[PATCH 01/32] netlink: Avoid memcpy() across flexible array boundary

[Gustavo A. R. Silva]

On Tue, May 03, 2022 at 06:44:10PM -0700, Kees Cook wrote:
> In preparation for run-time memcpy() bounds checking, split the nlmsg
> copying for error messages (which crosses a previous unspecified flexible
> array boundary) in half. Avoids the future run-time warning:
> 
> memcpy: detected field-spanning write (size 32) of single field "&errmsg->msg" (size 16)
> 
> Creates an explicit flexible array at the end of nlmsghdr for the payload,
> named "nlmsg_payload". There is no impact on UAPI; the sizeof(struct
> nlmsghdr) does not change, but now the compiler can better reason about
> where things are being copied.
> 
> Fixed-by: Rasmus Villemoes <linux at rasmusvillemoes.dk>
> Link: https://lore.kernel.org/lkml/d7251d92-150b-5346-6237-52afc154bb00@rasmusvillemoes.dk
> Cc: "David S. Miller" <davem at davemloft.net>
> Cc: Jakub Kicinski <kuba at kernel.org>
> Cc: Rich Felker <dalias at aerifal.cx>
> Cc: Eric Dumazet <edumazet at google.com>
> Cc: netdev at vger.kernel.org
> Signed-off-by: Kees Cook <keescook at chromium.org>
> ---
>  include/uapi/linux/netlink.h | 1 +
>  net/netlink/af_netlink.c     | 5 ++++-
>  2 files changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/include/uapi/linux/netlink.h b/include/uapi/linux/netlink.h
> index 855dffb4c1c3..47f9342d51bc 100644
> --- a/include/uapi/linux/netlink.h
> +++ b/include/uapi/linux/netlink.h
> @@ -47,6 +47,7 @@ struct nlmsghdr {
>  	__u16		nlmsg_flags;	/* Additional flags */
>  	__u32		nlmsg_seq;	/* Sequence number */
>  	__u32		nlmsg_pid;	/* Sending process port ID */
> +	__u8		nlmsg_payload[];/* Contents of message */
>  };
>  
>  /* Flags values */
> diff --git a/net/netlink/af_netlink.c b/net/netlink/af_netlink.c
> index 1b5a9c2e1c29..09346aee1022 100644
> --- a/net/netlink/af_netlink.c
> +++ b/net/netlink/af_netlink.c
> @@ -2445,7 +2445,10 @@ void netlink_ack(struct sk_buff *in_skb, struct nlmsghdr *nlh, int err,
>  			  NLMSG_ERROR, payload, flags);
>  	errmsg = nlmsg_data(rep);
>  	errmsg->error = err;
> -	memcpy(&errmsg->msg, nlh, payload > sizeof(*errmsg) ? nlh->nlmsg_len : sizeof(*nlh));
> +	errmsg->msg = *nlh;
> +	if (payload > sizeof(*errmsg))
> +		memcpy(errmsg->msg.nlmsg_payload, nlh->nlmsg_payload,
> +		       nlh->nlmsg_len - sizeof(*nlh));

They have nlmsg_len()[1] for the length of the payload without the header:

/**
 * nlmsg_len - length of message payload
 * @nlh: netlink message header
 */
static inline int nlmsg_len(const struct nlmsghdr *nlh)
{
	return nlh->nlmsg_len - NLMSG_HDRLEN;
}

(would that function use some sanitization, though? what if nlmsg_len is
somehow manipulated to be less than NLMSG_HDRLEN?...)

Also, it seems there is at least one more instance of this same issue:

diff --git a/net/netfilter/ipset/ip_set_core.c b/net/netfilter/ipset/ip_set_core.c
index 16ae92054baa..d06184b94af5 100644
--- a/net/netfilter/ipset/ip_set_core.c
+++ b/net/netfilter/ipset/ip_set_core.c
@@ -1723,7 +1723,8 @@ call_ad(struct net *net, struct sock *ctnl, struct sk_buff *skb,
                                  nlh->nlmsg_seq, NLMSG_ERROR, payload, 0);
                errmsg = nlmsg_data(rep);
                errmsg->error = ret;
-               memcpy(&errmsg->msg, nlh, nlh->nlmsg_len);
+               errmsg->msg = *nlh;
+               memcpy(errmsg->msg.nlmsg_payload, nlh->nlmsg_payload, nlmsg_len(nlh));
                cmdattr = (void *)&errmsg->msg + min_len;

                ret = nla_parse(cda, IPSET_ATTR_CMD_MAX, cmdattr,

--
Gustavo

[1] https://elixir.bootlin.com/linux/v5.18-rc5/source/include/net/netlink.h#L577