hv_netvsc: Fix a deadlock by getting rtnl lock earlier in netvsc_probe()

This patch fixes the race between netvsc_probe() and
rndis_set_subchannel(), which can cause a deadlock.

These are the related 3 paths which show the deadlock:

path #1:
    Workqueue: hv_vmbus_con vmbus_onmessage_work [hv_vmbus]
    Call Trace:
     schedule
     schedule_preempt_disabled
     __mutex_lock
     __device_attach
     bus_probe_device
     device_add
     vmbus_device_register
     vmbus_onoffer
     vmbus_onmessage_work
     process_one_work
     worker_thread
     kthread
     ret_from_fork

path #2:
    schedule
     schedule_preempt_disabled
     __mutex_lock
     netvsc_probe
     vmbus_probe
     really_probe
     __driver_attach
     bus_for_each_dev
     driver_attach_async
     async_run_entry_fn
     process_one_work
     worker_thread
     kthread
     ret_from_fork

path #3:
    Workqueue: events netvsc_subchan_work [hv_netvsc]
    Call Trace:
     schedule
     rndis_set_subchannel
     netvsc_subchan_work
     process_one_work
     worker_thread
     kthread
     ret_from_fork

Before path #1 finishes, path #2 can start to run, because just before
the "bus_probe_device(dev);" in device_add() in path #1, there is a line
"object_uevent(&dev->kobj, KOBJ_ADD);", so systemd-udevd can
immediately try to load hv_netvsc and hence path #2 can start to run.

Next, path #2 offloads the subchannal's initialization to a workqueue,
i.e. path #3, so we can end up in a deadlock situation like this:

Path #2 gets the device lock, and is trying to get the rtnl lock;
Path #3 gets the rtnl lock and is waiting for all the subchannel messages
to be processed;
Path #1 is trying to get the device lock, but since #2 is not releasing
the device lock, path #1 has to sleep; since the VMBus messages are
processed one by one, this means the sub-channel messages can't be
procedded, so #3 has to sleep with the rtnl lock held, and finally #2
has to sleep... Now all the 3 paths are sleeping and we hit the deadlock.

With the patch, we can make sure #2 gets both the device lock and the
rtnl lock together, gets its job done, and releases the locks, so #1
and #3 will not be blocked for ever.

Fixes: 8195b1396ec8 ("hv_netvsc: fix deadlock on hotplug")
Signed-off-by: Dexuan Cui <decui@microsoft.com>
Cc: Stephen Hemminger <sthemmin@microsoft.com>
Cc: K. Y. Srinivasan <kys@microsoft.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 10 insertions, 1 deletions

diff --git a/drivers/net/hyperv/netvsc_drv.c b/drivers/net/hyperv/netvsc_drv.c
index 1121a1ec407c..70921bbe0e28 100644
--- a/drivers/net/hyperv/netvsc_drv.c
+++ b/drivers/net/hyperv/netvsc_drv.c
@@ -2206,6 +2206,16 @@ static int netvsc_probe(struct hv_device *dev,
 
 	memcpy(net->dev_addr, device_info.mac_adr, ETH_ALEN);
 
+	/* We must get rtnl lock before scheduling nvdev->subchan_work,
+	 * otherwise netvsc_subchan_work() can get rtnl lock first and wait
+	 * all subchannels to show up, but that may not happen because
+	 * netvsc_probe() can't get rtnl lock and as a result vmbus_onoffer()
+	 * -> ... -> device_add() -> ... -> __device_attach() can't get
+	 * the device lock, so all the subchannels can't be processed --
+	 * finally netvsc_subchan_work() hangs for ever.
+	 */
+	rtnl_lock();
+
 	if (nvdev->num_chn > 1)
 		schedule_work(&nvdev->subchan_work);
 
@@ -2224,7 +2234,6 @@ static int netvsc_probe(struct hv_device *dev,
 	else
 		net->max_mtu = ETH_DATA_LEN;
 
-	rtnl_lock();
 	ret = register_netdevice(net);
 	if (ret != 0) {
 		pr_err("Unable to register netdev.\n");