Re: net/ipv4: deadlock in ip_ra_control

[Cong Wang]

On Wed, Mar 1, 2017 at 2:44 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> Hello,
>
> I've got the following deadlock report while running syzkaller fuzzer
> on linux-next/51788aebe7cae79cb334ad50641347465fc188fd:
>
> ======================================================
> [ INFO: possible circular locking dependency detected ]
> 4.10.0-next-20170301+ #1 Not tainted
> -------------------------------------------------------
> syz-executor1/3394 is trying to acquire lock:
>  (sk_lock-AF_INET){+.+.+.}, at: [<ffffffff838864cc>] lock_sock
> include/net/sock.h:1460 [inline]
>  (sk_lock-AF_INET){+.+.+.}, at: [<ffffffff838864cc>]
> do_ip_setsockopt.isra.12+0x21c/0x3540 net/ipv4/ip_sockglue.c:652
>
> but task is already holding lock:
>  (rtnl_mutex){+.+.+.}, at: [<ffffffff836fbd97>] rtnl_lock+0x17/0x20
> net/core/rtnetlink.c:70
>
> which lock already depends on the new lock.
>
>
> the existing dependency chain (in reverse order) is:
>
> -> #1 (rtnl_mutex){+.+.+.}:
>        validate_chain kernel/locking/lockdep.c:2265 [inline]
>        __lock_acquire+0x2149/0x3430 kernel/locking/lockdep.c:3338
>        lock_acquire+0x2a1/0x630 kernel/locking/lockdep.c:3753
>        __mutex_lock_common kernel/locking/mutex.c:754 [inline]
>        __mutex_lock+0x172/0x1730 kernel/locking/mutex.c:891
>        mutex_lock_nested+0x16/0x20 kernel/locking/mutex.c:906
>        rtnl_lock+0x17/0x20 net/core/rtnetlink.c:70
>        mrtsock_destruct+0x86/0x2c0 net/ipv4/ipmr.c:1281
>        ip_ra_control+0x459/0x600 net/ipv4/ip_sockglue.c:372
>        do_ip_setsockopt.isra.12+0x1064/0x3540 net/ipv4/ip_sockglue.c:1161
>        ip_setsockopt+0x3a/0xb0 net/ipv4/ip_sockglue.c:1264
>        raw_setsockopt+0xb7/0xd0 net/ipv4/raw.c:839
>        sock_common_setsockopt+0x95/0xd0 net/core/sock.c:2725
>        SYSC_setsockopt net/socket.c:1786 [inline]
>        SyS_setsockopt+0x25c/0x390 net/socket.c:1765
>        entry_SYSCALL_64_fastpath+0x1f/0xc2
>
> -> #0 (sk_lock-AF_INET){+.+.+.}:
>        check_prev_add kernel/locking/lockdep.c:1828 [inline]
>        check_prevs_add+0xa8f/0x19f0 kernel/locking/lockdep.c:1938
>        validate_chain kernel/locking/lockdep.c:2265 [inline]
>        __lock_acquire+0x2149/0x3430 kernel/locking/lockdep.c:3338
>        lock_acquire+0x2a1/0x630 kernel/locking/lockdep.c:3753
>        lock_sock_nested+0xcb/0x120 net/core/sock.c:2530
>        lock_sock include/net/sock.h:1460 [inline]
>        do_ip_setsockopt.isra.12+0x21c/0x3540 net/ipv4/ip_sockglue.c:652
>        ip_setsockopt+0x3a/0xb0 net/ipv4/ip_sockglue.c:1264
>        tcp_setsockopt+0x82/0xd0 net/ipv4/tcp.c:2721
>        sock_common_setsockopt+0x95/0xd0 net/core/sock.c:2725
>        SYSC_setsockopt net/socket.c:1786 [inline]
>        SyS_setsockopt+0x25c/0x390 net/socket.c:1765
>        entry_SYSCALL_64_fastpath+0x1f/0xc2
>

Please try the attached patch (compile only).

Thanks.
diff --git a/net/ipv4/ip_sockglue.c b/net/ipv4/ip_sockglue.c
index ebd953b..bda318a 100644
--- a/net/ipv4/ip_sockglue.c
+++ b/net/ipv4/ip_sockglue.c
@@ -591,6 +591,7 @@ static bool setsockopt_needs_rtnl(int optname)
 	case MCAST_LEAVE_GROUP:
 	case MCAST_LEAVE_SOURCE_GROUP:
 	case MCAST_UNBLOCK_SOURCE:
+	case IP_ROUTER_ALERT:
 		return true;
 	}
 	return false;
diff --git a/net/ipv4/ipmr.c b/net/ipv4/ipmr.c
index beacd02..932321b 100644
--- a/net/ipv4/ipmr.c
+++ b/net/ipv4/ipmr.c
@@ -1278,7 +1278,7 @@ static void mrtsock_destruct(struct sock *sk)
 	struct net *net = sock_net(sk);
 	struct mr_table *mrt;
 
-	rtnl_lock();
+	ASSERT_RTNL();
 	ipmr_for_each_table(mrt, net) {
 		if (sk == rtnl_dereference(mrt->mroute_sk)) {
 			IPV4_DEVCONF_ALL(net, MC_FORWARDING)--;
@@ -1289,7 +1289,6 @@ static void mrtsock_destruct(struct sock *sk)
 			mroute_clean_tables(mrt, false);
 		}
 	}
-	rtnl_unlock();
 }
 
 /* Socket options and virtual interface manipulation. The whole
@@ -1353,13 +1352,8 @@ int ip_mroute_setsockopt(struct sock *sk, int optname, char __user *optval,
 		if (sk != rcu_access_pointer(mrt->mroute_sk)) {
 			ret = -EACCES;
 		} else {
-			/* We need to unlock here because mrtsock_destruct takes
-			 * care of rtnl itself and we can't change that due to
-			 * the IP_ROUTER_ALERT setsockopt which runs without it.
-			 */
-			rtnl_unlock();
 			ret = ip_ra_control(sk, 0, NULL);
-			goto out;
+			goto out_unlock;
 		}
 		break;
 	case MRT_ADD_VIF:
@@ -1470,7 +1464,6 @@ int ip_mroute_setsockopt(struct sock *sk, int optname, char __user *optval,
 	}
 out_unlock:
 	rtnl_unlock();
-out:
 	return ret;
 }