Re: BPF selftest: mptcp subtest failing

[Matthieu Baerts <matttbe@kernel.org>]

Hi Harshvardhan,

On 08/08/2025 09:00, Harshvardhan Jha wrote:
> Hi Matthieu,
> 
> On 07/08/25 7:51 PM, Matthieu Baerts wrote:
>> Hi Harshvardhan,
>>
>> On 07/08/2025 05:50, Harshvardhan Jha wrote:
>>> Hi there,
>>> I have explicitly disabled mptpcp by default on my custom kernel and
>>> this seems to be causing the test case to fail. Even after enabling
>>> mtpcp via sysctl command or adding an entry to /etc/sysctl.conf this
>>> fails. I don't think this test should be failing and should account for
>>> cases where mptcp has not been enabled by default?
>> It looks like the test is failing because it expects MPTCP to be enabled
>> by default. Or, said differently, it doesn't expect the kernel to be
>> modified without adapting the corresponding tests :)
>>
>>> This is the custom patch I had applied on the LTS v6.12.36 kernel and
>>> tested it:
>>>
>>> diff --git a/net/mptcp/ctrl.c b/net/mptcp/ctrl.c
>>> index dd595d9b5e50c..bdcc4136e92ef 100644
>>> --- a/net/mptcp/ctrl.c
>>> +++ b/net/mptcp/ctrl.c
>>> @@ -89,7 +89,7 @@ const char *mptcp_get_scheduler(const struct net *net)
>>>  
>>>  static void mptcp_pernet_set_defaults(struct mptcp_pernet *pernet)
>>>  {
>>> -	pernet->mptcp_enabled = 1;
>>> +	pernet->mptcp_enabled = 0;
>>>  	pernet->add_addr_timeout = TCP_RTO_MAX;
>>>  	pernet->blackhole_timeout = 3600;
>>>  	atomic_set(&pernet->active_disable_times, 0);
>> First, I have the same question as the one I asked to RedHat devs: do
>> you still need to keep MPTCP disabled by default? If I remember well, on
>> RHEL side, they started to do that when they backported MPTCP on a
>> previous stable version, as an experimental feature. They left it like
>> that later mostly for internal process reasons I think. But honestly,
>> today, it no longer makes sense to do that and annoys users: all other
>> Linux distributions enable MPTCP by default without patching the kernel
>> like you did.
> 
> We had observed issues with mptcpd daemon failing before when we had
> this enabled by default. The mtpcpd userspace fix is yet to be integrated.

If net.mptcp.enabled is set to 0 by default, I guess most mptcpd unit
tests will be skipped. If you have issues when MPTCP is enabled by
default, there might be real issues in mptcpd that would need to be
fixed (or more likely, RHEL devs didn't noticed most tests were skipped,
and there are probably some dependences missing for these tests).

But that's a different topic.

> However, shouldn't the testcase be robust enough to handle that scenario
> regardless?

It should not be needed: these tests run in a dedicated netns where it
expects the kernel to have MPTCP enabled by default. If this behaviour
is changed without adapting the tests, that's normal to have issues.
This is not an issue with the upstream kernel.

>> If you don't want to revert this patch, I guess you can modify the BPF
>> selftests in 'prog_tests/mptcp.c' to set 'sysctl net.mptcp.enabled=1' in
>> each netns created by the test. But again, not changing the default
>> kernel behaviour sounds like a better solution.
> 
> Even after changing /etc/sysctl.conf which is supposed to keep mptcp
> enabled across reboots this issue occurs.

/etc/sysctl.conf needs a userspace daemon to load it and apply the
changes. In these tests, dedicated netns are created, and this file is
not read. That's the whole purpose of using netns: not to have to handle
default config changed on the host, just use the default values from the
kernel. Plus, it is cleaner to use netns: no need to revert changes
after, tests can be executed in parallel without impacting others, etc.

> I agree with what you have stated, mptcp should be enabled by default
> and the userspace fix should be incorporated ideally, however I still
> believe that the test case shouldn't be giving a false negative as it is
> in this case.

I don't think these tests have to handle every possible modifications
done by a custom kernel. If you have to modify the behaviour, then you
also need to adapt everything related to that. That's why I don't
recommend to change the behaviour.

> The false negatives seem to be occurring since this commit I believe:
> https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=df8d3ba55b4fa1d6aed8449971ee50757cb0732f

I don't think so, this commit changes MPTCP selftests, not the BPF ones
you have issues with. I guess you wanted to refer to this commit:

  02d6a057c7be ("selftests/bpf: run mptcp in a dedicated netns")

> This does the opposite of you have mentioned in certain functions.

The MPTCP selftests are setting net.mptcp.enabled=1, because RHEL devs
added them a while ago, while it was making sense, and we lefted them
there, even after the refactoring you mentioned. But we should probably
remove them indeed, because it is not needed.
> I suppose adding all these lines back might do the trick:
> 
> ip netns exec $netns sysctl -q net.mptcp.enabled=1

For your issue, yes. But again, not changing the default kernel
behaviour sounds like a better solution.

Cheers,
Matt
-- 
Sponsored by the NGI0 Core fund.