Hi Rafael,
On 12/17/18 10:53 AM, Rafael David Tinoco wrote:
Shuah,
I was recently investigating some errors coming out of our functional tests and we, Dan and I, came up with a discussion that might not be new for you, but, interests us, in defining how to better use kselftests as a regression mechanism/tool in our LKFT (https://lkft.linaro.org).
David / Willem,
I'm only using udpgso as an example for what I'd like to ask Shuah. Feel free to jump in in the discussion if you think its worth.
All,
Regarding: udpgso AND https://bugs.linaro.org/show_bug.cgi?id=3980
udpgso tests are failing in kernels bellow 4.18 because of 2 main reasons:
- udp4_ufo_fragment does not seem to demand the GSO SKB to be > than
the MTU for older kernels (4th test case in udpgso.c).
- setsockopt(...UDP_SEGMENT) support is not present for older kernels.
(commits "udp: generate gso with UDP_SEGMENT" and its fixes seem to be needed).
This case is easy right? Based on the test output below , I can see that the failure is due to
./udpgso: setsockopt udp segment: Protocol not available. setsockopt() is returning an error to clearly indicate that this options isn't supported. This will be a test change to say test is a skip as opposed to fail.
We have a solution for this - test should SKIP as opposed to FAIL.
With that explained, finally the question/discussion:
Shouldn't we enforce a versioning mechanism for tests that are testing recently added features ? I mean, some of the tests inside udpgso selftest are good enough for older kernels...
Right - we do have generic way to handle that by detecting if feature is supported and skip instead of using Kernel version which is going to be hard to maintain.
But, because we have no control over "kernel features" and "supported test cases", we, Linaro, have to end up blacklisting all selftests that have new feature oriented tests, because one or two test cases only.
This has already been solved in other functional tests projects: allowing to check the running kernel version and deciding which test cases to run.
I would like to see effort going into fixing tests to skip when a feature isn't supported. I think that is the solution that will be maintainable in the long run.
Would that be something we should pursue ? (We could try to make patches here and there, like this case, whenever we face this). Or... should we stick with mainline/next only when talking about kselftest and forget about LTS kernels ?
There is a middle of the road solution to run Kselftest from the same kernel release on LTS kernels and report the results as it is turning out be adding overhead in interpreting results when mainline/next Kselftest are run on LTS.
Kselftest mainline/next tends to be in a state where there could be bugs in tests like the one you are finding in the example you used to describe the problem. As we find them we fix them. That is just the nature of mainline/next
Maybe for LTS kernels it is better for you to stay with Kselftest from the same release or close to it. For example, running 4.20 Kselftest on 4.4 is going to result in more skips/(false fails) than running 4.4 Kselftest on 4.4 even though it might provide better coverage. It is a judgment call on the overhead vs. advantage running newer Kselftest from mainline/next on LTS.
I don't think versioning (skip or release based) can fully address the problem you are seeing considering the fluid nature of mainline/next.
thanks, -- Shuah