Re: [PATCH ipsec v2] xfrm: check MAC header is shown with both skb->mac_len and skb_mac_header_was_set()

[Eric Dumazet]

On Thu, Sep 12, 2024 at 9:17 AM En-Wei Wu <en-wei.wu@xxxxxxxxxxxxx> wrote:
>
> When we use Intel WWAN with xfrm, our system always hangs after
> browsing websites for a few seconds. The error message shows that
> it is a slab-out-of-bounds error:
>
> [ 67.162014] BUG: KASAN: slab-out-of-bounds in xfrm_input+0x426e/0x6740
> [ 67.162030] Write of size 2 at addr ffff888156cb814b by task ksoftirqd/2/26
>
> The reason is that the eth_hdr(skb) inside if statement evaluated
> to an unexpected address with skb->mac_header = ~0U (indicating there
> is no MAC header). The unreliability of skb->mac_len causes the if
> statement to become true even if there is no MAC header inside the
> skb data buffer.
>
> Check both the skb->mac_len and skb_mac_header_was_set(skb) fixes this issue.
>
> Fixes: 87cdf3148b11 ("xfrm: Verify MAC header exists before overwriting eth_hdr(skb)->h_proto")
> Signed-off-by: En-Wei Wu <en-wei.wu@xxxxxxxxxxxxx>
> ---
> Changes in v2:
> * Change the title from "xfrm: avoid using skb->mac_len to decide if mac header is shown"
> * Remain skb->mac_len check
> * Apply fix on ipv6 path too
> ---
>  net/xfrm/xfrm_input.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/net/xfrm/xfrm_input.c b/net/xfrm/xfrm_input.c
> index 749e7eea99e4..eef0145c73a7 100644
> --- a/net/xfrm/xfrm_input.c
> +++ b/net/xfrm/xfrm_input.c
> @@ -251,7 +251,7 @@ static int xfrm4_remove_tunnel_encap(struct xfrm_state *x, struct sk_buff *skb)
>
>         skb_reset_network_header(skb);
>         skb_mac_header_rebuild(skb);
> -       if (skb->mac_len)
> +       if (skb->mac_len && skb_mac_header_was_set(skb))
>                 eth_hdr(skb)->h_proto = skb->protocol;

I would swap the two conditions :
We might in the future debug kernels leave mac_len uninitialized if
mac_header was never set.

It would be nice to catch the issue sooner.
Something is calling skb_reset_mac_len() while the mac_header was not set ?
Considering the stack trace, I can not see why mac_header is not set.
Could you try the following patch, and compile your test kernel with
CONFIG_DEBUG_NET=y ?

diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 39f1d16f362887821caa022464695c4045461493..fb06dc81039253bafeb49f0b7228748e898f480f
100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -2909,9 +2909,19 @@ static inline void
skb_reset_inner_headers(struct sk_buff *skb)
        skb->inner_transport_header = skb->transport_header;
 }

+static inline int skb_mac_header_was_set(const struct sk_buff *skb)
+{
+       return skb->mac_header != (typeof(skb->mac_header))~0U;
+}
+
 static inline void skb_reset_mac_len(struct sk_buff *skb)
 {
-       skb->mac_len = skb->network_header - skb->mac_header;
+       if (!skb_mac_header_was_set(skb)) {
+               DEBUG_NET_WARN_ON_ONCE(1);
+               skb->mac_len = 0;
+       } else {
+               skb->mac_len = skb->network_header - skb->mac_header;
+       }
 }

 static inline unsigned char *skb_inner_transport_header(const struct sk_buff
@@ -3014,11 +3024,6 @@ static inline void
skb_set_network_header(struct sk_buff *skb, const int offset)
        skb->network_header += offset;
 }

-static inline int skb_mac_header_was_set(const struct sk_buff *skb)
-{
-       return skb->mac_header != (typeof(skb->mac_header))~0U;
-}
-
 static inline unsigned char *skb_mac_header(const struct sk_buff *skb)
 {
        DEBUG_NET_WARN_ON_ONCE(!skb_mac_header_was_set(skb));