Re: [PATCH] net: ethernet: fec: Add missing SPEED_

From: Heiner Kallweit
Date: Thu Oct 18 2018 - 16:41:25 EST


On 18.10.2018 22:10, Florian Fainelli wrote:
> On 10/18/2018 12:59 PM, LABBE Corentin wrote:
>> On Thu, Oct 18, 2018 at 12:38:32PM -0700, Florian Fainelli wrote:
>>> On 10/18/2018 12:16 PM, LABBE Corentin wrote:
>>>> On Thu, Oct 18, 2018 at 11:55:49AM -0700, Florian Fainelli wrote:
>>>>> On 10/18/2018 11:47 AM, LABBE Corentin wrote:
>>>>>> On Thu, Oct 18, 2018 at 11:39:24AM -0700, Florian Fainelli wrote:
>>>>>>> On 10/18/2018 08:05 AM, Corentin Labbe wrote:
>>>>>>>> Since commit 58056c1e1b0e ("net: ethernet: Use phy_set_max_speed() to limit advertised speed"), the fec driver is unable to get any link.
>>>>>>>> This is due to missing SPEED_.
>>>>>>>
>>>>>>> But SPEED_1000 is defined in include/uapi/linux/ethtool.h as 1000, so
>>>>>>> surely this would amount to the same code paths being taken or am I
>>>>>>> missing something here?
>>>>>>
>>>>>> The bisect session pointed your patch, reverting it fix the issue.
>>>>>> BUT since the fix seemed trivial I sent the patch without more test then compile it.
>>>>>> Sorry, I have just found some minutes ago that it didnt fix the issue.
>>>>>>
>>>>>> But your patch is still the cause for sure.
>>>>>>
>>>>>
>>>>> What you are writing is really lowering the confidence level, first
>>>>> Andrew is the author of that patch, and second "just compiling" and
>>>>> pretending this fixes a problem when it does not is not quite what I
>>>>> would expect.
>>>>>
>>>>> I don't have a problem helping you find the solution or the right fix
>>>>> though, even if it is not my patch, but please get the author and actual
>>>>> problem right so we can move forward in confidence, thanks!
>>>>
>>>> Sorry again, I wanted to acknoledge my error but I did it too fast and late.
>>>> And sorry to have confound you with Andrew.
>>>
>>> No worries, here to help, let us know what your bisection points to. THanks
>>
>> I have added printing of phydev->supported
>> My working kernel (on top of 58056c1e1b0e + revert patch) got:
>> [ 5.550838] fec_enet_mii_probe 2ff (gbit features)
>> [ 5.555848] fec_enet_mii_probe 2ef (without 1000baseT_Half)
>> [ 5.561620] fec_enet_mii_probe 22ef final (after pause)
>> [ 5.566914] Micrel KSZ9021 Gigabit PHY 2188000.ethernet-1:06: attached PHY driver [Micrel KSZ9021 Gigabit PHY] (mii_bus:phy_addr=2188000.ethernet-1:06, irq=POLL)
>> [ 8.730751] fec 2188000.ethernet eth0: Link is Up - 1Gbps/Full - flow control rx/tx
>> [ 8.788311] Sending DHCP requests ., OK
>> [ 8.832357] IP-Config: Got DHCP answer from 192.168.66.1, my address is 192.168.66.58
>>
>> the non-working kernel (next-20181015)
>> [ 7.308917] fec_enet_mii_probe 62ff after phy_set_max_speed
>> [ 7.314545] fec_enet_mii_probe 62ef after phy_remove_link_mode
>> [ 7.320418] fec_enet_mii_probe 62ef after pause
>> and then no link
>>
>> So it seems that phy_set_max_speed adds bit 14 (ETHTOOL_LINK_MODE_Asym_Pause_BIT)
>
> It's not masking it so it must be coming from phy_probe().
>
See df8ed346d4a8 ("net: phy: fix flag masking in __set_phy_supported").
phy_set_max_speed() used to (unintentionally) mask the pause bits
and it seems that the fec driver used this bug as a feature.

>>
>> I have patched by adding:
>> phy_remove_link_mode(phy_dev, ETHTOOL_LINK_MODE_Asym_Pause_BIT);

Instead of programmatically removing the feature bit it should be
possible to do this in the PHY driver configuration. See also
this part of phy_probe().

if (phydrv->features & (SUPPORTED_Pause | SUPPORTED_Asym_Pause)) {
phydev->supported &= ~(SUPPORTED_Pause | SUPPORTED_Asym_Pause);
phydev->supported |= phydrv->features &
(SUPPORTED_Pause | SUPPORTED_Asym_Pause);
} else {
phydev->supported |= SUPPORTED_Pause | SUPPORTED_Asym_Pause;
}

>> and got:
>> [ 7.310559] fec_enet_mii_probe 62ff after phy_set_max_speed
>> [ 7.316221] fec_enet_mii_probe 22ef after phy_remove_link_mode
>> [ 7.322128] fec_enet_mii_probe 22ef after pause
>> [ 7.326681] Micrel KSZ9021 Gigabit PHY 2188000.ethernet-1:06: attached PHY driver [Micrel KSZ9021 Gigabit PHY] (mii_bus:phy_addr=2188000.ethernet-1:06, irq=POLL)
>> [ 7.611276] Waiting up to 3 more seconds for network.
>> [ 7.881278] Waiting up to 2 more seconds for network.
>> [ 8.131277] Waiting up to 2 more seconds for network.
>> [ 8.401169] Waiting up to 2 more seconds for network.
>> [ 8.671269] Waiting up to 2 more seconds for network.
>> [ 8.941274] Waiting up to 1 more seconds for network.
>> [ 9.211181] Waiting up to 1 more seconds for network.
>> [ 9.481274] Waiting up to 1 more seconds for network.
>> [ 9.751275] Waiting up to 1 more seconds for network.
>> [ 10.021281] Waiting up to 0 more seconds for network.
>> [ 10.291274] Waiting up to 0 more seconds for network.
>> [ 10.381282] Sending DHCP requests .
>> [ 10.473000] fec 2188000.ethernet eth0: Link is Up - 1Gbps/Full - flow control rx/tx
>> [ 12.861267] ., OK
>> [ 12.903405] IP-Config: Got DHCP answer from 192.168.66.1, my address is 192.168.66.58
>>
>> So at least I got a link, but the link is still late to got
>
> The delay is likely something entirely different, it could be some of
> Heiner's recent changes to PHYLIB, Heiner do you have access to a system
> that polls the PHY?
>
I don't think there's anything wrong with phylib. Time difference
between the fec_enet_mii_probe messages and the "link up" message
is little bit more than 3s in both cases.
For a reason not visible here the fec_enet_mii_probe messages
come 2s later in the second case.
What happens after the "link up" message is out of control of phylib.