Re: [PATCH mptcp-next v8 13/17] selftests/bpf: Add bpf_burst scheduler

[Geliang Tang <geliang.tang@suse.com>]

On Tue, Jun 13, 2023 at 03:36:32PM +0200, Paolo Abeni wrote:
> On Tue, 2023-06-13 at 20:32 +0800, Geliang Tang wrote:
> > On Mon, Jun 12, 2023 at 04:22:31PM +0200, Paolo Abeni wrote:
> > > """
> > > Basically any access to:
> > > 
> > > btf_type.array_field[valid_and_validated_variable_index]
> > > 
> > > is not allowed.
> > > """
> > > 
> > > The above should be:
> > > 
> > > 	for (int i=0; i < data->number_of_subflows; i++) {
> > > 		context = mptcp_subflow_ctx_by_pos(i);
> > > 		// ...
> > > 	}
> > 
> > "for (int i = 0; i < data->subflows; i++)" will get an error too:
> > 
> >   The sequence of 8193 jumps is too complex.
> 
> Oops, the joy of loop unrolling! :( I almost forgot about that.
> 
> I guess/hope/think that changing this definition:
> 
> struct mptcp_sched_data {
>  	struct sock	*last_snd;
>  	int		snd_burst;
>  	bool		reinject;
> 	int		subflows;
>  	struct mptcp_subflow_context *contexts[MPTCP_SUBFLOWS_MAX];
>  };
> 
> to:
> 
> struct mptcp_sched_data {
>  	struct sock	*last_snd;
>  	int		snd_burst;
>  	bool		reinject;
> 	s8		subflow;
>  	struct mptcp_subflow_context *contexts[MPTCP_SUBFLOWS_MAX];
>  };
> 
> should "address" the issue (basically 'subflow' should have the minimum
> size possible.
> 
> Or possibly even better, simply:
> 
> 	for (int i = 0; i < < MIN(data->subflows, MPTCP_SUBFLOWS_MAX); i++)
> 
> 
> Or you could try using the bpf_loop() helper, even if that *looks*
> slow.

"for (int i = 0; i < min(data->subflows, MPTCP_SUBFLOWS_MAX); i++)"
doesn't work but this works:

  for (int i = 0; i < data->subflows && i < MPTCP_SUBFLOWS_MAX; i++)

So I just updated the patch and sent it to ML, named "mptcp: make the
contexts array of sched_data invisible in BPF". Please give me some
suggestions about it.

Thanks,
-Geliang

> 
> 
> Cheers,
> 
> Paolo
>