Re: [percpu] BUG: using __this_cpu_add() in preemptible [00000000] code: init/1

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 18 Mar 2014 09:43:27 -0500 (CDT) Christoph Lameter <cl@linux.com> wrote:

> Anyways the statistics in slub should be low impact and able to
> be placed anywhere. So I think we need this patch to avoid
> the preemption checks begin triggered.
> 
> 
> Subject: SLUB: Use raw_cpu_inc for incrementing statistics
> 
> Statistics are not critical to the operation of the allocation
> but should also not cause too much overhead.
> 
> __this_cpu_inc now checks if preemption is disabled and triggers.
> Use raw_cpu_inc to avoid the checks. Using this_cpu_ops may cause
> interrupt disable/enable sequences on various arches which may
> significantly impact allocator performance.
> 
> ...
>
> --- linux.orig/mm/slub.c	2014-03-07 14:10:30.142777022 -0600
> +++ linux/mm/slub.c	2014-03-18 09:39:15.675035324 -0500
> @@ -224,7 +224,7 @@
>  static inline void stat(const struct kmem_cache *s, enum stat_item si)
>  {
>  #ifdef CONFIG_SLUB_STATS
> -	__this_cpu_inc(s->cpu_slab->stat[si]);
> +	raw_cpu_inc(s->cpu_slab->stat[si]);
>  #endif
>  }

Please always consider how the code will appear to a naive reader. 
That reader won't know why raw_ is being used here, so we should
provide an explanation.

--- a/mm/slub.c~slub-use-raw_cpu_inc-for-incrementing-statistics-fix
+++ a/mm/slub.c
@@ -224,6 +224,10 @@ static inline void memcg_propagate_slab_
 static inline void stat(const struct kmem_cache *s, enum stat_item si)
 {
 #ifdef CONFIG_SLUB_STATS
+	/*
+	 * The rmw is racy on a preemptible kernel but this is acceptable, so
+	 * avoid this_cpu_add()'s irq-disable overhead.
+	 */
 	raw_cpu_inc(s->cpu_slab->stat[si]);
 #endif
 }
_