Re: Linux 3.4-rc1

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 03 April 2012, Kyungmin Park wrote:
> >
> >  - DMA-mapping framework. The tree now has a few more acks from
> > people, and it's largely in the same situation as HSI is: I'll
> > probably pull, but I really wanted the users who are impacted to
> > actually talk to me about it.
> >
> Hi,
> 
> Now Marek and Nvidia persons are used it for DMA mapping based IOMMU.
> 
> Are there other person who help to merge it?
> (+CC related persons).

I can give a little more background information.

We used to have separate implementations of dma_map_ops for each
architecture that had the need to abstract dma_alloc_coherent
and dma_map_* across both IOMMU and linear (virt_to_bus style)
mappings and/or swiotlb. Those dma_map_ops imlpementations were
already merged before the start of the git history and subsequently
got used on powerpc, ia64, x86_64, sparc, alpha, mips, unicore32
and hexagon, roughly in that order.

We are seeing lots of IOMMUs come up on ARM now, but unfortunately
the existing dma_map_ops do not cover the need on ARM to have
two different kind of coherent memory (write-combine and not
write-combine) that we have traditionally abstracted using
the dma_alloc_coherent and dma_alloc_writecombine functions on both
arm and avr32. In order to support IOMMUs on ARM using the same
API as everyone else, I've asked Marek to use dma_map_ops on
ARM, which requires slightly extending the include/linux/dma-mapping.h
interfaces so they cover this variation.

The actually interesting work to implement a better abstraction for
IOMMUs, building the dma-mapping.h API on top of the iommu.h API
in an architecture independent way has also been implemented but
requires these changes as a prerequisite.

	Arnd