[Qemu-discuss] Corruption on Windows VMs when running 4.10+ kernel

[Rene Bakkum]

Hi,

Sorry if I miss any information. If possible I'll try to add as much 
information as I can.

Problem:
Running Windows VMs with MSSQL on KVM causes SQL log corruption when 
running on Kernel 4.10 or higher.

Setup:
- Host OS: Ubuntu 16.04 LTS
- Host kernel: 4.13.0-39-generic #44~16.04.1-Ubuntu SMP
- QEMU/KVM version: 1:2.5+dfsg-5ubuntu10.25
- Virtual Disks in LVM
- Guest OS: Windows 2016 + SQL 2016
*included xml of a VM configuration*

Reproduce:
Running above setup (default Ubuntu 16.04 with hwe kernel installed) and 
a Windows VM running and run an SQL benchmark (we used Benchmark Factory 
Trail 8.0.1). After a couple of minutes the SQL logs will become corrupt.

Workaround:
- Change the disk cache from 'none' to 'writethrough'
- Revert back to 4.4 / 4.8 kernel

Background information:
We've done some testing and when we are running the 4.8 kernel from the 
Ubuntu repo, all is fine and we can run large tests on the SQL servers. 
When we upgrade to a 4.10, 4.13 or 4.15 kernel it breaks after 
generating roughly 500Mb of logs, than SQL servers start to report the 
ldf files are corrupt and SQL mirrors start to fail. We also see more 
often that SQL servers are putting databases in Suspect mode. However 
this is harder to reproduce and could be something else. However we 
believe the problems are related to each other.
With changing the disk cache from none to writethrough seems to solve 
the problem, but the side effect is that live migrations are not safe 
anymore to do. Reverting to a 4.4 (lts) or 4.8 (16.10 kernel) seems also 
to solve the issue, but with some hardware choices we can't really stay 
behind on the kernel versions that long.

We've done some testing on 18.04 as well, and the problem seems to be 
there as well. In that case we were running the 4.15 kernel.

Questions:
- Am I on the right list? ;)
- Does anyone else experience the same behavior?
- Better yet, does anyone have a better solution?

Thanks in advanced!