Re: [syzbot] [sound?] possible deadlock in __snd_pcm_lib_xfer (2)

[Takashi Iwai <tiwai@suse.de>]

On Thu, 04 Sep 2025 12:20:56 +0200,
Sebastian Andrzej Siewior wrote:
> 
> On 2025-08-29 17:06:02 [-0700], syzbot wrote:
> > syzbot has bisected this issue to:
> > 
> > commit d2d6422f8bd17c6bb205133e290625a564194496
> > Author: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> > Date:   Fri Sep 6 10:59:04 2024 +0000
> > 
> >     x86: Allow to enable PREEMPT_RT.
> > 
> > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=12db5634580000
> > start commit:   07d9df80082b Merge tag 'perf-tools-fixes-for-v6.17-2025-08..
> > git tree:       upstream
> > final oops:     https://syzkaller.appspot.com/x/report.txt?x=11db5634580000
> > console output: https://syzkaller.appspot.com/x/log.txt?x=16db5634580000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=e1e1566c7726877e
> > dashboard link: https://syzkaller.appspot.com/bug?extid=10b4363fb0f46527f3f3
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=10307262580000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=17110242580000
> > 
> > Reported-by: syzbot+10b4363fb0f46527f3f3@syzkaller.appspotmail.com
> > Fixes: d2d6422f8bd1 ("x86: Allow to enable PREEMPT_RT.")
> 
> This is unfortunate. There is nothing that sound did wrong, it is rather
> special softirq handling in this case. We don't see this often because
> it requires that a timer is cancelled at the time it is running.
> The assumption made by sound is that spin_lock_irq() also disables
> softirqs. This is not the case on PREEMPT_RT.
> 
> The hunk below avoids the splat. Adding local_bh_disable() to
> spin_lock_irq() would cure it, too. It would also result in random
> synchronisation points across the kernel leading to something less
> usable.
> The imho best solution would to get rid of softirq_ctrl.lock which has
> been proposed
> 	https://lore.kernel.org/all/20250901163811.963326-4-bigeasy@linutronix.de/
> 
> Comments?

Thank you for the detailed analysis!  It enlightened me.

It'd be appreciated if this gets fixed in the softirq core side.
If nothing else flies, we can take your workaround, sure...


Takashi

> 
> diff --git a/sound/core/pcm_native.c b/sound/core/pcm_native.c
> --- a/sound/core/pcm_native.c
> +++ b/sound/core/pcm_native.c
> @@ -84,19 +84,24 @@ void snd_pcm_group_init(struct snd_pcm_group *group)
>  }
>  
>  /* define group lock helpers */
> -#define DEFINE_PCM_GROUP_LOCK(action, mutex_action) \
> +#define DEFINE_PCM_GROUP_LOCK(action, bh_lock, bh_unlock, mutex_action) \
>  static void snd_pcm_group_ ## action(struct snd_pcm_group *group, bool nonatomic) \
>  { \
> -	if (nonatomic) \
> +	if (nonatomic) { \
>  		mutex_ ## mutex_action(&group->mutex); \
> -	else \
> -		spin_ ## action(&group->lock); \
> +	} else { \
> +		if (IS_ENABLED(CONFIG_PREEMPT_RT) && bh_lock)	\
> +			local_bh_disable();			\
> +		spin_ ## action(&group->lock);			\
> +		if (IS_ENABLED(CONFIG_PREEMPT_RT) && bh_unlock)	\
> +			local_bh_enable();			\
> +	}							\
>  }
>  
> -DEFINE_PCM_GROUP_LOCK(lock, lock);
> -DEFINE_PCM_GROUP_LOCK(unlock, unlock);
> -DEFINE_PCM_GROUP_LOCK(lock_irq, lock);
> -DEFINE_PCM_GROUP_LOCK(unlock_irq, unlock);
> +DEFINE_PCM_GROUP_LOCK(lock, 0, 0, lock);
> +DEFINE_PCM_GROUP_LOCK(unlock, 0, 0, unlock);
> +DEFINE_PCM_GROUP_LOCK(lock_irq, 1, 0, lock);
> +DEFINE_PCM_GROUP_LOCK(unlock_irq, 0, 1, unlock);
>  
>  /**
>   * snd_pcm_stream_lock - Lock the PCM stream
> 
> 
> Sebastian