Preemption question (4xx related)

Preemption question (4xx related)

[Felix Radensky]

Hi,

Preemption is disabled on the vast majority of powerpc
targets. Are there any known problems in this area ?
I'd like to enable CONFIG_PREEMPT_VOLUNTARY on
AMCC 405EX target, to increase application responsiveness.
Is it a bad idea ?
-- 
View this message in context: http://www.nabble.com/Preemption-question-%284xx-related%29-tp24271342p24271342.html
Sent from the linuxppc-dev mailing list archive at Nabble.com.

Re: Preemption question (4xx related)

[Benjamin Herrenschmidt]

On Tue, 2009-06-30 at 05:55 -0700, Felix Radensky wrote:
> Hi,
> 
> Preemption is disabled on the vast majority of powerpc
> targets. Are there any known problems in this area ?
> I'd like to enable CONFIG_PREEMPT_VOLUNTARY on
> AMCC 405EX target, to increase application responsiveness.
> Is it a bad idea ?

In theory it should work, but as you may have noticed, it's generally
disabled and thus not very well tested.

It would be a good idea for us, some of the core devs, to turn that on
in our usual test configs in fact.

Cheers,
Ben.

Re: Preemption question (4xx related)

[Josh Boyer]

On Thu, Jul 02, 2009 at 08:13:13AM +1000, Benjamin Herrenschmidt wrote:
>On Tue, 2009-06-30 at 05:55 -0700, Felix Radensky wrote:
>> Hi,
>> 
>> Preemption is disabled on the vast majority of powerpc
>> targets. Are there any known problems in this area ?
>> I'd like to enable CONFIG_PREEMPT_VOLUNTARY on
>> AMCC 405EX target, to increase application responsiveness.
>> Is it a bad idea ?
>
>In theory it should work, but as you may have noticed, it's generally
>disabled and thus not very well tested.
>
>It would be a good idea for us, some of the core devs, to turn that on
>in our usual test configs in fact.

I've toyed with that idea myself.  I keep coming back to the fact that you need
a workload that would really leverage it, and I don't have one at the moment.

josh

Re: Preemption question (4xx related)

[Benjamin Herrenschmidt]

On Wed, 2009-07-01 at 20:14 -0400, Josh Boyer wrote:

> I've toyed with that idea myself.  I keep coming back to the fact that you need
> a workload that would really leverage it, and I don't have one at the moment.

To some extent that's true but just turning full preemption including
kernel side with all the associated debug bits and lockdep should make a
whole bunch of things show up even with ordinary workloads.

For 440 tend to boot an ubuntu distro off NFS root with all X & DRI 3D
etc... and then run compiz :-)

Cheers,
Ben.

Re: Preemption question (4xx related)

[Michel Dänzer]

On Thu, 2009-07-02 at 08:13 +1000, Benjamin Herrenschmidt wrote:
> On Tue, 2009-06-30 at 05:55 -0700, Felix Radensky wrote:
> > Hi,
> >=20
> > Preemption is disabled on the vast majority of powerpc
> > targets. Are there any known problems in this area ?
> > I'd like to enable CONFIG_PREEMPT_VOLUNTARY on
> > AMCC 405EX target, to increase application responsiveness.
> > Is it a bad idea ?
>=20
> In theory it should work, but as you may have noticed, it's generally
> disabled and thus not very well tested.
>=20
> It would be a good idea for us, some of the core devs, to turn that on
> in our usual test configs in fact.

FWIW, on my PowerBook I've been running with CONFIG_PREEMPT_VOLUNTARY
enabled for a long time (more than a year). No problems so far, and it
definitely seems to help a lot for interactivity.


