[alsa-devel] [RFC PATCH 1/2] ASoC: core: don't increase component module refcount unconditionally

[Pierre-Louis Bossart]

On 2/4/19 10:25 PM, Vinod Koul wrote:
> On 01-02-19, 12:07, Ranjani Sridharan wrote:
>> On Fri, 2019-02-01 at 23:42 +0530, Vinod Koul wrote:
>>> On 01-02-19, 11:22, Pierre-Louis Bossart wrote:
>>>> The ASoC core has for the longest time increased the module
>>>> reference
>>>> counts, even before the transition to the component model. This is
>>>> probably fine on most platforms, but it introduces a deadlock case
>>>> on
>>>> Intel devices with the Skylake and SOF drivers which cannot be
>>>> removed
>>>> due to their reference counts being modified by the core.
>>>>
>>>> In these 2 cases, the PCI or ACPI driver .probe creates a platform
>>>> device to let the machine driver .probe register the audio
>>>> card. Conversely the PCI or ACPI driver .remove will unregister the
>>>> platform device which results in the card being removed by the
>>>> machine
>>>> driver .remove.
>>>>
>>>> With ascii art, this can be represented as
>>>>
>>>> modprobe
>>>> snd_soc_skl/
>>>> soc-pci-dev/sof-acpci-dev  ----------> pci/acpi probe
>>>>         ^                                    |
>>>>         |                     ---------------|
>>>>         |                    |               |
>>>>         |                    V               V
>>>>      increase            register        register machine
>>>>      refcount            component       platform_device
>>>>         ^                                    |
>>>>         |                                    |
>>>>         |                                    V
>>>>      component <----   register card  <---- probe
>>>>      probe
>>>>
>>>> The issue is that by playing with the component's module reference
>>>> counts during the card registration, it's no longer possible to
>>>> remove
>>>> the module which controls the component. This can be shown, e.g.
>>>> with
>>>> the following error:
>>>>
>>>> root at plb-XPS-13-9350:~# lsmod | grep snd_soc_skl
>>>> snd_soc_skl           110592  1
>>>>
>>>> root at plb-XPS-13-9350:~# rmmod snd_soc_skl
>>>> rmmod: ERROR: Module snd_soc_skl is in use
>>> Yup, that would be correct, the inuse is due to the fact the sound
>>> card
>>> is up and someone needs to unload the sound card to remove the
>>> reference.
>>>
>>> That can be done by doing the rmmod of machine driver first and IIRC
>>> that would remove the sound card and drop the reference and then
>>> snd_soc_skl can be unloaded.
>> This doesnt seem to be the case though. The machine driver module
>> cannot be removed either because its refcnt is also > 0.
> At least this used to be the case when I used to try removal of modules
> on skl, doing the reverse of load order seemed to work for me back then.

Unfortunately module unload is broken with the skylake driver (kernel 
oops left and right), so there's no way of verifying your 
assertion...Other folks are trying to restore the capability but it's 
not been working for a very long time.

Beyond the conceptual issue with the reference count, my other worry is 
that the topology is created by the Skylake driver driver but freed when 
you remove the card, so you end-up with non-sensical data structures and 
configurations when you remove the skl driver. it's *really* recommended 
to remove the component which instantiated the topology first before 
removing the card.