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

-----Original Message-----
From: Richard Cochran [mailto:richardcochran@gmail.com]
Sent: Thursday, May 10, 2012 10:15 PM
To: Keller, Jacob E
Cc: Kirsher, Jeffrey T; davem@davemloft.net; <redacted>;
gospo@redhat.com; sassmann@redhat.com
Subject: Re: [net-next 06/12] ixgbe: Hardware Timestamping + PTP Hardware
Clock (PHC)

On Thu, May 10, 2012 at 09:53:18PM +0000, Keller, Jacob E wrote:

quoted

quoted

quoted

+	/*
+	 * If this bit is set, then the RX registers contain the time

stamp. No

quoted

quoted

quoted

+	 * other packet will be time stamped until we read these

registers, so

quoted

quoted

quoted

+	 * read the registers to make them available again. Because only

one

quoted

quoted

quoted

+	 * packet can be time stamped at a time, we know that the

register

quoted

quoted

quoted

+	 * values must belong to this one here and therefore we don't

need to

quoted

quoted

quoted

+	 * compare any of the additional attributes stored for it.

I suspect that this assumption is wrong. What happens if the time
stamping logic locks a value but the packet is lost because the ring is

full?

quoted

quoted

BTW, the IGB driver also has this defect.

Note how I read the rx registers first? So it will always clear the value.
That should unlock the value for the next rx stamp packet.

1. Hw recognizes ptp event packet, locks time stamp 2. Hw drops packet because
queue is full 3. No more time stamps are ever generated

Can this happen? The docs seems to say it can.

Richard

It might be possible. I'm curious what's the best method to solve this. I don't
think this generation of hardware can get an interrupt when the RX timestamp is
locked, and I don't know how easy it would be to process the sequence number within
the driver. It might be the case that the next packet would still be marked in the
descriptor and therefor the previous timestamp would apply to the next packet. I am
not sure, but I know we haven't seen issues where timestamps drop completely. This
is something I am not sure what's the best solution.

- Jake