2005-10-27 02:28:40

by Chen, Kenneth W

Subject: Weird schedule delay time for cache_reap()

I can't convince myself that the 2nd argument in schedule_delayed_work
called from cache_reap() function make any sense:


static void cache_reap(void *unused)
{ ...

check_irq_on();
up(&cache_chain_sem);
drain_remote_pages();
/* Setup the next iteration */
schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
}


Suppose one have a lucky 1024-processor big iron numa box,
cpu0 will do cache_reap every 2 sec (REAPTIMEOUT_CPUC = 2*HZ).
cpu512 will do cache_reap every 4 sec,
cpu1023 will do cache_reap every 6 sec.

Is the skew intentional on different CPU? Why different interval for
different cpu#?

- Ken

2005-10-27 08:38:27

by Andi Kleen

Subject: Re: Weird schedule delay time for cache_reap()

On Thursday 27 October 2005 04:28, Chen, Kenneth W wrote:
> can't convince myself that the 2nd argument in schedule_delayed_work
>called from cache_reap() function make any sense:


>static void cache_reap(void *unused)
>{ ...
>schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC +
smp_processor_id());
>
> Suppose one have a lucky 1024-processor big iron numa box,
> cpu0 will do cache_reap every 2 sec (REAPTIMEOUT_CPUC = 2*HZ).
> cpu512 will do cache_reap every 4 sec,
> cpu1023 will do cache_reap every 6 sec.
>
> Is the skew intentional on different CPU? Why different interval for
> different cpu#?

It looks like a buggy attempt to make the timers not cluster.
The +smp_processor_id() should be probably only done on the first iteration.
start_cpu_timer() does this already, so removing it should be ok.

-Andi