randstruct: fix container_of() false positives after __mptr removal

Hi,

The GCC randstruct implementation specifically keyed on the __mptr temp variable name. With that removed, we need to check via a struct walk or a void * cast (which is what Clang already does).

I'm hoping to get these into v7.3 before release, but I want to put this through -next for a bit first.

-Kees

Kees Cook (2):
randstruct: fix container_of() false positives after __mptr removal
randstruct: report bad casts as warnings rather than notes

 scripts/gcc-plugins/randomize_layout_plugin.c | 63 +++++++++++++++++--
 1 file changed, 59 insertions(+), 4 deletions(-)

2.34.1

Commit f9e7a7564834 ("container_of: remove local __mptr variable") dropped the "void *__mptr" temporary from container_of(). The randstruct GCC plugin's find_bad_casts pass recognized the casts container_of() generates by that variable's name:

	const_tree ssa_name_var = SSA_NAME_VAR(rhs1);
	/* skip bogus type casts introduced by container_of */
	if (ssa_name_var != NULL_TREE && DECL_NAME(ssa_name_var) &&
	    !strcmp(DECL_NAME_POINTER(ssa_name_var), "__mptr"))
		continue;

With the variable gone the suppression never fires, so every container_of() whose container type is randomized now emits a note:

include/linux/container_of.h:23:9: note: randstruct: casting between
randomized structure pointer types (ssa): 'struct ocfs2_triggers' and 'struct jbd2_buffer_trigger_type'
fs/ocfs2/journal.c:524:16: note: in expansion of macro 'container_of'

The pass runs on GIMPLE, after folding, and for a member at offset 0 the whole expression collapses to a bare copy that is indistinguishable from an unsafe cast:

  to_ocfs2_trigger (struct jbd2_buffer_trigger_type * triggers)
  {
    _2 = triggers_1(D);   /* void * cast and subtraction gone */
    return _2;
  }

Match the type relationship instead. A cast from A * to B * is a container_of() if B reaches a field of type A at offset 0 through a chain of by-value members. The chain matters: container_of()'s member argument may be a dotted path, as in ceph_inode(), which is container_of(inode, struct ceph_inode_info, netfs.inode) and needs two levels. The search is depth-bounded to 4 just in case, since real paths are generally one or two members deep.

Requiring the cast happens at offset 0 is done because any other member offset the subtraction survives folding and the cast's rhs is still void *, which the pass already skips a few lines above.

A cast between two randomized types with no containment relationship is still reported. Verified with:

	struct cred *f(struct file *f) { return (struct cred *)f; }

which is still flagged with the patch applied.

Clang's implementation is unaffected. It checks the cast as written, and both the old and new macros cast from void *, which is always permitted; a genuinely bad cast is rejected there as a hard error rather than a note.

Build tested ARCH=x86_64 defconfig with CONFIG_RANDSTRUCT_FULL=y and GCC 14.2.0: randstruct notes 52 before, 0 after.

---
 scripts/gcc-plugins/randomize_layout_plugin.c | 61 ++++++++++++++++++-
 1 file changed, 58 insertions(+), 3 deletions(-)

index ff65a4f87f24..e2bd9ba08089 100644
@@ -698,6 +698,63 @@ static void handle_local_var_initializers(void)
 	}
 }
 
