Re: [PATCH v2] r8152: Suspend USB device before shutdown when WoL is enabled

From: Alan Stern
Date: Wed Aug 02 2023 - 11:26:43 EST

On Wed, Aug 02, 2023 at 02:56:43PM +0000, Gagniuc, Alexandru wrote:
> On Wed, Jul 19, 2023 at 02:36:25PM -0400, Alan Stern wrote:
> > On Wed, Jul 19, 2023 at 05:37:56PM +0000, Alexandru Gagniuc wrote:
> > > For Wake-on-LAN to work from S5 (shutdown), the USB link must be put
> > > in U3 state. If it is not, and the host "disappears", the chip will
> > > no longer respond to WoL triggers.
> > >
> > > To resolve this, add a notifier block and register it as a reboot
> > > notifier. When WoL is enabled, work through the usb_device struct to
> > > get to the suspend function. Calling this function puts the link in
> > > the correct state for WoL to function.
> >
> > How do you know that the link will _remain_ in the correct state?
> The objective is to get to xhci_set_link_state() with the USB_SS_PORT_LS_U3
> argument. This is achieved through usb_port_suspend() in drivers/usb/host/hub.c,
> and the function is implemented in drivers/usb/host/xhci-hub.c.
> This is the only path in the kernel that I am aware of for setting the U3 link
> state. Given that it is part of the USB subsystem, I am fairly confident it will
> show consistent behavior across platforms.

That does not answer my question. I agree that making this change will
put the link into the U3 state. But I don't have any reason to think
that some other software won't later put the link into some other state.

> > That is, how do you know that the shutdown processing for the USB host
> > controller won't disable the link entirely, thereby preventing WoL from
> > working?
> We are talking to the USB hub in order to set the link state. I don't see how
> specifics of the host controller would influence behavior.

Specifics of the host controller probably won't influence behavior.
However, specifics of the _software_ can make a big difference.

> I do expect a
> controller which advertises S4/S5 in /proc/acpi/wakeup to not do anything that
> would sabotage this capability. Disabling the link entirely would probalby
> violate that promise.

Not if the kernel _tells_ the controller to disable the link.

> Think of USB-C docks with a power button showing up as a HID class. The scenario
> herein would disable the power button. I would take that to be a bug in the host
> controller driver if the S4/S5 capability is advertised.

Indeed. And I am asking how you can be sure the host controller driver
(or some other part of the software stack) doesn't have this bug.

Alan Stern