Re: [Intel-wired-lan] Fwd: [PATCH] Use ixgbe_ptp_reset on linkup/linkdown for X550

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

On 8/1/2022 4:29 PM, Jacob Keller wrote:
> 
> 
> On 8/1/2022 4:00 PM, Ilya Evenbach wrote:
>>>> -----Original Message-----
>>>> From: achaiken@aurora.tech <achaiken@aurora.tech>
>>>> Sent: Monday, August 01, 2022 6:38 AM
>>>> To: Brandeburg, Jesse <jesse.brandeburg@intel.com>;
>>>> richardcochran@gmail.com
>>>> Cc: spayne@aurora.tech; achaiken@aurora.tech; alison@she-devel.com;
>>>> netdev@vger.kernel.org; intel-wired-lan@lists.osuosl.org
>>>> Subject: [PATCH] Use ixgbe_ptp_reset on linkup/linkdown for X550
>>>>
>>>> From: Steve Payne <spayne@aurora.tech>
>>>>
>>>> For an unknown reason, when `ixgbe_ptp_start_cyclecounter` is called
>>>> from `ixgbe_watchdog_link_is_down` the PHC on the NIC jumps backward
>>>> by a seemingly inconsistent amount, which causes discontinuities in
>>>> time synchronization. Explicitly reset the NIC's PHC to
>>>> `CLOCK_REALTIME` whenever the NIC goes up or down by calling
>>>> `ixgbe_ptp_reset` instead of the bare `ixgbe_ptp_start_cyclecounter`.
>>>>
>>>> Signed-off-by: Steve Payne <spayne@aurora.tech>
>>>> Signed-off-by: Alison Chaiken <achaiken@aurora.tech>
>>>>
>>>
>>> Resetting PTP could be a problem if the clock was not being synchronized with the kernel CLOCK_REALTIME,
>>
>> That is true, but most likely not really important, as the unmitigated
>> problem also introduces significant discontinuities in time.
>> Basically, this patch does not make things worse.
>>
> 
> Sure, but I am trying to see if I can understand *why* things get wonky.
> I suspect the issue is caused because of how we're resetting the
> cyclecounter.
> 
>>>
>>> and does result in some loss of timer precision either way due to the delays involved with setting the time.
>>
>>  That precision loss is negligible compared to jumps resulting from
>> link down/up, and should be corrected by normal PTP operation very
>> quickly.
>>
> 
> Only if CLOCK_REALTIME is actually being synchronized. Yes, that is
> generally true, but its not necessarily guaranteed.
> 
>>>
>>> Do you have an example of the clock jump? How much is it?
>>
>> 2021-02-12T09:24:37.741191+00:00 bench-12 phc2sys: [195230.451]
>> CLOCK_REALTIME phc offset        61 s2 freq  -36503 delay   2298
>> 2021-02-12T09:24:38.741315+00:00 bench-12 phc2sys: [195231.451]
>> CLOCK_REALTIME phc offset       169 s2 freq  -36377 delay   2294
>> 2021-02-12T09:24:39.741407+00:00 bench-12 phc2sys: [195232.451]
>> CLOCK_REALTIME phc offset 195213702387037 s2 freq +100000000 delay
>> 2301
>> 2021-02-12T09:24:40.741489+00:00 bench-12 phc2sys: [195233.452]
>> CLOCK_REALTIME phc offset 195213591220495 s2 freq +100000000 delay
>> 2081
>>
> 
> Thanks.
> 
> I think what's actually going on is a bug in the
> ixgbe_ptp_start_cyclecounter function where the system time registers
> are being reset.
> 
> What hardware are you operating on? Do you know if its an X550 board? It
> looks like this has been the case since a9763f3cb54c ("ixgbe: Update PTP
> to support X550EM_x devices").
> 
> The start_cyclecounter was never supposed to modify the current time
> registers, but resetting it to 0 as it does for X550 devices would give
> the exact behavior you're seeing.

I just posted an alternative fix which I believe resolves this issue.

Thanks,
Jake
_______________________________________________
Intel-wired-lan mailing list
Intel-wired-lan@osuosl.org
https://lists.osuosl.org/mailman/listinfo/intel-wired-lan