[RFC PATCH] watchdog: renesas_wdt: don't sleep in atomic context

[RFC PATCH] watchdog: renesas_wdt: don't sleep in atomic context

[Wolfram Sang]

In the restart handler, we hit the following OOPS because rwdt_start()
schedules because of usleep_range():

[   16.071855] BUG: scheduling while atomic: init/1184/0x00000002
...
[   16.159955] [<c05b6b3c>] (schedule) from [<c05b9884>] (schedule_hrtimeout_range_clock+0x108/0x15c)
[   16.168918]  r5:00000003 r4:bd855014
[   16.172498] [<c05b977c>] (schedule_hrtimeout_range_clock) from [<c05b9904>] (schedule_hrtimeout_range+0x2c/0x34)
[   16.182679]  r7:ea0edd48 r6:ea0ec000 r5:0000fa00 r4:00000000
[   16.188343] [<c05b98d8>] (schedule_hrtimeout_range) from [<c05b92dc>] (usleep_range+0x84/0xb8)
[   16.196973] [<c05b9258>] (usleep_range) from [<c049d4a8>] (rwdt_start+0x58/0xa8)

Convert usleep_range() to udelay() to avoid the bug. The downside is
busy looping when the watchdog is started or stopped (64us or 96us in my
case). However, this happens rarely enough, so it seems acceptable.

Fixes: b836005b4f95 ("watchdog: renesas_wdt: Add a few cycles delay")
Signed-off-by: Wolfram Sang <wsa+renesas@sang-engineering.com>
---

We could also add a new flag in the watchdog core "IS_RESTARTING" or
similar and check that in the driver. Depending on the flag, we use
udelay() or usleep_range(). Not sure, if this is feasible or
over-engineered, though. Looking forward for comments.

 drivers/watchdog/renesas_wdt.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/watchdog/renesas_wdt.c b/drivers/watchdog/renesas_wdt.c
index 00662a8e039c..2b0da9b76375 100644
--- a/drivers/watchdog/renesas_wdt.c
+++ b/drivers/watchdog/renesas_wdt.c
@@ -76,8 +76,7 @@ static void rwdt_wait_cycles(struct rwdt_priv *priv, unsigned int cycles)
 	unsigned int delay;
 
 	delay = DIV_ROUND_UP(cycles * 1000000, priv->clk_rate);
-
-	usleep_range(delay, 2 * delay);
+	udelay(delay);
 }
 
 static int rwdt_start(struct watchdog_device *wdev)
-- 
2.28.0

Re: [RFC PATCH] watchdog: renesas_wdt: don't sleep in atomic context

[Wolfram Sang]

[-- Attachment #1: Type: text/plain, Size: 185 bytes --]


>  	delay = DIV_ROUND_UP(cycles * 1000000, priv->clk_rate);
> -
> -	usleep_range(delay, 2 * delay);
> +	udelay(delay);

I should probably add a comment here explaining the situation.


[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [RFC PATCH] watchdog: renesas_wdt: don't sleep in atomic context

[Guenter Roeck]

On Fri, Oct 02, 2020 at 05:09:44PM +0200, Wolfram Sang wrote:
> In the restart handler, we hit the following OOPS because rwdt_start()
> schedules because of usleep_range():
> 
> [   16.071855] BUG: scheduling while atomic: init/1184/0x00000002
> ...
> [   16.159955] [<c05b6b3c>] (schedule) from [<c05b9884>] (schedule_hrtimeout_range_clock+0x108/0x15c)
> [   16.168918]  r5:00000003 r4:bd855014
> [   16.172498] [<c05b977c>] (schedule_hrtimeout_range_clock) from [<c05b9904>] (schedule_hrtimeout_range+0x2c/0x34)
> [   16.182679]  r7:ea0edd48 r6:ea0ec000 r5:0000fa00 r4:00000000
> [   16.188343] [<c05b98d8>] (schedule_hrtimeout_range) from [<c05b92dc>] (usleep_range+0x84/0xb8)
> [   16.196973] [<c05b9258>] (usleep_range) from [<c049d4a8>] (rwdt_start+0x58/0xa8)
> 
> Convert usleep_range() to udelay() to avoid the bug. The downside is
> busy looping when the watchdog is started or stopped (64us or 96us in my
> case). However, this happens rarely enough, so it seems acceptable.
> 
> Fixes: b836005b4f95 ("watchdog: renesas_wdt: Add a few cycles delay")
> Signed-off-by: Wolfram Sang <wsa+renesas@sang-engineering.com>
> ---
> 
> We could also add a new flag in the watchdog core "IS_RESTARTING" or
> similar and check that in the driver. Depending on the flag, we use
> udelay() or usleep_range(). Not sure, if this is feasible or
> over-engineered, though. Looking forward for comments.

Seems to be neither to me. I don't know if there are many drivers who sleep
in the stop function, but even if there are none (today) it would be a
generic problem.

On a side note, I am quite sure there is a system wide flag indicating
that the system is rebooting. Maybe we can just use that ?

> 
>  drivers/watchdog/renesas_wdt.c | 3 +--
>  1 file changed, 1 insertion(+), 2 deletions(-)
> 
> diff --git a/drivers/watchdog/renesas_wdt.c b/drivers/watchdog/renesas_wdt.c
> index 00662a8e039c..2b0da9b76375 100644
> --- a/drivers/watchdog/renesas_wdt.c
> +++ b/drivers/watchdog/renesas_wdt.c
> @@ -76,8 +76,7 @@ static void rwdt_wait_cycles(struct rwdt_priv *priv, unsigned int cycles)
>  	unsigned int delay;
>  
>  	delay = DIV_ROUND_UP(cycles * 1000000, priv->clk_rate);
> -
> -	usleep_range(delay, 2 * delay);
> +	udelay(delay);

I think there should be a comment such as "we can't sleep here because the
function may be called from the restart handler and thus must not sleep".
Either case, this happens so rarely and the delay is so minor that we
should just apply this patch (with comment added) and bother about the more
generic version if the problem is seen again.

Thanks,
Guenter

>  }
>  
>  static int rwdt_start(struct watchdog_device *wdev)
> -- 
> 2.28.0
> 

Re: [RFC PATCH] watchdog: renesas_wdt: don't sleep in atomic context

[Wolfram Sang]

[-- Attachment #1: Type: text/plain, Size: 1343 bytes --]


> > Convert usleep_range() to udelay() to avoid the bug. The downside is
> > busy looping when the watchdog is started or stopped (64us or 96us in my
> > case). However, this happens rarely enough, so it seems acceptable.
> > 
> > Fixes: b836005b4f95 ("watchdog: renesas_wdt: Add a few cycles delay")
> > Signed-off-by: Wolfram Sang <wsa+renesas@sang-engineering.com>
> > ---
> > 
> > We could also add a new flag in the watchdog core "IS_RESTARTING" or
> > similar and check that in the driver. Depending on the flag, we use
> > udelay() or usleep_range(). Not sure, if this is feasible or
> > over-engineered, though. Looking forward for comments.
> 
> Seems to be neither to me. I don't know if there are many drivers who sleep
> in the stop function, but even if there are none (today) it would be a
> generic problem.
> 
> On a side note, I am quite sure there is a system wide flag indicating
> that the system is rebooting. Maybe we can just use that ?

So, it took me a while to get back to this. Yes, we can use system_state
to determine which udelay to use. I am working on a helper for that.
However, as it turned out, this driver needs a different approach
because there were multiple issues with reusing the generic wdt-start
routine. Proper patch will come in some minutes.

Thanks for your input!


[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]