From owner-freebsd-stable@FreeBSD.ORG Sat Oct 31 09:16:35 2009 Return-Path: Delivered-To: freebsd-stable@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2001:4f8:fff6::34]) by hub.freebsd.org (Postfix) with ESMTP id 97D30106566C; Sat, 31 Oct 2009 09:16:35 +0000 (UTC) (envelope-from danny@cs.huji.ac.il) Received: from kabab.cs.huji.ac.il (kabab.cs.huji.ac.il [132.65.16.84]) by mx1.freebsd.org (Postfix) with ESMTP id 4964A8FC0C; Sat, 31 Oct 2009 09:16:35 +0000 (UTC) Received: from pampa.cs.huji.ac.il ([132.65.80.32]) by kabab.cs.huji.ac.il with esmtp id 1N4A4U-000Ay9-Kc; Sat, 31 Oct 2009 11:16:30 +0200 X-Mailer: exmh version 2.7.2 01/07/2005 with nmh-1.2 To: Rick Macklem In-reply-to: References: <20091027164159.GU841@twoquid.cs.ru.nl> Comments: In-reply-to Rick Macklem message dated "Wed, 28 Oct 2009 16:38:27 -0400." Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Date: Sat, 31 Oct 2009 11:16:30 +0200 From: Daniel Braniss Message-ID: Cc: freebsd-stable@freebsd.org, dfr@freebsd.org, freebsd-current@freebsd.org, Olaf Seibert Subject: Re: 8.0-RC1 NFS client timeout issue X-BeenThere: freebsd-stable@freebsd.org X-Mailman-Version: 2.1.5 Precedence: list List-Id: Production branch of FreeBSD source code List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Sat, 31 Oct 2009 09:16:35 -0000 > > First off, I know that cross posting is evil, but I wanted to try > and make sure developers saw it. > > On Tue, 27 Oct 2009, Olaf Seibert wrote: > > > I see an annoying behaviour with NFS over TCP. It happens both with nfs > > and newnfs. This is with FreeBSD/amd64 8.0-RC1 as client. The server is > > some Linux or perhaps Solaris, I'm not entirely sure. > > > > After trying to find something in packet traces, I think I have found > > something. > > > > The scenario seems to be as follows. Sorry for the width of the lines. > > > > > > No. Time Source Destination Protocol Info > > 2296 2992.216855 xxx.xxx.31.43 xxx.xxx.16.142 NFS V3 LOOKUP Call (Reply In 2297), DH:0x3819da36/w > > 2297 2992.217107 xxx.xxx.16.142 xxx.xxx.31.43 NFS V3 LOOKUP Reply (Call In 2296) Error:NFS3ERR_NOENT > > 2298 2992.217141 xxx.xxx.31.43 xxx.xxx.16.142 NFS V3 LOOKUP Call (Reply In 2299), DH:0x170cb16a/bin > > 2299 2992.217334 xxx.xxx.16.142 xxx.xxx.31.43 NFS V3 LOOKUP Reply (Call In 2298), FH:0x61b8eb12 > > 2300 2992.217361 xxx.xxx.31.43 xxx.xxx.16.142 NFS V3 ACCESS Call (Reply In 2301), FH:0x61b8eb12 > > 2301 2992.217582 xxx.xxx.16.142 xxx.xxx.31.43 NFS V3 ACCESS Reply (Call In 2300) > > 2302 2992.217605 xxx.xxx.31.43 xxx.xxx.16.142 NFS V3 LOOKUP Call (Reply In 2303), DH:0x61b8eb12/w > > 2303 2992.217860 xxx.xxx.16.142 xxx.xxx.31.43 NFS V3 LOOKUP Reply (Call In 2302) Error:NFS3ERR_NOENT > > 2304 2992.318770 xxx.xxx.31.43 xxx.xxx.16.142 TCP 934 > nfs [ACK] Seq=238293 Ack=230289 Win=8192 Len=0 TSV=86492342 TSER=12393434 > > 2306 3011.537520 xxx.xxx.16.142 xxx.xxx.31.43 NFS V3 GETATTR Reply (Call In 2305) Directory mode:2755 uid:4100 gid:4100 > > 2307 3011.637744 xxx.xxx.31.43 xxx.xxx.16.142 TCP 934 > nfs [ACK] Seq=238429 Ack=230405 Win=8192 Len=0 TSV=86511662 TSER=12395366 > > 2308 3371.534980 xxx.xxx.16.142 xxx.xxx.31.43 TCP nfs > 934 [FIN, ACK] Seq=230405 Ack=238429 Win=49232 Len=0 TSV=12431366 TSER=86511662 > > > > The server decides, for whatever reason, to terminate the > > connection and sends a FIN. > > > > 2309 3371.535018 xxx.xxx.31.43 xxx.xxx.16.142 TCP 934 > nfs [ACK] Seq=238429 Ack=230406 Win=8192 Len=0 TSV=86871578 TSER=12431366 > > > > Client acknowledges this, > > > > 2310 3375.379693 xxx.xxx.31.43 xxx.xxx.16.142 NFS V3 ACCESS Call, FH:0x008002a2 > > > > but tries to sneak in another call anyway. [A] > > > Probably not the best behaviour, but I think it is technically allowed by > TCP. (My TCP is very rusty, but I think the socket should be in > TCPS_CLOSE_WAIT at this point and the BSD code will have called > socantrcvmore(), but not socantsndmore().) > > > 2311 3375.474788 xxx.xxx.16.142 xxx.xxx.31.43 TCP nfs > 934 [ACK] Seq=230406 Ack=238569 Win=49232 Len=0 TSV=12431760 TSER=86875423 > > > > Server ACKs but doesn't send anything else... [B] > > > > Time passes... > > > This is where it seems interesting. It looks to me like the socket upcall > for receiving the FIN would have happened before this point, setting the > ct_error.re_status to RPC_CANTRECV, but the code in clnt_vc_call() doesn't > check for this. (It does check for it happening during and after the > sosend(), but not before it, from what I can see.) > > > > > [B] would be a bug of the server in my opinion. If it ACKs a call, it > > should send a reply. And if it can't, it shouldn't. > > > I'll leave this one for the TCP wizzards. I'm not sure what the > correct behaviour is when data is received on a connection. (I think > it is waiting for a FIN from the client side at this point.) > > If you could try the following patch and see if it helps, that would be > appreciated, rick > ps: I'll try to reproduce the situation here, but I'm not sure if I can. > --- rpc/clnt_vc.c.sav 2009-10-28 15:44:20.000000000 -0400 > +++ rpc/clnt_vc.c 2009-10-28 15:49:57.000000000 -0400 > @@ -413,6 +413,19 @@ > > cr->cr_xid = xid; > mtx_lock(&ct->ct_lock); > + /* > + * Check to see if the other end has already started to close down > + * the connection. If it happens after this point, it will be > + * detected below, when cr->cr_error is checked. > + */ > + if (ct->ct_error.re_status == RPC_CANTRECV) { > + if (errp != &ct->ct_error) { > + errp->re_errno = ct->ct_error.re_errno; > + errp->re_status = RPC_CANTRECV; > + } > + stat = RPC_CANTRECV; > + goto out; > + } > TAILQ_INSERT_TAIL(&ct->ct_pending, cr, cr_link); > mtx_unlock(&ct->ct_lock); Did a make buildworld -j8, using as server netapp, freebsd, and all seems ok. danny