Description of problem:

For containers with the "bridge" driver, Docker creates various veth devices, and one of them is enslaved to an existing network bridge.

When the container is stopped, those devices are removed, but a reference to the one enslaved is leaked, causing NetworkManager VSZ and RSS to slowly but steadily increase through time.


Version-Release number of selected component (if applicable):

Tested with NetworkManager-1.4.0-17.el7_3.x86_64.
Upstream (1e4f1892e052c69983245b14e17a88dec6e5d138 2017-03-17) is also affected.


How reproducible:

Always.


Steps to Reproduce:

1. Execute a bunch of containers (i.e. "n=0; while echo $((++n)); docker run --rm busybox /bin/true; do :; done")
2. Wait for a few iterations.
3. See NetworkManager's VSZ and RSS increase over time.


Actual results:

NetworkManager keeps allocating and using more and more memory.


Expected results:

NetworkManager memory usage should be kept reasonably stable over time.


Additional info:

At nm-manager.c:2252, when a link is being removed from a software device, nm_device_unrealize is called, instead of remove_device (used for hardware devices):

As a consequence, the device fails the condition at nm-manager.c:977 (nm_device_unrealize sets NMDevicePriv->real to FALSE), and nm_device_removed is not called, which is the function that would eventually remove the slave from its master, releasing the otherwise pseudo-leaked reference.

Unconditionally calling nm_device_removed, even for real == FALSE devices, seems to fix the problem, but I'm not sure if that's the proper solution.