qemu-1.4.0 and onwards, linux kernel 3.2.x, heavy I/O leads to kernel_hung_tasks_timout_secs message and unresponsive qemu-process

[Oliver Francke]

Hi,

after some testing I tried to narrow down a problem, which was initially reported by some users.
Seen on different distros - debian 7.1, ubuntu 12.04 LTS, IPFire-2.3 as reported by now.

All using some flavour of linux-3.2.x kernel.

Tried e.g. under Ubuntu an upgrade to "Linux 3.8.0-27-generic x86_64" which solves the problem.
Problem could be triggert with some workload ala:

spew -v --raw -P -t -i 3 -b 4k -p random -B 4k 1G /tmp/doof.dat
and in parallel do some apt-get install/remove/whatever.

That results in a somewhat stuck qemu-session with the bad "kernel_hung_task..." messages.

A typical command-line is as follows:

/usr/local/qemu-1.6.0/bin/qemu-system-x86_64 -usbdevice tablet -enable-kvm -daemonize -pidfile /var/run/qemu-server/760.pid -monitor unix:/var/run/qemu-server/760.mon,server,nowait -vnc unix:/var/run/qemu-server/760.vnc,password -qmp unix:/var/run/qemu-server/760.qmp,server,nowait -nodefaults -serial none -parallel none -device virtio-net-pci,mac=00:F1:70:00:2F:80,netdev=vlan0d0 -netdev type=tap,id=vlan0d0,ifname=tap760i0d0,script=/etc/fcms/add_if.sh,downscript=/etc/fcms/downscript.sh -name 1155823384-4 -m 512 -vga cirrus -k de -smp sockets=1,cores=1 -device virtio-blk-pci,drive=virtio0 -drive format=raw,file=rbd:1155823384/vm-760-disk-1.rbd:rbd_cache=false,cache=writeback,if=none,id=virtio0,media=disk,index=0,aio=native -drive format=raw,file=rbd:1155823384/vm-760-swap-1.rbd:rbd_cache=false,cache=writeback,if=virtio,media=disk,index=1,aio=native -drive if=ide,media=cdrom,id=ide1-cd0,readonly=on -drive if=ide,media=cdrom,id=ide1-cd1,readonly=on -boot order=dc

no "system_reset", "sendkey ctrl-alt-delete" or "q" in monitoring-session is accepted, need to hard-kill the process.

Please give any advice on what to do for tracing/debugging, because the number of tickets here are raising, and noone knows, what users are doing inside their VM.

Kind regards,

Oliver Francke.

[Stefan Hajnoczi (stefanha)]

On Fri, Aug 02, 2013 at 09:58:29AM -0000, Oliver Francke wrote:
> after some testing I tried to narrow down a problem, which was initially reported by some users.
> Seen on different distros - debian 7.1, ubuntu 12.04 LTS, IPFire-2.3 as reported by now.
>
> All using some flavour of linux-3.2.x kernel.
>
> Tried e.g. under Ubuntu an upgrade to "Linux 3.8.0-27-generic x86_64" which solves the problem.

Is that a guest kernel upgrade?

> Problem could be triggert with some workload ala:
>
> spew -v --raw -P -t -i 3 -b 4k -p random -B 4k 1G /tmp/doof.dat
> and in parallel do some apt-get install/remove/whatever.
>
> That results in a somewhat stuck qemu-session with the bad
> "kernel_hung_task..." messages.
>
> A typical command-line is as follows:
>
> /usr/local/qemu-1.6.0/bin/qemu-system-x86_64 -usbdevice tablet -enable-
> kvm -daemonize -pidfile /var/run/qemu-server/760.pid -monitor
> unix:/var/run/qemu-server/760.mon,server,nowait -vnc unix:/var/run/qemu-
> server/760.vnc,password -qmp unix:/var/run/qemu-
> server/760.qmp,server,nowait -nodefaults -serial none -parallel none
> -device virtio-net-pci,mac=00:F1:70:00:2F:80,netdev=vlan0d0 -netdev
> type=tap,id=vlan0d0,ifname=tap760i0d0,script=/etc/fcms/add_if.sh,downscript=/etc/fcms/downscript.sh
> -name 1155823384-4 -m 512 -vga cirrus -k de -smp sockets=1,cores=1
> -device virtio-blk-pci,drive=virtio0 -drive
> format=raw,file=rbd:1155823384/vm-760-disk-1.rbd:rbd_cache=false,cache=writeback,if=none,id=virtio0,media=disk,index=0,aio=native
> -drive
> format=raw,file=rbd:1155823384/vm-760-swap-1.rbd:rbd_cache=false,cache=writeback,if=virtio,media=disk,index=1,aio=native
> -drive if=ide,media=cdrom,id=ide1-cd0,readonly=on -drive
> if=ide,media=cdrom,id=ide1-cd1,readonly=on -boot order=dc
>
> no "system_reset", "sendkey ctrl-alt-delete" or "q" in monitoring-
> session is accepted, need to hard-kill the process.

