Re: [PATCH] usb: core: devio: add ioctls for suspend and resume

[Oliver Neukum <oneukum@suse.com>]

Am Dienstag, den 25.06.2019, 10:08 -0400 schrieb Alan Stern:
> On Tue, 25 Jun 2019, Mayuresh Kulkarni wrote:

> > Right, so user space should do the following when it determines the
> > device is idle from its point of view -
> > 
> > 1. Call ALLOW_SUSPEND ioctl

That is a race in principle. You should reverse steps 1 and 2

> > 2. Queue an URB and wait for its REAP. When the wait returns -EFAIL (or
> > something similar), that is the indication that the device is no longer
> > active (or suspended)
> > 3. Call WAIT_FOR_RESUME ioctl

It seems to me that there ought to be one API for that. Either an
ioctl or poll.

> > 4. When WAIT_FOR_RESUME ioctl returns, it is guaranteed that device is
> > active.
> > 5. Call FORBID_SUSPEND ioctl and read the cause of resume.
> > 6. Go to (1) when appropriate
> > 
> > Have I summarized this approach correctly from user-space point of view?
> 
> Yes, except for one thing: In step 4, it is _not_ guaranteed that the 
> device is active when WAIT_FOR_RESUME returns.  The only guarantee is 
> that a resume did occur sometime after step 1, but the device might 
> have gone back into suspend after that occurred.

Now is that a good API? Shouldn't we rather have an API for either
* resume the device now and bump the counter
* bump the counter when te device resumes

I don't see a value in not having a defined power state after resume.

> > 3. Extend usbdev_poll() to wait for udev->state == USB_STATE_SUSPENDED
> > when events == POLLPRI. Return POLLPRI when state = USB_STATE_SUSPENDED.
> > 4. As before, any ioctl != (ALLOW_SUSPEND or WAIT_FOR_RESUME)
> > calls usb_autoresume_device().
> 
> 3 sounds reasonable at first, but I'm not sure it would work.  
> Consider what would happen if the device is suspended very briefly and
> then wakes up.  The usbdev_poll() call might not return, because by the
> time it checks udev->state, the state has already changed back to
> USB_STATE_CONFIGURED.

Indeed. It seems to me that any power transition should be reported
back.

> In any case, we shouldn't do 4.  It would prevent the device from ever
> going into suspend, because the program would want to continue making
> usbfs ioctl calls while waiting for the suspend to occur.

Exactly.

> > The corresponding user-space calls would be -
> > A. When determined device is idle, call ALLOW_SUSPEND ioctl.
> > B. Call poll(device_fd, POLLPRI). When poll returns check revents
> > == POLLPRI.
> 
> What if the device never does go into suspend?  The poll() call
> wouldn't return and the program would just stop working.

Well, that is why you can poll for multiple events at the same
time and the syscall has a timeout.

> > 2. Is it safe to wait for udev->state to be of a particular value?
> 
> No, not really, because the state can change.

That is a general issue with poll. You cannot be sure that there
is still data to be read when you are ready to read.

	Regards
		Oliver