On Fri, 2007-03-09 at 19:38 +0000, Keir Fraser wrote:
> I've been reworking the guest timer support in Xen to make it more flexible.
> As a result of this the 100Hz VIRQ_TIMER event has gone away by default
> except for x86 PV guests (which I know will need it for backward compat, and
> can explicitly disable it if they don't need it). I'm not sure whether ia64
> or powerpc guests also require this event -- if so let me know and I'll add
> the necessary line to your vcpu_initialise() functions.
I guess your HVM domain APIC emulation creates a Xen timer when the
guest tries to set up an interrupt? Currently all PowerPC guests are
paravirtualized and create timer interrupts without Xen involvement
(either trap or hcall), so we would be unable to call Xen's set_timer()
as the guests expect.
Strictly speaking I don't think it needs the periodic timer: PowerPC
Linux uses the timer interrupt as a trigger to examine an independent
time source (the timebase) to see how many ticks to advance. However, on
a heavily-loaded Xen system where a guest vcpu doesn't get much time,
when it finally is scheduled I guess it would spend a lot of time
replaying the ticks it missed.
So in conclusion, I think we'll need the legacy behavior, though it
might be interesting for us in the future to modify Linux to use hcalls
to create timer events.
IBM Linux Technology Center
Xen-ppc-devel mailing list