Re: [PATCH] udp6: Fix race condition in udp6_sendmsg & connect

[Eric Dumazet]

On Fri, May 26, 2023 at 6:09 PM Vlad Efanov <vefanov@xxxxxxxxx> wrote:
>
> Eric,
>
>
> udp6_sendmsg() currently still locks the socket (on line 1595).
>

Not really, look more closely at lines 1580 -> 1594


>
> Best regards,
>
> Vlad.
>
>
> On 26.05.2023 18:29, Eric Dumazet wrote:
> > On Fri, May 26, 2023 at 5:08 PM Vladislav Efanov <VEfanov@xxxxxxxxx> wrote:
> >> Syzkaller got the following report:
> >> BUG: KASAN: use-after-free in sk_setup_caps+0x621/0x690 net/core/sock.c:2018
> >> Read of size 8 at addr ffff888027f82780 by task syz-executor276/3255
> > Please include a full report.
> >
> >> The function sk_setup_caps (called by ip6_sk_dst_store_flow->
> >> ip6_dst_store) referenced already freed memory as this memory was
> >> freed by parallel task in udpv6_sendmsg->ip6_sk_dst_lookup_flow->
> >> sk_dst_check.
> >>
> >>            task1 (connect)              task2 (udp6_sendmsg)
> >>          sk_setup_caps->sk_dst_set |
> >>                                    |  sk_dst_check->
> >>                                    |      sk_dst_set
> >>                                    |      dst_release
> >>          sk_setup_caps references  |
> >>          to already freed dst_entry|
> >
> >> The reason for this race condition is: udp6_sendmsg() calls
> >> ip6_sk_dst_lookup() without lock for sock structure and tries to
> >> allocate/add dst_entry structure to sock structure in parallel with
> >> "connect" task.
> >>
> >> Found by Linux Verification Center (linuxtesting.org) with syzkaller.
> >>
> >> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> > This is a bogus Fixes: tag
> >
> > In old times, UDP sendmsg() was using the socket lock.
> >
> > Then, in linux-4.0 Vlad Yasevich made UDP v6 sendmsg() lockless (and
> > racy in many points)
> >
> >
> >> Signed-off-by: Vladislav Efanov <VEfanov@xxxxxxxxx>
> >> ---
> >>   net/ipv6/udp.c | 3 +++
> >>   1 file changed, 3 insertions(+)
> >>
> >> diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c
> >> index e5a337e6b970..a5ecd5d93b0a 100644
> >> --- a/net/ipv6/udp.c
> >> +++ b/net/ipv6/udp.c
> >> @@ -1563,12 +1563,15 @@ int udpv6_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
> >>
> >>          fl6->flowlabel = ip6_make_flowinfo(ipc6.tclass, fl6->flowlabel);
> >>
> >> +       lock_sock(sk);
> >>          dst = ip6_sk_dst_lookup_flow(sk, fl6, final_p, connected);
> >>          if (IS_ERR(dst)) {
> >>                  err = PTR_ERR(dst);
> >>                  dst = NULL;
> >> +               release_sock(sk);
> >>                  goto out;
> >>          }
> >> +       release_sock(sk);
> >>
> >>          if (ipc6.hlimit < 0)
> >>                  ipc6.hlimit = ip6_sk_dst_hoplimit(np, fl6, dst);
> >> --
> >> 2.34.1
> >>
> > There must be another way really.
> > You just killed UDP performance.