Re: [PATCH 2/2] net: phy: Provide Module 4 KSZ9477 errata (DS80000754C)

[Oleksij Rempel]

On Wed, Aug 30, 2023 at 01:35:18PM +0100, Russell King (Oracle) wrote:
> On Wed, Aug 30, 2023 at 02:17:38PM +0200, Oleksij Rempel wrote:
> > On Wed, Aug 30, 2023 at 01:51:51PM +0200, Lukasz Majewski wrote:
> > > Hi Oleksij,
> > 
> > > It looks like the most optimal solution would be the one proposed by
> > > Tristam:
> > > https://www.spinics.net/lists/netdev/msg932044.html
> > 
> > In this case, please add the reason why it would work on this HW and
> > will not break by any changes in PHYlib or micrel.c driver.
> > 
> > If I remember it correctly, in KSZ9477 variants, if you write to EEE
> > advertisement register, it will affect the state of a EEE capability
> > register. Which break IEEE 802.3 specification and the reason why
> > ksz9477_get_features() actually exist. But can be used as workaround if
> > it is written early enough before PHYlib tried to read EEE capability
> > register.
> > 
> > Please confirm my assumption by applying your workaround and testing it
> > with ethtool --show-eee lanX.
> > 
> > It should be commented in the code with all kind of warnings:
> > Don't move!!! We use one bug to workaround another bug!!! If PHYlib
> > start scanning PHYs before this code is executed, then thing may break!!
> 
> Why would phylib's scanning cause breakage?
> 
> phylib's scanning for PHYs is about reading the ID registers etc. It
> doesn't do anything until the PHY has been found, and then the first
> thing that happens when the phy_device structure is created is an
> appropriate driver is located, and the driver's ->probe function
> is called.
> 
> If that is successful, then the fewatures are read. If the PHY
> driver's ->features member is set, then that initialises the
> "supported" mask and we read the EEE abilities.
> 
> If ->features is not set, then we look to see whether the driver
> provides a ->get_features method, and call that.
> 
> Otherwise we use the generic genphy_c45_pma_read_abilities() or
> genphy_read_abilities() depending whether the PHY's is_c45 is set
> or not.
> 
> So, if you want to do something very early before features are read,
> then either don't set .features, and do it early in .get_features
> before calling anything else, or do it in the ->probe function.

Let me summarize my view on the problem, so may be you can suggest a better
way to solve it.
- KSZ9477, KSZ8565, KSZ9893, KSZ9563, seems to have different quirks by
  the same PHYid. micrel.c driver do now know what exact HW is actually
  in use.
- A set of PHY workarounds was moved from dsa/microchip/ksz9477.c to
  micrel.c, one of this workaround was clearing EEE advertisement
  register, which by accident was clearing EEE capability register.
  Since EEE cap was cleared by the dsa/microchip/ksz9477.c code before
  micrel.c was probed, PHYlib was assuming that his PHY do not supports
  EEE and dint tried to use it.
  After moving this code to micrel.c, it is now trying to change EEE
  advertisement state without letting PHYlib to know about it and PHYlib
  re enables it as actually excepted.
- so far, only KSZ9477 seems to be broken beyond repair, so it is better
  to disable EEE without giving it as a choice for user configuration.

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |