Most recent kernel where this bug did not occur: Distribution: Can't remember, possibly FC2. Hardware Environment: Software Environment: Problem Description: Steps to reproduce: 1. Boot NFS v3 TCP client running iptables & mount NFS filesystem 2. Do a normal NFS client reboot & try mounting the same filesystem again 3. Experience intermittent failure to read superblock The cause of this problem is NFS server's improper response to SYN packet sent by the client. This occurs *after* successful client authorization, when the client tries to open the connection (i.e. sends SYN to the server's nfs port) to read the superblock. The server (sometimes) responds with a pure ACK without the SYN bit set. This is blocked by iptables -- thus, mount fails with a "could not read superblock" message. Here is an excerpt from ethereal log: 3 0.021733 client SERVER TCP 800 > nfs [SYN] Seq=0 Ack=0 Win=5840 Len=0 MSS=1460 TSV=24095 TSER=0 WS=2 4 0.021846 SERVER client TCP nfs > 800 [ACK] Seq=9138391 Ack=3580883479 Win=16022 Len=0 TSV=244936050 TSER=1149400 5 0.021864 client SERVER ICMP Destination unreachable (Host administratively prohibited) The above problem occurs with a very simple default+ssh iptables configuration. Disabling iptables on the client makes the problem go away. Even with iptables active, there is no problem when nfsd responds with a proper [SYN,ACK] instead of just pure ACK (this happens intermittently after the client reboot). Please fix nfsd so that it reliably responds to SYN packets with proper [SYN,ACK] packets instead of just ACK packets. Apparently, nfsd state doesn't get properly reset on client reboots. Other people have reported autofs failures which may be related (e.g. on remounts).
Olaf Kirch responds with: We've seen this previously, and submitted a fix to netfilter which supposedly went into mainline at some point. It seems to be gone from 2.6.14 though. The problem is with conntrack, and filtering on RELATED (I assume your netfilter config does that) What happens is that the client reboots, opens a new TCP connection with the same port as last time (say 800), sends SYN. Server still has an active TCB for this, and thus replies with an ACK containing its current sequence numbers. Now the client is supposed to RST the connection. Unfortunately, conntrack does not expect a lone ACK in this state and ignores it. So the client will retransmit the SYN until timeout. Then it picks a new port, and succeeds (maybe).
Created attachment 6666 [details] Keep netfilter from dropping ACK probes for half-open connections From: Jozsef Kadlecsik Patch-mainline: submitted by Joszef Kadlecsik References: SUSE46818 Mounting NFS file systems after a (warm) reboot could take a long time if firewalling and connection tracking was enabled. The reason is that the NFS clients tends to use the same ports (800 and counting down). Now on reboot, the server would still have a TCB for an existing TCP connection client:800 -> server:2049. The client sends a SYN from port 800 to server:2049, which elicits an ACK from the server. The firewall on the client drops the ACK because (from its point of view) the connection is still in half-open state, and it expects to see a SYNACK. The client will eventually time out after several minutes. The following patch corrects this, by accepting ACKs on half open connections as well. Acked-By: Olaf Kirch <okir@suse.de>
Created attachment 6667 [details] Keep netfilter from ingnoring all RST if the previous packet was an ACK From: Martin Josefsson Patch-mainline: 2.6.11-rc1 References: SUSE50484 This is incremental fix to netfilter-tcp-rst-ack-fix (#46818) The change was that an RST is ignored if the previous packet was an ACK. This is happens all the time. I know it was intended as a fix for the SYN - ACK probe - RST sequence but it breaks normal usage. The problem is that connections that end with RST never get their state changed and are left in ESTABLISHED state with a large timeout. The patch below adds a check for !test_bit(IPS_ASSURED_BIT, &conntrack->status) so your change will only be active for unassured connections. Acked-By: Karsten Keil <kkeil@suse.de>
Neil points to the bugzilla report at https://bugzilla.novell.com/show_bug.cgi?id=104379 Our conclusion appears to be that this is a netfilter issue. Reassigning the bug to the networking folks...
If netfilter fixes help, fine. However, to me the current situation seems fragile from NFS perspective alone. Normal client reboot ought to (?) close the TCP connection to the NFS server -- this could also fix the problem. Is there a technical reason why normal client shutdown doesn't cause the server to close the connection?
The updated netfilter patch which fixes the problem has been sent to the netdev and netfilter-devel lists for rewiev and kernel inclusion. > Normal client reboot ought to (?) close the TCP connection to the NFS server > -- this could also fix the problem. Yes, exactly, however netfilter blocked the packets required to tear down the half-open connection. NFS or even the TCP stack cannot "workaround" that, so it was completely a netfilter issue.