Re: [Intel-wired-lan] [PATCH iwl-next] igb: Retrieve Tx timestamp directly from interrupt

[Jacob Keller <jacob.e.keller@intel.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1988 bytes --]



On 8/18/2025 5:24 AM, Miroslav Lichvar wrote:
> On Fri, Aug 15, 2025 at 08:50:23AM +0200, Kurt Kanzenbach wrote:
>> Retrieve Tx timestamp directly from interrupt handler.
>>
>> 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, fetch the timestamp directly from the interrupt handler.
>>
>> The work queue code stays for the Intel 82576. Tested on Intel i210.
> 
> I tested this patch on 6.17-rc1 with an Intel I350 card on a NTP
> server (chrony 4.4), measuring packet rates and TX timestamp accuracy
> with ntpperf. While the HW TX timestamping seems more reliable at some
> lower request rates, there seems to be about 40% drop in the overall
> performance of the server in how much requests it can handle (falling
> back to SW timestamps when HW timestamp is missed). Is this expected
> or something to be considered? 
> 

Hm. The new steps do have to perform all the tasks for completing a Tx
timestamp within the interrupt, but that involves reading the register,
converting it to the appropriate format, clearing a bitlock, and
submitting the skb to the stack. I can't imagine those tasks being super
problematic within interrupt context..

The Tx timestamp interrupt occurs on the other IRQ which is used
primarily for non-traffic causes which aren't high volume (other than
the Tx interrupt itself). I suppose possibly freeing the SKB could be
something that is causing issues in interrupt context?

I'm having trouble interpreting what exactly this data shows, as its
quite a lot of data and numbers. I guess that it is showing when it
switches over to software timestamps.. It would be nice if ntpperf
showed number of events which were software vs hardware timestamping, as
thats likely the culprit. igb hardare only has a single outstanding Tx
timestamp at a time.

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 236 bytes --]