On Sat, Dec 14, 2024 at 8:14 AM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 12/13/24 8:02 AM, Jason Xing wrote:
> >>> +static u64 delay_tolerance_nsec = 5000000;
> >>
> >> If I count right, 5ms may not a lot for the bpf CI and the test could become
> >> flaky. Probably good enough to ensure the delay is larger than the previous one.
> >
> > You're right, initially I set 2ms which make the test flaky. How about
> > 20ms? We cannot ensure each delta (calculated between two tx points)
> > is larger than the previous one.
>
> or I was thinking the delay is always measured from sendmsg_ns.
>
> Regardless, whatever way the delay of a tx point is measured from (always from
> sendmsg_ns or from the previous tx point), it can also just check the measured
> delay is +ve or something like that instead of having a hard coded maximum delay
> here.
That makes things simpler. Got it.
>
> The following "struct delay_info" may not be the best. Feel free to adjust.
Okay.
>
> >> struct delay_info {
> >> u64 sendmsg_ns;
> >> u32 sched_delay; /* SCHED_OPT_CB - sendmsg_ns */
> >> u32 sw_snd_delay;
> >> u32 ack_delay;
> >> };
>
