Re: [PATCH mptcp-next v8 0/8] BPF packet scheduler

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

On Wed, 30 Mar 2022, Geliang Tang wrote:

> Hi Mat & Matt,
>
> On Tue, Mar 29, 2022 at 04:13:02PM -0700, Mat Martineau wrote:
>>
>> On Tue, 29 Mar 2022, Geliang Tang wrote:
>>
>>> v8:
>>> - use global sched_list instead of pernet sched_list.
>>
>> Yes, I think this is fine. I had initially asked about pernet configuration
>> with respect to registering BPF schedulers, but the important thing is that
>> the sysctl can be set per-namespace.
>>
>>> - drop synchronize_rcu() in mptcp_unregister_scheduler().
>>> - update mptcp_init_sched and mptcp_release_sched as Mat and Florian
>>>   suggested.
>>> - fix the build break in patch 8.
>>> - depends on: "add skc_to_mptcp_sock" v14.
>>> - export/20220325T055307
>>
>> Thanks for updating. Builds and runs fine here.
>>
>> Before we add this to the export branch, have you thought about how subflow
>> data (including the backup bit and throughput/latency data used by the
>> default scheduler) can be accessed in the BPF get_subflow hook? Do you think
>> that can be cleanly added after this series, or is there anything that may
>> need to be changed in this series?
>
> I plan to do that in the next series, BPF round-robin scheduler.
>
> I had implemented round-robin in kernel before:
>
> https://patchwork.kernel.org/project/mptcp/cover/cover.1631011068.git.geliangtang@xiaomi.com/
>
> This time I plan to implement it using BPF.
>
> In order to support bpf_rr, I think we need to do two more things
> based on this series:
>
> 1. Get the subflow data from BPF, maybe iterate over the subflows
> from mptcp_sock.
>

Like we discussed in the meeting, I think a helper function that's 
callable from the BPF code could gather the necessary information from the 
subflows in a safe way.

> 2. Make same members of struct mptcp_sock writable in BPF, say
> last_snd, then we can set it like this:
>
>        msk->last_snd = ssk.

The kernel C code can record which subflow was used last based on the 
previous value returned by the scheduler, or a BPF scheduler could track 
data in a BPF map (I think - I'm not a BPF expert).

For something like last_snd to work, I think we need some way to identify 
subflows that's not the raw ssk pointer. Are there rules about pointer 
usage in BPF that we need to consider?

--
Mat Martineau
Intel