clock_getres() and real resolution

clock_getres() and real resolution

[Dmitry Antipov]

IIUC, an idea behind clock_getres() is to give a hint about the resolution of
specified clock. This hint may be used by an application programmer to check whether
this clock is suitable for a some purpose. So why clock_getres() always returns
something like {0, 1} (if hrtimers are enabled) regardless of the underlying platform's
real numbers?

For example, OMAP4's real resolution of CLOCK_REALTIME is 30.5us for 32K timer and 26ns
for MPU timer. Such a difference definitely makes sense - but clock_getres(CLOCK_REALTIME,..)
always returns {0, KTIME_HIGH_RES}. Since this behavior causes a confusion like
http://lists.linaro.org/pipermail/linaro-dev/2012-February/010112.html, I'm considering
this as a stupid misfeature.

Dmitry

Re: clock_getres() and real resolution

[Richard Cochran]

On Wed, Feb 08, 2012 at 08:31:03AM -0800, Dmitry Antipov wrote:
> IIUC, an idea behind clock_getres() is to give a hint about the resolution of
> specified clock. This hint may be used by an application programmer to check whether
> this clock is suitable for a some purpose. So why clock_getres() always returns
> something like {0, 1} (if hrtimers are enabled) regardless of the underlying platform's
> real numbers?

I think "resolution" does not mean tick duration, but rather the finest timer unit.

HTH,
Richard

Re: clock_getres() and real resolution

[Dmitry Antipov]

On 02/08/2012 09:12 PM, Richard Cochran wrote:

> I think "resolution" does not mean tick duration, but rather the finest timer unit.

#include <stdio.h>
#include <time.h>

int main (int argc, char *argv[])
{
   int i;
   struct timespec rs, ts[10];

   clock_getres (CLOCK_REALTIME, &rs);
   printf ("res: %lus %luns\n", rs.tv_sec, rs.tv_nsec);

   for (i = 0; i < 10; i++)
     clock_gettime (CLOCK_REALTIME, ts + i);
   for (i = 0; i < 10; i++)
     printf ("%d: %lus %luns\n", i, ts[i].tv_sec, ts[i].tv_nsec);

   return 0;
}

=>

res: 0s 10000000ns
0: 1328779203s 975317500ns
1: 1328779203s 975317900ns
2: 1328779203s 975318200ns
3: 1328779203s 975318400ns
4: 1328779203s 975318600ns
5: 1328779203s 975318800ns
6: 1328779203s 975319000ns
7: 1328779203s 975319300ns
8: 1328779203s 975319500ns
9: 1328779203s 975319600ns

Old Sun Fire 880, SunOS 5.10 Generic_139555-08.

100ns precision with 10ms "finest timer unit"???

Dmitry

Re: clock_getres() and real resolution

[Richard Cochran]

On Thu, Feb 09, 2012 at 01:25:33AM -0800, Dmitry Antipov wrote:

> Old Sun Fire 880, SunOS 5.10 Generic_139555-08.
> 
> 100ns precision with 10ms "finest timer unit"???

I thought this list was about Linux kernel development, but now it
seems to be about Sun's old bugs.

Richard

Re: clock_getres() and real resolution

[Dmitry Antipov]

On 02/09/2012 10:40 AM, Richard Cochran wrote:

> I thought this list was about Linux kernel development, but now it
> seems to be about Sun's old bugs.

This Sun (probably) has ~100000x more accurate hrtimers than it's said,
and it's a bug. My panda board (with 32K timer enabled) has ~30000x
less accurate hrtimers than it's said, and it's not a bug. Great.

Dmitry

Re: clock_getres() and real resolution

[Richard Cochran]

On Thu, Feb 09, 2012 at 11:32:16AM -0800, Dmitry Antipov wrote:
> On 02/09/2012 10:40 AM, Richard Cochran wrote:
> 
> >I thought this list was about Linux kernel development, but now it
> >seems to be about Sun's old bugs.
> 
> This Sun (probably) has ~100000x more accurate hrtimers than it's said,
> and it's a bug. My panda board (with 32K timer enabled) has ~30000x
> less accurate hrtimers than it's said, and it's not a bug. Great.

