Re: hda codec unbind refcount hang

[Takashi Iwai <tiwai@suse.de>]

On Fri, 09 Sep 2022 21:39:19 +0200,
Ville Syrjälä wrote:
> 
> On Fri, Sep 09, 2022 at 05:59:47PM +0200, Takashi Iwai wrote:
> > On Fri, 09 Sep 2022 17:45:25 +0200,
> > Ville Syrjälä wrote:
> > > 
> > > Hi Takashi,
> > > 
> > > commit 7206998f578d ("ALSA: hda: Fix potential deadlock at codec
> > > unbinding") introduced a problem on at least one of my older machines.
> > > 
> > > The problem happens when hda_codec_driver_remove() encounters a
> > > codec without any pcms (and thus the refcount is 1) and tries to
> > > call refcount_dec(). Turns out refcount_dec() doesn't like to be
> > > used for dropping the refcount to 0, and instead if spews a warning
> > > and does its saturate thing. The subsequent wait_event() is then
> > > permanently stuck waiting on the saturated refcount.
> > > 
> > > I've definitely seen the same kind of pattern used elsewhere
> > > in the kernel as well, so the fact that refcount_t can't be used
> > > to implement it is a bit of surprise to me. I guess most other
> > > places still use atomic_t instead.
> > 
> > Does the patch below work around it?  It seem to be a subtle
> > difference between refcount_dec() and refcount_dec_and_test().
> 
> Aye, this works.
> 
> Tested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>

Good to hear.

I think the below is slightly safer, assuring the other *_disconnect()
calls.

Could you give it a try again?  Once after confirming it works, I'll
re-submit and merge to my tree.


thanks,

Takashi

-- 8< --
From: Takashi Iwai <tiwai@suse.de>
Subject: [PATCH] ALSA: hda: Fix hang at HD-audio codec unbinding due to refcount saturation
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

We fixed the potential deadlock at dynamic unbinding the HD-audio
codec at the commit 7206998f578d ("ALSA: hda: Fix potential deadlock
at codec unbinding"), but ironically, this caused another potential
deadlock.  The current code uses refcount_dec() and waits for the
pending task with wait_event for dropping the refcount to 0.  This
works fine when PCMs are assigned and actually waiting for the
refcount drop.

Meanwhile, when there was no PCM assigned, the refcount_dec() call
itself was supposed to drop to zero -- alas, it doesn't in reality;
refcount_dec() complains, spews kernel warning and it saturates
instead of dropping to 0, due to the nature of refcount_dec()
implementation.  This eventually blocks the wait_event() wakeup and
the code get stuck there.

For avoiding the problem, we call refcount_dec_and_test() and skips
the sync-wait if it already reaches to zero.

The patch does a slight code reshuffling to make sure to invoke other
disconnect calls before the sync-wait, too.

Fixes: 7206998f578d ("ALSA: hda: Fix potential deadlock at codec unbinding")
Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: <stable@vger.kernel.org>
Link: https://lore.kernel.org/r/YxtflWQnslMHVlU7@intel.com
Signed-off-by: Takashi Iwai <tiwai@suse.de>
---
 sound/pci/hda/hda_bind.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sound/pci/hda/hda_bind.c b/sound/pci/hda/hda_bind.c
index cae9a975cbcc..1a868dd9dc4b 100644
--- a/sound/pci/hda/hda_bind.c
+++ b/sound/pci/hda/hda_bind.c
@@ -157,10 +157,10 @@ static int hda_codec_driver_remove(struct device *dev)
 		return codec->bus->core.ext_ops->hdev_detach(&codec->core);
 	}
 
-	refcount_dec(&codec->pcm_ref);
 	snd_hda_codec_disconnect_pcms(codec);
 	snd_hda_jack_tbl_disconnect(codec);
-	wait_event(codec->remove_sleep, !refcount_read(&codec->pcm_ref));
+	if (!refcount_dec_and_test(&codec->pcm_ref))
+		wait_event(codec->remove_sleep, !refcount_read(&codec->pcm_ref));
 	snd_power_sync_ref(codec->bus->card);
 
 	if (codec->patch_ops.free)
-- 
2.35.3