Re: [PATCH net-next 3/3] net: phy: dp83869: Support 1000Base-X SFP modules

[Romain Gantois <romain.gantois@bootlin.com>]

[-- Attachment #1: Type: text/plain, Size: 2463 bytes --]

On Wednesday, 14 May 2025 14:32:22 CEST Andrew Lunn wrote:
> > > > +	/* Update advertisement */
> > > > +	if (mutex_trylock(&phydev->lock)) {
> > > > +		ret = dp83869_config_aneg(phydev);
> > > > +		mutex_unlock(&phydev->lock);
> > > > +	}
> > > 
> > > Just skimmed through this quickly and it's not clear to me why aneg is
> > > restarted only if there was no contention on the global phydev lock;
> > > it's not guaranteed a concurrent holder would do the same. If this is
> > > intended, a comment would be welcomed.
> > 
> > The reasoning here is that there are code paths which call
> > dp83869_port_configure_serdes() with phydev->lock already held, for
> > example:
> > 
> > phy_start() -> sfp_upstream_start() -> sfp_start() -> \
> > 
> > 	sfp_sm_event() -> __sfp_sm_event() -> sfp_sm_module() -> \
> > 	sfp_module_insert() -> phy_sfp_module_insert() -> \
> > 	dp83869_port_configure_serdes()
> > 
> > so taking this lock could result in a deadlock.
> > 
> > mutex_trylock() is definitely not a perfect solution though, but I went
> > with it partly because the marvell-88x2222 driver already does it this
> > way, and partly because if phydev->lock() is held, then there's a solid
> > chance that the phy state machine is already taking care of reconfiguring
> > the advertisement. However, I'll admit that this is a bit of a shaky
> > argument.
> > 
> > If someone has a better solution in mind, I'll gladly hear it out, but for
> > now I guess I'll just add a comment explaining why trylock() is being
> > used.
> The marvell10g driver should be the reference to look at.
> 
> As you say, phy_start() will eventually get around to calling
> dp83869_config_aneg(). What is more interesting here are the paths
> which lead to this function which don't result in a call to
> dp83869_config_aneg(). What are those?

Whenever you insert an SFP module, either sfp_irq() or sfp_poll() will will 
detect it and eventually call module_insert() from the SFP state machine. As 
far as I'm aware, this doesn't imply any calls to config_aneg().

Since the DP83869 remaps fiber registers to 0 when we switch it to RGMII-
to-1000Base-X mode, the value in MII_ADVERTISE won't be correct anymore,
which is why we have to reconfigure the advertisement after a module is 
inserted. It doesn't seem like the marvell10g driver does this, and I'm not 
sure why that's the case.

Thanks,

-- 
Romain Gantois, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]