Re: [RFC][PATCH] linux-2.5.64_monotonic-clock_A1

[john stultz <johnstul@us.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 1635 bytes --]

On Tue, 2003-03-11 at 13:47, george anzinger wrote:
> Some comments below on the scaling.

Thanks, I'll try to digest your comments and get back to you.


> On a related note, I would like to extend the CLOCK_MONOTONIC code to 
> the same res as CLOCK_REALTIME in the POSIX clocks and timers patch. 
> The patch uses jiffies_64 for CLOCK_MONOTONIC, so what I would like to 
> do is use get_offset() to fill in the sub_jiffies part. Is this 
> function available (i.e. timer->get_offset()) on all archs?


Nope, the timer_opts structure is i386 only. Further, the need for the
monotonic_clock() interface is because timer->get_offset() only returns
32bits of information, which on a 2Ghz cpu is only ~2 seconds worth of
time. We need multiple minutes worth of time to be returned, thus the 64
bit return of monotonic_clock. 

I considered making get_offset() return a 64bit value, but worried that
the cost of the 64bit math would hurt gettimeofday too much to be worth
it. So rather then complicate a heavily used function to handle a very
rare case, we decided to implement a new interface that doesn't need to
be as fast as gettimeofday, but can handle long periods of time w/o
interrupts.


> It seems to me that the lost jiffies should be rolled into 
> get_offset().  Have you considered doing this?

I'm not sure I'm following this? get_offset returns the amount of time
since mark_offset() was called(last interrupt). The lost-jiffies
compensation code I added uses get_offset() to detect how many jiffies
should have passed. How do you suggest rolling it into get_offset?

thanks
-john



[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 232 bytes --]