vhost/vsock: fix use-after-free in network stack callers
authorStefan Hajnoczi <stefanha@redhat.com>
Mon, 5 Nov 2018 10:35:47 +0000 (10:35 +0000)
committerMichael S. Tsirkin <mst@redhat.com>
Thu, 6 Dec 2018 19:28:38 +0000 (14:28 -0500)
If the network stack calls .send_pkt()/.cancel_pkt() during .release(),
a struct vhost_vsock use-after-free is possible.  This occurs because
.release() does not wait for other CPUs to stop using struct
vhost_vsock.

Switch to an RCU-enabled hashtable (indexed by guest CID) so that
.release() can wait for other CPUs by calling synchronize_rcu().  This
also eliminates vhost_vsock_lock acquisition in the data path so it
could have a positive effect on performance.

This is CVE-2018-14625 "kernel: use-after-free Read in vhost_transport_send_pkt".

Cc: stable@vger.kernel.org
Reported-and-tested-by: syzbot+bd391451452fb0b93039@syzkaller.appspotmail.com
Reported-by: syzbot+e3e074963495f92a89ed@syzkaller.appspotmail.com
Reported-by: syzbot+d5a0a170c5069658b141@syzkaller.appspotmail.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
drivers/vhost/vsock.c

index 731e2ea2aeca6e1deba84566c15680e544809e9f..98ed5be132c6a59a798346f182259b398cdfc11c 100644 (file)
@@ -15,6 +15,7 @@
 #include <net/sock.h>
 #include <linux/virtio_vsock.h>
 #include <linux/vhost.h>
+#include <linux/hashtable.h>
 
 #include <net/af_vsock.h>
 #include "vhost.h"
@@ -27,14 +28,14 @@ enum {
 
 /* Used to track all the vhost_vsock instances on the system. */
 static DEFINE_SPINLOCK(vhost_vsock_lock);
-static LIST_HEAD(vhost_vsock_list);
+static DEFINE_READ_MOSTLY_HASHTABLE(vhost_vsock_hash, 8);
 
 struct vhost_vsock {
        struct vhost_dev dev;
        struct vhost_virtqueue vqs[2];
 
-       /* Link to global vhost_vsock_list, protected by vhost_vsock_lock */
-       struct list_head list;
+       /* Link to global vhost_vsock_hash, writes use vhost_vsock_lock */
+       struct hlist_node hash;
 
        struct vhost_work send_pkt_work;
        spinlock_t send_pkt_list_lock;
@@ -50,11 +51,14 @@ static u32 vhost_transport_get_local_cid(void)
        return VHOST_VSOCK_DEFAULT_HOST_CID;
 }
 
-static struct vhost_vsock *__vhost_vsock_get(u32 guest_cid)
+/* Callers that dereference the return value must hold vhost_vsock_lock or the
+ * RCU read lock.
+ */
+static struct vhost_vsock *vhost_vsock_get(u32 guest_cid)
 {
        struct vhost_vsock *vsock;
 
-       list_for_each_entry(vsock, &vhost_vsock_list, list) {
+       hash_for_each_possible_rcu(vhost_vsock_hash, vsock, hash, guest_cid) {
                u32 other_cid = vsock->guest_cid;
 
                /* Skip instances that have no CID yet */
@@ -69,17 +73,6 @@ static struct vhost_vsock *__vhost_vsock_get(u32 guest_cid)
        return NULL;
 }
 
-static struct vhost_vsock *vhost_vsock_get(u32 guest_cid)
-{
-       struct vhost_vsock *vsock;
-
-       spin_lock_bh(&vhost_vsock_lock);
-       vsock = __vhost_vsock_get(guest_cid);
-       spin_unlock_bh(&vhost_vsock_lock);
-
-       return vsock;
-}
-
 static void
 vhost_transport_do_send_pkt(struct vhost_vsock *vsock,
                            struct vhost_virtqueue *vq)
@@ -210,9 +203,12 @@ vhost_transport_send_pkt(struct virtio_vsock_pkt *pkt)
        struct vhost_vsock *vsock;
        int len = pkt->len;
 
