Re: [PATCH][SMACK] add a socket_post_accept hook to fix netlabel issues with labeled TCP servers V1

From: Paul Moore
Date: Wed Feb 25 2009 - 12:21:53 EST


On Tuesday 24 February 2009 06:36:59 pm Paul Moore wrote:
> On Tuesday 24 February 2009 05:59:59 pm etienne wrote:
> > Paul Moore wrote:
> > > On Tuesday 24 February 2009 05:20:42 pm etienne wrote:
> > >> Paul Moore wrote:
> > >>> On Tuesday 24 February 2009 04:28:24 pm etienne wrote:
> > >>>> /**
> > >>>> + * smack_socket_post_access - post access check
> > >>>> + * @sock: the socket
> > >>>> + * @newsock : the grafted sock
> > >>>> + *
> > >>>> + * we have to match client IP against smack_host_label()
> > >>>> + */
> > >>>> +static void smack_socket_post_accept(struct socket *sock, struct
> > >>>> socket *newsock) +{
> > >>>> + char *hostsp;
> > >>>> + struct sockaddr_storage address;
> > >>>> + struct sockaddr_in *sin;
> > >>>> + struct sockaddr_in6 *sin6;
> > >>>> + struct in6_addr *addr6;
> > >>>> + struct socket_smack *ssp = newsock->sk->sk_security;
> > >>>> + int len;
> > >>>> +
> > >>>> + if (sock->sk == NULL)
> > >>>> + return;
> > >>>> +
> > >>>> + /* sockets can listen on both IPv4 & IPv6,
> > >>>> + and fallback to V4 if client is V4 */
> > >>>> + if (newsock->sk->sk_family != AF_INET && newsock->sk->sk_family
> > >>>> != AF_INET6) + return;
> > >>>> +
> > >>>> + /* get the client IP address **/
> > >>>> + newsock->ops->getname(newsock, (struct sockaddr *)&address, &len,
> > >>>> 2); +
> > >>>> + switch (newsock->sk->sk_family) {
> > >>>> + case AF_INET:
> > >>>> + sin = (struct sockaddr_in *)&address;
> > >>>> + break;
> > >>>> + case AF_INET6:
> > >>>> + sin6 = (struct sockaddr_in6 *)&address;
> > >>>> + addr6 = &sin6->sin6_addr;
> > >>>> + /* if a V4 client connects to a V6 listening server,
> > >>>> + * we will get a IPV6_ADDR_MAPPED mapped address here
> > >>>> + * we have to handle this case too
> > >>>> + * the test below is ipv6_addr_type()== IPV6_ADDR_MAPPED
> > >>>> + * without the requirement to have IPv6 compiled in
> > >>>> + */
> > >>>> + if ((addr6->s6_addr32[0] | addr6->s6_addr32[1]) == 0 &&
> > >>>> + addr6->s6_addr32[2] == htonl(0x0000ffff)) {
> > >>>> + __be32 addr = sin6->sin6_addr.s6_addr32[3];
> > >>>> + __be16 port = sin6->sin6_port;
> > >>>> + sin = (struct sockaddr_in *)&address;
> > >>>> + sin->sin_family = AF_INET;
> > >>>> + sin->sin_port = port;
> > >>>> + sin->sin_addr.s_addr = addr;
> > >>>> + } else {
> > >>>> + /* standard IPv6, we'll send unlabeled */
> > >>>> + smack_netlabel(newsock->sk, SMACK_UNLABELED_SOCKET);
> > >>>> + return;
> > >>>> + }
> > >>>> + break;
> > >>>> + default:
> > >>>> + /** not possible to be there **/
> > >>>> + return;
> > >>>> + }
> > >>>> + /* so, is there a label for the source IP **/
> > >>>> + hostsp = smack_host_label(sin);
> > >>>> +
> > >>>> + if (hostsp == NULL) {
> > >>>> + if (ssp->smk_labeled != SMACK_CIPSO_SOCKET)
> > >>>> + smack_netlabel(newsock->sk, SMACK_CIPSO_SOCKET);
> > >>>> + return;
> > >>>> + }
> > >>>> + if (ssp->smk_labeled != SMACK_UNLABELED_SOCKET)
> > >>>> + smack_netlabel(newsock->sk, SMACK_UNLABELED_SOCKET);
> > >>>> + return;
> > >>>> +}
> > >>>
> > >>> NAK, you can't ignore return values like that.
> > >>>
> > >>> I'm sorry I didn't get a chance to respond to your email from this
> > >>> morning, but the problem with the post_accept() hook is that you
> > >>> can't fail in this hook. There has been a _lot_ of discussion about
> > >>> this over the past couple of years on the LSM list. You should check
> > >>> the archives for all the details but the main problem is that the
> > >>> post_accept() hook is too late to deny the incoming connection so you
> > >>> can't reject the connection at that point in any sane manner.
> > >>
> > >> well, i don't want to reject the connection here :)
> > >>
> > >>> I think I'm going to draft a patch to remove the post_accept()
> > >>> hook since no one in-tree is using it and it's existence seems to
> > >>> cause more problems than it solves.
> > >>>
> > >>> Now, I understand that your patch doesn't actually enforce any access
> > >>> controls but it does call smack_netlabel() in several places and that
> > >>> function can fail
> > >>
> > >> The smack_netlabel(newsock->sk, SMACK_CIPSO_SOCKET) can failed, but
> > >> has no interest in this function (because the socket has already be
> > >> SMACK_CIPSO_SOCKET labeled by the policy) I can remove it.
> > >>
> > >> but smack_netlabel(SMACK_UNLABELED_SOCKET) cannot fail, and that's
> > >> what i'm interested in could this make the patch acceptable?
> > >
> > > Please elaborate a bit more on how you would intend a user to configure
> > > and make use of this. Also, in what cases would you remove the
> > > NetLabel from a socket? What cases would you keep it?
> >
> > well, i think it is simple : let's say i want to run a "smack-labelled
> > server" (apache, vsftpd, ...) clients connect from internet, so the
> > server admin/user will want to add a "0.0.0.0/0 @" entry in netlabel
> > that will _fail_ because the server will send back "labeled" packets.
>
> I had to go back and look at the address based labeling patches, I had
> somehow forgotten that the single label support in Smack can only be used
> for removing labels, not adding them. With that in mind your approach
> should work although you will still get really bizarre behavior in the
> following case:
>
> * Service not running at the ambient label
> * Only address based label loaded into Smack is "0.0.0.0/0 @" (everything
> unlabeled)
> * Client connects to service using labeled networking
>
> If you and Casey can live with labeled connection suddenly becoming
> unlabeled (I doubt the remote host will deal with it very gracefully) then
> go for it.

