libvirt will use rtl8139, but the qemu default changed a while
ago to e1000, and libvirt has diverged. e1000 is more modern and
considered the better default here, so use it
Don't rely on libvirt's default. This makes any XML changes more
explicit, and can help other parts of the code that may depend on
a machine type being encoded
- Remove anything for less than qemu 0.12 or libvirt 0.10, basically
rhel6 vintage stuff
- Open code some simple checks
- Remove some that are only used for unnecessary error reporting
Currently we hardcode these defaults for installing VMs:
<on_poweroff>destroy</on_poweroff>
<on_reboot>destroy</on_reboot>
<on_crash>destroy</on_crash>
And these for post-install config:
<on_poweroff>destroy</on_poweroff>
<on_reboot>restart</on_reboot>
<on_crash>restart</on_crash>
But this has some issues. on_crash=restart is problematic in that most
configs don't support it, and some like virtuozzo actively reject it.
Additionally this we should leave as much to libvirt defaults as we
can, which are more sensible in this case.
Drop everything except setting on_reboot=destroy for the install time
case (which is required for our config changes). Since on_poweroff
and on_crash already default to 'destroy', the install time config
is unchanged, and the post-install config now on_crash=destroy instead
of 'restart'.
The s390 on_crash=preserve workaround is dropped, which was intended
to avoid a reboot loop caused by on_crash=reboot
This feature is available in upstream Qemu since 2.0.0 but it was also
back-ported to downstream Qemu into qemu-kvm-1.5.3 released in RHEL-7.
Add a new check and enable hv_time also in RHEL systems if qemu new
enough.
Signed-off-by: Pavel Hrdina <phrdina@redhat.com>
We are just mirroring the behavior that virt-manager (and boxes) have
used for a while now.
In my experience the average user is confused by their VMs suspending,
so for our sake I'd rather make people opt into this feature.
Do not use the model specified by the OS if it is "virtio-net" and
virtionet is not supported.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1177113
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
The XML we use is:
<clock offset="utc">
<timer name="rtc" tickpolicy="catchup"/>
<timer name="pit" tickpolicy="delay"/>
<timer name="hpet" present="no"/>
</clock>
Which translates to the qemu commands:
-no-hpet -no-kvm-pit-reinjection -rtc driftfix=slew
The latter two bits are already used by openstack and gnome boxes by
default.
On RHEL hpet is compiled out so -no-hpet is the default,
but not everywhere else. Though recently the RH guys confirmed that
for regular usage it should be turned off because a) qemu support
is not that good, b) most users don't need it anyways c) it has
a performance penalty.
This default can be overridden from a virt-install command like
--clock rtc_tickpolicy=delay
When Guest() sees that a timer has already been defined, it won't
set the new defaults, and that setting above is the old implied
default rtc setting.
These bits only apply if virt-manager is running directly on a RHEL6.
Since latest virt-manager can't do that thanks to GTK3 conversion,
just drop it all.
It was required a long long time ago, before qemu supported -drive
and possibly ancient xen. Nowadays it should be pointless, and contributes
to some issues like bz 905439
This is where performance issues hit the most. This changes the
RHEL default to remove cache=None for all file volumes. If anyone
cares they can speak up, but it seems an overly unnecessary deviation
from upstream qemu.
The patches to drop /dev/loop0 meant get_blkdisk was inaccurate and
we weren't testing many things that were intended. Use a block disk
from the test driver.
- Unify looking up osdict values
- Have Guest.py pass in its defaults, not encoding them in osdict
- Explicitly specify every device parameter that osdict handles, rather
than have some psuedo magic dictionary format to handle it
The commit efab27afbf02743a3a2582e9a111eb1b7d985b26 in libvirt uses
/dev/%s for logical pools, change the tests to accommodate it.
(crobinso: additional fix to work with older libvirt)
We separate all the provisioning bits to diskbackend.py. VirtualDisk
users now need to explicitly opt in to storage creation by using
set_create_storage().
validation is no longer done automatically, users must call the
validation() command.
__init__ drops all extra parameters.
This will eventually get us to a point where we can unify the manual
XML building and XML parsing machinery, and get consistent validation
behavior between devices.
This patch adds and fixes test cases for cpu mode attribute.
Signed-off-by: Ken ICHIKAWA <ichikawa.ken@jp.fujitsu.com>
Signed-off-by: Hu Tao <hutao@cn.fujitsu.com>
(crobinso: Minor rebase tweaks)
In order to fix last tests that fail for me, I had to modify the
output xml, both modified machines will fit in the first numa node, so
the cpus from that node should be mentioned in the xml and that lead
me to the fact that the problem was in the test xml and not the code.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Cole Robinson