Re: [RFC PATCH] KVM: x86: Skip request checking branches in vcpu_enter_guest() more effectively

[Avi Kivity <avi@redhat.com>]

On 09/24/2012 04:32 PM, Takuya Yoshikawa wrote:
> On Mon, 24 Sep 2012 12:09:00 +0200
> Avi Kivity <avi@redhat.com> wrote:
> 
>> > while (vcpu->request) {
>> > 	xchg(vcpu->request, request);
>> > 
>> > 	for_each_set_bit(request) {
>> > 		clear_bit(X);
>> > 
>> > 		......
>> > 	}
>> > 	
>> > }
>> 
>> In fact I had something like that in one of the earlier versions, but it
>> was problematic.  I'll try to see what the issue was.
> 
> Unless there are many requests, the cost of for_each_set_bit() and a few
> added code may exceed that of the original code.
> (Looping using __ffs() is an alternative because requests is "long".)
> 
> So I wanted to know the most common requests pattern.

See other branch of this thread.  But in short, I now think you are
right and the special-case is warranted.

(not for STEAL_UPDATE - that's likely a bug, it should only happen on
overcommit)


-- 
error compiling committee.c: too many arguments to function