On Thu, Dec 07, 2017 at 11:52:46AM +0100, Martin Pieuchot wrote:
> On 07/12/17(Thu) 08:34, Landry Breuil wrote:
> > Hi,
> >
> > i've been having kvm VMs running 6.2 hardlocking/deadlocking since a
> > while, all those running on proxmox 5.1 using linux 4.13.8 & qemu-kvm
> > 2.9.1. There were hardlocks upon reboot which were 'solved' by disabling
> > x2apic emulation in kvm (args: -cpu=kvm64,-x2apic) or giving the host
> > cpu flags to the vm (args: -cpu host) but there still remains deadlocks
> > during normal operation.
> >
> > I'm now running a kernel with MP_LOCKDEBUG, so i'm collecting traces in
> > the vain hope that it might help someone interested in locking issues.
> > Here's the latest one:
>
> Let me add that when you had x2apic enabled the kernel 'froze' inside
> x2apic_readreg, trace below:
>
> ddb{0}> tr
> x2apic_readreg(10) at x2apic_readreg+0xf
> lapic_delay(ffff800022136900) at lapic_delay+0x5c
> rtcput(ffff800022136960) at rtcput+0x65
> resettodr() at resettodr+0x1d6
> perform_resettodr(ffffffff81769b29) at perform_resettodr+0x9
> taskq_thread(0) at taskq_thread+0x67
> end trace frame: 0x0, count: -6
>
> What you're seeing with a MP_LOCKDEBUG kernel is just a symptom. A CPU
> enters DDB because another one is 'frozen' while holding the
> KERNEL_LOCK(). What's interesting is that in both case the frozen CPU
> is trying to execute apic related code:
> - x2apic_readreg
> - lapic_delay
>
> I believe this issue should be reported to KVM developers as well.
*very* interestingly, i had a new lock, running bsd.sp.. So i think that
rules out openbsd mp.
ddb> tr
i82489_readreg(0) at i82489_readreg+0xd
lapic_delay(ffffffff81a84090) at lapic_delay+0x5c
rtcget(ffffffff81a84090) at rtcget+0x1a
resettodr() at resettodr+0x3a
perform_resettodr(ffffffff81659e99) at perform_resettodr+0x9
taskq_thread(0) at taskq_thread+0x57
end trace frame: 0x0, count: -6
Landry
