Re: [PATCH net v3 2/2] net: fec: Suspend the PHY on probe

[Florian Fainelli <f.fainelli@gmail.com>]

On 3/20/2024 9:54 AM, Russell King (Oracle) wrote:
> On Wed, Mar 20, 2024 at 03:25:54PM +0000, John Ernberg wrote:
>> Hi Russel,
> 
> Growl. Hi Peter.
> 
>> What we really want is the PHY to be suspended on suspend to RAM
>> regardless of us having had an initial link up or not.
> 
> So what you're asking is for the PHY to be suspended when the system
> is entering suspend, which is a long time after the system booted and
> thus phy_probe() was called, and could be some time before the system
> resumes.
> 
> I'm not sure what the relevance is of phy_probe() that was brought up
> previously then.
> 
>> This worked prior to 4c0d2e96ba05 ("net: phy: consider that suspend2ram
>> may cut
>> off PHY power") which was added in Linux 5.11, and 557d5dc83f68 ("net:
>> fec: use
>> mac-managed PHY PM") which was added in Linux 5.12.
> 
> Looking at the former commit, that looks to me like it is only
> affecting the resume paths, not the suspend paths, so wouldn't have
> any impact itself on what happens when suspend happens.
> 
> The latter commit states that it is a work around for an issue with a
> particular PHY. What happens if you revert just this commit, does your
> problem then go away?
> 
> Also, please clarify. It seems that you are reporting a regression -
> it used to work for you prior to 557d5dc83f68, but 557d5dc83f68 stops
> it working for you?
> 
>> Since FEC requires mac_managed_pm the generic PM suspend-resume paths
>> are not
>> taken. The resume sequencing with generic PM has been broken with the
>> FEC since
>> generic PM of the mdio bus was added, as the FEC will do phy_start()
>> (via FEC
>> resume) and then generic PM runs phy_init_hw() via mdio bus resume
>> (previously:
>> less damaging phy_resume()) due to how the FEC IP block works.
> 
> That suggests that even with 557d5dc83f68 reverted, it's broken.
> Digging into the history, what you're referring to dates from January
> 2016, so are you reporting a regression that occured 8 _years_ ago,
> at which point I'd question why it's taken 8 years.
> 
> Given the time that has passed, I don't think reverting commits is
> a sane approach. Quite what the right solution is though, I'm not
> sure.
> 
>  From the description and the commits pointed to, I just don't see
> that there is anything that could've changed with respect to the first
> boot - if that has changed, then I think more research into what caused
> it is needed.
> 
> If it's the subsequent state after a suspend-resume cycle, then yes,
> I would agree that its possible that these changes broke this for you.
> Would clearing ndev->phydev->mac_managed_pm just before
> phy_disconnect() in fec_enet_close() fix it for you, so the suspend/
> resume paths for the PHY get used when the network interface is down?
> 
> Maybe, however, that's something that should happen in any case inside
> phylib on phy_disconnect() as a matter of course, since the PHY will
> at that point be no longer under the control of the network driver for
> PM purposes. Could you give this idea a try please?
> 

On phy_disconnect() we will do a phy_detach() which calls phy_suspend(). 
Given that phy_disconnect() is called from fec_enet_close(), we still 
have a MDIO bus registered and we are not trying to suspend the MDIO 
bus, so we should have an effective phy_suspend() call here, what am I 
missing?
-- 
Florian