On Mon, 16 Oct 2023 17:00:27 +0200 Köry Maincent wrote:
> On Mon, 16 Oct 2023 07:22:04 -0700
> Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > > This is the main reason I changed this. This is Linux implementation
> > > purpose to know whether it should go through netdev or phylib, and then
> > > each of these drivers could use other timestamps which are hardware
> > > related.
> >
> > For an integrated design there's 90% chance the stamping is done
> > by the MAC. Even if it isn't there's no difference between PHY
> > and MAC in terms of quality.
>
> Ok, but there might be quality difference in case of several timestamp
> configuration done in the MAC. Like the timestamping precision vs frequency
> precision. In that case how ethtool would tell the driver to switch between
> them?
What's the reason for timestamp precision differences?
My understanding so far was the the differences come from:
1. different stamping device (i.e. separate "piece of silicon",
accessed over a different bus, with different PHC etc.)
2. different stamping point (MAC vs DMA)
I don't think any "integrated" device would support stamps which
differ under category 1.
> My solution could work for this case by simply adding new values to the enum:
>
> enum {
> NETDEV_TIMESTAMPING = (1 << 0),
> PHYLIB_TIMESTAMPING = (1 << 1),
> MAC_TS_PRECISION = (1 << 2)|(1 << 0),
> MAC_FREQ_PRECISION = (2 << 2)|(1 << 0),
> }
>
> Automatically Linux will go through the netdev implementation and could pass
> the enum value to the netdev driver.
We can add multiple fields to netlink. Why use the magic encoding?
> > But there is a big difference between MAC/PHY and DMA which would
> > both fall under NETDEV?
>
> Currently there is no DMA timestamping support right?
Kinda. Some devices pass DMA stamps as "HW stamps", and pretend they
are "good enough". But yes, there's no distinction at API level.
> And I suppose it fill fall under the net device management?
Yes.
> In that case we will have MAC and DMA under netdev and PHY under phylib and
> we won't have to do anything more than this timestamping management patch:
> https://lore.kernel.org/netdev/20231009155138.86458-14-kory.maincent@xxxxxxxxxxx/
Maybe we should start with a doc describing what APIs are at play,
what questions they answer, and what hard use cases we have.
I'm not opposed to the ethool API reporting just the differences
from my point 1. (in the first paragraph). But then we shouldn't
call that "layer", IMO, but device or source or such.
