Re: RFC: Non Power of 2 memory allocator

[Nick Piggin <nickpiggin@yahoo.com.au>]

Hugh Dickins wrote:
> On Sat, 25 Mar 2006, Andi Kleen wrote:
> 
>>Robin Getz <rgetz@blackfin.uclinux.org> writes:
>>
>>
>>>The  buddy system allocates things in power of 2 pages sizes (4k, 8k,
>>>16k, 32k, 64k, 128k, 256k, 512k, 1024k), which works fine on most
>>>systems, but an embedded system, which is running without a MMU (
>>>Memory Management Unit) - RAM is precious, and when you only need
>>>129k for an application, you don't want to allocate a power of 2,
>>>which gives you 256k -  an extra 127k, which can't be used by
>>>anything else.
>>
>>In 2.4 I solved this problem at some point by just returning
>>the excess pages to the buddy allocator. There was even
>>a nice function to do this (alloc_exact)
>>
>>That won't work for slab, but does for __get_free_pages() which
>>is better for large allocations anyways. slab imho doesn't make
>>sense for allocation anywhere bigger PAGE_SIZE/2. At some
>>point in 2.6 there was trouble with "compound pages" but I think
>>that has been resolved. 
>>
>>Just implementing alloc_exact again would be the simplest solution
>>for your problem.
> 
> 
> Nick has put a split_page function into the 2.6.16-git mm/page_alloc.c,
> which I believe is supposed to be a helper in this kind of operation.
> You'd best take a look at where and how it's used.  Perhaps Andi's
> alloc_exact should be reimplemented in terms of it.
> 

Indeed. nommu already uses the slab allocator for user allocations.
I guess the problem is that slab probably doesn't do this exact
allocation thing either.

I'd like to see nommu move over to using something like Andi's
alloc_exact (away from slab), because that's how the normal
kernel does it.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com