On 29/2/24 08:00, Takashi Iwai wrote:
On Wed, 21 Feb 2024 13:37:10 +0100, Richard Fitzgerald wrote:
snd_soc_card_get_kcontrol() must be holding a read lock on card->controls_rwsem while walking the controls list.
Compare with snd_ctl_find_numid().
The existing function is renamed snd_soc_card_get_kcontrol_locked() so that it can be called from contexts that are already holding card->controls_rwsem (for example, control get/put functions).
There are few direct or indirect callers of snd_soc_card_get_kcontrol(), and most are safe. Three require changes, which have been included in this patch:
codecs/cs35l45.c: cs35l45_activate_ctl() is called from a control put() function so is changed to call snd_soc_card_get_kcontrol_locked().
codecs/cs35l56.c: cs35l56_sync_asp1_mixer_widgets_with_firmware() is called from control get()/put() functions so is changed to call snd_soc_card_get_kcontrol_locked().
fsl/fsl_xcvr.c: fsl_xcvr_activate_ctl() is called from three places, one of which already holds card->controls_rwsem:
- fsl_xcvr_mode_put(), a control put function, which will already be holding card->controls_rwsem.
- fsl_xcvr_startup(), a DAI startup function.
- fsl_xcvr_shutdown(), a DAI shutdown function.
To fix this, fsl_xcvr_activate_ctl() has been changed to call snd_soc_card_get_kcontrol_locked() so that it is safe to call directly from fsl_xcvr_mode_put(). The fsl_xcvr_startup() and fsl_xcvr_shutdown() functions have been changed to take a read lock on card->controls_rsem() around calls to fsl_xcvr_activate_ctl(). While this is not very elegant, it keeps the change small, to avoid this patch creating a large collateral churn in fsl/fsl_xcvr.c.
Analysis of other callers of snd_soc_card_get_kcontrol() is that they do not need any changes, they are not holding card->controls_rwsem when they call snd_soc_card_get_kcontrol().
Direct callers of snd_soc_card_get_kcontrol(): fsl/fsl_spdif.c: fsl_spdif_dai_probe() - DAI probe function fsl/fsl_micfil.c: voice_detected_fn() - IRQ handler
Indirect callers via soc_component_notify_control(): codecs/cs42l43: cs42l43_mic_shutter() - IRQ handler codecs/cs42l43: cs42l43_spk_shutter() - IRQ handler codecs/ak4118.c: ak4118_irq_handler() - IRQ handler codecs/wm_adsp.c: wm_adsp_write_ctl() - not currently used
Indirect callers via snd_soc_limit_volume(): qcom/sc8280xp.c: sc8280xp_snd_init() - DAIlink init function ti/rx51.c: rx51_aic34_init() - DAI init function
I don't have hardware to test the fsl/*, qcom/sc828xp.c, ti/rx51.c and ak4118.c changes.
Backport note: The fsl/, qcom/, cs35l45, cs35l56 and cs42l43 callers were added since the Fixes commit so won't all be present on older kernels.
Signed-off-by: Richard Fitzgerald rf@opensource.cirrus.com Fixes: 209c6cdfd283 ("ASoC: soc-card: move snd_soc_card_get_kcontrol() to soc-card")
It would be great if people could test the fsl/, qcom/, ti/rx51 and ak4418 drivers.
This fix itself looks correct, and I merged Mark's PR now.
But in general, it'd be better to use snd_ctl_find_id() and snd_ctl_find_id_unlocked() if possible. Those standard APIs can use the fast Xarray lookup, and especially for the case like many ASoC drivers that expose hundreds of kcontrols, the performance gain becomes significant.
I see that there is no snd_ctl_find_mixer_id_unlocked() variant, but it should be trivial to add.
Yes, I'll have a look at that. I was thinking that it would be good to have all the code to find controls in one place, instead of a special case for ASoC. But I decided to make the bugfix with minimum changes.
thanks,
Takashi