Beth Kon wrote:
On Mon, 2008-10-27 at 12:49 +0200, Dor Laor wrote:
Beth Kon wrote:
On Fri, 2008-10-17 at 16:49 +0100, Jamie Lokier wrote:
Beth Kon wrote:
Clock drift on Linux is in the range of .017% - .019%, loaded and unloaded. I
haven't found a straightforward way to test on Windows and would appreciate
any pointers to existing approaches.
Is there any reason why there should be any clock drift, when the
guest is using a non-PIT clock?
I'm probably being naive, but with 32-bit or 64-bit HPET counters
available to the guest, and accurate values from the CMOS clock
emulation, I don't see why drift would accumulate over the long term
relative to the host clock.
I was measuring with ntpdate, so the drift is with respect to the ntp
server pool, not the host clock. But in any case, since timer interrupts
and reads of the hpet counter are at the mercy of the host scheduler
(i.e., the qemu process can be swapped out at any time during hpet read
or timer expiration), I'd guess there would always be some amount of
inaccuracy. Also, qemu checks for timer expiration (qemu_run_timers) as
part of a bigger loop (main_loop_wait), so the varying amounts of work
to do elsewhere in the loop from iteration to iteration would also
introduce irregular delays.
This is exactly why hpet as the other clock emulation in qemu (pit,
rtc, pm?) need
to check whether their irq was really injected. Gleb sent patches for
the rtc, pit.
The idea is to check with the irq chip if the injected irq was really
successful.
I assume these are the patches you're referring to?
http://thread.gmane.org/gmane.comp.emulators.kvm.devel/18974/focus=18977
Yap, Gleb just resent it.
Looks like they were never merged. Does anyone know the history on that?
Also, HPET generates edge-triggered interrupts (as dictated by Linux and
Windows) so I'm not sure if this scheme could work for it.
Right, I think if this time drift fix approach is accepted, it should
also be implemented
for qemu_irq_pulse too.
Dor
|