]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
netfilter: conntrack: resched in nf_ct_iterate_cleanup
authorFlorian Westphal <fw@strlen.de>
Wed, 20 Jan 2016 10:16:43 +0000 (11:16 +0100)
committerPablo Neira Ayuso <pablo@netfilter.org>
Sun, 31 Jan 2016 23:15:26 +0000 (00:15 +0100)
Ulrich reports soft lockup with following (shortened) callchain:

NMI watchdog: BUG: soft lockup - CPU#1 stuck for 22s!
__netif_receive_skb_core+0x6e4/0x774
process_backlog+0x94/0x160
net_rx_action+0x88/0x178
call_do_softirq+0x24/0x3c
do_softirq+0x54/0x6c
__local_bh_enable_ip+0x7c/0xbc
nf_ct_iterate_cleanup+0x11c/0x22c [nf_conntrack]
masq_inet_event+0x20/0x30 [nf_nat_masquerade_ipv6]
atomic_notifier_call_chain+0x1c/0x2c
ipv6_del_addr+0x1bc/0x220 [ipv6]

Problem is that nf_ct_iterate_cleanup can run for a very long time
since it can be interrupted by softirq processing.
Moreover, atomic_notifier_call_chain runs with rcu readlock held.

So lets call cond_resched() in nf_ct_iterate_cleanup and defer
the call to a work queue for the atomic_notifier_call_chain case.

We also need another cond_resched in get_next_corpse, since we
have to deal with iter() always returning false, in that case
get_next_corpse will walk entire conntrack table.

Reported-by: Ulrich Weber <uw@ocedo.com>
Tested-by: Ulrich Weber <uw@ocedo.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
net/ipv6/netfilter/nf_nat_masquerade_ipv6.c
net/netfilter/nf_conntrack_core.c

index 31ba7ca19757083a1a83f8388b87be5695b1ba98..051b6a6bfff6ca7c0dd42ceb43a417260a11bd23 100644 (file)
 #include <net/ipv6.h>
 #include <net/netfilter/ipv6/nf_nat_masquerade.h>
 
+#define MAX_WORK_COUNT 16
+
+static atomic_t v6_worker_count;
+
 unsigned int
 nf_nat_masquerade_ipv6(struct sk_buff *skb, const struct nf_nat_range *range,
                       const struct net_device *out)
@@ -78,14 +82,78 @@ static struct notifier_block masq_dev_notifier = {
        .notifier_call  = masq_device_event,
 };
 
+struct masq_dev_work {
+       struct work_struct work;
+       struct net *net;
+       int ifindex;
+};
+
+static void iterate_cleanup_work(struct work_struct *work)
+{
+       struct masq_dev_work *w;
+       long index;
+
+       w = container_of(work, struct masq_dev_work, work);
+
+       index = w->ifindex;
+       nf_ct_iterate_cleanup(w->net, device_cmp, (void *)index, 0, 0);
+
+       put_net(w->net);
+       kfree(w);
+       atomic_dec(&v6_worker_count);
+       module_put(THIS_MODULE);
+}
+
+/* ipv6 inet notifier is an atomic notifier, i.e. we cannot
+ * schedule.
+ *
+ * Unfortunately, nf_ct_iterate_cleanup can run for a long
+ * time if there are lots of conntracks and the system
+ * handles high softirq load, so it frequently calls cond_resched
+ * while iterating the conntrack table.
+ *
+ * So we defer nf_ct_iterate_cleanup walk to the system workqueue.
+ *
+ * As we can have 'a lot' of inet_events (depending on amount
+ * of ipv6 addresses being deleted), we also need to add an upper
+ * limit to the number of queued work items.
+ */
 static int masq_inet_event(struct notifier_block *this,
                           unsigned long event, void *ptr)
 {
        struct inet6_ifaddr *ifa = ptr;
-       struct netdev_notifier_info info;
+       const struct net_device *dev;
+       struct masq_dev_work *w;
+       struct net *net;
+
+       if (event != NETDEV_DOWN ||
+           atomic_read(&v6_worker_count) >= MAX_WORK_COUNT)
+               return NOTIFY_DONE;
+
+       dev = ifa->idev->dev;
+       net = maybe_get_net(dev_net(dev));
+       if (!net)
+               return NOTIFY_DONE;
 
-       netdev_notifier_info_init(&info, ifa->idev->dev);
-       return masq_device_event(this, event, &info);
+       if (!try_module_get(THIS_MODULE))
+               goto err_module;
+
+       w = kmalloc(sizeof(*w), GFP_ATOMIC);
+       if (w) {
+               atomic_inc(&v6_worker_count);
+
+               INIT_WORK(&w->work, iterate_cleanup_work);
+               w->ifindex = dev->ifindex;
+               w->net = net;
+               schedule_work(&w->work);
+
+               return NOTIFY_DONE;
+       }
+
+       module_put(THIS_MODULE);
+ err_module:
+       put_net(net);
+       return NOTIFY_DONE;
 }
 
 static struct notifier_block masq_inet_notifier = {
index 58882de06bd74f254033e0f06521f264c18c8d44..f60b4fdeeb8cc4fc600506ecc58e2bdc4cad1654 100644 (file)
@@ -1412,6 +1412,7 @@ get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
                }
                spin_unlock(lockp);
                local_bh_enable();
+               cond_resched();
        }
 
        for_each_possible_cpu(cpu) {
@@ -1424,6 +1425,7 @@ get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
                                set_bit(IPS_DYING_BIT, &ct->status);
                }
                spin_unlock_bh(&pcpu->lock);
+               cond_resched();
        }
        return NULL;
 found:
@@ -1440,6 +1442,8 @@ void nf_ct_iterate_cleanup(struct net *net,
        struct nf_conn *ct;
        unsigned int bucket = 0;
 
+       might_sleep();
+
        while ((ct = get_next_corpse(net, iter, data, &bucket)) != NULL) {
                /* Time to push up daises... */
                if (del_timer(&ct->timeout))
@@ -1448,6 +1452,7 @@ void nf_ct_iterate_cleanup(struct net *net,
                /* ... else the timer will get him soon. */
 
                nf_ct_put(ct);
+               cond_resched();
        }
 }
 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup);