Date: Fri, 11 Aug 2023 11:50:17 -0700
From: Martin KaFai Lau <martin.lau@...ux.dev>
To: Geliang Tang <geliang.tang@...e.com>
Cc: Alexei Starovoitov <ast@...nel.org>,
 Daniel Borkmann <daniel@...earbox.net>, Andrii Nakryiko <andrii@...nel.org>,
 Song Liu <song@...nel.org>, Yonghong Song <yhs@...com>,
 John Fastabend <john.fastabend@...il.com>, KP Singh <kpsingh@...nel.org>,
 Stanislav Fomichev <sdf@...gle.com>, Hao Luo <haoluo@...gle.com>,
 Jiri Olsa <jolsa@...nel.org>, Florent Revest <revest@...omium.org>,
 Brendan Jackman <jackmanb@...omium.org>,
 Matthieu Baerts <matthieu.baerts@...sares.net>,
 Mat Martineau <martineau@...nel.org>, "David S. Miller"
 <davem@...emloft.net>, Eric Dumazet <edumazet@...gle.com>,
 Jakub Kicinski <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>,
 John Johansen <john.johansen@...onical.com>, Paul Moore
 <paul@...l-moore.com>, James Morris <jmorris@...ei.org>,
 "Serge E. Hallyn" <serge@...lyn.com>,
 Stephen Smalley <stephen.smalley.work@...il.com>,
 Eric Paris <eparis@...isplace.org>, Mykola Lysenko <mykolal@...com>,
 Shuah Khan <shuah@...nel.org>, Simon Horman <horms@...nel.org>,
 bpf@...r.kernel.org, netdev@...r.kernel.org, mptcp@...ts.linux.dev,
 apparmor@...ts.ubuntu.com, linux-security-module@...r.kernel.org,
 selinux@...r.kernel.org, linux-kselftest@...r.kernel.org
Subject: Re: [PATCH bpf-next v11 2/5] selftests/bpf: Use random netns name for
 mptcp

On 8/11/23 2:29 AM, Geliang Tang wrote:
> On Thu, Aug 10, 2023 at 10:53:38PM -0700, Martin KaFai Lau wrote:
>> On 8/9/23 1:19 AM, Geliang Tang wrote:
>>> On Tue, Aug 08, 2023 at 11:03:30PM -0700, Martin KaFai Lau wrote:
>>>> On 8/6/23 11:40 PM, Geliang Tang wrote:
>>>>> On Fri, Aug 04, 2023 at 05:23:32PM -0700, Martin KaFai Lau wrote:
>>>>>> On 8/3/23 10:07 PM, Geliang Tang wrote:
>>>>>>> Use rand() to generate a random netns name instead of using the fixed
>>>>>>> name "mptcp_ns" for every test.
>>>>>>>
>>>>>>> By doing that, we can re-launch the test even if there was an issue
>>>>>>> removing the previous netns or if by accident, a netns with this generic
>>>>>>> name already existed on the system.
>>>>>>>
>>>>>>> Note that using a different name each will also help adding more
>>>>>>> subtests in future commits.
>>>>>
>>>>> Hi Martin,
>>>>>
>>>>> I tried to run mptcp tests simultaneously, and got "Cannot create
>>>>> namespace file "/var/run/netns/mptcp_ns": File exists" errors sometimes.
>>>>> So I add this patch to fix it.
>>>>>
>>>>> It's easy to reproduce, just run this commands in multiple terminals:
>>>>>     > for i in `seq 1 100`; do sudo ./test_progs -t mptcp; done
>>>>
>>>> Not only the "-t mptcp" test. Other tests in test_progs also don't support
>>>> running parallel in multiple terminals. Does it really help to test the bpf
>>>> part of the prog_tests/mptcp.c test by running like this? If it wants to
>>>> exercise the other mptcp networking specific code like this, a separate
>>>> mptcp test is needed outside of test_progs and it won't be run in the bpf
>>>> CI.
>>>>
>>>> If you agree, can you please avoid introducing unnecessary randomness to the
>>>> test_progs where bpf CI and most users don't run in this way?
>>>
>>> Thanks Martin. Sure, I agree. Let's drop this patch.
>>
>> Thanks you.
>>
>>>> I have a high level question. In LPC 2022
>>>> (https://lpc.events/event/16/contributions/1354/), I recall there was idea
>>>> in using bpf to make other mptcp decision/policy. Any thought and progress
>>>> on this? This set which only uses bpf to change the protocol feels like an
>>>> incomplete solution.
>>>
>>> We are implementing MPTCP packet scheduler using BPF. Patches aren't
>>> sent to BPF mail list yet, only temporarily on our mptcp repo[1].
>>>
>>> Here are the patches:
>>>
>>>    selftests/bpf: Add bpf_burst test
>>>    selftests/bpf: Add bpf_burst scheduler
>>>    bpf: Export more bpf_burst related functions
>>>    selftests/bpf: Add bpf_red test
>>>    selftests/bpf: Add bpf_red scheduler
>>>    selftests/bpf: Add bpf_rr test
>>>    selftests/bpf: Add bpf_rr scheduler
>>>    selftests/bpf: Add bpf_bkup test
>>>    selftests/bpf: Add bpf_bkup scheduler
>>>    selftests/bpf: Add bpf_first test
>>>    selftests/bpf: Add bpf_first scheduler
>>>    selftests/bpf: Add bpf scheduler test
>>>    selftests/bpf: add two mptcp netns helpers
>>>    selftests/bpf: use random netns name for mptcp
>>>    selftests/bpf: Add mptcp sched structs
>>>    bpf: Add bpf_mptcp_sched_kfunc_set
>>>    bpf: Add bpf_mptcp_sched_ops
>>>
>>> If you could take a look at these patches in advance, I would greatly
>>> appreciate it. Any feedback is welcome.
>>>
>>> [1]
>>> https://github.com/multipath-tcp/mptcp_net-next.git
>>
>> Thanks for sharing. I did not go into the details. iiuc, the scheduler is
>> specific to a namespace. Do you see if it is useful to have more finer
>> control like depending on what IP address it is connected to? BPF policy is
>> usually found more useful to have finer policy control than global or
>> per-netns.
>>
>> The same question goes for the fmod_ret here in this patch. The
>> progs/mptcpify.c selftest is as good as upgrading all TCP connections. Is it
>> your only use case and no need for finer selection?
> 
> This per-netns control is just the first step. We do need finer selection. The
> most ideal mode is to select one app to upgrade it's TCP connections only. So
> per-cgroup control is much better than per-netns. But we haven't found a good
> per-cgroup solution yet.

Selecting an app or cgroup can sort of be done by getting the current task or 
current cgroup (there is helper to do that). I am imagining eventually it will 
want to decide the protocol upgrade and/or the mptcp-scheduler when the 
destination IP is decided. This fmod_ret upgrade for all acts like a global knob 
(sysctl) and feels like a hack or at least incomplete. However, I also don't see 
a clean way to do that for now in the current shape.

Please respin another revision to address the earlier selftest comment on the 
netns name. Thanks.

Powered by blists - more mailing lists