-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
Problem
A stack_trap ... EXIT call in a test also runs the EXIT trap of the test script when the test ends. The script trap then runs once for each test that calls stack_trap, and again when the script exits.
In sanity-lnet.sh, the script trap is cleanup_testsuite. If the suite starts with Lustre mounted, restore_mounts is true, and cleanup_testsuite calls setupall in the middle of the suite. The later tests then find Lustre mounted and fail. For this reason, no sanity-lnet test uses stack_trap today.
Environment
- master at 2.17.60
- bash 5.1.8 (EL9) and bash 5.3. Bash 3.2 does not show the problem.
Reproduction
- Mount Lustre.
- Add a sanity-lnet test that calls stack_trap "echo cleanup" EXIT and prints trap -p EXIT.
- Run the test with ONLY.
The test prints:
trap -- 'echo test_30 cleanup; cleanup_testsuite' EXIT
After the test, the log shows "Starting client" from setupall.
Analysis
run_one_logged runs each test in a subshell. The subshell does not run the EXIT trap of the script. But trap -p in the subshell still prints that trap. stack_trap reads the trap -p output and adds the script trap to the test trap.
Other suites with a script-level EXIT trap have the same problem:
- large-lun.sh sets trap cleanupall EXIT. test_2 and test_5 call stack_trap, so each of them runs cleanupall when it ends.
- sanity-ec.sh sets stack_trap "unset LFS_EC_OK" EXIT. Each sanity-ec test that calls stack_trap runs that command again. The effect is harmless.
sanity.sh sets no script-level EXIT trap, so sanity.sh does not show the problem.
Proposed fix
Clear the EXIT trap at the start of run_one with trap - EXIT. Add a sanity-lnet test that checks that a test EXIT trap holds only the test commands.