On 2020-06-16 23:05, David Gow wrote:
On Wed, Jun 17, 2020 at 11:36 AM Kees Cook keescook@chromium.org wrote:
On Wed, Jun 17, 2020 at 02:30:45AM +0000, Bird, Tim wrote:
Agreed. You only need machine-parsable data if you expect the CI system to do something more with the data than just present it. What that would be, that would be common for all tests (or at least many test), is unclear. Maybe there are patterns in the diagnostic data that could lead to higher-level analysis, or even automated fixes, that don't become apparent if the data is unstructured. But it's hard to know until you have lots of data. I think just getting the other things consistent is a good priority right now.
Yeah. I think the main place for this is performance analysis, but I think that's a separate system entirely. TAP is really strictly yes/no, where as performance analysis a whole other thing. The only other thing I can think of is some kind of feature analysis, but that would be built out of the standard yes/no output. i.e. if I create a test that checks for specific security mitigation features (*cough*LKDTM*cough*), having a dashboard that shows features down one axis and architectures and/or kernel versions on other axes, then I get a pretty picture. But it's still being built out of the yes/no info.
*shrug*
I think diagnostic should be expressly non-machine-oriented.
So from the KUnit side, we sort-of have three kinds of diagnostic lines:
- Lines printed directly from tests (typically using kunit_info() or
similar functions): as I understand it, these are basically the equivalent of what kselftest typically uses diagnostics for -- test-specific, human-readable messages. I don't think we need/want to parse these much.
- Kernel messages during test execution. If we get the results from
scraping the kernel log (which is still the default for KUnit, though there is also a debugfs info), other kernel logs can be interleaved with the results. Sometimes these are irrelevant things happening on another thread, sometimes they're something directly related to the test which we'd like to capture (KASAN errors, for instance). I don't think we want these to be machine oriented, but we may want to be able to filter them out.
This is an important conceptual difference between testing a user space program (which is the environment that TAP initially was created for) and testing kernel code. This difference should be addressed in the KTAP standard. As noted above, a kernel test case may call into other kernel code, where the other kernel code generates messages that get into the test output.
One issue with the kernel issues is that they may be warnings or errors, and to anyone other than the test creator it is probably hard to determine whether the warnings and errors are reporting bugs or whether they are expected results triggered by the test.
I created a solution to report what error(s) were expected for a test, and a tool to validate whether the error(s) occurred or not. This is currently in the devicetree unittests, but the exact implementation should be discussed in the KUnit context, and it should be included in the KTAP spec.
I can describe the current implementation and start a discussion of any issues in this thread or I can start a new thread. Whichever seems appropriate to everyone.
-Frank
- Expectation failures: as Brendan mentioned, KUnit will print some
diagnostic messages for individual assertion/expectation failures, including the expected and actual values. We'd ideally like to be able to identify and parse these, but keeping them human-readable is definitely also a goal.
Now, to be honest, I doubt that the distinction here would be of much use to kselftest, but it could be nice to not go out of our way to make parsing some diagnostic lines possible. That being said, personally I'm all for avoiding the yaml for diagnostic messages stuff and sticking to something simple and line-based, possibly standardising a the format of a few common diagnostic measurements (e.g., assertions/expected values/etc) in a way that's both human-readable and parsable if possible.
I agree that there's a lot of analysis that is possible with just the yes/no data. There's probably some fancy correlation one could do even with unstructured diagnostic logs, so I don't think overstructuring things is a necessity by any means. Where we have different tests doing similar sorts of things, though, consistency in message formatting could help even if things are not explicitly parsed. Ensuring that helper functions that log and the like are spitting things out in the same format is probably a good starting step down that path.
Cheers, -- David