The more I thought about this last night the more it bothered me so I decided
to take a quick look to see if I could come up with something that would let
me sleep easier. The patch below is likely whitespace mangled and probably
won't apply cleanly but since I haven't done any testing I consider that a
good thing.

Take a look at the patch below and see if it accomplishes what you want/need;
I think this is a much better approach than the socket_post_accept() method.

diff --git a/security/smack/smack_lsm.c b/security/smack/smack_lsm.c
index 0278bc0..6419e83 100644
--- a/security/smack/smack_lsm.c
+++ b/security/smack/smack_lsm.c
@@ -20,6 +20,7 @@
#include <linux/ext2_fs.h>
#include <linux/kd.h>
#include <asm/ioctls.h>
+#include <linux/ip.h>
#include <linux/tcp.h>
#include <linux/udp.h>
#include <linux/mutex.h>
@@ -2559,21 +2560,40 @@ static void smack_sock_graft(struct sock *sk, struct
socket *parent)
static int smack_inet_conn_request(struct sock *sk, struct sk_buff *skb,
struct request_sock *req)
{
+ u16 family = sk->sk_family;
struct netlbl_lsm_secattr skb_secattr;
struct socket_smack *ssp = sk->sk_security;
char smack[SMK_LABELLEN];
int rc;

- if (skb == NULL)
- return -EACCES;
+ /* handle mapped IPv4 packets arriving via IPv6 sockets */
+ if (family == PF_INET6 && skb->protocol == htons(ETH_P_IP))
+ family = PF_INET;

netlbl_secattr_init(&skb_secattr);
+
rc = netlbl_skbuff_getattr(skb, sk->sk_family, &skb_secattr);
- if (rc == 0)
+ if (rc == 0) {
+ if (family == PF_INET &&
+ skb_secattr.type != NETLBL_NLTYPE_UNLABELED) {
+ struct iphdr *hdr = ip_hdr(skb);
+ struct sockaddr_in addr;
+
+ /* if we are going to treat the other side of this
+ * connection as a single label, unlabeled host we
+ * shouldn't allow it to initiate a labeled
+ * connection because we will end up confusing
+ * everyone when we suddenly drop the labeling later */
+ addr.sin_addr.s_addr = hdr->saddr;
+ if (smack_host_label(&addr) != NULL) {
+ rc = -EACCES;
+ goto inet_conn_request_return;
+ }
+ }
smack_from_secattr(&skb_secattr, smack);
- else
+ } else
strncpy(smack, smack_known_huh.smk_known, SMK_MAXLEN);
- netlbl_secattr_destroy(&skb_secattr);
+
/*
* Receiving a packet requires that the other end
* be able to write here. Read access is not required.
@@ -2585,9 +2605,45 @@ static int smack_inet_conn_request(struct sock *sk,
if (rc == 0)
strncpy(ssp->smk_packet, smack, SMK_MAXLEN);

+inet_conn_request_return:
+ netlbl_secattr_destroy(&skb_secattr);
return rc;
}

+/**
+ * smack_inet_conn_established - Setup a new inbound connection
+ * @sk: the new child socket
+ * @skb: the inbound packet
+ *
+ * Perform the setup of a new inbound stream connection; this basically means
+ * check to see if the other end of the connection is configured as a single
+ * or multi-label host and enure the new connection's socket is configured
+ * correctly.
+ */
+static void smack_inet_conn_established(struct sock *sk, struct sk_buff *skb)
+{
+ struct iphdr *hdr;
+ struct sockaddr_in addr;
+
+ /* we only need to bother with IPv4 since we don't do IPv6 labeling */
+ if (skb->protocol != htons(ETH_P_IP))
+ return;
+
+ hdr = ip_hdr(skb);
+ addr.sin_addr.s_addr = hdr->saddr;
+ if (smack_host_label(&addr) == NULL)
+ return;
+
+ /* the other end of this connection is configured as a single label,
+ * unlabeled host so we need to make sure we aren't going to label
+ * the socket */
+ /* NOTE: this is _very_ important - we can only _remove_ the label at
+ * this point, trying to add a label to the socket here could result
+ * in a failure which we can't safely catch here due to the inability
+ * to signal an error */
+ smack_netlabel(sk, SMACK_UNLABELED_SOCKET);
+}
+
/*
* Key management security hooks
*
@@ -2940,6 +2996,7 @@ struct security_operations smack_ops = {
.sk_free_security = smack_sk_free_security,
.sock_graft = smack_sock_graft,
.inet_conn_request = smack_inet_conn_request,
+ .inet_conn_established = smack_inet_conn_established,

/* key management security hooks */
#ifdef CONFIG_KEYS

--
paul moore
linux @ hp

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/