svn commit: r277900 - head/sys/x86/x86
John Baldwin
jhb at FreeBSD.org
Fri Jan 30 14:58:32 UTC 2015
On 1/29/15 10:51 PM, Ryan Stone wrote:
> On Thu, Jan 29, 2015 at 3:45 PM, John Baldwin <jhb at freebsd.org> wrote:
>> Also, note that we always use the TSC as the cpu ticker, even if it is not the
>> timecounter. Thus, if C2+ is enabled on the systems, the kernel idle threads
>> will have an artificially low runtime since their idle time is not properly
>> accounted for. This does not affect any other threads as only the idle
>> threads "run" while the CPU is in C2+ with the TSC effectively paused. The
>> new default means that the idle threads will now have the proper runtime,
>> however.
>
> However in a VM environment this means that runtimes for all threads
> can be very wrong (regardless of whether we enter C2 or lower or not).
> I have a coworker looking at using kvmclock in preference to TSC for
> the cpu ticker when the kvmclock is available.
I think that is orthogonal to older Core 2 CPUs which the above mail was
about. :) However, yes, that sounds sensible. You might want to talk
to bryanv at . He has a branch in projects (projects/paravirt I think?)
that provides a framework for hypervisor "drivers" that would be a good
place to plug this into.
--
John Baldwin
More information about the svn-src-head
mailing list