Re: [PATCH 0/2] eventfd: new EFD_STATE flag

[Gleb Natapov <gleb@redhat.com>]

On Wed, Aug 26, 2009 at 01:04:09PM -0700, Davide Libenzi wrote:
> On Wed, 26 Aug 2009, Gleb Natapov wrote:
> 
> > On Wed, Aug 26, 2009 at 10:42:05PM +0300, Avi Kivity wrote:
> > > On 08/26/2009 10:13 PM, Davide Libenzi wrote:
> > > >Ok, so why not using the eventfd counter as state?
> > > >On the device side:
> > > >
> > > >void write_state(int sfd, int state) {
> > > >	u64 cnt;
> > > >
> > > >	/* Clear the current state, sfd is in non-blocking mode */
> > > >	read(sfd,&cnt, sizeof(cnt));
> > > >	/* Writes new state */
> > > >	cnt = 1 + !!state;
> > > >	write(sfd,&cnt, sizeof(cnt));
> > > >}
> > > >
> > > >
> > > >On the hypervisor side:
> > > >
> > > >int read_state(int sfd) {
> > > >	u64 cnt;
> > > >
> > > >	read(sfd,&cnt, sizeof(cnt));
> > > >	return state - 1;
> > > >}
> > > >
> > > 
> > > Hadn't though of read+write as set.  While the 1+ is a little ugly,
> > > it's workable.
> > > 
> > It's two system calls instead of one to inject interrupt.
> 
> I guess that's going to completely throw off-chart your RT performance, 
> doesn't it?
> 
Do you consider interrupt injection path not worth of optimizing?

--
			Gleb.