Yesterday I saw a possibly related report on IRC. It was a Windows
guest running under OpenStack with images on Ceph.

They reported that the QEMU process would lock up - ping would not work
and their management tools showed 0 CPU activity for the guest.
However, they were able to "kick" the guest by taking a VNC screenshot
(I think). Then it would come back to life.

If you have a Linux guest that is reporting kernel_hung_task, then it
could be a similar scenario.

Please confirm that the hung task message is from inside the guest.

If you are able to reproduce this and have an alternative non-Ceph
storage pool, please try that since Ceph is common to both these bug
reports.

Stefan

[Oliver Francke (oliver-francke)]

Hi Stefan,

Am 02.08.2013 um 17:24 schrieb Stefan Hajnoczi <email address hidden>:

> On Fri, Aug 02, 2013 at 09:58:29AM -0000, Oliver Francke wrote:
>> after some testing I tried to narrow down a problem, which was initially reported by some users.
>> Seen on different distros - debian 7.1, ubuntu 12.04 LTS, IPFire-2.3 as reported by now.
>>
>> All using some flavour of linux-3.2.x kernel.
>>
>> Tried e.g. under Ubuntu an upgrade to "Linux 3.8.0-27-generic x86_64" which solves the problem.
>
> Is that a guest kernel upgrade?

yeah, sorry if that was not clear enough.

>
>> Problem could be triggert with some workload ala:
>>
>> spew -v --raw -P -t -i 3 -b 4k -p random -B 4k 1G /tmp/doof.dat
>> and in parallel do some apt-get install/remove/whatever.
>>
>> That results in a somewhat stuck qemu-session with the bad
>> "kernel_hung_task..." messages.
>>
>> A typical command-line is as follows:
>>
>> /usr/local/qemu-1.6.0/bin/qemu-system-x86_64 -usbdevice tablet -enable-
>> kvm -daemonize -pidfile /var/run/qemu-server/760.pid -monitor
>> unix:/var/run/qemu-server/760.mon,server,nowait -vnc unix:/var/run/qemu-
>> server/760.vnc,password -qmp unix:/var/run/qemu-
>> server/760.qmp,server,nowait -nodefaults -serial none -parallel none
>> -device virtio-net-pci,mac=00:F1:70:00:2F:80,netdev=vlan0d0 -netdev
>> type=tap,id=vlan0d0,ifname=tap760i0d0,script=/etc/fcms/add_if.sh,downscript=/etc/fcms/downscript.sh
>> -name 1155823384-4 -m 512 -vga cirrus -k de -smp sockets=1,cores=1
>> -device virtio-blk-pci,drive=virtio0 -drive
>> format=raw,file=rbd:1155823384/vm-760-disk-1.rbd:rbd_cache=false,cache=writeback,if=none,id=virtio0,media=disk,index=0,aio=native
>> -drive
>> format=raw,file=rbd:1155823384/vm-760-swap-1.rbd:rbd_cache=false,cache=writeback,if=virtio,media=disk,index=1,aio=native
>> -drive if=ide,media=cdrom,id=ide1-cd0,readonly=on -drive
>> if=ide,media=cdrom,id=ide1-cd1,readonly=on -boot order=dc
>>
>> no "system_reset", "sendkey ctrl-alt-delete" or "q" in monitoring-
>> session is accepted, need to hard-kill the process.
>
> Yesterday I saw a possibly related report on IRC. It was a Windows
> guest running under OpenStack with images on Ceph.
>
> They reported that the QEMU process would lock up - ping would not work
> and their management tools showed 0 CPU activity for the guest.
> However, they were able to "kick" the guest by taking a VNC screenshot
> (I think). Then it would come back to life.
>
> If you have a Linux guest that is reporting kernel_hung_task, then it
> could be a similar scenario.
>
> Please confirm that the hung task message is from inside the guest.
>

confirmed.

> If you are able to reproduce this and have an alternative non-Ceph
> storage pool, please try that since Ceph is common to both these bug
> reports.
>

I can reproduce it with: kernel 3.2.something + qemu-1.[456] ( never spent much time on 1.3) and high I/O.
I took this VM later this day and converted it to local-storage-qcow2, no prob with any kernel. I already asked on ceph-users-list for assistance, especially from Josh ( if he's not on summer holiday ;) )

What is strange, I have a session via VNC-console opened and have a loop ala:
...

[Oliver Francke (oliver-francke)]

Hi,

opened a ticket with the ceph-guys, and it turned out to be a bug in "librados aio flush".

With latest "wip-librados-aio-flush (bobtail)" I got no error even with _very_ high load.

Thnx for the attention ;)

Oliver.

[Thomas Huth (th-huth)]

Closing as "Invalid" since this was not a QEMU bug according to comment #3.