Hello Sergei,
On 06/03/2018 06:42 PM, Sergei Shtylyov wrote:
> Hello!
>
> Sorry for the delay replying, the management keeps me busy... :-(
>
> On 05/28/2018 12:51 PM, Vladimir Zapolskiy wrote:
>
>>>> The change replaces a custom implementation of .set_link_ksettings
>>>> callback with a shared phy_ethtool_set_link_ksettings(), this fixes
>>>> sleep in atomic context bug, which is encountered every time when link
>>>> settings are changed by ethtool.
>>>
>>> Seeing it now...
>
> And to say that this is *fixed* by removing the custom method is err...
> simply misleading. The sleep in atomic context is fixed solely by the removal
> of the spinlock grabbing before the phylib call.
>
As I've already said, "the removal of the spinlock grabbing before the phylib
call" is not a valid fix, but it will introduce another regression.
>>>> Now duplex mode setting is enforced in ravb_adjust_link() only, also
>>>> now TX/RX is disabled when link is put down or modifications to E-MAC
>>>> registers ECMR and GECMR are expected for both cases of checked and
>>>> ignored link status pin state from E-MAC interrupt handler.
>>>>
>>>> Signed-off-by: Vladimir Zapolskiy <vladimir_zapolskiy@xxxxxxxxxx>
>>>> ---
>>>> drivers/net/ethernet/renesas/ravb_main.c | 58 +++++++++-----------------------
>>>> 1 file changed, 15 insertions(+), 43 deletions(-)
>>>>
>>>> diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c
>>>> index 3d91caa44176..0d811c02ff34 100644
>>>> --- a/drivers/net/ethernet/renesas/ravb_main.c
>>>> +++ b/drivers/net/ethernet/renesas/ravb_main.c
>>>> @@ -980,6 +980,13 @@ static void ravb_adjust_link(struct net_device *ndev)
>>>> struct ravb_private *priv = netdev_priv(ndev);
>>>> struct phy_device *phydev = ndev->phydev;
>>>> bool new_state = false;
>>>> + unsigned long flags;
>>>> +
>>>> + spin_lock_irqsave(&priv->lock, flags);
>>>> +
>>>> + /* Disable TX and RX right over here, if E-MAC change is ignored */
>>>> + if (priv->no_avb_link)
>>>> + ravb_rcv_snd_disable(ndev);
>>>>
>>>> if (phydev->link) {
>>>> if (phydev->duplex != priv->duplex) {
>>>> @@ -997,18 +1004,21 @@ static void ravb_adjust_link(struct net_device *ndev)
>>>> ravb_modify(ndev, ECMR, ECMR_TXF, 0);
>>>> new_state = true;
>>>> priv->link = phydev->link;
>>>> - if (priv->no_avb_link)
>>>> - ravb_rcv_snd_enable(ndev);
>>>> }
>>>> } else if (priv->link) {
>>>> new_state = true;
>>>> priv->link = 0;
>>>> priv->speed = 0;
>>>> priv->duplex = -1;
>>>> - if (priv->no_avb_link)
>>>> - ravb_rcv_snd_disable(ndev);
>>>> }
>>>>
>>>> + /* Enable TX and RX right over here, if E-MAC change is ignored */
>>>> + if (priv->no_avb_link && phydev->link)
>>>> + ravb_rcv_snd_enable(ndev);
>>>> +
>>>> + mmiowb();
>>>> + spin_unlock_irqrestore(&priv->lock, flags);
>>>> +
>>>
>>> I like this part. :-)
>>>
>>
>> A weight off my mind :) And I hope that this change will remain the less
>> questionable one, other ones from the series are trivial.
>>
>> Anyway I hope it is understandable that this part of the change can not
>> be simply extracted from the rest one below, otherwise there'll be bugs of
>> another type intorduced.
>
> I never said I'd like to apply this part alone, my idea was more like removing
> the spinlock grabbing and the duplex handling down below.
>
As I've already said, "the removal of the spinlock grabbing" is not a valid fix,
but it will introduce another regression.
Please tell me, a removal of duplex handling change should be done before
a removal of the spinlock grabbing? Or after?
> [...]
