From: Stas Sergeev <[email protected]>
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.
And if it is, all other drivers registering a fixed phy link update
function need to be adjusted to do the same thing.
On 02/04/15 17:51, David Miller wrote:
> From: Stas Sergeev <[email protected]>
> 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
03.04.2015 03:51, David Miller пишет:
> From: Stas Sergeev <[email protected]>
> 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.
Hello David, sorry for this being discussed in a different thread,
so here you have a few pointers to why is this needed:
https://lkml.org/lkml/2015/3/30/367
> And if it is, all other drivers registering a fixed phy link update
> function need to be adjusted to do the same thing.
Or, for example, get rid of the callback and add an API:
http://www.spinics.net/lists/netdev/msg323517.html
Then you get the patch that avoids all this call-back singing and dancing:
http://www.spinics.net/lists/netdev/msg323518.html
In any case, I posted 3 different implementations and hope
people will choose the best one (any choice is fine with me).
From: Florian Fainelli <[email protected]>
Date: Thu, 02 Apr 2015 18:03:53 -0700
> On 02/04/15 17:51, David Miller wrote:
>> From: Stas Sergeev <[email protected]>
>> 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.
Ok I'll apply this mvneta patch to net-next then, thanks.
>> 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.
I just seems insane to me that phy_disconnect() doesn't stop the
callbacks from running.
Fixed PHY seems to me to suffer from a lack of proper integration
into the PHY layer.
From: David Miller <[email protected]>
Date: Fri, 03 Apr 2015 15:04:10 -0400 (EDT)
> Ok I'll apply this mvneta patch to net-next then, thanks.
Nevermind I just saw in my patchwork queue that there is a v3
of this change set that adds some adjustments to the fixed
PHY interfaces.
I think I'll apply that series instead of this patch.