Re: [PATCH v1 net] tipc: Fix use-after-free of kernel socket in cleanup_bearer().

["Eric W. Biederman" <ebiederm@xmission.com>]

Kuniyuki Iwashima <kuniyu@amazon.com> writes:

> From: Eric Dumazet <edumazet@google.com>
> Date: Tue, 26 Nov 2024 11:53:07 +0100
>> On Tue, Nov 26, 2024 at 7:14 AM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
>> >
>> > syzkaller reported a use-after-free of kernel UDP socket in
>> > cleanup_bearer() without repro. [0][1]
>> >
>> > When bearer_disable() calls tipc_udp_disable(), cleanup of the kernel
>> > UDP socket is deferred by work calling cleanup_bearer().
>> >
>> > Since the cited commit, however, the socket's netns might not be alive
>> > when the work is executed, resulting in use-after-free.
>> >
>> > Let's hold netns for the kernel UDP socket when created.
>> >
>> > Note that we can't call get_net() before scheduling the work and call
>> > put_net() in cleanup_bearer() because bearer_disable() could be called
>> > from pernet_operations.exit():
>> >
>> >   tipc_exit_net
>> >   `- tipc_net_stop
>> >      `- tipc_bearer_stop
>> >         `- bearer_disable
>> >
>> > [0]:
>> > ref_tracker: net notrefcnt@000000009b3d1faf has 1/1 users at
>> >      sk_alloc+0x438/0x608
>> >      inet_create+0x4c8/0xcb0
>> >      __sock_create+0x350/0x6b8
>> >      sock_create_kern+0x58/0x78
>> >      udp_sock_create4+0x68/0x398
>> >      udp_sock_create+0x88/0xc8
>> >      tipc_udp_enable+0x5e8/0x848
>> >      __tipc_nl_bearer_enable+0x84c/0xed8
>> >      tipc_nl_bearer_enable+0x38/0x60
>> >      genl_family_rcv_msg_doit+0x170/0x248
>> >      genl_rcv_msg+0x400/0x5b0
>> >      netlink_rcv_skb+0x1dc/0x398
>> >      genl_rcv+0x44/0x68
>> >      netlink_unicast+0x678/0x8b0
>> >      netlink_sendmsg+0x5e4/0x898
>> >      ____sys_sendmsg+0x500/0x830
>> >
>> > [1]:
>> > BUG: KMSAN: use-after-free in udp_hashslot include/net/udp.h:85 [inline]
>> > BUG: KMSAN: use-after-free in udp_lib_unhash+0x3b8/0x930 net/ipv4/udp.c:1979
>> >  udp_hashslot include/net/udp.h:85 [inline]
>> >  udp_lib_unhash+0x3b8/0x930 net/ipv4/udp.c:1979
>> >  sk_common_release+0xaf/0x3f0 net/core/sock.c:3820
>> >  inet_release+0x1e0/0x260 net/ipv4/af_inet.c:437
>> >  inet6_release+0x6f/0xd0 net/ipv6/af_inet6.c:489
>> >  __sock_release net/socket.c:658 [inline]
>> >  sock_release+0xa0/0x210 net/socket.c:686
>> >  cleanup_bearer+0x42d/0x4c0 net/tipc/udp_media.c:819
>> >  process_one_work kernel/workqueue.c:3229 [inline]
>> >  process_scheduled_works+0xcaf/0x1c90 kernel/workqueue.c:3310
>> >  worker_thread+0xf6c/0x1510 kernel/workqueue.c:3391
>> >  kthread+0x531/0x6b0 kernel/kthread.c:389
>> >  ret_from_fork+0x60/0x80 arch/x86/kernel/process.c:147
>> >  ret_from_fork_asm+0x11/0x20 arch/x86/entry/entry_64.S:244
>> >
>> > Uninit was created at:
>> >  slab_free_hook mm/slub.c:2269 [inline]
>> >  slab_free mm/slub.c:4580 [inline]
>> >  kmem_cache_free+0x207/0xc40 mm/slub.c:4682
>> >  net_free net/core/net_namespace.c:454 [inline]
>> >  cleanup_net+0x16f2/0x19d0 net/core/net_namespace.c:647
>> >  process_one_work kernel/workqueue.c:3229 [inline]
>> >  process_scheduled_works+0xcaf/0x1c90 kernel/workqueue.c:3310
>> >  worker_thread+0xf6c/0x1510 kernel/workqueue.c:3391
>> >  kthread+0x531/0x6b0 kernel/kthread.c:389
>> >  ret_from_fork+0x60/0x80 arch/x86/kernel/process.c:147
