Re: [PATCH 0/3] sched: newidle and RT wake-buddy fixes

["Gregory Haskins" <ghaskins@novell.com>]

>>> On Mon, Jun 30, 2008 at  9:15 AM, in message <20080630131511.GA7506@elte.hu>,
Ingo Molnar <mingo@elte.hu> wrote: 

> * Gregory Haskins <ghaskins@novell.com> wrote:
> 
>> Hi Ingo,
>>   The following patches apply to linux-tip/sched/devel and enhance the
>> performance of the kernel (specifically in PREEMPT_RT, though they do
>> not regress mainline performance as far as I can tell).  They offer
>> somewhere between 50-100% speedups in netperf performance, depending
>> on the test.
> 
> -tip testing found this boot hang:

Ok, I dug in a little bit here.  I haven't verified this out yet, but I think the problem is that
your config is PREEMPT_VOLUNTARY which NOPs the preempt_disable() in schedule() that
I rely on to allow the lock to be dropped.  (Doh!)

One way I can fix this is to fixup the newidle() code to only play the irq dropping tricks
ifdef CONFIG_PREEMPT == TRUE.  Does this sound reasonable, or is there a better way
to address this?

-Greg

> 
>  Linux version 2.6.26-rc8-tip (mingo@dione) (gcc version 4.2.3) #12917 
>  SMP Mon Jun 30 15:06:32 CEST 2008
>  [...]
>  CPU 1/1 -> Node 0
>  CPU: Physical Processor ID: 0
>  CPU: Processor Core ID: 1
>  CPU1: AMD Athlon(tm) 64 X2 Dual Core Processor 3800+ stepping 02
>  Brought up 2 CPUs
>  Total of 2 processors activated (8041.15 BogoMIPS).
>  [ hard hang ]
> 
> with this config:
> 
>   http://redhat.com/~mingo/misc/config-Mon_Jun_30_14_54_19_CEST_2008.bad
> 
> full bootlog:
> 
>   http://redhat.com/~mingo/misc/hang-Mon_Jun_30_14_54_19_CEST_2008.bad
> 
> it should continue with this bootup sequence:
> 
>   calling  net_ns_init+0x0/0x1a0
>   net_namespace: 376 bytes
>   initcall net_ns_init+0x0/0x1a0 returned 0 after 0 msecs
>   calling  init_smp_flush+0x0/0x60
> 
> i've bisected it down to:
> 
> --------------
> | commit cc8160c56843201891766660e3816d2e546c1b17
> | Author: Gregory Haskins <ghaskins@novell.com>
> | Date:   Fri Jun 27 14:29:50 2008 -0600
> |
> |     sched: enable interrupts and drop rq-lock during newidle balancing
> --------------
> 
> so i've reverted that change for now.
> 
> 	Ingo