Re: [PATCH net] net: phy: fix duplex out of sync problem while changing settings

[Zhang Changzhong]

On 2021/11/2 5:37, Heiner Kallweit wrote:
> On 01.11.2021 10:44, Zhang Changzhong wrote:
>> Before commit 2bd229df5e2e ("net: phy: remove state PHY_FORCING") if
>> phy_ethtool_ksettings_set() is called with autoneg off, phy_start_aneg()
>> will set phydev->state to PHY_FORCING, so adjust_link will always be
>> called.
>>
>> But after that commit, if phy_ethtool_ksettings_set() changes the local
>> link from 10Mbps/Half to 10Mbps/Full when the link partner is
>> 10Mbps/Full, phy_check_link_status() will not trigger the link change,
>> because phydev->link and phydev->state has not changed. This will causes
>> the duplex of the PHY and MAC to be out of sync.
>>
>> Fix it by re-adding the PHY_FORCING state to force adjust_link to be
>> called in fixed mode.
>>
>> Fixes: 2bd229df5e2e ("net: phy: remove state PHY_FORCING")
>> Signed-off-by: Zhang Changzhong <zhangchangzhong@xxxxxxxxxx>
>> ---
>>  drivers/net/phy/phy.c | 10 ++++++++--
>>  include/linux/phy.h   |  6 ++++++
>>  2 files changed, 14 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/net/phy/phy.c b/drivers/net/phy/phy.c
>> index a3bfb15..b114f15 100644
>> --- a/drivers/net/phy/phy.c
>> +++ b/drivers/net/phy/phy.c
>> @@ -49,6 +49,7 @@ static const char *phy_state_to_str(enum phy_state st)
>>  	PHY_STATE_STR(UP)
>>  	PHY_STATE_STR(RUNNING)
>>  	PHY_STATE_STR(NOLINK)
>> +	PHY_STATE_STR(FORCING)
>>  	PHY_STATE_STR(CABLETEST)
>>  	PHY_STATE_STR(HALTED)
>>  	}
>> @@ -724,8 +725,12 @@ static int _phy_start_aneg(struct phy_device *phydev)
>>  	if (err < 0)
>>  		return err;
>>  
>> -	if (phy_is_started(phydev))
>> -		err = phy_check_link_status(phydev);
>> +	if (phy_is_started(phydev)) {
>> +		if (phydev->autoneg == AUTONEG_ENABLE)
>> +			err = phy_check_link_status(phydev);
>> +		else
>> +			phydev->state = PHY_FORCING;
>> +	}
>>  
>>  	return err;
>>  }
>> @@ -1120,6 +1125,7 @@ void phy_state_machine(struct work_struct *work)
>>  		needs_aneg = true;
>>  
>>  		break;
>> +	case PHY_FORCING:
>>  	case PHY_NOLINK:
>>  	case PHY_RUNNING:
>>  		err = phy_check_link_status(phydev);
>> diff --git a/include/linux/phy.h b/include/linux/phy.h
>> index 736e1d1..e639729 100644
>> --- a/include/linux/phy.h
>> +++ b/include/linux/phy.h
>> @@ -446,6 +446,11 @@ struct phy_device *mdiobus_scan(struct mii_bus *bus, int addr);
>>   * - irq or timer will set @PHY_NOLINK if link goes down
>>   * - phy_stop moves to @PHY_HALTED
>>   *
>> + * @PHY_FORCING: PHY is being configured with forced settings.
>> + * - if link is up, move to @PHY_RUNNING
>> + * - if link is down, move to @PHY_NOLINK
>> + * - phy_stop moves to @PHY_HALTED
>> + *
>>   * @PHY_CABLETEST: PHY is performing a cable test. Packet reception/sending
>>   * is not expected to work, carrier will be indicated as down. PHY will be
>>   * poll once per second, or on interrupt for it current state.
>> @@ -463,6 +468,7 @@ enum phy_state {
>>  	PHY_UP,
>>  	PHY_RUNNING,
>>  	PHY_NOLINK,
>> +	PHY_FORCING,
>>  	PHY_CABLETEST,
>>  };
>>  
>>
> 
> I see your point, but the proposed fix may not be fully correct. You rely
> on the phylib state machine, but there are drivers that use
> phy_ethtool_ksettings_set() and some parts of phylib but not the phylib
> state machine. One example is AMD xgbe.

Yes, I notice that, thanks for remind!

> 
> Maybe the following is better, could you please give it a try?
> By the way: With which MAC driver did you encounter the issue?

We encounter the issue with the driver for our in-house MAC.

> 
> 
> diff --git a/drivers/net/phy/phy.c b/drivers/net/phy/phy.c
> index a3bfb156c..beb2b66da 100644
> --- a/drivers/net/phy/phy.c
> +++ b/drivers/net/phy/phy.c
> @@ -815,7 +815,12 @@ int phy_ethtool_ksettings_set(struct phy_device *phydev,
>  	phydev->mdix_ctrl = cmd->base.eth_tp_mdix_ctrl;
>  
>  	/* Restart the PHY */
> -	_phy_start_aneg(phydev);
> +	if (phy_is_started(phydev)) {
> +		phydev->state = PHY_UP;
> +		phy_trigger_machine(phydev);
> +	} else {
> +		_phy_start_aneg(phydev);
> +	}
>  
>  	mutex_unlock(&phydev->lock);
>  	return 0;
> 

We tested it and it works fine, thanks!