RE: [PATCH] spi: spi-rspi: Add force_dma variable to spi_ops

[Biju Das <biju.das.jz@xxxxxxxxxxxxxx>]

Hi Mark,

Thanks for the feedback.

> Subject: Re: [PATCH] spi: spi-rspi: Add force_dma variable to spi_ops
> 
> On Sat, Jul 16, 2022 at 04:39:34PM +0100, Biju Das wrote:
> > On RZ/G2L SoCs switching from DMA to interrupt mode, causes timeout
> > issue as we are not getting Rx interrupt even though SPRF bit is set
> > in the status register.
> >
> > But there is no issue if we don't switch between interrupt to DMA mode
> > or vice versa.
> 
> So we need this for correctness until someone can figure out what goes
> wrong :/

Yes, I am in touch with our HW engineers for the missing interrupt when
we switch from DMA to interrupt mode. I will update you once, I receive the
feedback from them.

> 
> > Performance comparison between interrupt and DMA mode on RZ/Five SMARC
> > platform connected to a display module shows that performance and CPU
> > utilization is much better with DMA mode compared to interrupt mode
> > (1->65 fps) and (98->8%).
> 
> How does that look for something doing lots of short transfers (like
> register I/O for example), and what's the throughput like?  DMA probably
> is the most sensible default even so since the impact of doing PIO for
> large transfers tends to be so bad, you have to be doing a *lot* of small
> I/O before it gets to be a problem whereas basically any large transfer
> will notice PIO.

Basically, this performance measurement done by running LVGL on RZ/Five
Connected to display module ili9341 by SPI interface.

Currently we are switch to DMA mode for transfer length greater than 8 bytes. 
Since we have an issue with DMA to interrupt mode switching, initially we tried with PIO mode and fps was very low 1 fps. So we switched to DMA mode and it shoots up the fps to 65.

Similar experiment done on RZ/G2UL connected to PMOD spi flash. In this case

1) The number of interrupts with PIO mode is 6 times greater than that of DMA.

2) The rd/wr test with DMA mode takes 3 minutes to finish whereas with PIO
mode it takes 4 minutes to finish.

> 
> > This patch introduces a variable force_dma to avoid switching between
> > DMA to interrupt mode for RZ platforms.
> 
> Not that it really matters but it's not a variable, it's a flag in the
> device configuration.  
OK.

> I'm wondering if we might want a way to override
> this at runtime in case someone does have an application that suffers
> badly from being forced into DMA (eg, some IIO thing), that could be done
> incrementally though.

OK. Sure. 

Cheers,
Biju