Re: [igt-dev] [PATCH i-g-t 1/2] lib/igt_kmod: Wait for a kmod to finish its probe before unloding it.

["Dixit, Ashutosh" <ashutosh.dixit@intel.com>]

On Thu, 31 Mar 2022 16:09:32 -0700, Ceraolo Spurio, Daniele wrote:
>
> On 3/31/2022 3:52 PM, Dixit, Ashutosh wrote:
> > On Wed, 30 Mar 2022 11:32:58 -0700, Daniele Ceraolo Spurio wrote:
> >> Modprobing i915 can result in other dependent modules being loaded
> >> automatically as a follow up. This means that by the time the i915 load
> >> function returns these modules may still be going through their init, so
> >> if we try to immediately unload i915 again we will fail with an
> >> -EBUSY.
> > Can you explain a bit where we see this happening and if any tests fail
> > because of this? My expectation would be if module A depends on module B,
> > loading A will cause B to load, but loading B will not cause A to load. So
> > what is causing these dependent modules to load when we load i915? Is there
> > e.g. something in /etc/modprobe.d or udev which is causing this to happen?
>
> This is a new scenario, introduced with the GSC series where i915 exposes
> the GSC as an aux device and the mei-gsc module binds to it (link to the
> kernel series in the cover letter). The kernel handles the addition of an
> aux device similarly to a device hot-plug, with the relevant driver
> (mei-gsc in this case) being automatically probed, without requiring an
> explicit modprobe. This aux device probe happens separately from the parent
> device probe, which means that the i915 probe can complete and return
> before the mei-gsc probe is done. The fact that mei-gsc binds to an i915
> aux device creates a dependency, even if mei-gsc doesn't actually
> explicitly use any function exported by i915, so we need to unload it to be
> able to unload i915.
>
> > Currently in IGT we maintain a list of modules depenedent on i915 which we
> > unload prior to unloading i915. An alternative approach to this patch may
> > be to maintain a similar list in igt_i915_driver_load() and load dependent
> > modules after loading i915 as long as they are present on the system. So we
> > would not need to wait for these dependent modules to complete loading when
> > we unload i915. Do you think this is feasible?
>
> No, as mentioned above the module is loaded automatically.

OK, thanks for the explanation, this corresponds to my understanding of how
these buses (in this case aux bus) work. I have a couple more questions
which let me ask separately. Thanks.