Re: [PATCH] Deferrable Timer

[Dave Winchell <dwinchell@virtualiron.com>]

Ke,

Yu, Ke wrote:

>Dave,
>
>Glad to see there is deferrable timer application. Please go ahead with
>that. And I will keep you updated if there is finding in my side.
>
ok.

> 
>
>BTW, Could you please elaborate more on the "guest-handles-missed-tick"
>case? Since there is no need to inject missed tick to guest, which timer
>would be used as deferrable timer?
>  
>
Hpet.c uses set_timer for hpet comparator/timer 0. When that timer expires,
a clock interrupt may be injected to the guest. This timer
is normally set to expire at the next period boundary.
We could, instead, have it expire over a range of say, several periods.

Vpt.c works in a similar fashion for its periodic timer. Other clocksources,
e.g. pit, rtc, are layred on vpt.c with interface create_periodic_timer.
I can imagine an option passed to create_periodic_timer signifying that
a deferrable timer may be used.

Ideally, the deferrable timer would have an option where a set of allowable
timeout values, rather than a range, could be provided. If it had this 
option, we could keep
the timeouts on the integer*period time line. Otherwise I need to warp 
the comparator
as discussed below. I anticipate that there may be some problems with
warping.

I realize that specifying a range gives you more options for combining 
timeouts.
I don't mind trying to solve the warping problem.

One further option would be a deferrable timer with a range fallowed by 
a non-deferrable
timer to get back on the integer*period timeline for interrupt delivery.

thanks,
Dave


>Best Regards
>Ke 
>
>Dave Winchell wrote:
>  
>
>>Ke,
>>
>>One would think that hpet or vpt support for the
>>guest-handles-missed-ticks policy would be a good application for a
>>deferrable timer. 
>>If a deferrable timer were used, then the comparator (cmp) would have
>>to 
>>be warped to a non-integer multiple of the period. This is because
>>Linux reads the comparator register to estimate the delay since the
>>interrupt 
>>was posted.
>>I don't think warping like this will be a problem. At some point, I
>>can test this.
>>
>>I think we could use the deferrable timer for the
>>guest-does-not-handle-missed-ticks
>>policy as well.
>>
>>Any investigation that you want to do in the platform timer area would
>>be fine.
>>Or I can do it, but that will probably be after I do the vpt.c/hpet.c
>>integration
>>work.
>>
>>thanks,
>>Dave
>>
>>    
>>
>>>Best Regards
>>>Ke
>>>
>>>_______________________________________________
>>>Xen-devel mailing list
>>>Xen-devel@lists.xensource.com
>>>http://lists.xensource.com/xen-devel
>>>      
>>>
>
>  
>