Re: [PATCH RFC v5 2/2] dmaengine: avalon-test: Intel Avalon-MM DMA Interface for PCIe test

From: Alexander Gordeev
Date: Thu Nov 14 2019 - 10:53:38 EST


On Thu, Nov 14, 2019 at 10:33:31AM +0530, Vinod Koul wrote:
> On 06-11-19, 20:22, Alexander Gordeev wrote:
> > This is a sample implementation of a driver using "avalon-dma" to
> > perform data transfers between target device memory and system memory:
> >
> > +----------+ +----------+ +----------+
> > | RAM |<-->| Avalon |<---PCIe--->| Host |
> > +----------+ +----------+ +----------+
> >
> > The target device is expected to use only Avalon-MM DMA Interface for
> > PCIe to initiate DMA transactions - without custom hardware specifics
> > to make such transfers possible.
> >
> > Unlike "dmatest" driver, the contents of DMAed data is not manipulated by
> > "avalon-test" in any way. It is basically pass-through and the the data
> > are fully dependent on the target device implementation. Thus, it is up
> > to the users to analyze received or provide meaningful transmitted data.
>
> Is this the only reason why you have not used dmatest. If so, why not
> add the feature to dmatest to optionally not check the DMAed data
> contents?

The main reason is that "dmatest" does not support DMA_SLAVE type of
transactions.

I considered adding DMA_SLAVE to "dmatest". But it would break the
current neat design and does not appear solving the issue of data
verification. Simply besause in general DMA_SLAVE case there is no
data integrity criteria easily available to the driver. I.e if the
data is a sensor image then verifying it in the driver would be
pointless.

So in case of "avalon-test" I offloaded the task of data verification
to the user. The driver itself just streams user data to/from device.

In fact, this approach is not "avalon-dma" specific and could be used
with any "dmaengine" compatible driver. Moreover, it would be quite
useful for bringing up devices in embedded systems. I.e one could
master a raw display frame in user space and DMA it via the driver -
without graphic stack involved.

The only missing functionality I could think of is using DMABUFs, but that
is very easy to add.

Actually, "avalon-test" is rather a presentation of how I tested
"avalon-dma". I understand "dmatest" is more easy to trust and I could
probably make it working with DMA_SLAVE type. But that would entail
hardware design requirements:

- DMA slave should both respond to read and write transactions;
- data read should always be the same as data written;

I have such version of hardware design, but I doubt majorify of devices
out there can honor the above requirements.

Summarizing - I would suggest not to change "dmatest" and bring in a
generalized version of "avalon-test" if you find it useful for a wider
audience.

Thanks!

> --
> ~Vinod