Re: [PATCH V4 1/3] x86: introduce CONFIG_X86_DEV_DMA_OPS

[Alessandro Rubini <rubini@gnudd.com>]

Peter Anvin:
> I looked over it, and at least this bit seems okay to me.
> 
> Acked-by: H. Peter Anvin <hpa@zytor.com>

Thanks.

> I have to express some concern about where this device may be going,
> though.

Well, this CONFIG_X86_DEV_DMA_OPS in particular should not be
concerning, as it's just separating out an x86-64 ifdef so it can be
used and activated by 32-bit systems as well.

As for "this device", it's an I/O Hub that can be used as main chipset
on Atom computers, mainly aimed at the automotive sector. It is a
typical ARM SoC with the usual set of internal peripherals, mounted on
a bridge between PCIe and AMBA, so the PCIe host can access the SoC
peripherals -- and SoC peripherals can DMA to PCIe memory.  Actually,
some SoC devices work well using the existing amba drivers, with
minimal changes that I plan to clean up and submit soon.

The strange DMA requirements here depend on the size limits of the
respective windows: SoC devices can only access a 512MB window of the
host memory, so we need to take care of this with special options.

The current approach is using swiotlb and it works pretty well, though
with a limit of 4MB of swiotlb area.  I've been considering use of the
DMA zone to this aim: internally I have some half-working thing that
resuses the ISA DMA zone for our own aims, raising the DMA limit to
512MB.

Do you think the approach may make sense? I use this basic
thing in Kconfig, to rely on GFP_DMA for the rest:

  config MAX_DMA_PADDR
          int
          default 536870912 if MAX_DMA_PADDR_512M
          default 16777216

  config MAX_DMA_PADDR_512M
          bool

(actually, we'd benefit from being able to use hex in defaults)

Is this worth exploring, to possibly submit a patch in this direction?

thanks
/alessandro