Slub support with PREEMPT_RT ?

Slub support with PREEMPT_RT ?

[Jean-Michel Hautbois]

Hi !

I am wondering why SLUB isn't supported with PREEMPT_RT ?
I cannot find any info on that except the fact it is not working yet.
What do we have to do in order to support it ?

Thanks in advance,
JM

Re: Slub support with PREEMPT_RT ?

[Darren Hart]

On 11/17/2010 08:18 AM, Jean-Michel Hautbois wrote:
> Hi !
>
> I am wondering why SLUB isn't supported with PREEMPT_RT ?
> I cannot find any info on that except the fact it is not working yet.
> What do we have to do in order to support it ?

Peter can answer this better than I can, but it has something to do with 
all the local_bh_disable() calls not being compatible with the goals of 
PREEMPT_RT - if I remember correctly.


-- 
Darren Hart
Yocto Linux Kernel

Re: Slub support with PREEMPT_RT ?

[Peter Zijlstra]

On Wed, 2010-11-17 at 17:23 -0800, Darren Hart wrote:
> On 11/17/2010 08:18 AM, Jean-Michel Hautbois wrote:
> > Hi !
> >
> > I am wondering why SLUB isn't supported with PREEMPT_RT ?
> > I cannot find any info on that except the fact it is not working yet.
> > What do we have to do in order to support it ?
> 
> Peter can answer this better than I can, but it has something to do with 
> all the local_bh_disable() calls not being compatible with the goals of 
> PREEMPT_RT - if I remember correctly.

Nah, its slub doens't use the softirq crap. But its close, you have to
iron out the per-cpu assumptions in the thing. I did a few slub-rt ports
at various times but the thing kept changing too much and I gave up.

I should probably do a new port and hope the thing is more stable now
that the excitement is down or somesuch.

Re: Slub support with PREEMPT_RT ?

[Jean-Michel Hautbois]

2010/11/18 Peter Zijlstra <peterz@infradead.org>:
> On Wed, 2010-11-17 at 17:23 -0800, Darren Hart wrote:
>> On 11/17/2010 08:18 AM, Jean-Michel Hautbois wrote:
>> > Hi !
>> >
>> > I am wondering why SLUB isn't supported with PREEMPT_RT ?
>> > I cannot find any info on that except the fact it is not working yet.
>> > What do we have to do in order to support it ?
>>
>> Peter can answer this better than I can, but it has something to do with
>> all the local_bh_disable() calls not being compatible with the goals of
>> PREEMPT_RT - if I remember correctly.
>
> Nah, its slub doens't use the softirq crap. But its close, you have to
> iron out the per-cpu assumptions in the thing. I did a few slub-rt ports
> at various times but the thing kept changing too much and I gave up.
>
> I should probably do a new port and hope the thing is more stable now
> that the excitement is down or somesuch.
>

My real question is the impact of SLUB compared to SLAB on the
determinism of kmalloc usage and all caches.
I did some tests using SLUB instead of SLAB on a small (MPC5200 /
400Mhz / 16k i-cache and 16k d-cache), and slub seems to be better.
But I don't know if there is a possibility to tune SLAB in order to
get good results.

Re: Slub support with PREEMPT_RT ?

[Peter Zijlstra]

On Thu, 2010-11-18 at 09:59 +0100, Jean-Michel Hautbois wrote:
> 
> My real question is the impact of SLUB compared to SLAB on the
> determinism of kmalloc usage and all caches.
> I did some tests using SLUB instead of SLAB on a small (MPC5200 /
> 400Mhz / 16k i-cache and 16k d-cache), and slub seems to be better.
> But I don't know if there is a possibility to tune SLAB in order to
> get good results. 

That's a pretty pointless idea, non of them are deterministic allocators
-- nor can they be, page allocation involves the whole vm reclaim path.

If you want determinism don't allocate memory.