Skip to contentSkip to conversation

kmemleak: Do not use off-slab management with SLAB_NOLEAKTRACE

PATCH In review Catalin Marinas 2 participants 4 messages Oct 14, 2009 – Oct 16, 2009 Memory management

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kmemleak: Do not use off-slab management with SLAB_NOLEAKTRACE

With the slab allocator, if off-slab management is enabled for the kmem_caches used by kmemleak, it leads to recursive calls into kmemleak_alloc(). Off-slab management can be triggered by other config options increasing the slab size, e.g. DEBUG_PAGEALLOC.

---
 mm/slab.c |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)
mm/slab.c +4 -2
index 7dfa481..646db30 100644
@@ -2261,9 +2261,11 @@ kmem_cache_create (const char *name, size_t size, size_t align,
 	/*
 	 * Determine if the slab management is 'on' or 'off' slab.
 	 * (bootstrapping cannot cope with offslab caches so don't do
-	 * it too early on.)
+	 * it too early on. Always use on-slab management when
+	 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
 	 */
-	if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
+	if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
+	    !(flags & SLAB_NOLEAKTRACE))
 		/*
 		 * Size is large, assume best to place the slab management obj
 		 * off-slab (should allow better packing of objs).
#1 lore raw

Hi Catalin,

Quoted 9 lines · Catalin Marinas

With the slab allocator, if off-slab management is enabled for the kmem_caches used by kmemleak, it leads to recursive calls into kmemleak_alloc(). Off-slab management can be triggered by other config options increasing the slab size, e.g. DEBUG_PAGEALLOC.

Forcing slabs to use on-slab management is pretty bad from memory consumption point of view. Wouldn't it be better to annotate the recursive calls somehow?

Pekka

Quoted 23 lines
---
 mm/slab.c |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)
mm/slab.c +4 -2
index 7dfa481..646db30 100644
@@ -2261,9 +2261,11 @@ kmem_cache_create (const char *name, size_t size, size_t align,
 	/*
 	 * Determine if the slab management is 'on' or 'off' slab.
 	 * (bootstrapping cannot cope with offslab caches so don't do
-	 * it too early on.)
+	 * it too early on. Always use on-slab management when
+	 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
 	 */
-	if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
+	if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
+	    !(flags & SLAB_NOLEAKTRACE))
 		/*
 		 * Size is large, assume best to place the slab management obj
 		 * off-slab (should allow better packing of objs).
#2 lore raw

Hi Pekka,

Quoted 6 lines · Pekka Enberg
Quoted 5 lines · Catalin Marinas

With the slab allocator, if off-slab management is enabled for the kmem_caches used by kmemleak, it leads to recursive calls into kmemleak_alloc(). Off-slab management can be triggered by other config options increasing the slab size, e.g. DEBUG_PAGEALLOC.

Quoted 8 lines
Quoted 4 lines

Forcing slabs to use on-slab management is pretty bad from memory consumption point of view. Wouldn't it be better to annotate the recursive calls somehow?

The are annotated using SLAB_NOLEAKTRACE but off-slab management uses a general cachep and we cannot use this flag on them as we end up ignoring kmalloc allocations. An alternative is to use GFP_ flag rather than a SLAB_ one to track the recursive calls.

I could also rework the way hooks were added to slab.c but currently they follow the same places as kmemcheck_slab_alloc and lockdep_trace_alloc.

Anyway, the kmemleak caches only need off-slab management when DEBUG_PAGEALLOC is enabled but in this case the memory consumption is high anyway.

Signature · 1 line
Catalin
Reviewed-by
#3 lore raw
Quoted 21 lines · Catalin Marinas
Quoted 6 lines · Pekka Enberg
Quoted 5 lines · Catalin Marinas

With the slab allocator, if off-slab management is enabled for the kmem_caches used by kmemleak, it leads to recursive calls into kmemleak_alloc(). Off-slab management can be triggered by other config options increasing the slab size, e.g. DEBUG_PAGEALLOC.

Quoted 8 lines
Quoted 4 lines

Forcing slabs to use on-slab management is pretty bad from memory consumption point of view. Wouldn't it be better to annotate the recursive calls somehow?

The are annotated using SLAB_NOLEAKTRACE but off-slab management uses a general cachep and we cannot use this flag on them as we end up ignoring kmalloc allocations. An alternative is to use GFP_ flag rather than a SLAB_ one to track the recursive calls.

Quoted 7 lines

I could also rework the way hooks were added to slab.c but currently they follow the same places as kmemcheck_slab_alloc and lockdep_trace_alloc.

Anyway, the kmemleak caches only need off-slab management when DEBUG_PAGEALLOC is enabled but in this case the memory consumption is high anyway.

Oh, right, I misread the patch, sorry. Please feel free to put the patch in kmemleak.git: