scatterlist.c or dma_map_sg() should coalesce?

scatterlist.c or dma_map_sg() should coalesce?

[Leon Woestenberg]

Hello,

in 2.6.27+, if I want to create a scatterlist of a large buffer in
virtual memory in a device driver, can I expect the scatterlist and/or
mapping API to provide me with a minimal list, each entry describing a
maximum bus-contiguous piece of memory? (i.e. can I assume coalescing
is done for me?).

In 2.6.24, dma_map_sg() *may* coalesce, but doesn't do so for my
platform, so for optimal DMA transfers the device driver needs to
coalesce the list entries.

Regards,
-- 
Leon

Re: scatterlist.c or dma_map_sg() should coalesce?

[FUJITA Tomonori]

On Sun, 28 Sep 2008 17:47:59 +0200
"Leon Woestenberg" <leon.woestenberg@gmail.com> wrote:

> Hello,
> 
> in 2.6.27+, if I want to create a scatterlist of a large buffer in
> virtual memory in a device driver, can I expect the scatterlist and/or
> mapping API to provide me with a minimal list, each entry describing a
> maximum bus-contiguous piece of memory? (i.e. can I assume coalescing
> is done for me?).

No, it's not the job of dma_map_sg. Before dma_map_sg, physical
continuous page frames in a scatter list are supposed be coalesced.
For example, the block layer creates a scatter list in which physical
continuous page frames are coalesced for SCSI drivers.


> In 2.6.24, dma_map_sg() *may* coalesce, but doesn't do so for my
> platform, so for optimal DMA transfers the device driver needs to
> coalesce the list entries.

With some of IOMMUs, dma_map_sg might coalesce virtual and physical
continuous page frames. If 2.6.24 doesn't do for you, 2.6.27 doesn't
too.