Re: mptcp splat

[Martin KaFai Lau <martin.lau@linux.dev>]

On 3/27/24 11:50 AM, Alexei Starovoitov wrote:
> On Wed, Mar 27, 2024 at 11:33 AM Paolo Abeni <pabeni@redhat.com> wrote:
>>
>> On Wed, 2024-03-27 at 10:00 -0700, Alexei Starovoitov wrote:
>>> On Wed, Mar 27, 2024 at 9:56 AM Paolo Abeni <pabeni@redhat.com> wrote:
>>>>
>>>> On Wed, 2024-03-27 at 09:43 -0700, Alexei Starovoitov wrote:
>>>>> I ffwded bpf tree with the recent net fixes and caught this:
>>>>>
>>>>> [   48.386337] WARNING: CPU: 32 PID: 3276 at net/mptcp/subflow.c:1430
>>>>> subflow_data_ready+0x147/0x1c0
>>>>> [   48.392012] Modules linked in: dummy bpf_testmod(O) [last unloaded:
>>>>> bpf_test_no_cfi(O)]
>>>>> [   48.396609] CPU: 32 PID: 3276 Comm: test_progs Tainted: G
>>>>> O       6.8.0-12873-g2c43c33bfd23 #1014
>>>>> #[   48.467143] Call Trace:
>>>>> [   48.469094]  <TASK>
>>>>> [   48.472159]  ? __warn+0x80/0x180
>>>>> [   48.475019]  ? subflow_data_ready+0x147/0x1c0
>>>>> [   48.478068]  ? report_bug+0x189/0x1c0
>>>>> [   48.480725]  ? handle_bug+0x36/0x70
>>>>> [   48.483061]  ? exc_invalid_op+0x13/0x60
>>>>> [   48.485809]  ? asm_exc_invalid_op+0x16/0x20
>>>>> [   48.488754]  ? subflow_data_ready+0x147/0x1c0
>>>>> [   48.492159]  mptcp_set_rcvlowat+0x79/0x1d0
>>>>> [   48.495026]  sk_setsockopt+0x6c0/0x1540
>>>>>
>>>>> It doesn't reproduce all the time though.
>>>>> Some race?
>>>>> Known issue?
>>>>
>>>> It was not known to me. Looks like something related to not so recent
>>>> changes (rcvlowat support).
>>>>
>>>> Definitely looks lie a race.
>>>>
>>>> If you could share more info about the running context and/or a full
>>>> decoded splat it could help, thanks!
>>>
>>> This is just running bpf selftests in parallel:
>>> test_progs -j
>>>
>>> The end of the splat:
>>> [   48.500075]  __bpf_setsockopt+0x6f/0x90
>>> [   48.503124]  bpf_sock_ops_setsockopt+0x3c/0x90
>>> [   48.506053]  bpf_prog_509ce5db2c7f9981_bpf_test_sockopt_int+0xb4/0x11b
>>> [   48.510178]  bpf_prog_dce07e362d941d2b_bpf_test_socket_sockopt+0x12b/0x132
>>> [   48.515070]  bpf_prog_348c9b5faaf10092_skops_sockopt+0x954/0xe86
>>> [   48.519050]  __cgroup_bpf_run_filter_sock_ops+0xbc/0x250
>>> [   48.523836]  tcp_connect+0x879/0x1160
>>> [   48.527239]  ? ktime_get_with_offset+0x8d/0x140
>>> [   48.531362]  tcp_v6_connect+0x50c/0x870
>>> [   48.534609]  ? mptcp_connect+0x129/0x280
>>> [   48.538483]  mptcp_connect+0x129/0x280
>>> [   48.542436]  __inet_stream_connect+0xce/0x370
>>> [   48.546664]  ? rcu_is_watching+0xd/0x40
>>> [   48.549063]  ? lock_release+0x1c4/0x280
>>> [   48.553497]  ? inet_stream_connect+0x22/0x50
>>> [   48.557289]  ? rcu_is_watching+0xd/0x40
>>> [   48.560430]  inet_stream_connect+0x36/0x50
>>> [   48.563604]  bpf_trampoline_6442491565+0x49/0xef
>>> [   48.567770]  ? security_socket_connect+0x34/0x50
>>> [   48.575400]  inet_stream_connect+0x5/0x50
>>> [   48.577721]  __sys_connect+0x63/0x90
>>> [   48.580189]  ? bpf_trace_run2+0xb0/0x1a0
>>> [   48.583171]  ? rcu_is_watching+0xd/0x40
>>> [   48.585802]  ? syscall_trace_enter+0xfb/0x1e0
>>> [   48.588836]  __x64_sys_connect+0x14/0x20
>>
>> Ouch, it looks bad. BPF should not allow any action on mptcp subflows
>> that go through sk_socket. They touch the mptcp main socket, which is
>> _not_ protected by the subflow socket lock.
>>
>> AFICS currently the relevant set of racing sockopt allowed by bpf boils
>> down to SO_RCVLOWAT only - sk_setsockopt(SO_RCVLOWAT) will call sk-
>>> sk_socket->ops->set_rcvlowat()
>>
>> So something like the following (completely untested) should possibly
>> address the issue at hand, but I think it would be better/safer
>> completely disable ebpf on mptcp subflows, WDYT?
>>
>> Thanks,
>>
>> Paolo
>>
>> ---
>> diff --git a/net/mptcp/sockopt.c b/net/mptcp/sockopt.c
>> index dcd1c76d2a3b..6e5e64c2cf89 100644
>> --- a/net/mptcp/sockopt.c
>> +++ b/net/mptcp/sockopt.c
>> @@ -1493,6 +1493,9 @@ int mptcp_set_rcvlowat(struct sock *sk, int val)
>>          struct mptcp_subflow_context *subflow;
>>          int space, cap;
>>
>> +       if (has_current_bpf_ctx())
>> +               return -EINVAL;
>> +
> 
> Looks fine to me.
> 
> Martin,
> 
> Do you have any better ideas?
> 
> The splat explains the race.
> In this case setget_sockopt test happen to run in parallel
> with mptcp/bpf test and the former one was TCP connect request
> but it was for subflow.
> 
> We can disable that callback when tcp flow is a subflow,
> but that doesn't feel right.

I am also not sure if we can disable all set/getsockopt for tcp subflows. Not 
clear if there is use case that depends on this to setsockopt the subflow.

> I think Paolo's targeted fix is cleaner.

I will also go with Paolo's fix. The radius is smaller. I don't have a better idea.

Unrelated, is there a way to tell if a tcp_sock is a subflow? bpf prog can use 
it to decide if it wants to setsockopt on a subflow or not.