Re: [PATCH] tracing/osnoise: fix potential deadlock in cpu hotplug

[Steven Rostedt]

On Tue, 24 Mar 2026 15:06:16 +0800 (CST)
<hu.shengming@xxxxxxxxxx> wrote:

> From: luohaiyang10243395 <luo.haiyang@xxxxxxxxxx>
> 
> The following sequence may leads deadlock in cpu hotplug:
> 
>   CPU0                        |  CPU1
>                               |  schedule_work_on
>                               |
>   _cpu_down//set CPU1 offline |
>   cpus_write_lock             |
>                               |  osnoise_hotplug_workfn
>                               |    mutex_lock(&interface_lock);
>                               |    cpus_read_lock();  //wait cpu_hotplug_lock
>                               |
>                               |  cpuhp/1
>                               |    osnoise_cpu_die
>                               |      kthread_stop
>                               |        wait_for_completion //wait osnoise/1 exit
>                               |
>                               |  osnoise/1
>                               |    osnoise_sleep
>                               |      mutex_lock(&interface_lock); //deadlock
> 
> Fix by swap the order of cpus_read_lock() and mutex_lock(&interface_lock).

So the deadlock is due to the "wait_for_completion"?

How did you find this bug? Inspection, AI, triggered?

Thanks,

-- Steve


> 
> Signed-off-by: Luo Haiyang <luo.haiyang@xxxxxxxxxx>
> ---
>  kernel/trace/trace_osnoise.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/trace/trace_osnoise.c b/kernel/trace/trace_osnoise.c
> index dee610e465b9..be6cf0bb3c03 100644
> --- a/kernel/trace/trace_osnoise.c
> +++ b/kernel/trace/trace_osnoise.c
> @@ -2073,8 +2073,8 @@ static void osnoise_hotplug_workfn(struct work_struct *dummy)
>  	if (!osnoise_has_registered_instances())
>  		return;
> 
> -	guard(mutex)(&interface_lock);
>  	guard(cpus_read_lock)();
> +	guard(mutex)(&interface_lock);
> 
>  	if (!cpu_online(cpu))
>  		return;
> @@ -2237,11 +2237,11 @@ static ssize_t osnoise_options_write(struct file *filp, const char __user *ubuf,
>  	if (running)
>  		stop_per_cpu_kthreads();
> 
> -	mutex_lock(&interface_lock);
>  	/*
>  	 * avoid CPU hotplug operations that might read options.
>  	 */
>  	cpus_read_lock();
> +	mutex_lock(&interface_lock);
> 
>  	retval = cnt;
> 
> @@ -2257,8 +2257,8 @@ static ssize_t osnoise_options_write(struct file *filp, const char __user *ubuf,
>  			clear_bit(option, &osnoise_options);
>  	}
> 
> -	cpus_read_unlock();
>  	mutex_unlock(&interface_lock);
> +	cpus_read_unlock();
> 
>  	if (running)
>  		start_per_cpu_kthreads();
> @@ -2345,16 +2345,16 @@ osnoise_cpus_write(struct file *filp, const char __user *ubuf, size_t count,
>  	if (running)
>  		stop_per_cpu_kthreads();
> 
> -	mutex_lock(&interface_lock);
>  	/*
>  	 * osnoise_cpumask is read by CPU hotplug operations.
>  	 */
>  	cpus_read_lock();
> +	mutex_lock(&interface_lock);
> 
>  	cpumask_copy(&osnoise_cpumask, osnoise_cpumask_new);
> 
> -	cpus_read_unlock();
>  	mutex_unlock(&interface_lock);
> +	cpus_read_unlock();
> 
>  	if (running)
>  		start_per_cpu_kthreads();