On Wednesday 03 June 2015 13:53:29 Timur Tabi wrote:

On 06/03/2015 01:25 PM, Guenter Roeck wrote:

quoted

In general the idea here would be to use a crashdump kernel, which,
when loaded, would reset the watchdog before it fires. This kernel
would then write a core dump to a specified location.

What is the mechanism for resetting the watchdog?  The only code that 
knows about the hardware registers is this driver.  Does the crashdump 
kernel call the watchdog stop function?

It might or might not, depending on what you want to achieve. In most
cases, I'd expect the crashdump kernel to have it enabled if we want
to let the user set a pretimeout, but that is a policy question that
is not for the kernel to decide.

quoted

If arm64 doesn't support a crashdump kernel, it might still be possible
to log the backtrace somewhere (eg in nvram using pstore if that is
supported via acpi or efi).

I think it's expected that the firmware support a crash dump mechanism 
of some kind.  But if we're talking about firmware support, then why 
bother with the panic() in the first place?

panic() is what triggers all the crash dump or pstore mechanisms, it
has to do that anyway.

quoted

Is there reason to believe that this all won't work on arm64 ?

No, but I'm still trying to figure out why pre-timeout is valuable.  If 
we don't disable WS1, then we risk having the hardware reset before we 
can take advantage of what the panic() offers.  In which case, what's 
the point of pre-timeout?

The timeouts are both configurable, so the sysadmin has to make sure that
the time between them is long enough to do whatever is necessary.

	Arnd