Re: BUG at mm/slab.c:662 - current 2.6.13-git (commit 87fc...) crashes on x86-64

[Petr Vandrovec <vandrove@vc.cvut.cz>]

Christoph Lameter wrote:
> On Sat, 10 Sep 2005, Petr Vandrovec wrote:
> 
> 
>>Christoph Lameter wrote:
>>
>>>Actually the kernel configuration you mentioned (SMP with K8 NUMA support
>>>booted on single processor) was the primary development platform for the
>>>patch. 
>>>CONFIG_DEBUG_SLAB=y
>>
>>Strange...  I had to apply patch below - after doing that everything seems to
>>be happy and running.  Though it is not right fix, it seems to work fine
>>here...
> 
> 
> Hmmm. That is strange indeed and would mean that one of the initial caches 
> has not correctly been initialized or we are using a smaller cache size 
> than the management caches. With your patch the system fell 
> back to a larger size slab (which seems to be present). Weird.
> 
> What is your setting for L1_CACHE_BYTES?

64.  Well, at least CONFIG_X86_L1_CACHE_BYTES has this value.  Last few lines in 
/proc/slabinfo in reverse order are:

kmem_cache
size-64
size-128
size-32
size-32(DMA)
size-64(DMA)
size-128(DMA)

so 64 (INDEX_AC) & 128 (INDEX_L3) slabs were created first.  But 
__find_general_cachep BUGs on *first* entry, which is 32 byte one on systems 
with 4KB pages.

It seems to me that problem is CONFIG_DEBUG_SLAB together with 
CONFIG_DEBUG_SPINLOCK and/or CONFIG_PREEMPT - in your configuration spinlock_t 
has 4 bytes, and whole arraycache_init 32 bytes - so it fits into size-32 slab, 
and you do not hit BUG that size-32 slab does not exist.

On SMP systems with CONFIG_DEBUG_SPINLOCK or CONFIG_PREEMPT spinlock_t is 8 
bytes, and so arraycache_init has 36 bytes, and it no longer fits into 
size-32...  And if you'll enable both SPINLOCK debugging and PREEMPT, you get 12 
bytes spinlock_t and 40 byte arraycache_init...

So I believe that if you'll rebuild your kernel with PREEMPT or DEBUG_SPINLOCK 
(or both), you'll get crash I'm seeing.  Probably size-32 slab needs to be 
special cased same way INDEX_AC and INDEX_L3 arrays are.
								Petr Vandrovec