kvm regression in 3.13.x affects host system

kvm regression in 3.13.x affects host system

[Thomas Sattler]

Hi there ...

I observe a regression in 3.13.x and initially reported it to Takuya
Yoshikawa, who told me to post to this list. (I'm neither subscribed
to this list nor LKML.)

I just compiled 3.13.3 and it happens here too. In this very moment.

All I do is running qemu to boot an ISO image like this

    $ qemu-kvm -cdrom systemrescuecd-x86-4.0.0.iso

and the guest eventually stops running causing a high system load
on the host system. Usually the guest hangs within 10-15 seconds,
but I once saw it hanging only after more than 90 seconds and
once even before printing the

    kvm: zapping shadow pages for mmiogeneration wraparound

message in dmesg. It seems there are absolutely no error messages.

Usually I can Ctrl-C the running qemu process and everything is
just fine again. Only in the one case where qemu hang very early
I had to 'kill -9' the process.

In case I do not kill qemu my system becomes unusable within a few
minutes. While I can start new terminals (via window manager hot-
keys) and run for example 'top' and 'vmstat' ... [freeze+reboot]

Thanks to Thunderbirds drafts autosave I can continue wrinting this
mail, I just hung it completely running 'xload' while qemu caused
a load above 6 (on a two core machine). So far I could always reco-
ver by killing qemu, but on the other hand I never tried to start
'xload' in a situation like this before. (Note: There is always a
copy of xload running on my desktop.)

I run a i686 Gentoo Linux with qemu-1.4.2 on Intel(R) Core(TM)2 Duo
CPU E8600 @ 3.33GHz. I tested 3.13.{0,1,2,3} and 3.12.{4,7,8,10}.
All 3.13.x are broken, all 3.12.x worked, even running 'qemu'
for several hours.

$ grep KVM config-3.12.10
CONFIG_HAVE_KVM=y
CONFIG_HAVE_KVM_IRQCHIP=y
CONFIG_HAVE_KVM_IRQ_ROUTING=y
CONFIG_HAVE_KVM_EVENTFD=y
CONFIG_KVM_APIC_ARCHITECTURE=y
CONFIG_KVM_MMIO=y
CONFIG_KVM_ASYNC_PF=y
CONFIG_HAVE_KVM_MSI=y
CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT=y
CONFIG_KVM=m
CONFIG_KVM_INTEL=m
# CONFIG_KVM_AMD is not set

$ grep KVM config-3.13.2
CONFIG_HAVE_KVM=y
CONFIG_HAVE_KVM_IRQCHIP=y
CONFIG_HAVE_KVM_IRQ_ROUTING=y
CONFIG_HAVE_KVM_EVENTFD=y
CONFIG_KVM_APIC_ARCHITECTURE=y
CONFIG_KVM_MMIO=y
CONFIG_KVM_ASYNC_PF=y
CONFIG_HAVE_KVM_MSI=y
CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT=y
CONFIG_KVM_VFIO=y
CONFIG_KVM=m
CONFIG_KVM_INTEL=m
# CONFIG_KVM_AMD is not set

"/usr/bin/qemu-kvm" (on gentoo) is a wrapper script containing

    #!/bin/sh

    exec /usr/bin/qemu-system-x86_64 -machine accel=kvm "$@"

I tried using /usr/bin/qemu-system-i386 (as I'm running i686) but the
same happend. (I must admit that I tested this only once, with one of
the 3.13.x kernels.)

Is there anything else I can to?

Thomas

Re: kvm regression in 3.13.x affects host system

[Paolo Bonzini]

Il 14/02/2014 13:03, Thomas Sattler ha scritto:
>
> I tried using /usr/bin/qemu-system-i386 (as I'm running i686) but the
> same happend. (I must admit that I tested this only once, with one of
> the 3.13.x kernels.)
>
> Is there anything else I can to?

Can you capture a trace of the hang?

You can follow the procedure in http://www.linux-kvm.org/page/Tracing 
(except that probably you want "-e kvm -e kvmmmu", not just "-e kvm").

Paolo