From mboxrd@z Thu Jan  1 00:00:00 1970
From: Alex Chiang <achiang@hp.com>
Date: Thu, 21 Jan 2010 23:05:51 +0000
Subject: Re: SLUB ia64 linux-next crash bisected to 756dee75
Message-Id: <20100121230551.GO17684@ldl.fc.hp.com>
List-Id: <linux-ia64.vger.kernel.org>
References: <alpine.DEB.2.00.1001151358110.6590@router.home> <1263587721.20615.255.camel@useless.americas.hpqcorp.net> <alpine.DEB.2.00.1001151730350.10558@router.home> <alpine.DEB.2.00.1001191252370.25101@router.home> <20100119200228.GE11010@ldl.fc.hp.com> <alpine.DEB.2.00.1001191427370.26683@router.home> <20100119212935.GG11010@ldl.fc.hp.com> <alpine.DEB.2.00.1001191545170.26683@router.home> <20100121214749.GJ17684@ldl.fc.hp.com> <alpine.DEB.2.00.1001211643020.20071@router.home>
In-Reply-To: <alpine.DEB.2.00.1001211643020.20071@router.home>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit
To: Christoph Lameter <cl@linux-foundation.org>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>, penberg@cs.helsinki.fi, linux-ia64@vger.kernel.org, linux-mm@kvack.org

* Christoph Lameter <cl@linux-foundation.org>:
> On Thu, 21 Jan 2010, Alex Chiang wrote:
> 
> > Here is another dump of dmesg. I tried to trim it a little bit
> > where it made sense.
> 
> Looks like percpu data is corrupted. One of my earlier fixes dimensioned
> the kmem_cache_cpu array correctly. That is missing here.

Ah, that was pilot error on my part. I didn't realize that the
second patch you sent was to be in combination with the first.
Sorry about that.

> Combined fix:
> 
> ---
>  mm/slub.c |    5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
> 
> Index: linux-2.6/mm/slub.c
> =================================> --- linux-2.6.orig/mm/slub.c	2010-01-21 16:39:26.000000000 -0600
> +++ linux-2.6/mm/slub.c	2010-01-21 16:40:35.000000000 -0600
> @@ -2086,7 +2086,7 @@ init_kmem_cache_node(struct kmem_cache_n
>  #endif
>  }
> 
> -static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[SLUB_PAGE_SHIFT]);
> +static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[KMALLOC_CACHES]);
> 
>  static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
>  {
> @@ -2176,7 +2176,8 @@ static int init_kmem_cache_nodes(struct
>  	int node;
>  	int local_node;
> 
> -	if (slab_state >= UP)
> +	if (slab_state >= UP && (s < kmalloc_caches ||
> +			s > kmalloc_caches + KMALLOC_CACHES))
>  		local_node = page_to_nid(virt_to_page(s));
>  	else
>  		local_node = 0;
> 

Yup, the two together finally got it.

Reported-and-tested-by: Alex Chiang <achiang@hp.com>

Thanks!
/ac