>> >  ret_from_fork_asm+0x11/0x20 arch/x86/entry/entry_64.S:244
>> >
>> > CPU: 0 UID: 0 PID: 54 Comm: kworker/0:2 Not tainted 6.12.0-rc1-00131-gf66ebf37d69c #7 91723d6f74857f70725e1583cba3cf4adc716cfa
>> > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014
>> > Workqueue: events cleanup_bearer
>> >
>> > Fixes: 26abe14379f8 ("net: Modify sk_alloc to not reference count the netns of kernel sockets.")
>> > Reported-by: syzkaller <syzkaller@googlegroups.com>
>> > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
>> > ---
>> > I'll remove this ugly hack by clearner API in the next cycle.
>> > see:
>> > https://lore.kernel.org/netdev/20241112001308.58355-1-kuniyu@amazon.com/
>> > ---
>> >  net/tipc/udp_media.c | 7 +++++++
>> >  1 file changed, 7 insertions(+)
>> >
>> > diff --git a/net/tipc/udp_media.c b/net/tipc/udp_media.c
>> > index 439f75539977..10986b283ac8 100644
>> > --- a/net/tipc/udp_media.c
>> > +++ b/net/tipc/udp_media.c
>> > @@ -673,6 +673,7 @@ static int tipc_udp_enable(struct net *net, struct tipc_bearer *b,
>> >         struct nlattr *opts[TIPC_NLA_UDP_MAX + 1];
>> >         u8 node_id[NODE_ID_LEN] = {0,};
>> >         struct net_device *dev;
>> > +       struct sock *sk;
>> >         int rmcast = 0;
>> >
>> >         ub = kzalloc(sizeof(*ub), GFP_ATOMIC);
>> > @@ -792,6 +793,12 @@ static int tipc_udp_enable(struct net *net, struct tipc_bearer *b,
>> >         if (err)
>> >                 goto free;
>> >
>> > +       sk = ub->ubsock->sk;
>> > +       __netns_tracker_free(net, &sk->ns_tracker, false);
>> > +       sk->sk_net_refcnt = 1;
>> > +       get_net_track(net, &sk->ns_tracker, GFP_KERNEL);
>> > +       sock_inuse_add(net, 1);
>> > +
>> >         return 0;
>> 
>> I think 'kernel sockets' were not refcounted to allow the netns to be removed.
>> 
>> Otherwise, what would tipc_bearer_stop() be needed ?
>
> Interestingly, the delayed cleanup exists since the udp media support
> was added in d0f91938bede2, and it's 2 months earlier than 26abe14379f8
> that drops netns refcnt for kernel sockets.

Just for reference commit 26abe14379f8 ("net: Modify sk_alloc to not
reference count the netns of kernel sockets.") doesn't ``drop'' the
netns refcnt for kernel sockets.  It changes the code so the refcnt is
not taken.  You will see in that commit a bunch of sk_change_net calls
which if memory serves are where the refcnt was previously dropped.

> So I thought the udp bearer did not assume bearer_disable() was called
> from the __net_exit path, it could be simply wrong though.
>
> At least, the __net_exit path works for other media types.

For the most part.  The network filesystem has been seeing similar
issues lately.  I suspect there is something (maybe just syzkaller)
that is making old bugs more likely to appear lately.

> @Erik Hugne, do you remember any context above ?
>
>
>> 
>> tipc_exit_net(struct net *net)  // can only be called when all refcnt
>> have been released
>>  -> tipc_net_stop()
>>   -> tipc_bearer_stop()
>>     -> bearer_disable()
>>      -> tipc_udp_disable()
>>        -> INIT_WORK(&ub->work, cleanup_bearer); schedule_work(&ub->work);
>> 

That schedule_work definitely looks like it will start running after
the network namespace and probably the entire kernel socket
has been released.

Eric