linux/tools/testing/kunit/kunit.py, branch v5.17

kunit: tool: Default --jobs to number of CPUs

2021-12-15T23:44:55Z

The --jobs parameter for kunit_tool currently defaults to 8 CPUs, regardless of the number available. For systems with significantly more (or less), this is not as efficient. Instead, default --jobs to the number of CPUs available to the process: while there are as many superstitions as to exactly what the ideal jobs:CPU ratio is, this seems sufficiently sensible to me. A new helper function to get the default number of jobs is added: get_default_jobs() -- this is used in kunit_tool_test instead of a hardcoded value, or an explicit call to len(os.sched_getaffinity()), so should be more flexible if this needs to change in the future. Signed-off-by: David Gow Signed-off-by: Daniel Latypov Reviewed-by: Daniel Latypov Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: make `build` subcommand also reconfigure if needed

2021-12-15T18:51:16Z

If I created a kunitconfig file that was incomplete, then $ ./tools/testing/kunit/kunit.py build --kunitconfig=my_kunitconfig would silently drop all the options with unmet dependencies! This is because it doesn't do the config check that `kunit.py config` does. So if I want to safely build a kernel for testing, I have to do $ ./tools/testing/kunit/kunit.py config $ ./tools/testing/kunit/kunit.py build It seems unlikely that any user of kunit.py would want the current `build` semantics. So make it effectively do `kunit.py config` + `kunit.py build`. Signed-off-by: Daniel Latypov Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: delete kunit_parser.TestResult type

2021-12-15T18:50:56Z

The `log` field is unused, and the `status` field is accessible via `test.status`. So it's simpler to just return the main `Test` object directly. And since we're no longer returning a namedtuple, which has no type annotations, this hopefully means typecheckers are better equipped to find any errors. Signed-off-by: Daniel Latypov Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: use dataclass instead of collections.namedtuple

2021-12-15T18:50:41Z

namedtuple is a terse way of defining a collection of fields. However, it does not allow us to annotate the type of these fields. It also doesn't let us have any sort of inheritance between types. Since commit df4b0807ca1a ("kunit: tool: Assert the version requirement"), kunit.py has asserted that it's running on python >=3.7. So in that case use a 3.7 feature, dataclasses, to replace these. Changes in detail: * Make KunitExecRequest contain all the fields needed for exec_tests * Use inheritance to dedupe fields * also allows us to e.g. pass a KUnitRequest in as a KUnitParseRequest * this has changed around the order of some fields * Use named arguments when constructing all request objects in kunit.py * This is to prevent accidentally mixing up fields, etc. Signed-off-by: Daniel Latypov Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: suggest using decode_stacktrace.sh on kernel crash

2021-12-13T20:59:19Z

kunit.py isn't very clear that 1) it stashes a copy of the unparsed output in $BUILD_DIR/test.log 2) it sets $BUILD_DIR=.kunit by default So it's trickier than it should be for a user to come up with the right command to do so. Make kunit.py print out a command for this if a) we saw a test case crash b) we only ran one kernel (test.log only contains output from the last) Example suggested command: $ scripts/decode_stacktrace.sh .kunit/vmlinux .kunit < .kunit/test.log | tee .kunit/decoded.log | ./tools/testing/kunit/kunit.py parse Without debug info a user might see something like [14:11:25] Call Trace: [14:11:25] ? kunit_binary_assert_format (:?) [14:11:25] kunit_try_run_case (test.c:?) [14:11:25] ? __kthread_parkme (kthread.c:?) [14:11:25] kunit_generic_run_threadfn_adapter (try-catch.c:?) [14:11:25] ? kunit_generic_run_threadfn_adapter (try-catch.c:?) [14:11:25] kthread (kthread.c:?) [14:11:25] new_thread_handler (:?) [14:11:25] [CRASHED] `tee` is in GNU coreutils, so it seems fine to add that into the pipeline by default, that way users can inspect the otuput in more detail. Note: to turn on debug info, users would need to do something like $ echo -e 'CONFIG_DEBUG_KERNEL=y\nCONFIG_DEBUG_INFO=y' >> .kunit/.kunitconfig $ ./tools/testing/kunit/kunit.py config $ ./tools/testing/kunit/kunit.py build $ This feels too clunky to include in the instructions. With --kconfig_add [1], it would become a bit less painful. [1] https://lore.kernel.org/linux-kselftest/20211106013058.2621799-2-dlatypov@google.com/ Signed-off-by: Daniel Latypov Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: add --kconfig_add to allow easily tweaking kunitconfigs

