Re: [XenPPC] Next profiling clarification question

[Jimi Xenidis <jimix@xxxxxxxxxxxxxx>]

On Mar 20, 2007, at 11:36 AM, Hollis Blanchard wrote:

> On Mon, 2007-03-19 at 15:30 +0100, Christian Ehrhardt wrote:
> > Hi,
> > I have one more questions regarding the oprofile extension to work in
> > xenppc guests.
> >
> > The plain linux implementation sets !always! the vector for the
> > performance interrupt 0xf00 to the function
> > "performance_monitor_exception" in head32.S/head64.S. Depending on
> > wether linux starts profiling this function is set to a dummy or the
> > "real" handler of e.g. oprofile.
> > Now in the virtualized environment each guest set's its own interrupt
> > handler for 0xf00 (Performance monitor interrupt).
Yes, each guest decided to set it to a dummy or a real handler.

> > Because we are
> > running with LPSE[1]=0 as Jimi wrote before the performance interrupt
> > while a Dom is running should be handled from the domain because  
> > nothing
> > sets MSR[HV] to 1.
> OK, so performance monitor interrupts go directly to the guest.
>
> > My question is now if the address translation & virtualization  
> > ensures
> > that each perf interrupt will be a) served by the right function  
> > adress
> > and b) in the domain. Maybe there is some exclusive part involved and
> > the different writes of the 0xf00 vector may overwrite each other or
> > something similar - that's why I want to ask to be sure.
On this architecure, all interrupts turn "translate off", MSR[IR|DR] 
=0b00.  The hypervisor design provided by using LPES[0]=1 allows us  
to create an area that the domain can consider "translate off" we  
call this the Real Mode Area (RMA) and each RMA must be unique to  
every guest.  So each guest is free to install its own unique  
interrupt handlers, since any interrupts that do not modify MSR[HV]  
will cause the the CPU to execute at the same effective address  
offset into the RMA.

> > Example of a system with two DomU guests that enabled the performance
> > counters.
> > Running:
> > Xen
> > 0xf00 -> undef?
> I'm not sure what to do about a perfmon interrupt in Xen. Since we're
> saving/restoring the perfmon register state on domain entry/exit,  
> can we
> just disable those interrupts while running Xen?
Yes with MMCR0[HFC]

> > Dom0
> > 0xf00 -> default_pmc_irq
> > DomU1
> > 0xf00 ->    pmc_irq_handler1
> > DomU2
> > 0xf00 ->    pmc_irq_handler2
> I don't understand the question about translation. The domains do not
> share code. For example, it is not possible for an interrupt in  
> Dom1 to
> trigger an exception handler in Dom0.
I think the Christian believes all domains share the same interrupt  
vectors, I hope the RMA discussion has clarified that.
-JX



_______________________________________________
Xen-ppc-devel mailing list
Xen-ppc-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ppc-devel