On 10/26/2023 11:37 PM, Roger Quadros wrote:
On 27/10/2023 03:07, Elson Serrao wrote:
While this takes care of runtime suspend case, what about system_suspend?
Should this check be moved to dwc3_suspend_common() instead?
Sure I can move these checks to dwc3_suspend_common to make it generic.
Before you do that let's first decide how we want the gadget driver to behave
in system_suspend case.
Current behavior is to Disconnect from the Host.
Earlier I was thinking on the lines that we prevent system suspend if
we are not already in USB suspend. But I'm not sure if that is the right
thing to do anymore. Mainly because, system suspend is a result of user
request and it may not be nice to not to meet his/her request.
Agree. Irrespective of whether USB is suspended or not it is better to honor the system suspend request from user.
Maybe best to leave this policy handling to user space?
i.e. if user wants USB gadget operation to be alive, he will not issue
system suspend?
Sure. So below two cases
Case1: User doesn't care if gadget operation is alive and triggers system suspend irrespective of USB suspend. Like you mentioned, current behavior already takes care of this and initiates a DISCONNECT
Case2: User wants gadget to stay alive and hence can trigger system suspend only when USB is suspended (there are already user space hooks that read cdev->suspended bit to tell whether USB is suspended or not for user to decide). Attempts to request system suspend when USB is not suspended, would result in a DISCONNECT.
For supporting Case2 from gadget driver point of view, we need to extend this series by having relevant checks in suspend_common()
Also, is it better to provide separate flags to control the gadget driver behavior for runtime suspend Vs system suspend when USB is suspended ? For example, what if we want to enable bus suspend handling for runtime suspend only and not for system suspend (Case1).
But you mentioned that for Case1, USB gadget would disconnect from Host. So USB will be in disconnected state and USB controller can be fully de-activated? Except maybe wakeup handling to bring system out of suspend on a USB plug/unplug event?
Why do we need separate flags for?
Sorry let me clarify. This is in reference to deciding how we want the dwc3 driver to behave in system_suspend case.
One option is to continue with the existing behavior where USB gadget would disconnect from Host irrespective of bus suspend state. We dont need any modification in this case and we can leave this series limited to runtime suspend only.
Second option is to stay connected IF we are in bus suspend state (U3/L2) otherwise DISCONNECT IF we are not in bus suspend state. The main motivation is to preserve the ongoing usb session
without going through a re-enumeration (ofcourse true only if we are in bus suspend state). This would need relevant checks in suspend_common().
The catch here is, what to do if the USB device is not in bus suspend state but user wants to put the system in suspend state? Do we still disconnect?
You might also want to refer to the discussion in [1]
[1] - https://lore.kernel.org/all/Y+z9NK6AyhvTQMir@xxxxxxxxxxxxxxxxxxx/
Thanks for the details. If we dont DISCONNECT when the USB link is NOT
in bus suspend, the other alternatives we have are below ones
1.) Abort system_suspend: In addition to not honoring the users request
to put the system in suspend, there is always a possibility of host
driver never sending bus suspend interrupt. If that happens we would be
denying system_suspend every time user requests it. So this is not a
right approach.
2.) Keep the gadget connected and proceed with system_suspend: Now the
system is suspended but from host point of view the USB link is active
and would continue the communication normally. The signalling probably
is not going to to be detected as a wakeup by the Rx hardware as there
is no explicit resume signal involved here . The only exception is if we
can somehow configure each and every event from the host as a wakeup
signal(not sure if this is even possible)
So IMO it is best to DISCONNECT from the host when USB link is NOT in
bus suspend state and user requests system_suspend.
Thanks
Elson
