ALSA: pcm: Disable bottom softirqs as part of spin_lock_irq() on PREEMPT_RT

[PATCH] ALSA: pcm: Disable bottom softirqs as part of spin_lock_irq() on PREEMPT_RT

Posted by Sebastian Andrzej Siewior 2 weeks, 2 days ago

snd_pcm_group_lock_irq() acquires a spinlock_t and disables interrupts
via spin_lock_irq(). This also implicitly disables the handling of
softirqs such as TIMER_SOFTIRQ.
On PREEMPT_RT softirqs are preemptible and spin_lock_irq() does not
disable them. That means a timer can be invoked during spin_lock_irq()
on the same CPU. Due to synchronisations reasons local_bh_disable() has
a per-CPU lock named softirq_ctrl.lock which synchronizes individual
softirq against each other.
syz-bot managed to trigger a lockdep report where softirq_ctrl.lock is
acquired in hrtimer_cancel() in addition to hrtimer_run_softirq(). This
is a possible deadlock.

The softirq_ctrl.lock can not be made part of spin_lock_irq() as this
would lead to too much synchronisation against individual threads on the
system. To avoid the possible deadlock, softirqs must be manually
disabled before the lock is acquired.

Disable softirqs before the lock is acquired on PREEMPT_RT.

Reported-by: syzbot+10b4363fb0f46527f3f3@syzkaller.appspotmail.com
Fixes: d2d6422f8bd1 ("x86: Allow to enable PREEMPT_RT.")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---

I don't see a way around this given the report. I don't see how to
address this within the softirq. Taking this lock as part of every
spin_lock_irq() would be a mess and while testing I didn't even manage
to boot the machine. So I probably missed a detail (but then I would
only know how mad it really is).

This can be an intermediate solution until
	https://lore.kernel.org/all/20250901163811.963326-4-bigeasy@linutronix.de/

gets merged and the !PREEMPT_RT_NEEDS_BH_LOCK case the default (i.e. not
a config option anymore).

 sound/core/pcm_native.c | 21 +++++++++++++--------
 1 file changed, 13 insertions(+), 8 deletions(-)

diff --git a/sound/core/pcm_native.c b/sound/core/pcm_native.c
index 1eab940fa2e5a..68bee40c9adaf 100644
--- 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, false, false, lock);
+DEFINE_PCM_GROUP_LOCK(unlock, false, false, unlock);
+DEFINE_PCM_GROUP_LOCK(lock_irq, true, false, lock);
+DEFINE_PCM_GROUP_LOCK(unlock_irq, false, true, unlock);
 
 /**
  * snd_pcm_stream_lock - Lock the PCM stream
-- 
2.51.0

Re: [PATCH] ALSA: pcm: Disable bottom softirqs as part of spin_lock_irq() on PREEMPT_RT

Posted by Takashi Iwai 2 weeks, 2 days ago

On Mon, 15 Sep 2025 17:28:51 +0200,
Sebastian Andrzej Siewior wrote:
> 
> snd_pcm_group_lock_irq() acquires a spinlock_t and disables interrupts
> via spin_lock_irq(). This also implicitly disables the handling of
> softirqs such as TIMER_SOFTIRQ.
> On PREEMPT_RT softirqs are preemptible and spin_lock_irq() does not
> disable them. That means a timer can be invoked during spin_lock_irq()
> on the same CPU. Due to synchronisations reasons local_bh_disable() has
> a per-CPU lock named softirq_ctrl.lock which synchronizes individual
> softirq against each other.
> syz-bot managed to trigger a lockdep report where softirq_ctrl.lock is
> acquired in hrtimer_cancel() in addition to hrtimer_run_softirq(). This
> is a possible deadlock.
> 
> The softirq_ctrl.lock can not be made part of spin_lock_irq() as this
> would lead to too much synchronisation against individual threads on the
> system. To avoid the possible deadlock, softirqs must be manually
> disabled before the lock is acquired.
> 
> Disable softirqs before the lock is acquired on PREEMPT_RT.
> 
> Reported-by: syzbot+10b4363fb0f46527f3f3@syzkaller.appspotmail.com
> Fixes: d2d6422f8bd1 ("x86: Allow to enable PREEMPT_RT.")
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
> 
> I don't see a way around this given the report. I don't see how to
> address this within the softirq. Taking this lock as part of every
> spin_lock_irq() would be a mess and while testing I didn't even manage
> to boot the machine. So I probably missed a detail (but then I would
> only know how mad it really is).
> 
> This can be an intermediate solution until
> 	https://lore.kernel.org/all/20250901163811.963326-4-bigeasy@linutronix.de/
> 
> gets merged and the !PREEMPT_RT_NEEDS_BH_LOCK case the default (i.e. not
> a config option anymore).

I applied now to for-next branch for 6.18.
It's already at a late stage for 6.17 release, and the issue doesn't
seem like an urgent regression to be fixed.


Thanks!

Takashi

Re: [PATCH] ALSA: pcm: Disable bottom softirqs as part of spin_lock_irq() on PREEMPT_RT

Posted by Sebastian Andrzej Siewior 2 weeks, 2 days ago

On 2025-09-16 11:43:53 [+0200], Takashi Iwai wrote:
> I applied now to for-next branch for 6.18.
> It's already at a late stage for 6.17 release, and the issue doesn't
> seem like an urgent regression to be fixed.

Sure. Thank you.

> Thanks!
> 
> Takashi

Sebastian