kvm unresponsive under io work

kvm unresponsive under io work

[Thomas Mueller]

hi

scenario:
- host: 2.6.26, kvm-79, Intel Xeon E5420, Disk is a hardware RAID-10 (4 
wd velociraptor's)
- host: /tmp is lvm volume with XFS mounted with option nobarriers
- host: kvm-img create -f qcow2 /tmp/10gb 10G
- host: kvm -drive fiel=/tmp/10gb,if=scsi,boot=on \
              -monitor socktet:/var/run/kvm.socket \
              -smp 1 -vnc :1
- guest: sudo -u nobody dd if=/dev/zero of=/tmp/fillit

while dd'ing in guest somtimes i can't connect to the monitor-socket. And 
sometimes on the guest vnc-display there are the following messages 
printed:


[ 1138.923666] sd 0:0:0:0: ABORT operation failed.
[ 1138.924186] sd 0:0:0:0: [sda] DEVICE RESET operation started
(messages above repeated a couple of times)
[ 1138.924845]  target0:0:0: control msgout: c.
[ 1138.925561]  target0:0:0: has been reset
[ 1138.926117] sd 0:0:0:0: DEVICE RESET operation complete.
[ 1138.926703] sd 0:0:0:0: M_REJECT received (0:0).

... and sometimes there are also kernel panics.  in case of kernel panic 
i see this on the host:

[1105705.188932] vcpu not ready for apic_round_robin
(repated many times)

there are fewer problems if:
- i do dd a second time (because the image space is already allocated?)
- if i give more io time on the host (aka ionice -c2 -n0 -p<pid>)
- i use if=ide , but throughput is really slow (1g data: guest=3.3MB/s 
host=316MB/s)

this behaviour i've also seen on 2 other hosts (1 intel, 1 amd, mostly on 
rsync'ing things)

what can i do to get reasonable throughput but no kernel warnings/panics?

- Thomas