tests: disable gRPC fork handlers to fix flaky fork() abort

The suite intermittently fails in CI due to (pexpect spawned)
labgrid-client being killed during teardown by SIGABRT and gRPC C-core
aborting after a fork:

  fork_posix.cc:71] Other threads are currently calling into gRPC, skipping fork() handlers
  ev_epoll1_linux.cc:1125] Check failed: next_worker->state == KICKED

gRPC documents that fork() is only safe if gRPC objects are created in
the child after the fork [1]. The tests do the opposite: the pytest
process spawns the labgrid-client commands via pexpect (using forkpty)
while gRPC is already initialized in it (imported at collection, and used
in-process by the coordinator tests). To partially support this, gRPC
installs a pthread_atfork handler that tries to quiesce its poller across
the fork. When another thread is in the C-core at fork time the handler
loses that race, skips the reset, and the forked child aborts on the
inconsistent poller before it can exec the client.

The problem could never be reproduced locally; it likely occurs more
often in CI because scheduling pressure makes the race more likely to
trigger.

Set GRPC_ENABLE_FORK_SUPPORT=0 to disable the handler. This is the
documented remedy for callers that always exec() after fork() and never
reuse a channel across fork(), which is what the test suite does. Set it
in conftest.py before pytest's collection imports grpc.

[1] https://github.com/grpc/grpc/blob/master/doc/fork_support.md

Signed-off-by: Bastian Krause <bst@pengutronix.de>

Author: Bastian-Krause

tests/conftest.py

@@ -1,4 +1,5 @@
 import logging
+import os
 from signal import SIGTERM
 import sys
 import threading
@@ -13,6 +14,15 @@
 
 psutil = pytest.importorskip("psutil")
 
+# gRPC is only safe to fork() if gRPC objects are created in the child after the fork. The tests do
+# the opposite: the pytest process spawns labgrid-client via pexpect (forkpty) while gRPC is
+# already initialized in it, so gRPC's pthread_atfork handler can lose a race and abort the child
+# before it execs. Set GRPC_ENABLE_FORK_SUPPORT=0 to disable the handler; the spawned clients
+# fork+exec and never reuse an inherited channel, so it is not needed anyway.
+# See also https://github.com/grpc/grpc/blob/master/doc/fork_support.md
+os.environ.setdefault("GRPC_ENABLE_FORK_SUPPORT", "0")
+
+
 @pytest.fixture(scope="session")
 def curses_init():
     """ curses only reads the terminfo DB once on the first import, so make