Driver to support the internal watchdog peripheral found on Atmel's

AT91SAM9 and AT91CAP9 processors (ARM).
Signed-off-by: Renaud Cerrato <r.cerrato@til-technologies.fr>

Signed-off-by: Andrew Victor <linux@maxim.org.za>

I thought I'd finally get around to trying this driver.

No luck.
The proximate cause is that Atmel's second stage "at91boot"

loader disables the watchdog.  That's because the reset state

of this watchdog is "reset after 16 seconds" ... so it must

be tended carefully until the Linux "watchdog" daemon takes

over.  Since there's no code in at91boot, U-Boot, or Linux to

do that, it gets disabled.  (How was this driver tested??)
Needless to say, this isn't exactly my model of how to make

a useful watchdog.  That "write once" rule is pure trouble.
I suggest merging the following patch, to reduce confusion.
- Dave
======== CUT HERE

The at91sam9 watchdog is a kind of annoying bit of hardware:  since its

mode register is write-once, it can't be reconfigured.  Moreover, Atmel's

standard second stage loader "at91boot" always this watchdog (at least

on Atmel's reference boards), ensuring that Linux normally can't use it.
This patch removes some confusion by not registering the watchdog device

on systems where it was disabled before Linux starts.
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
--- a/arch/arm/mach-at91/at91sam9260_devices.c	2008-06-06 15:00:06.000000000 -0700

+++ b/arch/arm/mach-at91/at91sam9260_devices.c	2008-06-06 20:40:38.000000000 -0700

@@ -21,6 +21,7 @@

 #include <asm/arch/at91sam9260.h>

 #include <asm/arch/at91sam9260_matrix.h>

 #include <asm/arch/at91sam9_smc.h>

+#include <asm/arch/at91_wdt.h>
 #include "generic.h"
@@ -666,6 +667,9 @@ static struct platform_device at91sam926
 static void __init at91_add_device_watchdog(void)

 {

+	/* WDT_MR is write-once; if it was disabled, we're stuck */

+	if (at91_sys_read(AT91_WDT_MR) & AT91_WDT_WDDIS)

+		return;

 	platform_device_register(&at91sam9260_wdt_device);

 }

 #else

--- a/arch/arm/mach-at91/at91sam9261_devices.c	2008-06-06 15:00:06.000000000 -0700

+++ b/arch/arm/mach-at91/at91sam9261_devices.c	2008-06-06 20:40:45.000000000 -0700

@@ -25,6 +2...
[ message continues ]

On Sun, 1 Jun 2008 18:40:46 +0200
--

Since the watchdog registers are Write-Once, if the user does not

specify a "wdt_timeout" parameter we'd still like the driver to load

even if it's not enabled.

The user can then always use ioctl(WDIOC_SETTIMEOUT:) to set the timeout later.
Regards,

  Andrew Victor

--

On Wed, 4 Jun 2008 09:31:50 +0200
Surely you can avoid doing the register writes until after you check the

validity of arguments ?

--

The validity of the timeout value is checked in at91_wdt_settimeout().

 It returns -EINVAL if the timeout value is invalid.
The issue above is rather that probe() should not fail if

at91_wdt_settimeout() returns -EINVAL, since we'd like the

driver/module to still load to allow the user to later specify a valid

timeout via ioctl(WDIOC_SETTIMEOUT).
Regards,

  Andrew Victor

--

Every other watchdog behaves as follows
- If you specify a bogus value it doesn't load

- If you specify no value you get a valid default

- If you specify a valid value you get that
I don't believe yours should be different.
Alan

--

I don't think any other in-kernel watchdog driver has to deal with

write-once hardware.  On these processors once the watchdog register

is programmed, it cannot be disabled or re-programmed.

If the above behaviour is required, then we might aswell remove the

ioctl(WDIOC_SETTIMEOUT) interface for this driver since if the user

wants anything other than the default timeout they would need to pass

it via the kernel command-line or module parameters.
Regards,

  Andrew Victor

--

Actually quit a few of them deal with various hardware limits by using a

software timer to maintain the hardware timer poking.
Alan

--

No locking.. so you could get two set timeout calls in parallel. Probably
At the moment those two are safe. When the open lock_kernel

goes away it will be possible to get
			misc_register

						open

						ioctl

			wdt_settimout()
So you may want to swap those two around (and disable the timer if the

register fails ?), or lock the open against the register routine.
Otherwise looks good.
Alan

--