[Fedora-xen] kernel-xen0-2.6.18-1.2200.fc5 amusingly unhappy

[Daniel P. Berrange]

On Sun, Oct 29, 2006 at 10:23:20AM +0800, Adrian Chadd wrote:
> On Sun, Oct 29, 2006, Adrian Chadd wrote:
> > On Sat, Oct 28, 2006, Daniel P. Berrange wrote:
> > 
> > > > Glad to know you've figured it out. Please let us know when there is something
> > > > in updates-testing we can try! Thanks!
> > > 
> > > It'll take me a little while to get a formal build made & pushed through the
> > > errata process, because I need to do some further testing before committing
> > > it to Fedora CVS. So if anyone is interested in testing in the meantime, I've
> > > put my unofficial prototype build up here http://people.redhat.com/berrange/xen/
> > > I'll remove RPMs once the formal build is in updates-testing.
> > 
> > Works for me - both on a FC5 machine thats been upgraded over time and on a
> > freshly-installed FC5 that's been upgraded just once. I'm running a slightly
> > modified config on one to support >1 vlan and physical interface; I've also
> > disabled the Xen TCP services and stuck to the UNIX services.
> 
> I spoke too soon: Now I'm seeing lots of these:
> 
> 4gb seg fixup, process named (pid 1266), cs:ip 73:00561378
> printk: 484562 messages suppressed.
> 4gb seg fixup, process poller.php (pid 1663), cs:ip 73:00540b0b
> printk: 1448 messages suppressed.
> 4gb seg fixup, process syslogd (pid 1235), cs:ip 73:00b0e81a
> 
> I'm running Ubuntu/Debian in some of my test Xen DomU's. This has apparently been
> fixed in 'etch' (with a "xen-aware" libc6) but why did this change between
> revision 2187 and revision 2200 of the xen kernels?

This issue has been present in Xen kernels ever since the 3.0.x series, if not
before. Basically libc uses -ve segment addressing, which means Xen hypervisor
has to do some magic tricks with segmentation to make things work. These tricks,
however, have a horrific performance impact on the guests so the Xen kernels
will warn about any process which does this -ve addressing. 

Since the vast majority of apps in Fedora are now fixed to not do -ve segment
addressing by default, we have the current kernels setup to print this warning
for any process which still uses the slow form of addressing, enabling easy
identification of apps which need fixing.

It sounds like you are using the Fedora kernels with a Debian distro which 
doesn't have any of the Xen fixups, hence you're seeing these warnings that
wouldn't normally be seen by Fedora users.

There's only really two solutions - either install a newer libc which has a 
'nosegneg' variant, or get rid of the existing 'tls' variant of libc and have
the guests use the older plain 'i686'  variant which doesn't do -ve segment
addressing.

Regards,
Dan.
-- 
|=- Red Hat, Engineering, Emerging Technologies, Boston.  +1 978 392 2496 -=|
|=-           Perl modules: http://search.cpan.org/~danberr/              -=|
|=-               Projects: http://freshmeat.net/~danielpb/               -=|
|=-  GnuPG: 7D3B9505   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505  -=|