Re: [PATCH net-next v2 0/2] net: dsa: mv88e6xxx: various hwstamp fixes

[Vladimir Oltean <olteanv@gmail.com>]

On Sun, Jul 19, 2026 at 09:22:39PM +1000, Luke Howard wrote:
> Hi Vladimir,
> 
> > ocelot_ptp_rx_timestamp() accesses MMIO-based registers, which can be
> > done atomically.
> > mv88e6xxx_ptp_clock_read() accesses MDIO bus registers, and the MDIO bus
> > is sleepable. Fundamental difference.
> > 
> > Your hardware only provides 32 bits of partial timestamp, so
> > mv88e6xxx_ptp_clock_read() will always be needed one way or another, to
> > recover the full 64 bits. Either through tstamp_{cc,tc} or through
> > direct calls.
> 
> This still happens from overflow_work().

Ok. My mistake.

> >> Deferring to the worker can reorder frames such that PTP general
> >> messages arrive before the timestamped event messages, which confuses
> >> some other PTP implementations such as gptp2d [1].
> > 
> > True, this is a caveat, but event messages and general messages can
> > already take different network paths, especially with PTP over IP where
> > they go through different UDP ports (even if for gPTP that is not the case).
> > The PTP user space implementation needs to be prepared to handle this.
> 
> Good point. So perhaps processing the embedded timestamp inline
> doesn’t confer much benefit. ptp4l (which we use) handles out-of-order
> messages fine.

If you're intending the in-band timestamps as a fix for gptp2d's
limitations, then no. As a general optimization - maybe.

> >> This optimisation of course only works for ArrTSMode because there is
> >> no MDIO read required.
> > 
> > I don't understand this comment given the partial 32-bit timestamp
> > limitation.
> 
> Better phrased as no MDIO read to recover the arrival timestamp.

Yeah, the timecounter/cyclecounter requires refreshes at least once
every 32-bit wraparound/2 time in order for the trick to work. But since
that is set up, it frees up the need to access MDIO per packet, since it
is guaranteed that cycle_now - tc->cycle_last is no larger than half the
wraparound time, which permits distinguishing timestamps taken by
hardware before tc->cycle_last from those taken after tc->cycle_last,
and correctly converting both to a valid 64-bit time base. Quite clever,
actually, I didn't fully understand the first time.

The only problem with timecounter/cyclecounter is that the hardware
clock remains free-running, which means you won't be able to apply a
time-synchronized taprio schedule (if the hw supports that).