Re: [Qemu-devel] [RFC][PATCH 01/12] coroutine: Add gtk-vnc coroutines library

[Anthony Liguori]

On 01/26/2011 10:13 AM, Avi Kivity wrote:
> On 01/26/2011 06:00 PM, Anthony Liguori wrote:
> > On 01/26/2011 09:25 AM, Avi Kivity wrote:
> > > On 01/22/2011 11:29 AM, Stefan Hajnoczi wrote:
> > > > Asynchronous image format code is becoming very complex.  Let's try
> > > > using coroutines to write sequential code without callbacks but use
> > > > coroutines to switch stacks under the hood.
> > > >
> > > >
> > > > +
> > > > +int cc_swap(struct continuation *from, struct continuation *to)
> > > > +{
> > > > +    to->exited = 0;
> > > > +    if (getcontext(&to->last) == -1)
> > > > +        return -1;
> > > > +    else if (to->exited == 0)
> > > > +        to->exited = 1;
> > > > +    else if (to->exited == 1)
> > > > +        return 1;
> > > > +
> > > > +    return swapcontext(&from->uc,&to->uc);
> > > > +}
> > >
> > > swapcontext() is very slow, involving the fpu and a syscall.
> > >
> > > A nice trick I've used in the past is to use getcontext/makecontext 
> > > for the initial setup and setjmp/longjmp for switching.  Of course 
> > > this can be done later, as an optimization.
> >
> > Yeah, there are further optimizations that can be had but really, 
> > it's not important here.  In fact, I'd rather start with the threaded 
> > version just because it's far more portable than ucontext.
>
> What do you mean by threaded version?

Stefan didn't post it, but the original code also has a GThread based 
implementation when ucontext isn't available (like on Windows).  It uses 
a mutex to control the execution of the coroutines.

Regards,

Anthony Liguori