Re: more iommu sg merging fallout

[FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>]

On Wed, 06 Feb 2008 15:18:55 -0800 (PST)
David Miller <davem@davemloft.net> wrote:

> From: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> Date: Thu, 07 Feb 2008 08:12:36 +0900
> 
> > Really sorry about it.
> 
> I am happy to test patches you send to me in the future :-)

Thanks a lot.


> > PARISC, Alpha, and IA64 IOMMUs use the two-pass algorithm like SPARC
> > but their first pass decides how to merge sg entires (and stores that
> > information in the sg entries), then the second pass simpliy follows
> > it (Hopefully I understand these IOMMUs correctly, or else I break
> > them too).
> 
> For now I've removed all of the merging code from the sparc64 IOMMU
> support so that other users do not get corrupt filesystems.  It
> basically mimicks how the intel-iommu code works, ie. no attempts to
> merge anything.

I've just saw it.


> I intend to put merging back in, perhaps something similar to
> powerpc's merging logic but without the expensive (in my opinion)
> IOMMU allocation every loop.  I think it is better to determine the
> segment breaks in one pass, allocate that many IOMMU entries in one
> allocation, then fill them all in.

I thought about asking you if I can modify the SPARC IOMMUs to do
allocation in every loop.

The reason why I need the allocation in every loop is that I also need
to fix the problem that IOMMUs allocate memory areas without
considering a low level driver's segment boundary limits.

http://linux.derkeiler.com/Mailing-Lists/Kernel/2007-11/msg07616.html
http://linux.derkeiler.com/Mailing-Lists/Kernel/2007-12/msg02286.html

As far as I know, all the IOMMUs except for SPARC allocate a free area
in every loop but if it's too expensive for SPARC, then we need to
find a different way to handle segment boundary limits.


> Ideally, we should have some generic code that does all of this.
> Then you would only need to test one implementation.
> 
> It is definitely doable and increasingly necessary as we have so
> many reimplementations of what is essentially identical core code.

Agreed though it's a very hard task.