[Linux-stm32] [PATCH net-next 1/2] dt-bindings: net: Add Siflower SF19A2890 GMAC

[Joseph Telaak]

> So i think we need a get out from backwards compatibility. We need to
> be able to break running boards with old DT blobs, in the future when
> we fix whatever is broken. Maybe the best way to do that it make the
> MAC glue driver and the DSA driver depend on BROKEN.

Before that, I measured from the switch's side. The switch driver sets
its RGMII delays on the CPU port in nominal 500 ps steps, so with the
MAC's delays at 0:

- Switch to MAC, with the MAC's RX delay line bypassed: the link works
  with the switch adding 0 to 1.5 ns, and fails from 2 ns.
- The same with the RX delay line on at step 0: at 1.5 ns from the
  switch, 4 more steps of the line still work, so step 0 adds at most
  about 0.25 ns over the bypass.
- MAC to switch, with the MAC's TX delay at 0: the link works with the
  switch adding 0 to 2 ns, and fails from 2.5 ns. The TX delay line has
  no bypass, so I can't check its step 0 the same way.

Measured against the MAC's delay lines, the switch's nominal 2 ns comes
out at about 2.4 ns in one direction and its 1.5 ns at about 1.3 ns in
the other, so its steps are only roughly 500 ps.

So on RX, 0 does mean about 0, and the receive direction still fails at
the standard 2 ns. That extra skew is outside the MAC's delay line: in
the switch's output, the board, or the MAC's input, which I can't tell
apart.

The DSA driver, mxl-gsw1xx, is already in mainline for the whole GSW1xx
family, so BROKEN there would also cover boards with other MACs. Would
BROKEN on the new glue alone do?