Re: [PATCH] mvneta: implement SGMII-based in-band link state signaling

[Florian Fainelli <f.fainelli@gmail.com>]

On 02/04/15 17:51, David Miller wrote:
> From: Stas Sergeev <stsp@list.ru>
> Date: Tue, 31 Mar 2015 16:24:59 +0300
> 
>> @@ -2590,6 +2651,7 @@ static int mvneta_mdio_probe(struct mvneta_port *pp)
>>
>>  static void mvneta_mdio_remove(struct mvneta_port *pp)
>>  {
>> +	fixed_phy_set_link_update(pp->phy_dev, NULL);
> 
> I do not see any other driver doing this on shutdown.
> Please show me why it is necessary.

The primary reason is that if you do not do that, past the point where
you call phy_disconnect(), we stop the PHY state machine, detach from
the net_device, such that it won't invoke the adjust_link callback
anymore. The fixed PHY driver, though will still keep calling the
fixed_link_update callback asking the driver whether the link parameters
need to be updated, and that will just cause a NULL pointer de-reference
phydev->attached_dev, since we are now in detached state.

I guess another way to fix that is to look for the PHY state in
fixed_mdio_read() and do nothing if it is PHY_HALTED.

> 
> And if it is, all other drivers registering a fixed phy link update
> function need to be adjusted to do the same thing.
> 

I think the bcmgenet driver is now doing this as a result of Petri's
latest changes, and I meant to comment on that before the patch got in.
drivers/net/dsa/bcm_sf2.c has a similar construct but does not invoke
phy_disconnect() nor can be rmmod'd, so a lesser issue.
-- 
Florian