tests: avoid test runner hanging on child test process join

In parallel test, the single process is spawning
a bunch of child processes running the tests,
and communicates to them. When the child process
signals that it has finished, the parent calls
child.join(). Sometimes this join never returns.
The result is a lot of defunct python processes,
and the test run just hangs.

I have seen this failure intermittently a fair bit
in a busy containerized environment, and by chance,
consistently reproduced it on a Thinkpad X280
with 8G of RAM and Ubuntu 19.04, which allowed
to diagnose it.

Type: test
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
Change-Id: If0a3110fc2d23e73d77c310d61c3ea90a2b53610
(cherry picked from commit 42693521f6046997133c8f63bcfc9d615d96f69d)

1 files changed, 13 insertions, 1 deletions

diff --git a/test/run_tests.py b/test/run_tests.py
index e6a182c016b..c9c5bdb8a6e 100644
--- a/test/run_tests.py
+++ b/test/run_tests.py
@@ -433,7 +433,17 @@ def run_forked(testcase_suites):
                         results) or stop_run
 
             for finished_testcase in finished_testcase_suites:
-                finished_testcase.child.join()
+                # Somewhat surprisingly, the join below may
+                # timeout, even if client signaled that
+                # it finished - so we note it just in case.
+                join_start = time.time()
+                finished_testcase.child.join(test_finished_join_timeout)
+                join_end = time.time()
+                if join_end - join_start >= test_finished_join_timeout:
+                    finished_testcase.logger.error(
+                        "Timeout joining finished test: %s (pid %d)" %
+                        (finished_testcase.last_test,
+                         finished_testcase.child.pid))
                 finished_testcase.close_pipes()
                 wrapped_testcase_suites.remove(finished_testcase)
                 finished_unread_testcases.add(finished_testcase)
@@ -726,6 +736,8 @@ if __name__ == '__main__':
 
     test_timeout = parse_digit_env("TIMEOUT", 600)  # default = 10 minutes
 
+    test_finished_join_timeout = 15
+
     retries = parse_digit_env("RETRIES", 0)
 
     debug = os.getenv("DEBUG", "n").lower() in ["gdb", "gdbserver"]