Re: [Intel-gfx] [PATCH 1/2] module: add a function to add module references

[Lucas De Marchi]

On Fri, Apr 29, 2022 at 11:23:51AM +0100, Mauro Carvalho Chehab wrote:
> Em Fri, 29 Apr 2022 12:10:07 +0200
> Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> escreveu:
>
> > On Fri, Apr 29, 2022 at 10:15:03AM +0100, Mauro Carvalho Chehab wrote:
> > > HI Greg,
> > >
> > > Em Fri, 29 Apr 2022 10:30:33 +0200
> > > Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> escreveu:
> > >
> > > > On Fri, Apr 29, 2022 at 09:07:57AM +0100, Mauro Carvalho Chehab wrote:
> > > > > Hi Daniel,
> > > > >
> > > > > Em Fri, 29 Apr 2022 09:54:10 +0200
> > > > > Daniel Vetter <daniel@xxxxxxxx> escreveu:
> > > > >
> > > > > > On Fri, Apr 29, 2022 at 07:31:15AM +0100, Mauro Carvalho Chehab wrote:
> > > > > > > Sometimes, device drivers are bound using indirect references,
> > > > > > > which is not visible when looking at /proc/modules or lsmod.
> > > > > > >
> > > > > > > Add a function to allow setting up module references for such
> > > > > > > cases.
> > > > > > >
> > > > > > > Reviewed-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > > > > > > Signed-off-by: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
> > > > > >
> > > > > > This sounds like duct tape at the wrong level. We should have a
> > > > > > device_link connecting these devices, and maybe device_link internally
> > > > > > needs to make sure the respective driver modules stay around for long
> > > > > > enough too. But open-coding this all over the place into every driver that
> > > > > > has some kind of cross-driver dependency sounds terrible.
> > > > > >
> > > > > > Or maybe the bug is that the snd driver keeps accessing the hw/component
> > > > > > side when that is just plain gone. Iirc there's still fundamental issues
> > > > > > there on the sound side of things, which have been attempted to paper over
> > > > > > by timeouts and stuff like that in the past instead of enforcing a hard
> > > > > > link between the snd and i915 side.
> > > > >
> > > > > I agree with you that the device link between snd-hda and the DRM driver
> > > > > should properly handle unbinding on both directions. This is something
> > > > > that require further discussions with ALSA and DRM people, and we should
> > > > > keep working on it.
> > > > >
> > > > > Yet, the binding between those drivers do exist, but, despite other
> > > > > similar inter-driver bindings being properly reported by lsmod, this one
> > > > > is invisible for userspace.
> > > > >
> > > > > What this series does is to make such binding visible. As simple as that.
> > > >
> > > > It also increases the reference count, and creates a user/kernel api
> > > > with the symlinks, right?  Will the reference count increase prevent the
> > > > modules from now being unloadable?
> > > >
> > > > This feels like a very "weak" link between modules that should not be
> > > > needed if reference counting is implemented properly (so that things are
> > > > cleaned up in the correct order.)
> > >
> > > The refcount increment exists even without this patch, as
> > > hda_component_master_bind() at sound/hda/hdac_component.c uses
> > > try_module_get() when it creates the device link.
> >
> > Ok, then why shouldn't try_module_get() be creating this link instead of
> > having to manually do it this way again?  You don't want to have to go
> > around and add this call to all users of that function, right?
>
> Works for me, but this is not a too trivial change, as the new
> try_module_get() function will require two parameters, instead of one:
>
> 	- the module to be referenced;
> 	- the module which will reference it.
>
> On trivial cases, one will be THIS_MODULE, but, in the specific case
> of snd_hda, the binding is done via an ancillary routine under
> snd_hda_core, but the actual binding happens at snd_hda_intel.
>
> Ok, we could add a __try_module_get() (or whatever other name that
> would properly express what it does) with two parameters, and then
> define try_module_get() as:
>
> 	#define try_module_get(mod) __try_module_get(mod, THIS_MODULE)

agree that this should be done at this level rather than open coding it
at every driver. Main improvement being fixed here regardless of the
snd-hda-intel issue is to properly annotate what is holding a module.

Right now we have 1) symbol module dependencies; 2) kernel references;
3) userspace references. With (2) and (3) being unknown to the user from
lsmod pov. Handling this any time try_module_get() is called would make
(2) visible to lsmod.

Paired with fixes to the (unreleased) kmod 30[1], this allows `modprobe
-r --remove-holders <module>` to also try removing the holders before
removing the module itself.

thanks
Lucas De Marchi

[1] https://lore.kernel.org/linux-modules/20220329090912.geymr6xk7taq4rtq@xxxxxxxxxxxxxxxxxxxxxxxxxxx/T/#t


> Would that work for you?
>
> Regards,
> Mauro