Re: [PATCH net-next] rhashtable: unnecessary to use delayed work

[Thomas Graf <tgraf@suug.ch>]

On 01/13/15 at 05:00pm, Ying Xue wrote:
> When we put our declared work task in the global workqueue with
> schedule_delayed_work(), its delay parameter is always zero.
> Therefore, we should define a normal work in rhashtable structure
> instead of a delayed work.
> 
> Signed-off-by: Ying Xue <ying.xue@windriver.com>
> Cc: Thomas Graf <tgraf@suug.ch>

> @@ -914,7 +914,7 @@ void rhashtable_destroy(struct rhashtable *ht)
>  
>  	mutex_lock(&ht->mutex);
>  
> -	cancel_delayed_work(&ht->run_work);
> +	cancel_work_sync(&ht->run_work);
>  	bucket_table_free(rht_dereference(ht->tbl, ht));
>  
>  	mutex_unlock(&ht->mutex);

I like the patch!

I think it introduces a possible dead lock though (see below). OTOH, it
could actually explain the reason for the 0day lock debug splash that
was reported.

Dead lock: The worker could already have been kicked off but was
interrupted before it acquired ht->mutex. rhashtable_destroy() is
called and acquired ht->mutex. cancel_work_sync() waits for worker to
finish while holding ht->mutex. Worker can't finish because it needs to
acquire ht->mutex to do so.

For the very same reason the reported warning could have been triggered.
Instead of the dead lock, it would have called bucket_table_free()
with a deferred resizer still underway.

What about we do something like this?

void rhashtable_destroy(struct rhashtable *ht)
{
        ht->being_destroyed = true;
	cancel_work_sync(&ht->run_work);

	mutex_lock(&ht->mutex);
	bucket_table_free(rht_dereference(ht->tbl, ht));
	mutex_unlock(&ht->mutex);
}

If you agree we can explain this shortly in the commit message and add:
Fixes: 97defe1 ("rhashtable: Per bucket locks & deferred expansion/shrinking")