Re: [PATCH 06/13] DMAENGINE: driver for the ARM PL080/PL081 PrimeCells

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 3103 bytes --]

On Sat, 2011-01-01 at 15:15 +0000, Russell King - ARM Linux wrote:
> On Wed, Dec 22, 2010 at 03:45:39PM -0800, Dan Williams wrote:
> > Support for the DMA_COMPL flags are necessary if the DMA_MEMCPY
> > capability is advertised, yes this driver got this wrong.  I'll update
> > the documentation to make this requirement clear, and audit the other
> > drivers.  With slave-only drivers the only usage model is one where
> > the client driver owns dma-mapping.  In the non-slave (opportunistic
> > memcpy offload) case the client is unaware of the engine so the driver
> > owns unmapping.  The minimal fix is to disable memcpy offload.
> 
> As a side note, the DMA mapping for slaves should be done using the
> DMA struct device, not the struct device of the peripheral making use
> of the DMA engine.
> 
> Why?  The slave device has no knowledge of how the DMA engine is
> connected into the system, or the DMA parameters associated with the
> device performing the DMA, such as the DMA mask and boundaries.  (If
> there are several generic DMA agents in the system, it can't know
> which is the correct one to use until a channel has been allocated.)
> The only struct device which has this information is the one for the
> DMA engine itself.
> 
> Therefore, the struct device which is passed into the DMA mapping APIs
> to prepare memory for DMA must always be the DMA engine struct device
> (chan->device->dev) and never the slave struct device.

That all seems eminently sensible. However, I wonder if it always has to
be true.

It is not impossible for the DMA controller to "delegate" transactions
so that (to the IOMMU) they appear to come from the individual slave
device rather than from itself.

The Intel IOMMU has now gained support for DMA mapping for devices
enumerated by ACPI — essentially the ACPI "DMAR" table just has a lookup
table of ACPI device paths, and tells us the PCI bus/devfn that their
DMA transactions will *appear* to be from.

What we've seen is that it is the individual slave devices that are
listed in these tables, *not* the DMA controller itself. It looks like
we are actually expected to set up the IOMMU mapping for the *slave*,
not the DMA controller. The system doesn't even *tell* me how to set up
DMA mappings for the DMA controller device; only the slaves.

I haven't looked at the datasheet for the DMA controller in question,
but it's probably the case that there's a per-channel configuration for
what PCIe source-id to put into the DMA transactions.¹

Of course, it's also possible that all these BIOSes are broken and they
*should* just list the DMA controller itself, instead of all the slave
devices. But while I'm always quick to jump to the conclusion that it's
the BIOS at fault, that doesn't necessarily seem likely here...

-- 
dwmw2

¹ We've seen plenty of multi-function devices where the DMA appears to
  come from the wrong function by *accident*, requiring quirks to make
  it work with the IOMMU. This would just be a case of doing the same
  kind of delegation on purpose.


[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5745 bytes --]