[Bug 204275] bluetoothd consumes 100% cpu on keyboard disconnect

[bugzilla-daemon@bugzilla.kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=204275

--- Comment #8 from Steven Newbury (steve@snewbury.org.uk) ---
Luiz, did you see (In reply to Luiz Von Dentz from comment #6)
> Ive sent a fix upstream:
> 
> https://lore.kernel.org/linux-bluetooth/20190724110151.4258-1-luiz.
> dentz@gmail.com/T/#u
> 
> Let me know if that works.

Luiz, did you see the patch I attached?  Perhaps it's better to remove the old
watcher first than not add another if it exists, but it shouldn't happen
anyway.  The explicit remove should happen to catch the case where the
connection is broken before being fully initialised but the callback should get
triggered once data starts to flow over the encrypted channel.

In my tests, allowing the channel to fully close by removing the sec_watch
prevents the keyboard from ever connecting, this is because the callback never
gets triggered.  I believe the bug is caused by the refcount sec_watch creates
not being automatically removed because the callback is left pending, it
doesn't get removed on channel shutdown, but has the side-effect of making it
work with the channel being already open when the device reconnects because it
triggers the callback.

What made it work every time is changing sec_watch to wait for G_IO_IN instead
of G_IO_OUT, but presumably it is G_IO_OUT for a reason?

-- 
You are receiving this mail because:
You are the assignee for the bug.