[PATCH v3] netlink: Bounds-check struct nlmsgerr creation

[PATCH v3] netlink: Bounds-check struct nlmsgerr creation

[Kees Cook]

For 32-bit systems, it might be possible to wrap lnmsgerr content
lengths beyond SIZE_MAX. Explicitly test for all overflows, and mark the
memcpy() as being unable to internally diagnose overflows.

This also excludes netlink from the coming runtime bounds check on
memcpy(), since it's an unusual case of open-coded sizing and
allocation. Avoid this future run-time warning:

  memcpy: detected field-spanning write (size 32) of single field "&errmsg->msg" at net/netlink/af_netlink.c:2447 (size 16)

Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: Jozsef Kadlecsik <kadlec@netfilter.org>
Cc: Florian Westphal <fw@strlen.de>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Cc: syzbot <syzkaller@googlegroups.com>
Cc: netfilter-devel@vger.kernel.org
Cc: coreteam@netfilter.org
Cc: netdev@vger.kernel.org
Signed-off-by: Kees Cook <keescook@chromium.org>
---
v3: add back ip_set_core patch
v2: https://lore.kernel.org/lkml/20220901064858.1417126-1-keescook@chromium.org
v1: https://lore.kernel.org/lkml/20220901030610.1121299-3-keescook@chromium.org
---
 net/netfilter/ipset/ip_set_core.c | 10 ++--
 net/netlink/af_netlink.c          | 81 +++++++++++++++++++------------
 2 files changed, 58 insertions(+), 33 deletions(-)

diff --git a/net/netfilter/ipset/ip_set_core.c b/net/netfilter/ipset/ip_set_core.c
index 16ae92054baa..43576f68f53d 100644
--- a/net/netfilter/ipset/ip_set_core.c
+++ b/net/netfilter/ipset/ip_set_core.c
@@ -1709,13 +1709,14 @@ call_ad(struct net *net, struct sock *ctnl, struct sk_buff *skb,
 		struct nlmsghdr *rep, *nlh = nlmsg_hdr(skb);
 		struct sk_buff *skb2;
 		struct nlmsgerr *errmsg;
-		size_t payload = min(SIZE_MAX,
-				     sizeof(*errmsg) + nlmsg_len(nlh));
+		size_t payload;
 		int min_len = nlmsg_total_size(sizeof(struct nfgenmsg));
 		struct nlattr *cda[IPSET_ATTR_CMD_MAX + 1];
 		struct nlattr *cmdattr;
 		u32 *errline;
 
+		if (check_add_overflow(sizeof(*errmsg), nlmsg_len(nlh), &payload))
+			return -ENOMEM;
 		skb2 = nlmsg_new(payload, GFP_KERNEL);
 		if (!skb2)
 			return -ENOMEM;
@@ -1723,7 +1724,10 @@ 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);
+		unsafe_memcpy(&errmsg->msg, nlh, nlh->nlmsg_len,
+			      /* "payload" was explicitly bounds-checked, based on
+			       * the size of nlh->nlmsg_len.
+			       */);
 		cmdattr = (void *)&errmsg->msg + min_len;
 
 		ret = nla_parse(cda, IPSET_ATTR_CMD_MAX, cmdattr,
diff --git a/net/netlink/af_netlink.c b/net/netlink/af_netlink.c
index f89ba302ac6e..1285779d9ab6 100644
--- a/net/netlink/af_netlink.c
+++ b/net/netlink/af_netlink.c
@@ -2400,35 +2400,44 @@ int __netlink_dump_start(struct sock *ssk, struct sk_buff *skb,
 }
 EXPORT_SYMBOL(__netlink_dump_start);
 
-static size_t
+/* Returns false on overflow */
+static bool __must_check
 netlink_ack_tlv_len(struct netlink_sock *nlk, int err,
-		    const struct netlink_ext_ack *extack)
+		    const struct netlink_ext_ack *extack,
+		    size_t *tlvlen)
 {
-	size_t tlvlen;
+	*tlvlen = 0;
 
 	if (!extack || !(nlk->flags & NETLINK_F_EXT_ACK))
-		return 0;
+		return true;
 
-	tlvlen = 0;
-	if (extack->_msg)
-		tlvlen += nla_total_size(strlen(extack->_msg) + 1);
-	if (extack->cookie_len)
-		tlvlen += nla_total_size(extack->cookie_len);
+	if (extack->_msg &&
+	    check_add_overflow(*tlvlen, nla_total_size(strlen(extack->_msg) + 1), tlvlen))
+		return false;
+
+	if (extack->cookie_len &&
+	    check_add_overflow(*tlvlen, nla_total_size(extack->cookie_len), tlvlen))
+		return false;
 
 	/* Following attributes are only reported as error (not warning) */
 	if (!err)
-		return tlvlen;
+		return true;
 
