From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andy Chittenden Subject: Re: nfs client hang Date: Tue, 27 Jul 2010 13:51:13 +0100 Message-ID: <4C4ED641.1030803@bluearc.com> References: <99613C19B13C5D40914FB8930657FA9303365708DE@uk-ex-mbx1.terastack.bluearc.com> <4C4E89D4.8040607@bluearc.com> <1280233276.2827.175.camel@edumazet-laptop> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: "Linux Kernel Mailing List (linux-kernel@vger.kernel.org)" , Trond Myklebust , netdev To: Eric Dumazet Return-path: In-Reply-To: <1280233276.2827.175.camel@edumazet-laptop> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On 2010-07-27 13:21, Eric Dumazet wrote: > Le mardi 27 juillet 2010 =C3=A0 11:53 +0100, Andy Chittenden a =C3=A9= crit : >>>>> IE the client starts a connection and then closes it again withou= t sending data. >>>> Once this happens, here's some rpcdebug info for the rpc module us= ing 2.6.34.1 kernel: >>>> >>>> ... lots of the following nfsv3 WRITE requests: >>>> [ 7670.026741] 57793 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026759] 57794 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026778] 57795 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026797] 57796 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026815] 57797 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026834] 57798 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026853] 57799 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026871] 57800 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026890] 57801 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7670.026909] 57802 0001 -11 ffff88012e32b000 (null) = 0 ffffffffa03beb10 nfsv3 WRITE a:call_reserveresult q:xprt_backlog >>>> [ 7680.520042] RPC: worker connecting xprt ffff88013e62d800 = via tcp to 10.1.6.102 (port 2049) >>>> [ 7680.520066] RPC: ffff88013e62d800 connect status 99 conne= cted 0 sock state 7 >>>> [ 7680.520074] RPC: 33550 __rpc_wake_up_task (now 4296812426) >>>> [ 7680.520079] RPC: 33550 disabling timer >>>> [ 7680.520084] RPC: 33550 removed from queue ffff88013e62db20 "xpr= t_pending" >>>> [ 7680.520089] RPC: __rpc_wake_up_task done >>>> [ 7680.520094] RPC: 33550 __rpc_execute flags=3D0x1 >>>> [ 7680.520098] RPC: 33550 xprt_connect_status: retrying >>>> [ 7680.520103] RPC: 33550 call_connect_status (status -11) >>>> [ 7680.520108] RPC: 33550 call_transmit (status 0) >>>> [ 7680.520112] RPC: 33550 xprt_prepare_transmit >>>> [ 7680.520118] RPC: 33550 rpc_xdr_encode (status 0) >>>> [ 7680.520123] RPC: 33550 marshaling UNIX cred ffff88012e002300 >>>> [ 7680.520130] RPC: 33550 using AUTH_UNIX cred ffff88012e002300 to= wrap rpc data >>>> [ 7680.520136] RPC: 33550 xprt_transmit(32920) >>>> [ 7680.520145] RPC: xs_tcp_send_request(32920) =3D -32 >>>> [ 7680.520151] RPC: xs_tcp_state_change client ffff88013e62d= 800... >>>> [ 7680.520156] RPC: state 7 conn 0 dead 0 zapped 1 >>> I changed that debug to output sk_shutdown too. That has a value of= 2 >>> (IE SEND_SHUTDOWN). Looking at tcp_sendmsg(), I see this: >>> err =3D -EPIPE; >>> if (sk->sk_err || (sk->sk_shutdown& SEND_SHUTDOWN)) >>> goto out_err; >>> which correlates with the trace "xs_tcp_send_request(32920) =3D -32= ". So, >>> this looks like a problem in the sockets/tcp layer. The rpc layer i= ssues >>> a shutdown and then reconnects using the same socket. So either >>> sk_shutdown needs zeroing once the shutdown completes or should be >>> zeroed on subsequent connect. The latter sounds safer. >> This patch for 2.6.34.1 fixes the issue: >> >> --- /home/company/software/src/linux-2.6.34.1/net/ipv4/tcp_output.c = 2010-07-27 08:46:46.917000000 +0100 >> +++ net/ipv4/tcp_output.c 2010-07-27 09:19:16.000000000 +0100 >> @@ -2522,6 +2522,13 @@ >> struct tcp_sock *tp =3D tcp_sk(sk); >> __u8 rcv_wscale; >> >> + /* clear down any previous shutdown attempts so that >> + * reconnects on a socket that's been shutdown leave the >> + * socket in a usable state (otherwise tcp_sendmsg() returns >> + * -EPIPE). >> + */ >> + sk->sk_shutdown =3D 0; >> + >> /* We'll fix this up when we get a response from the other = end. >> * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. >> */ >> >> As I mentioned in my first message, we first saw this issue in 2.6.3= 2 as supplied by debian (linux-image-2.6.32-5-amd64 Version: 2.6.32-17)= =2E It looks like the same patch would fix the problem there too. >> > CC netdev > > This reminds me a similar problem we had in the past, fixed with comm= it > 1fdf475a (tcp: tcp_disconnect() should clear window_clamp) > > But tcp_disconnect() already clears sk->sk_shutdown > > If NFS calls tcp_disconnect(), then shutdown(), there is a problem. > > Maybe xs_tcp_shutdown() should make some sanity tests ? > > Following sequence is legal, and your patch might break it. > > fd =3D socket(...); > shutdown(fd, SHUT_WR); > ... > connect(fd, ...); > > > Thanks for the response. From my reading of the RPC code, because=20 nothing clears the sk_shutdown flag, the RPC code goes into a loop when= =20 recovering from packets being lost: shutdown connect send fails so repeat My patch stops the NFS client hang that I was seeing but I'm not an=20 expert on either the socket layer, RPC code or NFS code so I'm happy fo= r=20 someone else to come up with the alternative, correct fix. --=20 Andy, BlueArc Engineering