On Sat, Sep 13, 2025 at 08:45:05AM +0200, Oleksij Rempel wrote:
On Fri, Sep 12, 2025 at 03:37:52PM +0100, Russell King (Oracle) wrote:
On Fri, Sep 12, 2025 at 10:29:47AM -0400, Alan Stern wrote:
On Fri, Sep 12, 2025 at 09:33:05AM +0100, Russell King (Oracle) wrote:
On Thu, Sep 11, 2025 at 10:30:09PM -0400, Alan Stern wrote:
The USB subsystem uses only one pair of callbacks for suspend and resume because USB hardware has only one suspend state. However, the callbacks do get an extra pm_message_t parameter which they can use to distinguish between system sleep transitions and runtime PM transitions.
Unfortunately, this isn't the case. While a struct usb_device_driver's suspend()/resume() methods get the pm_message_t, a struct usb_driver's suspend()/resume() methods do not:
static int usb_resume_interface(struct usb_device *udev, struct usb_interface *intf, pm_message_t msg, int reset_resume) { struct usb_driver *driver; ... if (reset_resume) { if (driver->reset_resume) { status = driver->reset_resume(intf); ... } else { status = driver->resume(intf);
vs
static int usb_resume_device(struct usb_device *udev, pm_message_t msg) { struct usb_device_driver *udriver; ... if (status == 0 && udriver->resume) status = udriver->resume(udev, msg);
and in drivers/net/usb/asix_devices.c:
static struct usb_driver asix_driver = { ... .suspend = asix_suspend, .resume = asix_resume, .reset_resume = asix_resume,
where asix_resume() only takes one argument:
static int asix_resume(struct usb_interface *intf) {
Your email made me go back and check the code more carefully, and it turns out that we were both half-right. :-)
The pm_message_t argument is passed to the usb_driver's ->suspend callback in usb_suspend_interface(), but not to the ->resume callback in usb_resume_interface(). Yes, it's inconsistent.
I suppose the API could be changed, at the cost of updating a lot of drivers. But it would be easier if this wasn't necessary, if there was some way to work around the problem. Unfortunately, I don't know anything about how the network stack handles suspend and resume, or what sort of locking it requires, so I can't offer any suggestions.
I, too, am unable to help further as I have no bandwidth available to deal with this. Sorry.
Thanks for all the valuable input.
I’ll process the feedback and investigate possible ways to proceed. As a first step I’ll measure the actual power savings from USB auto-suspend on AX88772 to see if runtime PM is worth the added complexity.
I ran quick power measurements to check whether USB autosuspend is worth the added complexity.
Meaning: - "admin up/down" = ip link set dev <if> up/down. - No link partner was attached, so the physical link was down in all tests.
Setups: - Debian 5.10 (USB autosuspend present, no phylib). - Debian 6.1 (phylib present, no regression). - Power meter: Fnirsi FNB58. - Env: QEMU 9.2.1 (USB passthrough) xHCI host Intel 100/C230 device ASIX AX88772B (0b95:772b)
Legend: - "RT: active" = runtime PM on; - "RT: suspended" = runtime PM auto (device suspended).
Results: - Kernel 5.10.237-1 admin up (link down): 0.453 W (RT: active) admin down: 0.453 W (RT: active) admin down: 0.453 W (RT: suspended)
- Kernel 6.1.148-1 admin up (link down): 0.453 W (RT: active) admin down: 0.248 W (RT: active) admin down: 0.248 W (RT: suspended)
Observations: In this setup, USB autosuspend did not reduce power further (admin-down power is identical with/without autosuspend).
The drop from ~0.453 W -> ~0.248 W on 6.1 appears to come from the phylib migration (PHY powered down on admin-down), not from autosuspend.
Proposal: Given autosuspend brings no measurable benefit here, and it hasn’t been effectively functional for this device in earlier kernels, I suggest a minimal -stable patch that disables USB autosuspend for ASIX driver to avoid the PM/RTNL/MDIO issues. If someone needs autosuspend-based low-power later, they can implement a proper device low-power sequence and re-enable it.
Would this minimal -stable patch be acceptable?
Best Regards, Oleksij