[PATCH] RDMA/core: Fix use-after-free when netns exit races compat dev removal

A compat device is removed from two places: disable_device() during the
ib device unregistration, and rdma_dev_exit_net() when the netns it
belongs to dies. remove_one_compat_dev() lets the xa_erase() decide
which of the two performs the removal, but drops compat_devs_mutex
before device_del(), so the caller that finds nothing to erase returns
without waiting for the removal the other one is running.

When that caller is rdma_dev_exit_net(), cleanup_net() carries on while
the compat device is still alive. Its kobject is tagged with the dying
net, so the device_del() still in progress dereferences net->uevent_sock
after uevent_net_exit() has freed it:

 CPU0 (ib-unreg-wq)                CPU1 (netns wq)
 ------------------                ---------------
 __ib_unregister_device()
   disable_device()
     remove_compat_devs()
       remove_one_compat_dev()
         xa_erase()
         device_del(cdev)
         ...
                                   cleanup_net()
                                     rdma_dev_exit_net()
                                       remove_one_compat_dev()
                                         xa_erase()
                                       // returns without waiting
                                     uevent_net_exit()
                                       kfree(net->uevent_sock)
         kobject_uevent_env()
           net->uevent_sock->sk    // UAF

  BUG: KASAN: slab-use-after-free in kobject_uevent_env+0xb6f/0xc80
  Read of size 8 at addr ffff888103806490 by task kworker/u16:1/41
  Workqueue: ib-unreg-wq ib_unregister_work
  Call Trace:
   kobject_uevent_env+0xb6f/0xc80
   device_del+0x737/0xc10
   disable_device+0x1ad/0x230
   __ib_unregister_device+0x229/0x3f0
  Freed by task 12:
   kfree+0x1b3/0x550
   ops_undo_list+0x273/0x8c0
   cleanup_net+0x3b3/0x720

Fix by holding compat_devs_mutex across the whole removal, so that once
remove_one_compat_dev() returns the compat device is gone no matter
which caller removed it.

Fixes: 4e0f7b907072 ("RDMA/core: Implement compat device/sysfs tree in net namespace")
Assisted-by: Claude:claude-opus-5
Signed-off-by: Serhat Kumral <[email protected]>
---
With this patch applied the report no longer shows up.

 drivers/infiniband/core/device.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
index d954eda63134..a80ac69ef986 100644
--- a/drivers/infiniband/core/device.c
+++ b/drivers/infiniband/core/device.c
@@ -1005,14 +1005,20 @@ static void remove_one_compat_dev(struct ib_device *device, u32 id)
 {
 	struct ib_core_device *cdev;
 
+	/*
+	 * Hold the lock across device_del(): the other remover may have won
+	 * the xa_erase() and still be inside device_del(), and the netns exit
+	 * path has to wait for it instead of letting cleanup_net() free the
+	 * netns state the compat device is still tagged with.
+	 */
 	mutex_lock(&device->compat_devs_mutex);
 	cdev = xa_erase(&device->compat_devs, id);
-	mutex_unlock(&device->compat_devs_mutex);
 	if (cdev) {
 		ib_free_port_attrs(cdev);
 		device_del(&cdev->dev);
 		put_device(&cdev->dev);
 	}
+	mutex_unlock(&device->compat_devs_mutex);
 }
 
 static void remove_compat_devs(struct ib_device *device)
-- 
2.53.0