On Tue, Sep 29, 2020 at 12:00:36PM -0400, Alan Stern wrote: > On Mon, Sep 28, 2020 at 06:43:55PM -0700, Matthias Kaehlcke wrote: > > > Have you tried manually unbinding and rebinding the two drivers a few > > > times to make sure they will still work? > > > > I went through a few dozen bund/unbind cycles for both drivers and things > > looked good overall, but then last minute I found that determining whether > > wakeup capable devices are connected doesn't always work as (I) expected. > > I didn't see this earlier, it seems to be reproduce more easily after > > unbinding and rebinding the platform driver. > > > > During development I already noticed that usb_wakeup_enabled_descendants() > > returns a cached value, which was a problem for an earlier version of the > > driver. The values are updated by hub_suspend(), my (flawed) assumption > > was that the USB driver would always suspend before the platform driver. > > This generally seems to be the case on my development platform after boot, > > but not necessarily after unbinding and rebinding the driver. Using the > > _suspend_late hook instead of _suspend seems to be a reliable workaround. > > Yes, for unrelated (i.e., not in a parent-child relation) devices, the > PM subsystem doesn't guarantee ordering of suspend and resume callbacks. > You can enforce the ordering by using device_pm_wait_for_dev(). But the > suspend_late approach seems like a better solution in this case. Thanks for the confirmation. Good to know about device_pm_wait_for_dev(), even if we are not going to use it in this case. > > > I'm a little concerned about all the devm_* stuff in here; does that > > > get released when the driver is unbound from the device or when the device > > > is unregistered? And if the latter, what happens if you have multiple > > > sysfs attribute groups going at the same time? > > > > The memory gets released when the device is unbound: > > > > device_release_driver > > device_release_driver_internal > > __device_release_driver > > devres_release_all > > > > Anyway, if you prefer I can change the driver to use kmalloc/kfree. > > No, that's fine. I just wasn't sure about this and wanted to check. I think the only concern would be a scenario where the USB devices are unbound and rebound over and over again, which would result in a struct udev_node being kept around for every bind until the platform device is removed. It seems unlikely and shouldn't be a big problem as long as the number of bind/unbind cycles is in the thousands rather than millions.
