Re: [PATCH net-next 2/4] hv_netvsc: protect nvdev->extension with RCU

Stephen Hemminger [off-list ref] writes:

On Tue, 31 Oct 2017 14:42:02 +0100
Vitaly Kuznetsov [off-list ref] wrote:

quoted

@@ -2002,7 +2002,9 @@ static int netvsc_probe(struct hv_device *dev,
 	device_info.recv_sections = NETVSC_DEFAULT_RX;
 	device_info.recv_section_size = NETVSC_RECV_SECTION_SIZE;
 
+	rtnl_lock();
 	nvdev = rndis_filter_device_add(dev, &device_info);
+	rtnl_unlock();

rtnl is not necessary here. probe can not be bothered by other changes.

Yes, this is only to support rtnl_dereference() down the stack.

quoted

--- a/drivers/net/hyperv/rndis_filter.c
+++ b/drivers/net/hyperv/rndis_filter.c

@@ -402,20 +402,27 @@ int rndis_filter_receive(struct net_device *ndev,
 			 void *data, u32 buflen)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(ndev);
-	struct rndis_device *rndis_dev = net_dev->extension;
+	struct rndis_device *rndis_dev;
 	struct rndis_message *rndis_msg = data;
+	int ret = 0;
+
+	rcu_read_lock_bh();
+
+	rndis_dev = rcu_dereference_bh(net_dev->extension);

filter_receive is already called only from NAPI only and has RCU lock and soft
irq disabled. This is not necessary.

quoted

-	net_dev->extension = NULL;
+	rcu_assign_pointer(net_dev->extension, NULL);
+
+	synchronize_rcu();

rcu_assign_pointer with NULL is never a good idea.
And synchronize_rcu is slow. Since net_device is already protected
by RCU (for deletion) it should not be necessary.

I thought we don't care that much about the speed of this patch as
rndis_filter_device_remove() is only called on device remove/mtu
change/... and we need to interact with the host -- and this is already
slow.

Thank you for trying to address these races. But it should be
done carefully not by just slapping RCU everywhere.

Ok, I may have missed something. I'll try reproducing the crash and
finding a better fine-grained solution.

Thanks for the feedback!

-- 
  Vitaly