On Tue, Mar 05, 2019 at 01:26:08AM +0000, Chris Packham wrote: > On 5/03/19 1:57 PM, Andrew Lunn wrote: > > On Tue, Mar 05, 2019 at 11:51:52AM +1300, Chris Packham wrote: > >> The orion watchdog can either reset the CPU or generate an interrupt. > >> The interrupt would be useful for debugging as it provides panic() > >> output about the watchdog expiry, however if the interrupt is used the > >> watchdog can't reset the CPU in the event of being stuck in a loop with > >> interrupts disabled or if the CPU is prevented from accessing memory > >> (e.g. an unterminated DMA). > >> > >> All of the orion based CPU cores (at least back as far as Kirkwood) have > >> spare timers that aren't currently used by the Linux kernel. > > Actually this appears to be incorrect Kirkwood does configure timer1 as > a clockevent timer. So I can't just grab timer1 for all platforms. > If you can't use it unconditionally, can you specify it (and use it) as clock ? > >> We can use > >> timer1 to provide a pre-timeout ahead of the watchdog timer and provide > >> the possibility of gathering debug before the reset triggers. > > > > Hi Chris > > > > I had a quick look at other drivers implementing pre-timeout. They > > seem to call watchdog_notify_pretimeout(). I don't see that here? What > > happens when timer1 fires? > > > > It invokes the regular orion_wdt_irq(). On Armada-385 prior to this > change the irq was not specified because the reset always kicked in so > there was no point. > I would suggest to update that function to actually call watchdog_notify_pretimeout() if a pretimeout is configured configured. After all, we do want to support the infrastructure, and that includes support for the various pretimeout governors (if enabled). > For correctness I could make the devicetree binding specify 2 > interrupts. One for the regular watchdog interrupt (which would never > usually get hit because the reset would kick in) and one for the > pretimeout/timer1. > Yes, if they are different interrupts and orion_wdt_irq() is only supposed to handle the real timeout. Thanks, Guenter > >> @@ -169,38 +174,46 @@ static int armadaxp_wdt_clock_init(struct platform_device *pdev, > >> } > >> > >> /* Enable the fixed watchdog clock input */ > >> - atomic_io_modify(dev->reg + TIMER_CTRL, > >> - WDT_AXP_FIXED_ENABLE_BIT, > >> - WDT_AXP_FIXED_ENABLE_BIT); > >> + val = WDT_AXP_FIXED_ENABLE_BIT | TIMER1_FIXED_ENABLE_BIT; > >> + atomic_io_modify(dev->reg + TIMER_CTRL, val, val); > >> > >> dev->clk_rate = clk_get_rate(dev->clk); > >> + > >> + > > > > One blank line is sufficient, > > > > > >> return 0; > >> } > > > > Andrew > > >
