Re: [PATCH net] igc: Fix LED-related deadlock on driver unbind

[Lukas Wunner <lukas@wunner.de>]

On Mon, Apr 22, 2024 at 04:46:28PM -0700, Jacob Keller wrote:
> To me, using devm from the PCI device should be equivalent to managing
> it manually within the igc_remove() function.

It is not equivalent because the ordering is different:

igc_remove() is called before device-managed resources are released:

__device_release_driver()
  device_remove()             # invokes igc_remove()
  device_unbind_cleanup()
    devres_release_all()      # releases device-managed resources

If you unregister LEDs explicitly in igc_remove() before unregistering
the netdev and disabling PCI device access, everything's fine.

If you instead use devm_led_classdev_register(), the LEDs would still
be registered and available in sysfs after igc_remove() has torn down
everything, which is bad.

You'd have to use devm_*() for all initialization steps in igc_probe()
to make this work.  With devm_*() it's generally all or nothing.

(There are exceptions:  Using devm_*() just for memory allocations is
fine as those can safely be freed even if everything else is torn down.)

Thanks,

Lukas