Re: [PATCH v14 net-next 08/23] net/tcp: Add AO sign to RST packets

[Dmitry Safonov]

Hi Oliver,

On 10/18/23 08:32, Oliver Sang wrote:
> hi, Dmitry Safonov,
> 
> On Wed, Oct 18, 2023 at 12:19:56AM +0100, Dmitry Safonov wrote:
>> On 10/17/23 09:37, kernel test robot wrote:
>>>
>>>
>>> Hello,
>>>
>>> kernel test robot noticed "BUG:sleeping_function_called_from_invalid_context_at_net/core/sock.c" on:
>>>
>>> commit: df13d11e6a2a3cc5f973aca36f68f880fa42d55f ("[PATCH v14 net-next 08/23] net/tcp: Add AO sign to RST packets")
[..]
>>>
>>> [  221.348247][ T7133] BUG: sleeping function called from invalid context at net/core/sock.c:2978
>>> [  221.349875][ T7133] in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 7133, name: trinity-c4
>>> [  221.351666][ T7133] preempt_count: 0, expected: 0
>>> [  221.352614][ T7133] RCU nest depth: 1, expected: 0
>>> [  221.353518][ T7133] 2 locks held by trinity-c4/7133:
>>> [ 221.354530][ T7133] #0: ed8b5660 (sk_lock-AF_INET6){+.+.}-{0:0}, at: tcp_sendmsg (net/ipv4/tcp.c:1336) 
>>> [ 221.374314][ T7133] #1: c27dbb18 (rcu_read_lock){....}-{1:2}, at: inet6_csk_xmit (include/linux/rcupdate.h:747 net/ipv6/inet6_connection_sock.c:129) 
>>> [  221.375906][ T7133] CPU: 1 PID: 7133 Comm: trinity-c4 Tainted: G        W       TN 6.6.0-rc4-01105-gdf13d11e6a2a #1
>>
>> Quite puzzling. It doesn't seem that it can happen to inet6_csk_xmit():
>>
>> :	rcu_read_lock();
>> :	skb_dst_set_noref(skb, dst);
>> :
>> :	/* Restore final destination back after routing done */
>> :	fl6.daddr = sk->sk_v6_daddr;
>> :
>> :	res = ip6_xmit(sk, skb, &fl6, sk->sk_mark, rcu_dereference(np->opt),
>> :		       np->tclass,  sk->sk_priority);
>> :	rcu_read_unlock();
>>
>> So, I presumed the calltrace was for nested rcu_read_lock() case.
>> Then I've looked at all return/goto cases, I couldn't find any
>> unbalanced rcu_read_{,un}lock().
>>
>> Is this reproducible by any chance?
> 
> do you mean how often it could be reproduced?
> 
> we run the tests upon this commit and parent both up to 15 times.
> from below, parent shows quite clean, but there are various issues while
> running tests upon this patch.
Thanks a lot for running your tests over it!

That's what a fresh brain does:
I see that tcp_v6_send_reset() with neither CONFIG_TCP_MD5SIG nor
CONFIG_TCP_AO set does grab rcu_read_lock() [as I wasn't sure at that
moment that nothing but md5 key does rely on that being held], but the
exit is covered with ifdefs:

: #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
: out:
:        if (allocated_traffic_key)
:                kfree(key.traffic_key);
:        rcu_read_unlock();
: #endif

Seems to have been correct versions ago (version 6), but on refactoring
when I've moved ifdeffs to tcp_parse_auth_options() declaration to
pollute less code - missed this.

Thanks again for your testing and reporting!


> 8468a6f4f3143 net/tcp: Add tcp_parse_auth_options()
> df13d11e6a2a3 net/tcp: Add AO sign to RST packets
> 
> 8468a6f4f3143ba2 df13d11e6a2a3cc5f973aca36f6
> ---------------- ---------------------------
>        fail:runs  %reproduction    fail:runs
>            |             |             |
>            :15          87%          13:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_arch/x86/entry/common.c
>            :15          33%           5:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_fs/dcache.c
>            :15          13%           2:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_include/linux/percpu-rwsem.h
>            :15          60%           9:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_include/linux/sched/mm.h
>            :15           7%           1:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_kernel/locking/mutex.c
>            :15           7%           1:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_kernel/locking/rwsem.c
>            :15           7%           1:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_kernel/sched/completion.c
>            :15          40%           6:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_kernel/task_work.c
>            :15          33%           5:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_lib/iov_iter.c
>            :15          33%           5:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_lib/strncpy_from_user.c
>            :15          20%           3:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_lib/usercopy.c
>            :15          13%           2:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_mm/vmscan.c
>            :15         100%          15:15    dmesg.BUG:sleeping_function_called_from_invalid_context_at_net/core/sock.c
>            :15         100%          15:15    dmesg.EIP:rcu_note_context_switch
>            :15         100%          15:15    dmesg.WARNING:at_kernel/rcu/tree_plugin.h:#rcu_note_context_switch
>            :15         100%          15:15    dmesg.WARNING:lock_held_when_returning_to_user_space
>            :15         100%          15:15    dmesg.is_leaving_the_kernel_with_locks_still_held
> 
[..]

Thanks,
           Dmitry