Re: [PATCH net] net: phy: allow phy_detach() before netdev registration

[Chenguang Zhao <chenguang.zhao@linux.dev>]

On Thu, Sep 17, 2026 at 02:04:51PM +0200, Andrew Lunn wrote:
> On Thu, Sep 17, 2026 at 03:06:46PM +0800, Chenguang Zhao wrote:
> > On Wed, Sep 16, 2026 at 02:12:27PM +0200, Andrew Lunn wrote:
> > > On Wed, Sep 16, 2026 at 10:15:05AM +0800, Chenguang Zhao wrote:
> > > > From: Chenguang Zhao <zhaochenguang@kylinos.cn>
> > > > 
> > > > phylink_connect_phy() may be called from probe without RTNL while the
> > > > net_device is still NETREG_UNINITIALIZED. If PHY bring-up fails,
> > > > phy_detach() uses rtnl_dereference(dev->hwprov) and lockdep reports
> > > > suspicious RCU usage.
> > > 
> > > What do you mean by PHY bring-up?
> > > 
> > >      Andrew
> > Hi Andrew
> > 
> > PHY bring-up refers to phylink_bringup_phy(). The call chain is as follows:
> > 
> > usbnet_probe
> >   -> ax88772_bind
> >        -> ax88772_init_phy
> >             -> phylink_connect_phy
> >                  -> phylink_attach_phy / phy_attach_direct   // suceess 
> >                  -> phylink_bringup_phy                      // fail 
> >                  -> phy_detach                               // No RTNL
> >                       -> rtnl_dereference(dev->hwprov)       // lockdep warning
> > 
> > Because RTNL is not held when phy_detach() is called, rtnl_dereference() triggers a lockdep warning.
> 
> Lets zoom out and look at the big picture....
> 
> Why is phy_detach() not symmetric with phy_attach()?
> 
> Why is phy_detach() touching dev->hwprov, when phy_attach() does not?
> 
> Where is dev->hwprov set and why is the mirror function not dealing
> with this?
The dev->hwprov is set by ETHTOOL_MSG_TSCONFIG_SET in
ethnl_set_tsconfig(), which is the only writer besides phy_detach().

phy_detach() clears hwprov only when the selected provider is this PHY,
so skb_clone_tx_timestamp() / skb_defer_rx_timestamp() do not keep a
dangling phydev pointer after the PHY is gone. That is not the reverse
of phy_attach(); it is lifetime cleanup for a pointer that ethtool
stashed in the netdev.

Chenguang
Thanks
> 
>      Andrew