Skip to contentSkip to conversation

slab: replace cache_from_obj() with inline checks

PATCH In review Vlastimil Babka 3 participants 3 messages Jan 21 – Jan 25

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 slab: replace cache_from_obj() with inline checks

Eric Dumazet has noticed cache_from_obj() is not inlined with clang and suggested splitting it into two functions, where the smaller inlined one assumes the fastpath is !CONFIG_SLAB_FREELIST_HARDENED. However most distros enable it these days and so this would likely add a function call to the object free fastpaths.

Instead take a step back and consider that cache_from_obj() is a relict from when memcgs created their separate kmem_cache copies, as the outdated comment in build_detached_freelist() reminds us.

Meanwhile hardening/debugging had reused cache_from_obj() to validate that the freed object really belongs to a slab from the cache we think we are freeing from.

In build_detached_freelist() simply remove this, because it did not handle the NULL result from cache_from_obj() failure properly, nor validate objects (for the NULL slab->slab_cache pointer) when called via kfree_bulk(). If anyone is motivated to implement it properly, it should be possible in a similar way to kmem_cache_free().

In kmem_cache_free(), do the hardening/debugging checks directly so they are inlined by definition and virt_to_slab(obj) is performed just once. In case they failed, call a newly introduced warn_free_bad_obj() that performs the warnings outside of the fastpath, and leak the object.

As an intentional change, leak the object when slab->slab_cache differs from the cache given to kmem_cache_free(). Previously we would only leak when the object is not in a valid slab page or the slab->slab_cache pointer is NULL, and otherwise trust the slab->slab_cache over the kmem_cache_free() argument. But if those differ, it means something went wrong enough that it's best not to continue freeing.

As a result the fastpath should be inlined in all configs and the warnings are moved away.

---
I'll put this to slab/for-next. Thanks for the reviews!

Changes in v2:
- Add a comment and mention in the commit log that we are leaking the
freed object also in cases where s != slab->slab_cache (Hao Li)
- Link to v1: https://patch.msgid.link/[email protected]

---
 mm/slub.c | 56 +++++++++++++++++++++++++++++++++-----------------------
 1 file changed, 33 insertions(+), 23 deletions(-)
mm/slub.c +33 -23
index 861592ac5425..fd915ea95121 100644
@@ -6738,30 +6738,26 @@ void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
 }
 #endif
 
-static inline struct kmem_cache *virt_to_cache(const void *obj)
+static noinline void warn_free_bad_obj(struct kmem_cache *s, void *obj)
 {
+	struct kmem_cache *cachep;
 	struct slab *slab;
 
 	slab = virt_to_slab(obj);
-	if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
-		return NULL;
-	return slab->slab_cache;
-}
-
-static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
-{
-	struct kmem_cache *cachep;
+	if (WARN_ONCE(!slab,
+			"kmem_cache_free(%s, %p): object is not in a slab page\n",
+			s->name, obj))
+		return;
 
-	if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
-	    !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
-		return s;
+	cachep = slab->slab_cache;
 
-	cachep = virt_to_cache(x);
-	if (WARN(cachep && cachep != s,
-		 "%s: Wrong slab cache. %s but object is from %s\n",
-		 __func__, s->name, cachep->name))
-		print_tracking(cachep, x);
-	return cachep;
+	if (WARN_ONCE(cachep != s,
+			"kmem_cache_free(%s, %p): object belongs to different cache %s\n",
+			s->name, obj, cachep ? cachep->name : "(NULL)")) {
+		if (cachep)
+			print_tracking(cachep, obj);
+		return;
+	}
 }
 
 /**
@@ -6774,11 +6770,25 @@ static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
  */
 void kmem_cache_free(struct kmem_cache *s, void *x)
 {
-	s = cache_from_obj(s, x);
-	if (!s)
-		return;
+	struct slab *slab;
+
+	slab = virt_to_slab(x);
+
+	if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) ||
+	    kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
+
+		/*
+		 * Intentionally leak the object in these cases, because it
+		 * would be too dangerous to continue.
+		 */
+		if (unlikely(!slab || (slab->slab_cache != s))) {
+			warn_free_bad_obj(s, x);
+			return;
+		}
+	}
+
 	trace_kmem_cache_free(_RET_IP_, x, s);
-	slab_free(s, virt_to_slab(x), x, _RET_IP_);
+	slab_free(s, slab, x, _RET_IP_);
 }
 EXPORT_SYMBOL(kmem_cache_free);
 
