On Mon, 1 Mar 2010 11:34:37 -0500
Konrad Rzeszutek Wilk [off-list ref] wrote:

On Sun, Feb 28, 2010 at 12:16:28AM -0800, David Miller wrote:

quoted

From: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Date: Sun, 28 Feb 2010 03:38:19 +0900

quoted

When I proposed such approach (always use swiotlb) before, IIRC,
the objections were:

- better to make allocation respect dma_mask. (I don't think that this
  approach is possible since we don't know which device handles data
  later when we allocate memory).

And such objects might end up being processed by multiple devices with
different DMA restrictions.

quoted

- swiotlb is not good for small systems since it allocates too much
  memory (we can fix this though).

Indeed.

What would be a good mechanism for this? Enumerating all of the PCI
devices to find out which ones are 32-bit and then allocate some chunk
of memory based on the amount of them? say, 1MB per card?

Or maybe a simpler one - figure out how many pages we have an allocate
based on some sliding rule (say, 8MB for under 512MB, 16MB between 512MB
and 2GB, and 32MB for 2GB to 4GB, and after that the full 64MB?)

Hmm, have you read the above objection from embedded system people ? :)

We can't pre-allocate such lots of memory (several MB). And we don't
need to.

We have to pre-allocate some for the block layer (we have to guarantee
that we can handle one request at least), however we don't need to
pre-allocate for the rest (including the network stack).

This guarantee is a bit difficult since the dma_map_* doesn't know who
is at the upper layer; the dma_map_* cannot know if it needs to
allocate from the pre-allocated pool or not. I thought about adding
'dma attribute flag' to struct device (to be exact, struct
device_dma_parameters).