Re: [PATCH 2/3] kvm hypervisor : Add hypercalls to support pv-ticketlock

[Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>]

On Fri, Jan 21, 2011 at 09:48:29AM -0500, Rik van Riel wrote:
> >>Why?  If a VCPU can't make progress because its waiting for some
> >>resource, then why not schedule something else instead?
> >
> >In the process, "something else" can get more share of cpu resource than its
> >entitled to and that's where I was bit concerned. I guess one could
> >employ hard-limits to cap "something else's" bandwidth where it is of real
> >concern (like clouds).
> 
> I'd like to think I fixed those things in my yield_task_fair +
> yield_to + kvm_vcpu_on_spin patch series from yesterday.

Speaking of the spinlock-in-virtualized-environment problem as whole, IMHO
I don't think that kvm_vcpu_on_spin + yield changes will provide the best
results, especially where ticketlocks are involved and they are paravirtualized 
in a manner being discussed in this thread. An important focus of pv-ticketlocks
is to reduce the lock _acquisition_ time by ensuring that the next-in-line 
vcpu gets to run asap when a ticket lock is released. With the way 
kvm_vcpu_on_spin+yield_to is implemented, I don't see how we can provide the 
best lock acquisition times for threads. It would be nice though to compare 
the two approaches (kvm_vcpu_on_spin optimization and the pv-ticketlock scheme) 
to get some real-world numbers. I unfortunately don't have access to a PLE
capable hardware which is required to test your kvm_vcpu_on_spin changes?

Also it may be possible for the pv-ticketlocks to track owning vcpu and make use
of a yield-to interface as further optimization to avoid the 
"others-get-more-time" problem, but Peterz rightly pointed that PI would be a 
better solution there than yield-to. So overall IMO kvm_vcpu_on_spin+yield_to
could be the best solution for unmodified guests, while paravirtualized
ticketlocks + some sort of PI would be a better solution where we have the
luxury of modifying guest sources!

- vatsa