Linux Ethernet Bridging - Jul 2019 - [Bridge] [PATCH net] net: bridge: delete local fdbs on device init failure

On initialization failure we have to delete all local fdbs which were
inserted due to the default pvid. This problem has been present since the
inception of default_pvid. Note that currently there are 2 cases:
1) in br_dev_init() when br_multicast_init() fails
2) if register_netdevice() fails after calling ndo_init()

This patch takes care of both since br_vlan_flush() is called on both
occasions. Also the new fdb delete would be a no-op on normal bridge device
destruction since the local fdbs would've been already flushed by
br_dev_delete(). This is not an issue for ports since nbp_vlan_init() is
called last when adding a port thus nothing can fail after it.

Reported-by: syzbot+88533dc8b582309bf3ee at syzkaller.appspotmail.com
Fixes: 5be5a2df40f0 ("bridge: Add filtering support for default_pvid")
Signed-off-by: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
---
Tested with the provided reproducer and can no longer trigger the leak.
Also tested the br_multicast_init() failure manually by making it always
return an error.

 net/bridge/br_vlan.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/net/bridge/br_vlan.c b/net/bridge/br_vlan.c
index 021cc9f66804..3e6a702e4c21 100644
--- a/net/bridge/br_vlan.c
+++ b/net/bridge/br_vlan.c
@@ -715,6 +715,11 @@ void br_vlan_flush(struct net_bridge *br)
 
 	ASSERT_RTNL();
 
+	/* delete auto-added default pvid local fdbs before flushing vlans
+	 * otherwise these will be leaked on bridge device init failure
+	 */
+	br_fdb_delete_by_port(br, NULL, 0, 1);
+
 	vg = br_vlan_group(br);
 	__vlan_flush(vg);
 	RCU_INIT_POINTER(br->vlgrp, NULL);
-- 
2.21.0

On initialization failure we have to delete the local fdb which was
inserted due to the default pvid creation. This problem has been present
since the inception of default_pvid. Note that currently there are 2 cases:
1) in br_dev_init() when br_multicast_init() fails
2) if register_netdevice() fails after calling ndo_init()

This patch takes care of both since br_vlan_flush() is called on both
occasions. Also the new fdb delete would be a no-op on normal bridge
device destruction since the local fdb would've been already flushed by
br_dev_delete(). This is not an issue for ports since nbp_vlan_init() is
called last when adding a port thus nothing can fail after it.

Reported-by: syzbot+88533dc8b582309bf3ee at syzkaller.appspotmail.com
Fixes: 5be5a2df40f0 ("bridge: Add filtering support for default_pvid")
Signed-off-by: Nikolay Aleksandrov <nikolay at cumulusnetworks.com>
---
v2: reworded the commit message and comment so they're not plural, we're
    talking about a single bridge local fdb added on the init vlan creation
    of the default pvid

Tested with the provided reproducer and can no longer trigger the leak.
Also tested the br_multicast_init() failure manually by making it always
return an error.

 net/bridge/br_vlan.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/net/bridge/br_vlan.c b/net/bridge/br_vlan.c
index 021cc9f66804..3e6a702e4c21 100644
--- a/net/bridge/br_vlan.c
+++ b/net/bridge/br_vlan.c
@@ -715,6 +715,11 @@ void br_vlan_flush(struct net_bridge *br)
 
 	ASSERT_RTNL();
 
+	/* delete auto-added default pvid local fdb before flushing vlans
+	 * otherwise it will be leaked on bridge device init failure
+	 */
+	br_fdb_delete_by_port(br, NULL, 0, 1);
+
 	vg = br_vlan_group(br);
 	__vlan_flush(vg);
 	RCU_INIT_POINTER(br->vlgrp, NULL);
-- 
2.21.0