IPVS netns shutdown/startup dead-lock
authorHans Schillstrom <hans@schillstrom.com>
Tue, 11 Oct 2011 01:54:35 +0000 (10:54 +0900)
committerPablo Neira Ayuso <pablo@netfilter.org>
Wed, 12 Oct 2011 16:32:15 +0000 (18:32 +0200)
ip_vs_mutext is used by both netns shutdown code and startup
and both implicit uses sk_lock-AF_INET mutex.

cleanup CPU-1         startup CPU-2
ip_vs_dst_event()     ip_vs_genl_set_cmd()
 sk_lock-AF_INET     __ip_vs_mutex
                     sk_lock-AF_INET
__ip_vs_mutex
* DEAD LOCK *

A new mutex placed in ip_vs netns struct called sync_mutex is added.

Comments from Julian and Simon added.
This patch has been running for more than 3 month now and it seems to work.

Ver. 3
    IP_VS_SO_GET_DAEMON in do_ip_vs_get_ctl protected by sync_mutex
    instead of __ip_vs_mutex as sugested by Julian.

Signed-off-by: Hans Schillstrom <hans@schillstrom.com>
Acked-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Simon Horman <horms@verge.net.au>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
include/net/ip_vs.h
net/netfilter/ipvs/ip_vs_ctl.c
net/netfilter/ipvs/ip_vs_sync.c

index 1aaf915656f3930b0b04253cd9171b3ba49a4d9c..8fa4430f99c14e025ad6a189a1116f72ae831317 100644 (file)
@@ -900,6 +900,7 @@ struct netns_ipvs {
        volatile int            sync_state;
        volatile int            master_syncid;
        volatile int            backup_syncid;
+       struct mutex            sync_mutex;
        /* multicast interface name */
        char                    master_mcast_ifn[IP_VS_IFNAME_MAXLEN];
        char                    backup_mcast_ifn[IP_VS_IFNAME_MAXLEN];
index 2b771dc708a30d212ed37fc1a76abb759279f159..3e5e08b78bfcfd53daf7adc85606db8e7d5e03cc 100644 (file)
@@ -2283,6 +2283,7 @@ do_ip_vs_set_ctl(struct sock *sk, int cmd, void __user *user, unsigned int len)
        struct ip_vs_service *svc;
        struct ip_vs_dest_user *udest_compat;
        struct ip_vs_dest_user_kern udest;
+       struct netns_ipvs *ipvs = net_ipvs(net);
 
        if (!capable(CAP_NET_ADMIN))
                return -EPERM;
@@ -2303,6 +2304,24 @@ do_ip_vs_set_ctl(struct sock *sk, int cmd, void __user *user, unsigned int len)
        /* increase the module use count */
        ip_vs_use_count_inc();
 
+       /* Handle daemons since they have another lock */
+       if (cmd == IP_VS_SO_SET_STARTDAEMON ||
+           cmd == IP_VS_SO_SET_STOPDAEMON) {
+               struct ip_vs_daemon_user *dm = (struct ip_vs_daemon_user *)arg;
+
+               if (mutex_lock_interruptible(&ipvs->sync_mutex)) {
+                       ret = -ERESTARTSYS;
+                       goto out_dec;
+               }
+               if (cmd == IP_VS_SO_SET_STARTDAEMON)
+                       ret = start_sync_thread(net, dm->state, dm->mcast_ifn,
+                                               dm->syncid);
+               else
+                       ret = stop_sync_thread(net, dm->state);
+               mutex_unlock(&ipvs->sync_mutex);
+               goto out_dec;
+       }
+
        if (mutex_lock_interruptible(&__ip_vs_mutex)) {
                ret = -ERESTARTSYS;
                goto out_dec;
@@ -2316,15 +2335,6 @@ do_ip_vs_set_ctl(struct sock *sk, int cmd, void __user *user, unsigned int len)
                /* Set timeout values for (tcp tcpfin udp) */
                ret = ip_vs_set_timeout(net, (struct ip_vs_timeout_user *)arg);
                goto out_unlock;
-       } else if (cmd == IP_VS_SO_SET_STARTDAEMON) {
-               struct ip_vs_daemon_user *dm = (struct ip_vs_daemon_user *)arg;
-               ret = start_sync_thread(net, dm->state, dm->mcast_ifn,
-                                       dm->syncid);
-               goto out_unlock;
-       } else if (cmd == IP_VS_SO_SET_STOPDAEMON) {
-               struct ip_vs_daemon_user *dm = (struct ip_vs_daemon_user *)arg;
-               ret = stop_sync_thread(net, dm->state);
-               goto out_unlock;
        }
 
        usvc_compat = (struct ip_vs_service_user *)arg;
@@ -2584,6 +2594,33 @@ do_ip_vs_get_ctl(struct sock *sk, int cmd, void __user *user, int *len)
 
