Hi,

as discussed, I have implemented a kfree() without virt_to_page(). The 
patch is a proof-of-concept, intended for developers.

- the kmalloc caches remained at power-of-two, i.e. the patch wastes 
lots of memory.
- alloc_threadinfo() assumed that kmalloc(8192) returned 8192 byte 
aligned memory. This was never guaranteed. I have replaced the kmalloc 
call with gfp() on i386. Probably all other archs are broken.
- get_colourspace() is incorrect, a bounds check is missing (the last 
object must never start beyond addr+PAGE_SIZE+sizeof(kmem_pagehdr)).

The patch boots on i386 with debugging enabled, I have not performed any 
other tests.

Btw, it might be possible to combine the kmem_pagehdr implementation and 
virt_to_page(): If the obj pointer is PAGE_SIZE-aligned, then use 
virt_to_page(). Otherwise use kmem_pagehdr.

--
    Manfred