patches-7.0: replace cdnsp-sky1 shutdown SError mitigations with host teardown

[npmccallum]

Remove 0061-DPTSW-24991 and 0062-DPTSW-25423 and replace them with 0061-usb-cdns3-cdnsp-sky1-tear-down-host-on-shutdown.patch.

cdnsp_sky1_shutdown() gates the USB clocks while the child xHCI host and its devices are still live above the controller. If anything touches the controller's registers after the clocks are gated, the access raises a fatal asynchronous SError that hangs shutdown/reboot. Three distinct paths can do this:

1. the gadget (cdnsp_gadget_pullup) - addressed by 0061 (release the gadget driver);
2. the xHCI IRQ handler (xhci_irq reading USBSTS) - addressed by 0062 (disable_irq + clear HCD_FLAG_HW_ACCESSIBLE);
3. a host device's async error recovery - not covered by either.

Instead of a third targeted mitigation, route ->shutdown() through the existing ->remove() path, which calls of_platform_depopulate() to remove the xHCI host first - disconnecting its devices and stopping all of their work (gadget, IRQ, and async recovery) - before the resets are asserted and the clocks gated. One patch supersedes both mitigations and covers the host async-recovery path they miss.

Reusing ->remove() for ->shutdown() is an established pattern for USB
dual-role / glue drivers in mainline, for example:

• dwc3-qcom: .shutdown = dwc3_qcom_remove
  (drivers/usb/dwc3/dwc3-qcom.c)
• dwc3-xilinx: .shutdown = dwc3_xlnx_remove
  (drivers/usb/dwc3/dwc3-xilinx.c)
• chipidea/imx: ci_hdrc_imx_shutdown() -> ci_hdrc_imx_remove()
  (drivers/usb/chipidea/ci_hdrc_imx.c)

[coffnix]

tested on 7.0.14. Thanks, your patch works.

[amazingfate]

This change uses of_platform_depopulate, which should not work with ACPI. The current patchset is focusing on ACPI now, DT mode support need further work.

[npmccallum]

You're right that of_platform_depopulate() is a no-op under ACPI — parent->of_node is NULL, so it returns without doing anything. My commit message was wrong to lean on it and I'll rewrite it.

The fix doesn't depend on it, though. The child xHCI device is already quiesced before the wrapper's shutdown runs, on either firmware:

• cdns_host_init() sets xhci->dev.parent = cdns->dev and registers the xHCI device with platform_device_add(), so it enters devices_kset after its parents.
• device_shutdown() walks that list from the tail, so children shut down before parents.
• xhci_plat_driver.shutdown = usb_hcd_platform_shutdown → xhci_shutdown() → xhci_halt(), which halts the controller and masks the interrupter.

By the time cdnsp_sky1's shutdown runs, the host is halted. That holds under ACPI and DT alike.

That's also the bug I'm fixing: 0061 and 0062 both run after that point. Releasing the gadget driver and calling disable_irq()/synchronize_irq() there blocks against a controller that is already stopped, and powering off with a USB device attached hangs.

One gap in my own patch, which your comment led me to. 0061 covered the gadget, and unlike the host the gadget has no child device the core tears down first. Under DT, of_platform_depopulate() removed the cdns3 core and unbound it; under ACPI nothing does. If the DPTSW-24991 SError reproduces in ACPI gadget mode, this patch would reintroduce it. Does anyone have a reproducer? If so I'll add a gadget teardown that doesn't depend on OF.