Use polling to detect the end of the rngc self test. This is much simpler
than using an interrupt and a completion.

The selftest should take approx. 450us. Keep the overhead to a minimum
by polling every 500us. (We've already lowered the timeout to 1.5ms.)

Signed-off-by: Martin Kaiser <martin@kaiser.cx>
---
v2:
- use shorter timeout and polling interval

 drivers/char/hw_random/imx-rngc.c | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

diff --git a/drivers/char/hw_random/imx-rngc.c b/drivers/char/hw_random/imx-rngc.c
index 8ff3d46674fd..09523936d2af 100644
--- a/drivers/char/hw_random/imx-rngc.c
+++ b/drivers/char/hw_random/imx-rngc.c
@@ -17,6 +17,7 @@
 #include <linux/hw_random.h>
 #include <linux/completion.h>
 #include <linux/io.h>
+#include <linux/iopoll.h>
 #include <linux/bitfield.h>
 
 #define RNGC_VER_ID			0x0000
@@ -101,22 +102,19 @@ static inline void imx_rngc_irq_unmask(struct imx_rngc *rngc)
 
 static int imx_rngc_self_test(struct imx_rngc *rngc)
 {
-	u32 cmd;
+	u32 cmd, status;
 	int ret;
 
-	imx_rngc_irq_unmask(rngc);
-
 	/* run self test */
 	cmd = readl(rngc->base + RNGC_COMMAND);
 	writel(cmd | RNGC_CMD_SELF_TEST, rngc->base + RNGC_COMMAND);
 
-	ret = wait_for_completion_timeout(&rngc->rng_op_done,
-					  usecs_to_jiffies(RNGC_SELFTEST_TIMEOUT));
-	imx_rngc_irq_mask_clear(rngc);
-	if (!ret)
-		return -ETIMEDOUT;
+	ret = readl_poll_timeout(rngc->base + RNGC_STATUS, status,
+				 status & RNGC_STATUS_ST_DONE, 500, RNGC_SELFTEST_TIMEOUT);
+	if (ret < 0)
+		return ret;
 
-	return rngc->err_reg ? -EIO : 0;
+	return readl(rngc->base + RNGC_ERROR) ? -EIO : 0;
 }
 
 static int imx_rngc_read(struct hwrng *rng, void *data, size_t max, bool wait)
-- 
2.39.2

Hi Martin,

thanks for splitting the series into single patches.

Am Donnerstag, 24. August 2023, 21:20:56 CEST schrieb Martin Kaiser:
> Use polling to detect the end of the rngc self test. This is much simpler
> than using an interrupt and a completion.

I'm still not convinced that using polling is simpler. By using 
readl_poll_timeout() you will also get an interrupt, the timer one. Why 
exactly is using polling much (!) simpler?

> The selftest should take approx. 450us. Keep the overhead to a minimum
> by polling every 500us. (We've already lowered the timeout to 1.5ms.)

I suppose these times only hold true for a specific peripheral clock 
frequency. Is it guaranteed that this frequency is fixed?
For using IRQ it's simpler, there is no guessing: you return once the self 
test finished. The timeout is identical anyway.

Best regards,
Alexander

> Signed-off-by: Martin Kaiser <martin@kaiser.cx>
> ---
> v2:
> - use shorter timeout and polling interval
> 
>  drivers/char/hw_random/imx-rngc.c | 16 +++++++---------
>  1 file changed, 7 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/char/hw_random/imx-rngc.c
> b/drivers/char/hw_random/imx-rngc.c index 8ff3d46674fd..09523936d2af 100644
> --- a/drivers/char/hw_random/imx-rngc.c
> +++ b/drivers/char/hw_random/imx-rngc.c
> @@ -17,6 +17,7 @@
>  #include <linux/hw_random.h>
>  #include <linux/completion.h>
>  #include <linux/io.h>
> +#include <linux/iopoll.h>
>  #include <linux/bitfield.h>
> 
>  #define RNGC_VER_ID			0x0000
> @@ -101,22 +102,19 @@ static inline void imx_rngc_irq_unmask(struct imx_rngc
> *rngc)
> 
>  static int imx_rngc_self_test(struct imx_rngc *rngc)
>  {
> -	u32 cmd;
> +	u32 cmd, status;
>  	int ret;
> 
> -	imx_rngc_irq_unmask(rngc);
> -
>  	/* run self test */
>  	cmd = readl(rngc->base + RNGC_COMMAND);
>  	writel(cmd | RNGC_CMD_SELF_TEST, rngc->base + RNGC_COMMAND);
> 
> -	ret = wait_for_completion_timeout(&rngc->rng_op_done,
> -					  
usecs_to_jiffies(RNGC_SELFTEST_TIMEOUT));
> -	imx_rngc_irq_mask_clear(rngc);
> -	if (!ret)
> -		return -ETIMEDOUT;
> +	ret = readl_poll_timeout(rngc->base + RNGC_STATUS, status,
> +				 status & RNGC_STATUS_ST_DONE, 500, 
RNGC_SELFTEST_TIMEOUT);
> +	if (ret < 0)
> +		return ret;
> 
> -	return rngc->err_reg ? -EIO : 0;
> +	return readl(rngc->base + RNGC_ERROR) ? -EIO : 0;
>  }
> 
>  static int imx_rngc_read(struct hwrng *rng, void *data, size_t max, bool
> wait)


-- 
TQ-Systems GmbH | Mühlstraße 2, Gut Delling | 82229 Seefeld, Germany
Amtsgericht München, HRB 105018
Geschäftsführer: Detlef Schneider, Rüdiger Stahl, Stefan Schneider
http://www.tq-group.com/

Hi Alexander,

Alexander Stein (alexander.stein@ew.tq-group.com) wrote:

> I'm still not convinced that using polling is simpler. By using 
> readl_poll_timeout() you will also get an interrupt, the timer one. Why 
> exactly is using polling much (!) simpler?

it requires much less code in the driver.

> > The selftest should take approx. 450us. Keep the overhead to a minimum
> > by polling every 500us. (We've already lowered the timeout to 1.5ms.)

> I suppose these times only hold true for a specific peripheral clock 
> frequency. Is it guaranteed that this frequency is fixed?

Good point. The lowest possible peripheral clock frequency is half of
what I used for the calculations, i.e. 33.25MHz. That would double the
durations. Should be ok for the selftest. But for the initial seed, we'd
get into a region where readl_poll_timeout (usleep_range) does no longer
make sense.

> For using IRQ it's simpler, there is no guessing: you return once the self 
> test finished. The timeout is identical anyway.

I've looked at other callers of readl_poll_timeout. It seems that none
of them is called in a driver's probe function or uses an overall
timeout of 200ms.

I'll keep the interrupt + completion and resubmit the patches for
adjusting the timeouts.

Thanks,
Martin