>>>> @@ -1096,44 +1106,6 @@ static int ravb_phy_start(struct net_device *ndev)
>>>> return 0;
>>>> }
>>>>
>>>> -static int ravb_set_link_ksettings(struct net_device *ndev,
>>>> - const struct ethtool_link_ksettings *cmd)
>>>> -{
>>>> - struct ravb_private *priv = netdev_priv(ndev);
>>>> - unsigned long flags;
>>>> - int error;
>>>> -
>>>> - if (!ndev->phydev)
>>>> - return -ENODEV;
>>>> -
>>>> - spin_lock_irqsave(&priv->lock, flags);
>>>> -
>>>> - /* Disable TX and RX */
>>>> - ravb_rcv_snd_disable(ndev);
>>>> -
>>>> - error = phy_ethtool_ksettings_set(ndev->phydev, cmd);
>>>> - if (error)
>>>> - goto error_exit;
>>>> -
>>>> - if (cmd->base.duplex == DUPLEX_FULL)
>>>> - priv->duplex = 1;
>>>> - else
>>>> - priv->duplex = 0;
>>>> -
>>>> - ravb_set_duplex(ndev);
>>>> -
>>>> -error_exit:
>>>> - mdelay(1);
>>>> -
>>>> - /* Enable TX and RX */
>>>> - ravb_rcv_snd_enable(ndev);
>>>> -
>>>> - mmiowb();
>>>> - spin_unlock_irqrestore(&priv->lock, flags);
>>>> -
>>>> - return error;
>>>> -}
>>>> -
>>>
>>> But this part is clearly lumping it all together...
>>
>> Please elaborate.
>
> My point is still that complete removal of the custom method was somewhat
> premature and completely unnecessary for fixing the issues we have.
>
>>> [...]
>>>> @@ -1357,7 +1329,7 @@ static const struct ethtool_ops ravb_ethtool_ops = {
>>>> .set_ringparam = ravb_set_ringparam,
>>>> .get_ts_info = ravb_get_ts_info,
>>>> .get_link_ksettings = phy_ethtool_get_link_ksettings,
>>>> - .set_link_ksettings = ravb_set_link_ksettings,
>>>> + .set_link_ksettings = phy_ethtool_set_link_ksettings,
>>>
>>> Should have been a part of the final patch in the fix/enhancement chain...
>>
>> Please elaborate.
>>
>> Do you mean that firstly I have to make erroneous ravb_set_link_ksettings()
>> to look similar to phy_ethtool_set_link_ksettings() and then remove it?
>
> Yes.
>
Then this change of "ravb_set_link_ksettings() looks similar to
phy_ethtool_set_link_ksettings()" will be a single commit, and it will be
a fix. Does it sound good?
>> As I see it in the current context (removal of ravb_set_duplex() call and
>> so on), the problem with this approach is that the actual fix change will
>> be done on top of a number of enchancement changes, thus it contradicts to
>
> Now I have to ask you to elaborate. I have no idea what you mean. :-(
>
My statement is based on the next facts:
1. ravb_set_duplex() call in ravb_set_link_ksettings() is unnecessary,
however its removal is an enchancement,
2. removal of the spinlock grabbing is just a *part* of the proper fix,
and the complete proper fix includes ravb_set_duplex() call removal,
adding spinlock grabbing to ravb_adjust_link() and other modifications
to ravb_adjust_link() from this commit.
3. commits with fixes must precede commits with enchancements in the
series, because enchancements are not backported.
The question remains the same, what do you ask me to do?
> And of course, sometimes the things are broken in a so subtle way, that
> only as pile of "cleanups" fixed them, we had that situation in e.g. the
> R-Car I2C driver -- *none* of AFAIR 9 patches was good as a -stable patch...
>
>> the accepted development/maintenace model "fixes first", and most probably
>> it won't be possible to backport the real fix, however this sole change can
>> be backported.
>
> My idea was to move the [G]ECMR writes to the adjust_link() callback and
> to stop grabbing the spinlock where it *was* grabbed in the same fix patch.
> Then just a single clean up, to start using the new phylib method.
>
It will be okay iff ravb_set_duplex() call removal is added to the first
part ("fixes" part) of two changes. Please confirm that our understanding
is aligned.
--
With best wishes,
Vladimir
