I did some digging in the mmcesfuncs to see what happens server side on fail over.
Basically the server losing the IP is supposed to terminate all sessions and the receiver server sends ACK tickles. My current supposition is that for whatever reason, the losing server isn't releasing something and the client still has hold of a connection which is mostly dead. The tickle then fails to the client from the new server. This would explain why failing the IP back to the original server usually brings the client back to life. This is only my working theory at the moment as we can't reliably reproduce this. Next time it happens we plan to grab some netstat from each side. Then we plan to issue "mmcmi tcpack $cesIpPort $clientIpPort" on the server that received the IP and see if that fixes it (i.e. the receiver server didn't tickle properly). (Usage extracted from mmcesfuncs which is ksh of course). ... CesIPPort is colon separated IP:portnumber (of NFSd) for anyone interested. Then try and kill he sessions on the losing server to check if there is stuff still open and re-tickle the client. If we can get steps to workaround, I'll log a PMR. I suppose I could do that now, but given its non deterministic and we want to be 100% sure it's not us doing something wrong, I'm inclined to wait until we do some more testing. I agree with the suggestion that it's probably IO pending nodes that are affected, but don't have any data to back that up yet. We did try with a read workload on a client, but may we need either long IO blocked reads or writes (from the GPFS end). We also originally had soft as the default option, but saw issues then and the docs suggested hard, so we switched and also enabled sync (we figured maybe it was NFS client with uncommited writes), but neither have resolved the issues entirely. Difficult for me to say if they improved the issue though given its sporadic. Appreciate people's suggestions! Thanks Simon ________________________________________ From: [email protected] [[email protected]] on behalf of Jan-Frode Myklebust [[email protected]] Sent: 25 April 2017 18:04 To: gpfsug main discussion list Subject: Re: [gpfsug-discuss] NFS issues I *think* I've seen this, and that we then had open TCP connection from client to NFS server according to netstat, but these connections were not visible from netstat on NFS-server side. Unfortunately I don't remember what the fix was.. -jf tir. 25. apr. 2017 kl. 16.06 skrev Simon Thompson (IT Research Support) <[email protected]<mailto:[email protected]>>: Hi, >From what I can see, Ganesha uses the Export_Id option in the config file (which is managed by CES) for this. I did find some reference in the Ganesha devs list that if its not set, then it would read the FSID from the GPFS file-system, either way they should surely be consistent across all the nodes. The posts I found were from someone with an IBM email address, so I guess someone in the IBM teams. I checked a couple of my protocol nodes and they use the same Export_Id consistently, though I guess that might not be the same as the FSID value. Perhaps someone from IBM could comment on if FSID is likely to the cause of my problems? Thanks Simon On 25/04/2017, 14:51, "[email protected]<mailto:[email protected]> on behalf of Ouwehand, JJ" <[email protected]<mailto:[email protected]> on behalf of [email protected]<mailto:[email protected]>> wrote: >Hello, > >At first a short introduction. My name is Jaap Jan Ouwehand, I work at a >Dutch hospital "VU Medical Center" in Amsterdam. We make daily use of IBM >Spectrum Scale, Spectrum Archive and Spectrum Protect in our critical >(office, research and clinical data) business process. We have three >large GPFS filesystems for different purposes. > >We also had such a situation with cNFS. A failover (IPtakeover) was >technically good, only clients experienced "stale filehandles". We opened >a PMR at IBM and after testing, deliver logs, tcpdumps and a few months >later, the solution appeared to be in the fsid option. > >An NFS filehandle is built by a combination of fsid and a hash function >on the inode. After a failover, the fsid value can be different and the >client has a "stale filehandle". To avoid this, the fsid value can be >statically specified. See: > >https://www.ibm.com/support/knowledgecenter/STXKQY_4.2.2/com.ibm.spectrum. >scale.v4r22.doc/bl1adm_nfslin.htm > >Maybe there is also a value in Ganesha that changes after a failover. >Certainly since most sessions will be re-established after a failback. >Maybe you see more debug information with tcpdump. > > >Kind regards, > >Jaap Jan Ouwehand >ICT Specialist (Storage & Linux) >VUmc - ICT >E: [email protected]<mailto:[email protected]> >W: www.vumc.com<http://www.vumc.com> > > > >-----Oorspronkelijk bericht----- >Van: >[email protected]<mailto:[email protected]> >[mailto:[email protected]<mailto:[email protected]>] > Namens Simon Thompson >(IT Research Support) >Verzonden: dinsdag 25 april 2017 13:21 >Aan: [email protected]<mailto:[email protected]> >Onderwerp: [gpfsug-discuss] NFS issues > >Hi, > >We have recently started deploying NFS in addition our existing SMB >exports on our protocol nodes. > >We use a RR DNS name that points to 4 VIPs for SMB services and failover >seems to work fine with SMB clients. We figured we could use the same >name and IPs and run Ganesha on the protocol servers, however we are >seeing issues with NFS clients when IP failover occurs. > >In normal operation on a client, we might see several mounts from >different IPs obviously due to the way the DNS RR is working, but it all >works fine. > >In a failover situation, the IP will move to another node and some >clients will carry on, others will hang IO to the mount points referred >to by the IP which has moved. We can *sometimes* trigger this by manually >suspending a CES node, but not always and some clients mounting from the >IP moving will be fine, others won't. > >If we resume a node an it fails back, the clients that are hanging will >usually recover fine. We can reboot a client prior to failback and it >will be fine, stopping and starting the ganesha service on a protocol >node will also sometimes resolve the issues. > >So, has anyone seen this sort of issue and any suggestions for how we >could either debug more or workaround? > >We are currently running the packages >nfs-ganesha-2.3.2-0.ibm32_1.el7.x86_64 (4.2.2-2 release ones). > >At one point we were seeing it a lot, and could track it back to an >underlying GPFS network issue that was causing protocol nodes to be >expelled occasionally, we resolved that and the issues became less >apparent, but maybe we just fixed one failure mode so see it less often. > >On the clients, we use -o sync,hard BTW as in the IBM docs. > >On a client showing the issues, we'll see in dmesg, NFS related messages >like: >[Wed Apr 12 16:59:53 2017] nfs: server >MYNFSSERVER.bham.ac.uk<http://MYNFSSERVER.bham.ac.uk> not >responding, timed out > >Which explains the client hang on certain mount points. > >The symptoms feel very much like those logged in this Gluster/ganesha bug: >https://bugzilla.redhat.com/show_bug.cgi?id=1354439 > > >Thanks > >Simon > >_______________________________________________ >gpfsug-discuss mailing list >gpfsug-discuss at spectrumscale.org<http://spectrumscale.org> >http://gpfsug.org/mailman/listinfo/gpfsug-discuss >_______________________________________________ >gpfsug-discuss mailing list >gpfsug-discuss at spectrumscale.org<http://spectrumscale.org> >http://gpfsug.org/mailman/listinfo/gpfsug-discuss _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org<http://spectrumscale.org> http://gpfsug.org/mailman/listinfo/gpfsug-discuss _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
