Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.31 as permitted sender) client-ip=23.128.96.31;
MIME-Version: 1.0
References: <3793723.1694795079@warthog.procyon.org.uk> <CANn89iLwMhOnrmQTZJ+BqZJSbJZ+Q4W6xRknAAr+uSrk5TX-EQ@mail.gmail.com>
 <0000000000001c12b30605378ce8@google.com> <3905046.1695031382@warthog.procyon.org.uk>
 <65085768c17da_898cd294ae@willemb.c.googlers.com.notmuch> <CANn89iJ39Hguu6bRm2am6J_u0pSnm++ORa_UVpC0+8-mxORFfw@mail.gmail.com>
 <ZQlbYVCjCyuPotdX@katalix.com>
In-Reply-To: <ZQlbYVCjCyuPotdX@katalix.com>
From:   Eric Dumazet <edumazet@google.com>
Date:   Tue, 19 Sep 2023 11:04:41 +0200
Message-ID: <CANn89i+fPjwonpWAXOCzrCG+hYH4kwMgJidwZf0CQaoCPOffXA@mail.gmail.com>
Subject: Re: [syzbot] [net?] WARNING in __ip6_append_data
To:     Tom Parkin <tparkin@katalix.com>
Cc:     Willem de Bruijn <willemdebruijn.kernel@gmail.com>,
        David Howells <dhowells@redhat.com>,
        syzbot <syzbot+62cbf263225ae13ff153@syzkaller.appspotmail.com>,
        bpf@vger.kernel.org, davem@davemloft.net, dsahern@kernel.org,
        kuba@kernel.org, linux-kernel@vger.kernel.org,
        netdev@vger.kernel.org, pabeni@redhat.com,
        syzkaller-bugs@googlegroups.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Tue, Sep 19, 2023 at 10:27=E2=80=AFAM Tom Parkin <tparkin@katalix.com> w=
rote:
>
> On  Mon, Sep 18, 2023 at 16:04:49 +0200, Eric Dumazet wrote:
> > On Mon, Sep 18, 2023 at 3:58=E2=80=AFPM Willem de Bruijn
> > <willemdebruijn.kernel@gmail.com> wrote:
> > >
> > > David Howells wrote:
> > > > David Howells <dhowells@redhat.com> wrote:
> > > >
> > > > > I think the attached is probably an equivalent cleaned up reprodu=
cer.  Note
> > > > > that if the length given to sendfile() is less than 65536, it fai=
ls with
> > > > > EINVAL before it gets into __ip6_append_data().
> > > >
> > > > Actually, it only fails with EINVAL if the size is not a multiple o=
f the block
> > > > size of the source file because it's open O_DIRECT so, say, 65536-5=
12 is fine
> > > > (and works).
> > > >
> > > > But thinking more on this further, is this even a bug in my code, I=
 wonder?
> > > > The length passed is 65536 - but a UDP packet can't carry that, so =
it
> > > > shouldn't it have errored out before getting that far?  (which is w=
hat it
> > > > seems to do when I try it).
> > > >
> > > > I don't see how we get past the length check in ip6_append_data() w=
ith the
> > > > reproducer we're given unless the MTU is somewhat bigger than 65536=
 (is that
> > > > even possible?)
> > >
> > > An ipv6 packet can carry 64KB of payload, so maxnonfragsize of 65535 =
+ 40
> > > sounds correct. But payload length passed of 65536 is not (ignoring i=
pv6
> > > jumbograms). So that should probably trigger an EINVAL -- if that is =
indeed
> > > what the repro does.
> >
> > l2tp_ip6_sendmsg() claims ip6_append_data() can make better checks,
> > but what about simply replacing INT_MAX by 65535 ?
>
> Slightly OT but I think the l2tp_ip6.c approach was probably cribbed
> from net/ipv6/udp.c's udpv6_sendmsg originally:
>
>
>     /* Rough check on arithmetic overflow,
>        better check is made in ip6_append_data().
>        */
>     if (len > INT_MAX - sizeof(struct udphdr))
>         return -EMSGSIZE;
>
>
> Should the udp code be modified similarly?
>

Unfortunately both l2tp and udp support CORK (MSG_MORE),
so a modified check like that will not be enough to prevent syzbot reports.

Better than nothing, of course.

I also note that ipv4 size of l2tp does not have any check,
an overflow seems possible with carefully chosen size.