+       rcu_read_lock();
+
        /* Find the vhost_vsock according to guest context id  */
        vsock = vhost_vsock_get(le64_to_cpu(pkt->hdr.dst_cid));
        if (!vsock) {
+               rcu_read_unlock();
                virtio_transport_free_pkt(pkt);
                return -ENODEV;
        }
@@ -225,6 +221,8 @@ vhost_transport_send_pkt(struct virtio_vsock_pkt *pkt)
        spin_unlock_bh(&vsock->send_pkt_list_lock);
 
        vhost_work_queue(&vsock->dev, &vsock->send_pkt_work);
+
+       rcu_read_unlock();
        return len;
 }
 
@@ -234,12 +232,15 @@ vhost_transport_cancel_pkt(struct vsock_sock *vsk)
        struct vhost_vsock *vsock;
        struct virtio_vsock_pkt *pkt, *n;
        int cnt = 0;
+       int ret = -ENODEV;
        LIST_HEAD(freeme);
 
+       rcu_read_lock();
+
        /* Find the vhost_vsock according to guest context id  */
        vsock = vhost_vsock_get(vsk->remote_addr.svm_cid);
        if (!vsock)
-               return -ENODEV;
+               goto out;
 
        spin_lock_bh(&vsock->send_pkt_list_lock);
        list_for_each_entry_safe(pkt, n, &vsock->send_pkt_list, list) {
@@ -265,7 +266,10 @@ vhost_transport_cancel_pkt(struct vsock_sock *vsk)
                        vhost_poll_queue(&tx_vq->poll);
        }
 
-       return 0;
+       ret = 0;
+out:
+       rcu_read_unlock();
+       return ret;
 }
 
 static struct virtio_vsock_pkt *
@@ -533,10 +537,6 @@ static int vhost_vsock_dev_open(struct inode *inode, struct file *file)
        spin_lock_init(&vsock->send_pkt_list_lock);
        INIT_LIST_HEAD(&vsock->send_pkt_list);
        vhost_work_init(&vsock->send_pkt_work, vhost_transport_send_pkt_work);
-
-       spin_lock_bh(&vhost_vsock_lock);
-       list_add_tail(&vsock->list, &vhost_vsock_list);
-       spin_unlock_bh(&vhost_vsock_lock);
        return 0;
 
 out:
@@ -585,9 +585,13 @@ static int vhost_vsock_dev_release(struct inode *inode, struct file *file)
        struct vhost_vsock *vsock = file->private_data;
 
        spin_lock_bh(&vhost_vsock_lock);
-       list_del(&vsock->list);
+       if (vsock->guest_cid)
+               hash_del_rcu(&vsock->hash);
        spin_unlock_bh(&vhost_vsock_lock);
 
+       /* Wait for other CPUs to finish using vsock */
+       synchronize_rcu();
+
        /* Iterating over all connections for all CIDs to find orphans is
         * inefficient.  Room for improvement here. */
        vsock_for_each_connected_socket(vhost_vsock_reset_orphans);
@@ -628,12 +632,17 @@ static int vhost_vsock_set_cid(struct vhost_vsock *vsock, u64 guest_cid)
 
        /* Refuse if CID is already in use */
        spin_lock_bh(&vhost_vsock_lock);
-       other = __vhost_vsock_get(guest_cid);
+       other = vhost_vsock_get(guest_cid);
        if (other && other != vsock) {
                spin_unlock_bh(&vhost_vsock_lock);
                return -EADDRINUSE;
        }
+
+       if (vsock->guest_cid)
+               hash_del_rcu(&vsock->hash);
+
        vsock->guest_cid = guest_cid;
+       hash_add_rcu(vhost_vsock_hash, &vsock->hash, guest_cid);
        spin_unlock_bh(&vhost_vsock_lock);
 
        return 0;