Re: RFC: Non Power of 2 memory allocator

Re: RFC: Non Power of 2 memory allocator

[Andi Kleen]

On Saturday 25 March 2006 17:21, uclinux-dev-owner@uclinux.org wrote:
> You are not allowed to post to this mailing list, and your message has
> been automatically rejected.  If you think that your messages are
> being rejected in error, contact the mailing list owner at
> uclinux-dev-owner@uclinux.org.

P.S.: It is considered quite rude to post with closed lists in Cc 
or even reply-to as it was the case here. Don't do that. If you
want comments on your emails please make sure they can be answered
by everybody.

-Andi

Re: RFC: Non Power of 2 memory allocator

[Andi Kleen]

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.

-Andi

Re: RFC: Non Power of 2 memory allocator

[Hugh Dickins]

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.

Hugh

Re: RFC: Non Power of 2 memory allocator

[Nick Piggin]

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