Re: [PATCH 05/15] kmemleak: Add the slub memory allocation/freeing hooks

[Pekka Enberg <penberg@cs.helsinki.fi>]

Hi Catalin,

On Fri, 2008-12-12 at 13:45 +0000, Catalin Marinas wrote:
> > Hmm, I'm not sure I understand why struct kmem_cache_cpu ->freelist is 
> > never scanned. 
> 
> Did you get any false positives? Or were you expecting false negatives
> because of freelist scanning which never occurred?

I haven't tested kmemleak so I'm just commenting on the code. I was
thinking about false negatives, not false positives.

On Fri, 2008-12-12 at 13:45 +0000, Catalin Marinas wrote:
> > For SMP, I suppose kmemleak doesn't scan the per-CPU 
> > areas?
> 
> It should scan the per-CPU areas in the memleak_scan() function:
> 
> #ifdef CONFIG_SMP
> 	/* per-cpu sections scanning */
> 	for_each_possible_cpu(i)
> 		scan_block(__per_cpu_start + per_cpu_offset(i),
> 			   __per_cpu_end + per_cpu_offset(i), NULL);
> #endif
> 
> >  But for UP, struct kmem_cache is allocated with kmalloc() and 
> > that contains struct kmem_cache_cpu as well.
> 
> They should be scanned as well.
>
> > And I suppose we never scan struct pages either. Otherwise ->freelist 
> > there would be a problem as well.
> 
> It was scanning the mem_map arrays in the past but removed this part and
> haven't seen any problems (on ARM).
> 
> Why would the ->freelist be a problem? I don't fully understand the slub
> allocator. Aren't objects added to the freelist only after they were
> freed? In __slab_alloc there seems to be a line:
> 
> c->page->freelist = NULL;
> 
> so the freelist won't count as a reference anymore. After freeing an
> object, kmemleak no longer cares about references to it.

I think we're talking about two different things here. Don't we then
have false negatives because we reach ->freelist of struct
kmem_cache_cpu which contains a pointer to an object that is free'd
(take a look at slab_free() fast-path)?

		Pekka