Re: high resolution timers, scheduling & sleep granularity

[Ric Wheeler <rwheeler@redhat.com>]

Peter Zijlstra wrote:
> On Fri, 2008-08-01 at 09:25 -0400, Josef Bacik wrote:
>
>   
>> Index: linux-2.6/fs/jbd/transaction.c
>> ===================================================================
>> --- linux-2.6.orig/fs/jbd/transaction.c
>> +++ linux-2.6/fs/jbd/transaction.c
>>     
>
>   
>> @@ -63,6 +65,32 @@ get_transaction(journal_t *journal, tran
>>  	return transaction;
>>  }
>>  
>> +static void precision_sleep(unsigned long time)
>> +{
>> +	struct hrtimer_sleeper t;
>> +
>> +	hrtimer_init_on_stack(&t.timer, CLOCK_REALTIME, HRTIMER_MODE_ABS);
>> +	hrtimer_init_sleeper(&t, current);
>> +	t.timer.expires = ktime_add_ns(ktime_get_real(), time);
>> +
>> +	do {
>> +		set_current_state(TASK_UNINTERRUPTIBLE);
>> +
>> +		hrtimer_start(&t.timer, t.timer.expires, HRTIMER_MODE_ABS);
>> +		if (!hrtimer_active(&t.timer))
>> +			t.task = NULL;
>> +
>> +		if (likely(t.task))
>> +			schedule();
>> +
>> +		hrtimer_cancel(&t.timer);
>> +	} while (t.task);
>> +
>> +	set_current_state(TASK_RUNNING);
>> +
>> +	destroy_hrtimer_on_stack(&t.timer);
>> +}
>> +
>>  /*
>>   * Handle management.
>>   *
>>     
>
> I was convinced we already had such a creature,. but I guess I was wrong
> as I can't find it ;-)
>
> Anyway, I'm thinking this function ought to live in kernel/hrtimer.c and
> possibly get renamed to something like hrtimer_sleep_ns() or some such
> (means you can also reuse the do_nanosleep helper in there).
>
> Also, have you considered the impact on platforms that do not support
> hrtimers, or don't have high resolution clock events available?
>   

I agree that this is definitely a much more generically interesting 
function.

With high res timers, falling back to the jiffie's granularity (or not 
sleeping at all in the case of the file system transaction batching) 
seems to be a reasonable fall back,

ric


>   
>> Index: linux-2.6/kernel/hrtimer.c
>> ===================================================================
>> --- linux-2.6.orig/kernel/hrtimer.c
>> +++ linux-2.6/kernel/hrtimer.c
>> @@ -1458,6 +1458,7 @@ void hrtimer_init_sleeper(struct hrtimer
>>  	sl->timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
>>  #endif
>>  }
>> +EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
>>  
>>  static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
>>  {
>>     
>
> That also gets rid of this export..
>
>