On Thu, Jan 9, 2020 at 1:36 PM Tzung-Bi Shih tzungbi@google.com wrote:
On Wed, Jan 8, 2020 at 7:50 PM Marek Szyprowski m.szyprowski@samsung.com wrote:
This fixes the following lockdep warning observed on Exynos4412-based Odroid U3 board: ====================================================== -> #1 (&card->controls_rwsem){++++}: snd_ctl_add_replace+0x3c/0x84 dapm_create_or_share_kcontrol+0x24c/0x2e0 snd_soc_dapm_new_widgets+0x308/0x594 snd_soc_bind_card+0x80c/0xad4 devm_snd_soc_register_card+0x34/0x6c odroid_audio_probe+0x288/0x34c platform_drv_probe+0x6c/0xa4
I noticed the stack is a little different than the last time (odroid_audio_probe vs. asoc_simple_probe). Did you use the same machine to test?
asoc_simple_probe+0x244/0x4a0 platform_drv_probe+0x6c/0xa4
(https://mailman.alsa-project.org/pipermail/alsa-devel/2019-December/160142.h...)
I would like to spend some time to find the root cause. It would be a little challenging though (I have no real runtime to test...).
After a few hours of study, I failed to find the reason to cause the possible circular locking. And would need more of your input.
Followed the information provided in the message (https://mailman.alsa-project.org/pipermail/alsa-devel/2019-December/160535.h...). As the message said "snd_soc_of_get_dai_link_codecs() return -EPROBE_DEFER". The snd_soc_of_get_dai_link_codecs( ) is before devm_snd_soc_register_card( ), and I didn't find any side effects in odroid_audio_probe( ).
Only a very minor issue: snd_soc_of_put_dai_link_codecs(codec_link) will be called twice. One in snd_soc_of_get_dai_link_codecs( ) when return -EPROBE_DEFER; another one is under the label "err_put_cpu_dai". (https://elixir.bootlin.com/linux/v5.5-rc5/source/sound/soc/samsung/odroid.c#...) The code doesn't generate any side effects because of snd_soc_of_put_dai_link_codecs( )'s robustness.
Another minor thing: odroid_card_dais is not immutable but odroid_audio_probe( ) would try to modify it (https://elixir.bootlin.com/linux/v5.5-rc5/source/sound/soc/samsung/odroid.c#...). Again, I don't think it would produce any troubles. I guess no machine would have multiple sound cards, share the same machine driver, and unbind/bind in runtime.
It is weird: userspace should not see things (e.g. no controlC0) until snd_card_register( ).
(based on today's broonie/sound.git/for-next) I would like to provide you more information about this statement to help you find further information. When userspace can see the control device? Ideally, snd_soc_bind_card( ) -> snd_card_register( ) -> snd_device_register_all( ) -> __snd_device_register( ) -> snd_ctl_dev_register( ) -> snd_register_device( ). If you look at the calling stack of possible circular locking, snd_soc_dapm_new_widgets( ) is before snd_card_register( ). That's why we think userspace should not see control devices (i.e. controlC0, controlC1, ...) and should not be able to set mixer control via ioctl( ).
As this may not directly be related to the issue, could you share the init script of alsactl in your system? Does it follow the convention? (i.e. sound card is ready when receives controlC* changed event in udev rule 78-sound-card.rules)
- when userspace init scripts (alsactl) enumerates devices
(https://mailman.alsa-project.org/pipermail/alsa-devel/2019-December/160535.h...)