2021-12-13T20:56:27Z

E.g. run tests but with KASAN $ ./tools/testing/kunit/kunit.py run --arch=x86_64 --kconfig_add=CONFIG_KASAN=y This also works with --kunitconfig $ ./tools/testing/kunit/kunit.py run --arch=x86_64 --kunitconfig=fs/ext4 --kconfig_add=CONFIG_KASAN=y This flag is inspired by TuxMake's --kconfig-add, see https://gitlab.com/Linaro/tuxmake#examples. Our version just uses "_" as the delimiter for consistency with pre-existing flags like --build_dir, --make_options, --kernel_args, etc. Note: this does make it easier to run into a pre-existing edge case: $ ./tools/testing/kunit/kunit.py run --arch=x86_64 --kconfig_add=CONFIG_KASAN=y $ ./tools/testing/kunit/kunit.py run --arch=x86_64 This second invocation ^ still has KASAN enabled! kunit.py won't call olddefconfig if our current .config is already a superset of the provided kunitconfig. Signed-off-by: Daniel Latypov Reviewed-by: David Gow Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: continue past invalid utf-8 output

2021-10-25T19:06:45Z

kunit.py currently crashes and fails to parse kernel output if it's not fully valid utf-8. This can come from memory corruption or just inadvertently printing out binary data as strings. E.g. adding this line into a kunit test pr_info("\x80") will cause this exception UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 1961: invalid start byte We can tell Python how to handle errors, see https://docs.python.org/3/library/codecs.html#error-handlers Unfortunately, it doesn't seem like there's a way to specify this in just one location, so we need to repeat ourselves quite a bit. Specify `errors='backslashreplace'` so we instead: * print out the offending byte as '\x80' * try and continue parsing the output. * as long as the TAP lines themselves are valid, we're fine. Fixed spelling/grammar in commit log: Shuah Khan < Signed-off-by: Daniel Latypov Reviewed-by: Brendan Higgins Tested-by: David Gow Signed-off-by: Shuah Khan

kunit: tool: improve compatibility of kunit_parser with KTAP specification

2021-10-19T20:22:02Z

Update to kunit_parser to improve compatibility with KTAP specification including arbitrarily nested tests. Patch accomplishes three major changes: - Use a general Test object to represent all tests rather than TestCase and TestSuite objects. This allows for easier implementation of arbitrary levels of nested tests and promotes the idea that both test suites and test cases are tests. - Print errors incrementally rather than all at once after the parsing finishes to maximize information given to the user in the case of the parser given invalid input and to increase the helpfulness of the timestamps given during printing. Note that kunit.py parse does not print incrementally yet. However, this fix brings us closer to this feature. - Increase compatibility for different formats of input. Arbitrary levels of nested tests supported. Also, test cases and test suites are now supported to be present on the same level of testing. This patch now implements the draft KTAP specification here: https://lore.kernel.org/linux-kselftest/CA+GJov6tdjvY9x12JsJT14qn6c7NViJxqaJk+r-K1YJzPggFDQ@mail.gmail.com/ We'll update the parser as the spec evolves. This patch adjusts the kunit_tool_test.py file to check for the correct outputs from the new parser and adds a new test to check the parsing for a KTAP result log with correct format for multiple nested subtests (test_is_test_passed-all_passed_nested.log). This patch also alters the kunit_json.py file to allow for arbitrarily nested tests. Signed-off-by: Rae Moar Reviewed-by: Brendan Higgins Signed-off-by: Daniel Latypov Reviewed-by: David Gow Signed-off-by: Shuah Khan

