Re: [net-next 06/12] ixgbe: Hardware Timestamping + PTP Hardware Clock (PHC)

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

On 05/11/2012 10:34 PM, Richard Cochran wrote:
> On Fri, May 11, 2012 at 07:23:44PM +0000, Keller, Jacob E wrote:
>>
>>
>> I believe this very rare case might be possible, but I don't think
>> that checking the ptp seqid will fix anything. In normal cases,
>> hardware latches Rx packet timestamp, then the ptp packet goes into
>> the queue and we process it shortly after. Before we process that
>> packet there will never be another packet in the queue that needs a
>> timestamp. We know this because the hardware stops timestamping
>> until we unlatch the RX registers. This should mean we don't need to
>> check the sequence ID, and spending time doing it would never fix
>> the issue you are talking about.
>>
>> The issue is for when a packet is timestamped and then never reaches
>> the queue. Then the rx stamp registers are locked for good, because
>> we never clear them, and hardware would never timestamp another
>> receive packet. I don't know a good solution to this, except to
>> clear the registers periodically. Do you have any suggestions?
>
> Well, one solution would be to check every received packet with the
> BPF in ptp_classify.h (whenever Rx time stamping is enabled).
>
> When the driver finds an event packet in the Rx queue, and
> TSYNCRXCTL[RXTT] is set, it reads out the time stamp along with
> RXSATRL/H. If the fields match, then add the time stamp to the skb.
>
> [ Or perhaps instead of using RXSATRL/H, just use the descriptor bit.
>    If *not* set, then the time stamp does not belong to this packet. ]
>
> HTH,
> Richard

Ok, this sounds like a good plan. Considering that the device already 
doesn't allow timestamping of other types of packets, so it doesn't need 
to be general purpose.

Am I correct in thinking all I need to do is check the type and if it 
matches the currently configured rx timestamp mode, then double check 
the bit for whether a timestamp is available, and whether the descriptor 
had a timestamp bit enabled?

Thanks

- Jake