networkdb: fix race in deleteNetwork

[LK4D4]

There are multiple places which read from that slice(i.e. bulkSync).

[aboch]

CI failure is legitimate:

# github.com/docker/libnetwork/networkdb
networkdb/networkdb.go:501: invalid operation: name == nodeName (mismatched types int and string)
networkdb/networkdb.go:504: cannot use name (type int) as type string in append

[LK4D4]

@aboch fixed! Sorry for that.

[aboch]

I see that func (nDB *NetworkDB) deleteNetworkNode() is always called under nDB lock. The same is true for the counterpart function func (nDB *NetworkDB) addNetworkNode().
Which makes me think this is why no precaution is taken inside these functions while manipulating the map content.

How would a race be possible, can you point me to the code path. Thanks.

[LK4D4]

@aboch Sure, nodes slice is not protected by anything, so it can be modified in deleteNetworkNode and read in https://github.com/docker/libnetwork/blob/master/networkdb/cluster.go#L416 for example. And it's not a single usage of that slice as I see.
I think that addNetworkNode is probably racy too, but it's not so obvious.

[aboch]

Oh I see, the code allows to get the nodes slice and manipulate it outside of these two functions.
Thanks for the pointer.

LGTM

[sanimej]

LGTM