-	if (extack->bad_attr)
-		tlvlen += nla_total_size(sizeof(u32));
-	if (extack->policy)
-		tlvlen += netlink_policy_dump_attr_size_estimate(extack->policy);
-	if (extack->miss_type)
-		tlvlen += nla_total_size(sizeof(u32));
-	if (extack->miss_nest)
-		tlvlen += nla_total_size(sizeof(u32));
+	if (extack->bad_attr &&
+	    check_add_overflow(*tlvlen, nla_total_size(sizeof(u32)), tlvlen))
+		return false;
+	if (extack->policy &&
+	    check_add_overflow(*tlvlen, netlink_policy_dump_attr_size_estimate(extack->policy),
+			       tlvlen))
+		return false;
+	if (extack->miss_type &&
+	    check_add_overflow(*tlvlen, nla_total_size(sizeof(u32)), tlvlen))
+		return false;
+	if (extack->miss_nest &&
+	    check_add_overflow(*tlvlen, nla_total_size(sizeof(u32)), tlvlen))
+		return false;
 
-	return tlvlen;
+	return true;
 }
 
 static void
@@ -2472,33 +2481,39 @@ void netlink_ack(struct sk_buff *in_skb, struct nlmsghdr *nlh, int err,
 	size_t payload = sizeof(*errmsg);
 	struct netlink_sock *nlk = nlk_sk(NETLINK_CB(in_skb).sk);
 	unsigned int flags = 0;
-	size_t tlvlen;
+	size_t alloc_size, tlvlen = 0;
 
 	/* Error messages get the original request appened, unless the user
 	 * requests to cap the error message, and get extra error data if
 	 * requested.
 	 */
-	if (err && !(nlk->flags & NETLINK_F_CAP_ACK))
-		payload += nlmsg_len(nlh);
+	if (err && !(nlk->flags & NETLINK_F_CAP_ACK) &&
+	    check_add_overflow(payload, (size_t)nlmsg_len(nlh), &payload))
+		goto failure;
 	else
 		flags |= NLM_F_CAPPED;
 
-	tlvlen = netlink_ack_tlv_len(nlk, err, extack);
+	if (!netlink_ack_tlv_len(nlk, err, extack, &tlvlen))
+		goto failure;
 	if (tlvlen)
 		flags |= NLM_F_ACK_TLVS;
 
-	skb = nlmsg_new(payload + tlvlen, GFP_KERNEL);
-	if (!skb) {
-		NETLINK_CB(in_skb).sk->sk_err = ENOBUFS;
-		sk_error_report(NETLINK_CB(in_skb).sk);
-		return;
-	}
+	if (check_add_overflow(payload, tlvlen, &alloc_size))
+		goto failure;
+
+	skb = nlmsg_new(alloc_size, GFP_KERNEL);
+	if (!skb)
+		goto failure;
 
 	rep = __nlmsg_put(skb, NETLINK_CB(in_skb).portid, nlh->nlmsg_seq,
 			  NLMSG_ERROR, payload, flags);
 	errmsg = nlmsg_data(rep);
 	errmsg->error = err;
-	memcpy(&errmsg->msg, nlh, payload > sizeof(*errmsg) ? nlh->nlmsg_len : sizeof(*nlh));
+	unsafe_memcpy(&errmsg->msg, nlh, payload > sizeof(*errmsg)
+					 ?  nlh->nlmsg_len : sizeof(*nlh),
+		      /* "payload" was bounds checked against nlh->nlmsg_len,
+		       * and overflow-checked as tlvlen was constructed.
+		       */);
 
 	if (tlvlen)
 		netlink_ack_tlv_fill(in_skb, skb, nlh, err, extack);
@@ -2506,6 +2521,12 @@ void netlink_ack(struct sk_buff *in_skb, struct nlmsghdr *nlh, int err,
 	nlmsg_end(skb, rep);
 
 	nlmsg_unicast(in_skb->sk, skb, NETLINK_CB(in_skb).portid);
+	return;
+
+failure:
+	NETLINK_CB(in_skb).sk->sk_err = ENOBUFS;
+	sk_error_report(NETLINK_CB(in_skb).sk);
+	return;
 }
 EXPORT_SYMBOL(netlink_ack);
 
-- 
2.34.1

Re: [PATCH v3] netlink: Bounds-check struct nlmsgerr creation

[Jakub Kicinski]

On Thu,  1 Sep 2022 00:13:36 -0700 Kees Cook wrote:
>  	rep = __nlmsg_put(skb, NETLINK_CB(in_skb).portid, nlh->nlmsg_seq,
>  			  NLMSG_ERROR, payload, flags);

All we should need here is __nlmsg_put() -> nlmsg_put(),
that's idiomatic for netlink.

