Re: [PATCH rdma-rc] IB/core: fix a UAF for netdev in netdevice_event process

On Tue, Oct 26, 2021 at 11:14:01AM +0800, Ziyang Xuan (William) wrote:

quoted

quoted

quoted

quoted

diff --git a/drivers/infiniband/core/roce_gid_mgmt.c b/drivers/infiniband/core/roce_gid_mgmt.c
index 68197e576433..063dbe72b7c2 100644
--- a/drivers/infiniband/core/roce_gid_mgmt.c
+++ b/drivers/infiniband/core/roce_gid_mgmt.c

@@ -621,6 +621,7 @@ static void netdevice_event_work_handler(struct work_struct *_work)
 {
 	struct netdev_event_work *work =
 		container_of(_work, struct netdev_event_work, work);
+	struct net_device *real_dev;
 	unsigned int i;
 
 	for (i = 0; i < ARRAY_SIZE(work->cmds) && work->cmds[i].cb; i++) {

@@ -628,6 +629,12 @@ static void netdevice_event_work_handler(struct work_struct *_work)
 					 work->cmds[i].filter_ndev,
 					 work->cmds[i].cb,
 					 work->cmds[i].ndev);
+		real_dev = rdma_vlan_dev_real_dev(work->cmds[i].ndev);
+		if (real_dev)
+			dev_put(real_dev);
+		real_dev = rdma_vlan_dev_real_dev(work->cmds[i].filter_ndev);
+		if (real_dev)
+			dev_put(real_dev);
 		dev_put(work->cmds[i].ndev);
 		dev_put(work->cmds[i].filter_ndev);
 	}

@@ -638,9 +645,10 @@ static void netdevice_event_work_handler(struct work_struct *_work)
 static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
 				struct net_device *ndev)
 {
-	unsigned int i;
 	struct netdev_event_work *ndev_work =
 		kmalloc(sizeof(*ndev_work), GFP_KERNEL);
+	struct net_device *real_dev;
+	unsigned int i;
 
 	if (!ndev_work)
 		return NOTIFY_DONE;

@@ -653,6 +661,12 @@ static int netdevice_queue_work(struct netdev_event_work_cmd *cmds,
 			ndev_work->cmds[i].filter_ndev = ndev;
 		dev_hold(ndev_work->cmds[i].ndev);
 		dev_hold(ndev_work->cmds[i].filter_ndev);
+		real_dev = rdma_vlan_dev_real_dev(ndev_work->cmds[i].ndev);
+		if (real_dev)
+			dev_hold(real_dev);
+		real_dev = rdma_vlan_dev_real_dev(ndev_work->cmds[i].filter_ndev);
+		if (real_dev)
+			dev_hold(real_dev);
 	}
 	INIT_WORK(&ndev_work->work, netdevice_event_work_handler);

Probably, this is the right change, but I don't know well enough that
part of code. What prevents from "real_dev" to disappear right after
your call to rdma_vlan_dev_real_dev()?

It is known that free the net_device until its dev_refcnt is one. The
detail realization see netdev_run_todo().The real_dev's dev_refcnt of
a vlan net_device will reach one after unregister_netdevice(&real_dev)
and unregister_vlan_dev(&vlan_ndev, ...) but the dev_refcnt of the vlan
net_device is bigger than one because netdevice_queue_work() will hold
the vlan net_device. So my solution is hold the real_dev too in
netdevice_queue_work().

              dev_hold(ndev_work->cmds[i].filter_ndev);
 +            real_dev = rdma_vlan_dev_real_dev(ndev_work->cmds[i].ndev);
 +            if (real_dev)
                  <------------ real_dev is released here.
 +                    dev_hold(real_dev);

At first, I thought the real_dev's dev_refcnt is bigger than one before
NETDEV_UNREGISTER notifier event of the vlan net_device because it calls
dev_put(real_dev) after calling unregister_netdevice_queue(dev, head).
I thought unregister_netdevice_queue() would issue NETDEV_UNREGISTER
notifier event of the vlan net_device, I can hold the real_dev in
NETDEV_UNREGISTER notifier event handler netdevice_queue_work().

But I read unregister_vlan_dev() again, found unregister_netdevice_queue()
in unregister_vlan_dev() just move the vlan net_device to a list to unregister
later. So it is possible the real_dev has been freed when we access in
netdevice_queue_work() although the probability is very small.

So the modification need to improve. For example set vlan->real_dev = NULL
after dev_put(real_dev) in unregister_vlan_dev() proposed by Jason Gunthorpe.

Do you have any other good ideas?

It is hard to tell, such implementation existed almost from day one.

Thanks

Thank you!