[Qemu-devel] Re: KVM call agenda for Feb 9

[Anthony Liguori <anthony@codemonkey.ws>]

On 02/09/2010 12:56 AM, Avi Kivity wrote:
> On 02/09/2010 03:28 AM, Chris Wright wrote:
>> Please send in any agenda items you are interested in covering.
>
> hpet overhead on large smp guests
>
> I measured hpet consuming about a half a core's worth of cpu on an 
> idle Windows 2008 R2 64-way guest.  This is mostly due to futex 
> contention, likely from the qemu mutex.
>
> Options:
> - ignore, this is about 1% of the entire system (but overhead might 
> increase greatly if a workload triggers more hpet accesses)
> - push hpet into kernel, with virtio-net, virtio-blk, and kernel-hpet, 
> there's little reason to exit into qemu

Security, shamurity, let's just stick all of qemu in the kernel :-)

> - rcuify/fine-grain qemu locks

Should be pretty straight forward.  It would start with removing the 
locking within kvm*.c such that qemu_mutex isn't acquired until we 
dispatch I/O operations.  Then we can add lockless paths for dispatch as 
we convert device models over.

Regards,

Anthony Liguori