nf-hook is not unregistered on STT device delete, But when
second time it was created it nf-hook is again registered.
which causes following softlockup.
Following patch fixes it by registering nf-hook only on very
first stt device.
---8<---
BUG: soft lockup - CPU#1 stuck for 22s! [ovs-vswitchd:11293]
RIP: 0010:[<ffffffffa0e48308>] [<ffffffffa0e48308>] nf_ip_hook+0xf8/0x180
[openvswitch]
Stack:
<IRQ>
[<ffffffff8163bf60>] ? ip_rcv_finish+0x350/0x350
[<ffffffff8163572a>] nf_iterate+0x9a/0xb0
[<ffffffff8163bf60>] ? ip_rcv_finish+0x350/0x350
[<ffffffff816357bc>] nf_hook_slow+0x7c/0x120
[<ffffffff8163bf60>] ? ip_rcv_finish+0x350/0x350
[<ffffffff8163c343>] ip_local_deliver+0x73/0x80
[<ffffffff8163bc8d>] ip_rcv_finish+0x7d/0x350
[<ffffffff8163c5e8>] ip_rcv+0x298/0x3d0
[<ffffffff81605f26>] __netif_receive_skb_core+0x696/0x880
[<ffffffff81606128>] __netif_receive_skb+0x18/0x60
[<ffffffff81606cce>] process_backlog+0xae/0x180
[<ffffffff81606512>] net_rx_action+0x152/0x270
[<ffffffff8106accc>] __do_softirq+0xec/0x300
[<ffffffff81710a1c>] do_softirq_own_stack+0x1c/0x30
Fixes: fee43fa2 ("datapath: Fix deadlock on STT device destroy.")
Signed-off-by: Pravin B Shelar <pshe...@nicira.com>
---
datapath/linux/compat/stt.c | 19 ++++++++++++++++++-
1 file changed, 18 insertions(+), 1 deletion(-)
diff --git a/datapath/linux/compat/stt.c b/datapath/linux/compat/stt.c
index 98d6d5b..7f736f0 100644
--- a/datapath/linux/compat/stt.c
+++ b/datapath/linux/compat/stt.c
@@ -153,6 +153,9 @@ struct stt_net {
struct list_head stt_list;
struct list_head stt_up_list; /* Devices which are in IFF_UP state. */
int n_tunnels;
+#ifdef HAVE_NF_REGISTER_NET_HOOK
+ bool nf_hook_reg_done;
+#endif
};
static int stt_net_id;
@@ -1553,12 +1556,23 @@ static int stt_start(struct net *net)
* rtnl-lock, which results in dead lock in stt-dev-create. Therefore
* use this new API.
*/
+
+ if (sn->nf_hook_reg_done)
+ goto out;
+
err = nf_register_net_hook(net, &nf_hook_ops);
+ if (!err)
+ sn->nf_hook_reg_done = true;
#else
+ /* Register STT only on very first STT device addition. */
+ if (!list_empty(&nf_hook_ops.list))
+ goto out;
+
err = nf_register_hook(&nf_hook_ops);
#endif
if (err)
goto dec_n_tunnel;
+out:
sn->n_tunnels++;
return 0;
@@ -1854,6 +1868,9 @@ static int stt_init_net(struct net *net)
INIT_LIST_HEAD(&sn->stt_list);
INIT_LIST_HEAD(&sn->stt_up_list);
+#ifdef HAVE_NF_REGISTER_NET_HOOK
+ sn->nf_hook_reg_done = false;
+#endif
return 0;
}
@@ -1868,7 +1885,7 @@ static void stt_exit_net(struct net *net)
/* Ideally this should be done from stt_stop(), But on some kernels
* nf-unreg operation needs RTNL-lock, which can cause deallock.
* So it is done from here. */
- if (!list_empty(&nf_hook_ops.list))
+ if (sn->nf_hook_reg_done)
nf_unregister_net_hook(net, &nf_hook_ops);
#endif
--
1.8.3.1
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
