Re: [PATCH net-next] net: ethtool: allow MAC drivers to override ethtool get_ts_info

[Russell King - ARM Linux admin <linux@armlinux.org.uk>]

On Thu, Jan 14, 2021 at 01:22:17PM +0000, Russell King - ARM Linux admin wrote:
> On Thu, Jan 14, 2021 at 04:55:06AM -0800, Richard Cochran wrote:
> > On Sun, Jan 10, 2021 at 11:13:44AM +0000, Russell King wrote:
> > > This allows network drivers such as mvpp2 to use their more accurate
> > > timestamping implementation than using a less accurate implementation
> > > in the PHY. Network drivers can opt to defer to phylib by returning
> > > -EOPNOTSUPP.
> > 
> > My expectation is that PHY time stamping is more accurate than MAC
> > time stamping.
> 
> PHY time stamping may be a "more accurate" location to get timestamps,
> in terms of the hardware, but when you consider the entire setup, that
> is not necessarily the case.
> 
> > > This change will be needed if the Marvell PHY drivers add support for
> > > PTP.
> > 
> > Huh? The mvpp2 appears to be a MAC.  If this device has integrated
> > PHYs then I don't see the issue.  If your board has the nvpp2 device
> > with the dp83640 PHYTER, then don't you want to actually use the
> > PHYTER?
> 
> You seem to be adding more information way beyond what I'm saying.
> 
> No, there aren't integrated PHYs. There's an external PHY - a Marvell
> 88e151x which has what I would call rudimentary stamping abilities,
> whereas the mvpp2 has advanced stamping abilities.
> 
> You implemented the Marvell timestamping in DSA, so you know what the
> Marvell offering there looks like and what it is capable of. That same
> hardware appears in some Marvell PHYs.
> 
> The mvpp2 hardware (which has support already merged after you acked
> the TAI part, and failed to provide comments on the mvpp2 part - so
> davem gave up waiting) is capable of:
> - stamping every received packet irrespective of its type
> - stamping any transmitted packet or only those we wish to stamp
> - inserting a timestamp into the packet (aka one-step, although that
>   isn't implemented that yet)
> - correcting the hardware time counter tick rate
> 
> There is considerable noise in reading the hardware timestamp counter
> from the PHYs - caused by latency over the MDIO bus, which makes the
> achievable accuracy lower. That noise is very much reduced when reading
> the hardware timestamp counter from mvpp2 - where we can implement
> mvpp22_tai_gettimex64(). Therefore, the achieved accuracy from mvpp2 is
> higher than from a PHY.
> 
> We had already discussed this patch last year, and you agreed with it
> then. What has changed?

See the discussion in this sub-thread:

https://lore.kernel.org/netdev/20200729105807.GZ1551@shell.armlinux.org.uk/

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!