[ kvm-Bugs-2019053 ] tbench fails on guest when AMD NPT enabled

["SourceForge.net" <noreply@sourceforge.net>]

Bugs item #2019053, was opened at 2008-07-16 00:10
Message generated for change (Comment added) made by awwy
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=893831&aid=2019053&group_id=180599

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: amd
Group: None
Status: Open
Resolution: None
Priority: 5
Private: No
Submitted By: Alex Williamson (alex_williamson)
Assigned to: Nobody/Anonymous (nobody)
Summary: tbench fails on guest when AMD NPT enabled

Initial Comment:
Running on a dual-socket system with AMD 2356 quad-core processors (8 total cores), 32GB RAM, Ubuntu Hardy 2.6.24-19-generic (64bit) with kvm-71 userspace and kernel modules.  With no module options, dmesg confirms: kvm: Nested Paging enabled

Start guest with:

/usr/local/kvm/bin/qemu-system-x86_64 -hda /dev/VM/Ubuntu64 -m 1024 -net nic,model=e1000,mac=de:ad:be:ef:00:01 -net tap,script=/root/bin/br0-ifup -smp 8 -vnc :0

Guest VM is also Ubuntu Hardy 64bit.  On the guest run 'tbench 16 <tbench server>'.  System running tbench_srv is a different system in my case.

The tbench client will fail randomly, often quietly with "Child failed with status 1", but sometimes more harshly with a glibc double free error.

If I unload the modules and reload w/o npt:

modprobe -r kvm-amd
modprobe -r kvm
modprobe kvm-amd npt=0

dmesg confirms: kvm: Nested Paging disabled

The tbench test now runs over and over successfully.  The test also runs fine on an Intel E5450 (no EPT).

----------------------------------------------------------------------

Comment By: Alexander Graf (awwy)
Date: 2008-07-24 05:45

Message:
Logged In: YES 
user_id=376328
Originator: NO

I'm seeing random segfaults when using NPT on a host kernel >= 2.6.23. So
far I have not been able to reproduce my test case breakages with an
openSUSE 10.3 kernel though, so could you please test that and verify if
tbench works for you on openSUSE 10.3? It does break with 11.0.

I have the feeling that we're seeing the same problem here.

----------------------------------------------------------------------

Comment By: Alex Williamson (alex_williamson)
Date: 2008-07-16 15:18

Message:
Logged In: YES 
user_id=333914
Originator: YES

No, I added mlockall(MCL_CURRENT | MCL_FUTURE) to qemu/vl.c:main() and it
makes no difference.  I'm only starting a 1G guest on an otherwise idle 32G
host, so host memory pressure is pretty light.

----------------------------------------------------------------------

Comment By: Avi Kivity (avik)
Date: 2008-07-16 14:19

Message:
Logged In: YES 
user_id=539971
Originator: NO

Strange.  If you add an mlockall() to qemu startup, does the test pass?

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=893831&aid=2019053&group_id=180599