>  	errmsg = nlmsg_data(rep);
>  	errmsg->error = err;
> -	memcpy(&errmsg->msg, nlh, payload > sizeof(*errmsg) ? nlh->nlmsg_len : sizeof(*nlh));
> +	unsafe_memcpy(&errmsg->msg, nlh, payload > sizeof(*errmsg)
> +					 ?  nlh->nlmsg_len : sizeof(*nlh),
> +		      /* "payload" was bounds checked against nlh->nlmsg_len,
> +		       * and overflow-checked as tlvlen was constructed.
> +		       */);

Re: [PATCH v3] netlink: Bounds-check struct nlmsgerr creation

[Kees Cook]

On Thu, Sep 01, 2022 at 12:13:36AM -0700, Kees Cook wrote:
> For 32-bit systems, it might be possible to wrap lnmsgerr content
> lengths beyond SIZE_MAX. Explicitly test for all overflows, and mark the
> memcpy() as being unable to internally diagnose overflows.
> 
> This also excludes netlink from the coming runtime bounds check on
> memcpy(), since it's an unusual case of open-coded sizing and
> allocation. Avoid this future run-time warning:
> 
>   memcpy: detected field-spanning write (size 32) of single field "&errmsg->msg" at net/netlink/af_netlink.c:2447 (size 16)

To get rid of the above warning...

> [...]
> -		memcpy(&errmsg->msg, nlh, nlh->nlmsg_len);
> +		unsafe_memcpy(&errmsg->msg, nlh, nlh->nlmsg_len,
> +			      /* "payload" was explicitly bounds-checked, based on
> +			       * the size of nlh->nlmsg_len.
> +			       */);

above is the "fix", since the compiler has no way to know how to bounds
check the arguments. But, to write that comment, I added all these
things:

> [...]
> -	if (extack->cookie_len)
> -		tlvlen += nla_total_size(extack->cookie_len);
> +	if (extack->_msg &&
> +	    check_add_overflow(*tlvlen, nla_total_size(strlen(extack->_msg) + 1), tlvlen))
> +		return false;

If that's not desirable, then I guess the question I want to ask is
"what can I put in the unsafe_memcpy() comment above that proves these
values have been sanity checked? In other words, how do we know that
tlvlen hasn't overflowed? (I don't know what other sanity checking may
have already happened, so I'm looking directly at the size calculations
here.)

I assume this isn't more desirable:

-	if (extack->cookie_len)
-		tlvlen += nla_total_size(extack->cookie_len);
+	if (extack->cookie_len) {
+		size_t len = nla_total_size(extack->cookie_len);
+
+		if (WARN_ON_ONCE(len > SIZE_MAX - tlvlen))
+			return 0;
+		tlvlen += len;
+	}

Or maybe wrap it nicely with a local macro and return 0 instead of
trying to pass an error up a layer?

+#define TLVADD(amount)	do { \
+	if (WARN_ON_ONCE(check_add_overflow(tlvlen, amount, &tlvlen))) \
+		return 0; \
+} while (0)

...
	if (extack->cookie_len)
-		tlvlen += nla_total_size(extack->cookie_len);
+		TLVADD(nla_total_size(extack->cookie_len));

-- 
Kees Cook

Re: [PATCH v3] netlink: Bounds-check struct nlmsgerr creation

[Jakub Kicinski]

On Fri, 2 Sep 2022 16:08:01 -0700 Kees Cook wrote:
> > -	if (extack->cookie_len)
> > -		tlvlen += nla_total_size(extack->cookie_len);
> > +	if (extack->_msg &&
> > +	    check_add_overflow(*tlvlen, nla_total_size(strlen(extack->_msg) + 1), tlvlen))
> > +		return false;  
> 
> If that's not desirable, then I guess the question I want to ask is
> "what can I put in the unsafe_memcpy() comment above that proves these
> values have been sanity checked? In other words, how do we know that
> tlvlen hasn't overflowed? (I don't know what other sanity checking may
> have already happened, so I'm looking directly at the size calculations
> here.)

The netlink helpers for adding attributes check whether there is enough
space left in the skb. So if the calculation overflows, so be it. We'll
hit EMSGSIZE in the writing phase and unwind. The writing should make
no assumptions about the skb size. In fact all dumps will routinely hit
EMSGSIZE as we try to fit as many objects into a skb as possible, so we
unwind the one that would go over. Unwinding is well exercised in a lot
of netlink code (not here, MSG_DONE/MSG_ERROR ain't a dump).

The pre-calculation is just an estimate, if the message size ends up
being insane it really doesn't matter if the calculation is 0, INT_MAX
or random(). User is not gonna get a response, anyway.

... unless someone uses the unsafe helpers like __nlmsg_put() rather
than nlmsg_put(), hence my suggestion in the other email.