Re: [avr-libc-dev] Hint For Watchdog Disable on Newer AVRs in the Manual

[Harald Kipp]

Joerg Wunsch wrote:

> > u_char mcusr_mirror __attribute__ ((section (".noinit")));
> >     
> Yes, that has been left out by accident.  Could you please fill in a
> bug report for this, so it won't be forgotten about?
>   
Done.

> What I don't quite understand, why does it not work for you to have
> that code in .init3?  After all, .init2 only contains clearing
> __zero_reg__, and the initialization of the stack pointer, so I don't
> really see why that should get you at risk with the default ~ 15 ms
> watchdog timer.  (Unless you've got loads of code in .init1, of
> course.)
>   

.init2 initializes the stack pointer, which may point into external 
memory. Some memory mapped chips (in our case the RTL8019AS Ethernet 
Controller) may interfere with the memory bus if not properly 
initialized. Resetting and initializing an Ethernet Controller can take 
much more time than 15 ms. Note, that this is an Open Source project and 
the sequence of .init3 code is beyond our control. So we have to use an 
earlier init segment.

Anyway, even our second mega2561 board, which got a different controller 
and doesn't require early initialization, failed with disabling the wdt 
in .init3 but runs fine when disabling it in .init0. Could that be a 
result of the fuses settings? CKSEL and SUT are all unprogrammed, which 
results in a clock start-up time of 16K cycles plus 65 ms. I assumed, 
that the watchdog won't start earlier than the CPU - but who knows?

Harald