Re: [PATCH v3 1/6] dmaengine: dw-edma: fix dw_edma_probe() can't be call globally

[Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>]

On Thu, Mar 10, 2022 at 11:41:12AM +0300, Serge Semin wrote:
> On Thu, Mar 10, 2022 at 11:52:42AM +0530, Manivannan Sadhasivam wrote:
> > On Wed, Mar 09, 2022 at 10:01:23PM +0300, Serge Semin wrote:
> > 
> > [...]
> > 
> > > > I'm afraid that this will not work for all cases (unless I miss something). As
> > > > Zhi Li pointed out, there are places where only chip pointer will be passed and
> > > > we'd need to extract the private data (dw_edma) from it.
> > > > 
> > > > Tbh I also considered your idea but because of the above mentioned issue and
> > > > also referring to other implementations like gpiochip, I settled with Frank's
> > > > idea of copying the fields.
> > > 
> > > What places are these? I see the only obstacle is the dw_edma_remove()
> > > method. But it's easily fixable.
> > 
> > Yeah, right. I overlooked that part.
> > 
> > > Except that, everything else is more
> > > or less straightforward (just a few methods need to have prototypes
> > > converted to accepting dw_edma instead dw_edma_chip).
> > > 
> > > In order to make the code design more coherent, we need to split up
> > > private data and device/platform info. As I see it dw_edma_chip is
> > > nothing but a chip info data. The eDMA driver is supposed to mainly
> > > use and pass it's private data, not the platform info. It will greatly
> > > improve the code readability and maintainability. Such approach will
> > > also prevent a temptation of adding new private data fields into the
> > > dw_edma_chip structure since reaching the pointer to dw_edma will be
> > > much easier that getting the dw_edma_chip data. In this case
> > > dw_edma_chip will be something like i2c_board_info in i2c.
> > > 
> > > Ideally dw_edma_chip could be a temporarily defined device info, which
> > > memory after the dw_edma_probe() method invocation could be freed. But
> > > in order to implement that we'd need a bit more modifications
> > > introduced.
> > > 
> > 
> 
> > While at it, we should also consider adding an ops structure for passing the
> > callbacks from controller drivers. Currently the eDMA driver has the callbacks
> > defined in v0-core.c but it is used directly instead of as a callback.
> 
> Are you saying about DBI/Native IOs? If so seems reasonable. Though in
> my case it isn't required.) The only problem was a dword-aligned access,
> which has been created in the DW eDMA driver by default.
> 

It is not causing any problem but it doesn't look correct to me.

Btw, do you have a patch for DWORD access? If so, please share. We are also
facing the problem and like to see how to are handling it.

Thanks,
Mani

> -Sergey
> 
> > 
> > This should anyway needs to be fixed when another version of the IP get's added.
> > 
> > Thanks,
> > Mani