Re: clock stability, TSC, and QEMU

[Ottavio Caruso <ottavio2006-usenet2012%yahoo.com@localhost>]

On 09/05/2021 07:23, nia wrote:
> hello,
>
> when running NetBSD as a QEMU guest, there seem to be problems
> with clock stability.

Interesting. I don't have this problem with NetBSD guest but I had a 
similar one with an OpenBSD guest (see below).

This is my command line (host is Debian Buster):

qemu-system-x86_64 \
-drive if=virtio,file=/home/oc/VM/img/netbsd.image,index=0,media=disk \
-drive if=virtio,file=/home/oc/VM/img/netbsd.image.old,index=1,media=disk \
-M q35,accel=kvm -m 250M -cpu host -smp $(nproc) \
-nic user,hostfwd=tcp:127.0.0.1:5555-:22,model=virtio-net-pci,ipv6=off  \
-daemonize -display none  -vga none \
-serial mon:telnet:127.0.0.1:6665,server,nowait \
-pidfile /home/oc/VM/pid/netbsd-pid -nodefaults



> i remember having this problem when using a commercial KVM-based VM
> host. erlang is very sensitive to monotonic clock drift, and
> will abort if it detects it. my application would infrequently
> abort regardless of kern.timecounter.hardware (although ACPI-Safe
> annecdotally made it less worse) due to the monotonic clock
> stepping backwards.
>
> when running NetBSD as a QEMU guest with NVMM, the following
> message is printed to dmesg on boot:
>
> autoconfiguration ERROR: 9225914709 cycle TSC drift observed
>
> (kern.timecounter.hardware defaults to hpet)

$ dmesg|grep drift
$ grep drift /var/log/messages
$ sysctl kern.timecounter.hardware
kern.timecounter.hardware = hpet0

> QEMU has a -rtc setting, but I'm unsure of best option
> to pick for NetBSD:
>
> -rtc [base=utc|localtime|<datetime>][,clock=host|rt|vm][,driftfix=none|slew]
> 	set the RTC base and clock, enable drift fix for clock ticks (x86 only)
>
> additionally, on my laptop running -curent, I get spammed
> with warnings about TSC drift when waking from suspend:
> http://gnats.netbsd.org/55694
> It is defaulting to using TSC as a timecounter -
> I wonder if NetBSD should be relying on TSC less.

I had a problem with an OpenBSD guest, where the clock would go mental 
after a prolonged period of inactivity and would trigger ddb(4) kernel 
panic. I was suggested to add "-kvmclock-stable-bit" to -"-cpu", so my 
OpenBSD boot is:

qemu-system-x86_64 \
-drive if=virtio,file=/home/oc/VM/img/openbsd.image,index=0,media=disk \
-drive if=virtio,file=/home/oc/VM/img/openbsd.image.old,index=1,media=disk \
-M q35,accel=kvm -m 250M -cpu host,-kvmclock-stable-bit -smp $(nproc) \
-nic user,hostfwd=tcp:127.0.0.1:5556-:22,model=virtio-net-pci \
-daemonize -display none   -vga none \
-serial mon:telnet:127.0.0.1:6666,server,nowait \
-pidfile /home/oc/VM/pid/openbsd-pid -nodefaults

telnet 127.0.0.1 6666

I don't know how all this fits with NVMM, though.



--
Ottavio Caruso