[PATCH] nscd: bump GC cycle during cache pruning (bug 26130)

[Florian Weimer]

* Andreas Schwab:

> While nscd prunes a cache it becomes inconsistent temporarily, which is
> visible to clients if that cache is shared.  Bump the GC cycle counter so
> that the clients notice the modification window.
> ---
>  nscd/cache.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
>
> diff --git a/nscd/cache.c b/nscd/cache.c
> index 94163d9ce3..2ceac94c23 100644
> --- a/nscd/cache.c
> +++ b/nscd/cache.c
> @@ -453,6 +453,11 @@ prune_cache (struct database_dyn *table, time_t now, int fd)
>  	  pthread_rwlock_wrlock (&table->lock);
>  	}
>  
> +      /* Now we start modifying the data.  Make sure all readers of the
> +	 data are aware of this and temporarily don't use the data.  */
> +      ++table->head->gc_cycle;
> +      assert ((table->head->gc_cycle & 1) == 1);
> +
>        while (first <= last)
>  	{
>  	  if (mark[first])
> @@ -493,6 +498,10 @@ prune_cache (struct database_dyn *table, time_t now, int fd)
>  	  ++first;
>  	}
>  
> +      /* Now we are done modifying the data.  */
> +      ++table->head->gc_cycle;
> +      assert ((table->head->gc_cycle & 1) == 0);
> +
>        /* It's all done.  */
>        pthread_rwlock_unlock (&table->lock);

I think this needs barriers after and before the increments.

Thanks,
Florian