Date: Mon, 30 Apr 2012 12:54:00 +0300 From: Daniel Braniss <danny@cs.huji.ac.il> To: Rick Macklem <rmacklem@uoguelph.ca> Cc: killing@multiplay.co.uk, Hiroki Sato <hrs@FreeBSD.org>, freebsd-stable@FreeBSD.org, ob@e-Gitt.NET Subject: Re: 9-STABLE, ZFS, NFS, ggatec - suspected memory leak Message-ID: <E1SOnIu-000Dhp-Tf@kabab.cs.huji.ac.il> In-Reply-To: <99719742.103996.1335732516952.JavaMail.root@erie.cs.uoguelph.ca> References: <99719742.103996.1335732516952.JavaMail.root@erie.cs.uoguelph.ca>
next in thread | previous in thread | raw e-mail | index | archive | help
> Daniel Braniss wrote: > > > Daniel Braniss wrote: > > > > > Daniel Braniss wrote: > > > > > > > ----Security_Multipart(Fri_Apr_27_13_35_56_2012_748)-- > > > > > > > Content-Type: Text/Plain; charset=us-ascii > > > > > > > Content-Transfer-Encoding: 7bit > > > > > > > > > > > > > > Rick Macklem <rmacklem@uoguelph.ca> wrote > > > > > > > in > > > > > > > <1527622626.3418715.1335445225510.JavaMail.root@erie.cs.uoguelph.ca>: > > > > > > > > > > > > > > rm> Steven Hartland wrote: > > > > > > > rm> > ---- Original Message ----- > > > > > > > rm> > From: "Rick Macklem" <rmacklem@uoguelph.ca> > > > > > > > rm> > > At a glance, it looks to me like 8.x is affected. > > > > > > > Note > > > > > > > that > > > > > > > the > > > > > > > rm> > > bug only affects the new NFS server (the > > > > > > > experimental > > > > > > > one > > > > > > > for 8.x) > > > > > > > rm> > > when exporting ZFS volumes. (UFS exported volumes > > > > > > > don't > > > > > > > leak) > > > > > > > rm> > > > > > > > > > rm> > > If you are running a server that might be affected, > > > > > > > just: > > > > > > > rm> > > # vmstat -z | fgrep -i namei > > > > > > > rm> > > on the server and see if the 3rd number shown is > > > > > > > increasing. > > > > > > > rm> > > > > > > > > rm> > Many thanks Rick wasnt aware we had anything > > > > > > > experimental > > > > > > > enabled > > > > > > > rm> > but I think that would be a yes looking at these > > > > > > > number:- > > > > > > > rm> > > > > > > > > rm> > vmstat -z | fgrep -i namei > > > > > > > rm> > NAMEI: 1024, 0, 1, 1483, 25285086096, 0 > > > > > > > rm> > vmstat -z | fgrep -i namei > > > > > > > rm> > NAMEI: 1024, 0, 0, 1484, 25285945725, 0 > > > > > > > rm> > > > > > > > > rm> ^ > > > > > > > rm> I don't think so, since the 3rd number (USED) is 0 here. > > > > > > > rm> If that # is increasing over time, you have the leak. > > > > > > > You > > > > > > > are > > > > > > > rm> probably running the old (default in 8.x) NFS server. > > > > > > > > > > > > > > Just a report, I confirmed it affected 8.x servers running > > > > > > > newnfs. > > > > > > > > > > > > > > Actually I have been suffered from memory starvation > > > > > > > symptom on > > > > > > > that > > > > > > > server (24GB RAM) for a long time and watching vmstat -z > > > > > > > periodically. It stopped working once a week. I > > > > > > > investigated > > > > > > > the > > > > > > > vmstat log again and found the amount of NAMEI leak was > > > > > > > 11,543,956 > > > > > > > (about 11GB!) just before the locked-up. After applying the > > > > > > > patch, > > > > > > > the leak disappeared. Thank you for fixing it! > > > > > > > > > > > > > > -- Hiroki > > > > > And thanks Hiroki for testing it on 8.x. > > > > > > > > > > > this is on 8.2-STABLE/amd64 from around August: > > > > > > same here, this zfs+newnfs has been hanging every few months, > > > > > > and > > > > > > I > > > > > > can see > > > > > > now the leak, it's slowly increasing: > > > > > > NAMEI: 1024, 0, 122975, 529, 15417248, 0 > > > > > > NAMEI: 1024, 0, 122984, 520, 15421772, 0 > > > > > > NAMEI: 1024, 0, 123002, 502, 15424743, 0 > > > > > > NAMEI: 1024, 0, 123008, 496, 15425464, 0 > > > > > > > > > > > > cheers, > > > > > > danny > > > > > Maybe you could try the patch, too. > > > > > > > > > > It's at: > > > > > http://people.freebsd.org/~rmacklem/namei-leak.patch > > > > > > > > > > I'll commit it to head soon with a 1 month MFC, so that > > > > > hopefully > > > > > Oliver will have a chance to try it on his production server > > > > > before > > > > > the MFC. > > > > > > > > > > Thanks everyone, for your help with this, rick > > > > > > > > I haven't applied the patch yet, but in the meanime I have been > > > > running some > > > > experiments on a zfs/nfs server running 8.3-STABLE, and don't see > > > > any > > > > leaks > > > > what triggers the leak? > > > > > > > Fortunately Oliver isolated this. It should leak when you do a > > > successful > > > "rm" or "rmdir" while running the new/experimental server. > > > > > but that's what I did, I'm running the new/experimental nfs server > > (or so I think :-), and did a huge rm -rf and nothing, nada, no leak. > > To check the patch, I have to upgrade the production server, the one > > with the > > leak, > > but I wanted to test it on a non production first. Anyways, ill patch > > the > > kernel > > and try it on the leaking production server tomorrow. > > > Well, I think the patch should be harmless. > > You can check which server you are running by doing: > # nfsstat -e -s > - and see if the numbers are increasing > if they're zero or not increasing, you are running the old (default on 8.x) > server was running the wrong nfsd, now all is ok, and the patch works (obviously :-) BTW, if the if the experimental is not running then # nfsstat -e -s nfsstat: experimental client/server not loaded # danny
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?E1SOnIu-000Dhp-Tf>