Re: [Qemu-devel] Re: [RFC: 0/2] patch for QEMU HPET periodic timer emulation to alleviate time drift

[Anthony Liguori]

On 02/07/2011 08:43 AM, Jan Kiszka wrote:
> On 2011-02-07 15:28, Anthony Liguori wrote:
>    
> > On 02/07/2011 08:10 AM, Jan Kiszka wrote:
> >      
> > > Again: please not in an ad-hoc fashion but as a generic services usable
> > > by _all_ periodic timer sources that want to implement compensation.
> > > This infrastructure should also be designed to once integrate IRQ
> > > coalescing information as well.
> > >
> > > The point why I'm insisting on a broader solution is that both sources
> > > for lost ticks (iothread and vcpu) end up in the same output: an
> > > adjustment of the injection frequency of the affected timer device.
> > > There is not "HPET" or "RTC" or "PIT" in this, all this may apply to the
> > > SoC timer of some emulated ARM board as well.
> > >
> > >        
> > Fair enough, how about:
> >
> > typedef struct PeriodicTimer PeriodicTimer;
> >
> > /**
> >    * @accumulated_ticks:  the number of unacknowledged ticks in total
> > since the creation of the timer
> >    **/
> > typedef void (PeriodicTimer)(void *opaque, int accumulated_ticks);
> >      
> I guess you mean PeriodicTimerFunc.

Yes.

>   Why the accumulated_ticks argument?
>    

Then the missing ticks is stored in the PeriodicTimer instead of storing 
it in the device state.  That means we won't forget to save it in vmstate.

It's convenient because then if we lose ticks in the PeriodicTimer 
layer, the devices have instance access to that info.  When you do a 
read() from timerfd, it returns the number of coalesced events.  That's 
the interface I had in my mind.

We could just add a getter for PeriodicTimer and it would serve the same 
purpose.

Regards,

Anthony Liguori

> > PeriodicTimer *periodic_timer_new(PeriodicTimerFunc *cb, void *opaque);
> >
> > void periodic_timer_mod(PeriodicTimer *timer, int64_t interval, TimeUnit
> > unit);
> >
> > /**
> >    * @policy: the drift catch-up policy
> >    *                DRIFT_COMP_FAST, deliver next tick as soon as any
> > tick is acknowledged if accumulated_ticks>  1
> >    *                DRIFT_COMP_NONE, do not change interval regardless of
> > accumulated ticks
> >    *                DRIFT_COMP_GRADUAL, shorten interval by half until
> > accumulated_ticks<= 1
> >    */
> > void periodic_timer_set_policy(PeriodicTimer *timer,
> > DriftCompensationPolicy policy);
> >
> > /**
> >    * @ticks: number of ticks to acknowledge that are currently outstanding.
> >    **/
> > void periodic_timer_ack(PeriodicTimer *timer, int ticks);
> >
> >      
> Looks reasonable otherwise.
>
> Jan
>
>