[Intel-gfx] [PATCH igt] module_reload: remove snd_hda_intel

[Ben Widawsky]

On Wed, Oct 23, 2013 at 09:42:06PM +0300, Jani Nikula wrote:
> On Wed, 23 Oct 2013, Paulo Zanoni <przanoni at gmail.com> wrote:
> > 2013/10/23 Ben Widawsky <ben at bwidawsk.net>:
> >> On Mon, Oct 21, 2013 at 11:56:09AM -0200, Paulo Zanoni wrote:
> >>> From: Paulo Zanoni <paulo.r.zanoni at intel.com>
> >>>
> >>> The audio driver uses the power well provided by our driver, so on
> >>> Haswell we can't "rmmod i915" if we don't "rmmod snd_hda_intel" first.
> >>>
> >>> The problem with removing snd_hda_intel is that we also need to kill
> >>> its users. On the specific machine I tested, the only user seem to be
> >>> alsactl, but on other machines this may change. IMHO we should leave
> >>> the "kill user space" step to whoever is running the script, but
> >>> Daniel asked me to put it here so we have a better chance of Just
> >>> Working on QA's machines.
> >>>
> >>> Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
> >>> ---
> >>>  tests/module_reload | 6 ++++++
> >>>  1 file changed, 6 insertions(+)
> >>>
> >>> diff --git a/tests/module_reload b/tests/module_reload
> >>> index 3d8cb85..c1fd395 100755
> >>> --- a/tests/module_reload
> >>> +++ b/tests/module_reload
> >>> @@ -18,6 +18,10 @@ if ! echo 0 > /sys/class/vtconsole/vtcon1/bind ; then
> >>>       exit 77
> >>>  fi
> >>>
> >>> +# The sound driver uses our power well
> >>> +pkill alsactl
> >>> +rmmod snd_hda_intel &> /dev/null
> >>> +
> >>>  #ignore errors in ips - gen5 only
> >>>  rmmod intel_ips &> /dev/null
> >>>  rmmod i915
> >>> @@ -38,6 +42,8 @@ fi
> >>>  modprobe i915
> >>>  echo 1 > /sys/class/vtconsole/vtcon1/bind
> >>>
> >>> +modprobe snd_hda_intel
> >>> +
> >>>  # try to run something
> >>>  $SOURCE_DIR/gem_exec_nop > /dev/null && echo "module successfully loaded again"
> >>>
> >>
> >> I don't really pretend to understand how module dependencies are
> >> calculated. However, I would think the dependency should be *somewhere*
> >> and so modprobe -r i915 should work.
> >
> > It's the opposite: "modprobe -r i915" will remove drm.ko,
> > drm_kms_helper, i2c-algo-bit and video, but won't remove snd_hda_intel
> > (and if you don't remove it first, it will complain that i915.ko is
> > still in use).
> >
> >>
> >> Don't read that as a nak. Rather, a wtf is with the existing stuff?
> >
> > I think the dependency is magically tracked by depmod when it sees the
> > Audio driver uses a symbol exported by i915.ko. But a "modinfo
> > snd_hda_intel" doesn't show i915.ko on the "depends" section, which is
> > strange. I'm also not an expert on this area...
> 
> This is because snd_hda_intel does not depend on i915, to make it
> possible to use snd_hda_intel without i915. Which makes sense. Instead,
> it uses symbol_request() to be able to call a couple of i915 functions
> when available.
> 
> I don't know if there's a way to check this runtime dependency from
> userspace. But obviously you can't remove i915 while snd_hda_intel holds
> function pointers to i915. (In fact, intel_ips is the same.)
> 
> BR,
> Jani.
> 
> 
> -- 
> Jani Nikula, Intel Open Source Technology Center

Looking around it seems a fairly common idiom is to do the
symbol_(get|request) when you need to call the function, and release it
when you're no longer using it. Since these functions we're talking
about require waking the GPU, I think speed shouldn't be a concern.

As you say, it makes sense to be able to load audio without i915 -
however looking at the code, it seems like audio will fail if i915 is
not loaded.

Also, it looks like there should be a way to make userspace aware of the
dependency since it calls back in to userspace (I find it funny that the
name of the function uses 4 underscores: ____call_usermodehelper).


-- 
Ben Widawsky, Intel Open Source Technology Center