Re: [net-next, v5, 02/11] ptp: support ptp physical/virtual clocks conversion

On Sat, Aug 07, 2021 at 07:22:59AM -0700, Richard Cochran wrote:

On Fri, Aug 06, 2021 at 06:15:29PM -0700, Vinicius Costa Gomes wrote:

quoted

quoted

 int ptp_clock_unregister(struct ptp_clock *ptp)
 {
+	if (ptp_vclock_in_use(ptp)) {
+		pr_err("ptp: virtual clock in use\n");
+		return -EBUSY;
+	}
+

None of the drivers (that I looked) expect ptp_clock_unregister() to
return an error.

So, what should we do?
 1. Fix all the drivers to return an error on module unloading (that's
 usually the path ptp_clock_unregister() is called)?
 2. Remove all the PTP virtual clocks when the physical clock is
 unregistered?

This:

 3. Let the vclocks hold a reference to the underlying posix dynamic clock.

So even if the vclock holds a reference to the underlying POSIX clock,
that won't prevent the hardware driver from unbinding, and further
gettime() calls on the vclock from faulting, will it?

What about:

4. Create a device link with the vclock being a consumer and the parent
   clock being a supplier? This way, ptp_vclock_unregister() is
   automatically called whenever (and before) ptp_clock_unregister() is.

https://www.kernel.org/doc/html/latest/driver-api/device_link.html