+/*
+ * Does @container reach a field of type @member_type by a chain of
+ * by-value members? That is the relationship container_of() expresses --
+ * its @member argument may be a dotted path, e.g.
+ * container_of(inode, struct ceph_inode_info, netfs.inode) -- so a cast
+ * from @member_type * to @container * is legitimate rather than a
+ * layout-confusing one.
+ *
+ * container_of() used to leave a "void *__mptr" temporary behind, and this
+ * pass recognised such casts by that name. Commit f9e7a7564834
+ * ("container_of: remove local __mptr variable") removed it to stop nested
+ * container_of() shadowing itself, and the cast now folds to a bare SSA
+ * copy when the member sits at offset 0, leaving nothing syntactic to key
+ * on. Match the type relationship instead.
+ *
+ * The depth bound keeps this cheap; container_of() paths are one or two
+ * members deep in practice.
+ */
+#define CONTAINER_OF_MAX_DEPTH 4
+
+static bool is_container_of_cast(const_tree container, const_tree member_type,
+				 int depth)
+{
+	const_tree field;
+
+	if (container == NULL_TREE || depth > CONTAINER_OF_MAX_DEPTH)
+		return false;
+
+	if (TREE_CODE(container) != RECORD_TYPE &&
+	    TREE_CODE(container) != UNION_TYPE)
+		return false;
+
+	for (field = TYPE_FIELDS(container); field; field = DECL_CHAIN(field)) {
+		const_tree field_type;
+
+		if (TREE_CODE(field) != FIELD_DECL)
+			continue;
+
+		/*
+		 * Only a member at offset 0 can reach here: for any other
+		 * offset container_of()'s subtraction survives folding, the
+		 * cast's rhs stays void *, and the caller skipped it above.
+		 */
+		if (!integer_zerop(byte_position(field)))
+			continue;
+
+		field_type = TYPE_MAIN_VARIANT(TREE_TYPE(field));
+		if (field_type == member_type)
+			return true;
+
+		if (is_container_of_cast(field_type, member_type, depth + 1))
+			return true;
+	}
+
+	return false;
+}
+
 /*
  * iterate over all statements to find "bad" casts:
  * those where the address of the start of a structure is cast

@@ -799,10 +856,8 @@ static unsigned int find_bad_casts_execute(void)
 #endif
 				MISMATCH(gimple_location(stmt), "op0", ptr_lhs_type, op0_type);
 			} else {
-				const_tree ssa_name_var = SSA_NAME_VAR(rhs1);
 				/* skip bogus type casts introduced by container_of */
-				if (ssa_name_var != NULL_TREE && DECL_NAME(ssa_name_var) && 
-				    !strcmp((const char *)DECL_NAME_POINTER(ssa_name_var), "__mptr"))
+				if (is_container_of_cast(ptr_lhs_type, ptr_rhs_type, 0))
 					continue;
 #ifndef __DEBUG_PLUGIN
 				if (lookup_attribute("randomize_performed", TYPE_ATTRIBUTES(ptr_rhs_type)))

2.34.1

find_bad_casts() reports a cast between two randomized structure pointer types with inform(), which renders as a "note:". It has done so since the plugin was originally added, while using error() freely for attribute misuse, UAPI structs, and version mismatches.

Clang's implementation of the same check has always been stricter:
it rejects such a cast as a full error. There is no reason for the GCC side to be effectively silent about the same problem.

Build tested ARCH=x86_64 with CONFIG_RANDSTRUCT_FULL=y and GCC 14.2.0:
allmodconfig clean, and defconfig clean under three different random seeds. A deliberate bad cast is still reported, now as a warning, at the correct line and column.

---
 scripts/gcc-plugins/randomize_layout_plugin.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index e2bd9ba08089..1e66f45fe29b 100644
@@ -22,7 +22,7 @@
 #define ORIG_TYPE_NAME(node) \
 	(TYPE_NAME(TYPE_MAIN_VARIANT(node)) != NULL_TREE ? ((const unsigned char *)IDENTIFIER_POINTER(TYPE_NAME(TYPE_MAIN_VARIANT(node)))) : (const unsigned char *)"anonymous")
 
-#define INFORM(loc, msg, ...)	inform(loc, "randstruct: " msg, ##__VA_ARGS__)
+#define INFORM(loc, msg, ...)	warning_at(loc, 0, "randstruct: " msg, ##__VA_ARGS__)
 #define MISMATCH(loc, how, ...)	INFORM(loc, "casting between randomized structure pointer types (" how "): %qT and %qT\n", __VA_ARGS__)
 
 __visible int plugin_is_GPL_compatible;

2.34.1

Hi Kees,

Yeah, it seems odd to mention that Clang's implementation has always treated this as an error then not match its behavior. If this cannot be made an error, I think the commit message should call out why.

Cheers,
Nathan

I guess I was thinking of moving from "invisible note" to "visible warning", but I'd agree: better to just match Clang on this. I'll send a v2.

Kees Cook