Quentin Barnes <qbarnes@urbana.css.mot.com> writes:
[...]
Ah, maybe there is some middle ground here. Instead of putting the
effort into figuring out some portable method for dynamic timeouts,
just change the behavior for a timeout to be user-settable [...]
It can be even easier than that. dejagnu's "timeout" tcl variable is
exactly the default timeout duration in seconds.
The "timeout" variable is an expect feature. It is already set
in stap_run.exp and stap_run2.exp, but timeouts are also manually
specified in various expect statements sprinkled through the
testsuite. Those are the ones that cause me the most headaches.
Otherwise, tinkering with just two files would be trivial.
The .exp files under
testsuite/config or even testsuite/lib could set this global variable
based on the "ishost" predicate - leave it for i686, double it for
s390x, dedicule (!) it for arm. Then we just need to police the test
cases to avoid messing with this value.
It's not that simple. For example, my setup is really, really slow
because it is using NFS mounted root and swap with a small amount of
RAM. Another ARM system could run easily 5x-10x faster than mine
with just more memory or a real hard disk.
Rather than create an "ishost" rule, I suggested that what would
probably be better is to use the MHz or BogoMIPS number from
/proc/cpuinfo. But even that's a heuristic because it only takes
in account the CPU speed, not the system speed that can be choked
due to I/O limitations.
What I'd like to know is if it is really necessary to have fatal
timeouts. How often does running the test suite truly hang up
where the timeout feature gets it unstuck?
I've found that if my system has taken too long, it's due to a bug
and the kernel is no longer stable. However, I don't work on the
stap translator. I suspect bugs in it are what causes recoverable
test hang ups to exist.