[Qemu-devel] Re: [PATCH 0/5] Refactor and enhance RTC configuration

[Jan Kiszka]

Dor Laor wrote:
> On 09/11/2009 11:54 AM, Jan Kiszka wrote:
>> Jamie Lokier wrote:
>>> Anthony Liguori wrote:
>>>>> Besides the interface thing, I'm also interesting in comments on the
>>>>> other core idea, the selectable RTC base clock. Do we want this
>>>>> knob? Do
>>>>> we want host_clock unconditionally? Or should the other RTC that
>>>>> currently use the host time already also gain vm_clock support over
>>>>> the
>>>>> time?
>>>>>
>>>> Hard to say.  Doesn't the rtc keep track of wallclock time even on
>>>> power
>>>> off?  I think using host_clock unconditionally does actually make
>>>> sense.
>>>
>>> Sometimes it's useful to offset the emulated clock for one reason or
>>> another, hence the -startdate options.  But having it run at the
>>> correct speed is usually useful :-)
>>
>> Indeed.
>>
>>>
>>> Also, sometimes (due to licenses with wallclock limits) it's useful
>>> for a guest to not see much time pass when the guest is powered off,
>>> although it still needs to be positive.
>>
>> I'm not sure if this is a common use case. And it currently only seems
>> to be support by very few RTCs, the MC146818 being the most prominent
>> one.
>>
>> I'm now a fan of converting the latter to the common scheme of using the
>> host's system time (here via host_clock) and watch out for the need of
>> adding -rtc clock=vm.
> 
> I'm in favor of sticking to clock=vm as a default.
> Most chances that the guest will have internet connect and the host (ala
> rom hypervisor) won't have.
> Furthermore, if the guest and the host are running ntp it might cause
> spurious ntp updates by the guest. Also on migration, you need to make
> sure that both src and dst are synchronized. When using clock=vm there
> is no such need.

clock=vm means that the RTC has no use as a reliable clock source, you
always need additional help by NTP etc. in the guest -- unless you don't
care about accurate time, of course.

Note that, if you don't trust the virtual RTC (e.g. because it's driven
by an isolated hypervisor) and you have NTP at hand, you typically
configure the guest to update the RTC according to NTP. Then migration
between two potentially unsynchronized hosts is also a none-issue.

> 
> Is the only case that clock=host is preferred is when the guest does not
> run ntp while the host does?

It is currently a must-have if you want to synchronize host and guest
clock (independent of the timezone) while NTP-like services are not an
option. This includes the case where none of both have NTP access.

Except for Jamie's case of extending some license runtime, I really see
no real use for RTC based on vm_clock anymore. There is some reason why
other RTCs are already based on the host clock.

Jan

signature.asc
Description: OpenPGP digital signature