[PATCH v2 7/9] spi: spi-s3c64xx: Ensure cur_speed holds actual clock value

[Tomasz Figa]

On Sat, Aug 22, 2020 at 5:18 PM Krzysztof Kozlowski <krzk at kernel.org> wrote:
>
> On Sat, Aug 22, 2020 at 04:52:40PM +0200, Tomasz Figa wrote:
> > On Sat, Aug 22, 2020 at 2:43 PM Krzysztof Kozlowski <krzk at kernel.org> wrote:
> > >
> > > On Fri, Aug 21, 2020 at 06:13:59PM +0200, Łukasz Stelmach wrote:
> > > > cur_speed is used to calculate transfer timeout and needs to be
> > > > set to the actual value of (half) the clock speed for precise
> > > > calculations.
> > >
> > > If you need this only for timeout calculation just divide it in
> > > s3c64xx_wait_for_dma().
> >
> > Division is not the point of the patch. The point is that
> > clk_set_rate() that was originally there doesn't guarantee that the
> > rate is set exactly.
>
> Unfortunately onlt that point of timeout is mentioned in commit msg. If
> the correction of timeout was not the point of the patch, then describe
> the real point...
>
> > The rate directly determines the SPI transfer
> > speed and thus the driver needs to use the rate that was actually set
> > for further calculations.
>
> Yep, makes sense.
>
> >
> > > Otherwise why only if (cmu) case is updated?
> >
> > Right, the !cmu case actually suffers from the same problem. The code
> > divides the parent clock rate with the requested speed to obtain the
> > divider to program into the register. This is subject to integer
> > rounding, so (parent / (parent / speed)) doesn't always equal (speed).
>
> It is not only this problem. The meaning of cur_speed is now changed.
> For !cmu_case this the requested in trasnfer SPI bus clock frequency,
> for cmu_case this is now real src_clk frequency. It's getting slightly
> messier.
>
> >
> > >
> > > You are also affecting here not only timeout but
> > > s3c64xx_enable_datapath() which is not mentioned in commit log. In other
> > > words, this looks wrong.
> >
> > Actually this is right and fixes one more problem, which I didn't spot
> > when looking at this code when I suggested the change (I only spotted
> > the effects on timeout calculation). The rounding error might have
> > caused wrong configuration there, because e.g. 30000000 Hz could be
> > requested and rounded to 28000000 Hz. The former is a threshold for
> > setting the S3C64XX_SPI_CH_HS_EN bit, but the real frequency wouldn't
> > actually require setting it.
>
> Wrong is here describing one thing and doing much more.

Yeah, agreed that the description could be improved. :)

Best regards,
Tomasz