Re: [Cocci] [PATCH] net: Zeroing the structure ethtool_wolinfo in ethtool_get_wol()

On Thu, 21 Nov 2019, Michal Kubecek wrote:

On Thu, Nov 21, 2019 at 11:23:34AM +0100, Enrico Weigelt, metux IT consult wrote:

quoted

On 26.10.19 21:40, Joe Perches wrote:

quoted

On Sat, 2019-10-26 at 15:54 +0800, zhanglin wrote:

quoted

memset() the structure ethtool_wolinfo that has padded bytes
but the padded bytes have not been zeroed out.

[]

quoted

diff --git a/net/core/ethtool.c b/net/core/ethtool.c

[]

quoted

@@ -1471,11 +1471,13 @@ static int ethtool_reset(struct net_device *dev, char __user *useraddr)

 static int ethtool_get_wol(struct net_device *dev, char __user *useraddr)
 {
-	struct ethtool_wolinfo wol = { .cmd = ETHTOOL_GWOL };
+	struct ethtool_wolinfo wol;

 	if (!dev->ethtool_ops->get_wol)
 		return -EOPNOTSUPP;

+	memset(&wol, 0, sizeof(struct ethtool_wolinfo));
+	wol.cmd = ETHTOOL_GWOL;
 	dev->ethtool_ops->get_wol(dev, &wol);

 	if (copy_to_user(useraddr, &wol, sizeof(wol)))

It seems likely there are more of these.

Is there any way for coccinelle to find them?

Just curios: is static struct initialization (on stack) something that
should be avoided ? I've been under the impression that static
initialization allows thinner code and gives the compiler better chance
for optimizations.

Not in general. The (potential) problem here is that the structure has
padding and it is as a whole (i.e. including the padding) copied to
userspace. While I'm not aware of a compiler that wouldn't actually
initialize the whole data block including the padding in this case, the
C standard provides no guarantee about that so that to be sure we cannot
leak leftover kernel data to userspace, we need to explicitly initialize
the whole block.

I'm not sure that it is likely that the compiler will do anything other
than ensure that all the fields are initialized.  Among the files that I
could compile, the only case with actual padding and no memset, memcpy,
copy_from_user or structure assignment that I could find was

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c

There is the code struct drm_amdgpu_info_device dev_info = {};

but I couldn't see any thing in the assembly language that seemed to zero
the structure.  gcc probably thought its job was done because all fields
are subsequently initialized.  But the size of the structure does not seem
to be the same as the sum of the size of the fields.

The set of fields was collected with Coccinelle, which could be unreliable
for this task.

julia

If the structure is not going to be copied to userspace (or otherwise
exposed), using the initializer is fully sufficient and looks cleaner.

Michal Kubecek
_______________________________________________
Cocci mailing list
Cocci@systeme.lip6.fr
https://systeme.lip6.fr/mailman/listinfo/cocci