> On Jan 5, 2020, at 00:20, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Sat, 4 Jan 2020, Kai-Heng Feng wrote:
>
>>>>>> @@ -3533,9 +3533,17 @@ int usb_port_resume(struct usb_device *udev, pm_message_t msg)
>>>>>> }
>>>>>>
>>>>>> /* see 7.1.7.7; affects power usage, but not budgeting */
>>>>>> - if (hub_is_superspeed(hub->hdev))
>>>>>> + if (hub_is_superspeed(hub->hdev)) {
>>>>>> + if (hub->hdev->quirks & USB_QUIRK_DISABLE_LPM_ON_U0) {
>>>>>> + usb_lock_device(hub->hdev);
>>>>>> + usb_unlocked_disable_lpm(hub->hdev);
>>>>>> + }
>>>>>> status = hub_set_port_link_state(hub, port1, USB_SS_PORT_LS_U0);
>>>>>> - else
>>>>>> + if (hub->hdev->quirks & USB_QUIRK_DISABLE_LPM_ON_U0) {
>>>>>> + usb_unlocked_enable_lpm(hub->hdev);
>>>>>> + usb_unlock_device(hub->hdev);
>>>>>
>>>>> The locking here seems questionable. Doesn't this code sometimes get
>>>>> called with the hub already locked? Or with the child device locked
>>>>> (in which case locking the hub would violate the normal locking order:
>>>>> parent first, child second)?
>>>
>>> I did a little checking. In many cases the child device _will_ be
>>> locked at this point.
>>>
>>>> Maybe introduce a new lock? The lock however will only be used by this specific hub.
>>>> But I still want the LPM can be enabled for this hub.
>>>
>>> Do you really need to lock the hub at all? What would the lock protect
>>> against?
>>
>> There can be multiple usb_port_resume() run at the same time for different ports, so this is to prevent LPM enable/disable race.
>
> But there can't really be an LPM enable/disable race, can there? The
> individual function calls are protected by the bandwidth mutex taken by
> the usb_unlocked_{en|dis}able_lpm routines, and the overall LPM setting
> is controlled by the hub device's lpm_disable_counter.
For enable/disable LPM itself, there's no race.
But the lock here is to protect hub_set_port_link_state().
If we don't lock the hub, other instances of usb_port_resume() routine can enable LPM and we want the LPM stays disabled until hub_set_port_link_state() is done.
Kai-Heng
>
> So I think you don't need to lock the hub here.
>
> Alan Stern
>
