Jason Thorpe writes:
> ...right, sort of a red-zone.  The trick is the "when"... not necessarily
> on context switch, since those may be "infrequent" (e.g. during
> autoconfiguration), or during some "frequent" event, such as the
> clock interrupt handler (which would just check curproc, if curproc
> wasn't NULL).

Since the point of all this is to find problems that would lead us to
want to expand the size of the kernel stack or fix our use of the
kernel stack in response to problems, I'd say that a red zone checked
by a periodic process would be sufficient. Zero fill it and have
something check every minute for processes with non-zero kernel red
zone pages. It won't tell you where you are failing, but the port
master can then use a trick like CGD's mcount hack to pinpoint the
problem.

Perry