ALSA: hda/cs_dsp_ctl: Fix mutex inversion when creating controls
Redesign the creation of ALSA controls so that the cs_dsp pwr_lock is not held when calling snd_ctl_add(). Instead of creating the ALSA control from the cs_dsp control_add callback, do it after cs_dsp_power_up() has completed. The existing functions are changed to return void instead of passing errors back - this duplicates the original behaviour, as cs_dsp does not abort firmware load if creation of a control fails. It is safe to walk the control list without taking any mutex provided that the caller is not trying to load a new firmware or remove the driver in parallel. There is no other situation that the list can change. So the caller can trigger creation of ALSA controls after cs_dsp_power_up() has returned. A cs_dsp control will have a non-NULL priv pointer if we have created an ALSA control. With the previous code the ALSA controls were created from the cs_dsp control_add callback. But this is called with pwr_lock held (as it is part of the DSP power-up sequence). The kernel lock checking will show a mutex inversion between this and the control creation path: control_add pwr_lock held, takes controls_rwsem (in snd_ctl_add) get/put controls_rwsem held, takes pwr_lock to call cs_dsp. This is not completely theoretical. Although the time window is very small, it is possible for these to run in parallel and deadlock the old implementation. Signed-off-by: Richard Fitzgerald <rf@opensource.cirrus.com> Signed-off-by: Stefan Binding <sbinding@opensource.cirrus.com> Link: https://lore.kernel.org/r/20221011143552.621792-4-sbinding@opensource.cirrus.comSigned-off-by: Takashi Iwai <tiwai@suse.de>
Showing
Please register or sign in to comment