Re: Help with the high res timers

[Chris Friesen <cfriesen@nortel.com>]

Nishanth Aravamudan wrote:

> If I understand your point correctly, I think this is achieved by
> TIMERINTERVAL_BITS in my patch (not to claim my patch is function, but
> conceptually). No matter what you actually request, the best you can do
> is 2^TIMERINTERVAL_BITS nanoseconds, and usually worse because the
> tick-rate and timerinterval length do not necessarily line up.

My point is simply that the timer for the next interval should start at 
the time the timer expires, not the time that userspace picks up the 
prior expiration.  Throttling the timer rate should be done at the time 
of timer request rather than timer expiry.

If I have usec-accuracy in the timer subsystem, I should be able to set 
a timer with an interval of 9.999ms and have it remain accurate over 
time (subject to scheduler jitter, of course).  N timer intervals later 
my timer should expire at (original_time + N*9.999ms + jitter).  In this 
case the error is roughly constant with time.

If the timer doesn't start counting the next interval until the user 
detects expiry, I'm going to get some non-zero addition to *each* 
interval such that my timers will not remain accurate over long periods 
of time.  In this case N timer intervals later my timer will expire at 
(original_time + N*(9.999ms + jitter)) which is a very different thing. 
  Since jitter will always be positive, the error increases with time.

Chris