On Fri, May 15, 2020 at 11:02:50PM +0300, Serge Semin wrote:
> On Fri, May 15, 2020 at 01:41:31PM +0100, Mark Brown wrote:

> > I guess we could, though it's really there because for historical
> > reasons we've got a bunch of different ways of specifying delays from
> > client drivers rather than for the executing a delay where you've
> > already got a good idea of the length of the delay.

> A beauty of spi_delay_exec() is that it provides a selective delay. I mean it
> checks the delay value and selects an appropriate delay method like ndelay,
> udelay and so on. That's the only reason I'd use it here. But It has got a few
> drawbacks:

Right, usually you'd have a good ideal how long the delay is and
therefore just be able to go directly for an appropraite delay function.

> - timeout value has type u16. It's too small to keep nanoseconds.

That could be increased, though obviously if you have a bigger delay you
can specify it in usecs instead.

> - semantically the xfer argument isn't optional and we can't fetch it that easy
>   in the dmaengine completion callbacks.

Not sure I follow this.

> So if there were an alternative method like _spi_transfer_delay_ns() I'd use it.
> Otherwise we'd need to locally implement the selective delay. Unless you know
> another alternative, which does it. If you don't and there isn't one then in
> order to not over-complicate a simple delay-loop code I'd simply leave the
> ndelay() here.

Not that I'm aware of.