Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 29 Apr 2009 21:01:23 +0200
From:      Andreas Tobler <andreast-list@fgznet.ch>
To:        Kostik Belousov <kostikbel@gmail.com>
Cc:        svn-src-head@freebsd.org, Jeff Roberson <jeff@freebsd.org>, src-committers@freebsd.org, svn-src-all@freebsd.org
Subject:   Re: svn commit: r191643 - in head/sys: kern sys
Message-ID:  <49F8A403.6070209@fgznet.ch>
In-Reply-To: <20090429165417.GG40751@deviant.kiev.zoral.com.ua>
References:  <200904290315.n3T3FiJW067558@svn.freebsd.org> <20090429165417.GG40751@deviant.kiev.zoral.com.ua>

next in thread | previous in thread | raw e-mail | index | archive | help
Kostik Belousov wrote:
> On Wed, Apr 29, 2009 at 03:15:44AM +0000, Jeff Roberson wrote:
>> Author: jeff
>> Date: Wed Apr 29 03:15:43 2009
>> New Revision: 191643
>> URL: http://svn.freebsd.org/changeset/base/191643
>>
>> Log:
>>    - Remove the bogus idle thread state code.  This may have a race in it
>>      and it only optimized out an ipi or mwait in very few cases.
>>    - Skip the adaptive idle code when running on SMT or HTT cores.  This
>>      just wastes cpu time that could be used on a busy thread on the same
>>      core.
>>    - Rename CG_FLAG_THREAD to CG_FLAG_SMT to be more descriptive.  Re-use
>>      CG_FLAG_THREAD to mean SMT or HTT.
>>   
>>   Sponsored by:   Nokia
>>
>> Modified:
>>   head/sys/kern/sched_ule.c
>>   head/sys/kern/subr_smp.c
>>   head/sys/sys/smp.h
> 
> Now I see a reason why it is better #ifdef SMP the code that uses CG_FLAG_*.
> Also, we should check for tdq_cg != NULL in one more place.
> 
> See the patch below, instead of exposing CG_FLAG_* for !SMP configs.
> 
> diff --git a/sys/kern/sched_ule.c b/sys/kern/sched_ule.c
> index 680572d..fe3a119 100644
> --- a/sys/kern/sched_ule.c
> +++ b/sys/kern/sched_ule.c
> @@ -891,6 +891,7 @@ tdq_move(struct tdq *from, struct tdq *to)
>  	return (1);
>  }
>  
> +#ifdef SMP
>  /*
>   * This tdq has idled.  Try to steal a thread from another cpu and switch
>   * to it.
> @@ -947,6 +948,7 @@ tdq_idled(struct tdq *tdq)
>  	spinlock_exit();
>  	return (1);
>  }
> +#endif
>  
>  /*
>   * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
> @@ -2525,7 +2527,9 @@ sched_idletd(void *dummy)
>  	struct thread *td;
>  	struct tdq *tdq;
>  	int switchcnt;
> +#ifdef SMP
>  	int i;
> +#endif
>  
>  	mtx_assert(&Giant, MA_NOTOWNED);
>  	td = curthread;
> @@ -2543,7 +2547,9 @@ sched_idletd(void *dummy)
>  		 * loops while on SMT machines as this simply steals
>  		 * cycles from cores doing useful work.
>  		 */
> -		if ((tdq->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0 &&
> +#ifdef SMP
> +		if (tdq->tdq_cg != NULL &&
> +		    (tdq->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0 &&
>  		    switchcnt > sched_idlespinthresh) {
>  			for (i = 0; i < sched_idlespins; i++) {
>  				if (tdq->tdq_load)
> @@ -2551,6 +2557,7 @@ sched_idletd(void *dummy)
>  				cpu_spinwait();
>  			}
>  		}
> +#endif
>  		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
>  		if (tdq->tdq_load == 0)
>  			cpu_idle(switchcnt > 1);



Fyi, with the above patch I can build amd64, sparc64 and powerpc kernel.

Andreas



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?49F8A403.6070209>