On Thu, May 07, 2015 at 05:45:13PM +0000, Opensource [Steve Twiss] wrote:
> On 06 May 2015 21:07 Guenter Roeck wrote:
>
> > > > > The DA9062 watchdog ping (register CONTROL_F) is "windowed" for
> > > > > protection
> > > > > against spurious writes -- i.e. the ping function cannot be called
> > > > > within a 250ms
> > > > > time limit or the PMIC will reset. This windowing protection also
> > > > > extends to altering
> > > > > the timeout scale in the CONTROL_D register -- in which case if the
> > > > > timeout
> > > > > register is altered and the ping() function is called within the
> > > > > 250ms limit, the
> > > > > PMIC will reset. The delay is there to stop that from happening.
> > > > >
> > > > > I realised my previous patch was over-sanitised: by putting the time
> > > > > delay into the
> > > > > ping() function I was protecting CONTROL_D in stop() and
> > > > > update_timeout_register(),
> > > > > but I was being too over-protective of the ping() function. Therefore
> > > > > if there was an
> > > > > "incorrect trigger signal", the watchdog would not be allowed to fail
> > > > > because the
> > > > > driver would have filtered out the errors.
> > > > >
> > > > Hi Steve,
> > > >
> > > > From your description, it sounds like the protection is only necessary
> > > > if there
> > > > was a previous write to the same register(s).
>
> Hi Guenter,
>
> A clarification from me. It is not the CONTROL_D register that needs
> protecting, but when
> the CONTROL_D register is altered the function call also performs a CONTROL_F
> watchdog
> ping. Too many pings close together would cause the PMIC reset.
>
> > > > If so, it might make sense to record the time of such writes, and only
> > > > add the delay
> > > > if necessary, and only for the remainder of the time.
>
> I've tried it several ways, but my previous suggestion of putting the delays
> in the stop() and
> update_timeout_register() functions just cause even more lengthy delays.
>
> So, I've followed your suggestion and used a variable delay inside the ping()
> function instead:
> this seems to cause a lot less delay. A debug message can be used to notify
> the user if the
> watchdog is trying to be kicked too quickly -- that would be more preferable
> than just shutting
> the PMIC down and still provide a notification that something wasn't quite
> right.
>
> > > > Would this be possible ?
>
> I'll run the tests overnight.
> I'm going to do something like this:
>
> diff --git a/linux-next/v4.0/drivers/watchdog/da9062_wdt.c
> b/linux-next/v4.0/drivers/watchdog/da9062_wdt.c
> index ad80261..d596910 100644
> --- a/gp_sparse/linux-next/v4.0/drivers/watchdog/da9062_wdt.c
> +++ b/gp_sparse/linux-next/v4.0/drivers/watchdog/da9062_wdt.c
>
> @@ -32,12 +33,37 @@ static const unsigned int wdt_timeout[] = { 0, 2, 4, 8,
> 16, 32, 65, 131 };
> #define DA9062_WDT_MIN_TIMEOUT wdt_timeout[DA9062_TWDSCALE_MIN]
> #define DA9062_WDT_MAX_TIMEOUT wdt_timeout[DA9062_TWDSCALE_MAX]
> #define DA9062_WDG_DEFAULT_TIMEOUT wdt_timeout[DA9062_TWDSCALE_MAX-1]
> +#define DA9062_RESET_PROTECTION_MS 300
>
> struct da9062_watchdog {
> struct da9062 *hw;
> struct watchdog_device wdtdev;
> + unsigned long j_time_stamp;
> };
>
> +static void da9062_set_window_start(struct da9062_watchdog *wdt)
> +{
> + wdt->j_time_stamp = jiffies;
> +}
> +
> +static void da9062_apply_window_protection(struct da9062_watchdog *wdt)
> +{
> + unsigned long delay = msecs_to_jiffies(DA9062_RESET_PROTECTION_MS);
> + unsigned long timeout = wdt->j_time_stamp + delay;
> + unsigned long now = jiffies;
> + unsigned int diff_ms;
> +
> + /* if time-limit has not elapsed then wait for remainder */
> + if (time_before(now, timeout)) {
> + diff_ms = jiffies_to_msecs(timeout-now);
> + dev_dbg(wdt->hw->dev,
> + "Delaying watchdog ping by %u msecs\n", diff_ms);
I would not bother about the dev_dbg, but that is your call.
> + mdelay(diff_ms);
Can you use usleep_range() ?
Othewise looks good. BTW, I had to do something similar in
drivers/hwmon/pmbus/zl6100.c; this is where the idea comes from.
Guenter
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
