Re: [PATCH v2 1/3] process cputimer is moving faster than its correspondingclock

From: KOSAKI Motohiro
Date: Mon Apr 29 2013 - 15:10:06 EST


(4/29/13 2:54 PM), Olivier Langlois wrote:
> On Mon, 2013-04-29 at 14:31 -0400, KOSAKI Motohiro wrote:
>> (4/29/13 2:20 PM), Olivier Langlois wrote:
>>>
>>>>
>>>>>> I'm confused. glibc's rt/tst-cputimer1 doesn't have thread exiting code. I have
>>>>>> no seen any issue in this accounting.
>>>>>
>>>>> glibc launch a helper thread to receive timer signal and will also
>>>>> create a new thread upon signal reception when a timer is created with
>>>>> sigev_notify = SIGEV_THREAD;
>>>>>
>>>>> please see:
>>>>>
>>>>> glibc-2.17/nptl/sysdeps/unix/sysv/linux/timer_create.c
>>>>> glibc-2.17/nptl/sysdeps/unix/sysv/linux/timer_routines.c
>>>>
>>>> I know. I taled thread exiting. not thread creating. And, as far as I can see, only test sig1 can fail,
>>>> not thr[12].
>>>>
>>> Apart from glibc helper thread, the threads created for handling timer
>>> firing all do exit immediatly as soon as user callback returns.
>>
>> And, libc ensure its exiting finished before starting actual tests. Why such thread exiting
>> affect timers code? It shouldn't. becuase signal.cputimer is initialized timer_settime().
>> The initialization is incorrect, we should fix initialization.
>
> It doesn't have anything to do with initialisation.
>
> Quick Quiz #1: How does the cputimer tick?
> Answer: With calls to account_group_exec_runtime()

Only account when cputimer->running. Quick Quiz: When turn on cputimer->running?


> Every task updates occuring after release_task() has been called in
> do_exit() (scheduler ticks or the task final schedule() call) will be
> lost because tasks stats are added to the global group stats located in
> the signal struct in release_task() So every update after release_task()
> will be lost but account_group_exec_runtime is still called.

tick lost doesn't occur an issue. because glibc only test posix conformance and
posix allow inacculacy. In other words, timer must not run faster than real clock.
but lost and makes slower are accepted in the test.


>
> Tasks that go in zombie state are fine because release_task() will be
> called later. Autoreap task (those with CLONE_THREAD ie: pthreads) calls
> release_task() before the last context switch. Please do read
> kernel/exit.c.
>
> Hence cputimer advance faster than the process clock.

Again the lost makes slower, not faster.


> Hence the POSIX compliance from your pseudo code does not hold
>
> sighandler(){
> t1 = clock_gettime()
> }
>
> t0 = clock_gettime()
> timer_settime(timeout);
> ... wait to fire
>
> assert (t1 - t0 >= timeout)
>
>>
>>
>>> I count 12 thread exits during tst-cputimer1 execution. The errors do
>>> add up hence you're more likely to see errors after 2.5 sec and up from
>>> start of execution. I have seen sig1, thr[12] fails. I see no reason why
>>> one could not fail.

I ran over 3000 times run this case. but I have no seen your described issue.






--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/