On 12/12/12 2:43 AM, Alan Bateman wrote:
On 11/12/2012 23:53, Stuart Marks wrote:
Please review the following gigantic webrev [1] to clean up @build tags in
RMI tests. Details underlying this change are in the bug report [2].
Briefly, if test classes listed in @build tags are in the wrong order, this
trips over a jtreg problem that in turn causes a cascade of subsequent tests
to fail. It's sensitive to the order in which tests run. The problem
currently occurs in the jdk7u repo. It doesn't happen in jdk8 right now, but
as things shift around it might occur in the future.
The changes looks okay to me. Also just to mention that I've seen this with
jdk8 too, but only with -concurrency where the tests are running concurrently
in different agent VMs to speed up the execution. It's very intermittent and
hopefully the underlying issue in jtreg has be resolved soon.
The problem occurs when a test with the "wrong" class order for @build gets run
first, after which it spoils a bunch of subsequent tests. If a "good" test is
run first the rest of the run is fine. When run without -concurrency the first
test in jdk8 is good, so we don't see the problem. But with -concurrency I can
easily see that a "bad" test runs first -- but only sometimes -- which then
causes the problem.
s'marks
