On Thu, Apr 26, 2012 at 02:37:50PM +0200, Juerg Haefliger wrote: > On Thu, Apr 26, 2012 at 12:38 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > On Tue, Apr 24, 2012 at 08:26:04PM +0200, Juerg Haefliger wrote: > >> On Tue, Apr 24, 2012 at 2:07 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > >> > On Tue, Apr 24, 2012 at 10:55:22AM +0200, Juerg Haefliger wrote: > >> >> > Alright, then I need all the usual information. I suspect an event > >> >> > trace is the only way I'm going to see what is happening. I just > >> >> > updated the FAQ entry, so all the necessary info for gathering a > >> >> > trace should be there now. > >> >> > > >> >> > http://xfs.org/index.php/XFS_FAQ#Q:_What_information_should_I_include_when_reporting_a_problem.3F > >> >> > >> >> Very good. Will do. What kernel do you want me to run? I would prefer > >> >> our current production kernel (2.6.38-8-server) but I understand if > >> >> you want something newer. > >> > > >> > If you can reproduce it on a current kernel - 3.4-rc4 if possible, if > >> > not a 3.3.x stable kernel would be best. 2.6.38 is simply too old to > >> > be useful for debugging these sorts of problems... > >> > >> OK, I reproduced a hang running 3.4-rc4. The data is here but it's a > >> whopping 2GB (yes it's compressed): > >> https://region-a.geo-1.objects.hpcloudsvc.com:443/v1.0/AUTH_9630ead2-6194-40df-afd3-7395448d4536/xfs-hang/report-2012-04-24.tar > > > > That's a bit big to be useful, and far bigger than I'm willing to > > download given that I'm on the end of a wet piece of string, not a > > big fat intarwebby pipe. > > Fair enough. > > > > I'm assuming it is the event trace > > that is causing it to blow out? If so, just the 30-60s either side of > > the hang first showing up is probaby necessary, and that should cut > > the size down greatly.... > > Can I shorten the existing trace.dat? No idea, but that's likely the problem - I don't want the binary trace.dat file. I want the text output of the report command generated from the binary trace.dat file... > I stopped the trace > automatically 10 secs after the the xlog_... trace showed up in syslog > so effectively some 130+ secs after the hang occured. Extract the text report from it, and compress that. For example, a trace i've just done: $ ~/trace-cmd/trace-cmd report > trace.out $ ls -ltr |tail -4 -rw-r--r-- 1 root root 21430272 Apr 27 08:36 trace.dat -rw-r--r-- 1 root root 10039296 Apr 27 08:36 trace.dat.cpu1 -rw-r--r-- 1 root root 10035200 Apr 27 08:36 trace.dat.cpu0 -rw-r--r-- 1 dave dave 48255670 Apr 27 08:37 trace.out $ gzip trace.out $ ls -ltr |tail -4 -rw-r--r-- 1 root root 21430272 Apr 27 08:36 trace.dat -rw-r--r-- 1 root root 10039296 Apr 27 08:36 trace.dat.cpu1 -rw-r--r-- 1 root root 10035200 Apr 27 08:36 trace.dat.cpu0 -rw-r--r-- 1 dave dave 2500733 Apr 27 08:37 trace.out.gz Has 200MB of binary trace data, which generates a 470MB text output file, which compresses really well - down to 2.5MB in this case. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
