skb for further processing by
usbnet_start_xmit().  If it doesn't, then we jump straight to
"not_drop", like we do when cdc_ncm_tx_fixup decides to eat the passed
skb.

But "funky" is i precise description of all this...  If someone feels
like it, then all that open coded skb queing inside cdc_ncm should be
completely rewritten.


Bj=C3=B8rn