Re: Q, slab, kmemleak_erase() and redzone?

[hooanon05@yahoo.co.jp]

Pekka Enberg:
> We are setting an element in the per CPU array to NULL so the the
> kmemleak code in ____cache_alloc() is safe. Red-zoning is done at the
> _object_ which is not touched by kmemleak. Looking at the oops, it
> does seem likely that you have a bug in your module (or in some other
> part of the kernel).

Thanks for reply.
In ____cache_alloc(), the variable 'ac' is assigned before
cache_alloc_refill() call, and it is used for the parameter of
kmemleak_erase(). The value may be changed by cache_alloc_refill(),
isn't it?
In this case, kmemleak_erase() receives the incorrect pointer and sets
NULL to somewhere else which may be redzone?

How about this fix?
If cpu_cache_get() call is heavy and we cannot ignore it when KMEMLEAK
is disabled, then a new wrapper may be necessary.

diff --git a/mm/slab.c b/mm/slab.c
index 71e0a1f..3f3e018 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -3104,6 +3104,7 @@ static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
 	} else {
 		STATS_INC_ALLOCMISS(cachep);
 		objp = cache_alloc_refill(cachep, flags);
+		ac = cpu_cache_get(cachep);
 	}
 	/*
 	 * To avoid a false negative, if an object that is in one of the



J. R. Okajima