[fedora-virt] Weekly virt status

Mark McLoughlin markmc at redhat.com
Fri Feb 6 18:41:30 UTC 2009

  2009-03-03 Feature freeze (25 days)
  2009-03-10 Beta Freeze (32 days)
  2009-04-14 Final freeze (67 days)

F11 Alpha, Take 3

The weird anaconda blocker which caused hang while moving to the
timezone screen was never figured out, but it turns out that later
rawhide didn't have this bug. Truly bizarre:


Because of this, rel-eng froze Alpha for the third time. After that
things got even more messy and even more confusing:


Another blocker cropped up, this time VNC installs failing:


In the end, after a mere two day slip, the installer gods were
appeased and Fedora 11 Alpha (blink) shipped to an eagerly waiting


On the virt side, the "disable kvmclock on machines with unsync TSC"
workaround ended up making it in at the last minute:


Unfortunately, what looks like a pvmmu bug didn't get fixed in time:


This causes some guest installs on an F11 Alpha host to oops during
heavy network activity.


Another edition of FWN, and another excellent virt section:


Thanks to Dale Bewley.

F11 Release Notes

Dale Bewley has begun work on the F11 release notes:


The "Unified Kernel Image" section provided some minior confusion on
the fedora-virt list:

  'The kernel package in Fedora 11 supports booting as a guest domU,
   but will not function as a dom0 until such support is provided
   upstream. The most recent Fedora release with dom0 support is
   Fedora 8.'

To which one user responded:

  As I write this, I am running F10 on a hardware virt enabled machine
  ... and have installed F10 and other distros as guests. Hence, the
  statement above is unclear to me for F10 is running as both dom0 and

  Am I not reading something right?

The confusion here is that folks use the terms "dom0" and domU to
refer to Xen, whereas the user in question is actually running
KVM. The equivalent terms for KVM tend to just be "KVM host" and "KVM


While on vacation, Daniel Veillard released libvirt 0.6.0:


  * New features:
  - thread safety of the API and event handling (Daniel Berrange)
  - allow QEmu domains to survive daemon restart (Guido Gunther)
  - extended logging capabilities (Daniel Veillard)
  - support copy-on-write storage volumes (Daniel Berrange)
  - support of storage cache control options for QEmu/KVM (Daniel Berrange)

Understandably, a number of post-release issues cropped up:

  SELinux issue causing libvirtd launched dnsmasq to fail

  This causes libvirt virtual networks to fail to start. The fix will
  involve a small update to libvirt and selinux-policy changes.

  libvir: Remote error : no call waiting for reply with serial 2

  Dan Berrange has a fix for this upstream

Also, there are reports that if QEMU is too slow starting up, it can


A lively update was had as to whether the libvirt-0.6.0 update in F9
and F10 updates-testing should be withdrawn until some of the issues
are resolved. The current plan is to just fix the issues ASAP with
further updates to libvirt and selinux-policy.

F11 Features

All F11 virt features are now available under:



Glauber posted an initial qemu package, and solicited feedback. This
package did two things:

  1) Updated to a snapshot of QEMU svn and

  2) Split the package into multiple sub-package, with a sub-package
     for each target architecture

There was some concern that updating to QEMU svn is not a good idea
until it is known that a QEMU release is due soon. By coincidence,
Anthony Liguori proposed doing a QEMU release at the end of February:


All involved were mightily relieved.

The discussion then moved on to a near-bikeshed argument about how to
split the package. In the end, it was agreed to go with:

  - qemu-common, containing core bits

  - qemu-system-* packages containing emulators for groups of
    architectures, e.g. qemu-system-x86

  - qemu-user, containing the Linux kernel ABI emulators for all

  - qemu, a meta package which requires all sub-packages for
    compatibility purposes

Further discussion was had on how to handle the building of BIOS
images. [Details redacted to avoid scaring the children]

The whole thread can be found here:


Also related, Eduardo added a kvm-tools sub-package containing
kvm_stat and kvmtrace:


Device Assignment

The CONFIG_DMAR issue lumbered on this week. Firstly, Kyle McMartin's
patch to make intel_iommu=off the default is now in Andrew Morton's
queue for 2.6.29.

