Re: [PATCH net 3/3] selftests: net: fib_nexthops: add test for group refcount imbalance bug

On 21/11/2021 19:53, Ido Schimmel wrote:

On Sun, Nov 21, 2021 at 05:24:53PM +0200, Nikolay Aleksandrov wrote:

quoted

From: Nikolay Aleksandrov <redacted>

The new selftest runs a sequence which causes circular refcount
dependency between deleted objects which cannot be released and results
in a netdevice refcount imbalance.

Signed-off-by: Nikolay Aleksandrov <redacted>
---
 tools/testing/selftests/net/fib_nexthops.sh | 56 +++++++++++++++++++++
 1 file changed, 56 insertions(+)

diff --git a/tools/testing/selftests/net/fib_nexthops.sh b/tools/testing/selftests/net/fib_nexthops.sh
index b5a69ad191b0..48d88a36ae27 100755
--- a/tools/testing/selftests/net/fib_nexthops.sh
+++ b/tools/testing/selftests/net/fib_nexthops.sh

@@ -629,6 +629,59 @@ ipv6_fcnal()
 	log_test $? 0 "Nexthops removed on admin down"
 }
 
+ipv6_grp_refs()
+{
+	run_cmd "$IP link set dev veth1 up"
+	run_cmd "$IP link add veth1.10 link veth1 up type vlan id 10"
+	run_cmd "$IP link add veth1.20 link veth1 up type vlan id 20"
+	run_cmd "$IP -6 addr add 2001:db8:91::1/64 dev veth1.10"
+	run_cmd "$IP -6 addr add 2001:db8:92::1/64 dev veth1.20"
+	run_cmd "$IP -6 neigh add 2001:db8:91::2 lladdr 00:11:22:33:44:55 dev veth1.10"
+	run_cmd "$IP -6 neigh add 2001:db8:92::2 lladdr 00:11:22:33:44:55 dev veth1.20"
+	run_cmd "$IP nexthop add id 100 via 2001:db8:91::2 dev veth1.10"
+	run_cmd "$IP nexthop add id 101 via 2001:db8:92::2 dev veth1.20"
+	run_cmd "$IP nexthop add id 102 group 100"
+	run_cmd "$IP route add 2001:db8:101::1/128 nhid 102"
+
+	# create per-cpu dsts through nh 100
+	run_cmd "ip netns exec me mausezahn -6 veth1.10 -B 2001:db8:101::1 -A 2001:db8:91::1 -c 5 -t tcp "dp=1-1023, flags=syn" >/dev/null 2>&1"

I see that other test cases in this file that are using mausezahn check
that it exists. See ipv4_torture() for example

Indeed, I'll adjust the test

quoted

+
+	# remove nh 100 from the group to delete the route potentially leaving
+	# a stale per-cpu dst

Not sure I understand the comment. Maybe:

"Remove nh 100 from the group. If the bug described in the previous
commit is not fixed, the nexthop continues to cache a per-CPU dst entry
that holds a reference on the IPv6 route."

?

Yes, that is the stale per-cpu dst.

quoted

+	run_cmd "$IP nexthop replace id 102 group 101"
+	run_cmd "$IP route del 2001:db8:101::1/128"
+
+	# add both nexthops to the group so a reference is taken on them
+	run_cmd "$IP nexthop replace id 102 group 100/101"
+
+	# if the bug exists at this point we have an unlinked IPv6 route

I would mention that by "the bug" you are referring to the bug described
in previous commit

since there is no commit id yet, I can give a brief description only
I'll may refer to it by subject though

quoted

+	# (but not freed due to stale dst) with a reference over the group
+	# so we delete the group which will again only unlink it due to the
+	# route reference
+	run_cmd "$IP nexthop del id 102"
+
+	# delete the nexthop with stale dst, since we have an unlinked
+	# group with a ref to it and an unlinked IPv6 route with ref to the
+	# group, the nh will only be unlinked and not freed so the stale dst
+	# remains forever and we get a net device refcount imbalance
+	run_cmd "$IP nexthop del id 100"
+
+	# if the bug exists this command will hang because the net device
+	# cannot be removed
+	timeout -s KILL 5 ip netns exec me ip link del veth1.10 >/dev/null 2>&1
+
+	# we can't cleanup if the command is hung trying to delete the netdev
+	if [ $? -eq 137 ]; then
+		return 1
+	fi
+
+	# cleanup
+	run_cmd "$IP link del veth1.20"
+	run_cmd "$IP nexthop flush"
+
+	return 0
+}
+
 ipv6_grp_fcnal()
 {
 	local rc

@@ -734,6 +787,9 @@ ipv6_grp_fcnal()
 
 	run_cmd "$IP nexthop add id 108 group 31/24"
 	log_test $? 2 "Nexthop group can not have a blackhole and another nexthop"
+
+	ipv6_grp_refs
+	log_test $? 0 "Nexthop group replace refcounts"
 }
 
 ipv6_res_grp_fcnal()

-- 
2.31.1