Re: [PATCH v4,bpf-next] bpf: Don't redirect packets with invalid pkt_len

[Stanislav Fomichev <sdf@google.com>]

"

On Thu, Nov 3, 2022 at 3:58 PM Stanislav Fomichev <sdf@google.com> wrote:
>
> On Thu, Nov 3, 2022 at 3:42 PM Martin KaFai Lau <martin.lau@linux.dev> wrote:
> >
> > On 11/3/22 2:36 PM, Stanislav Fomichev wrote:
> > > On Thu, Nov 3, 2022 at 2:07 PM Martin KaFai Lau <martin.lau@linux.dev> wrote:
> > >>
> > >> On 7/15/22 4:55 AM, Zhengchao Shao wrote:
> > >>> Syzbot found an issue [1]: fq_codel_drop() try to drop a flow whitout any
> > >>> skbs, that is, the flow->head is null.
> > >>> The root cause, as the [2] says, is because that bpf_prog_test_run_skb()
> > >>> run a bpf prog which redirects empty skbs.
> > >>> So we should determine whether the length of the packet modified by bpf
> > >>> prog or others like bpf_prog_test is valid before forwarding it directly.
> > >>>
> > >>> LINK: [1] https://syzkaller.appspot.com/bug?id=0b84da80c2917757915afa89f7738a9d16ec96c5
> > >>> LINK: [2] https://www.spinics.net/lists/netdev/msg777503.html
> > >>>
> > >>> Reported-by: syzbot+7a12909485b94426aceb@syzkaller.appspotmail.com
> > >>> Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
> > >>> ---
> > >>> v3: modify debug print
> > >>> v2: need move checking to convert___skb_to_skb and add debug info
> > >>> v1: should not check len in fast path
> > >>>
> > >>>    include/linux/skbuff.h | 8 ++++++++
> > >>>    net/bpf/test_run.c     | 3 +++
> > >>>    net/core/dev.c         | 1 +
> > >>>    3 files changed, 12 insertions(+)
> > >>>
> > >>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> > >>> index f6a27ab19202..82e8368ba6e6 100644
> > >>> --- a/include/linux/skbuff.h
> > >>> +++ b/include/linux/skbuff.h
> > >>> @@ -2459,6 +2459,14 @@ static inline void skb_set_tail_pointer(struct sk_buff *skb, const int offset)
> > >>>
> > >>>    #endif /* NET_SKBUFF_DATA_USES_OFFSET */
> > >>>
> > >>> +static inline void skb_assert_len(struct sk_buff *skb)
> > >>> +{
> > >>> +#ifdef CONFIG_DEBUG_NET
> > >>> +     if (WARN_ONCE(!skb->len, "%s\n", __func__))
> > >>> +             DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
> > >>> +#endif /* CONFIG_DEBUG_NET */
> > >>> +}
> > >>> +
> > >>>    /*
> > >>>     *  Add data to an sk_buff
> > >>>     */
> > >>> diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
> > >>> index 2ca96acbc50a..dc9dc0bedca0 100644
> > >>> --- a/net/bpf/test_run.c
> > >>> +++ b/net/bpf/test_run.c
> > >>> @@ -955,6 +955,9 @@ static int convert___skb_to_skb(struct sk_buff *skb, struct __sk_buff *__skb)
> > >>>    {
> > >>>        struct qdisc_skb_cb *cb = (struct qdisc_skb_cb *)skb->cb;
> > >>>
> > >>> +     if (!skb->len)
> > >>> +             return -EINVAL;
> > >>
> > >>   From another recent report [0], I don't think this change is fixing the report
> > >> from syzbot.  It probably makes sense to revert this patch.
> > >>
> > >> afaict, This '!skb->len' test is done after
> > >>          if (is_l2)
> > >>                  __skb_push(skb, hh_len);
> > >>
> > >> Hence, skb->len is not zero in convert___skb_to_skb().  The proper place to test
> > >> skb->len is before __skb_push() to ensure there is some network header after the
> > >> mac or may as well ensure "data_size_in > ETH_HLEN" at the beginning.
> > >
> > > When is_l2==true, __skb_push will result in non-zero skb->len, so we
> > > should be good, right?
> > > The only issue is when we do bpf_redirect into a tunneling device and
> > > do __skb_pull, but that's now fixed by [0].
> > >
> > > When is_l2==false, the existing check in convert___skb_to_skb will
> > > make sure there is something in the l3 headers.
> > >
> > > So it seems like this patch is still needed. Or am I missing something?
> >
> > Replied in [0].  I think a small change in [0] will make this patch obsolete.
> >
> > My thinking is the !skb->len test in this patch does not address all cases, at
> > least not the most common one (the sch_cls prog where is_l2 == true) and then it
> > needs another change in __bpf_redirect_no_mac [0].  Then, instead of breaking
> > the existing test cases,  may as well solely depend on the change in
> > __bpf_redirect_no_mac which seems to be the only redirect function that does not
> > have the len check now.
>
> Removing this check in convert___skb_to_skb and moving the new one in
> __bpf_redirect_no_mac out of (mlen) SGTM.
> Can follow up unless you or Zhengchao prefer to do it.
> There were some concerns about doing this len check at runtime per
> packet, but not sure whether it really affects anything..

I've implemented a simple selftest for both mac/no-mac cases, and I
feel like this explicit len==0 has to happen in both cases.
That "skb->mac_header >= skb->network_header" check is not enough :-(
I'll send the series early next week after another round of xdp metadata..

> > >> The fix in [0] is applied.  If it turns out there are other cases caused by the
> > >> skb generated by test_run that needs extra fixes in bpf_redirect_*,  it needs to
> > >> revisit an earlier !skb->len check mentioned above and the existing test cases
> > >> outside of test_progs would have to adjust accordingly.
> > >>
> > >> [0]: https://lore.kernel.org/bpf/20221027225537.353077-1-sdf@google.com/
> > >>
> > >>> +
> > >>>        if (!__skb)
> > >>>                return 0;
> > >>>
> > >>> diff --git a/net/core/dev.c b/net/core/dev.c
> > >>> index d588fd0a54ce..716df64fcfa5 100644
> > >>> --- a/net/core/dev.c
> > >>> +++ b/net/core/dev.c
> > >>> @@ -4168,6 +4168,7 @@ int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
> > >>>        bool again = false;
> > >>>
> > >>>        skb_reset_mac_header(skb);
> > >>> +     skb_assert_len(skb);
> > >>>
> > >>>        if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
> > >>>                __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
> > >>
> >