> Watchdog doesn't need to fork (except maybe when you start it) to do its
> job, which is simply to write something to /dev/watchdog on a regular
> interval. It will continue running just fine when the process table is
> full.
It doesn't _need_ to fork, but it does so to check the system status.
Maybe we're confused here. There is 'watchdog' support in the kernel,
with /dev/watchdog, and there is the user space application that uses
this device which was named 'Watchdog' (Whoever did that should be shot)
This user space application checks the system by using fork, allocatin
and freeing some memory, checking the load, etc etc.
> > still be effective? I can imagine it can't check the /dev/watchdog
> > device anymore, which is on a scsi disk.
>
> The only part of /dev/watchdog that is actually on the disk are the device
> numbers in the /dev directory. Once the device is opened by watchdog, it
> shouldn't need the disk any more (although I think it's still marked "in
> use" by reference counting).
Ah nod. I assume the user space program would still _check_ the disk
if or its existence as part of its 'checking the system' routine.
> Software watchdog will only catch a small subset of crashes. A crash must
> leave the kernel intact enough to execute the timer routine that checks if
> the watchdog timer has expired but the watchdog user program has stopped
> running but has left /dev/watchdog open (optional). If you need more than
> that, you'll have to look at a hardware solution.
*nod* I think it would be easier to disable all ncr-53c8xx features,
or to replace the scsi card itself.