Linux SLAB allocator issue

Linux SLAB allocator issue

[Xin Zhao]

Hi,

I am trying to check how many slabs are used for inode_cache, but
found that all slabs are added to slabs_full list, and slabs_partial
is always empty. Even if the active object number does not exactly
occupy all slabs.

Does that mean Linux 2.6 remove the use of slabs_partial?

Another question, the constructor transfered to the
kmem_cache_create() function is called for every object in a slab when
it is created. Is this true? Is there any way to call back a function
_only once_ when a new slab is allocated?

Thanks,
Xin

Re: Linux SLAB allocator issue

[Pekka Enberg]

On 6/6/06, Xin Zhao <uszhaoxin@gmail.com> wrote:
> I am trying to check how many slabs are used for inode_cache, but
> found that all slabs are added to slabs_full list, and slabs_partial
> is always empty. Even if the active object number does not exactly
> occupy all slabs.
>
> Does that mean Linux 2.6 remove the use of slabs_partial?

No. If slabs_partial is really empty, the number of active objects
should match the number of objects in a slab; otherwise you should see
an error message when you do cat /proc/slabinfo (see s_show in
mm/slab.c for details).

How are you verifying that the partial list is empty?

On 6/6/06, Xin Zhao <uszhaoxin@gmail.com> wrote:
> Another question, the constructor transfered to the
> kmem_cache_create() function is called for every object in a slab when
> it is created. Is this true? Is there any way to call back a function
> _only once_ when a new slab is allocated?

We don't have per-slab constructors. Only per-object. What do you need it for?

                                            Pekka

Re: Linux SLAB allocator issue

[Xin Zhao]

Thanks for your kind reply.

I did the following experiment.

I create my own cache with kmem_cache_create() and specify the
constructor to be init_once()-- a simple constructor like NFS does.

I checked the cache parameter and find that each slab has 1 page and
can hold 10 objects.

Then, I used kmem_cache_alloc() to allocate 128 objects. So it should
occupy 12 full slabs and 1 partial slab. Right?

But when I walk through the slabs_full and slabs_partial list, I found
that slabs_full returned 13 slabs but slabs_partial returned 0.

That's why I am confused. I am using 2.6.16 BTW.

Any further insight?

Thanks,
Xin

On 6/7/06, Pekka Enberg <penberg@cs.helsinki.fi> wrote:
> On 6/6/06, Xin Zhao <uszhaoxin@gmail.com> wrote:
> > I am trying to check how many slabs are used for inode_cache, but
> > found that all slabs are added to slabs_full list, and slabs_partial
> > is always empty. Even if the active object number does not exactly
> > occupy all slabs.
> >
> > Does that mean Linux 2.6 remove the use of slabs_partial?
>
> No. If slabs_partial is really empty, the number of active objects
> should match the number of objects in a slab; otherwise you should see
> an error message when you do cat /proc/slabinfo (see s_show in
> mm/slab.c for details).
>
> How are you verifying that the partial list is empty?
>
> On 6/6/06, Xin Zhao <uszhaoxin@gmail.com> wrote:
> > Another question, the constructor transfered to the
> > kmem_cache_create() function is called for every object in a slab when
> > it is created. Is this true? Is there any way to call back a function
> > _only once_ when a new slab is allocated?
>
> We don't have per-slab constructors. Only per-object. What do you need it for?
>
>                                             Pekka
>

Re: Linux SLAB allocator issue

[Christoph Lameter]

On Wed, 7 Jun 2006, Xin Zhao wrote:

> Then, I used kmem_cache_alloc() to allocate 128 objects. So it should
> occupy 12 full slabs and 1 partial slab. Right?

There may be additional objects in the various caches. If this is a UP 
system then some will certainly be in the per cpu cache.

You can push these back into the free lists by draining the array cache.

If you allocate objects on a slab that is fresh (no objects in it) then 
only full slabs will be used. The remaining objects will end up on the per 
cpu lists where they can be consumed without more work on the full/partial 
arrays.