On Sat, 1 Jun 2019 at 13:31, Vladimir Oltean [off-list ref] wrote:

On Sat, 1 Jun 2019 at 08:07, Richard Cochran [off-list ref] wrote:

quoted

On Fri, May 31, 2019 at 09:12:03PM +0300, Vladimir Oltean wrote:

quoted

It won't work unless I make changes to dsa_switch_rcv.

Or to the tagging code.

quoted

Right now taggers can only return a pointer to the skb, or NULL, case
in which DSA will free it.

The tagger can re-write the skb.  Why not reform it into a PTP frame?
This clever trick is what the phyter does in hardware.  See dp83640.c.

I think you're missing the point here.
If I dress the meta frame into a PTP frame (btw is there any
preferable event message for this purpose?) then sure, I'll make
dsa_skb_defer_rx_timestamp call my .port_rxtstamp and I can e.g. move
my state machine there.
The problem is that in the current DSA structure, I'll still have less
timestampable frames waiting for a meta frame than meta frames
themselves. This is because not all frames that the switch takes an RX
timestamp for will make it to my .port_rxtstamp (in fact that is what
my .can_timestamp patch changes). I can put the timestampable frame in
a 1-entry wait queue which I'll deplete upon arrival of the first meta
frame, but when I get meta frames and the wait queue is empty, it can
mean multiple things: either DSA didn't care about this timestamp
(ok), or the timestampable frame got reordered or dropped by the MAC,
or what have you (not ok). So I can't exclude the possibility that the
meta frame was holding a relevant timestamp.
Sure, I can dress the meta frame into whatever the previous
MAC-trapped frame was (PTP or not) and then I'll have .port_rxtstamp
function see a 1-to-1 correspondence with meta frames in case
everything works fine. But then I'll have non-PTP meta frames leaking
up the stack...

Actually maybe this is exactly what you meant and I didn't think it through.
If RX timestamping is enabled, then I can just copy all MAC-trapped
frames to a private skb in a per-driver data structure, and have DSA
drop them.
Then when their meta frame arrives, I can just morph them into what
the previous frame was, just that now I'm also holding the partial
timestamp in skb->cb.
PTP frames will reconstruct the full timestamp without waiting for any
meta (they are the meta), while other MAC-trapped frames (STP etc)
will just carry a meaningless skb->cb when passed up the stack.
In retrospect, it would have been amazing if the switch gave me the
meta frames *before* the actual link-local frames that needed the
timestamp.

Thanks!
-Vladimir

Regards,
-Vladimir

quoted

Thanks,
Richard