Progress on fixing the underlying issue is slow, but steady:


The focus is now on the fact that delaying for 2us after flushing the
I/O page tables from the CPU cache appears to resolve the
problem. Adar and Bhavesh from VMware are busily debugging the issue
on hardware which shows the problem. David Woodhouse and Chris Wright
have been working on patches to aid debugging. David has also included
Joerg Roedel's CONFIG_DMA_API_DEBUG patches in Fedora rawhide.

In other news, a minor circular locking issue was found and fixed:


Also, Weidong Han re-posted several device assignment fixes for QEMU
which had fallen through the cracks:


Some offline discussion was had around how to handle resetting PCI
devices before and after assignming them to guests. The details are
long and involved, but the core of the problem is that devices must be
reset before being assigned if they have been previously used in the

For example, with an e1000e NIC this manifests itself with "TX Unit
Hang" errors in the guest. Unfortunately, the preferred method for
resetting a PCI device (Function Level Reset, or FLR) is not
implemented by many devices. Alternatives include Secondary Bus Reset
and PCI PM D-state transitions. Both will need to be implemented in
addition to FLR support.

Xen Dom0

Some welcome interest was renewed in Xen Dom0 support on the
fedora-virt list this week with Gerd Hoffmann's addition of bzImage
loading support to the hypervisor:


Several fedora-virt users are now building and testing Dom0 kernels
using Jeremy Fitzhardinge's latest patch set.

kernel arches

Some changes to kernel architectures are on their way:


  The main user-visible changes would be:

    * 32-bit x86 would be built for i586 by default.
    * The x86_64 kernel would be installed on compatible hardware,
      even when installing a 32-bit OS
    * The PAE kernel would be installed on other 32-bit hardware,
      where it is supported

The proposal was accepted by FESCo and the kernel changes are already
in rawhide.

gcc 4.4

gcc-4.4 is coming in F11:


Jakub Jelinek did a test mass-rebuild of rawhide this week:


It seemed to go pretty well. No virt packages had any problems.


DOOM-O-METER: 191 open bugs last week, 192 this week. Awww!

  new_slab() oops from alloc_skb() with virtio_net, e1000 and rtl8139

  Marcelo jumped on this and began investigating. At first it
  looked like that it was a SLUB issue. However, the crafty
  Marcelo came up with a crafty LD_PRELOAD lib for disabling
  pvmmu for testing.

  Using Marcelo's hack, James Laska confirmed that the culprit is

  Both nVidia users who reported KVM MMU issues tryied to
  reproduce without nvidia loaded.

  The first reporter did indeed reproduce without nVidia and
  Marcelo logged into his machine. The problem turned out to
  be dodgy RAM though. Take home point - if something funny
  is going on, try running memtest.

  BUG_ON() in __shrink_dcache_sb() in rawhide x86_64 guest

  Another oops reported by James Laska. Not yet investigated.

  block-rw-range-check.patch breaks qcow2

  Fedora's single not-upstream kvm-userspace patch causes us to fail
  under kvm-autotest.

  qemu bios images source code not packaged

  The BIOS images in the qemu package aren't currently built
  from source. This is fixed in the kvm package, so should be
  resolved when we merge the two.

  QEmu uses 100% of two assigned cores for XP guest

  Perhaps the known issue with guest CPU usage accounting?

  virbr0 loses it's routing when managed by NetworkManager

  Reporter can't reproduce after an upgrade to f10, but it
  seems fairly certain that the problem was something
  twiddling ip_forward.

  The second request in recent times for libvirt autostart domains to
  be autoshutdown

  Sound under KVM requires exclusive access to the sound device

  The debate continues on how to integrate with a user's sound

  Issue with remote libvirt connections via SSH under virt-manager

  ambiguous dialog

  Jon McCann noticed a virt-manager dialog:

    Not Enough Free Space
    The requested volume capacity will exceed the available
    pool space when the volume is fully allocated. (4000M
    requested capacity >2118M available)
                                             [No] [Yes]

  I say "Yes!". Or maybe "No!". No, "Yes!" ... :-)

  xen pv_ops DomU BUG() after alpha install

More information about the Fedora-virt mailing list