Re: [PATCH net] usbnet: smsc95xx: Fix deadlock on runtime resume

[Alan Stern <stern@rowland.harvard.edu>]

On Wed, Apr 27, 2022 at 05:38:51PM +0200, Lukas Wunner wrote:
> On Wed, Apr 27, 2022 at 05:24:50PM +0200, Andrew Lunn wrote:
> > On Wed, Apr 27, 2022 at 08:41:49AM +0200, Lukas Wunner wrote:
> > > Commit 05b35e7eb9a1 ("smsc95xx: add phylib support") amended
> > > smsc95xx_resume() to call phy_init_hw().  That function waits for the
> > > device to runtime resume even though it is placed in the runtime resume
> > > path, causing a deadlock.
> > 
> > You have looked at this code, tried a few different things, so this is
> > probably a dumb question.
> > 
> > Do you actually need to call phy_init_hw()?
> > 
> > mdio_bus_phy_resume() will call phy_init_hw(). So long as you first
> > resume the MAC and then the PHY, shouldn't this just work?
> 
> mdio_bus_phy_resume() is only called for system sleep.  But this is about
> *runtime* PM.
> 
> mdio_bus_phy_pm_ops does not define runtime PM callbacks.  So runtime PM
> is disabled on PHYs, no callback is invoked for them when the MAC runtime
> suspends, hence the onus is on the MAC to runtime suspend the PHY (which
> is a child of the MAC).  Same on runtime resume.
> 
> Let's say I enable runtime PM on the PHY and use pm_runtime_force_suspend()
> to suspend the PHY from the MAC's ->runtime_suspend callback.  At that
> point the MAC already has status RPM_SUSPENDING.  Boom, deadlock.
> 
> The runtime PM core lacks the capability to declare that children should
> be force runtime suspended before a device can runtime suspend, that's
> the problem.

This might work out if you copy the scheme we use for USB devices and 
interfaces.

A USB interface is only a logical part of its parent device, and as such 
does not have a separate runtime power state of its own (in USB-2, at 
least).  Therefore the USB core calls pm_runtime_no_callbacks() for each 
interface as it is created, and handles the runtime power management for 
the interface (i.e., invoking the interface driver's runtime_suspend and 
runtime_resume callbacks) from within the device's runtime PM routines 
-- independent of the PM core's notion of what the interface's power 
state should be.

Similarly, you could call pm_runtime_no_callbacks() for the PHY when it 
is created, and manage the PHY's actual power state from within the 
MAC's runtime-PM routines directly (i.e., without going through the PM 
core).

Alan Stern