Re: Problems with remote-wakeup settings

From: Rafael J. Wysocki
Date: Sat Mar 06 2010 - 15:51:28 EST

On Saturday 06 March 2010, Alan Stern wrote:
> On Fri, 5 Mar 2010, Rafael J. Wysocki wrote:
> > > So the problem is that subsystems can't usefully set the can_wakeup
> > > flag before doing either device_initialize() or device_register().
> > > This can be fixed easily by removing the call in device_initialize().
> >
> > PCI depends on the flag being unset when pci_pm_init(dev) is called.
> >
> > If that's still valid after removing the call in device_initialize(), I'm fine
> > with the removal.
> It should still be valid. After all, there's nothing else to set the
> flag except for other parts of the PCI core.

All right, then.

> > > Agreed, ethtool and sysfs should affect the same flags.
> >
> > Yeah, but. Right now, if the setting is changed via sysfs, it doesn't modify
> > the WoL setting as visible by ethtool.
> > > I don't understand. Do you mean there's no way to update the
> > > _device's_ WoL setting when the sysfs attribute is changed?
> >
> > There's no code for that, that's the problem.
> >
> > > The device's WoL setting matters only at suspend time. So the network
> > > driver's suspend routine ought to test device_may_wakeup() to see
> > > whether or not WoL should be enabled. Maybe this can be centralized
> > > somewhere in the network stack.
> >
> > Maybe. The problem is people expect wakeup to work once WoL has been set
> > with the help of ethtool and they expect it to work if the WoL is set by
> > default.
> It's not difficult in theory to tie together the WoL setting and the
> wakeup flag:
> If ethtool changes the WoL setting, the driver's ioctl handler
> should make the corresponding change to the wakeup flag.
> If ethtool queries the WoL setting, the ioctl handler should
> check the wakeup flag. If the flag is off, it should report
> that WoL is disabled; if the flag is on, it should report that
> WoL is enabled. (The same check should be made in the suspend
> routine.)

That's done this way already in all drivers I know, but we need a hook
from wake_store() back to the driver.

> > > And also IMO, enabling WoL by default is very questionable. But that's
> > > a separate matter.
> >
> > That's been a common practice for years in the network adapter land and I don't
> > think we're able to change that now. Besides, if the WoL is set to g by
> > default, which also is common, that doesn't really lead to any problems.
> All right, we can declare that network drivers are allowed to enable
> WoL by default (like keyboard drivers). There shouldn't be any problem
> provided they initialize the wakeup setting before registering the
> network interface, so that the initialization doesn't override any
> action by udev.

That sounds reasonable.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at