Re: [PATCH bpf-next] mptcp: Add struct mptcp_sock definition when CONFIG_MPTCP is disabled

[Mat Martineau <mathew.j.martineau@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 2973 bytes --]

On Tue, 12 Jul 2022, Matthieu Baerts wrote:

> Hi Jiri,
>
> On 12/07/2022 14:12, Jiri Olsa wrote:
>> On Tue, Jul 12, 2022 at 11:06:38AM +0200, Matthieu Baerts wrote:
>>> Hi Jiri, Mat,
>>>
>>> On 11/07/2022 23:21, Mat Martineau wrote:
>>>> On Mon, 11 Jul 2022, Jiri Olsa wrote:
>>>>
>>>>> The btf_sock_ids array needs struct mptcp_sock BTF ID for
>>>>> the bpf_skc_to_mptcp_sock helper.
>>>>>
>>>>> When CONFIG_MPTCP is disabled, the 'struct mptcp_sock' is not
>>>>> defined and resolve_btfids will complain with:
>>>>>
>>>>>  BTFIDS  vmlinux
>>>>> WARN: resolve_btfids: unresolved symbol mptcp_sock
>>>>>
>>>>> Adding empty difinition for struct mptcp_sock when CONFIG_MPTCP
>>>>> is disabled.
>>>>>
>>>>> Signed-off-by: Jiri Olsa <jolsa@kernel.org>
>>>>> ---
>>>>> include/net/mptcp.h | 4 ++++
>>>>> 1 file changed, 4 insertions(+)
>>>>>
>>>>> diff --git a/include/net/mptcp.h b/include/net/mptcp.h
>>>>> index ac9cf7271d46..25741a52c666 100644
>>>>> --- a/include/net/mptcp.h
>>>>> +++ b/include/net/mptcp.h
>>>>> @@ -59,6 +59,10 @@ struct mptcp_addr_info {
>>>>>     };
>>>>> };
>>>>>
>>>>> +#if !IS_ENABLED(CONFIG_MPTCP)
>>>>> +struct mptcp_sock { };
>>>>> +#endif
>>>>
>>>> The only use of struct mptcp_sock I see with !CONFIG_MPTCP is from this
>>>> stub at the end of mptcp.h:
>>>>
>>>> static inline struct mptcp_sock *bpf_mptcp_sock_from_subflow(struct sock
>>>> *sk) { return NULL; }
>>>>
>>>> It's normally defined in net/mptcp/protocol.h for the MPTCP subsystem code.
>>>>
>>>> The conditional could be added on the line before the stub to make it
>>>> clear that the empty struct is associated with that inline stub.
>>>
>>> If this is required only for this specific BPF function, why not
>>> modifying this stub (or add a define) to return "void *" instead of
>>> "struct mptcp_sock *"?
>>
>> so btf_sock_ids array needs BTF ID for 'struct mptcp_sock' and if CONFIG_MPTCP
>> is not enabled, then resolve_btfids (which resolves and populate all BTF IDs)
>> won't find it and will complain
>>
>> btf_sock_ids keeps all socket IDs regardles the state of their CONFIG options,
>> and relies that sock structs are defined even if related CONFIG option is disabled
>
> Thank you for the explanation. I didn't know about that.
>
> Then it is fine for me to leave it in mptcp.h. If it is not directly
> linked to bpf_mptcp_sock_from_subflow(), I guess it can stay there but
> maybe better to wait for Mat's answer about that.
>
>> if that is false assumption then maybe we need to make btf_sock_ids values optional
>> somehow
>

I'd rather keep the full mptcp_sock definition in net/mptcp/protocol.h 
since moving it would require also moving a few other structs it depends 
on.

Defining the empty struct in mptcp.h is fine with me and it sounds like 
that meets the needs of btf_sock_ids - but I'd like a v2 of this patch 
that moves the new empty struct declaration next to the inline 
btf_mptcp_sock_from_subflow function in mptcp.h

--
Mat Martineau
Intel