Re: slab fragmentation ?

[Badari Pulavarty <pbadari@us.ibm.com>]

On Sat, 2004-10-02 at 23:04, Manfred Spraul wrote:
> Badari Pulavarty wrote:
> 
> >Yes. But the next allocations should be satisfied by filling in the
> >partial slabs, instead of getting a new slab.
> >
> >As you can see from my tests, we are allocating and freeing few
> >thousands every second. I can imagine this happening, if we allocated
> >150K objects and then freed 140K of them randomly.
> >
> >  
> >
> Could you check what is the maximum number of objects that were 
> allocated? Enable debugging and check /proc/slabinfo, it's listed in the 
> globalstat block.

Max "allocated" or Max "inuse" ? Its hard to tell whats the maximum
objects inuse, since the test (scsi-debug) allocates and frees lots of
these. But I have never seen max "inuse" anywhere close to the
allocated.  

I will enable slab debugging. Someone told me that, by enabling slab
debug, it fill force use of different slab for each allocation - there
by bloating slab usages and mask the problem. Is it true ?

> 
> >I modified "crash" kmem command to dump all the slabs in the cache. 
> >I am attaching the output.
> >
> >I am wondering why we are not filling up partial slabs, before
> >allocating new ones ?
> >  
> >
> It should be impossible. s_show checks that the slabs are in the correct 
> list (full/partial/empty). You didn't get any errors, thus everything 
> was filed correctly. And cache_alloc_refill only calls cache_grow if the 
> partial and empty lists are empty.
> 
> Wait - do you use kmem_cache_alloc_node()? If you use this function then 
> the fragmentation you have described can easily happen.

No. I am using scsi_debug, which does kmalloc() lots of structures.

Thanks,
Badari

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>