kunit: tool: support running each suite/test separately

2021-10-19T20:21:08Z

The new --run_isolated flag makes the tool boot the kernel once per suite or test, preventing leftover state from one suite to impact the other. This can be useful as a starting point to debugging test hermeticity issues. Note: it takes a lot longer, so people should not use it normally. Consider the following very simplified example: bool disable_something_for_test = false; void function_being_tested() { ... if (disable_something_for_test) return; ... } static void test_before(struct kunit *test) { disable_something_for_test = true; function_being_tested(); /* oops, we forgot to reset it back to false */ } static void test_after(struct kunit *test) { /* oops, now "fixing" test_before can cause test_after to fail! */ function_being_tested(); } Presented like this, the issues are obvious, but it gets a lot more complicated to track down as the amount of test setup and helper functions increases. Another use case is memory corruption. It might not be surfaced as a failure/crash in the test case or suite that caused it. I've noticed in kunit's own unit tests, the 3rd suite after might be the one to finally crash after an out-of-bounds write, for example. Example usage: Per suite: $ ./tools/testing/kunit/kunit.py run --kunitconfig=lib/kunit --run_isolated=suite ... Starting KUnit Kernel (1/7)... ============================================================ ======== [PASSED] kunit_executor_test ======== .... Testing complete. 5 tests run. 0 failed. 0 crashed. 0 skipped. Starting KUnit Kernel (2/7)... ============================================================ ======== [PASSED] kunit-try-catch-test ======== ... Per test: $ ./tools/testing/kunit/kunit.py run --kunitconfig=lib/kunit --run_isolated=test Starting KUnit Kernel (1/23)... ============================================================ ======== [PASSED] kunit_executor_test ======== [PASSED] parse_filter_test ============================================================ Testing complete. 1 tests run. 0 failed. 0 crashed. 0 skipped. Starting KUnit Kernel (2/23)... ============================================================ ======== [PASSED] kunit_executor_test ======== [PASSED] filter_subsuite_test ... It works with filters as well: $ ./tools/testing/kunit/kunit.py run --kunitconfig=lib/kunit --run_isolated=suite example ... Starting KUnit Kernel (1/1)... ============================================================ ======== [PASSED] example ======== ... It also handles test filters, '*.*skip*' runs these 3 tests: kunit_status.kunit_status_mark_skipped_test example.example_skip_test example.example_mark_skipped_test Fixed up merge conflict between: d8c23ead708b ("kunit: tool: better handling of quasi-bool args (--json, --raw_output)") and 6710951ee039 ("kunit: tool: support running each suite/test separately") Reported-by: Stephen Rothwell Shuah Khan Signed-off-by: Daniel Latypov Reviewed-by: David Gow Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan

kunit: tool: actually track how long it took to run tests

2021-10-19T20:18:50Z

This is a long standing bug in kunit tool. Since these files were added, run_kernel() has always yielded lines. That means, the call to run_kernel() returns before the kernel finishes executing tests, potentially before a single line of output is even produced. So code like this time_start = time.time() result = linux.run_kernel(...) time_end = time.time() would only measure the time taken for python to give back the generator object. From a caller's perspective, the only way to know the kernel has exited is for us to consume all the output from the `result` generator object. Alternatively, we could change run_kernel() to try and do its own book keeping and return the total time, but that doesn't seem worth it. This change makes us record `time_end` after we're done parsing all the output (which should mean we've consumed all of it, or errored out). That means we're including in the parsing time as well, but that should be quite small, and it's better than claiming it took 0s to run tests. Let's use this as an example: $ ./tools/testing/kunit/kunit.py run --kunitconfig=lib/kunit example Before: Elapsed time: 7.684s total, 0.001s configuring, 4.692s building, 0.000s running After: Elapsed time: 6.283s total, 0.001s configuring, 3.202s building, 3.079s running Signed-off-by: Daniel Latypov Reviewed-by: David Gow Reviewed-by: Brendan Higgins Signed-off-by: Shuah Khan