On Fri, 28 Jul 2023, Matthieu Baerts wrote:

> Hi Mat,
>
> On 28/07/2023 03:30, Mat Martineau wrote:
>> On Tue, 18 Jul 2023, Geliang Tang wrote:
>>
>>> Add mptcp_subflow_set_stale() helper.
>>>
>>
>> I was trying to remember the use case for this. The closest thing I can
>> find is this:
>>
>>     => the scheduler should be able to interact with the "stale" logic
>>        we have in the upstream kernel: mark a subflow as stale / back
>>        active.
>>
>> from
>> https://lore.kernel.org/mptcp/dd4363e7-d057-97e8-0b5f-8570f39aa538@tessares.net/
>>
>> I think that need is met by the change in patch 11 to export
>> mptcp_pm_subflow_chk_stale - if a particular scheduler wants to track
>> some stale-like metric, it can do so with its own sk_storage values.
>> There's also the matter of coordinating the values of stale_count and
>> stale_rcv_tstamp.
>>
>> My suggestion is to drop patches 8, 9, and 10 unless there's a
>> compelling reason to allow a BPF scheduler to control this bit. If
>> there's a need for it please speak up!
>
> If I'm not mistaken, the 'stale' logic is currently used by the core not
> to use a subflow (when calling mptcp_subflow_active()). A packet
> scheduler might want to proactively mark a subflow as stall, e.g. if the
> latency is too high, too many retransmissions, etc. and just set the
> subflow as stale.
>
> But maybe we cannot do that because the core will reset stale in some
> conditions? Maybe we need something similar but not reusing "stale"?
>

That's what I'm thinking, yes. Individual BPF schedulers could create 
their own "skip this subflow if possible" variables in their custom 
sk_storage structs, and those would not have the complex interactions with 
stale_count and stale_rcv_tstamp.


- Mat