vmalloc_32 and SLAB incompatible in 2.6.23.9

vmalloc_32 and SLAB incompatible in 2.6.23.9

[Loïc Grenié]

     Since GFP_DMA32 is not considered a valid flag, vmalloc_32 and
  SLAB are incompatible in kernel 2.6.23.9 (and probably later ones).
  It might be as simple as adding GFP_DMA32 everywhere where
  GFP_DMA is used in mm/slab.c, but I don't dare do it. I personally
  switched to SLUB.

      Thanks,

          Loïc Grenié

Re: vmalloc_32 and SLAB incompatible in 2.6.23.9

[Andi Kleen]

"Loïc Grenié" <loic.grenie@gmail.com> writes:

>      Since GFP_DMA32 is not considered a valid flag, vmalloc_32 and
>   SLAB are incompatible in kernel 2.6.23.9 (and probably later ones).
>   It might be as simple as adding GFP_DMA32 everywhere where
>   GFP_DMA is used in mm/slab.c, but I don't dare do it. I personally
>   switched to SLUB.

I don't think it would really work for slub either.

If you look at /proc/slabinfo you see that there are separate DMA
caches. That is because sl[abo]b all cache objects and if the 
objects have different requirements (like GFP_DMA, GFP_DMA32, GFP_KERNEL)
they would need different caches. Otherwise you risk getting
later objects originally allocated with the wrong flags.

To make GFP_DMA32 work would require adding another set of DMA32
caches to all kmalloc caches etc or a separate one for each
custom kmem cache that you use that only gets _DMA32 objects.

Better just use __get_free_pages() directly. 

-Andi

Re: vmalloc_32 and SLAB incompatible in 2.6.23.9

[Loïc Grenié]

23 Jan 2008 10:35:28 +0100, Andi Kleen <andi@firstfloor.org>:
> "Loïc Grenié" <loic.grenie@gmail.com> writes:
>
> >      Since GFP_DMA32 is not considered a valid flag, vmalloc_32 and
> >   SLAB are incompatible in kernel 2.6.23.9 (and probably later ones).
> >   It might be as simple as adding GFP_DMA32 everywhere where
> >   GFP_DMA is used in mm/slab.c, but I don't dare do it. I personally
> >   switched to SLUB.
>
> I don't think it would really work for slub either.
>
> If you look at /proc/slabinfo you see that there are separate DMA
> caches. That is because sl[abo]b all cache objects and if the
> objects have different requirements (like GFP_DMA, GFP_DMA32, GFP_KERNEL)
> they would need different caches. Otherwise you risk getting
> later objects originally allocated with the wrong flags.
>
> To make GFP_DMA32 work would require adding another set of DMA32
> caches to all kmalloc caches etc or a separate one for each
> custom kmem cache that you use that only gets _DMA32 objects.
>
> Better just use __get_free_pages() directly.

     vmalloc_32 is using __vmalloc. I am much too ignorant to try and change
  anything in mm/*.[ch]. If vmalloc_32 is incompatible with sl[aou]b, should'nt
  it be either modified or removed ?

        Thanks,

              Loïc Grenié

Re: vmalloc_32 and SLAB incompatible in 2.6.23.9

[Andi Kleen]

>      vmalloc_32 is using __vmalloc. I am much too ignorant to try and change
>   anything in mm/*.[ch]. If vmalloc_32 is incompatible with sl[aou]b, should'nt
>   it be either modified or removed ?

vmalloc does not use sl[aou]b for the data pages. That is why it can
support GFP_DMA32 (or even GFP_DMA if you wanted) 

-Andi