        if (copy_from_user(arg, user, copylen) != 0)
                return -EFAULT;
+       /*
+        * Handle daemons first since it has its own locking
+        */
+       if (cmd == IP_VS_SO_GET_DAEMON) {
+               struct ip_vs_daemon_user d[2];
+
+               memset(&d, 0, sizeof(d));
+               if (mutex_lock_interruptible(&ipvs->sync_mutex))
+                       return -ERESTARTSYS;
+
+               if (ipvs->sync_state & IP_VS_STATE_MASTER) {
+                       d[0].state = IP_VS_STATE_MASTER;
+                       strlcpy(d[0].mcast_ifn, ipvs->master_mcast_ifn,
+                               sizeof(d[0].mcast_ifn));
+                       d[0].syncid = ipvs->master_syncid;
+               }
+               if (ipvs->sync_state & IP_VS_STATE_BACKUP) {
+                       d[1].state = IP_VS_STATE_BACKUP;
+                       strlcpy(d[1].mcast_ifn, ipvs->backup_mcast_ifn,
+                               sizeof(d[1].mcast_ifn));
+                       d[1].syncid = ipvs->backup_syncid;
+               }
+               if (copy_to_user(user, &d, sizeof(d)) != 0)
+                       ret = -EFAULT;
+               mutex_unlock(&ipvs->sync_mutex);
+               return ret;
+       }
 
        if (mutex_lock_interruptible(&__ip_vs_mutex))
                return -ERESTARTSYS;
@@ -2681,28 +2718,6 @@ do_ip_vs_get_ctl(struct sock *sk, int cmd, void __user *user, int *len)
        }
        break;
 
-       case IP_VS_SO_GET_DAEMON:
-       {
-               struct ip_vs_daemon_user d[2];
-
-               memset(&d, 0, sizeof(d));
-               if (ipvs->sync_state & IP_VS_STATE_MASTER) {
-                       d[0].state = IP_VS_STATE_MASTER;
-                       strlcpy(d[0].mcast_ifn, ipvs->master_mcast_ifn,
-                               sizeof(d[0].mcast_ifn));
-                       d[0].syncid = ipvs->master_syncid;
-               }
-               if (ipvs->sync_state & IP_VS_STATE_BACKUP) {
-                       d[1].state = IP_VS_STATE_BACKUP;
-                       strlcpy(d[1].mcast_ifn, ipvs->backup_mcast_ifn,
-                               sizeof(d[1].mcast_ifn));
-                       d[1].syncid = ipvs->backup_syncid;
-               }
-               if (copy_to_user(user, &d, sizeof(d)) != 0)
-                       ret = -EFAULT;
-       }
-       break;
-
        default:
                ret = -EINVAL;
        }
@@ -3205,7 +3220,7 @@ static int ip_vs_genl_dump_daemons(struct sk_buff *skb,
        struct net *net = skb_sknet(skb);
        struct netns_ipvs *ipvs = net_ipvs(net);
 
-       mutex_lock(&__ip_vs_mutex);
+       mutex_lock(&ipvs->sync_mutex);
        if ((ipvs->sync_state & IP_VS_STATE_MASTER) && !cb->args[0]) {
                if (ip_vs_genl_dump_daemon(skb, IP_VS_STATE_MASTER,
                                           ipvs->master_mcast_ifn,
@@ -3225,7 +3240,7 @@ static int ip_vs_genl_dump_daemons(struct sk_buff *skb,
        }
 
 nla_put_failure:
-       mutex_unlock(&__ip_vs_mutex);
+       mutex_unlock(&ipvs->sync_mutex);
 
        return skb->len;
 }
@@ -3271,13 +3286,9 @@ static int ip_vs_genl_set_config(struct net *net, struct nlattr **attrs)
        return ip_vs_set_timeout(net, &t);
 }
 
-static int ip_vs_genl_set_cmd(struct sk_buff *skb, struct genl_info *info)
+static int ip_vs_genl_set_daemon(struct sk_buff *skb, struct genl_info *info)
 {
-       struct ip_vs_service *svc = NULL;
-       struct ip_vs_service_user_kern usvc;
-       struct ip_vs_dest_user_kern udest;
        int ret = 0, cmd;
-       int need_full_svc = 0, need_full_dest = 0;
        struct net *net;
        struct netns_ipvs *ipvs;
 
@@ -3285,19 +3296,10 @@ static int ip_vs_genl_set_cmd(struct sk_buff *skb, struct genl_info *info)
        ipvs = net_ipvs(net);
        cmd = info->genlhdr->cmd;
 
-       mutex_lock(&__ip_vs_mutex);
-
-       if (cmd == IPVS_CMD_FLUSH) {
-               ret = ip_vs_flush(net);
-               goto out;
-       } else if (cmd == IPVS_CMD_SET_CONFIG) {
-               ret = ip_vs_genl_set_config(net, info->attrs);
-               goto out;
-       } else if (cmd == IPVS_CMD_NEW_DAEMON ||
-                  cmd == IPVS_CMD_DEL_DAEMON) {
-
+       if (cmd == IPVS_CMD_NEW_DAEMON || cmd == IPVS_CMD_DEL_DAEMON) {
                struct nlattr *daemon_attrs[IPVS_DAEMON_ATTR_MAX + 1];
 
