Re: [PATCH] ocfs2/cluster: hold a reference on the heartbeat thread

[Joseph Qi]

On 9/23/26 9:01 AM, Karl Mehltretter wrote:
> Since commit 688bc88e2046 ("ocfs2/cluster: keep heartbeat local node
> stable"), o2hb_thread() leaves its loop and returns when the local node
> changes, for example after "echo 0 > node/<name>/local". The thread
> was started with kthread_run() and nothing holds a reference to its
> task_struct, so the task is freed once it exits, while reg->hr_task
> still points to it.
> 
> Reading the region's pid attribute then reads the freed task, and
> removing the region calls kthread_stop() on it:
> 
>   BUG: KASAN: slab-use-after-free in o2hb_region_pid_show+0xb3/0xc0
>   refcount_t: addition on 0; use-after-free.
>   Oops: Oops: 0000 [#1] SMP KASAN NOPTI
>   RIP: 0010:kthread_stop+0xb1/0x390
> 
> The thread could already return by itself before, when heartbeat start
> was aborted or on an unclean stop, but the local node change makes it
> reachable from userspace at any time.
> 
> Create the thread parked, take a reference on it and only then wake it,

Just a wording nit. kthread_create() doesn't leave the thread parked in
the kthread_park() sense. __kthread_parkme() only runs after the first
wake_up_process(), so "parked" here may be a little confused.

> so the reference cannot race with the thread exiting. Drop it with
> kthread_stop_put().
> 
> Fixes: 688bc88e2046 ("ocfs2/cluster: keep heartbeat local node stable")
> Reported-by: Sashiko <sashiko-bot@xxxxxxxxxx>
> Closes: https://sashiko.dev/#/patchset/20260616074931.3774929-1-zzzccc427%40gmail.com
> Assisted-by: LLM
> Signed-off-by: Karl Mehltretter <kmehltretter@xxxxxxxxx>

The change looks fine to me.
Reviewed-by: Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx>

> ---
> Reproduced under QEMU x86_64 with KASAN, using one node over configfs
> and a loop device as the heartbeat region:
> 
>   echo 0 > cluster/c1/node/n0/local     # heartbeat thread returns
>   cat cluster/c1/heartbeat/<uuid>/pid   # reads the freed task_struct
>   rmdir cluster/c1/heartbeat/<uuid>     # kthread_stop() on it
> 
> With the patch, that sequence, a normal region removal and a start
> interrupted by a signal ran clean on two CPUs, five rounds each, and
> every heartbeat task_struct was freed.
> 
>  fs/ocfs2/cluster/heartbeat.c | 12 ++++++++----
>  1 file changed, 8 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/ocfs2/cluster/heartbeat.c b/fs/ocfs2/cluster/heartbeat.c
> index 1c3def99bb076..a4c8ea695f5cf 100644
> --- a/fs/ocfs2/cluster/heartbeat.c
> +++ b/fs/ocfs2/cluster/heartbeat.c
> @@ -1966,18 +1966,22 @@ static ssize_t o2hb_region_dev_store(struct config_item *item,
>  	atomic_set(&reg->hr_unsteady_iterations, (live_threshold * 3));
>  	o2hb_set_region_stopping(reg, false);
>  
> -	hb_task = kthread_run(o2hb_thread, reg, "o2hb-%s",
> -			      reg->hr_item.ci_name);
> +	hb_task = kthread_create(o2hb_thread, reg, "o2hb-%s",
> +				 reg->hr_item.ci_name);
>  	if (IS_ERR(hb_task)) {
>  		ret = PTR_ERR(hb_task);
>  		mlog_errno(ret);
>  		goto out;
>  	}
> +	/* The thread may exit on its own, so pin it before it can run. */
> +	get_task_struct(hb_task);
>  
>  	spin_lock(&o2hb_live_lock);
>  	reg->hr_task = hb_task;
>  	spin_unlock(&o2hb_live_lock);
>  
> +	wake_up_process(hb_task);
> +
>  	ret = wait_event_interruptible(o2hb_steady_queue,
>  				atomic_read(&reg->hr_steady_iterations) == 0 ||
>  				reg->hr_node_deleted);
> @@ -2022,7 +2026,7 @@ static ssize_t o2hb_region_dev_store(struct config_item *item,
>  		spin_unlock(&o2hb_live_lock);
>  
>  		if (hb_task)
> -			kthread_stop(hb_task);
> +			kthread_stop_put(hb_task);
>  
>  		o2hb_unmap_slot_data(reg);
>  
> @@ -2208,7 +2212,7 @@ static void o2hb_heartbeat_group_drop_item(struct config_group *group,
>  	spin_unlock(&o2hb_live_lock);
>  
>  	if (hb_task)
> -		kthread_stop(hb_task);
> +		kthread_stop_put(hb_task);
>  
>  	if (o2hb_global_heartbeat_active()) {
>  		spin_lock(&o2hb_live_lock);