On 11/14/14 2:02 AM, Daniel P. Berrange wrote:
On Thu, Nov 13, 2014 at 01:55:06PM -0800, Surojit Pathak wrote:
Hi all,

[Issue observed]
If we issue 'nova reboot <server>', we get to have the console output of the
latest bootup of the server only. The console output of the previous boot
for the same server vanishes due to truncation[1]. If we do reboot from
within the VM instance [ #sudo reboot ], or reboot the instance with 'virsh
reboot <instance>' the behavior is not the same, where the console.log keeps
increasing, with the new output being appended.
This loss of history makes some debugging scenario difficult due to lack of
information being available.

Please point me to any solution/blueprint for this issue, if already
planned. Otherwise, please comment on my analysis and proposals as solution,
below -

[Analysis]
Nova's libvirt driver on compute node tries to do a graceful restart of the
server instance, by attempting a soft_reboot first. If soft_reboot fails, it
attempts a hard_reboot. As part of soft_reboot, it brings down the instance
by calling shutdown(), and then calls createWithFlags() to bring this up.
Because of this, qemu-kvm process for the instance gets terminated and new
process is launched. In QEMU, the chardev file is opened with O_TRUNC, and
thus we lose the previous content of the console.log file.
On the other-hand, during 'virsh reboot <instance>', the same qemu-kvm
process continues, and libvirt actually does a qemuDomainSetFakeReboot().
Thus the same file continues capturing the new console output as a
continuation into the same file.
Nova and libvirt have support for issuing a graceful reboot via the QEMU
guest agent. So if you make sure that is installed, and tell Nova to use
it, then Nova won't have to stop & recreate the QEMU process and thus
won't have the problem of overwriting the logs.
Hi Daniel,
Having GA to do graceful restart is nice option. But if it were to just preserve the same console file, even 'virsh reboot' achieves the purpose. As I explained in my original analysis, Nova seems to have not taken the path, as it does not want to have a false positive, where the GA does not respond or 'virDomain.reboot' fails later and the domain is not really restarted. [ CC-ed vish, author of nova/virt/libvirt/driver.py ]

IMHO, QEMU should preserve the console-log file for a given domain, if it exists, by not opening with O_TRUNC option, instead opening with O_APPEND. I would like to draw a comparison of a real computer to which we might be connected over serial console, and the box gets powered down and up with external button press, and we do not lose the console history, if connected. And that's what is the experience console-log intends to provide. If you think, this is agreeable, please let me know, I will send the patch to qemu-devel@.

-- 
Regards,
SURO