linux missing support for _POSIX_THREAD_CPUTIME?

linux missing support for _POSIX_THREAD_CPUTIME?

[Chris Friesen]

Hi all,

POSIX defines optional support for clock ID values obtained by invoking
pthread_getcpuclockid().  When passed to clock_gettime(), this allows
the caller to request the CPU-time clock of an arbitrary thread within
the same process as the caller.

It appears that linux doesn't support this functionality--is that
correct?  Is there any plan to enable this?

Is there any current way to obtain the runtime of a particular thread
without knowing its tid (since the thread_id to tid mapping is known
only to glibc)?

Thanks,

Chris

Re: linux missing support for _POSIX_THREAD_CPUTIME?

[Thomas Gleixner]

On Wed, 27 May 2009, Chris Friesen wrote:

That's more a question for the glibc folks. CC'ed.

> Hi all,
> 
> POSIX defines optional support for clock ID values obtained by invoking
> pthread_getcpuclockid().  When passed to clock_gettime(), this allows
> the caller to request the CPU-time clock of an arbitrary thread within
> the same process as the caller.
> 
> It appears that linux doesn't support this functionality--is that
> correct?  Is there any plan to enable this?
> 
> Is there any current way to obtain the runtime of a particular thread
> without knowing its tid (since the thread_id to tid mapping is known
> only to glibc)?
> 
> Thanks,
> 
> Chris
> 

Re: linux missing support for _POSIX_THREAD_CPUTIME ?

[Chris Friesen]

Chris Friesen wrote:

> It appears that linux doesn't support this functionality--is that
> correct?  Is there any plan to enable this?

In particular, it seems to fail the invalid_clockid() test.  Is that
function expecting the clock ID to be negative for this case?

Thanks,

Chris

Re: linux missing support for _POSIX_THREAD_CPUTIME?

[Chris Friesen]

Thomas Gleixner wrote:
> On Wed, 27 May 2009, Chris Friesen wrote:
> 
> That's more a question for the glibc folks. CC'ed.

Actually, I'm looking at the kernel code here.  It looks like
posix_cpu_clock_get() should be able to handle a per-task cpu clock for
another thread within the same process.  However, it's only ever called
by thread_cpu_clock_get(), which hardcodes a clock_id of THREAD_CLOCK,
which corresponds to the current thread.

Similarly, invalid_clockid() will say that the clock_id is invalid if
it's positive but greater than 15.  If my math is right, this means that
any pid > 2 will result in invalid_clockid() failing.

Chris

Re: linux missing support for _POSIX_THREAD_CPUTIME?

[Thomas Gleixner]

On Wed, 27 May 2009, Chris Friesen wrote:

> Thomas Gleixner wrote:
> > On Wed, 27 May 2009, Chris Friesen wrote:
> > 
> > That's more a question for the glibc folks. CC'ed.
> 
> Actually, I'm looking at the kernel code here.  It looks like
> posix_cpu_clock_get() should be able to handle a per-task cpu clock for
> another thread within the same process.  However, it's only ever called
> by thread_cpu_clock_get(), which hardcodes a clock_id of THREAD_CLOCK,
> which corresponds to the current thread.

No, it's also called via the CLOCK_DISPATCH magic in posix_timers.c
 
> Similarly, invalid_clockid() will say that the clock_id is invalid if
> it's positive but greater than 15.  If my math is right, this means that
> any pid > 2 will result in invalid_clockid() failing.

posix_cpu_clock_get() does:

  const pid_t pid = CPUCLOCK_PID(which_clock);

CPUCLOCK_PID is defined as:

#define CPUCLOCK_PID(clock)  ((pid_t) ~((clock) >> 3))

include/linux/posix-timers.h has the helper macros to build the
clock_id value for posix_cpu_clock_get()

#define MAKE_PROCESS_CPUCLOCK(pid, clock) \
        ((~(clockid_t) (pid) << 3) | (clockid_t) (clock))
#define MAKE_THREAD_CPUCLOCK(tid, clock) \
        MAKE_PROCESS_CPUCLOCK((tid), (clock) | CPUCLOCK_PERTHREAD_MASK)

Hope that helps,

     tglx

Re: linux missing support for _POSIX_THREAD_CPUTIME?

[Roland McGrath]

> POSIX defines optional support for clock ID values obtained by invoking
> pthread_getcpuclockid().  When passed to clock_gettime(), this allows
> the caller to request the CPU-time clock of an arbitrary thread within
> the same process as the caller.
> 
> It appears that linux doesn't support this functionality--is that
> correct?  Is there any plan to enable this?

glibc does implement pthread_getcpuclockid using the kernel support for it.
It uses MAKE_THREAD_CPUCLOCK(tid, CPUCLOCK_SCHED) (see linux/posix-timers.h).

> Is there any current way to obtain the runtime of a particular thread
> without knowing its tid (since the thread_id to tid mapping is known
> only to glibc)?

You need a tid to construct the clock ID with MAKE_THREAD_CPUCLOCK().
The pthread_getcpuclockid function exists to do this for you when you are
using pthreads.  If you are not using pthreads, then certainly you have to
figure out what a thread's tid is yourself.


Thanks,
Roland

Re: linux missing support for _POSIX_THREAD_CPUTIME?

[Chris Friesen]

Thomas Gleixner wrote:
> On Wed, 27 May 2009, Chris Friesen wrote:
> 
> 
>>Thomas Gleixner wrote:
>>
>>>On Wed, 27 May 2009, Chris Friesen wrote:
>>>
>>>That's more a question for the glibc folks. CC'ed.
>>
>>Actually, I'm looking at the kernel code here.  It looks like
>>posix_cpu_clock_get() should be able to handle a per-task cpu clock for
>>another thread within the same process.  However, it's only ever called
>>by thread_cpu_clock_get(), which hardcodes a clock_id of THREAD_CLOCK,
>>which corresponds to the current thread.
> 
> 
> No, it's also called via the CLOCK_DISPATCH magic in posix_timers.c

>>Similarly, invalid_clockid() will say that the clock_id is invalid if
>>it's positive but greater than 15.  If my math is right, this means that
>>any pid > 2 will result in invalid_clockid() failing.

Both invalid_clockid() and CLOCK_DISPATCH assume that the clock_id will
be negative in order to call posix_cpu_clock_get().

In pthread_getcpuclockid() in glibc, I couldn't see how we were ending
up with a negative clock_id.  However, I finally clued in to the fact
that there are multiple implementations of this function, and in one
code path in one of the implementations we call MAKE_THREAD_CPUCLOCK()
which can result in a negative clock_id.

Thanks for the help.

Chris