Re: [PATCH] dmaengine: at_hdmac: new driver for the Atmel AHB DMA Controller

[Nicolas Ferre <nicolas.ferre@atmel.com>]

Dan Williams :
> On Fri, Nov 14, 2008 at 9:34 AM, Nicolas Ferre <nicolas.ferre@atmel.com> wrote:
>>>>  include/linux/at_hdmac.h                |   26 +
>>> ...this header should be moved somewhere under arch/arm/include.
>> This is where dw_dmac.h resides. Moreover, if one day this IP is implemented
>> on a different architecture, it will be good not to reach it through
>> arch/arm path.
> 
> Ok, I won't gate acceptance on this since dw_dmac already set the
> precedent, but shouldn't the header move after the IP has been
> duplicated?  Just my 2cents.

Ok, I follow your advice.


>>>> +               memset(desc, 0, sizeof(struct at_desc));
>>>> +               dma_async_tx_descriptor_init(&desc->txd, chan);
>>>> +               async_tx_ack(&desc->txd);
>>> the DMA_CTRL_ACK bit is under control of the client.  It should be
>>> read-only to the driver (except for extra descriptors that the driver
>>> creates on behalf of the client).
>> This is precisely where the descriptors are been created so, I thought it
>> should be ok to initialize this bit. Am I right ?
>>
> 
> They will be acknowledged by client code.  Calls like async_memcpy
> assume that the the ack bit is clear by default so they can specify
> some actions to run at completion time.  By setting it early, at
> descriptor allocation time, async_tx will get confused.

This ack bit is annoying me : I cannot figure out how it is used for 
plain memcopy/slave offload calls...

Moreover, at recycle time, if I keep a descriptor chain as a whole, I 
have to introduce another state for my descriptors : consumed but not 
freed yet (with another linked list management).
If I only take care of the ACK flag for releasing descriptors, I loose 
the dependency in my descriptor chain (in a multi-descriptor memcpy case).

Can I only consider this information without taking care of the chaining 
dependency (and loose this information in a multi-descriptor operation) ?
Or, may I drop this DMA_CTRL_ACK bit management as I do not have the 
usefulness of redoing an operations on one descriptor (no xor engine) ?

Kind regards,
-- 
Nicolas Ferre