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

From: Lukasz Stelmach
Date: Tue Aug 25 2020 - 11:46:19 EST

It was <2020-08-25 wto 17:11>, when Tomasz Figa wrote:
> On Tue, Aug 25, 2020 at 11:02 AM Lukasz Stelmach <l.stelmach@xxxxxxxxxxx> wrote:
>>
>> It was <2020-08-24 pon 15:21>, when Tomasz Figa wrote:
>> > On Mon, Aug 24, 2020 at 3:17 PM Lukasz Stelmach <l.stelmach@xxxxxxxxxxx> wrote:
>> >>
>> >> It was <2020-08-22 sob 14:43>, when Krzysztof Kozlowski 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().
>> >>
>> >> I divide it here to keep the relationship between the value the variable
>> >> holds and the one that is inside clk_* (See? It's multiplied 3 lines
>> >> above). If you look around every single clk_get_rate() call in the file is
>> >> divided by two.
>> >>
>> >> > Otherwise why only if (cmu) case is updated?
>> >>
>> >> You are righ I will update that too.
>> >>
>> >> However, I wonder if it is even possible that the value read from
>> >> S3C64XX_SPI_CLK_CFG would be different than the one written to it?
>> >>
>> >
>> > It is not possible for the register itself, but please see my other
>> > reply, where I explained the integer rounding error which can happen
>> > when calculating the value to write to the register.
>>
>> I don't have any board to test it and Marek says there is only one that
>> doesn't use cmu *and* has an SPI device attached.
>>
>> Here is what I think should work for the !cmu case.
>>
>> --8<---------------cut here---------------start------------->8---
>> diff --git a/drivers/spi/spi-s3c64xx.c b/drivers/spi/spi-s3c64xx.c
>> index 18b89e53ceda..5ebb1caade4d 100644
>> --- a/drivers/spi/spi-s3c64xx.c
>> +++ b/drivers/spi/spi-s3c64xx.c
>> @@ -655,13 +655,18 @@ static int s3c64xx_spi_config(struct
>> s3c64xx_spi_driver_data *sdd)
>> return ret;
>> sdd->cur_speed = clk_get_rate(sdd->src_clk) / 2;
>> } else {
>> + int src_clk_rate = clk_get_rate(sdd->src_clk);
>
> The return value of clk_get_rate() is unsigned long.
>
>> + int clk_val = (src_clk_rate / sdd->cur_speed / 2 - 1);
>
> Perhaps u32, since this is a value to be written to a 32-bit register.
> Also if you could add a comment explaining that a negative overflow is
> impossible:
>
> /* s3c64xx_spi_setup() ensures that sdd->cur_speed <= src_clk_rate / 2. */

OK.

> But actually, unless my lack of sleep is badly affecting my brain
> processes, the original computation was completely wrong. Let's
> consider the scenario below:
>
> src_clk_rate = 8000000
> sdd->cur_speed = 2500000
>
> clk_val = 8000000 / 2500000 / 2 - 1 = 3 / 2 - 1 = 1 - 1 = 0
> [...]
> sdd->cur_speed = 8000000 / (2 * (0 + 1)) = 8000000 / (2 * 1) = 8000000
> / 2 = 4000000
>
> So a request for 2.5 MHz ends up with 4 MHz, which could actually be
> above the client device or link spec.
>
> I believe the right thing to do would be DIV_ROUND_UP(src_clk_rate /
> 2, sdd->cur_speed) - 1. It's safe to divide src_clk_rate directly,
> because those are normally high rates divisible by two without much
> precision loss.

This makes sense.

>> +
>> /* Configure Clock */
>> val = readl(regs + S3C64XX_SPI_CLK_CFG);
>> val &= ~S3C64XX_SPI_PSR_MASK;
>> - val |= ((clk_get_rate(sdd->src_clk) / sdd->cur_speed / 2 - 1)
>> - & S3C64XX_SPI_PSR_MASK);
>> + val |= (clk_val & S3C64XX_SPI_PSR_MASK);
>> writel(val, regs + S3C64XX_SPI_CLK_CFG);
>>
>> + /* Keep the actual value */
>> + sdd->cur_speed = src_clk_rate / (2 * (clk_val + 1));
>
> Also need to consider S3C64XX_SPI_PSR_MASK here, because clk_val could
> actually be > S3C64XX_SPI_PSR_MASK.

Good point, but this

clk_val = clk_val < 127 ? clk_val : 127;

seems more appropriate since masking may give very bogus value. Eg.

src_clk_rate = 80000000 // 80 MHz
cur_speed = 300000 // 300 kHz

clk_val = 80000000/300000/2-1 => 132
clk_val &= S3C64XX_SPI_PSR_MASK => 4
cur_speed = 80000000 / (2 * (4 + 1)) => 8000000 // 8 MHz


>> +
>> /* Enable Clock */
>> val = readl(regs + S3C64XX_SPI_CLK_CFG);
>> val |= S3C64XX_SPI_ENCLK_ENABLE;
>> --8<---------------cut here---------------end--------------->8---
>>
>>
>> >> > 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.
>> >>
>> >> Indeed, there is a reference too. I've corrected the message.
>> >>
>> >
>> > Thanks!
>> >
>> > Best regards,
>> > Tomasz
>> >
>> >> >>
>> >> >> Cc: Tomasz Figa <tfiga@xxxxxxxxxxxx>
>> >> >> Signed-off-by: Łukasz Stelmach <l.stelmach@xxxxxxxxxxx>
>> >> >> ---
>> >> >> drivers/spi/spi-s3c64xx.c | 1 +
>> >> >> 1 file changed, 1 insertion(+)
>> >> >>
>> >> >> diff --git a/drivers/spi/spi-s3c64xx.c b/drivers/spi/spi-s3c64xx.c
>> >> >> index 02de734b8ab1..89c162efe355 100644
>> >> >> --- a/drivers/spi/spi-s3c64xx.c
>> >> >> +++ b/drivers/spi/spi-s3c64xx.c
>> >> >> @@ -626,6 +626,7 @@ static int s3c64xx_spi_config(struct s3c64xx_spi_driver_data *sdd)
>> >> >> ret = clk_set_rate(sdd->src_clk, sdd->cur_speed * 2);
>> >> >> if (ret)
>> >> >> return ret;
>> >> >> + sdd->cur_speed = clk_get_rate(sdd->src_clk) / 2;
>> >> >> } else {
>> >> >> /* Configure Clock */
>> >> >> val = readl(regs + S3C64XX_SPI_CLK_CFG);
>> >> >> --
>> >> >> 2.26.2
>> >> >>
>> >> >
>> >> >
>> >>
>> >> --
>> >> Łukasz Stelmach
>> >> Samsung R&D Institute Poland
>> >> Samsung Electronics
>> >
>> >
>>
>> --
>> Łukasz Stelmach
>> Samsung R&D Institute Poland
>> Samsung Electronics
>

--
Łukasz Stelmach
Samsung R&D Institute Poland
Samsung Electronics

Attachment: signature.asc
Description: PGP signature