Re: bigphysarea in 2.2

Re: bigphysarea in 2.2

[David Mentre]

Peter Monta <pmonta@halibut.imedia.com> writes:

> > Is it too late to ask that the bigphysarea patch be included in the
> > 2.1-and-soon-to-be-2.2 kernel?
> 
> Seconded.  Thanks for offering to maintain it.

 With the new kernel memory manager, and if the defragmenting code which
is under development works, wouldn't it be more useful to use standard
kernel memory allocation. Static allocation like in bigphysarea is more
a work-around that a real solution.

 Maybe we should ask the memory kernel hackers (Stephen, Ben, Rick,
Werner?) to support big allocations. I personally need 512 Kbytes
contiguous blocks for a direct-from/to-memory network card. A possible
problem is that those blocks should 512 Kbytes aligned (Argh!! !*%&@
hardware). 

 Is it doable with current code in 2.1 (due to feature freeze) ? On the
other side, I agree that at least a work-around like bigphysarea should
be included in 2.2. More and more PCI cards make direct access to memory
for big transfers.

Regards,
d.
-- 
 David.Mentre@irisa.fr -- http://www.irisa.fr/prive/dmentre/
           == GNU & Linux: Change _our_ world ==

Re: bigphysarea in 2.2

[David Mentre]

"Stephen C. Tweedie" <sct@dcs.ed.ac.uk> writes:
  > me:
> >  With the new kernel memory manager, and if the defragmenting code which
> > is under development works, wouldn't it be more useful to use standard
> > kernel memory allocation. Static allocation like in bigphysarea is more
> > a work-around that a real solution.
> 
> Unfortunately, the current code simply doesn't grok areas larger than
> 128KB, and even if it did, it is unlikely that it could be made to
> work well in that case --- the existance of just one non-pagable
> allocation (slab, kmalloc, page table etc.) in any 512K block would
> render that entire region unreclaimable by the swapper.  If you need
> such large physically contiguous regions, then bigphysarea is still
> a better option.

 Ok, I have understood. The last problem is that driver using the
bigphysarea need a patched kernel. It's annoying (at least for a
production kernel). But it seems from recent discussions that a lobbying
group tries to incorporate it in the mainstream kernel. ;) I vote for
it, as this patch is more and more necessary (and because it seems to be
_the_ solution for big allocations).

Regards,
d.
-- 
 David.Mentre@irisa.fr -- Perso : http://www.irisa.fr/prive/dmentre/
            == GNU et Linux : Ameliorer _notre_ monde ==

Re: bigphysarea in 2.2

[Stephen C. Tweedie]

Hi,

On 11 Apr 1998 12:56:10 +0200, David Mentre <David.Mentre@irisa.fr> said:

> Peter Monta <pmonta@halibut.imedia.com> writes:
>> > Is it too late to ask that the bigphysarea patch be included in the
>> > 2.1-and-soon-to-be-2.2 kernel?
>> 
>> Seconded.  Thanks for offering to maintain it.

>  With the new kernel memory manager, and if the defragmenting code which
> is under development works, wouldn't it be more useful to use standard
> kernel memory allocation. Static allocation like in bigphysarea is more
> a work-around that a real solution.

>  Maybe we should ask the memory kernel hackers (Stephen, Ben, Rick,
> Werner?) to support big allocations. I personally need 512 Kbytes
> contiguous blocks for a direct-from/to-memory network card. A possible
> problem is that those blocks should 512 Kbytes aligned (Argh!! !*%&@
> hardware). 

Unfortunately, the current code simply doesn't grok areas larger than
128KB, and even if it did, it is unlikely that it could be made to
work well in that case --- the existance of just one non-pagable
allocation (slab, kmalloc, page table etc.) in any 512K block would
render that entire region unreclaimable by the swapper.  If you need
such large physically contiguous regions, then bigphysarea is still
a better option.

--Stephen