Re: [PATCH v2] bcm63xx_enet: fix internal phy IRQ assignment

[Heiner Kallweit <hkallweit1@gmail.com>]

On 25.02.2021 17:36, Daniel González Cabanelas wrote:
> El jue, 25 feb 2021 a las 8:22, Heiner Kallweit
> (<hkallweit1@gmail.com>) escribió:
>>
>> On 25.02.2021 00:54, Daniel González Cabanelas wrote:
>>> El mié, 24 feb 2021 a las 23:01, Florian Fainelli
>>> (<f.fainelli@gmail.com>) escribió:
>>>>
>>>>
>>>>
>>>> On 2/24/2021 1:44 PM, Heiner Kallweit wrote:
>>>>> On 24.02.2021 16:44, Daniel González Cabanelas wrote:
>>>>>> The current bcm63xx_enet driver doesn't asign the internal phy IRQ. As a
>>>>>> result of this it works in polling mode.
>>>>>>
>>>>>> Fix it using the phy_device structure to assign the platform IRQ.
>>>>>>
>>>>>> Tested under a BCM6348 board. Kernel dmesg before the patch:
>>>>>>    Broadcom BCM63XX (1) bcm63xx_enet-0:01: attached PHY driver [Broadcom
>>>>>>               BCM63XX (1)] (mii_bus:phy_addr=bcm63xx_enet-0:01, irq=POLL)
>>>>>>
>>>>>> After the patch:
>>>>>>    Broadcom BCM63XX (1) bcm63xx_enet-0:01: attached PHY driver [Broadcom
>>>>>>               BCM63XX (1)] (mii_bus:phy_addr=bcm63xx_enet-0:01, irq=17)
>>>>>>
>>>>>> Pluging and uplugging the ethernet cable now generates interrupts and the
>>>>>> PHY goes up and down as expected.
>>>>>>
>>>>>> Signed-off-by: Daniel González Cabanelas <dgcbueu@gmail.com>
>>>>>> ---
>>>>>> changes in V2:
>>>>>>   - snippet moved after the mdiobus registration
>>>>>>   - added missing brackets
>>>>>>
>>>>>>  drivers/net/ethernet/broadcom/bcm63xx_enet.c | 13 +++++++++++--
>>>>>>  1 file changed, 11 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/net/ethernet/broadcom/bcm63xx_enet.c b/drivers/net/ethernet/broadcom/bcm63xx_enet.c
>>>>>> index fd876721316..dd218722560 100644
>>>>>> --- a/drivers/net/ethernet/broadcom/bcm63xx_enet.c
>>>>>> +++ b/drivers/net/ethernet/broadcom/bcm63xx_enet.c
>>>>>> @@ -1818,10 +1818,19 @@ static int bcm_enet_probe(struct platform_device *pdev)
>>>>>>               * if a slave is not present on hw */
>>>>>>              bus->phy_mask = ~(1 << priv->phy_id);
>>>>>>
>>>>>> -            if (priv->has_phy_interrupt)
>>>>>> +            ret = mdiobus_register(bus);
>>>>>> +
>>>>>> +            if (priv->has_phy_interrupt) {
>>>>>> +                    phydev = mdiobus_get_phy(bus, priv->phy_id);
>>>>>> +                    if (!phydev) {
>>>>>> +                            dev_err(&dev->dev, "no PHY found\n");
>>>>>> +                            goto out_unregister_mdio;
>>>>>> +                    }
>>>>>> +
>>>>>>                      bus->irq[priv->phy_id] = priv->phy_interrupt;
>>>>>> +                    phydev->irq = priv->phy_interrupt;
>>>>>> +            }
>>>>>>
>>>>>> -            ret = mdiobus_register(bus);
>>>>>
>>>>> You shouldn't have to set phydev->irq, this is done by phy_device_create().
>>>>> For this to work bus->irq[] needs to be set before calling mdiobus_register().
>>>>
>>>> Yes good point, and that is what the unchanged code does actually.
>>>> Daniel, any idea why that is not working?
>>>
>>> Hi Florian, I don't know. bus->irq[] has no effect, only assigning the
>>> IRQ through phydev->irq works.
>>>
>>> I can resend the patch  without the bus->irq[] line since it's
>>> pointless in this scenario.
>>>
>>
>> It's still an ugly workaround and a proper root cause analysis should be done
>> first. I can only imagine that phydev->irq is overwritten in phy_probe()
>> because phy_drv_supports_irq() is false. Can you please check whether
>> phydev->irq is properly set in phy_device_create(), and if yes, whether
>> it's reset to PHY_POLL in phy_probe()?.
>>
> 
> Hi Heiner, I added some kernel prints:
> 
> [    2.712519] libphy: Fixed MDIO Bus: probed
> [    2.721969] =======phy_device_create===========
> [    2.726841] phy_device_create: dev->irq = 17
> [    2.726841]
> [    2.832620] =======phy_probe===========
> [    2.836846] phy_probe: phydev->irq = 17
> [    2.840950] phy_probe: phy_drv_supports_irq = 0, phy_interrupt_is_valid = 1
> [    2.848267] phy_probe: phydev->irq = -1
> [    2.848267]
> [    2.854059] =======phy_probe===========
> [    2.858174] phy_probe: phydev->irq = -1
> [    2.862253] phy_probe: phydev->irq = -1
> [    2.862253]
> [    2.868121] libphy: bcm63xx_enet MII bus: probed
> [    2.873320] Broadcom BCM63XX (1) bcm63xx_enet-0:01: attached PHY
> driver [Broadcom BCM63XX (1)] (mii_bus:phy_addr=bcm63xx_enet-0:01,
> irq=POLL)
> 
> Currently using kernel 5.4.99. I still have no idea what's going on.
> 
Thanks for debugging. This confirms my assumption that the interrupt
is overwritten in phy_probe(). I'm just scratching my head how
phy_drv_supports_irq() can return 0. In 5.4.99 it's defined as:

static bool phy_drv_supports_irq(struct phy_driver *phydrv)
{
	return phydrv->config_intr && phydrv->ack_interrupt;
}

And that's the PHY driver:

static struct phy_driver bcm63xx_driver[] = {
{
	.phy_id		= 0x00406000,
	.phy_id_mask	= 0xfffffc00,
	.name		= "Broadcom BCM63XX (1)",
	/* PHY_BASIC_FEATURES */
	.flags		= PHY_IS_INTERNAL,
	.config_init	= bcm63xx_config_init,
	.ack_interrupt	= bcm_phy_ack_intr,
	.config_intr	= bcm63xx_config_intr,
}

So both callbacks are set. Can you extend your debugging and check
in phy_drv_supports_irq() which of the callbacks is missing?

Last but not least: Do you use a mainline kernel, or is it maybe
a modified downstream kernel? In the latter case, please check
in your kernel sources whether both callbacks are set.



>> On which kernel version do you face this problem?
>>
> The kernel version 4.4 works ok. The minimum version where I found the
> problem were the kernel 4.9.111, now using 5.4. And 5.10 also tested.
> 
> Regards
> Daniel
> 
>>> Regards
>>>> --
>>>> Florian
>>