+               mutex_lock(&ipvs->sync_mutex);
                if (!info->attrs[IPVS_CMD_ATTR_DAEMON] ||
                    nla_parse_nested(daemon_attrs, IPVS_DAEMON_ATTR_MAX,
                                     info->attrs[IPVS_CMD_ATTR_DAEMON],
@@ -3310,6 +3312,33 @@ static int ip_vs_genl_set_cmd(struct sk_buff *skb, struct genl_info *info)
                        ret = ip_vs_genl_new_daemon(net, daemon_attrs);
                else
                        ret = ip_vs_genl_del_daemon(net, daemon_attrs);
+out:
+               mutex_unlock(&ipvs->sync_mutex);
+       }
+       return ret;
+}
+
+static int ip_vs_genl_set_cmd(struct sk_buff *skb, struct genl_info *info)
+{
+       struct ip_vs_service *svc = NULL;
+       struct ip_vs_service_user_kern usvc;
+       struct ip_vs_dest_user_kern udest;
+       int ret = 0, cmd;
+       int need_full_svc = 0, need_full_dest = 0;
+       struct net *net;
+       struct netns_ipvs *ipvs;
+
+       net = skb_sknet(skb);
+       ipvs = net_ipvs(net);
+       cmd = info->genlhdr->cmd;
+
+       mutex_lock(&__ip_vs_mutex);
+
+       if (cmd == IPVS_CMD_FLUSH) {
+               ret = ip_vs_flush(net);
+               goto out;
+       } else if (cmd == IPVS_CMD_SET_CONFIG) {
+               ret = ip_vs_genl_set_config(net, info->attrs);
                goto out;
        } else if (cmd == IPVS_CMD_ZERO &&
                   !info->attrs[IPVS_CMD_ATTR_SERVICE]) {
@@ -3536,13 +3565,13 @@ static struct genl_ops ip_vs_genl_ops[] __read_mostly = {
                .cmd    = IPVS_CMD_NEW_DAEMON,
                .flags  = GENL_ADMIN_PERM,
                .policy = ip_vs_cmd_policy,
-               .doit   = ip_vs_genl_set_cmd,
+               .doit   = ip_vs_genl_set_daemon,
        },
        {
                .cmd    = IPVS_CMD_DEL_DAEMON,
                .flags  = GENL_ADMIN_PERM,
                .policy = ip_vs_cmd_policy,
-               .doit   = ip_vs_genl_set_cmd,
+               .doit   = ip_vs_genl_set_daemon,
        },
        {
                .cmd    = IPVS_CMD_GET_DAEMON,
index 7ee7215b8ba048ffa94256a41b42b763b371fe71..3cdd479f9b5d179292182bb2711ec243b57736ee 100644 (file)
@@ -61,6 +61,7 @@
 
 #define SYNC_PROTO_VER  1              /* Protocol version in header */
 
+static struct lock_class_key __ipvs_sync_key;
 /*
  *     IPVS sync connection entry
  *     Version 0, i.e. original version.
@@ -1545,6 +1546,7 @@ int start_sync_thread(struct net *net, int state, char *mcast_ifn, __u8 syncid)
        IP_VS_DBG(7, "Each ip_vs_sync_conn entry needs %Zd bytes\n",
                  sizeof(struct ip_vs_sync_conn_v0));
 
+
        if (state == IP_VS_STATE_MASTER) {
                if (ipvs->master_thread)
                        return -EEXIST;
@@ -1667,6 +1669,7 @@ int __net_init ip_vs_sync_net_init(struct net *net)
 {
        struct netns_ipvs *ipvs = net_ipvs(net);
 
+       __mutex_init(&ipvs->sync_mutex, "ipvs->sync_mutex", &__ipvs_sync_key);
        INIT_LIST_HEAD(&ipvs->sync_queue);
        spin_lock_init(&ipvs->sync_lock);
        spin_lock_init(&ipvs->sync_buff_lock);
@@ -1680,7 +1683,9 @@ int __net_init ip_vs_sync_net_init(struct net *net)
 void ip_vs_sync_net_cleanup(struct net *net)
 {
        int retc;
+       struct netns_ipvs *ipvs = net_ipvs(net);
 
+       mutex_lock(&ipvs->sync_mutex);
        retc = stop_sync_thread(net, IP_VS_STATE_MASTER);
        if (retc && retc != -ESRCH)
                pr_err("Failed to stop Master Daemon\n");
@@ -1688,4 +1693,5 @@ void ip_vs_sync_net_cleanup(struct net *net)
        retc = stop_sync_thread(net, IP_VS_STATE_BACKUP);
        if (retc && retc != -ESRCH)
                pr_err("Failed to stop Backup Daemon\n");
+       mutex_unlock(&ipvs->sync_mutex);
 }