If I understand you correctly, what you are looking for is a way to
get a promise from the OS regarding a real time deadlines, right?
But that is a different question than the timer unit resolution:

Q: What is the finest timer duration that I may request?
A: One nanosecond (answer by getres).

Q: What kind of real time deadline will my system provide?
A: Nobody knows for sure.

Just because the OS claims timer resolution X does mean that your
application can assume, "okay, I'll just set my periodic task at
duration X and the OS will take care of the rest." The only thing the
user can count on is that the timer expiration will not come _before_
the deadline. The nanosleep man page puts its like this:

   If the interval specified in req is not an exact multiple of the
   granularity underlying clock (see time(7)), then the interval will
   be rounded up to the next multiple.  Furthermore, after the sleep
   completes, there may still be a delay before the CPU becomes free
   to once again execute the calling thread.

I agree that getres does not provide very useful information. Under
Linux, it merely indicates the present of high resolution timer
support.

The practical solution to what you are asking for is overall system
testing tuning, and this is unfortunately manual labor.

Richard

Re: clock_getres() and real resolution

[Thomas Gleixner]

On Wed, 8 Feb 2012, Dmitry Antipov wrote:

> IIUC, an idea behind clock_getres() is to give a hint about the resolution of
> specified clock. This hint may be used by an application programmer to check
> whether
> this clock is suitable for a some purpose. So why clock_getres() always
> returns
> something like {0, 1} (if hrtimers are enabled) regardless of the underlying
> platform's
> real numbers?
> 
> For example, OMAP4's real resolution of CLOCK_REALTIME is 30.5us for 32K timer
> and 26ns
> for MPU timer. Such a difference definitely makes sense - but
> clock_getres(CLOCK_REALTIME,..)
> always returns {0, KTIME_HIGH_RES}. Since this behavior causes a confusion
> like
> http://lists.linaro.org/pipermail/linaro-dev/2012-February/010112.html, I'm
> considering
> this as a stupid misfeature.

We had this discussion before. The point is that the accuracy of the
internal kernel timer handling is 1nsec in case of high resolution
timers. The fact that the underlying clock event device has a coarser
resolution does not change that. 

It would be possible to return the real resolution of the clock event
device, but we have systems, where the clockevent device is
dynamically changing. So which resolution do we expose to an
application? The one of the current active device or some magic number
of a device which might not even be initialized? That's more confusing
than telling user space that high resolution timers are active and the
kernel is trying to achieve the 1ns accuracy.

Thanks,

	tglx

Re: clock_getres() and real resolution

[Dmitry Antipov]

On 02/09/2012 02:12 AM, Thomas Gleixner wrote:

> It would be possible to return the real resolution of the clock event
> device, but we have systems, where the clockevent device is
> dynamically changing. So which resolution do we expose to an
> application? The one of the current active device or some magic number
> of a device which might not even be initialized? That's more confusing
> than telling user space that high resolution timers are active and the
> kernel is trying to achieve the 1ns accuracy.

First of all, it's not necessary to make unrealizable promises to
an application programmer. If it's known that _any_ hardware configuration
can't guarantee, for example, <20ns precision, it's better to return
{0, 20} than {0, 1} from clock_getres(...). If high-res subsystem
isn't active, just return -1 and set errno to EINVAL, regardless of
an arguments passed.

Second, it's very hard to deny that some applications really needs
precise time measurements. So, if the clockevent device is dynamically
changing, it would be nice to have a method to prevent the loss of
precision for such an application. For example, an application may
issue prctl(PR_SET_CLOCK_STABLE, 1) to make sure that hrtimer's resolution
isn't changed (or at least not changed with the loss of precision)
until prctl(PR_SET_CLOCK_STABLE, 0) or exit(); if some system-wide
event decreases hrtimer accuracy, such an application might receive
a signal, etc.

Dmitry