Re: FreeBSD timing issues and qemu (was: Re: [Qemu-devel] Re: Breakage with local APIC routing)

[John Baldwin]

On Friday 11 September 2009 1:03:17 pm Luigi Rizzo wrote:
> On Fri, Sep 11, 2009 at 11:22:59AM -0400, John Baldwin wrote:
> > On Thursday 10 September 2009 3:08:00 pm Luigi Rizzo wrote:
> ...
> > > > Index: sys/kern/kern_timeout.c
> > > > @@ -323,7 +323,7 @@ softclock(void *arg)
> > > >         steps = 0;
> > > >         cc = (struct callout_cpu *)arg;
> > > >         CC_LOCK(cc);
> > > > -       while (cc->cc_softticks != ticks) {
> > > > +       while (cc->cc_softticks-1 != ticks) {
> > > >                 /*
> > > >                  * cc_softticks may be modified by hard clock, so cache
> > > >                  * it while we work on a given bucket.
> > > > 
> > > 
> > > as mentioned in the followup message in that thread,
> > > you also need this change in callout_tick()
> > > 
> > >         mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
> > >      -  for (; (cc->cc_softticks - ticks) < 0; cc->cc_softticks++) {
> > >      +  for (; (cc->cc_softticks - ticks) <= 0; cc->cc_softticks++) {
> > >                 bucket = cc->cc_softticks & callwheelmask;
> > 
> > I would fix the style in the first hunk (spaces around '-') but I think 
you 
> > should commit this and get it into 8.0.  I think a per-CPU ticks might 
prove 
> > very problematic as 'ticks' is rather widely used (though I would find 
that 
> > cleaner perhaps).
> 
> i will ask permission to re -- i was hoping to get some feedback
> on the thread on -current but no response so far :(
> 
> Note that the per-cpu ticks i was proposing were only visible to the
> timing wheels, which don't use absolute timeouts anyways.
> So i think the mechanism would be quite safe: right now, when you
> request a callout after x ticks, the code first picks a CPU
> (with some criteria), then puts the request in the timer wheel for
> that CPU using (now) the global 'ticks'. Replacing ticks with cc->cc_ticks,
> would completely remove the races in insertion and removal.
> 
> I actually find the per-cpu ticks even less intrusive than this change.

Well, it depends.  If TCP ever started using per-CPU callouts (i.e. 
callout_reset_on()) it would probably need to start using the per-CPU ticks 
instead of the global ticks, etc.  You could have 'ticks' just be == to CPU 
0's ticks perhaps.

-- 
John Baldwin