lkml.org 
[lkml]   [2006]   [May]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: sched_clock() uses are broken
Nicolas Pitre wrote:
> On Tue, 2 May 2006, Russell King wrote:
>
>
>>On Tue, May 02, 2006 at 06:43:45PM +0200, Andi Kleen wrote:
>>
>>>Russell King <rmk+lkml@arm.linux.org.uk> writes:
>>>
>>>>However, this is not the case. On x86 with TSC, it returns a 54 bit
>>>>number. This means that when t1 < t0, time_passed_ns becomes a very
>>>>large number which no longer represents the amount of time.
>>>
>>>Good point. For a 1Ghz system this would happen every ~0.57 years.
>>>
>>>The problem is there is AFAIK no non destructive[1] way to find out how
>>>many bits the TSC has
>>>
>>>Destructive would be to overwrite it with -1 and see how many stick.
>>>
>>>
>>>>All uses in kernel/sched.c seem to be aflicted by this problem.
>>>>
>>>>There are several solutions to this - the most obvious being that we
>>>>need a function which returns the nanosecond difference between two
>>>>sched_clock() return values, and this function needs to know how to
>>>>handle the case where sched_clock() has wrapped.
>>>
>>>Ok it can be done with a simple test.
>
>
> Better yet the sched_clock() implementation just needs to return a value
> shifted left so the wrap around always happens on 64 bits and the
> difference between two consecutive samples is always right.
>
>
>>>>IOW:
>>>>
>>>> t0 = sched_clock();
>>>> /* do something */
>>>> t1 = sched_clock();
>>>>
>>>> time_passed = sched_clock_diff(t1, t0);
>>>>
>>>>Comments?
>>>
>>>Agreed it's a problem, but probably a small one. At worst you'll get
>>>a small scheduling hickup every half year, which should be hardly
>>>that big an issue.
>
>
> ... on x86 that is.
>
>
>>>Might chose to just ignore it with a big fat comment?
>>
>>You're right assuming you have a 64-bit TSC, but ARM has at best a
>>32-bit cycle counter which rolls over about every 179 seconds - with
>>gives a range of values from sched_clock from 0 to 178956970625 or
>>0x29AAAAAA81.
>>
>>That's rather more of a problem than having it happen every 208 days.
>
>
> Yet that counter isn't necessarily nanosecond based. So rescaling the
> returned value to nanosecs requires expensive divisions which could be
> done only once within sched_clock_diff() instead of twice as often in
> each sched_clock() calls.

Oh phooey!! Scaling can be done with a mpy and a shift. See the new clock
code where the TSC (or what ever) is scaled to ns.


--
George Anzinger george@wildturkeyranch.net
HRT (High-res-timers): http://sourceforge.net/projects/high-res-timers/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-05-04 05:55    [W:0.110 / U:0.252 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site