From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: WARNING: at include/net/sock.h:417 udp_lib_unhash Date: Thu, 02 Jul 2009 08:10:14 +0200 Message-ID: <4A4C4F46.2070701@gmail.com> References: Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: NetDev , "Brandeburg, Jesse" , "Kirsher, Jeffrey T" , Jiri Olsa , "David S. Miller" To: "Tantilov, Emil S" Return-path: Received: from gw1.cosmosbay.com ([212.99.114.194]:49430 "EHLO gw1.cosmosbay.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751639AbZGBGLg (ORCPT ); Thu, 2 Jul 2009 02:11:36 -0400 In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: Tantilov, Emil S a =E9crit : > I see the following trace during netperf stress mixed UDP/TCP IPv4/6 = traffic. This is on recent pulls from net-2.6 and net-next. >=20 > [45197.989163] ------------[ cut here ]------------ > [45197.994309] WARNING: at include/net/sock.h:417 udp_lib_unhash+0x81= /0xab() > [45197.994311] Hardware name: X7DA8 > [45197.994314] Modules linked in: e1000 [last unloaded: e1000] > [45197.994326] Pid: 7110, comm: netserver Tainted: G W 2.6.31= -rc1-net-next-e1000-06250902 #8 > [45197.994331] Call Trace: > [45197.994336] [] ? udp_lib_unhash+0x81/0xab > [45197.994344] [] warn_slowpath_common+0x77/0x8f > [45197.994349] [] warn_slowpath_null+0xf/0x11 > [45197.994352] [] udp_lib_unhash+0x81/0xab > [45197.994357] [] sk_common_release+0x2f/0xb4 > [45197.994364] [] udp_lib_close+0x9/0xb > [45197.994369] [] inet_release+0x58/0x5f > [45197.994374] [] inet6_release+0x30/0x35 > [45197.994383] [] sock_release+0x1a/0x6c > [45197.994386] [] sock_close+0x22/0x26 > [45197.994392] [] __fput+0xf0/0x18c > [45197.994395] [] fput+0x15/0x19 > [45197.994399] [] filp_close+0x5c/0x67 > [45197.994404] [] sys_close+0x7b/0xb6 > [45197.994412] [] system_call_fastpath+0x16/0x1b > [45197.994418] ---[ end trace 5acab6fc0afdaaa3 ]--- >=20 > Emil-- Thanks for this report Emil. I could not find a recent change in this area in last kernels. If struct sk is hashed (sk_hashed() true), then sk_refcnt was increment= ed in sk_nulls_add_node_rcu(), thus its value should be >=3D 2. Maybe we have a missing memory barrier somewhere or a list corruption. 1) Could you try CONFIG_DEBUG_LIST=3Dy ? 2) Could you give model of cpu, since it reminds me the ongoing discuss= ion raised by Jiri Olsa. CPU1 does an atomic_inc(&sk->sk_refcnt) : refcnt changes from 1 to 2 then CPU2 does an atomic_read(&sk->sk_refcnt) and reads 1 instead of 2 David, maybe this test is not safe and if we really want to do a check we need to use a stronger atomic function. If you can reproduce this problem easily could you try following patch = ? Thank you diff --git a/include/net/sock.h b/include/net/sock.h index 352f06b..96ab278 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -393,8 +393,9 @@ static __inline__ int sk_del_node_init(struct sock = *sk) =20 if (rc) { /* paranoid for a while -acme */ - WARN_ON(atomic_read(&sk->sk_refcnt) =3D=3D 1); - __sock_put(sk); + int res =3D atomic_dec_return(&sk->sk_refcnt); + + WARN_ON(res <=3D 0); } return rc; } @@ -413,9 +414,9 @@ static __inline__ int sk_nulls_del_node_init_rcu(st= ruct sock *sk) int rc =3D __sk_nulls_del_node_init_rcu(sk); =20 if (rc) { - /* paranoid for a while -acme */ - WARN_ON(atomic_read(&sk->sk_refcnt) =3D=3D 1); - __sock_put(sk); + int res =3D atomic_dec_return(&sk->sk_refcnt); + + WARN_ON(res <=3D 0); } return rc; }