Hi Jonas, On 03/07/2018 08:55 AM, Jonas Bonn wrote:
Here's a couple of fixups to the qmimodem atoms in order handle service lifetime.
I've taken the first two patches, but the whole unregister_all thing has horrible side-effects since you're force un-registering notifications that you don't even own. Ideally all of QMI core classes need to be fixed and handle cancellation of requests properly...
The QMI services are 'shared' between various atoms and this has some subtle problems. Patch 3, for the LTE atom, handles an issue where atoms from different ofono states share an atom. For things to work properly, only atoms from the same state should be sharing a service.
That needs to be fixed. The initial implementation of QMI assumed that a service would only live in a single atom. There were one or two exceptions, hence the shared service stuff was bolted on. However, these assumptions were proven to be false and resulted in lots of issues with lifetimes and cleanups. You're finding this out now. I believe I've commented on how QMI should be fixed before, but I'll take this opportunity to do so again :)
For things to work better, requests should probably be queued on the actual service instance that the atom is using and not on the underlying shared device queue. That way, the responses for a destroyed service instance could just be dropped instead of being propagated to a potentially free'd atom. This is a rather intrusive rework of the QMI core though, so the patches here will need to do for now.
So I would rather fix this cleanly than taking a 100+ line change that will be useless, confusing and completely different from the rest of the code if stuff inside qmi.c was done properly. Perhaps an easier solution would be to move gprs and radio-settings atoms to post_sim for now?
The patches here resolve a concrete segfault that arises when the modem is set offline.
The proper way of fixing this is to introduce the concept of groups. So that all requests for a particular atom belong to a single group. That way when the atom goes away, the entire group is canceled. GAtChat does this already, take a look at g_at_chat_clone and how various atoms use that. qmi_service might want to follow the same pattern and the shared vs non-shared qmi service distinction needs to go away.
Alternatively, you can simply support the group concept directly in the API. Take a look at how drivers/mbimmodem/mbim.c implements mbim_device_cancel_group.
Regards, -Denis _______________________________________________ ofono mailing list email@example.com https://lists.ofono.org/mailman/listinfo/ofono