Hi Pablo I've tested on mainline v5.0. The dangling pointer access of fake_rtable still exists. My env is like this: client0--box--client1. The box runs ubuntu with v5.0 kernel, br_netfilter and nfnetlink_queue are inserted. Reproduce steps: 1.Create a bridge on the box. 2.echo 1 >/proc/sys/net/bridge/bridge-nf-call-iptables 3.Add a netfilter hook function to queue the packets to nfqueuenum 0. The hook point must between <NF_BR_PRE_ROUTING,NF_BR_PRI_BRNF> and <NF_BR_FORWARD,NF_BR_PRI_BRNF - 1>. 4.Add a userspace process "nfqueue_rcv" to continuously read and set_verdict "NF_ACCEPT" to packets from queue 0. 5.Continuosly ping client1 from client0 6.Send "Ctrl + Z" to pause the "nfqueue_rcv" to simulate the queue congestion. 7.Using "ifconfig br0 down&&brctl delbr br0" to delete the bridge. 8.At this time the _skb_refdst of skbs in the nfqueue become dangling pointer. If we send "fg" to resume the "nfqueue_rcv", the kernel may try to access the freed memory. Debug log: Here I add debug logs in "netdev_freemem" and "dst_release" to prove the freed memory access. As the log shows, the "dst_release" accessed bridge's fake_rtable after the bridge was freed. Apr 8 22:25:14 raydon kernel: [62139.005062] netdev_freemem name:br0, fake_rtable:000000009d76cef0 Apr 8 22:25:21 raydon kernel: [62145.967133] dst_release dst:000000009d76cef0 dst->dev->name: řKU¡TH Apr 8 22:25:21 raydon kernel: [62145.967154] dst_release dst:000000009d76cef0 dst->dev->name: řKU¡TH Apr 8 22:25:21 raydon kernel: [62145.967180] dst_release dst:000000009d76cef0 dst->dev->name: řKU¡TH Apr 8 22:25:21 raydon kernel: [62145.967197] dst_release dst:000000009d76cef0 dst->dev->name: řKU¡TH The reason why the hook point should be after <NF_BR_PRE_ROUTING, NF_BR_PRI_BRNF> is skbs reference bridge's fake_rtable in "br_nf_pre_routing_finish" hooked at <NF_BR_PRE_ROUTING,NF_BR_PRI_BRNF>. And the reason why the hook point should be before <NF_BR_FORWARD, NF_BR_PRI_BRNF - 1> is "br_nf_forward_ip" will set the state.out to bridge dev. After this hook point, the "nfqnl_dev_drop" triggered by the bridge's NETDEV_DOWN event can flush the queued skbs before bridge's memory is freed, because the state.out now matches the bridge's dev. So the root cause is "nfqnl_dev_drop" didn't flush the skbs properly queued between <NF_BR_PRE_ROUTING,NF_BR_PRI_BRNF> and <NF_BR_FORWARD, NF_BR_PRI_BRNF - 1>. As you mentioned, hold the bridge dev for these skbs is not a proper solution. I will send another patch to let "nfqnl_dev_drop" can flush these skbs. Thanks Rundong Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> 于2019年4月4日周四 上午1:44写道： > > On Tue, Apr 02, 2019 at 12:56:09PM +0000, Rundong Ge wrote: > > Problem: > > When bridge-nf-call-iptables is enabled, skb_dst(skb) of packets that > > in the nfqueue may be a dangling pointer if user delete the bridge. > > Because packets go through the br_nf_pre_routing_finish will set the dst > > pointer to the br->fake_rtable. But the br struct will be freed > > without the reference check for these skbs. > > > > User impact: > > Kernel panic may happen when user delete the bridge if there are > > continuous traffics go through the nfqueue. > > Here is a panic in my device which using kernel v3.10. > > This kernel is _very old_. > > Could you provide the steps to reproduce this issue? > > Holding the device doesn't seem the way to go to me, we have a of > netdevice_notifier that is dropping packets for an interface that is > gone in nfnetlink_queue. We also drop packets whenever a hook in gone. > > So I wonder if this is still a problem in mainline kernels.