--=20
Earthling Michel D=C3=A4nzer           |                http://www.vmware.c=
om
Libre software enthusiast         |          Debian, X and DRI developer

Re: Preemption question (4xx related)

[Josh Boyer]

On Thu, Jul 02, 2009 at 05:33:12PM +1000, Benjamin Herrenschmidt wrote:
>On Wed, 2009-07-01 at 20:14 -0400, Josh Boyer wrote:
>
>> I've toyed with that idea myself.  I keep coming back to the fact that you need
>> a workload that would really leverage it, and I don't have one at the moment.
>
>To some extent that's true but just turning full preemption including
>kernel side with all the associated debug bits and lockdep should make a
>whole bunch of things show up even with ordinary workloads.

I can look at doing that for ppc44x_defconfig.  I'll be honest and say I don't
expect it to go well, particularly with lockdep :).

>For 440 tend to boot an ubuntu distro off NFS root with all X & DRI 3D
>etc... and then run compiz :-)

Yes.  Because that's a totally realistic workload for a 440.  I'm surprised you
don't have a p595 machine acting as your home router too!  ;)

josh

Re: Preemption question (4xx related)

[Benjamin Herrenschmidt]

On Thu, 2009-07-02 at 07:12 -0400, Josh Boyer wrote:
> On Thu, Jul 02, 2009 at 05:33:12PM +1000, Benjamin Herrenschmidt wrote:
> >On Wed, 2009-07-01 at 20:14 -0400, Josh Boyer wrote:
> >
> >> I've toyed with that idea myself.  I keep coming back to the fact that you need
> >> a workload that would really leverage it, and I don't have one at the moment.
> >
> >To some extent that's true but just turning full preemption including
> >kernel side with all the associated debug bits and lockdep should make a
> >whole bunch of things show up even with ordinary workloads.
> 
> I can look at doing that for ppc44x_defconfig.  I'll be honest and say I don't
> expect it to go well, particularly with lockdep :).
> 
> >For 440 tend to boot an ubuntu distro off NFS root with all X & DRI 3D
> >etc... and then run compiz :-)
> 
> Yes.  Because that's a totally realistic workload for a 440.  I'm surprised you
> don't have a p595 machine acting as your home router too!  ;)

It doesn't need to be realistic. In fact, a "realistic" workload is the
worst thing to test with because it won't exercise all the "uncommon"
code path which are the ones likely to bite.

So yesm it's not a "realistic" workload, but it's a good "torture"
workload to find bugs.

Cheers,
Ben.

Re: Preemption question (4xx related)

[Josh Boyer]

On Fri, Jul 03, 2009 at 08:41:00AM +1000, Benjamin Herrenschmidt wrote:
>On Thu, 2009-07-02 at 07:12 -0400, Josh Boyer wrote:
>> On Thu, Jul 02, 2009 at 05:33:12PM +1000, Benjamin Herrenschmidt wrote:
>> >On Wed, 2009-07-01 at 20:14 -0400, Josh Boyer wrote:
>> >
>> >> I've toyed with that idea myself.  I keep coming back to the fact that you need
>> >> a workload that would really leverage it, and I don't have one at the moment.
>> >
>> >To some extent that's true but just turning full preemption including
>> >kernel side with all the associated debug bits and lockdep should make a
>> >whole bunch of things show up even with ordinary workloads.
>> 
>> I can look at doing that for ppc44x_defconfig.  I'll be honest and say I don't
>> expect it to go well, particularly with lockdep :).
>> 
>> >For 440 tend to boot an ubuntu distro off NFS root with all X & DRI 3D
>> >etc... and then run compiz :-)
>> 
>> Yes.  Because that's a totally realistic workload for a 440.  I'm surprised you
>> don't have a p595 machine acting as your home router too!  ;)
>
>It doesn't need to be realistic. In fact, a "realistic" workload is the
>worst thing to test with because it won't exercise all the "uncommon"
>code path which are the ones likely to bite.
>
>So yesm it's not a "realistic" workload, but it's a good "torture"
>workload to find bugs.

It was a joke.  But yes, you make perfectly valid points :)

josh