On 4/13/21 9:05 AM, Heikki Krogerus wrote:
On Tue, Apr 13, 2021 at 03:20:45PM +0300, Heikki Krogerus wrote:
On Mon, Apr 12, 2021 at 03:36:20PM -0500, Pierre-Louis Bossart wrote:
I took the code and split it in two for BYT/CHT (modified to remove devm_) and SoundWire parts (added as is).
https://github.com/thesofproject/linux/pull/2810
Both cases result in a refcount error on device_remove_sof when removing the platform device. I don't understand the code well enough to figure out what happens, but it's likely a case of the software node being removed twice?
Right. Because you are injecting the node to an already existing device, the node does not get linked with the device in sysfs. That would increment the reference count in a normal case. It all happens in the function software_node_notify(). Driver core calls it when a device is added and also when it's removed. In this case it is only called when it's removed.
I think the best way to handle this now is to simply not decrementing the ref count when you add the properties, so don't call fwnode_handle_put() there (but add a comment explaining why you are not calling it).
No, sorry... That's just too hacky. Let's not do that after all.
We can also fix this in the software node code. I'm attaching a patch that should make it possible to inject the software nodes also afterwards safely. This is definitely also not without its problems, but we can go with that if it works. Let me know.
I tested manually on bytcr w/ RT5640 and used the SOF CI farm to test the SoundWire cases, I don't see any issues with your patch. The refcount issue is gone and the module load/unload cycles don't report any problems.
Would you queue it for 5.13-rc1, or is this too late already?
For a better solution you would call device_reprobe() after you have injected the software node, but before that you need to modify device_reprobe() so it calls device_platform_notify() (which it really should call in any case). But this should probable be done later, separately.
thanks,
P.S.
Have you guys considered the possibility of describing the connections between all these components by using one of the methods that we now have for that in kernel, for example device graph? It can now be used also with software nodes (OF graph and ACPI device graph are of course already fully supported).
I must admit I've never heard of a 'device graph'. Any pointers or APIs you can think of? It's a good comment since we are planning to rework the SOF clients and machine driver handling.