Fix testIPv[46]Remarking on 5.0+ kernels.
The undesired behaviour that this test expects was fixed upstream
in 5.0:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=50254256f382c56bde87d970f3d0d02fdb76ec70
Bug: 140851646
Test: ROOTFS=net_test.rootfs.amd64.buster.20200129 ../net_test/run_net_test.sh all_tests.sh
Change-Id: I7418b3262963bd2360efae060cd8cabee42f4fc0
diff --git a/net/test/multinetwork_test.py b/net/test/multinetwork_test.py
index 68d0ef4..2851e67 100755
--- a/net/test/multinetwork_test.py
+++ b/net/test/multinetwork_test.py
@@ -201,12 +201,15 @@
# check that the packets sent on that socket go out on the right network.
#
# For connected sockets, routing is cached in the socket's destination
- # cache entry. In this case, we check that just re-selecting the netid
- # (except via SO_BINDTODEVICE) does not change routing, but that
- # subsequently invalidating the destination cache entry does. Arguably
- # this is a bug in the kernel because re-selecting the netid should cause
- # routing to change. But it is a convenient way to check that
- # InvalidateDstCache actually works.
+ # cache entry. In this case, we check that selecting the network a second
+ # time on the same socket (except via SO_BINDTODEVICE, or SO_MARK on 5.0+
+ # kernels) does not change routing, but that subsequently invalidating the
+ # destination cache entry does. This is a bug in the kernel because
+ # re-selecting the netid should cause routing to change, and future
+ # kernels may fix this bug for per-UID routing and ucast_oif routing like
+ # they already have for mark-based routing. But until they do, this
+ # behaviour provides a convenient way to check that InvalidateDstCache
+ # actually works.
prevnetid = None
for netid in self.tuns:
self.SelectInterface(s, netid, mode)
@@ -223,15 +226,33 @@
s.sendto(UDP_PAYLOAD, (dstaddr, 53))
self.ExpectPacketOn(netid, msg, expected)
- if use_connect and mode in ["mark", "uid", "ucast_oif"]:
- # If we have a destination cache entry, packets are not rerouted...
- if prevnetid:
+ # Does this socket have a stale dst cache entry that we need to clear?
+ def SocketHasStaleDstCacheEntry():
+ if not prevnetid:
+ # This is the first time we're marking the socket.
+ return False
+ if not use_connect:
+ # Non-connected sockets never have dst cache entries.
+ return False
+ if mode in ["uid", "ucast_oif"]:
+ # No kernel invalidates the dst cache entry if the UID or the
+ # UCAST_OIF socket option changes.
+ return True
+ if mode == "oif":
+ # Changing SO_BINDTODEVICE always invalidates the dst cache entry.
+ return False
+ if mode == "mark":
+ # Changing the mark invalidates the dst cache entry in 5.0+.
+ return net_test.LINUX_VERSION < (5, 0, 0)
+ raise AssertionError("%s must be one of %s" % (mode, modes))
+
+ if SocketHasStaleDstCacheEntry():
ExpectSendUsesNetid(prevnetid)
# ... until we invalidate it.
self.InvalidateDstCache(version, prevnetid)
- ExpectSendUsesNetid(netid)
- else:
- ExpectSendUsesNetid(netid)
+
+ # In any case, future sends must be correct.
+ ExpectSendUsesNetid(netid)
self.SelectInterface(s, None, mode)
prevnetid = netid
