On Fri Aug 22 2025, Sebastian Andrzej Siewior wrote:
> On 2025-08-22 09:28:10 [+0200], Kurt Kanzenbach wrote:
>> The current implementation uses schedule_work() which is executed by the
>> system work queue to retrieve Tx timestamps. This increases latency and can
>> lead to timeouts in case of heavy system load.
>> 
>> Therefore, switch to the PTP aux worker which can be prioritized and pinned
>> according to use case. Tested on Intel i210.
>> 
>> Signed-off-by: Kurt Kanzenbach <[email protected]>
>> ---
>> Changes in v2:
>> - Switch from IRQ to PTP aux worker due to NTP performance regression 
>> (Miroslav)
>> - Link to v1: 
>> https://lore.kernel.org/r/[email protected]
>
> For the i210 it makes sense to read it directly from IRQ avoiding the
> context switch and the delay resulting for it. For the e1000_82576 it
> makes sense to avoid the system workqueue and use a dedicated thread
> which is not CPU bound and could prioritized/ isolated further if
> needed.
> I don't understand *why* reading the TS in IRQ is causing this packet
> loss.

Me neither. I thought it could be the irqoff time. On my test systems
the TS IRQ takes about ~16us with reading the timestamp. In the
kworker/ptp aux thread scenario it takes about ~6us IRQ time + ~10us run
time for the threads. All of that looks reasonable to me.

Also I couldn't really see a performance degradation with ntpperf. In my
tests the IRQ variant reached an equal or higher rate. But sometimes I
get 'Could not send requests at rate X'. No idea what that means.

Anyway, this patch is basically a compromise. It works for Miroslav and
my use case.

> This is also what the igc does and the performance improved
>       afa141583d827 ("igc: Retrieve TX timestamp during interrupt handling")
>
> and here it causes the opposite?

As said above, I'm out of ideas here.

Thanks,
Kurt

Attachment: signature.asc
Description: PGP signature

Reply via email to