hda codec unbind refcount hang

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.

Quoted 13 lines · Ville Syrjälä

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.

Quoted 4 lines

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().

thanks,

Takashi

-- 8< --

sound/pci/hda/hda_bind.c +5 -4

@@ -157,10 +157,11 @@ 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)) {
+		snd_hda_codec_disconnect_pcms(codec);
+		snd_hda_jack_tbl_disconnect(codec);
+		wait_event(codec->remove_sleep, !refcount_read(&codec->pcm_ref));
+	}
 	snd_power_sync_ref(codec->bus->card);
 
 	if (codec->patch_ops.free)

Quoted 20 lines · Takashi Iwai

Quoted 17 lines · Ville Syrjälä

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.

Quoted 2 lines

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 <[email protected]>

Quoted 22 lines · Ville Syrjälä

Quoted 20 lines · Takashi Iwai

Quoted 17 lines · Ville Syrjälä

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.

Quoted 6 lines

Quoted 2 lines

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 <[email protected]>

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 <[email protected]>
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 <[email protected]>

Cc <[email protected]>

Link https://lore.kernel.org/r/[email protected] open here

Signed-off-by Takashi Iwai <[email protected]>

---
 sound/pci/hda/hda_bind.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

sound/pci/hda/hda_bind.c +2 -2

index cae9a975cbcc..1a868dd9dc4b 100644
@@ -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)

Signature · 1 line

2.35.3