a question on DMA and remapping

a question on DMA and remapping

[Alessandro Rubini]

Hello.
This goes to the maintainers of x86::asm/dma-mapping.h and lib/swiotlb.c,
with Cc: to involved people.

I have an Intel evaluation board with the ST IO-Hub called STA2X11 and
I'm working to port the STA2X11 drivers to mainstream.  The code is
currently on sourceforge.  Since the device is based on a PCI-Amba
bridge, all DMA addresses are different from CPU addresses, even for
normal PCI devices, like EHCI.

Unfortunately, the current patch is changing 3 inlines to external
functions. They are dma_capable, phys_to_dma, dma_to_phys -- which
actually are only used in swiotlb.c .

I thought about the following two approaches towards a clean port:

  - using dma_supported(), which relies on dev->dma_ops->dma_supported
    and adding phys_to_dma and dma_to_phys to the dma operations. In
    the new fields, the default NULL may be used to select the current
    behaviour in an inline function.

  - copying lib/swiotlb.c to my own file, which will be almost
    identical to the existing one but for a few lines.

The former approach will have some tiny overhead on all users, besides
messing with dma_capable and dma_allowed, possibly introducing bugs in
some corner cases (but the current situation is quite messy, may I
say...)

The latter approach means code duplication, which is bad. Although
maybe over time I may be able to shrink the current swiotlb.c to a
much smaller snippet. I tend to prefer this one, but I'm not sure if
it's acceptable.

Any feedback is welcome. Thanks in advance.

/alessandro

Re: a question on DMA and remapping

[Andrew Morton]

On Fri, 18 Nov 2011 00:30:59 +0100
Alessandro Rubini <ru@gnudd.com> wrote:

> Hello.
> This goes to the maintainers of x86::asm/dma-mapping.h and lib/swiotlb.c,
> with Cc: to involved people.
> 
> I have an Intel evaluation board with the ST IO-Hub called STA2X11 and
> I'm working to port the STA2X11 drivers to mainstream.  The code is
> currently on sourceforge.  Since the device is based on a PCI-Amba
> bridge, all DMA addresses are different from CPU addresses, even for
> normal PCI devices, like EHCI.
> 
> Unfortunately, the current patch is changing 3 inlines to external
> functions. They are dma_capable, phys_to_dma, dma_to_phys -- which
> actually are only used in swiotlb.c .
> 
> I thought about the following two approaches towards a clean port:
> 
>   - using dma_supported(), which relies on dev->dma_ops->dma_supported
>     and adding phys_to_dma and dma_to_phys to the dma operations. In
>     the new fields, the default NULL may be used to select the current
>     behaviour in an inline function.

Sounds OK.  swiotlb.c isn't exactly super-fast code anyway.

>   - copying lib/swiotlb.c to my own file, which will be almost
>     identical to the existing one but for a few lines.

Don't do that ;)

> The former approach will have some tiny overhead on all users, besides
> messing with dma_capable and dma_allowed, possibly introducing bugs in
> some corner cases (but the current situation is quite messy, may I
> say...)
> 
> The latter approach means code duplication, which is bad. Although
> maybe over time I may be able to shrink the current swiotlb.c to a
> much smaller snippet. I tend to prefer this one, but I'm not sure if
> it's acceptable.
> 
> Any feedback is welcome. Thanks in advance.

Generalising the existing code to cover more cases isn't a bad thing to
do.  Others might be able to use it, and they surely won't be able to
use any cloned-and-owned swiotlb.c.

Please do carefully docment the new interfaces so others can understand
why they exist and can use them successfully.

Re: a question on DMA and remapping

[Alan Cox]

> Generalising the existing code to cover more cases isn't a bad thing
> to do.  Others might be able to use it, and they surely won't be able
> to use any cloned-and-owned swiotlb.c.

If you look on the kernel list there has also been some discussion and
work recently towards removing a lot of the duplication between the
iommu devices.

I think really you need to be an actual iommu instance because in
reality you are. Trying to fudge it any other way breaks the moment
someone tries to do a device to device DMA transfer because your
bus/phys mapping seen from your device and the bus/phys mapping seen
from another device into your device do not match as I understand it.

Alan