Re: 3.0+ NFS issues (bisected)

["J. Bruce Fields" <bfields@fieldses.org>]

On Fri, Aug 17, 2012 at 05:56:56AM +0400, Michael Tokarev wrote:
> On 12.07.2012 16:53, J. Bruce Fields wrote:
> > On Tue, Jul 10, 2012 at 04:52:03PM +0400, Michael Tokarev wrote:
> >> I tried to debug this again, maybe to reproduce in a virtual machine,
> >> and found out that it is only 32bit server code shows this issue:
> >> after updating the kernel on the server to 64bit (the same version)
> >> I can't reproduce this issue anymore.  Rebooting back to 32bit,
> >> and voila, it is here again.
> >>
> >> Something apparenlty isn't right on 32bits... ;)
> >>
> >> (And yes, the prob is still present and is very annoying :)
> > 
> > OK, that's very useful, thanks.  So probably a bug got introduced in the
> > 32-bit case between 2.6.32 and 3.0.
> > 
> > My personal upstream testing is normally all x86_64 only.  I'll kick off
> > a 32-bit install and see if I can reproduce this quickly.
> 
> Actually it has nothing to do with 32 vs 64 bits as I
> initially thought.  It happens on 64bits too, but takes
> more time (or data to transfer) to trigger.

That makes it sound like some kind of leak: you're hitting this case
eventually either way, but it takes longer in the case where you have
more (low) memory.

I wish I was more familiar with the tcp code.... What number exactly is
being compared against those limits, and how could we watch it from
userspace?

--b.

> 
> 
> > Let me know if you're able to narrow this down any more.
> 
> I bisected this issue to the following commit:
> 
> commit f03d78db65085609938fdb686238867e65003181
> Author: Eric Dumazet <eric.dumazet@gmail.com>
> Date:   Thu Jul 7 00:27:05 2011 -0700
> 
>     net: refine {udp|tcp|sctp}_mem limits
> 
>     Current tcp/udp/sctp global memory limits are not taking into account
>     hugepages allocations, and allow 50% of ram to be used by buffers of a
>     single protocol [ not counting space used by sockets / inodes ...]
> 
>     Lets use nr_free_buffer_pages() and allow a default of 1/8 of kernel ram
>     per protocol, and a minimum of 128 pages.
>     Heavy duty machines sysadmins probably need to tweak limits anyway.
> 
> 
> Reverting this commit on top of 3.0 (or any later 3.x kernel) fixes
> the behavour here.
> 
> This machine has 4Gb of memory.  On 3.0, with this patch applied
> (as it is part of 3.0), tcp_mem is like this:
> 
>   21228     28306   42456
> 
> with this patch reverted, tcp_mem shows:
> 
>   81216     108288  162432
> 
> and with these values, it works fine.
> 
> So it looks like something else goes wrong there,
> which lead to all nfsds fighting with each other
> for something and eating 100% of available CPU
> instead of servicing clients.
> 
> For added fun, when setting tcp_mem to the "good" value
> from "bad" value (after booting into kernel with that
> patch applied), the problem is _not_ fixed.
> 
> Any further hints?
> 
> Thanks,
> 
> /mjt
> 
> >> On 31.05.2012 17:51, Michael Tokarev wrote:
> >>> On 31.05.2012 17:46, Myklebust, Trond wrote:
> >>>> On Thu, 2012-05-31 at 17:24 +0400, Michael Tokarev wrote:
> >>> []
> >>>>> I started tcpdump:
> >>>>>
> >>>>>  tcpdump -npvi br0 -s 0 host 192.168.88.4 and \( proto ICMP or port 2049 \) -w nfsdump
> >>>>>
> >>>>> on the client (192.168.88.2).  Next I mounted a directory on the client,
> >>>>> and started reading (tar'ing) a directory into /dev/null.  It captured a
> >>>>> few stalls.  Tcpdump shows number of packets it got, the stalls are at
> >>>>> packet counts 58090, 97069 and 97071.  I cancelled the capture after that.
> >>>>>
> >>>>> The resulting file is available at http://www.corpit.ru/mjt/tmp/nfsdump.xz ,
> >>>>> it is 220Mb uncompressed and 1.3Mb compressed.  The source files are
> >>>>> 10 files of 1Gb each, all made by using `truncate' utility, so does not
> >>>>> take place on disk at all.  This also makes it obvious that the issue
> >>>>> does not depend on the speed of disk on the server (since in this case,
> >>>>> the server disk isn't even in use).
> >>>>
> >>>> OK. So from the above file it looks as if the traffic is mainly READ
> >>>> requests.
> >>>
> >>> The issue here happens only with reads.
> >>>
> >>>> In 2 places the server stops responding. In both cases, the client seems
> >>>> to be sending a single TCP frame containing several COMPOUNDS containing
> >>>> READ requests (which should be legal) just prior to the hang. When the
> >>>> server doesn't respond, the client pings it with a RENEW, before it ends
> >>>> up severing the TCP connection and then retransmitting.
> >>>
> >>> And sometimes -- speaking only from the behavour I've seen, not from the
> >>> actual frames sent -- server does not respond to the RENEW too, in which
> >>> case the client reports "nfs server no responding", and on the next
> >>> renew it may actually respond.  This happens too, but much more rare.
> >>>
> >>> During these stalls, ie, when there's no network activity at all,
> >>> the server NFSD threads are busy eating all available CPU.
> >>>
> >>> What does it all tell us? :)
> >>>
> >>> Thank you!
> >>>
> >>> /mjt
> >>> --
> >>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> >>> the body of a message to majordomo@vger.kernel.org
> >>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >>> Please read the FAQ at  http://www.tux.org/lkml/
> >>
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at  http://www.tux.org/lkml/
>