Re: [PATCH v15 23/26] sched: early boot clock

[Salvatore Bonaccorso]

Hi,

On Tue, Nov 06, 2018 at 06:35:36AM -0500, Steven Sistare wrote:
> Pavel has a new email address, cc'd - steve
> 
> On 11/6/2018 12:42 AM, Dominique Martinet wrote:
> > (added various kvm/virtualization lists in Cc as well as qemu as I don't
> > know who's "wrong" here)
> > 
> > Pavel Tatashin wrote on Thu, Jul 19, 2018:
> >> Allow sched_clock() to be used before schec_clock_init() is called.
> >> This provides with a way to get early boot timestamps on machines with
> >> unstable clocks.
> > 
> > This isn't something I understand, but bisect tells me this patch
> > (landed as 857baa87b64 ("sched/clock: Enable sched clock early")) makes
> > a VM running with kvmclock take a step in uptime/printk timer early in
> > boot sequence as illustrated below. The step seems to be related to the
> > amount of time the host was suspended while qemu was running before the
> > reboot.
> > 
> > $ dmesg
> > ...
> > [    0.000000] SMBIOS 2.8 present.
> > [    0.000000] DMI: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20180724_192412-buildhw-07.phx2.fedoraproject.org-1.fc29 04/01/2014
> > [    0.000000] Hypervisor detected: KVM
> > [    0.000000] kvm-clock: Using msrs 4b564d01 and 4b564d00
> > [283120.529821] kvm-clock: cpu 0, msr 321a8001, primary cpu clock
> > [283120.529822] clocksource: kvm-clock: mask: 0xffffffffffffffff max_cycles: 0x1cd42e4dffb, max_idle_ns: 881590591483 ns
> > [283120.529824] tsc: Detected 2592.000 MHz processor
> > ...
> > 
> > (The VM is x86_64 on x86_64, I can provide my .config on request but
> > don't think it's related)
> > 
> > 
> > It's rather annoying for me as I often reboot VMs and rely on the
> > 'uptime' command to check if I did just reboot or not as I have the
> > attention span of a goldfish; I'd rather not have to find something else
> > to check if I did just reboot or not.
> > 
> > Note that if the qemu process is restarted, there is no offset anymore.
> > 
> > I unfortunately just did that so cannot say with confidence (putting my
> > laptop to sleep for 30s only led to a 2s offset and I do not want to
> > wait longer right now), but it looks like the clock is still mostly
> > correct after reboot after disabling my VM's ntp client. Will infirm
> > that tomorrow if I was wrong.
> > 
> > 
> > Happy to try to help fixing this in any way, as written above the quote
> > I'm not even actually sure who is wrong here.

A user in Debian reported the same/similar issue (with 4.19.13):

https://bugs.debian.org/918036

Regards,
Salvatore