Re: [PATCH v3 nf-next 2/3] netfilter: nf_flow_table_core: teardown direct xmit when destination changed

Eric Woudstra [off-list ref] wrote:

+static int nf_flow_table_switchdev_event(struct notifier_block *unused,
+					 unsigned long event, void *ptr)
+{
+	struct switchdev_notifier_fdb_info *fdb_info;
+	struct nf_flowtable *flowtable;
+	struct flow_cleanup_data cud;
+
+	if (event != SWITCHDEV_FDB_DEL_TO_DEVICE)
+		return NOTIFY_DONE;
+
+	fdb_info = ptr;
+	cud.addr = fdb_info->addr;
+	cud.vid = fdb_info->vid;
+	cud.ifindex = fdb_info->info.dev->ifindex;
+
+	mutex_lock(&flowtable_lock);

Please always test your patches with lockdep enabled,
this doesn't work.  Switchdev notifiers are atomic.

BUG: sleeping function called from invalid context at kernel/locking/mutex.c:575
in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 1570, name: kworker/u32:8
preempt_count: 201, expected: 0
RCU nest depth: 1, expected: 0
6 locks held by kworker/u32:8/1570:
 #0: ffff88810032b948 ((wq_completion)netns){+.+.}-{0:0}, at: process_scheduled_works+0x8d7/0x1460
 #1: ffffc900021e7ba0 (net_cleanup_work){+.+.}-{0:0}, at: process_scheduled_works+0x912/0x1460
 #2: ffffffff848d1eb0 (pernet_ops_rwsem){++++}-{4:4}, at: cleanup_net+0xf8/0x7b0
 #3: ffffffff848de6e8 (rtnl_mutex){+.+.}-{4:4}, at: ops_undo_list+0x270/0x860
 #4: ffff88811f1f8bd8 (&br->hash_lock){+...}-{3:3}, at: br_fdb_delete_by_port+0x3b/0x290
 #5: ffffffff8457a800 (rcu_read_lock){....}-{1:3}, at: atomic_notifier_call_chain+0x27/0x150
Workqueue: netns cleanup_net
Call Trace:
 __mutex_lock+0xf5/0x14f0
  nf_flow_table_switchdev_event+0x104/0x270
 atomic_notifier_call_chain+0xc5/0x150
 br_switchdev_fdb_notify+0x2b2/0x330
...