RE: probe hrtimer testing

RE: probe hrtimer testing

[Stone, Joshua I]

On Wednesday, May 24, 2006 12:05 PM, William Cohen wrote:
> In the 2.6.17-rc4 kernel it looks like hrtimer is used by
> kernel/itimer.c. Is there a particular benefit explicitly using the
> hrtimer rather than going through the normal timer mechanism? Are
> there other modules that could use the hrtimer directly?

My understanding is that itimers are intended for user-mode processes.
They operate by sending a SIGALRM on expiration, which I don't think we
can use.

The advantage of hrtimers is just the potential for better resolution.
The normal timer mechanism is specified in terms of jiffies.  Currently
hrtimers are also jiffy limited, but since they are specified in
nanoseconds, there's the potential that future arch-specific
implementations will provide accuracy greater than the jiffy rate.

Whether sub-jiffy timers are useful for SystemTap is open for
discussion...


Josh

RE: probe hrtimer testing

[Roland McGrath]

Just to reiterate, itimer.c (and posix-timers.c) implement the user syscall
ABIs for timers, and the only things they can be used for are affecting the
user-visible ABI components.  Those are implemented in terms of the
kernel's internal timer facilities.  Anything in private kernel code such
as systemtap probes/runtime must be done using those facilities, and they
are easy to use for it.  In the latest kernels, the hrtimer interface is
the preferred internal timer facility and is now the canonical thing to use
in new kernel code (regardless of what granularity you are going to ask for).

RE: probe hrtimer testing

[Stone, Joshua I]

On Tuesday, May 23, 2006 10:57 AM, William Cohen wrote:
> Looking over the test coverage results recently posted, one of the
> places that is lacking in testing is the probes for hrtimer. Josh,
> does anyone have some tests for that area of the translator? It would
> do a lot to improve the test coverage of tapsets.cxx.

The hrtimer is not tested because it's not complete.  I ran into a
problem with needing exported symbols from the kernel, and LKML rejected
my request.  So until support is added to the kernel ABI for using
hrtimers, the code I added will be unusable.

I checked it in as part of normal development, in hopes that the ABI
would get resolved.  Since hrtimers are still not exported, perhaps I
should remove the code.  Would it be sufficient for your coverage tests
if I just blocked out the code with "#ifdef ENABLE_HRTIMER"?

If anyone has an interest in seeing the hrtimers work in SystemTap,
please speak up.  If there's enough interest, then maybe we can go back
to LKML to try to get it enabled in the ABI.


Josh

Re: probe hrtimer testing

[Frank Ch. Eigler]

"Stone, Joshua I" <joshua.i.stone@intel.com> writes:

> [...] If anyone has an interest in seeing the hrtimers work in
> SystemTap, please speak up.  If there's enough interest, then maybe
> we can go back to LKML to try to get it enabled in the ABI.

It's probably worth one more try.  In the mean time, you might add
some transok or semok tests to trigger the code in the translator
even without kernel support.

- FChE

Re: probe hrtimer testing

[William Cohen]

Stone, Joshua I wrote:
> On Tuesday, May 23, 2006 10:57 AM, William Cohen wrote:
> 
>>Looking over the test coverage results recently posted, one of the
>>places that is lacking in testing is the probes for hrtimer. Josh,
>>does anyone have some tests for that area of the translator? It would
>>do a lot to improve the test coverage of tapsets.cxx.
> 
> 
> The hrtimer is not tested because it's not complete.  I ran into a
> problem with needing exported symbols from the kernel, and LKML rejected
> my request.  So until support is added to the kernel ABI for using
> hrtimers, the code I added will be unusable.
> 
> I checked it in as part of normal development, in hopes that the ABI
> would get resolved.  Since hrtimers are still not exported, perhaps I
> should remove the code.  Would it be sufficient for your coverage tests
> if I just blocked out the code with "#ifdef ENABLE_HRTIMER"?

I wouldn't bother with the #ifdef.

> If anyone has an interest in seeing the hrtimers work in SystemTap,
> please speak up.  If there's enough interest, then maybe we can go back
> to LKML to try to get it enabled in the ABI.
> 
> 
> Josh

In the 2.6.17-rc4 kernel it looks like hrtimer is used by 
kernel/itimer.c. Is there a particular benefit explicitly using the 
hrtimer rather than going through the normal timer mechanism? Are there 
other modules that could use the hrtimer directly?

-Will

probe hrtimer testing

[William Cohen]

Looking over the test coverage results recently posted, one of the 
places that is lacking in testing is the probes for hrtimer. Josh, does 
anyone have some tests for that area of the translator? It would do a 
lot to improve the test coverage of tapsets.cxx.

http://people.redhat.com/wcohen/stap-combined-tests-20060518/src/index.html


-Will