Re: [PATCH] dmaengine: qcom_hidma: release the descriptor before the callback

[Vinod Koul]

On Mon, Aug 08, 2016 at 10:45:24AM -0400, Sinan Kaya wrote:
> On 8/8/2016 5:02 AM, Vinod Koul wrote:
> >> What Vinod is telling me that I need to set the cookie to complete 
> >> > whether the transaction is successful or not if the request was accepted
> >> > by HW. xyz_tx_status is just an indication that the transaction was accepted
> >> > by HW. An error can happen as a result of transaction execution.
> > Nope, if the txn is completed you mark it complete. If you can detect error
> > (can you??) then you can report DMA_ERROR.
> > 
> 
> Yes, the HW reports if a transaction failed or not. I have this information
> available in hidma_ll_status function for a limited amount of time until the
> descriptor gets reused.
> 
> > In that latter case do not use dma_async_is_complete() to check. You would
> > need to store and report that cookie 'x' failed which you report status in
> > .tx_statis()
> > 
> 
> I really don't like the idea of telling 'hey client I finished your work and I
> guarantee you it is complete. A month from now, by the way I actually didn't do
> the work that day and I did not tell you'

As i said previously, controllers cannot detect errors. In a system DMA burst
may go bad due to various different issues which controller has not handle
over. So from s system PoV we cannot declare success!

> That's why, I preferred not to call the callback when I observe an error which I
> think it makes more sense. 

That doesnt make sense. A client set a callback, it expect you to call one.
The result quried maybe txn completed or error. Since you have means, please
report..

> Where is the reliability in this? Some random bugs showing at random times. 
> I'd rather not call the callback and be safe. Especially, if you are talking about
> servers; this is plain unacceptable.

How does ignoring client wish caller solve this? You are really ona  wrong
path here. 

> As Lars-Peter and I indicated in my last email, I think we need to kill this 
> tx_status API and replace all the clients to use Dave's interface. It is practically
> impossible to implement a reliable tx_status function.
> 
> Once this transition happens, I can implement Dave's interface not before.
> 
> Again, it will be a different patch than this one. I think v2 of this patch
> needs to go in as it is.
> 
> https://lkml.org/lkml/2016/7/31/64

I havent looked at the patch. If it is not invoking callback set by user,
then I am not taking it. Sorry, we dont choose over client's wish.

Thanks
-- 
~Vinod