@@ -7305,7 +7315,7 @@ int build_detached_freelist(struct kmem_cache *s, size_t size,
 		df->s = slab->slab_cache;
 	} else {
 		df->slab = slab;
-		df->s = cache_from_obj(s, object); /* Support for memcg */
+		df->s = s;
 	}
 
 	/* Start new detached freelist */

---
base-commit: 0f61b1860cc3f52aef9036d7235ed1f017632193
change-id: 20260120-b4-remove_cache_from_obj-190fcaf16789

Best regards,

Signature · 1 line
Vlastimil Babka <[email protected]>
Acked-by on v2 1/1
#1 lore raw
Quoted 33 lines · Vlastimil Babka

Eric Dumazet has noticed cache_from_obj() is not inlined with clang and suggested splitting it into two functions, where the smaller inlined one assumes the fastpath is !CONFIG_SLAB_FREELIST_HARDENED. However most distros enable it these days and so this would likely add a function call to the object free fastpaths.

Instead take a step back and consider that cache_from_obj() is a relict from when memcgs created their separate kmem_cache copies, as the outdated comment in build_detached_freelist() reminds us.

Meanwhile hardening/debugging had reused cache_from_obj() to validate that the freed object really belongs to a slab from the cache we think we are freeing from.

In build_detached_freelist() simply remove this, because it did not handle the NULL result from cache_from_obj() failure properly, nor validate objects (for the NULL slab->slab_cache pointer) when called via kfree_bulk(). If anyone is motivated to implement it properly, it should be possible in a similar way to kmem_cache_free().

In kmem_cache_free(), do the hardening/debugging checks directly so they are inlined by definition and virt_to_slab(obj) is performed just once. In case they failed, call a newly introduced warn_free_bad_obj() that performs the warnings outside of the fastpath, and leak the object.

As an intentional change, leak the object when slab->slab_cache differs from the cache given to kmem_cache_free(). Previously we would only leak when the object is not in a valid slab page or the slab->slab_cache pointer is NULL, and otherwise trust the slab->slab_cache over the kmem_cache_free() argument. But if those differ, it means something went wrong enough that it's best not to continue freeing.

Quoted 8 lines

As a result the fastpath should be inlined in all configs and the warnings are moved away.

Thanks !

reviewer Tested-by
#2 lore raw
Quoted 35 lines · Eric Dumazet
Quoted 35 lines

Eric Dumazet has noticed cache_from_obj() is not inlined with clang and suggested splitting it into two functions, where the smaller inlined one assumes the fastpath is !CONFIG_SLAB_FREELIST_HARDENED. However most distros enable it these days and so this would likely add a function call to the object free fastpaths.

Instead take a step back and consider that cache_from_obj() is a relict from when memcgs created their separate kmem_cache copies, as the outdated comment in build_detached_freelist() reminds us.

Meanwhile hardening/debugging had reused cache_from_obj() to validate that the freed object really belongs to a slab from the cache we think we are freeing from.

In build_detached_freelist() simply remove this, because it did not handle the NULL result from cache_from_obj() failure properly, nor validate objects (for the NULL slab->slab_cache pointer) when called via kfree_bulk(). If anyone is motivated to implement it properly, it should be possible in a similar way to kmem_cache_free().

In kmem_cache_free(), do the hardening/debugging checks directly so they are inlined by definition and virt_to_slab(obj) is performed just once. In case they failed, call a newly introduced warn_free_bad_obj() that performs the warnings outside of the fastpath, and leak the object.

As an intentional change, leak the object when slab->slab_cache differs from the cache given to kmem_cache_free(). Previously we would only leak when the object is not in a valid slab page or the slab->slab_cache pointer is NULL, and otherwise trust the slab->slab_cache over the kmem_cache_free() argument. But if those differ, it means something went wrong enough that it's best not to continue freeing.

As a result the fastpath should be inlined in all configs and the warnings are moved away.

Quoted 8 lines
Quoted 5 lines