On Sun, Sep 25, 2022 at 09:17:40AM +0200, Paolo Abeni wrote:

On Fri, 2022-09-23 at 13:28 -0700, Kees Cook wrote:

quoted

All callers of APIs that allowed a 0-sized frag_size appear to be
passing actual size information already

AFAICS, not yet:

drivers/net/ethernet/qlogic/qed/qed_ll2.c:
	skb = build_skb(buffer->data, 0); // -> __build_skb(..., 0) 
		// ->  __build_skb_around()

drivers/net/ethernet/broadcom/bnx2.c:
	skb = build_skb(data, 0);

I guess some more drivers have calls leading to 

	__build_skb_around(...,  0)

there are several call path to checks...

Ah-ha! Thank you. I will try to hunt these down -- I think we can't
remove the "secret resizing" effect of ksize() without fixing these.

quoted

[...]

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 0b30fbdbd0d0..84ca89c781cd 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c

@@ -195,7 +195,11 @@ static void __build_skb_around(struct sk_buff *skb, void *data,
 			       unsigned int frag_size)
 {
 	struct skb_shared_info *shinfo;
-	unsigned int size = frag_size ? : ksize(data);
+	unsigned int size = frag_size;
+
+	/* All callers should be setting frag size now? */
+	if (WARN_ON_ONCE(size == 0))
+		size = ksize(data);

At some point in the future, I guess we could even drop this check,
right?

Alternatively, we might be able to ask the slab if "data" came from
kmalloc or a kmem_cache, and if the former, do:

	data = krealloc(kmalloc_size_roundup(ksize(data), ...)

But that seems ugly...

-- 
Kees Cook