lkml.org 
[lkml]   [2001]   [Mar]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: sys_sched_yield fast path

On 11-Mar-2001 Anton Blanchard wrote:
>
>> This is the linux thread spinlock acquire :
>>
>>
>> static void __pthread_acquire(int * spinlock)
>> {
>> int cnt = 0;
>> struct timespec tm;
>>
>> while (testandset(spinlock)) {
>> if (cnt < MAX_SPIN_COUNT) {
>> sched_yield();
>> cnt++;
>> } else {
>> tm.tv_sec = 0;
>> tm.tv_nsec = SPIN_SLEEP_DURATION;
>> nanosleep(&tm, NULL);
>> cnt = 0;
>> }
>> }
>> }
>>
>>
>> Yes, it calls sched_yield() but this is not a std wait for mutex but for
>> spinlocks that are hold a very short time. Real wait are implemented using
>> signals. More, with the new implementation of sys_sched_yield() the task
>> release all its time quantum so, even in a case where a task repeatedly
>> calls
>> sched_yield() the call rate is not so high if there is at least one process
>> to spin. And if there isn't one task with goodness() > 0, nobody cares
>> about
>> sched_yield() performance.
>
> The problem I found with sched_yield is that things break down with high
> levels of contention. If you have 3 processes and one has a lock then
> the other two can ping pong doing sched_yield() until their priority drops
> below the process with the lock. eg in a run I just did then where 2
> has the lock:
>
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 1
> 0
> 2
>
> Perhaps we need something like sched_yield that takes off some of
> tsk->counter so the task with the spinlock will run earlier.

2.4.x has changed the scheduler behaviour so that the task that call
sched_yield() is not rescheduled by the incoming schedule().
A flag is set ( under certain conditions in SMP ) and the goodness()
calculation assign the lower value to the exiting task ( this flag is cleared
in schedule_tail() ).
This could give the task owning the lock the opportunity to complete the locked
code.
But yes, if the locked code is rescheduled for some reason ( timeslice or I/O )
the yielding task will run again.
But this is a software design problem, not a sched_yield() one coz, if the time
path between lock ans unlock can be high the use of sched_yield() is not the
best way to wait.
Wait queue or user space equivalences are a better choice to do this.




- Davide

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

\
 
 \ /
  Last update: 2005-03-22 13:22    [W:0.055 / U:4.504 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site