Linux kernel RDMA NULL pointer dereference

While reading through Linux kernel’s ChangeLog I came across this one. I was unable to find any CVE ID for this vulnerability reported by Håkon Bugge on 6 December 2017. The vulnerability was discovered by Google’s syzkaller kernel fuzzer and reported in syzkaller719569. The vulnerable code starts in the setsockopt() system call implementation for RDS (Reliable Datagram Sockets). This code is located at net/rds/af_rds.c and here is the code path we are interested in.

static int rds_setsockopt(struct socket *sock, int level, int optname,
			  char __user *optval, unsigned int optlen)
{
	struct rds_sock *rs = rds_sk_to_rs(sock->sk);
	int ret;
   ...
	case RDS_GET_MR:
		ret = rds_get_mr(rs, optval, optlen);
		break;
	case RDS_GET_MR_FOR_DEST:
		ret = rds_get_mr_for_dest(rs, optval, optlen);
		break;
   ...
}

Before we check those two options, it is important to see some specific value in “rds_sock” structure. This structure is defined in net/rds/rds.h header file. The key here is that “rs_transport” pointer in “rds_sock” structure can be NULL.

/**
 * struct rds_transport -  transport specific behavioural hooks
   ...
 */
   ...
struct rds_transport {
   ...
};

struct rds_sock {
   ...
	/*
	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
	 * support.
	 */
   ...
	struct rds_transport    *rs_transport;
   ...
};

Back in rds_setsockopt(), there are two options with very similar code. The RDS_GET_MR and RDS_GET_MR_FOR_DES which will both result in a code that uses copy_from_user() to copy the “rds_get_mr_args” or “rds_get_mr_for_dest_args” structure to the kernel space, and then invoke __rds_rdma_map() routine to do the mapping.

int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
{
	struct rds_get_mr_args args;
   ...
	if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
			   sizeof(struct rds_get_mr_args)))
		return -EFAULT;

	return __rds_rdma_map(rs, &args, NULL, NULL);
}

int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
{
   ...
	if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
			   sizeof(struct rds_get_mr_for_dest_args)))
		return -EFAULT;
   ...
	return __rds_rdma_map(rs, &new_args, NULL, NULL);
}

If we now move to __rds_rdma_map() in net/rds/rdma.c we will quickly notice what’s the issue that triggers the vulnerability. The code will try to access the get_mr() callback function from the “rs_transport” pointer which as mentioned earlier can be unset (NULL), resulting in a NULL pointer dereference. This code path can be reached from the previously described setsockopt() system call options.

static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
				u64 *cookie_ret, struct rds_mr **mr_ret)
{
   ...
	if (!rs->rs_transport->get_mr) {
		ret = -EOPNOTSUPP;
		goto out;
	}
   ...
	return ret;
}

The patch was to add an additional check that ensures that “rs_transport” is not NULL. If it’s NULL the __rds_rdma_map() will exit with “ENOTCONN” (Transport endpoint is not connected) error code.

diff --git a/net/rds/rdma.c b/net/rds/rdma.c
index 8886f15abe90..bc2f1e0977d6 100644
--- a/net/rds/rdma.c
+++ b/net/rds/rdma.c
@@ -183,7 +183,7 @@  static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 	long i;
 	int ret;
 
-	if (rs->rs_bound_addr == 0) {
+	if (rs->rs_bound_addr == 0 || !rs->rs_transport) {
 		ret = -ENOTCONN; /* XXX not a great errno */
 		goto out;
 	}