Re: [PATCH v5 03/16] spi: dw: Locally wait for the DMA transactions completion

[Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>]

On Fri, May 29, 2020 at 12:26:10PM +0300, Andy Shevchenko wrote:
> On Fri, May 29, 2020 at 11:12:04AM +0300, Serge Semin wrote:
> > On Fri, May 29, 2020 at 10:55:32AM +0300, Andy Shevchenko wrote:
> > > On Fri, May 29, 2020 at 7:02 AM Serge Semin
> > > <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > Even if DMA transactions are finished it doesn't mean that the SPI
> > > > transfers are also completed. It's specifically concerns the Tx-only
> > > > SPI transfers, since there might be data left in the SPI Tx FIFO after
> > > > the DMA engine notifies that the Tx DMA procedure is done. In order to
> > > > completely fix the problem first the driver has to wait for the DMA
> > > > transaction completion, then for the corresponding SPI operations to be
> > > > finished. In this commit we implement the former part of the solution.
> > > >
> > > > Note we can't just move the SPI operations wait procedure to the DMA
> > > > completion callbacks, since these callbacks might be executed in the
> > > > tasklet context (and they will be in case of the DW DMA). In case of
> > > > slow SPI bus it can cause significant system performance drop.
> > > 
> > 
> > > I read commit message, I read the code. What's going on here since you
> > > repeated xfer_completion (and its wait routine) from SPI core and I'm
> > > wondering what happened to it? Why we are not calling
> > > spi_finalize_current_transfer()?
> > 
> > We discussed that in v4. You complained about using ndelay() for slow SPI bus,
> > which may cause too long atomic context execution. We agreed. Since we can't wait
> > in the tasklet context and using a dedicated kernel thread for waiting would be too
> > much, Me and Mark agreed, that
> 

> > even if it causes us of the local wait-function
> > re-implementation the best approach would be not to use the generic
> > spi_transfer_wait() method, but instead wait for the DMA transactions locally
> > in the DMA driver and just return 0 from the transfer_one callback indicating
> > that the SPI transfer is finished and there is no need for SPI core to wait. As
> > a lot of DMA-based SPI drivers do.
> 
> The above is missed in the commit message.
> 
> > If you don't understand what the commit message says, just say so. I'll
> > reformulate it.
> 
> See above. A bit of elaboration would be good. Thank you!

Agreed. I'll create a more detailed commit description, which will have the
info you cited.

-Sergey

> 
> -- 
> With Best Regards,
> Andy Shevchenko
> 
>